fn
string | text
string | doi
string | title
string | authors
string | __index_level_0__
int64 |
|---|---|---|---|---|---|
10.48550_arXiv.0709.0346
|
###### Abstract
We analytically derive the lower bound of the total conformational energy of a protein structure by assuming that the total conformational energy is well approximated by the sum of sequence-dependent pairwise contact energies. The condition for the native structure achieving the lower bound leads to the contact energy matrix that is a scalar multiple of the native contact matrix, i.e., the so-called Go potential. We also derive spectral relations between contact matrix and energy matrix, and approximations related to one-dimensional protein structures. Implications for protein structure prediction are discussed.
_Keywords_ protein structure prediction; spectral relations; one-dimensional structures
pacs: 87.15.Cc, 87.15.-v, 87.14.Ee
## I Introduction
Proteins' biological functions are made possible by their precise three-dimensional (3D) structures, and each 3D structure is determined by its amino acid sequence through the laws of thermodynamics. Therefore, predicting protein structures from their amino acid sequences is important not only for inferring proteins' biological functions, but also for understanding how 3D structures are encoded in such one-dimensional information as amino acid sequence. The problem of protein structure prediction is naturally cast as an optimization problem where a potential function is minimized. Given an appropriate potential function, conformational optimization should yield the native structure as the unique global minimum conformation of the potential function. Thus, the problem has been traditionally divided into two sub-problems: One is to establish an appropriate potential function, and the other is to develop the methods to efficiently search the vast conformational space of a protein. Among various forms of effective energy functions, statistical contact potentials have been widely used. In this Letter, we exclusively treat a class of such contact potentials, neglecting other contributions such as electrostatics and local interactions. Accordingly, a protein conformation is represented as a contact matrix in which the \((i,j)\) element is 1 if the residues \(i\) and \(j\) are in contact in space, otherwise it is 0. Although the contact matrix is a coarse-grained representation of protein conformation, it has been known that the contact matrix contains sufficient information to recover the three-dimensional (native) structure of proteins. It is noted that, for the lattice model of proteins, these representations of protein conformation and energy function are exact.
## II Theory
### Lower bound of contact energy
Our fundamental assumption is that the conformational energy of a protein can be somehow expressed in terms of a contact matrix. Now let us assume that the total energy of a protein can be well approximated by the sum of pairwise contact energies between amino acid residues, and that each pairwise contact energy can be decomposed into a sequence-dependent term and a conformation-dependent term. The sequence-dependent term is expressed as a matrix \(\mathcal{E}(S)=(\mathcal{E}_{ij})\) which we call the contact energy matrix, or \(E\)-matrix for short. Each element \(\mathcal{E}_{ij}\) of the \(E\)-matrix represents the energy between the residues \(i\) and \(j\) when they are in contact. This form of the \(E\)-matrix is a very general one: Each element, \(\mathcal{E}_{ij}\), may depend on the entire sequence, \(S\), or it may depend only on the types of the interacting amino acid residues, \(i\) and \(j\), as in the conventional contact potentials. The conformation-dependent term is expressed as another matrix \(\Delta(C)=(\Delta_{ij})\) which we call the contact matrix, or \(C\)-matrix. Each element \(\Delta_{ij}\) of the \(C\)-matrix assumes a value of either 1 or 0, depending on the residues \(i\) and \(j\) are in contact or not, respectively.
\[E(C,S) = \frac{1}{2}\sum_{i=1}^{N}\sum_{j=1}^{N}\mathcal{E}_{ij}(S)\Delta _{ij}(C) \tag{1}\] \[= \frac{1}{2}[\mathcal{E}(S),\Delta(C)] \tag{2}\]
Based on this assumption, we derive the lower bound for the conformational energy and the conditions for the native structure and \(E\)-matrix toachieve the bound.
The Frobenius inner product leads to the matrix \(l_{2}\) norm defined as, for a matrix \(M\), \(\|M\|\equiv[M,M]^{1/2}=(\sum_{i,j}M_{ij}^{2})^{1/2}\).
\[\|\Delta(C)\|^{2}=2N_{c}(C) \tag{3}\]
As for any inner products, the Frobenius inner product satisfies the Cauchy-Schwarz inequality (\(||[A,B]|\leq\|A\|\|B\|\)) from which we have
\[[\mathcal{E},\Delta]\geq-\|\mathcal{E}\|\|\Delta\| \tag{4}\]
where the equality holds if and only if
\[\mathcal{E}=\varepsilon\Delta \tag{5}\]
Although the inequality (Eq. 4) holds for any pair of matrices, we now regard it as the lower bound for conformational energy for a given \(E\)-matrix. For simplicity, we first consider the energy minimization problem for conformations with \(\|\Delta(C)\|\) fixed to the value of the native conformation. It is desirable for the native conformation to satisfy the lower bound and hence its condition Eq.. If the native conformation indeed satisfies the condition Eq., then the elements of the \(E\)-matrix is either \(0\) or \(\varepsilon\) so that only the contacts present in the native conformation are stabilizing. Thus, the native conformation satisfying Eq. is actually a GMEC among any conformations with arbitrary values of \(\|\Delta(C)\|\). An \(E\)-matrix that satisfies Eq. for the native \(C\)-matrix is a kind of the so-called Go potential which has been essential for studying the protein folding problem. At this point, it is still possible that the native structure is not the unique GMEC. For example, if a conformation contains all the native contacts together with some other contacts, this conformation has the same energy as the native conformation. In order for a native conformation to be the unique GMEC, it is required that the total number of contacts of the native conformation is larger than that of any other conformations that contain all the native contacts. From the relation Eq., maximizing the total number of contacts is equivalent to maximizing the norm of the \(C\)-matrix, which in turn implies the minimization of the right-hand side of Eq.. To summarize, for a given \(E\)-matrix, \(\mathcal{E}(S)\), of a protein, its native conformation, \(C_{n}\), achieves the lower bound in Eq. if and only if \(\mathcal{E}(S)=\varepsilon\Delta(C_{n})\) for some \(\varepsilon<0\), and such native structure is the unique GMEC if and only if \(\|\Delta(C_{n})\|\) is the maximum of all possible conformations that contain all the native contacts. Note that the former condition is a relation between \(E\)-matrix and \(C\)-matrix whereas the latter is a condition for a native structure to satisfy. The magnitude of \(\varepsilon\) is not specified here, but it should be determined by other factors such as the folding temperature. It should be noted that a native structure can be the unique GMEC without achieving the lower bound of Eq.. Such a case is made possible either by the limitation of the conformational space imposed by other steric factors such as chain connectivity or excluded volumes, or by inherent inconsistencies of the \(E\)-matrix so that no plausible conformations are allowed to satisfy the lower bounds.
### Spectral relations
To examine more closely how the lower bound can be achieved, we next derive a more generous lower bound in a more restricted case.
\[\Delta=\sum_{\alpha=1}^{N}\sigma_{\alpha}\mathbf{u}_{\alpha}\mathbf{v}_{\alpha }^{T} \tag{6}\]
\(U=(\mathbf{u}_{1},\cdots,\mathbf{u}_{N})\) and \(V=(\mathbf{v}_{1},\cdots,\mathbf{v}_{N})\) are orthogonal matrices. The singular components are sorted in decreasing order of the singular values: \(\sigma_{1}\geq\cdots\geq\sigma_{N}(\geq 0)\). Since \(\Delta\) is real symmetric, the singular values are the absolute values of the eigenvalues of \(\Delta\), and the singular vectors are such that \(\mathbf{u}_{\alpha}=\pm\mathbf{v}_{\alpha}\) where the sign corresponds to that of the respective eigenvalue.
\[\mathcal{E}=\sum_{\alpha=1}^{N}\tau_{\alpha}\mathbf{x}_{\alpha}\mathbf{y}_{ \alpha}^{T} \tag{7}\]
Since \(\mathcal{E}\) is also real symmetric, the singular components have the same properties as the \(C\)-matrix \(\Delta\).
\[(\mathbf{u}_{\alpha}^{T}\mathbf{x}_{\beta})(\mathbf{v}_{\alpha}^{T}\mathbf{y} _{\beta})=-\delta_{\alpha,\beta} \tag{9}\]
We now regard this inequality as a lower bound for the conformational energy for a given \(E\)-matrix. For a fixed set of the singular values \(\sigma_{\alpha}\) (\(\alpha=1,\cdots,N\)), if and only if there exists such a conformation that satisfies the condition in Eq., then that conformation is the lowest possible energy conformation. Let \(\lambda_{\alpha}\) and \(\varepsilon_{\alpha}\) (\(\alpha=1,\cdots,N\)) be the eigenvalues of the \(C\)-matrix and \(E\)-matrix, respectively, sorted in the decreasing order of their absolute values. Then \(\sigma_{\alpha}=|\lambda_{\alpha}|\) and \(\tau_{\alpha}=|\varepsilon_{\alpha}|\) for \(\alpha=1,\cdots,N\), and \({\bf u}_{\alpha}\) and \({\bf x}_{\alpha}\) are the eigenvectors of the corresponding matrices. Thus, in terms of eigenvalues and eigenvectors, the lower bound in Eq. is equal to \(\sum_{\alpha}\lambda_{\alpha}\varepsilon_{\alpha}\) with \(\lambda_{\alpha}\varepsilon_{\alpha}\leq 0\) for \(\alpha=1,\cdots,N\). In addition to the condition Eq. for the lower bound of Eq., if \(\Delta\) and \({\cal E}\) are of the same rank, then the numbers of positive, negative, and zero eigenvalues of \(\Delta\) and \(-{\cal E}\) are the same and \({\bf u}_{\alpha}=\pm{\bf x}_{\alpha}\).
\[{\cal E}=-S\Delta S^{T}, \tag{10}\]
If the conformation that satisfy the condition Eq. is the native structure, the \(E\)-matrix is consistent in the sense that the contributions from all the eigencomponents are stabilizing the native structure (\(\lambda_{\alpha}\varepsilon_{\alpha}\leq 0\)). Since the matrix \(S\) is non-singular, we can "predict" the native structure from the \(E\) matrix as \(\Delta=-S^{-1}{\cal E}S^{-T}\) (if we can construct the appropriate matrix \(S\)). At this point, however, the native structure may not be the GMEC since other conformations with a different set of singular values may have lower energies.
In order to compare the energies of conformations with different sets of singular values, we use another inequality:
\[-\sum_{\alpha=1}^{N}\sigma_{\alpha}\tau_{\alpha}\geq-\|{\cal E}\|\|\Delta\| \tag{11}\]
We note that, in terms of singular values, the matrix norms are expressed as \(\|\Delta\|=(\sum_{\alpha}\sigma_{\alpha}^{2})^{1/2}\) and \(\|{\cal E}\|=(\sum_{\alpha}\tau_{\alpha}^{2})^{1/2}\). Hence, it is clear that the equality in Eq. holds if and only if, in addition to the condition in Eq., there exists a scalar constant \(c\) such that \(\tau_{\alpha}=c\sigma_{\alpha}\) for all \(\alpha=1,\cdots,N\). These conditions are equivalent to Eq..
### One-dimensional approximations
To connect the present results with previous studies, we next introduce two approximations. First, we consider the case where the \(E\)-matrix is well approximated by its principal eigencomponent, that is, \({\cal E}\approx\varepsilon_{1}{\bf x}_{1}{\bf x}_{1}^{T}\). This approximation is motivated by the eigenvalue analysis of the Miyazawa-Jernigan (MJ) contact potential performed by Li et al., and has been employed by others. In this case, the lower bound Eq. is achieved if and only if \({\bf x}_{1}=\pm{\bf u}_{1}\) and \(\varepsilon_{1}\lambda_{1}<0\). This result was previously derived by Cao et al. who subsequently showed that the vector \({\bf x}_{1}\) constructed by using the components of the principal eigenvector of the MJ contact potential is indeed highly correlated with the principal eigenvector of the native contact matrices. Bastolla et al. obtained a similar result, but they also showed that taking the average of such \({\bf x}_{1}\) over evolutionarily related proteins greatly improved the correlation. Since the rank of the contact matrix is in general not 1, Eq. does not hold and the equality in Eq. cannot be satisfied. Consequently, there are attractive interactions between non-native contacts even when \({\bf x}_{1}={\bf u}_{1}\) holds exactly. Nevertheless, Porto et al. have demonstrated that the knowledge of \({\bf u}_{1}\) alone is practically sufficient for reconstructing the native contact matrix of small single-domain proteins. Therefore, construction of effective rank-1 \(E\)-matrices is of great interest. Based on the Porto et al.'s result, it is tempting to postulate that the satisfaction of the lower bound by a rank-1 \(E\)-matrix is sufficient for the native conformation to be the unique GMEC. At present, however, there is no clear connection between the present formulation (energy minimization) and the Porto et al.'s combinatorial algorithm.
Another approximation is a kind of mean-field approximations in which the matrix element \({\cal E}_{ij}\) is replaced by its average over column \(\langle{\cal E}_{i\bullet}\rangle\equiv\sum_{j=1}^{N}{\cal E}_{ij}/N\). Let us define \({\bf e}=(\langle{\cal E}_{1\bullet}\rangle,\cdots,\langle{\cal E}_{N\bullet }\rangle)^{T}\) and \({\bf n}=(n_{1},\cdots,n_{N})^{T}\) where \(n_{i}\equiv\sum_{j=1}^{N}\Delta_{ij}\) is the contact number of the \(i\)-th residue. Then, we have the following approximation and the lower bound:
\[E(C,S) \approx \frac{1}{2}{\bf e}^{T}{\bf n} \tag{12}\] \[\geq -\frac{1}{2}\|{\bf e}\|\|{\bf n}\| \tag{13}\]
This lower bound condition is analogous to Eq., and can be regarded as another kind of the Go potential for one-dimensional protein structure. It has been suggested that contact number vector can significantly constrain the conformational space. Together with other one-dimensional structures, contact number vector is also used for recovering the native structures, and can be accurately predicted. It has been pointed out that the contact number vector is highly correlated with the principal eigenvector of the \(C\)-matrix, which suggests that this mean-field approximation is qualitatively similar to the principal eigenvector approximation introduced above.
## Discussion
Using a more restricted, but conventional, form of the \(E\)-matrix where each element \({\cal E}_{ij}\) depends only on the types of \(i\)-th and \(j\)-th residues (e.g., the MJ potential),Vendruscolo et al. have shown that it is impossible for such \(E\)-matrices to stabilize all the native structures in a database. The conventional \(E\)-matrices such as those they studied do not take into account the sequence-dependence beyond a summation of the contributions from residue pairs. In the present study, we assumed a more general form for the \(E\)-matrix, allowing each element \(\mathcal{E}_{ij}\) to depend on the whole amino acid sequence. In practical situations of protein structure prediction, we want to optimize an energy function so that the native conformations of arbitrary proteins achieve the lower bound. Now let us impose this as a requisite for the \(E\)-matrix. Then, there should exist a function, namely \(\mathcal{E}\), that maps each amino acid sequence to the corresponding optimal \(E\)-matrix, that is, the Go potential. Thus, the problem of structure prediction becomes a trivial matter. Currently, most efforts for developing energy functions seem to be focused on accurate estimation of a fixed set of parameters for a given functional form. The present analysis suggests that inferring the function \(\mathcal{E}\) that can generate the Go-like \(E\)-matrices from amino acid sequences is essential if a contact potential is used. The lower bound inequality (Eq. 4) and its condition for the equality (Eq. 5) will serve as the guiding principle for inferring such a function. This approach to structure prediction is apparently similar to machine-learning approaches to contact matrix prediction. Although conventional machine-learning methods are not directly targeted at the optimization of the form of Eq., their prediction accuracy should be indicative of the possibility for identifying the function \(\mathcal{E}\).
In the preceding paragraph, we have assumed the existence of the function \(\mathcal{E}\) to construct the optimal contact potential from a given amino acid sequence. What if, however, there is no such function? In fact, the limited success of current contact matrix prediction strongly suggests that this is more likely the case. Such a case implies either that there are proteins for which the lower bound energy cannot be achieved, or that the total energy cannot be sufficiently accurately approximated by Eq.. The former case indicates that some proteins are inherently frustrated, but to a good approximation such proteins should be rather exceptional for natural proteins. The latter case may indicate that multi-body contact interactions and/or other energy components than contact energies are more important.
In summary, we have shown that the requirement for the native structure to achieve the lower bound naturally leads to the Go potential and the requirement for such a conformation to be the unique GMEC leads to the native conformation being the most compact one among those containing all the native contacts. These results suggest that protein structure prediction should be possible simply by constructing the optimal energy matrices or that the contact potential alone is not suitable for the problem. Although not yet definitive, the current state of contact prediction as well as recent studies on local interactions suggest that the latter may be the case. Nevertheless, the present results may be useful for evaluating the optimality of potential functions in either case.
|
10.48550/arXiv.0709.0346
|
On the optimal contact potential of proteins
|
Akira R. Kinjo, Sanzo Miyazawa
| 2,531
|
10.48550_arXiv.0906.5274
|
###### Abstract
We report a numerical simulation of the rate of crystal nucleation of sodium chloride from its melt at moderate supercooling. In this regime nucleation is too slow to be studied with "brute-force" Molecular Dynamics simulations. The melting temperature of ("Tosi-Fumi") NaCl is \(\sim 1060\)K. We studied crystal nucleation at \(T\)=800K and 825K. We observe that the critical nucleus formed during the nucleation process has the crystal structure of bulk NaCl. Interestingly, the critical nucleus is clearly faceted: the nuclei have a cubical shape. We have computed the crystal-nucleation rate using two completely different approaches, one based on an estimate of the rate of diffusive crossing of the nucleation barrier, the other based on the Forward Flux Sampling and Transition Interface Sampling (FFS-TIS) methods. We find that the two methods yield the same result to within an order of magnitude. However, when we compare the extrapolated simulation data with the only available experimental results for NaCl nucleation, we observe a discrepancy of nearly 5 orders of magnitude. We discuss the possible causes for this discrepancy.
## 1 Introduction
Crystallization of salts is a phenomenon of great practical relevance. In fact, it is one of the most important industrial separation processes. But it also plays a crucial role in geological processes that occur on an altogether different time scale. The crystallization process consists of two steps: nucleation and growth. If nucleation is slow compared to the time it takes a crystal to grow to a size comparable to the size of the container, large single crystals will form (an example is rock salt). When nucleation is fast, the resulting solid will form as a fine powder. It is clearly important to be able to predict the rate of nucleation of salts and - at a later stage - to understand the factors that influence nucleation. In the present paper we aim to demonstrate that, with current simulation techniques and currently available force-fields, it is indeed possible to compute the rate of nucleation of a real salt crystal (in the present case NaCl from its melt) This opens the way to "ab-initio" predictions of nucleation rates of many ionic substances.
Solutions or melts can often be cooled well below their freezing temperature. The reason is that the formation of small nuclei of the stable crystal phase is an activated process that may be extremely slow. Intuitively, it is easy to understand why crystal nucleation is an activated process, i.e. why there is a free-energy barrier separating the metastable parent phase (the liquid) from the stable crystal phase. The point is that, initially, the formation of small crystalline nuclei costs free energy. But once the crystal nucleus exceeds a critical size, its free energy decreases as it grows. The rate at which crystal nuclei form depends strongly on \(\Delta G_{crit}\), the free-energy required to form a critical nucleus. Classical Nucleation Theory (CNT) is commonly used to estimate the height of the nucleation barrier and to predict the rate of crystal nucleation. According to CNT, the total free energy of a crystallite that forms in a supersaturated solution or melt contains two terms: the first is a "bulk" term that expresses the fact that the solid is more stable than the supersaturated fluid - this term is negative and proportional to the volume of the crystallite. The second is a "surface" term that takes into account the free-energy cost of creating a solid-liquid interface. This term is positive and proportional to the surface area of the crystallite.
\[\Delta G=\frac{4\pi}{3}R^{3}\rho_{S}\Delta\mu+4\pi R^{2}\gamma, \tag{1}\]
The function \(\Delta G\) goes through a maximum at \(R=2\gamma/(\rho_{S}|\Delta\mu|)\) and the height of the nucleation barrier is
\[\Delta G_{crit}=\frac{16\pi}{3}\gamma^{3}/(\rho_{S}|\Delta\mu|)^{2}. \tag{2}\]
The crystal-nucleation rate per unit volume, \(I\), depends strongly on \(\Delta G_{crit}\):
\[I=\kappa\exp(-\Delta G_{crit}/k_{B}T). \tag{3}\]
Here \(\kappa\) is a kinetic prefactor, \(T\) is the absolute temperature and \(k_{B}\) is Boltzmann's constant.
\[I=\kappa\,\exp\left[-\frac{16\pi}{3}\gamma^{3}/(\rho_{S}|\Delta\mu|)^{2}\right]. \tag{4}\]
Under experimental conditions, nucleation is infrequent on the time scale of typical molecular processes. Yet, when it happens, it proceeds rapidly. This makes it difficult to study the structure and dynamics of crystal nuclei of atoms or small molecules in experiments. In the case of NaCl, the experiments are also complicated by the fact that crystallization occurs at high temperatures. This may explain why there is a scarcity of experimental data on the nucleation of NaCl. To our knowledge, the only data are those of Buckle and Ubbelhode from the 1960's. In these experiments, crystallization in NaCl micro-droplets was observed visually. As the droplet size ((\(\mathcal{O}(3\mu)\)) and time window for the measurement (1-30 seconds - after which the droplets sedimented out of view) - were fixed, the nucleation rate could be determined at one temperature only (905 K for NaCl). At this temperature, the nucleation rate was such that, on average, one nucleus would form during the observation time (\({\cal O}(10s)\)) in a droplet with a volume of order \(10^{-17}m^{3}\). Hence the experimental nucleation rate per unit volume was \(O(10^{16})\)\(m^{-3}s^{-1}\).
For experimental nucleation rates of this order of magnitude, brute-force MD simulations are out of the question. The average time it would take for nuclei to form spontaneously in a system consisting of several thousands of particles is of order of \(10^{20}\) seconds. Clearly, this is beyond the scope of MD simulations. The standard solution to circumvent this problem is to perform simulations at much larger undercooling than used in the experiments. Huang et al. performed MD simulations of melting and freezing of a droplet composed of 216 NaCl ions in vacuum: to this end, they performed temperature quenches down to 550K (i.e. approximately half the melting temperature) and found nucleation rates of the order of \(O(10^{36})m^{-3}s^{-1}\), which is 20 orders of magnitude higher than the experimental rate at \(905K\).
Another effort to study nucleation at less severe supercooling was made by Koishi et al.. These authors performed an MD simulation of 125000 ions system in vacuum, at temperatures of 740K (i.e. approximately 0.7 the melting temperature \(T_{m}\)) and 640K (i.e. approximately 0.6 the melting temperature \(T_{m}\)). Both free and periodic boundary conditions were used. The estimated nucleation rate at \(740K\) was \(O(10^{35})m^{-3}s^{-1}\), which is virtually the same value as found by Huang et al. at a much larger supercooling. This is surprising because nucleation rates tend to depend very strongly on temperature. This suggests that, at least at the lowest temperatures, the barrier for crystal nucleation is negligible. More in general, crystal nucleation under extreme supercooling need not proceed following the same path as under moderate supercooling.
In what follows, we use the technique of refs. based on a combination of umbrella sampling (to determine the barrier height) and a dynamical simulation (to determine the crossing rate). The computing time required for this scheme does not scale exponentially with the nucleation barrier, but it does increase with increasing nucleus size.
We also compute the nucleation rate using an algorithm based on the "forward flux sampling" (FFS) and the "Transition Path Samplig" (TIS) techniques and compare it with the one obtained using the method previously mentioned.
In the present work we study homogeneous crystal nucleation in the Tosi-Fumi NaCl model for NaCl at two different temperatures, viz. \(T_{1}\)=800K and \(T_{2}\)=825K, corresponding to 25% and 22% undercooling. For this system, we computed the nucleation barrier, examined the structure and shape of the critical nucleus and computed the nucleation rate.
## 2 Methods
The Tosi-Fumi rigid-ion interaction potential for NaCl is of the following form,
\[U_{ij}(r)=A_{ij}e^{[B(\sigma_{ij}-r)]}-\frac{C_{ij}}{r^{6}}-\frac{D_{ij}}{r^{8}}+ \frac{q_{i}q_{j}}{r}\;, \tag{5}\]
This pair potential is written as the sum of a Born-Mayer repulsion, two attractive van der Waals contributions and a Coulomb interaction term. In our simulations, we calculated the Coulomb interactions using the Ewald summations method with a real space cutoff of 10 \(\dot{A}\) and a real space damping parameter of 0.25 \(\dot{A}^{-1}\). We truncated the Van Der Waals part of the potential at 9 \(\dot{A}\), assuming the \(g(r)=1\) beyond this cutoff.
The computed number density of ions in the bulk solid at 800K and 825K at \(10^{5}\) Pa was 0.041\(\dot{A}^{-3}\), in agreement with experiment. The density of ions in the supercooled liquid at the same temperature and pressure was 0.034\(\dot{A}^{-3}\).
We prepared under cubic boundary conditions a supercooled system of\({}^{3}\) NaCl ion pairs at ambient pressure by cooling it down below the melting temperature. For the present model, Anwar et al. have computed the melting temperature: \(T_{m}\)=(1064 \(\pm\) 14)K, which is very close to the experimental melting temperature (\(T_{m}^{exp}\)=1072K). Using constant-pressure Monte Carlo simulations, we cooled the system down to the temperatures where we studied nucleation: \(T_{1}\)=800K and \(T_{2}\)=825K, corresponding to 25% and 22% undercooling. Note that the experiments on NaCl nucleation where performed at a somewhat higher temperature (16% supercooling). The reason why we could not perform simulations at these higher temperatures is that the critical nucleus would be about twice the size of the nucleus that could be studied without spurious finite-size effects for the system sizes that we employed. At temperatures below 750K, spontaneous nucleation occurred during the simulations. We therefore kept the temperature above this lower limit.
Nucleation is an activated process. In steady-state, the nucleation rate per unit volume and time is given by eq. where \(\exp(-\beta\Delta G_{crit})\) is the equilibrium probability per nucleus to find a critical nucleus in the metastable parent phase. \(\kappa\) is a kinetic prefactor. In the case of a diffusive barrier crossing, \(\kappa\) can be expressed as:
\[\kappa=\sqrt{\frac{|\Delta\mu|}{6\pi k_{B}Tn_{crit}}}\;\rho_{liq}f^{+}_{n^{crit}} \tag{6}\]
\begin{table}
\begin{tabular}{||c|c|c|c|c|c||} \hline & \(A_{ij}\) & \(B\) & \(C_{ij}\) & \(D_{ij}\) & \(\sigma_{ij}\) \\ & \([kJ/mol]\) & \([\dot{A}^{-1}]\) & \([\dot{A}^{6}kJ/mol]\) & \([\dot{A}^{8}kJ/mol]\) & \([\dot{A}]\) \\ \hline Na-Na & 25.4435 & 3.1546 & 101.1719 & 48.1771 & 2.340 \\ \hline Na-Cl & 20.3548 & 3.1546 & 674.4793 & 837.0770 & 2.755 \\ \hline Cl-Cl & 15.2661 & 3.1546 & 6985.6786 & 14031.5785 & 3.170 \\ \hline \end{tabular}
\end{table}
Table 1: Potential parameters for NaCl.
It is important to distinguish between \(f^{+}_{ncrit}\), which is the rate at which particles are added to a nucleus with the critical size, and the net flux across the nucleation barrier. In steady state, this net flux is equal the number of nuclei that go from \(ncrit\) to \(ncrit+1\) minus the number that go from \(ncrit+1\) to \(ncrit\). Hence, the actual nucleation rate is a combination of the forward rate \(f^{+}\) and the backward rate \(f^{-}\). However, because of detailed balance, knowledge of \(f^{+}_{ncrit}\) (combined with knowledge of the barrier height and shape) is enough to compute the nucleation rate.
Assuming a diffusive attachment or detachment of single particles from the critical nucleus, the forward rate \(f^{+}_{crit}\) at the top of the barrier can be related to the spontaneous fluctuations in the number of particles in a nucleus at the top of the nucleation barrier:
\[f^{+}_{crit}=\frac{1}{2}\frac{\langle\Delta n^{2}_{crit}(t)\rangle}{t} \tag{7}\]
To estimate \(f^{+}_{crit}\) a series of dynamical trajectories were necessary: after generating a set of uncorrelated configurations at the top of the barrier, we carried out NVT MD simulations using the \(DL\_POLY\) package with a timestep of 0.5 fs. We computed the nucleation rate using equation.
Moreover, we computed the nucleation rate per unit time and volume using an algorithm based on path-sampling techniques and compared the results with those obtained using equation.
## 3 Results
We computed the free energy barrier for crystal nucleation at \(T_{1}\)=800K and \(T_{2}\)=825K, corresponding to \(\beta\Delta\mu_{1}\)=0.54 and \(\beta\Delta\mu_{2}\)=0.48. The values of \(\beta\Delta\mu\) were estimated numerically by thermodynamic integration from the coexistence temperature and free energies reported by Anwar et al.. Fig.1 shows the computed nucleation barriers as a function of \(n\).
As expected, \(\Delta G\) decreases with supersaturation. Around T\(<\)750K the barrier gets sufficiently low that spontaneous nucleation can take place on the time scale of a simulation. The size of the critical nucleus, \(N_{c}\), was estimated according to a fit of the functional form of the CNT, and we found \(N_{c}\approx 120\) ions at \(T_{1}\) and \(N_{c}\approx 150\) ions at \(T_{2}\). Koishi et al. estimated \(N_{c}=120-130\) ions at 640K and 740K, which is surprising in view of the CNT prediction that the size of the critical nucleus scales as \((\gamma/|\Delta\mu|)^{3}\). If we make the usual assumption that \(\gamma\sim\Delta h\) and \(\Delta\mu\approx(\Delta h/T_{m})(T-T_{m})\), where \(\Delta h\) is the entropy of fusion per ion pair, then we would expect that \(N_{c}\sim(1-T/T_{m})^{-3}\) and we would predict that the critical nucleus at 640K should contain a quarter as many particles as those at 800K.
Next, we consider the structure and shape of the critical nucleus. Fig.2 shows a snapshot of the critical nucleus at \(T_{2}\)=825K. Note that the crystal presents rudimentary low-index facets. In experiments the existence of such facets was postulated, as they may act as sites for subsequent heterogeneous nucleation.
As can be seen from Fig.2, the critical nucleus already shows the charge-ordered rock-salt structure of the bulk phase.
Snapshot of the critical nucleus at \(T_{2}\)=825K: the bulk NaCl structure is already evident. The critical crystal nucleus seems to have rudimentary facets, which is in agreement with the interpretation of the experiments in ref..
Free energy barriers \(\Delta G\) as a function of the nucleus size \(n\) for \(T_{1}\)=800K(\(\beta\Delta\mu_{1}=0.54\)) and \(T_{2}\)=825K(\(\beta\Delta\mu_{2}=0.48\)). Error-bars on \(\beta\Delta G\) are of the order of 1 \(k_{B}T\). The dashed curves are fits to the functional form given by CNT.
Fig.2 also shows that, in the temperature range that we studied, the critical nucleus is non-spherical. In order to quantify the degree of non-sphericity of the critical nucleus, we expanded its density with respect to the center of mass in rank-four-spherical-harmonics and constructed the quadratic invariant \(S_{4}\): we obtained \(S_{4}(T_{1})\)=0.115 and \(S_{4}(T_{2})\)=0.110. For a simple cube \(S_{4}\)=0.172 and for a sphere \(S_{4}\)=0, therefore the shape of the critical nucleus is closer to a cube than to a sphere. In other words, the critical nucleus already exhibits the morphology of macroscopic NaCl crystals.
In order to make sure that there were not finite size effects that resulted in interactions between image cluster, we visually checked that the critical nuclei didn't show a preferred spacial orientation and that the minimum distance between them was bigger than half box.
Moreover we calculated the Debye-Huckle screening lenght and found that it was smaller than 1 \(\dot{A}\), showing that the critical nuclei did not even electrostatically interact. We could therefore conclude that there was no induced nucleation due to the interaction between a critical cluster and its own periodic image.
Using the computed height of the nucleation barrier and the values of \(\Delta\mu\) as input, we can estimate the surface free-energy density \(\gamma_{ls}\). To this end, we make use of the CNT expression for the barrier height (eqn. 2). However, this expression assumes that the critical nucleus is spherical. It is easy to derive the corresponding expression for a cubical nucleus. The results for both estimates of \(\gamma_{ls}\) are given in Table 2.
There exist experimental estimates of \(\gamma_{ls}\) at 905K. These estimates are based on a somewhat questionable CNT expression for the nucleation rate. Moreover, in ref. it is assumed that the critical nucleus is spherical. The experimental estimate of \(\gamma_{ls}\) (\(\gamma_{exp}\)=84.1[\(erg\)\(cm^{-2}\)]) is therefore not based on a direct determination. Nevertheless, in the absence of other experimental data, this is the only number we can compare to.
As the table shows, there is a fair agreement between simulation and experiment. The experimental estimate for \(\gamma\) is based on the assumption that the critical nucleus is spherical. If it is cubic, one would obtain the number in the lower right-hand corner. In view of the many uncertainties in the analysis of the experimental data, it is impossible to tell whether the discrepancy between simulation and experiment is significant.
\begin{table}
\begin{tabular}{||c|c|c||} \hline \hline
## T[K]
& \(\gamma_{sphere}\) & \(\gamma_{cube}\) \\ \hline
## 800
& \(98\pm 2\) & \(80\pm 1\) \\ \hline
## 825
& \(99\pm 1\) & \(79\pm 1\) \\ \hline
## 905
& \(84.1^{(e)}\) & \(67.8^{(e)}\) \\ \hline \hline \end{tabular}
\end{table}
Table 2: Surface free energy density (in \(erg\)\(cm^{-2}\)) assuming spherical and cubical shape for critical nuclei. At 905K we report the experimental value. The entry in the lower right-hand corner is based on the experimental estimate, but assuming a cubical nucleus.
Huang et al. estimated the solid-liquid surface free-energy density of NaCl from the nucleation rate at 550K. To achieve this, Huang et al. assumed that the CNT expression for the nucleation rate is valid. Under those assumptions, they obtained: \(\gamma\)=119.6[\(erg~{}cm^{-2}\)] for a spherical nucleus.
In the present work, we can compute absolute nucleation rates without making use of CNT. The only assumption we make is that the barrier crossing is diffusive and that the Zeldovitch (pre)factor is well approximated by the form given in eqn.. The Zeldovitch factors were found to be respectively \(Z_{1}\)=0.016 for \(T_{1}\) and \(Z_{2}\)=0.013 for \(T_{2}\). The true Zeldovitch factor may be slightly different, but is in any event expected to be of \({\cal O}(10^{-2})\). From our MD simulations we obtained the following estimates for the forward rates (eq. 7) : \(f_{crit}^{+}\)=0.013 ps\({}^{-1}\) for \(T_{1}\) and \(f_{crit}^{+}\)=0.033 ps\({}^{-1}\) for \(T_{2}\). Combining this information, we can compute the kinetic prefactor of eq. 6: \(\kappa\)(\(T_{1}\))= 6.9\(\times\)10\({}^{36}\)\(m^{-3}s^{-1}\) and \(\kappa\)(\(T_{2}\))= 1.5\(\times\)10\({}^{37}\)\(m^{-3}s^{-1}\). As is to be expected, the kinetic prefactor depends only weakly on temperature. Using eqn. we then calculated the nucleation rate. The results are: \(I(T_{1})\)=3\(\times\)10\({}^{26\pm 1}\)\(m^{-3}s^{-1}\) and \(I(T_{2})\)=4\(\times\)10\({}^{24\pm 1}\)\(m^{-3}s^{-1}\).
These nucleation rates are about ten orders of magnitude higher than the estimated experimental rate at 905K (\(O(10^{16})m^{-3}s^{-1}\)). Such a difference is hardly surprising because the nucleation rate is expected to increase rapidly with increasing supercooling.
We also computed the nucleation rate using an algorithm based on the path-sampling techniques of refs. (Forward-Flux Sampling (FFS) and Transition-Interface Sampling (TIS)). The value obtained at \(T_{1}=800\)K is \(I_{FFS-TIS}(T_{1})=O(10^{27\pm 2})\)\(m^{-3}s^{-1}\), which agrees surprisingly well with the one obtained using the diffusive barrier-crossing approach.
As the path-sampling method does not depend on the choice of the reaction coordinate and does not require prior knowledge of the phase space density, we can conclude that the method based on the free-energy calculation gives us a good estimate for the nucleation rate at this temperature.
We have also computed the nucleation rate at T=750 K, using the FFS-TIS method. The computed nucleation rate is \(I_{FFS-TIS}(T=740K)\)=\(O(10^{35})\)\(m^{-3}s^{-1}\). This is of the same order of magnitude as the nucleation rate obtained calculated by Koishi et al. using "brute-force" MD. We cannot use the diffusive barrier crossing method at this temperature, as the barrier is too low to avoid spontaneous nucleation during long runs. However, if we assume that the kinetic pre-factor and the surface free-energy density do not vary much with temperature, we can use CNT to extrapolate the nucleation rate from \(800K\) to \(740K\). We find: \(I_{extrap}(T=740K)=O(10^{30})\)\(m^{-3}s^{-1}\) which is considerably lower than the results of the direct calculations. This suggests that an extrapolation procedure based on CNT is not reliable. A summary of our numerical results for the nucleation barriers and rates are given in Table 3.
A similar problem occurs if we try to extrapolate our numerical data at \(800K\) and \(825K\) to \(905K\), the temperature of the experiments of ref.. If we can extrapolate our simulation results to the experimental temperature of 905 K (\(\beta\Delta\mu\)=0.3) we obtain an estimated nucleation at 905 K that is \(O(6\times 10^{11})m^{-3}s^{-1}\). This is some nearly five orders of magnitude less than the experimentally observed rate.
The discrepancy between simulation and experiment can be due to several reasons. a) There might be an appreciable (but unspecified) error in the experimental estimates (e.g. due to residual heterogeneous nucleation). b) The estimated error in the computed melting temperature of the Tosi-Fumi model is \(\pm 20\)K. Such an uncertainty again easily translates into a variation of the nucleation rate by several orders of magnitude. c) In view of the extreme sensitivity of nucleation rates to the details of the intermolecular potential (see, e.g.), the Tosi-Fumi potential may be inadequate to model nucleation in NaCl, even though it can reproduce the static properties of the solid and liquid NaCl. d) Finally, it it is not quite correct to assume that the kinetic prefactor, the surface free energy and the latent heat of fusion are temperature-independent.
We can also compare our calculated nucleation rate at \(800K\) to the rate estimated with CNT. In order to do that, we need to compute the kinetic pre-factor \(\kappa_{CNT}\) that, using the CNT approximations, is:
\[\kappa_{CNT}=Z\;\rho_{liq}\;\frac{24D_{S}n_{crit}^{2/3}}{\lambda^{2}}. \tag{8}\]
The attachment rate of particles to the critical nucleus (\(f_{n^{crit}}^{+}\)) takes into account the number of available attachment sites on the surface of a spherical nucleus (\(n_{crit}^{2/3}\)) and depends on the jump frequency for bulk diffusion (\(D_{S}/\;\lambda^{2}\)), where \(\lambda\) is the atomic jump distance. Since the functional form of the nucleation barrier can be fitted to the corresponding CNT expression, the computed Zeldovitch factor (Z) coincides with the predicted one. We computed the self-diffusion coefficient with MD simulations using the \(DL\_POLY\) package in the supercooled liquid at \(T_{1}\)=800K. We found \(D_{S}^{Na}=3.4\times 10^{-5}cm^{2}s^{-1}\), in good agreement with an estimate (\(D_{S}^{Na}=2.3\times 10^{-5}cm^{2}s^{-1}\)) based on extrapolation of the available experimental data of Ref. to the temperature \(T_{1}\). Since \(D_{S}^{Na}/D_{S}^{Cl}\approx 1\), we only considered the self-diffusion of the Na\({}^{+}\) ions. We estimated \(\lambda\) as a fitting parameter from the \(f_{n^{crit}}^{+}\) previously calculated; we obtained \(\lambda(T_{1})=10^{2}\dot{A}\). However, considering that the ion size is \(\sigma_{Na}\sim 1.1\;\dot{A}\), this value for the jump distance seems unphysical (\(\lambda\sim 100\sigma\)). Typically, one would expect \(\lambda\) to be of the order of a mean free path. In a molten salt, the mean free path of an ion is certainly less than a particle diameter. This discrepancy also suggests that the CNT picture is inadequate to describe crystal nucleation of NaCl.
In summary, we have computed the crystal nucleation rate of sodium chloride from the melt using two independent methods: one based on calculations of the free
\begin{table}
\begin{tabular}{||c|c|c|c|c|c||} \hline
## T[K]
& \(\beta\Delta\mu\) & \(\beta\Delta G_{crit}\) & \(f_{ncrit}^{+}\) & \(I[m^{-3}s^{-1}]\) & \(I_{FFS-TIS}[m^{-3}s^{-1}]\) \\ \hline
## 800
& 0.54 & 24 & 0.013 & 3 \(\times 10^{26\pm 1}\) & \(10^{27\pm 2}\) \\ \hline
## 825
& 0.48 & 29 & 0.033 & 4\(\times 10^{24\pm 1}\) & \(--\) \\ \hline \end{tabular}
\end{table}
Table 3: Summary of the simulation results for the calculation of the free-energy barrier and the nucleation rate for Tosi-Fumi NaCl.
We have found that, to within an order of magnitude, the two approaches yield the same value for the nucleation rate. When we use Classical Nucleation Theory to extrapolate our numerical data to lower temperatures, we observe serious discrepancies with the results of direct calculations. When we use CNT to extrapolate to high temperatures, we find serious discrepancies with the nucleation rates found in experiments. Several factors may contribute to this discrepancy but, at present, it is not yet known which factor is most important.
The work of the FOM Institute is part of the research program of FOM and is made possible by financial support from the Netherlands Organization for Scientific Research (NWO). C.V. gratefully acknowledges the financial support provided through the European Community Human Potential Program under contract HPRN-CT-1999-00025, (Nucleus). E.S. gratefully acknowledges the Spanish government for the award of a FPU Ph.D. grant, and the FOM Institute for the hospitality during the period in which this work was carried out.
C.V. and E.S. thanks Angelo Cacciuto and Rosalind Allen for valuable discussions and suggestions, and Georgios Boulougouris and Josep Pamies for a critical reading of the manuscript.
## Appendix A Identification of crystalline clusters
To distinguish between solid-like and liquid-like particles and identify the particles belonging to a solid cluster, we used the local bond order parameter introduced by Ten Wolde el al.. Although the method we used is the same as the one proposed by Ten Wolde, the definition of a'solid-like' particle is not rigorously the same. The q vector and the thresolds selected were optimazed for the NaCl Tosi-Fumi model.
First we computed a normalized complex vector \(q_{4}\) for every particle \(i\). Each component of this vector was given by:
\[\vec{q}_{4,m}(i)=\frac{\frac{1}{N_{b}(i)}\sum_{j}^{N_{b}(i)}\Upsilon_{4,m}( \theta_{i,j},\phi_{i,j})}{\vec{q}_{4,m}(i)\cdot\vec{q}_{4,m}^{\,*}(i)},\qquad m =[-4,4] \tag{1}\]
Where \(N_{b}(i)\) is the number of neighbours of the particle \(i\) within a cut-off radius of \(4\dot{A}\) (the first minimum in the Na-Cl radial distribution function).
Then we computed a scalar product \(q_{4}(i)\cdot q_{4}^{*}(j)\) for every particle \(i\) with each of its neighbours particle \(j\). A particle was considered to be'solid-like' when at least 6 of the scalar products were bigger than 0.35. Finally two'solid-like' particles were considered to be neighbours in the same cluster if they were closer than 3.4 \(\dot{A}\).
Ionic fluids are more ordered than Lennard-Jones or Hard Spheres ones, as the radial distribution function shows. However, with the method implemented by TenWolde, we were able to clearly distinguish between solid-like and liquid-like particles in the NaCl Tosi-Fumi model. We enclose a plotA1 that shows the distributions of the number of scalar products bigger than 0.35 at T=800 K for the solid and for the liquid: for values bigger than 6 a particle was considered to be solid-like.
|
10.48550/arXiv.0906.5274
|
Rate of Homogeneous Crystal Nucleation in molten NaCl
|
C. Valeriani, E. Sanz, D. Frenkel
| 4,593
|
10.48550_arXiv.1905.06168
|
## 1 Introduction
Synthetic methods allowing one-step C-C bond formation through homogenous-catalyst-mediated transformation of C-H bonds have become increasingly important in both industry and academia. Potential routes for selective C-H bond activation and subsequent C-C bond formation in alkenes and aromatic compounds include hydroarylation by addition of aromatic C-H bonds across an unsaturated C=C bond, or an oxidative coupling that preserves the double bond. The latter reaction, while a highly desirable industrial goal, is challenging from the synthetic point of view. Pioneering examples of Ru-catalyzed coupling of aromatic carbon-hydrogen bonds with olefins were reported two decades ago. Since then (for a review, see Ref.), an increasing number of examples catalyzed by Rh, Ru and Pd have been published, but the mechanistic aspects of the reactions have only been addressed by experimental methods.
Motivated by the experimental results of Milstein and coworkers, over a decade ago we started to explore the mechanisms of the concurrent reactions of oxidative coupling and hydroarylation of methyl acrylate (MA) catalyzed by Ru carbonyl complexes (Scheme 1) using hybrid and double hybrid DFT families. These calculations showed that either proton elimination by chloride ion, or hydrogen transfer to coordinated olefin, can serve as the initial step of aromatic C-H bond activation, while oxidative addition mechanism could be excluded. Proton elimination proceeds via transition state **TS1** and initiates oxidative coupling with olefin (**TS2**) according to Scheme 2a. Next, interaction of Ru hydride intermediate **Int3** with MA and HCl regenerates the initial RuCl2 carbonyl (**TS4a**). Alternatively, the catalytic cycle could be closed by interaction of **Int3** with MA and benzene, yielding **Int1 (TS4b)**. Hydrogen transfer to MA (**TS1a** and **TS1b**) causes coexistence of phenyl and one of the two isomeric alkyl ligands in the Ru coordination sphere (**Int4** and **Int5**). This mainly leads to the hydroxylation products via **TS3** and **TS3a** (Scheme 2,b), but coordination of the second olefin molecule followed by oxidative coupling is also possible.
## Scheme 2. Mechanisms of MA interactions with benzene in presence of Ru(II) chloride carbonyl complexes: oxidative coupling (a) and hydroxylation (b).
We found that the activation barriers, the relative energies of the key intermediates, and the overall direction of the catalytic reaction strongly depend on the composition of the Ru coordination sphere. Ru complexes that could form in the reaction mixture, and serve as initial species of catalytic cycles,are shown in Scheme 3. In all the complexes, chloride anions are strongly bound to the metal atom; the only exception is RuCl2(CO)4(benzene) (rightmost complex in the 2\({}^{\text{nd}}\) row of Scheme 3) showing CO insertion into one of the Ru-Cl bonds. Moreover, the energetic results, particularly calculated activation barriers, differ between computational approaches and were hard to reconcile with experimental observations.
Therefore, for this specific case, we will attempt to obtain rigorous first-principles results by means of coupled cluster theory near the complete basis set limit, and use these results to assess the performance of more affordable computational methods. We believe our results have broader relevance for modeling mechanisms of catalytic reactions mediated by transition metal complexes.
The CCSD(T) method5 is considered the "gold standard" of quantum chemistry. However, the computational cost of canonical CCSD(T) calculations scales as O(N7) and becomes prohibitively high for mechanistic studies of practical transition metal catalysis problems. Recently developed domain pair natural orbital methods, such as DLPNO-CCSD(T) of Neese and coworkers6 and PNO-LCCSD(T) of Werner and coworkers,7 scale almost linearly with system size (at least for closed-shell cases) and, in the main group, provide similar accuracy to the corresponding canonical calculation. Recently, benchmark studies of the performance of density functionals for transition metal problems, using DLPNO-CCSD(T) for calibration, have started appearing for reaction energies8 and barrier heights.9 Since the Ru complexes shown in Schemes 2 and 3 are still barely tractable by canonical methods, this enables us to assess the "domain error" for real-size transition metal complexes. In this paper, we assess both DFT and PNO methods against canonical CCSD(T) for the hydroxylation and oxidative coupling of benzene and methyl acrylate (MA) catalyzed by RuCl2-carbonyl complexes, as a representative example for the complex mechanisms of homogeneous catalytic reactions.
Having in mind specific computational problems usually addressed in such mechanistic studies, the calculations were divided into four groups: (i) overall reaction energies (Scheme 1) that do not involve transition metals; (ii) relative energies of stable RuCl2 complexes with CO, benzene and MA (Scheme 3) that should reproduce breaking and formation of the metal-ligand bonds and deep alterations of the coordination sphere and electronic structure of the metal atom; (ii) energies of key intermediates along reaction pathways catalyzed by different Ru complexes (Scheme 2) and (iv) barrier heights along these reaction pathways. It would be natural to calculate relative energies of the carbonyl complexes relative to RuCl\({}_{2}\), however, our calculations revealed that it has a triplet ground state, and that the singlet is essentially purely biradical (which is also reflected in the pathological D\({}_{1}\) diagnostic value of 0.435). Discussion of the open-shell calculations are beyond the scope of this preliminary report; therefore, we used RuCl\({}_{2}\)(benzene) as a reference, as RuCl\({}_{2}\) will anyway have no independent existence under the experimental conditions (benzene solvent). At all levels, 1:1 exchange of benzene with CO or MA is energetically unfavorable; all other complexes are exothermic with respect to the reference. The relative energies of key intermediates and transition states along each reaction path were calculated relative to the initial form of the catalyst, (C\({}_{6}\)H\({}_{6}\))(CO)\({}_{n}\)RuCl\({}_{2}\) (n=0-4).
## Computational Methods
The Weigend-Ahlrichs basis set family def2-TZVP, def2-TZVP, and def2-QZVPP was used throughout.
Reference geometries were optimized at the PBE0-D3BJ/def2-TZVP level using Gaussian 09; identities of transition states were verified by frequency and intrinsic reaction coordinate calculations.
At the final geometries, canonical CCSD(T)/def2-TZVPP single-point energy calculations were performed using MOLPRO 2018, both using default frozen cores and including Ru(4s,4p) subvalence orbitals (which _are_ correlated by default in ORCA. We found in this work that the mean absolute effects of Ru(4s,4p) subvalence correlation on carbonyl ligand energies and transition states are both a nontrivial 1.0 kcal/mol.) DLPNO-CCSD(T) and the version with improved iterative triples, DLPNO-CCSD(T1), were calculated with the def-TZVPP basis set using ORCA, likewise DLPNO-CCSD(T)/def2-QZVPP calculations were done for basis set extrapolation using the simple L\({}^{-3}\) formula. TightPNO cutoffs were used to reduce domain discretization error; we found in the present work that DefaultPNO causes errors up to 3.5 kcal/mol in energy differences, and hence do not recommend its use.
In addition, single-point DFT calculations with a number of DFT functionals were carried out using ORCA. Aside from PBE0 already mentioned, these include: (a) the Berkeley "combinatorially optimized" B97M-V, oB97X-V and oB97M-V; (b) the M06 family: M06-L, M06 and M06-2X; (c) TPSS and two different hybrids thereof, namely, TPSSh and TPSS0 (10% and 25% HF exchange, respectively); (d) both the original double-hybrid DSD-PBEP86-D3BJ and its reparametrized version revDSD-PBEP86-D4, in the latter, D3BJ also replaced with the very recently published next-generation D4 model. As basis set convergence of double hybrids tends to be dominated by the MP2-like term, we carried out def2-TZVP and def2-QZVPP calculations and applied L\({}^{-3}\) basis set extrapolation, for the remaining DFT functionals we applied def2-TZVPP except for oB97, which were accurate enough that we also tried def2-QZVPP. (Changes are on the order of 1 kcal/mol.) GRID6 was used in all DFT calculations.
In all Orca calculations, the RIJCOSX approximation was employed, as well as the RI-MP2 approximation for the double hybrids, in conjunction with the respective appropriate auxiliary basis sets for the def2 family.
## Results and Discussion
MAD (mean absolute deviation) and RMSD (root mean square deviation) error statistics with respect to our best CCSD(T) basis set limit estimates are shown in for the four types of energy differences. As expected, the smallest deviations were found for the overall reaction energies, which involve only main group elements. The hybrid functionals of the M06 family (MAD=0.68 and 0.57 kcal/mol for M06 and M06-2X, respectively) and double hybrid revDSD-PBEP86 functionals (MAD=0.64) show the best performance in this group. However, overall for the four criteria, the best results were obtained using oB97M-V and oB97X-V range-separated hybrids as well as by the revDSD-PBEP86 double hybrid, with accuracy between DLPNO-CCSD and DLPNO-CCSD(T). On the 3\({}^{\rm rd}\) rung (meta-GGA) of the Jacob's Ladder, B97M-V performed best for reaction energies and carbonyl complex stabilities; however, M06-L showed similar performance, and TPSS outperformed them for barrier heights (MAD 3.69 vs. 4.54 for M06-L and 5.47 for B97M-V). The revDSD-PBEP86 functional outperforms the original DSD-PBEP86 for all four groups of calculations, whereas DSD-SCAN-D4 improved on DSD-PBEP86-D3BJ only for the carbonyl complexes. The DLNPO-CCSD(T) approach shows very close agreement (MAD=0.35 kcal/mol) with canonical CCSD(T) for the reaction energies, while for other groups there is a bit more daylight between them (MAD=1.04 for carbonyls, 1.56 for intermediates, and 1.58 kcal/mol for TSe). Using the DLNPO-CCSD(T1) approach with improved perturbative triples decreases these statistics to 0.60, 0.80, and 1.02 kcal/mol, respectively. The RMSD/MAD ratios are close to the theoretical value (for a normal distribution) of \(\sqrt[n(\pi)\)\(\approx\)1.2533, for carbonyls and intermediates but much larger for TSes, indicating an outlier (RMSD=0.80, 0.86, 2.05 kcal/mol): we note that TS2-CO2 is essentially biradical (and has D1\(>\)0.4). If we exclude it, MAD and RMSD drop to quite pleasing values of 0.63 and 0.73 kcal/mol, respectively.
At all levels except for DLPNO-MP2, the highest RMSD and MAD values were found for the first reaction, i.e. dissociation of benzene and association of CO and MA ligands. Most density functionals tend to overbind the ligands; detailed analysis shows that for one group of functionals (PBE0-D3BJ, SCAN-D3BJ, TPSS, M06L and DSD-PBEP86 family, Fig. 2a) the error becomes greater with increasing number of CO ligands; the other group (M06, M06-2X, TPSS0, TPSSh, and B97, Fig. 2b) exhibits a less pronounced opposite trend, especially in the presence of coordinated benzene.
Our findings are consistent (see also our companion paper in the present volume) with the findings of Najibi and Goerigk and ourselves for the very large GMTKN55 main-group benchmark and of Iron and Janes for the MOBH35 transition metal reaction benchmark: Notably, that the range-separated hybrids \(\omega\)B97X-V and \(\omega\)B97M-V acquit themselves particularly well, that revDSD represents an improvement over the original DSD not just for the main group but also transition metals, and that unlike for the main group where empirical double hybrids are clearly superior, they offer no clear advantage over \(\omega\)B97M-V for transition metal reactions. Unlike Iron and Janes, however, who found the new DSD-SCAN double hybrid to be among the best performers for MOBH35, we find it to be inferior to revDSD-PBEP86 and the \(\omega\)B97\(n\)-V family for the present problem.
This research was supported by the Israel Science Foundation (grant 1358/15), the Minerva Foundation, and the Helen and Martin Kimmel Center for Molecular Design (Weizmann Institute of Science).
RMSD and MAD relative to CCSD(T)/CBS for the four types of energetics considered.
Energy deviation relative to CCSD(T)/CBS in different RuCl\({}_{2}\) (CO)\({}_{\text{n}}\) complexes as a function of n.
|
10.48550/arXiv.1905.06168
|
Coupled Cluster Benchmark of New Density Functionals and Domain Pair Natural Orbital Methods: Mechanisms of Hydroarylation and Oxidative Coupling Catalyzed by Ru(II) Chloride Carbonyls
|
Irena Efremenko, Jan M. L. Martin
| 4,128
|
10.48550_arXiv.2205.03691
|
###### Abstract
The time-dependent exchange-correlation potential has the unusual task of directing fictitious non-interacting electrons to move with exactly the same probability density as true interacting electrons. This has intriguing implications for its structure, especially in the non-perturbative regime, leading to step and peak features that cannot be captured by bootstrapping any ground-state functional approximation. We review what has been learned about these features in the exact exchange-correlation potential of time-dependent density functional theory in the past decade or so, and implications for the performance of simulations when electrons are driven far from any ground-state.
## I Introduction
Time-resolved dynamics of electrons in molecules and solids have become increasingly relevant over the past decades. A description beyond equilibrium electronic structure and excitation spectra is necessary in many applications of fundamental and technological importance: photovoltaic processes, photocatalysis, radiation damage in biomolecules, nanoscale conductance devices, time-resolved pump-probe spectroscopies, and strong light-matter coupling for quantum-based technologies. Scalable theoretical methods to computationally model coupled electron, ion, and photon dynamics help in interpreting and predicting experiments, and to suggest new systems with improved functionalities. Arguably, the electron-component of the problem is the most challenging: there are many of them, they interact with each other as well as with the nuclei and light fields, and they require a quantum mechanical description, unlike nuclei and photons for which, for different reasons, a classical description may suffice.
To this end, time-dependent density functional theory (TDDFT) has emerged as a method of choice. TDDFT is inherently a "real-time" method in that dynamics is woven into the formulation from the very beginning with the foundational theorem developed for general time-dependent evolution of arbitrary initial states. While perturbations around the ground-state yield a formalism for excited state energies, their couplings, and response, it is particularly the generality and practical efficiency of the real-time formulation that has led to the possibility of applications on large systems which could not be done otherwise, e.g..
In a sense, TDDFT is an extension of its ground-state counterpart (DFT). The central object is the one-body density, \(n(\mathbf{r},t)\), which is the probability density of finding any one electron at point \(\mathbf{r}\) in space at time \(t\): \(n(\mathbf{r},t)=N\sum_{\sigma_{1}\ldots\sigma_{N}}\int d^{3}r_{2}...d^{3}r_{N} |\Psi(\pi\sigma_{1},\mathbf{r}_{2}\sigma_{2}...r_{N}\sigma_{N})|^{2}\) where \(\sigma_{i}\) represents the spin coordinate, and \(N\) is the number of electrons in the system. Just like DFT, TDDFT allows one to bypass having to solve the analytically and computationally complicated many-body time-dependent Schrodinger equation (TDSE) and instead only requires one to solve a set of time-dependent single-particle equations involving a modified one-body potential whose solutions yield the same time-dependent one-body density as that of the true system. This drastically reduces the amount of computational effort in obtaining quantum many-body dynamics.
This is an audacious concept: the idea that a set of non-interacting electrons reproduces the density of interacting electrons seems fantastical, especially in the time-dependent case, where the motion of the electrons driven by some external perturbation or by nuclear motion is affected by their mutual repulsion in an intricate dance. The one-body potential somehow directs the non-interacting electrons to evolve with the same one-body density as the interacting electrons. Recent work has shown that this choreography results in dynamical step and peak structures that have a non-local density-dependence in time and in space, and are completely missed by the commonly used "adiabatic" approximations. Such approximations utilize ground-state exchange-correlation (xc) functionals, and although ground-state potentials also may feature steps and peaks, they appear only in particular situations such as static correlation or fractionally charged systems, while the TDDFT steps appear quite generically.
In this review, we revisit what is known about non-adiabatic features of the exact xc potential in TDDFT, focussing on the dynamical steps and peaks. We begin with a brief reminder of the fundamental theory of TDDFT in Sec. II, including a discussion of memory-dependence. Section III demonstrates the dynamical peaks and steps that arise from memory-dependence, by numerical examples as well as analysis of the equations, identifying a "kinetic component" in the xc potential as responsible for these features, and discussing the role of the local accelerations in the system. In Section IV we make a case study on the helium atom, showing that these features, which have largely been discussed in one-dimensional (1D) systems, persist just as strongly in three-dimensions, and verifying the relevance of the analyses previously made. We summarize in Section V.
## II TddFT formalism
At the heart of TDDFT lies the Runge-Gross theorem which, for time-dependent problems, plays an analogous role to the Hohenberg-Kohn theorem for ground-state problems. Namely, for a fixed particle-particle interaction and statistics, it establishes a one-to-one mapping between the possibly time-dependent external potential acting on the electrons and the time-dependent density for a given initial state:
\[\Psi:n\leftrightarrow v_{\rm ext} \tag{1}\]
As a consequence, all physical observables can be expressed as functionals of the density and initial state: for a given \(\Psi\), \(n\) points to a unique \(v_{\rm ext}\) which points to a unique unique Hamiltonian, which in theory points to a unique time-dependent wavefunction, from which any observable can be extracted in principle.
But this can be of little more than theoretical interest unless we know how to obtain the density of the interacting system and the observables of interest. As in ground-state DFT, one maps the interacting system to a fictitious non-interacting system, the Kohn-Sham (KS) system, whose orbitals are required to reproduce the density at any instant. The KS orbitals evolve in the one-body KS potential, \(v_{\rm S}({\bf r},t)\), according to a single-particle TDSE:
\[\left(-\frac{\nabla^{2}}{2}+v_{\rm S}({\bf r},t)\right)\phi_{i}({\bf r},t)=i \partial_{t}\phi_{i}({\bf r},t)\,. \tag{2}\]
(Atomic units are used throughout). The KS potential is written as the sum of three terms in a similar way to DFT:
\[v_{\rm S}({\bf r},t)=v_{\rm ext}({\bf r},t)+v_{\rm H}[n]({\bf r},t)+v_{\rm XC}[n;\Psi_{0},\Phi_{0}]({\bf r},t) \tag{3}\]
We note that there is much freedom in selecting the initial KS state \(\Phi_{0}\): it can be any wavefunction that reproduces the density of the true interacting initial state and its first time-derivative.
The vast majority of applications of TDDFT are for linear response where a perturbative limit of Eqs.- yields a Dyson-like equation, or matrix equations, in the frequency domain, whose solution gives excitation energies and oscillator strengths. The success of TDDFT in this regime, with the available approximate xc functionals, is incontrovertible, e.g. Refs, however not without plemishes for certain classes of excitations, e.g. double-excitations and charge-transfer excitations, which can create havoc for the black-box use of the method in coupled electron-ion dynamics. But the fully non-perturbative regime is of particular interest in TDDFT, due to the much harsher computational scaling of alternative methods.
Armed with a good approximation to the xc potential, solving Eq. with Eq. yields a good approximation to the density of the physical interacting system, obtained from non-interacting KS electrons, through \(n({\bf r},t)=\sum_{i=1}^{N}|\phi_{i}({\bf r},t)|^{2}\). Observables that are directly related to the density, such as the dipole moment, can be directly extracted, while further approximations would be needed to extract other observables, such as momentum distributions, double-ionization cross-sections, and current-densities, where the rotational part differs in general from that of the KS system. More often than not, these observables are approximated by those of the KS system; an open question is how errors from the xc functional itself compare with the (usually unacknowledged) errors from the observable evaluation.
An exact expression for the xc potential can be derived by equating the second time-derivative of the density, \(\tilde{n}({\bf r},t)\), for the KS system with that of the interacting system. While the Heisenberg equation of motion for the density gives the continuity equation \(\dot{n}=-\nabla\cdot{\bf j}\), \(\tilde{n}({\bf r},t)\) can then be obtained from the equation of motion for the current-density \({\bf j}({\bf r},t)\). Subtracting the equation obtained for the interacting system from that for the KS system results in the decomposition of the xc potential, \(v_{\rm XC}({\bf r},t)\) into kinetic (T) and interaction (W) terms \(v_{\rm XC}({\bf r},t)=v_{\rm C}^{\rm T}({\bf r},t)+v_{\rm XC}^{\rm W}({\bf r},t)\), that have the following structure :
\[\nabla\cdot\left(n\nabla v_{\rm XC}^{\rm W}\right)=\nabla\cdot\left(n({\bf r },t)\int n_{\rm XC}({\bf r},{\bf r}^{\prime},t)\nabla w(|{\bf r}^{\prime}-{ \bf r}|)d^{3}{\bf r}^{\prime}\right) \tag{4}\]
\[\nabla\cdot\left(n\nabla v_{c}^{\rm T}\right)=\nabla\cdot\left({\cal D}_{{\bf r }^{\prime},{\bf r}}\Delta\rho_{1}({\bf r}^{\prime},{\bf r},t)|_{{\bf r}^{ \prime}={\bf r}}\right)\,, \tag{5}\]
(with the \(({\bf r},t)\)-dependences on the left-hand-side understood) where \(n_{\rm XC}({\bf r},{\bf r}^{\prime},t)\) is the xc hole, defined through the diagonal two-body density matrix \(\rho_{2}({\bf r},{\bf r}^{\prime};{\bf r},{\bf r}^{\prime})=N(N-1)\int dr_{3}...dr _{N}|\Psi({\bf r},{\bf r}^{\prime},{\bf r}_{3}..r_{N})|^{2}=n({\bf r}^{\prime}, t)\left(n({\bf r},t)+n_{\rm XC}({\bf r},{\bf r}^{\prime},t)\right)\) and \(\Delta\rho_{1}({\bf r}^{\prime},{\bf r},t)=\rho_{1}({\bf r}^{\prime},{\bf r},t )-\rho_{1,S}({\bf r}^{\prime},{\bf r},t)\) is the difference between the spin-summed one-body density matrix of the true interacting system \(\rho_{1}({\bf r}^{\prime},{\bf r},t)\) and that of the Kohn-Sham system, \(\rho_{1,S}({\bf r}^{\prime},{\bf r},t)\). The differential operator \({\cal D}_{{\bf r}^{\prime},{\bf r}}=\frac{1}{4}(\nabla^{\prime}-\nabla)(\nabla^ {2}-\nabla^{\prime 2})\). Eqs.- give an exact expression for the xc potential in terms of the exact xc hole, and the exact and KS one-body reduced matrices. They are useful for analysis of TDDFT and understanding errors in approximations. Further, they offer a starting point for approximations: although \(\rho_{1,S}({\bf r}^{\prime},{\bf r},t)\) is accessible in a KS evolution, \(\rho_{1}\) and \(n_{\rm XC}\) need to be approximated in terms of KS quantities. We will also return to this in Sec. III.
### Memory: History and Initial-State Dependence
In traditional wavefunction-based quantum mechanics, knowing the wavefunction \(\Psi(\mathbf{r}_{1}...\mathbf{r}_{N},t)\) at any time \(t\) is enough to know all properties of the system at that time; its value at earlier times is not required. The same cannot be said about the one-body density. While the TDDFT reformulation of the many-body problem in terms of its one-body time-dependent density \(n(\mathbf{r},t)\) drastically simplifies the problem from the computational viewpoint, it inevitably introduces some complications. In particular, the xc potential is memory-dependent in that the functional \(v_{\mathrm{xc}}[n,\Psi_{0},\Phi_{0}](\mathbf{r},t)\) depends not only on the instantaneous density \(n(\mathbf{r},t)\) but also on the history of the density \(n(\mathbf{r},t^{\prime}<t)\) and the initial states \(\Psi_{0},\Phi_{0}\). This follows directly from the Runge-Gross theorem Eq.: the mapping is between the density- and potential- functions over space and time, and is one-to-one for a given initial state. This would mean \(v_{\mathrm{ext}}(\mathbf{r},t)\) functionally depends on the density over all times but due to causality it depends only the history of the density, not its future. The initial-state dependence means that the same time-dependent density can be obtained by propagating in two different potentials if the systems begin in different initial states. Applying Eq. to the KS system yields that \(v_{\mathrm{s}}(\mathbf{r},t)\) functionally depends on the history of the density and the KS initial state, and thus \(v_{\mathrm{xc}}(\mathbf{r},t)=v_{\mathrm{s}}(\mathbf{r},t)-v_{\mathrm{ext}}( \mathbf{r},t)-v_{\mathrm{ir}}(\mathbf{r},t)\) functionally depends on both the true and KS states: \(v_{\mathrm{xc}}[n;\Psi,\Phi](\mathbf{r},t)\). As mentioned earlier, one can begin in any initial KS state that reproduces the density of the initial interacting state and its first time-derivative; the structure of the exact xc potential has a strong dependence on this choice.
There is an intimate connection between history-dependence and initial-state-dependence that can be useful. What we take as the "initial" time can be re-set, and if we know the wavefunctions at the reset time, then we can evaluate the xc potential on the domain of those wavefunctions and a truncated history, i.e.
\[v_{\mathrm{xc}}[n;\Psi_{0},\Phi_{0}](\mathbf{r},t)=v_{\mathrm{xc}}[n_{t^{\prime }};\Psi_{t^{\prime}},\Phi_{t^{\prime}}](\mathbf{r},t),\ \ \ t^{\prime}\leq t \tag{6}\]
A useful consequence is that if the xc potential is found at some time \(t\) for a particular dynamics of a system, then in _any_ dynamics where the exact same interacting and KS states happen to be reached at some time, then the xc potential at that time will be the same. We will exploit this in Section IV to emphasize the generality of the features of the xc potential found there.
Almost all calculations today however completely neglect memory. They use an adiabatic approximation, in which the instantaneous density is input into a ground-state approximation: \(v_{\mathrm{xc}}^{A}[n;\Psi_{0},\Phi_{0}](\mathbf{r},t)=v_{\mathrm{xc}}^{ \mathrm{g.s.}}\left[n(t)\right](\mathbf{r})\). Such an approximation has two sources of error: one arising from the approximation made for the ground-state functional, and the other from the adiabatic approximation itself. To isolate the error from the adiabatic approximation itself, the adiabatically-exact approximation is defined as \(v_{\mathrm{xc}}^{\mathrm{adia-xc}}[n;\Psi_{0},\Phi_{0}](\mathbf{r},t)=v_{ \mathrm{xc}}^{\mathrm{exact-g.s.}}[n(t)](\mathbf{r})\) which can be a useful analysis tool in cases where the exact ground-state xc potential can be computed (usually model 1D systems). Due to the lack of memory, the adiabatic approximation leads to large errors in some applications, sometimes failing completely, but in other cases it has been found to yield good predictions, even when the system is far from a ground-state. It is not completely understood why: possible reasons include, that the adiabatic approximation satisfies a number of exact conditions that are important in the time-dependent case, that in some applications a strong external field dominates over xc effects in driving the dynamics and that partial compensation of self-interaction in the Hartree potential, even at the ground-state level, is enough, especially when the observables involve averaging over the details of the density distribution. Recently it was further argued that an indicator of the expected success or failure of the adiabatic approximation lies in the natural orbital occupation numbers: if the initial KS state has a configuration close to that of the true initial state, and the natural orbital occupation numbers of the true system do not evolve significantly in time, the adiabatic approximation may make good predictions even for strongly non-perturbative dynamics.
### Examples showing the relevance of memory
Model systems have been crucial in understanding cases where the adiabatic approximation fails, because numerically exact solutions are available and because the exact xc potential can be extracted to compare with approximations. shows the errors that adiabatic approximations make in a variety of studies.
The top panel shows the electronic dipole for field-free evolution of a state that is a 50:50 superposition of the ground and first-excited singlet state of a 1D soft-Coulomb-interacting He atom; such a state may be reached from a ground state driven by a field that is then turned off, for example. In that scenario, the natural choice for the KS initial state is a Slater determinant, which for our two-electron system, is a doubly-occupied spatial orbital. As seen in the figure, the adiabatic exact-exchange (AEXX) and local density approximations (ALDA) do not get the period of oscillations correct, and display additional beat frequencies. In Sec. III, we will see that the exact xc potential has non-adiabatic step and peak features that even the adiabatically-exact approximation lacks.
Such features have an even more notable effect in electron-scattering, illustrated in panel b in This shows the reflection probability of an electron initially in a gaussian wavepacket moving towards a target 1D H atom. The approximations severely underestimate the reflection of the wavepacket (\(N_{R}\) is the numberof electrons on the right of the atom). The dashed and solid lines represent different choices of initial KS state, both of which have the same density as the true state: In one, a Slater determinant is chosen for example to simulate the return of an ionized electron to its parent, while in the other, the initial KS state is a two-orbital state which has the same configuration as the interacting state with one electron in the bound target ion, and the other in a gaussian wavepacket. Although there are significant differences in the details of the two time-dependent densities ensuing from these states when propagated under ALDA or AEXX, with the Slater determinant demonstrating spurious oscillations and reproducing the exact dynamics less accurately initially, neither of them capture the eventual reflection even qualitatively. The exact xc potential shows a step and valley structure that is essential for this effect, and is missing in the adiabatic approximations (see also Sec. III). This may explain the underestimated predictions of scattering probabilities and energy transfer that has been observed in real systems.
Panel c of gives an example of a resonantly-driven charge-transfer out of the ground-state. A molecule is modeled via an asymmetric double-well, in which the ground-state has two electrons in the left-well. Applying a weak field that is resonant with an excitation to a charge-transfer state triggers a Rabi oscillation, as evident by the large change in the dipole seen as the molecule reaches the charge-transfer state. Although the adiabatic TDDFT approximations shown yield excellent values for the charge-transfer excitation energies, they completely fail to capture the charge-transfer dynamics. Again, dynamical step features develop as the electron transfers (see Sec. III). But adiabatic TDDFT generally fails at resonant driving, even to local excitations, because it violates the fundamental condition that resonant excitation frequencies of a system should not shift with the instantaneous state; with adiabatic approximations, the density-dependence of the KS potential leads to the response of a non-equilibrium state having spuriously-shifted poles, while the exact generalized xc kernel requires a frequency-dependence to correct this spurious shift. Although it has not yet been explicitly shown, it is likely this frequency-dependence is related to dynamical steps and peaks in the time-domain. This could lead to some unreliability in TDDFT simulations of pump-probe spectroscopy; absorption peak-shifts have been observed in a number of molecules.
## III Dynamical steps and peaks
One common thread across all the examples in Sec. II.1, is the presence of prominent step and peak features in the exact xc potential, which are significant on the scale of the total KS potential.
Three examples illustrating substantial errors in adiabatic TDDFT propagation. a) Dipole moment in field-free propagation of a 50:50 superposition of the ground and first excited state in a 1D He atom, showing the exact, against AEXX and ALDA. The initial KS orbital is chosen to be a Slater determinant. b) Scattering of an electron off a 1D H-atom; the top panel shows the density upon approach, while the lower panel shows the integral of the electron density to the right of the atom (\(x\geq-5\)); dashed curves for the Slater determinant initial state, and solid for a two-orbital state. c) Resonantly-driven charge-transfer out of the ground-state of the double-well shown. Neither self-interaction-corrected LDA (SIC-LDA) nor AEXX capture the dynamics, despite yielding good approximations for their excitations.
The exact xc potential was computed by numerically inverting the propagation operator viewed as a function of the potential to target the exact density. In the case of a Slater determinant formed with a doubly-occupied orbital, the KS orbital is directly related to the exact density, and then the xc potential can alternatively be computed using analytical formulas (see also Sec. IV).
A salient feature common to all the exact xc potentials over the range of different dynamics are step- and peak-like structure which are absent in the adiabatic approximations, including the adiabatically-exact (although partially reproduced in case c)). These features are therefore non-adiabatic and their absence in adiabatic approximations yield inaccurate density-dynamics.
Given that in this case the KS state is a doubly-occupied orbital, \(v_{\rm x}=-v_{\rm n}/2\) has a simple well structure that smoothly cradles the density (the exact \(n({\bf r},t)\) is shown as blue dotted), the step and peak are features of the correlation potential which often dominates the KS potential. In the case shown, the KS state has a fundamentally different structure to the true state; instead, if the KS state is chosen with the same configuration of the true state, the step features are smaller but still appear, although their impact on the ensuing dynamics is less.
In we display the xc potential and density for the scattering example of Fig 1b, for the case when \(\Phi=\Psi\). The middle panel shows again a prominent peak and step feature just to the right of \(x=-10\)a.u. at the shoulder of the electron density (top panel), which persists over time and contains the effective correlation needed for reflecting part of the density of the two KS electrons. Adiabatic approximations miss this structure and fail to reflect: the top right and middle panels show the density and xc potential for the ALDA in red. The feature appears in the \(v_{\rm C}^{\rm T}\) component of the exact potential, and the lower panels show that while the adiabatically-exact approximation accurately captures \(v_{\rm XC}^{\rm W}\), it hardly resembles \(v_{\rm C}^{\rm T}\) at all. The figure also plots a non-adiabatic approximation, \(v_{\rm XC}^{\rm S}\), which replaces the exact xc hole and one-body density-matrix in Eqs.- with their KS counterparts; this yields a reasonable approximation to \(v_{\rm XC}^{\rm W}\) but gives zero for \(v_{\rm C}^{\rm T}\).
Finally, turning to the example of resonantly-driven charge-transfer of Fig. 1c, we plot in the exact correlation potential at the time the charge-transfer state is reached. A step appears in the adiabatically-exact potential, however it has the wrong height. Ref. shows that the heights of the two steps in the limit of large separation between the donor (D) and acceptor (A) can be given by the differences in ionization potential and electron affinity indicated in the figure; the adiabatically-exact step has a size equal to the \((N_{D}-1)\)-electron derivative discontinuity, and is smaller than that of the exact.
Non-adiabatic step and peak features in \(v_{\rm XC}\) for the examples of (a) Snapshots of \(v_{\rm XC}(x,t)\) and \(n(x,t)\) for the field-free dynamics of the superposition state of the 1D He atom, over a half-Rabi cycle, whose dipole appears in The adiabatically-exact and ALDA potentials lack the prominent step and peak structures of the exact. (b) Snapshots of the density and potentials for the e-H scattering problem of The top panel shows the density at a time when the electron begins to reflect from the atom and partially transmit to the left, with the grey being the initial density for reference; the exact (black, left panel), ALDA (red, right panel), \(v_{\rm XC}^{\rm S}\) (blue, right panel). The middle panels show the exact \(v_{\rm XC}(x,t)\) at that time (black), ALDA (red), \(v_{\rm XC}^{\rm S}\) (blue), and grey shows the exact \(v_{\rm XC}(x,0)\). The lower left shows the adiabatically-exact approximation (orange) for the exact \(v_{\rm XC}^{\rm W}\) component (black), and lower right \(v_{\rm C}^{\rm T}\) component. Reproduced from Ref. with permission from the Royal Society of Chemistry.) The exact and adiabatically-exact correlation potentials at the final time when the charge-transfer state is reached, for the resonantly-driven charge-transfer dynamics of
transfer, there is also an oscillatory dynamical step similar in nature to that in part (a).
Thus these non-adiabatic step and peak features have been demonstrated on a wide range of dynamics, and we now briefly mention some beyond those above. Ref. found them in the exact KS potential for a model describing the propagation of a single electron through an infinite semiconductor wire, using a quasi-particle wavepacket of nonzero crystal momentum added to the ground-state of a semiconductor. Refs. studied field-induced tunneling in a system of two or three spinless electrons in 1D. Ref. showed barrier structures that were essential for autoionization processes in a 1D He atom. Refs. considered Hubbard models, while Ref. demonstrated their importance for single-electron transport through a quantum dot using an Anderson model. Although computationally more convenient to demonstrate on 1D model systems, in Section IV we use the example of the real (three-dimensional) Helium atom to illustrate how the dynamical non-adiabatic features in the exact \(v_{\rm XC}\) shown in the 1D helium atom persist just as strongly in the three-dimensional case.
Step and peak features are no strangers to DFT. In the ground-state, they appear in situations associated with "fractional charge" regions, for example when a molecule is near a metal surface, related to the derivative-discontinuity. They appear in interatomic regions of dissociating molecules where they are signatures of static correlation. Without the interatomic step, which is equal to the ionization potential difference between the two fragments, the KS system dissociates to unphysical fractional charges. Associated with the onset of the interatomic step is a peak, and recently a secondary peak was found to appear in the very low density region far to the side of the molecule that heralds the descent of the step back to zero asymptotically. These step and peak features appear in the correlation potential and can be analyzed in terms of changes in the conditional amplitudes of the interacting and KS systems, defined through a factorization, \(\Psi({\bf r}_{1},{\bf r}_{2}...{\bf r}_{N})=\sqrt{\frac{n({\bf r}_{1})}{N}} \Psi_{\rm cond}({\bf r}_{2}...{\bf r}_{N})\). Although the exact exchange potential displays step features when the orbital dominating the density switches to one with a different asymptotic decay, it does not capture the interatomic step in dissociating diatomic molecules which is a strong correlation effect. In the absence of nodal planes of the highest occupied molecular orbital (HOMO), the steps go back down to zero asymptotically, which is a striking difference with the non-adiabatic steps in the time-dependent case. Even when there are nodal planes, steps, peaks, and diverging behavior appear as one traverses across the nodal plane of the HOMO in cases where the density has isotropic decay away from the system. We stress here that the exact ground-state xc potential captures these features but does not capture the non-adiabatic steps in non-equilibrium TDDFT.
In the linear response regime, steps have appeared that counter the electric field across long-range molecules, and in the xc kernel near charge-transfer excitations associated with a derivative-discontinuity. Steps were found to be essential in Coulomb blockade phenomena, again related to fractional charges and the derivative-discontinuity. These cases involve perturbations around the ground-state, and the steps are associated with fractional charge effects. Beyond the response regime, dynamical steps were found in ionization processes, where an adiabatic approximation that depended on the fraction of charge remaining locally near the parent atom was able to approximate them.
The non-adiabatic steps we are discussing here are a distinct phenomenon to all these cases: they are not related to fractional charges, ionization, nor is an external field necessary, and they are missing from any adiabatic approximation, that is, they cannot be captured by using any ground-state approximation. They appear in the kinetic component of the xc potential, Eq., and are instead related to the local KS velocities and accelerations in the system, as we will discuss next.
### Relationship to local velocity and acceleration
Earlier analysis on the two-electron systems have shown that the dynamical steps tend to be associated with the spatial integral of the local KS acceleration in the system, while the peak structures are associated with a maximum in the local KS velocity, or at maximum curvature in the density in regions of low density. To see this, consider first the two-electron spin-singlet where the KS initial state is chosen to be a Slater determinant.
\[\nabla\cdot{\bf j}=\nabla\cdot(n({\bf r},t)\nabla\alpha({\bf r},t))=-\frac{ \partial}{\partial t}n({\bf r},t)\,. \tag{8}\]
We note that the KS current-density \({\bf j}_{\rm s}({\bf r},t)=n\nabla\alpha({\bf r},t)\) may differ from the true current-density \({\bf j}({\bf r},t)\) by a rotational component (although in 1D they are identical, \({\bf j}_{\rm s}({\bf r},t)={\bf j}({\bf r},t)\)). Inverting Eq. yields the exact KS potential:
\[v_{\rm s}({\bf r},t) = \frac{\nabla^{2}\sqrt{n({\bf r},t)}}{2\sqrt{n({\bf r},t)}}-\frac{| \nabla\alpha({\bf r},t)|^{2}}{2}-\frac{\partial\alpha({\bf r},t)}{\partial t}\] \[= \frac{\nabla^{2}\sqrt{n({\bf r},t)}}{2\sqrt{n({\bf r},t)}}-\frac{1 }{2}u^{2}({\bf r},t)-\int_{{\bf r}_{0}}^{{\bf r}}\frac{\partial{\bf u}({\bf s} ^{\prime},t)}{\partial t}\cdot\hat{\bf s}^{\prime}ds^{\prime}\]where the second line is written in terms of local KS velocities and accelerations and is defined up to a global spatial constant. The second term in Eq. is directly related to the local KS velocity, \(\mathbf{u}(\mathbf{r},t)=\frac{\mathbf{\dot{s}}_{\mathbf{\mathbf{\mathbf{\mathbf{ \mathbf{\mathbf{\mathbf{\mathbf{\mathbf{\mathbf{\mathbf{\mathbf{\mathbf{\mathbf{\mathbf{ \mathbf{\mathbf{\mathbf{\mathbf{\mathbf{ \mathbf{ }}}}}}}}}}}}}}}{\mathbf{n}( \mathbf{r},t)}\), and this term tends to produce peak structures, especially near density minima. The third term is related to a spatial-integral of the local KS acceleration, and relative to some reference position \(\mathbf{r}_{0}\), it is given by \(\dot{\alpha}(\mathbf{r},t)-\dot{\alpha}(\mathbf{r}_{0},t)=\int_{\mathbf{r}_{0 }}^{\mathbf{r}}\partial_{t}\mathbf{u}(\mathbf{s}^{\prime},t)\cdot\dot{\mathbf{ s}}^{\prime}ds^{\prime}\) where \(\mathbf{s}^{\prime}\) represents a path from \(\mathbf{r}_{0}\) to \(\mathbf{r}\). This shows that when there is a localized peak in the KS acceleration this term develops a step-like feature there. It should however be noted that the first term of Eq. may also have structures that one can associate with local peaks and local steps: these can partially cancel or enhance the velocity and acceleration terms. For example, in the trivial one-electron case one may get peaks in the local velocity and acceleration but \(v_{\textsc{xc}}\) is zero, since the local curvature of the density can also display local steps and peaks especially near density-minima. But steps that straddle across one side of a localized density to the other, leading to different asymptotic values of the potential in different directions, have a height given by the spatial integral of the localized KS acceleration across the system. We will show an example of this Sec. IV for the (three-dimensional) He atom.
Eq. holds for the special case of two-electron singlet systems where the KS initial state is chosen to be a Slater determinant. For a general choice and for \(N\)-electrons, the same expression holds but where the density, velocity, and acceleration are instead those of one of the occupied orbitals. Different orbitals yield different terms individually but the sum of the terms must be the same for each orbital; an example can be found in Ref..
### Approximations for the dynamical steps and peaks: Importance of \(v_{\textsc{xc}}^{\mathrm{T}}\)
Although it is tempting to build an approximation directly from Eq., it is not clear how to, because hiding in the expression is the external potential, whose functional dependence would somehow need to be subtracted out. In practise, the external potential is input as the potential function (_not_ density functional) that is physically applied to the system, and only the xc and Hartree potentials are considered as functionals of the density and initial states. In fact this is essential, since without explicit input of the external potential one could not propagate with Eq.: it defines the physics of the problem being solved, and if instead it was represented by a density functional, causality and predictivity issues arise. We want an approximation for only the xc potential, not for the entire KS potential. Still, it is interesting to note that an approximation for the non-adiabatic part of the full KS potential itself has been explored, by transforming to the local instantaneous rest frame of the density, taking the adiabatic approximation to be exact in such a frame, and transforming back. This results in an approximation that adds precisely the last term of Eq. to an adiabatic approximation.
\[-\frac{\partial}{\partial t}\int^{x}\frac{j(x^{\prime},t)}{n(x^{\prime},t)}dx ^{\prime}=v_{\mathrm{ext}}(x,t)-\int^{x}\frac{\frac{\partial j^{2}(x^{\prime},t)}{\partial x^{\prime}}}{2n^{2}(x^{\prime},t)}+\frac{\mathcal{Q}(x^{\prime},t)}{n(x^{\prime},t)}dx^{\prime} \tag{10}\]
Instead, Eqs.- can be used to develop approximations. Moreover, analyses based on this decomposition have revealed that the non-adiabatic dynamical features figure more prominently in the kinetic term as compared to the interaction term. As demonstrated in one of the examples in Fig. 2, the step structures are a feature of the kinetic component \(v_{\textsc{c}}^{\mathrm{T}}\) and are absent in adiabatic approximations to this term. While adiabatic approximations may tend to approximate \(v_{\textsc{xc}}^{\mathrm{W}}\) well, they do a poor job of capturing the structure in \(v_{\textsc{c}}^{\mathrm{T}}\). This is not surprising given the form of the two terms: While the interaction term benefits from smoothing through the integral, the appearance of multiple gradients in the kinetic component can create large step and peak features that evolve in time, especially in regions where the density has local minima.
One such class of approximations is the density-matrix coupled approximation of Refs.. Here, the correlated one-body density matrix needed in \(v_{\textsc{c}}^{\mathrm{T}}\) is replaced by an approximated one computed from the first equation of the BBGKY hierarchy, and propagated alongside the KS calculation. In the BBGKY equation the two-body density matrix is approximated by that of the KS system. This ensures that the diagonal part of this density matrix stays equal to the density of the KS system at all times, while the off-diagonal terms can differ from the KS one-body density matrix (as would be the case in an exact calculation).
Even though this approach satisfies most exact conditions known in TDDFT, and is able to capture the elusive dynamical steps and peaks, it remains numerically too unstable to be of practical use. An improved approach within this class of approximations would include an approximation for two-body density matrix as a time-dependent functional of either or both the KS and correlated one-body density matrices.
Eq. suggests that the steps are a probe of the difference between the local character of the true and KS density-matrices near the diagonal (Eq. 5). Choosing initial KS states to minimize this difference may improve the performance of the adiabatic approximations at short times. The idea that the performance of an adiabatic approximation is connected with how small the variation of natural orbital occupation numbers is was investigated in Ref.. Ref. found that the largest step structures appeared at local minima of the largest occupation numbers. However it should also be noted that the impact of a large localized dynamical step or peak on the dynamics may be relatively small if the structure oscillates rapidly in time or is very localized in a region of small density.
## IV Case study: dynamics in the helium atom
The examples illustrated in the previous section involved 1D systems, and a question then arises: are the non-adiabatic steps and peaks equally prominent in real three-dimensional systems? It is often said that correlation effects are enhanced in reduced dimensionality. A sensible comparison would be to compare the steps and peaks in a three-dimensional system undergoing analogous dynamics to one of the 1D examples.
We find the exact interacting state at time \(t\) in terms of the eigenstates of the He atom: \(1^{1}S_{0}\), denoted \(\Psi_{0}\), and singlet first excited state \(2^{1}P_{1}\) that has angular quantum numbers \(L=1\) and \(M=0\), denoted here \(\Psi_{1}\).
\[|\Psi(t)\rangle=\frac{1}{\sqrt{1+|a|^{2}}}\left(|\Psi_{0}\rangle+ae^{-i\omega t }|\Psi_{1}\rangle\right) \tag{11}\]
The parameter \(a\) allows us to tune the proportion of the excited state \(\Psi_{1}\) that goes into the superposition. For the analog to the 1D He example in Figs. 1a and 2a, \(a=1\), the 50:50 superposition.
\[n({\bf r},t)=\frac{1}{1+|a|^{2}}\left(n_{0}({\bf r})+|a|^{2}n_{1}({\bf r})+2an _{01}({\bf r})\cos(\omega t)\right) \tag{12}\]
We note that although this may seem to be a very special case of dynamics, the results for the xc potential at any given time \(t\) in fact apply to a far wider class of dynamics: due to Eq. the xc potentials we will find apply in any situation where the instantaneous true state is given by Eq. 11 and an initial Slater determinant is chosen for the KS state. This means that independently of how the system reached instantaneous state at some time \(t\) the xc potential at that time is the same.
To find the exact xc potential, we will need to invert the time-dependent KS equation for the three-dimensional case, but first we need to find the exact densities of the interacting eigenstates and the transition density in order to construct the target density Eq.. We use the results of Refs. to find the exact density. There, the system is solved using the close-coupling method, in which the angular part of the exact wavefunction is written in terms of coupled spherical harmonics while the radial part is obtained by the finite element discrete variable representation method.
### Extracting the exact xc potential
Here we choose the initial KS state as a Slater determinant: this is the natural choice if the state Eq. is reached from applying an external field to a ground-state and then turning the field off. One would use ground-state DFT to find the initial KS orbitals, and by the ground-state theorems, this is a Slater determinant. Since the KS evolution involves a one-body Hamiltonian, the state remains a single Slater determinant.
This choice of initial KS state is the most relevant choice for simulations of general dynamics of the atom when beginning in the ground state. This is often the case for simulating experiments involving laser-driven dynamics, for example, or in photovoltaics where one models the initial photo-excitation process itself; and, as mentioned above, the xc potential we find at time \(t\) is the same for any situation in which the true state Eq. 11 is reached and a KS Slater determinant is chosen. For the case of the He atom, the KS state is then a doubly-occupied orbital. The exact xc potential is then obtained from subtracting \(v_{\rm ext}=-2/|{\bf r}-{\bf r}^{\prime}|\) and \(v_{\rm n}({\bf r},t)\) from Eq. 3. Further, one can isolate the correlation component by noting that for this KS state, \(v_{\rm x}({\bf r},t)=-v_{\rm n}({\bf r},t)/2\).
Thus, finding the exact xc potential reduces to solving Eq. for \(\alpha({\bf r},t)\) and inserting it into Eq.. We note that for a different choice of initial KS state, e.g. using a two-configuration state that is more similar to that of the actual interacting state, the inversion to find \(v_{\rm xc}\) involves an iterative numerical procedure; some examples for the 1D analog of the He dynamics can be found in Refs. (see also discussion in Sec. III). This could be a more natural state to begin the KS calculation in some situations, e.g. if the state was prepared in such a superposition at the initial time. The importance of judiciously choosing the KS initial state when using an adiabatic approximation has been realized and exploited in strong-field charge-migration simulations.
### Numerical Details
We observe that the second equality in Eq. has the form of a Sturm-Liouville type equation, which yields a unique solution for \(\alpha({\bf r},t)\) for a given boundary condition. Furthermore, thanks to the azimuthal symmetry of our density (\(M=0\) at all times), we need solve this in effectively two dimensions, \((r,\theta)\).
\[\alpha({\bf r}\rightarrow\infty,t)=0\quad\mbox{and}\quad\frac{\partial}{\partial \theta}\alpha({\bf r},t)|_{\theta=\pi,0}=0\,. \tag{13}\]
We use a rectangular computational domain for the grid (\(r,\theta\)), extending from \(0\to R=30\) a.u. in \(r\) (the density is negligible this far from the nucleus) and \(0\rightarrow\pi\) in \(\theta\), and the boundary conditions, Eq.
\[\alpha(r=R,\theta,\varphi,t) = 0\] \[\frac{\partial}{\partial\theta}\alpha(r,\theta,\varphi,t)|_{\theta =\pi} = 0\quad\quad\frac{\partial}{\partial\theta}\alpha(r,\theta,\varphi,t)|_{ \theta=0}=0 \tag{14}\]
With the finite-difference scheme the matrix representation of the derivative operator is highly sparse, and consequently the semi-local operator \(\nabla\cdot(n({\bf r},t)\nabla)\) is also, coupling only a few nearby grid points. The dimensions of the matrix are \((N_{\theta}N_{r})\times(N_{\theta}N_{r})\) where \(N_{\theta}=50\) and \(N_{r}=301\) are the number of grid points in the (\(\theta,r\)) computational domain. The high sparsity also allows for efficient matrix inversion. Despite the computational efficiency, caution is required to avoid numerical inaccuracies especially where the density becomes small. To ensure that our conclusions are robust, we restrict our analysis to regions where the inversion is fairly accurate. This is ensured by checking that the action of the matrix representing \(\nabla\cdot n({\bf r},t)\nabla\) on the solution vector \(\alpha({\bf r},t)\) agrees with the right-hand-side of Eq..
In addition to the uniqueness entering the solution of Eq. as stated before, choosing the initial condition as \(\alpha({\bf r},0)=0\) fixes our initial state as \(\phi({\bf r},0)=\sqrt{n({\bf r},0)/2}\). The Runge-Gross theorem then ensures that there is a unique \(v_{\rm xC}({\bf r},t)\) that reproduces the exact \(n({\bf r},t)\) and yields a unique \(\alpha({\bf r},t)\) at later times. It is not _a priori_ obvious that the unique solution to Eq. with the boundary-condition Eq. applied with time \(t\) as a parameter is compatible with the TDKS evolution, but the results do evolve smoothly in time. Moreover, the rapid decay of the density at large \(r\) in the left-hand side of Eq. implies that this leads to the matrix elements corresponding to these points being killed off which renders the boundary conditions at \(r=R\) superfluous.
The numerical inversion of the matrix operator \(\nabla\cdot n({\bf r},t)\nabla\) subject to the boundary conditions Eq. 13 described above, produces the solution of Eq. for \(\alpha({\bf r},t)\). When used in Eq. this yields the KS potential, \(v_{\rm s}({\bf r},t)\).
\[\nabla^{2}v_{\rm H}({\bf r},t)=-4\pi n({\bf r},t)\,, \tag{15}\]
### Results
The problem possesses several symmetry features which we leverage to simplify the analysis. The azimuthal symmetry mentioned earlier and the fact that the states that enter the superposition are those for which \(L=0\) and \(L=1\), imply that the density, current, and potentials in the lower half-plane (\(\pi/2<\theta<\pi\)) evolve in exactly the same way as those in the upper half-plane (\(0<\theta<\pi/2\)) but a half-cycle out of phase, i.e. if \(O(r,\theta,t)\) represents the above quantities, then \(O(r,\pi-\theta,t)=O(r,\theta,t+T/2)\). Another consequence of the simple form of the superposition is the fact \(O({\bf r},T-t)=O({\bf r},t)\). It is therefore sufficient to look at time-snapshots only over a half cycle in one of the octants; the discussion will carry over to the other octants in accordance with the above conditions.
In we plot the snapshots of the xc potential, \(v_{\rm xC}({\bf r},t)\) at different fractions of the period of oscillation, \(T=2\pi/\omega=8.057\) a.u.in the octant spanned by \([r=]\times[\theta=(\pi/2,\pi)]\). We observe the xc-potential displays a prominent peak and a step across \(r\) at \(t=0\) in the region swept by \(\pi/2<\theta<\pi\) that decrease in magnitude over the first half-period time until vanishing from the octant and appearing on the other side of \(\theta=\pi/2\). These features have been shown to be completely missing in adiabatic approximations for 1D cases (e.g. Fig. 1) yet they often dominate the KS potential. Here we find these non-adiabatic features persist just as prominently, in analogous dynamics of the real three-dimensional He atom. This justifies that conclusions drawn from previous studies involving 1D systems do apply to real systems as well and that these strong correlation effects do not arise from dimensional reduction. We expect that the lack of the step and peak features in adiabatic approximations lead to less structured density profiles, as was seen in 1D cases.
It is important to mention here that at any instant of
Exact xc potential, \(v_{\rm xC}({\bf r},t)\) for field-free evolution of the 50:50 superposition state (\(a=1\)) of the He atom in the range \(\pi/2<\theta<\pi\) at times \(t=0,T/6,T/4\), and \(T/2\)
The flatness of the KS potential asymptotically in each octant is consistent with a zero current-density there. Our inversion to find the potential very close to \(\theta=\pi/2\) is not reliable near the density's local minimum along this line: Looking across the \(\theta-\)direction, we observe a sharp step and peak in \(v_{c}\) at \(\theta=\pi/2\) at large \(r\), which yields a force that prevents the KS current vectors from crossing the \(xy\)-plane, in a manner consistent with the behavior of the true current vectors. These features make a numerical inversion near \(\theta=\pi/2\) difficult. At \(\theta=\pi/2\), the density of the \(P\)-state \(\Psi_{1}\) vanishes, and the large and sharp change in the potential reminds us of the divergent behavior along the highest-occupied molecular orbital nodal plane found in the asymptotic region of the ground-state potential mentioned in Sec. III for cases where the density has the same decay in all directions. But in our case the density does decay differently along this plane than in other directions, and it cannot be captured by any adiabatic approximation.
As discussed in Sec. III, we find that a significant contribution to the peak is related to the local KS velocity, while the step when a cut is taken across a fixed \(\theta\) is related to the radial integral of the local acceleration, \(\dot{\alpha}(r,t)=\int^{r}\partial_{t}\mathbf{u}(\mathbf{r}^{\prime},t).\hat{ \mathbf{r}}^{\prime}dr^{\prime}\), and likewise for the step at \(\pi/2\) in the \(\theta\)-direction. We will explicitly demonstrate this in Sec. IV.4.
Taking different superpositions of the ground and excited states is further evidence that the step and peak features are universally present in real three-dimensional systems. The first and third columns of shows the KS and correlation potentials at the initial time, when \(a\) in Eq. 11 is changed through \(0,1,2,\infty\), going from a pure ground-state, through to the purely excited state; these two limits are of course time-independent. The correlation potential in the purely excited case (top panel, third row) has a barrier with a structure that is not dissimilar to the 1D case in magnitude and in shape (Figs. 3 and 4 in Ref.); the KS potential must be such to maintain the constant excited \({}^{1}P\) density at all times with a non-interacting doubly-occupied orbital. (For all cases, \(v_{\mathrm{s}}\) goes asymptotically like a constant \(-1/r\); the different scales makes this behavior less apparent in some cases.)
### "Non-Interacting Helium"
As observed in Sec. III, it is the kinetic component \(v_{\mathrm{C}}^{\mathrm{T}}\) that is largely responsible for the non-adiabatic step and peak features, and that even the best adiabatic approximation to this term, the adiabatically-exact, does not capture them. This component of the xc potential depends on spatial derivatives of the difference between the true and KS one-body density matrix (Eq.), and so is significantly affected by the choice of initial KS state, as discussed earlier. In fact, even considering the (unphysical) limit of zero electron-interaction in the He atom, \(v_{\mathrm{xc}}=v_{\mathrm{C}}^{\mathrm{T}}\) is non-zero when the initial KS state is chosen differently to that of this non-interacting atom. This raises the theoretical question of how do the steps and peaks compare in such a system to the physical situation, that is, how are these affected by electron-interaction itself, as opposed to configurational effects.
To this end, we consider here a "noninteracting He" atom, and prepare our system in a superposition of the ground state and an excited state such that the configuration is the same as that of the actual interacting He atom considered in the previous subsection, except for the fact that the wavefunctions that go into the superposition are products of the ordinary hydrogenic wavefunctions. That is, in Eq., instead of the interacting ground state, \(|\Psi_{0}\rangle=|1^{1}S_{0}\rangle\) and the first excited state
Comparison between interacting and non-interacting helium atom. Columns 1 & 3 correspond to the interacting case while columns 2 & 4 correspond to the non-interacting case. Top panel shows \(n(\mathbf{r},t)\) (dashed red), \(v_{\mathrm{s}}(\mathbf{r},t)\) (dashed blue) and correlation \(v_{\mathrm{C}}(\mathbf{r},t)\) (solid blue) for purely ground state (\(a=0\), columns 1 & 2) and purely excited state (\(a=\infty\), columns 3 & 4), at a cross section taken at \(\theta=0.75\pi\)\({}^{\mathrm{Noted}}\). Middle panel: \(n(\mathbf{r},t)\), \(v_{\mathrm{S}}(\mathbf{r},t)\) and \(v_{\mathrm{C}}(\mathbf{r},t)\) plotted for superposition states 50:50 (\(a=1\), columns 1 & 2) and 20:80 (\(a=2\), columns 3 & 4) at time \(t=0\). Lower panel: Same quantities as middle panel for \(t=T/8\).
\(|\Psi_{1}\rangle=|2^{1}P_{1}\rangle\), we prepare our system in the following superposition
\[|\Psi^{}(t)\rangle=\frac{1}{\sqrt{1+|a|^{2}}}\left(|\Psi^{}_{0}\rangle+ae^{ -i\omega t}|\Psi^{}_{1}\rangle\right) \tag{16}\]
where \(\Psi^{}_{0}(r_{1},r_{2})=\phi_{1}(r_{1})\phi_{1}(r_{2})\) and \(\Psi^{}_{1}=\frac{1}{\sqrt{2}}[\phi_{1}(r_{1})\phi_{2}(r_{2})+\phi_{1}(r_{ 2})\phi_{2}(r_{1})]\) are composed in terms of the products of hydrogenic wavefunctions \(\phi_{1}\) and \(\phi_{2}\)
\[\phi_{1}(r)=\sqrt{\frac{8}{\pi}}e^{-2r},\;\;\;\phi_{2}(r)=\frac{\cos\theta}{ \sqrt{\pi}}re^{-r} \tag{17}\]
Following the same steps as involved in the solution of Eq. we analyse the exact KS potential \(v_{\rm s}\) along with the exact correlation potential \(v_{\rm c}({\bf r},t)\) for this non-interacting version of the Helium atom. The second and fourth columns of Figure present these potentials for \(a=0,1,2\) and \(\infty\). Comparing with the first and third columns we see that they display structures that resemble the ones found in the interacting case. We find that the peaks and the steps follow a similar trend to that displayed by the interacting system when the proportions of ground and excited states are changed.
The appearance of such dominating steps in the correlation potential here is fundamentally linked to the difference in configurations of the interacting superposition state and KS Slater determinant rather than being a direct consequence of electron interaction. Tuning down the electron-interaction dampens the peak but the step remains. Both these features have a role in somehow nudging non-interacting electrons to be in separate lobes of the density, with the peak containing dynamical Coulomb interaction effects further enhancing the separation. An accurate approximation for the kinetic component \(v_{\rm c}^{\rm T}\) is required to capture these features.
Finally, we demonstrate with this non-interacting system, the decomposition of the exact KS potential in the second line of Eq.. In particular, the second term depends on the local velocity, and yields a peak structure, while the third term dependent on a spatial integral of the local acceleration is primarily responsible for the step. Having analytic expressions for the "true" wavefunction gives analytic expressions for \(n({\bf r},t)\) and \({\bf j}({\bf r},t)\), and so allow us calculate the local acceleration and velocity in a straightforward way. verifies the assertions above. We note that the first term in Eq., that depends on the curvature of the instantaneous density, also can yield peaks near the minimum of the density; in fact at the initial time, where the current-density is zero, the peak structure arises predominantly from the first term. The first term can also yield localized step-like structures, but it is only the third term that can yield a step structure that asymptotes to different constants away from a localized density. Interestingly, in this case, \({\bf j}_{\rm s}({\bf r},t)={\bf j}({\bf r},t)\), in contrast to the physical interacting He atom where snapshots of their difference can be found in Ref..
## V Conclusions and outlook
TDDFT tasks the potential driving the non-interacting electronic system with reproducing the exact time-dependent density of interacting electrons at all times. As a result, the exact time dependent xc potential must choreograph a rather unusual dance through a landscape of dynamical steps and peaks that nudge non-interacting electrons to evolve with the same density as Coulomb-interacting ones. In this work, we have reviewed aspects of what is known about these steps and demonstrated them on a real three-dimensional system, the He atom. They are distinct from step and peak features that arise in the ground-state case which tend to be associated with fractional charge and static correlation issues. The dynamical steps and peaks instead appear generically in dynamics far from the ground-state, associated with a kinetic component to the xc potential that is sensitive to the difference of the exact and KS one-body density matrices near the diagonal. They have a non-local dependence on the density in time, and are completely absent in any adiabatic approximation. They also have a non-local dependence on the density in space, as can be seen e.g. from the different asymptotes of the potential. The entanglement of spatial and time non-locality in TDDFT has been pointed out in extended systems in linear response, where approximations that build in frequency-dependence while remaining local in space violate the zero-force theorem, and here we see the time- and space- entanglement is a generic feature in non-perturbative dynamics of finite systems as well. The lack of these structures in approximations in use today lead to errors, including spurious peak-shifting that can muddle interpretation of pump-probe spectra.
(a) Decomposition of \(v_{\rm s}({\bf r},t)\): Terms 1-3 are the three terms in Eq. plotted as a function of \(\theta\) at r=2.0 a.u. (b) Same three terms now plotted across as a function of \(r\) at \(\theta=0.75\pi\). In both (a) and (b), the density, \(n(r,t)\) (dotted magenta) as a function of \(\theta\) is plotted on a different scale at the same r.
Going beyond the adiabatic approximations, functionals that have explicit dependence on the instantaneous orbitals incorporate memory (e.g. time-dependent exact exchange but the orbital-dependence of these commonly only involves exchange which is inadequate to capture these dynamical steps, as they appear in \(v_{c}^{\mathrm{T}}\). Building practical non-adiabatic approximations for these features have so far proven elusive, although the recent density-matrix coupled approximations show it may be possible. The hope is that such approximations would increase the reliability of real-time TDDFT for dynamics of electrons in non-perturbative fields. The challenge is certainly worth it.
|
10.48550/arXiv.2205.03691
|
The Exact Exchange-Correlation Potential in Time-Dependent Density Functional Theory: Choreographing Electrons with Steps and Peaks
|
Davood Dar, Lionel Lacombe, Neepa T. Maitra
| 3,624
|
10.48550_arXiv.1110.5423
|
###### Abstract
We consider the flow-driven translocation of single polymer chains through nanochannels. Using analytical calculations based on the de Gennes blob model and mesoscopic numerical simulations, we estimate the threshold flux for the translocation of chains of different number of monomers. The translocation of the chains is controlled by the competition between entropic and hydrodynamic effects, which set a critical penetration length for the chain before it can translocate through the channel. We demonstrate that the polymers show two different translocation regimes depending on how their length under confinement compares to the critical penetration length. For polymer chains longer than the threshold, the translocation process is insensitive to the number of monomers in the chain as predicted in Sakaeue _et al._, _Euro. Phys. Lett._, **72** 83. However, for chains shorter than the critical length we show that the translocation process is strongly dependent on the length of the chain. We discuss the possible relevance of our results to biological transport.
## I Introduction
The passage of polymer chains through nanochannels is an ubiquitous process in nature. Biopolymers, such as DNA and RNA, have to cross a multitude of barriers to perform different biological functions, for example, in translocating through cellular membrane pores or when ejecting from viral capsids. Considerable interest has arisen in the details of the translocation process due to a vast array of practical applications that include the potential sequencing of DNA chains, the sorting of biopolymers using smart entropic traps or in sieving processes in the pharmaceutical and food industries.
The translocation of a polymer chain through a nanochannel can be described as a three-stage process: the chain must first find the pore, then enter it and finally move through it. It is the second stage, of a polymer entering a long nanochannel, that we concentrate on here. Recent theoretical progress has shown that the entry of a polymer chain into a narrow channel driven by a fluid flow can be regarded as a tunnelling phenomenon, in which the entropic cost of squeezing the chain into the pore is opposed by the energy gain provided by the driving hydrodynamic force. The outcome of such competition is a free energy barrier, which the polymer has to surpass in order that translocation takes place.
According to the de Gennes blob model for confined chains, the barrier is overcome once the chain has been pushed a distance \(y^{*}\) into the channel, at which point, the hydrodynamic force wins over the entropic pressure. Therefore, the strength and position of the the free energy barrier can be controlled by varying the driving volumetric flux, \(J\). Increasing \(J\) has the effect of shifting the position of the barrier closer to the channel entrance, up to a critical flux, \(J_{c}\), at which the barrier height becomes comparable to the thermal energy and translocation takes place.
\[J_{c}\sim\frac{k_{B}T}{\eta}, \tag{1}\]
Remarkably, the threshold flux given by eq. is independent of the degree of polymerisation of the chain, \(N\). Physically, this occurs because the free energy barrier is overcome when a major portion of the chain is still outside the pore. This behaviour thus belongs to a long-chain regime, where the polymer always reaches the position of the barrier before being completely brought into the channel. Conversely, there is a short-chain regime, in which the whole chain is pushed inside the pore without reaching \(y^{*}\). In this case, the barrier still arises from the competition of the entropic pressure and the hydrodynamic force as before, but with the crucial difference that its magnitude and position are determined by the length of the confined chain. The practical consequence is that the translocation process becomes \(N\)-dependent, a feature of potential interest, for example, in sorting chains according to their number of monomers.
In this paper we shall focus on the existence of two regimes of translocation between long and short chains, and will examine the effect of the size of the chain on the threshold flux in each regime. Applying the energy barrier approach of Ref. we obtain a different scaling relation for the threshold translocation flux for long and short chains and test our predictions against numerical simulations.
The numerical modelling of polymer-solvent dynamics has made considerable progress during the last few years. In particular, coarse-grained models for the fluid that offer a physical coupling with polymer chains have been developed, leading to a reliable representation of the dynamics of the system whilst retaining an efficient numerical performance. Given that the large-\(N\) limit relevant to the blob model is difficult to access directly, we shall adopt a coarse-grained representation introduced by Dunweg _et al._ and described by Usta _et al._, which is based on a lattice-Boltzmann model for the fluid coupled to a bead-spring representation of the polymer chain through an effective Stokes drag. This approach has been used to study the lateral migration of chains in channel flows under external forces and pressure gradients. More recently, Markesteijn _et al._ used the same model to study the forced translocation of long chains into narrow pores in the limit where the bead-spring model is expected to give the blob-model behaviour. As expected, they confirmed the scaling of eq. showing that the coarse-grained model is indeed able to capture the main mechanisms at play in the translocation process.
The rest of this paper is organised as follows: in Section II we present scaling arguments for the position and height of the energy barrier, and for the threshold translocation flux in the long-chain and short-chain regimes. In Section III we describe the lattice-Boltzmann algorithm and the bead-spring model for the polymer chain, and list the set of parameters used to carry out the numerical simulations. Our numerical results are presented in Section IV. After describing the simulation setup in Section IV.1, in Section IV.2 we study the translocation process for long chains, demonstrating the presence of the energy barrier. We then focus on the cross-over to the short-chain regime in Section IV.3, and on the short-chain translocation process which gives rise to a length-dependent threshold flux, in Section IV.4. Finally, in Section V we present the discussion and conclusions of this work.
## II Entry of a polymer chain into a nanopore: blob model
In this section we present scaling arguments to predict the threshold flux allowing the translocation of linear polymer chains through a narrow channel. As we have anticipated above, we will examine two regimes for the translocation corresponding to long and short chains.
### Confined chains in equilibrium
We start by reviewing the scaling argument of Sakaue _et al._ for long, blob-like linear chains. Consider a linear polymer chain of ideal radius \(R_{0}\simeq aN^{1/2}\), composed of \(N\) monomers of size \(a\). In equilibrium, the Flory radius of the free chain is \(R\simeq aN^{3/5}\).
The size of the chain changes when it is confined in a channel of width \(D<R\). As depicted in Fig. 1, the chain stretches due to the effect of confinement.
\[\frac{F}{k_{B}T}\simeq\frac{L^{2}}{R_{0}^{2}}+\frac{N^{2}a^{3}}{LD^{2}}, \tag{2}\]
and reads,
\[L\simeq D\left(\frac{a}{D}\right)^{5/3}N\simeq D\left(\frac{R}{D}\right)^{5/3}. \tag{3}\]
In the de Gennes' blob picture, the confined chain accommodates itself into \(M\) blobs of uniform size \(\xi\). The number of blobs then obeys the relation \(M=LD^{2}/\xi^{3}\). Within each blob the effect of confinement is unimportant. Hence, the size of the blob scales as \(\xi\simeq aP^{3/5}\), where \(P\) is the number of monomers in each blob.
\[\phi=\frac{Na^{3}}{LD^{2}}=\frac{Pa^{3}}{\xi^{3}},\]
from which it follows that for a linear chain, the size of the blob is comparable to the pore size, _i.e.,_
\[\xi\simeq D. \tag{4}\]
### Long-chain translocation
We now consider a partly confined chain in the presence of a driving flow. For a weak driving flux, the conformation of the chain outside the channel is close to equilibrium. Hence, the translocation process is controlled by the forces acting on the confined part of the chain. This corresponds to the inside approach in the terminology of Sakaue _et al._ and is valid as long as the length of the pore, \(L_{p}\), is larger than its thickness, \(D\).
Schematic representation of a long blob-like polymer chain under confinement. The chain is confined up to a distance \(y\), adopting a configuration of \(M\) stacked blobs of size \(\xi\). For a linear chain, the size of a blob scales as \(\xi\sim D\). When subject to a driving flux, \(J\), the competition between the hydrodynamic drag and the entropic pressure sets a barrier to translocation, located at \(y^{*}\). Once it has reached the position of the barrier the polymer is able to translocate.
The confinement of the chain has an entropic penalty of the order of \(k_{B}T\) per blob. Thus, for a chain composed of \(M\) blobs which penetrates a distance \(y\) into the channel, as shown in Fig. 1, there is an energy cost
\[F_{S}\simeq k_{B}T\frac{yD^{2}}{\xi^{3}}=Ak_{B}T\left(\frac{y}{D}\right), \tag{5}\]
As expected, the entropic cost increases with \(y\), given that a larger number of monomers are pushed into the channel.
Countering the entropic cost is the fluid flow, which tends to drag the chain further into the channel. The hydrodynamic drag per blob scales as \(\eta u\xi\), where \(u=J/D^{2}\) is the typical velocity in the translocation direction inside the channel. For \(M\) blobs, the change in free energy corresponds to the work done by the fluid to displace the chain up to a distance equal to \(y\),
\[F_{H}\simeq-\eta u\int_{0}^{y}M(y^{\prime})\xi(y^{\prime})\mathrm{d}y^{\prime} =-\frac{B\eta J}{2}\left(\frac{y}{D}\right)^{2}, \tag{6}\]
As before, we introduce a proportionality constant, \(B\).
Adding the contributions given by eqs.
\[\frac{\Delta F}{k_{B}T}(y)=F_{S}+F_{H}=Ak_{B}T\left(\frac{y}{D}\right)-\frac{B \eta J}{2}\left(\frac{y}{D}\right)^{2}. \tag{7}\]
The competition between the entropic and hydrodynamic terms gives an energy barrier, \(\Delta F^{*}\), located at \(y=y^{*}\). Differentiation of eq.
\[y^{*}=\left(\frac{A}{B}\right)\frac{k_{B}T}{\eta J}D, \tag{8}\]
to obtain its magnitude,
\[\frac{\Delta F^{*}}{k_{B}T}=\frac{A}{2}\left(\frac{y^{*}}{D}\right)=\frac{C}{2 }\frac{k_{B}T}{\eta J}, \tag{9}\]
In order that translocation proceeds, the chain must overcome this barrier.
\[P=\kappa_{D}\tau_{m}\exp{\left(-\frac{\Delta F^{*}}{k_{B}T}\right)}, \tag{10}\]
We define the threshold flux by
\[P(J_{c})=P_{c}, \tag{11}\]
Inverting eq.
\[J_{c}\sim\frac{k_{B}T}{\eta}.\]
As we had anticipated, \(J_{c}\) does not depend on \(N\). This is because the polymer reaches the position of the barrier, \(y^{*}\), before it is completely confined in the channel, _i.e._, the translocation is triggered when \(N^{*}<N\) monomers have been pushed into the channel. To obtain the critical number of monomers, \(N^{*}\), it suffices to replace \(L\) by \(y^{*}\) in eq., from which
\[N^{*}\simeq\frac{y^{*}}{D}\left(\frac{D}{a}\right)^{5/3}\simeq\left(\frac{A}{B }\right)\left(\frac{D}{a}\right)^{5/3}\frac{k_{B}T}{\eta J}, \tag{12}\]
We note that in Ref. the scaling for the threshold flux was obtained by setting \(\Delta F^{*}\simeq k_{B}T\) in eq.. This corresponds to a free energy barrier whose height is overcome by pushing only one blob into the pore, with a corresponding number of monomers \(N_{D}\equiv N^{*}(y^{*}=D)\simeq(D/a)^{5/3}\).
### Short-chain translocation
The short-chain regime corresponds to the limit \(L<y^{*}\), or equivalently, to \(N<N^{*}\), where the chain is completely confined _before_ overcoming the energy barrier given by eq.. From eqs. and, for a given chain length, this corresponds to weak fluxes and/or wide pores. Instead of being located at \(y^{*}\), the energy barrier corresponds to a penetration \(y=L\), given that there is no extra cost for pushing the chain into the channel any further than its own length. In terms of the number of beads, the free energy barrier obeys \(\Delta F^{*}=\Delta F(L(N))\), and follows from eq.,
\[\frac{\Delta F^{*}}{k_{B}T}=A\left(\frac{a}{D}\right)^{5/3}\left(1-\frac{1}{2 }\frac{N}{N^{*}}\right)N, \tag{13}\]
and to write the dependence on \(N\) and \(N^{*}\) explicitly. According to this result, for a given constant value of the flux, the barrier can be reached more easily by decreasing the number of beads. Conversely, increasing \(N\) has the effect of increasing the strength of the barrier, up to \(N=N^{*}\), where one crosses over to the long-chain regime, and eq. reduces to eq. as expected.
The threshold flux can be calculated from the probability of translocation, which follows after combining eqs. and. Using the criterion given by eq., we obtain
\[\frac{\eta J_{c}}{k_{B}T} =2\left(\frac{A}{B}\right)\left(\frac{D}{a}\right)^{10/3} \tag{14}\] \[\times\left(\frac{\left(\frac{a}{D}\right)^{5/3}N+\frac{\log(P_{c })}{A}-\frac{\log(\kappa_{D}\tau_{m})}{A}}{N^{2}}\right).\]The scaling should be valid in the range \(N_{D}<N<N^{*}\), which corresponds to \(D<R\) and \(L<y^{*}\). In this range, the threshold flux increases monotonically with the number of monomers. This behaviour can be traced back to the \(N\)-dependent terms in eq., where the linear term, corresponding to the entropic cost, grows faster than the hydrodynamic gain, provided that \(N/N^{*}<1\).
## III Numerical method
To further study the cross-over from long- to short-chain translocation discussed in Section II, we use the numerical method introduced in Ref.. This is a hybrid scheme that couples a lattice-Boltzmann fluid to a bead-spring model for the polymer chain. Our work is based on the original implementation of the code, _susp3d_, which was kindly provided by the developers. Here we present the main features of the numerical method. For a more detailed description, the reader is referred to the original papers.
### Lattice Boltzmann Method
In the lattice-Boltzmann algorithm the fluid dynamics follows from the evolution of the particle velocity distribution function \(f_{i}\), which is defined for a discretised velocity set \(\{\vec{c}_{i}\}\). For a given velocity vector, \(f_{i}\) is proportional to the average number of particles moving in the direction of \(\vec{c}_{i}\).
The dynamics of \(f_{i}\) is given by the lattice-Boltzmann equation,
\[f_{i}(\vec{r}+\vec{c}_{i}\Delta t,t+\Delta t)-f_{i}(\vec{r},t)=\Delta_{i}(\vec {r},t)+F_{i}^{\rm ext}, \tag{15}\]
Here we use the D3Q19 model, which consists of a cubic lattice with a set of nineteen velocity vectors in three dimensions. The lattice spacing, \(\Delta x\), is uniform along the lattice axes. The model has three possible magnitudes of the velocity vectors, \(\{|\vec{c}_{i}|\}=(0,1,\sqrt{2})\Delta x/\Delta t\). Accordingly, the \(f_{i}\) have a corresponding weight \(a_{c_{i}}\) that satisfies the condition \(\sum_{i}a_{c_{i}}=1\). A suitable choice of the weights for the lattice model used here is \(a_{0}=1/3\), \(a_{1}=1/18\) and \(a_{\sqrt{2}}=1/36\).
The dynamics expressed by eq. is composed of two steps. First, the distribution function undergoes a collision step, where fluid particles exchange momentum according to the collision operator, \(\Delta_{i}\), and are driven by the term \(F_{i}^{\rm ext}\), which plays the role of a body force. Following this collision stage, the \(f_{i}\) are propagated to neighbouring sites in a streaming step, corresponding to the left-hand side of the equation.
The mapping between the lattice-Boltzmann scheme and the hydrodynamic equations follows from the definition of the hydrodynamic variables as moments of the \(f_{i}\).
\[\sum_{i}f_{i}=\rho,\quad\sum_{i}f_{i}\vec{c}_{i}=\rho\vec{v},\quad\mbox{and} \quad\sum_{i}f_{i}\vec{c}_{i}\vec{c}_{i}=\Pi, \tag{16}\]
In the absence of external forces, the system relaxes towards equilibrium through the collision stage in eq.. For the lattice-Boltzmann model used in this paper, this is done by defining the post-collision distribution function, \(f_{i}^{*}=f_{i}+\Delta_{i}\), which can be expressed as an expansion in the fluid velocity:
\[f_{i}^{*}=a_{c_{i}}\left(\rho+\frac{\rho\vec{v}\cdot\vec{c}_{i}}{c_{s}^{2}}+ \frac{(\rho\vec{v}\vec{v}+\Pi^{\rm neq,*}):(\vec{c}_{i}\vec{c}_{i}-c_{s}^{2} \mathbf{1})}{2c_{s}^{4}}\right), \tag{17}\]
Similarly, the forcing term, \(F_{i}^{\rm ext}\), can be expanded in powers of the fluid velocity (cf. Ref.).
The post-collisional momentum flux tensor, \(\Pi^{\rm neq,*}\), describes the relaxation towards the equilibrium momentum flux tensor, \(\Pi^{\rm eq}\), according to
\[\Pi^{\rm neq,*}=(1+\lambda)\bar{\Pi}^{\rm neq}+\frac{1}{3}(1+\lambda_{\nu})( \Pi^{\rm neq}:\mathbf{1})\mathbf{1}, \tag{18}\]
The parameters \(\lambda\) and \(\lambda_{\nu}\) characterise the relaxation timescales of the lattice-Boltzmann fluid. By performing a Chapman-Enskog expansion of eq., corresponding to the limit of long length and timescales compared to the lattice spacing and the relaxation timescale of the fluid, the lattice-Boltzmann scheme leads to the continuity equation for the fluid density and the Navier-Stokes equations with second order corrections in the velocity for the fluid momentum.
\[\eta=-\rho c_{s}^{2}\Delta t\left(\frac{1}{\lambda}+\frac{1}{2}\right), \tag{19}\]
and
\[\eta_{\nu}=-\frac{2\rho c_{s}^{2}}{3}\Delta t\left(\frac{1}{\lambda}+\frac{1} {2}\right), \tag{20}\]
Solid boundaries in the simulation box are implemented using the well-known bounce-back rules. These correspond to a reflection of any distribution function propagating to a solid node back to the fluid node it came from at the streaming stage in eq.. As a consequence, a stick condition for the velocity is recovered approximately halfway between the fluid node and the solid node.
### Polymer chain
The linear polymer is modelled by a chain composed of \(N\) beads joined by freely rotating bonds.
\[U_{\rm el}(r)=k(r-a)^{2}, \tag{21}\]
The short-range excluded-volume interactions are modelled by a truncated DLVO potential,
\[U_{\rm DLVO}=U_{0}\frac{\exp(-\kappa_{\rm DH}r)}{r}, \tag{22}\]
The position vector of the \(i\)-th polymer bead, \(\vec{x}_{i}\), evolves in time according to
\[m\ddot{\vec{x}}_{i}=-\sum_{j\neq i}\vec{\nabla}_{ij}U+\vec{F}_{i}, \tag{23}\]
The term \(\vec{F}_{i}\) contains the viscous hydrodynamic force that couples the bead to the lattice-Boltzmann fluid,
\[\vec{F}_{i}=-\xi_{0}(\vec{x}_{i}-\vec{v}(\vec{x}_{i}))+\vec{F}_{i}^{\rm r}. \tag{24}\]
This expression includes the Stokes drag, \(\xi_{0}=6\pi\eta r_{\rm H}\), with \(r_{H}\) being the hydrodynamic radius of the beads, and a random force \(F_{i}^{\rm r}\) that satisfies the fluctuation-dissipation relation,
\[\langle\vec{F}_{i}^{\rm r}(t)\vec{F}_{i}^{\rm r}(t^{\prime})\rangle=2k_{B}T\xi _{0}\delta(t-t^{\prime})\mathbf{1}. \tag{25}\]
While the polymer beads move in continuous space, the lattice-Boltzmann fluid is only defined at the lattice nodes. In order to calculate the coupling force, \(\vec{F}_{i}\), the model uses a linear interpolation scheme to estimate the fluid velocity, \(\vec{v}\), at the position of the bead, \(\vec{x}_{i}\). The discretisation of the lattice gives an _effective_ hydrodynamic radius, \(r_{\rm H}^{\rm eff}\), which differs from the input hydrodynamic radius, \(r_{\rm H}\). Here we follow the same procedure as in Ref., and choose \(r_{\rm H}\) in order to obtain the desired value of \(r_{\rm H}^{\rm eff}\). Once the hydrodynamic force has been exerted on the monomer, momentum conservation is enforced by exerting a force of equal magnitude back onto the fluid.
### Parameter Values
Our objetive is to carry out numerical simulations of polymer chains in the limit where the bead-spring model gives the blob-model behaviour presented in Section II. Such regime has been validated when the monomer size matches the blob size, \(a\simeq\xi\), for linear chains, where \(\xi\simeq D\). Therefore, the bead-spring model should give the blob-limit behaviour for \(a\simeq D\). On the other hand, the hydrodynamic radius of the beads, \(r_{\rm H}^{\rm eff}\), is limited to small values, corresponding to the point-particle coupling between the chain and the lattice-Boltzmann fluid. With these conditions in mind, we fix the bead diameter to \(2r_{\rm H}^{\rm eff}=\Delta x/2\), while \(a=\Delta x\) and \(D=2\Delta x\).
The remaining model parameters are chosen as follows: we work at a fixed temperature, \(k_{B}T=0.1\). To prevent chain crossings the spring constant is taken as \(k=300k_{B}T/\Delta x^{2}\). Parameter values for the excluded-volume potential are fixed to \(U_{0}=k_{B}T\Delta x\) and \(\kappa_{\rm DH}=80/\Delta x\), which ensure that inter-monomer repulsions are larger than \(k_{B}T\) at distances comparable to \(r_{\rm H}^{\rm eff}\). To match the hydrodynamic diameter to the effective bead diameter, we fix \(r_{\rm H}=0.32\Delta x\) following the calibration procedure presented in Ref..
In order to resolve the hydrodynamics correctly, one needs to ensure that the distribution function relaxes on a faster timescale than the diffusive timescale of the polymer chain. This condition can be satisfied by setting the parameters in the collision operator in eq. to \(\lambda=-1\) and \(\lambda_{\nu}=-1\). We set the fluid density to \(\rho=36\), and the timestep and lattice spacing in the lattice-Boltzmann fluid to \(\Delta t=1\) and \(\Delta x=1\). Using these values, the dynamic viscosity in simulation units is \(\eta=6\).
## IV Numerical results
### System Geometry and Initial Conditions
The geometry of the system is depicted in We consider two identical rectangular ducts of dimensions \(L_{x}=108\), \(L_{y}=14\) and \(L_{z}=10\), separated by a square pore of width \(D=2\) and length \(L_{p}=24\). Solid walls are imposed in the \(y\) and \(z\) directions, while periodic boundary conditions are enforced in the \(x\) direction.
For small Reynolds numbers, one expects that the volumetric flux, \(J\), caused by applying a uniform body force
Schematic representation of the simulation box.
to the fluid, \(f\), obeys Darcy's Law:
\[J=\frac{kS}{\eta}f, \tag{26}\]
In order to verify the validity of eq., we measure the local flux for different applied forcings. The results, illustrated in Fig. 3, show a linear dependence of \(J\) on \(f\) as expected. A linear fit to the data gives \(kS\simeq 238\). For a given forcing, we find that the flux does not vary significantly whether it is measured inside or outside the pore thus confirming that, for the range of forcings considered here, the lattice-Boltzmann fluid behaves as an incompressible liquid. Given that we are interested in the scaling of the critical flux, we use values of the forcing that give a ratio between hydrodynamic and thermal effects in the range \(0<\eta J/k_{B}T<1.5\).
We initially tether the polymer chain at a position \(y=y_{0}\) inside the channel as depicted in the inset of Fig. 4, and let the system equilibrate for 5\(\times 10^{4}\) time steps. The equilibration period allows the relaxation of the chain. Evidence for this is presented in the same figure, which shows the number of beads inside the pore, \(N_{y_{0}}\), as a function of \(y_{0}\) for \(N=128\). For a confined chain in equilibrium, we have, from eq., \(N_{y_{0}}\simeq(y_{0}/D)(D/a)^{5/3}\). Our results show that the number of beads increases linearly with the tethering position, as expected.
### Energy Barrier
In order to confirm that the polymer translocation is controlled by surpassing an energy barrier, we first perform simulations of long polymer chains subject to a fixed driving flux while varying the initial tethering position, \(y_{0}\). We consider different values of \(y_{0}\) in the range \(0<y_{0}<L_{p}\), and set the number of monomers to \(N=128\), thereby ensuring that the major part of the chain lies outside the pore.
Simulations are carried out by letting the chain equilibrate as before. Once the chain has equilibrated, it is released from its tethering point and the body force is applied uniformly to the fluid. The chain then is either carried down the pore by the underlying fluid flow and eventually translocates to the opposite duct, or is ejected from the pore back into the original chamber. Simulations are run for \(5\times 10^{5}\) timesteps, which is a sufficiently long timescale to identify successful or failed translocation events.
We observe a clear transition from non-translocating (\(P=0\)) to translocating (\(P=1\)) chains as \(y_{0}\) is increased. This confirms the presence of an energy barrier, which is progressively approached as the chain is pushed further into the channel. The probability curves are shifted to the left as one increases the imposed flux, \(J\). This indicates a shift of the position of the barrier, \(y^{*}\), closer to the pore entrance caused by a higher driving hydrodynamic force.
As a criterion we define the position of the barrier as \(P(y^{*})\approx 1\), and plot the measured values of \(y^{*}\) as a function of \(k_{B}T/\eta J\) in the inset of As expected from eq., we observe a linear growth. A fit of the data shown in the figure to the function \(y^{*}(x)=(A/B)Dx\) gives an estimate of the numerical prefactor in eq., \(A/B\simeq 4.4\).
The imposed flux sets the position and height of the energy barrier. Given that we fix the chain to a prescribed position inside the channel, the height of the barrier is reduced by an amount determined by \(y_{0}\).
Volumetric flux, \(J\), as a function of the body force, showing that Darcy’s law holds.
Number of beads in the confined part of the chain as a function of the initial tethering position. Inset: schematic representation of the tethered chain.
\(y-y_{0}\) in eq. and performing the integration in eq. from \(y^{\prime}=y_{0}\) to \(y^{\prime}=y\).
\[\frac{\Delta F^{*}}{k_{B}T}=\frac{A}{2}\left(\frac{y^{*}}{D}\right)\left(1+g \left(\frac{y_{0}}{y^{*}}\right)\right). \tag{27}\]
we can calculate the probability of translocation. This predicts \(\log P\sim\Delta F^{*}/k_{B}T\) as observed in As expected, all data points collapse onto the same curve, which is linear in the free energy. A fit to the data gives an estimate for the numerical prefactor in eq., \(A\simeq 1.8\), which together with the measured value of \(A/B\), gives the amplitude of the hydrodynamic contribution to the free energy in eq., \(B\simeq 0.41\).
The inset in shows how the theoretical prediction accurately captures the main features of the translocation process. The probability of translocation increases to unity as \(y_{0}/y^{*}\to 1\). The spread of the curves is dictated by the \(\sim k_{B}T/\eta J\) prefactor in eq., increasing the likelihood of translocation at a given \(y_{0}/y^{*}\) for larger fluxes, or lower temperatures.
### Cross-over from long to short chains
So far we have discussed the translocation of long chains, which can reach the position of the barrier while a significant amount of monomers remain outside the pore. The threshold flux is thus controlled only by the number of monomers that it takes to reach the barrier, \(N^{*}\), and not by the total number of monomers in the chain, \(N\).
This picture breaks down for shorter chains, where our theory predicts that the scaling of the critical flux becomes dependent on \(N\). In order to explore this effect, we have carried out simulations at a fixed initial tethering position, \(y_{0}\), and driving flux, while decreasing the total number of beads in the chain, \(N\). We fix the initial tethering position to \(y_{0}=L_{p}/2=12\), for which the number of beads inside the pore after equilibration is \(N_{y_{0}}\simeq 18\). We carry out the simulations in the range \(16<N<64\), where we expect to observe the cross-over.
The probability always increases with the applied flux as expected.
Logarithm of the translocation probability as a function of the free energy barrier magnitude. Symbols correspond to numerical simulations. The solid line is a linear fit to the data. Inset: translocation probability as a function of \(y_{0}/y^{*}\). Numerical results (symbols) show a good agreement with the theoretical prediction (solid lines).
Probability of translocation as a function of the number of beads in the chain at different values of the imposed flux. Simulation results correspond to symbols, while the theoretical model is indicated by the solid lines where we set \(N_{y_{0}}=20\) and \(A=1.8\).
Probability of translocation as a function of \(y_{0}\). Inset: linear growth of the position of the energy barrier with \(k_{B}T/\eta J\).
For smaller \(N\) the probability of translocation increases strongly with decreasing \(N\) until it saturates at \(N\simeq 20\) to \(P\to 1\), where all polymer chains translocate even under the weakest flow.
This behaviour is in agreement with our theoretical model. For large \(N\), the critical number of beads to overcome the barrier follows from eq., and obeys \(N^{*}\sim(D/a)^{5/3}(k_{B}T/\eta J).\) Therefore, the cross-over to the short-chain regime (\(N<N^{*}\)) is observed at larger chain lengths as \(J\) is decreased. From eq. we also have \(N^{*}\sim y^{*}\), from which it is possible to estimate the critical number of beads corresponding to each driving flux by interpolating the data shown in Once each value of \(N^{*}\) is known, we can determine the energy barrier as a function of \(N\) from eq.. In order to include the effect of \(y_{0}\), we follow the same procedure as that used to obtain eq..
\[\frac{\Delta F^{*}}{k_{B}T}=A\left(\frac{a}{D}\right)^{5/3}\left[N\left(1- \frac{1}{2}\frac{N}{N^{*}}\right)+\frac{1}{2}N^{*}g\left(\frac{N_{y_{0}}}{N^ {*}}\right)\right]. \tag{28}\]
Using this expression for the free energy barrier an estimate for the probability of translocation follows from eq.. The observed saturation length \(N\simeq 20\) is indeed in good agreement with our estimate \(N_{y_{0}}\simeq 18\), at which the free energy barrier given by eq. vanishes. As shown in Fig. 7, we obtain a good agreement between the theory and simulations.
### Critical flux for small chains
We now turn our attention to the dependence of the critical flux, \(J_{c}\), on the number of monomers in the short-chain regime. To find the threshold flux, we carry out simulations of the translocation process at a fixed number of beads while increasing the flux. We fix \(y_{0}=12\) as before, and consider the range \(0<\eta J/k_{B}T<1\), for which the probability curves cross over from complete rejection of the chains to complete translocation.
For \(N>32\) the \(N\)-independence of the long-chain regime is recovered, as curves fall on top of each other. Our results in the long-chain regime are consistent with recent experimental studies of polymer translocation in nanochannels carried out by Beguin _et al._. In their experiments they measure the rejection coefficient of chain translocation, which is related to the probability of translocation by \({\cal R}\sim 1-P\). Their experiments give the same smooth transition from rejection to translocation of the chains in the range \(0<\eta J/k_{B}T<1\), very close to our simulation results. For \(N<32\) the probability curves shift upwards and become increasingly plateau-like as the number of monomers is decreased, indicating the cross-over to the short-chain regime.
In order to quantify the cross-over from the short- to the long-chain regime, we use the criterion given by eq. to interpolate \(J_{c}\) from each of the curves shown in The threshold flux is shown in as a function of \(N\) for six different values of the threshold probability. As expected, the overall behaviour does not depend on the particular choice of \(P_{c}\). The threshold flux saturates for large \(N\), corresponding to the long-chain regime, and shows a marked decrease as \(N\to N_{y_{0}}\), where the chain becomes completely confined inside the channel.
Using eqs., and, we can estimate the threshold flux as a function of the number of monomers in the chain,
\[\frac{\eta J_{c}}{k_{B}T} \simeq 2\frac{A}{B}\left(\frac{D}{a}\right)^{10/3}\] \[\times\left(\frac{\left(\frac{a}{D}\right)^{5/3}(N-N_{y_{0}})+ \frac{\log(P_{c})}{A}-\frac{\log(\kappa_{D}T_{m})}{A}}{N^{2}-N_{y_{0}}^{2}} \right).\]
Both prefactors, \(A\) and \(B\), depend on the intrinsic properties of the chain and not on the particular translocation regime. We therefore set them to the values that we have measured previously.
Remarkably, we find a good quantitative agreement by using no new fit parameters in our theoretical prediction.
## V Summary and discussion
The results presented in this paper show that the translocation of polymer chains through narrow pores exhibits two different regimes depending on the length of the chain. As previously proposed in Ref., our numerical results show that the translocation process is controlled by overcoming a free energy barrier characterised by a critical penetration of the polymer inside the pore, \(y^{*}\). For long chains, this length scale always outruns the polymer length under confinement, \(L\).
Probability of translocation as a function of the imposed flux for different number of monomers in the chain.
Conversely, for small chains, where the confined length of the polymer is smaller than \(y^{*}\), we have shown that the translocation process is controlled by \(L\), and the process becomes \(N\)-dependent. A qualitatively similar conclusion should apply to the passage of the chains through two dimensional narrow slits.
For even smaller chains, where \(L\) is comparable to the size of the pore, \(D\), and for very small fluxes (\(\eta J/k_{B}T\simeq 10^{-1}\)), we have observed a reduction of the probability of translocation of the chains. This is caused by the effect of diffusion of the chain inside the pore, which can favour the ejection of small chains at very low forcings. However, this regime falls out of the free energy barrier picture presented in this paper, and we therefore leave it for future exploration.
An application of the dependence of the translocation probability of short chains on \(N\) is the potential sorting of the chains according to their number of monomers. In the short-chain regime the polymer must be completely pushed in before translocating through the pore. Given that only a small number of blobs can be pushed into the pore by fluctuations, one expects that only very short chains translocate by virtue of thermal effects in this regime. For longer chains, but still in the short-chain regime, the translocation must be assisted, for example, by a molecular motor that is able to push the chain deeper into the pore.
In an aqueous solution at room temperature, the critical flux in the long chain regime follows the scaling of eq. and is of order \(J_{c}\sim 10^{-18}\) m\({}^{3}\) s\({}^{-1}\). For a micrometre sized channel the corresponding velocity, \(v_{c}\simeq J_{c}/D^{2}\), is of order \(10^{-6}\) m s\({}^{-1}\). Such a small value suggests that chains can easily translocate through microfluidic chambers at normal operation velocities, which are typically of centimetres to metres per second. However, recent experimental measurements of cytoplasmic streaming in micrometre cell channels show that the streaming velocities are in the range of 10 \(\mu\)m s\({}^{-1}\). For weaker fluxes than \(J_{c}\), one enters the short-chain translocation regime. Therefore, it is feasible that the translocation regime proposed in this paper might be at play in biological systems as a regulator of selective translocation.
Here we have considered polymers in a good solvent using the DLVO potential to implement short-ranged excluded-volume interactions with the wall. Understanding the details of electrostatic effects, with regards both to the interaction with the wall and the interactions between segments, and how they are affected by the salt concentration, is an important issue for future work, relevant, in particular, to the dynamics of DNA in practical applications.
Finally, we comment on the feasibility of an experimental confirmation of our prediction. Beguin _et at._ have recently performed an experimental study of the flow-driven translocation of hydrosoluble polymers into nanopores. In their experiments, they consider chains whose radius of gyration is \(R\approx 84\) nm, and which are forced across pores \(23-35\) nm in radius and 6.5 \(\mu\)m in length. Their results for the rejection coefficient, \({\cal R}=1-\exp{(\Delta F^{*}/k_{B}T)}\), show a good agreement with the long-chain limit of the de Gennes model, where they use a similar expression to eq.. According to our prediction, the short chain regime would be observable for chains whose length under confinement is smaller than the position of the barrier. This traduces into a cross-over radius of gyration \(R^{*}\simeq D(k_{B}T/\eta J)^{3/5}\). Taking \(D=70\) nm and \(\eta J/k_{B}T\approx 0.8\), which correspond to experimental conditions reported in ref., we estimate \(R^{*}\approx 80\) nm, which is close to the radius of gyration of the polymers used in the experiments. This supports the feasibility of future experimental work to verify the theoretical predictions presented in this paper.
|
10.48550/arXiv.1110.5423
|
Length-dependent translocation of polymers through nanochannels
|
Rodrigo Ledesma-Aguilar, Takahiro Sakaue, Julia M. Yeomans
| 4,204
|
10.48550_arXiv.2305.08835
|
###### Abstract
We describe a multiple electronic state adaptation of the mapping approach to surface hopping introduced recently by Mannouch and Richardson (J. Chem. Phys. 158, 104111). Our modification treats populations and coherences on an equal footing and is guaranteed to give populations in any electronic basis that tend to the correct quantum-classical equilibrium values in the long-time limit (assuming ergodicity). We demonstrate its accuracy by comparison with exact benchmark results for three- and seven-state models of the Fenna-Matthews-Olson complex, obtaining electronic populations and coherences that are significantly more accurate than those of fewest switches surface hopping and at least as good as those of any other semiclassical method we are aware of. Since these results were obtained by adapting the scheme of Mannouch and Richardson, we go on to compare our results with theirs for a variety of problems with two electronic states. We find that their method is sometimes more accurate, and especially so in the Marcus inverted regime. However, in other situations the accuracies are comparable, and since our scheme can be used with multiple electronic states it can be applied to a wider variety of electronically nonadiabatic systems.
## I Introduction
Fewest switches surface hopping (FSSH) is the dominant method for simulating electronically nonadiabatic dynamics in chemistry. Originally proposed by Tully in the early 1990's,Tully it has become a standard tool for running _ab initio_ trajectories and is now routinely used to study various aspects of photochemical dynamics including both energy and charge transfer processes.Tully; Tully; Tully and Zubairy; Tully It remains popular for three (good) reasons: (i) it is easy to use; (ii) it is robust when used with _ab initio_ potentials; and (iii) it has been found to work reasonably well in many applications. However, while there have been attempts to derive it from first principles,Tully and Zubairy FSSH is still widely considered to be an _ad hoc_ algorithm. It also suffers from a well-known 'overcoherence' problem that has led to the development of a number of different 'decoherence' corrections,Tully and Zubairy none of which has become universally accepted.
A second school of thought advocates treating the electronic and nuclear degrees of freedom on an equal footing by morphing the multi-state electronic system onto a classical phase space. The simplest approach of this type is (multiple-trajectory) Ehrenfest dynamics, which is based on a mean-field coupling of the nuclear and electronic motions. However, this is known to produce various unphysical results including an overheating of the electronic subsystem.Tully and Zubairy Semiclassical mapping methods represent the multiple electronic states as harmonic oscillatorsStein; Manolopoulos or generalized spins,Tully and Zubairy and use either initial phase space sampling or symmetrical quasi-classical binningTully and Zubairy; Manolopoulos to calculate the dynamics. These mapping methods typically give better accuracy than either FSSH or Ehrenfest dynamics in applications to system-bath models of condensed-phase systems. However, they do so at the cost of allowing the electronic-state populations to become negative, which can result in the nuclei evolving (and potentially even diverging) on inverted potentials. While this problem may not be visible in simple models, it does help to explain why mapping methods are only rarely combined with _ab initio_ potentials and applied to realistic chemical systems.
Given the complementary strengths and weaknesses of surface hopping and phase-space mapping, it is natural to ask whether they could be combined to give a better method. Very recently, Mannouch and Richardson have explored this idea and used it to develop a new'mapping approach to surface hopping' (MASH).Mansolopoulos As in surface hopping, this method uses trajectories that hop between the physical adiabats, thereby eliminating concerns about dynamics on inverted mapped potentials. But rather than employing stochastic hops, the active surface is obtained deterministically from the electronic wavefunction, as it is in the phase-space mapping approach. This has several distinct advantages, not least of which is that it avoids the need for heuristic 'decoherence' corrections. The results can instead be improved (if necessary) through a careful resampling of the electronic wavefunction (the quantum-jump procedureStein). Even without this resampling, MASH has been found to give better results than either pure surface hopping or pure mapping for a wide range of two-state problems, with no more computational effort.Mansolopoulos It is thus arguably the most promising method that has yet been proposed for nonadiabatic dynamics.
However, the formulation of MASH that Mannouch and Richardson presented was restricted to two coupled electronic states.Mansolopoulos It is not obvious how to extend it to more states because they based their formulation on correlation functions involving specific two-state prescriptions for the electronic populations and coherences that were specialized to the adiabatic basis (see Sec II). A general nonadiabatic dynamics method should be applicable to an arbitrary number of electronic states, it should treat populations and coherences on an equal footing, it should provide results that transform correctly underthe unitary rotations that take one electronic basis to another, and it should be able to provide these results directly in any chosen basis. In Sec. III, we describe such a multi-state adaptation of MASH, and show that it is guaranteed to give the correct quantum-classical equilibrium populations of the electronic states in any basis in the long-time limit (assuming ergodicity). In Sec. IV, we demonstrate that this adaptation works just as well for a standard multi-state exciton energy transfer problem as Mannouch and Richardson have shown MASH to work for two-state problems. In Sec. V we compare our adaptation with Mannouch and Richardson's original version of MASH for a variety of problems with just two electronic states, and in Sec. VI we conclude this paper.
## II Nonadiabatic dynamics
Consider a general nonadiabatic system defined by the Hamiltonian
\[\hat{H}(p,q)=\sum_{j=1}^{f}\frac{p_{j}^{2}}{2m_{j}}+\hat{V}(q), \tag{1}\]
In model systems, it is usually most convenient to express the potential in a given diabatic (\(q\)-independent) basis \(\{|n\rangle\}\), i.e.,
\[\hat{V}(q)=\sum_{nm}V_{nm}(q)|n\rangle\langle m|. \tag{2}\]
Although it would in principle be possible to evolve surface hopping dynamics directly in the diabatic basis, it is in practice almost always run in the adiabatic basis (the local eigenbasis of \(\hat{V}(q)\)), where
\[\hat{V}(q)=\sum_{a}V_{a}(q)|a(q)\rangle\langle a(q)|. \tag{3}\]
The two bases are related through their transformation matrix elements \(U_{na}(q)=\langle n|a(q)\rangle\). In this article, we will consider model systems defined in a diabatic basis, but our method can also be used directly with potentials in the adiabatic basis, such as those provided by _ab initio_ electronic structure calculations.
For most photochemical applications, it is a reasonable approximation to treat the nuclei as classical particles. However, to consistently evolve a coupled system of classical and quantum degrees of freedom is a long-standing problem in semiclassical dynamics. To see why, consider a particle at configuration \(q\) with momentum \(p\) and with electronic state \(|\psi\rangle=\sum_{n}c_{n}|n\rangle\). The natural starting point is to evolve the electronic state according to Schrodinger's equation of motion.
\[\dot{c}_{n}=-{\rm i}\sum_{m}V_{nm}(q)c_{m}, \tag{4}\]
or in the adiabatic basis
\[\dot{c}_{a}=-{\rm i}V_{a}(q)c_{a}-\sum_{j}\frac{p_{j}}{m_{j}}\sum_{b}d^{j}_{ab }(q)c_{b}, \tag{5}\]
(Throughout this paper we use units where \(\hbar=1\).) The main difficulty arises when constructing the nuclear dynamics, and in particular when considering the 'back-action' of the electrons on the nuclei. There is a 'force operator' \(\hat{F}_{j}(q)=-\nabla_{j}\hat{V}(q)\), but this is not yet useful to run classical dynamics.
\[\dot{q}_{j} =p_{j}/m_{j} \tag{6a}\] \[\dot{p}_{j} =F_{j}(q), \tag{6b}\]
Existing schemes for nonadiabatic dynamics differ mainly in the way they construct this force. In the following subsections we briefly summarize and comment on some of the most important strategies.
### Fewest switches surface hopping
In surface hopping, the instantaneous force on the nuclei is taken to be that of a single adiabatic state, called the _active surface_.
\[F_{j}(q)=-\langle a(q)|\nabla_{j}\hat{V}(q)|a(q)\rangle. \tag{7}\]
If the trajectories enter a region with non-zero nonadiabatic coupling, they can switch active surface (or 'hop').
\[P_{a\to b}=2\Delta t\,{\rm Re}\left(\frac{c_{b}}{c_{a}}\right)\sum_{j}\frac{p_ {j}}{m_{j}}d^{j}_{ab}. \tag{8}\]
Tully chose this probability such that the fraction of trajectories evolving on surface \(a\) would approximate the average population \(\langle|c_{a}|^{2}\rangle\) with a minimal number of switches. If a hop occurs, the momentum is rescaled along the nonadiabatic coupling vector such that the total energy is conserved. If there is not sufficient kinetic energy to overcome the difference in potential energy between the pre- and post-hop adiabatic surfaces, the standard (although not universally accepted) practice is to reject the hop and reverse the momentum in the direction of the nonadiabatic coupling vector.
Despite the considerations behind the choice of hopping probability, the fraction of trajectories evolving on surface \(a\) does not strictly agree with \(\langle|c_{a}|^{2}\rangle\). This inconsistency is at the root of many of the issues present in surface hopping. Traditionally, the problem has been identified as an 'overcoherence' of the electronic coefficients that can be overcome with (more or less heuristic)'decoherence corrections'. Many such corrections have been proposed, but despite much effort, there is as yet no consensus as to whether any of them has solved the underlying problem. Furthermore, surface hopping does not generally guarantee relaxation to the correct long-time equilibrium in condensed-phase systems, although it does often provide a better approximation than Ehrenfest dynamics.
### Semiclassical mapping approaches
A rather different strategy is to construct the force to be a coherent average over contributions from multiple electronic states.
\[F_{j}(q)=-\langle\psi|\nabla_{j}\hat{V}(q)|\psi\rangle=-\sum_{nm}\nabla_{j}V_{ nm}(q)c_{n}^{*}c_{m}. \tag{9}\]
Ehrenfest dynamics has several severe drawbacks, of which the most important are that it violates detailed balance (it relaxes the system to an overheated equilibrium) and fails to capture wavepacket branching in scattering models. Nevertheless, it has the advantage of being invariant to a unitary transformation of the electronic basis, and the deterministic nature of the force allows an ergodic analysis of its long-time limit. Explicitly, the real and imaginary parts of the electronic coefficients can be regarded as phase-space variables on the same footing as the nuclear degrees of freedom. In terms of these variables, it is clear that Eq. is equivalent to the dynamics of a set of \(N\) harmonic oscillators.
The last of these observations has inspired a more formal mapping of the electronic states to quantum harmonic oscillators, which is now known as the Meyer-Miller-Stock-Thoss (MMST) mapping. Taking the classical limit of this oscillator model leads to a classical phase-space theory with its own force as well as expressions for population and coherence estimators. The new force is also of coherent-average type, but differs from the Ehrenfest force in that the instantaneous populations can be negative (or larger than one). Despite this seemingly unphysical behaviour, the weighted average over many trajectories has for many model problems been found to be more accurate with the MMST mapping than in (multi-trajectory) Ehrenfest dynamics.
Harmonic oscillators are not the only way to map the electronic coefficients onto classical variables. For two-level systems, another choice would be to use the well-known isomorphism to a spin-1/2 system. This approach was recently used to develop a'spin mapping' analogous to the MMST mapping, which leads to a subtly different definition of the force and the estimators. For \(N\)-level systems, the spin mapping has a natural generalization in terms of the so-called Stratonovich-Weyl transformation. This has been shown to at least partly solve the overheating problem of Ehrenfest dynamics, in the sense that the long-time equilibrium reduces to phase-space averages that agree with quantum mechanics up to first order in \(\beta=1/k_{\rm B}T\).
However, just like the MMST mapping, spin mapping can predict (unphysical) negative populations at low temperatures. For an individual trajectory, an instantaneous negative population can cause it to evolve and diverge on an inverted potential, which is completely unphysical. A recent attempt to overcome this problem was to construct a mapping to an anisotropic spin. This does remove the issue of inverted potentials and can (at least for two-level systems) be constructed so as to give the correct long-time equilibrium populations at any temperature, as long as the classical nuclear assumption is valid. However, the timescale of the relaxation to equilibrium was found to be worse in several cases than that of the original spin mapping.
### Mapping approach to surface hopping
It should be clear from what we have said so far that the surface hopping and mapping approaches each have their own advantages and disadvantages. Given this, the natural question is whether it is possible to develop a method that combines the strengths and eliminates the weaknesses of the two strategies.
Mannouch and Richardson have recently shown that, at least for two-level systems, this may indeed be possible. Their'mapping approach to surface hopping' (MASH) uses the nuclear force of a single active adiabatic surface as in surface hopping, but instead of treating the active surface \(V_{a}(q)\) separately from the electronic wavefunction, they set it to be that of the adiabatic state with the _largest instantaneous population_\(|c_{\rm a}|^{2}\). In this way, there is no need to introduce a stochastic hopping probability, because the active state is always uniquely determined by the electronic wavefunction.
Adiabatic states 1 and 2 correspond to the opposite poles along the axis parallel to \(\mathbf{V}=(V_{x},V_{y},V_{z})\), where \(V_{\alpha}=\mathrm{Tr}[\hat{V}\hat{\sigma}_{\alpha}]\) and the \(\{\hat{\sigma}_{\alpha}\}\) are the Pauli spin operators. In MASH, the instantaneous active surface is set to 1 when the Bloch vector \(\mathbf{\sigma}=\langle\psi|\mathbf{\hat{\sigma}}|\psi\rangle\) is on the hemisphere closest to adiabatic 1, and 2 when it is on the hemisphere closest to adiabatic state 2.
In addition to providing a deterministic alternative to stochastic surface hopping, the MASH approach has several other appealing features, a more detailed discussion of which can be found in Ref.:
1. it satisfies detailed balance in the sense that it gives the correct long-time populations of the adiabatic electronic states (assuming ergodicity);
2. the momentum rescaling and momentum reversal arise naturally in the deterministic surface hopping algorithm without any ambiguity;3. in place of heuristic decoherence corrections, there is a way to systematically improve the results by carefully resampling the electronic wavefunction along the trajectory (the quantum jump procedure).
Mannouch and Richardson have demonstrated by comparison with quantum benchmark calculations that MASH is more accurate for typical system-bath models than both FSSH and state-of-the-art mapping approaches. Their method can treat wavepacket branching just at least as well as FSSH, which is currently only possible in mapping via a more expensive cancellation of positive and negative phase-space contributions. They also found MASH to be the most accurate classical-trajectory method considered so far in describing ultrafast internal conversion in pyrazine.
However, their formulation of the method was restricted to two-level systems, and it is not obvious how to generalize it to more electronic states because Mannouch and Richardson specifically constructed their observables for the case of two.
\[C_{AB}(t)=\langle\mathcal{W}_{AB}(c_{0})A(c_{0})B(c_{t})\rangle, \tag{10}\]
Here the estimators \(A(c)\) and \(B(c)\) are constructed from the following complete set of mappings for two-level populations and coherences defined in the adiabatic representation,
\[|1\rangle\langle 1| \mapsto h(|c_{1}|^{2}-|c_{2}|^{2}) \tag{11a}\] \[|2\rangle\langle 2| \mapsto h(|c_{2}|^{2}-|c_{1}|^{2})\] (11b) \[|a\rangle\langle b| \mapsto c_{a}^{*}c_{b}\quad\;\;(a\neq b), \tag{11c}\]
The weight function \(\mathcal{W}_{AB}(c)\) is also specified in the adiabatic representation, as \(\mathcal{W}_{AB}(c)=3\) if \(A\) and \(B\) are both coherences, 2 if one is a coherence and the other a population, and \(2|(|c_{1}|^{2}-|c_{2}|^{2})|\) if they are both populations. Diabatic observables first have to be converted to the adiabatic representation, as in conventional surface hopping, in order to calculate these quantities.
Although this scheme has been shown to work well for a multitude of problems, it is worth asking whether there exists a simpler procedure, especially if one aims to generalize it to more than two electronic states. Firstly, in our view, there should be no reason to give coherences a special treatment. Since these are just population differences in a rotated basis, one should be able to treat them on an equal footing to the populations. Secondly, it would be preferable to calculate the observables in any diabatic basis directly from the wavefunction coefficients in that basis, without having to first convert them to a linear combination of adiabatic observables. The adaptation of MASH presented in the next section achieves both of these goals while at the same time generalising the method to an arbitrary number of coupled electronic states.
## III Multi-state mapping
Having set the context of what we are about to do, we are now ready to propose a multi-state generalisation of MASH. We will separately address the issues of how to run the dynamics, how to measure observables, and how to set up the initial conditions for a typical photochemically initiated nonadiabatic problem.
### Dynamics
The dynamics extends quite naturally from the two-level case by using the same expression for the nuclear force as in Eq. and picking the active surface to be that of the adiabatic state with the largest instantaneous population.
\[P_{n} =|c_{n}|^{2} \tag{12a}\] \[\phi_{n} =\arg c_{n}. \tag{12b}\]
If we also define the classical state projectors
\[\Theta_{n}=\begin{cases}1&\text{if }P_{n}>P_{m}\;\forall\,m\neq n\\ 0&\text{otherwise},\end{cases} \tag{13}\]
As in two-level MASH, we take the dynamically active potential energy surface - the surface that is used to define the nuclear force in Eq. - to be that of the unique
Bloch sphere representation of a two-level system. Any selection of two opposite fixed points on the sphere (e.g., the black dots) defines a diabatic basis, whereas the adiabatic basis corresponds to the opposite points on an axis that follows \((V_{x},V_{y},V_{z})\) (blue and red dots). In MASH, the nuclei evolve on the adiabatic state that is closest to the instantaneous direction of the Bloch vector, \(\mathbf{\sigma}=\langle\psi|\mathbf{\sigma}|\psi\rangle\).
This choice of the adiabatic potential for the nuclear dynamics is consistent with experience from decades of successful surface-hopping calculations.
Assuming the initial electronic wavefunction is normalized, the populations will always sum up to one in any basis, \(\sum_{n=1}^{N}P_{n}=\sum_{n=1}^{N}|c_{n}|^{2}=1\), and they will always be non-negative, \(P_{n}\geq 0\). (The electronic evolution in Eq. is unitary and preserves these two conditions for all time.) In the two-level case, one can visualize \((P_{1},P_{2})\) as a point evolving on the line segment between \(\) and \(\). Similarly, in the three-level case \((P_{1},P_{2},P_{3})\) is a point on the triangle with vertices \(\), \(\), and \(\). illustrates the regions of different active surfaces for these two cases. In general, for \(N\) levels, \((P_{1},P_{2},\ldots,P_{N})\) is a point on the _simplex_
\[\{P\in\mathbb{R}^{N}|\sum_{n}P_{n}=1,P_{n}\geq 0\ \forall\,n\} \tag{14}\]
When a trajectory passes from the region associated with adiabatic state \(a\) to that associated with state \(b\), there is a 'hop', and we rescale the momentum along a direction determined by the nonadiabatic coupling vector as explained in detail in Appendix E. If the kinetic energy is insufficient to hop, we reverse the momentum along the same direction. So far, everything is just a direct extension of the two-level case. In the three-level case, however, we need to address a new issue, which is that hops can occur between states that are uncoupled from each other. For example, if states 1 and 2 are coupled but both are uncoupled from state 3, then trajectories that start in the blue region but are close enough to the green can hop to state 3 before they reach state 2. It is not obvious how to deal with this situation. The simplest option is to accept it. Another option would be to reject the hop if the coupling between the two relevant states (in this case 1 and 3) is less than a given threshold. As in the case of insufficient kinetic energy, one would in such a situation also reverse the momentum. Here we will adopt the simplest option and accept the hop regardless of how strongly coupled the states are. This avoids the ambiguity of defining a coupling threshold and it finishes our discussion of the dynamics. What remains to be decided is how to measure observables. Since our aim is a method that gives the correct quantum-classical equilibrium populations in the long-time limit, we shall begin by discussing what happens when the dynamics has reached equilibrium.
### Equilibrium
Because MASH is deterministic (in contrast to standard surface hopping), one can make more powerful statements about its long-time limit.
\[\rho(p,q,c)\propto\mathrm{e}^{-\beta E(p,q,c)}, \tag{15}\]
where the energy function is
\[E(p,q,c)=T(p)+\sum_{a}V_{a}(q)\Theta_{a} \tag{16}\]
This means that one can predict the long-time values of all observables in terms of phase-space averages with respect to this distribution. Such an analysis has recently been done in the two-state case for a variety of mapping methods, including MASH. In the following, we consider equilibrium expectation values in MASH for a general \(N\)-level system.
For a coupled electronic-nuclear system, the quantum-mechanical partition function with classical nuclei is
\[Z=\int\frac{\mathrm{d}p\mathrm{d}q}{(2\pi)^{f}}\operatorname{Tr}[\mathrm{e}^{ -\beta\hat{H}(p,q)}]. \tag{17}\]
This expression mixes a classical phase-space integral with a quantum trace.
Schematic depiction of a MASH trajectory for two (left) and three (right) states. In any given basis, the set of \(N\) populations is a point on a simplex (generalized triangle) with the \(N\) unit vectors as vertices. In MASH, the active surface is set to be the adiabatic state with the highest instantaneous population. Thereby, each adiabatic state corresponds to the region on the simplex closest to a given vertex. Hops occur whenever the system crosses a border between two of these regions.
over all normalized electronic wavefunctions \(c\),
\[Z_{\rm MASH}=\int\frac{{\rm d}p{\rm d}q}{(2\pi)^{f}}\int_{|c|=1}\frac{{\rm d}c}{ \mathcal{N}}\,{\rm e}^{-\beta E(p,q,c)}, \tag{18}\]
Because each \(\Theta_{a}\) is an idempotent projector onto a different region of \(c\) space, and it has a unit integral \(\int_{|c|=1}{\rm d}c/\mathcal{N}\,\Theta_{a}=1\), it follows that
\[\int_{|c|=1}\frac{{\rm d}c}{\mathcal{N}}\,{\rm e}^{-\beta\sum_{a}V_{a}(q) \Theta_{a}}=\sum_{a}{\rm e}^{-\beta V_{a}(q)}, \tag{19}\]
By the same argument, the thermal equilibrium population of each adiabatic quantum state,
\[\langle P_{a}\rangle=\frac{1}{Z}\int\frac{{\rm d}p{\rm d}q}{(2\pi)^{f}}\,{\rm Tr }[{\rm e}^{-\beta\hat{H}(p,q)}|a(q)\rangle\langle a(q)|], \tag{20}\]
will be equal to the corresponding MASH expression
\[\langle\Theta_{a}\rangle=\frac{1}{Z}\int\frac{{\rm d}p{\rm d}q}{(2\pi)^{f}} \int_{|c|=1}\frac{{\rm d}c}{\mathcal{N}}\,{\rm e}^{-\beta E(p,q,c)}\Theta_{a}, \tag{21}\]
Hence, if one were to choose to use the state projectors as population estimators, this would be guaranteed to give the correct equilibrium populations in the adiabatic basis.
However, it would not work more generally, because \(\langle\Theta_{n}\rangle\) does not equal \(\langle P_{n}\rangle\) in other bases.
\[\langle P_{n}\rangle =\frac{1}{Z}\int\frac{{\rm d}p{\rm d}q}{(2\pi)^{f}}\,{\rm Tr}[{ \rm e}^{-\beta\hat{H}(p,q)}|n\rangle\langle n|]\] \[=\frac{1}{Z}\int\frac{{\rm d}p{\rm d}q}{(2\pi)^{f}}\,\sum_{a}{\rm e }^{-\beta[T(p)+V_{a}(q)]}|\langle a(q)|n\rangle|^{2}, \tag{22}\]
whereas the former is
\[\langle\Theta_{n}\rangle =\frac{1}{Z}\int\frac{{\rm d}p{\rm d}q}{(2\pi)^{f}}\int_{|c|=1} \frac{{\rm d}c}{\mathcal{N}}\,{\rm e}^{-\beta E(p,q,c)}\Theta_{n}\] \[=\frac{1}{Z}\int\frac{{\rm d}p{\rm d}q}{(2\pi)^{f}}\sum_{a}{\rm e }^{-\beta[T(p)+V_{a}(q)]}\int_{|c|=1}\frac{{\rm d}c}{\mathcal{N}}\,\Theta_{a} \Theta_{n}. \tag{23}\]
This is in general _not_ equal to \(\langle P_{n}\rangle\) because
\[\int\frac{{\rm d}c}{\mathcal{N}}\,\Theta_{a}\Theta_{n}\neq|\langle a(q)|n \rangle|^{2}. \tag{24}\]
In other words, the state projection defined by Eq. does not transform correctly as a population estimator under unitary basis rotations.
### Populations
A population estimator \(\Phi_{n}\) can in fact be constructed such that \(\langle\Phi_{n}\rangle=\langle P_{n}\rangle\) in _any_ basis of \(N\) orthonormal electronic states, whether they are adiabatic or diabatic. Here we shall give an overview of the argument that establishes this and leave the technical details to Appendices B and C.
Since the Ehrenfest populations \(P_{n}=|c_{n}|^{2}\)_do_ transform correctly (i.e., quantum mechanically) under unitary basis rotations, as do scalars, our ansatz is an equivariant population estimator of the form
\[|n\rangle\langle n|\mapsto\Phi_{n}=\alpha_{N}\,P_{n}+\beta_{N}, \tag{25}\]
Our goal is to choose these parameters such that
\[\int_{|c|=1}\frac{{\rm d}c}{\mathcal{N}}\,\Theta_{a}\Phi_{n}=|\langle a(q)|n \rangle|^{2} \tag{26}\]
This will suffice to ensure - by virtue of the argument in Eqs. and - that the expectation value of \(\Phi_{n}\) will be correct at equilibrium. Appendix B shows that Eq.
\[\int_{|c|=1}\frac{{\rm d}c}{\mathcal{N}}\,\Theta_{a}\Phi_{b}=\delta_{ab} \tag{27}\]
The constraints that this imposes on \(\alpha_{N}\) and \(\beta_{N}\) are then investigated in Appendix C, which obtains the unique solution
\[\alpha_{N}=\frac{N-1}{H_{N}-1}\quad\text{and}\quad\beta_{N}=\frac{1-\alpha_{N} }{N}, \tag{28}\]
Hence we arrive at the estimator
\[\Phi_{n}=\frac{1}{N}+\alpha_{N}\left(P_{n}-\frac{1}{N}\right), \tag{29}\]
### Coherences
In contrast to Ref., as well as most other literature on surface hopping, we shall begin by rewriting coherences as population differences between rotated states so that they can be treated in the same way as populations. Note that we are in an ideal position to do this here because our population estimator in Eq. transforms correctly under unitary basis rotations.
The coherences between states \(|n\rangle\) and \(|m\rangle\) are linear combinations of the Hermitian operators \(\hat{\sigma}^{x}_{nm}=|n\rangle\langle m|+|m\rangle\langle n|\) and \(\hat{\sigma}^{y}_{nm}=-\mathrm{i}\left(|n\rangle\langle m|-|m\rangle\langle n|\right)\),
\[|n\rangle\langle m| =\frac{1}{2}\left(\hat{\sigma}^{x}_{nm}+\mathrm{i}\,\hat{\sigma}^ {y}_{nm}\right) \tag{30a}\] \[|m\rangle\langle n| =\frac{1}{2}\left(\hat{\sigma}^{x}_{nm}-\mathrm{i}\,\hat{\sigma}^ {y}_{nm}\right), \tag{30b}\]
and we can rewrite \(\hat{\sigma}^{x}_{nm}\) and \(\hat{\sigma}^{y}_{nm}\) as differences between the population operators of the rotated states \(|x^{\pm}_{nm}\rangle=\sqrt{\frac{1}{2}}\left(|n\rangle\pm|m\rangle\right)\) and \(|y^{\pm}_{nm}\rangle=\sqrt{\frac{1}{2}}\left(|n\rangle\pm\mathrm{i}|m\rangle\right)\):
\[\hat{\sigma}^{x}_{nm} =|x^{+}_{nm}\rangle\langle x^{+}_{nm}|-|x^{-}_{nm}\rangle\langle x ^{-}_{nm}| \tag{31a}\] \[\hat{\sigma}^{y}_{nm} =|\,y^{+}_{nm}\rangle\langle y^{+}_{nm}|-|\,y^{-}_{nm}\rangle \langle y^{-}_{nm}|. \tag{31b}\]
Since these involve pollution differences, we can measure them with classical estimators in the same way as we measure populations,
\[\hat{\sigma}^{x}_{nm} \mapsto\Sigma^{x}_{nm}=\alpha_{N}\left(P_{x^{\pm}_{nm}}-P_{x^{ \mp}_{nm}}\right) \tag{32a}\] \[\hat{\sigma}^{y}_{nm} \mapsto\Sigma^{y}_{nm}=\alpha_{N}\left(P_{y^{\pm}_{nm}}-P_{y^{ \mp}_{nm}}\right), \tag{32b}\]
Our mapping for the coherences is therefore
\[|n\rangle\langle m| \mapsto\frac{1}{2}\left(\Sigma^{x}_{nm}+\mathrm{i}\,\Sigma^{y}_{ nm}\right) \tag{33a}\] \[|m\rangle\langle n| \mapsto\frac{1}{2}\left(\Sigma^{x}_{nm}-\mathrm{i}\,\Sigma^{y}_{ nm}\right), \tag{33b}\]
can be written in the neater form
\[|n\rangle\langle m|\mapsto\alpha_{N}c^{*}_{n}c_{m}. \tag{34}\]
Together with the population estimator in Eq., this leads to a general estimator for electronic observables that will give the same result in any unitarily rotated basis:
\[\sum_{nm}O_{nm}|n\rangle\langle m|\mapsto\sum_{nm}O_{nm}\left[\alpha_{N}c^{*} _{n}c_{m}+\beta_{N}\delta_{nm}\right]. \tag{35}\]
### Initial conditions
To simulate a non-equilibrium process starting in a pure electronic state \(|n\rangle\langle n|\), such as the bright state of a molecule that has just been photoexcited, we need to define an initial electronic distribution, \(\rho_{n}(c)\), such that
\[\int_{|c|=1}\frac{\mathrm{d}c}{\mathcal{N}}\rho_{n}\Phi_{m}=\delta_{nm}. \tag{36}\]
In this way, the initial estimate of the population of state \(m=n\) is 1, and that of all other states is zero. There are many choices of \(\rho_{n}\) that fulfil this condition.
\[\rho_{n}(c)=\Theta_{n}(c), \tag{37}\]
2). This choice fulfils Eq. by the same argument as in Appendix C, regardless of whether the initial state is a diabat or an adiabat. For a two-state problem in which the initial state is specified in the adiabatic basis, Eq. is equivalent to the prescription for \(A(c)\) that Mannouch and Richardson used in Eq.. However, our prescription for \(B(c)\) is different from theirs, and our formulation avoids their weight function \(\mathcal{W}_{AB}(c)\).
In practice, one can sample points \(c\) from \(\Theta_{n}(c)\) is as follows. First, sample \(N\) pairs \((x_{k},y_{k})\), where \(x_{k}\) and \(y_{k}\) are Gaussian deviates with zero mean and unit variance. Then set \(c_{k}=(x_{k}+\mathrm{i}y_{k})/\sqrt{\sum_{l}(x_{l}^{2}+y_{l}^{2})}\). This has \(|c|=1\) and, due to the rotational invariance of the multi-dimensional normal distribution, is a point with uniform probability on the simplex. Finally, check which state has the largest \(|c_{k}|^{2}\) - if it is the initial state \(n\), then accept the point, and if not, resample a new point in the same way until it is accepted.
## IV Application to exciton energy transfer
To test the multi-state algorithm introduced in Sec. III, we have applied it to a Frenkel-exciton model of energy transfer in the Fenna-Matthews-Olson complex. This has become a standard benchmark for nonadiabatic dynamics and it allows a comparison with a variety of other trajectory-based methods. The problem is challenging for conventional surface hopping due to the presence of 'trivial' crossings with low hopping probabilities, which can require an extremely small time step to converge. Several phase-space methods have been reported to perform well for FMO at high temperatures, including spin mapping, MMST mapping, and the symmetric quasi-classical trajectory method of Cotton and Miller. However, these methods struggle to recover the correct long-time equilibrium populations at low temperatures, and they only avoid unstable inverted potentials because all of the potential-energy surfaces in the model are harmonic with the same frequencies.
The standard FMO model is comprised of seven sites, but to facilitate a comparison with a previous surface hopping study we have also considered a reduced model with three sites. In both cases, the sites are coupled independently to identical harmonic baths. The full Hamiltonian and all model parameters are given in Appendix D. The initial condition for our simulations is a pure electronic state in an uncoupled classical Boltzmann bath.
The initial electronic state can be in the site or the exciton basis and we shall report results for both below. Each simulation was averaged over \(10^{5}\) trajectories for good statistical convergence, although rough results can typically be obtained with \(10^{3}\). We used a time step of 0.25 fs with the trajectory integrator and hopping protocol described in Appendix E. Fully quantum mechanical HEOM benchmark results were computed for comparison using the pyrho open source software.
### Three-state model
Sindhu and Jain have recently used a three-site FMO model to compare a variety of different decoherence corrections to fewest switches surface hopping. To assess the performance of MASH against FSSH, we shall use the most accurate of the methods they considered in their comparison as a reference. This method, the augmented surface hopping (A-FSSH), employs a parameter-free decoherence correction that has been found to improve upon the original FSSH algorithm for condensed-phase problems (for example, it has been found to recover Marcus theory rates in the golden-rule limit). It should be noted that Sindhu and Jain quantized one of the nuclear modes in their FMO calculation to give a manifold of vibronic states, whereas our MASH calculations treat all nuclear modes classically. Since their quantization improved the agreement of A-FSSH with HEOM rather than harmed it, we still feel this provides a fair comparison.
In Figure 3, we compare MASH to the A-FSSH results from Ref. as well as to HEOM. We find that MASH almost perfectly captures the initial coherent oscillations in the HEOM polulation dynamics, as well as the long-time relaxation. In contrast, A-FSSH leads to overly damped oscillations and too rapid thermalization. These results are interesting not only because they demonstrate that MASH provides a clear improvement over A-FSSH, but because it does so without decoherence corrections. This suggests that overcoherence may not be such a useful way to understand the problems of FSSH as previously thought. The issue may instead be more closely related to the inconsistency between \(\langle|c_{a}|^{2}\rangle\) and the fraction of trajectories on the active surface. This is precisely the problem that MASH was designed to resolve by uniquely defining the active surface in terms of the wavefunction coefficients.
Another interesting observation is that, even at \(77\,\mathrm{K}\) where \(k_{\mathrm{B}}T=53.5\,\mathrm{cm}^{-1}\) is roughly two times smaller than the characteristic phonon energy of \(\hbar\omega_{\mathrm{c}}=106\,\mathrm{cm}^{-1}\), MASH gives accurate results despite treating the nuclei classically. This indicates that nuclear quantum effects are less important for this system than the quantization of one of the nuclear modes in Ref. would suggest them to be.
### Seven-state model
We are not aware of any calculations using conventional surface hopping for the seven-state FMO model, but it is expected that they would suffer from the same difficulties as in the three-state case. In Figure 4, we compare our MASH results for this model directly to those of the quantum HEOM benchmark. Panel (a) shows the population dynamics in the site basis at \(300\,\mathrm{K}\), starting from site 1, and panel (b) shows the dynamics in the exciton basis at \(300\,\mathrm{K}\), starting from exciton state 1. The exciton basis is defined as the eigenbasis of the system Hamiltonian in Eq.. In both cases, MASH is seen to agree almost perfectly with the fully quantum benchmark. The agreement is at least as good as or better than that obtained with previously reported mapping approaches at the same temperature.
Panel (c) of shows the same situation as in panel (a), but at \(77\,\mathrm{K}\). This low-temperature regime is more challenging for mapping approaches, several of which predict negative populations. However, MASH continues to agree well with HEOM at short times, except perhaps for the small deviation in the transfer between sites 1 and 2, and it gives the correct quantum-mechanical equilibrium populations in the long-time limit. It is worth repeating that, as in the case of the three-state model considered in Figure 3, these long-time populations are obtained correctly without including nuclear quantum effects in the calculation.
Comparison of HEOM (dash-dotted), MASH (solid), and A-FSSH (dotted) for population dynamics in a three-site FMO model. Results are reported in the site basis at \(77\,\mathrm{K}\), starting in site 1. The A-FSSH results are from Ref..
(Including nuclear quantum effects might perhaps improve on the MASH description of the short-time population transfer between sites 1 and 2 in panel (c) of Figure 4, but this would only be a small correction.)
### Coherence dynamics
To assess the accuracy of the coherence estimator introduced in Sec. III.4, we have used it to calculate the coherence between sites 1 and 2 of the seven-state FMO model with the initial population on site 1. To facilitate a comparison with previous spin-mapping simulations, the timescale of the bath was set to \(\tau_{\mathrm{c}}=100\,\mathrm{fs}\) in these calculations, as it was in the HEOM calculations of Sarovar _et al._. Since these earlier HEOM calculations included a trapping rate that is not present in our FMO model, we recomputed the HEOM results without any trapping, and found that they differ only very slightly (by around 1%) from the results reported in Ref..
The agreement is not perfect, but it is clearly very good for the imaginary part of the coherence and reasonably good for the real part, especially at short times. Our coherence estimator has therefore been validated for a multi-state problem.
## V Application to two-level systems
Since our scheme for measuring electronic observables differs from that of Mannouch and Richardson even for two-level systems, it is important to check the performance of our approach also in this case. For this purpose, we have applied our method to some of the two-state Tully, spin-boson, and pyrazine models considered previously by Mannouch and Richardson. All model definitions and parameters are the same as in Ref., where further comparisons to FSSH, Ehrenfest, and spin-mapping results can be found. Unless otherwise stated, the results were computed from \(10^{5}\) trajectories and integrated with a timestep of \(0.25\,\mathrm{fs}\).
Comparison of MASH (solid) and HEOM (dash-dotted) for population dynamics in the seven-site FMO model at (a) \(300\,\mathrm{K}\) in the site basis, (b) \(300\,\mathrm{K}\) in the exciton basis, and (c) \(77\,\mathrm{K}\) in the site basis. The initial state was site 1 for panels (a) and (c) and exciton 1 for panel (b).
Comparison of MASH (solid) and HEOM (dash-dotted) results for the dynamics of the dominant coherence of the seven-state FMO model in the site basis at \(300\,\mathrm{K}\), starting from site 1.
### Tully models
First, in Fig. 6, we consider the problem of wavepacket splitting in Tully's single avoided crossing model. Mannouch and Richardson have shown that their scheme reproduces the correct final momentum distribution at both low and high incoming kinetic energies. Our scheme uses a different trajectory weighting that reproduces their results at high energy but gives slightly less accurate peak heights at low energy. The area under each peak is proportional to the fraction of trajectories emerging on each adiabat. In the original MASH, this fraction is \(\langle\Theta_{a}(t)\rangle\), where \(t\) is a time at which the crossing is complete. In our scheme, the time-dependent populations \(\langle\Phi_{a}(t)\rangle\) do give the correct adiabatic state branching ratio in the low energy example. However, since \(\langle\Phi_{a}(t)\rangle\) is not the same as the fraction of trajectories emerging on each adiabat, our final momentum distribution is slightly incorrect.
Here, our results are similar to those of the original MASH method, and arguably slightly better at high momenta. For low momenta, neither method can reproduce the Stuckelberg oscillation, which is due to electronic interference. Note also that our estimator predicts negative transmission probabilities at low momenta where the upper product adiabat is energetically inaccessible, whereas Mannouch and Richardson's transmission probability goes correctly to zero.
Despite the issues that these tests have identified, the overall impression we get from the comparisons in Figs. 6 and 7 is that our method does not perform significantly worse for these models than the original MASH method of Mannouch and Richardson. Especially when one considers that these non-ergodic, one-dimensional, microcanonical models do not satisfy the assumptions we made when deriving our population estimator, and are not therefore the sort of problems for which our method was designed. The present MASH results in are certainly more accurate than those of either Ehrenfest dynamics or spin mapping, for example, both of which fail to describe the wavepacket bifurcation.
### Spin-boson model
Next, we consider the spin-boson model, for which
\[\hat{H}=H_{0}+\Delta\,\hat{\sigma}_{x}+\Big{(}\varepsilon+\sum_{j}\xi_{j}q_{j} \Big{)}\hat{\sigma}_{z}, \tag{38}\]
where \(H_{0}=\frac{1}{2}\sum_{j}(p_{j}^{2}+\omega_{j}^{2}q_{j}^{2})\) is a bath of oscillators with spectral density
\[J(\omega)=\frac{\Lambda}{2}\frac{\omega\omega_{\rm c}}{\omega^{2}+\omega_{\rm c }^{2}}. \tag{39}\]
For this model, we initialized the nuclei from a Wigner distribution to be consistent with Ref..
Final momentum distribution after a single avoided crossing in Tully’s first model, for two initial energies. The original MASH results and the model parameters are from Ref.. These histograms were generated from \(10^{6}\) trajectories.
Transmission probability from the lower to the upper adiabat in Tully’s double avoided crossing model. The original MASH results and the model parameters are from Ref..
The integration time step (in units of \(\Delta^{-1}\)) was 0.01 for panels (a), (b) and (d), and 0.002 for panel (c). In the coherent examples [panels (a) and (b)], our scheme is slightly more accurate at long times than Mannouch and Richardson's. For the case of activationless electron transfer (\(2\varepsilon=\Lambda\)) [panel (c)], both schemes yield similar results. However, in the Marcus inverted regime [panel (d)], our relaxation is too slow, even though it does eventually reach the correct limit (not shown). This regime is known to be challenging for trajectory-based methods and it is remarkable that Mannouch and Richardson's scheme is so accurate. The fact that the present version of MASH does not capture the inverted regime correctly is probably the most serious deficiency of the method we have found so far. It is not clear how to solve this problem without abandoning the basis-set independence of our population estimator, but it is conceivable that a quantum-jump procedure might improve our results.
### Pyrazine
Finally, we consider ultrafast dynamics through the conical intersection in a full-dimentional (24-mode) model of pyrazine. This model includes bilinear couplings which make it challenging for traditional mapping methods. It is also typical of the sort of photochemical problems to which one might expect methods like MASH to be applied. As shown in Fig. 9, our results for this final two-state problem are of comparable quality to those of the original MASH scheme.
## VI Concluding remarks
In this article, we have shown how to extend the MASH methodology of Mannouch and Richardson to general \(N\)-level systems.
Population dynamics of the spin-boson model in various regimes: (a) coherent dynamics at a high temperature; (b) coherent dynamics at a low temperature; (c) activationless electron transfer; (d) electron transfer deep in the inverted regime. The original Mash results, the exact results, and the model parameters are all from Ref..
Population transfer in a 24-mode pyrazine model. The original MASH results and the model parameters are from Ref.. The exact (MCDTH) results are from Ref..
Our applications to FMO exciton models have shown that the resulting multi-state MASH method is at least as accurate as any previous semiclassical mapping method, and significantly more accurate than an up-to-date implementation of fewest-switches surface hopping.
Regarding the restriction to classical nuclear motion, we would point out that the majority of interesting nonadiabatic dynamics problems follow a photoexcitation step in which a vast amount of energy (significantly larger than \(k_{\rm B}T\)) has been deposited in the system. The resulting nuclear motion will often be sufficiently fast that the classical nuclear motion approximation is well justified, as it certainly seems to be in all of the FMO calculations we have presented here. However, nuclear quantum effects are expected to be important in some other contexts. For example, nuclear tunnelling is known to have a significant impact on the rates of electron transfer reactions in the Marcus inverted regime. It might therefore be interesting to add nuclear quantum effects to the present methodology, perhaps by adapting the ring-polymer molecular dynamics techniques that have already been developed for standard surface hopping.
Regarding how much more successful MASH is for the FMO problem than fewest switches surface hopping (as we have shown in Figure 3), we would say that Mannouch and Richardson's idea of deterministically tying the active surface to the adiabatic state with the largest population was quite inspired. This allowed them to _derive_ the momentum rescaling and momentum reversal stages of their MASH surface hopping algorithm from first principles, to avoid the inconsistency between the stochastically averaged \(\langle|c_{\alpha}|^{2}\rangle\) and the active adiabatic surface \(V_{a}(q)\), and thereby to eliminate the need for _ad hoc_ 'decoherence' corrections, in a single stroke. All we have done here is to show that their idea can be adapted to treat an arbitrary number of coupled electronic states in a straightforward and internally consistent way. While this has come at the cost of sacrificing some of the advantages of the original MASH scheme for two-state problems, as we have shown in Figures 6 to 8 and discussed in Sec. V, we feel that the extension to more electronic states is worth this sacrifice because it opens up the possibility of applying the method to a far wider variety of interesting nonadiabatic problems.
|
10.48550/arXiv.2305.08835
|
A multi-state mapping approach to surface hopping
|
Johan E. Runeson, David E. Manolopoulos
| 1,918
|
10.48550_arXiv.1208.0397
|
###### Abstract
We report a study of the structure of droplets of colloidal gels containing dissolved sodium chloride. The components segregate and form intricate patterns. The salt crystalizes in fractal and multi-fractal dendritic forms which are determined by the material which forms the colloidal gel. Here potato starch, gelatine and carboxymethyl cellulose have been used. The substrate also plays a role in some cases. Photographs and micrographs at different level of magnification are shown.
## Introduction
This study demonstrates how a very simple experiment of drying different colloidal solutions containing sodium chloride provides a fascinating array of rich and diverse patterns. Various aspects of evaporation of sessile drops on solid surfaces have long been a topic of interest. The shape and time development of the contact line, the contour of the three dimensional liquid-vapour surface, the convection currents set up in the fluid, segregation of different components and other interesting features have been observed and reported. The present paper deals with evaporation and drying of three component aqueous solutions, one of which is sodium chloride, leading to segregation and crystallization of the salt. The interesting point is that the morphology of the host colloid controls the aggregation process and the crystal deposits range through; simple cubic shapes, to faceted cuboids, highly branched multi-fractal dendritic patterns to extremely intricate designs resembling Julia sets depending upon the solvent and other conditions such as water content, temperature and humidity. The drying droplets have been observed at different scales of magnification and photographed by camera (Nikon CoolPixL120), polarizing microscope and by SEM imaging, each scale revealing new features. The solid substrate on which the sample dries also affects the pattern and the adhesion of the dried film to the substrate. We have used two different solids as the substrate - glass and polypropylene(PP).
The subject is not of academic interest only, there are practical applications as well, e.g. drying paints and coatings must be designed to avoid segregation of components, biomedical science utilizes the process for pathological investigation of biological fluids.
## Materials and Methods
In the present paper we report work on potato starch gel (with varying concentration), gelatine and carboxymethyl cellulose as the colloidal host fluid and sodium chloride as the dissolved salt. For comparison we also observe dried drops of sodium chloride in water and drying of starch films without salt.
## Two component solution: Salt in water
When a drop of a solution of sodium chloride in distilled water is left to dry at room temperature (30degC), one observes the well-known coffee stain effect. Salt crystals form around the outer boundary of the drop (Fig.1). When a colloidal solution is used instead of pure water, the result is strikingly different. We discuss the different cases below.
## Two component solution: Starch in water
Drying a drop of starch solution without any salt produces a transparent film, which detaches easily from the PP substrate, but sticks to the glass. During drying the outer fluid-solid contact line remains pinned but the boundary of the still fluid portion recedes inward. This is shown in the series of successive photographs.
## Experiments with three-component solutions
## Potato Starch Gel:
Colloidal gels of different proportion of the three components (x(salt) of NaCl, x(ps) of potato starch, MW (162.14)\({}_{\text{a}}\) and 50 ml of distilled water) have been prepared. The chemicals are procured from Loba Chemie, Mumbai, x(salt) mol of NaCl is dissolved in the water and slightly heated. Then x(ps) g of potato starch are added and the mixture is heated to 90degC and stirred to get a homogeneous gel. Stirring is stopped and the gel is allowed to cool for 1 h.
We show results for x(salt) = 0.01 mol, x(ps) = 0.5 g and for x(salt) = 0.15 mol, x(ps) = 0.5 g. A drop of a 0.5 ml is deposited on a solid substrate (glass or PP) and left to dry. The dried film, which has a diameter of about 1.4 cm is observed under different magnifications.
## Images observed by the naked eye
The appearance of the dried drops are very striking. For lower salt concentration a transparent band of starch separates along the outer periphery. On moving inward a number of narrower concentric rings are to be seen. In the central region the salt crystalizes forming fractal dendritic patterns. The films are photographed by Nikon CoolpixL120.
As the salt concentration increases, the band of starch around the edge grows successively narrower and finally disappears. Now the whole circular patch is covered by salt aggregates with a prominent cubic symmetry, but surrounded by fine dendritic structures.
Changing the substrate from glass to polypropylene (PP) does not change the patterns much, but on PP the film detaches cleanly from the substrate.
shows the drops for two different compositions, on different substrates as photographed in close-up
## Microscope images
Micrographs taken with Leica DM750 reveal very intricate and interesting details of each of the distinct rings seen with the naked eye. Shows the detailed texture of each region of the complete film shown in (A), which is mapped schematically in the centre. Region is the outermost layer comprising exclusively of starch. The granules are seen in the micrograph. As one focusses inward regions contain more and more salt. The thickness of these rings are shown magnified in the map for clarity. In the innermost circle salt dendrites are seen (they look dark) on a background of starch (which appears white).
## SEM images
SEM images with EDAX show structural details at higher magnification and the composition of each of the layers. Two regions are shown in figures5. The outermost layer clearly consists mostly of starch. The micrographs and SEM images both show the typical starch grains and EDAX shows strong lines corresponding to C and O, Na and Cl lines are very weak here. On entering into the circular rings the composition reverses. Here Na and Cl are abundant and C, O much weaker. The SEM image shows intricate formations of salt.
## Gelatine and Salt
The next set of experiments is done with gelatine in water as the colloidal medium. 0.03 mole of NaCl is dissolved in 50 cc of water, then 0.5 g gelatine is mixed with the solution and stirred for 1 h at 70degC until a homogeneous gel forms.
As in the previous set, drops are allowed to dry on two different substrates. The dried films look very different from the potato starch films shown in There is no clear segregation of the salt either along the edge of the drop, or towards the centre. Instead, cubic crystals of salt form more or less uniformly throughout the film (Fig.6). Here the appearance of the dried films looks similar on both surfaces.
## Microscopic images with Leica DM750
Under a microscope we can see interesting details shown in the two square frames marked in the Focussing on the large crystals shows a regular step-like structure of the crystal. This is similar to what the crystals in Fig.1 look like under the microscope. The region in between the large crystals, now shows beautiful patterns of aggregation of the salt. The structure on analysis is found to be multi-fractal, further work in this direction is in progress and will be reported elsewhere.
The result for a slightly different composition is also shown in Here 0.03 mole NaCl is taken with 2g gelatine and 50 cc water and the gel is prepared in the same way as described before. For this thicker gel the salt aggregation pattern has a strong resemblance to a Julia set.
## A possible scenario for multi-fractal growth.
Geometric multi-fractals are formed by strongly inhomogeneous growth processes. We suggest an inhomogeneous process which may lead to a formation as shown in Fig.7, this is illustrated schematically in After formation of the large crystals, the surrounding region is left depleted of dissolved salt. The drying process continues, leaving a very thick viscous gel, through which the remaining salt ions move very slowly towards microscopic nucleating crystallites, which are also square in shape. The aggregation is driven by the concentration gradient, which is strongest at the corners of the cubic crystal. The remaining ions therefore tend to attach preferentially at the corners of the seed crystal.
**Carboxymethyl cellulose(CMC) and salt**Our last set of samples contains CMC as the gel former. Here 0.01 mole of NaCl and 0.1 g CMC are stirred in 50 cc water for 1 h at room temperature (32 degC).
## Photographs and micrographs
Corresponding micrographs are shown in figures (C) and (D).
Unlike the two earlier samples, where a strongly directed
Possibly, the effect of the gel is strong enough to compete with the interaction between the salt ions leading to rock-salt structure formation. The micrographs show a more directed growth at short ranges, but the pattern is not like the multi-fractals observed earlier.
The effect of the substrate seems to play a role here. The structure of the film on glass does not have a ring like crystalline border which is seen in the film cast on the PP substrate.
## Discussion and Conclusions
This simple experiment reveals a rich variety of pattern formation processes at work in an evaporating drop of colloid with an added salt. The coffee stain effect is now well studied. It has also been shown that Marangoni effect can suppress this effect. Electro-wetting under different conditions has been used to control the coffee stain effect. Strong phase segregation in albumen and salt solution has been observed by Tarasevich and Pravoslavnova. However, a satisfactory explanation for such phenomena is still lacking.
Our observation is that the width of the band of starch drying out first along the boundary for the potato starch-salt gels, depends on the details of the composition and also ambient humidity. In our experiments it is unlikely that Marangoni effects which produce very weak temperature gradients will have a significant role. Moreover, as the colloid material changes from potato starch to gelatine to CMC, the overall structure of the film as well as the detailed structure of the salt crystal aggregates changes remarkably. We may infer that the visco-elastic properties of the colloids and their microstructure affect the migration and aggregation pattern as the salt moves through the thickening gel. The subject appears to be a promising area for future study, particularly as it has already been proved to be a useful diagnostic technique.
Further work to understand the drying process, which is itself an interesting topic of research is necessary as well as a theoretical analysis of the segregation phenomena under different conditions. The formation of fractal and multi-fractal aggregates may be simulated through random walks or Laplacian growth processes similar to viscous fingering, driven in this case by the concentration gradient.
|
10.48550/arXiv.1208.0397
|
Aggregation Patterns of Salt Crystalizing in Drying Colloidal Solvents
|
Moutushi Dutta Choudhury, Sayanee Jana, Sruti Dutta, Sujata Tarafdar
| 5,250
|
10.48550_arXiv.1209.0195
|
###### Abstract
We calculate accurate eigenvalues and eigenfunctions of the Schrodinger equation for a two-dimensional quantum dipole. This model proved useful for the study of elastic effects of a single edge dislocation. We show that the Rayleigh-Ritz variational method with a basis set of Slater-type functions is considerably more efficient than the same approach with the basis set of point-spectrum eigenfunctions of the two-dimensional hydrogen atom used in earlier calculations.
## 1 Introduction
In a recent paper Dasbiswas et al discussed the bound-state spectrum for a straight edge dislocation oriented along the \(z\) axis. Within the continuum model, the authors reduced the problem to the two-dimensional Schrodinger equation for a quantum dipole. They obtained very accurate eigenvalues and eigenfunctions by means of a discretization of the \(x-y\) space. Since this real-space diagonalization method (RSDM) is not so practical for highly excited states, those authors also carried out a Rayleigh-Ritz (RR) variational calculation with the basis set of eigenfunctions of the two-dimensional Coulomb problem.
The variational RR eigenvalues are known to approach the exact ones from above. However, in the present case the ground-state energy \(\epsilon_{1}^{RR}=-0.0970\) calculated with as many as 400 basis functions exhibits a considerable discrepancy with respect to the same eigenvalue obtained by RSDM \(\epsilon_{1}^{RSDM}=-1.39\). Later on, Amore carried out a more accurate calculation with 3600 hydrogen eigenfunctions and obtained \(\epsilon_{1}^{RR}=-0.128\) as well as \(\epsilon_{1}^{RRS}=-0.132\) by fitting and extrapolating the outcome of a Shanks transformation. The authors of both articles resorted to a variational parameter (decaying parameter or length scale) that considerably improves the result. The remaining disagreement between RR and RSDM is probably due to the well known fact that the basis set used in those calculations is not complete because it does not include the continuous spectrum (see, for example Ref. and the references therein). For this reason the RR calculation proposed by those authors is expected to have alimited accuracy no matter how large the dimension of the basis set of discrete states. It is surprising, however, that the lack of the continuous wavefunctions appears to be more noticeable for the ground state. In addition to it, at first sight it seems that there is a better agreement for the odd states.
The purpose of this paper is to carry out a RR calculation with a nonorthogonal basis set of square-integrable functions that in principle does not require the continuous spectrum. In section 2 we describe the RR variational method with such an improved basis set. In section 3 we compare present results with those obtained earlier by Dasbiswas et al and Amore. Finally, in section 4 we summarize the main results and draw conclusions.
## 2 Rayleigh-Ritz variational method
The linearized model for the Ginzburg-Landau theory leads to the Schrodinger equation
\[-\frac{\hbar^{2}}{2m}\nabla^{2}\psi+p\frac{\cos\theta}{r}\psi=E\psi \tag{1}\]
Choosing the units of length \(\hbar^{2}/(2mp)\) and energy \(2mp^{2}/\hbar^{2}\) we obtain the dimensionless eigenvalue equation
\[-\nabla^{2}\psi+\frac{\cos\theta}{r}\psi=\epsilon\psi \tag{2}\]
Since the potential \(V(r,\theta)=\cos\theta/r\) is invariant under reflection about the \(x\) axis \(V(r,-\theta)=V(r,\theta)\), then the wavefunctions are either even \(\psi(r,-\theta)=\psi(r,\theta)\) or odd \(\psi(r,-\theta)=-\psi(r,\theta)\) under such coordinate transformation.
Dasbiswas et al and Amore resorted to the RR variational method with the basis set of eigenfunctions of the planar hydrogen atom:
\[\psi_{n,l}^{H}(r,\theta)=\frac{1}{\sqrt{\pi}}R_{n,l}(r)\times\left\{\begin{array} []{c}\cos(l\theta),\,1\leq l\leq n\\ \frac{1}{\sqrt{2}},l=0\\ \sin(l\theta),\,-n\leq l\leq-1\end{array}\right. \tag{3}\]where \(n=1,2,\ldots\) and \(R_{n,l}(r)\) is the normalized solution to the radial equation. However this basis set is incomplete if one does not include the eigenfunctions for the continuous spectrum (see, for example, and references therein).
If, on the other hand, the chosen basis set of square-integrable functions does not require the continuous spectrum to be complete then one expects the accuracy of the RR variational results to be determined only by the basis dimension.
\[\left\{\phi_{j}^{e},\,j=1,2,\ldots\right\}=\left\{e^{-\alpha r},r^{i+1}\cos^{j }\theta e^{-\alpha r},\,i,j=0,1,\ldots\right\} \tag{4}\]
for the even states and
\[\left\{\phi_{j}^{o},\,j=1,2,\ldots\right\}=\left\{r^{i+1}\sin\theta\cos^{j} \theta e^{-\alpha r},\,i,j=0,1,\ldots\right\} \tag{5}\]
This basis set resembles the Slater orbitals commonly used in quantum chemistry calculations of atomic and molecular electronic structure.
The RR method with the variational ansatz
\[\psi=\sum_{j=m}^{N}c_{m}\phi_{m} \tag{6}\]
leads to the generalized eigenvalue problem
\[{\bf HC}=\epsilon{\bf SC} \tag{7}\]
Note that it is possible to obtain explicit expressions for both kinds of matrix elements in terms of the variational parameter \(\alpha\); even more important is the fact that, with the help of a computer algebra software, like Mathematica, we can obtain the inverse of \({\bf S}\) explicitly for all the cases considered in the present paper.
\[{\bf S}^{-1}{\bf HC}=\epsilon{\bf C} \tag{8}\]
## 3 Results
For brevity we write both the exact and approximate RR eigenvalues and eigenfunctions as \(\epsilon_{1}<\epsilon_{2}<\ldots\) and \(\psi_{1},\psi_{2},\ldots\), respectively. We express the rate of convergence of the RR results in terms of the largest parameter \(K=i+j\) in the wavefunction expansion, where \(i\) and \(j\) are the exponents of \(r\) and \(\cos\theta\) in either equation or. Thus, for a given value of \(K\) there are \(N=(K^{2}+K+2)/2\) basis functions of either even or odd symmetry. The optimal value of the nonlinear variational parameter \(\alpha\) depends on both the chosen eigenvalue and the number of terms \(N\) in the wavefunction expansion. For example, shows that for the ground state \(\alpha\) increases with \(K\) oscillating about a straight line. shows the RR eigenvalue \(\epsilon_{1}\) for a range of values of \(K\). The rate of convergence is remarkably greater than the one for the Coulomb basis set.
Table 1 shows the RR eigenvalues obtained with the nonorthogonal basis sets and (\(\epsilon_{n}^{NB}\)) on the one side and with the basis sets of even and odd Coulomb functions (\(\epsilon_{n}^{CB}\)) on the other. We appreciate the following facts: the accuracy of present nonlinear basis set is always greater (\(\epsilon_{n}^{NB}\leq\epsilon_{n}^{CB}\)) in spite of the fact that the RR calculations have been carried out with with \(N=211\) nonorthogonal basis functions (\(K=20\)) and \(N=3600\) Coulomb functions. The discrepancy is greater for the lowest states and, it is less noticeable for the odd ones. We are presently unable to provide a rigorous proof for the last two facts; however, we may conjecture that the omitted continuous spectrum is not so relevant in those cases where there is agreement between the NB and CB variational results. The number of digits in the entries of this table is dictated by comparison purposes and does not reflect the estimated accuracy of the calculation.
Both the analytical calculation of the matrix elements and the analytical inversion of the matrix \({\bf S}\) are time consuming but we do them only once for all the states. On the other hand, the optimization of the variational parameter \(\alpha\) for every state is a time consuming calculation that we should repeat several times. Earlier and present calculations suggest that the RR variational method is less efficient for the ground state.
For this reason we have attempted variational calculations with considerably greater basis sets only for this state. For example, we have obtained \(\epsilon_{1}^{NB\;even}=-0.13774677227\) with \(N=466\) NB functions (\(K=30\)) and \(\epsilon^{NB\;even}=-0.13774778205\) with \(N=821\) ones (\(K=40\)). These results suggest that the first 6 digits remain stable and, consequently, that the RR calculation with the Slater-type basis functions does not appear to approach the RSDM results any more closely. However, it is worth noticing that present RR eigenvalues agree with the RSDM ones within the 2% error estimated by Dasbiswas et al.
By means of the approximate ground-state wavefunction \(\psi_{1}(x,y)\) in terms of the Slater-like orbitals we have also calculated the effective dimensionless coupling constant
\[g=\int dxdy|\psi_{1}(x,y)|^{4} \tag{9}\]
Dasbiswas et al obtained \(g=0.017\) by means of a simple variational function constructed from the first three elements of the even nonorthogonal basis set: \(\{e^{-\alpha r},re^{-\alpha r},r\cos\theta e^{-\alpha r}\}\) and \(g=0.0194\) by means of the RSDM. Amore obtained \(g=0.017\) by means of a reduced RR basis set of Coulomb functions (\(N=345\)). Note that just \(N=3\) NB functions yield the same result as \(N=345\) Coulomb ones. The RR method with \(N=211\) functions (\(K=20\)) of the basis set yields \(g=0.0193\) that is quite close to the RSDM result.
Figures 3, 4 (left panels), 5 and 6 show the contour plots for \(\psi(x,y)^{2}\) for the first five even and odd states obtained by means of the \(N=211\) NB functions. One clearly realizes that there are two types of nodal lines \(\psi(x,y)^{2}=0\) and that the energy depends differently on each of them. The right panels of Figures 3 and 4 show 3D plots of the probability densities of the first even and odd states.
## 4 Conclusions
Present results clearly show that the basis set of Slater-type orbitals is preferable to the Coulomb basis set. With just a few functions of the former set one obtains results that are considerably more accurate than those arising from much larger sets of the latter. As argued above, the reason is that any linear combination of discrete-spectrum Coulomb eigenfunctions is orthogonal to the continuous-spectrum eigenfunctions. On the other hand, no continuous-spectrum functions are required when using the Slater-type basis set. The contribution of the continuous spectrum appears to be more relevant for the lowest states and for the even ones. However, we have not proved that the RR variational method with the Slater-type functions converges towards the actual eigenvalues as \(K\) increases. Present variational eigenvalues converge to limits that are slightly larger than the RSDM ones and we cannot safely state that all the stable digits of our results agree with those of the exact eigenvalues. Accurate lower bounds are required for that purpose and we have not yet been able to obtain them. In spite of this fact, it is encouraging that present RR results agree with the RSDM ones within the reported 2% accuracy of the latter. If, as argued by Dasbiswas et al, the RR variational method is more convenient than the RSDM for highly excited states, then present contribution is relevant because there is no doubt that the basis set proposed in this paper is preferable to the Coulomb one.
P.A. acknowledges support of Conacyt through the SNI fellowship and also of PIFI. F.M.F acknowledges support of PIFI and of UNLP through the "subsidio para vajes y/o estadias"
|
10.48550/arXiv.1209.0195
|
Bound states for the quantum dipole moment in two dimensions
|
Paolo Amore, Francisco M Fernández
| 3,558
|
10.48550_arXiv.2411.00990
|
### Active-space
Modeling electronic structure problems with quantum computing is a complex task, primarily due to the linear scaling of the necessary qubits with the number of spin orbitals, which is twice the number of spatial orbitals in the basis set. This challenge is further compounded by the need to use small basis sets and the active space approximation, which involves selecting a subset of Molecular Orbitals (MOs) for the calculation. The classification of molecular orbitals into core, active, and external orbitals adds another layer of complexity, with core orbitals typically fully occupied by two electrons, active orbitals potentially occupied by zero, one, or two electrons, and external orbitals generally not included in the calculation. Since simulating the entire molecule \(C_{7}H_{6}O_{4}\) requires 122 qubits, a number that exceeds the capabilities of our simulator (which can handle up to 35 qubits), we have reduced the size of the system through active-space selection.
In our investigation, we are focusing on a crucial aspect of the acid group's conjugation with the aromatic ring, leading to internal resonance.
Chemical structure of protocatechuic acid (\(C_{7}H_{6}O_{4}\)), also known as 3,4-dihydroxybenzoic acid. The molecule consists of a benzene ring substituted with two hydroxyl groups (-OH) at the 3rd and 4th positions and a carboxyl group (-COOH) at the 1st position. This arrangement classifies it as a dihydroxybenzoic acid, contributing to its antioxidant properties commonly found in various plants. C: brown, O: red, H: white.
Our initial focus is on a 4-electron, 4-orbital active space corresponding to an 8-qubit case. We are also exploring larger active spaces within the capacity of our hardware, considering 6 and 8 orbitals for the active space, which correspond to 12 and 16 spin orbitals, represented by 12 and 16 qubits, respectively.
|
10.48550/arXiv.2411.00990
|
Exploiting the Variational Quantum Eigensolver for Determining Ground State Energy of Protocatechuic Acid
|
Gleydson Fernandes de Jesus, Erico Souza Teixeira, Lucas Queiroz Galvão, Maria Heloísa Fraga da Silva, Mauro Queiroz Nooblath Neto, Bruno Oziel Fernandez, Clebson dos Santos Cruz
| 4,743
|
10.48550_arXiv.2407.18842
|
## Appendix C RT-NEO-TDDFT Ehrenfest Dynamics with Traveling Proton Basis
#### c.3.1 Traveling Proton Basis Approach
In RT-NEO-TDDFT Ehrenfest dynamics, the classical nuclei move according to the EOM exerted by the time-dependent density in the context of RT-NEO-TDDFT. Conventionally, the basis function centers for the electrons are centered on the atomic nuclei and therefore move with the nuclei. In principle, the basis function centers corresponding to quantum protons can stay stationary even in RT-NEO-TDDFT Ehrenfest dynamics. When the quantum protons move far away from their original location, however, a single localized basis function center for each quantum proton is not sufficient. One strategy is to introduce multiple localized basis functions centered on fictitious "ghost" atoms. However, this scheme is not always convenient because an approximate trajectory of the quantum proton needs to be known ahead of time to keep the number of extra proton basis function centers manageable.
Alternatively, Zhao _et al._ proposed an approach called the traveling proton basis (TPB) such that fictitious classical degrees of freedom are added to propagate the basis function centers for the quantum protons. In this original TPB approach, the equations of motion for the basis functions associated with the quantum protons are given by the energy gradient with respect to their basis function center positions using a fictitious proton mass. This TPB idea entails extending the original NEO Lagrangian (Eq. 1) as
\[L^{\text{NEO-TPB}}(t,\mathbf{R},\dot{\mathbf{R}})=L^{\text{NEO}}(t,\mathbf{R},\dot{\mathbf{R}}) +\sum_{N}\frac{1}{2}M^{p}\dot{\mathbf{R}_{N}^{p}}^{2}, \tag{19}\]
For brevity, we omit the explicit dependence of the Lagrangian on the proton and electron KS wavefunctions here. Given the Lagrangian of the system, the classical EOM (analogous to Eq. 18) is formally given by the Euler-Lagrange equation (Eq. 8) which becomes
\[M^{p} \ddot{\mathbf{R}}_{N}^{p}(t)=-\frac{\partial E_{\text{tot}}}{\partial \mathbf{R}_{N}^{p}(t)}\] \[+\sum_{n,ijlm}\int\text{d}\mathbf{k}c_{n\mathbf{k},i}^{e}\,{}^{*} \left(\mathbf{B}_{N,il}^{e}{}^{*}\text{S}_{lm}^{e}{}^{-1}\text{H}_{mj}^{e}+ \text{H}_{il}^{e}\text{S}_{lm}^{e}{}^{-1}\mathbf{B}_{N,mj}^{e}\right)c_{n \mathbf{k},j}^{e}\] \[+\sum_{n,ijlm}c_{n,i}^{p}\,{}^{*}\left(\mathbf{B}_{N,il}^{p}\,{} ^{*}\text{S}_{lm}^{p}{}^{-1}\text{H}_{mj}^{p}+\text{H}_{il}^{p}\text{S}_{lm} ^{p}{}^{-1}\text{B}_{N,mj}^{p}\right)c_{n,j}^{p}\] \[+i\sum_{n,ijlm}\int\text{d}\mathbf{k}c_{n\mathbf{k},i}^{e}\,{}^{*} \left(\text{B}_{il}^{e}{}^{*}\text{S}_{lm}^{e}{}^{-1}\text{B}_{N,mj}^{e}- \mathbf{B}_{N,il}^{e}{}^{*}\text{S}_{lm}^{e}{}^{-1}\text{B}_{mj}^{e}+\langle \frac{\partial\phi_{i}^{e}}{\partial\mathbf{R}_{N}^{p}}|\frac{\text{d}\phi_{j}^{e} }{\text{d}t}\rangle-\langle\frac{\text{d}\phi_{i}^{e}}{\text{d}t}|\frac{ \partial\phi_{j}^{e}}{\partial\mathbf{R}_{N}^{p}}\rangle\right)c_{n\mathbf{k},j}^{e}\] \[+i\sum_{n,ijlm}c_{n,i}^{p}\,{}^{*}\left(\text{B}_{il}^{p}{}^{*} \text{S}_{lm}^{p}{}^{-1}\text{B}_{N,mj}^{p}-\mathbf{B}_{N,il}^{p}{}^{*}\text{S }_{lm}^{p}{}^{-1}\text{B}_{mj}^{p}+\langle\frac{\partial\phi_{i}^{p}}{\partial \mathbf{R}_{N}^{p}}|\frac{\text{d}\phi_{j}^{p}}{\text{d}t}\rangle-\langle\frac{ \text{d}\phi_{i}^{p}}{\text{d}t}|\frac{\partial\phi_{j}^{p}}{\partial\mathbf{R}_{N }^{p}}\rangle\right)c_{n,j}^{p}, \tag{20}\]
In the original TPB scheme, the EOM was given by the derivative of the total energy and thus the four additional terms do not appear. As discussed above for Eq. 18, the last two summation terms with \(i\sum_{n,ijlm}\) are omitted in our implementation. In the following discussions, we refer to this Lagrangian-derived TPB scheme as l-TPB scheme. The pros and cons of the TPB and the new Lagrangian-derived l-TPB schemes are discussed using a numerical example in Sec. III.1. In both cases, the fictitious mass of the proton basis function center is typically chosen such that the position expectation value corresponding to the quantum proton density and the proton basis function center closely follow each other. The physical proton mass is shown to be the correct mass for describing the dynamics of the expectation value of the proton position in a harmonic potential.Alternative Formulation of TPB Approach
In addition to the original TPB approach discussed above, herein we propose an alternative TPB approach by applying a semiclassical approximation to the Lagrangian instead of introducing an extra classical kinetic energy term for the quantum proton basis function center. For a given proton KS orbital \(n\), its translational degree of freedom can be expressed using the explicit phase with the momentum \(\mathbf{k}_{n}^{p}\),
\[\psi_{n}^{p}(\mathbf{r}^{p},t)=e^{i\mathbf{k}_{n}^{p}\cdot\mathbf{r}^{p}}\psi_{n}^{\prime p }(\mathbf{r}^{p},t). \tag{21}\]
This does not alter the proton density (i.e., \(\rho^{p}(\mathbf{r}^{p},t)=\rho^{p\prime}(\mathbf{r}^{p},t)\)) nor the quantum dynamics in RT-TDDFT.
\[\begin{split}\langle\psi_{n}^{p}|\,\frac{1}{2M^{p}}\nabla_{\mathbf{r} ^{p}}^{2}\,|\psi_{n}^{p}\rangle=&\quad\langle\psi_{n}^{\prime p} |\,\frac{1}{2M^{p}}\nabla_{\mathbf{r}^{p}}^{2}\,|\psi_{n}^{\prime p}\rangle-\frac{ \mathbf{k}_{n}^{p}{}^{2}}{2M^{p}}+i\frac{\mathbf{k}_{n}^{p}}{M^{p}}\,\langle\psi_{n}^{ \prime p}|\,\nabla_{\mathbf{r}^{p}}\,|\psi_{n}^{\prime p}\rangle\\ \langle\psi_{n}^{p}|\,i\frac{\mathrm{d}}{\mathrm{d}t}\,|\psi_{n}^ {p}\rangle=&\quad-\,\mathbf{k}_{n}^{p}\,\langle\psi_{n}^{\prime p}|\, \hat{\mathbf{r}}^{p}\,|\psi_{n}^{\prime p}\rangle+\langle\psi_{n}^{\prime p}|\,i \frac{\mathrm{d}}{\mathrm{d}t}\,|\psi_{n}^{\prime p}\rangle\end{split} \tag{22}\]
Then the NEO Lagrangian (Eq. 1) reads
\[L^{\mathrm{NEO}}= L^{\prime\mathrm{NEO}}-\sum_{n,p}\left(\mathbf{k}_{n}^{p}\,\langle \psi_{n}^{\prime p}|\,\hat{\mathbf{r}}^{p}\,|\psi_{n}^{\prime p}\rangle+\frac{\mathbf{k }_{n}^{p}{}^{2}}{2M^{p}}-i\frac{\mathbf{k}_{n}^{p}}{M_{p}}\,\langle\psi_{n}^{\prime p }|\,\nabla_{\mathbf{r}^{p}}\,|\psi_{n}^{\prime p}\rangle\right) \tag{23}\]
1 but expressed in terms of \(\psi_{n}^{\prime p}\) instead of \(\psi_{n}^{p}\) (see Eq. 21).
Next we make a semiclassical approximation to the translational degree of freedom associated with the proton KS orbital so that it can be treated separately as a classical degree of freedom in the RT-NEO-TDDFT Ehrenfest dynamics in the same spirit as in the original TPB approach. This entails substituting \(\mathbf{k}_{n}^{p}\) with the classical momentum \(\mathbf{k}_{n}^{p}=M^{p}\mathbf{k}_{N}^{p}\), where we assume that the proton orbital \(\psi_{n}^{p}\) is localized on the proton basis function center \(\mathbf{R}_{N}^{p}\). In making this semiclassical approximation, we must take into account that the Lagrangian is defined with a negative sign for the quantum mechanical kinetic energy and with a positive sign for the classical mechanical kinetic energy. Therefore, the sign of this kinetic energy term is changed as we make this semiclassical approximation to the translational kinetic energy associated with the proton KS orbitals. Additionally, this semiclassical approximation results in a term with \(\ddot{\mathbf{R}}_{N}^{p}\) in the corresponding Lagrangian. Because such a higher derivative term in the Lagrangian is known to lead to instability (i.e., Ostrogradskyinstability), we simply neglect the \(\mathbf{R}_{N}^{p}\) dependent term, \(-M^{p}\dot{\mathbf{R}_{N}^{p}}\left\langle\psi_{n}^{p\prime}\right|\hat{\mathbf{r}}^{p} \left|\psi_{n}^{p\prime}\right\rangle\).
\[L^{\text{NEO-sc-TPB}}= L^{\prime\text{NEO}}+\sum_{N}\left(\frac{1}{2}M^{p}\dot{\mathbf{R}_{N }^{p}}^{2}-i\dot{\mathbf{R}_{N}^{p}}\sum_{n}\left\langle\psi_{n}^{p\prime}\right| \nabla_{\mathbf{R}_{N}^{p}}\left|\psi_{n}^{p\prime}\right\rangle\right). \tag{24}\]
In this new NEO-sc-TPB Lagrangian obtained by converting the translational degrees of freedom for the proton KS orbitals to classical degrees of freedom, the EOM for the proton KS orbital coefficients and the classical dynamics of the proton basis function centers can be derived via the Euler-Lagrange equation in the usual manner. The EOM for the quantum proton degrees of freedom (corresponding to Eq. 14) is
\[\dot{c}_{n,i}^{p}=-i\sum_{jl}\text{S}_{ij}^{p}{}^{-1}\text{H}_{jl}^{p}c_{n,l}^ {p}. \tag{25}\]
The EOM for the proton basis function centers, which are now classical degrees of freedom, is
\[M^{p} \dot{\mathbf{R}_{N}^{p}}(t)=-\frac{\partial E_{\text{tot}}}{\partial \mathbf{R}_{N}^{p}(t)}\] \[+\sum_{n,ijlm}\int\text{d}\mathbf{k}c_{n\mathbf{k},i}^{e}{}^{*} \left(\mathbf{B}_{N,il}^{e}{}^{*}\text{S}_{lm}^{e}{}^{-1}\text{H}_{mj}^{e}+ \text{H}_{il}^{e}\text{S}_{lm}^{e}{}^{-1}\text{B}_{N,mj}^{e}\right)c_{n\mathbf{ k},j}^{e}\] \[+i\sum_{n,ijlm}\int\text{d}\mathbf{k}c_{n\mathbf{k},i}^{e}{}^{*} \left(\text{B}_{il}^{e}{}^{*}\text{S}_{lm}^{e}{}^{-1}\text{B}_{N,mj}^{e}- \mathbf{B}_{N,il}^{e}{}^{*}\text{S}_{lm}^{e}{}^{-1}\text{B}_{mj}^{e}+\left\langle \frac{\partial\phi_{i}^{e}}{\partial\mathbf{R}_{N}^{p}}\right|\frac{\text{d}\phi_{ j}^{e}}{\text{d}t}\right\rangle-\left\langle\frac{\text{d}\phi_{i}^{e}}{\text{d}t} \frac{\partial\phi_{j}^{e}}{\partial\mathbf{R}_{N}^{p}}\right\rangle\right)c_{n \mathbf{k},j}^{e}. \tag{26}\]
Note that the terms with proton nonadiabatic couplings (the \(B_{jl}^{p}\) term in Eq. 13 and the \(i\dot{\mathbf{R}_{N}^{p}}\sum_{n}\left\langle\psi_{n}^{p\prime}\right|\nabla_{\mathbf{ R}_{N}^{p}}\left|\psi_{n}^{p\prime}\right\rangle\) term in Eq. 24) cancel each other for propagating both the proton wavefunctions and the proton basis function centers in this alternative formulation of the TPB approach.
Let us now discuss the issue of energy conservation.
\[H\equiv\sum_{i}\dot{q}_{i}\frac{\partial L}{\partial\dot{q}_{i}}-L, \tag{27}\]
In mixed quantum-classical systems, we have \(q_{i}=(\{\mathbf{R}_{I}\},\{\psi_{i}^{*}\})\). In most cases, the Lagrangian has no explicit dependence on time (i.e. \(\partial_{t}L=0\)) and the constant of motion is therefore given by the corresponding Hamiltonian. However, in the RT-NEO-TDDFT approach, the term \(i\partial_{t}(=\hat{H}^{KS}(t))\) in the Lagrangian (Eq. 1) leads to \(\partial_{t}L\neq 0\) because of its explicit time dependence, i.e., for a general electronic or protonic orbital, \(\left\langle\psi_{i}(t)\right|\hat{H}^{KS}(t)\left|\psi_{i}(t)\right\rangle= \epsilon_{i}^{KS}(t)\). Thus, the conserved quantity is no longer given by the Hamiltonian alone. For the NEO Lagrangian defined in Eq. 1, the conserved quantity is given by
\[\frac{dH}{dt}+\frac{\partial L^{NEO}}{\partial t}=\frac{d}{dt}(E_{\text{tot}}+ \sum_{N}\frac{1}{2}M^{p}\mathbf{\dot{R}_{N}}^{p}{}^{2})=0. \tag{28}\]
In practical simulations, a few numerical simplifications are used, as discussed above, such as neglecting the velocity-dependent terms in Eq. 18. The energy conservation is presented and discussed in Sec. III.1. The semiclassical approximation of the TPB approach discussed in this section is referred to as the sc-TPB approach for simplicity.
### Numerical Implementation
#### iii.4.1 Energy Derivatives
In our previous work, the total energy for multicomponent NEO-DFT (as defined by Eq. 16) is calculated with a NEO-induced correction to the conventional DFT energy,
\[\begin{split} E_{\text{tot}}^{\text{NEO}}=&\sum_{l} f_{l}\epsilon_{l}^{\text{NEO}}-\int v_{xc}^{e-\text{NEO}}(\mathbf{r}^{e})\rho^{e}( \mathbf{r}^{e})\text{d}\mathbf{r}^{e}+E_{xc}^{e-\text{NEO}}[\rho^{e}(\mathbf{r}^{e})]-\\ &\frac{1}{2}\int v_{\text{es}}^{e-\text{NEO}}(\mathbf{r}^{e})\rho^{e} (\mathbf{r}^{e})\text{d}\mathbf{r}^{e}+E_{\text{nuc-nuc}}+\Delta E_{\text{nuc-nuc}}^{ \text{NEO}},\end{split} \tag{29}\]
\(v_{xc}^{e-\text{NEO}}\) and \(E_{xc}^{e-\text{NEO}}\) are the exchange-correlation potential and energy of the NEO electron KS Hamiltonian, which includes both electron-electron exchange-correlation and electron-proton correlation. The fourth term with \(v_{\text{es}}^{e-\text{NEO}}\) corrects for the double counting in the electrostatic interactions. \(E_{\text{nuc-nuc}}\) is the energy contributed from interactions between classical nuclei, as in conventional DFT.
\[\Delta E_{\text{nuc-nuc}}^{\text{NEO}}=T^{p}+\frac{1}{2}\left(J^{p}-K^{p} \right)-E^{cpp}+\sum_{I}^{\text{nuclei}}Z_{I}\left[V_{\text{es}}^{p}(\mathbf{R}_{ I})-V_{\text{es}}^{cp}(\mathbf{R}_{I})\right], \tag{30}\]
In the last term, \(V_{\text{es}}^{p}\) represents the Coulombinteraction between the quantum protons and the classical nuclei \(I\) with charge \(Z_{I}\), and \(V_{\rm es}^{cp}\) is the analogous term in conventional DFT calculations to remove double counting. RT-NEO-TDDFT Ehrenfest dynamics requires the evaluation of the partial derivative of the NEO total energy \(E_{\rm tot}^{\rm NEO}\) with respect to the classical nuclear coordinate \(\mathbf{R}_{I}\) (see Eq. 18), defining the force on nucleus \(I\) as
\[\mathbf{F}_{I}=-\frac{\partial}{\partial\mathbf{R}_{I}}E_{\rm tot}^{\rm NEO}=-\frac{ \partial}{\partial\mathbf{R}_{I}}(E^{\rm e\text{-}NEO}+\Delta E_{\rm nuc\text{-} nuc}^{\rm NEO}). \tag{31}\]
This force can be classified as having an electron contribution and a quantum proton contribution.
The electron contribution \(\mathbf{F}_{I}^{e}=-\partial E^{\rm e\text{-}NEO}/\partial\mathbf{R}_{I}\) has the same form as the conventional DFT force. For example, in the FHI-aims code, the forces are expressed as having several components including Hellmann-Feynman forces \(\mathbf{F}_{I}^{\rm HF}\), Pulay forces \(\mathbf{F}_{I}^{\rm P}\), electrostatic multipole derivatives \(\mathbf{F}_{I}^{\rm MP}\), scalar-relativistic derivatives \(\mathbf{F}_{I}^{\rm at.ZORA}\), and so forth. In our periodic NEO-DFT implementation, these terms are evaluated by the FHI-aims package in the same manner as for conventional DFT, but using the updated electron density \(\rho^{e}\) and effective potential \(v_{\rm eff}^{e}\) that include the NEO effects.
The quantum proton contribution of \(\mathbf{F}_{I}^{p}=-\partial\Delta E_{\rm nuc\text{-}nuc}^{\rm NEO}/\partial\mathbf{R }_{I}\) needs to be explicitly evaluated for the individual terms in \(\Delta E_{\rm nuc\text{-}nuc}^{\rm NEO}\) (Eq. 30). A combination of numerical real-space grid integration and analytical integration is used to evaluate the proton Hamiltonian as discussed in Ref., and the energy gradients are also computed accordingly. For a given quantum proton basis function center \(\mathbf{R}_{N}^{p}\), the gradient of the kinetic energy \(T^{p}\) and exchange energy \(K^{p}\) (see Eq. 30) are evaluated with analytical integrals,
\[\frac{\partial T^{p}}{\partial\mathbf{R}_{N}^{p}}= \sum_{ij}\left[\frac{\partial\text{D}_{ij}^{p}}{\partial\mathbf{R}_{ N}^{p}}\text{T}_{ij}^{p}+\text{D}_{ij}^{p}\frac{\partial\text{T}_{ij}^{p}}{ \partial\mathbf{R}_{N}^{p}}\right], \tag{32}\] \[\frac{\partial K^{p}}{\partial\mathbf{R}_{N}^{p}}= \sum_{ijkl}\left[2\frac{\partial\text{D}_{ij}^{p}}{\partial\mathbf{R }_{N}^{p}}\left(ik|jl\right)\text{D}_{kl}^{p}+\text{D}_{ij}^{p}\frac{\partial \left(ik|jl\right)}{\partial\mathbf{R}_{N}^{p}}\text{D}_{kl}^{p}\right], \tag{33}\]
12, and \((ik|jl)\) is the 'two electron integral'. All the above stated analytical integrals are evaluated using the libcint package. The electrostatic energy is evaluated with numerical integrals. For faster convergence with respect to distance, the electrostatic potential of a quantum proton orbital \(\psi_{n}^{p}\) and the corresponding classical proton at \(\mathbf{R}_{N}^{p}\) are always evaluated together as a correction potential \(V_{\text{es-corr}}^{p}\left(\mathbf{R}_{N}^{p};\mathbf{r}\right)=V_{\text{es},N}^{p}\left( \mathbf{R}_{N}^{p}-\mathbf{r}\right)-1/|\mathbf{R}_{N}^{p}-\mathbf{r}|\), with \(V_{\text{es},N}^{p}\) denoting the electrostatic potential of the quantum proton centered on \(\mathbf{R}_{N}^{p}\).
As a part of \(E_{\text{nuc-nuc}}^{\text{NEO}}\) (Eq. 30), we define \(E_{\text{es}}^{p}\) as the sum of the electrostatic energies contributed by the quantum protons and the classical protons
\[E_{\text{es}}^{p}=\frac{1}{2}J^{p}-E^{cpp}=\frac{1}{2}\int d\mathbf{r}\rho^{p}(\bm {r})\sum_{N}V_{\text{es-corr}}^{p}\left(\mathbf{R}_{N}^{p};\mathbf{r}\right)+\frac{1}{ 2}\sum_{N,N^{\prime}}V_{\text{es-corr}}^{p}\left(\mathbf{R}_{N}^{p};\mathbf{R}_{N^{ \prime}}^{p}\right). \tag{34}\]
The gradient of the above energy contribution \(E_{\text{es}}^{p}\) can be evaluated as
\[\begin{split}\frac{\partial E_{\text{es}}^{p}}{\partial\mathbf{R}_{N }^{p}}=&\sum_{ij,N^{\prime}}\frac{\partial\text{D}_{ij}^{p}}{ \partial\mathbf{R}_{N}^{p}}\int d\mathbf{r}\phi_{i}^{p}(\mathbf{r})\phi_{j}^{p}(\mathbf{r})V_{ \text{es},N^{\prime}}^{p}\left(\mathbf{R}_{N^{\prime}}^{p}-\mathbf{r}\right)+\sum_{ij, N^{\prime}}\text{D}_{ij}^{p}\int d\mathbf{r}\phi_{i}^{p}(\mathbf{r})\frac{\partial \phi_{j}^{p}(\mathbf{r})}{\partial\mathbf{R}_{N}^{p}}V_{\text{es-corr}}^{p}\left(\mathbf{R} _{N^{\prime}}^{p};\mathbf{r}\right)\\ &+\frac{1}{2}\int d\mathbf{r}\rho^{p}\left(\mathbf{r}\right)\left.\frac{ \partial V_{\text{es-corr}}^{p}\left(\mathbf{R}_{N}^{p};\mathbf{r}\right)}{\partial\bm {R}_{N}^{p}}\right|_{D^{p}}+\frac{1}{2}\sum_{N^{\prime}\neq N}\frac{\partial V _{\text{es-corr}}^{p}\left(\mathbf{R}_{N^{\prime}}^{p};\mathbf{R}_{N}^{p}\right)}{ \partial\mathbf{R}_{N}^{p}}.\end{split} \tag{35}\]
The gradient of the correction potential \(\partial V_{\text{es-corr}}^{p}/\partial\mathbf{R}_{N}^{p}\) is evaluated numerically on real space grids using a multipole expansion. Similarly, the gradient of the electrostatic energy between the quantum protons and the classical nuclei, i.e., the last term in Eq. 30, can be written as:
\[\frac{\partial E_{\text{nuc}}^{p}}{\partial\mathbf{R}_{N}^{p}}=\sum_{ij,I}\frac{ \partial\text{D}_{ij}^{p}}{\partial\mathbf{R}_{N}^{p}}\int d\mathbf{r}\phi_{i}^{p}(\bm {r})\phi_{j}^{p}(\mathbf{r})\frac{Z_{I}}{|\mathbf{R}_{N}^{p}-\mathbf{R}_{I}|}+\sum_{I}Z_{I} \left.\frac{\partial V_{\text{es-corr}}^{p}\left(\mathbf{R}_{N}^{p};\mathbf{R}_{I} \right)}{\partial\mathbf{R}_{N}^{p}}\right|_{\text{D}^{p}}. \tag{36}\]
For the classical nuclei \(I\), the gradients of the first four terms in Eq.30 are zero, leading to
\[\frac{\partial\Delta E_{\text{nuc-nuc}}^{\text{NEO}}}{\partial\mathbf{R}_{I}}= \frac{\partial E_{\text{nuc}}^{p}}{\partial\mathbf{R}_{I}}=Z_{I}\sum_{N}\frac{ \partial V_{\text{es-corr}}^{p}\left(\mathbf{R}_{N}^{p};\mathbf{R}_{I}\right)}{ \partial\mathbf{R}_{I}}. \tag{37}\]
As discussed in our recent work, the electrostatic potential for the quantum proton \(V_{\text{es}}^{p}\) and electron-proton correlation potential \(V_{\text{epc}}\) are passed to the electron Hamiltonian \(\hat{H}^{e}\) in the numerical implementation for evaluating the total energy. Thus, \(E_{\text{es}}^{e}\) and \(E_{\text{epc}}\) are evaluated as a part of the electron contribution \(E^{\text{e-NEO}}\) in Eq. 29. Because these two terms have contributions from both the electrons and the quantum protons, their energy gradients with respect to both the electron and proton degrees of freedom need to be addressed separately.
\[\frac{\partial E_{\text{es}}^{e}}{\partial\mathbf{R}_{N}^{p}}= \sum_{ij}\frac{\partial\text{D}_{ij}^{p}}{\partial\mathbf{R}_{N}^{p}} \int d\mathbf{r}\phi_{i}^{p}(\mathbf{r})\phi_{j}^{p}(\mathbf{r})V_{\text{es}}^{e}\left(\bm {r}\right)+\int d\mathbf{r}\rho^{e}\left(\mathbf{r}\right)\left.\frac{\partial V_{\text {es-corr}}^{p}\left(\mathbf{R}_{N}^{p};\mathbf{r}\right)}{\partial\mathbf{R}_{N}^{p}}\right| _{\text{D}^{p}}, \tag{38}\] \[\frac{\partial E_{\text{epc}}}{\partial\mathbf{R}_{N}^{p}}= \sum_{ij}\frac{\partial\text{D}_{ij}^{p}}{\partial\mathbf{R}_{N}^{p}} \int d\mathbf{r}\phi_{i}^{p}(\mathbf{r})\phi_{j}^{p}(\mathbf{r})V_{\text{epc}}^{p}\left( \mathbf{r}\right)+\sum_{ij}2\text{D}_{ij}^{p}\int d\mathbf{r}\phi_{i}^{p}(\mathbf{r}) \frac{\partial\phi_{j}^{p}(\mathbf{r})}{\partial\mathbf{R}_{N}^{p}}V_{\text{epc}}^{p} \left(\mathbf{r}\right), \tag{39}\]
In summary, Eqs. 32-39 provide all the necessary expressions for evaluating the total energy gradient within the RT-NEO-TDDFT scheme. Note that, unlike the total derivative, the partial derivative \(\sum_{ij}\partial\text{D}_{ij}^{p}/\partial\mathbf{R}_{N}^{p}=0\), because the proton density matrix does not have an explicit dependence on the coordinates of the quantum proton basis function center, and this simplifies several equations for implementation. See the Appendix for further information relevant to the ground-state NEO-DFT calculations.
#### ii.1.2 TD-KS Orbitals in RT-NEO-TDDFT Ehrenfest Dynamics
As discussed in Sec. II.1, the dynamics of the electron and proton KS wave functions are governed by the coupled TD-KS equations within the RT-NEO-TDDFT framework,
\[i\frac{\text{d}}{\text{d}t}\psi_{i,\mathbf{k}}^{e}(\mathbf{r}^{e},t)= \hat{H}^{e}\psi_{i,\mathbf{k}}^{e}(\mathbf{r}^{e},t)=\left[-\frac{1}{ 2}\nabla^{2}+U_{\text{eff}}^{e}(\mathbf{r}^{e},t)+\hat{v}_{\text{ext}}(t)\right] \psi_{i,\mathbf{k}}^{e}(\mathbf{r}^{e},t) \tag{40}\] \[i\frac{\text{d}}{\text{d}t}\psi_{i}^{p}(\mathbf{r}^{p},t)= \hat{H}^{p}\psi_{i}^{p}(\mathbf{r}^{p},t)=\left[-\frac{1}{2M^{p}} \nabla^{2}+U_{\text{eff}}^{p}(\mathbf{r}^{p},t)-\hat{v}_{\text{ext}}(t)\right]\psi _{i}^{p}(\mathbf{r}^{p},t). \tag{41}\]
6. Although the exchange-correlation potentials depend on all former time in principle, we adopt the adiabatic approximation that ignores this history dependence and thus removes the explicit time dependence of \(U_{\text{eff}}\). Because the effective potential contains coupled contributions from the electrons and protons, the electron and proton orbitals need to be propagated at the same time. When needed, \(\hat{v}_{\text{ext}}(t)\) can describe the (time-dependent) external electromagnetic field, which can be represented either in the length gauge \(\hat{v}_{\text{ext}}(t)=\hat{\mathbf{r}}\cdot\mathbf{E}(t)\) or in the velocity gauge \(\hat{v}_{\text{ext}}(t)=-i\mathbf{A}(t)\cdot\nabla+\mathbf{A}(t)^{2}\) with \(\mathbf{E}(t)=-\partial_{t}\mathbf{A}(t)\) denoting the external electric field and \(\mathbf{A}(t)\) denoting the corresponding vector potential.
The electron KS wave functions are propagated according to the electronic TD-KS equation using the existing RT-TDDFT module in FHI-aims with the NEO-modified effective potential as described in Ref.. The electronic effective potential is updated on-the-fly following the NEO-DFT approach to include NEO contributions. The proton degrees of freedom are propagated as follows in the RT-NEO-TDDFT Ehrenfest dynamics. As described in Eqs. 9 and 10 in Sec. II.2, the time-dependent proton KS orbitals are expanded using a set of periodic Gaussian-type basis functions, each of which are associated with a center \(I\).
\[\frac{\mathrm{d}}{\mathrm{d}t}\mathbf{C}(t)=-i\mathbf{S}^{-1}\mathbf{H}(t) \mathbf{C}(t), \tag{42}\]
12. This differential equation can be solved using an exponential integration scheme, i.e., by propagating the proton KS wave functions with a given time step \(\Delta t\): \(\mathbf{C}(t+\Delta t)=\exp(\hat{O})\mathbf{C}(t)\). The \(\hat{O}\) here is any generic quantum mechanical operator.
\[\mathbf{C}(t+\Delta t)=\mathbf{S}^{-\frac{1}{2}}\mathrm{exp}\left(-i\Delta t \mathbf{S}^{-\frac{1}{2}}\mathbf{H}(t+\frac{\Delta t}{2})\mathbf{S}^{-\frac{1 }{2}}\right)\mathbf{S}^{\frac{1}{2}}\mathbf{C}(t). \tag{43}\]
Since the TD-KS equations are a set of non-linear equations, where the Hamiltonian depends on the KS wave functions, an iterative predictor/corrector method is used to obtain the Hamiltonian matrix at time \(t+\frac{\Delta t}{2}\).
\[\mathrm{exp}\left(\mathbf{A}\right)=\mathbf{V}\mathrm{diag}\left(e^{\lambda_{ 1}},...e^{\lambda_{n}}\right)\mathbf{V}^{-1}, \tag{44}\]
Alternative approaches, such as the explicit fourth-order Runge-Kutta method, are also implemented for solving Eq. 42.
In Ehrenfest dynamics, the nonadiabatic couplings between the quantum electrons/protons and the classical nuclei need to be computed. With the TPB approaches, the positions of the classical nuclei, as well as the positions of the basis function centers for the quantum protons, are propagated on the fly as discussed in Sec. II.3. The electron and proton KS wave functions are propagated according to the EOM in Eq. 14 and/or 25, depending on the particular TPB scheme chosen. In the case of the sc-TPB scheme, the EOM for the quantum proton is given by Eq. 25. In the case of the TPB and l-TPB schemes, the nonadiabatic coupling (NAC) term is needed for the quantum proton propagation, as given by Eq. 14, yielding
\[\frac{\mathrm{d}}{\mathrm{d}t}\mathbf{C}(t)=-i\mathbf{S}^{-1}[\mathbf{H}(t)-i \sum_{I}\dot{\mathbf{\mathbf{\dot{R}}}}_{I}(t)\overrightarrow{\mathbf{\mathbf{\dot{B}} }}_{I}]\mathbf{C}(t), \tag{45}\]
12. Because of the anti-Hermitian nature of the \(\overrightarrow{\mathbf{\mathbf{\dot{B}}}}_{I}\) matrix, adding the NAC term does not change the hermiticity when evaluating the matrix exponential in Eq. 43. The classical EOM for propagating the classical nuclei and the quantum proton basis function centers is given by Eq. 20 for the TPB and l-TPB schemes or Eq. 26 for the sc-TPB scheme. As discussed above, some of the terms are neglected in our implementation. In our Ehrenfest dynamics implementation, the positions of the classical nuclei and quantum proton basis function centers are propagated using the velocity Verlet method, and all NEO contributions are taken into account as an additional effective Newtonian force on top of the conventional Ehrenfest forces as described in Sec. II.4.1.
Algorithm 1 outlines the numerical implementation scheme for the NEO Ehrefest dynamics. Because of the heavier mass of the quantum proton, the time step of the proton KS wave function propagation \(\Delta t_{\mathrm{wf}}^{p}\) can be larger than that associated with the electrons \(\Delta t_{\mathrm{wf}}^{e}\). Similarly, the classical nuclei can be propagated with an even larger time step \(\Delta t_{\mathrm{nuc}}\).
## III Results and Discussion
The above-described RT-NEO-TDDFT with Ehrenfest dynamics using the TPB schemes is implemented in the FHI-aims package. We first validated our implementation against the existing implementation in the Q-Chem code using isolated systems (see Supplementary Material). For all the RT-NEO-TDDFT calculations discussed in this section, we used the PBE exchange-correlation functional for electrons, the exact exchange for quantum protons, and the epc17-2 functional for electron-proton correlation. We used the intermediate numerical atomic orbital (NAO) basis set for electrons and the PB4-F2 basis set for the quantum protons. All NEO-Ehrenfest dynamics simulations used the same electronic/protonic timestep, \(\Delta t_{\mathrm{wf}}^{p}=\Delta t_{\mathrm{wf}}^{e}=0.0048\) fs, and a larger timestep for the classical nuclei, \(\Delta t_{\mathrm{nuc}}=0.048\) fs.
```
0:\(\Delta t_{\text{wf}}^{e}\), \(\Delta t_{\text{wf}}^{p}=N^{p}*\Delta t_{\text{wf}}^{e}\), \(\Delta t_{\text{nuc}}=N^{c}*\Delta t_{\text{wf}}^{e}\) for\(i=1,2,...,\)do\(\triangleright\) Ehrefnest step for\(j=1,2,...,N^{c}\)do\(\triangleright\) Proton step Propagate proton wavefunction \(\psi^{p}(t)\rightarrow\psi^{p}(t+\Delta t_{\text{wf}}^{p})\) based on Eq. 14 or 25 for\(k=1,2,...,N^{p}\)do\(\triangleright\) Electron step Propagate electron wavefunction \(\psi^{e}(t)\rightarrow\psi^{e}(t+\Delta t_{\text{wf}}^{e})\) based on Eq. 14 \(t=t+\Delta t_{\text{wf}}^{e}\) endfor endfor Propagate classical nuclei \(\mathbf{R}_{I}(t)\rightarrow\mathbf{R}_{I}(t+\Delta t_{\text{nuc}})\) based on Eq. 18 if Traveling proton basis then Propagate proton basis function centers \(\mathbf{R}_{N}^{p}(t)\rightarrow\mathbf{R}_{N}^{p}(t+\Delta t_{\text{nuc}})\) based on Eq. 20 or 26 endif Update (numerical/analytical) atomic integrals based on new coordinates endfor
```
## Algorithm 1
NEO Ehrenfest dynamics implemented in this work
### Traveling Proton Basis schemes in RT-NEO-TDDFT Ehrenfest Dynamics
We discuss here the performance of the three TPB schemes discussed in Sec. II.3 by comparing to the standard RT-NEO-TDDFT Ehrenfest dynamics with fixed proton basis function centers (FPB) on "ghost" atoms. In the limit of having enough basis function centers to represent any quantum proton dynamics, the FPB scheme is considered the most rigorous formulation of RT-NEO-TDDFT Ehrenfest dynamics. However, in practice, the FPB scheme can be used only with a finite number of proton basis function centers along a pre-defined spatial region of expected quantum proton dynamics. In the following, we use the results from the sc-TPB simulation to choose the locations of the proton basis function centers for the FPB simulation. In Supplementary Material, we present the FPB simulation results with the proton basis function centers placed according to the original TPB simulation, but no noticeable differences are found.
Following previous work, the excited state intramolecular proton transfer (ESIPT) pro cess in an o-hydroxybenzaldehyde (oHBA) molecule was used for the comparison of different TPB schemes. The molecule was initially optimized at the ground-state geometry. Here, only the transferring proton is treated quantum mechanically using the NEO method. At \(t=0\), one electron was "manually" promoted from the highest occupied molecular orbital (HOMO) to the lowest unoccupied molecular orbital (LUMO) to induce the ESIPT process. Figure. 1 (a) shows the trajectories of the ESIPT of the oHBA molecule using the TPB approaches and the standard FPB approach. In (a), the solid lines and dashed lines are the expectation value of the quantum proton position operator and the location of the proton basis function center (for TPB schemes), respectively. For simplicity, the proton transfer is considered to have occurred when the expectation value of the proton position operator is calculated. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator to ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator to the ground state energy. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator. The calculated value of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator and the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of the proton position operator is calculated as the ratio of proton position operator is the same distance from the donor and acceptor oxygen atoms, i.e., when the two lines cross in (a). Qualitatively, all TPB and FPB approaches show the proton transfer process on a similar time scale. Quantitatively, the standard TPB scheme shows the best agreement with the reference FPB approach. The sc-TPB scheme is somewhat worse, and the l-TPB scheme predicts a transfer time that is \(\sim 20\%\) longer. With the l-TPB scheme, the position of the proton basis function center (dashed blue lines) tends to drag behind the quantum proton position expectation value (the solid blue lines), and this artificial friction somewhat slows down the proton transfer.
Another aspect of examining the dynamics from the mathematical viewpoint is the conservation of the constant of motion for the dynamics governed by the Lagrangian. The conserved quantity for the l-TPB and sc-TPB approaches is given by Eq. 28. We will denote this quantity as \(E_{\rm cons}\), which is the sum of the total system energy \(E_{\rm tot}\) and the kinetic energy of the proton basis function centers. Note that we have \(E_{\rm tot}=E_{\rm cons}\) for the FPB approach since the kinetic energy of the proton basis function centers is zero. The conserved quantity for the original TPB approach was presumed to be \(E_{\rm tot}\) because the moving basis function center is not part of the physical system but rather is just a mathematical tool to allow the use of a smaller basis set. (b) shows \(E_{\rm cons}\) and \(E_{\rm tot}\) during the real-time propagation using the FPB and TPB approaches. Having been derived directly from the NEO Lagrangian, both the l-TPB and sc-TPB schemes show stable behavior for \(E_{\rm cons}\), similar to the FPB scheme. None of the TPB approaches preserve the total energy \(E_{\rm tot}\) as well as the FPB approach, as discussed in previous work.
The TPB and sc-TPB schemes are appealing for different reasons. The original TPB scheme shows better performance in terms of how closely it agrees with the quantum dynamics of the reference FPB simulation, as determined by the motion of the expectation value of the proton position operator. At the same time, the sc-TPB scheme addresses the previously noted issue of energy conservation, and the expected conserved quantity is identified by deriving the dynamics from the NEO Lagrangian. In the most rigorous approach, as when a very accurate description of the quantum dynamics with energy conservation is necessary, the FPB scheme can be performed with pre-defined proton basis function center positions based on simulations with the original TPB or sc-TPB schemes.
### Applications with Periodic Boundary Conditions
Finally, we demonstrate the utility of the newly developed extension of RT-NEO-TDDFT Ehrenfest dynamics with periodic boundary conditions (PBCs). In particular, we study how the explicit dynamics of solvating water molecules can significantly impact the ESIPT process. For this purpose, we study the ESIPT of the oHBA molecule that is solvated in explicit liquid water molecules. To obtain equilibrium configurations of the oHBA molecule in water, we performed first-principles molecular dynamics (FPMD) simulations in a periodic cubic simulation cell containing the oHBA molecule and 57 water molecules at room temperature (300 K). The SCAN meta generalized gradient approximation to the XC functional was used for FPMD, as it has been shown to provide a reasonably accurate structure of water. The oHBA molecule was fixed at its ground state geometry during the FPMD simulations so that the same starting oHBA geometry could be used for all subsequent Ehrenfest dynamics simulations to enable the study of the effects from the dynamics of the solvating water molecules.
(a) A snapshot of the oHBA molecule solvated in liquid water from FPMD simulation. The isosurfaces show the molecular HOMO state (yellow) and the molecular LUMO state (blue). (b) The distances between the position expectation value of the transferring proton and the donor and acceptor oxygen atoms in the oHBA molecule as a function of time using the sc-TPB approach. The circles represent the isolated oHBA molecule in vacuum, and the colored solid lines represent the trajectories started from four randomly selected snapshots from the FPMD simulation.
We then performed NEO Ehrenfest dynamics simulations using the TPB and sc-TPB approaches on four randomly selected snapshots from the FPMD trajectory (Fig. 2(a)). The transferring proton in the oHBA molecule and all nearest protons of the surrounding H\({}_{2}\)O molecules that exhibit hydrogen bonding with the two oxygen atoms of the oHBA molecule in the initial configuration were treated as quantum protons using NEO. At \(t=0\), a single electron was excited from the molecular HOMO to the molecular LUMO to initiate the proton transfer. The molecular HOMO and LUMO can be hybridized with liquid water states. Thus, the hybrid states that preserve the spatial characteristics of the molecular HOMO/LUMO most closely were chosen for the initial excitation. Figure 2(b) shows the results of the NEO Ehrenfest dynamics using the sc-TPB approach, and these results are compared to the case where the oHBA molecule is in the gas phase. The NEO Ehrenfest dynamics trajectories using the TPB approach show similar results, and they can be found in Supplementary Material. No proton transfer between the electronically excited oHBA molecule and its surrounding water molecules was observed, which is consistent with our previous study.
Previously, the ESIPT dynamics in oHBA were found to be sensitive to the configuration of the surrounding waters at the time the electronic excitation takes place using a hybrid QM/MM approach for the RT-NEO-TDDFT Ehrenfest dynamics. Similarly, our results for periodic QM systems indicate that the presence of surrounding water molecules qualitatively impacts the proton transfer process, and it could even suppress ESIPT entirely, as shown by the green lines in Fig.2(b). Even though simulating only four instances is not enough to provide a comprehensive understanding of the solvation effect on the ESIPT in the oHBA molecule, these results show that solvent dynamics can play a central role in these ESIPT processes. Additionally, we performed an analogous simulation with one of these initial molecular configurations using the conventional RT-TDDFT Ehrenfest dynamics without NEO such that all nuclei were treated classically. In this case, the proton transfer takes place but at a much slower rate (see Figure S4 in Supplementary Material). This observation further emphasizes the importance of the quantum treatment of the transferring protons in proton transfer processes.
Conclusion
In this work, we presented a systematic method for integrating the NEO method with periodic RT-TDDFT and Ehrenfest dynamics through the Lagrangian for multicomponent quantum-classical systems. The NEO Lagrangian allows us to derive the equations of motion for both quantum particles and classical nuclei on an equal footing. Ehrenfest dynamics with RT-NEO-TDDFT was introduced with three distinct schemes for the traveling proton basis approach, in which the centers of the quantum proton basis functions are propagated as classical degrees of freedom. The three schemes are compared in terms of the quantum proton dynamics and the corresponding constant of motion for the dynamics.
Building on our original periodic NEO-DFT work, further technical details regarding the evaluation of the energy gradients and real-time propagators are provided. The numerical implementation demonstrated that NEO Ehrenfest dynamics represents a powerful approach for investigating the coupled quantum dynamics of electrons and protons in heterogeneous environments, in which the dynamics of other nuclei are important. As a proof-of-principle example, we showed how the dynamics of solvating water molecules significantly impacts the photoinduced intramolecular proton transfer in an o-hydroxybenzaldehyde molecule. We anticipate that this new first-principles method will facilitate the investigation of a broad range of photocatalytic reactions, including PCET in extended condensed matter systems.
|
10.48550/arXiv.2407.18842
|
Lagrangian Formulation of Nuclear-Electronic Orbital Ehrenfest Dynamics with Real-time TDDFT for Extended Periodic Systems
|
Jianhang Xu, Ruiyi Zhou, Tao E. Li, Sharon Hammes-Schiffer, Yosuke Kanai
| 6,619
|
10.48550_arXiv.2303.13851
|
## 1 Introduction
Liquid water under high pressure has been the subject of intensive studies due to its importance in physics, chemistry, and life sciences. Among them, one particular focus is the distortion of the hydrogen bond (HB) network with respect to external pressure. In the last two decades, a number of X-ray and neutron diffraction experiments have been conducted to determine the static structure factor (SSF) and radial distribution function (RDF) of water under high pressures. Through many efforts, the influence of pressure on the SSF of liquid water is generally converged as follows. Below 1 GPa, as pressure increases, the first peak of SSF rises significantly. In contrast, the second peak decreases monotonically in height, resulting in the fact that the highest second peak of SSF at ambient pressure shrinks into the shoulder of the first peak at 1 GPa. The trend continues to higher pressures, resulting in the fusion of the first two peaks into a single first peak at 4 GPa. Recent works attribute the changes in the SSF to the damage of the tetrahedral structure of the inner shell.
Meanwhile, the RDF of oxygen atoms \(\mathrm{g}_{\mathrm{OO}}(\mathrm{r})\) also undergoes substantial changes as the pressure increases. While the position of the first peak, which stands for the hydrogen-bonded water molecules, remains almost unchanged as pressure rises up to 1 GPa, the right shoulder of the first peak rises monotonically, and the second peak decreases to a local minimum. Despite the converged experimental results, the influence of the hydrogen-bond network on the high-pressure water is still inconclusive, for instance, the following questions arise: (i) How do the changes in the HB network affect the first two peaks of SSF in the high-pressure water? (ii) In terms of the number and the directionality of HBs, how does the pressure influence the accepting and donating ends of water molecules? (iii) How are the changes in the tetrahedral structure of water related to the HB network?
From a theoretical perspective, molecular dynamics (MD) serves as an important tool to investigate liquid water, bridging the gap between theory and experimentally measured SSF/RDF. Past MD works generally yielded that the typical tetrahedral HB structure is barely influenced by pressures under 1 GPa, while the HB structure outside the first shell is changed and occasionally one or two non-hydrogen-bonded water are pushed into the first shell. In recent years, _ab initio_ molecular dynamics (AIMD) with first-principles accuracy is vastly applied in place of classical MD, lifting the accuracy of simulation to a quantitative agreement with experiments. Although AIMD exhibits prediction power to explain the pressure effects on liquid water to a large extent, the small system size and short trajectory adopted by AIMD could only yield SSF with insufficient accuracy.
Workflow for calculations of static structure factors of liquid water at three densities (1, 1.115, 1.24 g/cm\({}^{3}\)). Training data are obtained from AIMD simulations of 64 water molecules. (a-c) show the workflow of our calculation of static structure factor (SSF) under the three different pressures. (d) shows the formula used to calculate the SSF, in which \(f_{i}\) stands for the shape factor of atom \(i\) in X-ray diffraction. For S\({}_{\rm{OO}}\)(Q) calculation, all \(f_{i}\) are set to 1. (e-g) show the calculated SSFs compared with experiment results, which in 1 and 1.115 g/cm\({}^{3}\) are S\({}_{\rm{OO}}\)(Q) and in 1.24 g/cm\({}^{3}\) is the total S(Q), since only total SSF in 1.24 g/cm\({}^{3}\) is provided in Ref..
In this regard, using AIMD to directly compute the SSF of liquid water remains a challenging issue.
## 2 Methods
Our work aims to directly compute the static structure factors of liquid water at high pressures with first-principles accuracy and elucidate the related H-bond structures in detail. We perform AIMD simulations with the meta-GGA exchange-correlation functional, SCAN, which is known to satisfy all 17 restrictions on semi-local exchange-correlation functional and excellent for water properties. For instance, the SCAN functional substantially improves the strong H-bonds description provided by GGA functionals such as PBE. Meanwhile, the added intermediate-ranged van der Waals interactions in SCAN pull the second-shell water molecules closer to the interstitial area, causing a denser and more disordered interstitial area around water molecules closer to the experimental data. Consequently, the SCAN functional accurately predicts the H-bonded structure of liquid water. In this work, we perform three AIMD simulations of bulk water (64 water molecules) at ambient temperature with fixed densities of 1.0, 1.115, and 1.24 g/cm\({}^{3}\), which correspond to water densities under ambient pressure, 360 MPa, and 1 GPa, respectively.
A converged SSF needs MD simulations with a large system and a long trajectory, which is typically limited due to the high computational costs of AIMD. In this regard, we adopt the Deep Potential Molecular Dynamics (DPMD) to train neural-network-based potentials for water with different densities; the workflow is illustrated in Figs. 1(a)-(c). The resulting neural networks are several orders of magnitude faster than AIMD and scale linearly with the number of atoms. In previous works, DPMD has been validated to learn AIMD accuracy from both SCAN and PBE0-TS functionals. In this work, the training sets, including the atomic positions, forces, and energy, are obtained from AIMD simulations with the SCANfunctional in a 64-water cell. The energy predicted by DPMD is in quantitative agreement with the AIMD result (1 meV/atom). As a result, DPMD simulations are performed for 512 water molecules for 300 ps. More details of the simulations and training can be found in Supplemental Material Section I.
## 2 Results and Discussions
Fig. 1(d) lists the formula to compute SSF (S(Q)), and the results for the three densities are shown in Figs. 1(e-g). We find the DP model quantitatively reproduces the experimental data for all three densities. As the pressure increases from ambient to 1 GPa, the first peak of S\({}_{\text{OO}}\)(Q) significantly rises while the second peak decreases into a shoulder. Correspondingly, as the position and shape of the first peak of g\({}_{\text{OO}}\)(r) generally remain the same, the second peak under ambient pressure decreases substantially as pressure rises. The first minimum in between rises to form a bump beside the first peak. The third peak also moves inward significantly as pressure rises. Furthermore, we find that the height change of the first two peaks of SSF has nothing to do with the hydrogen-bonded first shell but purely results from the inward movement of non-hydrogen-bonded interstitial water molecules. To further reveal the relations between the HBs in water and the detailed changes in SSF under pressure, we decompose S\({}_{\text{OO}}\)(Q) into contributions of different water molecule pairs with the (a-c) Static structure factor (SSF) S\({}_{\rm OO}\)(Q), SSF contributed by hydrogen-bonded water molecule pairs, and SSF contributed by non-hydrogen-bonded water molecule pairs. (d-f) Radial distribution function (RDF) g\({}_{\rm OO}\)(r), RDF contributed by hydrogen-bonded water molecule pairs, and RDF contributed by non-hydrogen-bonded water molecule pairs. Water systems with a density of 1 g/cm\({}^{3}\), 1.115 g/cm\({}^{3}\), and 1.24 g/cm\({}^{3}\) are respectively colored in red, blue, and green.
formula as follows
\[\begin{split} S_{\text{OO}}(\text{Q})&=\frac{1}{N}\sum_ {j=1}^{N}\sum_{k=1}^{N}e^{i(\mathbf{r_{j}}-\mathbf{r_{k}})\cdot\mathbf{Q}}\\ &=\frac{1}{N}\left\{N+\sum_{j=1}^{N-1}\sum_{k=j+1}^{N}\left[e^{i( \mathbf{r_{j}}-\mathbf{r_{k}})\cdot\mathbf{Q}}+e^{i(\mathbf{r_{k}}-\mathbf{r_{ j}})\cdot\mathbf{Q}}\right]\right\}\\ &=\frac{1}{N}\left\{N+2\sum_{j=1}^{N-1}\sum_{k=j+1}^{N}\left[\cos \left(\mathbf{r_{j}}-\mathbf{r_{k}}\right)\cdot\mathbf{Q}\right]\right\}\\ &=1+\frac{2}{N}\left\{\sum_{(j,k)\in HB}\cos\left(\mathbf{r_{j}}- \mathbf{r_{k}}\right)\cdot\mathbf{Q}+\sum_{(j,k)\notin HB}\cos\left(\mathbf{r _{j}}-\mathbf{r_{k}}\right)\cdot\mathbf{Q}\right\},\end{split} \tag{1}\]
Specifically, \(\text{S}_{\text{OO}}(\text{Q})\) of water is decomposed into two components by classifying the atom pairs into hydrogen-bonded pair and non-hydrogen-bonded pairs, as displayed in Figs. 2(b) and (c), respectively. Notably, the HB criteria are chosen with the O-O distance less than 3.5 A and the O-O-H angle smaller than 30\({}^{\circ}\); although new criteria for defining HBs have been proposed for both ambient and high-pressure water systems, our conclusions are insensitive to the hydrogen-bond criteria, more details of which are shown in Supplemental Material Section IV. We find the changes in \(\text{S}_{\text{OO}}(\text{Q})\) across the three densities mainly come from the non-hydrogen-bonded water molecules. On the one hand, Fig. 2(b) shows that the three \(\text{S}_{\text{OO}}(\text{Q})\) contributed by the hydrogen-bonded water molecules in the first solvation shell at different pressures are almost identical, indicating that \(\text{S}_{\text{OO}}(\text{Q})\) is insensitive to the changes in the hydrogen-bonded water molecules when the density rises from 1 to 1.24 g/cm\({}^{3}\). On the other hand, the three \(\text{S}_{\text{OO}}(\text{Q})\) contributed by the non-hydrogen-bonded water molecules, as illustrated in Fig. 2(c), exhibit substantially different features especially when the wave vector \(\text{Q}<4\) A\({}^{-1}\). In detail, the first peak in Fig. 2(c) decreases slightly when the water density changes from 1.0 to 1.115 g/cm\({}^{3}\), but increases from 1.115 to 1.24 g/cm\({}^{3}\). Meanwhile, the peak moves towards a larger wave vector Q when the density increases. During such a rightward move, the peak gradually merges into the first peak of hydrogen-bonded S\({}_{\rm OO}\)(Q) at 2.7 A\({}^{-1}\), as illustrated in Fig. 2(b). In addition, we observe that the decrease of the second peak at 2.9 A\({}^{-1}\) in Fig. 2(a) corresponds to the decrease of the small peak at the same wave vector in Fig. 2(c), suggesting this experimental feature of S\({}_{\rm OO}\)(Q) is also related to the non-hydrogen-bonded water molecules. In summary, the above changes in S\({}_{\rm OO}\)(Q) from both hydrogen-bonded and non-hydrogen-bonded water molecules fully explain the increase of the first peak and decrease of the second peak in S\({}_{\rm OO}\)(Q) (Fig. 2(a)) when the density increases from 1.0 to 1.24 g/cm\({}^{3}\).
The features of S\({}_{\rm OO}\)(Q) in Fig. 2(a) can be further explained by its Fourier transform counterpart, which is the radial distribution function g\({}_{\rm OO}\)(r) shown in Fig. 2(d). By adopting the same criterion for decomposing S\({}_{\rm OO}\)(Q), we decompose g\({}_{\rm OO}\)(r) into hydrogen-bonded and non-hydrogen-bonded terms and the results are shown in Figs. 2(e) and (f). The main peaks in Fig. 2(e) are contributed by hydrogen-bonded water molecules at different densities. We observe the unaffected positions of the three peaks as the density changes from 1.0 to 1.24 g/cm\({}^{3}\), which is due to the relative preservation of local HB order. We also notice that the amplitude of the peak decreases slightly at a larger density, which is caused solely by the increase in water density but not affected by the hydrogen-bonded tetrahedral structure. Furthermore, as illustrated in Fig. 2(f), we find the water molecules in the second and third shells move towards the inner shell under pressure, which is evidenced by the disappearance of the second peak at 4.6 A, and the leftward movement of the third peak from 6.7 to 6.1 A. In particular, when the density increases from 1.0 to 1.24 g/cm\({}^{3}\), the non-hydrogen-bonded water molecules in the second shell move towards the region of the first shell, filling the O-O void space, as evidenced by the emergence of the peak at 3.4 A. As a result, a shoulder of the first peak emerges in g\({}_{\rm OO}\)(r) when the density is 1.24 g/cm\({}^{3}\), as shown in Fig. 2(d). In conclusion, the features of S\({}_{\rm OO}\)(Q) under pressure are mainly affected by the movements of the non-hydrogen-bonded water molecules from the interstitial region and beyond towards the inner hydrogen-bonded tetrahedral structure.
The S\({}_{\rm OO}\)(Q) and g\({}_{\rm OO}\)(r) yield HB information only in the radial direction. Since more ex perimental information is limited, first-principles simulations without empirical parameters play an important role in obtaining more HB network information for water under pressure. Earlier literature reported that the H-bonded first shell of water molecules is almost unaffected, and the average number of HBs is not significantly influenced. However, when the water density changes from 1 to 1.24 g/cm\({}^{3}\), the tetrahedral structure has been substantially altered according to our simulations. Fig. 3(a) illustrates the distributions of the tetrahedral parameter \(q\) at the three densities from DPMD trajectories. The parameter \(q\) measures the local structural order of liquid water in the sense of tetrahedrality. In our results, \(q\) distribution possesses a skewed peak at \(q\)=0.8 and a shoulder at \(q\)=0.5 under ambient pressure. The height of the shoulder keeps rising while the peak decreases as the density changes from 1 to 1.24 g/cm\({}^{3}\), forming a platform in the distribution between \(q\)=0.5 and \(q\)=0.8 at 1.24 g/cm\({}^{3}\), which is in consistent with earlier studies. To further investigate the changes in the distribution of the tetrahedral parameter \(q\), we decompose the distribution by the number of H-bonded water molecules (denoted as n\({}_{\rm HB}\)) within the four nearest neighbors of a water molecule. The distributions contributed by n\({}_{\rm HB}\) =4, 3, and 2 are shown in Figs. 3(b), (c), and (d), respectively. We find that the distribution involving four H-bonds exhibits a single peak at \(q\)=0.8 that represents a typical skewed tetrahedral HB structure in liquid water, while the n\({}_{\rm HB}=3\) distribution exhibits a platform between \(q\)=0.5 and \(q\)=0.8, and n\({}_{\rm HB}=2\) forms a single-peak distribution centered at \(q\)=0.5. As pressure increases, the percentage of n\({}_{\rm HB}=4\) drops substantially from 52.2% (1 g/cm\({}^{3}\)) to 37.1% (1.24 g/cm\({}^{3}\)) while percentages of n\({}_{\rm HB}=2\) and 3 increase, resulting in the significant deformation of \(q\).
Fig. 4(a) shows the classification of the five main H-bond structures, where a hydrogen-bond structure that accepts \(m\) hydrogen bonds and donates \(n\) hydrogen bonds is denoted as '\(m_{\rm A}n_{\rm D}\)'. We notice that the pressure mainly affects the accepting end rather than the donating end of water molecules. For example, when the density rises from 1 to 1.24 g/cm\({}^{3}\), the percentage of water molecules accepting 2 HBs substantially decreases from 64.7% to 58.8%, while the percentage of donating 2 HBs only slightly decreases from 78.8% to 78.3%. As a result, Fig. 4(a) shows that the percentage of the 2\({}_{\rm A}\)2\({}_{\rm D}\) configurations decreases from 54.2% at 1 g/cm\({}^{3}\) to 48.7% at 1.24 g/cm\({}^{3}\); around 5.5% of water molecules deviate from the tetrahedral HB structure by gaining or losing an accepting HB. Consequently, the percentage of the 1\({}_{\rm A}\)2\({}_{\rm D}\) and 3\({}_{\rm A}\)2\({}_{\rm D}\) configurations increases monotonically as pressure increases.
Pressure mainly affects the HB distribution of the accepting end of a water molecule rather than the donating end. Figs. 4(b), (c), and (d) provide more details about the distributions of O-O-O angles for the tetrahydroral 2\({}_{\rm A}\)2\({}_{\rm D}\) water at a density of 1, 1.115, and 1.24 g/cm\({}^{3}\), respectively. Here, the O-O-O angles are divided into two components as contributed by the accepting and donating ends.
(a) Distribution of the tetrahedral parameter \(q\) at 1 g/cm\({}^{3}\) (red), 1.115 g/cm\({}^{3}\) (blue) and 1.24 g/cm\({}^{3}\) (green). (b-d) Distributions of \(q\) contributed by water molecules whose 4, 3 and 2 neighbors out of the four closest neighbors are H-bonded, denoted by \({}_{\rm HB}=4,3,2\), respectively. Inset in each subplot depicts a typical water with its four closest neighbors of the corresponding nHB.
(a) Percentages of five major stable HB structures observed in liquid water with densities being 1, 1.115, and 1.24 g/cm\({}^{3}\). (b-d) Distributions of the O\({}_{\rm A}\)-O-O\({}_{\rm A}\) angle (dashed line) and O\({}_{\rm D}\)-O-O\({}_{\rm D}\) angle (solid line) of the 2\({}_{\rm A}\)2\({}_{\rm D}\) structure at the three different densities. Vertical dashed and solid lines in corresponding colors mark the peak positions of the O\({}_{\rm A}\)-O-O\({}_{\rm A}\) and O\({}_{\rm D}\)-O-O\({}_{\rm D}\) angle distribution, while the grey dashed line mark 109\({}^{\circ}\)28\({}^{\prime}\). The three insets are the spatial distribution function (SDF) of the 2\({}_{\rm A}\)2\({}_{\rm D}\) structure.
We further add in the O-O spatial distribution function (SDF), which provides a 3-dimensional view of the distribution of H-bonded O atoms close to the O of a water molecule. At the donating end, we find the \(\rm O_{D}\)-\(\rm O\)-\(\rm O_{D}\) angle distribution barely changes with pressure, as the peak position (vertical solid line in corresponding color) stays almost unchanged, close to the typical \(109^{\circ}28^{{}^{\prime}}\) of tetrahedral structure (vertical grey line). As shown in the insets of Figs. 4(b-d), the SDF forms two separate disk-like distributions and remains unchanged as pressure increases. On the contrary, the \(\rm O_{A}\)-\(\rm O\)-\(\rm O_{A}\) angle distribution moves leftward as pressure increases, the peak of which (dashed line in corresponding color) decreases from \(93.6^{\circ}\) to \(83.0^{\circ}\). Correspondingly, the inset of Fig. 4(b) shows that the SDF at the accepting end exhibits two separate distributions at 1 g/cm\({}^{3}\), but the two distributions spread more loosely and begin to merge at 1.115 g/cm\({}^{3}\), and the trend becomes more clear at 1.24 g/cm\({}^{3}\).
We summarize our points as follows. A higher pressure decreases the percentage of standard tetrahedral HB structures and increases the percentage of non-tetrahedral HB structures, like \(3_{\rm A}2_{\rm D}\). For the tetrahedral \(2_{\rm A}2_{\rm D}\) HB structure, a higher pressure largely changes the \(\rm O_{A}-O-O_{A}\) distribution on the accepting end but leaves the \(\rm O_{D}-O-O_{D}\) distribution on the donating end almost unchanged (Fig. 4(b-d)). Increasing pressure also results in a higher chance of non-H-bonded water molecules moving into the inner shell, thereby further degrading the structural tetrahedrality. However, the average HB number per water molecule and the radial distribution of H-bonded water are hardly affected by the pressure. The increase in the density of the inner shell is mainly contributed by the increase in the non-H-bonded water inserted into the inner shell. Overall, these features form a complete picture of the pressure influence on the structure of liquid water.
Furthermore, the deformation of the tetrahedral structure at the accepting end, along with the movement of water molecules from the interstitial region and beyond into the inner shell, could potentially result in a higher proportion of weak and distorted hydrogen bonds compared to water under ambient conditions. The distortion of the HB geometry is also revealed by Fig. S3 (c-e) in the Supplemental Material. Since the nuclear quantum effects (NQEs) are suggested to generally strengthen the strong HBs, and weaken the weak ones, it is likely that a higher number of weakened HBs and a more disordered HB network would be observed in water under high pressure when taking NQEs into account. However, to what extent NQEs affect the structure of liquid water under pressure still requires further study involving the NQEs.
## 4 Conclusions
In conclusion, we systematically investigated the changes in H-bond structures of liquid water under pressure by performing neural-network-based DPMD simulations with the first-principles accuracy. The use of machine-learning-assisted DPMD models largely boosted the efficiency of AIMD, enabling us to simulate a large cell with 512 water molecules and a trajectory length of 300 ps. Besides, the use of the SCAN functional provided first-principles accuracy for liquid water without empirical parameters. As a result, we directly computed the SSF of water at different densities and proposed a new method that links the H-bond structure information to the decomposed SSF. The results and analyses provided in this work may help better understand the influences of pressures on the H-bond network of water. Furthermore, it can be readily applied to study similar scientific problems that involve H-bond information in the X-ray or neutron diffraction experiments.
|
10.48550/arXiv.2303.13851
|
Characterization of the Hydrogen-Bond Network in High-Pressure Water by Deep Potential Molecular Dynamics
|
Renxi Liu, Mohan Chen
| 4,883
|
10.48550_arXiv.2307.06837
|
###### Abstract
We study the energetics of quasi-particle excitations in CsPbBr\({}_{3}\) perovskite nanocrystals using path integral molecular dynamics simulations. Employing detailed molecular models, we elucidate the interplay of anharmonic lattice degrees of freedom, dielectric confinement, and electronic correlation on exciton and biexciton binding energies over a range of nanocrystal sizes. We find generally good agreement with some experimental observations on binding energies, and additionally explain the observed size dependent Stokes shift. The explicit model calculations are compared to simplified approximations to rationalize the lattice contributions to binding. We find that polaron formation significantly reduces exciton binding energies, whereas these effects are negligible for biexciton interactions. While experimentally, the binding energy of biexcitons is uncertain, based on our study we conclude that biexcitons are bound in CsPbBr\({}_{3}\).
Lead halide perovskite nanocrystals are currently the subject of significant interest due to their exceptionally high photoluminescence quantum yields, which make them ideal materials for light emission, lasing and photodetection. The electronic properties of the lead halides depend strongly on the coupling between charges and their surrounding soft, polar lattices. This coupling has been implicated in a number of novel phenomena including photo-induced phase transitions, long radiative recombination rates, and anomalous temperature dependent mobilities. In nanocrystals, optical properties are largely determined by the behavior of exciton complexes. Recently, Dana et al. proposed a potential anti-binding of biexcitons in perovskite nanocrystals, implicating the potential role of the lattice in mediating this interaction. Here, we study the energetics of quasiparticle excitations including excitons and biexcitons with an explicit description of the lattice, over a range of perovskite nanocrystal sizes. We find that while polaron formation weakens the exciton binding energy, biexciton energetics are largely unaffected, leading to an expectation that they are bound in nanocrystals and in bulk.
Excitons and biexcitons are both characterized by a binding energy, the energy required to dissociate the pair of quasiparticles-free charges or excitons. While there is reasonable consensus on the bulk exciton binding energy, its dependence on nanocrystal size is less certain. Further, a number of biexciton binding energies have been reported experimentally for lead halide perovskites, but their values span a large range, and even disagree in sign. Experimental measurements are hampered by the nanocrystal polydispersity, spectral drift, and thermal broadening of spectral lines. These ambiguities could be clarified theoretically, however there are few suitable approaches available. Unlike traditional semiconductors where structural fluctuations can be ignored or described within a harmonic approximation, the perovskite lattice structure with its anharmonic tilting and rocking motions result in significant renormalization of electronic properties. Given that incorporating the effects from the lattice is important in understanding excitonic properties, there have been some attempts to include them into a theoretical description. For example, extensions to GW and Bethe-Salpeter equations have been developed to include phonon effects on excitonic properties perturbatively. At the same time, quasiparticle excitations requires a balanced description of electron correlation, making _ab initio_ models difficult to apply in nanocrystals where the number of atoms are large. Tight binding and pseudo-potential models have been developed to study excitonic structure, but these have not yet been unified with approaches to describe electron-phonon effects.
Here, we use quasiparticle path integral molecular dynamics with an explicit atomistic description of the lattice, which allows us to include all orders of anharmonicity from the lattice, and treat electron correlation exactly. This approach has been previously successful at describing the excitonic properties of bulk systems, and the lattice model we employ has been used extensively to describe both vibrational and nonlinear properties of a variety of lead-halide systems. In this study, we consider a system of a single biexciton interacting with CsPbBr\({}_{3}\) cubic perovskite nanocrystals. The model Hamiltonian consists of three pieces, \(\mathcal{H}=\mathcal{H}_{\rm el}+\mathcal{H}_{\rm lat}+\mathcal{H}_{\rm int}\).
\[\mathcal{H}_{\text{el}}=\sum_{i}\frac{\hat{\mathbf{p}}_{i}^{2}}{2m_{i}}+\sum_{i \neq j}\frac{q_{i}q_{j}}{4\pi\varepsilon_{\infty}|\hat{\mathbf{x}}_{i}-\hat{ \mathbf{x}}_{j}|} \tag{1}\]
We consider a singlet biexciton, so electrons and holes are distinguishable particles.
For the lattice, we model the CsPbBr\({}_{3}\) nanocrystals explicitly using a previously validated _ab initio_ derived forcefield.
\[\mathcal{H}_{\text{lat}}=\sum_{i=1}^{N}\frac{\mathbf{p}_{i}^{2}}{2m_{i}}+U_{ \text{lat}}(\mathbf{x}_{\text{lat}}) \tag{2}\]
The atomistic forcefield,\(U_{\text{lat}}\), is the sum of pairwise interactions with distance \(x_{ij}=|\mathbf{x}_{i}-\mathbf{x}_{j}|\) consisting of a Coulomb potential and Lennard-Jones potential
\[U_{\text{lat}}=\sum_{i,j=1}^{N}\frac{q_{i}q_{j}}{4\pi\varepsilon_{0}x_{ij}}+4 \varepsilon_{ij}\left[\left(\frac{\sigma_{ij}}{x_{ij}}\right)^{12}-\left( \frac{\sigma_{ij}}{x_{ij}}\right)^{6}\right] \tag{3}\]
The interaction between the quasiparticles and lattice is written as the sum of pseudopotentials between each quantum particle \(i\) and the lattice particle \(j\),
\[\mathcal{H}_{\text{int}}=\sum_{i}\sum_{j=1}^{N}\frac{q_{i}q_{j}}{4\pi \varepsilon_{0}\sqrt{r_{\text{cut}}^{2}+|\hat{\mathbf{x}}_{i}-\hat{\mathbf{x} }_{j}|^{2}}} \tag{4}\]
The quasiparticle interactions are screened by the optical dielectric constant of the nanocrystal, which changes discontinuously at the boundary between the perovskite and surrounding solution.
\[U_{\text{wall}}=\sum_{i,k}\frac{q_{i}^{2}}{8\pi\varepsilon_{0}|\hat{\mathbf{x }}_{i}-\mathbf{x}_{\text{wall},k}|}\cdot\frac{\varepsilon_{\infty}-\varepsilon _{0}}{\varepsilon_{\infty}+\varepsilon_{0}} \tag{5}\]
This external potential results in a dielectric confinement. Additional details on the atomistic force field, pseudopotentials, and wall potentials are described in the Supporting information (SI).
For electrons and holes, we use imaginary time path integrals, while the heavy atoms of the lattice are treated classically.
\[\mathcal{Z}=\int\mathcal{D}[\mathbf{x}_{eh}(\tau),\,\mathbf{x}_{\text{lat}}] \,e^{-\int_{\tau=0}^{gh}\mathcal{H}[\mathbf{x}_{eh}(\tau),\,\mathbf{x}_{\text {int}}]/\hbar} \tag{6}\]
Discretizing the path action renders each quantum particle isomorphic to a ring polymer, where neighboring timeslices are connected by harmonic springs. We use molecular dynamics simulations with a second order discretization of the path integral to compute expectation values of this system, for which each quant particle is represented by 1000 timeslices. Representative simulation snapshots of both the biexciton and exciton are shown in Throughout we consider nano-crystalline cubes, where \(L\) is the edge length.
For comparison, we also consider two approximate models that incorporate harmonic and static lattice effects. To incorporate effects from harmonic phonons, we adopt a _dynamic_ screening model described previously, which is a model for charges that are coupled linearly with the polarization field generated by a collection of harmonic modes.
Simulation snapshot of biexciton (left) and exciton (right) interacting with CsPbBr\({}_{3}\) perovskite nanocrystal where \(L=3.56\) nm represents the edge length of the nanocrystal.
We parameterize the influence functional with a Frohlich coupling for the electrons and holes as \(\alpha_{e}=2.65\) and \(\alpha_{h}=2.76\) and an optical phonon mode with energy \(\hbar\omega=16.8\) m\(e\)V. We also compare a _static_ lattice, where only the electronic Hamiltonian and confinement effects are considered. All simulations are done in LAMMPS and the details of the discretized Hamiltonian can be found in the SI.
With an explicit lattice, the exciton binding energy is
\[\Delta_{X}=\lim_{T\to 0}\langle E\rangle_{\rm e}+\langle E\rangle_{\rm h}- \langle E\rangle_{\rm ex}-\langle U_{\rm lat}\rangle_{\rm ex}\,,\]
while the biexciton binding energy is
\[\Delta_{XX}=\lim_{T\to 0}2\langle E\rangle_{\rm ex}-\langle E\rangle_{\rm biex }-\langle U_{\rm lat}\rangle_{\rm ex}\,,\]
Simulations are performed at 50K, which is low enough to extract the ground state energy.
\[\langle E\rangle=-\frac{\partial}{\partial\beta}\ln\mathcal{Z}[\mathbf{x}_{eh} (\tau),\mathbf{x}_{\rm lat}(\tau)] \tag{7}\]
The derivative above produces two terms, an average kinetic energy and an average potential energy, where we use a virial estimator to efficiently estimate the kinetic energy. For the evaluation of binding energies from the simulations with only quasiparticles, in the static or dynamic approximation, the same definitions are used without the lattice relevant terms. Within the static approximation, the binding energies are determined by only \(\mathcal{H}_{e}\) and the confining potential. In order to efficiently extract the binding energies with dynamic approximation, we use a thermodynamic perturbative theory approach (see SI).
(a) shows the exciton binding energy \(\Delta_{\rm X}\) computed from our molecular model, in addition to those approximated from with the static or dynamics models for the nanocubes considered, ranging from \(L=2.4\) nm to \(6\) nm. Given the reduced mass of the exciton, \(\mu=0.11\) and optical dielectric constant, the Bohr radius of the exciton is \(R_{X}=2.07\)nm. Thus for the range of nanocrystals considered, we are in a moderate to strong confinement regime. As a consequence, all binding energy estimates exhibit a strong \(L\) dependence, which we model in Fig. 2(a) as \(\Delta_{X}=\Delta_{X}^{\circ}(1+\ell_{1}/L+\ell_{2}^{2}/L^{2})\) where \(\Delta_{X}^{\circ}\) is the bulk binding energy and \(\ell_{1}\) and \(\ell_{2}\) are fit parameters.
Over the full range of nanocrystal sizes considered, the static approximation yields a binding energy much larger than either the explicit model calculation or the dynamic approximation. The value from the static approximation with the largest nanocrystal gives a reasonable agreement with the expectation from a Wannier-Mott model \(\Delta_{X,\rm WM}=R_{H}\mu/\varepsilon_{\infty}^{2}=84.5\) m\(e\)V, where \(R_{H}\) is a Rydberg energy. The values at finite \(L\) agree well with recent pseudo-potential based calculations of CsPbI\({}_{3}\), up to a shift of the bulk binding energy by 20 m\(e\)V to account for the change in halide.
The explicit model calculations are well reproduced by the dynamic approximation. The suppression of the binding energy due to lattice effects could arise because of two distinct mechanisms. First the lattice could screen the electron-hole interactions, weakening them. Second the lattice can lower the self-energies of the free charges, bringing them closer to the exciton self-energy, effectively reducing the binding energy. In agreement with previous calculations on bulk MAPbI\({}_{3}\), we find the latter effect is more prominent.
Energetics of exciton formation. (a) Exciton binding energy under static (\(\Delta_{X,S}\) black circles) dynamic (\(\Delta_{X,D}\) blue triangles) and explicit lattice (red stars) models. Comparison to Weinberg et al is shown in green squares. (b) The difference in the exciton binding energies from dynamic and static screenings, which is defined as \(\delta_{\rm X}=\Delta_{\rm X,S}-\Delta_{\rm X,D}\) compared to Stokes shift measurements from Brennan et al and Protesescu et. al. Solid lines are fits to \(\Delta_{X}=\Delta_{X}^{\circ}(1+\ell_{1}/L+\ell_{2}^{2}/L^{2})\) and \(\delta_{X}=\delta_{X}^{\circ}(1+\ell^{\prime}/L)\).
With the explicit perovskite lattice, the extrapolated value to the large nanocrystal size limit is in good agreement with bulk CsPbBr\({}_{3}\) exciton binding energy of 40 meV, reflecting slight anharmonic weakening of the optical phonon.
Shown in (b) is the difference between the dynamic and static screening models, \(\delta_{X}=\Delta_{X,S}-\Delta_{X,D}\). This energy is due to polaron formation. The strong nanocrystal size dependence reflects the increasing confinement of the charges, which lower their self-energy by increasing their localization, as the energy of a charge in a dielectric will decrease like \(1/L\), which fits \(\delta_{X}\) well. We find the polaron formation energy agrees well with size-dependent Stokes shift measurements made from the difference between absorption and emission spectra. Results from two different experiments on CsPbBr\({}_{3}\) nanocrystals, compare quantitatively well with our computed \(\delta_{X}\), increasing from 10 m\(\,\)eV to over 450 m\(\,\)eV. Previously, a lattice origin of the Stokes shift had been disregarded due to the relatively small polaron formation energy in bulk. However, we find the polaron formation energy can increase substantially in small nanocrystals.
With the excitonic energetics understood, we now turn to the binding of biexcitons. The biexciton binding energy \(\Delta_{XX}\) for the explicit model, as well as the static and dynamic approximations, are shown in Fig. 3(a). All three models agree quantitatively, indicating that contrary from expectations from the exciton calculations, lattice effects are unimportant for biexciton binding. We rationalize this by noting that the primary contribution from the lattice for the exciton was the stabilization of the free charges due to the polaron formation, which does not contribute to the biexciton energy. In (b) is the electron-hole and electron-electron radial probability distributions, \(p(r)\), from the simulations with explicit lattice where electron-electron repulsion makes the average distance slightly larger than the average electron-hole distance. However this increase is small and the their size is largely determined by confinement rather than electron correlation. The similar masses for the electron and hole, produce a very weak dipole in the exciton or quadrapole in the biexcton that is not large enough to generate a polarization response from the lattice. Further, the similar results in both binding energies from explicit lattice and dynamic models imply that the anharmonicity coming from the explicit perovskite lattice does not play a crucial role in determining the behavior of biexcitons, likely due to the relatively small amplitude distortion. However, \(\Delta_{XX}\) is relatively large, 1/2-1/4 of \(\Delta_{X}\), indicating the importance of electron correlation.
As with \(\Delta_{X}\), we find a strong system size dependence of \(\Delta_{XX}\), increasing strongly with decreasing \(L\). We model this dependence as \(\Delta_{XX}=\Delta_{XX}^{\diamond}(1+\tilde{\ell}_{1}/L+\tilde{\ell_{2}}^{2}/L^ {2})\), with \(\tilde{\ell}_{1}\) and \(\tilde{\ell}_{2}\) employed as fitting parameters. While experimentally values of the \(\Delta_{XX}\) have been reported between -100 and 100 m\(\,\)eV, recent size dependent measurements on CsPbBr\({}_{3}\) nanocubes from two different groups are shown in Fig. 3(a) and agree well with our results.
In summary, we employed path integral molecular dynamics simulations to study quasiparticle energetics in perovskite nanocrystals. Using atomic models and simple approximations, we are able to systematically explore the roles of the lattice on effects on exciton and biexciton binding energies. We found that the effects from the lattice are large in exciton binding energy. but negligible in biexciton interaction, attributed to the weak coupling between the lattice and the neutral exciton complex. For all nanocrystals, regardless of the size and the type of screening employed, we found that biexcitons are bound, strongly suggestive that reports of anti-binding are misinterpreted.
_Acknowledgements._ This work was supported by the
Energetics and structure of biexcitons. (a) Biexciton binding energies for our explicit model (red stars) and its dynamical (blue squares) and static (black circles) approximations. Comparisons to Cho. et al and Amara et. al are shown in green squares. The solid line is a fit to \(\Delta_{XX}=\Delta_{XX}^{\diamond}(1+\tilde{\ell}_{1}/L+\tilde{\ell_{2}}^{2}/L^ {2})\). (b) Electron-hole (blue) and electron-electron (red) distributions of biexciton from the simulations with explicit lattice.
U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, Materials Sciences and Engineering Division, under Contract No. DEAC02-05-CH11231 (Grant No. KC3103). Y.P. acknowledges the Kwanjeong Educational Foundation. D.T.L. acknowledges the Alfred P. Sloan Foundation.
|
10.48550/arXiv.2307.06837
|
Biexcitons are bound in CsPbBr3 Perovskite Nanocrystals
|
Yoonjae Park, David T. Limmer
| 2,642
|
10.48550_arXiv.2011.02797
|
###### Abstract
Two alternative routes are taken to derive, on the basis of the dynamics of a finite number of dumbbells, viscoelasticity in terms of a conformation tensor with fluctuations. The first route is a direct approach using stochastic calculus only, and it serves as a benchmark for the second route, which is guided by thermodynamic principles. In the latter, the Helmholtz free energy and a generalized relaxation tensor play a key role. It is shown that the results of the two routes agree only if a finite-size contribution to the Helmholtz free energy of the conformation tensor is taken into account. Using statistical mechanics, this finite-size contribution is derived explicitly in this paper for a large class of models; this contribution is non-zero whenever the number of dumbbells in the volume of observation is finite. It is noted that the generalized relaxation tensor for the conformation tensor does not need any finite-size correction.
Keywords:Thermal fluctuations - Viscoelastic fluids - Dumbbell model - Conformation tensor - Finite-size effects +
Footnote †: journal: EPJ
## 1 Introduction
Fluctuations are particularly important when studying small systems. This also holds for fluids, including complex fluids, e.g., macromolecular and polymeric liquids. Small scales are involved, e.g., in microrheology and micro- and nanofluidic devices. For Newtonian fluids, i.e., fluids with a deformation-independent viscosity and a lack of memory, the dynamics on small scales could be described in terms of the fluctuating Newtonian fluid dynamics developed by Landau and Lifshitz. However, this is not sufficient for complex fluids, and thus extensions are needed. For example, the stress tensor has been related to the rate-of-strain tensor by a memory kernel, and correspondingly colored noise has been introduced on the stress tensor. Another approach towards modeling fluctuating effects in complex fluids has been taken by Vazquez-Quesada, Ellero, and Espanol and applied to microrheology, in which smoothed-particle hydrodynamics is extended by a conformation tensor that describes the conformation of the small number of polymer chains per volume element. The concept of fluctuating dynamics for the conformation tensor has been extended recently, to make it applicable not only to the Maxwell model, as in, but to a wider class of models, e.g. the FENE-P model and the Giesekus model. In the approach taken in, the Helmholtz free energy in terms of the conformation tensor plays an essential role.
The dynamics of the conformation tensor roots in a finer description, in particular, it can be related to the kinetic theory of dumbbells (e.g., see chapter 13 in). The question addressed in this paper is what lessons can be learned from deriving the dynamics for the conformation tensor with fluctuations from an underlying kinetic description for a finite number of dumbbells. It is pointed out that the dumbbell description already contains the relaxation and fluctuation effects that are relevant also on the conformation-tensor level. This is in contrast to coarse graining from an atomistic description to bead-spring chains or directly to the conformation tensor, e.g., see the work of Underhill and Doyle and of Ilg _et al._, respectively, without fluctuations on the conformation-tensor level.
The paper is organized as follows. In sect. 2, a certain class of kinetic dumbbell models is introduced, based on which a description of fluctuating viscoelasticity in terms of the conformation tensor is derived via a direct route,for a finite number of dumbbells. This route is paralleled in sect. 3, where a thermodynamic approach is taken to arrive at fluctuating viscoelasticity. As part of that, the finite-size correction to the Helmholtz free energy is calculated, and this is found to be essential for finding agreement between the two approaches. In the appendix, three different calculation methods for this free energy are detailed, each of which arrives at the same result. In sect. 4, the relation between the dumbbell models and the multiplicative decomposition of the conformation tensor, which has been discussed recently in the literature, is examined. The paper ends with conclusions and a discussion in sect. 5.
Throughout this paper, the following notation will be used. All summations are spelled out, i.e., no Einstein summation-convention is used for repeated indices. While the symbol \(\cdot\) denotes a contraction of one pair of indices, we use \(\odot\) for a double contraction: For an order-four tensor \(\mathbf{A}^{}\) and an order-two tensor \(\mathbf{B},\left[\mathbf{A}^{}\odot\mathbf{B} \right]_{\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!with two constants \(\beta_{1}\) and \(\beta_{2}\), and which accounts for the finite extensibility in a mean-field sense. Typically, in the FENE-P approximation (see also), \(\hat{\mathbf{c}}\) in eq. is used for infinitely many dumbbells. However, since in this paper the focus is on studying systems with a finite number \(N\) of dumbbells, we generalize this by using eq. for finite \(N\).
\[\Phi=\frac{N\beta_{2}k_{\mathrm{B}}T}{2}\ln f\,, \tag{8}\]
Therefore, using the approximation \(\Phi\), given by eq. with eq., instead of the exact energy \(\Phi_{\mathrm{exact}}\) is safe for small deformations. For other deformations, it is noted that the potential \(\Phi\) does include finite extensibility, albeit in a different way than if applied to each dumbbell individually. It is a topic of future research to examine the foundations of the mean-field approximation \(\Phi\) for finite \(N\) thoroughly. In this paper, this approximate expression for the energy forms part of defining the kinetic models that are examined in the following, and it is suitable for deriving closed dynamics for \(\mathbf{c}\), in the same spirit as the FENE-P approximation has been introduced earlier for infinite \(N\).
As far as the mobility tensor \(\mathbf{M}\) is concerned, its dependence on the dumbbell vectors \(\{\mathbf{Q}_{\mu}\}_{\mu=1,\ldots,N}\) is restricted to a dependence on the instantaneous conformation tensor,
\[\mathbf{M}=\mathbf{M}(\hat{\mathbf{c}})\,. \tag{9}\]
A particular realization of that is
\[\mathbf{M}=\frac{2}{\zeta}\left((1-\alpha)\mathbf{1}+\alpha\hat{\mathbf{c}} \right)\,, \tag{10}\]
In case of imposed deformation, a non-zero value for \(\alpha\) thus results in anisotropy of the friction tensor, where the anisotropy is introduced in a mean-field sense. The form corresponds to the widely used Giesekus model for anisotropic drag. For the Giesekus model, one typically uses eq. for infinitely many chains, i.e., \(N\to\infty\), to render the model solvable. However, the mobility of a dumbbell is affected primarily by the other dumbbells in its vicinity, and therefore the finite-\(N\) generalization is reasonable.
Allowing for the dumbbell potential energy \(\Phi_{\mathrm{d}}(\mathrm{tr}\left(\ldots\right))\) in eq. to be nonlinear and/or the dumbbell mobility tensor \(\mathbf{M}\) to depend on the conformation tensor effectively introduces mean-field type couplings of the individual dumbbells. In practice this implies that either (a) the \(N\) dumbbells must be in the vicinity of each other so they can interact, or (b) they diffuse rapidly enough in space to effect such interactions. In the case of the mobility tensor, the implied physics is reasonably clear: the assumption is that the average orientation of surrounding dumbbells affects the mobility of any given test dumbbell. Having this rationale in mind suggests that this mean-field mobility makes more sense for a finite number \(N\) of dumbbells, than it does for infinitely many. In the case of the potential energy, the microscopic physics of the implied coupling between dumbbells is less clear. Still, we find it worthy of note that a potential energy function exists from which the FENE-P model can be derived exactly.
In order to highlight the overall structure of the modeling in the remainder of this paper, the general forms eq. for the potential \(\Phi\) and eq. for the mobility tensor \(\mathbf{M}\) will be used. These general results can then be reduced to the Hookean dumbbell model, the FENE-P model, and the Giesekus model, respectively, by appropriate choices for the forms and parameters for the potential \(\Phi\) and the mobility tensor \(\mathbf{M}\), see table 1.
### Transition from dumbbells to the conformation tensor
Given the definition of the instantaneous conformation tensor \(\hat{\mathbf{c}}\), eq., and using the Ito interpretation of stochastic calculus, one has in general
\[\mathrm{d}\hat{\mathbf{c}}=\frac{1}{N}\sum_{\mu=1}^{N}\left((\mathrm{ d}\mathbf{Q}_{\mu})\mathbf{Q}_{\mu}+\mathbf{Q}_{\mu}(\mathrm{d}\mathbf{Q}_{\mu})+\langle( \mathrm{d}\mathbf{Q}_{\mu})(\mathrm{d}\mathbf{Q}_{\mu})\rangle^{\mathrm{Ito}}_{ \mathrm{d}t}\right)\,, \tag{11}\]
where \(\langle\ldots\rangle^{\mathrm{Ito}}_{\mathrm{d}t}\) implies that in \(\mathrm{d}\mathbf{Q}_{\mu}\) only terms involving the Wiener increments are kept and subsequently reduced according to the rule (see Table 3.1 in)
\[\mathrm{d}\mathbf{W}_{\mu}\mathrm{d}\mathbf{W}_{\mu}\to\mathrm{d}t\mathbf{1}\,. \tag{12}\]
Applied to the above class of models, the SDEs for the dumbbell vectors \(\{\mathbf{Q}_{\mu}\}_{\mu=1,\ldots,N}\) can be transformed into an SDE for the conformation tensor,
\[\mathrm{d}\hat{\mathbf{c}} = \left[\mathbf{\kappa}\cdot\hat{\mathbf{c}}+\hat{\mathbf{c}}\cdot\mathbf{\kappa}^{ \mathrm{T}}-2\mathbf{M}\cdot\left(2\frac{\partial\Phi_{\mathrm{d}}}{\partial( \mathrm{tr}\,\hat{\mathbf{c}})}\hat{\mathbf{c}}-k_{\mathrm{B}}T\mathbf{1}\right)\right] \mathrm{d}t \tag{13}\] \[+\frac{2k_{\mathrm{B}}T}{N}\left[\hat{\mathbf{c}}\cdot\frac{\partial} {\partial\hat{\mathbf{c}}}\cdot\mathbf{M}+\left(\hat{\mathbf{c}}\cdot\frac{\partial}{ \partial\hat{\mathbf{c}}}\cdot\mathbf{M}\right)^{\mathrm{T}}\right]\mathrm{d}t+ \mathrm{d}\hat{\mathbf{c}}^{\mathrm{f}}\,,\]
in terms of the dumbbell potential energy \(\Phi_{\mathrm{d}}\) and the dumbbell mobility tensor \(\mathbf{M}\), where it has been assumed
\begin{table}
\begin{tabular}{l|c c} \hline \hline Dumbbell & Parameters: & \\ model & \(\Phi\) & \(\mathbf{M}\) \\ \hline Hookean & \(\beta_{1}=1\), \(\beta_{2}\to\infty\) (\(f=1\)) & \(\alpha=0\) \\ FENE-P & \(\beta_{1}=1\), \(\beta_{2}\) finite (\(f=1\)) & \(\alpha=0\) \\ Giesekus & \(\beta_{1}=1\), \(\beta_{2}\to\infty\) (\(f=1\)) & \(\alpha>0\) \\ \hline \hline \end{tabular}
\end{table}
Table 1: Overview of the parameters in the potential \(\Phi\) given by eq. and eq., and the mobility tensor \(\mathbf{M}\), eq., for the three modelsthat \(\mathbf{M}\cdot\hat{\mathbf{c}}=\hat{\mathbf{c}}\cdot \mathbf{M}\). It is pointed out that the symmetry of \(\hat{\mathbf{c}}\) must be taken into account explicitly when calculating the partial derivatives of \(M\) (see for details).
\[{\rm d}\hat{\mathbf{c}}^{\rm f}=\frac{\sqrt{2k_{\rm B}T}}{N}\sum_{\mu= 1}^{N}\left(\left(\mathbf{B}\cdot{\rm d}\mathbf{W}_{\mu} \right)\mathbf{Q}_{\mu}+\mathbf{Q}_{\mu}\left(\mathbf{B}\cdot{\rm d}\mathbf{W}_{\mu}\right)\right)\,, \tag{14}\]
for a general mobility tensor \(M\), eq..
\[\langle{\rm d}\hat{\mathbf{c}}^{\rm f}_{ij}\rangle = 0\,, \tag{15}\] \[\langle{\rm d}\hat{\mathbf{c}}^{\rm f}_{ij}{\rm d}\hat{ \mathbf{c}}^{\rm f}_{kl}\rangle = \frac{2k_{\rm B}T}{N}\left(\hat{c}_{ik}M_{jl}+\hat{c}_{il}M_{jk}\right.\] \[\left.\hskip 28.452756pt+\hat{c}_{jk}M_{il}+\hat{c}_{jl}M_{ik} \right){\rm d}t\,.\]
The SDE for the conformation tensor with fluctuations obeying the statistical properties given by eq. and eq. is the benchmark to which the thermodynamic treatment further below will be compared, for the general class of models described by energy \(\Phi\), eq., and mobility tensor \(M\), eq.. If the potential \(\Phi\) and the mobility tensor \(M\) are of forms more general than eq. and eq., respectively, the dynamics for the conformation tensor would not close automatically, in which case one would have to employ procedures of coarse graining. The procedure presented in this sec. 2.2 has also been followed in for deriving the \(\hat{\mathbf{c}}\)-dynamics for the Hookean and FENE-P models.
It is noted that, in the limit of many dumbbells, i.e., \(N\to\infty\), not only do the fluctuations \({\rm d}\hat{\mathbf{c}}^{\rm f}\) become insignificant. In addition, also the "thermal drift", i.e., the second-last contribution on the right-hand side (r.h.s.) of eq. vanishes as well, and thus the conventional deterministic dynamics for the conformation tensor is recovered.
The Hookean dumbbell model - which results in the Maxwell model for the conformation tensor -, the FENE-P model, and the Giesekus model are sub-cases of the SDE with noise when choosing the parameters \(\beta_{1}\), \(\beta_{2}\), and \(\alpha\) appropriately, see table 1.
## 3 Fluctuating viscoelasticity derived using thermodynamics
The main idea in taking a thermodynamic approach to modeling dynamical systems is that one can concentrate on key ingredients for the static and dynamic properties, and the thermodynamic approach makes sure that these ingredients are processed towards the final model in a consistent way. For the dumbbell dynamics, the key ingredients are the potential \(\Phi\) and the mobility tensor \(M\). In contrast, for the dynamics of \(c\), the key ingredients are the Helmholtz free energy density \(\psi\) and the generalized relaxation tensor \(\mathbf{A}^{}\) (see for details).
\[\mathbf{B}^{}\odot\mathbf{B}^{,{\rm T}}=\mathbf{A}^{}\,, \tag{18}\]
The structure of the SDE with eq. is completely analogous to the one for the dumbbell models, eq. with eq..
In the following, we derive expressions for \(\psi\) and \(\mathbf{A}^{}\), based on those for \(\Phi\) and \(M\), respectively.
### Free energy density \(\psi(c)\) for finite \(N\)
The Helmholtz free energy \(\Psi=\Psi(c)\) for the symmetric conformation-tensor \(c\) for a finite number \(N\) of dumbbells is given by \(\Psi=-k_{\rm B}T\ln Z\), with the canonical partition-function
\[Z(\mathbf{c})=\int e^{-\Phi(\hat{\mathbf{c}})/(k_{\rm B}T)} \,\delta^{(K)}\left(\hat{\mathbf{c}}-\mathbf{c}\right)\,d^{DN }Q\,, \tag{19}\]
The \(K\)-dimensional Dirac \(\delta\)-function makes sure that only those states in \(\{\mathbf{Q}_{\mu}\}\)-space are accounted for that are compatible with the conformation tensor \(c\). Since \(\hat{\mathbf{c}}\) is symmetric by definition, see eq., only \(K=D(D+1)/2\) independent conditions are needed (instead of \(D^{2}\)); no more conditions are required for properly restricting the integration in \(\{\mathbf{Q}_{\mu}\}\)-space.
It is pointed out that \(\delta^{(K)}\) is actually a \(\delta\)-function in \(c\)-space in the sense that \(\int\delta^{(K)}(\hat{\mathbf{c}}-\mathbf{c})d^{K}c=1\).
\[\Xi=\int Z(\mathbf{c})d^{K}c=\int e^{-\Phi(\hat{\mathbf{c}} )/(k_{\rm B}T)}\,d^{DN}Q\,, \tag{20}\]
As we restrict our attention to energy functions \(\Phi\) which depend on \(\{\mathbf{Q}_{\mu}\}_{\mu=1,\ldots,N}\) only by way of \(\hat{\mathbf{c}}\), see eq., the canonical partition-function \(Z\) is related to the micro-canonical partition-function \(\Gamma\) by way of
\[Z(\mathbf{c})=e^{-\Phi(\mathbf{c})/(k_{\rm B}T)}\,\Gamma(\mbox {\boldmath$c$})\,, \tag{21}\]
with
\[\Gamma(\mathbf{c})=\int\,\delta^{(K)}\left(\hat{\mathbf{c}}- \mathbf{c}\right)\,d^{DN}Q\,. \tag{22}\]
Different procedures for calculating the dependence of \(\Gamma\) on \(c\) explicitly are discussed in Appendix B, one based on deriving a differential equation for \(\Gamma\), another one with a more geometrical interpretation, and a third one using a scaling argument.
\[\Gamma=\Gamma_{0}\left(\det\mathbf{c}\right)^{(N-D-1)/2}\,, \tag{23}\]
Based on eq. with eq., the Helmholtz free energy density \(\psi=\Psi/V\) per volume \(V\) becomes
\[\psi=\frac{\Phi}{V}-\frac{nk_{\rm B}T}{2}\ln\left(\det\mathbf{c} \right)+\Delta\psi\,, \tag{24}\]
with
\[\Delta\psi=\frac{k_{\rm B}T}{2V}(D+1)\ln\left(\det\mathbf{c}\right)\,, \tag{25}\]
Since \(\Phi\) is proportional to \(N\), it is evident that the first two contributions on the r.h.s. of eq. are independent of the size of the system, for given number density \(n\). In contrast, the third contribution, \(\Delta\psi\), does depend on the size of the system, in particular it becomes more relevant the smaller the system. To the best of our knowledge, this finite-size correction to the Helmholtz free energy (density) has not been derived earlier.
Using \(H=k_{\rm B}T\) and \(G=nk_{\rm B}T\), the Helmholtz free energy density with eq. for the three models discussed in table 1 agrees with standard literature (e.g., see), and with what has been used in the fluctuating-viscoelasticity approach in, with the important difference of the finite-size correction \(\Delta\psi\). Using, in contrast to our procedure, a continuous (representative of \(N\to\infty\)) distribution for the dumbbell vector \(Q\) (e.g., see), the thereby-derived Helmholtz free energy density corresponds to eq. where the finite-size correction \(\Delta\psi\) is absent.
### Relaxation tensor \(\mathbf{\Lambda}^{}\)
In the dumbbell dynamics, structural relaxation is expressed as \(-\mathbf{M}\cdot\left(\partial\Phi/\partial\mathbf{Q}_{\mu }\right)\mathbf{\rm d}t\). In the \(\mathbf{c}_{\mu}\)-dynamics, structural relaxation is expressed as \(-\mathbf{\Lambda}^{}\odot\left(\partial\psi/\partial\mathbf{c}\right)\mathbf{\rm d}t\), with an order-four relaxation tensor \(\mathbf{\Lambda}^{}\). In the translation from \(M\) to \(\mathbf{\Lambda}^{}\), a reduction of variables and the volume of the system \(V\) are involved, the latter being necessary since \(\Phi\) is an energy while \(\psi\) is an energy density. The relation between \(M\) and \(\mathbf{\Lambda}^{}\) is given by (see also, and sect. 6.4 in)
\[\mathbf{\Lambda}^{}=V\left(\sum_{\mu=1}^{N}\frac{\partial\hat{ \mathbf{c}}}{\partial\mathbf{Q}_{\mu}}\cdot\mathbf{M }\cdot\frac{\partial\hat{\mathbf{c}}}{\partial\mathbf{Q}_{ \mu}}\right)\,, \tag{26}\]
Using \(\partial\hat{c}_{ij}/\partial Q_{\mu,m}=(\delta_{im}Q_{\mu,j}+\delta_{jm}Q_{\mu,i})/N\), one finds
\[\Lambda^{}_{ijkl}=\frac{V}{N}\left(\hat{c}_{jl}M_{ik}+\hat{c}_{jk}M_{il}+ \hat{c}_{il}M_{jk}+\hat{c}_{ik}M_{jl}\right)\,. \tag{27}\]
Since \(M\) depends on the positions \(\{\mathbf{Q}_{\mu}\}_{\mu=1,\ldots,N}\) only by way of \(\hat{\mathbf{c}}\), taking the average is thus equivalent to replacing \(\hat{\mathbf{c}}\) by \(c\) everywhere in eq.. It is to be noted that there is _no_ finite-size correction in \(\mathbf{\Lambda}^{}\).
Using again \(G=nk_{\rm B}T\) and with \(\zeta=4k_{\rm B}T\lambda\), the relaxation tensor \(\mathbf{\Lambda}^{}\) given by eq. for the three models in table 1 turns out to agree with the standard expressions in the literature (e.g., see), and with what has been used in in the context of fluctuating viscoelasticity.
### Application to fluctuating viscoelasticity
According to the general procedure in, represented in eq., and using the Helmholtz free energy density with eq.
\[\mathbf{\rm d}\mathbf{c}\big{|}_{\Delta\psi} = -\mathbf{\Lambda}^{}\odot\frac{\partial(\Delta\psi)} {\partial\mathbf{c}}\mathbf{\rm d}t \tag{28a}\] \[= -2k_{\rm B}T((D+1)/N)\mathbf{M}\mathbf{\rm d}t\,. \tag{28b}\]
In particular, one obtains for the complete fluctuating dynamics
\[\mathbf{\rm d}\mathbf{c}=\left[\mathbf{ \kappa}\cdot\mathbf{c}+\mathbf{c}\cdot\mathbf{ \kappa}^{\rm T}-2\mathbf{M}\cdot\left(2\frac{\partial\Phi_{\rm d}}{ \partial(\mathbf{\rm tr}\,\mathbf{c})}\mathbf{c}-k_{ \rm B}T\mathbf{1}\right)\right]\!\mathbf{\rm d}t\] \[+\frac{2k_{\rm B}T}{N}\left[\hat{\mathbf{c}}\cdot\frac{ \partial}{\partial\hat{\mathbf{c}}}\cdot\mathbf{M}+\left( \hat{\mathbf{c}}\cdot\frac{\partial}{\partial\hat{\mathbf{c} }}\cdot\mathbf{M}\right)^{\rm T}\right]\!\mathbf{\rm d}t+ \mathbf{\rm d}\mathbf{c}^{\rm f}\,, \tag{29}\]
See Appendix A for explicit exemplary applications of this equation.
\[\mathbf{\rm d}\mathbf{c}^{\rm f}=\sqrt{\frac{2k_{\rm B}T}{N}} \left(\mathbf{b}\cdot\mathbf{\rm d}\tilde{\mathbf{W}} \cdot\mathbf{B}^{\rm T}+\mathbf{B}\cdot\mathbf{\rm d }\tilde{\mathbf{W}}^{\rm T}\cdot\mathbf{b}^{\rm T}\right)\,, \tag{30}\]
where \(b\) satisfies the condition
\[\mathbf{c}=\mathbf{b}\cdot\mathbf{b}^{\rm T}\,. \tag{31}\]
In general, \(B\) in eq. has dimensions \(D\times P\) with \(P\geq D\) (as described in sect. 2.1), \(b\) has dimensions \(D\times P^{\prime}\) with \(P^{\prime}\geq D\), and therefore \(\mathbf{\rm d}\tilde{\mathbf{W}}\) has dimensions \(P^{\prime}\times P\); for practical purposes, one may choose \(P=P^{\prime}=D\).
\[\langle\mathbf{\rm d}\tilde{W}_{ij}(t)\rangle = 0\,, \tag{32}\] \[\langle\mathbf{\rm d}\tilde{W}_{ij}(t)\,\mathbf{\rm d }\tilde{W}_{kl}(t^{\prime})\rangle = \delta_{ik}\,\delta_{jl}\,\delta(t-t^{\prime})\,dt\,dt^{\prime}\,. \tag{33}\]
When using component notation, eq. implies that any two of the components of \(\mathbf{\rm d}\tilde{\mathbf{W}}\) are independent from each other.
A direct comparison shows that the SDE derived directly from the dumbbell model and the SDE derived via the thermodynamic route, respectively, agree. Itis noted that the expressions for the fluctuations, \({\rm d}\hat{\mathbf{c}}^{\rm f}\) in eq. and \({\rm d}{\mathbf{c}}^{\rm f}\) in eq., respectively, have a different form. However, it can be shown that they have the same statistical properties. First, both representations are linear superpositions of increments of Wiener processes, and second, for both representations the average is given by eq. and the covariance by. The difference in the expressions for the fluctuations is not a short-coming of the approach; it rather reflects the non-uniqueness of the decompositions and. The non-uniqueness of the decomposition can actually be utilized for relating the expressions and for the noise in even more explicit terms. Specifically, choosing \(\mathbf{b}\) to be \(D\times N\) with the column vectors of \(\mathbf{b}\) equal to \(\mathbf{Q}_{\mu}/\sqrt{N}\) (\(\mu=1,\ldots,N\)) (see sect. 4 for a further elaboration), and setting the row vectors of \({\rm d}\tilde{\mathbf{W}}\) equal to \({\rm d}\mathbf{W}_{\mu}\) (\(\mu=1,\ldots,N\)), one finds that the expressions and are identical.
In deriving the relaxation term in eq., i.e. the first term on the r.h.s. that is proportional to \(\mathbf{M}\), from the general form eq., one notices the following: The prefactor \(1/N\) in \(\mathbf{\Lambda}^{}\) given by eq. is cancelled by the prefactor \(N\) in the derivative \(\partial(\psi-\Delta\psi)/\partial\mathbf{c}\), for \(\psi\) given by eq. with \(\Phi\) according to eq.. If one chose to not cancel these factors, one would observe that \((1/N)\mathbf{M}\) (with factor \(1/N\)) is the relevant mobility on the conformation-tensor level not only for the thermal drift and the fluctuations (by way of the covariances), but also for the relaxation.
The importance of the finite-size correction \(\Delta\psi\) in the free energy density, eq., for the evolution equation is pointed out. In particular, \({\rm d}{\mathbf{c}}\big{|}_{\Delta\psi}\) exactly cancels those contributions from \({\rm div}_{\mathbf{c}}\mathbf{\Lambda}^{}\) that are related to the derivative of the explicit factors \(\hat{\mathbf{c}}\) in \(\mathbf{\Lambda}^{}\), eq.. If \(\Delta\psi\) was neglected, agreement between the SDEs and could not be achieved.
In Appendix A, the dynamics for the conformation tensor with fluctuations, eq. with eq., is presented explicitly for three models, namely for the Hookean dumbbell / Maxwell model, the FENE-P model, and the Giesekus model.
## 4 Comments on the multiplicative decomposition of \(\mathbf{c}\)
### Eliminating degrees of freedom
Above, the relation has been established between the dynamics formulated in terms of dumbbell vectors, on the one hand, and in terms of the conformation tensor, on the other hand. In this section, the relation of the dumbbell-vector description to the multiplicative decomposition of the conformation tensor (e.g., see),
\[\hat{\mathbf{c}}=\hat{\mathbf{b}}_{P^{\prime}}\cdot\hat{\mathbf{b}}_{P^{\prime}}^{\rm T}\,, \tag{34}\]
This decomposition can be written in the form
\[\hat{\mathbf{c}}=\sum_{\nu=1}^{P^{\prime}}\hat{\mathbf{b}}_{P^{\prime},\nu}\hat{\mathbf{b} }_{P^{\prime},\nu}\,, \tag{35}\]
In general, one obviously must require \(P^{\prime}\geq D\) for this decomposition to be complete for arbitrary conformation tensor \(\hat{\mathbf{c}}\).
In view of the expression for the conformation tensor, a natural choice is \(P^{\prime}=N\) with \(\hat{\mathbf{b}}_{N,\mu}=\mathbf{Q}_{\mu}/\sqrt{N}\), relating the dynamics of \(\hat{\mathbf{b}}_{N}\) directly to that of the dumbbell vectors \(\mathbf{Q}_{\mu}\). In the following, we focus on Hookean dumbbells, i.e., the Maxwell model, for illustrative purposes.
\[{\rm d}\mathbf{Q}_{\mu}=\left[\mathbf{\kappa}\cdot\mathbf{Q}_{\mu}-\frac{1}{2\lambda}\mathbf{Q }_{\mu}\right]{\rm d}t+\frac{1}{\sqrt{\lambda}}{\rm d}\mathbf{W}_{\mu}\,, \tag{36}\]
Eq. translates directly into the dynamics of \(\hat{\mathbf{b}}_{N}\),
\[{\rm d}\hat{\mathbf{b}}_{N}=\left[\mathbf{\kappa}\cdot\hat{\mathbf{b}}_{N}-\frac{1}{2 \lambda}\hat{\mathbf{b}}_{N}\right]{\rm d}t+\frac{1}{\sqrt{N\lambda}}{\rm d}\hat{ \mathbf{W}}_{N}\,, \tag{37}\]
Let us now compare this result with the dynamics for the "square root" \(\mathbf{b}_{D}\) of the conformation tensor \(\mathbf{c}\), i.e. \(\mathbf{c}=\mathbf{b}_{D}\cdot\mathbf{b}_{D}^{\rm T}\) where \(\mathbf{b}_{D}\) has dimensions \(D\times D\), as derived in and amended in,
\[{\rm d}\mathbf{b}_{D}=\left[\mathbf{\kappa}\cdot\mathbf{b}_{D}-\frac{1}{2 \lambda}\big{(}\mathbf{b}_{D}-\mathbf{b}_{D}^{\rm-1,T}\big{)}-\frac{D}{2N\lambda}\mathbf{ b}_{D}^{\rm-1,T}\right]{\rm d}t\] \[+\frac{1}{\sqrt{N\lambda}}{\rm d}\tilde{\mathbf{W}}_{D}\,. \tag{38}\]
The close relation between \(\mathbf{b}_{D}\) and the elastic, i.e., recerable, part of the deformation gradient in solid mechanics has been discussed in. For \(N\to\infty\), the dynamics of the column vectors of \(\mathbf{b}_{3}\) agree with the treatment proposed in.
Two major differences between eq. and eq. are apparent. First, the relaxation in eq. drives the column vectors of \(\hat{\mathbf{b}}_{N}\) to zero, while, considering e.g. \(D=3\), according to eq. \(\mathbf{b}_{3,1}\) relaxes to \(\mathbf{b}_{3,2}\times\mathbf{b}_{3,3}/\det\mathbf{b}_{3}\), \(\mathbf{b}_{3,2}\) to \(\mathbf{b}_{3,3}\times\mathbf{b}_{3,1}/\det\mathbf{b}_{3}\), and \(\mathbf{b}_{3,3}\) to \(\mathbf{b}_{3,1}\times\mathbf{b}_{3,2}/\det\mathbf{b}_{3}\), respectively, i.e., the column vectors of \(\mathbf{b}_{3}\) become orthonormal in the course of relaxation. The second major difference between eq. and eq. relates to the absence of the thermal drift (the third term on the r.h.s. of eq.) in eq.. Both differences, in relaxation and thermal drift, are a hallmark of eliminating degrees of freedom when going to a reduced description of the dynamics. More specifically, both of these contributions are tightly related to the entropy, i.e., to the counting of states in configuration space, see sect. 3.1. It is pointed out that the difference in relaxation does not depend on the value of \(N\), while the thermal drift clearly is a finite-\(N\) effect, i.e., it is related to the fluctuations. However, despite these differences between eq. and eq., one should keep in mind that they both result in the same dynamics for the conformation tensor.
### Rotational dynamics
In, it has been discussed that, due to the non-uniqueness of the decomposition \(\mathbf{c}=\mathbf{b}_{D}\cdot\mathbf{b}_{D}^{\mathsf{L}}\), there is on-going rotational dynamics in eq. even at equilibrium, with a relaxation time that depends approximately linearly on \(N\). In the following, an attempt is made to rationalize this \(N\)-dependence of the rotational relaxation time in terms of the dynamics for the dumbbell vectors \(\mathbf{Q}_{\mu}\), eq..
\[\mathbf{R}=\sum_{\mu=1}^{N}\gamma_{\mu}\mathbf{Q}_{\mu}\,. \tag{39}\]
Based on eq., its dynamics is given by
\[\mathrm{d}\mathbf{R}=\left[\mathbf{\kappa}\cdot\mathbf{R}-\frac{1}{2\lambda}\mathbf{R}\right] \mathrm{d}t+\frac{1}{\sqrt{\lambda}}\mathrm{d}\mathbf{\bar{W}}\,, \tag{40}\]
where the Wiener process increments defined by \(\mathrm{d}\mathbf{\bar{W}}=\sum_{\mu=1}^{N}\gamma_{\mu}\mathrm{d}\mathbf{W}_{\mu}\) satisfy
\[\langle\mathrm{d}\mathbf{\bar{W}}\rangle =\mathbf{0}\,, \tag{41}\] \[\langle\mathrm{d}\mathbf{\bar{W}}\mathrm{d}\mathbf{\bar{W}}\rangle =\overline{\gamma^{2}}\mathrm{d}t\mathbf{1}\,, \tag{42}\]
Let us consider equilibrium, \(\mathbf{\kappa}=\mathbf{0}\). In order to study the orientation dynamics, we write \(\mathbf{R}=R\mathbf{n}\), with \(R\) and \(\mathbf{n}\) the length and the orientation vector of \(\mathbf{R}\), respectively.
\[\mathrm{d}R =\left(\frac{\overline{\gamma^{2}}}{2R\lambda}\left(D-1\right)- \frac{1}{2\lambda}R\right)\mathrm{d}t+\frac{1}{\sqrt{\lambda}}\mathbf{n}\cdot \mathrm{d}\mathbf{\bar{W}}\,, \tag{43}\] \[\mathrm{d}\mathbf{n} =-\frac{\overline{\gamma^{2}}}{2R^{2}\lambda}\left(D-1\right)\mathbf{ n}\mathrm{d}t+\frac{1}{R\sqrt{\lambda}}\left(\mathbf{1}-\mathbf{n}\mathbf{n}\right)\cdot \mathrm{d}\mathbf{\bar{W}}\,. \tag{44}\]
The contributions proportional to \(\overline{\gamma^{2}}\) originate from the second-order term in the Ito calculus 2.
Footnote 2: Note: It can be shown, again by using Ito calculus, that indeed \(\mathrm{d}|\mathbf{n}|=0\).
\[\lambda_{n}=\frac{2R^{2}}{\overline{\gamma^{2}}\left(D-1\right)}\lambda\,. \tag{45}\]
As an example, consider the case \(D=3\). To get an idea about the rotational dynamics of the column vectors \(\mathbf{b}_{3,\mu}\) of \(\mathbf{b}_{3}\), which satisfy \(\mathbf{c}=\sum_{\mu=1}^{3}\mathbf{b}_{3,\mu}\mathbf{b}_{3,\mu}\), we make a particular choice for \(\mathbf{R}\), i.e., for the linear combination. To that end, think of a Voronoi tessellation on the sphere, generated by the six "poles" that themselves are generated as intersections of three orthogonal axes with the sphere. Let us now consider two of these Voronoi sectors, \(\mathcal{V}_{i}\) and \(\mathcal{V}_{i}\), which are on opposite sides of the sphere. It is reasonable to assume that the number of vectors \(\mathbf{Q}_{\mu}\) which have their orientation in these two Voronoi sectors together can be written as \(\varphi N\), where \(\varphi\) is independent of \(N\). After mapping all vectors \(\mathbf{Q}_{\mu}\) with orientation in \(\mathcal{V}_{i}\) into \(\mathcal{V}_{i}\) by way of \(\mathbf{Q}_{\mu}\rightarrow-\mathbf{Q}_{\mu}\), which is just making use of the symmetry of dumbbell description, all \(\varphi N\) vectors with orientation in \(\mathcal{V}_{i}\) are averaged to obtain \(\mathbf{R}\), i.e., \(\gamma_{\mu}=1/(\varphi N)\) for these vectors and \(\gamma_{\mu}=0\) for all others. In this case, one finds \(\overline{\gamma^{2}}=1/(\varphi N)\), and thus the relaxation time for rotation is increased w.r.t. \(\lambda\), namely as \(\lambda_{n}\propto N\lambda\), in agreement with the observation in.
For the construction of a suitable vector \(\mathbf{R}\), the Voronoi sectors have been used, instead of, e.g., considering a random selection of vectors \(\mathbf{Q}_{\mu}\), for two reasons: First, the construction with Voronoi sectors allows to construct three such vectors (representative of the column vectors of \(\mathbf{b}_{3}\)) that are clearly linearly independent from each other. And second, the length of \(\mathbf{R}\) is proportional to the length of the dumbbell vectors \(\mathbf{Q}_{\mu}\) with a prefactor that is of order \(\mathcal{O}(1/2)\), i.e. independent of \(N\).
## 5 Discussion and conclusions
The focus of this paper has been on deriving viscoelasticity in terms of a conformation tensor with fluctuations, based on the kinetic theory of dumbbells. This has been achieved by identifying the conformation tensor with the arithmetic average over a finite number \(N\) of dumbbells, eq., and using two alternative routes for deriving the dynamics: a direct approach using stochastic calculus, and a thermodynamic approach, in which the Helmholtz free energy plays a key role. It has been shown that these two approaches agree only if a finite-size contribution to the Helmholtz free energy of the conformation tensor is taken into account. The main messages of this paper are therefore the following:
* If the number \(N\) of dumbbells is _finite_, the commonly employed expressions for the thermodynamic potentials need to be corrected: Using statistical mechanics (see Appendix B), one finds that the conformational entropy must be corrected by replacing \(N\) by \(N-D-1\) (with \(D\) the number of spatial dimensions), which in turn modifies the Helmholtz free energy, see eq. with correction term eq..
* The thereby obtained finite-size correction in the free energy is crucial for guaranteeing compatibility between the dynamics of the conformation tensor and of the underlying dumbbells.
While these general conclusions have been established in general terms for a large class of models, they have also been exemplified for the Hookean (Maxwell) model, the FENE-P model, and the Giesekus model (see Appendix A);the dynamics for the conformation tensor with fluctuations for these three models is summarized in table 2.
When discussing a model with fluctuations, the deterministic counterpart serves as a benchmark. For all models discussed in this paper, one recovers the known deterministic models in the thermodynamic limit \(N\to\infty\), i.e., if the number of dumbbells in the volume of interest \(V\) diverges, keeping the number density \(n=N/V\) constant. Beyond that thermodynamic limit, however, there is also an interest in the behavior of the average conformation tensor (\(\hat{\mathbf{c}}\)) for finite \(N\), i.e., in the presence of fluctuations. For most models studied in this paper, the nonlinearities do not allow to obtain a closed form equation for (\(\hat{\mathbf{c}}\)) based on the stochastic differential equation (SDE), eq. and eq., for the fluctuating conformation tensor \(\hat{\mathbf{c}}\). The notable exception to this rule is the Hookean dumbbell (Maxwell) model, with potential energy \(\Phi_{\rm d}(\operatorname{tr}\left(\mathbf{QQ}\right)=(H/2)\operatorname{tr} \left(\mathbf{QQ}\right)\) and mobility tensor \(\mathbf{M}=(2/\zeta)\mathbf{1}\). Taking the average of the SDE over different realizations of the fluctuations, one observes that the average conformation tensor (\(\hat{\mathbf{c}}\)) obeys the same differential equation as its deterministic (\(N\to\infty\)) counterpart - an observation that generally does not hold for nonlinear models. In particular, one finds for the Hookean dumbbell model at equilibrium (\(\hat{\mathbf{c}}\))\({}_{\rm eq}=\mathbf{1}\), where \(H=k_{\rm B}T\) has been used. As a word of caution, it is pointed out that the conformation tensor \(\hat{\mathbf{c}}\) that minimizes the Helmholtz free energy density, eq. with eq., is given by \(\hat{\mathbf{c}}_{\rm min}=\left(1-(D+1)/N\right)\mathbf{1}\), i.e., it does depend on the finite size (\(N\)) of the system. The fact that \(\hat{\mathbf{c}}_{\rm min}*(\hat{\mathbf{c}})_{\rm eq}\) is not a contradiction; it merely points out that the distribution of thermal fluctuations around the minimum is not symmetric.
The relevance of the finite-size correction of the Helmholtz free energy, \(\Delta\psi\) given by eq., has been discussed primarily in the context of formulating dynamics with fluctuations for the conformation tensor \(\mathbf{c}\). However, it is also of immediate consequence for the formulation of fluctuating viscoelasticity in terms of the "square root" \(\mathbf{b}_{3}\), where \(\mathbf{c}=\mathbf{b}_{3}\cdot\mathbf{b}_{3}^{\rm T}\). In, a thermodynamic approach has been taken towards deriving the dynamics of \(\mathbf{b}_{3}\), based on the dynamics of \(\mathbf{c}\). Therefore, if the thermodynamic potential for the \(\mathbf{c}\)-dynamics contains a finite-size contribution, the same holds true also for the thermodynamic potential for the \(\mathbf{b}_{3}\)-dynamics, see sect. 4.3 in for details. Beyond these implications for the thermodynamics of a \(\mathbf{b}_{3}\)-formulation, the kinetic models for \(N\) dumbbells have also been employed in this paper to give an explanation for the existence of a rotational relaxation time proportional to \(N\) in the fluctuating dynamics of \(\mathbf{b}_{3}\), which has been observed earlier.
|
10.48550/arXiv.2011.02797
|
Fluctuating viscoelasticity based on a finite number of dumbbells
|
Markus Hütter, Peter D. Olmsted, Daniel J. Read
| 1,090
|
10.48550_arXiv.1404.7213
|
## Abstract:
The rate coefficient formulae of unimolecular reactions are generalized to the systems with the power-law distributions based on nonextensive statistics, and the power-law rate coefficients are derived in the high and low pressure limits, respectively. The numerical analyses are made of the rate coefficients as functions of the \(\nu\)-parameter, the threshold energy, the temperature and the number of degrees of freedom. We show that the new rate coefficients depend strongly on the \(\nu\)-parameter different from one (thus from a Boltzmann-Gibbs distribution). Two unimolecular reactions, CH\({}_{3}\)CO\(\rightarrow\)CH\({}_{3}\)+CO and CH\({}_{3}\)NC\(\rightarrow\)CH\({}_{3}\)CN, are taken as application examples to calculate their power-law rate coefficients, which obtained with the \(\nu\)-parameters slightly different from one can be exactly in agreement with all the experimental studies on these two reactions in the given temperature ranges.
|
10.48550/arXiv.1404.7213
|
The rate coefficients of unimolecular reactions in the systems with power-law distributions
|
Cangtao Yin, Ran Guo, Jiulin Du
| 6,333
|
10.48550_arXiv.2410.11392
|
###### Abstract
Recent progress in machine learning (ML) has made high-accuracy quantum chemistry (QC) calculations more accessible. Of particular interest are multifidelity machine learning (MFML) methods where training data from differing accuracies or fidelities are used. These methods usually employ a fixed scaling factor, \(\gamma\), to relate the number of training samples across different fidelities, which reflects the cost and assumed sparsity of the data. This study investigates the impact of modifying \(\gamma\) on model efficiency and accuracy for the prediction of vertical excitation energies using the QeMFi benchmark dataset. Further, this work introduces QC compute time informed scaling factors, denoted as \(\theta\), that vary based on QC compute times at different fidelities. A novel error metric, error contours of MFML, is proposed to provide a comprehensive view of model error contributions from each fidelity. The results indicate that high model accuracy can be achieved with just 2 training samples at the target fidelity when a larger number of samples from lower fidelities are used. This is further illustrated through a novel concept, the \(\Gamma\)-curve, which compares model error against the time-cost of generating training samples, demonstrating that multifidelity models can achieve high accuracy while minimizing training data costs.
Introduction
Machine learning (ML) and quantum chemistry (QC) have become increasingly interlinked over the recent times. Both have seen rapid development in tandem allowing for quick prediction of QC properties in place of the costly conventional calculations. This has allowed researchers to perform preliminary examination of complex QC problems with much speed. The ML-QC pipeline first identifies a QC property of interest. Next, a training set is calculated for a desired QC method, say Density Functional Theory (DFT); this is also referred to as a _fidelity_, that is, the level of accuracy of the method with respect to what would be considered ground truth. Once a training dataset is computed, a ML model of choice is trained.
The bottleneck in such a pipeline is often the high cost of generating training data. A ML model can only be as good as the data it is trained on. It is often noted that a larger number of training samples results in a more accurate ML model. This observation implies that either one needs to use a less accurate, and thereby less expensive QC method to train the ML model, or have less training samples at a higher fidelity thereby, resulting in a less accurate ML model, when it comes to the prediction error relative to the data. Several methodological improvements over the single fidelity ML methods for QC have been proposed to overcome this hurdle in the ML-QC pipeline, including \(\Delta\)-ML, where one trains on the difference between two fidelities. This method has been shown to reduce the number of training samples needed at the expensive fidelity and has since been modified in various flavors including hierarchical ML, multifidelity ML (MFML), and optimized MFML (o-MFML). MFML and its variant of o-MFML, systematically combine several \(\Delta\)-ML like models with more than two fidelities. This method has been shown to be superior in predicting excitation energies along molecular trajectories. A recent work has also introduced the use of multitask Gaussian processes to harness heterogeneous multifidelity data in order to predict three-body interaction energy in water trimer with coupled cluster (CC) accuracy. MFML differs from the conventional \(\Delta\)-ML method not just in terms of the number of fidelities that are used but also in the number of training samples used at each fidelity. Conventionally, the \(\Delta\)-ML method uses the same number of samples at both the fidelities. In MFML, these training samples are scaled down as one increases the fidelity of the training data. This further reduced the number of costly training samples needed at the highest fidelity, also called the _target fidelity_. This _scaling factor_, in previous studies was set to be 2, meaning that at each subsequently lower fidelity, the number of training samples would be scaled up by a factor of 2. Ref. discusses that the scaling factor of 2 for MFML was decided based on previous work related to sparse grid combination techniques (SGCT).
The _scaling factor_, the ratio of training samples used at two consecutive fidelities, or levels, directly controls the total number of training samples used for MFML and thereby the cost of generating a training set for the approach. Understanding the effect of this parameter in the efficiency and accuracy of the MFML approach would potentially provide opportunities to further improve the overall multifidelity approach for QC. Previously, ref. has studied a two-fidelity MFML model with varying the number of training samples at the cheaper fidelity. By increasing the number of training samples at the lowest fidelity in an additive manner, the model error has been shown to decrease for the prediction of polymer bandgaps. A similar study has been performed in ref. for the study of bandgaps in solids. However, these studies lack any systematic assessment of the scaling factor itself but rather loosely study the effect of training data size within a two-fidelity data structure. This work assesses scaling factors that are different from those used thus far in literature. Several fixed scaling factors, that is, the same scaling factors across the different fidelities are systematically tested. These are evaluated on the recently released benchmarking multifidelity dataset, QeMFi, which consists of 135,000 geometries of nine complex molecules. Since the QeMFi dataset also provides the compute time for each fidelity for each molecule type, two time-cost informed scaling factors are also assessed.
Studying model accuracy in relation to the cost of generating the training set for the model also provides a robust measure of how the diverse MFML models behave with respect to the single fidelity models as has been shown in refs.. Therefore, assessment of model accuracy and time-cost of generating corresponding training data is made for the diverse scaling factors. In interest of a complete investigation not only into the scaling factors but also into better understanding the multifidelity data structure, this work further introduces a new error metric for multifidelity methods for QC, namely _error contours_ of MFML. Error contours describe the model error with respect to training samples used at two fidelities thereby giving a more comprehensive analysis of the contribution of each fidelity to the overall accuracy of the MFML model. The study of the error contours of MFML indicates that using much lower training samples at the costlier fidelity while increasing the number of training samples at the lowest fidelity results in an MFML model of high accuracy at a much lower cost than the conventional MFML approach with a fixed scaling factor. To systematically assess this, this work studies multifidelity models built with a small number of fixed training samples at the target fidelity and increasing the scaling factor. This gives rise to the notion of the \(\Gamma\)-curve as delineated in The \(\Gamma\)-Curve. The models that are build in such a manner are shown to be superior to the conventional MFML approach in terms of model error for a given cost of generating the training data.
The rest of this manuscript is structured as follows: all the methodological pre-requisites including dataset details are presented in Methods. The concepts of scaling factors and the tools used in this work are also explained in detail. This is followed by the results of MFML and o-MFML models for the prediction of excitation energies with the different scaling factors in Results. In addition to the time-cost of the different MFML models in Time Benefit Analysis, the error contours of MFML and the \(\Gamma\)-curves are studied in Multifidelity Error Contours and \(\Gamma\)-curves respectively. Inferences on these results are made followed by a discussion and outlook of this work along with its implications for future work in multifidelity methods.
Methods
This section discussed the various methodological pre-requisites needed to appreciate the results and inferences of this work. First, the dataset used is introduced and the multifidelity structure explained. This is followed by the ML details including KRR, MFML, and o-MFML. The section also discusses the concept of scaling factors in detail, to establish the conceptual motivation behind this work.
### Dataset
The QeMFi dataset is a recently released benchmark dataset of diverse molecules of varying chemical conformations. These molecules include urea, 2-nitrophenol, and thymine among others. It contains a total of 135,000 point geometries with diverse QC properties such as the first 10 vertical excitation energies in addition to the compute time for each molecule for each fidelity. These properties are calculated with the DFT formalism with five different basis sets, which in turn form the ordered multifidelity hierarchy. In increasing order of accuracy, these are STO3G, 321G, 631G, def2-SVP, and def2-TZVP. In this work, these fidelities are hereon referred to only with the basis set names, and often with the shortened notation such as TZVP for def2-TZVP.
The first excitation energies were taken from QeMFi as the property to be studied with the different data scaling applied. While QeMFi provides various other QC properties such as ground state energies and molecular dipole moments, the excitation energies are chosen since they are generally more challenging for ML model than ground state energies. Although there is increasing interest in vector properties such as molecular dipole moments, these are not studied here since they would require an extensive discussion of equivariance and invariance of molecular descriptors, that is the map between the Cartesian coordinates and machine learnable input features. This discussion lies out of the scope of this current work. Thus the excitation energies of QeMFi are used herein. The energies are given in cm\({}^{-1}\)As a first step, from the 135,000 total geometries in QeMFi, 120,000 were randomly sampled to form the diverse training sets. From the remaining 15,000 geometries, a validation set of 2,000 samples was set aside to be used in the o-MFML method. Finally, the remaining 13,000 samples were set aside as a test set. Both the validation set and the test set are not changed over the course of this entire study, that is, all errors are reported on the same test set. It is to be noted that the test set is never used during the training or validation phases of any of the models. Therefore the error reported by the models is on a set of truly unseen data. The validation set and the test set are fixed and not changed during the course of the experiments in this work.
In order to assess the efficiency of the multifidelity models vis-a-vis the single fidelity model, the time-cost of generating the training data versus model error are studied. For this purpose, the QC-compute time of each fidelity is considered from the QeMFi dataset. As ref. 21 notes, the QC-compute time provided in the dataset is the time of a single-core QC-calculation of a fidelity for a given molecule. To compute the training dataset generation time, these calculation times are used.
### Kernel Ridge Regression
Given a fidelity, \(f\), consider a training set of molecular representations (denoted as \(\boldsymbol{X}_{q}\)) and corresponding excitation energies (denoted as \(y_{q}^{(f)}\)) given by \(\mathcal{T}^{(f)}:=\{(\boldsymbol{X}_{i},y_{i}^{(f)})\}_{i=1}^{N_{\text{rain}}}\). The molecular representations used in this assessment are the unsorted Coulomb Matrix(CM) representations.
\[C_{i,j}:=\begin{cases}\frac{Z_{i}^{2.4}}{2}\,&i=j\\ \frac{Z_{i}\cdot Z_{j}}{\|\boldsymbol{R}_{i}-\boldsymbol{R}_{j}\|}\,&i\neq j \,\end{cases} \tag{1}\]
The resulting representation for a molecule is then flattened into a 1-D array. Since the CM representation is symmetric for the atomic indices, a molecule with \(m\) atoms is therefore represented by a 1-D array of size \(m(m+1)/2\). The QeMFi dataset consists of molecules of differing number of atoms, therefore the CM are padded with zero to match with the size of the CM representation of the largest molecule of the dataset which is o-HBDI with 23 atoms. This corresponds to a padded representation size of 253 entries per geometry.
For a query CM representation given as \(\mathbf{X}_{q}\), the prediction of energies at fidelity \(f\) are given by:
\[P_{\text{KRR}}^{(f)}\left(\mathbf{X}_{q}\right):=\sum_{i=1}^{N_{\text{train}}^{(f) }}\alpha_{i}^{(f)}k\left(\mathbf{X}_{q},\mathbf{X}_{i}\right)\, \tag{2}\]
In this work, the Matern kernel of first order and discrete \(L_{2}\) norm is used across the KRR models built.
\[k\left(\mathbf{X}_{i},\mathbf{X}_{j}\right)=\exp\left(-\frac{\sqrt{3}}{\sigma}\left\| \mathbf{X}_{i}-\mathbf{X}_{j}\right\|_{2}^{2}\right)\cdot\left(1+\frac{\sqrt{3}}{ \sigma}\left\|\mathbf{X}_{i}-\mathbf{X}_{j}\right\|_{2}^{2}\right)\, \tag{3}\]
The coefficients of KRR, \(\mathbf{\alpha}^{(f)}\) from Eq., are computed by solving
\[(\mathbf{K}+\lambda\mathbf{I})\mathbf{\alpha}^{(f)}=\mathbf{y}^{(f)} \tag{4}\]
for each molecule pair in the training set. In Eq., \(\lambda\) is a parameter that penalizes overfitting.
### Multifidelity Machine Learning
Refs. have shown that the multifidelity machine learning (MFML) model for some QC property can be built by systematically combining individual _sub-models_ that are themselves trained for a given fidelity, \(f\). Within this approach, each sub-model is then identified by the fidelity and the corresponding number of training samples used, \(N_{\text{train}}^{(f)}\). This is carried out by a composite index, \(\mathbf{s}=(f,\eta_{f})\), where \(2^{\eta_{f}}=N_{\text{train}}^{(f)}\).
In mathematical formalism, a MFML model is given as:
\[P_{\text{MFML}}^{(F,\eta_{F};f_{b})}\left(\mathbf{X}_{q}\right):=\sum_{\mathbf{s}\in \mathcal{S}^{(F,\eta_{F};f_{b})}}\beta_{\mathbf{s}}P_{\text{KRR}}^{(\mathbf{s})}\left( \mathbf{X}_{q}\right)\, \tag{5}\]
This selection is decided by the choice of the _baseline fidelity_, \(f_{b}\), and the number of training samples at the highest fidelity (also called the target fidelity), \(N_{\text{train}}^{(F)}=2^{\eta_{F}}\). The sub-models for a MFML model with a given \(F\), \(\eta_{F}\), and \(f_{b}\) are identified as discussed previously in ref.. In Eq., \(\beta_{\mathbf{s}}\) are the coefficients of the linear combination of these selected sub-models.
For MFML, the \(\beta_{\mathbf{s}}\) are selected as
\[\beta_{\mathbf{s}}^{\text{MFML}}=\begin{cases}+1,&\text{if }f+\eta_{f}=F+\eta_{F} \\ -1,&\text{otherwise}\end{cases}\, \tag{6}\]
The value of \(\mathbf{\beta}_{s}\) prescribed in the above equation can be better understood with the following example. Consider a 3 fidelity dataset, that is \(f\in\{1,2,3\}\) with \(f=3=F\), and \(f=1=f_{b}=1\). If one uses \(N_{\text{train}}^{}=2^{2}\) then \(\eta_{F}=2\). Therefore, following Eq. and Eq.
\[P_{\text{MFML}}^{}(\mathbf{X}_{q}):=\left(P_{\text{KRR}}^{}(\mathbf{X}_{q} )-P_{\text{KRR}}^{}(\mathbf{X}_{q})\right)+\left(P_{\text{KRR}}^{}(\mathbf{X }_{q})-P_{\text{KRR}}^{}(\mathbf{X}_{q})\right)+P_{\text{KRR}}^{}(\mathbf{X }_{q})\,\]
Recently, a methodological development over conventional MFML was proposed and shown to be superior to MFML in ref.. Termed optimized MFML (o-MFML), it optimizes the combination of the sub-models of MFML. In other words, it considers coefficient values different from those defined in Eq. for the different sub-models by optimizing over a validation set \(\mathcal{V}_{\text{val}}^{F}:=\{(\mathbf{X}_{q}^{\text{val}},y_{q}^{\text{val}})\}_{q=1} ^{N_{\text{val}}}\).
\[\beta_{\mathbf{s}}^{\text{opt}}=\arg\min_{\beta_{\mathbf{s}}}\left\|\sum_{v=1}^{N_{ \text{val}}}\left(y_{v}^{\text{val}}-\sum_{\mathbf{s}\in S^{(F,\eta_{F};f_{b})}} \beta_{\mathbf{s}}P_{\text{KRR}}^{(\mathbf{s})}\left(\mathbf{X}_{v}^{\text{val}}\right) \right)\right\|_{p}\, \tag{8}\]
The optimization procedure used in this work is ordinary least squares (OLS) which uses a \(p=2\) norm.
### Scaling Factors
Thus far, in both MFML and o-MFML, the number of training samples used for each fidelity are scaled by a _scaling factor_ of \(\gamma=\)2 based on research on SGCT. For example, if one has \(N_{\text{train}}^{(F)}=32\) training samples for fidelity \(F\) then it is that \(N_{\text{train}}^{(F-1)}=64\) training samples for fidelity \(F-1\), \(N_{\text{train}}^{(F-2)}=128\) training samples on fidelity \(F-2\), and so on. The scaling up of the training samples as one decreases the fidelities can be thought of intuitively from a perspective of sparseness of data. As one increases the fidelity, the cost of QC calculation increases. This results in lower number of point-calculations that need to be made at this fidelity.
The scaling factor can itself be varied to assess its effect on the model error. This work tests five such scaling factors, in particular, \(\gamma\in\{2,3,4,5,6\}\). For each of these scaling factors, the training set size increases exponentially as one goes down the fidelities. If one starts with \(N_{\text{train}}^{F}\) samples at the target fidelity, then at each lower fidelity, \(f<F\), the number of training samples would be \(N_{\text{train}}^{f}=\gamma^{F-f}\times N_{\text{train}}^{F}\). As an example, if \(F=5\), \(N_{\text{train}}^{F}=2\)and \(\gamma=3\), then for \(f_{b}=1\) the training set size for each fidelity, in increasing order of the fidelity, would be \(\{3^{4}\cdot 2,3^{3}\cdot 2,3^{2}\cdot 2,3\cdot 2,2\}\). The variation of the training set sizes with increasing values of \(\gamma\) is represented pictorially in Figure 1(a)-(b).
This work studies two additional approaches for QC-cost adapted selection of scaling factors. This approach takes into account the compute times for each fidelity before adaptively selecting the ratio of training samples between two consecutive fidelities. Previous literature for MFML constructs these ML models with a scaling factor of 2 for each consecutive fidelity resulting from work on SGCT. Theoretical work in SGCT arrives at such a factor by analyzing the cost-benefit ratio of the different levels, or fidelities.
A hypothetical comparison of training data used across fidelities for the different kinds of scaling factors used in this work. a) The multifidelity training data structure used in MFML with a small fixed scaling factor (\(\gamma\)). b) Multifidelity training data structure for a large fixed scaling factor (\(\gamma\)) results in a larger number of training samples being used at the cheaper fidelities. c) The structure of multifidelity training data used for scaling factors that are decided based on the QC-time cost, explained in Scaling Factors as \(\theta_{f}^{F}\) and \(\theta_{f-1}^{f}\). d) Comparison of training data structure evolution for conventional MFML and the \(\Gamma\)-curve introduced in The \(\Gamma\)-Curve. Notice how the number of training samples used at the target (the costliest) fidelity remain same across the data structure for the \(\Gamma\)-curve while they increase for the conventional MFML method.
Simple put, there is an inherent assumption that \(T_{f+1}/T_{f}\approx 2\). Therefore, one can consider the time-cost informed scaling factors between consecutive fidelities to be a generalization of this assumption. The QeMFi dataset provides the QC compute time in seconds for each of the five fidelities when computed on a single core. This information can be used to determine a time-informed scaling factor for each fidelity as opposed to setting a single scaling factor for all the fidelities.
While there could be different ways to determine these time-informed scaling factors, the most trivial approach is to take the nearest integer value of the ratio of the compute times for the subsequent fidelity. That is, one can define \(\theta_{f-1}^{f}:=\lfloor T^{f}/T^{f-1}\rceil\), where \(\lfloor\cdot\rceil\) denotes integer rounding. This specific choice of scaling factors is made to take into account the relative time-cost of the fidelities used in the MFML model. It is reasonable to assume that the number of training samples used at consecutive fidelities should be based on the ratio of the cost of those fidelities. Since QeMFi is a collection of different molecules, this approach was carried out with reference to the compute times for the largest molecule in the database: o-HBDI. This results in scaling factors \(\theta_{f-1}^{f}=\{3,1,2,1\}\) for increasing fidelity. That is, at SVP, the same number of training samples as TZVP are used while at the 631-G fidelity, it is twice, and so on. However, MFML models are built in such a way that subsequently cheaper fidelities have some more training samples than the previous fidelity so that the difference between the sub-models can be taken. In order to achieve this, after the number of training samples are decided by the scaling factors, if fidelity \(f-1\) has the same number of training samples as fidelity \(f\), then one additional sample is added to the sub-model at fidelity \(f-1\). As an example, if \(N_{\text{train}}^{\text{TZVP}}=2\), then the training samples for the different fidelities would be \(\{12,4+1,4,2+1,2\}\). Hereon, the MFML models built with this approach of scaling factors are referred by \(\theta_{f-1}^{f}\).
A second approach of implicitly incorporating the time-cost of the fidelities is to take the ratio of the compute times with respect to the target fidelity. This approach is motivated by posing the question, what amount of training data used at a specific fidelity would cost the same as the training data used at the target fidelity. Once again, the nearest integer value is considered. This leads to the definition of \(\theta_{F}^{f}:=\lfloor T^{F}/T^{f}\rceil\) for all \(f<F\). As for the case of \(\theta_{f-1}^{f}\), the reference molecule was chosen to be o-HBDI. This leads to scaling factors \(\theta_{f}^{F}=\{9,3,2,1\}\) for increasing fidelity. Since the SVP fidelity is scaled by a factors of 1, as discussed earlier, one additional training samples was added each time to maintain the multifidelity structure required for MFML. As an example, consider the case for \(N_{\text{train}}^{\text{TZVP}}=2\). Then the training samples at the different fidelities would be \(\{108,12,4,2+1,2\}\). This formulation of scaling factors is hereon associated with \(\theta_{f}^{F}\). Scaling factors based on QC-compute cost as diagrammatically depicted in Figure 1(c).
### Error Contours of MFML
The prediction error of the different ML and MFML models assessed in this work are given as relative mean absolute errors (RMAE) which is calculated over the holdout test set as:
\[\text{RMAE}=\frac{1}{N_{\text{test}}}\sum_{q=1}^{N_{\text{test}}}\left|\frac{P _{\text{ML}}\left(\mathbf{X}_{q}^{\text{ref}}\right)-y_{q}^{\text{ref}}}{y_{q}^{ \text{ref}}}\right|. \tag{9}\]
Due to the multifidelity data structure a simple cross validation approach cannot be used. Instead, the RMAE of the MFML models is calculated for a 10-run average as proposed and implemented in ref. 13, which accounts for the nested multifidelity data structure. For most ML approaches, the RMAE is reported for increasing training set sizes, thereby resulting in a learning curve. The learning curves is used as an indicator of the ability of the ML model to predict over unseen data. Learning curves form a major part of the analysis offered in this work. In addition to the usual RMAE versus training samples learning curves, this work also studies the RMAE versus cost of generating training data for the multifidelity model as first proposed and implemented in ref. 12 for excitation energies.
The analysis of the learning curves for the different values of \(\gamma\) (see Learning Curves and Time Benefit Analysis) indicate that the MFML training data structure needs only very little training samples at the higher fidelity. The contribution of each fidelity and the number of training samples at each fidelity is more complex than just the QC-time cost of the fidelity. In interest of studying the individual contribution of each fidelity and the diverse training samples choices at each fidelity, this work introduces a new error metric, namely, error contours. As a conceptual extension of learning curves, error contours of MFML report MFML model error as a function not simply of training samples chosen at a single fidelity but as a function of training samples, \(N_{\text{train}}^{f}\) and \(N_{\text{train}}^{f+1}\), chosen at two consecutive fidelities, \(f\) and \(f+1\). This form of analysis helps better understand the contribution of these fidelities to the overall model accuracy. Since the error contours can be studied for all consecutive fidelity pairs, these provide an in depth understanding of the contribution of each fidelity \(f_{b}\leq f\leq F\) to the MFML model \(P_{\text{MFML}}^{(F,\eta_{F};f_{b})}\) in terms of model accuracy. Consecutive fidelities are studied instead of an arbitrary pair of fidelities in order to systematically assess what happens to the model error as one adds a cheaper fidelity while increasing the training set sizes. In other words, the error contour is the RMAE of the MFML model by varying the training samples at two fidelities simultaneously. Since this work uses the QeMFi dataset that contains five fidelities, the error contours of MFML are studied for the following fidelity pairs: TZVP-SVP, SVP-631G, 631G-321G, and 321G-STO3G. Since the error contours are a function of two variables, \(N_{\text{train}}^{f}\) and \(N_{\text{train}}^{f+1}\), they are reported as contour plots. Herein, error contours of MFML are discussed only for \(\gamma=2\). The error contours give a better view into the contribution of a multifidelity data structure to model accuracy for a given target fidelity. The investigation of error contours for each fidelity pair indicates, in some sense, the weighted contribution of those fidelities to the overall model. A better understanding of this contribution will aid the choice of \(N_{\text{train}}^{f}\) for each fidelity that constitutes the MFML model.
### The \(\Gamma\)-Curve
The study of the error contours in Multifidelity Error Contours indicates that the multifidelity data structure can provide a high-accuracy model with a much lower number of costly training samples than the conventional MFML data structure approach. Coupled with the results of studying the learning curves for different scaling factors (see Time Benefit Analysis and Multifidelity Error Contours), a new multifidelity approach is proposed: the \(\Gamma\)-curve.
In this approach, a fixed number of training samples are chosen at the highest fidelity, \(N_{\rm train}^{\rm TZVP}\). An o-MFML model is built with \(\gamma=2\). The cost of the training data is noted along with the model error over the holdout test set. For the next step of this curve, instead of varying \(N_{\rm train}^{\rm TZVP}\), \(\gamma\) is increased by an integer value. This curve is identified as \(\Gamma(N_{\rm train}^{\rm TZVP})\)-curve and is a measure of RMAE versus time-cost of training data of the multifidelity model for varying \(\gamma\). For example, if one were to set \(N_{\rm train}^{\rm TZVP}=2\), \(f_{b}:321\)G, then the \(\Gamma\)-curve would be built with the first multifidelity training data structure (in increasing fidelity) as \(\{2^{3}\cdot 2,2^{2}\cdot 2,2^{1}\cdot 2,2\}\). The next point would be built with a training data structure of \(\{3^{3}\cdot 2,3^{2}\cdot 2,3^{1}\cdot 2,2\}\) and so on. In general, for a \(\Gamma(N_{\rm train}^{\rm TZVP})\)-curve, the training data structure for \(f:b\ 321\)G is given as \(\{\gamma^{3}\cdot N_{\rm train}^{\rm TZVP},\gamma^{2}\cdot N_{\rm train}^{\rm TZVP },\gamma^{1}\cdot N_{\rm train}^{\rm TZVP}\}\). For reasons explained in Time Benefit Analysis and Multifidelity Error Contours, the STO3G baseline is not considered. Herein, the \(\Gamma(\cdot)\)-curve is studied for \(N_{\rm train}^{\rm TZVP}\in\{2,8,64\}\) to further assess the multifidelity structure of training data and evaluate the limits of the multifidelity approach. Since there is no trivial way to express the number of training samples used at a certain fidelity and relate it to the RMAE of the model, the \(\Gamma(\cdot)\)-curve is plotted only as RMAE versus multifidelity training data generation cost. The variation of the number of training samples used at each fidelity for the \(\Gamma\)-curve is contrasted with that for the learning curve of MFML in Figure 1(d). Notice the wider base of the pyramid for \(\Gamma\)-curve signifying a much larger number of training samples being used at the cheapest fidelities.
Results
This section presents the analysis of varying the scaling factor for MFML and o-MFML. The results are presented in two major formats. First, standard learning curves of RMAE versus number of training samples used at the highest fidelity of TZVP are presented. Following this, the model error is assessed a function of the time-cost of generating the training data for the model. This assessment from ref. 12 informs of the effectiveness of the diverse models that are studied in this work. Once these results are interpreted, error contours of MFML as described in Error Contours of MFML are studied.
### Learning Curves
The primary assessment of the effect of different scaling factors is carried out using learning curves for the resulting MFML and o-MFML models. These learning curves are shown in for different scaling factors. In all cases, the scaling factors are constant across the different fidelities as explained in Scaling Factors.
Multifidelity learning curves for the prediction of excitation energies taken from the QeMFi dataset. The top row corresponds to the MFML models while the bottom row is for the o-MFML models. Different fixed scaling factors are used to scale the data across each fidelity in the multifidelity models as explained Scaling Factors. The scaling factors are reported on the top of each column.
The learning curves are shown for varying baseline fidelities. A single fidelity KRR learning curve is also shown for reference. The RMAEs are reported as unitless quantities.
In Figure 2, the first column shows the learning curves for the scaling factor of 2. This is the original scaling factor used in refs. and is used as a reference to evaluate the other scaling factors against. Within these reference results, one observes that the addition of cheaper baselines results in a constantly lowered offset of the learning curves. The interpretation from the lowered offsets is that similar models errors can be achieved with lower number of training samples at the target fidelity with the addition of cheaper fidelities. With the cheapest fidelity, STO3G, being added to the multifidelity model, one observes RMAE of 0.4 with around 200 training samples at TZVP. In comparison to this, an increase of of the scaling factor, \(\gamma\), provides MFML models that achieve similar errors for lower number of training samples at TZVP. For example with a scaling factor of 3, the STO3G baseline MFML model achieves an RMAE of 0.4 with \(N_{\text{train}}^{\text{TZVP}}=32\). With a scaling factor of 6, the number of training samples at TZVP needed to achieve this same error is lowered further to about 4. The learning curves for o-MFML also indicate the same across varying scaling factors. There is little difference between the learning curves for MFML and o-MFML. This could be due to the MFML combination being already optimal for this multifidelity data structure as has been argued in ref.. There is however, slight improvement in all cases of o-MFML and it does result in reduced RMAEs across the different scaling factors.
From these results it appears that a higher number of training samples with cheaper fidelities improves the predictive capabilities of the MFML and o-MFML models. One possible reason for this could be that the use of larger data at the cheaper fidelities results in more information about the overall multifidelity structure being included into the MFML models.
In addition to the fixed scaling factors across fidelities, two special cases of scaling factors were introduced in Scaling Factors based on the QC compute time of each fidelity. Thesewere denoted by \(\theta_{f-1}^{f}\) and \(\theta_{f}^{F}\) as explained in Scaling Factors in detail. Learning curves were generated for both MFML and o-MFML models for both these cases. The results are shown in Figures 3 and 4 for both scaling factor cases with various baseline fidelities. The single fidelity KRR learning curves is also depicted for reference.
As explained in Scaling Factors, these scaling factors are based on the ratio of QC-compute times of subsequent fidelities. Between some fidelities - namely between TZVP and SVP, and between 631G and 321G - this scaling was observed to be 1. It is anticipated that these fidelities will not significantly improve the MFML models since there is very little additional information that is being added to the model. Indeed, as seen in Figure 3, the multifidelity model built with SVP baseline does not provide any improvement over the single fidelity KRR. This is due to the fact that the number of training samples at both fidelities are nearly identical, only different by 1 sample, due to the scaling factor. This same observation can be made for the learning curves with 321G as baseline fidelity. With 631G and STO3G baselines, however, one observes improvement of the MFML and o-MFML models. With the STO3G baseline, MFML and o-MFML reach an RMAE of 0.03 with roughly 500 training samples at TZVP.
MFML and o-MFML learning curves for scaling factors, \(\theta_{f-1}^{f}\), between fidelities chosen as ratios of the QC compute time of subsequent fidelities. Single fidelity KRR at TZVP is also shown for reference. Single fidelity KRR learning curves are also provided for reference. The legend describes the baseline fidelity, \(f_{b}\), of the multifidelity model.
Similarly, reports the learning curves for \(\theta_{f}^{F}\). The left-pane shows the results for MFML, while the right pane shows those for o-MFML. The SVP baseline fidelity once again shows very little improvement over the single fidelity KRR due to the scaling factor being unity (see Scaling Factors). However, each additional cheaper baseline fidelity, results in lowered offsets of the corresponding learning curves. With \(N_{\rm train}^{\rm TZVP}=256\), the multifidelity models with the STO3G baseline result in RMAE of \(\sim\)0.03.
To aid comparison of the different scaling factors discussed so far, depicts the
\begin{table}
\begin{tabular}{|c|c|c|} \hline
## Factor** & **MFML** & **o-MFML
\\ \hline \hline
2 & 0.0790 & 0.0754 \\
3 & 0.0486 & 0.0472 \\
4 & 0.0371 & 0.0297 \\
5 & 0.0312 & 0.0264 \\
6 & 0.0290 & 0.0264 \\ \(\theta_{f-1}^{f}\) & 0.1201 & 0.1143 \\ \(\theta_{f}^{F}\) & 0.0844 & 0.0854 \\ \hline \end{tabular}
\end{table}
Table 1: RMAE rounded off to 4 decimal points for MFML and o-MFML models built with the STO3G baseline fidelity for \(N_{\rm train}^{\rm TZVP}=2^{5}\). This allows for a uniform comparison of the model accuracy not just between MFML and o-MFML but also across the scaling factors that are studied in this work. Notice that the learning curve for \(\gamma=6\) only goes up to \(N_{\rm train}^{\rm TZVP}=2^{5}\) and therefore this is chosen as a comparison point for all other curves.
MFML and o-MFML learning curves for scaling factors, \(\theta_{f}^{F}\), between fidelities selected as ratios of the QC compute time of that fidelity to the compute time of TZVP, that is the target fidelity. Single fidelity KRR learning curves are also provided for reference. The legend describes the baseline fidelity, \(f_{b}\), of the multifidelity model.
The various factors are delineated in the legend of the plot. This plots shows that increasing values of \(\gamma\) result in a lowered constant offset of the learning curves. In contrast, the multifidelity models built with time-informed scaling factors, \(\theta_{f-1}^{f}\) and \(\theta_{f}^{F}\) both show the highest model error. This observation is consistent for both MFML and o-MFML models as can be seen from the two plots shown in Furthermore, the o-MFML models show lower errors than the MFML counterparts for all the cases as seen in table 1 which reports the RMAEs for the MFML and o-MFML models with various scaling factors for the STO3G baseline for \(N_{\text{train}}^{\text{TZVP}}=2^{5}\) for ready reference. This training set size is chosen so that there is uniform comparison between the different scaling factors. The behavior of the MFML models with increasing \(\gamma\) is in some sense expected since an increasing value of the scaling factor implies an increased amount of training data, albeit only at the cheaper fidelities. This could be one potential reason to explain the lowered offsets that are observed. An increased amount of training samples at the lowered fidelities, due to the nested structure of the multifidelity training data, could impart meaningful information about the conformational phase-space and its relation to the excitation energies. The limited improvement of the scaling factor is also observed in the same figure.
Comparison of learning curves for fixed scaling factors \(\gamma\), \(\theta_{f-1}^{f}\), and \(\theta_{f}^{F}\) with \(f_{b}\): STO3G. The x-axis reports the number of training samples used at the highest fidelity, that is, TZVP. Both MFML and o-MFML models are compared. Increasing values of \(\gamma\) result in a constant lowered offset of the learning curves. The cost informed scaling factors show a higher value of MAE.
Furthermore, the value of \(\theta_{f-1}^{f}\) for the SVP and 321G fidelities was 1 which did not provide any additional information to the MFML model as was pointed out in the discussion for This in turn affects the overall model that is built with the STO3G baseline fidelity. A similar argument can be made for why the MFML model with \(\theta_{f}^{F}\) has limitations. Regardless, the MFML model built with \(\theta_{f}^{F}\) does in fact achieve model RMAE that are comparable to the MFML model built with \(\gamma=2\).
The results for fixed scaling factors, \(\gamma\), indicate that a higher \(\gamma\) results in a lower model error, or, smaller number of training samples at the costly target fidelity are needed. For the time-informed scaling factors, it was seen that these do not perform as well as was anticipated. However, one must be cautious about the results from and Figures 3 and 4 before considering them to be improvements over the conventional MFML method with \(\gamma=2\). Since one uses much more training samples at the cheaper fidelities as one increases the value of \(\gamma\), the cost of generating training data needs to be assessed to better understand the cost-accuracy trade-off in these multifidelity models. In interest of such an analysis, the time-cost of generating training data versus model RMAE are discussed in the next section. Although the MFML and o-MFML models show similar RMAEs, only o-MFML models are discussed hereon. This is due to the observation of ref. that the o-MFML method provides a superior model even in cases of poor data distribution of the cheaper fidelities. That is, in the case of a poor data distribution, the o-MFML provides the better model in comparison to MFML. Since all o-MFML models use the same validation set, the cost of generating the validation set is not included in the time-analysis plots.
### Time Benefit Analysis
As presented in refs., a good assessment of multifidelity methods is the study of model error versus the time to generate the training samples for the model. In interest of such a study, the RMAE versus training data generation time are studied for the just discussed test cases. The time to generate data for a multifidelity model is the sum over the times for generation of all the training samples used at all fidelities that form the multifidelity model. That is, \(T_{\rm train-data}^{\rm MFML}:=\sum_{f_{b}\leq f\leq F}N_{\rm train}^{f}\cdot T_ {QC}^{f}\) where \(N_{\rm train}^{f}\) is the number of training samples used at some fidelity \(f\), and \(T_{QC}^{f}\) is the corresponding single-point QC-compute time for that fidelity. The QC-compute times recorded in the QeMFi dataset are those for a single-core computation and are provided for each fidelity for each molecule type.
Time to generate training data versus RMAE of the corresponding o-MFML model for the diverse scaling factors studied. The different scaling factors used are denoted as sub-titles. The RMAE is unitless while the time-cost is in minutes. The single fidelity KRR case is also depicted for reference. As one increases the scaling factors across the fidelities, one observes that the learning curves of the MFML models shifts further due to the larger amount of training samples used. The two cases of \(\theta_{f-1}^{f}\) and \(\theta_{f}^{F}\) are explained in Scaling Factors. The bottom-right corner plot compares the o-MFML curves for the 321G baseline for the two time-informed scaling factors and the case of \(\gamma=2\).
The RMAE versus \(T_{\rm train-data}^{\rm MFML}\) plots for the various scaling factors are shown in for o-MFML. Only o-MFML is shown since it has a lower error compared to MFML for all the cases (see Figures 2,3, and 4). The RMAE and the time axes are both presented in log-scaled values. The axes of the plots are scaled identically for easy comparison among the different scaling factors. The bottom-right corner plot compares the time-cost based scaling factors to the case of \(\gamma=2\) for the MFML model built with \(f_{b}\): 321G and not for the cheapest STO3G baseline for reasons discussed below.
For the different cases of scaling factors shown in Figure 6, it can be seen that the addition of cheaper baselines helps achieve a specific model accuracy with less time cost to generate the training data. In general, fixing a specific MAE, one can see that the curves of the cheaper baseline achieve this error earlier with respect to the time axis. Alternatively, if one were to set a time budget and draw a vertical line at that value (as on the x-axis), then the cheaper baseline models result in lower RMAEs than the single fidelity KRR model. The case of STO3G baseline is an exception. For all scaling factors, the addition of the STO3G baseline does not provide significant improvement of the model. In fact, it increases the training data generation cost. The STO3G energies do not provide major improvement to the o-MFML model over the 321G baseline. This could be due to poor data distribution that has previously been noted for the STO3G fidelity for excitation energies of molecules. The time-cost versus RMAE plots make this evident. Although the analysis of conventional learning curves from Learning Curves indicated that the STO3G baseline fidelity improved the MFML model, these time-cost plots indicate that this comes at a cost which supersedes the RMAE improvement that is observed. However, consider the case of the special scaling factors \(\theta_{f}^{F}\), which are decided by the ratio of the QC-compute times of a fidelity \(f\) to the QC-compute time of the target fidelity \(F\). For some portions of learning curve for \(f_{b}=\) STO3G, o-MFML does provide lower errors as is expected from such multifidelity models. This could indicate that the use of o-MFML could improve the model accuracy even for the cases of poor data distribution as seen in the STO3G fidelity. For the o-MFML modelsthat are built with the 32G baseline fidelity, the time benefit of the multifidelity approach becomes all the more perceptible across the various scaling factors. For instance, in the case of \(\gamma=3\), the o-MFML model results in an RMAE of \(0.1\) with a time cost of \(\sim 200\) minutes. The KRR model achieves a similar error with a time-cost of \(\sim 1,000\) minutes. This indicates a time benefit of about \(5\) times with this baseline fidelity for \(\gamma=3\). Similarly, for \(\gamma=5\), the KRR model achieves an RMAE of \(0.07\) with a time cost of \(\sim 2,000\) minutes while the o-MFML model achieves a similar error for a time cost of \(\sim 300\) minutes resulting in a time benefit of about \(6\) times. Similar observations can be made for the other values of \(\gamma\). The time benefit is less pronounced for the cases of \(\theta_{f-1}^{f}\) and \(\theta_{f}^{F}\) but is still present for \(f_{b}:321\)G.
While each scaling factor does improve the time-cost needed to achieve a certain RMAE vis-a-vis the single fidelity KRR, it is also important to see which scaling factor performs better with respect to the others for a given baseline fidelity. The bottom-right plot of compares the time-cost versus RMAE curves of MFML models for \(\gamma=2\), \(\theta_{f-1}^{f}\), and \(\theta_{f}^{F}\) for the baseline fidelity of 321G. The STO3G baseline is not considered due to its poor distribution. These specific scaling factors are chosen to better understand the standing of the time-informed scaling factors with respect to the fixed scaling factors. This comparison in for the 321G fidelity shows that the fixed scaling factor of \(\gamma=2\) performs better than both the time-cost informed scaling factors (see Scaling Factors). The MFML model built with \(\theta_{f}^{F}\) does perform only as well as that built with \(\gamma=2\), which is the default set-up for MFML. This could indicate that just the QC-compute time-cost information might not suffice to select the training samples at each fidelity. It could be that the model accuracy and multifidelity training structure relation is more complex than just accounting for the QC-compute cost. To better understand how each fidelity and the number of training samples at each fidelity contribute to the overall model error, the next section studies a new error metric, error contours (see Error Contours of MFML for details). This is intended to give a better view into the inner mechanisms of the multifidelity data structure in building a MFML model.
### Multifidelity Error Contours
The time versus RMAE results for different scaling factors hint that one might not necessarily need many training samples at the target fidelity. This would imply that one could build a cheap multifidelity model with a large number of training samples at the cheaper fidelities and then 'raise' it to the target fidelity with an exceptionally small number of training samples at the target fidelity. This can be further studied with the error contours of multifidelity. These contours involve studying the model prediction error by varying the training sizes along fidelity \(f\) and \(f-1\) for all \(f\leq F\). As was discussed in Scaling Factors, this is performed for consecutive fidelity pairs TZVP-SVP, SVP-631G, 631G-321G, and 321G-STO3G.
Consider the top-left plot corresponding to the TZVP-SVP fidelity pair. The y-axis denotes the number of training samples used at TZVP, while the x-axis depicts the number of training samples used at the SVP fidelity. The colors of the plot itself correspond to the MAE. In the usual o-MFML approach, the number of training samples used at SVP with respect to the number of training samples used at TZVP would be scaled by the factor \(\gamma\) (in this case by 2). However, for this set-up there is no trivial scaling of training data that is carried out. Instead, the multifidelity model is built with a specific selection of training samples. For example, take the case for \(N_{\rm train}^{\rm TZVP}=2\) and \(N_{\rm train}^{\rm SVP}=2^{10}\) which is marked by the top-corner red circle. The RMAE reported here, 0.025, is for a multifidelity model that is built with the following multifidelity training structure (with increasing fidelity): \(\{2^{3}\cdot 2^{10},2^{2}\cdot 2^{10},2\cdot 2^{10},2^{10},2\}\). In other words, the scaling factor is only applied for the fidelities that are not studied as part of the error contour. In contrast, in the usual o-MFML the training data structure would be \(\{2^{4}\cdot 2,2^{3}\cdot 2,2^{2}\cdot 2,2\cdot 2,2\}\), this is the block that corresponds to \((N_{\rm train}^{\rm SVP}=2^{2},N_{\rm train}^{\rm TZVP}=2)\) on the plot. The accompanying color-bar depicts that this regular o-MFML model results in a higher RMAE than 0.025. In general, the diagonal of the contour plot depicts the regular o-MFML model which is identified in the learning curves of for \(\gamma=2\). The RMAE for \((N_{\rm train}^{\rm SVP}=2^{6},N_{\rm train}^{\rm TZVP}=2^{5})\)Excitation energy prediction errors with o-MFML for different training samples at different fidelities. The details of the method are explained in Scaling Factors for each case. In each plot, the vertical axis depicts the number of training samples used at the costlier fidelity, \(f\), while horizontal axis reports the training samples used at the cheaper fidelity \(f-1\). The resulting error for the o-MFML model with the specific choice of training samples used at fidelity \(f\) and \(f-1\) are depicted as the error contours. Here, the RMAE are depicted as contour plots for different training samples spanned across two fidelities. Two specific RMAEs are enumerated for all 4 cases: first, that for the smallest training set size at the higher fidelity, \(f\), and the largest training set size at \(f-1\); second, for the case where the training sample at \(f\) and \(f-1\) have the scaling factor of 2.
This is a remarkable observation in that simply using two training samples at TZVP while increasing the training size at the lower fidelities results in a model that is more than twice as accurate. Furthermore, the RMAE for the block \((N_{\rm train}^{\rm SVP}=2^{10},N_{\rm train}^{\rm TZVP}=2)\) is similar to the model block \((N_{\rm train}^{\rm SVP}=2^{10},N_{\rm train}^{\rm TZVP}=2^{9})\). In general, it is seen that a lower number of TZVP training samples with a larger training set size at the cheaper fidelities results in more accurate multifidelity model.
Similar observations and inferences can be made for the error contour for the SVP-631G fidelity pair as seen on the top-right plot of In this set-up, consider the top right corner which is marked with a circle. This is identified as \((N_{\rm train}^{631G}=2^{11},N_{\rm train}^{\rm SVP}=2^{2})\) and has the following multifidelity data structure (with increasing fidelity): \(\{2^{13},2^{12},2^{11},2^{2},2\}\). As in the previous case, the training data scaling is only applied to the fidelities that are not studied as part of the error contour. This mode reports 0.030 as the RMAE. Once again, the diagonal of the contour plot corresponds to the regular o-MFML model. Consider then, the block identified by \((N_{\rm train}^{631G}=2^{6},N_{\rm train}^{\rm SVP}=2^{7})\) which has a training data structure (in increasing order of fidelity): \(\{2^{9},2^{8},2^{7},2^{6},2^{5}\}\). This regular o-MFML model reports an RMAE 0.064, over twice as much as for the previous one. The overall contour plot reveals that the use of very few training samples at SVP paired with a larger number of training samples at the lower fidelities results in RMAEs that are comparable to the cases where one would use a lot more training samples at the SVP fidelity. In particular, this form of _flattening_ out the multifidelity training structure by using few training samples at the top fidelities and increasing the training samples at the cheaper fidelities, outperforms the regular o-MFML model (which are the diagonal blocks of the error contour).
Similar observations are made for the 631G-321G and 321G-STO3G pairs of fidelities which are seen in the bottom row of It is interesting to note, however, that the 321G-STO3G error contours do not follow the same trend as the others. Using very little 321G training samples and increasing the training samples at the STO3G fidelity does not result in lower RMAE as seen from the top-corner red marker error being 0.088 while the center marker reporting RMAE of 0.066. This is once again explained by the poor data distribution that has previously been reported for the STO3G fidelity.
The error contours for the multifidelity model hint at an interesting mechanism in the MFML approach. Based on the behavior of the model error as discussed above, it appears that one does not necessarily need to use many training samples at the higher fidelities, in particular at the target fidelity. This is indeed something that has been previously been hinted at in ref. using the optimization procedure for h-ML albeit with a larger number of training samples at the target fidelity. However, a thorough investigation of the multifidelity structure such as that performed in Fig.7 reveals that not only can a multifidelity model be built with low number of training samples at the costlier fidelities, but that this number is far smaller than what would be anticipated in the general MFML and similar methods. The error contours indicate that there is still a great deal of information available at the cheaper fidelities which only need to be 'raised up' to the target fidelity with a surprisingly small number of training samples. With such an understanding of the multifidelity training structure, one can begin to think of ways to select training samples at the different fidelities that need not necessarily follow the concept of a scaling factor between the fidelities. Furthermore, the results of varying \(\gamma\) from hint at a possible approach which is detailed in the following section.
### \(\Gamma\)-curves
The contour plots of provide an interesting observation about MFML. One can potentially build cheaper multifidelity models by limiting the training samples used at the expensive fidelities and then proceeding to add cheap fidelity data to the multifidelity model.
Consider first the left-hand side plot of which shows the RMAE curves of o-MFML with \(f_{b}:321\)G for the different \(\gamma\) studied in this work for comparison. An inset plot is provided which zooms into the region between 1,000-3,000 minutes to show the different curves clearly. In addition, a new curve as introduced in The \(\Gamma\)-Curve, the \(\Gamma\)-curve is depicted in the plot. The \(\Gamma\)-curve is essentially the case where the number of training samples at TZVP are constrained to \(2\) but the remaining multifidelity data structure is allowed to grow as per the scaling factor \(\gamma\). That is, the \(\Gamma\)-curve is built with the first point of the curves for the different \(\gamma\) values. In some sense, this translates into it being a learning curve not as a function of \(N_{\rm train}^{\rm TZVP}\) but rather of \(\gamma\). From Figure 8, it becomes evident that even for as little as \(2\) training samples at the highest fidelity, if one adds cheaper data to the multifidelity model - which corresponds to increasing the value of \(\gamma\) without increasing \(N_{\rm train}^{\rm TZVP}\) - the error of the o-MFML model decreases. For the same time-cost of a conventional o-MFML model built with \(\gamma=2\), if one were to chose the models along the \(\Gamma\)-curve, a lower RMAE can be achieved. The \(\Gamma\)-curve in shows data points for up to \(\gamma=10\) where the multifidelity training data structure (for increasing fidelity) is: \(\{20000,2000,200,20,2\}\). In the inset of the plot, one observes that the \(\Gamma\)-curve results in errors that are lower than the o-MFML learning curves for fixed \(\gamma\). However, the \(\Gamma\)-curve converges to the o-MFML model built with \(\gamma=6\). One potential reason for this could be the saturation of the multifidelity model built along the \(\Gamma\)-curve. Due to a very large number of training samples at the cheaper fidelities (for instance, \(2\cdot 10^{4}\) at 321G for the last point on the \(\Gamma\)-curve), the model is no longer able to clearly learn the correction between SVP and TZVP.
The right-hand plot of further investigates this saturation by comparing the \(\Gamma(N_{\rm train}^{\rm TZVP})\)-curves for \(N_{\rm train}^{\rm TZVP}\in\{2,8,32\}\). The y-axis reports the RMAE while the x-axis denotes the time-cost in minutes to generate the training data used in the multifidelity models. An inset is provided for the interval between \(10^{3}-10^{4}\) minutes for a better view of the \(\Gamma(\cdot)\)curves. The o-MFML learning curve for \(\gamma=2\) is provided for reference.
Since the model errors throughout this work were reported in unitless RMAE, in order to understand how this translates to actual energy prediction, predictions are made for the holdout test set. For this purpose, the multifidelity model corresponding to \(\Gamma\) is used with \(\gamma=10\). Using this model, the prediction of the first vertical excitation energies is made on the holdout test set. The absolute error values are then computed as \(|y_{\text{ref}}-y_{\text{pred}}|\). The resulting values are reported in Table 2. This includes the mean of the absolute error, minimum absolute error, maximum absolute error, and the standard deviation of the absolute error. For all molecules, it can be seen that the model error is nearly identical indicating that the final multifidelity model built with the \(\Gamma\)-curve is not affected by the difference of the molecule. This is of course due to the fact that the training data consists of these molecules. In all cases, the model reports a mean absolute error close to 1 kcal/mol.
### Transferability Assessment
Transferability in ML refers to the concept of ML models being trained on specific type of dataset and having it predict the QC property for an out of sample dataset.
(a) Time to generate training data and corresponding o-MFML model error as RMAE for constant scaling factors, \(\gamma\) used in this study. An inset between 1,500-3,000 minutes is provided for the comparison of the curves for all \(\gamma\) studied in this work to readily compare in regions that are too crowded to be observed in the main plot. (b) RMAE versus time-cost for different \(\Gamma(N_{\text{train}}^{\text{TZVP}})\)-curves. Increasing the number of training samples at TZVP improves the model accuracies along the \(\Gamma(\cdot)\)-curves with a saturation observed towards the end of each curve.
The transferability of ML models in QC has long been a challenging issue. In general, since ML is a statistical method, transferability is restricted by the type of data the model is trained on. That is, if a model is trained only on, say, benzene configurations, it is not expected to perform well in predicting for methanol or acrolein. Not only so, the type of molecular representation that is used in the model makes a major difference to the overall transferability of the model. However, in this subsection, the robustness of the \(\Gamma\)-curve method is investigated for transferability in order to complete the discussion on the development of this method.
The QUESTDB database is a collection of several small molecules for which high accuracy excitation energies are available. The energies for these molecules are computed with mostly CC levels of theory. In order to properly assess the transferability, which is challenging as is, the excitation energies of 90 molecules from QUESTDB were computed with the DFT method using CAMB3LYP functional with the def2-TZVP basis set in the exact same manner as was done for the QeMFi dataset. Therefore, the predictions made on the molecules from QUESTDB are compared to the correct reference fidelity.
\begin{table}
\begin{tabular}{|c|c|c|c|c|} \hline
## Molecule** & **mean** & **min** & **max** & **Std.
\\ \hline \hline
## urea
& 1.0178 & 0.9940 & 1.0264 & 0.0035 \\
## acrolein
& 1.0145 & 0.9928 & 1.0220 & 0.0034 \\
## alanine
& 1.0145 & 0.9928 & 1.0222 & 0.0035 \\
## SMA
& 1.0147 & 0.9928 & 1.0224 & 0.0035 \\
## 2-nitrophenol
& 1.0147 & 0.9928 & 1.0224 & 0.0035 \\
## urocanic
& 1.0146 & 0.9928 & 1.0224 & 0.0035 \\
## DMABN
& 1.0145 & 0.9928 & 1.0220 & 0.0034 \\
## thymine
& 1.0145 & 0.9928 & 1.0224 & 0.0035 \\
## o-HBDI
& 1.0146 & 0.9928 & 1.0224 & 0.0035 \\ \hline \end{tabular}
\end{table}
Table 2: Absolute error values in kcal/mol between prediction and reference of excitation energies for the nine different molecules using \(\Gamma\) for \(\gamma=10\). The mean, range, and standard deviation of the absolute differences are presented.
as relative error (RE) which are computed as
\[\text{RE}:=\left|\frac{P_{\text{ML}}\left(\mathbf{X}_{q}\right)-y_{q}^{\text{ref}}}{y _{q}^{\text{ref}}}\right| \tag{10}\]
The geometries from the QUESTDB database have an additional challenge in testing for transferability. This is the issue of index invariance in generating the unsorted CM molecular descriptor. Since the geometries of the molecules from QeMFi are arranged in such a way to ensure index permutation invariance, one can use unsorted CM to train and evaluate ML models. However, this is not the case with the QUESTDB database. To overcome this fundamental issue, the following tests are performed using row-norm sorted CM representations wherein the regular CM representation is built and then the rows are sorted based on \(L_{2}\) norm. That is, all models are trained and tested using sorted CM representations unlike the unsorted CM used in the preceding sections.
* Generate row-norm sorted CM for 90 molecules from QUESTDB
* Train single fidelity KRR model on QeMFi with \(2^{10}\) training samples, Matern kernel from Eq.
* Predict energies for the 90 molecules from QUESTDB using KRR and \(\Gamma\)-MFML models from step 3 and step 4
* Compare prediction to reference computed excitation energies from step 1A scatter plot of reference excitation energies versus ML predicted excitation energies is shown in The scatter plot is shown for \(\Gamma\) and \(\Gamma\) MFML models. The markers labeled in the legend as \(\Gamma\)-MFML correspond to the prediction versus reference of all 90 molecules chosen from QUESTDB. The best 10, that is those with the lowest RE, are highlighted in green along with the worst 10 in red. For each of the \(\Gamma\)-curve models, for the best and worst predictions, the corresponding predictions of the single fidelity KRR model are also presented. This is done for two reasons. Firstly to indicate that it is a challenge even for single fidelity ML models to handle transferability tests. Secondly, it is interesting to assess how well the MFML models performed with respect to the single fidelity KRR models.
Consider the left pane of for \(\Gamma\) MFML model. The overall prediction of the \(\Gamma\) MFML model is poor for the QUESTDB database. There is a wide scatter of the points across the identity line (dashed black line).
Scatter plot of reference and ML predicted excitation energies for transferability tests of \(\Gamma\)-curve on molecules from the QUESTDB database. The best 10 and worst 10 predictions are highlighted along with their predictions made using the single fidelity KRR model.
Consider the predictions made by the single fidelity KRR model for these very geometries which corresponds to the translucent green square markers. These are much further away from the identity line. In other words, the \(\Gamma\)-curve method is somewhat better than the single fidelity KRR method. Even for the 10 worst predictions of the \(\Gamma\) model, the corresponding KRR predictions given as translucent red square markers, lie further away from the identity mapping.
Similar observations can be made for the case of \(\Gamma\) on the right-hand side pane of The best 10 predictions of the \(\Gamma\)-curve model are close to the identity line while the predictions of the KRR model are more loosely scattered. However, in this case, the poorest predicted molecules for single fidelity KRR do have some points close to the identity mapping line. Certainly, the overall scatter plot for \(\Gamma\) MFML looks closer to the identity mapping line as compared to the \(\Gamma\) MFML model. Based on these results, one can hold on to what was stated before we started this assessment, the transferability of ML models remains a challenge, even for multifidelity approaches. Certainly the \(\Gamma\) MFML model is more efficient, and more accurate as has been sufficiently established in preceding sections. The key takeaway from this discussion on transferability is not the effectiveness of one model over the other but rather the fact that transferability is a difficult task for both single fidelity and multifidelity ML models.
To complete this discussion on transferability one can take a closer look into the errors of the \(\Gamma\) MFML model. presents the molecules which comprise the 10 best predictions and 10 worst predictions along with the RE. It is unsurprising to notice that the DMABN and acrolein geometries from the QUESTDB database have well predicted energies since the QeMFi dataset, and by extension the training dataset, contains geometries of these molecules. As argued previously, this is due to the fact that the ML model trained on specific geometries does well in predicting the energies of similar geometries from 'unseen' datasets. In addition, several other aromatic compounds such as naphthalene and phthalazine show low REs. Consider the molecules that are not well predicted, the bottom two rows of Figure10. One observes radicals and molecules with elements such as sulfur and silicon which are not present in the QeMFi dataset. Once again, it becomes clear how important the initial training dataset is to the ability of a ML model in predicting QC properties.
Concluding this digression on testing the \(\Gamma\)-curve for transferability, some final remarks are made here. As has become evident, the task of transferability of a ML model is challenging and demands investigation in its own right. The analysis of this short subsection is not indicative of an issue in the \(\Gamma\)-curve approach itself but rather ties into the larger picture of training general purpose ML models for QC.
Best 10 (green) and worst 10 (red) predictions of the \(\Gamma\)-MFML model over the QUESTDB dataset. The numbers under the name of the molecules indicate the relative error.
## 4 Conclusions and Outlook
This work discusses the concept of scaling for the number of training data across different fidelities for MFML and o-MFML in the prediction of excitation energies of the QeMFi dataset. Constant scaling factors, \(\gamma\), were studied along with QC-calculation time-cost informed scaling factors, \(\theta_{f-1}^{f}\) and \(\theta_{f}^{F}\). It is seen in the results that the use of constant scaling factors,\(\gamma\), is effective with a higher value of \(\gamma\) resulting in lower model errors for reasonable time-cost of generating training data. A new error metric, the error contour of MFML, was introduced and results discussed for the prediction of first vertical excitation energies of the QeMFi dataset. Such an analysis revealed that the data requirements for MFML-like methods is not as trivial as has been previously employed. In fact, one can achieve similar model accuracies with much less costly training samples if one increases the number of training samples at the lower end of the multifidelity data structure. The error contours depicted that one could potentially use as little as 2 training sample at the target fidelities and achieve exceptional model accuracy if the subsequent fidelities used a larger number of training samples in comparison to MFML models built with some \(\gamma\). This was systematically studied with the newly introduced \(\Gamma\)-curve for a fixed number of training samples at the target fidelity and an increasing value of \(\gamma\). The models built in this fashion were shown to be time-cost efficient over conventional MFML approach. A brief digression was made to thoroughly analyze transferability of the \(\Gamma\)-MFML model on completely unseen data from the QUESTDB database concluding that transferability is challenging for both single fidelity and multifidelity ML models.
These results provide a window into the inner mechanisms of MFML-like methods allowing for a better understanding of how they can be employed for accurate predictions of excitation energies with low cost of training data generation. The development of the \(\Gamma\)-curveapproach in this work in its current form is only benchmarked for a specific DFT functional and this could be a potential limitation of this work. A possible extension of this work could be the study of the \(\Gamma\)-curve approach for a wider range of DFT functionals. At the moment this is inhibited by the lack of compute cost times in most large-scale multifidelity datasets. Another interesting area of research can be the use of approaches developed in this work to assess the efficiency of multifidelity approaches for fidelity structures built on CC level of theory and would be of particular interest since CC is considered the gold standard in QC. Future methodological development over this could include algorithms that optimally select the number of training samples to be used at each fidelity. Further work could involve extending this to other time-based scaling factors which also take into account the error reduction that each additional training sample contributes to the overall multifidelity model. This would further improve the scope of application for the multifidelity methods discussed herein.
|
10.48550/arXiv.2410.11392
|
Investigating Data Hierarchies in Multifidelity Machine Learning for Excitation Energies
|
Vivin Vinod, Peter Zaspel
| 378
|
10.48550_arXiv.1011.4564
|
## I. Introduction.
Density-Functional Theory (DFT) is the leading theoretical framework for studying the electronic properties of matter. Most practical applications rely on the Kohn-Sham (KS) formulation, which was originally formulated assuming that the ground state density of an interacting \(N\)-electron system in an external potential \(v(\vec{r})\) can be represented by the ground state density of a non-degenerate non-interacting \(N\)-electron system in a reference potential \(v_{\mathit{eff}}(\vec{r})\). This assumption is known as pure-state \(v\)-representability of the density in the non-interacting system, and is denoted PSVR below.
For PSVR systems, the \(N\)-electron wave function \(\Psi(\vec{r}_{1},...,\vec{r}_{N})\) of the reference system can be written as a single Slater determinant of the one-electron wave functions \(\psi_{i}(\vec{r})\), which are solutions of the non-interacting Schrodinger equation \(\Big{(}(-\hbar^{2}\,/\,2m_{e}\,)\nabla^{2}+v_{\mathit{eff}}(\vec{r})\Big{)} \psi_{i}=\varepsilon_{i}\psi_{i}\). The index \(i\) refers to the relevant set of quantum numbers, e.g. \(i=\{n,l,m,\sigma\}\) for spherical atoms. The density is given by \(n(\vec{r})=\sum\limits_{i}^{N}\left|\psi_{i}(\vec{r})\right|^{2}\), where the sum is over the \(N\) lowest-energy eigenstates of the system. The effective potential \(v_{\mathit{eff}}[n]=v(\vec{r})+v_{\mathit{H}}[n]+v_{\mathit{xc}}[n]\) is a functional of the density, where \(v_{\mathit{H}}[n]=e^{2}\int n(\vec{r}^{\prime})/\left|\vec{r}-\vec{r}^{\prime} \right|d^{3}r^{\prime}\) and \(v_{\mathit{xc}}[n]=\delta E_{\mathit{xc}}\)/\(\delta n\) are the Hartree and the exchange-correlation potentials, respectively. The density and the total energy \(E[n]=T_{s}[n]+\int n(\vec{r})v(\vec{r})\)\(d^{3}r+E_{\mathit{H}}[n]+E_{\mathit{xc}}[n]\) of the interacting system are found by solving self-consistently the equations above. Here \(T_{s}[n]=-\hbar^{2}\)/\((2m_{e})\sum\limits_{i}^{N}\left\langle\psi_{i}\left|\nabla^{2}\left| \psi_{i}\right.\right\rangle\right.\) is the kinetic energy and \(E_{\mathit{H}}[n]\) and \(E_{\mathit{xc}}[n]\) are the Hartree and exchange-correlation energy terms. It was assumed, and later shown to be true, that for ground state densities the exchange-correlation energy derivative \(\delta E_{\mathit{xc}}\)/\(\delta n\) exists. Note, however, that the self-consistent solution of the Kohn-Sham equations requires the existence of the exchange-correlation energy derivative \(\delta E_{\mathit{xc}}\)/\(\delta n\) for any density considered in the iterative process, not only for the ultimate density.
Two questions arise immediately from this formulation, the first of which has been considered extensively and the second far less so. (i) Are all ground state densities PSVR? (ii) For which densities does \(v_{\mathit{xc}}[n]=\delta E_{\mathit{xc}}\)/\(\delta n\) exist and how can the self-consistent solution process ensure that only such densities are considered? From early on it was known that there are physical systems (e.g., Fe andCo atoms in the spherical approximation), for which the density was not represented by a pure state. For these atoms, no self-consistent result was found for which integrally occupied orbitals were also the lowest in energy. The absence of a self-consistent solution expressed itself in the so-called 'Fermi statistics' problem: the energy levels of the effective potential cross repeatedly during the iterations, and being occupied according to the ground-state rule, yield radically different orbital densities, so that the self-consistent cycle cannot converge.
A possible stopgap measure is to suspend the requirement of ground state occupation and to allow the system to relax while freezing the energy level occupation, resulting in convergence to a solution in which the representing non-interacting system is in an excited state. In this spirit, Janak proposed an extension of the Kohn-Sham scheme to include occupation numbers \(\{g_{i}\}\) (termed the electronic configuration) as parameters in the definitions of the density \(n(\vec{r})=\sum_{i}g_{i}\left|\psi_{i}(\vec{r})\right|^{2}\) and the kinetic energy \(\widetilde{T}_{s}=\sum_{i}g_{i}\left\langle\psi_{i}\right|-\frac{\hbar^{2}}{2m _{e}}\nabla^{2}\left|\psi_{i}\right\rangle\). To determine the occupation numbers, Janak minimized the total energy \(\widetilde{E}[\{g_{i}\}]=\widetilde{T}_{s}+\int n(\vec{r})v(\vec{r})\ d^{3}r+ E_{H}[n]+E_{xc}[n]\) with respect to \(\{g_{i}\}\). It was shown that the minimum of the total energy \(\widetilde{E}\) is achieved only for a "proper" electronic configuration, i.e. one in which the energy levels are occupied without gaps. This includes not only the PSVR systems treated by Kohn and Sham, but also cases where the representing non-interacting \(N\)-electron system possesses a ground-state degeneracy, with the highest electron orbitals degenerate and partially occupied. For such a configuration the density is ensemble \(v\)-representable in the non-interacting reference system, and this is denoted by NI-EVR below. However, Janak questioned the validity of his results in the non-PSVR cases, especially in situations with unequal occupations of the degenerate orbitals.
Addressing the problem of \(v\)-representability, a recent study by Ullrich and Kohn on topologies of the \(v\) and \(n\) spaces examined for lattice systems showed that potentials which generate NI-EVR densities are not isolated points in \(v\)-space, because not every perturbation potential imposed on the system will lift the degeneracy of its energy levels. Even more significant is that NI-EVR densities occupy manifolds of finite volume in \(n\)-space. Thus, systems with NI-EVR densities are not rare, but form a significant physical class that should be treated with proper care.
Few examples of EVR densities have been studied in detail. Some of these studies are analytical, and thus relate to the exact (but unknown) exchange-correlation functional, while others involve direct computations using available approximate functionals. Recently, two analytical examples were introduced: a finite lattice problem, which produced either a PSVR or an NI-EVR solution, depending on the value of the potential on the different lattice points, and the Be ion series.
Upon application of the Janak formalism, NI-EVR densities have been identified in several physical systems, such as Fe and Co atoms. The occupations of the \(3d^{\downarrow}\) and \(4s^{\downarrow}\) degenerate levels are found to be fractional and unequal, contrary to what one might expect from elementary statistical mechanics considerations. This surprising result made the authors question their findings and speculate that they might be a consequence of the spherical symmetry approximation or the local-density approximation employed.
Other authors have rejected fractional occupation altogether. Averill and Painter devised a numerical scheme in which the minimization over the occupations in Janak's functional is performed as part of the iterative process, but took the view that the physical representation of the system would necessarily involve an electronic configuration with integral occupation numbers, even if it is "improper". Alternatively, in a well-known comprehensive study in atomic physics performed by Kotochigova et al., the experimental electronic configurations were used throughout the calculations. Although for some atoms the occupations thus obtained are "improper", the validity of such a choice of electronic configuration for different atoms/ions was not discussed.
The Kohn-Sham scheme requires differentiability of both the kinetic and the exchange-correlation functionals. After the generalization of density-functional theory in the constrained-search formalism to all \(N\)-representable densities, it was found that differentiability exists only for densities which are both interacting ensemble \(v\)-representable (I-EVR) and non-interacting ensemble \(v\)-representable (NI-EVR) (see Sec. II for the definitions). However, during the iterative convergence of the Janak-Kohn-Sham equations, the densities generated are not restricted to be EVR. This is unlike the situation in the self-consistent solution of the Kohn-Sham equations, where if the densities are generated from ground states of the trial effective potentials, then they are NI-EVR by construction. Therefore, the Janak scheme is expected to be invalid for the exact form of \(E_{xc}[n]\) and is formally applicable only for approximations, which are differentiable at all densities, such as the local-density approximation (LDA).
Therefore, in consequence of the extension of the DFT-KS formalism to include NI-EVR densities, it is desirable to have an algorithm which will guarantee that only EVR densities are considered in the self-consistent solution of the Kohn-Sham equations. To the best of our knowledge, such an algorithm hasnot yet been suggested. We note that it was proved that for lattice systems all the densities are EVR. Therefore, in computational implementations on a grid all densities are implicitly EVR. However, this result has not been extended to the continuum, and this indicates the benefit of an algorithm which is restricted to explicitly EVR densities.
Surprisingly, in all of these developments the finite-temperature version of DFT (T-DFT) is scarcely mentioned. The finite-temperature theory can be represented as an extension of the standard theory of thermal ensembles to non-uniform systems. It is well-known that finite-temperature DFT does not suffer from the \(v\)-representability problem, and connects continuously to the ground-state theory in the limit \(T\to 0\). This has previously allowed the resolution or reinterpretation of several theoretical issues in DFT.
In the present work we present three contributions to the ongoing discussion. First, we show that densities generated by "improperly" occupying the orbitals in the reference system cannot be ground state densities of the given external potential, complementing the known result that the ground state can be described by properly occupying the reference system. Consequently, we present a new algorithm which allows the solution of Kohn-Sham systems with ground-state densities that are not PSVR, while considering only EVR densities in the process and overcoming the "Fermi statistics" convergence difficulties. We apply the algorithm to the well-known case of Fe and show that the results obtained are converged, stable numerically and the emergence of NI-EVR density is insensitive to the choice of the exchange-correlation functional approximation. We also demonstrate that the subspace of external potentials which yield NI-EVR densities is finite, by considering variations of the nuclear charge about \(Z=26\). In passing, we discuss the solution of the Be ion series in the LDA. Finally, our third contribution is an analysis of the EVR problem within T-DFT, where we expose the physical significance of the fractional occupations and resolve the question of why degenerate states can be unequally occupied.
## 2 Theory.
The extension of the Kohn-Sham formalism to EVR densities is briefly presented below for completeness. An interacting ensemble-\(v\)-representable (I-EVR) density is defined as an integrable non-negative function \(n(\widetilde{r})=\sum_{k}\lambda_{k}n^{(k)}(\widetilde{r})\), with \(\lambda_{k}\geq 0\), \(\sum_{k}\lambda_{k}=1\), where \(n^{(k)}(\vec{r})=N\int...\int\left|\Psi^{(k)}(\vec{r},\vec{r}_{2},...,\vec{r}_{N })\right|^{2}d\vec{r}_{2}...d\vec{r}_{N}\) and \(\Psi^{(k)}\) are orthonormal degenerate ground states of the Hamiltonian of an interacting many-electron system :
\[\hat{H}=\hat{T}+\hat{V}+\hat{W} \tag{1}\]
Here \(\hat{T}=-\hbar^{2}\)/(2\(m_{e}\))\(\sum_{i}\nabla_{i}^{2}\) is the kinetic energy operator, \(\hat{V}=\sum_{i}v(\vec{r}_{i})\) is the external potential operator and \(\hat{W}=(e^{2}\)/2)\(\sum_{i\neq j}1/\left|\vec{r}_{i}-\vec{r}_{j}\right|\) is the electron-electron repulsion operator. The pure-state case is a particular case with only one \(\lambda_{k}=1\), and the others are zero. A non-interacting ensemble-\(v\)-representable (NI-EVR) density is defined similarly, with a non-interacting Hamiltonian \(\hat{H}_{0}=\hat{T}+\hat{V}_{eff}\) (\(\hat{W}=0\)).
According to the constrained search formulation of Hohenberg-Kohn theorems, the ground state energy is the minimum value of the functional of the density \(E_{v}[n]=\int v(\vec{r})\,n(\vec{r})\,d^{3}r+F_{L}[n]\), for a given external potential \(v(\vec{r})\). \(F_{L}[n]\) is the Lieb functional (, p. 14 and references therein,):
\[F_{L}[n]=\mathop{\min}\limits_{\sum_{i}d_{i}=n}\left[\sum_{i}d_{i}\left\langle \Phi_{i}\left|\hat{T}+\hat{W}\left|\Phi_{i}\right.\right\rangle\right]\right. \tag{2}\]
Similarly to the pure-state case, the density is represented through a reference system of non-interacting electrons, defined by \(v_{eff}(\vec{r})\).
\[T_{L}[n]=\mathop{\min}\limits_{\sum_{i}d_{i}=n}\left[\sum_{i}d_{i}\left\langle \Phi_{i}\left|\hat{T}\right|\Phi_{i}\right.\right\rangle\right] \tag{3}\]
\(E_{H}[n]\) is the Hartree energy term and the exchange-correlation term is defined as \(E_{xc,L}[n]=F_{L}[n]-E_{H}[n]-T_{L}[n]\).
To obtain the density from the reference system, the effective potential is determined, up to a constant, via the differential Euler relations for the interacting and the non-interacting systems :
\[\delta T_{L}\,/\,\delta n+v(\vec{r})+v_{H}[n]+\delta E_{xc,L}\,/\,\delta n=const\, \tag{4}\]
\[\delta T_{L}\,/\,\delta n+v_{eff}\,(\vec{r})=const\.\]It follows that \(v_{\mathit{eff}}(\vec{r})=v(\vec{r})+v_{\mathit{H}}[n]+v_{\mathit{xc},L}[n]\), only if both \(\partial T_{L}/\partial n\) and \(\partial E_{\mathit{xc},L}/\partial n=v_{\mathit{xc},L}[n]\) exist. Otherwise, no relation between the interacting system and the reference system can be established.
The differentiability conditions for both \(F_{L}[n]\) and \(T_{L}[n]\) were formulated by Englisch and Englisch: the functional \(T_{L}[n]\) is differentiable at all NI-EVR densities, and only there; the functional \(E_{\mathit{xc},L}[n]\) is differentiable at densities, which are both NI-EVR and I-EVR, and only there. While the equivalence of these two sets of densities has the status of a conjecture, it can be proved that for any I-EVR density there exists a NI-EVR density arbitrarily close. Therefore, a Kohn-Sham scheme can always be set up. Regarding condition, note that for explicit approximations employed for \(E_{\mathit{xc},L}[n]\), such as the LDA, mathematical differentiability is always assured, even if for the given density the exact functional would not be differentiable.
As to condition, we note that the ground state wave-functions \(\Phi_{i}\) of a non-interacting system must obviously be constructed as Slater determinants of the \(N\) lowest eigenfunctions \(\psi_{i}(\vec{r})\) that emerge from the Schrodinger equation \(\left(-(\hbar^{2}/2m_{e})\nabla^{2}+v_{\mathit{eff}}(\vec{r})\right)\psi_{i} =\varepsilon_{i}\psi_{i}\).
\[T_{L}[n]=\min_{\sum_{i}\varepsilon_{i}|\psi_{i}|^{2}=n}\left[\sum_{i}g_{i}\left< \psi_{i}\left|-\frac{\hbar^{2}}{2m_{e}}\nabla^{2}\left|\psi_{i}\right.\right. \right\rangle\right] \tag{5}\]
and the density is given by
\[n(\vec{r})=\sum_{i}g_{i}\left|\psi_{i}(\vec{r})\right|^{2}\text{,} \tag{6}\]
where the occupation numbers \(g_{i}\) are linearly related to the coefficients \(d_{i}\) and are restricted to be :
\[g_{i}=\begin{cases}D_{i}&:\text{ }\varepsilon_{i}<\varepsilon_{F}\\ x_{i}&:\text{ }\varepsilon_{i}=\varepsilon_{F}\\ 0&:\text{ }\varepsilon_{i}>\varepsilon_{F}\end{cases}\text{.} \tag{7}\]
Here \(x_{i}\in[0,D_{i}]\), \(\varepsilon_{F}\) is the highest occupied energy level and \(D_{i}\) is the degeneracy of the \(i\)-th level, which can now be greater than 1, depending on the special symmetries the problem may have, e.g. spherical symmetry for atoms. The restriction is a straightforward consequence of the fact that only ground states of a reference system are considered. We denote occupations consistent with as "proper".
It can be seen from that fractional occupation can occur only if for more than one energy level \(\varepsilon_{i}=\varepsilon_{F}\). The physical interpretation of such a situation is discussed by us in Sec. IV.
Now we consider the use of "improper" densities, i.e. those which are excited states of their own effective potential. As frequently happens in Janak-type calculations, a trial density \(n(\vec{r})\) is represented in the form with an "improper" (excited) occupation of levels of a non-interacting system, defined by \(v_{\mathit{eff}}(\vec{r})\), which obeys \(v_{\mathit{eff}}(\vec{r})=v(\vec{r})+v_{\mathit{H}}[n]+v_{\mathit{xc}}[n]\). It is not evident that in such a case the derivative \(\delta T_{L}/\delta n\) must exist. However, if it exists, then there exists another non-interacting system defined by \(\tilde{v}_{\mathit{eff}}[n]=\tilde{v}(\vec{r})+v_{\mathit{H}}[n]+v_{\mathit{xc }}[n]\), for which \(n(\vec{r})\) is the ground-state density or a linear combination of ground-state densities. Thus, \(n(\vec{r})\) can be represented in terms of the \((\tilde{\ })\)-system as \(n=\sum_{i}\tilde{g}_{i}\big{|}\tilde{\psi}_{i}\big{|}^{2}\) with "proper" occupation. Then, however, \(n(\vec{r})\) cannot be the ground state of the original system of interest because Euler's relations cannot be satisfied for \(v_{\mathit{eff}}[n]\) if they were satisfied for \(\tilde{v}_{\mathit{eff}}[n]\). Therefore, we arrive at the conclusion that an "improper" configuration cannot be used in to construct the density, even if \(\delta T_{L}/\delta n\) exists, as could be misunderstood from Ref. 19. As a corollary, an algorithm that relies on the generalized form \(T_{L}[n]\) must be restricted to "proper" densities only.
Considering the validity of Janak's theorem questioned by Valiev and Fernando, we comment that Janak's theorem is valid, but only for Janak's \(\tilde{E}\), which is an extension of the energy functional to include systems with variable and fractional number of electrons, which are beyond the scope of the present work. In the domain of systems with integer occupation, the ground state densities are restricted to be "proper", as in and the \(\{g_{i}\}\) are determined uniquely by the self-consistency requirement. In this formalism, Janak's theorem, in the sense of a derivative of the energy functional with respect to occupation numbers, is inapplicable and this, we believe, should be the context of Ref. 19.
The above formalism can be extended to spin-polarized cases via representing a spin-polarized system as two distinct (but coupled) electronic subsystems, denoted by the index \(\sigma=\big{\{}\uparrow,\downarrow\big{\}}\).
\[T_{L}[n_{\sigma},0]=\min_{\sum_{i}\varepsilon_{i}\omega_{i}\big{|}\psi_{i} \big{|}=-n_{\sigma}}\sum_{i}g_{i\sigma}\big{\langle}\psi_{i\sigma}\big{|}- \frac{\hbar^{2}}{2m_{\varepsilon}}\nabla^{2}\big{|}\psi_{i\sigma}\big{\rangle}.\]Each of the subsystems is restricted to be "properly" occupied by, but there can be gaps in the occupation of the whole system if the first vacant level of one of the sub-systems lies below the last occupied level of the another one. We call such a system "proper in a broad sense". This situation arises because we are constrained from having partial occupation of spin states by the physical requirement that the total spin in the reference system is half-integer, i.e. that the spin occupations remain integral.
To summarize, the only densities which can be considered as candidate densities in the constrained search formalism are those with "proper" sets of occupation numbers. In the spin-polarized case, each of the spin subsystems must be properly occupied, even though the combined system may be "improperly" occupied due to the constraint of integral spin occupations. It remains to be shown in Sec. III how to restrict the self-consistent search for the density to NI-EVR densities only.
## III. Numerical algorithm and applications.
In this section we present an algorithm to solve the Kohn-Sham equations that is restricted to searching only over explicitly NI-EVR densities in accordance with the results of Sec. II. We have shown above that Janak's approach involves minimizing the energy over a set of densities which are not necessarily "proper", and therefore the existence of \(\,v_{xc}[n]\!=\!\delta E_{xc}\,/\,\delta n\,\) is not assured. We apply our algorithm to the spherical Fe atom which is a simple, well-established non-PSVR system, which is NI-EVR.
As an example of a NI-EVR system, we investigate some of its properties. In passing we show that the Be ion series which has been shown to be a NI-EVR system is instead found to be PSVR in the LDA.
In the case of the spherical Fe atom, the use of integral occupation numbers does not lead to a "proper" solution of the Kohn-Sham equations, a fact that was troubling from early days. Allowing fractional occupation numbers determined by the minimization of the total energy in Janak's approach yields a NI-EVR density for the spin \(\,S\!=\!2\,\). E.g., using the Vosko, Wilk and Nusair (VWN) parameterization of the LSDA for the exchange-correlation functional and a spherical approximation for the density yields \(\,E\!=\!-\!1261.229308\,\) hartree as the total energy for Fe, with an electronic configuration of \(\,[\mbox{Ar}]3d_{1,398}^{5}4s_{0,602}^{1}\,\). This result is in agreement with earlier publications on Fe.
Our alternative algorithm for treating NI-EVR densities is to start with a reasonable guess for \(v_{eff,\sigma}(\vec{r})\) and occupy the energy levels of the reference systems "properly" given the total spin value \(S\). Therefore, we are assured that the density generated by the reference system in each iteration is NI-EVR and by the EVR conjecture is also I-EVR. This allows us to calculate the effective potential from this density. If we employ this potential directly to calculate a new density, then as expected, the process starts alternating between two electronic configurations, e.g. for \(S=2\) between [Fe] = [Ar]\(3d_{1}^{5}4s_{1}^{1}\) and [Fe] = [Ar]\(3d_{2}^{5}4s_{0}^{1}\).
Instead, as the first new step in the algorithm, we reduce the linear mixing coefficient for the effective potential each time the electronic configuration changes. The reduction factor can vary around 0.5. By repeating this procedure, the eigenvalues eventually stop crossing and slowly approach each other. It is possible then to find an effective potential, with energy levels (\(\varepsilon_{3d\downarrow}\) and \(\varepsilon_{4s\downarrow}\) in this example) that coincide, to any required accuracy (\(10^{-6}\) hartree was required for Fe) with the original integer configuration. However, this solution is not necessarily a self-consistent one, and additional iterations in the domain of NI-EVR densities are required to obtain the final result.
The second new step in the algorithm continues the search in the subspace of potentials that generate NI-EVR densities while relaxing the constraint on integer occupation. Once an effective potential with two degenerate levels is found, it can produce a range of legitimate, "proper" NI-EVR densities by fractionally occupying the degenerate levels. Utilizing the fact that potentials generating NI-EVR densities are not, in general, points in \(v_{eff}\)-space, we choose the occupation that will generate a new potential which preserves the degeneracy of the energy levels for the next iteration.
\[\left\langle\psi_{a}^{(k)}\left|w\right|\psi_{a}^{(k)}\right\rangle-\left\langle \psi_{b}^{(k)}\left|w\right|\psi_{b}^{(k)}\right\rangle=0\,. \tag{8}\]
Here \(v_{eff}^{(k)}\), \(n^{(k)}\) and \(\psi_{m}^{(k)}\) are the effective potential, the density and the wavefunction of the \(m\)-th level, in the \(k\)-th iteration, respectively, and \(\{g_{i}\}\) symbolizes the dependence of the density on fractional occupations. Note that the presence of a Fermi-statistics problem implies that equality is fulfilled within the available range, as the left-hand side changes sign as a function of the occupations, with the degeneracy of the levels \(a\) and \(b\) lifted in one direction for one extreme occupation, and in the other direction for the other extreme. Numerically, equality is enforced by searching for the electronic configuration \(\{g_{i}\}\) that, once employed to obtain a new density \(n^{(k+1)}(\{g_{i}\})\), and consequently \(v_{eff}^{(k+1)}[n^{(k+1)}(\{g_{i}\})]\), will nullify the left-hand side of eq.. The search is made using Newton-Raphson's method, which converges within 3-5 steps for Fe, where we have required a convergence of \(10^{-6}\) for the occupation numbers. This search is repeated for each iteration in this part of the convergence process, affording us to preserve the degeneracy in the reference system.
Continuing the iterations described above, while preserving the degeneracy of the energy levels (\(\varepsilon_{{}_{3d\downarrow}}\) and \(\varepsilon_{{}_{4s\downarrow}}\) in Fe), but not constraining their value, we eventually converge to a self-consistent result, which is confirmed by increasing the mixing coefficient. This result for Fe coincides, within the convergence accuracy of \(\Delta_{E}=10^{-6}\) hartree, with the result reached by Janak's approach.
A schematic description of convergence of Janak's algorithm and the algorithm proposed above in the \(v_{eff}-n\) plane is shown in Janak's algorithm passes through "improper" densities. A full minimization of occupation numbers brings it, however, to a "proper" fractionally occupied configuration. The algorithm proposed here uses only "proper" configurations, and by reduction of the mixing constant, enters the NI-EVR range of potentials, where it converges to the desired density.
The process above is performed for all values of the total spin \(S\). In the example of Fe, the energy levels for \(S=0,1\) are occupied "properly in a broad sense" and possess higher total energies, as shown in Table 1. As mentioned above, the stability of the results obtained was verified by increasing the mixing coefficient once the iterations converged. No separation of the degenerate energy levels was observed. The influence of the initial guess for the effective potential on the final result was also checked. It was found that the eigenvalue at which the energy levels first became degenerate depends significantly on the guess. However, during the subsequent iterative process within the degenerate subspace, the eigenvalues converge to the expected value, along with the total energy and other quantities. Therefore, we conclude that the initial condition affects only the number of iterations needed to reach self-consistency, and not the converged self-consistent values.
In order to demonstrate that an emergence of a NI-EVR solution for Fe is not an artifact related to the local-density approximation for the exchange-correlation functional as speculated in Refs. and, we repeated the calculation adding the PBE-GGA to the VWN-LSDA. We obtained a NI-EVR density in this approximation as well, with \(E=-1263.446126\) hartree for the total energy and \([\mathrm{Ar}]3d_{1\,349}^{5}4s_{0\,651}^{1}\)for the electronic configuration, as can be seen from Table 1. Moreover, we have applied our algorithm successfully to additional atoms and ions using the same procedures described here. The details of these calculations are omitted for brevity.
It was emphasized recently that NI-EVR densities are not points in \(v\)-space, i.e. given an external potential \(v(\vec{r})\) that produces a NI-EVR density, there are potentials in its environment that also produce NI-EVR densities. We use the Fe atom to demonstrate such a situation. Changing the nucleus' charge \(Z\), we obtain a region \(25.82<Z<26.25\) for which NI-EVR densities occur, whereas for \(Z<25.82\) or \(Z>26.25\) PSVR solutions are obtained. The energy levels of the reference system and the occupation numbers are plotted in and 3, as a function of \(Z\).
Finally, we comment on the Be ion series that has been previously presented as an example of a NI-EVR system. It was proposed to solve the Be ion series in the limit of large \(Z\) analytically, by mapping it onto a four-electron system with an electron-electron interaction, which becomes negligible in the limit \(Z\rightarrow\infty\). The mapping is performed by a scaling \(\vec{r}=Zr/4\) for the distances and \(\vec{e}^{2}=4e^{2}/Z\) for the electron charge, in the Hamiltonian. When \(Z\rightarrow\infty\), the electron-electron interaction in the (\(\sim\))-system vanishes and the energy levels \(2s\) and \(2p\) become degenerate, with a first-order correction proportional to \(1/Z\). If the perturbation applied is of the exact form \(e^{2}/\left|\vec{r}_{1}-\vec{r}_{2}\right|\), then the correct zeroth order wave functions mix the \((2s)^{2}\) and \((2p)^{2}\) configurations and generate the NI-EVR density as reported by Ullrich and Kohn and in agreement with the CI-based procedure of Morrison. However, Kohn-Sham calculations do not support this result, for reasons unknown. We propose that this disagreement arises from symmetry considerations. If a Hartree type approximation modified by the LDA (or an extension) is the perturbation, then its spherical character will cause the relevant wave functions to separate into the \((2s)^{2}\) and \((2p)^{2}\) configurations, thus leading to a PSVR type density. Such a separation can be seen in Fig. 4, which illustrates the difference between the energy levels \(\varepsilon_{2s}\) and \(\varepsilon_{2p}\) of the Be ion series calculated for \(Z=4,40,100,200\) with VWN-LSDA.
## IV. DISCUSSION. T-DFT ANALYSIS.
It can be shown (see and references therein) that at finite temperatures the grand canonical potential
\[\Omega[v(\vec{r})]=-k_{{}_{B}}T\ln\left[\text{Tr}\left(\exp\left(-\frac{\hat{T }+\hat{W}+\hat{\rho}v}{k_{{}_{B}}T}\right)\right)\right] \tag{9}\]
II). Here \(\hat{\rho}=\sum_{i=1}^{N}\delta(\vec{r}-\vec{r}_{i})\) is the density operator, whose statistical average yields the electron density \(n(\vec{r})=\left<\hat{\rho}\right>\). The first functional derivative of \(\Omega\) equals the density: \(\delta\Omega/\delta v(\vec{r})=n(\vec{r})\). As a result, the Hohenberg-Kohn free-energy functional \(F_{{}_{HK}}[n(\vec{r})]=\Omega[v(\vec{r})]-\int n(\vec{r})v(\vec{r})\,d^{3}r\) is differentiable with respect to the density, and its first derivative is \(\delta F_{{}_{HK}}/\delta n(\vec{r})=-v(\vec{r})\). Since for a non-interacting electron system, \(F_{{}_{HK}}[n(\vec{r})]\) reduces to the kinetic energy functional, this last functional is differentiable as well. Therefore, for electronic systems at finite temperatures the \(v\)-representability question does not exist.
In T-DFT, a many-electron interacting system at a temperature \(T\) is described by reference to a non-interacting system with the same temperature. It follows then that the occupation numbers in the reference system are determined by the Fermi-Dirac distribution and the entropy is given by \(S=-k_{{}_{B}}\sum_{i}D_{i}\left(f_{i}\ln(f_{i})+(1-f_{i})\ln(1-f_{i})\right)\), where the sum is over all energy levels and \(f_{i}=g_{i}/D_{i}\). The effective potential preserves the form \(v_{{}_{eff}}(\vec{r})=v(\vec{r})+v_{{}_{H}}[n]+v_{{}_{xc}}[n]\), but now the exchange-correlation potential depends also on \(T\).
At finite temperature the occupancy \(g_{i}=D_{i}(1+\exp[(\varepsilon_{i}-\mu)/k_{{}_{B}}T])^{-1}\) is determined by the combination \((\varepsilon_{i}-\mu)/k_{{}_{B}}T\), which would lead us to expect that the fractional occupations of degenerate states will be equal. However, closer observation shows that in the limit of zero temperature the Fermi-Dirac distribution becomes multi-valued for \(\varepsilon_{i}=\mu\), taking on all values \(0\leq g_{i}\leq D_{i}\).
For cases with several degenerate states at the Fermi level having different occupations, the degeneracy is lifted linearly with temperature, with the combinations \((\varepsilon_{i}-\mu)/k_{{}_{B}}T\) left essentially unchanged. This can be seen in detail from the following relations. Expanding the energy and the chemical potential for low temperatures, we obtain \(\varepsilon_{i}(T)=\varepsilon_{i}^{}+\varepsilon_{i}^{}T+\dfrac{1}{2} \varepsilon_{i}^{}T^{2}\), \(\mu(T)=\mu^{}+\mu^{}T+\dfrac{1}{2}\mu^{}T^{2}\) up to terms \(O(T^{3})\), where \(\varepsilon_{i}^{(n)}=(d^{n}\varepsilon_{i}/dT^{n})_{T=0}\) and \(\mu^{(n)}=(d^{n}\mu/dT^{n})_{T=0}\). If at \(T=0\) there are two or more degenerate partially-occupied levels, whose energy is \(\varepsilon^{}\), then obviously \(\mu^{}=\varepsilon^{}\). Then, the partial occupation numbers of the \(i\)-th level at zero temperature are determined by the relation:
\[g_{i}^{}=D_{i}\lim_{T\to 0}\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!
## V Conclusions.
In conclusion, in the present contribution we introduce a new algorithm for the self-consistent solution of the Kohn-Sham equations in cases of non-PSVR densities. Since we have shown that densities which are constructed from excited states of the effective potential cannot be ground state densities of the interacting system, our algorithm solves for non-PSVR systems while considering NI-EVR densities only. It is perhaps surprising that such algorithms have hitherto scarcely been discussed. Avenues for future research include further development of such algorithms, e.g., based on studies of the efficiency of their performance.
The Fe atom was used as an illustration. By considering variations in the nuclear charge \(Z\) in Fe, we have also provided a demonstration that the subspace of external potentials which yields NI-EVR densities in Fe is finite. In addition, the fact that NI-EVR densities appear in Fe was found to be insensitive to the particular approximation of the exchange-correlation functional.
Lastly, the physical meaning of fractional occupation numbers, and in particular the unequal occupation of degenerate levels, has been clarified through the finite-temperature density-functional formalism. As the occurrence of ensembles in this formalism is natural, it may be expected that it will aid progress on additional issues as well, including the abovementioned issue of seeking improved algorithms.
|
10.48550/arXiv.1011.4564
|
Fractional occupation in Kohn-Sham density-functional theory and the treatment of non-pure-state v-representable densities
|
Eli Kraisler, Guy Makov, Nathan Argaman, Itzhak Kelson
| 712
|
10.48550_arXiv.2402.04039
|
###### Abstract
Ionic liquids (ILs) are an extremely exciting class of electrolytes for energy storage applications because of their unique combination of properties. Upon dissolving alkali metal salts, such as Li or Na based salts, with the same anion as the IL, an intrinsically asymmetric electrolyte can be created for use in batteries, known as a salt-in-ionic liquid (SiIL). These SiILs have been well studied in the bulk, where negative transference numbers of the alkali metal cation have been observed from the formation of small, negatively charged clusters. The properties of these SiILs at electrified interfaces, however, have received little to no attention. Here, we develop a theory for the electrical double layer (EDL) of SiILs where we consistently account for the thermoreversible association of ions into Cayley tree aggregates. The theory predicts that the IL cations first populate the EDL at negative voltages, as they are not strongly bound to the anions. However, at large negative voltages, which are strong enough to break the alkali metal cation-anion associations, these IL cations are exchanged for the alkali metal cation because of their higher charge density. At positive voltages, we find that the SiIL actually becomes _more aggregated while screening the electrode charge_ from the formation of large, negatively charged aggregates. Therefore, in contrast to conventional intuition of associations in the EDL, SiILs appear to become more associated in certain electric fields. We present these theoretical predictions to be verified by molecular dynamics simulations and experimental measurements.
Alkali metal salt doped ionic liquids, referred to as salt-in-ionic liquids (SiILs) here, have attracted attention as an electrolyte for applications in batteries/energy storage devices, owing to the combination of the active Li/Na cations coupled with the highly desirable properties of ionic liquids (ILs). ILs have extremely low vapour pressures, are effectively universal solvents, and are non-flammable making them a much safer alternative to carbonate based solvents typically used in conventional battery electrolytes. Unlike conventional battery electrolytes, SiILs are solely composed of ions, where the anion of the alkali metal salt and IL are often the same, e.g., Li[TFSI] dissolved in [EMIM][TFSI] resulting in there being more anions than either type of the cation. Therefore, these electrolytes are _inherently asymmetric_, and moreover, strongly correlated because of the super-concentration of salts.
One area where the asymmetry and correlation manifests is in the transport properties of the SiILs. In electrophoretic NMR experiments and MD simulations of SiILs, it has been revealed that, at low mole fractions of the alkali metal salt, the alkali metal cation has a _negative_ transference number, indicating that it moves in the opposite direction of the electric field. This property has been explained by the alkali metal cation being strongly solvated by anions. Owing to the larger number of anions, these species can form small, but high negatively charged clusters that are transported in a vehicular way.
These strong interactions are also evident at higher mole fractions. With increasing mole fraction, the sizes of the clusters increase from only containing 1 cation to containing several cations in an extended cluster connected by anions, as found to occur in MD simulations. These clusters continue to increase in size with the mole fraction and, eventually, a percolating ionic network of alkali metal cations and anions occurs. A simple, analytical theory for the ionic clusters of SiILs was developed by McEldrew _et al._, based on the analogy between polymers and concentrated electrolytes, which was able to rationalise the changes to the cluster distribution with mole fraction. Drawing from the famous theories of Flory, Stockmayer and Tanaka, it was predicted that the percolating ionic network of ions should behave like a gel, and in fact, experimental indications of this gel have been found in water-doped SiILs. Therefore, it is clear that ionic aggregation in SiILs is an extremely important phenomena.
One aspect of SiILs that is not well understood at all, however, is the electrical double layer (EDL), i.e. how the ions arrange themselves at an electrified interface. To our knowledge, Haskins _et al._ is one of the only papers investigating this topic in detail, presumably because of the complexity of the electrolyte. It found that the Li\({}^{+}\) induces disorder in the EDL and do not prefer to reside right at the electrode interface, opting instead to reside within the anion-rich layers. Recently, Goodwin _et al._ extended the theories of McEldrew _et al._ to make predictions in the EDL finding that the electric field induces cracking of clusters. However, this work has only been applied to simple symmetric ILs and further development is necessary to extend the theories of McEldrew and Goodwin _et al._ to SiILs.
In this paper, we develop a theory for the EDL of SiILs based on the work of McEldrew and Goodwin and co-workers. Conventional theories for the EDL are not able to correctly capture the associations which occur in the EDL nor the extended ionic aggregates which occur in SiILs. Here we present a theory for the EDL of SiILs which accounts for ionic associations in a consistent way and allows us to investigate these associations in the EDL of a strongly correlated and intrinsically asymmetric electrolyte. We find two main observations from this theory. Firstly, a cation exchange occurs at negative potentials, where the IL cation is found to be dominant at small negative potentials but is then replaced by the smaller alkali metal cation at increasingly negative potentials. This exchange occurs because of the competition between different sizes of species and the associations between the alkali metal cations and anions. Secondly, in positive potentials, we find there is some intermediate range of potentials where the SiIL _becomes more associated than in the bulk_.
## II Theory
We consider our system to be an incompressible lattice-gas model, which is composed of alkali metal cations, denoted by \(+\), and IL cations and anions, denoted by, respectively, \(\oplus\) and \(-\). In the bulk, the total volume fractions, \(\phi_{j}\), of each species are known, where \(j\) is \(+\), \(-\) or \(\oplus\). The lattice site is assumed to be the size of the alkali metal cation, \(v_{+}\), with the volumes occupied by the other species being determined by \(v_{+}\xi_{i}\), where \(\xi_{i}=v_{i}/v_{+}\) with being \(-\) or \(\oplus\) here. The dimensionless concentrations of each species are given by \(c_{j}=\phi_{j}/\xi_{j}\), with \(\xi_{+}=1\) for the alkali metal cations.
Similar to previous work, it is assumed that the anions and metal cations may form associations; but, the IL cations only interact with the open association sites of anions through a regular solution term. The metal cations can form a maximum of \(f_{+}\) associations and the anions a maximum of \(f_{-}\) associations, referred to as the functionality of each species. The number of associations is typically bounded by the integer number of coordination of an associating species within the first solvation shell of each species. Typically these associations are determined via a cutoff distance between the key moieties in the molecules, such as the oxygens' in a TFSI anion and a lithium ion. Additionally, they can be determined by studying the spatial distribution functions for the associating species to identify how many "hot-spot" regions exist for the species, which corresponds to the functionality of that species. Alternatively, kinetic criteria or machine learning methods can be used to define associations between ions in electrolytes. As these functionalities are larger than 1, a polydisperse cluster distribution can form clusters of rank \(lm\), where \(l\) is the number of cations and \(m\) is the number of anions in a cluster with rank \(lm\), with dimensionless concentration \(c_{lm}\), but we assume that only Cayley-tree-like clusters can form, i.e. clusters with no loops. This Cayley tree assumption of the clusters is required for our approach to remain analytical and physically transparent, but the approximation is known to break down sometimes.
For functionalities larger than 2, a percolating ionic network can emerge.
Schematic of the modulation of aggregations occurring in the EDL of SiILs near a positively (left) and negatively (right) charged electrodes. Here the metal cations can form up to 4 associations and the IL anions up to 3 associations. Ion associations are shown by touching vertices.
In the gel regime, we employ Flory's convention to determine the volume fractions of each species in the sol (\(\phi_{+/-}^{sol}\), which contains the clusters and free species) and gel phase (\(\phi_{+/-}^{gel}\), which only contains the percolating network of ions), where \(\phi_{+/-}=\phi_{+/-}^{sol}+\phi_{+/-}^{gel}\). The total dimensionless concentration of each species is given by a sum over all possible clusters, \(c_{+}=\sum_{lm}lc_{lm}+c_{+}^{gel}\), \(c_{-}=\sum_{lm}mc_{lm}+c_{-}^{gel}\).
The free energy functional (\(\mathcal{F}\)) is proposed to take the following form
\[v_{+}\mathcal{F}= \int_{V}d\mathbf{r}-\frac{v_{+}\epsilon_{0}\epsilon_{r}}{2}\big{(} \nabla\Phi\big{)}^{2}+v_{+}\rho_{e}\Phi+k_{B}Tc_{\oplus}\ln\phi_{\oplus}+\sum_{ lm}\left(k_{B}Tc_{lm}\ln\phi_{lm}+c_{lm}\Delta_{lm}\right)\] \[+\int_{V}d\mathbf{r}\chi\phi_{\oplus}\sum_{lm}\left(f_{-}m-m-l+1 \right)c_{lm}+\Delta_{+}^{gel}c_{+}^{gel}+\Delta_{-}^{gel}c_{-}^{gel}\] \[+\int_{V}d\mathbf{r}k_{B}T\Lambda\left(1-\xi_{\oplus}c_{\oplus}- \sum_{lm}(l+\xi_{-}m)c_{lm}\right) \tag{1}\]
Here the electrostatic potential, \(\Phi(\mathbf{r})\), charge density, \(\rho_{e}(\mathbf{r})\), and volume fractions/dimensionless concentrations, \(\phi(\mathbf{r})/c(\mathbf{r})\), all vary in space from the interface and we integrate over the entire electrolyte region. The first two terms represent the electrostatic energy, where \(\epsilon_{0}\) and \(\epsilon_{r}\) are, respectively, the permittivity of free space and the relative dielectric constant, \(\Phi\) is the electrostatic potential and \(\rho_{e}\) is the charge density, given by \(\rho_{e}=\frac{e}{v_{+}}(c_{\oplus}+c_{+}-c_{-})\), with \(e\) denoting the elementary charge. The third term is the ideal entropy of the IL cations, and the fourth term is the ideal entropy from clusters of rank \(lm\). The fifth term is the free energy of forming clusters, where \(\Delta_{lm}\) is the free energy of forming each cluster of rank \(lm\), with more details following later.
\[\sum_{lm}(f_{-}m-m-l+1)c_{lm}=f_{-}c_{-}(1-p_{-+}). \tag{2}\]
Note that this expression only holds in the pre-gel regime; where the left hand side assumes free anions or anions in clusters can interact with IL cations but the right hand side assumes all anions can interact with IL cations, even those in the gel. The seventh and eighth terms come from the free energy of species associating to the gel, \(\Delta_{j}^{gel}\), which is a function of \(\phi_{\pm}\) for thermodynamic consistency. Finally, the last term is a Lagrange multiplier to enforce the incompressibility more explicitly than previously reported. Note this term introduces another unknown, \(\Lambda\), which needs to be found.
We consider the free energy of forming a cluster of rank \(lm\) to have two contributions
\[\Delta_{lm}=\Delta_{lm}^{comb}+\Delta_{lm}^{bind}, \tag{3}\]
In the context of polymers, Stockmayer solved the combinatorial entropy for Cayley tree associations
\[\Delta_{lm}^{comb}=k_{B}T\ln\{f_{+}^{l}f_{-}^{m}W_{lm}\}, \tag{4}\]
where
\[W_{lm}=\frac{(f_{+}l-l)!(f_{-}m-m)!}{l!m!(f_{+}l-l-m+1)!(f_{-}m-m-l+1)!}. \tag{5}\]
Moreover, the binding free energy is simply given by
\[\Delta_{lm}^{bind}=(l+m-1)\Delta f_{\pm}, \tag{6}\]
We can calculate the chemical potential of the bare ions in the bulk and the EDL, where an overbar will indicate that the variable is the EDL version and \(\Phi\) is non-zero.
\[\beta\bar{\mu}_{\oplus}=\beta e\Phi+\ln\bar{\phi}_{\oplus}+\beta\chi\xi_{ \oplus}f_{-}\bar{c}_{-}\left(1-\bar{p}_{-+}\right)-\xi_{\oplus}\Lambda, \tag{7}\]
and for the clusters the chemical potential, up to arbitrary constants, is given by
\[\beta\bar{\mu}_{lm}=(l-m)\beta e\Phi+\ln\bar{\phi}_{lm}+\Delta_{lm}+\beta\chi \bar{\phi}_{\oplus}(f_{-}m-m-l+1)+(l+\xi_{-}m)(d^{\prime}-\Lambda), \tag{8}\]
In the bulk by asserting the clusters are in equilibrium with the bare species
\[l\mu_{10}+m\mu_{01}=\mu_{lm}, \tag{9}\]we can predict the cluster distribution from the bare species
\[c_{lm}=\frac{W_{lm}}{\lambda}\left(\lambda f_{+}\phi_{10}\right)^{l}\left(\lambda f _{-}\phi_{01}/\xi_{-}\right)^{m}, \tag{10}\]
where \(\lambda\) is the ionic association constant given by
\[\lambda=\exp\left\{\beta\left(-\Delta f_{\pm}+\chi\phi_{\oplus}\right)\right\}= \lambda_{0}\exp\left\{\beta\chi\phi_{\oplus}\right\}. \tag{11}\]
By establishing the chemical equilibrium between free species and clusters _within_ the EDL
\[l\bar{\mu}_{10}+m\bar{\mu}_{01}=\bar{\mu}_{lm}, \tag{12}\]
we obtain an analogous solution for the cluster distribution given the volume fractions of the bare species
\[\bar{c}_{lm}=\frac{W_{lm}}{\bar{\lambda}}\left(\bar{\lambda}f_{+}\bar{\phi}_{1 0}\right)^{l}\left(\bar{\lambda}f_{-}\bar{\phi}_{01}/\xi_{-}\right)^{m}, \tag{13}\]
where
\[\bar{\lambda}=\lambda_{0}\exp\left\{\beta\chi\bar{\phi}_{\oplus}\right\}. \tag{14}\]
Following the work of Goodwin _et al._Goodwin _et al._, we connect the bulk and EDL cluster distributions to the Poisson equation through closure relations. These closure relations are based on the pre-gel regime and naively extended to the post-gel regime (this assumes the additional chemical potential contribution is not significant, with further research being required to quantify its contribution). Equating the chemical potential of the free ions in bulk and in the EDL allows us to write down 3 additional equations.
\[\bar{\phi}_{10}=\phi_{10}\exp(-e\beta\Phi+\Lambda), \tag{15}\]
Second, for the anions
\[\bar{\phi}_{01}=\phi_{01}\exp(e\beta\Phi-\beta\chi f_{-}(\bar{\phi}_{\oplus}- \phi_{\oplus})+\xi_{-}\Lambda), \tag{16}\]
Finally, for the IL cations\[\bar{\phi}_{\oplus}=\phi_{\oplus}\exp\left(-e\beta\Phi+\beta\chi f_{-}\xi_{\oplus} \left\{c_{-}(1-p_{-+})-\bar{c}_{-}(1-\bar{p}_{-+})\right\}+\xi_{\oplus}\Lambda \right), \tag{17}\]
Note in previous versions, an additional parameter, \(\alpha\), was used in the closure relations which accounts for additional short-ranged correlations between ions. It is known these simple mean-field theories underestimate the correlations between ions, which causes the screening lengths to be too small, and this parameter is a way of correcting this deficiency. This \(\alpha\) parameter could also be introduced here, but for simplicity, we won't show results with this parameter.
Lastly to allow us to solve these Boltzmann closure relationships, we need to introduce the idea of association probabilities, conservation of associations and the law of mass action on associations. As \(\phi_{10}\) and \(\phi_{01}\) are, in principle, experimentally inaccessible. Instead, it is natural to express the cluster distribution in terms of the overall volume fractions of each species, \(\phi_{i}\), which is an experimentally and computationally controllable parameter through the mole fraction of the alkali metal salt. This connection is established by introducing ion association probabilities, \(p_{ij}\), which is the probability that an association site of species \(i\) is bound to species \(j\), where \(i\) and \(j\) are either the alkali cation (\(+\)) or the IL anion (\(-\)). Therefore, the volume fraction of free alkali cations can be written as \(\phi_{10}=\phi_{+}(1-p_{+-})^{f_{+}}\) and free anions as \(\phi_{01}=\phi_{-}(1-p_{-+})^{f_{-}}\).
The association probabilities can be determined through the conservation of associations and a mass action law between open and occupied association sites.
\[p_{+-}\psi_{+}=p_{-+}\psi_{-}, \tag{18}\]
The mass action law between open and occupied association sites is
\[\lambda\zeta=\frac{p_{+-}p_{-+}}{(1-p_{+-})(1-p_{-+})}, \tag{19}\]where \(\zeta=\psi_{-}p_{-+}=\psi_{+}p_{+-}\) is dimensionless concentration of associations (per lattice site). The definition of \(\lambda\) as the ionic _association constant_ becomes clear from its appearance in the association mass action law. It sets the equilibrium for association sites to be occupied or open.
\[\psi_{-}p_{-+}=\psi_{+}p_{+-}=\frac{1+\lambda(\psi_{-}+\psi_{+})-\sqrt{\left[1+ \lambda(\psi_{-}+\psi_{+})\right]^{2}-4\lambda^{2}\psi_{-}\psi_{+}}}{2\lambda}. \tag{20}\]
Note, analogous expression for the association probabilities occur in the EDL but where EDL quantities are utilised.
As shown in the works of McEldrew _et al._, the association constants can be found from fitting Eq. using association probabilities obtained from MD simulations. For the specific case of SiILs, there are two contributions to the association constant, \(\lambda_{0}\) and \(\chi\), which can be extracted if various salt compositions are studied. As the association constant in the EDL only depends on these bulk parameters, and variables in the EDL, we do not need to further determine \(\bar{\lambda}\). In Ref. 22, it was found that \(\lambda_{0}\approx 50\) and \(\chi\approx-2\) k\({}_{B}\)T, which shall mainly be used in this study. Therefore, our theory has no free fitting parameters in terms of the associations in the SiILs. This feature is achieved as all the parameters needed to predict the EDL structure can be determined from the bulk SiIL. Explicitly one can extract \(\lambda_{0}\), \(\chi\), and integer bounds on \(f_{+}\) and \(f_{-}\) from MD simulations of the SiIL system at various salt compositions and the volume ratios are backed out by molecular data of the species composing the SiIL. Alternatively, one can use integral equations and Wertheim's formalism to determine the association constants without fitting association constants from MD simulations. However, this is only accessible for certain cases, such as charged hard spheres with known Mayer f-functions for interacting species, and not for specific chemistries.
Taking the functional derivative of the free energy with respect to the electrostatic potential, we arrive at the Poisson equation,
\[\epsilon_{0}\epsilon_{r}\nabla^{2}\Phi=-\rho_{e}=-\frac{e}{v_{+}}(\bar{c}_{ \oplus}+\bar{c}_{+}-\bar{c}_{-}). \tag{21}\]
From this equation, we can define the inverse Debye length, \(\kappa=\sqrt{e^{2}\beta(c_{\oplus}+c_{+}+c_{-})/v_{+}\epsilon_{0}\epsilon_{r}}\), which will be used to to convert distances from the electrode to dimensionless values. The Poisson-Boltzmann (PB) equation follows from a mean-field approximation of Coulomb interactions, which is technically only valid in the dilute limit of point-like ions in a uniform dielectric continuum. Many modified PB equations are available which involve corrections for finite ion sizes, Coulomb correlations, and non-local dielectric response. In our simple model, such corrections are captured indirectly through their short-range associations, which is to promote the formation of ionic clusters with "spin glass" ordering favoring oppositely charged neighbors.
To solve the system of equations and the Poisson equation, we use the following procedure. First, we calculate bulk properties from Eq. using the association constant. Next we solve our closure relations, Eqs.-, with Eq. and the incompressibility condition in the EDL (\(1=\bar{\phi}_{+}+\bar{\phi}_{-}+\bar{\phi}_{\oplus}\)), as we need 6 equations to solve for our 6 unknowns, to obtain the relationships between \(\bar{\phi}_{+},\bar{\phi}_{-},\bar{\phi}_{\oplus}\) and \(\Phi\), on a regular grid of \(\Phi\). We then use this mapping between \(\Phi\) and \(\bar{\phi}\)'s to numerically solve the Poisson equation. Note that this procedure is similar to that of Ref., but where we now also solve the Lagrange multiplier, \(\Lambda\), to account for the different sizes of each species.
An informative quantity to understanding the EDL can be the differential capacitance, C, also known as the double layer capacitance,
\[C=\frac{dq_{s}}{d\Phi_{s}}. \tag{22}\]
Here, \(\mathrm{q}_{s}\) is the surface charge at the interface and \(\Phi_{s}\) is the electrostatic potential at the charged interface, this is equivalent to the potential drop across the EDL. Using our previous procedure to solve the Poisson equation over range of \(\mathrm{q}_{s}\) we can numerically solve for how the \(\mathrm{q}_{s}\) is a function of \(\Phi_{s}\). Following this, we can numerically take the derivative of \(\mathrm{q}_{s}\) with respect to \(\Phi_{s}\) to determine the differential capacitance. For additional details on setting up the system of equations and reproducing the calculations see the Supplemental Material.
## III Results
Here we describe the predictions of the model in detail. As the theory heavily relies on previous work for SiILs in the bulk, we suggest the reader is first acquainted with Ref.. Moreover, reading Ref. might help the reader put into context the predictions for the EDL. In all of the results shown in the main text, we choose the following parameters.
A mole fraction of alkali metal salt of \(x_{s}=0.01\) is chosen, which is quite small, and was motivated by ensuring that no gelation occurs for all potentials for the rest of the parameters used. For the functionalities, we choose \(f_{+}=5\) and \(f_{-}=3\), inspired by previous work. We have chosen \(\chi=-2\) k\({}_{B}\)T and \(\lambda_{0}=50\), which correspond to values previously used for SiILs. Finally, we choose \(\xi_{+}\)=1, \(\xi_{-}\)=7, \(\xi_{\oplus}\)=7. Some predictions of the model highly depend on these parameters. We have included calculations in the Supplemental Material for how the results depend on these parameters.
In Fig. 2, we display the solutions to our equations as a function of electrostatic potential.
Properties of the EDL of SiILs as a function of applied electrostatic potential. a) Association probabilities. b) Proximity to gelation, 1-\(\bar{p}_{+-}\bar{p}_{-+}\)(f\({}_{+}\)-1)(f\({}_{-}\)-1). c) Total volume fractions of each species. d) Volume fraction of free cations, anions and clusters. Here we use \(x_{s}=0.01\), \(f_{+}=5\), \(f_{-}=3\), \(\xi_{+}=1\), \(\xi_{-}=7\), \(\xi_{\oplus}=7\), \(\chi=-2\) k\({}_{B}\)T, and \(\lambda_{0}=50\).
First, we focus on the changes in the total volume fractions of each species, \(i=\oplus,+,-\), as a function of potential as seen in Fig. 2.c). For positive applied potentials, there is a large increase in \(\bar{\phi}_{-}\), which plateaus at 1 at 0.2 V, while the IL cations, \(\bar{\phi}_{\oplus}\), are monotonically expelled from the EDL. Interestingly, there is an increase in the volume fraction of alkali metal cations, \(\bar{\phi}_{+}\), at modest applied positive potentials, before the volume fraction decreases to 0 at large applied positive voltages. For negative potentials, the anions are monotonically expelled from the EDL but the response of the cations is more complex. At small to moderate voltages, the IL cation dominates the EDL, and saturates at \(\bar{\phi}_{\oplus}\approx 1\) at a potential of -0.25 V. At potentials more negative than -0.3 V, a cation exchange occurs, where the IL cations are replaced by metal cations, \(\bar{\phi}_{\oplus}\approx 0\) and \(\bar{\phi}_{+}\approx 1\).
A cation exchange occurs at negative voltages because of the competition between the different sizes of the cations, and its position is dependent on the strong associations between the metal cations and the anions. At small potentials, the IL cations first populate the EDL because they are not strongly interacting with the anions, which are pushed out of the EDL and take the alkali metal cations with them through the associations. At large negative voltages, it would be expected that the smallest cation populates the EDL, as this maximises the charge density in the EDL and reduces its length. Therefore, it is expected that the smaller alkali metal cation saturates the EDL at large negative voltages instead of the larger IL cation. If both cations are the same size, this cation exchange does not occur as the IL cation remains in the EDL. There is a critical voltage at which this cation exchange occurs, which depends not only on the sizes of the cations but also the association constant between the alkali metal cation and the anion. For larger association constants, the alkali metal cation is more strongly bound to the anions, which means it requires larger fields to break these associations to obtain the cation exchange. In the Supplemental Material we show how the cation exchange depends on these parameters in more detail.
Next we describe the association probabilities as a function of applied voltage, as seen in Fig. 2.a). At large negative potentials, \(\bar{p}_{+-}=0\) and \(\bar{p}_{-+}=1\), which is expected if there is \(\bar{\phi}_{+}=1\) and \(\bar{\phi}_{-}=0\). This can be understood by considering trace amounts of IL anions in alkali metal cations. In this case, it would expected the alkali metal picked at random to rarely be associated to an IL anion; but, any IL anion picked at random would be associated with alkali metal cations. Similarly, for large positive potentials \(\bar{p}_{-+}=0\) and \(\bar{p}_{+-}=1\) because of \(\bar{\phi}_{-}=1\) and \(\bar{\phi}_{+}=0\). These limits are also discussed in detail in Ref.. At a voltage of \(\sim 0.1\) V, however, there is also a peak in \(\bar{p}_{-+}\), which occurs because there is an increase in the volume fraction of alkali metal cations. These cations are brought to the EDL from the strong associations with the anions, and therefore, results in _electric field induced associations in the EDL_. This can be further seen in Fig. 2.b), which displays the proximity to gelation, 1-\(\bar{p}_{+-}\)\(\bar{p}_{-+}\)(f\({}_{+}\)-1)(f\({}_{-}\)-1). Here the proximity to gelation measures how close the clusters are to having the capacity to potentially form an infinite cluster, i.e. proximity to percolation. The criterion for percolation for clusters with a cation-anion associated backbone has been previously derived to occur when 1=(f\({}_{+}\)-1)(f\({}_{-}\)-1)\(p^{*}_{+-}p^{*}_{-+}\), where \(p^{*}_{ij}\) indicate the critical association probability for gelation. From this, we can define the proximity to gelation as 1-\(\bar{p}_{+-}\)\(\bar{p}_{-+}\)(f\({}_{+}\)-1)(f\({}_{-}\)-1) as when it approaches zero the system will approach gelation and as it becomes negative more of the system will be in the gel phase. At most voltages especially large ones, the system is far from being a gel. At no applied field, the system is not close to the gel point; but as a positive potential is applied, the system gets closer and closer to forming a gel. At an applied potential of \(\sim 0.1\) V, the system almost reaches the gel point before the _electric field induced associations in the EDL_ are destroyed in favour of the saturation regime of anions at large potentials. There have been suggestions of similar effects in other electrolytes EDLs.
Finally, we describe the response of free ions and aggregates as seen in Fig. 2.d). As expected, the volume fraction of free alkali metal cations are practically 0 until the cation exchange has occurred at voltages more negative than -0.3 V. There are still many free anions at no applied voltage owing to the highly asymmetric mixture of the electrolyte. As positive potentials are applied, the volume fraction of free anions increases until 0.1 V, when a plateau occurs. This plateau occurs at the same voltage as the bump in \(\bar{p}_{-+}\), as the additional anions accumulating in the EDL are not free, but are associating to the additional alkali metal cations in the EDL. For larger potentials, the saturation regime of anions is reached, and the volume fraction of anions reaches 1.
The description of the EDL properties displayed in allows us to now understand the distributions of ions and association probabilities in the EDL. In Fig. 3, we show properties of the EDL for negatively charged interfaces. In Fig. 4, we display the similar figures but for positively charged interfaces. All plots are shown as a function from the electrified interface, in dimensionless units of distance, normalised by the inverse Debye length, \(\kappa\), here \(1/\kappa\) is \(\lambda_{D}\sim 0.3\) A for the EDL plots presented here. These plots were produced from numerically solving the grids shown in in the Poisson equation, shown in Eq..
For negative voltages, the electrostatic potential decays in an exponential-like way to 0 in the bulk, as seen in Fig. 3.a). The charge density starts from 0 in the bulk, and first reaches a plateau of \(\sim\)1/7, in units of the number of lattice sites, which corresponds to the saturation regime of the IL cation, before the cation exchange occurs and the dimensionless charge density saturates at 1, i.e. the alkali metal cation saturates the EDL. As seen in Fig. 3.c), the cation exchange is now separated in space. Close to the interface, we find alkali metal cations are saturating the high field. After this dense layer of alkali metal cations, there is an additional saturation layer of the IL cations where the anions are also depleted. Naturally, there is a strong increase of \(\bar{\phi}_{10}\) from the free cations right near the interface and a depletion of free anions \(\bar{\phi}_{01}\) occurs further out from the interface. At negative potentials, there is only a monotonic destruction of aggregates, which can also be understood from \(\bar{p}_{+-}\bar{p}_{-+}\), which can be used to quantify the gel transition, which also monotonically decreases.
For positive voltages, the behaviour is more complex as seen in The electrostatic potential decays in a monotonic way from the interface and the charge density has the typical saturation curve of a Fermi-like function. The structure of the electrolyte is complex despite these relatively simple appearing electrostatic potential and charge density distributions. As seen in Fig. 4.c), the total volume fraction of anions monotonically increases and the IL cation volume fraction monotonically decreases towards the interface. The alkali metal cations, however, have a strongly non-monotonic dependence, as shown in where the volume fraction has been scaled up by a factor of 100 to make it more visible. Approaching the interface from the bulk, \(\bar{\phi}_{+}\) dramatically increases, before reaching a peak, to then dramatically decrease in the saturation regime of the anions. Interestingly, as seen in Fig. 4.d), this increase in \(\bar{\phi}_{+}\) exclusively corresponds to the formation of large aggregates in the EDL at moderate voltages, as can be seen from \(\bar{\phi}_{10}\) monotonically decreasing towards the interface and a large peak in \(\bar{\phi}_{11>}\). This large peak in \(\bar{\phi}_{11>}\) also appears as a peak in \(\bar{p}_{-+}\) in the EDL, as well as \(\bar{p}_{-+}\bar{p}_{+-}\) approaching the critical gel line. Therefore, there are intermediate regions in space of the EDL at positive voltages where the SiIL becomes more aggregated than in the bulk. This finding is in contrast to previous theories of the EDL with associations, which only predicted the destruction of associations. This finding comes generally from the asymmetry of the interactions, functionality, size and numbers of species found in SiIL. However, even with these asymmetries, there are specific compositions of SiIL and parameters such that the electric field induced associations are not present. The system will be most aggregated when the product of the association probabilities is maximized, which is not necessarily guaranteed to occur at zero voltage when asymmetries are present. The observation of electric field induced associations has been seen before in MD simulations of other electrolyte systems; a similar theory to the one presented here may be capable of explaining those observations.
Overall, the EDL of SiILs is more complex than ILs. At negative voltages, the EDL
Distributions of properties of SiILs in the EDL as a function from the interface, in dimensionless units, where \(\kappa\) is the inverse Debye length. a) Electrostatic potential. b) Dimensionless charge density, in units of \(e/v_{+}\). c) Total volume fractions of each species. d) Volume fractions of free cations, free anions and aggregates. e) Association probabilities. f) Product of association probabilities, where the dashed line indicates the critical line. Here we use \(x_{s}\) = 0.01, \(f_{+}\) = 5, \(f_{-}\) = 3, \(\xi_{+}\) = 1, \(\xi_{-}\) = 7, \(\xi_{\oplus}\) = 7, \(\chi\) = \(-\)2 k\({}_{B}\)T, \(\lambda_{0}\) = 50, \(\epsilon_{r}\) = 5, and q\({}_{s}\) = -0.3 C/m\({}^{2}\).
for the more general case); moderate distances in the EDL, where the IL cation dominates; and small distances a saturation layer of alkali metal cations. Similarly for positive voltages, there are three regimes in space: the bulk, where associations occur to create small negative aggregates; moderate distances in the EDL, where the change in composition occurs large aggregates to appear from electric field induced associations; and the saturation regime close to the interface.
Distributions of properties of SiILs in the EDL as a function from the interface, in dimensionless units, where \(\kappa\) is the inverse Debye length. a) Electrostatic potential. b) Dimensionless charge density, in units of \(e/v_{+}\). c) Total volume fractions of each species. d) Volume fractions of free cations, free anions and aggregates. e) Association probabilities. f) Product of association probabilities, where the dashed line indicates the critical line. Here we use \(x_{s}=0.01\), \(f_{+}=5\), \(f_{-}=3\), \(\xi_{+}=1\), \(\xi_{-}=7\), \(\xi_{\oplus}=7\), \(\chi=-2\) k\({}_{B}\)T, \(\lambda_{0}=50\), \(\epsilon_{r}=5\), and q\({}_{s}=0.18\) C/m\({}^{2}\).
Moreover, the link between associations and overscreening was established in Ref., and further discussed in Ref.. These associations occur without any visible feature in the charge density, and therefore, must be probed by computing the associations in the EDL through more sophisticated theories or through experimental means.
Next we turn to how the screening length, \(\lambda_{s}\), varies with the mole fraction of alkali metal salt, \(x_{s}\), in the pre-gel regime. The screening length was extracted from fitting an exponential function to the electrostatic potential at small potentials for various \(x_{s}\). In Fig. 5.a), we observe that as \(x_{s}\) increases the screening length decreases. It is typically expected, however, that the screening length increases as the electrolyte becomes more associated, and as \(x_{s}\) increases the SiILs becomes more associated. Here we find that the screening length actually decreases with \(x_{s}\) because of the formation of small, highly charged aggregates, which is the opposite of conventional intuition.
Lastly, we predict the differential capacitance for SiILs as a function of potential, as shown in Fig. 5.b). At low potentials the differential capacitance takes on a slightly asymmetry bell shape. At large positive potentials, we observe the typical decay in differential capacitance.
b) Differential capacitance of SiIL with \(x_{s}=0.01\), as a function of electrostatic potential. Here we use \(f_{+}=5\), \(f_{-}=3\), \(\xi_{+}=1\), \(\xi_{-}=7\), \(\xi_{\oplus}=7\), \(\chi=-2\) k\({}_{B}\)T, \(\lambda_{0}=50\), and \(\epsilon_{r}=5\). b.) Here the red dashed line at -0.43 V indicates the cation exchange voltage.
However, at a critical potential we observe a large peak in differential capacitance which can be attributed to the cation exchange shortening the EDL. This satellite peak can be used to verify the predictions of our theory.
In the Supplemental Material, we display additional plots and discuss the role of changing parameter and variable values from the values chosen in the main text.
## IV Discussion
In Ref., Haskins _et al._ performed molecular dynamics simulations of Li[TFSI] in [pyr14][TFSI] at electrified interfaces. For positive voltages, they found that the anions dominate the first layer of ions with some IL cations. After this layer, there is a sharp peak in alkali metal cations residing between layers of anions. The second peak of anions does not have an associated second peak of Li\({}^{+}\) ions, which could indicate that the Li\({}^{+}\) ions in between these peaks are associating to anions in multiple layers of the overscreening structure. Whereas at negative potentials, the IL cation dominates the first layer of the EDL, with no Li\({}^{+}\) ions. In the second layer of ions, there is pronounced overscreening with an increase in the number density of anions accompanied by a dramatic increase in Li\({}^{+}\) ions. In this second layer, large aggregates could be forming. These MD results are promising in confirming the possible electric field induced associations, but further MD simulations must be done to confirm the predictions of our theory. The EDL of SiILs is an extremely understudied problem and we hope our theory will provide a framework in which it can be understood.
A few studies have looked at the differential capacitance of SiILs. In Ref., Yamagata _et al._ experimentally measured the differential capacitance of Li[TFSI] in [EMIM][FSI] and Li[TFSI] in [EMIM][TFSI] at negative potentials. In their experiments, they observe a significant difference between these two SiILs differential capacitance at large negative potentials in comparison to the neat ILs. For Li[TFSI] in [EMIM][FSI], they observe that at low potential the differential capacitance is lower than the neat IL; however for larger negative potentials, the SiILs differential capacitance grows significantly. The sharp rise in differential capacitance seen in the Li[FSI] in [EMIM][FSI] could be coming from a cation exchange occurring in the system. They proposed a schematic of their EDL for this system where the Li\({}^{+}\) ion is the cation present at the interface, while still being associated with an FSI anion. In our theory, this peak in the differential capacitance comes from the cation exchange, which is a similar origin to that suggested by Yamagata _et al._. In contrast in the Li[TFSI] in [EMIM][TFSI] system, they do not see this differential capacitance growth at large negative potentials. They consider the EDL for this case to have the interface occupied by the EMIM cations and, therefore, has not undergone the cation exchange yet. This schematic is similar to what is predicted from the MD results in Ref., as discussed above. For certain parameters of our theory, we do not obtain this satellite peak from the cation exchange as shown in the Supplemental Material.
Experimentally, it has been found that ILs and other concentrated electrolytes have extremely long force decay lengths as measured from surface force measurements. The assumption of Gebbie _et al._ was that these forces were electrostatic in origin, owing to a large renormalisation in the concentration of free charge carriers. While this topic remains controversial and unresolved, we would like to briefly highlight that the aggregation of the alkali metal cations and anions in SiILs was able to successfully explain the transport properties in these super-concentrated electrolytes. Therefore, aggregates could shed some light onto these measurements of surface force interactions in SiILs and long ranged interactions in other concentrated electrolytes. Recently, experimental evidence for the connection between the cluster-like (soft) structure in a SiIL composed of NaTFSI and EMIM TFSI, at various mole ratios, and long-range (non-exponentially decaying) surface forces measured with a Surface Forces Apparatus. Furthermore, the long-ranged interactions point at the presence of aggregates with a high aspect ratio at the mica interfaces that contribute to the surface force, rather than a purely electrostatic origin. We believe that the theory described here could shed some light not only into these measurements, but also into the long ranged interactions in other concentrated electrolytes.
We have chosen not to include the \(\alpha\)-parameter introduced in Ref. and adopted in Ref.. Therefore, the voltages in our curves are smaller than what is expected from MD simulations and experiments which should be kept in mind when interpreting the results. As outlined in Ref., there are also some conceptual and technical limitations with our presented theory. Firstly, overscreening, the alternating layers of cations and anions, is represented through the thermo-reversible associations between the ions; but, the internal structure of these aggregates is not explicitly defined here. Therefore, we cannot explicitly obtain decaying oscillations in the charge density but need to recognise that overscreening is included through these associations. In SiILs, the electric field induced associations at positive voltages further complicates this problem. If the potential is oscillating, there will be positive potential values at both the cathode and anode. Thus, we might, in fact, predict that there are electric field induced associations at both charged interfaces. This will need to be further studied through MD simulations. In addition, our theory can describe well the diffuse EDL, but not the interfacial layer of ions, where a breakdown of the cluster distribution should occur and specific interactions with the electrode could dominate. This latter point could be predicted by the MD simulations of Haskins _et al._ which found that the IL cation prefers to reside in the first layer of ions, where we expect the alkali metal cations to dominate the saturation regime. However, this could also be a problem with short equilibration times of the simulations as discussed later.
The largest limitation of our theory is that we find the screening length is less than 1 A. Since \(\lambda_{s}\) is smaller than an individual alkali metal cation (\(\sim 2.8\) A, as used in this work), and much less than the size of the small clusters predicted to be in the SiIL solution. This points to the theory not being self-consistent in terms of its electrostatic predictions, as is generally known to be a problem of mean-field lattice gas models which is discussed in detail in Ref.. This common failing of mean-field theories can be corrected with more involved modified Poisson-Boltzmann equations that can account for higher-order correlations and/or non-local effects. This issue can be addressed to some extent in local mean field models such as by accounting for higher order local terms such as done in BSK theory or via modifying the Coulomb interactions. However typically to overcome this inconsistency, one must turn to a non-local model that can capture the entropic effects of excluded volume in a significantly better fashion. The issue from the current inconsistency indicates caution should be used when comparing the model's predictions of the spatial profile of the ions and clusters near the surface. Importantly even with this inconsistency, mean-field models can have valuable predictive power in integrated quantities such as the capacitance of the double layer, the general trends in what accumulates in the EDL, and importantly, the new predictions for how association probabilities change within the EDL.
Finally, we have presented an analysis of the equilibrium properties of SiILs. One of the noteworthy properties of these systems is their transport properties. As outlined in Ref., the dynamics of this model can also be investigated, which could yield some interesting effects. Moreover as this system has an active cation for intercalation, developing the theory of coupled ion electron transfer reactions to include microscopic solvation effects could be a promising place to start.
## V Conclusions
Here we have developed a theory for the electrical double layer (EDL) of salt-in-ionic liquids (SiILs). These electrolytes are interesting for applications in energy storage but their behavior at electrified interfaces is not well studied. Here, we found a highly asymmetric response of the SiILs in the EDL. At negative potentials, associations are destroyed in the EDL; but, a cation exchange can occur when the IL cation initially saturates in the EDL but is later replaced by the smaller, associating alkali metal cation. In contrast, the positive potentials predict there to be electric field induced associations, which may lead to gelation. At large positive voltages we obtain the saturation regime for anions. All of these EDL features appear to be captured in the predicted differential capacitance profile for SiILs. Overall, we hope that the predictions made by this theory will inspire further work from molecular dynamics simulations to understand these SiILs, or for experimental measurements to be performed to probe the cation exchange or electric field induced associations.
|
10.48550/arXiv.2402.04039
|
Electric Field Induced Associations in the Double Layer of Salt-in-Ionic-Liquid Electrolytes
|
Daniel M. Markiewitz, Zachary A. H. Goodwin, Michael McEldrew, J. Pedro de Souza, Xuhui Zhang, Rosa M. Espinosa-Marzal, Martin Z. Bazant
| 5,513
|
10.48550_arXiv.2307.09565
|
## Abstract:
The solvation structures and ion dynamics of CaCl\({}_{2}\) aqueous electrolytes have been investigated using _ab initio_ molecular dynamics simulations and molecular dynamics simulations with deep learning potentials. We found multiple solvation structures around the Ca\({}^{2+}\) ion, including fully hydrated single Ca\({}^{2+}\) ion, Ca-Cl contact ion pair, and Ca-2Cl bridged ion pair, could coexist. The ion-pairing condition plays an important role in the translational and orientational distribution of water molecules in the solvation shell. And the local ordering introduced by the Ca\({}^{2+}\) ion can extend to the second solvation shell. Furthermore, we found the lifetime of water molecules in the solvation shell is sensitive to the ion-pairing conditions. The self-diffusivities of ions and water molecules, as calculated in molecular dynamics simulations with deep learning potentials, are in good agreement with experimental measurements. Finally, we elucidate the transition of Ca\({}^{2+}\) ion dynamics between different regimes by analyzing angle probability distribution histograms and van Hove correlation function.
Introduction
Aqueous electrolytes are ubiquitous components in chemistry, energy, biology, and environmental sciences. Over the years, numerous noteworthy studies have emerged aimed at comprehending the solvation structure of ions, dynamics of ions and water, and macroscopic thermodynamics, all of which play crucial roles in the design and application of aqueous electrolytes. For example, investigations employing techniques such as dielectric relaxation and femtosecond infrared spectroscopy have revealed that monovalent cations and anions can exhibit a cooperative effect on slowing down the reorientation dynamics of the water molecules in hydration shells. Far-infrared (terahertz) and X-ray absorption studies also provide direct evidence of the formation of contact ion pair. Nevertheless, the solvation structures captured through diverse scattering and spectroscopic techniques remain inconclusive. Additionally, debates persist regarding the range of ions' effect on the water. X-ray absorption suggests a local effect, limited to one hydration shell, and dielectric relaxation support a long-range effect, extending to several hydration shells. For a more detailed understanding, interested readers are encouraged to refer to several comprehensive reviews discussing the experimental investigations on the structure and dynamics of hydration ions.
From a theoretical standpoint, two computational methodologies have been extensively employed to investigate the physiochemical properties of aqueous electrolytes: hybrid cluster/continuum quantum chemistry calculations and molecular dynamics (MD) simulations. The former is typically employed to explore the solvation structure and thermodynamics (e.g., solvation energy, redox potential, etc.). This approach heavily relies on ion charge localization and the relevance of the infinite dilution limit. MD simulations, encompassing classical and _ab initio_ methods, can provide ensemble-averaged solvation structure, dynamics, and thermodynamics. Classical MD (CMD), in comparison to _ab initio_ MD (AIMD), offers a longer time scale (\(\sim\)100 ns) and larger length scale (\(\sim\)10 nm). Therefore, CMD has been broadly applied in complex systems composing of intricate and bulky ions and solvent molecules. However, the accuracy of CMD highly depends on the reliability of the potential/force field. Some recent studies emphasize the importance of considering the electronic diversity of water molecules and charge transfer between an ion and its hydration shell to accurately reproduce the dynamics in aqueous electrolytes. AIMD simulations, which utilize forces derived from the ground state of electrons, can be used to capture the accurate structure, dynamics, and thermodynamics. However, some transitions between metastable solvation structures are hindered kinetically and cannot be observed within the short-time period currently accessible to AIMD (\(\sim\)100 ps). Therefore, AIMD simulations are primarily applied in some simple scenarios, such as single-ion systems (a single cation/anion mediated in solvents) and monovalent ion pair systems. Recent advancements in machine learning force fields have partially resolved the accuracy versus efficiency dilemma. A deep learning potential can be trained using a large database comprising corresponding atomic coordinates, forces, and energies calculated from AIMD simulations or density functional theory (DFT) calculations. Recent studies have demonstrated that MD simulations employing deep learning potentials can accurately reproduce the structural and dynamic properties of several condensed phase systems obtained from AIMD simulations while significantly reducing computational costs by a factor of 10\({}^{4}\)\(\sim\)10\({}^{5}\).
In this work, we employ a combination of metadynamics AIMD and machine learning MD simulations to investigate the solvation structure and ion dynamics in a representative divalent aqueous electrolyte, specifically, CaCl\({}_{2}\). To the best of our knowledge, this is the first investigation to systematically examine solvation and dynamics in a multivalent aqueous solution using the combination of metadynamics and machine learning techniques. Our findings reveal the coexistence of multiple metastable solvation structures within the solution. Furthermore, we highlight the crucial role of ion pairing in shaping the translational and orientational distribution of water molecules within the solvation shell, as well as the lifetime of water molecules in that shell. Additionally, we demonstrate that MD simulations employing a deep learning potential are highly effective in capturing ion dynamics with accuracy comparable to AIMD simulations.
## 2 Computational Methods
### Simulation systems
The metadynamics AIMD simulations were performed to explore the free energy surface (FES) as a function of the coordination environment of Ca\({}^{2+}\) ion in the CaCl\({}_{2}\) aqueous solution. 1 CaCl\({}_{2}\) and 64 water molecules were randomly packed in a cubic simulation box with a side length of 12.51 A using the PACKMOL code. The concentration is 0.848 M (mol/L) and the density is 1.073 g/ml, which is consistent with the experiments. The initial configuration was then relaxed using classical MD simulations with generic OPLS-aa force fields under the NVT ensemble for 1 ns. Then, the AIMD metadynamics simulations were performed for 100 ps. The normal AIMD without biased potential and machine learning MD simulations were performed starting from three local metastable solvation structures of Ca\({}^{2+}\) ion, which are fully hydrated single Ca\({}^{2+}\) ion, Ca-Cl contact ion pair, and Ca-2Cl bridged ion pair, denoted as FHS, CIP, and BIP system, respectively. The dimension and composition of these three simulations are consistent with metadynamics AIMD simulations. Three AIMD and MD simulations were performed for 25 ps and 1 ns, respectively.
### Simulation methods
The AIMD simulations were carried out using the Quickstep module implemented in the CP2K package under the NVT ensemble. Core electrons were described by norm-conservative Goedecker-Teter-Hutter (GTH) pseudopotentials. The valence electrons were described at the DFT level using the revPBE parametrization functional and an atom-centered Gaussian double-zeta basis set augmented with a set of d-type or p-type polarization functions (DZVP-MOLOPT-SR-GTH). Several prior studies have found the revPBE functional has great performance in reproducing the dynamics properties in water-based systems. The DZVP-MOLOPT-SR-GTH has also been broadly applied in many liquid electrolytes. The energy grid cutoff was set as 400 Ry, which was determined through the energy converge test (shown in Figure S1). DFTD3 empirical Grimme correction is considered for the long-range dispersive forces. The MD simulation with the deep learning potential was performed using LAMMPS under the NVT ensemble. Nose-Hoover thermostat was used to stabilize the temperature at 303 K. The simulation box is periodic, and the time step is 0.5 fs for both AIMD and MD.
Two collective variables, the coordination number (CN) of Cl\({}^{\cdot}\) ions and oxygen atoms in water molecules around the Ca\({}^{2+}\) ion, were used in metadynamics AIMD simulations. Small repulsive Gaussian hills with a height of \(\sim\)0.054 eV (0.002 Hartree) and a width of 0.2 CN were added at a frequency of once per every 100 time steps for these two collective variables. The low Gaussian hill, which is \(\sim\)1/60 of the depth of the deepest valley in the free energy surface (FES), and the narrow width make sure we can capture precise FES as a function of the solvation structure. Meanwhile, the system can relax between two biased potentials, and the energy drift is less than 1 meV atom-1 ps-1. The representative solvation structures (e.g., FHS, CIP, and BIP) in the FESmaintained in the last 10 ps of the simulations (see Figure S2), which proves the AIMD metadynamics simulations were converged. The CN values collected in the metadynamics simulation were defined as \(CN=\sum_{l=1}^{N}\frac{1-(r_{l}/r_{o})^{p}}{1-(r_{l}/r_{o})^{q}}\), where \(p\)=12, \(q\)=24, \(r_{o}\) is the cutoff distance (i.e., 3.4 A), which corresponds to the first valley in the radial distribution function (RDF) from Ca\({}^{2+}\) to Cl\({}^{\cdot}\) ion and O in water within the AIMD simulations (see Figure S3a and Figure 2a); and \(r_{l}\) is the distance from Ca\({}^{2+}\) to the \(i^{\text{th}}\) Cl\({}^{\cdot}\) ion or O atom.
### Deep learning potential
The deep learning potential was constructed following the method proposed by Zhang et al.. The potential energy of a system can be decomposed into a sum of atomic energy contributions, and the atomic energy is fully determined by the coordinates of this atom and its neighboring atoms within a smooth cutoff radius. Explicitly, the deep learning potential was constructed in two steps, including preparing data and training the model. First, a large database composed of 50000 data sets has been prepared. One data set includes the atomic coordinates, corresponding atomic forces, and static energy. Twenty thousand coordinates were extracted from 100 ps metadynamics simulations at a time interval of 5 fs. The static force and energy were calculated on each coordinate with the same DFT method as the AIMD simulation using the Quickstep module in the CP2K package. The other 30000 data sets were from the last 20 ps of the three AIMD simulations on the local solvation structures at a time interval of 2 fs. Then, the whole database was passed to Tensorflow for deep neural network training using the DeepMD-kit package. At the end of the training, the root mean square energy and force errors over the whole testing set are 2.5\(\times\)10\({}^{-4}\) eV and 6.6\(\times\)10\({}^{-2}\) eV/A shown in Figure S4. We refer the reader to the recent work from Wang and E's group for the detailed theoretical discussion.
## 3 Results and Discussion
We employed metadynamics AIMD sampling for the CaCl\({}_{2}\) aqueous solution to explore the solvation environment composing of O in water and Cl\({}^{+}\) ion around Ca\({}^{2+}\) ion. Our analysis of the free energy landscapes revealed the presence of distinct local free energy minima, designated as 1, 2, and 3 in These minima correspond to specific configurations: the fully hydrated single Ca\({}^{2+}\) ion (FHS), Ca-Cl contact ion pair (CIP), and Ca-2Cl bridged ion pair (BIP). Although the FHS configuration has the lowest free energy, the mild free energy barrier (\(<\) 1 eV) for Ca\({}^{2+}\) to exchange through these free energy minima implies that both ion pair and fully dissociated configurations likely coexist in \(\sim\)0.85 M CaCl\({}_{2}\) aqueous solution. To gain a clearer understanding of the free energy profiles along these configurations, we analyzed the data along the three white dotted lines in Figure 1a, as illustrated in For the FHS configuration, the local free energy minima locate around 9 CN of O, which is consistent with the experimental hydration number through vapor pressure measurements. When one Cl\({}^{+}\) ion involves in the first solvation shell around the Ca\({}^{2+}\) ion, the local free energy minima is with \(\sim\)7 CN of O. Meanwhile, a plateau appears in the range of 5 to 7 CN of O, which is consistent with the observation in a recent work that multiple stable CN of O (i.e., 5 and 6) in the solvation shell of CIP can coexist. When two Cl\({}^{+}\) ions are in the first solvation shell of the Ca\({}^{2+}\) ion, a wider plateau exists from 2 to 6 CN of O. The energy wells and profiles of both CIP and BIP configurations are shallower and more dispersed, indicating less-defined solvation structures.
(a) Free energy surface contour plots of the CaCl\({}_{2}\) aqueous solution as a function of CN of Cl (\(x\)-axis) and O (\(y\)-axis) around the Ca\({}^{2+}\) ion from metadynamics AIMD simulations. (b) Free energy profiles as a function of CN of O around the Ca\({}^{2+}\) ion with 0, 1, and 2 CN of Cl\({}^{+}\) ions, corresponding to three white dotted lines in (a).
These findings underscore the diversity of solvation structures present in CaCl\({}_{2}\) aqueous solution and corroborate the theory of multiple stable solvation structures.
Following 25 ps of normal AIMD simulations without biased potentials, we found that the initial solvation structures (FHS, CIP, and BIP) remained stable across all three systems. This observation serves as a cautionary note for researchers regarding the selection of initial configurations in AIMD simulations. Within the limited simulation time, solvation structures can easily become trapped in metastable conditions. Therefore, careful consideration should be given to the initial setup of simulations. reveals that the first peak of \(g_{\text{Ca-O}}\) locates at \(\sim\)2.45 A in all three cases and becomes lower and narrower when there are more Cl\({}^{\cdot}\) ions in the first solvation shell around the Ca\({}^{2+}\) ion. A distinct second solvation shell can be observed at \(\sim\)4.6 A from the center Ca\({}^{2+}\) ion. Slight differences on the second peak may originate from the free Cl\({}^{\cdot}\) ion, which also affects the water distribution. A recent study found the local ordering introduced by the Na\({}^{+}\) ion in the first solvation shell does not extend further into the liquid since a highly consistent distribution occurs between \(g_{\text{O-O}}\) and \(g_{\text{Na-O}}\) beyond the first peak. However, the local ordering introduced by Ca\({}^{2+}\) ion can extend to a much longer distance, since \(g_{\text{O-O}}\) and \(g_{\text{Ca-O}}\) exhibit unique distribution at the second solvation shell. Similar long-distance distribution has also been observed around Mg\({}^{2+}\) ions in aqueous solution in our recent work. This observation highlights the multivalent cation has the capacity to lead a longer range ordering compared to the monovalent cation.
RDF from (a) Ca to O (\(g_{\text{Ca-O}}\)) and (b) Cl to O (\(g_{\text{Cl-O}}\)) in FHS, CIP, and BIP system. Magenta curves in (a-b) are the RDF among O atoms in bulk water (\(g_{\text{O-O}}\)). Scheme of the interference on the distribution of O around Ca and Cl ions in (c1) CIP system and (c2) BIP system. The red and blue circle denote the first peak of O around Ca and Cl in FHS system, respectively. The green circle represents the interference region.
The \(g_{\text{Cl-O}}\) shifts to a larger distance (see the arrows in Figure 2b) beyond the first peak at \(\sim\)3.3 A when there are more Cl' ions in the first solvation shell around the Ca\({}^{2+}\) ion. This can be understood through the interference between the water distribution around Ca\({}^{2+}\) and Cl' ion shown in When Ca\({}^{2+}\) and Cl' ion form the CIP configuration, the hydration shell will overlap to form an interference region shown in green circles in Since the first peak in \(g_{\text{Ca-O}}\) is much higher than that in \(g_{\text{Cl-O}}\), more free energy is needed to be overcome for a water molecule escaping from the Ca\({}^{2+}\) ion's hydration shell than Cl' ion's. In other words, the first hydration shell around the Ca\({}^{2+}\) ion is more rigid than that around the Cl' ion. Consequently, the interference will have a more significant effect on the distribution of water around the Cl' ion. For instance, in Figure 2c1, the positive interference within the green circle between the first hydration shell of the Ca\({}^{2+}\) ion and the Cl' ion results in a slight compression of the surrounding water molecules. In the BIP system, there is a positive interference between the Ca\({}^{2+}\) ion and the two attached Cl' ions. Additionally, given that the preferred \(\angle\)ClCaCl is approximately 90\({}^{\circ}\) (see Figure S3b), another interference occurs between two Cl' ions in the BIP system. Consequently, \(g_{\text{Cl-O}}\) can extend to a longer distance. This interference theory has already been successfully used to explain colloid and polymer interaction in inorganic salts and ionic liquids.
2D Angular-radial distribution functions of \(\angle\)XO\({}_{\text{c}}\)H\({}_{2}\) and XO\({}_{\text{c}}\). (X is Ca\({}^{2+}\) ion in (a) and Cl' ion in (b), O\({}_{\text{c}}\)H\({}_{2}\) represents the bisector of the water molecules, and O\({}_{\text{c}}\) represents O in water molecules).
The definition of angle and distance is also shown in the set of (a1) and (b1) for clarify. The color bar denotes the time averaged frequency.
The 2D angular-radial distribution function provides information on the preferred orientation of water dipoles near Ca\({}^{2+}\) and Cl' ion in different solvation shells. In our analysis, we calculated the angular distribution between O\({}_{\rm c}\)X and the bisector of the water molecules H\({}_{2}\)O\({}_{\rm c}\), as well as the radial distribution between X and O\({}_{\rm c}\). Here, X denotes an ionic center, and O\({}_{\rm c}\) represents the oxygen atom within a water molecule. From Figure 3(a1-a3), we can see the angular distribution of water molecules in the first solvation shell (\(r\)=2.0-3.0 A) around the Ca\({}^{2+}\) ion is very similar, which mainly locates around cos(\(\theta\)) = -1 (\(\theta\)=180\({}^{\circ}\)). The influence of the Ca\({}^{2+}\) ion on neighboring water molecules resembles that of a hydrogen bond donor in bulk water. For the water molecules in the second solvation shell (\(r\)=4.0-6.0 A) around the Ca\({}^{2+}\) ion, the angular distribution becomes broader. Moving from FHS to CIP and subsequently to BIP, we observe a decrease in the frequency of large angle distribution (-1.0\(<\)cos(\(\theta\))\(<\)0), while new features emerge in the distribution at small angles (0\(<\)cos(\(\theta\))\(<\)1.0). These changes primarily arise from the effects of the attached Cl' ion on the surrounding water molecules.
From Figure 3(b1-b3), we observe distinct features in the angular distribution of water molecules in the first solvation shell (\(r\)=3.0-5.0 A). For instance, the peak appears in the range of cos(\(\theta\)) = 1.0 to 0.5 in the FHS system. This corresponds to the value of \(\theta\) from 0 to 60\({}^{\circ}\), indicating that the influence of the single Cl' ion on the neighboring water molecules resembles that of a hydrogen bond acceptor in bulk water. This peak becomes less prominent, and another board preferential region appears around cos(\(\theta\)) =0 to -1.0 (see Figure 3b2-b3) when Cl' ion approaches the Ca\({}^{2+}\) ion to form an ion pair. This originates from the significant effects of the Ca\({}^{2+}\) ion on the angular distribution of the water molecules. Regarding the angular distribution of water molecules within the second solvation shell around the Cl' ion (\(r\)\(>\)5.0 A), it becomes smeared due to the combined effects of preferred orientation around the Ca\({}^{2+}\) ion and hydrogen bonding. Consequently, no clear preference for the tilt angle is observed for these water molecules.
### 3.2 Dynamics of ions and solvation shell
To overcome the time-consuming nature of AIMD simulations, MD simulations with the deep learning potential were conducted to investigate the dynamics in the CaCl\({}_{2}\) aqueous solution. The reliability of the MD simulations with deep learning potential has been thoroughly validated against AIMD simulations, encompassing various structural and dynamic parameters. Further details can be found in the supporting information section 3. To characterize the influence of ion pairing on the dynamics of water molecules, we calculate the residence correlation function of water molecules within the first solvation shell surrounding the Ca\({}^{2+}\) and Cl\({}^{\cdot}\) ions. The frequency of water exchange around ions is widely acknowledged as a crucial factor for the reactions involving these ions in aqueous solutions. The residence correlation function is defined as \(\langle cc(t)\rangle\). \(c(t)\) is equal to 1.0 if the water molecule within the solvation shell at \(t=0\) continuously resides in this shell until time \(t\). Otherwise, \(c(t)\) equals 0. By examining the time evolution of the residence correlation function illustrated in Figure 4, we observe that it becomes more challenging for water molecules to escape from the first solvation shell around the Ca\({}^{2+}\) ion as more Cl\({}^{\cdot}\) ions become involved in the solvation shell. Meanwhile, water molecules can escape more readily from the single Cl\({}^{\cdot}\) ion compared to the Cl\({}^{\cdot}\) ions involved in ion pairing.
By examining the trajectories of machine learning MD simulations in three distinct systems, we found both CIP and BIP systems transform to the FHS condition after tens of picoseconds. This transformation occurs because the FHS condition exhibits a lower free energy state compared to the other structures shown in While the CIP and BIP systems can only maintain their initial configurations for a brief period, our subsequent focus primarily centers on the dynamic properties observed in the FHS system. First, we calculated the time evolution of mean square displacement (MSD) of Ca\({}^{2+}\), Cl\({}^{\cdot}\), and water in the FHS utilizing a 500 ps equilibrated trajectory.
Residence correlation function for water in the first solvation shell around Ca\({}^{2+}\) and Cl\({}^{\cdot}\) ion in (a) FHS, (b) CIP, and (c) BIP system. The cutoff distance for water molecules in the first solvation shell corresponds to the first valley in \(g_{\text{Ca-O}}\) and \(g_{\text{Cl-O}}\) shown in Figure 2a-b.
The diffusion coefficients of Ca\({}^{2+}\), Cl\({}^{\cdot}\), and water in the MD simulation is 0.52 \(\times\) 10\({}^{9}\), 1.87 \(\times\) 10\({}^{9}\), and 2.28 \(\times\) 10\({}^{9}\) m\({}^{2}\)/s. Compared to experimental diffusion coefficients of Ca\({}^{2+}\) (0.66 \(\times\) 10\({}^{9}\) m\({}^{2}\)/s), Cl\({}^{\cdot}\) (1.59 \(\times\) 10\({}^{9}\) m\({}^{2}\)/s), and water (1.89 \(\times\) 10\({}^{9}\) m\({}^{2}\)/s), the differences are less than 22%. Given the possibility of the presence of a small fraction of ion pair conditions in the solution and the temperature disparity between the experimental setup (296.15-298.15 K) and our simulations (303.15 K), these discrepancies in the diffusion coefficients are considered negligible. This remarkable agreement underscores the effectiveness of MD simulations employing deep learning potential as a formidable tool for studying ion dynamics over extended trajectories, surpassing the traditional time scale limitations of AIMD (\(\sim\)100 ps).
Furthermore, it is crucial to acknowledge the significant influence of the solvation shell on ion dynamics. Initially, the motion of the ion remains unhindered by its solvation shell, resulting in a ballistic regime of ion diffusion (characterized by a slope of 2 in MSD).
(a) MSD curves of Ca\({}^{2+}\), Cl\({}^{\cdot}\), and water in the FHS system. (b) Slope of MSD of the Ca\({}^{2+}\) ion. (c) Histograms of the angle probability distribution as a function of the time interval \(\Delta\) for 0-0.5 ps of Ca ion in the FHS system. (d) Self-part van Hove correlation function 4\(\pi r^{2}G_{\mathrm{s}}(r)\) for the Ca ion in the FHS system.
Finally, as the solvation shell either moves along with the solvated ion or undergoes rupture due to the cleavage caused by the solvated ion, the ion diffusion enters the diffusive regime (manifested by a slope of 1 in MSD). The distinct progression through these three stages is aptly illustrated by the MSD plot of the Ca\({}^{2+}\) ion depicted in Notably, the transition between ballistic and subdiffusive regimes occurs at \(\sim\)0.1 ps. At \(\sim\)10 ps, the diffusion of the Ca\({}^{2+}\) ion experiences the transition between subdiffusive and diffusive regimes.
In order to gain further insights into the transition dynamics of the Ca\({}^{2+}\) ion within different regimes, we conducted additional calculations. Specifically, we investigated the angle probability distribution histograms and self-part van Hove correlation function \(4\pi r^{2}G_{\mathrm{s}}(r)\) for the Ca\({}^{2+}\) ion in the FHS system. To analyze the angle probability distribution, we calculated the ensemble average of the angle between two vectors as a function of the time interval. These vectors, denoted as \(V_{1}(t)\) and \(V_{2}(t)\), were obtained by evaluating \(X(t+\Delta)-X(t)\) and \(X(t+2\Delta)-X(t+\Delta)\), respectively, where \(X(t)\) represents the coordinate of the Ca\({}^{2+}\) ion at time \(t\), and \(\Delta\) represents the time interval. From Figure 5c, we can see that, initially, a prominent angle distribution emerges at a small angle (\(\sim\)25\({}^{\circ}\)). This finding suggests that the ion tends to move in a similar direction as the previous time step, with its motion mildly impeded by the solvation shell. The motion of the Ca\({}^{2+}\) ion is considered a forward ballistic motion. As the time interval increases, the angle distribution gradually expands, indicating a progressive hindrance of the Ca\({}^{2+}\) ion's motion by the solvation shell. After approximately 0.1 ps, the angle distribution stabilizes around 100\({}^{\circ}\), indicating that the ion's movement becomes obstructed by other atoms, causing a change in direction. Intriguingly, this time interval aligns with the transition observed between the ballistic and subdiffusion regimes, as depicted in the MSD curve of Figure 5ab.
Additionally, we examined the self-part of the van Hove correlation function, denoted as \(G_{\mathrm{s}}(r,t)\), for the Ca\({}^{2+}\) ion. This function measures the probability that a particle has moved to a distance \(r\) away from its initial position at time \(t=0\). Computed as \(G_{\mathrm{s}}(r,t)=\frac{1}{N}\sum_{i=1}^{N}(\delta[\mathbf{r}+\mathbf{r}_{i}( 0)-\mathbf{r}_{i}(t)])\), where \(\delta(\cdot)\) represents the delta function, \(\mathbf{r}_{i}\) denotes the position of ion \(i\), this function provides valuable insights into the ion's displacement behavior. illustrates the resulting \(4\pi r^{2}G_{\mathrm{s}}(r,t)\) for the Ca\({}^{2+}\) ions at t = 10 ps. Notably, the graph reveals a high probability of the Ca\({}^{2+}\) ion moving within a range of 0.1\(\sim\)0.15 nm. In fact, the ion can even traverse distances up to 0.25 nm from its initial position. Interestingly, this range corresponds to the first peak observed in the \(g_{\rm Ca\text{-}O}\) plot shown in Figure 2a, suggesting that the transition from the subdiffusive to the diffusive regime occurs when the motion of the Ca\({}^{2+}\) ion begins to approach the location of its original solvation shell. Overall, by examining both the angle probability distribution and the self-part van Hove correlation function, we have gained valuable insights into the transition dynamics of the Ca\({}^{2+}\) ion, shedding light on its behavior within different regimes.
## 4 Conclusions
Using a combination of AIMD and MD simulations with deep learning potentials, we conducted an extensive investigation into the solvation structure and ion dynamics within CaCl\({}_{2}\) aqueous electrolytes. Our findings shed light on key aspects of this complex system. Initially, our analysis revealed the coexistence of multiple metastable solvation structures around the Ca\({}^{2+}\) ion in approximately 0.85 M CaCl\({}_{2}\) aqueous solution. Interestingly, we observed that the translational and orientational distribution of water molecules within the solvation shell surrounding the Ca\({}^{2+}\) ion exhibited sensitivity to the ion pairing conditions, namely FHS, CIP, and BIP. Moreover, our study elucidated the crucial role played by ion pairing conditions in determining the lifetime of water molecules within the solvation shell. Specifically, we discovered that the presence of more Cl\({}^{\prime}\) ions within the solvation shell rendered it more challenging for water molecules to escape the immediate vicinity of the Ca\({}^{2+}\) ion. Importantly, our investigation underscored the effectiveness of machine learning MD simulations as a powerful approach for studying ion dynamics in a highly efficient manner. Notably, the self-diffusivity of ions and water molecules calculated through our MD simulations yielded results comparable to experimental measurements, highlighting the fidelity of our approach. To gain a comprehensive understanding of the transition dynamics of the Ca\({}^{2+}\) ion, we employed angle probability distribution histograms and van Hove correlation functions. Through these analyses, we made notable observations. The transition from ballistic to subdiffusive motion occurred when the forward motion of the ion became impeded, while the transition from subdiffusive to diffusive motion manifested when the ion's movement approached the location of its original solvation shell. In short, our study provides valuable insights into the solvation structure and ion dynamics within CaCl\({}_{2}\) aqueous electrolytes. By combining AIMD and MD simulations with deep learning potentials, we have unraveled the coexistence of solvation structures, the influence of ion pairing conditions on water molecule lifetimes, and the transitions in Ca\({}^{2+}\) ion dynamics. Our findings highlight the efficacy of machine learning MD simulations as a robust tool for investigating complex ion dynamics.
|
10.48550/arXiv.2307.09565
|
Solvation Structures and Ion Dynamics of CaCl$_2$ Aqueous Electrolytes Using Metadynamics and Machine Learning Molecular Dynamics Simulations
|
Zhou Yu, Lei Cheng
| 3,409
|
10.48550_arXiv.2404.09078
|
###### Abstract
We develop a static quantum embedding scheme that utilizes different levels of approximations to coupled cluster (CC) theory for an active fragment region and its environment. To reduce the computational cost, we solve the local fragment problem using a high-level CC method and address the environment problem with a lower-level Moller-Plesset (MP) perturbative method. This embedding approach inherits many conceptual developments from the hybrid MP2 and CC works by Nooijen and Sherrill (J. Chem. Phys. 111, 10815, J. Chem. Phys. 122, 234110). We go beyond those works here by primarily targeting a specific localized fragment of a molecule and also introducing an alternative mechanism to relax the environment within this framework. We will call this approach MP-CC. We demonstrate the effectiveness of MP-CC on several potential energy curves, and a set of thermochemical reaction energies, using CC with singles and doubles as the fragment solver, and MP2-like treatments of the environment. The results are substantially improved by the inclusion of orbital relaxation in the environment. Using localized bonds as the active fragment, we also report results for N=N bond breaking in azomethane and for the central C\(-\)C bond torsion in butadiene. We find that when the fragment Hilbert space size remains fixed (e.g., when determined by an intrinsic atomic orbital approach), the method achieves comparable accuracy with both a small and a large basis set. Additionally, our results indicate that increasing the fragment Hilbert space size systematically enhances the accuracy of observables, approaching the precision of the full CC solver.
## I Introduction
Coupled cluster (CC) theory provides the best balance between efficiency and accuracy for calculating the electronic structure of weak to moderately correlated molecules. However, CC theory with even the simplest singles and doubles truncation, still has computational cost scaling as \(\mathcal{O}(O^{2}V^{4})\), where \(O\) and \(V\) are the numbers of occupied and virtual orbitals being correlated. This precludes calculations on large molecular systems using large basis sets. At a lower level of accuracy, such problems are often addressed with composite methods. There is a long history of such methods in quantum chemistry, including the Gaussian-\(n\) methods (G1-G4), the Weizmann-\(n\) methods (W1-W4), the HEAT protocol, the correlation consistent Composite Approach (ccCA), etc. As a simple example, the basis set correction to an expensive coupled cluster CC calculation such as CCSD(T) might be approximated using a far less demanding approach such as second-order Moller-Plesset theory (MP2) that exhibits similar basis set dependence via:
\[E_{\text{Composite}}\approx E_{\text{CC/small}}+\left(E_{\text{MP2/large}}-E_{ \text{MP2/small}}\right). \tag{1}\]
In a similar additive fashion, the ONIOM model treats different parts of a system with different levels of theory to tackle the system size scaling.
Local correlation methods are a more advanced treatment of the system-size scaling that exploits the locality of electron correlations to approach linear scaling of compute effort with size. In contrast, our goal is to approximate the total energy using a high-level treatment for a specific local region of a big molecule and to control the computational expenses of such calculations by employing a lower-level approximation for the environment. This "QM-in-QM" approach is useful for addressing the chemistry of an active region of a catalyst, or point defects in materials, for example. Or, when the active fragment spans all the atoms in the molecule, a similar approach can be used to address the basis set convergence of the total problem. This will serve as an alternative method to overcome the bottleneck of large basis set calculations with CC, for which many methods have been already proposed, such as explicitly correlated R12/F12 methods, diagrammatic separation-based extrapolation, and cardinal number based extrapolation, among many others.
Our effort to address the challenges of large system size and large basis set calculations can also be viewed from a quantum embedding perspective. In quantum embedding, one identifies an important (preferably local) region of the system, which is denoted as the fragment (F). The remainder of the system is referred to as the environment (E) and is treated by a low-scaling method, such as Hartree-Fock, density functional theory, or a perturbative method. For the fragment problem, one first evaluates a coupling potential (\(\Delta\)) between the environment and the fragment. This coupling potential is then incorporated into the fragment Hamiltonian treating the fragment problem as an open-quantum system. The fragment problem is solved with a high-accuracy method (typically called a solver). CC methods have proven to be successful as a solverfor many embedding methods, especially when the fragment size needs to be large. This highly accurate solution on the fragment is subsequently used to self-consistently update \(\Delta\). The fragment problem will subsequently be described by a renormalized interaction, and solved by a CC method truncated at a certain excitation rank. The environment problem is then treated by a lower-scaling method such as MP2 theory. The cost of the fragment problem remains constant with the increase of system size/basis set size and thus reduces the overall cost.
Here we combine the fragment and the environment problems using two sets of coupled projective equations: one for the environment and one for the fragment. Furthermore, we invoke a self-consistency condition between the fragment and the environment based on the "global" CC amplitudes. This approach inherits many essential ideas from the hybrid MP2 and CCSD methods developed by Nooijen and Sherrill et al. Both Nooijen and Sherrill _et al._ presented two versions of the theory. Their simplest version did not relax orbitals in the environment; however, in an improved version such contributions were included. In that version, certain classes of environment amplitudes (external and semi-internal) were updated using the same amplitude equations as CCSD. These amplitudes were chosen such that the computational cost remains low. This idea has its origin in earlier work by Adamowicz, Piecuch, and coworkers published between 1993 and 1998, which introduced partitioning of the cluster operator into internal and external components.
Our approach differs from Nooijen and Sherrill _et al._'s approach in two crucial aspects. First, instead of choosing canonical molecular orbital (CMOs) as a basis for the active space, we choose localized MO (LMO)s, which lets us define the active fragment as a bond or region. Second, the relaxation of the environment orbitals is performed differently (vide infra) with the objective in mind that it does not increase the cost of the low-level method even when we increase the excitation rank of the high-level CC method. Another related approach is the multi-level CC (MLCC) theory of Koch and co-workers. This approach also utilizes a localized active space and also defines a mechanism for relaxing the environment. Differences between our approach and the MLCC models in terms of the general framework and computational implications will be elaborated in Sec. IV and Appendix A.
The remainder of the manuscript is organized as follows. In Section II, we describe the general framework, then outline various choices for the environmental projection equation, and finally discuss some theoretical aspects of the fragment amplitude equations. We will use CCSD as the fragment solver, although this choice can be extended to higher-level CC models. As in all embedding methods, it is important to choose a suitable local basis that describes the fragment problem well, as the results are not invariant with respect to this choice. Our choice of localized representation is described in Section III. To follow up on the review of existing embedding methods given above, in Section IV we compare some formal properties of our current method with related methods. Implementation details are briefly summarized in Section V. Finally, in Section VI we will illustrate various aspects of the theory via a set of prototypical numerical examples. We will compare different versions of our embedding theory relative to the full high-level method, and a composite approach, which can be regarded as an uncoupled version of the embedding method.
## II Embedded coupled cluster formalism
Coupled cluster theory is a wave function approach that expresses the ground-state wave function using an exponential parametrization. Given a single Slater determinant reference state like the Hartree-Fock (HF) determinant \(|\Phi_{0}\rangle\), any intermediately normalized state \(|\Psi\rangle\) obeying \(\langle\Phi_{0}|\Psi\rangle=1\) can be uniquely written as \(e^{T}|\Phi_{0}\rangle\), where \(T\) is the corresponding cluster operator. To define the cluster operators, we introduce the notation that \(i\), \(j,k,l\)... label the occupied orbitals and \(a\), \(b\), \(c\), \(d\),... the virtual orbitals.
\[T=\sum_{\mu\in\mathcal{I}}t_{\mu}X_{\mu}=\sum_{k=1}^{N}\sum_{ \begin{subarray}{c}\mu\in\mathcal{I}\\ |\mu|=k\end{subarray}}t_{\mu}X_{\mu}, \tag{2}\]
where
\[\mathcal{I}=\left\{\begin{pmatrix}a_{1},...,a_{k}\\ i_{1},...,i_{k}\end{pmatrix}\ :\ 1\leq k\leq N\right\} \tag{3}\]
The CC amplitudes, denoted by \(\mathbf{t}=(t_{\mu})_{\mu\in\mathcal{I}}\), are determined through projective equations:
\[\langle\Phi_{\mu}|\overline{H}|\Phi_{0}\rangle=0\qquad\forall\mu\in\mathcal{I}, \tag{4}\]
where
\[\overline{H}=e^{-T}He^{T} \tag{5}\]
The corresponding CC energy is then given by
\[E_{\text{CC}}=\langle\Phi_{0}|\overline{H}|\Phi_{0}\rangle. \tag{6}\]
Since the untruncated CC equations rapidly become numerically intractable, truncations are commonly employed. The subject of this work is the CCSD variant, where \(T\) comprises excitation operators up to a maximum excitation rank of two, i.e.,
\[T=\sum_{i,a}t_{i}^{a}X_{i}^{a}+\frac{1}{4}\sum_{i,j,a,b}t_{ij}^{ ab}X_{ij}^{ab}. \tag{7}\]
However, even CCSD has computational effort scaling as \(\mathcal{O}(Q^{2}V^{4})\), where \(O\) and \(V\) denote the number of occupied and virtual orbitals included in the correlation treatment. Even CCSD becomes computationally demanding when naively applied to larger systems.
Here we explore an approximation to CCSD that is inspired by quantum embedding theory. The underlying core concept is well-known: it is the recognition that certain CC amplitudes possess greater physical significance than others if expressed in a suitable basis, such as localized or natural orbitals. The proposed approach aims to identify and accurately evaluate these critical amplitudes, while the less significant amplitudes are approximated via many-body perturbation theory with reduced precision (vide infra). We therefore name this approach the many-body perturbation coupled-cluster (MP-CC) method. It aims to maintain high accuracy while simultaneously reducing overall computational complexity.
Underlying the MP-CC approach is a partition of the total orbital space into two sets: the active fragment (F) orbitals and the environment (E) orbitals. The fragment orbitals describe the chemically relevant region that requires higher accuracy. The fragment orbitals are formed from intrinsic atomic orbitals (IAOs) which span only a small chemically relevant subspace of the full basis set of the calculation. The environment orbitals, on the other hand, will describe additional virtual orbitals associated with the full atomic orbital basis, and all occupied orbitals not contained in the fragment orbital space. This allows the use of all valence occupied and some virtual orbitals as the fragment. Alternatively, the fragment can be just a single local site of particular interest (e.g. a catalytic active site), or a set of local sites.
A valid partition splits the occupied and virtual orbitals into environment and fragment subsets, such that
\[O=O_{\mathrm{E}}+O_{\mathrm{F}}\quad\text{and}\quad V=V_{\mathrm{E}}+V_{ \mathrm{F}}. \tag{8}\]
Once a partition is made, there are also two classes of cluster operators: \(T_{\mathrm{F}}\), and \(T_{\mathrm{E}}\). Note that the \(T_{\mathrm{E}}\) operators contain amplitudes that involve only environment orbitals, as well as mixed amplitudes that contain both fragment and environment orbitals. The latter describes correlations between electrons in fragment and environment occupied levels, or between two occupied fragment levels using environment virtuals, etc.
In the spirit of quantum embedding, we will compute the cluster operators \(T_{\mathrm{F}}\), and \(T_{\mathrm{E}}\) using different levels of sophistication. More precisely, we propose to determine the fragment cluster operator \(T^{\mathrm{F}}\) using the standard CC formalism (the high-level theory), involving a full expansion of \(\overline{H}\) (see Eq.) to yield \(\overline{H}_{\mathrm{F}}^{\mathrm{high}}=\overline{H}\). In contrast, the calculation of \(T_{\mathrm{E}}\) is based on a lower-order perturbative expansion of \(\overline{H}\), resulting in \(\overline{H}_{\mathrm{E}}^{\mathrm{low}}\) (vide infra). This lower-order expansion offers various options for gaining a computational advantage, as will be elaborated in Sec. II.1. The MP-CC approach results in the following set of coupled projective equations:
\[\langle\Phi_{\mu}^{\mathrm{F}}|\overline{H}_{\mathrm{F}}^{\mathrm{ high}}|\Phi_{0}\rangle =0, \tag{9}\] \[\langle\Phi_{\mu}^{\mathrm{E}}|\overline{H}_{\mathrm{E}}^{\mathrm{ low}}|\Phi_{0}\rangle =0. \tag{10}\]
We emphasize that both \(\overline{H}_{\mathrm{F}}^{\mathrm{high}}\) and \(\overline{H}_{\mathrm{E}}^{\mathrm{low}}\) in Eqs. and contain the full set of \(T\)-amplitudes i.e., \(\mathbf{t}_{\mathrm{F}}\) and \(\mathbf{t}_{\mathrm{E}}\). In other words, although Eq. solves only for \(\mathbf{t}_{\mathrm{F}}\), the \(\mathbf{t}_{\mathrm{E}}\) amplitudes still enter the equations. Similarly, Eq. solves for \(\mathbf{t}_{\mathrm{E}}\) in the presence of \(\mathbf{t}_{\mathrm{F}}\). Therefore, Eqs. and should be solved self-consistently, unless they are uncoupled by the perturbative approximations. The total CC energy remains as Eq..
Moreover, the energy and amplitude equations of the MP-CC approach can be combined into a Lagrangian:
\[\begin{split}\mathcal{L}(\mathbf{t},\boldsymbol{\lambda})& =\langle\Phi_{0}|\overline{H}|\Phi_{0}\rangle+\langle\Phi_{0}| \Lambda_{\mathrm{E}}\overline{H}_{\mathrm{E}}^{\mathrm{low}}|\Phi_{0}\rangle \\ &+\langle\Phi_{0}|\Lambda_{\mathrm{E}}\overline{H}_{\mathrm{F}}^{ \mathrm{high}}|\Phi_{0}\rangle,\end{split} \tag{11}\]
and are satisfied; \(\Lambda_{\mathrm{E}}\) and \(\Lambda_{\mathrm{F}}\) are defined in accordance with \(T_{\mathrm{E}}\) and \(T_{\mathrm{F}}\). The working equations then correspond to a first-order optimality condition of the Lagrangian, i.e.,
\[\begin{cases}\dfrac{\partial\mathcal{L}}{\partial\mathbf{t}_{\alpha}}=0,& \alpha\in\{\mathrm{E},\mathrm{F}\}\\ \dfrac{\partial\mathcal{L}}{\partial\boldsymbol{\lambda}_{\alpha}}=0,& \alpha\in\{\mathrm{E},\mathrm{F}\}\,.\end{cases} \tag{12}\]
Compared to other embedding approaches, the Lagrangian in Eq. enables us to access a broad range of observables beyond the energy, by employing analytic energy derivative techniques.
### Environment amplitude equations
We employ two distinct approaches to the lower-order expansion \(\overline{H}_{\mathrm{E}}^{\mathrm{low}}\), both based on perturbative arguments.
#### ii.1.1 Unrelaxed approach
In the first approach, we use the simplest possible MP2-like projection equations for the environment amplitudes, i.e.,
\[\langle\Phi_{\mu}^{\mathrm{E}}|V+[F,T]|\Phi_{0}\rangle=0. \tag{13}\]
While \(T\) contains all cluster amplitudes, Eqs. are used to determine only \(T_{\mathrm{E}}\). In defining these equations, we have used the partitioned Hamiltonian: \(H=F+V\), where \(F\) is the mean-field Hamiltonian or Fock operator from HF theory and \(V\) is the two-particle fluctuation potential, expressed in the HF basis. Note that we use normal-ordered operators for \(F\) and \(V\), though we do not denote it explicitly. By the Brillouin theorem, \(\langle\Phi_{s}^{\mathrm{E}}|V|\Phi_{0}\rangle=0\) for all singles, \(|\Phi_{s}^{\mathrm{E}}\rangle\) so the environment orbitals are unrelaxed by singles amplitudes. We therefore refer to Eqs. 13 as the unrelaxed MP-CCSD method. Denoting fragment singles as \(S_{\mathrm{F}}\), fragment doubles as \(T_{\mathrm{F}}\) and environment doubles as \(T_{\mathrm{E}}^{}\), this is a wavefunction ansatz of the form:
\[|\Psi_{\mathrm{MP-CC(unrelaxed)}}\rangle=e^{S_{\mathrm{F}}}e^{T_{\mathrm{E}}^{}+T_{\mathrm{F}}}|\Phi_{0}\rangle. \tag{14}\]
The unrelaxed method yields fixed environment amplitudes that influence the fragment amplitudes but not vice-versa.
#### ii.1.2 Relaxed approach
In the second approach, we relax the wave function in the sense of Thouless relaxation, wherein the HF determinant \(\Phi_{0}\) can be transformed into another determinant \(\Phi_{0}^{\prime}\) by the exponential of a single substitution operator, denoted as \(S\): \(\Phi_{0}^{\prime}\propto e^{S}\Phi_{0}\). Higher-body amplitudes describing correlation, such as doubles will still be considered only to first order.
\[|\Psi\rangle=e^{S}e^{T_{\rm E}^{}+T_{\rm F}}|\Phi_{0}\rangle. \tag{15}\]
More precisely, we will assign the formal perturbation parameter \(\eta\) to the interaction term of the similarity transformed Hamiltonian: \(\tilde{H}=e^{-S}He^{S}\), that is, \(\tilde{H}=E_{cl}+\tilde{F}^{}+\eta\tilde{V}^{}\). Here, \(\tilde{F}\) and \(\tilde{V}\) are the one-particle and two-particle parts of the transformed Hamiltonian, respectively, and \(E_{cl}\) is a closed scalar contribution, which will not contribute to the projective equations. Keeping the terms up to first order in the perturbation, \(\eta\), yields the following projective equations for the singles (s) and doubles (d) amplitudes:
\[\langle\Phi_{s}^{\rm E}|\tilde{F}+[\tilde{F},T_{\rm E}^{}]+[ \tilde{F},T_{\rm F}]|\Phi_{0}\rangle =0, \tag{16}\] \[\langle\Phi_{d}^{\rm E}|\tilde{V}+[\tilde{F},T_{\rm E}^{}]+[ \tilde{F},T_{\rm F}]|\Phi_{0}\rangle =0. \tag{17}\]
In this work, \(T_{\rm E}^{}\) contains the first-order environment doubles amplitudes, while the fragment doubles amplitudes are in \(T_{\rm F}\). In the derivation of Eq. and we used the fact that the similarity transformation with \(e^{S}\) does not increase the operator rank of the Hamiltonian. By comparing Eqs. and with Eq., we obtain the definition of \(\overline{H}_{\rm E}^{\rm low}\).
In Fig. 1, we schematically show the workflow of the proposed method (\(W_{\rm F}\) is defined in II.2, and not relevant for the following discussion). Note that when employing the relaxed method (i.e., the second approach outlined above) in step 3 (see Fig. 1), the first iteration uses \(S=0\) implying that Eq. and reduce to Eq.. In particular, in the first iteration, the two procedures outlined above yield the same inactive amplitudes. However, in the subsequent iterations, non-zero active singles amplitudes generated from the \(T_{\rm F}\)-amplitude equations will contribute to Eq. and, thus yielding non-zero inactive singles amplitudes.
The computational scaling of Eq. and is MP2-like, using a factorized two-body Hamiltonian, as only low-rank singles amplitudes contract with them. The factorization of the two-body Hamiltonian can be carried out using either the density fitting (DF) approach or a cholesky decomposition (CD) approach. Following Mester _et. al._, this yields \(\mathcal{O}(O^{2}V^{2}X)\) computational scaling for our relaxed approach, where \(X\) is the dimension of the auxiliary basis used in the DF or CD scheme.
### Fragment amplitude equations
An important observation is that in the fragment amplitude equations, the Hamiltonian \(\overline{H}_{\rm F}^{\rm high}\) is screened by the environment amplitudes \(\mathbf{t}^{\rm E}\). We may rewrite Eq.
\[\begin{split} W_{\rm F}&=[e^{-T_{\rm E}}He^{T_{ \rm E}}]_{\rm F}\\ &=W_{1{\rm B},{\rm F}}+W_{2{\rm B},{\rm F}}+W_{3{\rm B},{\rm F}}+...\end{split} \tag{19}\]
In Appendix B we show the working equations arising from Eq..
In Eq., we renormalize the bare Hamiltonian by applying a similarity transformation via \(e^{T_{\rm E}}\) and then restrict the orbital indices to the fragment. Note that the similarity transformation increases the rank of \(W_{\rm F}\) compared to the bare Hamiltonian. However, when solving Eq., we order the operation of the tensors such that we avoid generating higher rank tensors of \(W_{\rm F}\). This particular construction is conceptually important because it is the mechanism by which the low-energy fragment (or active region) Hamiltonian is screened by the "high-energy" environment amplitudes, which is a crucial requirement for accurate quantum embedding theories. We note that in a similar manner, Kowalski _et. al._ constructed a CC downfolded Hamiltonian by using sub-system embedding sub-algebras (SES). In Fig. 1, we schematically show that in step 4 and step 5, we first construct \(W_{\rm F}\), and then iteratively solve Eq.. The maximum complexity of solving Eq. for the fragment amplitudes is \(\mathcal{O}(O_{\rm F}^{2}V_{\rm F}^{4})\). If there are multiple (disjoint) fragments, it also scales linearly with the number of fragments. The construction of \(W_{\rm F}\), on the other hand, is more computationally demanding, and scales as \(\mathcal{O}(O^{2}O_{\rm F}V^{2}V_{\rm F})\); however, this cost is non-iterative.
Schematic view of the current embedding scheme.
## III Choice of basis set for active space
The choice of a suitable basis to define the active space is crucial for embedding methods. The dimensionality and accuracy of the fragment impurity problem are dependent on this choice, because the basis determines spatial locality on the one hand, and the most suitable virtual orbitals to capture the main electron correlation effects in the fragment region on the other hand. To capture spatial locality, localized orbitals are of course essential. To capture the most relevant virtuals for electron correlation, without performing an active space calculation, we adopt the Atomic Valence Active Spaces (AVAS) approach. AVAS straightforwardly defines valence occupied and virtual orbitals for a fragment, and is trivially able to incorporate higher angular momentum functions into the virtual space as discussed below. AVAS has important advantages over simply using canonical MOs, as illustrated in Appendix C.
The objective of AVAS is to isolate a set of valence/semi-valence atomic orbitals (\(A\)) denoted by \(p,q\), which has overlap with the HF occupied (\(|i\rangle\)) and virtual orbitals (\(|a\rangle\)) of the full molecule.
To this end, we define the projection operator
\[P_{A}=\sum_{p,q\in A}[\sigma^{-1}]_{pq}|p\rangle\langle q|\quad\text{where} \quad\sigma_{pq}=\langle p|q\rangle. \tag{20}\]
With this projection operator, we then construct two overlap matrices
\[[S^{\rm q}]_{ij}=\langle i|P_{A}|j\rangle\quad\text{and}\quad[S^{\rm q}]_{ab}= \langle a|P_{A}|b\rangle \tag{21}\]
To isolate the active orbitals, \([S^{\rm q}]_{ij}\) and \([S^{\rm q}]_{ab}\) are diagonalized and the orbitals \(|i\rangle\) and \(|a\rangle\) are rotated by the diagonalizing matrices to generate two new sets of orbitals: \(|\hat{i}\rangle\) and \(|\hat{a}\rangle\). We then choose the active orbitals to be the vectors in \(|\hat{i}\rangle\) and \(|\hat{a}\rangle\) corresponding to non-zero eigenvalues. In the construction of \([S^{\rm q}]_{ij}\) matrix, we freeze the core occupied orbitals, which will then be added to the inactive (environment) orbitals.
The atomic orbitals defining the fragment must be specified before the above procedure can be invoked. Commonly, a minimal atomic orbital basis (e.g., such as MINAO or autogenerated via AutoSAD) is employed for constructing \(A\). An important characteristic of the AVAS scheme is that one can alternatively employ a slightly larger reference basis set so that the active space includes orbitals of higher angular momentum. This aspect is important, for instance, in first-row transition metals, where often "double-shell" or 4d orbitals are required. We will project the AVAS reference basis such that it exactly spans an integer number of fragment occupied and virtual orbitals equal to its dimension in our system. This restriction is especially crucial for the occupied space.
## IV Comparison with existing approaches
In the hybrid MP2 and CCSD approaches developed by Nooijen and Sherrill et al., two very similar unrelaxed schemes were introduced, namely, A-CC/PT and MP2-CCSD(I). Our unrelaxed scheme, as defined in Sec. II.1.1, is identical to those approaches for a given choice of fragment orbitals. However, in contrast to the canonical molecular orbitals (CMOs) chosen in their implementations, we choose localized molecular orbitals (LMOs) from the AVAS scheme (Sec. III) to construct the active space, which has the advantage of treating a spatially localized fragment. Furthermore, both Nooijen and Sherrill _et al._ defined two improved versions of their methods, namely R-CC/PT and MP2-CCSD(II), respectively by including selected sets of external and semi-internal singles and semi-internal doubles amplitudes. In our approach, on the other hand, we relax the whole set of environment amplitudes (which corresponds to external plus semi-internal amplitudes) with a low-level method, leading to Eq. and.
In the Multi-Level CC (MLCC) work by Koch and co-workers, a CC2-like approach was adopted to define the inactive amplitude equations. The CC2-like equations incorporate \(T_{1}\) amplitude equations to all orders while treating the \(T_{2}\) amplitude equations only up to the first order terms in conjunction with a second-order \([V,T^{F}]\) term. By contrast, our approach consistently applies a first-order perturbative treatment to both the \(T_{1}\) and \(T_{2}\) amplitude equations to determine inactive environment CC amplitudes. Similar to the MLCC approach, the computational complexity of the inactive MP-CC amplitude equations is \(\mathcal{O}(O^{2}V^{2}X)\) at the singles and doubles level. However, the future inclusion of \(T_{3}\) amplitudes in the MP-CC framework - where active non-zero \(T_{3}\) amplitudes induce further relaxation of the environment orbitals - will not increase the complexity of the low-level projection equation.
In contrast, with the MLCC approach, the complexity of the low-level projection equations would have been higher. However, Koch _et al._ approximate the triple excitations to include only active indices to reduce the complexity. For a more detailed discussion on the differences between MLCC and MP-CC, we refer the interested reader to Appendix A. We have summarized all the comparisons in Table 1. In that table we use uppercase letters (\(I,J,K,L,...\) for holes; \(A,B,C,D,...\) for particles) to denote orbitals in the fragment, and lowercase letters (\(i,j,k,l,...\) for holes; \(a,b,c,d,...\) for particles) for orbitals in the environment. A side-by-side numerical comparison of MP2-CCSD(II), MLCC and MP-CC is left for future work.
## V Implementation
The computational results are obtained using a pilot implementation of the MP-CC method at the singles-doubles truncated level, that is MP-CCSD, based on the Python-based Simulations of Chemistry Framework (PySCF). In this implementation, the amplitude equations are separated into different classes for the fragment and environment, allowing us to experiment with the different heuristics mentioned in Sec. II.1. We conduct all-electron calculations using CCSD as the computational method in both spin-restricted and spin-unrestricted formulations. For the spin-unrestricted version of MP-CCSD, unrestricted Hartree-Fock (UHF) orbitals willbe used. Since all calculations employ localized orbitals, the MP2 equations, as in Eq., are of the non-canonical type.
We want to comment on a few aspects of the implementation of steps- in We solve the projection equations in both Step 5 and Step 6 iteratively. After solving Step 6 we rebuild \(W^{F}\), and solve Step 5 and 6 again, thereby introducing a global outer loop. The entire set of equations are then solved in a double-loop fashion. Similar to the inner loop that involves step 5 and step 6, we employ the DIIS scheme for the outer loop as well. The number of iterations in the outer loop is usually much smaller than that in the inner loop.
## VI Results
### Diatomic potential energy curves
We apply our embedding method to the potential energy curves (PECs) of N\({}_{2}\) and CO to study the following aspects of the theory:
1. To assess the accuracy of different approaches, we analyze the PECs with different approaches, namely MP2, both relaxed and unrelaxed versions of the MP-CCSD method and also a composite method. For the composite method, we modify Eq. to make it an uncoupled approximation to the current embedding scheme: \[\begin{split} E_{\text{Composite}}&=E_{\text{MP2, large}}+E_{\text{CCSD,small(fc)}}\\ &\quad\quad-E_{\text{MP2,small(fc)}},\end{split}\] where large and small stand for large, small AO basis sets considered for the current calculation, and fc stands for the frozen core calculation (only if it is used). We compare the error in energy with respect to the CCSD method, using two metrics: the maximum absolute error (MAE) and the non-parallelity error (NPE). The spin unrestricted energy at dissociation will be the zero of energy. We furthermore evaluate several spectroscopic parameters, namely, equilibrium bond distance (\(R_{e}\)), harmonic frequency (\(\omega_{c}\)), and first anharmonicity constant (\(\omega_{c}\),\(c_{e}\)) by the least square fit of the PEC with a polynomial with high enough degree such that the parameter values converge.
2. We investigate the method's convergence as we systematically increase the size of the active space aiming to assess systematic improvability.
3. We show consistent behavior of the MP-CCSD method with increased AO basis set size when we keep the active space fixed. This observation is significant, considering the challenges often encountered in achieving similar enhancements within many quantum embedding schemes, for example, with Schmidt decomposition-based DMET and Bootstrap Embedding (BE)-DMET methods.
We will employ three different choices for the small basis (MINAO) used for IAO construction, giving three different active spaces as summarized in Table 2. Note that we employ unrestricted orbitals that can break the \(S^{2}\) symmetry of the electronic Hamiltonian, while allowing us to reach lower variational energies than with restricted orbitals. Notably, UHF shows the correct qualitative behavior in the dissociation limit.
#### vi.1.1 The nitrogen molecule (N\({}_{2}\))
The PEC of N\({}_{2}\) often serves as a benchmark problem in electronic structure theory due to the complexity involved in the dissociation of its triple bond. UCCSD for the whole system will serve as the reference method by which we test the embedded MP-CCSD approximations, employing the aug-cc-pCVDZ and aug-cc-pCVTZ basis sets. While not strictly relevant for assessing embedding approaches, we note that UCCSD is not very accurate relative to a higher rank CC method, UCCSDT. While UCCSD shows correct dissociative behavior, its MAE and NPE versus UCCSDT are 29 mH and 20 mH, respectively with the aug-cc-pCVDZ basis set.
We observe that the PECs obtained from various methods are not of similar quality. This is especially true close to the Coulson-Fischer point at 1.2A where
\begin{table}
\begin{tabular}{c|c|c|c c} & name & active orbitals & MINAO & (\(n_{e},n_{\text{orb}}\)) \\ \hline A & minimal & 1s 2s 2p & STO-3G & \\ B & split valence (SV) & 1s 2s 2p 3s 3p & SV & \\ C & SV + polarization & 1s 2s 2p 3s 3p 3d & cc-pVDZ & \\ \end{tabular}
\end{table}
Table 2: Choices of active space for the N\({}_{2}\) and CO diatomic molecules, and the small basis (denoted as MINAO) used for IAO construction.
\begin{table}
\begin{tabular}{c c c c c} method & MP-CCSD & R-CC/PT & MP2-CCSD(II) & MLCCSD \\ & (Relaxed) & & & \\ \hline reference & this work & 28 & 30 & 33 \\ \hline \(T_{I}^{\mu}+T_{J}^{\mu}\) & All & All & All & All \\ \hline Classes of & All & All & All & All \\ \(T_{I}^{\text{E}}\) relaxed & & & & \\ \hline Classes of & All & \(\{T_{J}^{\mu B},T_{J}^{\mu B}\}\) & \(\{T_{J}^{\mu B},T_{J}^{\mu B}\}\) & All \\ \(T_{J}^{\text{E}}\) relaxed & & \(\{T_{J}^{\mu B}\}\) & & \\ \hline Active space & AVAS & CMO & CMO & CD \\ basis & & & & \\ \hline Relaxation & Sec. II.1 & CCSD & CCSD & CC2 \\ method & & & & \\ \hline Scaling of & & & & \\ relaxation & & & & \\ method & & & & \\ \end{tabular}
\end{table}
Table 1: Comparison of the MP-CCSD embedding scheme with three other related approaches. Note the relaxation method scaling cost is per iteration, and CD stands for Cholesky decomposition based localized orbitals.
UMP2 and the composite method show a first derivative discontinuity. Its origin lies in the fact that UMP2 is incapable of removing UHF's spin contamination error. Towards the dissociative limit, all methods are qualitatively satisfactory, due to using unrestricted orbitals.
To focus on the relative errors, plots the differences between each approximation and the UCCSD reference, as well as showing embedding results in the minimal active space A. shows that UMP2 errors peak near the Coulson-Fischer point and its derivative exhibits discontinuity. The UMP-CCSD method reduces the MAE from 80 mH (UMP2) to 38 mH / 30 mH, and the NPE from 78 mH (UMP2) to 24 mH / 17 mH for A / B choice of active spaces, respectively. Similarly, the composite scheme reduces MAE and NPE to 24 mH and 24 mH, respectively. However, the composite method contains _two_ derivative discontinuities which result from the different Coulson-Fischer points in the small and large basis sets, and is clearly undesirable! The unrelaxed variant of UMP-CCSD shows a derivative discontinuity at the same bond distance as UMP2. Very encouragingly, this discontinuity is lifted using the relaxed version of UMP-CCSD, showing the importance of the _environment singles amplitudes_ in symmetry restoration.
To further improve the UMP-CCSD energies, we augment the active space with N 3d orbitals (the SVP active space C). We report the relaxed and unrelaxed UMP-CCSD results in aug-cc-pCVDZ (DZ) and aug-cc-pCVTZ (TZ) basis set for active space choices B and C in Comparing the left and right panels of shows that MAE and NPE have significantly improved in the SVP active space C compared to the smaller SV active space B. With active space C, the error in energy is well within chemical accuracy ( \(<\) 1 kcal/mol) for the relaxed method, and it is close (\(\approx\) 2 kcal/mol) with the unrelaxed method.
We will now analyze the AO basis set convergence of embedding errors with the aid of We have enlarged the basis set from aug-cc-pCVDZ (54 basis functions) to aug-cc-pCVTZ (118 basis functions) for the SV (B) and SVP (C) active spaces. For relaxed UMP-CCSD, we observe a nearly constant error as we increase the AO basis from double-\(\zeta\) to triple-\(\zeta\) quality. However, for the unrelaxed UMP-CCSD variant, we observe that the error increases substantially from double-\(\zeta\) to triple-\(\zeta\). We also observed a few instabilities in the error curve of the unrelaxed UMP-CCSD method, corresponding to at least two solutions. With the relaxed approach, we do not see such instabilities, meaning no evidence of multiple solutions, as well as no loss of accuracy when enlarging the basis set.
The quality of the PECs and the difference curve with various methods, as described so far, influences the spectroscopic parameters as shown in Table 3. Compared to UCCSD, we observe that UMP-CCSD reproduces \(R_{e}\), \(\omega_{e}\) and \(\omega_{e}x_{e}\) better than the composite method and UMP2.
N\({}_{2}\) electronic energy error with respect to UCCSD in aug-cc-pCVDZ (denoted as DZ) and aug-cc-pCVTZ (denoted as TZ) basis sets. B and C active spaces are as described in Table 2. We set the energy at dissociation as the zero energy for each method.
Electronic energy error with respect to UCCSD along a PEC for N\({}_{2}\) in the aug-cc-pCVDZ basis set. A and B active spaces are as described in Table 2. DZ stands for the aug-cc-pCVDZ basis set. We set the energy at dissociation as the zero energy for each method.
PEC of N\({}_{2}\) molecule in aug-cc-pCVDZ basis set with various methods. B active space is as described in Table 2.
Although the relaxed and unrelaxed variants of UMP-CCSD yield similar MAE and NPE along the PEC, the energy difference plots around the equilibrium region become much smoother when orbital relaxations are considered. This improvement is reflected in the spectroscopic parameters \(\omega_{e}\) and \(\omega_{e}\),\(x_{e}\), which are coefficients of higher order terms in the polynomial. They now become significantly closer to the corresponding UCCSD results when orbital relaxations are taken into account. Hence, the self-consistency cycle improves the behavior around the equilibrium region significantly.
#### iii.1.2 Carbon monoxide (CO)
We now consider the isoelectronic carbon monoxide molecule. The full PECs for UMP2, UCCSD, relaxed and unrelaxed UMP-CCSD, and the composite method are given in Appendix Beginning with the cc-pVDZ basis and the large SVP active space C, the energy difference curves for all methods are shown in (top panel). There are two derivative discontinuities in the UMP2 curve and multiple derivative discontinuities in the composite method curve. The latter arises because the composite method inherits discontinuities from its individual computations; here multiple discontinuities appear at different bond distances for different basis sets, see "UMP2 (large)" and "UMP2 (small)" graphs in Appendix Using the unrelaxed version of UMP-CCSD, we observe one derivative discontinuity, which can be lifted by including orbital relaxations.
The bottom panel of Fig. 5, plots the basis set dependence of the error in the relaxed embedding method and the composite method. As for N\({}_{2}\), we observe less than 1 kcal/mol loss of accuracy when enlarging the basis set from aug-cc-pCVDZ to aug-cc-pCVTZ. For the uncoupled composite scheme, the extent of agreement varies strongly across the PEC, and is qualitatively worse than the relaxed UMP-CCSD method. This issues the value of electronic embedding.
We also calculate the spectroscopic parameters \(R_{e}\), \(\omega_{e}\) and \(\omega_{e}\),\(x_{e}\) for the carbon monoxide molecule, which are listed in Table 4. Compared to the preceding N\({}_{2}\) example, the composite method shows noticeable improvement over UMP2 for all computed quantities. We attribute this to the qualitative agreement of the composite method's PEC with the PEC of UCCSD near the equilibrium region, see Appendix Fig. 15, and the fact that the discontinuities appear further away from equilibrium. We still observe that \(R_{e}\) and \(\omega_{e}\) are best reproduced with the UMP-CCSD method, showing a consistent improvement over the composite method.
### W4 Dataset
We now investigate the accuracy of the embedding approaches on thermochemical energy differences using the well-known W4-11 dataset. The cc-pCVTZ basis set is used.
\begin{table}
\begin{tabular}{c|c c c} Method & \(\Delta R_{e}\) & \(\Delta\omega_{e}\) & \(\Delta\omega_{e}x_{e}\) \\ & (in Å) & (in cm\({}^{-1}\)) & (in cm\({}^{-1}\)) \\ \hline UMP2 & 0.0101 & -105.82 & 1.221 \\ Composite & 0.0017 & -12.76 & -0.0878 \\ UMP-CCSD (C) & 0.0007 & -6.73 & -0.1200 \\ UMP-CCSD (C)(relaxed) & 0.0005 & -4.65 & 0.0817 \\ \end{tabular}
\end{table}
Table 4: Spectroscopic parameters of CO evaluated in aug-cc-pCVTZ basis set. We reported the difference to the UCCSD results for all spectroscopic parameters. Note that “(C)” in “UMP-CCSD (C)”, stands for the active choice defined in Table 2.
Electronic energy differences w.r.t. UCCSD along a PEC for CO. DZ and TZ are abbreviations for aug-cc-pCVDZ and aug-cc-pCVTZ basis sets respectively. We set the energy at dissociation as the zero energy for each method.
\begin{table}
\begin{tabular}{c|c c c} Method & \(\Delta R_{e}\) & \(\Delta\omega_{e}\) & \(\Delta\omega_{e}x_{e}\) \\ & (in Å) & (in cm\({}^{-1}\)) & (in cm\({}^{-1}\)) \\ \hline UMP2 & 0.0059 & 1230.06 & 120.3261 \\ Composite & 0.0539 & 327.78 & 320.5781 \\ UMP-CCSD (C) & 0.0003 & 86.85 & -6.8039 \\ UMP-CCSD (C)(relaxed) & 0.0006 & 2.60 & -1.7131 \\ \end{tabular}
\end{table}
Table 3: Spectroscopic parameters of N\({}_{2}\) evaluated in aug-cc-pCVTZ basis set. We reported the difference w.r.t. the UCCSD method for all the spectroscopic parameters. In UMP-CCSD (C), C stands for the active space defined in Table 2.
All atoms are again included in the fragment. We use only the W4-11 single reference (SR) systems, for which UCCSD serves as the reference to test the MP-CCSD embedding models as well as the composite model and MP2. The multi-reference (MR) systems are less accurately described than the SR systems via UCCSD and are excluded. Altogether, we compute 140 total atomization energies (TAEs) and bond dissociation energies (BDEs), 20 isomerization energies (ISO), 505 heavy atom transfer (HAT), and 13 nucleophilic substitution (SN) reactions.
We observe that the root mean square error (RMSE) in UMP-CCSD is significantly improved compared to UMP2 for all classes of reactions. Moreover, relaxed UMP-CCSD provides a significant improvement over the composite results. It is encouraging that the RMSE of UMP-CCSD is at or within chemical accuracy for all the reactions considered in this study. Additionally, we observe that the maximal spread of the deviation is also much more favorable for relaxed UMP-CCSD than for the other methods considered.
### Rotational isomers of 1,3-butadiene
In this section, we report on the torsional potential around the central C\(-\)C bond of 1,3-butadiene, CH\({}_{2}\)\(=\)CH\(-\)CH\({}_{2}\). This conjugated molecule has been extensively studied in the literature, as its torsional potential is an archetype of the effect of loss of conjugation between the two double bonds at torsional angles (\(\tau\)) close to 90\({}^{\circ}\). More specifically, we have studied five different rotamers as shown in The optimized structures for all those isomers were taken from Feller _et al._, optimized at the highly accurate CCSD(T)/aug-cc-pVQZ level of theory. Following our previous examples, we have compared the UMP-CCSD and UMP2 methods with the all-electron UCCSD method, using the cc-pCVTZ basis set. Unlike the previous examples, we have chosen a fragment that is localized spatially as well as limited in its active orbitals. Thus we use the minimal STO-3G basis on only the 4 carbon atoms and consider only the C 2s 2p orbitals to comprise the active space. While the molecule has 30 electrons, 22 of them remain in the fragment, which has a total of 25 orbitals. To construct IAOs within the AVAS scheme, we have considered the STO-3G basis set as our minimal basis.
While it is not widely known (and indeed we were surprised to observe it), HF results exhibit spin-polarization all along the torsional potential (see Appendix Table 5), so UHF energies lie below RHF energies. This indicates the presence of some biradicaloid character in butadiene, which can be interpreted as arising from the \({}^{\bullet}\)CH\({}_{2}\)\(-\)CH\({}_{2}\)\(=\)CH\(-\)\({}^{\bullet}\)CH\({}_{2}\) resonance structure. It is well-known that UMP2 cannot recover symmetries that are broken at the UHF level, due to the absence of orbital rotations, sometimes leading to poor results. Indeed this artificial HF symmetry-breaking turns the torsional potential of butadiene from an easy problem for RMP2 into a challenging problem for UMP2. This is evident in where the relative energy of all the isomers was plotted relative to the _trans_ isomer. UMP2 inverts the correct form of the torsional potential, predicting \(\tau=90^{\circ}\) as the most stable structure, whereas it should be very close to the barrier.
By contrast, UCCSD relative energies agree quite well in comparison to the benchmark theore
Violin plot of the errors of UMP2, relaxed UMP-CCSD, and the composite method compared to the corresponding CCSD results for various thermochemical reactions with W4-11 dataset in the cc-pCVTZ basis set. The different panels show a) the total atomization energy, b) the isomerization energy, c) the bond dissociation energy, d) the nucleophilic substitution reaction energy, and e) the heavy atom transfer energy. All atoms are included in the fragment, with active orbitals defined via the use of cc-pVDZ in the AVAS scheme.
Rotational isomers of 1,3-butadiene.
Feller _et al._, due to the restoration of spin symmetry breaking to a large extent (see Appendix Table 5) in the presence of singles amplitudes. Turning to the UMP-CCSD methods, due to the absence of inactive singles in the unrelaxed version of the UMP-CCSD method, it is still qualitatively incorrect, although improved relative to UMP2. By contrast, upon inclusion of orbital rotations via \(T_{1}\) the relaxed version of the UMP-CCSD method improves significantly upon both the UMP2 method and the unrelaxed UMP-CCSD method and shows results that are qualitatively and quantitatively very similar to the UCCSD method. The interpretation is straightforward: \(T_{1}\) has largely removed the symmetry-breaking seen at the UHF level. This is further evidenced in Fig. 9, where the orbital-optimized MP2 (OOMP2) method based on an unrestricted reference also removes the inverted trend of the UMP2 method.
### Azomethane
In order to further test the efficiency of the UMP-CCSD method in capturing system relative energies using a spatially localized active fragment we have studied the PEC of azomethane, H\({}_{3}\)C\(-\)N\(=\)N\(-\)CH\({}_{3}\), along the N\(=\)N double bond stretching coordinate. All the geometric parameters, except the N\(=\)N bond distance of this molecule, were taken from Hermes _et. al_. We use the aug-cc-pCVDZ basis set, with all electrons correlated. For the active fragment, we have chosen only the 2s and 2p orbitals of the 2 nitrogen atoms, as obtained via the AVAS scheme. Also, to construct the IAOs within the AVAS scheme, we have chosen the minimal STO-3G basis set, resulting in a very small fragment consisting of 16 electrons in 16 orbitals.
In we have plotted the PEC with various methods and also the difference plot with respect to the UCCSD method. First we identify the Coulson-Fischer (CF) point by plotting the PEC with the RHF and UHF methods, which separate at that point. The UMP2 method shows a pronounced first derivative discontinuity (kink) at the CF geometry. The unrelaxed version of the UMP-CCSD method numerically improves upon the UMP2 method, but it cannot remove the kink. However, the relaxed UMP-CCSD method produces a smooth PEC, illustrating the key role of environment \(T_{1}\) amplitudes.
These characteristics of the PEC w.r.t. the UCCSD method are even more pronounced in the difference plot (right panel of Fig. 10). Interestingly, we observe apparent derivative discontinuities at bond lengths longer than the CF point with both UMP2 and unrelaxed UMP-CCSD, suggesting that the UHF solution changes character there. The relaxed version of the UMP-CCSD method, on the other hand, removes all discontinuities very satisfactorily. This analysis suggests a qualitatively correct behavior of the relaxed version of the UMP-CCSD method for local chemistry. However, the NPE of the PEC remains quite high with the current (very small) choice of active space, which presumably is coming from the larger error towards the dissociation regime.
To reduce the NPE, we expanded the size of the active space, by using N 2s 2p 3s 3p 3d orbitals, where cc-pVDZ basis set was considered as the reference basis set for IAO construction. shows that this choice significantly reduces the NPE. We then further increased the size of the basis set for the actual calculation to cc-pCVTZ, and chose the latter (larger) active space. We observed that when the relaxed scheme is employed, the error remains almost constant w.r.t. the cc-pCVDZ basis set, which corroborates our previous observations for diatomic dissociation.
## VII Conclusion
We have proposed and numerically scrutinized the new MPC-CC quantum embedding scheme, where the active fragment
Relative energy of different rotational isomers along the torsional potential around the central C\(-\)C bond of the butadiene molecule, H\({}_{2}\)C\(=\)CH\(-\)CH\({}_{2}\), with various unrestricted methods. The _trans_, that is, the isomer at \(\tau=180^{\circ}\) has been considered as the zeroth of energy. See text for computational details.
Relative energy of different rotational isomers of the butadiene molecule (w.r.t. the trans isomer) with both restricted and unrestricted methods. See text for computational details.
Instead of constructing a density matrix or Green's function at the CC level, MP-CC relies only on the cluster amplitudes to separate the environment and the fragment problem. Self-consistency is ensured through amplitude equations coupling the fragment and the environment. In particular, we do not need to construct a hybridization term between these two layers to take "environment fluctuations" into account. Furthermore, the interaction term for the fragment problem is renormalized by the low-level environment amplitudes.
MP-CC provides a general framework allowing the use of different low-level methods for solving the environment problem, while the high-level method for the fragment problem remains fixed as a CC method. In this work, we fixed the CC method as CCSD, and proposed and compared two different MP2-like environment models: one without orbital relaxation and one with orbital relaxation. The relaxed version significantly improves upon the unrelaxed one by removing a derivative discontinuity in the error curve relative to UCCSD results. By contrast, a composite method (or the uncoupled embedding method) numerically improves upon UMP2, but inherits first derivative discontinuities of UMP2, making it significantly inferior to relaxed UMP2-CCSD. Unrelaxed UMP2-CCSD is intermediate in performance between the composite theory and the preferable relaxed UMP2-CCSD approach. For relaxed UMP2-CCSD, a significant finding is that we achieve similar accuracy for a fixed active space when the basis set size is increased, which is generally not guaranteed for embedding methods.
The efficacy of relaxed MP-CCSD has been further established by studying rotational isomerism in 1,3-butadiene. The UMP2 method predicts an anomalous trend in the relative energies of the rotamers, which the unrelaxed version of our method cannot rectify. However, the relaxed UMP-CCSD method can recover the correct trend even with a very small active space.
We have investigated the reliability of the proposed orbital relaxed UMP-CCSD method for single reference systems in the W4-11 thermochemistry dataset. Our investigations show that relaxed UMP-CCSD reproduces, on average, the UCCSD results to chemical accuracy.
Furthermore, we have studied the central nitrogen double bond stretching of azomethane as a test case of a localized fragment. The relaxed UMP-CCSD method shows an almost constant energy difference relative to the all-electron UCCSD energy along the PEC, whereas the UMP2 method and the unrelaxed UMP-CCSD methods show non-smooth behavior near the symmetry-breaking region of this molecule.
There is ample scope for future work. The implementation reported and used here is not yet fully optimal, because of an unfavorable \(\mathcal{O}(V^{2}O^{2}O_{\text{F}}V_{\text{F}})\) scaling when the renormalized interaction is built before solving the fragment amplitude equations. Approximation schemes to construct the renormalized interaction could achieve lower polynomial scaling and also facilitate an efficient multi-fragment approach. This could be an alternative to existing embedding approaches such as LASSCF, DMET, SEET, DMFT or g-RISB. Moreover, there is also the possibility of utilizing efficient local correlation methods to reduce the complexity further. With respect to the high-level CC method, it would be very desirable to extend the theory to CCSDT; preliminary analysis has shown that MP-CC has a favorable scaling when including triples amplitudes. With respect to the low-level MP
Energy difference w.r.t. the UCCSD method for azomethane molecule. DZ and TZ stand for the cc-pCVDZ and cc-pCVTZ basis sets, respectively. A larger local active space on the 2 N atoms, as described in the text, is used for this plot. We set the energy at dissociation as the zero energy for each method.
(left panel) PEC of N=N double bond stretching of azomethane in the cc-pCVDZ basis set. (right panel) Difference plot w.r.t. the UCCSD method. The smallest (valence) active space on the 2 N atoms, as described in the text, is used for this plot.
# Appendix E Measure of spin symmetry breaking in 1,3-butadiene
In this section, we have tabulated the \(\langle S^{2}\rangle\) value as a measure of spin symmetry breaking in various rotational isomers of 1,3-butadiene molecule.
Correlation energy differences w.r.t. UCCSD in aug-cc-pCVDZ basis set along a PEC for CO. large and small basis sets stand for aug-cc-pCVDZ and cc-pVDZ respectively.
Total energy differences w.r.t. UCCSD in aug-cc-pCVDZ basis set along a PEC for CO.
|
10.48550/arXiv.2404.09078
|
A static quantum embedding scheme based on coupled cluster theory
|
Avijit Shee, Fabian M. Faulstich, Birgitta Whaley, Lin Lin, Martin Head-Gordon
| 1,219
|
10.48550_arXiv.2409.12438
|
###### Abstract
While collision lifetimes are a fundamental property of few-body scattering events, their behavior at ultralow temperatures is not completely understood. We derive a general expression for the Smith lifetime Q-matrix using multichannel quantum defect theory, which allows us to obtain the average time delay in the incident \(s\)-wave collision channel in the limit of zero collision energy \(E\), \(Q_{11}=A/\sqrt{E}+B\), where \(A\) and \(B\) are constants. We show that the time delay is dominated by elastic scattering, and contains an additional multichannel contribution independent of the two-body scattering length. We also obtain the expressions for the collision lifetime using the universal model of Idziaszek and Julienne [Phys. Rev. Lett. **104**, 113202]. The lifetime acquires an imaginary part in the presence of inelastic loss due to the lack of unitarity of the S-matrix, and shortens with increasing the short-range loss parameter \(y\).
## I Introduction
Ultracold few-body collisions and chemical reactions are responsible for a wide range of phenomena, which occur in ultracold atomic and molecular gases, such as thermalization, inelastic losses, and cluster formation. Particularly intriguing are two-body collisions of ultracold molecules, which are characterized by the formation of long-lived collision complexes with lifetimes ranging from nanoseconds to milliseconds. These sticky collisions lead to inelastic losses, which are commonly described by the universal model (UM) using a single parameter \(y\in\), which quantifies the extent of loss of the collision flux once the collision partners have reached the short-range region of the interaction potential. The value of the ultracold reaction rate in the universal limit (\(y=1\)) does not depend on the details of short-range interactions. Numerically exact coupled-channel calculations go beyond the UM and enable rigorous insights into short-range dynamics of the collision complex by providing the full S-matrix for a given interaction potential (see, e.g., Refs.).
A complementary approach to characterizing bimolecular collisions is based on the concept of collision lifetime \(Q=d\eta/dE\), where \(\eta\) is the scattering phase shift and \(E\) is the collision energy (\(\hbar=1\) units are used throughout). As shown by Wigner and Eisenbud, \(Q\) gives the time delay experienced by a quantum wavepacket in the short-range interaction region as compared to the free wavepacket in the absence of interactions. The concept of time delay was generalized by Smith to multichannel scattering. The collision lifetime of a multichannel quantum scattering event is characterized in terms of the Q-matrix, whose diagonal elements give average time delays in each incident collision channel. The Q-matrix approach has proven fruitful for characterizing scattering resonances in atomic, molecular, and electron collisions. Importantly, the trace of the Q-matrix is related to the excess density of states of the collision complex (with respect to the unperturbed density of states) via the time delay theorem (see Ref. and references therein). The zero-temperature limit of the time delay in single-channel elastic collisions is known to be \(Q=-a\mu/k\), where \(a\) is the scattering length, \(\mu\) is the reduced mass of the collision partners, and \(k\) is the collision wavevector. (Alternative definitions of the time delay are also possible, such as \(Q^{\prime}=Q-\frac{1}{2E}\sin\eta\) with \(Q^{\prime}\to 0\) in the \(s\)-wave limit.) However, ultracold collisions are typically multichannel in nature, i.e., they are accompanied by inelastic scattering, in which the internal states of the collision partners change. To our knowledge, the zero-temperature behavior of multichannel collision lifetimes is unknown.
Here, we use multichannel quantum defect theory (MQDT) to derive the threshold laws for collision lifetimes in multichannel scattering. MQDT takes advantage of the separation of distance and energy scales in ultracold collisions to obtain their theoretical description in terms of energy-independent short-range parameters. MQDT is a powerful tool in the theory of ultracold atom and atom-molecule collisions. It has recently been combined with a frame transformation for greatly enhanced computational efficiency. Here, we obtain a general MQDT expression for the multichannel Q-matrix and show that the average time delay in the incident collision channel (hereafter referred to as time delay or lifetime) scales as \(Q_{11}=A/k+B\), where \(A\) and \(B\) are constants, an analog of the Wigner threshold law for ultracold time delays. We also examine the role of elastic and inelastic contributions to the time delay, and find that the former usually dominates. Finally, we use the universal model of Idziaszek and Julienne to derive the universal limit for ultracold two-body collision lifetimes in the presence of inelastic losses.
## II MQDT theory of collision lifetimes
The lifetime Q-matrix may be expressed in terms of the scattering S-matrix as
\[{\bf Q}=i{\bf S}\frac{d{\bf S}^{\dagger}}{dE}. \tag{1}\]The diagonal matrix elements of \(\mathbf{Q}\), \(Q_{ii}\), represent the average time delays in the \(i\)-th channel, with \(i=1\) corresponding to the incident channel.
\[\mathbf{S}=e^{i\mathbf{\xi}}[\mathbf{I}+i\bar{\mathbf{R}}][\mathbf{I}-i\bar{ \mathbf{R}}]^{-1}e^{i\mathbf{\xi}}, \tag{2}\]
Combining this expression with Eq.
\[\bar{\mathbf{R}}=\mathbf{C}^{-1}[\bar{\mathbf{Y}}^{-1}-\tan\mathbf{\lambda}])^{-1 }\mathbf{C}^{-1}. \tag{4}\]
Here, the diagonal matrices \(\tan\mathbf{\lambda}\) and \(\mathbf{C}^{-1}\) account for threshold effects and the relationship between the amplitude of the WKB-normalized and energy-normalized reference functions. The matrix \(\bar{\mathbf{Y}}\) is expressed via the open-open (oo), open-closed (oc), and closed-closed (cc) subblocks of yet another matrix \(\mathbf{Y}\) as \(\bar{\mathbf{Y}}=\mathbf{Y}_{oo}-\mathbf{Y}_{oc}[\tan\mathbf{\nu}+\mathbf{Y}_{cc}] ^{-1}\mathbf{Y}_{co}\), where \(\tan\mathbf{\nu}\) is a diagonal matrix of closed-channel phases. The elements of \(\bar{\mathbf{Y}}\) can be assumed independent of the collision energy \(E\). The matrix \(\bar{\mathbf{Y}}\) contains closed-channel effects (which give rise to, e.g., Feshbach resonances) and can therefore vary rapidly with \(E\). Here, we are interested in threshold effects on multichannel collision lifetimes, so we will neglect these variations and set \(\bar{\mathbf{Y}}=\mathbf{Y}_{oo}\), which is equivalent to assuming that we are far away from any resonances. The term \(\mathbf{Y}_{\lambda}=[\mathbf{Y}_{oo}^{-1}-\tan(\mathbf{\lambda})]^{-1}\) is energy-independent because \(\tan\mathbf{\lambda}\) is energy-independent near \(s\) and \(p\)-wave thresholds. As a result, Eq. becomes \(\bar{\mathbf{R}}=\mathbf{C}^{-1}\mathbf{Y}_{\lambda}\mathbf{C}^{-1}\), so \(\bar{\mathbf{R}}^{\prime}=0\) and the last term in Eq. vanishes.
We are interested in evaluating the average time delay in the incident collision channel (\(Q_{11}\)). To this end, we assume that the energy gap between the incident channel and all other inelastic channels is large compared to the collision energy. This allows us to neglect the energy dependence of all but the leading MQDT parameters in the diagonal matrices \(\tan\mathbf{\lambda}\) and \(\mathbf{C}\). To isolate the energy dependent terms, we decompose the \(N\times N\)\(\bar{\mathbf{R}}\) matrix into \(1\times 1\) and \((N-1)\times(N-1)\) blocks and use the Schur block inversion method to calculate \([\mathbf{I}-i\bar{\mathbf{R}}]^{-1}\) in Eq. (see the Supplemental Material).
\[a_{1} =iY_{\lambda_{11}}+i\frac{Y_{\lambda_{12}}}{C_{2}}(i\tilde{D}_{11 }\frac{Y_{\lambda_{12}}}{C_{2}}+i\tilde{D}_{12}\frac{Y_{\lambda_{13}}}{C_{3}} +...) \tag{6}\] \[\quad+i\frac{Y_{\lambda_{13}}}{C_{3}}(i\tilde{D}_{12}\frac{Y_{ \lambda_{12}}}{C_{2}}+i\tilde{D}_{22}\frac{Y_{\lambda_{13}}}{C_{3}}+...)+...\] \[M_{N} =-iY_{\lambda_{11}}+\frac{1}{C_{2}}Y_{\lambda_{12}}(\frac{\tilde{ D}_{11}}{C_{2}}Y_{\lambda_{12}}+\frac{\tilde{D}_{12}}{C_{3}}Y_{\lambda_{13}}+...)\] \[\quad+\frac{1}{C_{3}}Y_{\lambda_{13}}(\frac{\tilde{D}_{21}}{C_{2 }}Y_{\lambda_{12}}+\frac{\tilde{D}_{22}}{C_{3}}Y_{\lambda_{13}}+...)+...\,,\]
The first column of the S-matrix is given by
\[S_{n1}=e^{i\xi_{1}}e^{i\xi_{n}}\frac{C_{1}^{-1}}{1+C_{1}^{-2}M_{N}}b_{n}, \tag{7}\]
where \(n\geq 2\) and
\[b_{n}=i\frac{Y_{\lambda_{1n}}}{C_{n}}+(i\tilde{D}_{1(n-1)}\frac {Y_{\lambda_{12}}}{C_{2}}+i\tilde{D}_{2(n-1)}\frac{Y_{\lambda_{13}}}{C_{3}}+...)\] \[\quad\quad\quad\quad\quad\quad\quad+i\frac{Y_{\lambda_{2n}}}{C_{ 2}C_{n}}(i\tilde{D}_{11}\frac{Y_{\lambda_{12}}}{C_{2}}+i\tilde{D}_{12}\frac{Y_{ \lambda_{13}}}{C_{3}}+...)\] \[\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \quad\\(C^{-2}=k\bar{a}[1+(s-1)^{2}]\), \(\tan\xi=-ka\), and \(\tan\lambda=1-s\). Here, \(s=a/\bar{a}\) is the dimensionless scattering length, and \(\bar{a}=4\pi C_{6}/\Gamma(\frac{1}{4})^{2}\) is the mean scattering length expressed in terms of the long-range dispersion coefficient \(C_{6}\). Substituting these values in Eq.
\[Q_{11}=-\frac{2a\mu}{k(1+a^{2}k^{2})}\frac{1-2M_{R}k\bar{a}[1+(s- 1)^{2}]}{1+2M_{R}k\bar{a}[1+(s-1)^{2}]} \tag{12}\] \[-2\mu\frac{1}{k}\frac{M_{I}(\bar{a}[1+(s-1)^{2}])}{1+4M_{R}k\bar {a}[1+(s-1)^{2}]}\] \[-\frac{a\mu}{1+a^{2}k^{2}}\frac{4M_{R}\bar{a}[1+(s-1)^{2}]}{1+2M_ {R}k\bar{a}[1+(s-1)^{2}]}\] \[-4\mu\frac{M_{R}M_{I}\bar{a}^{2}[1+(s-1)^{2}]^{2}}{1+4M_{R}k\bar {a}[1+(s-1)^{2}]}.\]
We observe that the energy dependence of the average time delay is determined by four contributions. The first two contributions dominate in the limit \(k\to 0\), where the first term reduces to \(-2a\mu/k\), in agreement with prior single-channel results. The second term in Eq. gives an intrinsically multichannel contribution to the average time delay, which vanishes for single-channel elastic scattering. The third and the fourth terms are also of multichannel origin, but approach constant values in the limit \(k\to 0\). We analyze these contributions in more detail below.
To gain more insight into the threshold scaling of the average time delay, consider a two-channel model described by a single short-range parameter \(Y_{12}\). The choice \(Y_{11}=Y_{22}=0\) simplifies the calculations without loss of generality. The corresponding elastic and inelastic cross sections are plotted as a function of collision energy in At low energies, the cross sections follow the expected Wigner threshold laws \(\sigma(k)\simeq\mathrm{const}\) (\(\sigma(k)\simeq 1/k\)) for s-wave elastic (inelastic) scattering.
While one can derive closed-form analytic expressions for the average time delay in the two-channel model, they are too cumbersome to analyze. Instead, we plot in the four contributions to \(Q_{11}\) [see Eq.] as a function of collision energy for different values of \(Y_{12}\). As noted above, the first and the second terms in Eq. arise from elastic collisions, while the third and the forth terms arise from inelastic collisions. We observe that the elastic contributions to the average time delay dominate at all energies and \(Y_{12}\) values. This holds regardless of the value of \(Y_{12}\), and is further emphasized in Fig. 2, where the total inelastic contribution is compared with the total value of \(Q_{11}\) and found to be significantly smaller in comparison. This is consistent with Eq., which shows that in the \(k\to 0\) limit, the elastic contributions scale as \(Q_{11}^{\mathrm{el}}(k)\simeq 1/k\), whereas the inelastic ones remain constant, so the former are expected to dominate.
Physically, the predominance of the elastic scattering contribution to the time delay can be understood by noting that, in the limit of zero collision energy, the inelastically scattered wavepacket escapes from the collision region with a much higher velocity than either the incident or the elastically scattered wavepacket. As a result, the lifetime of the inelastically scattered wavepacket makes a negligible contribution to the average time delay, consistent with the threshold law in Eq..
The four contributions to the average time delay \(Q_{11}\) in the two-channel model [see Eq.] plotted as a function of collision energy. Solid blue and red lines show the elastic contributions [the first two terms in Eq.]. Solid green and black lines show the inelastic contributions [the last two terms in Eq.]. Blue circles – the total \(Q_{11}\), red circles – the total inelastic contribution to \(Q_{11}\).
Elastic and inelastic cross sections as a function of collision energy in the two-channel model.
## III Universal model of collision lifetimes
The \(S\)-matrix expression in the UM takes the form
\[S_{11}=\frac{1+i(\tan\xi(E)-\frac{yC^{-2}(E)}{i+y\tan\lambda(E)})}{1-i(\tan\xi(E) -\frac{yC^{-2}(E)}{i+y\tan\lambda(E)})}, \tag{13}\]
Taking the energy derivative of Eq. we obtain for the average collision time delay \(Q_{11}=-iS_{11}^{1}\frac{d\xi_{11}}{dE}\)
\[-i\frac{1-i(\tan\xi-\frac{yC^{-2}(E)}{-i+y\tan\lambda})}{1+i(\tan\xi-\frac{yC^ {-2}(E)}{-i+y\tan\lambda})}\frac{2i(\frac{d\tan\xi}{dE}+2\frac{y}{\frac{y}{ \frac{y}{\frac{y}{\sqrt{2}}}{C^{2}(E)}}})}{(1-i(\tan\xi-\frac{yC^{-2}(E)}{i+y \tan\lambda}))^{2}} \tag{14}\]
Using the values of MQDT functions in the \(s\)-wave limit Eq.
\[Q_{11}^{R} = \frac{2\mu}{k}[aF_{1}(s,y)+y\bar{a}(2-s)F_{2}(s,y)],\] \[Q_{11}^{I} = \frac{2\mu}{k}[aF_{2}(s,y)-y\bar{a}(2-s)F_{1}(s,y)], \tag{15}\]
The time delay given by Eq. is a complex quantity because the scattering matrix in the UM is unitary only in the limit \(y\to 0\) (no loss), where \(F_{1}(s,y)\rightarrow-1\) and \(F_{2}(s,y)\to 0\), and the expressions reduce to the prior result obtained for one-dimensional potential scattering \(Q_{11}=-2\mu a/k\). Thus, the real part of the complex time delay (\(Q_{11}^{R}\)) can be interpreted as arising from elastic scattering, and the imaginary part (\(Q_{11}^{R}\)) as due to inelastic scattering. This is similar to the imaginary part of the scattering length, which is proportional to the inelastic scattering cross section. The universal limit for the average time delay may be obtained by setting \(y\to 1\) in Eq..
For small and moderate \(s\leq 10\), the real (elastic) part of the time delay always dominates. The imaginary contribution \(Q_{11}^{I}\) is very small for small \(y\), but increases with \(y\), which is consistent with the interpretation of \(Q_{11}^{I}\) as due to inelastic loss. Interestingly, we observe a single maximum for \(s=2\) (and a maximum followed by a minimum for \(s\geq 10\)) in the imaginary time delay as a function of \(y\). These maxima and minima occur near \(y=1\) for small \(s\) but shift to smaller \(y\) for larger \(s\). We attribute these variations to the quantum interference between the flux partially reflected from the short-range region (for \(y<1\)) and the incident flux.
Interestingly, as shown in Fig. 3, the ratio \(Q_{11}^{R}(y)/Q_{11}(y=0)<1\) is smaller than unity for all values of \(s\) and \(y\) studied. Thus, the average time delay in the presence of inelastic loss (\(y>1\)) tends to be smaller in absolute magnitude than the single-channel time delay [\(Q_{11}(y=0)=2\mu|a|/k\). The latter may therefore be regarded as an upper limit to the time delay that can be reached in a multichannel collision (in the absence of resonant scattering).
## IV Summary
In summary, we have developed the MQDT theory of collision lifetimes based on Smith's Q-matrix formalism. This enabled us to derive the Wigner threshold law for multichannel collision lifetimes, extending previous single-channel results. We found that the zero-energy behavior of the average time delay in a simple two-channel MQDT model is dominated by elastic scattering in most regimes. Finally, we have extended the UM of lossy collisions to calculate the collision lifetimes, and explored their dependence on the short-range loss parameter \(y\), again finding the predominance of elastic scattering in all regimes, except at large \(y\simeq 1\) and \(s\gg 10\). In future work, it would be interesting to investigate multichannel collision lifetimes in the presence of closed channels, which give rise to Feshbach resonances.
|
10.48550/arXiv.2409.12438
|
Universality and threshold laws for collision lifetimes at ultralow temperatures
|
Rebekah Hermsmeier, Timur V. Tscherbul
| 3,539
|
10.48550_arXiv.2407.21366
|
## 1 Introduction
Quantum mechanical tunneling is a central concept in chemistry and physics. One of the experimentally observable hallmarks of tunneling processes is the splitting of energy levels. These tunneling splittings are measurable and can be probed, _e.g._, using microwave or infrared spectroscopy. The splittings are exquisitely sensitive to the energetics and shape of the underlying potential energy surface (PES) and provide valuable benchmarks to validate state-of-the-art theoretical methodology.
Accurate PESs allow to carry out quantitative molecular simulations that link directly to experiments. One of the most successful recent approaches to devise such PESs uses a combination of electronic structure calculations and machine learning techniques to give a machinelearning-based PES. Coupled cluster-based techniques such as CCSD(T) are among the most accurate methods for closed-shell systems. However, due to the unfavourable scaling of such calculations (seventh power of the number of basis functions), the necessary number of reference calculations required for training a full-dimensional PES (\(\sim 10^{4}\) or more) at this level of theory is only feasible for small molecules with \(\lesssim 5\) heavy atoms. However, if a small number of calculations can be afforded at such high levels, a considerable boost in performance can be achieved using transfer learning (TL) techniques. The essence of TL is that the overall topographies of molecular PESs at different levels of theory are often comparable which can be exploited by retraining a low-level (LL) PES with small amounts of data computed at a high level. Such an approach is explored in the present work for two systems that have so far been intractable at the CCSD(T) level of theory.
Computing accurate tunneling splittings for multidimensional systems from fully quantum-mechanical methods is a formidable task. The ring-polymer instanton (RPI) approach provides a semiclassical approximation for tunneling calculations which scales well with system size. As instanton theory makes a local harmonic approximation around the optimal tunneling pathway, it typically deviates from quantum-mechanical benchmarks by up to 20%. This error can, however, be significantly reduced using a new procedure to perturbatively correct the results using information about third and fourth derivatives along the path.
The present work reports tunneling splittings for 15-atom tropolone and 12-atom (propiolic acid)-(formic acid) dimer (PFD) (see Figure 1) following the TL+RPI approach. These molecules are related to malonaldehyde and the formic acid dimer (FAD), which have been studied experimentally and theoretically in great detail. Both molecules have been treated with RPI, which was particularly successful for malonaldehyde. Using perturbatively corrected RPI on two different malonaldehyde PESs gave a tunneling splitting for hydrogen (H-)transfer of 22.1 cm\({}^{-1}\) whereas the experimental value is 21.6 cm\({}^{-1}\). Compared with malonaldehyde and FAD, far less theoretical and experimental work has been devoted to their "larger siblings", tropolone and PFD, mostly due to the difficulty of obtaining CCSD(T) quality PESs for these systems of many electrons.
The experimentally reported ground vibrational state (H/D) tunneling splittings for tropolone are (0.974/0.051) cm\({}^{-1}\). Theoretical works include a permutationally invariant polynomial PES based on 6604 B3LYP reference energies and forces. This was recently improved to near-CCSD(T) quality using the molecular tailoring approach (MTA) to obtain approximate CCSD(T) energies and \(\Delta\)-machine learning based on 2044 energies. The (H/D) tunneling splitting as evaluated on the \(\Delta\)-machine learned PES was (0.68/0.031) cm\({}^{-1}\), which was subsequently improved by scaling the action using further MTA calculations to give (0.92/0.042) cm\({}^{-1}\).
For PFD, the experimental energy splittings are (0.009721/0.0001117) cm\({}^{-1}\) for H-H and D-D exchange, respectively. Theoretically, PFD has, for example, been investigated using second-order vibrational perturbation theory and a one-dimensional model has been employed to estimate tunneling splittings, which, however, were not in quantitative agreement with experiment.
Global minimum energy structures for tropolone (A) and PFD (B).
## 2 Results and Discussion
For both tropolone and PFD, a sufficiently large number of CCSD(T)/aug-cc-pVTZ calculations is currently out of reach (_e.g._, for tropolone: "Ideally, we would learn the difference between the DFT and the "gold standard" CCSD(T) level. However, this is not feasible (for us) owing to \(N^{7}\) scaling..."). In an attempt to reach CCSD(T) quality, the present work employs TL strategies for tropolone and PFD. First, LL reactive PESs were determined, see Supporting Information. Schematics of the two PESs at the MP2 (low-)level including molecular structures and the accuracy of the energy on stationary points are shown in Figures S2 and S3. Comparing the energies of the stationary points of the PESs with direct _ab initio_ calculations yield maximum deviations of 0.05 and 0.03 kcal/mol for tropolone and PFD, respectively. Additional validations on a hold-out (test) set, see Table 1 and Figure S4, feature mean absolute errors on energies (MAE(\(E\))) well below 0.1 kcal/mol.
The barriers for H-transfer from the LL-PESs agree to within 0.001 and 0.01 kcal/mol with those from direct _ab initio_ calculations at the MP2 level which are 4.29 and 6.47 kcal/mol for tropolone and PFD, respectively. Harmonic frequencies constitute an additional measure for the quality of a PES around a stationary point. For the global minimum and the transition state (TS) for H-transfer, the MAEs are \(\sim 1\) cm\({}^{-1}\) and \(\sim 0.4\) cm\({}^{-1}\) for tropolone and PFD, respectively. Tables S1 and S2 compare the individual frequencies for tropolone and PFD. This validates the quality of the LL-PESs compared with reference electronic structure calculations at the MP2 level of theory.
For the present work, the main target observable is the tunneling splitting \(\Delta^{\rm H}_{\rm RPI}\) for H-transfer. Despite the remarkable accuracy of the two LL-PESs for tropolone and PFD when compared with the reference MP2 data, the computed tunneling splittings overestimate the experimental results by an order of magnitude: The H-transfer tunneling splittings on the LL-PESs are 10.94 and 0.11 cm\({}^{-1}\), whereas experimental measurements give 0.974 and 0.0097211 cm\({}^{-1}\) for tropolone and PFD, respectively. This is a clear indication that MP2 is not sufficient for quantitative calculations of tunneling splittings. Hence, transfer learning to the CCSD(T) level was explored next based on comparatively small numbers of additional high-level calculations. The iterative data selection procedure, which was identical for both systems, is described in detail in the Supporting Information, with only a concise summary provided here.
Based on an initial set of 25 judiciously selected structures, the data set was iteratively extended by selecting \(n_{i}\) structures along the refined minimum energy path (MEP) and in
\begin{table}
\begin{tabular}{c c c} \hline kcal/mol(/Å) & **Tropolone** & **PFD** \\ \hline MAE(\(E\)) & 0.067 & 0.034 \\ RMSE(\(E\)) & 0.094 & 0.165 \\ MAE(\(F\)) & 0.069 & 0.133 \\ RMSE(\(F\)) & 0.336 & 1.793 \\ \hline \(E_{\rm B}\) & 4.284 & 6.460 \\ \(E_{\rm B}^{\rm Ref.}\) & 4.285 & 6.470 \\ \hline MAE(\(\omega\))\({}^{\rm min}\) (cm\({}^{-1}\)) & 1.02 & 0.37 \\ MAE(\(\omega\))\({}^{\rm TS}\) (cm\({}^{-1}\)) & 1.06 & 0.41 \\ \hline \hline \end{tabular}
\end{table}
Table 1: PhysNet performance for the LL-PESs: Mean absolute and root mean squared errors (MAE and RMSE, respectively) on energies and forces are given for predicting the test sets of 2000 and 4000 random structures for tropolone and PFD, respectively. The energy barrier for H-transfer is obtained from PhysNet (\(E_{\rm B}\)) and from _ab initio_ calculations (\(E_{\rm B}^{\rm Ref.}\)) at the LL of theory. The MAEs for the harmonic frequencies calculated from the two LL-PESs for the global minimum and the H-transfer TS as compared to their reference _ab initio_ values are also reported.
reports sets \(\rm TL_{0}\) to \(\rm TL_{2}\) for tropolone together with the MEPs (dot-dashed) and IPs (dashed) on the LL- (black) and a representative TL-PESs (purple). All steps involving the LL- and TL-PESs are rapid whereas the high-level _ab initio_ calculations are computationally demanding (\(\sim 50\) h for tropolone and \(\sim 100\) h for tropolone). The high-level _ab initio_ calculations are computationally demanding (\(\sim 50\) h for tropolone) and \(\sim 100\) h for tropolone, respectively. The high-level _ab initio_ calculations are computationally demanding (\(\sim 50\) h for tropolone) and \(\sim 100\) h for tropolone, respectively.
However, the high-level _ab initio_ calculations for multiple structures can be run in parallel. Due to the rather small high-level data set sizes, TL was repeated 10 times on different splits of the data.
The H-transfer barrier heights for tropolone and PFD from all TLs are shown in For tropolone, the barrier height at the MP2 level is 4.28 kcal/mol, which increases to 6.62 kcal/mol for TL\({}_{2}\), _i.e._, after TL with 100 high-level data points. Remarkably, the result with only 25 data points (TL\({}_{0}\)) appears converged as further addition of data points reduces the uncertainty (the error bars) but not the average barrier height. Similarly, for PFD the MP2 energy barrier of 6.46 kcal/mol increases to 7.76 kcal/mol on the TL\({}_{3}\) PES. Again, going from TL\({}_{0}\) to TL\({}_{3}\) reduces the uncertainty and adjustments of the energy barrier progressively decrease. It should be noted that obtaining optimized geometries and harmonic frequencies from brute-force electronic structure calculations at the CCSD(T)/aug-cc-pVTZ level is not feasible for systems of the sizes studied here, but straightforward when using the TL-PES. From the two TL-PESs the best estimate for the barrier heights are 6.62 kcal/mol and 7.76 kcal/mol which compare with 6.63 kcal/mol and 7.76 kcal/mol from single-point CCSD(T) calculations using the optimized and TS structures obtained from the TL\({}_{2}\) and TL\({}_{3}\) PESs for tropolone and PFD, respectively.
Previous work reported barrier heights of 6.69 kcal/mol for tropolone from a single-point CCSD(T)/aug-cc-pVTZ calculation using a MP2/aug-cc-pVTZ optimized structure. Alternatively, using \(\Delta\)-machine learning to the MTA-CCSD(T)/aug-cc-pVTZ level yields a barrier height of 7.18 kcal/mol. For PFD, preliminary calculations at the CCSD(T)-F12b/aug-cc-pVDZ level found \(E_{\rm B}\sim 8.0\) kcal/mol for double H-transfer in PFD.
Ring polymer instanton simulations do not involve random numbers or statistical errors which clearly distinguishes them from path-integral molecular dynamics (PIMD) methods or Diffusion Monte Carlo (DMC) simulations. Hence, the instanton (once it is converged with \(\tau\rightarrow\infty\) and \(N/\tau\rightarrow\infty\)) is in principle uniquely determined by a given PES. As shown in the IP differs from the MEP. Representative IPs on the LL- and TL-PESs for both systems are shown in and as a movie in the Supporting Information. For tropolone, the instanton is mainly characterized by a hydrogen motion and slight adjustments in the CC and CO bond lengths due to single-/double-bond migration. For PFD, more pronounced displacements of the oxygen atoms, the monomer separation and wagging of the CCH tail occur during double H-transfer.
The H- and D-transfer tunneling splittings as obtained from the different TL-PESs for tropolone and PFD are shown in For tropolone, the instanton result on the LL-PES of 10.94 cm\({}^{-1}\) is brought much closer to the experimental result of 0.974 cm\({}^{-1}\) using the TL-PESs. In particular, the experimental result is within the standard deviation of the calculations. Note that the excellent agreement of TL\({}_{0}\) and its rather limited fluctuation around the mean probably appears by chance.
Energy barriers \(E_{\rm B}\) for H-transfer in tropolone (A) and PFD (B) from all ten TL-PESs (transparent circles). The corresponding averages (opaque circle) and standard deviations (error bars as \(\pm\sigma\)) are also indicated. An _ab initio_ optimization and subsequent determination of the energy barrier at the CCSD(T) level was out of reach for the two molecular systems.
For TL\({}_{2}\), which is arguably the most accurate PES for tropolone, \(\Delta_{\rm RPI}^{\rm H}=1.01\) cm\({}^{-1}\) and \(\Delta_{\rm RPI}^{\rm D}=0.055\) cm\({}^{-1}\), which both overestimate experimental values by \(\sim 5\%\), and indicate that further improvements should be possible which is addressed further below.
For PFD, gradually increasing the data set size from TL\({}_{0}\) to TL\({}_{3}\) led to smaller uncertainty and decreasing changes in \(\Delta_{\rm RPI}\). The final transfer learning, TL\({}_{3}\), was performed to capture the potential along the dissociation of the dimer (see Figure S5). As anticipated this only marginally affected the tunneling splitting. The values of \(\Delta_{\rm RPI}^{\rm H}=0.014\) cm\({}^{-1}\) and \(\Delta_{\rm RPI}^{\rm D}=0.00024\) cm\({}^{-1}\) overestimate the experimentally reported splittings by \(\sim 50\%\) and a factor of 2, respectively.
Energy profiles (top) and corresponding instantons (bottom) with \(N=4096\) beads from the LL- and (a representative) TL-PES for tropolone (A) and PFD (B). The energy profiles are given as a function of cumulative mass-weighted path length \(r\).
The tunneling splittings for the two systems considered here as determined on the largest available TL data sets are summarized in Table 2. For tropolone, the tunneling splitting of the symmetric \({}^{13}\)C isotopologue has been determined experimentally to be 0.97 cm\({}^{-1}\), compared with \(\Delta^{\rm SP}_{\rm RPI}=1.00\) cm\({}^{-1}\) from the present simulations. For the \({}^{18}\)OH\({}^{18}\)O isotopologue only an experimentally estimated value \(\Delta^{{}^{18}\rm O}_{\rm Expt.}\sim 0.83\) cm\({}^{-1}\) is available so far. Hence, the splitting was computed and yields \(\Delta^{{}^{\rm SP,\,18}\rm O}_{\rm RPI}=0.95\) cm\({}^{-1}\).
Tunneling splittings for hydrogen (left) and deuterium (right) transfer in tropolone (top row) and PFD (bottom row) from all TL-PESs (transparent circles). The corresponding averages (opaque circle) and standard deviations (error bars as \(\pm\sigma\)) are also reported.
Thus, the ratio of calculated results can be used to obtain a new estimate of the true tunneling splitting as \(\Delta^{\rm 18O}_{\rm Expt.}\sim\Delta^{\rm H}_{\rm Expt.}\cdot\Delta^{\rm SP,\, \rm 18O}_{\rm RPI}/\Delta^{\rm SP,\rm H}_{\rm RPI}=0.974\cdot 0.95/1.01=0.916\) cm\({}^{-1}\). This value is considered an improved estimate for \(\Delta^{\rm 18O}_{\rm Expt.}\) over \(\Delta^{\rm 18O}_{\rm Expt.}\sim 0.83\) cm\({}^{-1}\) and is a testable prediction for future experiments from the present simulations. In all cases, the computed splittings are larger than those reported experimentally. This is at least partly due to inherent approximations of instanton theory.
Recently, a correction scheme for RPI-based tunneling splittings using third- and fourth-order derivatives was proposed. For cases where the higher-order derivatives are carried out numerically, it was shown that numerically stable applications of this correction scheme require double-precision arithmetics for training and inference of a neural network based PES. This goes beyond standard procedures because machine learning models typically operate at 32-bit accuracy (single-precision) to reduce the computational cost for training and inference, and to lower memory requirements. It was found that single-precision can have detrimental effects particularly if higher-order derivatives of the energy are required. For that reason, TL was repeated using double-precision for the largest TL data set sizes and \(\Delta^{\rm DP}_{\rm RPI}\) was determined together with the perturbatively corrected value \(\Delta^{\rm DP}_{\rm PC}\).
\begin{table}
\begin{tabular}{l|c c c|c c} \hline & \multicolumn{2}{c|}{Tropolone \(\langle\)TL\({}_{2}\rangle\)} & \multicolumn{2}{c}{PFD \(\langle\)TL\({}_{3}\rangle\)} \\ & H & D & H(\({}^{13}\)C) & H & D \\ \hline \(\Delta^{\rm SP}_{\rm RPI}\) & 1.01 & 0.055 & 1.00 & 0.0142 & 0.000235 \\ \(\Delta^{\rm DP}_{\rm RPI}\) & 1.07 & 0.058 & n.d. & 0.0165 & 0.000269 \\ \(\Delta^{\rm DP}_{\rm PC}\) & 0.94 & 0.047 & n.d. & 0.0147 & 0.000217 \\ Expt. & 0.974 & 0.051 & 0.97 & 0.009721 & 0.0001117 \\ \hline \end{tabular}
\end{table}
Table 2: Tunneling splittings \(\Delta^{\rm SP}_{\rm RPI}\) and \(\Delta^{\rm DP}_{\rm RPI}\) determined on the single- and double-precision PESs and \(\Delta^{\rm DP}_{\rm PC}\) including perturbative corrections (PC) for tropolone and PFD. All simulations used the largest available TL data sets. For comparison, results from experiments are also reported. ”n.d.” is ”not determined”.
The averaged double-precision tunneling splittings \(\Delta^{\rm DP}\) (averaged over 10 independent TLs) for tropolone and PFD are presented in Table 2. It is interesting to note that even the uncorrected tunneling splittings \(\Delta^{\rm SP}_{\rm RPI}\) and \(\Delta^{\rm DP}_{\rm RPI}\) differ slightly between single- and double-precision arithmetics. These differences are probably because single-precision does not give the Hessians to sufficient accuracy, leading to small errors in the fluctuation factors \(\Phi\). Comprehensive results to support this, including energy barriers, actions \(S\), fluctuation factors \(\Phi\), uncorrected tunneling splittings and correction factors \(c_{\rm PC}\) for the individual TLs are given in Tables S4, S5, S7 and S8.
For tropolone, the uncorrected (\(\Delta^{\rm DP}_{\rm RPI}\)) and perturbatively corrected (\(\Delta^{\rm DP}_{\rm PC}\)) tunneling splittings are in excellent agreement with experiment and are the most accurate results reported so far. Empirically, standard RPI theory for \(\Delta_{\rm RPI}\) was found to give splittings within \(\sim 20\%\) of fully quantum-mechanical calculations using the same PES for typical molecular systems, as long as the barrier height is significantly higher than the zero-point energy along the tunneling mode. For malonaldehyde and tropolone, the deviations from experiment are 16% and 10%, respectively, which is within the expected range. If the results are perturbatively corrected (\(\Delta^{\rm DP}_{\rm PC}\)), they are within 2% to 3% for malonaldehyde and tropolone, respectively. As such, contrary to earlier calculations that do not include perturbative corrections, the present calculations yield the "right answer for the right reason".
For PFD, however, the calculated splitting for H-transfer is roughly 50% larger than that found in the experiments, even after the perturbative correction is applied. This is reminiscent of results for FAD for which \(\Delta^{\rm H}_{\rm RPI}=0.014\) cm\({}^{-1}\) was found. The calculation compares with measured splittings between 0.011 and 0.017 cm\({}^{-1}\) (with 0.011 cm\({}^{-1}\) from microwave spectroscopy probably being the most reliable value), which also differs by more than 20%. In search of possible explanations, additional calculations were carried out to probe (i) limitations in the quantum chemical methods used (ii) PES representation errors and (iii) limitations inherent in RPI theory and the perturbative correction.
Item (i) can be partly assessed by rescaling the action \(S\) with a more accurate estimation of the instanton barrier height according to \(S_{\rm new}=S\cdot\sqrt{E_{\rm inst}^{\rm new}/E_{\rm inst}}\). The CCSD(T)/aug-cc-pVQZ instanton barrier for PFD optimized on a representative TL-PES is \(E_{\rm B}^{\rm new}=12.609\) kcal/mol and compares to a barrier of 12.222 kcal/mol on the TL-PES. Rescaling yields \(\Delta_{\rm RPI}\) (\(\Delta_{\rm PC}\)) of 0.0132 (0.0117) cm\({}^{-1}\) if \(c_{\rm PC}=0.89\) from \(\langle\)TL\({}_{3}\rangle\) (see Table S8) is used. Although the results are closer to the measurements, they still differ by more than 20% in particular for the leading order estimation. To put this into context, \(S-\)scaling was also applied to malonaldehyde. Using a representative PES (\(E_{\rm inst}=5.398\) kcal/mol, \(\Delta_{\rm RPI}=24.2\) cm\({}^{-1}\), \(\Delta_{\rm PC}=22.0\) cm\({}^{-1}\)) and rescaling the action using \(E_{\rm inst}^{\rm new}=5.608\) kcal/mol from CCSD(T)/aug-cc-pVQZ calculations yields \(\Delta_{\rm RPI}\) (\(\Delta_{\rm PC}\)) of 21.9 (19.8) cm\({}^{-1}\). Comparing the results with experimental measurements of 21.583 cm\({}^{-1}\) this is a clear deterioration for \(\Delta_{\rm PC}\) and shows that this simple scaling approach can only be considered as a rough estimate of the error. While inaccuracies in the level of theory cannot be ruled out completely, the errors for PFD remain significantly higher than for, _e.g._, malonaldehyde.
For point (ii) - deficiencies in representing the PES - it is noted that PhysNet was used with the same hyperparameters for both tropolone and PFD. Training on MP2 reference data yields comparable test set errors MAE(\(E\))/MAE(\(F\)) of 0.067/0.069 kcal/mol(/A) for tropolone and 0.034/0.133 kcal/mol(/A) for PFD (see Table 1). Hence, it appears unlikely that representation errors lead to larger deviations in the tunneling splittings only for PFD. The quality of the PES is also influenced by the choice of structures included. Although the structures for both tropolone and PFD were chosen following the same procedure and although \(\Delta_{\rm RPI}\) is converged, it cannot be ruled out that due to the soft intermolecular modespresent in PFD, additional sampling is required. To assess this, scans of representative soft modes and their comparison with _ab initio_ CCSD(T) energies are given in Figures S5 and S6 for PFD, which show good agreement.
Finally, computed tunneling splittings can be influenced by the semiclassical approximations underlying RPI theory, see point (iii). Given the excellent agreement between computations and experiment for malonaldehyde and tropolone and the inferior performance for the hydrogen-bonded dimer PFD, supplementary calculations were performed for FAD. Using a CCSD(T)/aug-cc-pVTZ level of theory PES for FAD and supplementing it with 667 additional structures along and around the instanton path yields \(\Delta_{\text{RPI}}^{\text{DP}}=0.0158\) cm\({}^{-1}\) and \(\Delta_{\text{PC}}^{\text{DP}}=0.0186\) cm\({}^{-1}\). Again and similar to PFD, this is consistently higher than the measured value of 0.011 cm\({}^{-1}\). Given that the soft modes are well captured by the PES (a comparison of the PhysNet energies with their _ab initio_ CCSD(T) energies is given in Figure S7) it seems likely that RPI theory faces challenges when applied to H-bonded dimers, for example, due to strongly anharmonic out-of-plane modes. This hypothesis can be tested in future work using non-perturbative approaches such as PIMD to calculate the tunneling splittings for a given PES.
## 3 Conclusion
The present work demonstrates that combining state-of-the-art electronic structure and machine learning techniques with RPI theory yields H- and D-transfer tunneling splittings in very good agreement with experiments. The calculations for tropolone are the most accurate to date due to the high-quality PES and because they include a perturbative correction, leading to the "right answer for the right reason." In contrast, the best previous calculation of 0.92 cm\({}^{-1}\), which was carried out without these corrections, probably benefited somewhat from cancellation of errors (i) incurred by the instanton approximation, and (ii) arising from the fragmentation scheme for obtaining a near-CCSD(T) quality PES. Key to the success in the present work is the combination of transfer learning and including the perturbative correction to RPI. Importantly, brute-force calculation of the required number of reference data (energies and forces) at the CCSD(T) level of theory for training a machine learning PES is unfeasible at present and for the foreseeable future. Likewise, the Gaussian Process Regression instanton approach would probably also not be possible as it requires computation of a few Hessians at the CCSD(T) level.
Given the excellent performance of TL combined with RPI theory and perturbative corrections for tunneling splittings of intramolecular H-transfer (malonaldehyde and tropolone) it is anticipated that the methods used in the present work can also be used to _quantitatively predict_ tunneling splittings for similar systems. This is particularly relevant because experimental information is only available for a handful of systems and reliable predictions facilitate exploration of as of now uncharacterized systems. In a yet broader context, together with complementary experimental information such as vibrational spectroscopic data, tunneling splittings provide valuable information to characterize the PES around and away from the minimum energy structure for further refinements, for example by using "morphing" techniques.
The strategy followed here is generic and applicable to molecules for which a few dozen very expensive _ab initio_ calculations for judiciously chosen geometries can be carried out. Application to yet larger molecules necessitates advancements in either hardware capabilities or electronic structure theory. Concerning the latter, considerable effort is put into developing linear scaling CCSD(T) methods, which are becoming competitive alternatives for accurate computations of molecules exceeding 30 atoms. Alternatively, similarly to the GEMS approach, machine learning PESs can be built from molecular fragments that allow the generation of _ab initio_ reference data at higher levels of theory. Since these approximate methods are inherently difficult to test, the framework and predictions for, _e.g._, tropolone provided here serve as a valuable benchmark for future applications. Finally, the present approach - with suitable adaptations in the data selection procedure - can also be applied to other observables that are computationally expensive to determine (_e.g._, quantum bound states of molecules or scattering cross sections) or be combined with other quantum dynamics methods including wavepacket propagation or PIMD simulations.
|
10.48550/arXiv.2407.21366
|
Accurate Tunneling Splittings for Ever-Larger Molecules from Transfer-Learned, CCSD(T) Quality Energy Functions
|
Silvan Käser, Jeremy O. Richardson, Markus Meuwly
| 5,656
|
10.48550_arXiv.1709.01611
|
## 1 Introduction
Solar energy is by far the most abundant renewable energy source, and harvesting it to produce electricity and "solar fuels" (e.g., molecular hydrogen) seems to be the most promising route in the transition to an energetically sustainable future. Silicon-based solar cells are already making a significant impact on worldwide energy production, but other photovoltaic technologies are being actively researched for the medium and long term. Organic photovoltaic (OPV) devices represent one of the alternatives, which could be attractive for some applications in view of the wide availability of raw materials, low production costs and printability on mechanically flexible substrates. Their key component is a thin semiconducting active layer consisting of a blend of an electron-donor (D) and an electron-acceptor (A), which may be conjugated polymers and/or small molecules. Today, OPV devices with10% power conversion efficiency (PCE) have been produced by several groups, and a record 12% PCE has been achieved with a ternary blend. This has been possible thanks to a careful selection of the materials--synthetic possibilities are almost limitless--and optimization of the blend structure and morphology by controlling the deposition methods and post-deposition treatments.
As early as 2009, the group of Lee and Heeger achieved near-100% internal quantum efficiency (IQE) in "bulk heterojunction" cells having 6% PCE, based on a low-bandgap donor copolymer and a fullerene-based acceptor. IQE measurements are somewhat difficult and therefore they are not usually performed in experimental studies, but it seems likely that IQE's exceeding 90% should be achieved in all the current state-of-the-art devices. Such high values imply that virtually every absorbed photon--the IQE is actually a function of their wavelength--is successfully converted into a negatively charged electron (transported through the A material to the cell's cathode) and a positively charged hole (transported through the D material to the cell's anode). In turn, this implies a near-100% success in each of the processes which follow the formation of an exciton by photon absorption within either phase. According to the conventional wisdom, these are the diffusion of the exciton to the D-A interface, its dissociation into a "bound" electron-hole pair or charge-transfer (CT) exciton, their separation into free charge carriers and the migration/collection of the latter at the electrodes. A truly remarkable result, which apparently defies simple "classical" explanations: The attractive interaction between a positive and a negative point charge at 1-2 nm distance in a medium with relative permittivity \(\varepsilon_{r}\)=3-4 is 0.2-0.5 eV, which is much greater than \(k_{B}T\)=0.025 eV at room temperature. This and other observations have prompted the suggestion that to understand organic photovoltaics it is essential to invoke general quantum mechanical principles of delocalization, coherence and uncertainty, and that it might be possible to enhance the performance of OPV devices by properly harnessing them.
Spectroscopic experiments have been carried out with a range of methods, allowing the characterization of the relevant species--excitons and polaronic charge carriers, as well as interfacial CT states--with increasing detail. Ultrafast pump-probe experiments showed that high-energy, "hot" CT states (CT\({}_{n}\)) tend to dissociate faster than the lower-energy ones (CT\({}_{1}\)). A higher rate of charge separation was linked to a higher degree of quantum mechanical coherence and delocalization and was assumed to translate into a higher overall dissociation efficiency. However, other experiments could be interpreted more conventionally in terms of a slow (on the ps scale, in comparison with the fs scale of the previous ultrafast studies), diffusive dissociation of "classical" charge carriers. Salleo, Neher and coworkers have reported that "cold" CT\({}_{1}\) states produced by direct, weakly allowed absorption from the ground state dissociate just as efficiently (or as inefficiently, depending on the D:A combination) as the higher energy ones. On the theoretical front, charge photogeneration has been modelled by accurate excited-state or time-dependent calculations on few-molecule systems, microelectrostatic or quantum chemical calculations on larger D:A aggregates produced by molecular dynamics simulations, or Kinetic Monte Carlo (KMC) and Master Equation (ME) simulations at the scale of whole OPV devices. These methods have complementary strengths and weaknesses, but overall it has proved difficult to combine them to provide a general, fully satisfactory answer to the long-standing question: "Why is exciton dissociation so efficient at the interface between a conjugated polymer and an electron acceptor?".
To sum up, several candidates have been identified as likely "facilitators" of CT dissociation: built-in electric fields at D:A interfaces, delocalization of the excitons and of the charge carriers, high charge mobility, energetic variability and structural disorder, domain size and degree of intermixing of the D and A "phases", non-linearity or inhomogeneity of the dielectric medium, excess energy of the photogenerated excitons. All these factors seem have some importance, but probably not equally so. Besides, some of them are likely to be incompatible with each other (e.g., disorder and delocalization/mobility of the charges).
Here we present fresh theoretical insights based on our effective two-orbital quantum chemical model, which provides a "minimal" but theoretically sound description of OPV materials. It is similar in spirit to those of Troisi, Bittner and Silva and Ono and Ohno, but it can be applied to much larger systems. Thus, the model can provide a meaningful description of OPV operation at the mesoscale (10 nm or higher), which is crucial to account for the effect of blend morphology. Unlike Bittner and Silva, the present version of our model does not account the coupled electron-nuclear dynamics, which are responsible for decoherence phenomena. On the other hand, several of the previous facilitators of CT dissociation may be readily introduced in a calculation, allowing a systematic and unbiased assessment of their relative importance.
## 2 The model
We model portions of a photoactive layer consisting of equal number of D and A sites. Overall there are \(M=12\times 12\times 12=1728\) sites, arranged on a simple cubic lattice with a spacing of 1.0 nm. Figure S.1 in the Supporting Information shows the structure of our six model heterojunctions. The simplest one has a planar interface between the D and A sites, the others present some interpenetration between the phases to give a "comb" morphology (the systems have been named N\(n\)T\(t\), according to the number \(n\) of pillars and the thickness \(t\) of the intermixing region). In general, each site may represent a whole molecule, or a \(\pi\)-conjugated section of a long polymer chain. There are two electrons and two orbitals per site, representing its highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO). This picture is similar to the one adopted in KMC simulations of OPV devices, but here the electronic states are derived from a proper quantum-mechanical description, without any assumption about their localization or delocalization.
The on-site parameters of our model Hamiltonian (the HOMO and LUMO energies, as well as the Coulomb and exchange interactions among the electrons) can be chosen to reproduce exactly the energies of the main electronic states of the individual sites/molecules:lowest singlet excitation energy (\(SX\)), lowest triplet excitation energy (\(TX\)), ionization energy to form a cation (\(IE\)) and electron affinity to form an anion (\(EA\)). In our calculations, we employ a set of on-site parameters which correspond to C\({}_{70}\) for the acceptor and pentacene for the donor. Taking the gas-phase experimental data as a starting point, the effect of the surrounding dielectric on the ionized states has been obtained from the Born formula for the solvation free energy of an ion. All these energies have been collected in Table 1. Alternatively, this information about the single-molecule states could be obtained by conventional quantum chemical calculations, which can account for the surrounding dielectric by a polarizable continuum model.
The inter-site part of the Hamiltonian consists of a one-electron part, describing off-diagonal orbital couplings and the interaction with the positively charged cores of the other sites, and a two-electron part. The orbital couplings are assumed to be essentially random and decay exponentially with inter-site distance. The inter-site electron-electron and electron-core interactions are approximated by the electrostatic interaction between two spherical Gaussian charge distributions, embedded in a dielectric medium with relative permittivity \(\varepsilon_{r}=3.5\). As side note, we point out that situations with degenerate or near-degenerate HOMO and LUMO levels occur frequently in fullerene-based and other materials. These could be modelled within our coarse-grained model by connecting three or four sites into "super-molecules" with \(D_{3h}\) or \(T_{d}\) symmetry. Two sites could be considered to be connected when their coupling is roughly one order of magnitude larger than that between unconnected ones. Orbital degeneracy might indeed have an effect on charge separation and transport, and we hope to study it in the future.
Both diagonal and off-diagonal disorder can be introduced in a controlled way, by admitting local deviations in the orbital energies and couplings. Our method allows independent, essentially unrestricted variations in their relative sizes. Here we model them as random numbers, drawn from Gaussian distributions with standard deviations \(\sigma_{w}=0.08\)eV (for diagonal energetic disorder) and \(\sigma_{t}=0.08\)eV (for off-diagonal coupling disorder). Thesevalues are comparable to those which arise in the calculation of charge mobilities in organic semiconductors. For a given arrangement of D and A sites, different realizations of the disorder can be generated by simply re-assigning these energies and couplings, starting from a different random number seed. Typically, in order to extract systematic trends from our calculations, we consider one hundred independent realizations of the disorder.
The ground state energies, wavefunctions and charge distributions have been obtained by self-consistent-field, restricted Hartree-Fock (HF) calculations. The analogous excited state properties have been obtained by configuration interaction calculations including all the single excitations (CIS) from the \(M\) occupied to the \(M\) virtual HF orbitals. CIS can be considered an excited-state extension of the HF method, as both of them neglect electron correlation effects. Note that the singly excited configuration do not contribute to the ground-state wavefunction, as a consequence of the variational nature of the HF solution (Brillouin's theorem). We have also evaluated the effect of non-dynamical electron correlation, by comparing the results of CIS and high-level excited-state calculations (equation-of-motion coupled-cluster singles plus doubles, or EOM-CCSD). These results will be presented at the end of the following section. All calculations have been carried out with a modified version of GAMESS-US. Further details are given in the Supporting Information.
We consider the lowest energy states obtained from the CIS calculations to be close theoretical relatives of the cold CT\({}_{1}\) states mentioned in the Introduction. Our states are coherent or "pure", being described by a stationary electronic wavefunction instead of a density matrix.
\begin{table}
\begin{tabular}{|l|c c|c c|} \hline Energies & \multicolumn{2}{c|}{Vacuum} & \multicolumn{2}{c|}{Dielectric} \\ & A & D & A & D \\ \hline \(IE_{r}\) & 7.48 & 6.61 & 6.45 & 5.58 \\ \(EA_{r}\) & 2.68 & 1.35 & 3.71 & 2.38 \\ \(SX_{r}\) & 2.44 & 2.28 & 2.44 & 2.28 \\ \(TX_{r}\) & 1.56 & 1.76 & 1.56 & 1.76 \\ \hline \end{tabular}
\end{table}
Table 1: Single-molecule energies (in eV) for the acceptor A (C\({}_{70}\)) and the donor D (pentacene), in the gas phase or within a dielectric with \(\epsilon_{r}=3.5\).
These phenomena have been modelled at the atomistic level by calculating explicitly the time-dependence of the nuclear coordinates and electronic wave function. Clearly, this approach cannot be directly applied to the present model. However, a time-dependent extension could be attempted by introducing a dependence of the orbital energies and couplings on some generalized intra- and intermolecular phonon coordinates.
## 2 Results and discussion
As a preliminary step to the discussion of the electronic states at the heterojunctions, we provide some data on the pure materials. These have been obtained by HF and CIS calculations on blocks of \(M\)=1728 D-only or A-only sites. Considering for example the \(i\)-th realization of a donor block (\(D_{i}\)), the relevant energies are:
\[\begin{array}{rcl}SX(D_{i})&=&E_{\rm CIS}(D_{i})-E_{\rm HF}(D_{i}),\\ IE(D_{i})&=&E_{\rm HF}(D_{i}^{+})-E_{\rm HF}(D_{i}),\\ EA(D_{i})&=&E_{\rm HF}(D_{i})-E_{\rm HF}(D_{i}^{-}),\end{array} \tag{1}\]
Table 2 gives the averages and standard deviations of these energy differences, obtained by calculation on several independent realizations of the disorder. The \(SX\) energies of A and D compare favourably with the optical band gaps of solid C\({}_{70}\) (1.66 eV) and pentacene (1.85 eV), respectively. At the same time, these energies are substantially lower than those of the single-molecule, on-site excitations (see again Table 1). We take this as an indication that delocalization effects are significant and they are reasonably well described with our choice of the orbital coupling parameters.
The \(IE\) and \(EA\) data in Table 2 are also interesting, as they allow us to estimate the average energy of an electron-hole pair at infinite separation. Putting the hole on the donor and the electron on the acceptor, we find:
\[E_{eh}^{\infty}=IE(D)-EA(A). \tag{2}\]
The result is \(E_{eh}^{\infty}=0.94\pm 0.24\) eV. This is roughly one half of the energy which could be estimated from the single-molecule data in Table 1, assuming fully localized charges (1.87 eV). This confirms the importance of delocalization effects, despite of the sizeable amount disorder which has been included in the model.
Each of them
\begin{table}
\begin{tabular}{|l|c c c|} \hline Material & \(SX\) & \(IE\) & \(EA\) \\ \hline A & 1.62\(\pm\)0.17 & 5.99\(\pm\)0.21 & 4.17\(\pm\)0.15 \\ D & 1.87\(\pm\)0.22 & 5.11\(\pm\)0.19 & 2.85\(\pm\)0.25 \\ \hline \end{tabular}
\end{table}
Table 2: Averages and standard deviations of the singlet excitation energies, ionization energies and electron affinities (in eV) of the pure materials.
Distributions of the first and second excitation energies (in eV) for all the heterojunction morphologies. The horizontal axis is the number of counts within an energy bin. The horizontal black line marks the average first excitation energy. The dashed red line marks the average energy of an electron-hole pair at infinite separation (\(E_{eh}^{\infty}\)).
The lowest excitation energies are of the order of 0.90 eV. The average excitation energies (continuous black lines in the Figure) do not seem to depend much on the interface morphology. These band gaps are substantially lower than those of the pure phases. At the same time, the CT\({}_{1}\) states are within only 0.04 eV from the average \(E_{eh}^{\infty}\) (dashed red lines in the Figure). Note also that histograms of first and second excitations overlap strongly, each having a width of the same order of magnitude (comparable to \(k_{B}T\)). Thus, the second (or third) excitation energy of one system may be lower than the lowest excitation energy of another. In the following, we will concentrate on the discussion of the CT\({}_{1}\) states, but one should always keep in mind that there is actually a near-continuum of states above them.
We have just established that, using a resonable set of Hamiltonian parameters, the lowest interfacial excited states are within \(k_{B}T\) from the infinitely separated charges, contrary to the "classical" expectations. Before we look at this finding in greater depth, it is necessary to discuss its general validity. Clearly, by a proper choice of the Hamiltonian parameters, our calculations can be tuned to reproduce the vertical excitation and ionization energies of the don or and the acceptor. Since the model does not account for geometrical relaxation following photoexcitation, our lowest excited states are "electronically cold" but "vibrationally hot", and we are certainly overestimating the transition energies that would be measured by fluorescence spectroscopy. On the other hand, the electron-hole pairs tend to be already well separated in these excited states (see below). It reasonable to assume that their vibrational relaxation energies are comparable to the sum of those of an electron within the A and a hole within the D. Thus, even though we are overestimating the energies of the relaxed interfacial excitons with respect to the ground state, we believe that the energies with respect to the fully separated charges should be roughly correct.
(upper panel) illustrates the charge distribution in the ground and the first excited states, for one realization of the flat bilayer.
At first sight, the charges in the two states seem rather similar. In fact, they seem to be almost random, with several negative values on the D side and several positive ones on the A side of the interface. This randomness of the charge distribution reflects the randomness of the underlying Hamiltonian, of course.
A clear picture starts to emerge by plotting the charge difference between the states, \(q_{k}^{X}-q_{k}^{0}\). (lower panel) shows several examples of the charge differences between the excited and the ground states, one for each instance of the model heterojunctions.
Upper panel: charge distributions in one realization of the bilayer, respectively for (a) ground state, (b) first excited singlet state. Lower panel: (c) charge differences between the excited and ground state, for one realization of each system. The red and blue shading of the background indicate the D and A sides of the heterojunctions, respectively, whereas the colouring of the sites indicates positive (red) or negative (blue) charges. For clarity, (c) shows only the sites with a charge difference greater than \(0.002e\) (in modulus).
We see two distinct charge pockets, which are well separated and reside entirely on the expected phases (D for positive, A for negative). Summing these charge differences we always obtain:
\[\sum_{k\in D}(q_{k}^{X}-q_{k}^{0})=-\sum_{k\in A}(q_{k}^{X}-q_{k}^{0})=1.00e\]
These photogenerated charges appear to be delocalized over 10-20 sites, but this estimate can increase to 40-50 sites depending on the threshold adopted for their individual values. Of course, this sizeable delocalization reflects the situation immediately following photoexcitation, before the nuclear motions set in to produce decoherence and further localization.
Because of the large variability in the charge distributions associated with different realizations of the Hamiltonian, it is necessary to average hundreds of them in order to extract further systematic trends.
Two-dimensional plots of the average ground state charge distributions, \(\langle q_{k}^{0}\rangle\). The colour scale is such that saturation of the red and blue occurs when the charges exceed \(0.0025e\), in modulus.
These have been extracted from the six layers at the center of the blocks, excluding those at the top and at the bottom (in the orientation of Fig. 2) in order to minimize boundary effects. These charge density plots may also be taken to represent the result of an incoherent superposition of many quantum mechanical states, as all phase information is discarded when averaging their charge distributions. The ground state maps demonstrate the formation of sharp, well-defined interfacial dipoles. The average charge on a site sitting at the D/A interface is \(0.007e\) (in modulus). Remembering that our sites are spaced by 1 nm, this is at least comparable with the interfacial charge densities of \(0.02e/\mathrm{nm}^{2}\), as estimated by Bassler and Kohler on the basis of the typical voltage drop at a heterojunction. The charges located on the sites away from the interface, although appreciable within a single realization of the disorder (see again Figure 2), tend to cancel each other upon averaging.
These plots tend to be noisier than those for the ground states, but even so it is possible to draw some interesting conclusions. Naively, considering that we are looking at the lowest energy excitations, one would have expected the photogenerated charges to be "squeezed" at the interfaces by their mutual attraction.
Two-dimensional plots of the average photogenerated excess charge distributions, \(\langle q_{k}^{X}-q_{k}^{0}\rangle\).
densities tend to be spread almost everywhere, except perhaps at the external boundaries of the blocks (good news, as this implies that finite-size effects are reasonably under control). In particular, there is a significant photogenerated charge density within the first 2-3 layers from the interface. The subsequent time-dependent evolution will produce some localization of the photogenerated changes. Since we cannot simulate it with the current model, we can only speculate that their average distribution will still resemble the one in In any case, the figure highlights that a truly unbiased model of charge photogeneration should include hundreds of donor and acceptor molecules, and this in currently incompatible with a detailed atomistic description of the coupled electron-nuclear dynamics.
The electrostatic repulsion between the newly generated charges and the ground state interfacial dipoles--or, in other words, the built-in electric field--seems to be a key factor in enhancing this electron-hole separation. Exploratory calculations with other parameter sets show that an increased conjugation (introduced in the form of higher inter-obital couplings) and a lower disorder within the phases tend to produce an even greater electron-hole separation, up to a point where the photogenerated charge densities are almost zero on the D/A sites which are directly in contact. Thus, the calculations confirm the idea that both interfacial dipoles (a classical effect) and delocalization (a quantum-mechanical effect) can act as "facilitators" of charge dissociation.
Here the positions of the electrons and holes were identified by computing the center-of-charge of the excess photogenerated charges (negative and positive, respectively). The plots show that increasing the interfacial thickness from two (N4T2) to four or six nm's (N4T4 and N4T6) should have a beneficial effect on charge separation, by shifting the distribution of electron-hole distances toward larges values. If confirmed, this observation could turn into a useful design rule for the choice of photovoltaic materials and their assembly within the active layer.
Finally, we validate the use of the CIS method by comparing it with EOM-CCSD excited state calculations. We have carried out benchmark calculations on two systems: A two-dimensional heterojunction consisting of 15 D and 15 A sites, and a three-dimensional heterojunction consisting of 18 D and 18 A sites. shows the results for two independent realizations of each system. In comparison with EOM-CCSD, CIS systematically underestimates the excitation energies by about 0.05 eV. This applies to all the lowest excited states. As a results, one excitation spectrum is simply a shifted version of the other. Note also the significant changes in the excitation spectra, produced by simply changing the random seed for the assignment of the orbital couplings. Once more, this confirms the importance of performing extensive statistics before drawing any conclusions. The insets in show also the EOM-CCSD charge distribution in the lowest excited state of each system. In all cases, the CIS charge distributions (not shown) are virtually indistinguishable from them. Thus, the neglect of electron correlation in the description of the excited states of these systems appears to have solid foundations. This is certainly good news also for those performing atomistic (ab initio or semiempirical) excited-state calculations.
Distribution of the electron-hole distances for selected interfacial morphologies.
## Conclusions
In conclusion, we have presented the results of mesoscale quantum mechanical calculations of the lowest excited states at organic photovoltaic interfaces. Despite of its relative simplicity--or perhaps thanks to it--the model allows an unbiased exploration of the effect of several possible facilitators of charge photogeneration, such as built-in or external electric fields, charge delocalization, disorder, nanostructuring of the donor-acceptor blends. The nature of the problem requires large numbers of calculations on large systems, as it is necessary to avoid assumptions about the localization of the photogenerated charges, minimize finite-size effects and perform adequate statistical averaging of the results. We have explored only a small fraction of the parameter space, for example using C\({}_{70}\) as the only acceptor and pentacene as the only donor. Nonetheless, the results are already very encouraging, as they show that fully separated charges (electron and hole) are easily within 1-2 \(k_{B}T\) from the lowest excited states, in agreement with much experimental evidence but contrary to simple "classical" arguments.
Comparison of CIS and EOM-CCSD calculations of the first ten excited states, in two-dimensional (above) and three-dimensional (below) model heterojunctions. The insets show the EOM-CCSD charge distributions in the first excited states.
We find that a significant degree of delocalization of the excited charge transfer states (tens of molecules/sites) may occur also in presence of some energetic and coupling disorder. On the other hand, nuclear relaxation phenomena are not included in our treatment, and as a result the real situation might be somewhere in between our description and the fully localized one which is assumed in classical KMC simulations of OPV devices.
Our model is somewhat generic, as it retains a minimum of physically important features and neglects most molecular details. In this sense, it is analogous to many classical coarse-grained models of soft materials (polymers, liquid crystals, colloids, etc.), in which atomic level details are sacrificed in favour of generality, lower computational cost and greater interpretability of the results. Considering the relative ease of achieving \(>90\%\) IQE's (relative to the difficulty of achieving high charge mobilities in organic semiconductors, for example), it seems that the photogeneration of charges is precisely a situation where a somewhat generic explanation is required. This contrasts with other aspects of the behaviour of organic photovoltaic materials. Compare it, for example, with the difficulty of achieving high charge mobilities in molecular or polymeric semiconductors. This has eventually been possible thanks to a lengthy and painstaking selection of molecules and processing conditions, providing the optimal combinations of molecular structure and supramolecular organization. Clearly, the latter is a situation where detailed molecular-level modelling is necessary in order to support and guide the experimental efforts. In any case, our calculations provide guidelines also for those performing more conventional, atomistic excited state calculation at D/A heterojunctions. There is a clear need to move towards models incorporating hundreds of molecules in order to remove any bias about the degree of localization. This is extremely challenging, considering also the need to average over many disordered configurations, but at least our coupled-cluster calculations show that these calculations should not require the inclusion of dynamical electron correlation.
We conclude with a brief perspective on future developments. Although here we have considered idealized on-lattice models of the blends interfaces, more realistic off-lattice models are fully within our reach. It should also be interesting to examine the consequences of a polymer-like connectivity of donor or acceptor sites, to form long conjugated chains. Bittner and Silva have already considered a two-dimensional system in which the both the donor and acceptor are polymeric. In their lattice model, all the chains are parallel to the interface and this produces low-energy charge transfer states in which the electron and hole are delocalized but "pinned" to the interface. However, the situation in a real system will be more complicated, and even a partial orientation or \(\pi\)-stacking of the chains in the orthogonal direction might have a strong beneficial effect on charge separation. In the longer term, the present model should be extended by incorporating electron-phonon coupling and nuclear relaxation effects, within an explicitly time-dependent picture of system. One day, with further developments along the lines of the present study, it might even become possible to simulate the operation of whole OPV devices by adding a quantum mechanical description of exciton diffusion, charge transport and charge extraction/injection at the electrodes.
|
10.48550/arXiv.1709.01611
|
Origin of Charge Separation at Organic Photovoltaic Heterojunctions: A Mesoscale Quantum Mechanical View
|
Mosè Casalegno, Raffaele Pastore, Julien Idé, Riccardo Po, Guido Raos
| 3,655
|
10.48550_arXiv.1108.2034
|
###### Abstract
We investigate the accuracy and efficiency of the semiclassical Frozen Gaussian method in describing electron dynamics in real time. Model systems of two soft-Coulomb-interacting electrons are used to study correlated dynamics under non-perturbative electric fields, as well as the excitation spectrum. The results show that a recently proposed method that combines exact-exchange with semiclassical correlation to propagate the one-body density-matrix holds promise for electron dynamics in many situations that either wavefunction or density-functional methods have difficulty describing. The results also however point out challenges in such a method that need to be addressed before it can become widely applicable.
## I Introduction
Accurately capturing the correlated motion of electrons in atoms, molecules, and solids remains an active research area today. In ground-state electronic structure problems, the difficulties of solving many-electron systems are rolled up into the correlation energy, a.k.a. _the stupidity energy_. It has been examined in many different scenarios and various limits, but its complexity still haunts us today. Essentially representing the deviation of the many-electron wavefunction from an antisymmetrized product of single-particle orbitals, correlation plays a significant role in time-dependent problems as well. When atoms and molecules are exposed to either perturbative or non-perturbative external fields, fascinating and subtle electron interaction effects mean one must go beyond the "single active electron" picture, yet to solve the time-dependent Schrodinger equation (TDSE) for more than two electrons in strong fields pushes today's computational limits.
The advent of time-dependent density functional theory (TDDFT) in 1984 opened up the possibility of using single-particle orbitals to describe the dynamics of interacting electrons exactly. In TDDFT a single slater determinant (SSD) in a non-interacting system (the Kohn-Sham (KS) system) is propagated in time, such that it reproduces the exact time-dependent density of the true interacting system. Clearly the KS determinant is not the true interacting wavefunction, nor is it supposed to be an approximation to it; nevertheless by the rigorous theorems of TDDFT all observables of the true, correlated, electronic system can be in principle extracted from it exactly. In practise, approximations are needed for the exchange-correlation functionals, limiting the accuracy of the method. TDDFT is a method of choice in the linear-response regime where the photo-spectra of atoms, molecules and clusters can be found accurately. The computational effort involved scales relatively well, so the spectrum of large systems such as the green fluorescent protein, or even candidates for solar cells, have been studied.
Although it has proved successful for a wide variety of time-dependent problems, its progress for real-time dynamics in non-perturbative fields has been slower. Three major obstacles involve the lack of memory in the commonly-used functional approximations, lack of good functional approximations to extract observables not directly related to the density, and the inaccuracy of the usual functional approximations when the true wavefunction evolves far from a SSD.
Recently a method that accounts for electron correlation semiclassically has been proposed that will, in principle, remedy the aforementioned problems of TDDFT. The method operates within the context of time-dependent density-matrix functional theory (TDDMT), where the one-body density matrix is propagated in time. The second-order density-matrix enters into the equation for the one-body density-matrix, so is approximated as a functional of the one-body density-matrix in TDDMT. However it is still difficult to find approximations in TDDFT that work well when the true wavefunction evolves from being close to a SSD to being far from one, i.e. those developed so far cannot dynamically change "occupation numbers". In the proposed method of Ref., a separate semiclassical propagation is done from which the correlation potential is extracted, and used to drive the actual propagation. The approach naturally contains memory effects carried along by classical trajectories, all one-body observables may be immediately extracted, and occupation numbers do dynamically evolve.
Whether the proposed method is accurate for a given application however depends on how close the semiclassical dynamics is to the true dynamics. In the method, only the correlation component of the dynamics is treated semiclassically, while all other terms (including Hartree and exchange), are treated exactly. However, if the semiclassical dynamics deviates far from the true dynamics, the proposed method is unlikely to be accurate. Therefore it is worth investigating how well a semiclassical description of the _entire_ dynamics is for the problems of interest. This is the goal of the present paper. We compute the semiclassical dynamics (within the Frozen Gaussian approximation) of model two-electron systems where the exact dynamics can be numerically exactly computed, so that we cantest the accuracy of the semiclassical dynamics against this. We can also compare with the exact KS system: that is, using the exact exchange-correlation potential (that yields the exact density), but comparing observables such as momentum-densities, computed directly from the KS SSD (instead of using the unknown appropriate observable-functionals as in exact TDDFT). Our examples model the following phenomena: states of double-excitation in spectra, dynamics in a strong oscillating field, and population transfer to excited states via resonant driving, or via an optimal field. The chosen systems are one-dimensional models of a quantum dot and a helium atom.
We show both successes and failures of the semiclassical propagation for these applications. We expect however that the proposed method of Ref. is more accurate than such a semiclassical treatment of the entire problem since there it is only the correlation component that is approximated semiclassically. So the results of our studies are expected to improve once applied in the context of the method of Ref..
Our results also have relevance for semiclassical studies of dynamics in their own right (i.e. outside of density-functional-type methods). Although semiclassical treatments of dynamics have seen much interest for nuclear dynamics (eg. Refs.), there have only been a few applications to electron dynamics and almost all involved a single electron. Part of the reason is that in atoms there is a significant probability that classical trajectories will autoionize after a few cycles in the atomic well: through the electron interaction, one electron gains energy from the other and ionizes while the other drops below its zero-point energy; a process that is classically allowed but quantum-mechanically forbidden. This is a manifestation of what has been called the "zero-point-energy" problem, and happens also when there are coupled vibrational modes, for example. These trajectories create a lot of noise in the semiclassical sum, making the convergence very difficult, and so are typically terminated. For example, in Ref., accurate spectra of the He atom in both its collinear \(Zee\) and \(eZe\) configurations were obtained, by discarding trajectories where one electron reaches a certain threshold distance. Theoretically, a question is whether fundamentally these trajectories should be excluded from the semiclassical sum. That is, if the semiclassics could be done with an infinite number of trajectories, whether these trajectories incorrectly contribute to the semiclassical sum. We find that indeed their contribution to the semiclassical sum decreases as more trajectories are added: in this sense, they are valid classical trajectories, and via phase-interference, the semiclassical dynamics restores the zero-point-energy, eliminating their contribution from the semiclassical sum. In practise, nevertheless, we must deal with a finite number of trajectories, so sensible ways to discard these trajectories must be devised. Our results on the model atoms, as well as those on a model quantum dot, lend support to the conclusions of Ref. that semiclassical dynamics is useful for electrons, for both spectra and dynamics in non-perturbative fields (in some cases). A draw of semiclassics is the interpretive power it carries and we hope to use it also to to interpret mechanisms for processes involving interacting electrons, e.g. the role of correlation in multiphoton ionization and multielectron ionization.
The "flavor" of semiclassics explored in the current work is the Frozen Gaussian (FG) dynamics, originally proposed by Heller. This is not one of the rigorous semiclassical propagators that satisfy the TDSE up to order \(\hbar\); it is however closely related to the rigorous Heller-Herman-Kluk-Kay (HHKK) propagator, which is based on a coherent-state representation of the semiclassical propagator. In Section II, after a brief review of TDDFT and TDDMFT that explains the motivation behind Ref., we review some background regarding FG dynamics and how it is proposed to be used in the method of Ref. in density-matrix propagation. Section III presents results on one-dimensional model systems of two electrons, using FG for dynamics and spectra. Finally, we make some conclusions in Sec. IV.
## II Background
In this section, we briefly review TDDFT, the motivation behind going from TDDFT to a density-matrix approach, why we consider a semiclassical approach to correlation, and semiclassical dynamics.
### TDDFT and its challenges
TDDFT is an exact reformulation of the non-relativistic time-dependent quantum mechanics of many-body systems. It operates by mapping the true problem of interacting electrons into a fictitious, non-interacting system of fermions, called the Kohn-Sham (KS) system, whose one-body density is, in principle, exactly that of the true system, and from which, in principle, all properties of the true system can be obtained. The many-body effects of the interacting system are modelled by a one-body potential, called the exchange-correlation potential \(v_{\textsc{xc}}\). Similar thus in flavor to ground-state density functional theory, its functionals are nevertheless quite different. For example, \(v_{\textsc{xc}}\) functionally depends on the entire history of the density, as well as the initial-state of the true system and the initial state of the KS system, \(v_{\textsc{xc}}[n,\Psi_{0},\Phi_{0}](\mathbf{r},t)\). This _memory-dependence_ is however lacking in most of the applications today: they use an "adiabatic" approximation, which simply feeds the time-evolving density into a ground-state approximation:
\[v_{\textsc{xc}}^{\rm adia}[n,\Psi_{0},\Phi_{0}](\mathbf{r}t)=v_{\textsc{xc}}^ {\rm gs}[n](\mathbf{r})|_{n(\mathbf{r})=n(\mathbf{r}t)}. \tag{1}\]With this simple approximation, TDDFT has had great success in calculating spectra and response: computationally it scales similar to methods such as time-dependent Hartree-Fock (TDHF) or configuration-interaction singles, but with accuracy that is usually far greater. Yet, this simple approximation is also behind why in some cases approximate TDDFT fails: e.g. to capture states of double-excitation character a frequency-dependent kernel is required, but the adiabatic approximation yields a frequency-independent kernel.
The lack of memory in the usual functional approximations is also one of the reasons why the application of TDDFT to real-time dynamics applications has not progressed so fast. Model systems for which exact results are available indicate that memory-dependence can sometimes be a crucial factor in the exchange-correlation potential. Another difficulty in real-time dynamics, is the problem of "observable-functionals": although in principle all properties of interest can be extracted from the KS system, it is not known _how_ to extract those that are not directly related to the density. One may write these observables as operators on the exact wavefunction, however substituting the KS wavefunction is usually a poor approximation. For example, for momentum distributions, it was shown for a model system with one electron ionizing that the KS momentum distribution incorrectly develops strong oscillations as the electron moves away. In the case of ion-momentum-recoil upon ionization of a model He atom, the KS momentum-distributions were found to be drastically wrong, displaying a single maximum instead of the characteristic two hump structure, and with a significantly overestimated magnitude. Double-ionization probabilities are another example where a more sophisticated observable functional is needed.
A third problem arises when the true wavefunction evolves far from an SSD. An illustration of this is in certain quantum control problems, where a laser pulse is found to take the system to a specified target e.g. populating an excited state. For example, a pulse can be found to move the \(1s^{2}\) ground state of Helium to the \(1s2p\) excited state. As TDDFT stays in a SSD, with a single doubly-occupied orbital, for all times, it has great difficulty in describing a state that fundamentally is best described by two SSDs. It should be emphasized here that in principle, TDDFT can describe this situation, however the exchange-correlation potential and observable-functionals are extremely difficult to approximate.
### Time-Dependent Density Matrix Functional Theory (TDDFT)
To overcome some of the difficulties of TDDFT, we may attempt to use the one-body reduced density matrix, defined below, as the fundamental variable. This has the advantage of containing more information than the density while retaining similar concepts and system-size scaling as TDDFT. An advantage is that functionals for the momentum distribution and kinetic energy are given exactly in terms of the density matrix. The first-order spin-summed reduced density matrix is defined as:
\[\rho({\bf r^{\prime}},{\bf r},t)=N\sum_{\sigma_{1}..\sigma_{N}} \int d^{3}r_{2}..d^{3}r_{N}\] \[\Psi^{*}({\bf r^{\prime}}\sigma_{1},{\bf r}_{2}\sigma_{2}...{\bf r }_{N}\sigma_{N},t)\Psi({\bf r}\sigma_{1},{\bf r}_{2}\sigma_{2}...{\bf r}_{N} \sigma_{N},t) \tag{2}\]
and can be diagonalized by the so-called natural orbitals, \(\xi_{j}({\bf r},t)\):
\[\rho({\bf r^{\prime}},{\bf r};t)=\sum_{j}\eta_{j}(t)\xi_{j}^{*}({\bf r^{\prime }},t)\xi_{j}({\bf r},t) \tag{3}\]
We can now interpret the quantum control example discussed at the end of Sec. II.1 as an issue with the NO numbers. The NOs of the true system will begin with one NO occupation near \(2\) and the rest close to zero, and will end with two occupation numbers close to \(1\). The KS system, beginning in the ground-state SSD, has only a single natural orbital which is doubly occupied and time-evolution in the one-body KS Hamiltonian means that the NO occupation number always remains at \(2\). By using the density matrix of the true system as basic variable instead, we allow the NO occupations to change and thus it should be better than KS at capturing the correct behavior with simpler functional approximations.
The equation of motion for the density matrix is known as part of the BBKGY hierarchy of equations:
\[i\dot{\rho}({\bf r^{\prime}},{\bf r};t) = \left(-\frac{\nabla^{2}}{2}+v_{\rm ext}({\bf r};t)+\frac{\nabla^ {\prime 2}}{2}-v_{\rm ext}({\bf r^{\prime}};t)\right)\rho({\bf r^{\prime}},{ \bf r};t) \tag{4}\] \[+\int d{\bf r}_{2}f_{\rm ss}({\bf r^{\prime}},{\bf r}_{2})\rho_{2} ({\bf r^{\prime}},{\bf r}_{2},{\bf r},{\bf r}_{2};t)\]
where \(f_{\rm ss}({\bf r^{\prime}}{\bf r},{\bf r}_{2})=1/|{\bf r}-{\bf r}_{2}|-1/|{\bf r ^{\prime}}-{\bf r}_{2}|\) and \(\rho_{2}\) is the second-order spin-summed reduced density matrix, conveniently decomposed as
\[\rho_{2}({\bf r^{\prime}},{\bf r}_{2},{\bf r},{\bf r}_{2};t) = n({\bf r}_{2})\rho({\bf r^{\prime}},{\bf r};t)-\rho({\bf r^{ \prime}},{\bf r}_{2};t)\rho({\bf r}_{2},{\bf r};t)/2 \tag{5}\] \[+\rho_{\rm c}({\bf r^{\prime}},{\bf r}_{2},{\bf r},{\bf r}_{2};t)\;.\]
Atomic units (\(e^{2}=\hbar=m_{e}=1\)) are used throughout this paper. Here the terms play the role of Hartree, exchange, and correlation respectively. Setting \(\rho_{\rm c}=0\) results in the TDHF equations. One could look to close this expression by expressing \(\rho_{\rm c}\) as a functional of \(\rho\) and this is the usual approach in TDDFT. One naturally seeks an adiabatic approximation in which the time-dependent density-matrix is fed into a \(\rho_{2}\)-functional bootstrappedfrom ground-state density-matrix functional theory Unfortunately, neither TDHF nor the adiabatic functionals lead to NO occupation numbers changing in time. Thus we must look further afield to try and fix this.
A possible solution to this was proposed in Ref. where \(\rho_{c}\) is given by an auxiliary semiclassical dynamics calculation running parallel to the density-matrix evolution. That is, we time-evolve \(\rho\) using Eq. 4 with Eq. 5, where all terms are treated exactly quantum-mechanically _except_ the last term \(\rho_{c}\) of Eq. 5. This last term is treated as a driving term in this equation, its value obtained from the parallel semiclassical propagation of the system. At each time, the second-order density-matrix, the density matrix, and the density are calculated in the semiclassical system, which then uses Eq. 5 to extract a semiclassical \(\rho_{c}\). This term is then inserted in Eq. 4 as a driving term. In Ref., it was argued that such an approach addresses the main obstacles TDDFT and adiabatic TDDFT approaches for real-time dynamics: memory is naturally carried along by the classical trajectories, initial-state dependence is automatically accounted for, all one-body observables may be obtained without the need for further observable-functionals, and NO occupation numbers are time-evolving. We can expect this method to most improve upon the purely semiclassical calculation when in the high-density limit, as the dominating exchange affects will be treated exactly.
Various candidates for the semiclassical method to use in this scheme are possible, with their numerical efficiency inversely related to their semiclassical rigor and accuracy. Most efficient and least accurate is to use a quasiclassical propagation of the one-body Wigner function, while least efficient but most rigorous, solving TDSE exactly to order \(\hbar\), is to use the Heller-Hermann-Kluk-Kay (HHKK) propagator. The Frozen Gaussian (FG) method may be viewed as an approximation to HHKK, and was argued to be a good candidate for use in a "forward-backward" scheme for \(\rho_{c}\). We focus in this paper therefore on FG dynamics, and our present goal is to investigate how accurate it alone is for electron dynamics. Although in the scheme of Ref., semiclassics is used only for the correlation component of the dynamics, the scheme is likely to be most accurate when the semiclassical propagation of the whole system is reasonably accurate. The purpose of the studies in Section III is therefore to study how accurate FG dynamics on model systems is.
### Frozen Gaussian Dynamics
Since the earliest days of quantum mechanics semiclassical methods have been explored, for the purpose of interpreting and understanding quantum mechanics via the more intuitive classical dynamics, and also for the purpose of approximation. These methods construct an approximate quantum propagator utilizing classical trajectory information alone. Although there are a variety of forms (e.g.), their essential structure is a sum over classical trajectories: \(\sum_{\mathrm{cl.traj}}G_{t}(t)e^{iS_{i}(t)/\hbar}\) where \(S_{i}(t)\) is the classical action along the \(i\)th trajectory, and the prefactor \(C_{i}(t)\) captures fluctuations around the classical path. Miller showed the equivalence of different semiclassical representations within stationary-phase evaluation of the transformations. Semiclassical formulae have been derived both from largely intuitive arguments (e.g. Heller's frozen and thawed gaussians Ref.) as well as from rigorous asymptotic analyses of the quantum propagator (see e.g. Refs.) that satisfy TDSE to order \(\hbar\). The latter are based on taking the semiclassical limit of Feynman's path integral for exact quantum dynamics. This yields a propagator of van Vleck form, that requires solving a boundary value problem to find the classical paths eg. from \(x^{\prime}\) to \(x\) in time \(t\); transforming these to "initial-value" representations where instead a sum over an initial coordinate-momentum phase-space is performed, makes the numerics significantly more feasible, especially for longer times and more degrees of freedom.
Semiclassical dynamics captures quantum effects such as interference, zero-point energy, tunneling (to some extent), while generally scaling favorably with the number of degrees of freedom. As the propagator is constructed from classical trajectories, intuition about the physical mechanisms underlying the dynamics can be gained. Although mostly used for nuclear dynamics in molecules, and condensed phases, there have been a handful of applications to electrons. The reasons for the reluctance of applying semiclassics to electrons were recently discussed in Ref.. One is that the classical dynamics of interacting electrons are typically mixed with regular and chaotic regions. This makes the semiclassical sum over trajectories increasingly difficult to converge as time evolves. Another is the classical autoionization problem discussed in the introduction; we return to this in Sec. III.3. A third problem is that the classical equations of motion become singular at the nucleus, where the potential diverges Coulombically, requiring the need for regularization methods to be applied. As Ref. showed however, this is apparently only a problem when systems are treated in reduced dimensionality: in three-dimensions, no special techniques are required to deal with the Coulomb potential except for a set of measure zero trajectories which hit the nucleus head-on (and these trajectories can be safely discarded). Ref. showed that all these obstacles can be overcome to make semiclassical calculations of atoms and molecules quite practical.
The FG propagation we are exploring can be expressed mathematically as a simplified version of the HHKK propagator where the \(N\)-particle wavefunction at time \(t\) as a function of the \(3N\) coordinates we denote \(\underline{\underline{r}}\) is:
\[\Psi^{\rm FG}(\underline{\underline{r}},t)=\int\frac{d{\bf q}_{0}d{\bf p}_{0}}{(2 \pi\hbar)^{N}}\langle\underline{\underline{r}}|{\bf q}_{t}{\bf p}_{t}\rangle e^{ iS_{t}/\hbar}\langle{\bf q}_{0}{\bf p}_{0}|\Psi_{i}\rangle \tag{6}\]
In Eq. 6, \(\langle\underline{\underline{r}}|{\bf q}{\bf p}\rangle\) denotes the coherent state:
\[\langle\underline{\underline{r}}|{\bf q}{\bf p}\rangle=\prod_{j=1}^{3N}\left( \frac{\gamma_{j}}{\pi}\right)^{1/4}e^{-\frac{\gamma_{j}}{2}(r_{j}-q_{j})^{2}+ ip_{j}(r_{j}-q_{j})/\hbar} \tag{7}\]
In HHKK, each trajectory in the integrand is weighted by a complex prefactor based on the monodromy (stability) matrix. The pre-factor is time-consuming to compute, and scales cubically with the number of degrees of freedom, but in the FG approximation, it is set to unity. As a consequence, although HHKK solves TDSE exactly to order \(\hbar\), the FG propagator does not; also, although it can be shown that HHKK results are independent of the choice of width parameter \(\gamma_{j}\), FG results are not. For all our calculations we take \(\gamma_{j}=1\). Neither the HHKK propagation nor FG are unitary; typically we find the norm of the FG wavefunction decreases with time, and so we renormalize at every time-step. (In fact for the typically chaotic dynamics of interacting systems, the HHKK prefactor, and norm, grows with time).
The phase-space integral in Eq. 6 is done by Monte Carlo, with the distribution of initial phase-space points weighted according to the initial distribution, \(|\langle{\bf q}_{0}{\bf p}_{0}|\Psi_{i}\rangle|^{2}\). Although Monte-Carlo methods scale as \(\sqrt{N}\) for positive integrands, the oscillatory phase can make the FG propagation very difficult to converge. The scheme of Ref. takes advantage of the "forward-backward" nature of the propagation of the second-order density-matrix (i.e in the second-order density matrix, there is both a \(\Psi(t)\) and a \(\Psi^{*}(t)\)), which leads to some cancellation of phase for more than two electrons. We also observe that the spatial-symmetry of the initial-state is preserved during the evolution (since the Hamiltonian is for identical particles, exchanging coordinate-momentum pairs of two electrons does not change the action). As we study here two-electron singlet states, the wavefunction is spatially-symmetric under exchange of particles.
## III Results+Discussion
We study the accuracy of FG propagation for several different systems involving two soft-Coulomb-interacting electrons in a spin-singlet in one-dimension. Because the TDSE for two interacting electrons in one-dimension can be mapped onto a TDSE for a single electron in two dimensions, the exact problem is easily numerically solvable, and we can compare our FG results to exact ones. The exact ones are obtained either using the approximate enforced time-reversal symmetry (aerts) algorithm coded in the octopus code, or our own exponential mid-point rule code (both use a fourth order Taylor expansion for the exponential). We also comment on how these compare with results using various common approximations in TDDFT for either the exchange-correlation functional (Sec. III.1), or for observable-functionals (Secs. III.2 and III.4).
In the latter case, we compute observables from usual operators acting directly on the exact KS wavefunction: this enables us to isolate errors arising from the observable-functional approximation alone (i.e. as opposed to errors from the approximation used for the exchange-correlation potential). For two electrons in a spin-singlet, given the exact time-dependent density, it is straightforward to obtain the exact KS wavefunction. We briefly review how.
\[\frac{\partial n(x,t)}{\partial t}+\frac{\partial j(x,t)}{\partial x}=0\, \tag{8}\]
and since the exact KS system reproduces the exact density at all times,
\[j(x,t)=j_{\rm KS}(x,t)=2\;\Re{\bf e}\;\phi^{*}_{\rm KS}(x,t)\frac{1}{i}\frac{ \partial\phi_{\rm KS}(x,t)}{\partial x}. \tag{9}\]
The last equality follows from the fact that for two electrons in a spin-singlet, there is just one spatial KS orbital \(\phi_{\rm KS}\) (doubly-occupied).
\[\phi_{KS}(x,t)=\sqrt{\frac{n(x,t)}{2}}\exp\left(i\int^{x}\frac{j(x^{\prime},t) }{n(x^{\prime},t)}dx^{\prime}\right). \tag{10}\]
So, the procedure is as follows: we solve the TDSE exactly numerically, finding the exact two-electron correlated wavefunction. The density and current are then extracted from \(n(x,t)=2|\phi_{KS}(x,t)|^{2}\) and Eq., and then inserted into Eqs.. Usual approximations for observables in TDDFT evaluate the operator corresponding to the observable directly on the KS state; although this is exact for observables directly related to the density, there are errors for other observables, e.g. the momentum-density, which is what we shall look at in examples III.2 and III.4.
In all cases, we will utilize the exact ground-state wavefunction to construct our initial state for propagation. This is to separate out any errors due to a poor ground-state. For a general case where the initial wavefunction is unknown exactly, an appropriate approximate initial state (e.g. sum of a few KS SSD's, or from a high-level wavefunction calculation) would be used.
### Hooke Model: Spectra
We begin with the Hooke model for a one-dimensional two-electron quantum dot defined by the Hamiltonian
\[\hat{H}_{0}=\sum_{i=1}^{2}\left(-\frac{1}{2}\frac{d^{2}}{dx_{i}^{2}}+\frac{1}{2}x _{i}^{2}\right)+\frac{1}{\sqrt{(x_{1}-x_{2})^{2}+1}} \tag{11}\]
The Hamiltonian becomes separable when written in relative \(r=x_{1}-x_{2}\) and center of mass \(R=(x_{1}+x_{2})/2\) coordinates, and we use the quantum numbers of these systems to label the excitations of the full system. In Table 1, we give the frequencies of the lowest \(8\) singlet excitations of this system, along with labels in \(R,r\). The \(r\) label must be even in order to have the correct spatial symmetry for a singlet state, while even/odd values for \(R\) determine the parity of the state. We have grouped the levels into multiplets whose members become degenerate if the electron-interaction is turned off. Each of the two states in the second and third multiplets are states of double-excitation character: they are mixtures of a single excitation (where, in a non-interacting reference, only one electron is promoted from its ground-state orbital to an excited one), and a double-excitation, where both the electrons occupy excited orbitals. The three states in the fourth multiplet are mixtures of a single-excitation and two double-excitations.
Such states of double-excitation character are well-recognized to be poorly described with the adiabatic approximation in TDDFT, which is used in almost all calculations: instead of producing a multiplet of states, only one state is obtained in the adiabatic approximation, and it has only single-excitation character. In order to find the correct frequencies, TDDFT requires a frequency-dependent kernel to mix the KS single and double-excitations, as was derived in Ref.. The kernel is an _a posteriori_ correction to an adiabatic approximation and although it has been recently tested on a variety of systems, it is not widely used; almost all codes still rely on the adiabatic approximation that fails to capture these. This is the motivation behind our first exploration of the FG dynamics: is semiclassical correlation able to capture states of mixed single and double-excitation character? The answer, as shall see shortly, is yes, at least approximately.
We compute the FG spectra by Fourier-transforming a time-propagation.
\[\Psi_{k}(x_{1},x_{2})=e^{ik(x_{1}+x_{2})}\Psi_{0}(x_{1},x_{2}) \tag{12}\]
where \(\Psi_{0}\) is the exact ground-state wavefunction, in order to populate the excited states and then to Fourier transform the time-dependent dipole moment, \(d(t)\), to find the dipole-spectra:
\[d(\omega)=\int dte^{i\omega t}d(t) \tag{13}\]
The _kick_ is equivalent to applying an impulse electric field, \(\delta v_{\rm ext}(x,t)=k\delta(t)x\), to the system. For The spectra found this way consists of peaks at the excitation frequencies, the widths of which becomes smaller the longer you propagate the system. For Hooke's model, this method needs a slight adjustment. In linear response, the _kick_ yields a perturbation that is proportional to \(R\), which only couples the ground state \(\) to the \(\) state, as the Hamiltonian is simply a harmonic oscillator in this coordinate; the harmonic potential has the special property that a dipole can only connect two states differing by one quantum number. This can also be seen as a consequence of the Harmonic Potential Theorem. So in the exact system, we only see a single peak in the dipole moment when using a linear kick. In order to see higher excitations, we use quadratic and cubic kicks and plot the quadrupole and third-order moments respectively. For example the quadratic kick, \(e^{ik(x^{2}+y^{2})}\), in the linear-response regime will be proportional to both \(R^{2}\) and \(r^{2}\), which, using the symmetry and parity discussion earlier, will couple most strongly to the \(\) and \(\) states.
We begin the FG calculation in the quadratically _kicked_ state (using the exact ground state) with \(k=0.01\) and use \(120000\) classical trajectories. These trajectories are propagated using the standard leapfrog verlet algorithm for a total propagation time of \(T=200\) a.u with the dipole moment calculated every \(0.1\) a.u. The power spectrum (absolute value squared) of the quadrupole moment is shown as the solid line in Fig. 1; we use the power spectrum as it is usually more stable than either the purely real or imaginary part, and we also apply the \(3^{rd}\) order polynomial filter given in Ref.
\begin{table}
\begin{tabular}{r|r r} \hline \hline Label (R,r) & \(\omega_{n}\) & \(\omega_{n}^{FG}\) \\ \hline
1,0 & 1.000000 & 1.0 \\
0,2 & 1.734522 & 1.6 \\
2,0 & 2.000000 & 2.0 \\
1,2 & 2.734522 & 2.6 \\
3,0 & 3.000000 & 3.0 \\
0,4 & 3.648334 & 3.5 \\
2,2 & 3.734522 & 3.6 \\
4,0 & 4.000000 & 4.0 \\ \hline \hline \end{tabular}
\end{table}
Table 1: The singlet excitation frequencies \(\omega_{n}=E_{n}-E_{0}\), where the ground-state energy \(E_{0}=1.774040\) a.u., for the Hooke model. For each excitation, the energy can be written \(E_{n}=E_{R}+E_{r}\), where \(R\),\(r\) are labels in the center of mass and relative coordinates respectively.
We observe two peaks, representing each of the two states in the second multiplet (,). The FG therefore does capture states of double-excitation character. The exact values of the and excitations are indicated on the graph as dashed lines. We see that FG gives the peak at \(\omega=2\) a.u. exactly right and the other is shifted lower to approximately \(\omega=1.58\) a.u.. The peak at 2 a.u. is purely an excitation in the center of mass coordinate, where, as mentioned, the system is purely harmonic. For harmonic systems, most semiclassical methods, including FG, are exact, and so it is expected that FG works well for this peak.
We reiterate that TDDFT in its usual adiabatic approximation, only yields _one_ peak; Ref., examined this pair of excitations using adiabatic exact exchange within the single-pole approximation and found, as expected, a single peak (of frequency 1.87 a.u.).
A quadratic perturbation yields zero transition probability to states of odd parity, such as in the third multiplet. To access these states, we cubicly _kick_ the system and find the third-order moment, also plotted in (dashed line). This is a harder test for FG as we are amplifying any errors far from the center in the decaying part of the wavefunction. Again we see that FG works well, giving three peaks corresponding to the first excitation, and the mixed single- and double-excitations in the third multiplet, with again the center-of-mass excitations given exactly. Looking at Table 1, we see that the FG correctly reproduces the property that the pair of peaks in the third multiplet differs from those in the second by a single excitation in the center-of-mass coordinate.
The peaks in the fourth multiplet are most clearly resolved by applying a quartic kick to the system. These three states involve mixtures of a single-excitation with two double-excitations, and FG provides a good approximation.
Our spectra could be made cleaner by running for longer times or using harmonic inversion methods for example. However they are adequate for our current purposes of illustrating that the FG method captures correlated states of mixed single and double-excitation character.
In conclusion, although the original motivation for a semiclassical description of electron correlation was to address challenges TDDFT has for real-time dynamics in non-perturbative fields, rather than for spectra, this study shows that semiclassical dynamics may nevertheless also be useful in the linear response regime. In particular, FG dynamics does capture states of double-excitation character, albeit approximately, missing in the usual adiabatic approximation of TDDFT. In comparison to the "dressed TDDFT" that uses a frequency-dependent kernel derived in Ref., the FG results are not as accurate, e.g. for the (,) multiplet, dressed TDDFT gives 1.712 a.u. and 2.000 a.u. On the other hand, in the dressed TDDFT approach, the procedure to obtain the double-excitations involves an identification by the user of zeroth-order single- and double-excitations that are likely to interact strongly, while in the present semiclassical approach, they emerge naturally from the dynamics.
We make two remarks at this point. First, a full HHKK treatment would likely yield more accurate semiclassical spectra, given that the prefactor missing in FG is complex, so its phase contributes interference effects important in determining the resonant frequencies. Second it is important to note that the proposed method of Ref. treats _only_ the correlation component of the density-matrix dynamics semiclassically, rather than of the entire dynamics as we have done here; this may result in improved accuracy of the spectra and we shall investigate this in the future. Given the more favorable scaling with system-size of FG than of HHKK, and given that it will ultimately be used in conjunction with exact kinetic, Hartree, and exchange components of the evolution, it is the FG method we are most interested in at present.
### Hooke Model: Natural Orbitals and Momentum Distributions
Next, we apply a resonant driving perturbation to the Hooke model dot, to investigate whether FG yields time-dependent NO occupation numbers accurately. Due to the harmonic potential theorem, applying an electric field does not change the natural orbital occupation numbers; instead we apply a quadratic perturbation, \(\delta v_{\rm ext}(x,t)=k(t)x^{2}/2\), with a time-dependent spring constant:
\[k(t)=0.05\sin(2\;t) \tag{14}\]
The power spectra of the Frozen Gaussian quadratic, cubic, and quartic moments for the Hooke’s model dot. The positions of the exact frequencies are shown as the vertical dashed lines.
In Ref., the NO occupation numbers for the Moshinsky model were calculated, however the Moshinsky Hamiltonian is completely quadratic and so, as mentioned, FG is exact. In the Hooke model case, the soft-Coulomb interaction between the electrons makes it a little more realistic model of a quantum dot, and a harder test for FG. In the following FG calculations for the system being driven with Eq., only \(60000\) classical trajectories were needed for convergence.
In Fig. 2, we plot the exact and FG NO occupation numbers. Recall that for both exact TDDFT and all common approximations in TDDFT, the occupation numbers are fixed by their initial values and cannot change. It is quite clear that despite not being perfect, FG works very well here, tracking the large changes in the occupation values. The quadrupole moment is shown in Fig. 3, showing that the FG density is also very accurate.
In an exact TDDFT KS calculation, the density would be exact, but the NO occupation numbers would remain at integers 2 (occupied orbital), and 0 for all the rest. We now show how this feature drastically affects observables that are not directly related to the coordinate-space density.
\[n(p;t)=\frac{2}{(2\pi)^{2}}\int dp^{\prime}\left|\int dxdx^{\prime}e^{i(px+p^{ \prime}x^{\prime})}\Psi(x^{\prime},x;t)\right|^{2}, \tag{15}\]
it is the probability of finding any one electron with momentum \(p\). Although it is known that the momentum distribution of the exact KS wavefunction is not that of the true system, little is known about how to extract the latter from the former. In the absence of a good observable-functional for momentum, one resorts to simply using the exact KS momentum distribution.
In Fig. 4, we show the exact, exact KS (computed directly from the exact KS orbital obtained via the procedure in Sec III), and FG momentum distributions at four snapshots in time. The first moment of the KS and exact distributions must integrate to the same value for all one-dimensional systems (in this case zero), due to the facts that \(\langle p(t)\rangle=\int dxj(x,t)=\int dxx\hat{n}(x,t)\), and that the KS exactly tracks the density. Despite this, the KS distribution oscillates wildly. The reason for this is not dissimilar to that seen in Ref.: the KS system becomes strongly non-classical as it attempts to describe the system of two-electrons using a single (doubly-occupied) orbital. For short times, the KS \(n(p)\) behaves well, however as the breathing motion of the system becomes more pronounced, moving out further and faster, it becomes an increasingly non-classical dynamics for a single orbital, such as in KS, to describe. The underlying phase-space distribution of the KS system develops strong oscillations into negative regions, signifying the non-classical description of different "parts" of one electron moving in different directions. The momentum-distribution represents one observable that the occupation numbers strongly influence.
### Soft-Coulomb Helium: The Problem of Classical Autoionization
Now we move from model quantum dots to model atoms: a soft-Coulomb interaction is used for the nuclear potential, as a one-dimensional model of a Helium atom (see e.g. Refs.).
The quadrupole moment for Hooke model dot driven at resonant frequency. Due to its highly oscillatory behavior, we show the exact values only as square points for clarity.
NO occupation numbers of the Hooke model dot driven at a resonant frequency of 2.0a.u.
With our electronic system no longer everywhere bound, we encounter the problem of classical autoionization, mentioned in the introduction. Even when _no_ field is present, this haunts the dynamics, as we will now show. We begin in the exact ground-state of the Hamiltonian Eq. with \(\epsilon(t)=0\). plots the positions \(x_{1},x_{2}\) of the trajectories in the initial distribution, and where they have evolved to at \(T=10\)a.u. The initial distribution centered at the origin increasingly evolves into a cross, signifying the classical autoionization: one electron zipping off to infinity after stealing the energy from the other left near the origin. Using just 20000 trajectories to compute the FG sum, we plot in the density at four snapshots in time. Autoionizing wavepackets are clearly seen, evolving out quasiclassically on either side of the central distribution. (Note that in the exact problem, nothing happens; the distribution remains centered at the origin, as the initial state is an eigenstate.) Although it may appear that the autoionizing packets are growing as a function of time, this is an artifact of the way the FG dynamics is renormalized (see Sec. II.3). Before renormalization, the FG norm decreases with time as more trajectories from the central distribution are lost to autoionization and other incoherent effects. Why some of the classically autoionizing trajectories seem to add coherently, forming the wavepackets on the sides, remains to be understood. However choosing a different initial seed to generate the initial distribution results in autoionizing wavepackets centered at different positions moving with different speeds. That is, certainly the results in are unconverged. In fact, in we show the effect of increasing the number of trajectories is to significantly decrease the classical autoionization. We deduce that by phase interference in the semiclassical sum, contributions from classically autoionizing trajectories cancel each other out. Considering that HHKK-semiclassics gives the correct dynamics to order \(\hbar\), this is as it should be: Although FG dynamics is not correct to order \(\hbar\) a similar effect happens here.
Although these results show that classical autoionization is not a fundamental problem of the semiclassical method, but rather a convergence issue, it does remain an important practical one for most applications; impossibly large numbers of trajectories could be required to make the classical autoionization effect small enough for the purposes. In our following studies we terminate trajectories in one of two ways. In the applications we consider, we do not expect significant amplitude beyond a certain distance. So, (i) we simply terminate any classical trajectory where the coordinate of one electron has reached a prescribed distance, \(L\), away from the nucleus. This is similar to what is done in Ref. in obtaining semiclassical spectra (where the termination distance was additionally dependent on the total energy of the trajectory). Another alternative is to use purely the energy terms to determine autoionization such as was done in Ref.. At all times, this means that we may still observe autoionizing wavepackets peeling off from the central distribution, that eventually reach \(L\), only after which they are discarded; therefore at the earlier times, these may erroneously contribute to the FG. We therefore also consider (ii) prescreening all trajectories, such that if they, at some point during the duration of interest, reach \(L\), they are discarded from the very start of the propagation. (Note this prescreening process is very fast, as only classical dynamics needs to be run; therefore many initial candidate trajectories may be explored). In this way no autoionizing trajectories (as defined by \(L\)) ever contribute to the FG sum. In our particular studies, there was some difference between the results obtained with (i) and (ii) especially at shorter times, with (ii) leading to narrower distributions than (i).
The exact (solid line), KS (short-dashed line), and FG (long-dashed line) momentum densities calculated at times \(75\), \(135\). \(160\). and \(200\) au for the on-resonant hooke system.
The kinetic energy for the exact, KS, and FG calculations, when the Hooke dot is driven on resonance. Although the KS system must reproduce the exact momentum, the kinetic energy becomes worse for longer times, while the error in the FG remains about the same.
In particular, does the earlier behavior of a trajectory that eventually autoionizes, contribute in a non-cancelling, sensible way to the dynamics at times prior to its autoionizing event? This is an important question we leave for future work.
### Soft-Coulomb Helium:Strong field
We now apply a relatively strong field to the model He atom:
\[\epsilon(t)=\frac{1}{\sqrt{2}}\cos(0.5t)\left\{\begin{array}{ll}\frac{t}{20} &t\leq 20\\ 1&t>20\end{array}\right. \tag{17}\]
The value of this field is large enough that classical autoionization effects are relatively small in comparison to the large oscillations in the dynamics induced by the field. We use the procedure (i) described above and terminate trajectories once \(|x_{i}|>30\)a.u. We found the results were essentially converged starting with \(2\times 10^{6}\) trajectories and finishing with \(5\times 10^{5}\), the rest being discarded by the termination condition at some point during the run. We plot the NO occupation numbers in Fig. 9, up to a time of 35 a.u. Notice that the number of trajectories needed for convergence in the soft-Coulomb He atom is much larger than those needed in the Hooke-model quantum dot; the latter is almost a best case scenario for semiclassics, because of the quadratic nature of the external potential. The FG captures the change in occupation numbers, perhaps a little too enthusiastically; shows the dipole is also reasonably well approximated. In Fig. 11, we plot the momentum distributions at several snapshots of time, comparing the FG with the exact, and with the exact-KS, as in Sec. III.2. We see that the error of the FG remains about the same throughout the evolution, while the error of the KS increases with time. This is a direct consequence of the change in NO occupation numbers: the KS momentum distribution is that of a SSD with NO occupation numbers 2 for one orbital and zero for all others, while that of the true and, roughly captured by FG, changes dramatically in time, as shown in Consequently, typical observables not directly related to the density, when evaluated using the usual operators on the KS wavefunction become badly approximated; the unknown observable-functionals of exact TDDFT are likely quite complicated.
The FG density at time \(t=25\) a.u. (with no field applied to the soft-Coulomb He atom). Adding more trajectories reduces the classical autoionization.
The positions, \(x_{1}\) and \(x_{2}\), of both particles plotted against eachother at the initial time (square points) and at time \(t=10\) a.u. (triangle points). The problem of classical autoionization is characterized by the cross-shape distribution at the later time: one particle falls to the center while the other flies away from the atom.
Snapshots of the FG density at various times when no field is applied to the soft-Coulomb He atom. The lobes on either side of the central distribution represent autoionizing wavepackets. The number of trajectories in the monte carlo sum was \(M=20000\).
### Soft-Coulomb Helium: Towards Optimal Control Theory
Our final example is also the hardest one: we apply an optimal control field to the soft-Coulomb He atom, to try to get it to evolve from the ground-state to the first-excited state. The occupation numbers in the ground-state are close to 2 for the highest, and close to zero for all others, while in the target state, the highest two occupation numbers are both close to 1, having a double Slater determinant character. (See also Sec. II, and). However, instead of optimizing the field for the FG evolution, we simply take an optimal field found for the exact dynamics, and ask how well does it work for the FG problem. (Of course, the former is what we ideally would do, however, we leave the coding of optimal control problems for future work.) We find the optimal field using the optimal control functionality of the octopus code. Moreover, we consider a relatively short duration for the optimal pulse, \(T=35\) a.u., as we will find this illustrates already the challenges the FG approach has for such problems. Because there are only a few cycles of the laser field in this short time, the yield in the exact problem is not very large: the optimal field found in yields a final state population of \(0.73\).. The field is significantly weaker than in the previous section, and we really do not expect much ionization, classical or real. To avoid the situation of Fig. 7, we therefore use method (ii) to terminate the trajectories, prescreening them such that no trajectories where one coordinate reaches a distance \(L=10\) a.u. at any time in the run are included. We find that beginning with \(5\times 10^{6}\) candidate trajectories, the prescreening process discards many such that \(2391950\) remain for the FG calculation.
The result for the NO in in are disappointing. The FG results are converged, as different seeds, and increasing the number of candidate trajectories did not change the results very much. Also, simply running unscreened calculations with increasing numbers of trajectories, appeared to converge, more or less, to these results. Although initially the largest FG NO occupation number tracks the exact reasonably well, we see that after about \(T=15\), it turns upwards instead of continuing down; the FG state appears to reduce to a SSD instead of developing more of a double -SD character. Turning to Figs. 14 and 15, we see that after beginning to spread, the density then contracts and settles into a narrow distribution in the well.
The momentum density n(p) at times \(10,15,20,25,30\) for the strong field case. Solid line is exact, dotted is KS, and short-dashed is FG.
The optimal electric field to maximally populate the first excited singlet-state of soft-coulomb Helium within a time of 35au.
NO occupation numbers for the soft-Coulomb He atom dynamics induced by Eq. 17.
We also note that the density appears to tighten as time evolves; this is a consequence of autoionization and occurs even if we do not apply any termination to the trajectories, in the limit of convergence; the predominant trajectories that autoionize are those in the tails of the distribution.
Why the FG results are poor at larger times is, we believe, due to the FG resonant frequencies being so off-set from the true ones, that a field that is optimal for population transfer for one is off-resonant for the other, leading to little transfer. The exact excitation frequency to the first excited state is \(0.53\), while the FG one is about \(0.68\) (found by applying a kick to the initial ground-state, as in Sec. III.1). For example, in the exact case, if the frequency of the applied field is shifted by as little as 0.05au away from resonance, there is no longer the strong change in NO numbers that we saw in
Thus the optimal control problem presents an extremely tough test of the frozen gaussian dynamics as it depends critically on subtle interference effects of on-resonance oscillations. As stated in Sec.. III.1, we expect the FG excitation frequencies to improve when coupled with the TDDM propagation, and hence also improve the optimal control results. This will be investigated in future work.
## IV Conclusions and Outlook
The semiclassical FG method provides an intuitive picture of quantum mechanics, as based on classical trajectories, each smeared out into a little fuzzy ball in phase-space, evolving classically in time, and added together with phases determined by their classical action. It, and its more rigorous HHKK version, has also provided useful numerical tools in quantum molecular dynamics. Its application to interacting electronic systems has been hesitant, although recent work clearly shows its promise for electronic spectra. The results of the present work suggest that for systems of interacting electrons in external fields, FG dynamics can also be useful.
In relation to TDDFT and adiabatic TDDMFT, we have shown that FG dynamics overcomes some of the problems these methods have; capturing states of double-excitation character, changing occupation numbers, and accurate momentum distributions. There are nevertheless issues related to its convergence, efficiency and accuracy, that require further study. The issues mainly concern the computational details of the classical trajectories, in particular, the large number of trajectories needed for convergence and also the issue of classical autoionization.
Although a large number of trajectories is required for convergence for our two-electron model atoms, we expect that the "forward-backward" nature of the semiclassical computation in Ref. for the two-body reduced density-matrix will lead to favorable scaling with the number of electrons. Because the prescription of Ref.
The density of the prescreened FG calculation with the optimal field at various times compared to the exact.
The highest natural orbital occupation for a prescreened FG calculation with the optimal electric field of The exact NO value is nearly \(1\) at the end, reflecting the double ssd character of the target state.
The prescreened FG dipole moment versus the exact as a function of time. Although initially the FG mimics the true at short times, the delicate nature of quantum control means FG is too inaccurate for this problem.
How to deal with classical autoionization in a consistent and practical way needs further study. When a strong field is present and electrons are unbound at least for some time, the effect of classical autoionization may be relatively small enough for times of interest that a simple termination of trajectories when they get "too far away" makes sense. However in other situations prescreening trajectories by discarding at the start those which at any time reach a given boundary, may be a better approach to avoid autoionizing trajectories "on their way out" from distorting earlier dynamics. Yet there is a delicate balance: if prescreening is done over too long time, too many trajectories that are important at smaller times get discarded. The solution of this problem also depends on what is the quantity of interest, e.g. whether it is time-resolved densities, or time-averaged spectra.
By including correlation semiclassically, the scheme of Ref. is likely to provide more accurate dynamics than the FG treatment of the entire dynamics shown here, as well as improving over the TDHF and adiabatic TDDMFT results. The present results are encouraging for this next step.
# 75
012506.
* K. Pemal, K. Giesebertz, O. Gritsenko, and E.J. Baerends, J. Chem. Phys. **127** 214101.
* K. Giesebertz, E.J. Baerends, O. Gritsenko, Phys. Rev. Lett. **101**, 033004.
* R. Requist and O. Pankratov, Phys. Rev. A **81**, 042519.
* E. J. Heller, J. Chem. Phys. **94**, 2723.
* W.H. Miller, J. Chem. Phys. **53**, 3578.
* W. H. Miller, Adv. Chem. Phys. **25**, 69.
* J.H. van Vleck, Proc. Natl. Acad. Sci. USA **14**, 178;
* M. C. Gutzwiller, J. Math. Phys. **8**, 1979.
* E. J. Heller, Acc. Chem. Res. **14**, 368.
*_Techniques and Applications of Path Integration,_L. S. Schulman, (Wiley & Sons, Inc., 1981).
* N. Rohringer, S. Peter, J. Burgdorfer, Phys. Rev. A **74**, 042512.
* M. Thiele, E. K. U. Gross, S. Kummel, Phys. Rev. Lett. **100**, 153004.
* M. Lein and S. Kummel, Phys. Rev. Lett. **94**, 143003.
* R.J. Cave, F. Zhang, N.T. Maitra, and K. Burke, Chem. Phys. Lett. **389**, 39.
* G. Mazur, and R. Wlodarczyk, _J. Comput. Chem._, **30**, 811
* Mazur, G., M. Marcin, R. Wlodarczyk, and Y. Aoki, _Int. J. Quant. Chem._**111**, 810.
* M. Huix-Rotllant, A. Ipatov, A. Rubio, and M. E. Casida, _Chem. Phys._ in press.
* K. Yabana and G. F. Bertsch, Phys. Rev. B **54**, 4484.
* K. Yabana, T. Nakatsukasa, J.-I. Iwata, and G. F. Bertsch, Phys. Stat. Sol. (b) **243**, 1121.
*[http://www.tddft.org/programs/octopus](http://www.tddft.org/programs/octopus); M.A.L. Marques, A. Castro, G. F. Bertsch, A. Rubio, Comput. Phys. Commun. **151**, 60.
* M. Thiele and S. Kummel, Phys. Chem. Chem. Phys. **11**, 4631.
* J. F. Dobson, Phys. Rev. Lett. **73**, 2244.
* J. Javanainen, J. Eberly, Q. Su, Phys. Rev. A **38**, 3430.
* D. Lappas and R. van Leeuwen, J. Phys. B: At. Mol. Opt. Phys. **31**, L249.
* A. Bandrauk, H. Ngyuen, Phys. Rev. A **66**, 031401(R).
* M. Lein, E. K. U. Gross, and V. Engel, Phys. Rev. Lett. **85**, 4707.
* K. Richter, G. Tanner, and D. Wintgen, Phys. Rev. A **48**, 4182.
* A. Castro, J. Werschnik, E. K. U. Gross, submitted; also arXiv:1009.2241v1.
|
10.48550/arXiv.1108.2034
|
Electron Correlation via Frozen Gaussian Dynamics
|
Peter Elliott, Neepa T. Maitra
| 1,022
|
10.48550_arXiv.1706.05270
|
## 1 Introduction
Essential part of biological development and adaptation on this planet is the conversion of nitrogen into ammonia (nitrogen fixation). This is due to the fact that nitrogen is needed for the utilization of biogenesis, which is an enzyme-catalyzed process where substances are converted into more complex products for basic building blocks of living organisms. Eventhough nitrogen is one of the most abundant elements on Earth, dominantly in the form of nitrogen gas (N\({}_{2}\)) in the Earths atmosphere, living organisms can only utilize reduced forms of nitrogen. Organisms utilize the reduced form of nitrogen through several forms, for example: through (i) ammonia (NH\({}_{3}\)) and/or nitrate (NO\({}_{3}^{-}\)) fertilizer, (ii) utilization of released compounds during organic matter decomposition, (iii) the conversion of atmospheric nitrogen by natural processes, such as lightening, and (iv) biological nitrogen fixation (BNF).
This study aims to depict key aspects that characterize BNF processes for an alloyed sulfur desorbed photocatalytic ternary system MMoS\({}_{2}\) M=Mo,Fe,Co. The MoS\({}_{2}\) based sulfur desorbed system tends to decrease the coordination of the Mo sites, which in contrast to the desorption of sulfur atoms the surface would lack affinity for nitrogen species. Experimentally, the desorption of sulfur has been achieved during ultra-high vacuum (UHV) annealing of this transition metal dichalcogenides(TMD). TMD or TMDC share similar properties to graphene, which makes it an advantageous photocatalyst. From a morphology perspective TMDs form two-dimensional structures, similar to graphene. However, differences are realized when expresing TMDs and graphene's electronic properties. A commonly studied TMD and the focus of this study is a Molybdenite TMD, which is commonly refered to as MoS\({}_{2}\) monolayers. The optical band gap properties for MoS\({}_{2}\) range from 1.2eV to 1.8eV depending on the monolayer environment. There are several experimental approaches to achieving the monolayer phase that include both mechanical and chemical exfoliation. It was found that during chemical exfoliation and a post anneal phase at UHV that sulfur desorption is realized and sulfur vacancies are formed. With this experimental method being known, the approach taken is to study the photocatalytic properties of a desorbed MoS\({}_{2}\) monolayer.
### Biological Nitrogen Fixation (BNF)
In nature, the BNF process occurs naturally in soil by nitrogen-fixing bacteria that are affiliated with some plants. Biological nitrogen fixation is carried out by specific groups of prokaryotes, these organisms utilize the enzyme nitrogenases (enzymes that are produced by certain bacteria, and are responsible for the reduction of nitrogen to ammonia). The nitrogen molecule is comprised of two nitrogen atoms joined by a triple covalent bond, making the molecule highly inert and nonreactive. Microorganisms that fix nitrogen (nitrogenases) require 16 moles of adenosine triphosphate (ATP) to reduce each mole of nitrogen (approximately 5 eV). Reduction of atmospheric nitrogen takes place in nitrogenases by initially weakening the N-N bond by successive protonation until the dissociation barrier is low enough that the N-N bond breaks (later in the reaction sequence). This particular reaction sequence is referred to as the associative mechanism. The energy required by the microorganisms to reduce nitrogen is obtained by oxidizing organic molecules. Meaning that non-photosynthetic living microorganisms must obtain these molecules from another organisms, while photosynthetic capable microorganisms use sugars produce by photosynthesis to obtain the essential energy to oxidize other organic molecules.
The active site in the enzyme is a cluster of FeMo\({}_{7}\)S\({}_{9}\)N, the FeMo-cofactor, with an electrochemical reaction: \(N_{2}+8(H^{+}+e^{-})\to 2NH_{3}+H_{2}\). These clusters have shown great stability in various configurations and sustainability in natural environments. Thus, most synthetically researched nitrogen fixation processes of transition metal configurations are based in some variation to molybdenum (Mo) structures.
A key insight into the nitrogen fixation process has been realized through a series of study on the biological nitrogen fixation processes. First, it was determined that nitrogenase of the BNF process is a two component system comprised of MoFe protein and the electron transfer Fe protein. Second, a reducing source and MgATP are needed for catalysis and that the Fe protein and MoFe protein associate and dissociate in a catalytic cycle involving single electron transfer and MgATP hydrolysis. Third, the MoFe protein contains two metal clusters, the iron-molybdenum cofactor (FeMo-co), which provides an active site for substrate binding and reduction, and P-cluster (which involves the electron transfer from the Fe protein to FeMo-co). Fourth, crystallographic structure are important for both Fe and MoFe proteins. These catalytic advancements have greatly assisted in the reduction of reaction pressure and temperature of the Haber Process. However, no significant energy consumption reduction has been shown, thus yielding the catalytic mechanism to remain incomplete.
### Haber-Bosch Process
A typical method of synthetically producing ammonia is through the Haber-Bosch process. This process essentially reduces nitrogen the same way as biological systems, however, with much greater energy demand. In this process conversion of atmospheric nitrogen (N\({}_{2}\)) to ammonia (NH\({}_{3}\)) is done by reactions with hydrogen (H\({}_{2}\)), coupled with metal catalysts (typically an iron based catalyst). The reaction takes place under high temperature and pressure, where some energy can be recuperated. During the process N and H gas molecules are heated to approximately 400 to 450 \({}^{\circ}\)C, while continuously being kept at a pressure of 150 to 200 atm. Prior to the exposure to the catalyst, it is necessary to remove as much of the oxygen as possible to avoid oxidation of the catalyst. The dissociated N and H mixture is passed over a Fe-based catalyst to form ammonia.
\[N_{2}(g)+3H_{2}(g)\rightleftharpoons 2NH_{3}(g) \tag{1}\]
In contrast to BNF, the Haber-Bosch process initially dissociates the N\({}_{2}\) bond on the first step and then protonates each nitrogen atom, referred to as the dissociative mechanism. The nitrogen and hydrogen atoms do not react until the strong N\({}_{2}\) triple bond and H\({}_{2}\) bond have been broken. Eventhough the reaction is reversible and an exothermic process, relatively high temperature and pressure are still needed to make the reaction evolve quickly. However, this shifts the equilibrium point towards the reactants thus resulting in the lower conversion of ammonia. To correct for this shift of the reactants equilibrium, high pressure is thus needed to shift the equilibrium in favor of the reactions products.
## 2 Methodology
There is a growing concerns and need for synthetically producing ammonia that will need to be addressed with the growing population. What this study aims to understand is the possibility of utilizing concepts of electrochemistry for ammonia production. Excellent insight into obstacles faced while developing catalytic materials for nitrogen reduction have been demonstrated in the past several years from a theoretical and experimental point of view, which provides a great starting point for further investigation on these catalytic materials. Previous studies also show that ammonia synthesis is very structure sensitive on metal surfaces and primarily occurs on surfaces steps of Fe and Ru, and potently could expect the associative mechanism to be even more structure sensitive. Thus, looking at the highly under coordinated MoS\({}_{2}\) structure, which provides a means to investigate this structure sensitive associative mechanism.
Development of density functional theory (DFT) allows for more feasible means of reliably modeling chemical reactions of expressed MoS\({}_{2}\) structure for investigation of the electrocatalytic production of ammonia. This study focuses on the highly under coordinated MoS\({}_{2}\) structures, which has shown oxidation reactions performed on transition metal nanoparticles consisting of metals at lower temperatures, and has shown to dramatically change the reactivity of inert metals. Also, due to the fact that various experimental work has demonstrated this structure to have good: light absorption, band gap variability, charge mobility, thermo stabilaty, and structure mobility. While previous studies primarily looked at step metal surfaces of MoS\({}_{2}\), this study performed calculations on consistent basis for configurations of MoS\({}_{2}\) and M\({}_{x}\)Mo\({}_{y}\)S\({}_{2}\) (M=Fe, Co) sub configurations. This allows the investigation of the reaction intermediates for the dissociative and associative mechanisms on MoS\({}_{2}\) structure for Mo, Co, and Fe (M\({}_{x}\)Mo\({}_{y}\)S\({}_{2}\) (M=Fe, Co)) transition metals, and develop stepped and closed packed metal surface relations. These calculations use statistical mechanics to construct Gibbs free energy diagrams for both reaction pathways based on entropy, zero point energy corrections, and vibrational calculations. These diagrams are then used to determine the lowest potential across the investigated metal configurations, which will make each part an exothermic reaction step.
### Nitrogen Reduction
Nitrogen reduction is a product of oxidation and reduction reactions, the loss and gain of electrons in a chemical reaction. Biochemical cycling of C, H, and O are the foundations in which oxidation and reduction reactions take place in chemical elements with which nitrogen reduction is most critically associated with. Were nitrogen valence range undergoes a biochemical cycling from which nitrogen can lose all five of its outer shell electrons to surrounding elements to gain three electrons from other elements to complete all orbitals of its outer-most electron shells.
Potentially in this reduction, nitrogen takes two cycling forms, first N atom can lose outer shell electrons if the surrounding element has a higher electron affinity (to be more electronegative). For example, if the surrounding element is to be oxygen which is a more electronegative element (affinity for electrons is greater than nitrogen), thus N will eventually lose all five outer shell electrons and thus N can eventually become fully oxidized as nitrate (NO\({}_{3}^{-}\)). In contrast, N can eventually fill all outer shell electron orbitals (gain three electrons) from elements such as H and C elements, which are less electronegative than N. With this gain of electrons, N can be fully reduced to ammonia (NH\({}_{3}\)), and potentially N can be full to partially reduced in various organic compounds. Same as C, H, and O, the N cycling involves a zero valence form (N\({}_{2}\) gas), for which the charge of the seven protons in N's nucleus are balanced by seven electrons orbiting the nucleus by the two electrons of the inner electron shell and the five of the outer shell. However, the depicted N reduction is greatly idealized and can therefore contain many intermediate oxidation/reduction forms that N can assume that are not expressed but are discussed in great detail in references.
Furthermore, the oxidation/reduction reactions of N are mediated by biological, chemical, and physical factors. These factors do not only influence the form that N takes, but also the flux and accumulation of nitrogen, as well as the nature and extent of the reactions in which N contributes. The importance of these mediating factors are expressed from a thermodynamic examination of N reduction, for which N exists in forms other than its most thermodynamically stable form (NO\({}_{3}^{-}\)). This indicates that external sources of energy that are directed by mediating chemical, physical and biological factors truly contribute to the possibility of driving nitrogen reduction.
### Photocatalytic
The main issue of synthetically reducing nitrogen is the energy required for the process to successfully be implemented. Typically this energy is attained by burning fossil fuels that yield environmental pollution as discussed in previous sections. This prompted researchers to investigate better pathways for ammonia production through the mimicry of biological organisms, thus allowing various researchers to try and develop catalysts for this process. This yielded many researchers to investigate the mechanism of nitrogen fixation and produce photocatalytic and photoelectrochemical processes to utilize solar energy conversion.
The idea is to allow photoreaction to be accelerated by the presence of a catalyst, meaning that the photocatalytic activity is dependent on the catalyst to create hole pairs (electron charge carriers), which generate free unpaired valence electrons that can contribute to secondary reactions. Typically, photocatalysts are transition metal oxides and/or semiconductors that have void energy region (band gap) where no energy levels are present to undergo an electron and hole recombination produced by photoactivation (light/photon absorption). When the energy of an absorbed photon is equal to or greater than that of the materials band gap, an electron becomes excited from the valence band (bottom of band gap) to the conduction band (top of band gap), thus producing a positive hole in the valence band. This excited electron and hole become recombined releasing heat from the energy gained by the photon exciting the surface of the material. The idea is to have a reaction that produces an oxidized product by reaction of generated holes with a reducing agent and to have a reduced product by the interaction of the excited electron and an oxidant. Simply-put, photoelectrochemical and photocatalytic processes rely on semiconductor materials to absorb sunlight and generate excited charge carriers that allow for reactions to take effect without any external electricity input.
Thus yielding a photocatalytic and electrochemical investigation for sunlight-driven nitrogen fixation to reduce energy consumption, which will also provide flexibility in designing materials for nitrogen conversion. This work prompts to gain a crucial understanding of nitrogen to ammonia reaction process on the photocatalytic and electrochemical level and to develop means of approaching a sustainable solar driven conversion process of nitrogen to ammonia.
### Material Phase
The material of interest is a MMoS\({}_{2}\) (Figure1) that has received tremendous attention due to the earth-abundant composition and attractive optical, catalytic, and electronic properties, and high chemical stability. However, the focus of MoS\({}_{2}\) has mainly been focused on the hydrogen evolution reaction (HER). The bulk crystal is indirect gap semiconductor that has an energy gap of approximately 1.29 eV and is built up of van der Walls bonding of the S-Mo-S units. Due to their large surface areas and highly dense active sites along edges, MoS\({}_{2}\) are potentially promising material for electrochemical reaction studies. However, MoS\({}_{2}\) have proven to have poor conductivity, which has limited their electrochemical response. The fact that MoS\({}_{2}\) monolayers is known to have two phases, the trigonal prismatic (2H) and octahedral (1T), makes it a unique structure to investigate. The 2H phase has proven to be very stable, but with poor conductivity. However, the 1T phase is particularly favoring a stable structure at room temperature, but is metallic and better conductivity in nature. This is why the combination of Fe and Co into the MoS\({}_{2}\) is being looked at in this study, to allow ideal properties to be demonstrated for distinct structures. If the properties of the two phases are combined in a monolayer structure, such as higher stability of the 2H phase and the high conductivity of the 1T phase, this will result in the very favorable structure for N reduction. The resulting structure would have both large specific surface area and high charge transfer abilities that can be achieved simultaneously.
The structural transformation between semiconducting (2H) and metallic (1T) phases of MoS\({}_{2}\) has been a topic of interest in the past. These phase transitions (1T/2H) is made up of Mo and or S atomic glide planes that require a precursor intermediate phase (\(\alpha\)-phase). Also, the phase transition is found to be associated with movement between B and y-boundaries. In a thermodynamic system, the phase transition is fundamental phenomena and of great technological importance in the material science research, due to the fact that properties of a material are able to be altered without the need on additional atoms into the system.
The MoS\({}_{2}\) crystal (Figure1) structure is made of atomic layers stacked by van der Walls forces, each layer is comprised of strong in-plane binding of S-Mo-S triple atomic planes. Depending on the arrangements of the S atom, the crystal structure can appear in two distinct symmetry(2H and 1T) as expressed by Figure1, the phases can be easily converted to each other through an inner layer atomic plane glide, which involves a transverse displacement of one of the S-planes. The 2H phase demonstrates hexagonal lattice with a threefold symmetry with an atomic stacking sequence of ABA (S-M-S), whereas the 1T phase similarly shows atomic stacking sequence of ABA (S-M-S) however, the S-plane occupies the hollow center of the 2H hexagonal lattice. This S plane glide occupies the hollow center site of the 2H hexagon, thus resulting in a 2H\(\rightarrow\)1T phase transition (Figure1). Due to weak interactions between these monolayers coupled with much stronger intralayer interactions, the formation of ultrathin crystals of MoS\({}_{2}\) by micromechanical cleavage technique has been demonstrated. Systematic studies of the evolution of the optical properties and electronic structure of ultrathin MoS\({}_{2}\) crystals have been demonstrated and shown to be as a function of layer numbers. To further modify the activity of the structure towards nitrogen, sulfur atoms have been desorbed from the top surface. This has been achieved experimentally using a UHV anneal process.
### Material Design Space
Typically a good strategy to design new catalyst for nitrogen reduction is by combining components that lie from both the dissociative and associative side. Initial DFT calculations show that MoS\({}_{2}\) to favor the dissociative side and with the addition of Fe and or Co, which typically favor the associative mechanism, the material of interest takes the form of M\({}_{x}\)Mo\({}_{y}\)S\({}_{2}\) (M=Mo,Fe,Co). The main structural modification that was done is the substitution of the non-organic ion with various combinations in all possible positions in the unit cell (MMoS\({}_{2}\), M = Mo,Co,Fe). The 8 (1T) and 7 (2H) atomic positions that were modified are clearly shown in of the unit cell for the 1T and 2H structures.
Illustration of the sulfur desorbed MMoS\({}_{2}\) 1T and 2H supercell structure. A 1T unit cell made up of 16 atoms 8 transition metal atoms and 8 S atoms, whereas the 2H unit cell is comprised of 18 total atoms 7 transition metal atoms and 11 S atoms. The transition metal atoms in the structures are alloyed with all combinations of Mo, Fe, Co atoms.
These configurations are classified under one element case (pure Mo, Co, Fe structure), two element cases (combination of Co and Fe, Co and Mo, Fe and Mo), and three element cases (combination of Co and Fe and Mo).
It is noted that while the MoS\({}_{2}\) material is coupled with Fe and Co, all combination of the three elements are and will be evaluated in this study to better vary the mechanism of interest for each distinct adsorption species (N\({}_{2}\)H\({}_{x}\), NH\({}_{x}\)). For cases of elemental combination (two element and three element cases), there are different configurations possible for same elemental ratios. This makes a total of 6561 (3 elements for 8 positions 38) unique configurations for 1T and 2187 (3 elements for 7 positions 37) unique configurations for 2H structures. Only the best configurations (lowest energy of the relaxed structure) for each distinct structure will be evaluated for the electrochemical reactions. For example, configuration of 1T structure of 1 atom Co and 7 atom Mo (Co1Mo7)can have 8 distinct configurations of Co being in any of the 8 metal positions, so only the lowest energy structure after the relaxation will be implemented in the electrochemical reactions, and this is the same for every other configuration for 1T and 2H (Co2Mo6, Co3Mo5,...).
Also, its is noted that for each configuration there are various operations that had to be done in sequence to obtain proper data, this resulted in a total of 45 configurations for 1T and 36 configurations for 2H that had the lowest energy structure for a specific configuration. The reader should note that this is a tremendous amount of computational and analytically challenging process. This is an area that needs to be investigated from a statistical point of view. Then only the 45 (1T) and 36 (2H) unique structures went on to be implemented in the electrochemical reactions. This will significantly allow every configuration of 1T and 2H structure to be explored and give the best configuration of MMoS\({}_{2}\), M = Mo,Co,Fe structure as a means of nitrogen reduction catalyst.
### Associative and Dissociative Reaction Pathways
In the process of electrochemically forming ammonia, it is conventional to reference the source of protons and electrons to model the anode reaction
\[H_{2}\rightleftharpoons 2(H^{+}+e^{-}) \tag{2}\]
First protons are introduced into the proton conducting electrolyte to sustain the equilibrium and diffuse into the cathode, while an external circuit is used to transport electrons to the cathode side through a wire.
\[N_{2}+6(H^{+}+e^{-})\to 2NH_{3}, \tag{3}\]
Theoretically, the reaction can take the form of two different possible types of pathways in which there are two different possible types of mechanism for each pathway in order to synthesize ammonia electrochemically. The respected pathways are the associative and the dissociative pathway, where the possibility of either adsorbed N\({}_{2}\)H\({}_{x}\) or NH\({}_{x}\) species can be hydrogenated. These pathways correspond to the Tafel-type mechanism and the Heyrovsky-type mechanism.45 Solvated protons from the solution first adsorb46 on the surface and combine with electrons, then the hydrogen adatoms react with the adsorbed species(N\({}_{2}\)H\({}_{x}\) or NH\({}_{x}\)) in the Tafel-type mechanism. This type of mechanism can only have an indirect effect through interchangeable concentrations of the reactants.45 So due to the fact that this study focuses on room temperature processes and also the activation barriers for Tafel-type reactions are about 1 eV or higher for most transition metal surfaces,47,54 this type of mechanism will most likely prove to be very slow. Also, this type of mechanism requires hydrogenation steps54 of the reaction barriers to be overcome, and thus will also require higher temperature as well to drive the process forward. This is due to the requirement of the reaction to merge proton and electrons to form hydrogen adatom on the surface first.46 Thus the process will therefore either go through an associative and or dissociative Heyrovsky-type reaction.
In Heyrovsky-type of reaction45, the adsorbed species are directly protonated so that a coordinate bond to the proton and the species are formed. Molecules from the electrolyte get directly attached by protons and then electrons from the surface merge with the protons to form a hydrogen bonded to the molecule. Also, by applying a bias in the latter cases of the mechanism, a thermochemical barrier can be directly affected. Initially, this study considers the possibility of the reaction to take an associative Heyrovsky mechanism, similar to that of the mechanism for the BNF, where the N-N bond is initially weakened by successive protonations until the dissociation barrier is low enough so that the N-N bond can be broken later in the reaction. For the Heyrovsky mechanism, nitrogen molecule is first attached to the surface and is then protonated before N-N bond dissociates. In the bellow equations of the associative Heyrovsky mechanism (asterisk '*', denotes a site on the surface):
\[N_{2}(g)+6(H^{+}+e^{-})+*\rightleftharpoons N_{2}*+6(H^{+}+e^{-}) \tag{4}\]
\[N_{2}*+6(H^{+}+e^{-})\rightleftharpoons N_{2}H*+5(H^{+}+e^{-}) \tag{5}\]
\[N_{2}H*+5(H^{+}+e^{-})\rightleftharpoons N_{2}H_{2}*+4(H^{+}+e^{-}) \tag{6}\]
\[N_{2}H_{2}*+4(H^{+}+e^{-})\rightleftharpoons N_{2}H_{3}*+3(H^{+}+e^{-}) \tag{7}\]
\[N_{2}H_{3}*+3(H^{+}+e^{-})+*\rightleftharpoons 2NH_{2}*+2(H^{+}+e^{-}) \tag{8}\]\[2NH_{2}*+2(H^{+}+e^{-})\rightleftharpoons NH_{3}*+NH_{2}*+(H^{+}+e^{-}) \tag{9}\]
\[NH_{3}*+NH_{2}*+(H^{+}+e^{-})\rightleftharpoons 2NH_{3}* \tag{10}\]
\[2NH_{3}*\rightleftharpoons NH_{3}*+NH_{3}(g)+* \tag{11}\]
\[NH_{3}*+NH_{3}(g)\rightleftharpoons 2NH_{3}(g)+* \tag{12}\]
The addition of the fourth H to the N\({}_{2}\)H\({}_{3}\)* molecule weakens the N-N bond to readily dissociates molecule into the species (NH\({}_{x}\)) on the surface. Also, there is a possibility of reaction to split into NH and NH\({}_{2}\) on the surface and has been observed on some metals.
The second mechanism, dissociated Heyrovsky mechanism, is also considered and compared to the associative mechanism.
For a deep understanding of the reactions, free energy correction is needed to be determined and included in the analysis for each reaction intermediate. From DFT calculations, a reasonable approximation to the free energy for the adsorbed species relative to the gas phase molecular N and H can be obtained from the expression:
\[\Delta G=\Delta E+\Delta E_{ZPE}-T\Delta S, \tag{23}\]
In this study, only the ZPE is considered for the gas phases.
The applied reference potential driving the electrochemical reaction is set to be that of the standard hydrogen electrode (SHE), which this study also takes into account in addition to zero point energy and entropy. By using the computational SHE, this study is able to include the effect of the potential on the expressed reactions for surface sites. The standard hydrogen electrode has given great incite to describe numerous electrochemical reactions, such as trends in CO\({}_{2}\) (carbon dioxide), nitrogen, and oxygen reduction.
Thus this study uses the standard hydrogen electrode as the reference potential, which expresses the free energy per H (chemical potential of (H\({}^{+}\)+e\({}^{-}\)) as related to that of \(\frac{1}{2}\)H\({}_{2}\)(g), which is equation in equilibrium. This implies that the pH = 0, the potential is that of U = 0 V relative to the SHE, and a pressure of 1 bar of H\({}_{2}\) in gas phase at 298 K, thus reaction's free energy is equal to that of net reactions of- or- at an electrode. Also, it is noted that all presented calculations in this study are for that of pH equal to 0.
### Computational Details
Ground state thermodynamic properties were predicted for thermodynamic steps that helped this study analyze various configuration reaction steps by means of density functional theory (DFT) approach. DFT calculations used functional form of the pseudo-wave function that was based on Perdew-Burke-Ernzerhof (PBE) exchange-correlation function at potentials with a cut-off wave function energy of 1224 eV (90 Ry), which provided most accurate and stable for the intended unit cell. Also, a pseudized wave function allowed the reduction of computational expense for which this study greatly benefited due to the various configurational combinations this structure had that this study hoped to explore. In addition, computational expense was greatly reduced when only a single primitive cell was simulated for each configuration. Monkhorst-Pack with a k-point mesh sampling 2x2x2 grid with an offset of 1/2,1/2,1/2 and a 7 A vacuum around the mono-layer has been applied. Van der Waals correction term was incorporated to account for the Van der Waals interaction, which allowed some empiricism into the calculation. Scaling parameters were specified to be 0.7 and cut-off radius for the dispersion interaction was 900 A.
When solving electronic density self-consistently, adsorbates (N, H, NH, etc.) sitting on the structure and unit cell geometries were relaxed to a relative total energy less than 1x10\({}^{-10}\) and overall cell pressure of less than 0.5kBar. The unit cell was relaxed in the vacuum before any absorbates were implemented in the structure, then once relaxation of the structure was done the atoms were kept fixed so that the adsorbates were allowed to relax on the surface of the structure in the 7 A vacuum. The reader should note that pure DFT predictions of energetics, band gap, etc. are often under predicted due to the over-analyticity of the functionals and exchange-correlation terms. Therefore, the reported energies in this study should not be used as absolutes but used to study the trends for various structures.
## 3 Results and Discussion
Adsorption energies of all reaction states in- are calculated for various combinations of M\({}_{\rm x}\)Mo\({}_{\rm y}\)S\({}_{2}\) (M=Fe, Co) in a 1T and 2H structure. Results are used to estimate the free energy change in elementary reactions- for the associative mechanism and- for the dissociative mechanism. The theoretical over-potential needed for reactions to take effect is approximated using free energy difference of elementary steps. Linear approximating relations for adsorption energies that correspond to adsorption species (N\({}_{2}\)H\({}_{\rm x}\), NH\({}_{\rm x}\)) assist in depicting trends in the catalytic activity of the implemented transition metals (Mo, Co, Fe). Its known that adsorption energies of simple hydrogen containing species such as N\({}_{2}\)H\({}_{\rm x}\), NH\({}_{\rm x}\) depend linearly on the adsorption energy of the nitrogen atom. This paper establishes linear relations for N\({}_{2}\)H\({}_{\rm x}\), NH\({}_{\rm x}\) species using absorption energies of N adatom and also express the relation of metal transition linear scheme for the expressed structure (1T, 2H).
### Adsorption Sites
Adsorption sites illustrated in Figure2 of MoS\({}_{2}\) are used to demonstrate characteristics of adsorption states for all metals studied in this report because these sites have very similar trends; however, this is not to say that the binding energy is the same for different structures. It's sufficient to say that combinations of M\({}_{\rm x}\)Mo\({}_{\rm y}\)S\({}_{2}\) (M=Fe, Co) 1T and 2H structures produce very small variations in adsorption sites. These sites are usually classified as adsorption sites of hollow, bridge, and on top, and typically each classification will be slightly geometrically different. Thus each structure (M\({}_{\rm x}\)Mo\({}_{\rm y}\)S\({}_{2}\) (M=Fe, Co)) will yield unique electronic properties that can be analyzed and structures that prove more stable can be found and investigated.
Figure2 demonstrates these adsorption sites on MoS\({}_{2}\) monolayer, first hydrogen molecule proved to be quite unstable on the surface and tended to lie on the crystal surface. Thus hydrogen proved to have a bonding site of bridge classification; however, a very small difference of energies was found between adsorption sites. Whereas nitrogen adsorbed to the surface on a hollow site and preferred to bind in a di-sigma (strong covalent bond, formed by overlapping between atomic orbitals) bond on the surface with two metal atoms, thus proving edge sites to be the most stable. Similar in the absorption of N\({}_{2}\), however it is noted that N\({}_{2}\) molecule proved to have few stable adsorption configurations, due to the strong N-N bond.
The NH molecule binds to hollow sites, NH\({}_{2}\) to bridge sites, and NH\({}_{3}\) on-top sites and all proved to be stable structures (bottom row). However, in the case of N\({}_{2}\)H, N\({}_{2}\)H\({}_{2}\), and N\({}_{2}\)H\({}_{3}\) species (Figure2 middle row), they typicality preferred to bond in a bridge site, and each nitrogen atom bonded to a metal atom similarly to the di-sigma bonding expressed for N and N\({}_{2}\) species (Figure2 top row). Based on the absorption sites, species' configuration orientation changed slightly as more hydrogen atoms are implemented in the surface species. Weakening of the N-N bond is demonstrated when looking carefully at Figure2, as a visual representation of the internal bonding length dramatically increases as more hydrogen atoms are implemented in the species. It is very apparent when looking at NH\({}_{3}\) absorbed on the structure, were the internal bonding length is much greater to that of other species. Also, it is very apparent that nitrogen atom becomes further away from the metal atom as more hydrogen atoms become bonded to the respected nitrogen atom, this is demonstrated for BNF for the associative mechanism and shows the weakening of bonds between nitrogen and metal atoms.
### Ammonia Formation on Surface
For each surface composition, two morphologies (1T,2H) with several adsorbate species that conform to the two reaction pathways were investigated to predict the electrochemical reaction of each step for different mechanism configurations. Figure3 illustrate the electrochemical reaction which refers to the free energy of reaction steps- for the associative mechanism and- for the dissociative mechanism of 1T and 2H structures.
DFT calculations show that a N\({}_{2}\) molecule binds to all surfaces of 1T and 2H structures with an adsorption energy that is always slightly more negative when compared to all other reaction steps, (Figure3(A,C) dissociative \(\Delta\)G14 and Figure3(B,D)
Absorption sites of the MoS\({}_{2}\) 1T structure. The absorption sites presented for the MoS\({}_{2}\) 1T structure is only used for the adsorption types and not for actual bonding for all other metals in this study. This means that the species of NH\({}_{\rm x}\) could bind more strongly to Mo, Fe, Co structure configurations studied.
The large loss in enthalpy in going from gas phase N\({}_{2}\) to a surface bonded molecule, this corresponds to a slightly negative free energy under ambient conditions for the 1T and 2H structures. For the most part, the associative mechanism for both structures (1T and 2H) tend to have the potential determining steps (largest positive step to overcome) in the reduction to form ammonia, this is the addition of one hydrogen atom to transition from N\({}_{2}\) to N\({}_{2}\)H (Figure3 from \(\Delta\)G5 to \(\Delta\)G6), and addition of one hydrogen atom to go from NH\({}_{2}\) NH\({}_{3}\) (Figure3 from \(\Delta\)G9 to \(\Delta\)G10). This hydrogenation step corresponds to a positive step that is needed to overcome from free energy formation of N\({}_{2}\) and NH\({}_{2}\) on the surface, which is also deemed as a large energy barrier that is associated with the binding of hydrogen from the gas phase.
Typically the addition of hydrogen atoms promotes a positive step in free energy for most surfaces investigated and the first hydrogen step (from \(\Delta\)G5 to \(\Delta\)G6) appears to be the most uphill in free energy for all hydrogen steps. However, the overpotential in the dissociative mechanism for both morphologies (1T and 2H) is most significant when initially trying to break the N-N bonding going from N\({}_{2}\) to N (Figure3 from \(\Delta\)G14 to \(\Delta\)G15), this is very likely to occur due to N-N bond which binds in a di-sigma bond that is a very strong valent bond formed by overlapping between atomic orbitals that are very hard to overcome. Even thou N\({}_{2}\) binds to the surface with a negative adsorption energy from the significant loss in entropy from the gas phase, this is not enough to overcome this initial potential to brake N-N bonding without applying an external potential. It is also noted that for both the associative and dissociative mechanism of 1T and 2H structures (Figure3) the last three steps (associative(B,D): \(\Delta\)G10 to \(\Delta\)G11 to \(\Delta\)G12 and dissociative(A,C): \(\Delta\)G20 to \(\Delta\)G21 to \(\Delta\)G22) are typically positive steps in free energy associated with the reduction of NH\({}_{3}\) on the surface to NH\({}_{3(g)}\) this is due to a large gain of entropy going from surface bounded molecule to a gas phase.
### Adsorption of N\({}_{2}\)H\({}_{x}\) and NH\({}_{x}\) Species on Fe and Co Configurations
The associative mechanism initially relies on absorbed configuration steps of N\({}_{2}\)H\({}_{x}\) molecules whereas the dissociative mechanism initially relies on absorbed configuration steps of NH\({}_{x}\) molecules. The associated Gibbs energies of these reactions depends on both the morphology and composition of the surface. In this study stable configurations of the surface, which is determined based on the formation energy predictions, are presented. The solid solution stability is important in the MoS\({}_{2}\) structure because both CoS\({}_{2}\) and FeS\({}_{2}\) forms pyrite structures. This means that only concentrations below fifty percent are stable in the MMoS\({}_{2}\) structure before pyrite is formed. Therefore, as shown in the results are depicted for only the stable compositions of MMoS\({}_{2}\) and not the pyrite phase.
First, it is known that the catalytic activity of the plane parallel to the lateral or horizontal axis(basal plane) of a pure 1T and 2H MoS\({}_{2}\) configurations mainly becomes apparent from the structure's affinity for binding hydrogen at surface S sites. Meaning that one would expect configurations comprised of more Mo atoms to have hydrogenation steps (associative: \(\Delta\)G5 to \(\Delta\)G6, \(\Delta\)G9 to \(\Delta\)G10, and dissociative: \(\Delta\)G15 to \(\Delta\)G16) which corresponds to the largest positive step that needs to be overcome from free energy formations of N\({}_{2}\) to N\({}_{2}\)H (\(\Delta\)G5 to \(\Delta\)G6), NH\({}_{2}\) to NH\({}_{3}\) (\(\Delta\)G9 to \(\Delta\)G10), and N to NH (\(\Delta\)G15 to \(\Delta\)G16). This can be seen in Figure3, where more concentrations of Mo in the structure yields positive steps at the depicted formation energy regions (associative: \(\Delta\)G5 to \(\Delta\)G6, \(\Delta\)G9 to \(\Delta\)G10, and dissociative: \(\Delta\)G15 to \(\Delta\)G16). Also, one would expect structures to have hydrogen as a large positive step, due to internal bonding length of N-N dramatically increases as more hydrogen atoms are implemented in the species (Figure2).
Second, nitrogen binds in a di-sigma bond (that is formed by overlapping between atomic orbitals) on the surface with metal atoms, and that Co and Fe atoms have a higher affinity to nitrogen than Mo atoms. Thus causing structures having more concentration of Fe and Co to have the positive determining step associated with braking N-N bonding (dissociative: \(\Delta\)G14 to \(\Delta\)G15 going from N\({}_{2}\) to N) because nitrogen binds very strongly to Fe and Co atoms compared to Mo. This can be seen when looking at lower concentrations of Mo atom in conjunction with a higher concentration of Fe and Co atoms in Figure3 of the dissociative mechanism going from N\({}_{2}\) to N (\(\Delta\)G14 to \(\Delta\)G15).
Based on the two aforementioned ideas, one would expect a region where there is a shift between which positive step to be more dominant, either hydrogenation step or N-N bonding breaking step. In fact based on the dissociative mechanism illustrated by Figure 3(A,C), one can see the largest step going from \(\Delta\)G14 to \(\Delta\)G15 (N\({}_{2}\) to N) to dramatically increase when more of Fe and or Co is implemented in the structure, because the strong di-sigma bonding with N-N and metal atoms. However, when the ratio of Mo to Fe/Co shifts from a structure containing more Mo atoms, the largest positive steps are affiliated with the hydrogenation as expressed by the associative mechanism of Figure 3(B,D) going from \(\Delta\)G5 to \(\Delta\)G6 (N\({}_{2}\) to N\({}_{2}\)H), \(\Delta\)G7 to \(\Delta\)G8 (N\({}_{2}\)H\({}_{2}\) to N\({}_{2}\)H\({}_{3}\)), and \(\Delta\)G9 to \(\Delta\)G10 (NH\({}_{2}\) to NH\({}_{3}\)) due to the first expressed idea of this section.
### Onset Maximum Positive Determining Step
Figure4A and Figure4B illustrate the largest energy barrier that every composition for both morphologies must overcome. If this energy can be overcome by either an applied electric field in the case of electrocatalyst or a photogenerate potential in the case of photocatalyst, the reaction is free to evolve to completion. These values are in the absence of any kinetic considerations that may increase the required over-potential. The white area of the ternary plots corresponds to the formation of a pyrite structure. This is associated with a high concentration of Fe/Co to Mo atoms are incorporated in the structure. The points are associated with each DFT predictions. The contours are interpolation between the DFT predictions. It is reasoned that the colored region is a continuous design space and all compositions should be obtainable. The method of interpreting these plots is to recognize that Mo, Fe, and Co bound the space at the corners of the triangle. Along the edges of the triangle there is a binary composition of transition metals and within the triangle is a ternary composition. Also, it is important to note that these plots are only representing the largest positive step attained from Figure3 by taking the difference between each step and only recording maximum positive step. This step is noted as the largest energy barrier to overcome, therefore, the lowest value in will be the best to implement because these structures will require the least amount of implemented potential.
Based on findings mentioned in the previous section, one can visually confirm in the shift in energy from a composition that has an affinity for hydrogenation to a composition with an affinity for N-N splitting. Where this shift is very clear for the dissociative mechanism of 1T and 2H structures, going from a pure Mo structure (bottom left corner) along the Fe/Co line. This region is saturated by higher energies toward the middle of the lower triangle for both the dissociative and associative mechanisms, but definitely more apparent in the dissociative mechanism. It is also noted that the highest energy for the associative 2H mechanism is attained for a pure Mo structure because the 2H structure has more Mo atoms and in turn, has higher affinity to H. The largest positive potential is associated with the hydrogenation, first point discussed in the previous section. This can be further convinced by the accept performance of 2H MoS\({}_{2}\) for HER.
Focusing on the minimum positive potential for the 1T and 2H associative (Figure 4(B,D)), where the associative mechanism tends to have more hydrogenation step to gradually weaken
Cumulative free energy for the dissociative (A,C) and associative (B,D) mechanism on both the 2H and 1T morphology. Energies are attained from DFT calculations for an electrolyte with pH=0 at 300 K. The energies correspond to reaction steps- for the associative mechanism and- for the dissociative mechanism relative. Each grouping can be associated with the potential-energy curve for successive reactions. The best performing structure and morphology is (Mo\({}_{6}\)Fe\({}_{2}\))S\({}_{2}\) 1T phase (C).
One would expect a structure having more Mo atoms and few Co/Fe atoms to be more dominant. This is seen for the 1T and 2H associative mechanisms, where the mechanism preferred Co atom due to the fact that Co is more electronegative than Fe. That is way higher concentration of Fe that is needed to satisfy the same energy levels as fewer Co implemented in the structure. Of course, this is not a unary or binary transition metal composition because the mechanisms still need to have hydrogenation and N-N bond breaking steps to fully form NH\({}_{2}\) from N\({}_{2}\).
Ternary plots of the maximum reaction barrier for all successive reactions for the three inorganic components and the 2H (top) and 1T (bottom) morphologies. The contour colors corresponds to the energy of the most difficult step to overcome for reaction steps- for the associative mechanism and- for the dissociative mechanism. The white region of ternary plots corresponds to pyrite structure formation, which was not considered. The star in the figures represents the best possible configuration for the given pathway and morphology. The best points are tabulated in Table 1.
Therefore, structures have to have a balance of atomic components that both have an affinity for hydrogenation and N-N bond breaking. The structure that proves to have the best performance for the associative mechanism is a combination of 2Co and 6Mo atoms (Mo\({}_{0.75}\)Co\({}_{0.25}\)S\({}_{2}\)) in the 8 metal positions for 1T structure and 3Mo and 4Fe atoms (Mo\({}_{0.43}\)Fe\({}_{0.57}\)S\({}_{2}\)) in the 7 metal positions for the 2H structure that resulted in approximately 1.23 eV for the positive limiting step (over-potential ignoring kinetic considerations). The 1.23eV over-potential is within the band gap energy of MoS\({}_{2}\). Further investigation is required to determine the change in and gap energy for the compositional space.
Contrary to the associative mechanism, for the 1T and 2H dissociative mechanisms, the N-N bonding is broken on the second reaction and the latter steps in the reaction are responsible for hydrogenations. Based on the claims for the associative mechanism, a structure with lower affinity to nitrogen will be associated with Fe-containing composition and reaction associated with a hydrogen containing species tend towards Mo incorporated compositions. This requires a balance structure also needs to effectively associate the hydrogenation that happens late in the reaction. The 1T dissociative mechanism preferred to have 6Mo and 2Fe atoms (Mo\({}_{0.75}\)Fe\({}_{0.25}\)S\({}_{2}\)) in the 8 metal positions and the 2H preferred to have 4Fe and 3Mo atoms (Mo\({}_{0.43}\)Fe\({}_{0.57}\)S\({}_{2}\)) in the 7 metal positions that resulted in approximately 1 eV for the positive determining step. It is also noted that these results should only be used as trends and not exact results because of approximation in the DFT method. The surface can be expected to be covered with hydrogen atom that forms a natural environment that can't be accounted for using the current approach. Also, the typical formation of hydrogen gas can end up being very fast unless the surface is covered with N adatoms rather than H adatoms, which contributes to possible discrepancies.
## 4 Conclusion
In this study a DFT approach was implemented to explore the design space of MMoS\({}_{2}\) M=Mo,Fe,Co, which would otherwise be time and cost prohibitive from a pure experimental point of view. Moreover, the first principle approach allow additional physics to be understood that would otherwise be difficult to determine experimentally. The study limits the investigation to only the thermodynamic energy barriers for two reaction pathways in the absence of kinetic considerations. The study found that strong affinity of hydrogen species towards Mo and a similarly strong affinity of nitrogen towards Fe. The resulting structure was determined to be an alloy that combined both Mo and Fe. While both reactions paths could be optimized to obtain over-potentials below 2eV the most desired and the lowest of the pathways was the dissociative pathway. The optimal structure was a Mo\({}_{0.75}\)Fe\({}_{0.25}\)S\({}_{2}\) with a over-potential of approximately 0.1eV in the absence of kinetic consideration. An additional benefit of the 1T phase, which is associated with higher electonic mobility. Future studies will investigate the optics, electron mobility, and the intermediate transitions on this phase.
|
10.48550/arXiv.1706.05270
|
Ab initio Screening of a Sulfur Desorbed MoS$_2$ Photocatalyst for Nitrogen Fixation
|
Alhassan S. Yasin, Nianqiang Wu, Terence Musho
| 6,548
|
10.48550_arXiv.1411.1579
|
## Computational details.
Equilibrium lattice constants, bulk moduli, and cohesive energies have been calculated for 24 solids, including Al, Ca, K, Li, Na (simple metals); Ag, Cu, Pd, Rh, V (transition metals); LiCl, LiF, MgO, NaCl, NaF (ionic solids); AlN, BN, BP, C (insulators); GaAs, GaP, GaN, Si, SiC (semiconductors). Reference data to construct this set were taken from Ref.. These calculations have been performed with the VASP program using PBE-PAW pseudopotentials. All Brillouin zone integrations were performed on \(\Gamma\)-centered symmetry-reduced Monkhorst-Pack \(k\)-point meshes, using the tetrahedron method with Bloch corrections. For all the calculations a \(24\times 24\times 24\)\(k\)-mesh grid was applied and the plane-wave cutoff was chosen to be 30% larger than maximum cutoff defined for the pseudopotential of each considered atom.
Calculations for Table 2 have been performed with the TURBOMOLE program package using a def2-TZVP basis set. For details about the molecular test sets, see Refs..
Hyper-GGA calculations are non-self-consistent, using accurate exact exchange orbitals and densities.
|
10.48550/arXiv.1411.1579
|
Gradient-dependent upper bound for the exchange-correlation energy and application to density functional theory
|
L. A. Constantin, E. Fabiano, A. Terentjevs, F. Della Sala
| 4,260
|
10.48550_arXiv.1703.10018
|
## Abstract
Macrocycles have attracted much attention due to their specific "endless" topology, which results in extraordinary properties compared to related linear (open-chain) molecules. However, challenges still remain in their controlled synthesis with well-defined constitution and geometry. Here, we report the first successful application of the (pseudo-)high dilution method to the conditions of on-surface synthesis in ultrahigh vacuum (UHV). This approach leads to high yields (up to 84%) of cyclic hyperbenzene (-honeycombene) _via_ an Ullmann-type reaction from 4,4''-dibromo-_meta_-terphenyl (DMTP) as precursor on a Ag surface. The mechanism of macrocycle formation was explored in detail using scanning tunneling microscopy (STM) and X-ray photoemission spectroscopy (XPS). We propose that hyperbenzene (MTP)\({}_{6}\) forms majorly by stepwise desilverization of an organometallic (MTP-Ag)\({}_{6}\) macrocycle, which preforms _via_ cyclisation of (MTP-Ag)\({}_{6}\) chains under pseudo-high dilution condition. The high probability of cyclisation on the stage of the organometallic phase results from the reversibility of the C-Ag bond. The case is different from that in solution, in which cyclisation typically occurs on the stage of covalently bonded open-chain precursor. This difference in the cyclisation mechanism on a surface compared to that in solution stems mainly from the 2D confinement exerted by the surface template, which to a large extent prevents the flipping of chain segments necessary for cyclisation.
Macrocycle, Pseudo-high dilution, On-surface synthesis, Ullmann reaction, Organometallic, Scanning tunneling microscopy, X-ray photoemission spectroscopy Tailoring the topology of polymers has intrigued chemists for many years, because a polymer's macroscopic properties depend inherently on its nanoscopic topology. Due to their specific "endless" topology, cyclic polymers bear significantly different characteristics compared to related linear (_i.e.,_ open-chain) molecules, including higher density, lower intrinsic viscosity, and higher thermostability. To date, the two most frequently used synthetic routes to cyclic polymers are high-dilution and molecular template methods performed in solution. The high-dilution principle, according to which low concentrations of the starting precursor favor cyclization over chain formation, was first developed by Paul Ruggli and Karl Ziegler for the cyclization of small organic molecules. The template method was developed to guarantee that the target molecule has the desired shape and thus was employed to synthesize various complicated macrocycles. An example for the latter is the recently reported Vernier templating synthesis of \(\pi\)-conjugated porphyrin nanorings with different diameters. The main drawback of these solution-based methods is that the target cyclic polymer or macrocycles cannot be arbitrarily designed due to the solubility requirements of the reactants. This drawback can be compensated by the on-surface synthetic approach under solvent-free conditions, which enable the polymerization and cyclisation of even insoluble reactive organic molecules directly on a solid surface. As an additional advantage of the on-surface approach, the properties of the target macrocycles can be explored _in situ_ with powerful surface science techniques, such as scanning tunneling microscopy/spectroscopy (STM/STS) and photoemission spectroscopy (PES). Recently, several macrocyles were successfully produced _via_ the on-surface Ullmann reaction, including honeycombenes, sexiphenylene, oligothiophene nanorings, and the (DAD)\({}_{6}\) cycle with bis(5-bromo-2-thienyl)-benzobis(1,2,5-thiadiazole) as the precursor monomer. Nevertheless, the yields of these macrocycles are still limited. This is mainly because chain-growth dominates over ring-closure under the typically employed reaction conditions on surfaces, _e.g._, precursor deposition rate in the range of 1 to 100 monolayers per hour (ML/h).
To enhance the probability of ring-closure in the ring/chain competition, we have transferred the high-dilution principle, which is well proven for synthesis in solution, to the solvent-free conditions of on-surface synthesis. This principle uses the fact that cyclization of a chain is a first-order reaction, while the chain growth by attachment of another precursor is a second-order reaction. With decreasing concentration, the first-order cyclization competes more and more successfully against the second-order chain growth. We show that rigorous application of the high-dilution principle to the synthesis of hyperbenzene (-honeycombene) from 4,4''-dibromo-_meta_-terphenyl (DMTP) on Ag results in considerably higher yield (84%) compared to that obtained under the usual conditions, _e.g._, in previous works. In addition, the mechanism of the high-yield formation of hyperbenzene under pseudo-high dilution condition is revealed. It is found that the cyclisation happens on the stage of an intermediate organometallic phase, which reversibly forms (MTP-Ag)\({}_{6}\) macrocycles. These organometallic macrocycles then undergo stepwise demetcalation forming hyperbenzene. This mechanism is different from the typical cyclisation mechanism in solution, which occurs on the stage of irreversibly (typically covalently) bonded open-chain precursors. Therefore, this study provides important general guidance for the high-yield on-surface synthesis of macrocycles.
## Results and Discussion
## On-surface pseudo-high dilution synthesis of hyperbenzene.
For an initial comparison between the pseudo-high dilution condition and typical high-concentration conditions, deposition of 0.6 monolayer (ML) 4,4'-dibromo-_meta_-terphenyl (DMTP) onto Ag surface at 353 K was first performed with high flux (5 ML/h) to obtain a high initial concentration of the DMTP monomers. This leads to the formation of ordered islands of zigzag chains as shown in The detailed structure of the zigzag chains is revealed by the magnified STM image in the inset, in which alternating corners and bright dot-like features are observed. According to our previous studies, the corner motifs and bright dots are assigned to the \(m\)-terphenyl (MTP) units and Ag atoms, respectively, as illustrated by the superimposed molecular model in the inset of In combination with the C-Br bond scission at 353 K, as evidenced by Br 3d XP spectra (Figure S1), the corner-to-corner distance of 16.0 A strongly indicates, by bond length considerations, that C-Ag-C bonds are formed between the MTP units. This organometallic phase can be viewed as a reservoir of MTP precursors with high initial concentration. The organometallic chains can undergo transformation into hydrocarbon chains or macrocycles _via_ elimination of the bridging Ag atoms and subsequent C-C coupling.
Subsequent annealing of the sample in to 463 K leads to the formation of both zigzag chains and hexagonal cycles as marked by the blue and red rectangles in Figure 1b, respectively. Their detailed structures are revealed in the zoom-in STM images in the inset of and 1c. For the zigzag chains, the corner-to-corner distance is reduced to 13.3 A, indicating that C-C bonds are now formed between the MTP units (see the overlaid molecular model). A similar shrinking occurs for the hexagonal cycles in Figure 1c, in which the corner-to-corner distance is 13.1 A. This suggests that hyperbenzene (_i.e._,-honeycombene) molecules are formed, as illustrated by the superimposed molecular model. Noteworthy, some of the hyperbenzene molecules have inclusions with bent and star-shaped features, which are assigned to single MTP units or diffusing Ag adatoms (see Figure S2 in the Supporting Information for structural details of the inclusions).
(a) STM image taken after deposition (flux \(f\) = 5 ML/h) of 0.6 ML DMTP onto Ag held at 353 K. The inset shows a magnified view of the green framed region. (b) STM image taken after annealing the sample in panel (a) to 463 K. The inset shows the blue framed region. (c) Zoom-in STM image of the red framed region in panel (b). Molecular models are overlaid. Grey spheres represent carbon atoms; white, hydrogen; cyan, silver. (d) Large-scale STM image taken after deposition of 0.5 ML DMTP onto Ag held at 463 K with a deposition rate of 0.05 ML/h. Tunneling parameters: (a) \(U\) = 1.4 V, \(I\) = 0.08 nA; (b), (c) \(U\) = 1.5 V, \(I\) = 0.13 nA; (d) \(U\) = 0.29 V, \(I\) = 0.19 nA.
The formation of zigzag oligophenylene chains with high yield (94%) and hyperbenzene with low yield (6%) by post-annealing of the organometallic phase is typical for high-concentrationconditions, and indeed the initial coverage of the MTP precursors (0.6 ML) is relatively high compared to the range that favors cyclization reaction: The two-dimensional (2D) coverage of 0.6 ML MTP corresponds to a three-dimensional (3D) concentration of approximately 0.6 mol/L. This value is much higher than the typical precursor concentrations of \(<10^{-4}\) mol/L previously used for high-dilution cyclization reactions performed in solution. To obtain pseudo-high dilution conditions in the surface reaction, a very low rate of \(f\)= 0.05 ML/h was now used during the deposition of DMTP onto the Ag surface held at 463 K. In this way, the two main requirements for the pseudo-high dilution condition were satisfied: The first requirement (I) is that the reaction rate should be higher than the rate of addition of the precursor, such that the precursor concentration will not build up during the reaction. In fact, it can be estimated that the stationary concentration of DMTP is lower than 5 \(\times\) 10\({}^{-4}\) ML for a flux of 0.05 ML/h (see part (d) in the SI for details). The second requirement (II) is that the reaction products should be stable under the reaction conditions. The requirement (I) was demonstrated by the fact that no excess of intact DMTP molecules exists, which was confirmed by STM and XPS measurements. The requirement (II) has been verified by previous studies on both Ag and Au, in which the covalent products are very stable under the reaction conditions. Remarkably, this procedure leads to a high yield (84%) of hyperbenzene macrocycles, as shown by the large-scale STM image (and Figure S3a2) of the as-prepared sample. The result indicates that the pseudo-high dilution method on a surface indeed works similarly successful to that in solution.
In addition, the relationship between the yield of hyperbenzene and the deposition rate, \(f\), of DMTP has been investigated in detail. Figure 2a-e shows overview STM images taken after deposition of 0.5 ML DMTP onto the Ag surface at 463 K with different deposition rates of 0.08, 0.5, 5, 50, and 150 ML/h. The image series shows that the yield of hyperbenzene (greenareas) decreases dramatically with increasing deposition rate \(f\). plots the variation of the hyperbenzene yield with \(f\), as derived from statistical analysis of the STM image series in Figure S3. The data points can be well fitted with an exponential curve (red) and illustrate that a low concentration of the DMTP precursors strongly favors the formation of hyperbenzene rings.
Overview STM image taken after deposition of 0.5 ML DMTP onto Ag held at 463 K with different evaporation rates: (a) 0.08 ML/h, (b) 0.5 ML/h, (c) 5 ML/h, (d) 50 ML/h, and (e) 150 ML/h. Tunneling parameters: (a) \(U=1.2\) V, \(I=0.13\) nA; (b) \(U=1.3\) V, \(I=0.21\) nA; (c) \(U=1.6\) V, \(I=0.12\) nA; (d) \(U=1.5\) V, \(I=0.16\) nA; (e) \(U=1.6\) V, I = 0.13 nA. The islands of hyperbenzene and zigzag oligophenylene chains are indicated by green and grey shading, respectively. (f) Hyperbenzene yield versus DMTP deposition rate \(f\). See the text for the meaning of the blue and red curves.
## Cyclisation on the stage of organometallic intermediates.
To obtain deeper insight into the mechanism of the hyperbenzene formation, the reaction temperature was lowered by 20 K to 443 K during the deposition of DMTP. In this way, we were able to observe intermediate species for the formation of hyperbenzene. shows the overview STM image taken after deposition of 0.5 ML DMTP onto the Ag surface at 443 K with a flux of \(f\)= 0.5 ML/h. Domains of two different hexagonal species were observed on the surface as labelled by the green and blue frames. The structure of the hexagon with larger cavity was revealed by the zoom-in STM image in The hexagons contain alternatingly arranged corners and protrusions, which are assigned to MTP units and Ag atoms, respectively, similar to those in the zigzag organometallic chains. In addition, the corner-to-corner distances of the hexagon are 16.0 A, in agreement with those of the zigzag organometallic chains.
Besides the organometallic (MTP-Ag)\({}_{6}\) species, hexagons with a lower number of Ag atoms, (MTP)\({}_{6}\)Ag\({}_{\mathrm{x}}\) (x\(<\)6), were also observed on the sample in shows an area with a mixture of hyperbenzenes (MTP)\({}_{6}\) and two-fold metalated hexagons (MTP)\({}_{6}\)Ag\({}_{2}\) (molecular model overlaid in Figure 3c). The hyperbenzene is marked with a white dotted hexagon. Its identity is verified by the uniform lengths of the six sides of the hexagon motifs, in line with that in The (MTP)\({}_{6}\)Ag\({}_{2}\), labelled with distorted yellow dotted hexagons in Figure 3c, appears as an elongated hexagon with four short and two long sides. Its structural assignment has been confirmed by the following considerations. First, protrusions were observed in the center of the two long sides. These protrusions are attributed to Ag atoms. In addition, the corner-to-corner distances (15.9 A) of the longer sides indicate that the involved two MTP units are linked withC-Ag-C bonds, in line with the distances (16.0 A) in the organometallic chains. Second, the four short sides with corner-to-corner distances of 13.1 A suggest a direct covalent C-C linkage between the two MTP units.
(a) STM image taken after deposition (\(f\) = 0.5 ML/h) of 0.5 ML DMTP onto Ag held at 443 K. (b),(d) Zoom-in of the green and blue framed region in panel (a). (c) Section of partially metalated hyperbenzene molecules observed on the sample in panel (a). Tunneling parameters: (a), (b) \(U\) = 1.4 V, \(I\) = 0.15 nA; (c) \(U\) = 0.44 V, \(I\) = 0.19 nA; (d) \(U\) = 0.88 V, \(I\) = 0.17 nA. (e) Scheme for the evolution of cyclic (MTP-Ag)\({}_{6}\) to hyperbenzene (MTP)\({}_{6}\). Grey spheres represent carbon atoms; white, hydrogen; cyan, silver.
Furthermore, singly metalated hexagons (MTP)\({}_{6}\)Ag with only one Ag atom were observed, as can be seen in the blue-framed domain (bottom right) in The magnified STM image shows, apart from regular hyperbenzene (example marked with white dotted hexagon), also irregular hexagons (marked by green dotted hexagons) with one long and five short sides. The single long side of the irregular hexagon has a corner-to-corner distance of 15.6 A, similar to that of two long sides in cyclic (MTP)\({}_{6}\)Ag\({}_{2}\). This strongly implies that the irregular hexagons are (MTP)\({}_{6}\)Ag, which is proposed to be the final intermediate that leads to the formation of hyperbenzene by elimination of the last remaining Ag atom. Therefore, we propose that the macrocycle formation occurs at the stage of organometallic phase under pseudo-high dilution condition.
Cyclization reactions in solution can often be treated with the conventional Jacobson-Stockmayer (J-S) model, which describes the yield of cyclic products as a function of the reactant concentrations in high-dilution systems. According to the J-S model, the mechanism for the formation of (MTP-Ag)\({}_{6}\) macrocycles can be derived as shown in Scheme 1. Six DMTP monomers initially react with six Ag atoms forming a six-membered (MTP-Ag)\({}_{6}\) organometallic chain with five C-Ag-C bridges. This chain then undergoes cyclisation forming the (MTP-Ag)\({}_{6}\) macrocycle or grows further into a seven-membered (MTP-Ag)\({}_{7}\) chain by addition of another MTP-Ag unit. The competition between the cyclisation and chain growth finally determines the yield ratio of (MTP-Ag)\({}_{6}\) macrocycles and (MTP-Ag)\({}_{7}\) chains. Since this competition depends strongly on the precursor concentration according to the high-dilution principle, the yield of (MTP-Ag)\({}_{6}\) macrocycles can be derived as a function of the precursor deposition rate:\[\%\ (\text{MTP-Ag})_{6}\ \text{cycles}=\frac{1}{1+\frac{K_{L}}{6K_{C}}f\tau} \times 100 \tag{1}\]
## Scheme 1.
Growth (light blue shaded) and cyclisation (light brown shaded) of all-_trans_ configured organometallic (MTP-Ag)\({}_{6}\) Chain 1. The isomerization of organometallic (MTP-Ag)\({}_{6}\) Chain 1 into Chain 6 is achieved by three steps: detachment, in-plane rotation, reattachment of MTP units. \(K_{C}\) and \(K_{L}\) denote the rate coefficients for cyclization and chain growth process, respectively. The details of the 4 isomerization steps from (MTP-Ag)\({}_{6}\) Chain 1 to 5 are shown in Scheme S1 in the supporting information.
For the formation of the (MTP-Ag)\({}_{6}\) macrocycle, _K\({}_{C}\)_ and _K\({}_{L}\)_ are constants depending solely on the mean square end-to-end distance of the six-membered (MTP-Ag)\({}_{6}\) chain.\(\tau\) is the reaction time for C-Ag-C bond formation, which should be constant at a particular reaction temperature. Considering the observation that hyperbenzene forms from the (MTP-Ag)\({}_{6}\) macrocycle, the yield of hyperbenzene should be proportional to the yield of (MTP-Ag)\({}_{6}\) macrocycles _i.e._, it should depend reciprocally on \(f\), according to the J-S model and Equation 1. In reality, however, the reciprocal function does not fit to the obtained experimental results well (see the blue curve in Figure 2e). Especially at high deposition rates, the yield of hyperbenzene is higher than that expected by the J-S model.
We attribute the observed deviation from the J-S model to the following differences between on-surface cyclization and cyclization in solution: First, in the J-S model, chains can only grow or remain unchanged, but they cannot become shorter. This is a realistic scenario for solution synthesis with typically direct covalent coupling between the precursor units. On the surface, however, cyclization occurs on the stage of the organometallic species. Since the C-Ag-C bond formation is reversible, chains that have already grown too long for hyperbenzene formation (_i.e._, (MTP-Ag)\({}_{\rm x}\) with x \(>\) 6) have a certain probability of dissociation into shorter chains (x \(\leq\) 6) that can still contribute to the formation of hyperbenzene. This is one of the reasons why the yield of hyperbenzene is higher than predicted by the J-S model, especially in the range of higher fluxes, where the chains growth dominates. Second, the (MTP-Ag)\({}_{6}\) open chains have with 20 possible configurations due to the random arrangement of the _cis_ or _trans_ connections between the MTP units (see Scheme S1 in the SI for selected examples). One of these configurations, the all-_cis_ Chain 6 (Scheme 1), is the direct conformational precursor for the cyclisation into the (MTP-Ag)\({}_{6}\) macrocycle. If the Chain 6 configuration is formed directly from MTP units and Ag,rapid formation of the (MTP-Ag)\({}_{6}\) macrocycle is possible without prior isomerization. Since the isomerization requires high-dilution conditions (because otherwise further chain growths is favored), this direct cyclization of Chain 6 increases the macrocycle formation rate compared to the J-S model for high-flux conditions. Note that this effect is caused by the 2D confinement of the chains, _i.e._, it results from the templating influence of the surface.
For the (MTP-Ag)\({}_{6}\) chains with other configurations, _e.g._, the all-_trans_ configured (MTP-Ag)\({}_{6}\) Chain 1 (Scheme 1), their cyclisation demands the preceding isomerization into (MTP-Ag)\({}_{6}\) Chain 6. This is only realistically possible by the detachment, in-plane rotation, and re-attachment of the MTP units (_e.g._, from (MTP-Ag)\({}_{6}\) Chain 5 to 6, Scheme 1) under the prerequisite of pseudo-high dilution condition. (The alternative flipping of chain segments through 3D space is highly unlikely, because it would require the intermediate desorption of units larger than (MTP)\({}_{2}\)Ag\({}_{2}\), see also the Scheme S2 in the SI for a related all-covalent chain.) Without the detachment/re-attachment mechanism, the chain growth would dominate due to the high probability of attaching an additional MTP unit (light blue shaded region in Scheme 1).
Scheme S1 shows the possible six steps for the isomerization of the all-_trans_ configured (MTP-Ag)\({}_{6}\) chain 1 into the all-_cis_ configured Chain 6. In each step, the cleavage of a single C-Ag bond enables the detachment of the MTP units from the chain. Subsequently, the in-plane rotation and recombination of the MTP units eventually result in the isomerization of the (MTP-Ag)\({}_{6}\) chains.
Noteworthy, once the (MTP-Ag)\({}_{6}\) Chain 6 cyclize into the (MTP-Ag)\({}_{6}\) macrocycle, the detachment of MTP units is energetically unfavorable compared to that from a chain, because the simultaneous scission of two C-Ag-C bridges is required. This selection process finally leads to the building up of the yield of (MTP-Ag)\({}_{6}\) macrocycles.
An additional significant aspect is the template effect of the surface beyond the simple 2D confinement mentioned above. The surface lattice favors the formation of products with matching symmetry. Such template effects have previously been reported for related systems, in particular in comparing the topology of DMTP-derived species on six-fold symmetric Cu and two-fold symmetric Cu. Therefore, the formation of six-fold symmetric (MTP-Ag)\({}_{6}\) cycle is favored in the initial reaction of DMTP monomers with Ag atoms. This also contributes to the higher than expected hyperbenzene yield at high deposition rates \(f\).
## Cyclisation on the stage of covalent species.
Figure 4a, 4b and Figure S4a show STM images taken after deposition of 0.5 ML DMTP (_f_ = 0.5 ML/h) onto the Ag surface held at elevated temperatures of 483, 503, and 523 K. It can be seen that, with increasing substrate temperatures, the yield of hyperbenzene (green region) decreases along with an increase of both yield and lengths of oligophenylene chains (grey regions). shows the corresponding length distributions of the oligophenylene chains formed on Ag held at 463, 483, 503, and 523 K. All the fitted curves shows a Poisson type length distribution, indicating a chain-growth polymerization process of DMTP. In addition, the most probable length of the oligophenylene chains rises with increasing substrate temperature. These phenomena can be interpreted as follows.
The temperature of 463 K represents the threshold for C-C coupling of DMTP on Ag. If the reaction is performed at this temperature, there is sufficient time for the cyclisation of the organometallic (MTP-Ag)\({}_{6}\) chains, because demetcalation of cyclic (MTP-Ag)\({}_{6}\) to hyperbenzene is slow. With increasing substrate temperatures, the life-times of the organometallic species become shorter, which undermines the cyclisation process on the stage of organometallic phase. Then, the formation of hyperbenzene proceeds more and more likely through the cyclisation of preformed (MTP)\({}_{6}\) oligophenylene chains (Scheme 2) rather than stepwise desilverization of preformed (MTP-Ag)\({}_{6}\) macrocycle.
However, due to the irreversibility of the C-C bond formation, the isomerization of oligophenylene chains necessary for cyclisation must follow a different mechanism with higher
STM images taken at 300 K after deposition (\(f\) = 0.5 ML/h) of 0.5 ML DMTP onto Ag held at (a) 483 K, (b) 503 K, and (c) 523 K. The islands of hyperbenzene and zigzag oligophenylene chains were marked out by green and grey shading. (d) Magnified view of a section of the macrocycle species formed on the sample in panel (c). Tunneling parameters: (a) \(U\) = 1.6 V, \(I\) = 0.12 nA; (b) \(U\) = 1.2 V, \(I\) = 0.13 nA; (c) \(U\) = 0.97 V, \(I\) = 0.24 nA; (d) \(U\) = 0.76 V, \(I\) = 0.12 nA. (e) Length distributions of the zigzag oligophenylene chains on the samples prepared by deposition of 0.5 ML DMTP onto Ag held at 463 (black), 483 (red), 503 (green), and 523 (blue) with Poisson fits. (f) Ring size distributions of the formed macrocycles on the sample in panel (c).
This leads to a considerably lower probability of cyclisation on the stage of covalent species and hence the lower hyperbenzene yield, as explained in the following: For the organometallic (MTP-Ag)\({}_{6}\) chains, their isomerization into an all-_cis_ configured chain is achieved by C-Ag bond scission and reformation. This process allows the detached MTP units undergo in-plane rotation (Scheme 1), _i.e._, the reaction can proceed in 2D confinement. For the (MTP)\({}_{6}\) oligophenylene chains, the isomerization occurs exclusively by the flipping of chain moiety (red colored) through 3D space, as illustrated in Scheme 2. This means parts of the (MTP)\({}_{6}\) oligophenylene chain desorbs from the surface. Because of this intermediate desorption, hydrocarbons with large molecular weight have the flipping diffusion barriers that are typically higher than the in-plane sliding diffusion barriers. This can easily be understood considering that the flipping and sliding diffusion barriers on a surface relate to the total depth (adsorption energy) and the corrugation of the surface potential, respectively. Estimates of the flipping barriers can be derived from temperature programmed desorption data. For example, the desorption temperature for a submonolayer of _para_-quaterphenyl on Au is around 550 K. Because of the typically larger adsorption energy of phenyl units on Ag than Au, the desorption temperature of _para_-quaterphenyl on Ag should be even higher. This means it is highly unlikely that the pentaphenyl moiety (red colored) in the (MTP)\({}_{6}\) Chain 4 (Scheme 2) can flip over to form chain 5 on Ag at reaction temperatures below 550 K. Thus, except for the (MTP)\({}_{6}\) Chains 5 and 6, for which only small units need to flip, other (MTP)\({}_{6}\) chains are completely blocked for cyclisation under the employed reaction temperatures (483 K, 503 K, and 523 K) due to the high barrier for flipping over of moieties bigger than _para_-quaterphenyl. However, the probability of the formation of (MTP)\({}_{6}\) Chain 5 and 6 (Scheme 2) by C-C coupling of six MTP units is only \(\frac{3}{32}\).
Therefore, even under the pseudo-high dilution condition, the probability of cyclisation on the stage of covalent species is low. This finally leads to the reduced yield of hyperbenzene at high reaction temperatures.
## Scheme 2.
Growth (light blue shaded) and cyclisation (light brown shaded) of all-_trans_ configured (MTP)\({}_{6}\) oligophenylene chain 1. The isomerization process of (MTP)\({}_{6}\) chains necessary for cyclisation is achieved by flipping over of the chain moieties (red colored). \(K_{C}\) and \(K_{L}\) denote the rate coefficients for cyclization and chain growth process, respectively. The details of the 3 isomerization steps from (MTP)\({}_{6}\) Chain 1 to 4 are shown in Scheme S2 in the supporting information.
Besides, with the increasing of the substrate temperature, higher sliding diffusion rates of the oligophenylene chains were achieved. This enhanced the probability for the meeting of the ends of two different chains, resulting in the formation of longer chains. Moreover, the formed longer chains cannot dissociate into shorter chains because of the irreversibility of C-C bond formation. As one of the consequences, oligophenylene macrocycles with different sizes are formed at a reaction temperature of 523 K,, while hyperbenzene (marked green) only appears as a minority species. According to the statistics acquired from the large-scale STM image in Figure S4a, the total yield of macrocycles is around 5%, to which hyperbenzene contributes only 0.5%. The magnified STM image in reveals the detailed structures of the products. Apart from the hexagonal hyperbenzene, the larger macrocycles (MTP)\({}_{10}\), (MTP)\({}_{12}\), (MTP)\({}_{14}\), and (MTP)\({}_{16}\) were formed, as indicated by the white labels. The size distribution of the macrocycles is given in As can be seen, the cyclic decamer (MTP)\({}_{10}\) reaches the highest relative yield. The formation of larger macrocycles can again be attributed to the fact that the lengths of oligophenylene chains ready for cyclization has been enlarged due to the dominance of an irreversible chain growth process at high temperatures, as was discussed above. Noteworthy, no (MTP)\({}_{\text{x}}\) cycles with x = 1-5, 7-9, 13, and 15 are formed on the surface, probably because these (MTP)\({}_{\text{x}}\) macrocycles would have ring strain. The major source of overall strain for these macrocycles would be Baeyer strain arising from the deformation of C-C \(\sigma\) bonds between the phenyl units. (MTP)\({}_{\text{x}}\) cycles with x = 6, 10, 12, 14, and 16 adopt a conformation with less Baeyer strain (no C-C \(\sigma\) bonds deformation) and thus are energetically more favorable to form.
Our observation that cyclisation on the stage of covalent species undermines the yield of hyperbenzene was further supported by otherwise identical experiments on the Au surface as discussed in the following: shows an STM image taken after low-flux deposition (\(f\) = 0.5 ML/h) of 0.5 ML DMTP onto Au held at 463 K. The yield of hyperbenzene (green region) is only 9%, which is a surprisingly low value compared to the high yield of 72.5% achieved on Ag at 463 K. Note that the 463 K is again the threshold temperature of C-C coupling on Au. This is evidenced by the coexistence of oligophenylene chains and intact DMTP on Au at 443 K (see Figure S5c). The most probable factor accounting for this large yield difference under otherwise identical conditions on Au and Ag is the intrinsically shorter life-time of the organometallic species with C-Au-C bonds comparing to those with C-Ag-C bonds. The short life-time of the C-Au bond is supported by the fact that no stable organometallic species formed from DMTP on the Au surface as confirmed by Figure S5 (and is in agreement with the rarely reported gold-organic intermediates). With the short life-time of C-Au-C bond, cyclisation is more likely to occur on the stage of covalent species, which results in the low cyclisation probability and hence the low yield of hyperbenzene.
Similar to the situation on the Ag surface, increasing the temperature of Au during deposition (\(f\)= 0.5 ML/h) of submonolayer of DMTP leads to even more reduced yields of hyperbenzene due to the further decrease of the life-time of the organometallic species. This is manifested as reduced number of hyperbenzene molecules (marked in green) in the STM image series: (463 K), 5b (483 K), 5c (503 K), and Figure S6 (523 K). In addition, the chain length distribution in shows an increase of the most probable lengths of the oligophenylene chains with increasing temperatures of the Au substrate. As can be seen in Figure 5f, the most probable chain length correlates with the substrate temperature on both metal surfaces, and longer chains are observed on Au under otherwise identical conditions. This can be attributed to the higher sliding diffusion rates of the oligophenylene chains achieved on Au than Ag under otherwise identical conditions, which favors the chain growth of the \(\sim\) 1.8 V, \(I\) = 0.10 nA; (c) \(U\) = 1.9 V, \(I\) = 0.16 nA. (e) Length distributions of zigzag oligophenylene chains on the samples prepared by deposition of submonolayer DMTP onto Au held at 463 K (black), 483 K (red), 503 K (green), and 523 K (blue) with Schulz-Flory and Poisson fits. (f) Dependence curve: most probable lengths of zigzag oligophenylene chains versus substrate temperatures for both Au and Ag.
STM images taken after deposition (\(f\)= 0.5 ML/h) of 0.5 ML DMTP onto Au held at (a) 463 K, (b) 483 K, and (c) 503 K. The islands of hyperbenzene and zigzag oligophenylene chains were marked out by green and grey shading. (d) A section of the sample in panel (b) with large macrocycles (marked with cyan). Tunneling parameters: (a) \(U\) = 1.2 V, \(I\) = 0.14 nA; (b), (d) \(U\) = 1.8 V, \(I\) = 0.10 nA; (c) \(U\) = 1.9 V, \(I\) = 0.16 nA. (e) Length distributions of zigzag oligophenylene chains on the samples prepared by deposition of submonolayer DMTP onto Au held at 463 K (black), 483 K (red), 503 K (green), and 523 K (blue) with Schulz-Flory and Poisson fits. (f) Dependence curve: most probable lengths of zigzag oligophenylene chains versus substrate temperatures for both Au and Ag.
Furthermore, the obtained longer chains have certain probabilities of cyclisation into larger macrocycles. This is supported by the observation of larger covalent oligophenylene macrocycles (marked with cyan in Figure 5d) on Au than Ag at 483 K, including (MTP)\({}_{10}\) and (MTP)\({}_{14}\), (MTP)\({}_{18}\), (MTP)\({}_{19}\), and (MTP)\({}_{48}\).
## Conclusions
In conclusion, we have experimentally explored the possibilities for (pseudo-)high dilution conditions in macrocycle synthesis on metal surfaces for the first time. By employing an extremely low deposition rate of 0.05 ML/h, we have successfully achieved the high-yield (84%) formation of hyperbenzene (-honeycombene) from DMTP on Ag at 463 K. The mechanism for the formation of hyperbenzene was explored in detail. In the range of the threshold temperature for C-C coupling, e.g. at 463 K, the hyperbenzene is formed majorly by stepwise desilverization of (MTP-Ag)\({}_{6}\) macrocycles, which preformed _via_ cyclisation of (MTP-Ag)\({}_{6}\) chains under pseudo-high dilution conditions. Above the threshold temperature for C-C coupling, the cyclisation process increasingly proceeds by cyclisation of covalent (MTP)\({}_{6}\) oligophenylene chains due to the reduced life-time of organometallic species. The cyclisation probability of (MTP)\({}_{6}\) oligophenylene chains is considerably lower than that of the (MTP-Ag)\({}_{6}\) organometallic chains because of the higher conformational isomerization energy barrier of the oligophenylene precursor chain. Therefore, with the shortening of the life-time of organometallic species, the probability of cyclisation decreases, which finally leads to the reduced yield of hyperbenzene. This observation is further supported by the considerably lower yield of hyperbenzene obtained under otherwise identical conditions on Au, in which system the organometallic species with C-Au-C bonds has intrinsically shorter life-time than that with C-Ag-C bonds. Therefore, we conclude that the high yield macrocycle formation on surfaces reliesstrongly on the efficient cyclisation process on the stage of a reversible (e.g., organometallic) species and a highly diluted precursor. Compared to macrocycle synthesis in solution, the additional reversibility condition as a direct consequence of the 2D confinement makes the detailed understanding of the reaction mechanism and the optimization of the reaction temperature particularly critical.
## Experimental Section
The experiments were performed in two separate UHV systems. The STM measurements were performed in a two-chamber UHV system, which has been described previously, at a background pressure below 10\({}^{\text{-}10}\) mbar. The STM probe is a SPECS STM 150 Aarhus with SPECS 260 electronics. All voltages refer to the sample and the images were recorded in constant current mode. Moderate filtering (Gaussian smooth, background subtraction) has been applied for noise reduction. The Ag and Au single crystals with an alignment of better than 0.1\({}^{\circ}\) relative to the nominal orientation were purchased from MaTecK, Germany. Preparation of a clean and structurally well controlled Ag and Au surface was achieved by cycles of bombardment with Ar\({}^{+}\) ions and annealing at 800 K and 850 K, respectively. As described elsewhere, 4,4\({}^{\text{"}}\)-dibromo-1,1':3',1'-terphenyl (4,4\({}^{\text{"}}\)dibromo-_meta_-terphenyl, DMTP) was made from 4-bromophenylacetylene in a short reaction sequence utilizing a Grubbs-enyne metathesis reaction and a regioselective cobalt-catalyzed Diels-Alder reaction followed by mild oxidation. DMTP was vapor-deposited from a commercial Kentax evaporator with a Ta crucible held at different temperatures for different deposition rates: 329 K (0.05 ML/h), 333 K (0.08 ML/h), 339 K (0.5 ML/h), 367 K (5 ML/h), 389 K (150 ML/h). All STM images were recorded at a sample temperature of 300 K. Coverages were derived from STM images. The DMTP coverage of 1 monolayer (ML) refers to a densely packed layer of zigzag oligophenylene chains and equals 0.052 DMTP molecules per surface Ag atom. The XPS measurements were performed on the Catalysis and Surface Science Endstation located in National Synchrotron Radiation Laboratory (NSRL), Hefei. The detailed description of the endstation can be found elsewhere. The XPS spectra were collected at an emission angle of 60\({}^{\circ}\) with respect to the surface normal.
|
10.48550/arXiv.1703.10018
|
On-Surface Pseudo-High Dilution Synthesis of Macrocycles: Principle and Mechanism
|
Qitang Fan, Tao Wang, Jingya Dai, Julian Kuttner, Gerhard Hilt, J. Michael Gottfried, Junfa Zhu
| 3,039
|
10.48550_arXiv.0806.4141
|
###### Abstract
In an effort to develop a chemically reactive interaction potential suitable for application to the study of conventional, organic explosives, we have modified the diatomic AB potential of Brenner _et al._ such that it exhibits improved detonation characteristics. In particular, equilibrium molecular dynamics (MD) calculations of the modified potential demonstrate that the detonation products have an essentially diatomic, rather than polymeric, composition and that the detonation Hugoniot has the classic, concave-upward form. Non-equilibrium MD calculations reveal the separation of scales between chemical and hydrodynamic effects essential to the Zel'dovitch, von Neumann, and Doring theory.
## I Introduction
For over 15 years, the Reactive Empirical Bond-Order (REBO) potential of Brenner _et al._, or variations thereof, have been used for Molecular Dynamics (MD) simulations of detonation. It has been shown to follow the Chapman-Jouguet (CJ) theory and even that of Zel'dovich, von Neumann, and Doring (ZND). Being an MD model, it has at least two major shortcomings, namely its spatial and temporal scales. As computers become more powerful and codes and algorithms become more advanced, some of the restrictions that limit these scales can be loosened, allowing for more realistic behavior to be modeled.
While providing an atomic scale model of a reactive material, the empirical and classical model of Brenner _et al._ falls short of including all aspects of an accurate representation. Additionally, the spatial and temporal scales of real explosions make MD simulations a large computational task. But MD is still a useful tool for probing the characteristics of the detonation phenomenon, and the REBO potential is one of the best at balancing realism in the potential with accessibility to the large-scale; yet there is room for improvement.
MD has certain conceptual advantages over hydrodynamics approaches that are parameterized to match the behavior of real high explosives. Even though the latter models can mimic real experiments on the proper spatial and temporal scales, they make assumptions about the reaction rate and multiphase equation of state (EOS) to do so. Some of these assumptions are parameter fittings, which may not elucidate any new physical intuition about detonation of the high explosive (HE) in question. In comparison, MD simulations depend on the parameterization of an interaction potential, which is arguably easier to connect directly to physical considerations than is a multiphase and multi-species EOS.
REBO has been used by many groups to model a variety of parameterizations and experimental configurations. However, a major criticism of this potential is that it has a thin reaction zone (\(\sim 100\) A) relative to typical real high explosives (\(\sim 1\) mm). In previous work several unconventional characteristics of the default parameterization of Brenner _et al._ (called Modell as in) are made evident:
* ModelI displays nearly instantaneous dissociation upon compression by an unsupported detonation. Its entire reaction zone is characterized by a dissociative state.
* Its reaction zone is thin.
* It readily allows for clustering.
* Its CJ state is highly compressed.
* It has clearly non-hyperbolic equilibrium Hugoniots (\(\mathcal{H}\)) in \(P\)-\(v\) space.
None of these characteristics are proven to be unrealistic. In fact for some primary high explosives, the plasma-like state seen in Modell at CJ may be realistic. Clustering, particularly of carbon, is a real phenomenon in the thermal decomposition of some HEs. However, successful models of conventional explosives have either assumed or suggested that more molecular states with hyperbolic \(\mathcal{H}\)s in \(P\)-\(v\) space are typical and that compressions at CJ conditions should be \(\approx 25\%\) (for example PETN) rather than \(\approx 43\%\) as with Modell. Given that Modell can model a dissociated state, it would be useful to show that a more molecular state can also be modeled. The task of this work is to make modifications to the Modell potential in order to accomplish this goal. Although, for the last feature in the above list, it is difficult to predict what changes would adjust it, an attempt to address the former features is made by making adjustments based on physical reasoning (see Sec. II). In Sec. IV the effects that the changes to the model of Brenner _et al._ have on the thermodynamic properties are investigated by the methods outlined in Sec. III. Notable results of non-equilibrium MD simulations are studied in the companion paper.
## II Changing Rebo: Physical and aesthetic reasoning
In this section we motivate changes to the Modell potential to form a new potential (called ModelIV after the naming scheme of White _et al._), the total bonding energy of which takes the form
\[\begin{split} E_{b}=\sum_{i}^{N}\sum_{j>i}^{N}&\{f_ {c}(r_{ij})[(2-\overline{B}_{ij})V_{R}(r_{ij})-\overline{B}_{ij}V_{A}(r_{ij} )]+\\ & V_{vdW}(r_{ij})\}.\end{split} \tag{1}\]The parameters for Eq.1 can be found in Table 1. Before mentioning the physical basis for the modifications made to Modell, let us list the features that ModelI already takes into account. In tests using Modell, the parameterization is set to a valence of one, which should prefer dimers because one atom's sharing of electrons with one other fills both atoms' valence shells.
\[E_{b}=\sum_{i}^{N}\sum_{j>i}^{N}\{f_{c}(r_{ij})[V_{R}(r_{ij})-\overline{B}_{ij} V_{A}(r_{ij})]+V_{vdW}(r_{ij})\}. \tag{2}\]
See Table 1 in the errata of Brenner _et al._ for the functions and components of Eq. 2. In ModelI this bond-order behavior is implemented by reducing the strength of the attractive term of a Morse potential used to model such covalent bonding. This is accomplished with a bond-order coefficient (\(\overline{B}_{ij}\)) that varies from one to zero with increasing proximity and number of neighbors.
Rice _et al._ found that Modell allowed for trimer formation and subsequently strengthened this bond-order effect by increasing the repulsive term as well in order to better model bond saturation so that a particle became less likely to bond to more than one other. The contribution of Rice _et al._ is a simplified adjustment for what can be a complicated interaction. For instance, the \(\overline{B}_{ij}\) coefficient does not take spin or degeneracy into account. However, this type of correction is similar to the "bond saturation" term that is used in the more sophisticated and calibrated ReaxFF potential of van Duin _et al._.
In combination with this bond-order functionality, the variation of bond well depths allows for reactive chemistry. If \(i\) and \(k\) are different types of atoms, their respective well depths with \(j\) will be different. If \(j\)'s is deeper with \(k\) than with \(i\), \(k\) will require less kinetic energy to displace \(i\) than \(i\) would have, had their rolls been reversed. That difference in energy will be converted into the kinetic energy of the system, an exothermic reaction. This is the type of reaction modeled by ModelII, \(2\text{AB}\rightarrow\text{A}_{2}+\text{B}_{2}+2Q\), where \(Q\equiv D_{e}^{\text{ANABB}}-D_{e}^{\text{AB}}\) is the exothermicity of the reaction. The amount of energy needed to dissociate AB is \(D_{e}^{\text{AB}}\). To dissociate AA or BB requires \(D_{e}^{\text{ANABB}}\). These numbers ignore the small contribution of the \(V_{vdW}\) term in Eq. 2.
In Modell the van der Waals (vdW) interaction is represented by a Lennard-Jones (LJ) form. It is connected with a terminating third-order polynomial with negative curvature in the short range domain so that the overly repulsive twelfth-order term does not compete with the Morse potential, which handles the covalent bonding and Pauli Exclusion repulsion. The LJ is parameterized to allow for solid lattice formation at low temperature. It turns out that the third-order inner spline introduces an artifact in Modell since it has a section of positive slope, which allows it to trap particles (see Fig. 1) because of the resulting attractive force. It is only first-derivative continuous at the spline point and is cropped at the cutoff point.
Modell simulates well the repulsion between individual atoms separated by less than their equilibrium distance--which depends on the number and proximity of neighbors to the pair. It does not, however, represent the electrostatic repulsion between dimers that occurs as their charge clouds overlap on approaching each other while still being too far away to rearrange bonds, \(\sim 3\) A. As a result there is no significant interaction between molecules at the range and magnitude that gives rise to high pressure dense molecular HE product fluid mixtures. There are some good models for this, for example, ZBL and ReaxFF. In an attempt to improve upon Modell and its results, we design a new model (Mode-IIV), which incorporates, along with the aforementioned contribution of Rice _et al._, the following modifications:
The inner spline is replaced with a repulsive core, which smoothly connects the negative (the zero of the potential energy is the dissociated state) LJ curve to a constant plateau in the region in which the Morse potential dominates. This inner core has no sections of positive slope--that is, sections of attractive force--and therefore cannot trap atoms, and it crudely models electrostatic repulsion.
The reason for connecting to a plateau is that it allows one to dictate the depth of the bonding potential. It also simplifies the definition of a covalent bond. Any two particles that are within the defined bond distance (see the \(f_{c}\) term in Table 1) with a radial component of kinetic energy less than the height of the plateau are considered bonded.
\begin{table}
\begin{tabular}{l l l} \(V_{R}(r)\) & \(=\) & \(\frac{\mathcal{D}_{e}}{S-1}\exp\left[-\beta\sqrt{2S}(r-r_{e})\right]\) \\ \(V_{A}(r)\) & \(=\) & \(\frac{S\mathcal{D}_{e}}{S-1}\exp\left[-\beta\sqrt{\frac{2}{S}}(r-r_{e})\right]\) \\ \(B_{ij}\) & \(=\) & \(\left\{1+G\sum_{k\neq i,j}f_{c}(r_{ik})\exp[m(r_{ij}-r_{ik})]\right\}^{-n}\) \\ \(y(r)\) & \(=\) & \(\frac{r-\gamma_{1}}{\gamma_{2}-\gamma_{1}}\) \\ \(f_{c}(r)\) & \(=\) & \(\left\{\begin{array}{ll}1&y(r)\ <0\\ (1-y(r))^{3}(1+3y(r)+6y^{2}(r))&0\leq y(r)\ <1\\ 0&1\leq y(r)\end{array}\right.\) \\ \(V_{vdW}(r)\) & \(=\) & \(\left\{\begin{array}{ll}\epsilon ec\\ \epsilon\left(c+\sum_{i=3}^{5}P_{i}(r-\alpha_{1})^{i}\right)&\alpha_{1}\leq r \ <\alpha_{2}\\ \epsilon\left[\left(\frac{r^{*}}{r}\right)^{12}-2\left(\frac{r^{*}}{r}\right)^{6 }\right]&\alpha_{2}\leq r\ <\alpha_{3}\\ \epsilon(C_{0}+C_{1}r+C_{2}r^{2}+C_{3}r^{3})&\alpha_{3}\leq r\ <\alpha_{4}\\ 0&\alpha_{4}\leq r\end{array}\right.\) \\ \hline \end{tabular} \(\begin{array}{l}\text{Mass}_{\text{AA}}=14.008\text{~{}ann}\text{~{}max}\text{~{}} \text{}\fined by neighbors, and it can be determined at any instant. The height of the plateau, 1 eV, is chosen to be comparable to the thermal energy at the CJ state for ModelI, an addition that should decrease the reaction cross section, desensitizing the HE to initiation and, perhaps, adjusting reaction time and, thus, the width of the reaction zone. Perhaps, a more physical function would continue to increase monotonically to a finite value at \(r=0\) and have \(V_{vdW}\) be particle-type dependent.
All of the spline points in ModelIV have, at least, continuous second derivatives, with the exception of the outermost cutoff point. To terminate the \(V_{vdW}\) term in the long range, a Holian-Evans spline, which is second derivative continuous at its inner spline point but only first derivative continuous at its outer one, is used. A computational advantage for using the Holian-Evans spline is that the reach of the ModelIV potential is shorter (\(\approx 5.19\) A) than ModelI's (\(\approx 7.32\) A). shows a comparison of the two vdW potentials and shows the corresponding forces. One can see that the ModelIV \(V_{vdW}\) term is still smooth in the force.
As with the spline points in the \(V_{vdW}\) term, the cutoff function in the bond-order function and Morse potential is also replaced so that it is second-order continuous. Second derivative continuity helps energy conservation in the MD simulations (and allows longer time steps to be taken).
Another difference between ModelIV and ModelI is the depth of the metastable covalent wells. With the addition of the repulsive core, ModelIV fails to detonate with a well as deep as the default value for ModelI. The metastable well is raised by 1 eV, such that \(D_{e}^{\text{AB}}=1.0\) eV, \(D_{e}^{\text{A\scriptsize{VBB}}}=5.0\) eV, and therefore, \(Q=4.0\) eV. Comparing the potential energy surface for the linear, symmetric, metastable configuration for ModelI (see Fig. 4) to the one for ModelIV (see Fig. 5), we notice that the depression near (1.2 \(r_{e}\), 1.2 \(r_{e}\)) diminishes somewhat for ModelIV and that the barrier to be overcome for one particle to displace another via an end-on attack is significantly increased.
All of these changes have notable effects on the EOS and detonation properties of the model HE. The final difference is that the masses of the particles are changed so that they are different from each other. The changes to ModelI to form ModelIV may be summarized as follows:
* A bond-order coefficient is applied to the repulsive term in the Morse part of the interaction potential.
* The inner spline of the \(V_{vdW}\) term is replaced with a repulsive core, which comprises a plateau connected to
Total bond energy in a 1D, three-particle, A-B-A, ModelII interaction in which the distance between the B and one of the A’s is fixed at the bond’s equilibrium distance vs. the distance between B and the remaining A (short dash). \(V_{vdW}\) for any two atoms (solid). Total bonding energy less the \(V_{vdW}\) contributions (long dash). All curves are subtracted by their corresponding values at the position of the local minimum of the first curve for comparison purposes. This shows that the section of positive slope in the REBO \(V_{vdW}\) inner spline can trap particles. Without it, there is no trap.
(Color online) Forces that correspond to the potentials in
(Color online) van der Waals (\(V_{vdW}\)) potential vs. the interatomic distance (\(r_{ij}\)). Spline points are indicated by changes in color–dashing. ModelI \(V_{vdW}\) is made of a third-order polynomial (cyan–solid) connected to a Lennard–Jones (LJ) form (mauve–short dash), which is cropped at a finite distance at which the slope and value of \(V_{vdW}\) are nearly zero. ModelIV \(V_{vdW}\) starts as a constant (red–long dash) and is connected with a fifth-order polynomial (green–long short dashes) so that the spline points are second-derivative smooth, to a LJ form (blue–short short long dashes) at zero, below which it coincides with the ModelI LJ until the inflection point, at which it is connected to a curve (black–long long short dashes) that is brought smoothly to zero within a finite distance, a Holian–Evans spline.
* The order of all of the splines is increased such that all but one spline point (the outermost cutoff) have smooth second derivatives.
* The masses of the atom types are changed so that they are no longer equal.
* The depth of the binding energy for the metastable bonds is made shallower.
## III Methods
Investigating the two REBO potentials of which this paper is a study, ModelI and ModelIV, we use SPaSM 3.0 to carry out the MD simulations, of which there are two main types, microcanonical ensembles (NVE) and non-equilibrium MD (NEMD). The NVE simulations are used for testing the thermochemical properties of a model and NEMD are used to investigate its detonation behavior. Both utilize the leapfrog Verlet method to advance the atoms' positions and momenta.
We conduct two sub-types of NVE simulations. The first is meant to find the equilibrium Hugoniots and are similar to those described by others. The new model requires a shorter time step \(\delta t\approx 0.16\) fs than does that of Brenner _et al._, \(\delta t\approx 0.25\) fs because of the relatively high curvature of the fifth-order spline in ModelIV's \(V_{vdW}\) term. Its zero-pressure configuration is also slightly different. The length of the sides of the rectangular unit lattice cell are \(l_{x}=6.19122\) A and \(l_{z}=4.20538\) A. The cell contains two dimers in a herringbone configuration. The angles that the dimers make with the horizontal are \(\pm 27.7109^{\circ}\). The atoms are placed 0.59976 A from the centers of their respective dimers, which are positioned at (1/4 \(l_{z}\), 1/4 \(l_{x}\)) and (3/4 \(l_{z}\), 3/4 \(l_{x}\)) from the lower left corner of the cell. These values are determined by isothermal-isobaric Monte Carlo simulations.
The second type of NVE simulation, the cookoff, is used to determine the reaction rate. The initial conditions are the same as with the first type of NVE except that the constituents are 100\(\times\)100 cells\({}^{2}\) of AB not 25\(\times\)25 of AA and BB. The time step is also drastically reduced so that good statistics can be found for measurements taken during rapid reactions. The time step varies among cookoff simulations, \(6\times 10^{-4}\) fs \(<\delta t<0.02\) fs.
One of the purposes of creating a new potential is to expand the reaction zone by lowering the reaction cross section. This requires larger NEMD simulations to be performed. As Rice _et al._ have done for computational efficiency, we try to reduce the amount of pre-shocked material modeled by tacking initial state material onto the end of the sample as the shock front approaches. To reduce surface effects, we introduce at the surface a one-cell-thick layer of two new particles, C and D, that have all of the properties of A and B, respectively, except that they start frozen and their velocities are not updated. When the shock front nears the frozen layer, the frozen layer is converted to A and B and is thermalized along with the new material. Another frozen layer is tacked onto the end. The size of an NEMD simulation is discretely dependent on time. The largest simulation contained over three million atoms.
Another effect of decreasing the reaction cross section is that the incubation time, before which detonation occurs, extends. The sensitivity to impact is thus lessened. To overcome this, we initially overdrive the simulation with a piston that moves into the material at a velocity \(u_{pstn}\approx 4.9\) km/s. After a detonation front has seemingly been established, it is smoothly backed off by following a sine-shaped deceleration over 200 time steps to a desired velocity. The time step is the same as with the first type of NVE simulation, \(\delta t\approx 0.16\) fs.
Potential energy surface for the linear, symmetric, metastable configuration of ModelI atoms. The contours are spaced every 0.1 eV. \(r_{\text{A-B}}\) is the distance between and A and B atom. \(r_{\text{B-A}}\) is the the distance between the same A atom and a different B atom. \(r_{e}\) is the equilibrium distance for the metastable covalent interaction.
Potential energy surface for the linear, symmetric, metastable configuration of ModelIV atoms. The contours are spaced every 0.1 eV.
## IV Comparison of Thermodynamic Properties
We start the comparison of the properties of ModelIV by analyzing how the ModelIV material behaves when compressed by the passage of an under-driven detonation front. From the snapshot of an NEMD simulation using ModelIV (see Fig. 6), one can see on the left half that many unreacted dimers are lining up horizontally. This is representative of uniaxial compression. Farther back it has melted, and farther back still reacted products are evident. On the right half, the dimers do not seem to line up before reacting. It is made evident in the companion article that this is an effect of detonation instability and the propagation of transverse waves. When comparing this to a snapshot of detonation in ModelII (see in the paper of Heim _et al_.), one notices that, even at the right side of the sample, ModelIV seems to hold its molecular identity better than ModelII during compression by a detonation front that is not overdriven. It seemingly displays a greater resilience to dissociation. Both the right half of the ModelIV snapshot and the ModelI snapshot lack a significant induction zone.
To better compare the widths of the detonation fronts of the two models, profiles of the average \(z\)-component of particle velocity from a critically supported detonation of ModelIV are plotted. indicates that the reaction zone is about 700 A wide since that is the position at which the curves settle down to the constant \(u_{pj}\). With the same method, the width of the reaction zone for ModelI was determined to be about 300 A.
Although ModelIV's reaction zone is still small compared to that of real explosives, it can be widened. To widen the reaction zone farther, one can create a reaction that requires more steps. One also can have the reaction increase the number of moles of material. The resulting expansion in volume would be another driver for the shock wave. Yet another effect of this would be to introduce an entropic penalty to back-reaction and make the model more closely approximate the ZND assumption of irreversibility.
To find the critical piston velocity in the preceding analysis, the CJ state for ModelIV is found by conducting NVE simulations and seeking the values of \(E\) and \(v\) that satisfy the Hugoniot jump conditions. For Model \(v_{j}/v_{0}\approx 0.57\) and \(u_{sj}\approx 9.7\) km/s. From the results (see Table 2), one should notice that for ModelIV the CJ state is less compressed than for ModelI and that \(u_{sj}\) is faster. Compared to conventional explosives, which typically have specific volumes \(v_{j}/v_{0}\approx 0.75\) and shock velocities \(u_{sj}\approx 6\) km/s (for example PETN), ModelIV detonation fronts are over twice as fast and its CJ state is slightly less compressed. Note, however, that it has been shown for these REBO potentials that 3D simulations have lower, and therror more realistic, velocities than do 2D ones. Heim _et al_. show how closely the conditions of the final state in a detonation of a ModelI material match those of the CJ conditions. For contrast in Fig. 8,
\begin{table}
\begin{tabular}{|l|l|} \hline \(v_{j}/v_{0}\) & 0.789851 \\ \hline \(u_{sj}\) & 13.87868 \\ \hline \(u_{pj}\) & 2.91658 \\ \hline \(\langle k_{B}T_{j}\rangle\) (eV) & 0.9185 \\ \hline \(\langle U_{j}\rangle\) (eV) & -0.8631 \\ \hline \(\langle E_{j}\rangle\) (eV) & 0.0554 \\ \hline \(P_{j}\) (eV/Å\({}^{2}\)) & 0.83846 \\ \hline \(\lambda_{j}\) & 0.8548516 \\ \hline \end{tabular}
\end{table}
Table 2: Determined thermodynamic values at CJ for ModelIV, where \(v\) is the specific volume, \(u_{s}\) is the shock velocity, \(u_{p}\) is the particle velocity, \(T\) is the temperature, \(U\) is the potential energy, \(E\) is the internal energy, \(P\) is the pressure, and \(\lambda\) is the degree of reaction. Subscript \(j\) indicates the CJ value, Subscript \(0\) indicates the value at the initial state. Averages are per particle.
Snapshot of a section of a shock front for an under-supported detonation using the ModelIV potential. Shock is propagating upward. Particles are shaded by atom type.
Overlap of profiles of the \(z\)-component of the particle velocity for a critically supported NEMD simulation using ModelIV parameterized such that \(D_{\text{s}}^{\text{sk}}=1.0\) eV and \(Q=4.0\) eV. A constant line is drawn at the CJ value of the particle velocity (\(u_{pj}\)), at which the driving piston is moving. Using a critically supported NEMD as opposed to an unsupported one drastically reduces the transient time from initiation to steady state.
In ZND theory this would require that the final state for ModelIV is either at a strong point or a weak point. The strong point is unstable and will only be a solution to the conservation equations if the detonation is overdriven. In the strict ZND theory there is no path to the weak point. ZND is, however, based on assumptions not realized by ModelIV. ModelIV has a reversible reaction, the pathway of which includes an endothermic dissociation. Endothermic steps can be responsible for weak point final states. Perhaps, the repulsive core introduced in ModelIV, by sheltering the dimers, prevents the reduction of the activation energy of the dissociative step and, hence, increases that step's reaction time, thus making it a stronger contributer to the overall reaction rate and causing the final state to be at a weak point. As is shown in the accompanying paper and can be inferred from Fig. 6, however, a 1D theory is not totally appropriate for this model.
In order to confirm that the modifications to the REBO model in the ModelIV potential do, indeed, maintain a dimerized state, the radial distribution function (RDF) at the CJ state is plotted. In the CJ state of ModelIV is compared to that of ModelI. From the RDF for ModelIV, one can see that, after the peak at \(r=r_{e}\), the curve drops to nearly zero while the analogous region in the RDF of ModelII is closer to unity. This is indicative of a molecular state for ModelIV and a dissociative state for ModelI. From Fig. 10, it is clear that the snapshot of the ModelIV CJ state supports the finding of the corresponding RDF. The snapshot for ModelI shows many more clusters and dissociated atoms.
For these images, a bond is defined for each potential. Two particles, \(i\) and \(j\), are considered bonded if \(E_{b,ij}+KE_{\parallel,ij}<E_{c}\) and \(r_{ij}<r_{c}\), where \(E_{b,ij}\) is as in Eq. 1 for ModelIV and Eq. 2 for ModelI except that the outer sum is over only \(i\) and \(j\). \(KE_{\parallel,ij}\equiv\frac{1}{2}\mu\left|(\vec{v}_{i}-\vec{v}_{j})\cdot\vec {r}_{ij}/r_{ij}\right|^{2}\). \(\mu\) is the reduced mass of \(i\) and \(j\). For ModelIV \(E_{c}=ec\), and \(r_{c}=\gamma_{2}\) (see Table 1). For ModelI \(E_{c}\) is the peak of the inner spline on the \(V_{vdW}\) term and \(r_{c}\) is the location of that peak (see Fig. 2). It can be shown that for ModelI there exists no minimum above and within these cutoff values for all values of \(\overline{B}_{ij}\). To be considered part of a dimer, a particle must be bonded to only one other particle that is bonded to no other. If the particle types are the same, it is a product dimer; if different, a reactant dimer. To be part of a cluster, a particle must be bonded to either multiple particles or to one other that is bonded to multiple particles. To be dissociated, a particle must be bonded to no other.
With a bond defined, cookoffs that will reveal the reaction rate and the activation energy (\(E_{a}\)) for each model can be simulated. This is done for two reasons, to compare ModelIV to continuum models and to show another fortuitous difference between ModelI and ModelIV.
Hugoniots for ModelI (circles) and ModelIV (squares). Solid lines are a guide to the eye for ModelI and a fit for ModelIV. Rayleigh lines are determined by the initial conditions and slopes. Dotted lines represent Rayleigh lines the slope of which are determined by the CJ value of the detonation velocity. Slopes of dash-dotted lines are determined by the average velocity of the shock waves in unsupported detonations. Boxes are magnifications.
Snapshots of the CJ state for ModelIV (upper) and ModelII (lower). Particles are marked by bond type. Gray atoms are in unreacted dimers; striped are unbonded, dissociated atoms; black are clustered atoms; and white are reacted dimers.
Radial distribution functions for the CJ states of ModelIV and of ModelI.
As a simulation is started, thermal energy is partitioned among the different modes. After this has occurred and the reaction has progressed somewhat, the reaction rate is sampled over a short time and the temporal average of the rate and its standard error are recorded for the current simulation. In the data are plotted for a series of simulations using the ModelIV potential.
\[\dot{\lambda}=(1-\lambda)A\exp\left[\frac{-E_{a}}{k_{B}T}\right], \tag{3}\]
\(\dot{\lambda}\) is the reaction rate and \(A\) is the frequency factor.
When contrasting the two figures, one notices that, over a similar range of volumes, ModelIV maintains a better fit to the Arrhenius form than ModelI and that the calculation of \(E_{a}\) remains relatively constant when compared to the \(E_{a}\) of ModelI. A probable contribution to the former observation is the manner in which a reaction is defined. At compressions typical in detonations, ModelI has a larger number of atoms in clusters and free states, which are not counted as reacted. Only stable dimers are. One may consider a cluster of two A atoms tightly bonded together with a third, B-type particle loosely bonded to one of them as a state more reacted than one in which A is tightly bonded to B and loosely bonded to the other A. However, by our definition of reaction, all particles in a cluster are considered unreacted and only contribute to the denominator of the calculation of \(\lambda\).
\(E_{a}\) for ModelI decreases with decreasing \(v\). This is because at higher compressions an atom can have more neighbors within the bonding distance than with ModelIV. The bond-order coefficient then lowers the attraction between it and its neighbors, making it easier to break apart any existing bonds. For ModelIV, the repulsive core keeps neighbors away from dimers so that the bond-order coefficient and, thus, the dissociation energy remain unaffected by the same level of compression that would otherwise cause ModelI's dissociation energy to be reduced. For ModelIV there is a higher tendency than for ModelI for the compression to use up the space between dimers than that between the dimers' constituents. This is because for the former case a third particle would have to climb the repulsive core that was added to the \(V_{vdW}\) term before it got within the range of the cutoff function \(f_{c}\), where it can diminish the bond-order coefficient \(\overline{B}_{ij}\). In the latter case a third particle would only need to overcome the barrier of ModelI's inner spline (see Fig. 2), which is small compared to the temperature for most situations considered in reaction and detonation. For ModelIV the activation energy goes up slightly with decreasing volume. The exact reason for this is not known at present, but simulations with ReaxFF on the high explosive RDX also show this trend. \(E_{a}\) for ModelIV is roughly 1 eV greater than for ModelI. This is commensurate with the height of the repulsive core.
Even though we show that ModelIV fits an Arrhenius reaction rate in NVE cookoff simulations, we would like to know that this is the type of reaction that takes place in the reaction zone. We can use that reaction rate along with the von Neumann spike (vNs) state to predict a \(\lambda\) profile. Since we do not have the full EOS throughout the reaction zone, we make some assumptions that will tend to shorten the width of the profile. Let's assume that \(u_{p}\) increases linearly and that \(v\) and \(P\) are constant such that \(T(\lambda)\). We conduct a series of shock tube simulations of a substance similar to ModelIV except that \(Q\) is set to zero. We find the value of piston velocity that produces the constant zone the state of which falls on the theoretical \(\mathcal{R}\) from This is the vNs state. The value of the specific volume at the vNs state is \(v_{n}\approx 0.558\,v_{0}\) and temperature \(T_{n}\approx 0.25/k_{B}\). From Fig.11, we estimate
Reaction rate of ModelIV for various internal energies and volumes vs. temperature. Lines are fits to the Arrhenius form for constant volume. Their slopes are the negative of the activation energy. \(t_{u}\) is a unit of time and equals 10.180505 fs.
Reaction rate of ModelI for various internal energies and volumes vs. temperature. Lines are fits to the Arrhenius form for constant volume. Slope is the negative of the activation energy. \(t_{u}\) is a unit of time.
3) at the vNs state to be \(A_{n}\approx 3.6\) and \(E_{an}\approx 2.0\).
\[T(\lambda)=T_{n}+(T_{j}-T_{n})\frac{\lambda}{\lambda_{j}} \tag{4}\]
3 and using the change of variable
\[\frac{d\lambda}{dz}=\frac{d\lambda}{dt}\frac{dt}{dz}=\dot{\lambda}\left(u_{sj} -\frac{u_{pj}}{2}\right)^{-1}, \tag{5}\]
The variables with the subscript \(j\) are the CJ values listed in Table 2.
The theoretical \(\lambda\) profile reaches \(\lambda_{j}\) at \(z\approx-70\) A, much closer to the front than the \(-700\) A that we measure for the CJ value of the particle velocity in This does not prove that the reaction rate is Arrhenius in the reaction zone, but it does show that the order of magnitude of the measured reaction zone width is significant compared to the lower bound of the theoretical width. Had it been otherwise, this would have shown the reaction rate to not be Arrhenius and thus a different form, for example, a flame front.
Replacing the assumptions in the analysis with more accurate ones, for example, variable \(v\), would tend to lengthen the theoretical width because the expansion of the material would tend to cool it and slow the reaction rate. In lieu of the full EOS, we could have used data from an NEMD to fill in the state within the reaction zone. However, we would only be testing internal consistency. This comparison also suffers from an assumption of steady-state one-dimensionality. We have seen that the NEMD simulations are 2D. As we show in the companion paper, transverse waves traverse the reaction zone. This causes the profile to be time and \(x\) dependent. The EOS throughout the reaction zone is currently being calculated for normal mode stability analysis, which assumes an Arrhenius reaction rate.
## V Conclusion
In this paper it was shown that the changes to the Modell version of the REBO potential have reduced the amount of clustered and dissociated atoms at the CJ state (by RDF) if not throughout the reaction zone (by a snapshot of an NEMD simulation). The CJ state of ModelIV is at a less compressed state and ModelIV's reaction zone is wider than Modell's is. Modell's unsupported detonation velocity is much better predicted by CJ theory, where ModelIV's propagates significantly faster. In the companion paper we investigate ModelIV's relative disparity more closely. The cookoff simulations reveal that ModellIV fits an Arrhenius reaction rate through a wider domain of compressions than does Modell and it has a higher and less density-dependent activation energy.
It was put forth that this higher \(E_{a}\) causes the final state to be at a weak point since it might make more significant in the reaction rate the endothermic step of dissociation. Given, however, the 2D shape of the front in Fig. 6, such a 1D explanation is probably insufficient. Higher \(E_{a}\)s should increase the likelihood of 1 and 2D instabilities in detonation waves. The current value of \(\approx 2\) eV is consistent with the presumed activation energies for many conventional HEs.
The goal of the current research is the improvement of REBO, which we believe we have achieved, mainly, by introducing increased atomic repulsion, which reduces the reaction cross section. It was shown that ModelIV's reaction zone is wider than with ModelI and that the CJ state is a molecular one. REBO's insensitivity to initiation is increased, and a thicker induction zone is introduced. The reaction rate now behaves in a more Arrhenius manner, and the product EOS behaves more like a polytropic gas as indicated by the hyperbolic shape of its equilibrium \(\mathcal{H}\). Some measurements, for example, the adiabatic \(\gamma\)--as is shown in the companion paper--and the amount of compression at CJ, are more commensurate with conventional explosives. However, some are not too close, for example, the reaction temperature and shock velocity. 3D simulations are needed to investigate how the changes truly compare to real experiments.
|
10.48550/arXiv.0806.4141
|
An Interaction Potential for Atomic Simulations of Conventional High Explosives
|
Andrew J. Heim, Niels Gronbech-Jensen, Edward M. Kober, Jerome J. Erpenbeck, Timothy C. Germann
| 1,912
|
10.48550_arXiv.1304.2464
|
## I Introduction
This and the subsequent paper (termed 'paper II') describe the results of an ab initio quantum dynamical study of the absorption spectrum and the non-adiabatic dissociation mechanisms of carbon dioxide photoexcited with the ultraviolet (UV) light between 120 nm and 160 nm. A brief account of this work has already been published. Paper II gives the motivation behind the study and discusses its main result -- the quantum mechanical absorption spectrum and its interpretation in terms of wave functions of metastable resonance states. The present paper sets the stage for paper II and describes the ab initio calculationsand the topography of the potential energy surfaces (PESs) involved in photodissociation. This information, which is often stashed away in supplementary online sections, is a central and, indeed, an indispensable ingredient of a reliable dynamics calculation. The construction of ab initio PESs and their diabatization -- which, without much exaggeration, amounts to learning the topography of PESs and their intersections by heart -- is often more challenging than the subsequent quantum dynamical calculation.
Photoabsorption from the ground electronic state \(\tilde{X}^{1}\Sigma_{g}^{+}\) of linear CO\({}_{2}\) at wavelengths 120 nm -- 160 nm is due to the first five excited singlet valence states \(1^{1}\Sigma_{u}^{-}\), \(1^{1}\Pi_{g}\), and \(1^{1}\Delta_{u}\). In the \(C_{s}\) group notation, appropriate for bent molecule, these states are \(2,3^{1}\!A^{\prime}\) and \(1,2,3^{1}\!A^{\prime\prime}\). UV light excites CO\({}_{2}\) into the region of multiple electronic degeneracies, nuclear motion through which induces strong non-adiabatic couplings between electronic states. These couplings directly affect the observed absorption spectrum of the valence states and control distributions of photofragments over the final states. Their indirect influence apparently extends to shorter wavelengths where Rydberg transitions dominate: The combined experimental and theoretical analysis indicates that the manifold of coupled valence states acts as a'sink' for the optically bright Rydberg states and affects their dissociation lifetimes.
Electronic degeneracies in the Franck-Condon (FC) region, which have been the focus of several studies in the past, are of two types. Glancing intersections occur in the orbitally doubly degenerate \(1^{1}\Pi_{g}\) and \(1^{1}\Delta_{u}\) states which upon bending split into \(A^{\prime}\) and \(A^{\prime\prime}\) components. Conical intersections (CIs) arise at the accidental \(1^{1}\Pi_{g}/1^{1}\Delta_{u}\) crossing inside \(A^{\prime}\) and \(A^{\prime\prime}\) symmetry blocks. In fact, CIs between valence states are ubiquitous and found far outside the near-linear FC region, at strongly bent geometries. Together with local minima and saddles, these crossings are the principal features shaping the topography of the singlet valence states.
The outline of the paper is the following. Section II sketches the technical details of ab initio calculations. Next, the constructed adiabatic PESs are presented for the ground (Sect. III.1) and the excited (Sect. III.2) electronic states. Their intersections are discussed for near linear (Sect. III.3) and bent geometries (Sect. III.4), and put into perspective by a review of a network of closely spaced valence and Rydberg states (Sect. III.5).
If the solution of the Schrodinger equation for nuclei is out of reach, the description of adiabatic surfaces in Sect. III would be the last step in theoretical ab initio analysis. If, on the other hand, one intends to treat nuclear dynamics quantum mechanically, ab initio PESsfeaturing CIs have to be diabatized. This is especially desirable if -- as in paper II -- a discrete grid is used to represent the nuclear Hamiltonian, because the diabatized potential matrix is free from either divergent off-diagonal couplings or non-differentiable potential cusps. While general schemes for constructing approximate adiabatic-to-diabatic transformations are established (see, for instance, Ref.), their application to the valence states of CO\({}_{2}\) is complicated by the number of CIs to be simultaneously treated. Simplifications are called for, as described in Sect. IV, in which the diabatic representation is constructed separately for bent CIs (Sect. IV.1) and linear CIs (Sect. IV.2). Section V concludes.
## II Electronic Structure Calculations
All ab initio calculations are carried out with the MOLPRO package. The Gaussian atomic basis sets used in this work are due to Dunning. Previous studies indicate that diffuse functions should be added to the basis sets on oxygen and carbon atoms in order to account for the mixed valence-Rydberg character of the \(\Pi\) state. A series of tests was conducted, in which \(s\), \(p\), \(d\) etc. basis functions of triple and quadrupole zeta quality were selectively augmented by one or two diffuse functions. The pre-computed doubly augmented correlation consistent polarized valence quadrupole zeta (d-aug-cc-pVQZ) basis set, as implemented in MOLPRO, was found computationally most stable and selected for calculations of global PESs.
Three-dimensional (3D) PESs of states \(1,2,3^{1}\!A^{\prime}\) and \(1,2,3^{1}\!A^{\prime\prime}\) are calculated at the internally-contracted multireference configuration interaction singles and doubles (MRD-CI) level, based on state-averaged full-valence complete active space self-consistent field (CASSCF) calculations with 16 electrons in 12 active orbitals and 6 electrons in three fully optimized closed-shell inner orbitals. The electronic configuration of the ground state \(\tilde{X}^{1}\Sigma_{g}^{+}\) is ([core]\(2\sigma_{u}^{2}3\sigma_{g}^{2}4\sigma_{g}^{2}3\sigma_{u}^{2}1\pi_{u}^{4}1\pi_{g }^{4}\)). The dominant electronic excitations, leading to the lowest excited states, include \(1\pi_{g}^{4}\to 2p2\pi_{u}^{1}\) (giving states \(1^{1}\Delta_{u}\) and \(1^{1}\Sigma_{u}^{-}\)) and \(1\pi_{g}^{4}\to 5s5\sigma_{g}^{1}\) (giving state \(1^{1}\Pi_{g}\)). Active orbitals in CASSCF comprised \(2\sigma_{u}-4\sigma_{u}\), \(3\sigma_{g}-5\sigma_{g}\),\(1\pi_{u}-2\pi_{u}\), and \(1\pi_{g}\). In \(C_{s}\) symmetry, used in the calculations, these are \(4a^{\prime}-12a^{\prime}\) and \(1a^{\prime\prime}-3a^{\prime\prime}\). In the MRD-CI step all 16 valence electrons were correlated. The maximum numbers of open shells allowed in the MRD-CI calculations were 8 in the reference space and 12 in the internal space. This lead to 38159928 contracted configurations. The Davidson correction was applied in order to account for higher-level excitations and size-extensivity.
Adiabatic energies are calculated on a 3D grid of the two C-O bond lengths \(R_{1,2}\) and the OCO bond angle \(\alpha_{\mathrm{OCO}}\): \(R_{2}\in[1.9\,a_{0},3.0\,a_{0}]\) (step size equals \(0.1\,a_{0}\)), \(R_{1}\in[R_{2},6.2\,a_{0}]\) (step size varies between \(0.1\,a_{0}\) and \(0.4\,a_{0}\)), \(\alpha_{\mathrm{OCO}}\in[70^{\circ},179.9^{\circ}]\) (step size varies between \(2^{\circ}\) and \(10^{\circ}\)). Additionally, many cuts in the \((R_{1},R_{2})\) plane are computed for angles \(\alpha_{\mathrm{OCO}}\) between \(60^{\circ}\) and \(0^{\circ}\) within the continuing effort to construct a balanced description of both CO + O and C + O\({}_{2}\) arrangement channels. At present, the grid comprises 4800 symmetry distinguishable points. The resulting energies were scanned in one and two dimensions for obvious errors. The list of corrected adiabatic energies was subsequently interpolated using 3D cubic splines and also used for constructing the quasi-diabatic representation. Missing energies for \(\alpha_{\mathrm{OCO}}<70^{\circ}\) in the dissociation channels were obtained from those for \(\alpha_{\mathrm{OCO}}>70^{\circ}\) using trigonometric extrapolation.
Absolute intensity calculations of paper II require transition dipole moments (TDMs) with the ground state \(\tilde{X}\). Components \((\mu_{x},\mu_{y},\mu_{z})\) of the TDM vector are calculated, for each electronic state, on a 3D grid \(R_{1,2}=[1.9\,a_{0},2.4\,a_{0}]\) and \(\alpha=[165^{\circ}-179^{\circ}]\) covering the spot over which the vibrational ground state in \(\tilde{X}\) is delocalized. The molecular axes in these calculations are chosen such that \(x^{\prime}\) is orthogonal to the molecular plane, \(z^{\prime}\) runs along one of the CO bonds and \(y^{\prime}\perp z^{\prime}\). For \(A^{\prime}\) states, the in-plane components are generally non-zero, while for \(A^{\prime\prime}\) states, it is the \(\mu_{x^{\prime}}\) component which carries the transition.
## III Properties of the valence PESs and their crossings
### Ground electronic state
The \(\sim 7.5\,\mathrm{eV}\) deep adiabatic ground state PES supports three structural isomers: The familiar linear OCO molecule is the global equilibrium, while the carbene-like bent OCO and the linear COO are the two local ones. Table 1 summarizes the characteristic features of the \(\tilde{X}^{1}\Sigma_{g}^{+}\) state at the three equilibria and compares them with the previous ab initio studies and with the available experimental data.
The vicinity of the global minimum is of capital importance for the environmental chemistry. The calculated equilibrium CO bond distance in linear OCO, \(R_{e}=2.1991\,a_{0}\), agrees well with the experimental value of \(2.1960\,a_{0}\). The accuracy of the vibrational zero point energy (ZPE) and the vibrational transitions frequencies is assessed in Table 2 which compares energies of the low lying vibrationally excited states in rotating CO\({}_{2}\) (the total angular momentum \(N_{\text{CO}2}\geq 0\)) with experiment and with recent electronic coupled cluster/vibrational configuration interaction calculations. Each eigenstate \((v_{s},v_{b}^{l},v_{a})\) is labeled using the quantum numbers of the symmetric stretch \(v_{s}\), the bend \(v_{b}^{l}\) (with \(l\) indicating the vibrational angular momentum, \(N_{\text{CO}2}\geq l\)), and the antisymmetric stretch \(v_{a}\). The calculated fundamental frequencies of the infrared active bend (\(\omega_{b}=668.6\,\text{cm}^{-1}\)) and antisymmetric stretch (\(\omega_{a}=2350.6\,\text{cm}^{-1}\)) are accurate to within \(1.5\,\text{cm}^{-1}\). The zeroth order symmetric stretch frequency\(\omega_{s}^{0}\approx 1333\,\text{cm}^{-1}\) is about twice as large as the bending frequency \(\omega_{b}\), and the two modes are involved in the accidental Fermi resonance. As a result, the vibrational spectrum is organized in polyads with the polyad quantum number \(P=2v_{s}+v_{b}\); states with \(P=2\), 3 and 4 are given in Table 2. In the original version of the PES, called 'PES1' in Table 2, the energies of states \((1,0^{0},0)\) and \((0,2^{0},0)\), belonging to the lowest polyad \(P=2\), are underestimated by \(20\,\text{cm}^{-1}\) and the difference with the observed energies grows rapidly with \(P\).
\[R_{+} \rightarrow \sqrt{2}R_{1e}+(R_{+}-\sqrt{2}R_{1e})\cdot 1.023\] \[\alpha_{\text{OCO}} \rightarrow 180^{\circ}+(\alpha_{\text{OCO}}-180^{\circ})\cdot 1.0035\,.\]
The vibrational energies in the scaled 'PES2' agree with their experimental counterparts to within \(7\,\text{cm}^{-1}\) and for the most states below \(3000\,\text{cm}^{-1}\) the accuracy is better than \(3\,\text{cm}^{-1}\). The results outperform even the highly accurate calculations of Refs. shown in the second column of Table 2, making 'PES2' one of the best available ab initio potentials of the \(\tilde{X}^{1}\Sigma_{g}^{+}\) state. Since the coordinate dependent dipole moment \(\mu_{\tilde{X}}\) has also been calculated, the ab initio intensities of the infrared rovibrational transitions can be directly evaluated.
The other two isomers in Table 1 have never been detected in the gas phase, and the only reference data stem from the previous ab initio studies. For the bent OCO, discovered by Xantheas and Ruedenberg, the present calculations confirm the \(C_{2v}\) symmetric equilibrium with the CO bond lengths of \(2.51\,a_{0}\) and the OCO bond angle of \(73.2^{\circ}\). This minimum is located \(6.03\,\text{eV}\) above the global one, again in good agreement with the previous findings. The fundamental excitations in the OCO well, calculated for \(N_{\text{CO}2}=0\), are \(\omega_{a}=680\,\text{cm}^{-1}\), \(\omega_{b}=720\,\text{cm}^{-1}\), and \(\omega_{s}=1550\,\text{cm}^{-1}\). For the linear COO, the calculated CO and OObond lengths are identical to the ones given in Ref.; both are elongated compared to free diatoms (\(2.2\,a_{0}\) vs. \(2.14\,a_{0}\) for CO and \(2.45\,a_{0}\) vs \(2.28\,a_{0}\) for OO). The calculations place the COO minimum at \(7.35\,\)eV, about \(0.1\,\)eV below the lowest dissociation threshold.
The ground electronic state correlates adiabatically with two dissociation channels,
\[\mathrm{CO}_{2}+h\omega(E_{\mathrm{ph.}}\geq\ 7.41\,\mathrm{eV}) \rightarrow \mathrm{O}(^{1}\!D)+\mathrm{CO}(X^{1}\Sigma^{+}) \tag{1}\] \[\mathrm{CO}_{2}+h\omega(E_{\mathrm{ph.}}\geq\ 11.52\,\mathrm{eV}) \rightarrow \mathrm{C}(^{3}\!P)+\mathrm{O}_{2}(X^{3}\Sigma^{-})\,, \tag{2}\]
In channel, the calculated \(D_{0}\) is \(0.13\,\)eV less than the experimental value. The deviation might reflect a large basis set superposition error introduced by the diffuse functions and as such is the downside of the highly accurate vibrational spectrum in Table 2. The error is independent of the arrangement channel, and \(D_{0}\) in channel [closed between \(120\,\)nm and \(160\,\)nm] is equally underestimated. The calculations of Hwang and Mebel, using a noticeably smaller basis set, perfectly agree with the experimental dissociation energy for this channel.
One-dimensional (1D) cuts through the ground state PES are given for several \(\alpha_{\mathrm{OCO}}\) angles in panels (a,c,e) of Figs. 1 and 2. Black solid circles are the raw adiabatic energies. The O + CO limit is reached smoothly and no barrier is detected towards the asymptote for any orientation of the CO diatom. The same is true for the C + O\({}_{2}\) channel, as illustrated in Fig. 2(e); the potential well in Fig. 2(e) is the COO isomer. Angular dependence of the \(\tilde{X}^{1}\!A^{\prime}\) state is shown in panels (a,c,e) of Figs. 3 and 4 for two sets of fixed CO bonds. In Fig. 3, \(R_{1}\) is fixed at the FC value; in Fig. 4, it is fixed close to the equilibrium of the bent OCO. Consequently, although the carbene-type minimum is perceptible in all panels, it is best seen in As CO\({}_{2}\) bends, the adiabatic \(\tilde{X}^{1}\!A^{\prime}\) state (black dots) forms a sharp narrowly avoided crossing with the state \(2^{1}\!A^{\prime}\) around \(\alpha_{\mathrm{OCO}}=100^{\circ}\) (see, for example, Fig. 3). A dynamically meaningful representation for such nearly degenerate pairs is diabatic rather than adiabatic. In fact, the black line in Figs. 1 -- 4 depicts the \(\tilde{X}^{1}\!A^{\prime}\) state locally diabatized at bent geometries as described in Sect. IV.1. This is the reason why the black dots sometimes switch away from the black line and why lines of different colors cross. In this locally diabatic picture, the bent OCO minimum correlates with the state \(2^{1}\!A^{\prime}\) [purple line]. The transition state separating the bent and the linear minima, analyzed by Xantheas and Ruedenberg and Hwang and Mebel is thus the signature of this two-state intersection.
### Overview of the excited electronic states
Potentials of the excited electronic states are shown in Figs. 1 and 2 along one CO bond for \(A^{\prime}\) and \(A^{\prime\prime}\) symmetries. As with the \(\tilde{X}\) state, the solid circles indicate ab initio adiabatic energies. In the CO + O arrangement channel, which is the focus of the present investigation, all calculated states but one converge to the dissociation threshold.
\[\mathrm{CO}_{2}+h\omega(E_{\mathrm{ph.}}\geq\ 11.46\,\mathrm{eV})\to \mathrm{O}(^{3}\!P)+\mathrm{CO}(a^{3}\Pi)\,. \tag{3}\]
In the C + O\({}_{2}\) arrangement channel [Fig. 2(e,f)], three electronic states, the \(\tilde{X}\) state and the two components of the \(\Pi\) state, correlate with channel. Three other states (\(\Sigma^{-}\) and \(\Delta\)) converge to the electronically excited fragments:
\[\mathrm{CO}_{2}+h\omega(E_{\mathrm{ph.}}\geq\ 13.75\,\mathrm{eV})\to \mathrm{C}(^{1}\!D)+\mathrm{O}_{2}(^{1}\Delta)\,. \tag{4}\]
Topographic hallmarks of the excited states can be exemplified using the states \(2,3^{1}\!A^{\prime}\). In the FC region near linearity [Fig. 1(a)], the states \(2^{1}\!A^{\prime}\) and \(3^{1}\!A^{\prime}\) form two sharp avoided crossings near \(R_{1}=2.2\,a_{0}\) and \(R_{1}=2.8\,a_{0}\). These crossings are in fact two CIs between states \(1^{1}\Pi_{g}\) and \(1^{1}\Delta_{u}\). The CIs are not independent: They are connected into a whole line, called a 'CI seam', unusual properties of which are discussed in Sect. III.3. As the molecule bends, the gap between the adiabatic states grows. In the lower state \(2^{1}\!A^{\prime}\), the intersection cone first turns into a broad barrier along the dissociation path [\(\alpha_{\mathrm{OCO}}>170^{\circ}\), Fig. 1(c)]. As \(\alpha_{\mathrm{OCO}}\) decreases further, the local minimum near \(2.3\,a_{0}\) deepens and the dissociation barrier disappears [Fig. 1(e) and Fig. 2(a)]. In linear COO [\(\alpha_{\mathrm{OCO}}=0^{\circ}\), Fig. 2(e)], the sharp avoided crossing between states \(2,3^{1}\!A^{\prime}\) reappears again, this time at \(R_{1}=3.5\,a_{0}\). Similar to \(\tilde{X}\), the state \(2^{1}\!A^{\prime}\) supports a COO intermediate, although the local minimum lies \(0.7\,\mathrm{eV}\) above the C + O\({}_{2}\) threshold. The evolution of the uppermost state \(3^{1}\!A^{\prime}\) with decreasing angle is different, because its topography is very much influenced by a pronounced barrier located outside the FC zone near \(R_{1}=3.6\,a_{0}\) and separating the flat inner region from a steep decline towards the asymptotic limit. This barrier is distinct over a broad angular range and its sharpness suggests a CI with a higher lying state. The nature of this intersection becomes apparent in Sect. III.5 discussing valence/Rydberg crossings.
The \(A^{\prime\prime}\) states are similar in many respects. Near linearity, both symmetries mirror the topography of the orbitally degenerate states \(1^{1}\Pi_{g}\) and \(1^{1}\Delta_{u}\), the states \(1,3^{1}\!A^{\prime\prime}\) are involved in the same CIs in the FC region [Fig. 1(b)], and their behavior in the bent molecule up to \(\alpha_{\mathrm{OCO}}=120^{\circ}\) closely follows that of \(2,3^{1}\!A^{\prime}\). The \(1^{1}\!A^{\prime\prime}\) state stabilizes with decreasing \(\alpha_{\mathrm{OCO}}\), while the \(3^{1}\!A^{\prime\prime}\) state features a pronounced barrier around \(R_{1}=3.6\,a_{0}\) [Fig. 1(d,f) and Fig. 2(b)]. Similarities persist to \(\alpha_{\mathrm{OCO}}=0^{\circ}\) [Fig. 2(f)]: The states \(1,3^{1}\!A^{\prime\prime}\) form a CI at \(R_{1}=3.5\,a_{0}\), and the state \(1^{1}\!A^{\prime\prime}\) supports a local COO minimum. Differences with the \(A^{\prime}\) symmetry are due to the state \(2^{1}\!A^{\prime\prime}\), which in the FC region correlates with \(1^{1}\Sigma_{u}^{-}\) and which has no counterpart among the calculated \(A^{\prime}\) states. This state, which at linearity is accidentally degenerate with \(1^{1}\Delta_{u}\), has a single minimum near \(2.4\,a_{0}\) and approaches the dissociation limit without a barrier. This simple shape is preserved through most cuts in Figs. 1 and 2. The state \(2^{1}\!A^{\prime\prime}\) also supports a COO minimum [Fig. 2(f)].
Potential cuts along bending angle are shown in Figs. 3 and 4. The doubly degenerate \(1^{1}\Pi_{g}\) and \(1^{1}\Delta_{u}\) states split into \(A^{\prime}\) and \(A^{\prime\prime}\) components for \(\alpha_{\mathrm{OCO}}<180^{\circ}\). Evolution of the electronic energies with decreasing \(\alpha_{\mathrm{OCO}}\) was analyzed by Spielfiedel et al. using Walsh rules. Based on this analysis, the adiabatic excited states are commonly classified as bent or linear. The energy of states \(2^{1}A^{\prime}\) and \(1^{1}A^{\prime\prime}\) lowers as the molecule bends, while the energy of states \(3^{1}A^{\prime}\) and \(3^{1}A^{\prime\prime}\) grows [see Fig. 3(a,b)]. The global minima of the 'bent' states lie near \(120^{\circ}\) (\(2^{1}A^{\prime}\)) and \(130^{\circ}\) (\(1^{1}A^{\prime\prime}\)) and are located below the dissociation limit [Fig. 3(a,b) and Fig. 4(a,b)]. Bent equilibrium of the state \(2^{1}A^{\prime}\), which at \(C_{2v}\) geometries becomes \({}^{1}\!B_{2}\), was predicted by Dixon in his analysis of the CO flame emission bands. Finally, the state \(2^{1}A^{\prime\prime}\), which together with \(3^{1}A^{\prime}\) and \(3^{1}A^{\prime\prime}\) is 'linear', is the least anisotropic of all states in a broad vicinity of \(\alpha_{\mathrm{OCO}}\sim 180^{\circ}\).
Properties of various stationary points in the PESs of excited electronic states are given in Table 3. The \(C_{s}\) point group notation is used to label adiabatic states; \(D_{\infty h}\) labels refer to the diabatic states at the FC point. Experimental reference data for the excited electronic states are scarce and fit into the footnote \(a\). For the bent states, the experimental equilibrium CO distances and the bending angle are reproduced within \(0.1\,a_{0}\) and \(3^{\circ}\), respectively; the bending frequencies are accurate to within \(30\,\mathrm{cm}^{-1}\) or better, and the calculated band origins lie within the experimental uncertainties. Results of the previous ab initio studies are also shown in Table 3. Agreement in the equilibrium geometries is excellent, with the exception of the intersection-ridden uppermost states \(3^{1}A^{\prime}\) and \(3^{1}A^{\prime\prime}\), for which \(C_{2v}\)-restricted calculations of Refs. and miss the local \(C_{s}\) minima in the upper CI cones. Vertical excitation energies \(T_{v}\) are close to those of Ref.. Slight differences are not surprising for near degenerate states whose ordering is sensitive to the details of the ab initio set up. Peculiarity of the spectral region \(120\,\)nm -- \(160\,\)nm is that \(T_{v}\) values are poor approximations to the positions of the absorption maxima because all transitions are electronically forbidden and the TDMs at the FC point are strictly zero. Finally, the ab initio dissociation energies \(D_{0}\) in channels,,, and, also given in Table 3, agree with the experimental values within \(\sim 0.15\,\)eV irrespective of the particular arrangement or electronic channel.
### Topography of state intersections in the FC zone
Intersections of electronic states in linear CO\({}_{2}\) follow simple'symmetry rules' which severely restrict the intersection topography. Two types of intersections with regard to their symmetry properties are prominent in Figs. 1(a,b) and 3(a,b):
Renner-Teller (RT) glancing intersections involve the \(A^{\prime}\) and \(A^{\prime\prime}\) components of the \(1^{1}\Pi_{g}\) or the \(1^{1}\Delta_{u}\) state. In the rotating molecule, the \(A^{\prime}/A^{\prime\prime}\) interaction \(\sim\Lambda\Omega/\sin^{2}\alpha_{\rm OCO}\) is proportional to the projections \(\Lambda\) and \(\Omega\) of the electronic (\(\hat{L}\)) and the total angular momentum (\(\hat{J}=\hat{N}_{\rm CO2}+\hat{L}\)) on the molecular axis and diverges for \(\alpha_{\rm OCO}\to 180^{\circ}\). The \(1^{1}\Pi_{g}\leftarrow\tilde{X}^{1}\Sigma_{g}^{+}\) transition is best classified as linear-linear, and the \(1^{1}\Delta_{u}\leftarrow\tilde{X}^{1}\Sigma_{g}^{+}\) transition is linear-bent. The 'degeneracy manifold' for the intersecting states is the whole \((R_{1},R_{2})\) plane defined by the condition \(\alpha_{\rm OCO}=180^{\circ}\).
Two nested CIs involve the \(2,3^{1}\!A^{\prime}\) and \(1,2^{1}\!A^{\prime\prime}\) or \(2,3^{1}\!A^{\prime\prime}\) states and stem from the accidental \({}^{1}\Pi_{g}/^{1}\Delta_{u}\) crossing. Both degeneracies are lifted linearly along the tuning and coupling modes spanning the common branching space. According to Fig. 1(a,b), the tuning mode is \(R_{\rm CO}\), i.e. a combination of the symmetric and antisymmetric stretch (their irreps are \(\sigma_{g,u}^{+}\) in the \(D_{\infty h}\) group); \(\alpha_{\rm OCO}\) breaking the linear symmetry (irrep \(\pi_{u}\)) is the coupling mode. As a consequence of the'symmetry rules', the CIs occur along a line \(F_{\rm CI}(R_{1}^{*},R_{2}^{*})=0\) in the \((R_{1},R_{2})\) plane at \(\alpha_{\rm OCO}=180^{\circ}\): The 'degeneracy manifold' is a 1D seam.
The CI seam, constructed separately for the \(2,3^{1}\!A^{\prime}\) and \(1,3^{1}\!A^{\prime\prime}\) states on a fine \((R_{1},R_{2})\) grid, is depicted in It has two remarkable properties. First, the intersection along the seam is fivefold. Two CIs and two RT intersections imply four degenerate states. The hitherto ignored state \(1^{1}\Sigma_{u}^{-}\) closely follows \(1^{1}\Delta_{u}\): The \(\Delta/\Sigma\) energy gap falls consistently below \(300\,\mathrm{cm}^{-1}\) which is at the limit of the ab initio accuracy. Thus, the \(1^{1}\Pi_{g}/1^{1}\Delta_{u}\) and the \(1^{1}\Delta_{u}/1^{1}\Sigma_{u}^{-}\) pairs cross at the same CO bond distances and the total degeneracy is five. Second, the calculated seam traces out a closed loop. Closed CI seems are rarely encountered, although arguments have been devised to prove their ubiquity. A technical implication of closed or strongly curved seams is that local diabatization schemes fail and global or semiglobal diabatization becomes necessary. The impact of closed seams on photodissociation dynamics has never been systematically investigated. In CO\({}_{2}\), only a small portion of the loop is directly accessible to UV light. Nevertheless, paper II demonstrates that the seam topology deeply affects the observed absorption spectrum.
The curved intersection seam is responsible for a peculiar shape of the potentials in the FC region. The adiabatic PESs of the states \(2,3^{1}\!A^{\prime}\) in linear CO\({}_{2}\) are shown in The lower adiabatic state \(2^{1}\!A^{\prime}\) has a \(C_{2\mathrm{v}}\) minimum at \(R_{1}=R_{2}=2.41\,a_{0}\), and two \(C_{\mathrm{s}}\) symmetric saddles near \(R_{1}\approx 2.8\,a_{0}\) or \(R_{2}\approx 2.8\,a_{0}\), separating the \(C_{2\mathrm{v}}\) minimum from the dissociation asymptotics. Outside the area enclosed by the seam line, the \(2^{1}\!A^{\prime}\) state has the \(\Pi\), inside the \(\Delta\) character. The saddles hide portions of cusp lines, along which the states intersect and which are washed out by the spline interpolation. The topography of the upper adiabatic \(3^{1}\!A^{\prime}\) state is a literal mirror image of the lower state; one finds a \(C_{2\mathrm{v}}\) saddle at \(R_{1}=R_{2}=2.41\,a_{0}\) and two \(C_{\mathrm{s}}\) symmetric minima near \((R_{1},R_{2})=(2.84\,a_{0},2.27\,a_{0})\) and \((2.27\,a_{0},2.84\,a_{0})\). The state character changes from \(\Delta\) outside the seam line to \(\Pi\) inside. The two minima are cusp-like and correspond to the upper cones of the CIs. The two other saddle points in the \(3^{1}\!A^{\prime}\) state, lying outside the FC zone near \((R_{1},R_{2})=(3.6\,a_{0},2.2\,a_{0})\) and \((2.2\,a_{0},3.6\,a_{0})\), result from an avoided crossing with a higher lying state.
Changes of the electronic character of \(A^{\prime\prime}\) states across the intersection can be monitored using matrix elements of the electronic angular momentum \(\hat{L}_{z}\). The matrix elements \(\left|\langle iA^{\prime\prime}|\hat{L}_{z}|jA^{\prime}\rangle\right|\) along the line passing twice through the closed seam are depicted in for \(i=1,2,3\) and \(j=2,3\). The states \(|2^{1}\!A^{\prime}\rangle\) and \(|3^{1}\!A^{\prime}\rangle\) act as 'probes' whose assignment in terms of \(\Pi\) or \(\Delta\) is known. Outside the seam area, most matrix elements are close to integer values of \(0\), \(1\), and \(2\). For \(R_{1}\leq 2.2\,a_{0}\) for example, \(\left|\langle 1A^{\prime\prime}|\hat{L}_{z}|2A^{\prime}\rangle\right|=1\) and \(\left|\langle 2A^{\prime\prime}|\hat{L}_{z}|2A^{\prime}\rangle\right|=0\) so that \(|1A^{\prime\prime}\rangle\) is a \(\Pi\) state, while \(|2A^{\prime\prime}\rangle\) is a \(\Sigma\) state. \(\left|\langle 3A^{\prime\prime}|\hat{L}_{z}|2A^{\prime}\rangle\right|\) vanishes too, but the state \(|3A^{\prime\prime}\rangle\) is a \(\Delta\) state, as confirmed by the matrix element \(\left|\langle 3A^{\prime\prime}|\hat{L}_{z}|3A^{\prime}\rangle\right|=2\). As the CI seam is crossed at \(R_{1}=2.25\,a_{0}\) into the area interior to the seam loop, the three-stateintersection induces violent changes in the electronic labels of the adiabatic states. The state \(|3A^{\prime\prime}\rangle\) becomes a \(\Pi\) state (\(\left|\langle 3A^{\prime\prime}|\hat{L}_{z}|3A^{\prime}\rangle\right|=1\)); the state \(|2A^{\prime\prime}\rangle\) acquires \(\Delta\) character (\(\left|\langle 2A^{\prime\prime}|\hat{L}_{z}|2A^{\prime}\rangle\right|\approx 2\)), while the state \(|1A^{\prime\prime}\rangle\) acquires \(\Sigma\) character (\(\left|\langle 1A^{\prime\prime}|\hat{L}_{z}|2A^{\prime}\rangle\right|\approx 0.5\)), and the non-integer values of the latter two reflect strong mixing of the near degenerate \(\Delta_{u}/\Sigma_{u}^{-}\) pair. The next reshuffling occurs as the seam loop is crossed outwards at \(R_{1}\approx 2.8\,a_{0}\). The state \(|1A^{\prime\prime}\rangle\) becomes \(\Pi\) again, while the matrix elements involving \(|2A^{\prime\prime}\rangle\) and \(|3A^{\prime\prime}\rangle\) vary until \(R_{1}\approx 3.6\,a_{0}\) is reached, where \(\left|\langle 3A^{\prime\prime}|\hat{L}_{z}|3A^{\prime}\rangle\right|\) finally vanishes indicating that \(|3A^{\prime\prime}\rangle\) emerges from the crossing region as a \(\Sigma\) state converging towards the upper threshold, while \(|2A^{\prime\prime}\rangle\) becomes \(\Delta\) state correlating with threshold.
The positions of cusps, minima, and saddles in the PESs of the \(A^{\prime\prime}\) states at linearity are identical to those in \(A^{\prime}\) states. Exception is the PES of the state \(2^{1}\!A^{\prime\prime}\), correlating in the FC region with \(1^{1}\Sigma_{u}^{-}\). Its appearance, illustrated in Fig. 8, is very much simplified by it being tightly linked with the diabatic \(1^{1}\Delta_{u}\) state. The PES has a single \(C_{\rm 2v}\) minimum at \(R_{1}=R_{2}=2.41\,a_{0}\), is free from cusps seen in other adiabatic PESs, and provides a nature's illustration of the shape of the diabatic \(1^{1}\Delta_{u}\) PES.
CIs at linearity can also be recognized in the potentials plotted in the \((R_{1},\alpha_{\rm OCO})\) plane. An example including four electronic states at \(R_{2}=2.2\,a_{0}\) is shown in Although the characteristic features are no longer \(C_{2v}\) symmetric, a maximum in the \(2^{1}\!A^{\prime}\) state at \(\alpha_{\rm OCO}=180^{\circ}\) and \(R_{1}=2.8\,a_{0}\) and two minima in the \(3^{1}\!A^{\prime}\) state at \(\alpha_{\rm OCO}=180^{\circ}\) and \(R_{1}=2.3\,a_{0}\) and \(R_{1}=2.8\,a_{0}\) are recognizable in panels (a) and (b). Close to linearity, the \(A^{\prime\prime}\) states [panels (c) and (d)] maintain the same contour maps as \(A^{\prime}\) states.
### Intersections and avoided crossings in bent CO\({}_{2}\)
As CO\({}_{2}\) bends and all above degeneracies are removed, the primary topographic features (cusps, minima, and saddles) remain clearly visible up to \(\alpha_{\rm OCO}\sim 160^{\circ}\). This is demonstrated in the \((R_{1},\alpha_{\rm OCO})\) plane in and in the \((R_{1},R_{2})\) plane in Further decrease of \(\alpha_{\rm OCO}\) spawns new avoided crossings, a detailed map of which is given in the angular cuts in Figs. 3 and 4. All pairs of calculated states become near degenerate at various geometries. For example, the \(\tilde{X}\) state forms avoided crossings successively with states \(2^{1}\!A^{\prime}\) at \(\alpha_{\rm OCO}=100^{\circ}\) and \(3^{1}\!A^{\prime}\) at \(\alpha_{\rm OCO}=80^{\circ}\). The state \(3^{1}\!A^{\prime\prime}\) approaches closely \(2^{1}\!A^{\prime\prime}\) at \(\alpha_{\rm OCO}\sim 120^{\circ}-130^{\circ}\) and \(1^{1}\!A^{\prime\prime}\) at \(\alpha_{\rm OCO}=90^{\circ}\), and the states \(1^{1}\!A^{\prime\prime}\) and \(2^{1}\!A^{\prime\prime}\) cross between \(70^{\circ}\)and 80\({}^{\circ}\). Solid lines in Figs. 3 and 4 cross because they refer to the states locally diabatized through the 'bent' avoided crossings (see Sect. IV.1). Potential curves of the states 3\({}^{1}\!A^{\prime}\) and 3\({}^{1}\!A^{\prime\prime}\) bear evidence of strong interactions with the next higher states. Especially pronounced in Fig. 3(d,f) and Fig. 3(b,d) are the sharp near-intersections in 3\({}^{1}\!A^{\prime\prime}\) around 155\({}^{\circ}\).
As a result of multiple avoided crossings, local bent equilibria are found in all calculated states. The diabatic origin of a particular local minimum is invariably different from the adiabatic one. Consider the carbene-like bent OCO with \(\alpha_{\rm OCO}\sim 70^{\circ}\) in Fig. 4(a). Purely adiabatically (solid dots), this minimum belongs to the ground electronic state. In the locally diabatic picture, bent OCO belongs the state 2\({}^{1}\!A^{\prime}\) (purple line). An avoided crossing between states 2\({}^{1}\!A^{\prime}\) and 3\({}^{1}\!A^{\prime}\), recognizable in panel (a) around \(\alpha_{\rm OCO}\sim 95^{\circ}\), implies that another diabatization might re-assign the bent OCO minimum to the 3\({}^{1}\!A^{\prime}\) state (brown line). Furthermore, the broad barrier in 3\({}^{1}\!A^{\prime}\) near \(\alpha_{\rm OCO}\approx 140^{\circ}\) indicates another avoided crossing with the next higher \(A^{\prime}\) state which thus would be the true owner of the bent OCO minimum. Similar analysis applies to the bent OCO in the \(A^{\prime\prime}\) states in Fig. 4(b). Depending on the chosen representation, it can be ascribed to the fully adiabatic state 1\({}^{1}\!A^{\prime\prime}\) (dots), to the locally diabatic state 3\({}^{1}\!A^{\prime}\) (brown line), or -- via the sharp near-intersection around 155\({}^{\circ}\) -- to the next higher \(A^{\prime\prime}\) state. Other bent conformations in the excited states result from state interactions, too. One such isomer with the valence angle of 100\({}^{\circ}\) is created via an avoided crossing in 3\({}^{1}\!A^{\prime}\) [Fig. 3(c,e)]. Crossing of the same 3\({}^{1}\!A^{\prime}\) state with the next higher state between 60\({}^{\circ}\) and 70\({}^{\circ}\) leads to another high-energy bent minimum [Fig. 4(a,e)].
### Beyond the first six states: Valence/Rydberg crossings
Many local barriers and minima in the states 3\({}^{1}\!A^{\prime}\) and 3\({}^{1}\!A^{\prime\prime}\) are due to intersections with 'invisible' higher lying electronic states. Previous detailed ab initio studies exposed these 'invisible' intersection partners as mainly Rydberg states. The aim of this section is to illustrate how the Rydberg/valence interaction affects the potential profiles along the dissociation path and to outline the geometries at which valence states can effectively drain population from the Rydberg manifold. To this end, CASSCF calculations of the first 10 states, 1--5\({}^{1}\!A^{\prime}\) and 1--5\({}^{1}\!A^{\prime\prime}\), have been performed with the d-aug-cc-pVQZ basis set for the CO bond distances \(R_{2}\in[1.7\,a_{0},5.0\,a_{0}]\) and the OCO bond angles \(\alpha_{\rm OCO}\in[100^{\circ},179^{\circ}]\). One CO bond is kept fixed at \(R_{1}=2.4\,a_{0}\), so that only CO + O arrangement channel is covered.
The CASSCF potentials along \(R_{2}\) are shown in All five \(A^{\prime\prime}\) states and four \(A^{\prime}\) states converge to thresholds or. The states correlating with the lowest threshold are of \(\Sigma^{+}\), \(\Delta\), and \(\Pi\) symmetry. The states correlating with the highest threshold are \(\Sigma^{\pm}\), \(\Delta\), and \(\Pi\).
\[\mathrm{CO}_{2}+h\omega(E_{\mathrm{ph.}}\geq\ 9.64\,\mathrm{eV})\ \rightarrow\ \mathrm{O}(^{1}\!S)+\mathrm{CO}(X^{1}\Sigma^{+})\,. \tag{5}\]
At linearity, this state is the Rydberg state \(1^{1}\Sigma_{u}^{+}(3\pi_{u})\). Compared to Fig. 1, many new avoided crossings are found in Fig. 11(a,b). The \(D_{\infty h}\) notation of a state \(|i\rangle\) in these panels is related to the expectation value \(\langle i|L_{z}^{2}|i\rangle\) which a diabatic state preserves across the intersections. Potential curves in Fig. 11(a,b) are color coded according to the \(\langle L_{z}^{2}\rangle\) values: \(\Sigma\) states (\(\langle L_{z}^{2}\rangle=0\)) are shown in green; \(\Pi\) states (\(\langle L_{z}^{2}\rangle=1\)) are blue, and \(\Delta\) states (\(\langle L_{z}^{2}\rangle=4\)) are red/orange. Clearly, most adiabatic curves change color more than once as the CO bond stretches: They are tailored out of several diabatic states.
Familiar from the discussion in Sect. III.3 are the fivefold crossings involving \(1^{1}\Pi_{g}(5\sigma_{g})\) and the accidentally degenerate pair \(1^{1}\Delta_{u}(2\pi_{u})/1^{1}\Sigma_{u}^{-}(2\pi_{u})\). These crossings are marked with arrows in Fig. 11(b). As in the MRD-CI calculations, the \(\Delta_{u}/\Sigma_{u}^{-}\) pair is easily recognizable using matrix elements of \(L_{z}\). At linearity, \(\langle L_{z}^{2}\rangle\) for these two states is either 0 or 4, but already a tiny deviation of \(2^{\circ}\) causes \(\langle L_{z}^{2}\rangle\) to collapse towards an average value of \(\sim 2\). This state mixing is stressed in Fig. 11(a,b) with the \(\Delta_{u}/\Sigma_{u}^{-}\) label and with the red/green color. The second edition of the mixed \(\Delta_{u}/\Sigma_{u}^{-}\) pair is found around 11.5 eV. The Rydberg state \(2^{1}\Delta_{u}(3\pi_{u})\), shown with an orange line in Fig. 11(a,b), carries as a satellite the state \(2^{1}\Sigma_{u}^{-}(3\pi_{u})\) [green line in Fig. 11(b)]. Again, \(\langle L_{z}^{2}\rangle\) for these states assumes a non-integer value between 1.5 and 2.5 in even slightly bent molecule.
The valence pair \(1\Delta_{u}/1\Sigma_{u}^{-}\) converges towards the uppermost threshold and successively traverses the higher lying states. The prominent dissociation barrier in the \(3^{1}\!A^{\prime}\) and \(3^{1}\!A^{\prime\prime}\) states near \(R_{2}=3.6\,a_{0}\) is a clear signature of the sixfold valence/Rydberg crossing involving \(1\Delta_{u}/1\Sigma_{u}^{-}\), the \(1^{1}\Sigma_{u}^{+}\) state [green line in Fig. 11(a)] and a repulsive \(R^{1}\Delta_{u}\) state [orange line in Fig. 11(a,b)]. \(1^{1}\Sigma_{u}^{+}\) is the optically bright Rydberg state responsible for the strong absorption band around 11.1 eV (111.7 nm). The state \(R^{1}\Delta_{u}\) corresponds to a pair of strongly repulsive \({}^{1}\!A^{\prime}/^{1}\!A^{\prime\prime}\) states descending towards threshold from very high energies. Two sections of this repulsive potential curve are seen in Fig. 11(a,b) in the intervals \([3.2\,a_{0},3.7\,a_{0}]\) and \([3.7\,a_{0},5.0\,a_{0}]\). Thus, the asymptotic repulsive portions of the \(3^{1}\!A^{\prime}\) and \(3^{1}\!A^{\prime\prime}\) states diabatically belong to \(R^{1}\Delta_{u}\). The sixfold crossing \(1\Sigma_{u}^{+}/R^{1}\Delta_{u}/1\Delta_{u}/1\Sigma_{u}^{-}\) near \(R_{2}=3.6\,a_{0}\) is not the only one involving \(R^{1}\Delta_{u}\) state. At higher energies and shorter CO bonds, \(R^{1}\Delta_{u}\) intersects the Rydberg pair \(2\Delta_{u}/2\Sigma_{u}^{-}\) (cf. Fig. 11(a,b) near \(R_{2}\approx 3.0\,a_{0}\)). The diabatic \(R^{1}\Delta_{u}\) state distinctly stands out because it strictly preserves the projection \(\langle L_{z}^{2}\rangle\approx 4\) even for \(\alpha_{\rm OCO}=165^{\circ}\), while \(\langle L_{z}^{2}\rangle\) values for the other states become non-integer.
The above discussion is valid for both \(A^{\prime}\) and \(A^{\prime\prime}\) states. The \(A^{\prime\prime}\) symmetry block in Fig. 11(b) contains one more state, namely the Rydberg state \(2^{1}\Pi_{g}(6\sigma_{g})\) missing among the \(A^{\prime}\) states where it would have been \(6A^{\prime}\). This state, materializing out of nowhere at \(R_{2}=3.0\,a_{0}\), is involved in the \(R^{1}\Delta_{u}/2\Delta_{u}/2\Sigma_{u}^{-}\) crossing -- making it a sevenfold intersection.
While the adiabatic gaps in the valence/valence intersection \(1\Pi_{g}/1\Delta_{u}/1\Sigma_{u}^{-}\) grow as \(\alpha_{\rm OCO}\) deviates from \(180^{\circ}\), all Rydberg/Rydberg and Rydberg/valence intersections not only remain recognizable in bent CO\({}_{2}\), but clearly sharpen up. This is illustrated in panels (c) -- (f) of drawn for \(\alpha_{\rm OCO}=175^{\circ}\) and \(160^{\circ}\). As a side result, the repulsive \(R^{1}\!A^{\prime}\) and \(R^{1}\!A^{\prime\prime}\) states, deriving from \(R^{1}\Delta_{u}\), remain distinct at all angles, despite new 'bent' intersections in the \(3,4,5^{1}\!A^{\prime}\) and \(3,4,5^{1}\!A^{\prime\prime}\) states [see, for example, Fig. 11(e,f)]. Via these intersections, carbon dioxide excited with vacuum UV light can -- at bent geometries -- reach any of the shown dissociation channels. Although a detailed analysis of these multiple pathways is beyond the scope of the present work, a brief description of intersections involving the optically bright \({}^{1}\Sigma_{u}^{+}\) state is added at the end of this section.
Potential curves along \(\alpha_{\rm OCO}\) are shown in The crossing patterns, expecially numerous among \(A^{\prime\prime}\) states, explain intricate topography of states \(3^{1}\!A^{\prime}\) and \(3^{1}\!A^{\prime\prime}\) in For example, the sharp barrier in the state \(3^{1}\!A^{\prime\prime}\) at angles between \(150^{\circ}\) and \(170^{\circ}\) [Fig. 3(d,f) and Fig. 3(b,d)] originates from a complicated three state crossing around \(165^{\circ}\) involving states \(3,4,5^{1}\!A^{\prime\prime}\) [Fig. 12(b)]. Diabatically, the decreasing branch of the \(3^{1}\!A^{\prime\prime}\) state at \(\alpha_{\rm OCO}<160^{\circ}\) belongs to \(4^{1}\!A^{\prime\prime}\) -- the state which at linearity merges into the Rydberg pair \(2\Delta_{u}/2\Sigma_{u}^{-}\). This diabatic state can be followed to even smaller angles through another intersection, this time with \(2^{1}\!A^{\prime\prime}\) near \(130^{\circ}\). Ultimately, as indicated with an arrow in Fig. 12(b), it is this diabatic state which the carbene-like OCO minimum belongs to. The avoided crossings in the \(A^{\prime}\) symmetry block are similar and simpler. The three state crossing of \(3,4,5^{1}\!A^{\prime\prime}\) is a mere ghost because the adiabatic gap is almost \(2\,\)eV wide and the barrier in the \(3^{1}\!A^{\prime}\) state is broad and low. The state \(4^{1}\!A^{\prime}\), originating from \(2^{1}\!\Delta_{u}\), stabilizes upon bending and is easily traced to smaller angles through the intersection with \(3^{1}\!A^{\prime}\) near \(\alpha_{\rm OCO}=150^{\circ}\). Again, the carbene-like OCO correlates diabatically with the state originating from \(2^{1}\!\Delta_{u}\) at linearity.
The analysis of broadening and splittings in the strong absorption band of the Rydberg state \(1^{1}\Sigma_{u}^{+}\) focused on the interactions of \(1^{1}\Sigma_{u}^{+}\) with valence states. Present calculations reveal numerous avoided crossings with both valence and Rydberg states inside and outside the FC zone. At linearity, the \(1^{1}\Sigma_{u}^{+}\) state cuts twice through the Rydberg state \(2^{1}\!\Delta_{u}\) [Fig. 11(a)], and both crossings persist to smaller angles [Fig. 11(c,e)]. Another crossing occurs near \(R_{2}=3.7\,a_{0}\) (for all calculated angles) and mixes \(1^{1}\Sigma_{u}^{+}\) with \(R^{1}\Delta_{u}\). This interaction strongly perturbs the ab initio \(\langle L_{z}^{2}\rangle\) values but vanishes beyond \(R_{1}=4.0\,a_{0}\). The potential cut along \(\alpha_{\mathrm{OCO}}\) in Fig. 12(a) exposes another avoided crossing at \(\sim 120^{\circ}\) which involves the states \(4,5^{1}\!A^{\prime}\) and leads to a local minimum in \(5^{1}\!A^{\prime}\) (the state, correlating with \(1^{1}\Sigma_{u}^{+}\) at these R\({}_{\mathrm{CO}}\)). Finally, the barrier in \(5^{1}\!A^{\prime}\) near \(150^{\circ}\) implies interaction with the next higher state, which according to the analysis of Ref. correlates with \(2^{1}\Sigma_{u}^{+}\) at linearity. While all these crossings can redirect population from the optically bright \(1^{1}\Sigma_{u}^{+}\) state along various linear and bent routes, clearly demonstrates that the diabatic \(1^{1}\Sigma_{u}^{+}\) state, calculated at the CASSCF level, is repulsive at all geometries and thus can dissociate directly.
## IV Quasi-diabatization of the valence states
As explained in the Introduction, the diabatic representation, although not generally indispensable, is best suited for the particular implementation of nuclear quantum dynamics used in paper II. Rigorously speaking, diabatization has to be performed simultaneously on all six calculated valence states, because each pair of states intersects either in the FC zone or at bent geometries. Simplifications to this 'Herculean task' stem from the expectation that the two groups of intersections influence photodissociation in different ways: While the 'bent' CIs affect later stages of the product formation, the electronic branching ratios, and/or the rovibrational photofragment distributions, the 'FC' CIs are directly responsible for the shape of the observed absorption spectrum. This distinction guides the practical construction of the diabatic states: The 'bent' CIs and the 'FC' CIs are analyzed at two different levels of detail, commensurate with their expected impact on the absorption spectrum. As a result, the complete multistate problem splits into several steps and the need for a global diabatization of six multiply intersecting states is obviated.
### Intersections at bent geometries: Local diabatization
The vicinity of 'bent' CIs is diabatized locally, using the energy-based scheme as described, for example, by Koppel. All considered CIs are shown in Figs. 3 and 4 and include
1. the \(\tilde{X}^{1}\!A^{\prime}/2^{1}\!A^{\prime}\) pair at \(\alpha_{\rm OCO}=90^{\circ}-110^{\circ}\);
2. the \(2^{1}\!A^{\prime\prime}/3^{1}\!A^{\prime\prime}\) pair at \(\alpha_{\rm OCO}=120^{\circ}-150^{\circ}\);
3. the \(1^{1}\!A^{\prime\prime}/3^{1}\!A^{\prime\prime}\) pair, with \(3^{1}\!A^{\prime\prime}\) state diabatized in step (b), at \(\alpha_{\rm OCO}=70^{\circ}-100^{\circ}\).
The \(\tilde{X}^{1}\!A^{\prime}/2^{1}\!A^{\prime}\) pair is considered as an example. At \(C_{2v}\) geometries, these states belong to \(A_{1}\) and \(B_{2}\) irreps, so that their CI is'symmetry allowed'. Its branching space includes \(\alpha_{\rm OCO}\) as the tuning mode (irrep \(A_{1}\)) and the antisymmetric stretch \(R_{-}\) as the symmetry breaking coupling mode (irrep \(B_{2}\)). The \(2\times 2\) diabatic potential matrix \({\bf V}^{d}\) is constructed on the ab initio grid of bond distances \((R_{1},R_{2})\) from the diagonal adiabatic potential matrix \({\bf V}^{a}\) using the orthogonal adiabatic-to-diabatic transformation (ADT)
\[{\bf V}^{d}={\bf S}^{T}{\bf V}^{a}{\bf S}\,, \tag{6}\]
with the transformation matrix \({\bf S}\) parameterized by the mixing angle \(\Theta\),
\[\Theta(R_{1},R_{2},\alpha_{\rm OCO})=-\frac{1}{2}\arctan\left(\frac{\alpha_{ \rm OCO}-\alpha_{\rm OCO}^{\star}(R_{1},R_{2})}{W_{0}(R_{1},R_{2})}\right), \tag{7}\]
which varies between 0 and \(\pi/2\); \(\alpha_{\rm OCO}^{\star}(R_{1},R_{2})\) denotes the crossing point of the diabatic potentials (\(\Theta=\pi/4\)); the width \(W_{0}\) of the function \(\Theta(\alpha_{\rm OCO})\) is evaluated from the adiabatic energy gaps \(\Delta V^{a}/2\) and the average slope of the diabatic potentials \(\overline{F}\),
\[W_{0}=\left.\frac{\Delta V^{a}}{2\overline{F}}\right|_{\alpha_{\rm OCO}^{\star }}\,, \tag{8}\]
With the off-diagonal coupling in Eq. neglected, the ground state \(\tilde{X}^{1}\!A^{\prime}\) becomes completely decoupled from other \(A^{\prime}\) states. The remaining diagonal matrix element in \({\bf V}^{d}\) is the potential waiting to be further diabatized at linearity. The diabatization is kept local by restricting the interval, over which \(\Theta(\alpha_{\rm OCO})\) varies from 0 to \(\pi/2\), to the vicinity of \(\alpha_{\rm OCO}^{\star}\):
\[\tilde{\Theta}(R_{1},R_{2},\alpha_{\rm OCO})=g_{1}\left[g_{2}\Theta(R_{1},R_{2 },\alpha_{\rm OCO})+\frac{\pi}{2}(1-g_{2})\right]\,, \tag{9}\]The two switching functions \(g_{1,2}(\alpha_{\rm OCO})\),
\[g_{1,2}(\alpha_{\rm OCO})=\left[1+\exp(-(\alpha_{\rm OCO}-\alpha_{1,2})/\lambda_{ 1,2})\right]^{-1}\,, \tag{10}\]
If \((\alpha_{1}-\alpha_{2})\to 0\) in \(g_{1,2}\), the so-called 'diabatization by eye' is recovered, which corresponds to a relabeling of ab initio energies at the crossing angle \(\alpha_{\rm OCO}^{\star}\).
The main advantage of local diabatization, the results of which are shown in 1D in Figs. 3 and 4, is that smooth potentials are created for strongly bent geometries at relatively low cost. The quality of the resulting PESs in 2D, either in the \((R_{1},R_{2})\) or in the \((R_{1},\alpha_{\rm OCO})\) plane is illustrated for all states in Figs. 5, 8, 9, 10, and 13. For the lowest excited states in each symmetry block, \(2^{1}\!A^{\prime}\) and \(1^{1}\!A^{\prime\prime}\), the global bent equilibrium at \(120^{\circ}\) or \(130^{\circ}\) [Fig. 9(a,c)] is not affected by the local diabatization. In contrast, the carbene-like bent OCO minima change the owner: They appear in the state \(2^{1}\!A^{\prime}\) [local minimum near \(70^{\circ}\) barely visible in Fig. 9(a) but pronounced in Fig. 13(a)] and in the state \(3^{1}\!A^{\prime\prime}\) [Fig. 13(d)]. Barriers and local minima arising in the avoided crossings with Rydberg states are not diabatized within this scheme. Examples are the sharp barrier in \(3^{1}\!A^{\prime\prime}\) near \(150^{\circ}\) [Fig. 9(d) and Fig. 13(d)] and the deep local minimum in \(3^{1}\!A^{\prime}\) near \(100^{\circ}\) [Fig. 9(b) and Fig. 13(b)].
### Intersections at linear geometries: Regularized diabatic states
In the second step, the CIs in the FC region are diabatized; kinematic singularities due to glancing intersections between \(A^{\prime}\) and \(A^{\prime\prime}\) states are not considered. Accurate diabatization schemes rely on direct numerical differentiation of the electronic wave functions with respect to nuclear coordinates or on the orbital rotation method as implemented in MOLPRO. Both types of calculations become prohibitively expensive with the d-aug-cc-pVQZ basis set, and further approximations are needed in order to find the diabatic potential matrix, consisting of a \(2\times 2\) block of \(A^{\prime}\) and a \(3\times 3\) block of \(A^{\prime\prime}\) states:
\[{\bf V}^{d}=\left(\begin{array}{c|c|c}V_{\Delta^{\prime}}&V_{\Pi^{\prime} \Delta^{\prime}}\\ V_{\Pi^{\prime}\Delta^{\prime}}&V_{\Pi^{\prime}}\end{array}\right. \tag{11}\]The five diagonal matrix elements are the five diabatic states \(\Sigma^{\prime\prime}\), \(\Pi^{\prime}\), \(\Pi^{\prime\prime}\), \(\Delta^{\prime}\), and \(\Delta^{\prime\prime}\), with the smooth intersecting potentials in the \((R_{1},R_{2}|\alpha_{\rm OCO}=180^{\circ})\) plane constructed to coincide at \(\alpha_{\rm OCO}=180^{\circ}\) with the ab initio PESs for \(1^{1}\Sigma^{-}\), \(1^{1}\Pi_{g}\), and \(1^{1}\Delta_{u}\), respectively. The semi-global diabatization scheme, akin to the vibronic coupling model and adjusted to the topography of the closed fivefold CI seam, has already been introduced in Ref.. Diabatization proceeds in two steps. First, a model diabatic matrix of the form Eq. with elements \(V_{ij}^{}\) is constructed. Due to orbital degeneracy, \(V_{\Pi^{\prime}}^{}(R_{1},R_{2})=V_{\Pi^{\prime\prime}}^{}(R_{1},R_{2})\) and \(V_{\Delta^{\prime}}^{}(R_{1},R_{2})=V_{\Delta^{\prime\prime}}^{}(R_{1},R _{2})\). Moreover, the accidental near-degeneracy of states \(1^{1}\Sigma_{u}^{-}\) and \(1^{1}\Delta_{u}\) implies \(V_{\Sigma^{\prime\prime}}^{}(R_{1},R_{2})\approx V_{\Delta^{\prime\prime}}^ {}(R_{1},R_{2})\) in a broad vicinity of the closed CI seam -- the property which substantially simplifies diabatization of \(A^{\prime\prime}\) states. Deviations from linearity, measured by the coordinate \(Q_{u}\sim\sin\alpha_{\rm OCO}\), are included in the model via off-diagonal matrix elements represented as symmetry adapted expansions in \(Q_{u}\):
\[V_{\Pi^{\prime}\Delta^{\prime}}^{} = \sum_{k=0}^{N^{\prime}}\alpha_{k}(R_{1},R_{2})Q_{u}^{2k+1} \tag{12}\] \[V_{\Sigma^{\prime\prime}\Pi^{\prime\prime}}^{} = V_{\Pi^{\prime\prime}\Delta^{\prime\prime}}^{}=\sum_{k=0}^{N^ {\prime\prime}}\beta_{k}(R_{1},R_{2})Q_{u}^{2k+1}\,. \tag{13}\]
Couplings of the accidentally degenerate \(\Sigma^{\prime\prime}\) and \(\Delta^{\prime\prime}\) states to \(\Pi^{\prime\prime}\) are set equal, while the matrix element \(V_{\Sigma^{\prime\prime}\Delta^{\prime\prime}}^{}\) for the RT-like \(\Sigma^{\prime\prime}/\Delta^{\prime\prime}\) interaction is neglected. The model is complete after the expansion coefficients in \(V_{ij}^{}\) are calculated from a non-linear least-squares fit to ab initio energies. In the second step, the regularized diabatic states approach is invoked, and the matrix elements \(V_{ij}^{}\) are used to define the orthogonal ADT matrix, which is applied to the adiabatic matrix \({\bf V}^{a}\) via Eq. giving the desired diabatic matrix elements of Eq. on the full ab initio grid. Final interpolation is performed using 3D splines.
The diabatization is localized to the vicinity of the CI seam by modifying matrix elements \(V_{ij}^{}\) (\(i\neq j\)) in Eq.,
\[\tilde{V}_{ij}^{}=\frac{V_{ij}^{}}{1+\exp\left[(\rho-\rho_{0})/\lambda_ {0}\right]}\,, \tag{14}\]
Adiabatic and diabatic states are forced to coincide if either bond becomes longer than \(\sim 3.8\,a_{0}\), so that diabatization regions for linear and bent CIs are cleanly kept apart.
The constructed representation describes best the vicinity of linearity in which non-adiabatic transitions occur. The exact range of validity is determined by the length of the expansion in Eq.. The choice \(N^{\prime}=N^{\prime\prime}=1\) gives a model which fits ab initio data with a root mean square error of \(\sim 180\,\mathrm{cm}^{-1}\) for angles \(\alpha_{\mathrm{OCO}}\geq 150^{\circ}\). The ADT, constructed using this model, is guaranteed to remove kinematic singularities at the CIs, but leaves the strength of residual non-adiabatic couplings unspecified. The ultimate test of the scheme is the quantum mechanical absorption spectrum described in paper II. In order to assess the accuracy of the truncated \(Q_{u}\) expansion, three diabatic representations are constructed, based on the expansion coefficients obtained from fitting in three different angular ranges \(180^{\circ}-170^{\circ}\), \(180^{\circ}-160^{\circ}\), and \(180^{\circ}-150^{\circ}\). The corresponding absorption spectra are virtually identical. The spectra are also insensitive to small variations in \(\rho_{0}\) and \(\lambda_{0}\) in Eq. -- the modifications take place too far away from the crossing seam to affect nuclear dynamics.
Another test of the constructed ADT is given in Fig. 14, in which the model mixing angle \(\tilde{\Theta}\) for states \(2,3^{1}A^{\prime}\) is compared with the ab initio one calculated using MOLPRO with a smaller cc-pVQZ atomic basis set. The dependence \(\tilde{\Theta}(R_{1})\) on the CO bond length has a characteristic bell shape: The closed CI seam is intersected twice giving rise to the ascending and the descending branch. The curve \(\tilde{\Theta}(R_{1})\) flattens out as \(\alpha_{\mathrm{OCO}}\) decreases and \(\mathrm{CO}_{2}\) leaves the degeneracy plane at \(180^{\circ}\). Agreement between the model and the ab initio results is satisfying for all angles. A constant shift of \(25^{\circ}\) applied to the ab initio mixing angle has no effect on the strength of non-adiabatic coupling proportional to \(\partial\tilde{\Theta}/\partial R_{1}\).
One-dimensional cuts through the diabatic PESs (diagonal matrix elements corresponding to states \(1^{1}\Pi_{g}\) and \(1^{1}\Delta_{u}\)) are shown in They cross at all geometries and can be directly compared with adiabatic states in The off-diagonal coupling matrix elements are large in the intersection region vanishing off towards the asymptotic channels. Diabatic matrix elements are further illustrated in in the \((R_{1},R_{2})\) plane and in in the \((R_{1},\alpha_{\mathrm{OCO}})\) plane. In all cuts, the diabatic PESs smoothly depend on internal coordinates. In the \((R_{1},R_{2})\) plane, the off-diagonal diabatic coupling stays localized in the inner region. In contrast, the coupling along bending coordinate is delocalized across a substantial \(\alpha_{\mathrm{OCO}}\) range in the \((R_{1},\alpha_{\mathrm{OCO}})\) plane. As a result, the angular shape of the diabatic potentials is distorted compared to the adiabatic case: Diabatic potentials along the coupling mode \(\alpha_{\mathrm{OCO}}\) are close to the average adiabatic potential \(\frac{1}{2}(V_{i}^{a}+V_{j}^{a})\).
Conclusions
This paper describes properties of global PESs of the first six singlet electronic states of CO\({}_{2}\) constructed from about 5000 symmetry unique ab initio points calculated with the d-aug-cc-pVQZ basis set using the MRD-CI method. The main results can be summarized as follows:
1. Calculations accurately reproduce the known benchmarks for all states and establish missing benchmarks for the future calculations: Bond distances and bond angles are accurate to within 0.1%, known fundamental frequencies (mainly ground state \(\tilde{X}^{1}\!A^{\prime}\)) are accurate to within 1.5 cm\({}^{-1}\), the accuracy of the vertical excitation energies is expected to be better than 0.05 eV; dissociation energies agree with experimental thresholds within 0.15 eV for four covered arrangement channels.
2. Local equilibria are abundant in the calculated states. Bent OCO isomer is found in the adiabatic states \(1,3^{1}\!A^{\prime}\) and \(1,3^{1}\!A^{\prime\prime}\). Linear COO is found in \(1,2^{1}\!A^{\prime}\) and \(1,2^{1}\!A^{\prime\prime}\). Their diabatic electronic origin is clarified, and the properties, including equilibrium geometries, excitation energies, and vibrational frequencies, are established.
3. Near degeneracies can be found for each pair of six valence states, at linear or bent geometries, or at both. Avoided crossings and conical and glancing intersections literally shape the observed topography of the excited electronic states. Detected intersections are not limited to the valence manifold and the search for electronic origins of local minima and barriers involves valence/Rydberg and Rydberg/Rydberg intersections.
4. Characteristic for state intersections in CO\({}_{2}\), both conical and glancing, is that they include several states. In the FC region, a fivefold intersection between \(1^{1}\Sigma_{u}^{-}\), \(1^{1}\Pi_{g}\), and \(1^{1}\Delta_{u}\) states is found. The seam of this intersection forms a closed loop, spectroscopic manifestations of which are discussed in paper II. Outside the FC region at linearity, six- and sevenfold intersections are predicted, some of which persist over extended angular range in the bent molecule.
5. Diabatic \(6\times 6\) potential matrix, with all elements smoothly depending on internal coordinates, is constructed using two-step local diabatizations of linear and bent conical intersections.
It is tempting to try to infer the course of photodissociation and the principal features of the absorption spectrum -- the outcome of a complicated quantum mechanical calculation -- from the constructed PESs alone. Two issues have to be resolved if one deals with five interacting states. The first is the strength of diabatic (intra-symmetry) and RT (inter-symmetry) coupling. If the off-diagonal vibronic coupling is weak, the diabatic potentials should be chosen as 'zeroth order guides'. If the vibronic coupling is strong, it is the adiabatic description which becomes relevant -- and the difference between the two pictures is striking, especially along the bending angle as Figs. 9 and 17 demonstrate. As discussed in paper II, the vibronic coupling is strong, the RT interaction between \(A^{\prime}\) and \(A^{\prime\prime}\) states is to a large extent quenched, and the adiabatic potentials can be used for qualitative analysis. The global minima of states \(2^{1}\!A^{\prime}\) or \(1^{1}\!A^{\prime\prime}\) are bent and lie \(\sim 2.0\,\)eV below the dissociation threshold or \(\sim 4.5\,\)eV below the FC point. The route connecting the FC region with these bent equilibria is barrierless. In contrast, there is a barrier to dissociation near linearity -- the leftover of the lower cone of the \(\Pi/\Delta\) CI. Thus, one expects the low energy bands in the absorption spectrum, associated with \(2^{1}\!A^{\prime}\) and \(1^{1}\!A^{\prime\prime}\) states, to reflect highly excited bending motion. This interpretation is commonly given in the literature: Instead of dissociating directly, the molecule bends first. With growing photon energy, the contribution of direct dissociation through linear geometries will certainly grow, because the barrier is only about 0.2 eV high and is located \(\sim 0.4\,\)eV below the FC point. Above \(\sim 9\,\)eV, the valence states \(3^{1}\!A^{\prime}\) and \(3^{1}\!A^{\prime\prime}\) will contribute to the observed spectrum. These 'linear' states are separated from the dissociation asymptote by a \(\sim 1\,\)eV high and broad barrier, with the implication that CO\({}_{2}\) in these states can decay only through non-adiabatic interactions with the dissociation continuum of the lower states. In other words, one expects to see a resonance-dominated absorption spectrum. The next qualitative change in the absorption spectrum within the valence manifold can be expected after the photon energy reaches the top of the dissociation barrier in the upper valence states and allows direct dissociation from linearity.
The above discussion is based on 1D and 2D potential cuts and the data in Table 3 -- given the adiabatic representation is the adequate one. However, there is another important piece of information still missing, namely the TDMs with the ground state. As has already been mentioned in Sect. III.2, the electronic transitions in the wavelength range of 160 nm -- 120 nm are forbidden. The bands are observed only because the TDMs are not constant but strongly change with molecular coordinates as one moves away from the high-symmetryFC point. This dependence is a manifestation of the Herzberg-Teller effect which plays the leading role in shaping the absorption bands, is at least as important as the potential profiles discussed above, and has to be considered on equal footing with the state intersections. The discussion of the coordinate dependence of the TDMs is deferred to paper II.
|
10.48550/arXiv.1304.2464
|
Photodissociation of carbon dioxide in singlet valence electronic states. I. Six multiply intersecting ab initio potential energy surfaces
|
Sergy Yu. Grebenshchikov
| 1,328
|
10.48550_arXiv.1412.3847
|
###### Abstract
Since the advent of quantum mechanics different approaches to find analytical solutions of the Schrodinger equation have been successfully developed. Here we follow and generalize the approach pioneered by Natanzon and others by which the Schrodinger equations can be transformed into another well-known equation for transcendental function (e.g., the hypergeometric equation). This sets a class of potentials for which this transformation is possible. Our generalization consists in finding potentials allowing the transformation of the Schrodinger equation into a triconfluent Heun equation. We find the energy eigenvalues of this class of potentials, the eigenfunction and the exact superpartners.
pacs: XXX
## I Introduction
The study of exactly solvable Schrodinger equations can be traced back to the beginning of quantum mechanics. The earliest examples are represented by the harmonic oscillator, Coulomb, Morse, Poschl-Teller, Eckart and Manning-Rosen potentials. By looking closer at these examples one can start to conjecture that exact solvability of the Schrodinger equation depends on the fact that such an equation can be suitably reduced to the hypergeometric or the confluent hypergeometric equation. The problem of deriving the most general class of potentials such that the Schrodinger equation can be transformed into the hypergeometric equation has been solved by. Furthermore, studied solutions regular at infinity for the basic SUSY ladder of Hyperbolic Poschl-Teller potentials that admit representations in terms of confluent Heun polynomials. derived new classes of potentials such that the one-dimensional Schrodinger equation can be turned into the Heun equation and its confluent cases. The generalized Heun equation has been considered as well in.
\[y(1-y)\frac{d^{2}v}{dy^{2}}+\left[c-(a+b+1)y\right]\frac{dv}{dy}-abv(y)=0,\quad y \in I\subset\mathbb{R} \tag{1}\]
with \(a,b,c\in\mathbb{R}\) is a special case of the Heun equation
\[\frac{d^{2}v}{dy^{2}}+\left(\frac{\gamma}{y}+\frac{\delta}{y-1}+\frac{\epsilon }{y-a}\right)\frac{dv}{dy}+\frac{\alpha\beta y-q}{y(y-1)(y-a)}v(y)=0,\quad y \in I\subset\mathbb{R}, \tag{2}\]
We start by reviewing the method developed by allowing the construction of the most general potential such that the Schrodinger equation can be reduced to the triconfluent Heun equation. Such a potential contains six free parameters.
Reduction of the radial Schrodinger equation to a triconfluent Heun equation
We consider a quantum particle in a central field in three and two spatial dimensions, respectively. In the three dimensional case the behaviour of the particle is described by the Schrodinger equation (\(\hbar^{2}=2m=1\))
\[i\partial_{t}\Psi(t,\mathbf{r})=H\Psi(t,\mathbf{r}),\quad H=-\Delta+U(r)\]
Here, \(\Delta\) denotes the Laplacian. Since the Hamilton operator commutes with the angular momentum operator \(\mathbf{L}=(L_{x},L_{y},L_{z})\), it is sufficient to solve the eigenvalue problem for \(H\), i.e. the time-dependent Schrodinger equation, on the subspace of \(L^{2}(\mathbb{R}^{3})\) spanned by a basis of common eigenvectors of \(|\mathbf{L}^{2}|\) and \(L_{z}\). Such a subspace is made of functions having the form \(\Phi_{\ell m}(r,\theta,\phi)=\Phi(r)Y_{\ell m}(\theta,\phi)\) where \(\Phi\in L^{2}(\mathbb{R}_{+},r^{2}dr)\) and \(Y_{\ell m}(\theta,\phi)\) denotes the spherical harmonics.
\[\Delta=\frac{1}{r^{2}}\partial_{r}r^{2}\partial_{r}-\frac{|\mathbf{L}^{2}|}{r ^{2}}.\]
If \(\Phi_{\ell m}\) is an eigenfunction of \(H\) relative to the eigenvalue \(E\), then \(\Phi\) must satisfy the radial Schrodinger equation
\[\left[-\left(\frac{d^{2}}{dr^{2}}+\frac{2}{r}\frac{d}{dr}-\frac{\ell(\ell+1)}{ r^{2}}\right)+U(r)\right]\Phi(r)=E\Phi(r).\]
The above equation can be further simplified if we bring it to its canonical form by means of the transformation \(\Phi(r)=\psi(r)/r\) with \(\psi\in L^{2}(\mathbb{R}_{+},r^{2}dr)\) and we obtain
\[\frac{d^{2}\psi}{dr^{2}}+\left[E-V_{eff}(r)\right]\psi(r)=0,\quad V_{eff}(r)= \frac{\ell(\ell+1)}{r^{2}}+U(r). \tag{3}\]
The region where the dynamics of the particle can take place is constrained to the half axis \(r>0\). Moreover, it will be assumed that the effective potential does not depend on the energy of the particle. Two-dimensional quantum systems appear in solid state physics in connection with the fractional quantum Hall effect and high temperature superconductivity. For the two-dimensional case we consider a system of two anyons, i.e. particles with fractional statistics, in a spherically-symmetric potential. The time-independent Schrodinger equation governing this system is (\(\hbar^{2}=2\mu=1\) with \(\mu\) the reduced mass of the system)
\[\left[-\left(\frac{d^{2}}{dr^{2}}+\frac{1}{r}\frac{d}{dr}-\frac{\nu^{2}}{r^{2} }\right)+U(r)\right]\Phi(r)=E\Phi(r),\quad\nu=m+\frac{\theta}{\pi},\]
Note that \(r\neq 0\) to avoid that the two particles overlap.
\[V_{eff}(r)=\frac{\nu^{2}-\frac{1}{4}}{r^{2}}+U(r). \tag{4}\]
We want to construct the most general potential \(U\) such that the radial Schrodinger equation with effective potential given by with \(\ell=0\) or with fixed \(\nu\) can be transformed into the triconfluent Heun equation
\[\frac{d^{2}v}{d\rho^{2}}+I(\rho)v(\rho)=0,\quad I(\rho)=A_{0}+A_{1}\rho+A_{2} \rho^{2}-\frac{9}{4}\rho^{4} \tag{5}\]
Note that can be obtained by means of a confluence process of the singularities involved in the Heun equation. To appreciate the role of the TCH equation in physics, we refer to where this equation appears in the treatment of anharmonic oscillators in quantum mechanics. A short exposition of known results concerning anharmonic oscillators can be found in Ch. I of.
\[r=\rho,\quad\psi=v,\quad E=A_{0},\quad\ell=0,\quad U(r)=-A_{1}r-A_{2}r^{2}+ \frac{9}{4}r^{4},\]i.e. \(U\) is a quartic potential. In what follows we are interested in non trivial transformations of the dependent variable \(\psi\) and the radial coordinate \(r\) transforming into.
\[\frac{d^{2}v}{dr^{2}}-\frac{\rho^{{}^{\prime\prime}}}{\rho^{{}^{\prime}}}\frac{ dv}{dr}+(\rho^{{}^{\prime}})^{2}I(\rho(r))v(\rho(r))=0,\quad^{{}^{\prime}}:= \frac{d}{dr}.\]
The standard form of the above equation can be achieved by using the Liouville transformation
\[v(\rho(r))=\exp\left(\frac{1}{2}\int\frac{\rho^{{}^{\prime\prime}}(r)}{\rho^{{} ^{\prime}}(r)}\ dr\right)\psi(r)=\sqrt{\rho^{{}^{\prime}}(r)}\psi(r),\]
where we must require that
\[\rho^{{}^{\prime}}(r)>0 \tag{6}\]
\(\rho\) must be an increasing function of the variable \(r\). Hence, we end up with the linear ODE
\[\frac{d^{2}\psi}{dr^{2}}+J(r)\psi(r)=0,\quad J(r)=(\rho^{{}^{\prime}})^{2}I( \rho(r))+\frac{1}{2}S(\rho). \tag{7}\]
Here, \(S(\rho)\) denotes the Schwarzian derivative of the coordinate transformation \(\rho\) evaluated at \(r\) and it is given by
\[S(\rho)=\frac{\rho^{{}^{\prime\prime\prime}}}{\rho^{{}^{\prime}}}-\frac{3}{2} \left(\frac{\rho^{{}^{\prime\prime}}}{\rho^{{}^{\prime}}}\right)^{2}.\]
It can be easily checked that solutions of will be expressed in terms of the solutions of according to
\[\psi(r)=\frac{1}{\sqrt{\rho^{{}^{\prime}}(r)}}v(\rho(r)).\]
Moreover, equation will reduce to the radial Schrodinger equation when \(J(r)=E-V_{eff}(r)\). Therefore, the effective potential is completely determined by the Bose invariant \(I\) and the Schwarzian derivative of the coordinate transformation.
* the Bose invariant admits a decomposition of the form \[I(\rho)=I_{0}(\rho)+EI_{1}(\rho).\] According to Theorem IV.6 in, this will be the case if the parameters entering in the triconfluent Heun equation can be written as \(A_{i}=a_{i}+Eb_{i}\) with \(i=0,1,2\) and \(a_{i},b_{i}\in\mathbb{R}\) for any \(i=0,1,2\). Then, we have \[I_{0}(\rho)=a_{0}+a_{1}\rho+a_{2}\rho^{2}-\frac{9}{4}\rho^{4},\quad I_{1}(\rho) =b_{0}+b_{1}\rho+b_{2}\rho^{2}.\]
* The coordinate transformation satisfies the differential equation \[\left(\rho^{{}^{\prime}}\right)^{2}I_{1}(\rho(r))=1.\] The condition expressed by implies that we have to take the positive square root of, that is \[\rho^{{}^{\prime}}(r)=\frac{1}{\sqrt{I_{1}(\rho(r))}}.\] With the help of and the effective potential can be written as \[V_{eff}(r)=-\frac{I_{0}(\rho(r))}{I_{1}(\rho(r))}-\frac{1}{2}S(\rho).\]Rewriting the Schwarzian derivative in terms of \(I_{1}\) and its derivatives with respect to \(\rho\) we obtain the most general form for the effective potential such that the radial Schrodinger equation can be transformed into a triconfluent Heun equation, namely
\[V_{eff}(r)=-\frac{I_{0}(\rho(r))}{I_{1}(\rho(r))}+\frac{4I_{1}(\rho(r))\tilde{I _{1}}(\rho(r))-5\dot{I_{1}}^{2}(\rho(r))}{16I_{1}^{3}(\rho(r))},\quad\cdot:= \frac{d}{d\rho}. \tag{11}\]
Replacing the corresponding expressions for \(I_{0}\) and \(I_{1}\) into we can write the effective potential as
\[V_{eff}(r)=-\frac{4a_{0}+4a_{1}\rho+4a_{2}\rho^{2}-9\rho^{4}}{4(b_{0}+b_{1}\rho +b_{2}\rho^{2})}-\frac{12b_{2}^{2}\rho^{2}+12b_{1}b_{2}\rho+5b_{1}^{2}-8b_{0}b _{2}}{16(b_{0}+b_{1}\rho+b_{2}\rho^{2})^{3}},\quad\rho=\rho(r). \tag{12}\]
Note that represents a class of potentials depending upon the six real parameters \(a_{i},b_{i}\) with \(i=0,1,2\). Concerning the solution of the differential equation governing the coordinate transformation we need to require that \(I_{1}(\rho(r))>0\) for \(r\in(0,\infty)\). Let \(\Delta=b_{1}^{2}-4b_{0}b_{2}\).
1. \(b_{0},b_{1},b_{2}\neq 0\) and \(b_{2}>0\). 1. If \(\Delta<0\), the coordinate transformation \(\rho\) maps the interval \((0,\infty)\) of the radial variable into the whole real line. Using 2.261 and 2.262 in we find that the solution of can be written as \[r(\rho)=\frac{1}{2}\left(\rho+\frac{b_{1}}{2b_{2}}\right)\sqrt{I_{1}}-\frac{ \Delta}{8b_{2}\sqrt{b_{2}}}\mathrm{arcsinh}\bigg{(}\frac{2b_{2}\rho+b_{1}}{ \sqrt{-\Delta}}\bigg{)}.\] As \(\rho\rightarrow\infty\) we see that \(r\approx(\sqrt{b_{2}}/2)\rho^{2}\). 2. Let \(\Delta=0\). Since \(b_{1}^{2}>0\), \(b_{2}>0\), and \(b_{1}^{2}=4b_{0}b_{2}\), then \(b_{0}>0\) and \(b_{1}=\pm 2\sqrt{b_{0}b_{2}}\). Hence, \(I_{1}=(\sqrt{b_{2}}\rho\pm\sqrt{b_{0}})^{2}\) and equation becomes \[\rho^{{}^{\prime}}(r)=\frac{1}{\sqrt{b_{2}}\rho\pm\sqrt{b_{0}}},\] where the plus sign must be taken when \(b_{1}>0\). Note that \(\rho^{{}^{\prime}}>0\) whenever \(\rho>\mp\sqrt{b_{0}/b_{2}}\) where the minus sign must be taken for \(b_{1}>0\). Integrating the above equation we obtain \[r(\rho)=\frac{\sqrt{b_{2}}}{2}\rho^{2}\pm\sqrt{b_{0}}\rho.\] The above equation is quadratic in \(\rho\) and therefore we can solve it in order to express \(\rho\) as a function of the radial variable. In the case \(b_{1}>0\) and requiring that \(\rho\) is an increasing function of the radial variable we find that \[\rho(r)=\frac{-\sqrt{b_{0}}+\sqrt{b_{0}+2\sqrt{b_{2}}r}}{\sqrt{b_{2}}}.\] In this case \(\rho\) maps the interval \((0,\infty)\) into itself and the condition \(\rho>-\sqrt{b_{0}/b_{2}}\) is automatically satisfied. If \(b_{1}<0\), by a similar reasoning we find the solution \[\rho(r)=\frac{\sqrt{b_{0}}+\sqrt{b_{0}+2\sqrt{b_{2}}r}}{\sqrt{b_{2}}}.\] The image of the interval \((0,\infty)\) under the transformation \(\rho\) is the interval \((2\sqrt{b_{0}/b_{2}},\infty)\) and also in this case the condition \(\rho>\sqrt{b_{0}/b_{2}}\) is satisfied. If \(b_{0}=0\), then \(\Delta=0\) implies \(b_{1}=0\) and \(\rho\) is given by the expression \[\rho(r)=\sqrt{\frac{2r}{\sqrt{b_{2}}}}.\] 3. If \(\Delta>0\), then \(I_{1}>0\) on the interval \((-\infty,\rho_{1})\cup(\rho_{2},\infty)\) where \(\rho_{1}\) and \(\rho_{2}\) denote the roots of \(I_{1}\). In this case we can express the radial variable in terms of \(\rho\) as \[r(\rho)=\frac{1}{2}\left(\rho+\frac{b_{1}}{2b_{2}}\right)\sqrt{I_{1}}-\frac{ \Delta}{8b_{2}\sqrt{b_{2}}}\ln\Big{(}2\sqrt{b_{2}I_{1}}+2b_{2}\rho+b_{1} \Big{)}.\]2. \(b_{0},b_{1},b_{2}\neq 0\) and \(b_{2}<0\). In this case only if \(\Delta>0\) we can make \(I_{1}>0\) on an open interval \((\rho_{1},\rho_{2})\) where \(\rho_{1}\) and \(\rho_{2}\) denote the roots of \(I_{1}\). Employing 2.261 and 2.262 in yield \[r(\rho)=\frac{1}{2}\left(\rho+\frac{b_{1}}{2b_{2}}\right)\sqrt{I_{1}}+\frac{ \Delta}{8b_{2}\sqrt{-b_{2}}}\text{arcsin}\left(\frac{2b_{2}\rho+b_{1}}{\sqrt{ \Delta}}\right).\]
3. If \(b_{2}=0\) and \(b_{1}\neq 0\), then \(I_{1}(\rho)=b_{1}\rho+b_{0}\) and we must require that \(\rho>-b_{0}/b_{1}\). The solution of the ODE governing the coordinate transformation is \[r(\rho)=\frac{2}{3b_{1}}(b_{1}\rho+b_{0})^{3/2}.\] Moreover, we can express \(\rho\) in terms of the radial coordinate as \[\rho(r)=-\frac{b_{0}}{b_{1}}+\frac{1}{b_{1}}\left(\frac{3b_{1}r}{2}\right)^{2 /3}.\] We can observe that \(\rho\) is an increasing function of \(r\) and it maps the interval \((0,\infty)\) to \((-b_{0}/b_{1},\infty)\). If \(b_{0}=0\) the above expression simplifies to \[\rho(r)=\left(\frac{3}{2}\right)^{2/3}b_{1}^{-1/3}r^{2/3}.\]
## III Analysis of the potentials
We analyze those potentials arising from the cases when the coordinate transformation \(\rho\) can be written as an explicit function of the real variable \(r\).
### The case \(\Delta=0\), and \(b_{0},b_{1},b_{2}>0\)
It is straightforward to verify that \(I_{1}=b_{0}+2\sqrt{b_{2}}r\) and the effective potential can be written as
\[V_{1}(r)=c_{0}+\frac{c_{1}}{\sqrt{b_{0}+2\sqrt{b_{2}}r}}+\frac{c_{2}}{b_{0}+2 \sqrt{b_{2}}r}+\frac{c_{3}}{(b_{0}+2\sqrt{b_{2}}r)^{2}}+c_{4}\sqrt{b_{0}+2 \sqrt{b_{2}}r}+c_{5}r \tag{13}\]
with
\[c_{0}=\frac{63b_{0}}{4b_{2}^{2}}-\frac{a_{2}}{b_{2}},\quad c_{1}=\frac{2a_{2} \sqrt{b_{0}}}{b_{2}}-\frac{a_{1}}{\sqrt{b_{2}}}-\frac{9b_{0}\sqrt{b_{0}}}{b_{ 2}^{2}},\quad c_{2}=-a_{0}+\frac{a_{1}\sqrt{b_{0}}}{\sqrt{b_{2}}}+\frac{9b_{0}^ {2}}{4b_{2}^{2}}-\frac{a_{2}b_{0}}{b_{2}},\]
\[c_{3}=-\frac{3}{4}b_{2},\quad c_{4}=-\frac{9\sqrt{b_{0}}}{b_{2}^{2}},\quad c_{ 5}=\frac{9}{2b_{2}\sqrt{b_{2}}}\]
Note that \(c_{0}=c_{1}=c_{2}=0\) if we choose
\[a_{0}=9\left(\frac{b_{0}}{b_{2}}\right)^{2},\quad a_{1}=\frac{45}{2}\left( \frac{b_{0}}{b_{2}}\right)^{3/2},\quad a_{2}=\frac{63b_{0}}{4b_{2}}.\]
We observe that
1. Since \(r>0\) and \(b_{0}>0\), it follows that the denominators appearing in can never vanish and therefore the potential is never singular on the open interval \((0,\infty)\).
2. If \(r=0\) the potential is finite there while asymptotically for \(r\to\infty\) it grows linearly as \[V_{1}(r)\approx\frac{9r}{2b_{2}\sqrt{b_{2}}}.\]3. There are choices of the parameters such that \(V_{1}\) has at least a global minimum on the positive real line.
The study of the critical points of this potential leads to a polynomial equation of degree eight. Applying Descartes' rule of signs we find that the potential has four, two, or no positive critical points whenever \(c_{1}\) and \(c_{3}\) are negative and \(c_{2}\) is positive; three or one positive critical point if \(c_{1}<0\) and \(c_{2},c_{3}>0\), and only one positive critical point if \(c_{1}\), \(c_{2}\), and \(c_{3}\) are positive. For other choices of the signs of the coefficients \(c_{1}\), \(c_{2}\), and \(c_{3}\) there are two or no positive critical point.
### The case \(\Delta=0\), \(b_{0},b_{2}>0\), and \(b_{1}<0\)
The function \(I_{1}\) is formally given as in the previous case and the effective potential has the same form as with coefficients \(d_{0},\cdots,d_{5}\) given by
\[d_{0}=c_{0},\quad d_{1}=-\frac{2a_{2}\sqrt{b_{0}}}{b_{2}}-\frac{a_{1}}{\sqrt{b _{2}}}+\frac{9b_{0}\sqrt{b_{0}}}{b_{2}^{2}},\quad d_{2}=a_{0}-\frac{a_{1}\sqrt {b_{0}}}{\sqrt{b_{2}}}+\frac{9b_{0}^{2}}{4b_{2}^{2}}-\frac{a_{2}b_{0}}{b_{2}}, \quad d_{3}=c_{3},\quad d_{4}=-c_{4},\quad d_{5}=c_{5}.\]
Note that \(d_{0}=d_{1}=d_{2}=0\) if we choose
\[a_{0}=-36\left(\frac{b_{0}}{b_{2}}\right)^{2},\quad a_{1}=-\frac{45}{2}\left( \frac{b_{0}}{b_{2}}\right)^{3/2},\quad a_{2}=\frac{63b_{0}}{4b_{2}}.\]
We observe that
1. Since \(r>0\) and \(b_{0}>0\), it follows that \(I_{1}\) can never vanish and therefore the potential is never singular.
2. The potential is finite at \(r=0\) while asymptotically for \(r\to\infty\) it grows linearly as.
3. There are again choices of the parameters such that the potential has at least one minimum.
Behaviour of the potential for different choices of the parameters.
### The case \(\Delta=0\), \(b_{2}>0\), and \(b_{0}=b_{1}=0\)
The potential takes the form
\[V_{3}(r)=e_{0}+\frac{e_{1}}{\sqrt{r}}+\frac{e_{2}}{r}+\frac{e_{3}}{r^{2}}+e_{4}r \tag{14}\]
where
\[e_{0}=-\frac{a_{2}}{b_{2}},\quad e_{1}=-\frac{a_{1}}{\sqrt{2}b_{2}^{3/4}},\quad e _{2}=-\frac{a_{0}}{2\sqrt{b_{2}}},\quad e_{3}=-\frac{3}{16},\quad e_{4}=c_{5}.\]
Note that it becomes singular at \(r=0\) and asymptotically at infinity it behaves as. There are choices of the parameters such that this potential has a maximum very close to \(r=0\) and a minimum. See The potential is singular at \(r=0\) and it grows linearly for \(r\to\infty\). We have choices of the parameters such that the potential exhibits maxima and minima.
\[2e_{4}\tau^{6}-e_{1}\tau^{3}-2e_{2}\tau^{2}-4e_{3}=0. \tag{15}\]
If \(a_{0}\) and \(a_{1}\) are both positive, there is no change of sign in and Descartes' rule of signs implies that the sextic equation will not have any positive real root, and hence the potential has no extrema. For all other choices of the signs of the coefficients \(a_{0}\), and \(a_{1}\), there are only two sign changes in and hence the potential will admit two positive critical points or none. For a complete root classification of a sextic equation we refer to.
### The case \(b_{2}=0\)
It can be easily checked that \(I_{1}(r)=(3b_{1}r/2)^{2/3}\) and the potential can be written as
\[V_{4}(r)=v_{0}+\frac{v_{1}}{r^{2/3}}+\frac{v_{2}}{r^{2}}+v_{3}r^{2/3}+v_{4}r^{ 4/3}+v_{5}r^{2} \tag{16}\]
Behaviour of the potential for the case \(\Delta=0\), \(b_{0},b_{2}>0\), and \(b_{1}<0\).
with
\[v_{0}=-\frac{a_{1}}{b_{1}}+\frac{2a_{2}b_{0}}{b_{1}^{2}}-\frac{9b_{0}^{3}}{b_{1}^{ 4}},\quad v_{1}=\left(\frac{2}{3b_{1}}\right)^{2/3}\left(\frac{a_{1}b_{0}}{b_{1 }}-a_{0}-\frac{a_{2}b_{0}^{2}}{b_{1}^{2}}+\frac{9b_{0}^{4}}{4b_{1}^{4}}\right), \quad v_{2}=-\frac{5}{36},\]
\[v_{3}=\frac{1}{b_{1}^{2/3}}\left(\frac{2}{3b_{1}}\right)^{2/3}\left[\frac{27}{ 2}\left(\frac{b_{0}}{b_{1}}\right)^{2}-a_{2}\right],\quad v_{4}=-\frac{3^{10/3 }b_{0}}{2^{4/3}b_{1}^{8/3}},\quad v_{5}=\frac{81}{16b_{1}^{2}}.\]
We can set \(v_{0}=v_{1}=v_{3}=0\) if we choose
\[a_{0}=\frac{27}{4}\left(\frac{b_{0}}{b_{1}}\right)^{4},\quad a_{1}=18\left( \frac{b_{0}}{b_{1}}\right)^{3},\quad a_{2}=\frac{27}{2}\left(\frac{b_{0}}{b_{1 }}\right)^{2}.\]
The potential is singular at \(r=0\) and it grows quadratically for \(r\rightarrow\infty\). We have choices of the parameters such that the potential exhibits maxima and minima.
\[3v_{5}\tau^{6}+2v_{4}\tau^{5}+v_{3}\tau^{4}-v_{1}\tau^{2}-3v_{2}=0. \tag{17}\]
If \(v_{1}<0\) and \(v_{3},v_{4}>0\), we do not have any change of sign in and Descartes' rule of signs implies that the sextic equation will not have any positive real root, and therefore the potential has no extrema. If \(v_{1},v_{2}>0\) and \(v_{4}<0\), there are four changes of sign and the potential may have four, two or no positive critical point. For all other choices of the signs of the coefficients \(v_{1}\), \(v_{3}\), and \(v_{4}\) there are only two sign changes in and hence the potential will admit two positive critical points or none.
\[V_{5}(r)=u_{0}+\frac{u_{1}}{r^{2/3}}+\frac{u_{2}}{r^{2}}+u_{3}r^{2/3}+u_{4}r^{2} \tag{18}\]
where
\[u_{0}=-\frac{a_{1}}{b_{1}},\quad u_{1}=-a_{0}\sqrt{\frac{4}{9b_{1}^{2}}}, \quad u_{2}=v_{2},\quad u_{3}=-a_{2}\sqrt{\frac{4}{9b_{1}^{4}}},\quad u_{4 }=\frac{81}{16b_{1}^{2}}.\]
If we make the substitution \(r=\tau^{3/4}\), it is not difficult to verify that the critical points of the potential must satisfy the cubic equation
\[3u_{4}\tau^{3}+u_{3}\tau^{2}-u_{1}\tau-3u_{2}=0. \tag{19}\]
If \(a_{2}<0\) and \(a_{0}>0\), we do not have any change of sign in and Descartes' rule of signs implies that the cubic equation will not have any positive real root. Hence, the potential has no positive extrema. For any other choice of the signs of the parameters \(a_{2}\) and \(a_{1}\), we will always have two sign changes in and therefore the potential will admit two positive critical points or none.
## IV Solutions of the radial Schrodinger equation
We study the behaviour of the solutions of the radial Schrodinger equation for the general class of potentials and for the potentials discussed in the previous section.
\[\psi(r)=\sqrt{b_{0}+b_{1}\rho+b_{2}\rho^{2}}v(\rho),\]
It is in general convenient to bring into its canonical form with the help of the substitution
\[v(\rho)=e^{\frac{A_{2}}{3}\rho-\frac{e^{3}}{2}}y(\rho),\]
where the unknown function \(y\) must satisfy the equation
\[\frac{d^{2}y}{d\rho^{2}}-(\gamma+3\rho^{2})\frac{dy}{d\rho}+[\alpha+(\beta-3) \rho]=0 \tag{20}\]with
\[\alpha=A_{0}+\frac{A_{2}^{2}}{9},\quad\beta=A_{1},\quad\gamma=-\frac{2}{3}A_{2}. \tag{21}\]
Hence, the general solution of the radial Schrodinger equation can be cast in the form
\[\psi(r)=\sqrt{b_{0}+b_{1}\rho+b_{2}\rho^{2}}e^{\frac{A_{2}}{3}\rho-\frac{ \delta^{2}}{r}}y(\rho). \tag{22}\]
Note that the function multiplying \(y\) in decays exponentially in \(r\) as \(r\rightarrow+\infty\) because in the case \(b_{0},b_{1},b_{2}\neq 0\) we have
\[\rho(r)\approx\left\{\begin{array}{ll}\sqrt{\frac{2r}{\sqrt{b_{2}}}}&\mbox{ if }b_{2}>0\\ \rho_{2}+\frac{1}{\sqrt{|b_{2}|(\rho_{2}-\rho_{1})}}\left(\frac{3r}{2} \right)^{2/3}&\mbox{if }b_{2}<0.\end{array}\right.\]
with \(\rho_{1}\) and \(\rho_{2}\) denoting the roots of the quadratic polynomial \(I_{1}\) introduced in Section II whereas if \(b_{2}=0\) and \(b_{1}\neq 0\) we get
\[\rho(r)\approx\frac{1}{b_{1}}\left(\frac{3b_{1}r}{2}\right)^{2/3}.\]
Since equation has no finite singular points, we can construct Taylor series for the solutions of the triconfluent Heun equation by adopting the method outlined in.
\[\delta^{2}y-(1+\gamma\rho+3\rho^{3})\delta y+(\beta-3)\rho^{3}y=0.\]
Observing that \(\rho^{n}\) with \(n\in\mathbb{N}\) are eigenfunctions of the Euler operator, one finds the following two linearly independent solutions, namely
\[T_{1}(\alpha,\beta,\gamma;\rho)=\sum_{n=0}^{\infty}e_{n}(\alpha,\beta,\gamma) \rho^{n},\quad T_{2}(\alpha,\beta,\gamma;\rho)=\sum_{n=0}^{\infty}s_{n}(\alpha,\beta,\gamma)\rho^{n+1} \tag{23}\]
Behaviour of the potential for different values of the parameters.
with
\[e_{0}(\alpha,\beta,\gamma)=1,\quad e_{1}(\alpha,\beta,\gamma)=0,\quad e_{2}(\alpha,\beta,\gamma)=-\frac{\alpha}{2},\]
\[e_{n}(\alpha,\beta,\gamma)=\frac{(n-1)\gamma e_{n-1}(\alpha,\beta,\gamma)- \alpha e_{n-2}(\alpha,\beta,\gamma)-(\beta+6-3n)e_{n-3}(\alpha,\beta,\gamma)}{ n(n-1)},\quad n\geq 3\]
and
\[s_{0}(\alpha,\beta,\gamma)=1,\quad s_{1}(\alpha,\beta,\gamma)=\frac{\gamma}{2},\quad s_{2}(\alpha,\beta,\gamma)=\frac{\gamma^{2}-\alpha}{6},\]
\[s_{n}(\alpha,\beta,\gamma)=\frac{n\gamma s_{n-1}(\alpha,\beta,\gamma)-\alpha s _{n-2}(\alpha,\beta,\gamma)-(\beta+3-3n)s_{n-3}(\alpha,\beta,\gamma)}{n(n+1)},\quad n\geq 3.\]
We will refer to as the general solution to contrast it with the polynomial one which will be discussed below. Worth mentioning is also the convergence of as shown in.
In order to find polynomial solutions of we start AGAIN by assuming a power solution of the form
\[y(\rho)=\sum_{n=0}^{\infty}w_{n}\rho^{n}. \tag{24}\]
Substituting into we obtain the following recurrence relation for the coefficients \(w_{n}\)
\[(\beta-3n)w_{n-1}+\alpha w_{n}-\gamma(n+1)w_{n+1}+(n+1)(n+2)w_{n+2}=0 \tag{25}\]
The above recursion relation is the same as the recursion relation for \(e_{n}\) (see above) with a shift for \(n\). In order to have a polynomial solution of degree \(N\) we require that \(w_{n}=0\) for all \(n>N\). Moreover, for \(n=N+1\) the recurrence relation above gives the condition \(\beta=3(N+1)\).
\[E_{N}=\frac{3(N+1)-a_{1}}{b_{1}},\quad b_{1}\neq 0\]
Furthermore, for \(0\leq n\leq N\) the recursion formula generates a system of \(N+1\) homogenous linear equations for the coefficients \(w_{0},w_{1},\cdots,w_{N}\).
\[D_{N+1}=\begin{pmatrix}\alpha&-\gamma&2&0&0&\cdots&&&0\\ 3N&\alpha&-2\gamma&2\cdot 3&0&\cdots&&&\\ 0&3(N-1)&\alpha&-3\gamma&3\cdot 4&\cdots&&&\\ \vdots&\vdots&\vdots&\vdots&\vdots&\ddots&&&\vdots\\ &&&&3\cdot 3&\alpha&-(N-1)\gamma&N(N-1)\\ &&&&0&3\cdot 2&\alpha&-N\gamma\\ 0&&&\cdots&0&0&3\cdot 1&\alpha\end{pmatrix}\]
The first polynomials are given by
* \(N=0\): \(\beta=3\), \(\alpha=0\) and \(p_{0}(\rho)=1\);
* \(N=1\): \(\beta=6\), \(\alpha^{2}+3\gamma=0\) and \(p_{1}(\rho)=\rho-\alpha/3\);
* \(N=2\): \(\beta=9\), \(\alpha^{3}+12\alpha\gamma+36=0\) and \[p_{2}(\rho)=\rho^{2}-\frac{\alpha}{3}\rho+\frac{\alpha^{2}}{36}-\frac{1}{ \alpha};\]
* \(N=3\): \(\beta=12\), \(\alpha^{4}+30\alpha^{2}\gamma+216\alpha+81\gamma^{2}=0\) and \[p_{3}(\rho)=\rho^{3}-\frac{\alpha}{3}\rho^{2}+\left(\frac{\gamma}{2}+\frac{ \alpha^{2}}{18}\right)\rho-\frac{\alpha^{3}}{162}-\frac{7}{54}\alpha\gamma- \frac{2}{3};\]* \(\underline{N=4}\): \(\beta=15\), \(\alpha^{5}+60\alpha^{3}\gamma+756\alpha^{2}+576\alpha\gamma^{2}+5184\gamma=0\) and \[p_{4}(\rho)=\rho^{4}-\frac{\alpha}{3}\rho^{3}+\left(\frac{\alpha^{2}}{18}+\frac{ 2}{3}\gamma\right)\rho^{2}-\left(\frac{\alpha^{3}}{162}+\frac{5}{27}\alpha \gamma+\frac{4}{3}\right)\rho+\frac{\alpha^{4}}{1944}+\frac{2}{81}\alpha^{2} \gamma+\frac{5}{18}\alpha+\frac{\gamma^{2}}{9}.\]
Note that the above list extends the one given in.
\[-\frac{d^{2}v}{d\rho^{2}}+P(\rho)v=Ev,\quad v(-\infty)=v(+\infty)=0,\quad P( \rho)=\frac{9}{4}\rho^{4}-a_{2}\rho^{2}-a_{1}\rho-a_{0} \tag{26}\]
Note that the boundary condition is equivalent to the requirement \(v\in L^{2}(\mathbb{R})\). According to the spectrum of the above problem is discrete, all eigenvalues are real and simple and they can be arranged into an increasing sequence \(E_{0}<E_{1}<\cdots\).
\[E_{n}\approx\left[\frac{4\pi n}{2B(3/2,1/4)}\right]^{4/3}=\left[\frac{3\Gamma^ {2}(3/4)}{\sqrt{2\pi}}n\right]^{4/3},\]
Furthermore, all non-real zeros of the eigenfunctions of the problem belong to the imaginary axis (see Theorem 1 in).
We conclude this section by deriving a formula for the solution of the recurrence relation.
\[w_{n+3}=-\pi_{1}(n,\gamma)w_{n+2}-\pi_{2}(n,\alpha)w_{n+1}-\pi_{3}(n,\beta)w_{n} \tag{27}\]
with \(n\geq-2\) and
\[\pi_{1}(n,\gamma)=-\frac{\gamma}{n+3},\quad\pi_{2}(n,\alpha)=\frac{\alpha}{(n+ 2)(n+3)},\quad\pi_{3}(n,\beta)=\frac{\beta-3(n+1)}{(n+2)(n+3)},\quad w_{-2}=w_ {-1}=0.\]
By letting \(n=0\) we obtain \(w_{3}\) in terms of \(w_{2},w_{1}\), and \(w_{0}\), namely
\[w_{3}=\frac{\gamma}{3}w_{2}-\frac{\alpha}{2\cdot 3}w_{1}-\frac{\beta-3}{2\cdot 3 }w_{0}.\]
Furthermore, for \(n=-2\) and \(n=-1\) we get, respectively
\[w_{1}=\gamma w_{0},\quad w_{2}=\frac{1}{2}(\gamma^{2}-\alpha)w_{0}\]
from which it follows that
\[w_{3}=\frac{1}{3!}\left[\gamma^{3}-2\alpha\gamma-(\beta-3)\right]w_{0},\quad w _{4}=\frac{1}{4!}\left[\gamma^{4}-3\alpha\gamma^{2}+\alpha^{2}-3\gamma(\beta- 5)\right]w_{0}.\]
Hence, \(w_{n}\) can be evaluated once \(w_{0}\) is assigned.
\[w_{0}=1,\quad w_{1}=\gamma,\quad w_{2}=\frac{\gamma^{2}-\alpha}{2}. \tag{28}\]
Since our difference equation is linear and of third order, it will have a unique solution which can be expressed as a linear combination of three linear independent solutions \(w_{n}^{}\), \(w_{n}^{}\), and \(w_{n}^{}\).
\[W_{n}=\det\left(\begin{array}{ccc}w_{n}^{}&w_{n}^{}&w_{n}^{}\\ w_{n+1}^{}&w_{n+1}^{}&w_{n+1}^{}\\ w_{n+2}^{}&w_{n+2}^{}&w_{n+2}^{}\end{array}\right)\]does not vanish for any value of \(n\). On the other hand, Abel's lemma allows us to compute the Casoratian even without knowing explicitly the solutions \(w_{n}^{(i)}\) with \(i=1,2,3\). Applying (2.2.10) in we find
\[W_{n}=(-)^{3n}\prod_{k=0}^{n-1}\pi_{3}(k,\beta)W_{0}=\frac{2(-)^{3n}(n+2)}{(n+2 )!^{2}}(\beta-3)(\beta-6)\cdots(\beta-3n)W_{0}.\]
It is clear that we will have three linearly independent solutions provided that \(\beta\neq 3(k+1)\) for all \(k\in\mathbb{N}\) and \(W_{0}\neq 0\).
\[w_{n}=C_{1}w_{n}^{}+C_{2}w_{n}^{}+C_{3}w_{n}^{}\]
In order to solve the recurrence relation we first rewrite it as a system of first order equations of dimension 3, namely
\[Z_{n+1}=A(\alpha,\beta,\gamma;n)Z_{n} \tag{29}\]
with
\[Z_{n}=\left(\begin{array}{c}Z_{n}^{}\\ Z_{n}^{}\\ Z_{n}^{}\end{array}\right)=\left(\begin{array}{c}w_{n}\\ w_{n+1}\\ w_{n+2}\end{array}\right),\quad A(\alpha,\beta,\gamma;n)=\left(\begin{array}{ ccc}0&1&0\\ 0&0&1\\ -\pi_{3}(n,\beta)&-\pi_{2}(n,\alpha)&-\pi_{1}(n,\gamma)\end{array}\right),\]
Clearly, this matrix is non-singular if its determinant does not vanish and this condition translates into the requirement \(\beta\neq 3(n+1)\) for any \(n\in\mathbb{N}\).
\[Z_{0}=\left(\begin{array}{c}w_{1}\\ w_{1}\\ w_{2}\end{array}\right)=\left(\begin{array}{c}1\\ \gamma\\ \frac{\gamma^{2}-\alpha}{2}\end{array}\right).\]
By Theorem 3.4 in this initial value problem has a unique solution \(U_{n}\) such that \(U_{0}=Z_{0}\).
\[U_{1}=A(\alpha,\beta,\gamma;0)Z_{0},\quad U_{2}=A(\alpha,\beta,\gamma;1)Z_{1}= A(\alpha,\beta,\gamma;1)A(\alpha,\beta,\gamma;0)Z_{0}\]
and by induction we conclude that
\[U_{n}=\left(\prod_{k=0}^{n-1}A(\alpha,\beta,\gamma;k)\right)Z_{0},\]
where
\[\prod_{k=0}^{n-1}A(\alpha,\beta,\gamma;k)=\left\{\begin{array}{ll}A(\alpha, \beta,\gamma;n-1)A(\alpha,\beta,\gamma;n-2)\cdots A(\alpha,\beta,\gamma;0)& \mbox{if }n\geq 0\\ \mathbb{I}_{3}&\mbox{if }n=0.\end{array}\right..\]
Even though the above formula contains the solution of the recurrence relation, it is difficult to extract from it the behaviour of the solution \(w_{n}\) as \(n\to\infty\). The next result circumvents this problem by investigating the asymptotic behaviour of \(w_{n}\) with the help of the so-called Z-transform method which reduces the study of a linear difference equation to an examination of a corresponding complex function.
## Theorem IV.1
## For the recurrence relation with the initial condition it holds
\[\lim_{n\to\infty}w_{n}=0.\]
## Proof.
Applying the definition of the Z-transform to our recurrence relation, i.e.
\[\widetilde{x}(z)=Z(w_{n})=\sum_{j=0}^{\infty}\frac{w_{j}}{z^{j}},\quad z\in \mathbb{C}\backslash\{0\},\]we find with the help of the properties of the Z-transform listed in Ch. 6 in that the original recurrence relation gives rise to the following second order, linear, non homogeneous complex differential equation
\[z^{5}\frac{d^{2}\widetilde{x}}{dz^{2}}+z(2z^{3}+\gamma z^{2}+3)\frac{d\widetilde {x}}{dz}+(\alpha z+\beta-3)\widetilde{x}=(2w_{2}-\gamma w_{1}+\alpha w_{0})z. \tag{30}\]
The initial condition implies that \(2w_{2}-\gamma w_{1}+\alpha w_{0}=0\) and therefore, we are left with the following homogeneous differential equation
\[z^{5}\frac{d^{2}\widetilde{x}}{dz^{2}}+z(2z^{3}+\gamma z^{2}+3)\frac{d \widetilde{x}}{dz}+(\alpha z+\beta-3)\widetilde{x}=0. \tag{31}\]
By means of the substitution \(\omega=1/z\) the above equation can be transformed into the triconfluent Heun equation. Hence, a particular solution is given by \(\widetilde{x}_{1}(\omega)=T_{1}(\alpha,\beta,\gamma;\omega)\) with \(T_{1}\) defined as in. In order to construct a linearly independent set we make use of Preposition 4.2 with \(j=1\) in.
\[\gamma T_{2}(\alpha,\beta,\gamma;\omega)+T_{1}(\alpha,\beta,\gamma;\omega)=e^{ \omega^{3}+\gamma\omega}T_{1}(\alpha,-\beta,\gamma;-\omega)\]
Hence, the general solution of for \(\gamma\neq 0\) is
\[\widetilde{x}(z)=C_{1}T_{1}(\alpha,\beta,\gamma;1/z)+C_{2}e^{\frac{\gamma z^{ 2}+1}{z^{3}}}T_{1}(\alpha,-\beta,\gamma;-1/z),\]
Then, the Final Value Theorem for the Z-transform implies that
\[\lim_{n\to\infty}w_{n}=\lim_{z\to 1}(z-1)\widetilde{x}(z)=0.\]
If \(\gamma=0\), we can still use the particular solution \(\widetilde{x}_{1}\) found before together with Abel's formula. Let \(\widetilde{x}_{2}\) denote another particular solution of the triconfluent Heun equation arising from after the transformation \(\omega=1/z\). Then, the Wronskian is given by \(W(T_{1}(\alpha,\beta,0;\omega);\widetilde{x}_{2}(\omega))=C\text{exp}(\omega^ {3})\) for some \(C\neq 0\).
\[\frac{d\widetilde{x}_{2}}{d\omega}-\frac{T^{{}^{\prime}}_{1}(\alpha,\beta,0; \omega)}{T_{1}(\alpha,\beta,0;\omega)}\widetilde{x}_{2}=C\frac{e^{\omega^{3}} }{T_{1}(\alpha,\beta,0;\omega)}.\]
The corresponding integrating factor is given by \(\mu(\omega)=T_{1}^{-1}(\alpha,\beta,0;\omega)\) and hence
\[\widetilde{x}_{2}(\omega)=T_{1}(\alpha,\beta,0;\omega)\int\frac{e^{\omega^{3} }}{T_{1}^{2}(\alpha,\beta,0;\omega)}\;d\omega.\]
Finally, the general solution of for \(\gamma=0\) is represented by
\[\widetilde{x}(z)=T_{1}(\alpha,\beta,0;1/z)\left(D_{1}+D_{2}\int\frac{e^{1/z^ {3}}}{z^{2}T_{1}^{2}(\alpha,\beta,0;1/z)}\;dz\right).\]
Taking into account that the following expansion holds for \(|z-1|<1\)
\[\int\frac{e^{1/z^{3}}}{z^{2}T_{1}^{2}(\alpha,\beta,0;1/z)}\;dz=\frac{e(z-1)}{T _{1}^{2}(\alpha,\beta,0;1)}+\mathcal{O}(z-1)^{2},\]
\(\Box\)
We conclude the analysis of the recurrence relation by constructing asymptotic expansions valid as \(n\to\infty\) consisting of an exponential leading term multiplied by a descending series. These kind of expansions are called Birkhoff series.
\[w_{n}=\mathcal{E}_{n}K_{n},\quad\mathcal{E}_{n}=e^{\mu_{0}n\ln n+\mu_{1}n}n^{ \theta},\quad K_{n}=e^{\alpha_{1}n^{\widetilde{\beta}}+\alpha_{2}n^{\widetilde {\beta}-1/\widetilde{\rho}}+\cdots} \tag{32}\]
Then, for \(k=1,2,3\) we have
\[\frac{w_{n+k}}{w_{n}}=n^{\mu_{0}k}\lambda^{k}\left[1+\frac{A}{n}+\frac{B}{n^{2 }}+\frac{C}{n^{3}}+\cdots\right]e^{\alpha_{1}\widetilde{\beta}kn^{\widetilde {\beta}-1}+\alpha_{2}\left(\widetilde{\beta}-\frac{1}{\widetilde{\rho}} \right)kn^{\widetilde{\beta}-(1/\widetilde{\rho})-1}+\cdots},\quad\lambda=e^{ \mu_{0}+\mu_{1}}. \tag{33}\]where
\[A = k\theta+\frac{\mu_{0}k^{2}}{2},\] \[B = \frac{\mu_{0}^{2}k^{4}}{8}-\frac{\mu_{0}k^{3}}{6}+\frac{1}{2}\mu_{0 }k^{3}\theta+\frac{k^{2}}{2}\theta(\theta-1),\] \[C = \frac{\mu_{0}^{3}k^{6}}{48}-\frac{\mu_{0}^{2}k^{5}}{12}+\frac{\mu_ {0}k^{4}}{12}+\theta k\left(\frac{\mu_{0}^{2}k^{4}}{8}-\frac{\mu_{0}k^{3}}{6} \right)+\frac{\mu_{0}k^{4}}{4}\theta(\theta-1)+\frac{k^{3}}{6}\theta(\theta-1)( \theta-2).\]
To construct Birkhoff series for the recurrence relation it turns out to be convenient to rewrite it as
\[(\beta-3-3n)w_{n}+\alpha w_{n+1}-\gamma(n+2)w_{n+2}+(n+2)(n+3)w_{n+3}=0.\]
Substituting and into the above equation, we find the equation to be formally satisfied is
\[\beta-3-3n+\alpha\lambda n^{\mu_{0}}\left[1+\frac{\theta+\mu_{0}/2}{n}+ \mathcal{O}(n^{-2})\right]e^{\alpha_{1}\widetilde{\beta}n^{\widetilde{\beta}- 1}+\alpha_{2}\left(\widetilde{\beta}-\frac{1}{p}\right)n^{\widetilde{\beta}-(1 /\widetilde{\beta})-1}+\cdots}\]
\[-\gamma\lambda^{2}n^{2\mu_{0}+1}\left[1+\frac{2(\theta+\mu_{0}+1)}{n}+ \mathcal{O}(n^{-2})\right]e^{2\alpha_{1}\widetilde{\beta}n^{\widetilde{\beta}- 1}+2\alpha_{2}\left(\widetilde{\beta}-\frac{1}{p}\right)n^{\widetilde{\beta}-( 1/\widetilde{\beta})-1}+\cdots}\]
\[+\lambda^{3}n^{3\mu_{0}+2}\left[1+\frac{3\theta+\frac{9}{2}\mu_{0}+5}{n}+ \mathcal{O}(n^{-2})\right]e^{3\alpha_{1}\widetilde{\beta}n^{\widetilde{\beta}- 1}+3\alpha_{2}\left(\widetilde{\beta}-\frac{1}{p}\right)n^{\widetilde{\beta}-( 1/\widetilde{\beta})-1}+\cdots}=0.\]
Obviously, it must be \(\mu_{0}=-1/3\) and this implies \(\lambda=3^{1/3}\). This further requires that \(\widetilde{\rho}=3\) and \(\widetilde{\beta}=1/3\).
\[(3\alpha_{1}-3^{2/3}\gamma)n^{1/3}+\beta-3+3\left(3\theta+\frac{7}{2}\right)+ \mbox{(lower order terms)}=0.\]
Thus, \(\alpha_{1}=\gamma/3^{1/3}\) and \(\theta=-5/6-\beta/9\).
\[w_{n}^{}\sim\left(\frac{3e}{n}\right)^{n/3}n^{-\frac{5}{6}-\frac{\beta}{9}} e^{\gamma\sqrt{\frac{7}{3}}}\left[1+\frac{c_{1}}{n^{1/3}}+\frac{c_{2}}{n^{2/3}}+ \frac{c_{3}}{n}+\mathcal{O}(n^{-4/3})\right]\]
with
\[c_{1} = \frac{1}{\sqrt{9}}\left(\alpha-\frac{\gamma^{2}}{6}\right),\] \[c_{2} = \frac{1}{2\sqrt{9}}\left[\gamma\left(1-\frac{\beta}{3}\right) +\left(\alpha-\frac{\gamma^{2}}{6}\right)^{2}\right],\] \[c_{3} = \frac{1}{324}\left[-108c_{1}^{3}+324c_{1}c_{2}+(8\sqrt{1089}-3 6\sqrt{9})\gamma c_{1}+6\beta^{2}+18\beta-2\gamma^{3}-81\right].\]
According to other two formal series solutions can be obtained by letting \(n\to ne^{2\pi ki}\) with \(k=0,1,\cdots,\widetilde{\rho}-1\).
\[w_{n}^{} \sim \left(\frac{3e}{n}\right)^{n/3}n^{-\frac{5}{6}-\frac{\beta}{9}}e^ {-\frac{2}{3}\pi in-\frac{1}{2}(1-i\sqrt{3})\gamma}\sqrt{\frac{7}{3}}\left[ 1-\frac{(1+i\sqrt{3})c_{1}}{2n^{1/3}}-\frac{(1-i\sqrt{3})c_{2}}{2n^{2/3}}+ \frac{c_{3}}{n}+\mathcal{O}(n^{-4/3})\right],\] \[w_{n}^{} \sim \left(\frac{3e}{n}\right)^{n/3}n^{-\frac{5}{6}-\frac{\beta}{9}}e ^{-\frac{4}{3}\pi in-\frac{1}{2}(1+i\sqrt{3})\gamma}\sqrt{\frac{7}{3}} \left[1-\frac{(1-i\sqrt{3})c_{1}}{2n^{1/3}}-\frac{(1+i\sqrt{3})c_{2}}{2n^{2/3 }}+\frac{c_{3}}{n}+\mathcal{O}(n^{-4/3})\right].\]
Last but not least, note that the above asymptotic results for \(w_{n}^{(i)}\) with \(i=1,2,3\) tend to zero for \(n\rightarrow\infty\) as we would expect from Theorem IV.1.
## V Analysis of the triconfluent Heun equation using supersymmetric quantum mechanics
For the following analysis it is convenient to rewrite the triconfluent Heun equation as \((L_{T}v)(\rho)=0\) where
\[L_{T}=-\frac{d^{2}}{d\rho^{2}}+\Omega(\rho),\quad\Omega(\rho)=-A_{0}-A_{1}\rho- A_{2}\rho^{2}+\frac{9}{4}\rho^{4}.\]
We look for a factorization of the operator \(L_{T}\) having the form
\[L_{T}=AA^{\dagger}=\left(\frac{d}{d\rho}+W(\rho)\right)\left(-\frac{d}{d\rho}+ W(\rho)\right).\]
This will be the case if the unknown function \(W\) satisfies the Riccati equation
\[\frac{dW}{d\rho}+W^{2}(\rho)=\Omega(\rho).\]
We solved the above equation with the help of the software package Maple 18 and we found that the superpotential \(W\) is given by
\[W(\rho)=\frac{1}{6}\frac{e^{f(\rho)}F(\alpha,\beta,\gamma;\rho)-c_{1}e^{-f( \rho)}F(\alpha,-\beta,\gamma;-\rho)}{e^{f(\rho)}T_{1}(\alpha,\beta,\gamma;\rho )+c_{1}e^{-f(\rho)}T_{1}(\alpha,-\beta,\gamma;-\rho)}\]
where \(c_{1}\) is an arbitrary integration constant and
\[f(\rho)=-\frac{\gamma}{2}\rho-\frac{\rho^{3}}{2},\quad F(\alpha,\beta,\gamma; \rho)=-3(\gamma+3\rho^{2})T_{1}(\alpha,\beta,\gamma;\rho)+6\hat{T}_{1}(\alpha, \beta,\gamma;\rho).\]
Note that as a byproduct result we also obtained a new factorization of the triconfluent Heun equation since it does not coincide with the one offered in.
Combining the operators \(A\) and \(A^{\dagger}\) we can construct Hamiltonians
\[\mathcal{H}_{-}=A^{\dagger}A=-\frac{d^{2}}{d\rho^{2}}+V_{-}(\rho),\quad \mathcal{H}_{+}=AA^{\dagger}=L_{T}=-\frac{d^{2}}{d\rho^{2}}+V_{+}(\rho)\]where \(V_{+}\) and \(V_{-}\) are the supersymmetric partner potentials and they are given by
\[V_{\pm}(\rho)=W^{2}(\rho)\pm\frac{dW}{d\rho}.\]
Clearly, \(V_{+}(\rho)=\Omega(\rho)\) and moreover the other partner potential can be expressed as
\[V_{-}(\rho)=2W^{2}(\rho)-\Omega(\rho)=\Omega(\rho)-2\frac{dW}{d\rho}.\]
If we denote the eigenstates of the Hamiltonians \({\cal H}_{-}\) and \({\cal H}_{+}\) by \(v_{n}^{(-)}\) and \(v_{n}^{(+)}\), respectively, then \(v_{n}^{(\pm)}\) must satisfy the eigenvalue equations
\[{\cal H}_{-}v_{n}^{(-)}=E_{n}^{(-)}v_{n}^{(-)},\quad{\cal H}_{+}v_{n}^{(+)}=E_{ n}^{(+)}v_{n}^{(+)}.\]
In order to establish whether or not the supersymmetry is unbroken we need to find out if the ground state of \({\cal H}_{-}\) has zero energy, that is \(E_{0}^{(-)}=0\).
## VI Analysis of the zero energy state
Let \(\psi_{0}\) denote the zero energy solution, i.e. \(E=0\), associated to an Hamiltonian \(H_{0}\) such that \(H_{0}\psi_{0}=0\) and \(H_{0}=d^{2}/dr^{2}+V_{eff}(r)\) where the effective potential has been already computed in Section II.
\[\psi_{0}(r)=\sqrt{r}\left[C_{1}J_{\sqrt{7}/6}\left(\frac{2\sqrt{e_{4}}}{3}r^{3 /2}\right)+C_{2}Y_{\sqrt{7}/6}\left(\frac{2\sqrt{e_{4}}}{3}r^{3/2}\right)\right]\]
This solution is not square integrable since it oscillates asymptotically as a plane wave. Therefore, it is obvious that we are handling here the scattering solution for a spacial case of the energy. Note that from the definition of \(e_{4}\) we can conclude that it is always strictly positive. Another case that can be solved exactly is the following. Consider the potential \(V_{4}\) and let \(a_{0}=a_{2}=b_{0}=0\). Then, the coefficients \(v_{1}\), \(v_{3}\), and \(v_{4}\) vanishes whereas \(v_{0}=-a1/b_{1}\).
\[V_{4}(r)=-\frac{5}{36r^{2}}+v_{5}r^{2}+v_{0}\]with \(v_{5}>0\).
\[\frac{d^{2}\psi}{dr^{2}}+V_{4}(r)\psi(r)=E\psi(r)\]
can be exactly solved in terms of a linear combination of Whittaker functions given by
\[\psi_{E}(r)=\frac{1}{\sqrt{r}}\left[C_{1}M_{\alpha,\beta}(i\sqrt{v_{5}}r^{2})+C _{2}W_{\alpha,\beta}(i\sqrt{v_{5}}r^{2})\right]\]
with
\[\alpha=\frac{i(E-v_{0})}{4\sqrt{5}},\quad\beta=\frac{\sqrt{14}}{12}.\]
Observe that if \(E=v_{0}\) or \(E=0\) and \(v_{0}=0\) we find that
\[\psi_{E=v_{0}}(r)=\sqrt{r}\left[C_{1}J_{\beta}\left(\frac{\sqrt{v_{5}}}{2}r^{2 }\right)+C_{2}Y_{\beta}\left(\frac{\sqrt{v_{5}}}{2}r^{2}\right)\right].\]
|
10.48550/arXiv.1412.3847
|
Potentials of the Heun class: the triconfluent case
|
D. Batic, D. Mills-Howell, M. Nowakowski
| 5,780
|
10.48550_arXiv.1211.4884
|
###### Abstract
We show that classical molecular density functional theory (MDFT), here in the homogeneous reference fluid approximation in which the functional is inferred from the properties of the bulk solvent, is a powerful new tool to study, at a fully molecular level, the solvation of complex surfaces and interfaces by polar solvents. This implicit solvent method allows for the determination of structural, orientational and energetic solvation properties that are on a part with all-atom molecular simulations performed for the same system, while reducing the computer time by two orders of magnitude. This is illustrated by the study of an atomistically-resolved clay surface composed of over a thousand atoms wetted by a molecular dipolar solvent. The high numerical efficiency of the method is exploited to carry a systematic analysis of the electrostatic and non-electrostatic components of the surface-solvent interaction within the popular CLAYFF force field. Solvent energetics and structure are found to depend weakly upon the atomic charges distribution of the clay surface, even for a rather polar solvent. We conclude on the consequences of such findings for force-field development.
pacs: 61.20.Gy, 61.20.Ja, 61.25.Em, 68.03.Hj, 68.08.-p, 68.08.De
## I Introduction
Liquids at solid-liquid interfaces behave differently from bulk liquids. Interfacial phenomena play key roles in various applications like, for instance, heterogeneous catalysis, electrochemistry, adsorption and transport in porous media. It is our goal to improve the description of such interfaces at the microscopic scale.
Experimentally, specific techniques have been developed, such as the vibrational sum frequency spectroscopy or quasi-elastic neutron scattering techniques, that now give access to information on the structural and dynamic properties of the liquid on few molecular layers on top of the surface. At these scales, theoretical modeling can clarify experimental observations by linking them to atomistic phenomena. For instance, Rotenberg _et al._ showed by molecular simulations that wetting properties of talc surfaces, that behave as hydrophilic or hydrophobic depending on the relative humidity, are a consequence of the competition between surface-solvent adhesion, favorable entropy in the gas phase and molecular cohesion in the liquid phase.
This kind of theoretical modeling is typically based on molecular dynamics (MD) or Monte Carlo simulations where each atom is considered individually. For systems that are intrinsically multiscale, e.g. porous media, one has to deal with thousands of atoms. At this point, numerical heaviness becomes critical and hinders systematic studies. To overcome this difficulty, various mesoscale models have been proposed, that are less computationally demanding. One can for instance forget the molecular nature of the solvent to only account for a Polarizable Continuum Medium (PCM), or simplify it to a hard sphere, with or without dipole. Coarse-grained strategies have also been successfully applied to solid-liquid interfaces. While these methods and others proved efficient on a panel of problems, the applicability of their underlying approximation to an unknown system remains to be carefully and systematically checked.
In the second half of the last century, liquid state theories like the classical density functional theory (DFT) have blossomed. Classical DFT, like other methods such as integral equations or classical field theories have shown to give thermodynamic and structural results comparable to all-atom simulations at attractive numerical cost, but they have for now been limited to highly symmetric systems, bulk fluids or simple model interfaces, e.g. hard and soft walls. A current challenge lies in the development of three-dimensional theories and implementations to describe molecular liquids, solutions, mixtures, in complex environments such as biomolecular media or atomically-resolved surfaces and interfaces. Recent developments in this direction use lattice field or Gaussian field theoretical approaches, or the 3D reference interaction site model (3D-RISM), an appealing integral equation theory that has proven recently to be applicable to, e.g., structure prediction in complex biomolecular systems. Integral equations are, however, restricted by the choice of a closure relation (typically HyperNetted Chain (HNC), Percus-Yevick, or Kovalenko-Hirata approximation), and despite their great potential, they remain difficult to control and improve, and can prove difficult to converge.
Recently, a molecular density functional theory (MDFT) approach to solvation has been introduced. It relies on the definition of a free-energy functional depending on the full six-dimensional position and orientation solvent density. In the so-called homogeneous reference fluid (HRF) approximation, the (unknown) functional can be inferred from the properties of the bulk solvent. Compared to reference molecular dynamics calculations, such approximation was shown to be accurate for polar, non-protic fluids, but to require corrections for water. Until now, only simple three-dimensional solutes have been considered in this framework including ions and polar or apolar molecules. In ref., the authors have shown that the MDFT scheme was applicable to the computation of electron-transfer properties such as reaction free energies and solvent reorganization free energies. In this paper, we show how the most recent developments of the molecular density functional theory, described here in the homogeneous reference fluid approximation (HRF-MDFT), can be implemented to study the solvation of a complex atomically-resolved clay surface, the pyrophyllite, that contains over a thousand atoms. In order to uncouple the difficulties intrinsic to the method to those due to the particular correlations in water, we restrict ourselves here to a generic model of dipolar solvent, the Stockmayer fluid - with eventually parameters adjusted to mimick a few properties of water.
In section II, we present the most recent developments in molecular density functional theory in the homogeneous reference fluid approximation. We also describe our reference molecular dynamics simulations and the system under investigation. In section III, we discuss the structural and orientational properties of the complex interface between the solid and the dipolar liquid. There, we compare HRF-MDFT with reference molecular dynamics results. Finally, in section III.3, we illustrate the possibilities offered by the method in terms of numerical efficiency to analyze the electrostatic and non-electrostatic components of the state-of-the-art force field for clays, CLAYFF. We conclude on the consequences of our analysis on force field development.
## II Methods and Model System
### Molecular Density Functional Theory of Solvation
#### ii.1.1 Theoretical Aspects
Recent developments of a molecular density functional theory (MDFT) unlocked the study of the solvation of three-dimensional molecular solutes in arbitrary solvents using classical density functional theory. In MDFT, the solvent is composed of rigid molecules and described in terms of a position and orientation density, \(\rho\left(\mathbf{r},\Omega\right)\).
\[F[\rho]=\Phi\left[\rho\right]-\Phi[\rho_{0}], \tag{1}\]
Following the theoretical framework introduced by Evans, the density functional \(F\) can be rewritten as the sum of an ideal contribution (\(F_{id}\)), an excess term (\(F_{exc}\)) and an external (solute) contribution (\(F_{ext}\)),
\[F[\rho]=F_{id}[\rho]+F_{ext}[\rho]+F_{exc}[\rho]. \tag{2}\]
The ideal and external contributions \(F_{id}\) and \(F_{ext}\) can be formally and exactly expressed as
\[F_{id}[\rho] = k_{B}T\int d\mathbf{r}d\Omega[\rho\left(\mathbf{r},\Omega\right) \ln\left(\frac{\rho\left(\mathbf{r},\Omega\right)}{\rho_{0}}\right) \tag{3}\] \[-\rho\left(\mathbf{r},\Omega\right)+\rho_{0}],\] \[F_{ext}[\rho] = \int d\mathbf{r}d\Omega\ V_{ext}(\mathbf{r},\Omega)\rho\left( \mathbf{r},\Omega\right), \tag{5}\]
\(V_{ext}\) is the sum of the electrostatic and Lennard-Jones interactions between the solute and one solvent molecule located at \(\mathbf{r}\) with orientation \(\Omega\):
\[V_{ext}(\mathbf{r},\Omega) = \sum_{j\in\text{solvent}}\left\{q_{j}\,V_{q}(\mathbf{r}_{j})\right. \tag{6}\] \[+\left.\sum_{i\in\text{solute}}4\epsilon_{ij}\left[\left(\frac{ \sigma_{ij}}{r_{ij}}\right)^{12}-\left(\frac{\sigma_{ij}}{r_{ij}}\right)^{6} \right]\right\}.\]
If \(\mathbf{R}(\Omega)\) is the rotation matrix associated to \(\Omega\) and \(\mathbf{s}_{j}\) is the position of the solvent site \(j\) in the molecular frame, then \(\mathbf{r}_{j}=\mathbf{r}+\mathbf{R}(\Omega)\mathbf{s}_{j}\) denotes its absolute position in space and \(\mathbf{r}_{ij}=\mathbf{r}_{j}-\mathbf{r}_{i}\) its relative position with respect to the solute site \(i\) located at \(\mathbf{r}_{i}\). \(\epsilon_{ij}\), \(\sigma_{ij}\) are the Lennard-Jones parameters between solute site \(i\) and solvent site \(j\). \(q_{j}\) is the partial charge carried by site \(j\) and \(V_{q}(\mathbf{r}_{j})\) is the electrostatic potential created by the solute at \(\mathbf{r}_{j}\).
The excess functional is unknown, but can be expressed formally as
\[F_{exc}[\rho] = -\frac{1}{2}k_{B}T\iint d\mathbf{r}_{1}d\mathbf{r}_{2}d\Omega_{1 }d\Omega_{2}\Delta\rho(\mathbf{r}_{1},\Omega_{1}) \tag{7}\] \[\times c(\mathbf{r}_{2}-\mathbf{r}_{1},\Omega_{1},\Omega_{2})\, \Delta\rho(\mathbf{r}_{2},\Omega_{2})\] \[+F_{B}[\Delta\rho],\]
The first term represents the homogeneous reference fluid approximation where the excess free-energy density is written in terms of the angular-dependent direct correlation of the pure solvent. It amounts to a second order Taylor expansion of the excess free-energy functional around the homogeneous solvent. It is equivalent to the HNC approximation in integral equation theories when the solute is taken as a solvent particle. The second term represents the unknown correction to that term (or bridge term) that can be expressed as of a systematic expansion of the solvent correlations in terms of the three-body,\(\ldots\) n-body terms direct correlation functions. We will consider below the case \(F_{B}=0\). This approximation was shown in Ref. to be accurate for several polar aprotic solvents. Difficulties were found, however, for water, certainly the most interesting but also the most complex solvent, in part due to its high-order correlations, if not interactions. We have proposed several corrections for water entering in the definition of \(F_{B}\): an empirical three-body correlation term inspired by the water model by Molinero _et al._, and a bridge function extracted from a hard sphere functional. The first one enforces the missing tetrahedral order, while the second one introduces the \(N\)-body repulsion terms (\(N>2\)) of a hard-sphere fluid.
Since water adds some extra complexity in the functional that deserve to be tackled separately, it is not our purpose to study hydration here, but rather to show the advantages/disadvantages of MDFT and its practical implementation for the molecular solvation of complex atomically-resolved surfaces. We will thus consider a generic dipolar fluid, the Stockmayer model, with parameters adjusted to mimick a few properties of water (particle size, molecular dipole and dielectric constant), but no hydrogen bonds, and certainly not a "drinkable" water.
For dipolar models, each orientation \(\Omega\) can be described by the normalized orientation vector \(\mathbf{\Omega}\) and the direct correlation function in Eq. 7 may be developed on a basis of three rotational invariants
\[c\left(\mathbf{r_{12}},\mathbf{\Omega_{1}},\mathbf{\Omega_{2}}\right) =c_{S}\left(r_{12}\right)\Phi^{100}\] \[+c_{\Delta}\left(r_{12}\right)\Phi^{110}\left(\mathbf{\Omega_{1} },\mathbf{\Omega_{2}}\right)\] \[+c_{D}\left(r_{12}\right)\Phi^{112}\left(\mathbf{\Omega_{1}}, \mathbf{\Omega_{2}}\right), \tag{8}\]
where
\[\Phi^{100} =1, \tag{9}\] \[\Phi^{110} =\mathbf{\Omega_{1}}\cdot\mathbf{\Omega_{2}},\] \[\Phi^{112} =3\left(\mathbf{\Omega_{1}}\cdot\mathbf{\hat{r}_{12}}\right) \left(\mathbf{\Omega_{2}}\cdot\mathbf{\hat{r}_{12}}\right)-\mathbf{\Omega_{1} }\cdot\mathbf{\Omega_{2}}, \tag{11}\]
If we also define the molecular density \(n\left(\mathbf{r}\right)\) and the polarization density \(\mathbf{P}\left(\mathbf{r}\right)\)
\[n\left(\mathbf{r}\right) =\int\rho\left(\mathbf{r},\mathbf{\Omega}\right)d\mathbf{\Omega}, \tag{12}\] \[\mathbf{P}\left(\mathbf{r}\right) =\int\mathbf{\Omega}\cdot\rho\left(\mathbf{r},\mathbf{\Omega} \right)d\mathbf{\Omega}, \tag{13}\]
the excess free-energy density functional \(F_{exc}\left[\rho\right]\) can be rewritten as a functional of \(n\) and \(\mathbf{P}\) instead of \(\rho\left(\mathbf{r},\Omega\right)\):
\[F_{exc}\left[\rho\left(\mathbf{r},\mathbf{\Omega}\right)\right] =-\frac{1}{2}k_{B}T\int d\mathbf{r}_{1}d\mathbf{r}_{2}\{c_{S}(r _{12})\,\Delta n(\mathbf{r}_{1})\,\Delta n(\mathbf{r}_{2})\] \[-c_{\Delta}(r_{12})\,\mathbf{P}(\mathbf{r}_{1})\cdot\mathbf{P}( \mathbf{r}_{1}) \tag{14}\] \[-c_{D}(r_{12})[3(\mathbf{P}(\mathbf{r}_{1})\cdot\mathbf{\hat{r} _{12}})(\mathbf{P}(\mathbf{r}_{2})\cdot\mathbf{\hat{r}_{12}})\] \[-\mathbf{P}(\mathbf{r}_{1})\cdot\mathbf{P}(\mathbf{r}_{1})]\}.\]
This significantly increases numerical efficiency, as will be discussed below. The same reduction is true for the ideal and external contributions if the solute-solvent electrostatic interaction is strictly restricted to charge-dipole interactions. Even in that case, however, it remains advantageous, both for convergence reasons and for keeping the generality of the code, to stick to the expression of the ideal free-energy in terms of the angular distribution \(\rho\left(\mathbf{r},\mathbf{\Omega}\right)\), eq. 4, and to minimize the functional in the full position-angle space. This is now described.
#### ii.1.2 Numerical Aspects of HRF-MDFT
Here, we give insight into numerical details associated with the variational minimization of the density functional \(F\), _i.e._ the resolution of the Euler-Lagrange equation
\[\frac{\delta F[\rho]}{\delta\rho}=k_{B}T\ln\left(\frac{\rho\left(\mathbf{r}, \mathbf{\Omega}\right)}{\rho_{0}}\right)+V_{ext}\left(\mathbf{r},\mathbf{ \Omega}\right)+\frac{\delta F_{exc}[\rho]}{\delta\rho}=0. \tag{15}\]
The inhomogeneous density \(\rho\) is projected onto an orthorhombic position grid of \(N_{x}\times N_{y}\times N_{z}\) nodes with periodic boundary conditions. To each node is associated an angular grid on which is discretized the orientation vector \(\mathbf{\Omega}\). The variational density \(\rho\left(\mathbf{r},\mathbf{\Omega}\right)\) which minimizes \(F\) is optimized numerically by the limited memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) method as implemented by Byrd, Lu, Nocedal and Zhu. This quasi-Newton algorithm only requires the knowledge of the functional \(F\) and its first derivative with respect to the density at each node, which are known analytically. The Hessian, needed in Newton-derived algorithms, is approximated using gradients at previous self-consistent iterations. This makes a notable difference with other optimizers: faster than Piccard iterations, without requiring the second variation of the functional with respect to the density as typically required by pseudo-Newton algorithms. Convolutions in \(F_{exc}\) are calculated in reciprocal space by fast Fourier transforms (FFT) as implemented in the FFTW3 library. Rewriting \(F\) as a functional of \(n\) and \(\mathbf{P}\) reduces considerably the number of angular summations and of FFTs to be performed. As a summary, the high performance of our three-dimensional HRF-MDFT is due to (\(i\)) a quasi-Newton implementation, (\(ii\)) the calculation of convolutions in reciprocal space, associated to the extreme performance of FFTW3, (\(iii\)) the rewriting of the functional in a form optimal for our purpose.
For the given solvent molecular model, the direct correlation function of the solvent is extracted from all-atom simulations from which are generated the pair distribution functions. The corresponding correlation function can then be deduced by solving the Ornstein-Zernike equation. This can be done in Fourier space but raises numerical issues at large \(r\), _i.e._ at small \(k\), the conjugate of \(r\). The direct space method of Baxter combined with the variational method of Dixon and Hutchinsonwas used accordingly. This method enjoins that \(c\) vanishes beyond a radius, set to 8.7 A in the present system. The projection of the direct correlation function of the Stockmayer fluid on rotational invariants \(c_{S}\), \(c_{\Delta}\) and \(c_{D}\) as expressed in Eqs. 8 to 11 are plotted as a function of \(k\) in Note that the exact bridge function of the Stockmayer fluid from explicit Monte Carlo simulation data has been published recently by Puibasset and Belloni.
The external potential \(V_{exc}(\mathbf{r},\mathbf{\Omega})\) is computed only once, at the beginning of the simulation: At each node, an isolated solvent molecule is considered with a given molecular orientation, for which is tabulated the total interaction energy between the solvent molecule sites and the solute sites, according to Eq. 6. To accelarate the computation, the value of the electrostatic potential at each solvent site is interpolated from its values on the grid. The grid electrostatic potential itself is obtained by first extrapolating the solute charge density on the grid and then solving the resulting Poisson equation by FFT's. The calculation of the Lennard-Jones external potential takes advantage of cut-off distances for each solute site.
In this work, we used typically \(4\times 4\times 4\) nodes per A\({}^{3}\), which, for the system considered below, corresponds to roughly \(150^{3}\) 3D-grid points. The molecular orientation \(\mathbf{\Omega}\) was discretized over 18 angles and integrated over the whole sphere by Legendre quadratures. This implies \(4\times 4\times 4\times 18=1152\) variables per A\({}^{3}\) to be optimized. All results presented here were carefully checked with respect to the number of nodes per A\({}^{3}\) and per orientation. The convergence of the results as a function of grid resolution can be quantified as follows: with respect to calculations with a very fine grid resolution of \(5^{3}\) grid points per A\({}^{3}\), the relative difference in solvation free energy amounts to 1.2 % with \(2^{3}\) grid points per A\({}^{3}\) and 0.3 % with \(3^{3}\) grid points per A\({}^{3}\). It is less than 0.1 % with \(4^{3}\) grid points per A\({}^{3}\). Values given above are found to be adequate for all observables presented in this paper. In order to illustrate the high efficiency of the method, the iterative convergence of \(F\) with iteration steps is illustrated in for the grid described above. After 15 iterations, the convergence in solvation free energy is of the order of \(10^{-5}\). We also illustrate in the CPU time per iteration for 18 angles per node and \(2^{3}\), \(3^{3}\), \(4^{3}\) and \(5^{3}\) nodes per A\({}^{3}\) for our model system containing several hundred atoms. One observes a linear scaling in \(N_{x}\times N_{y}\times N_{z}\), which leads to convergence in less than 15 minutes for the \(4\times 4\times 4\) grid per A\({}^{3}\) described above on an ordinary laptop without parallelization.
### Molecular Dynamics
Molecular dynamics simulations were generated for the same surface and solvent molecular models in order to compare with the MDFT results. All simulations were done in the canonical NVT ensemble, with a Nose-Hoover thermostat. After a phase of equilibration, all structural quantities were collected and averaged over 5 ns.
Projections \(c_{S}\) (black line), \(c_{\Delta}\) (red dashed line) and \(c_{D}\) (blue dotted line) of the direct correlation function of the Stockmayer fluid on the first three rotational invariants.
The grid is composed of \(4\times 4\times 4\) nodes per Å\({}^{3}\) and 18 discrete orientations per node. A typical minimization is performed with a precision of \(10^{-5}\) in the solvation energy in less than 15 iterations. b) CPU time in seconds per iteration for grid meshes \(N_{x}\times N_{y}\times N_{z}=2^{3}\), \(3^{3}\), \(4^{3}\) and \(5^{3}\) per Å\({}^{3}\) and 18 discrete orientations per node. The calculation of the external potential for a given grid mesh is done only once at the beginning of the simulation and takes a CPU time comparable to 1 to 2 iterations. The solvated solute consists of 1280 point charges and 1152 Lennard-Jones sites, for which parameters are extracted from the popular CLAYFF force field and given in table 2. The supercell volume is approximately 68 nm\({}^{3}\).
The densities perpendicularly to the clay layers can be calculated by averaging the above density in the \((x,y)\) planes. The spatial densities of the average dipoles were calculated in the same way. Molecular dynamics simulations were performed with the DLPOLY package.
### System Description
Pyrophyllite is neutral clay. It is monoclinic of space group \(2/m\), perfectly cleaved along orientation {001}. Two views of a pyrophyllite sheet are provided in It consists of a stack of 6 atomic layers. The top layer consists of oxygen (O) and silicon (Si) atoms with a lateral hexagonal symmetry. The central layer consists of aluminum (Al) atoms in 2/3 of the octahedral sites. Between these two layers is a oxygen-hydrogen (O-H) layer, with O atoms at the center of the hexagons formed by the top Al-O sites. The O-H axis is oriented in the direction of the empty octahedral site. The coordinate \(z\) of each site along the normal to the clay surface is given in table 1. The simulation box contains two half clay layers of lateral dimensions \(L_{x}\times L_{y}=41.44\times 35.88\) A\({}^{2}\), which corresponds to 32 clay unit cells of formula A\({}_{\rm 14}\)[Si\({}_{\rm 02}\)](OH)\({}_{4}\). The distance between the surfaces is \(L_{z}=45.57\) A, chosen to recover the bulk density of the Stockmayer fluid at the center of the pore, _i.e._\(n_{0}=0.0289\) molecule per A\({}^{3}\). The simulation supercell is thus \(\approx 68\) nm\({}^{3}\) in periodic boundary conditions, and contains 1280 clay atoms (640 per half clay layer). This pore contains 1600 solvent molecules in the reference all-atom simulations.
We used the CLAYFF force field for both molecular dynamics simulations and MDFT minimization. It is a general purpose force field for simulations involving multicomponent mineral systems and their interfaces with solutions. Related information is tabulated in table 2. The Stockmayer fluid is described by a 3-sites molecule. The neutral central site interacts via Lennard-Jones potentials only. The external sites have opposite charges of \(\pm 1.91\) e, both distant of 0.1 A from the central site. It results in a solvent molecular dipole of 1.84 D (incidentally, approximately that of a water molecule in the gas phase), and, with the chosen Lennard-Jones parameters that match the reduced parameter set of Pollock and Alder, in a dielectric constant of roughly 80, comparable to that of bulk water at room conditions. The parameters for the Stockmayer fluid force field can also be found in table 2. Once again, it is not our purpose here to study the more complex solvation by water, which requires extra terms in the functional, but to demonstrate the possibilities of MDFT in the HRF approximation for a generic polar solvent whose functional is of a good quality.
## III Results
We first compare the predictions of HRF-MDFT for the solvent density and orientation to reference all-atom simulations, which are two orders of magnitude slower to acquire. We then analyze the role of electrostatic interactions on these quantities.
### Density Profiles
The main structural observable describing solvated interfaces is the number density profile \(n(z)\), defined as the average of the number density on plane \(z\),
\[n(z)=\frac{1}{L_{x}L_{y}}\iint\frac{n(\mathbf{r})}{n_{0}}dxdy. \tag{16}\]
\begin{table}
\begin{tabular}{|c|c|c|c|c|c|c|} \hline \multicolumn{2}{|c|}{ Molecule} & Atom & \(\epsilon\) (kJ/mol) & \(\sigma\) (Å) & \(q\) (e) \\ \hline \hline Pyrophyllite & Al & 5.56388e-6 & 4.27120 & 1.575 \\ \cline{2-7} & Si & 7.7005e-6 & 3.30203 & 2.100 \\ \cline{2-7} & O\({}_{G}\) & 0.650190 & 3.16554 & \(-1.050\) \\ \cline{2-7} & O\({}_{H}\) & 0.650190 & 3.16554 & \(-0.950\) \\ \cline{2-7} & H\({}_{G}\) & 0.0 & 0.0 & 0.425 \\ \hline Stockmayer & central & 1.847 & 3.024 & 0.0 \\ \cline{2-7} & side 1 & 0.0 & 0.0 & 1.91 \\ \cline{2-7} & side 2 & 0.0 & 0.0 & \(-1.91\) \\ \hline \end{tabular}
\end{table}
Table 2: Force field used to model the pyrophyllite solute and the Stockmayer fluid in both molecular dynamics simulations and HRF-MDFT minimization. Pyrophyllite parameters are extracted from the CLAYFF force field.
Microscopic clay structure, side and top views (red, O; white, H; yellow, Si; cyan, Al). In the top view, only layers above the Al atoms are shown to highlight the hexagonal symmetry of the first O-Si layer, the hydroxyl group (-OH) parallel to the sheet, and the Al position in 2/3 of the octahedral sites.
The density profiles extracted from explicit molecular dynamics and from HRF-MDFT are given in Good overall agreement is found between the two methods. A shoulder followed by two main peaks are found in MD. In HRF-MDFT, the shoulder looks more like a weak peak. The two strongest peaks are followed by long-ranged oscillations. The first feature (shoulder or weak peak) is found at \(z=8.3\) A, 1.76 A away of the surface layer (see Table 1 to identify layers coordinates). Its intensity is limited to \(\approx 0.7\) in MD and \(\approx 1\) in HRF-MDFT. The largest and main peak is at \(z=9.5\) A, 2.96 A after the top surface layer. An intermediately intense and broad peak is also found at \(z=12.3\) A, 5.76 A away of the surface. These three structures will later be called "prepeak", "main peak" and "secondary peak". Further weak oscillations are found up to 15 A away from the surface. At the center of the supercell, the density is flat at \(n=n_{0}\), namely at the bulk density, which means for solvent molecules to be in bulk, homogeneous, conditions.
The only noticeable difference lies in the weak prepeak seen in HRF-MDFT observed as a shoulder in molecular dynamics. The HRF-MDFT is known to slightly overestimate the height of the first peak in polarized systems. The localized nature of the prepeak can be seen in where is presented the density map in the plane defined by the maximum of the prepeak, as calculated by MD and HRF-MDFT. The same position and overall shape is found with the two methods. It is slightly broader in-plane in HRF-MDFT for approximately the same intensity, which induces the higher value once averaged in the plane. The prepeak is localized at the center of the hexagons formed by surface Si and O atoms. One may note the high maximum value of the normalized number density \(n/n_{0}\) there (up to \(\approx 30\)). The integral of the density in this peak is the total number of particles in it. While a strict deconvolution of the three-dimensional prepeak is arbitrary, no consistent definition results in more than one solvent molecule inside each prepeak.
In figure 5, we also plot the number density in the plane of the main and secondary peaks. Again, the shapes are very similar in MD and HRF-MDFT. The main peak is localized on top of Si atoms. On the contrary, a depletion is found on top of O atoms of the surface layer. The broad secondary peak can be found again on top of the center of hexagons, _i.e._ on top of O atoms.
We now have a clear three-dimensional view of the solvent structure on the pyrophyllite surface: rare solvent molecules are adsorbed very close to the surface, at the center of hexagons formed by Si and O atoms of the clay surface layer. On top of these molecules is stacked a strongly structured hexagonal layer of solvent molecules, above Si atoms. An additional, more diffuse layer is found once again on top of the center of the hexagons. The relatively small distance between the first two layers demonstrates significant interactions, and thus cohesion, between layers. This conclusion is in agreement with Rotenberg _et al._ who showed recently how the competition between adhesion and cohesion determines the hydrophobicity of these surfaces.
As a partial conclusion about structural properties, HRF-MDFT results are in quantitative agreement with the reference molecular dynamics simulations, with a speed-up of _two orders of magnitude_.
### Orientational Properties
We now turn to the orientational properties of the molecules solvating the pyrophyllite surface.
Normalized number density profile of the solvent between two pyrophyllite layers as calculated by molecular dynamics (open circles) and HRF-MDFT (black line).
Maps of the solvent normalized number density \(n(\mathbf{r})/n_{0}\) in planes of the prepeak (left), main peak (center) and secondary peak (right), as calculated by molecular dynamics (top) and HRF-MDFT (bottom).
This problem is recurrent in numerous simulation techniques where orientational degrees of freedom (from electronic spins to molecular orientations) induces new and rich behaviors. For its part, HRF-MDFT gives a direct access to this quantity through the full orientational density \(\rho\left(\mathbf{r},\mathbf{\Omega}\right)\), and also the polarization density \(\mathbf{P}(\mathbf{r})\) defined in Eq. 13, as one of the two natural observables of the theory.
The polarization density is found to be aligned in the \(z\) direction, both in MD and HRF-MDFT. In figure 6, we report the projection of \(\mathbf{P}\) on the \(z\) axis (noted \(P_{z}\)) as a function of coordinate \(z\) (_i.e._ it is averaged in each plane \(z\)). Quantitative agreement is again found between the reference molecular dynamics simulations and the HRF-MDFT. A first peak is found at the location of the pre-peak, another one at the main peak, _etc._ Between each maximum, the sign of \(P_{z}\) changes. Maps of \(P_{z}\) in the planes of the prepeak, main peak and secondary peak are given in Similarly to the density, the overall shapes from both methods are in very good agreement. The polarization density is overestimated by HRF-MDFT in the prepeak, which is expected from previous work on small polarized systems where the polarization close to the solute is often slightly overestimated. In the main and secondary peaks, the polarization densities are found larger in MD than in HRF-MDFT, although not significantly. This may be due to a balance of the layer-by-layer polarizations in HRF-MDFT.
In order to get more information on the orientational properties per molecule, we also define several orientation order parameters.
\[\cos\theta_{P}(\mathbf{r})=\frac{P_{z}\left(\mathbf{r}\right)}{\left\|\mathbf{ P}(\mathbf{r})\right\|}. \tag{17}\]
Another is the averaged molecular orientation
\[\cos\theta_{\mu}(\mathbf{r})=\frac{\left\|P_{z}\left(\mathbf{r}\right)\right\| }{n\left(\mathbf{r}\right)}. \tag{18}\]
We plot the average molecular orientation \(\cos\theta_{\mu}(z)=\left(\cos\theta_{\mu}(\mathbf{r})\right)_{xy}\) in The solvent molecules show strong average orientation in the first layer. On the contrary, the high polarization density in the second, more cohesive layer reflects the large number of molecules inside, each one with a relatively small preferential orientation along \(z\).
One can get more insight into the orientation properties of the solvent molecules by looking at the order parameter \(\cos\theta_{P}\) defined in Eq. 17. When \(\cos\theta_{P}=1\) (\(-1\)), the dipole is oriented against (toward) the surface on the left of the supercell, and vice versa for the other surface. In figure 9, we plot \(\cos\theta_{P}\) in the plane of the first three peaks identified earlier and in an intermediate coordinate \(z\) between the prepeak and the main peak. We observe that solvent molecules in the prepeak, strongly oriented, point against the surface. In the main peak, dipoles are preferentially oriented toward the surface when localized on top of O atoms. In the secondary layer, preferential polarization is found in the whole plane against the surface.
Component along the normal \(z\) to the surface of the dimensionless solvent polarization density, between two porphyllite layers, as calculated by molecular dynamics (open circles) and HRF-MDFT (black line). The density profile is also presented in arbitrary units (red dashed line). Maxima in the polarization density correspond to maxima in the solvent number density.
Maps of the polarization density projected on the normal \(z\) to the surface axis, \(P_{z}\), in the planes of the prepeak (left), main peak (center) and secondary peak (right), as calculated by molecular dynamics simulations (top) and HRF-MDFT (bottom).
### The Role of Electrostatics
The computational efficiency of the molecular density functional theory unlocks systematic studies of the solvent structure and thermodynamics, such as the relative roles of van der Waals and electrostatic interactions to the CLAYFF force field. They are modeled respectively by Lennard-Jones interactions and a point charge distribution reported in table 2. reports the number density profile along the \(z\) axis for modified systems where the CLAYFF charges of the clay atoms have been scaled by a factor of 1, 0.5 and 0. It is observed that, surprisingly, only the shape of the prepeak is modified when the point charges are turned off. This part of the density profile evolves from a localized peak (with electrostatics on) to a shoulder of the main peak when quenched. The rest of the number density profile is unchanged. As might be expected, the polarization vanishes when charges are turned off. In is also plotted the relative change in solvation free energy \(F\left[\rho\right]\) for scale factors between 0 (turned off) and 1 (completely turned on) of the charges of the clay atoms. The solvation energy is affected by less than 6 % when charges are turned off. The contribution of the electrostatics is thus small compared to the Lennard-Jones one, which is not an obvious result even for a surface with a zero net charge.
## IV Conclusion
We have shown that the molecular density functional theory in the homogeneous reference fluid approximation (HRF-MDFT) is able to handle the study of solvation properties of a complex clay surface of several hundred atoms. The description of its structural and orientational properties is quantitatively comparable to reference all-atom molecular dynamics, while reducing the CPU-time by two orders of magnitude. This means that the computation of these properties is now accessible on a simple workstations within minutes. The HRF-MDFT calculations allowed to accurately describe the subtle structure of the first molecular layers at the surface, in particular the orientational properties. The only noticeable difference lies in a localized zone of high density at the center of hexagons, a defect due to a small overestimation of the polarization close to the solute inherent to the method. Some improvements in this respect are in progress.
The numerical efficiency of our approach made it possible to analyze the relative contributions of the CLAYFF force field to the solvation properties. The density profile and the solvation energy are found to be insensitive to the electrostatic contribution. This may imply that the role of electrostatics in charged clays may be reasonably reduced to their charged defects.
Profile along the normal to the surface \(z\) of the average molecular orientation, \(\cos\theta_{\mu}(z)\) (black line). The density profile of the solvent is also given in arbitrary units (red dashed line).
Only the prepeak is affected. b) Relative change in solvation free energy as a function of the scale factor. The relative changes stay below 5.5 %.
From left to right, order parameter \(\cos\theta_{P}\) in the prepeak, at the intermediate state between the prepeak and the main peak, in the main peak, and in the secondary peak.
Finally, the next step of this work will be to consider the solvation of clay surfaces by a realistic model of water, at either a dipolar or multipolar level, and by ionic solutions.
|
10.48550/arXiv.1211.4884
|
Solvation of complex surfaces via molecular density functional theory
|
Maximilien Levesque, Virginie Marry, Benjamin Rotenberg, Guillaume Jeanmairet, Rodolphe Vuilleumier, Daniel Borgis
| 1,726
|
10.48550_arXiv.1911.06976
|
###### Abstract
The electronic structure of benzene in the presence of a high-intensity high-frequency circularly polarized laser supports a middle-of-the-ring electron localization. Here, the laser polarization coincides with the ring plane of benzene. The high-frequency oscillating electric field creates circular currents centered at each atom with a circle radius equal to the maximum field amplitude of the laser. All six carbons have six such rings. For a maximum field amplitude of 1.42 A, which is the carbon-carbon bond distance, all six dynamic current circles intersect to create a deep vortex in the middle, which supports a bound state of a pair of electrons. Such states for benzene can be realized in experiments using a circularly polarized XUV-laser in a range of intensities 10\({}^{16}\)-10\({}^{17}\) W/cm\({}^{2}\) and frequencies 16 eV to 22 eV. Electronic dynamics calculations predict a minimal ionization of benzene when the rise-time of the laser pulse is sudden, indicating a possible experimental realization of these states characterized by a large cut-off in the harmonic generation spectra. This stable electronic structure of the light-dressed benzene is doubly-aromatic due to an extra aromaticity from a D\({}_{6h}\) symmetric circular distortion of the \(\sigma\)-framework while the \(\pi\)-electrons, with low density in the ring-plane, are least affected.
## 1 Introduction
Contrary to intuition and expectation, high-frequency laser fields with intensities comparable to internal electric fields of atoms or molecules create long-lived quasi-bound states, which are almost stable to decay, dissociation or ionization. Chemistry in these strong high-frequency fields is not far from reality with the maximum achievable laser intensity now being recorded at \(10^{21}\) W/cm\({}^{2}\) consisting of photons that are almost in the relativistic regime. The existence of quasi-bound states (metastable, but slow-ionizing with a long lifetime) have been a theoretical prediction, with an experimental realization in the case of noble gas atoms such as neon, helium and argon. Over the decades, many theoretical studies have looked at such exotic states in the case of diatomic molecules. Several interesting effects such as conformational changes for small molecules, linear Stark shifts of laser-dressed atoms etc. have been theoretically proposed. A question that always springs up for a chemist is what happens with polyatomic molecules? How would their light-dressed stable states look like?
With several theoretical proposals of driving electronic currents in molecules such as porphyrin, and some latest path-breaking experiments on graphene, a simple molecule of thought for a chemist would be benzene. Given its \(D_{6h}\) geometry and aromaticity, circularly polarized lasers (CPLs) would be an apt choice to understand its electronic dynamics in a strong high-frequency oscillating field. For porphyrin, the proposal was to induce a unidirectional electron flow through its structure with a CPL. A quantum control of electronic fluxes charge migration in the attosecond regime was demonstrated for a benzene molecule, using a combination of linearly and circularly polarized lasers. The laser preparation was found to impart electronic phases to the model molecule, affecting the charge migration mechanism.
Since attosecond, strong pulses of circularly polarized light are the call of the day, strong-field effects in benzene invoke ample curiosity, for both experiments and theory. The latest experiments on graphene have come up with new topological states created by laserpulses. Even though the electronic structure of benzene is different from that of graphene, as the latter comprises a band structure that lacks a band gap, benzene can be considered a _beginner's choice_ apropos model for graphene. A study of benzene in CPL fields of \(10^{14}\)-\(10^{15}\) W cm\({}^{-2}\) by Baer et al. shows a preservation of the D\({}_{6h}\) symmetry, interpreted from a higher harmonic generation spectra. The CPL-induced dynamics and laser-dressed electronic structure of benzene need to be understood in perspectives of contrast as well as similarities with graphene.
In a circularly polarized laser, the time-dependent electric field with \(x\)- and \(y\)-polarization in Cartesian directions (and a propagation direction of \(z\)) will be,
\[\vec{\epsilon}(t)=f(t)\epsilon_{0}\left[\cos(\omega t)\hat{e}_{x}+\sin(\omega t )\hat{e}_{y}\right], \tag{1}\]
For a continuous wave (CW) laser we set \(f\left(t\right)=1\), with an instantaneous rise-time. For benzene, in this work, a CW-laser form (as a CPL) is chosen initially instead of a pulse, to understand its light-dressed state and possible light-dressed effects. The Time-Dependent Schrodinger Equation (TDSE) for such an \(n\)-electron, \(M\)-atom molecule is (in atomic units, a.u.):
\[i\frac{\partial}{\partial t}\Psi\left(\{\vec{r}_{j}\},t\right)=\left[\frac{1} {2}\sum_{j}^{n}\left[\hat{p}_{j}+\hat{\vec{A}}(t)\right]^{2}-\sum_{A}^{M}\sum_ {j}^{n}\frac{Z_{A}}{\left|\vec{r}_{j}-\vec{R}_{A}\right|}+\sum_{j>k}^{n}\frac{ 1}{\left|\vec{r}_{j}-\vec{r}_{k}\right|}\right]\Psi\left(\{\vec{r}_{j}\},t \right), \tag{2}\]
Moving on to the Kramers-Henneberger (KH) frame of reference, which is an oscillating frame of reference, is achieved via a time-dependent unitary transformation:
\[\hat{\Omega}=\exp\Bigg{[}i\int_{-\infty}^{t}\Bigg{(}\vec{A}\left(\tau\right) \cdot\sum_{j=1}^{n}\vec{p}_{j}+\frac{1}{2}\left|\vec{A}\left(\tau\right) \right|^{2}\Bigg{)}\,d\tau\Bigg{]}. \tag{3}\]In the KH frame, the wave function \(\Phi\left(\left\{\vec{r}_{j}\right\}\!,t\right)\) is \(\hat{\Omega}\Psi\left(\left\{\vec{r}_{i}\right\}\!,t\right)\), a unitarily transformed quantity. Here, \(\hat{\Omega}=\hat{\Omega}_{1}\hat{\Omega}_{2}\), with \(\hat{\Omega}_{1}=\exp\left[i\sum_{j=1}^{n}\int_{-\infty}^{t}\vec{A}\left(\tau \right)\cdot\vec{p}_{j}d\tau\right]\). This gives a TDSE of the form:
\[i\frac{\partial}{\partial t}\Phi\left(\left\{\vec{r}_{j}\right\}\!,t\right)= \left[\frac{1}{2}\sum_{j}^{n}\hat{p}_{j}^{2}+\sum_{j=1}^{n}V_{C}\left[\vec{r}_ {j}-\vec{\alpha}\left(t\right)\right]+\sum_{j>k}^{n}\frac{1}{\left|\vec{r}_{j }-\vec{r}_{k}\right|}\right]\Phi\left(\left\{\vec{r}_{j}\right\}\!,t\right). \tag{4}\]
The effect of \(\hat{\Omega}_{2}\) is to generate a purely time-dependent phase factor while \(\hat{\Omega}_{1}\) acts as a time-dependent displacement operator, modifying the nuclear-electron attraction potential to the expression:
\[V_{C}\left[\vec{r}-\vec{\alpha}\left(t\right)\right]=-\sum_{A}^{M}\frac{Z_{A} }{\left|\vec{r}-\alpha_{0}\hat{e}_{x}\cos(\omega t)-\alpha_{0}\hat{e}_{y}\sin( \omega t)-\vec{R}_{A}\right|}. \tag{5}\]
The inter-electron repulsion remains the same. Explicitly, \(\vec{r}_{j}\rightarrow\vec{r}_{j}-\vec{\alpha}\left(t\right)\) and \(\vec{r}_{k}\rightarrow\vec{r}_{k}-\vec{\alpha}\left(t\right)\), and consequently in \(\left|\vec{r}_{j}-\vec{r}_{k}\right|\), the time-dependent shift cancels. In simple intuitive terms, the inter-electronic (particle) distance will remain the same, on a change in the frame of reference. The factor \(\alpha_{0}\) has units of distance and is the maximum field amplitude of the laser, in a.u.:
\[\alpha_{0}=\frac{\epsilon_{0}}{\omega^{2}}. \tag{6}\]
Thus, in this oscillating frame of reference the binding Coulomb potential is time dependent. For an isolated single atom, the binding potential has a form where the Coulombic well (singularity) traverses in a circular path with the radius of the circle being \(\alpha_{0}\), the maximum field amplitude of the CPL. This is the _oscillating_ frame of reference from the point of view of the electron. For such a spatially circular and temporally periodic traversal of a real potential, the harmonics can be understood via a time-dependent Fourier expansionsymbolized as:
\[V_{C}\left[\vec{r}-\vec{\alpha}\left(t\right)\right]=V_{0}^{KH}\left(\vec{r} \right)+\sum_{n=1}^{+\infty}V_{n}^{KH}\left(\vec{r}\right)\cos\left(n\omega t \right). \tag{7}\]
In the break-up of the sum, the individual Fourier components (harmonics) have the form:
\[V_{n}^{KH}\left(\vec{r}\right)=\frac{1}{T}\int_{0}^{T}V_{C}\left[\vec{r}- \alpha_{0}\left(\hat{e}_{x}\cos\omega t+\hat{e}_{y}\sin\omega t\right)\right] \cos\left(n\omega t\right)dt. \tag{8}\]
For a high-frequency CPL, on a time averaging (over the faster, high-frequency \(n>0\) components), the temporally oscillating terms, \(V_{n}^{KH}\left(\vec{r}\right)\) with \(n>0\) would contribute less. Hence, the potential term, \(V_{0}^{KH}\left(\vec{r}\right)\), will now be the effective binding potential instead of the field-free Coulomb potential. The zeroeth order term \(V_{0}^{KH}\left(\vec{r}\right)\) is time independent (the first term in a Fourier expansion is a simple time average) and the _dominant_ one, which gives a time-independent Schrodinger equation in the effective potential. It was shown by solving Equation 4 using high frequency Floquet theory (HFFT) in successive lowest order of \(\omega^{-1}\), that Equation 4 reduces to a time-independent Schrodinger equation with the potential given by \(V_{0}^{KH}\). In the limit of a high-frequency, an adiabatic Floquet theory ensures that the time-dependence is rendered inconsequential, as the higher order terms in the KH potential are negligible. Thus, for short laser pulses, ionization is suppressed due to a laser-induced resonance state. Using a laser intensity as high as \(10^{16}\) W/cm\({}^{-2}\) and frequency 806 nm, experiments have demonstrated that the electron scattering is "switched off" and ionization and molecular fragmentation are suppressed in these regimes. Currently, lasers with higher intensities, frequency of 26.5 eV and pulse that lasts upto 10 ns are available. The high frequency conditions also ensure that the ionization is minimized and the bound states of the time-independent Schrodinger equation in the effective potential determine the chemistry.
In this zeroeth order time-independent effective potential, the Coulombic binding potential transforms from spherical to toroidal. This is portrayed in Figure 1**(a)** where the usual spherical Coulomb potential for a single, isolated atom and the corresponding high-frequency laser-dressed potential (see Figure 1**(a)**, second and third plots) have been depicted. The isosurfaces for the Coulomb potential assume a _doughnut-like_ toroidal shape with a saddle point at the atomic position instead of the Coulombic singularity. Thus, when the laser intensity is high enough (high electric field strength), the electric field of the CPL induces a circular motion of the electrons in the plane of polarization, the circle being centered at the atom. The radius of this torus is equal to \(\alpha_{0}\) given in Equation 6,which in turn is proportional to the electric field strength, \(\epsilon_{0}\), for a fixed \(\omega\), frequency.
In our work, the effect of a CPL on a benzene molecule has been examined with the laser polarization taken along \(xy\)-plane which is the molecular plane. The field-free (without laser) Coulombic potentials and electron densities of benzene molecule are presented in Figure 1**(b)** and 1**(c)**. The electric field of the CPL is proportional to the electric field strength, \(\epsilon_{0}\), for a fixed \(\omega\), frequency. The electric field of the CPL is proportional to the electric field strength, \(\epsilon_{0}\), for a fixed \(\omega\), frequency. The electric field of the CPL is proportional to the electric field strength, \(\epsilon_{0}\), for a fixed \(\omega\), frequency. The electric field of the CPL is proportional to the electric field strength, \(\epsilon_{0}\), for a fixed \(\omega\), frequency.
Motivation and Introduction : **(a)** 1-D plots for a without-field and with-CPL isolated single atom. The spherical Coulomb potential loses its singularity upon switching on the laser and traverses a circular path with the radius of the circle being \(\alpha_{0}\), the maximum field amplitude of the CPL, depicted as a ’split’ in the potential, revolution of which would form the figure adjacent in 2-D. A 3-D plot for an effective KH-potential. Isosurface plots for field-free **(b)** bare-nuclear potential (V\({}^{field-free}\)) and **(c)** Electron Density (\(\rho\)).
For the one-electron density there are maxima at the nuclear positions and saddle points at bond-centers, bringing out the structural signatures of field-free benzene.
The gist of the questions addressed in this work (and a hint of the results), is pictorially represented in showing a two-dimensional depiction of the effective potential \(V_{0}^{KH}\left(\vec{r}\right)\) for benzene. If a single atom is toroidal or circular in the plane of laser polarization, then what are the effects, unusual or interesting, which one could have for benzene, a molecule where each of the carbon atoms are arranged on the vertices of a regular hexagon? Figure 2**(a)** gives a schematic portrayal of an intensity variation of the CPL, which effectively amounts to an increasing \(\alpha_{0}\), which in turn is the radius of each of the circles centered at each carbon atom.
**(a)** A schematic representation of circles centered at vertices of a hexagon, intersecting at varying \(\alpha_{0}\), given in the panels **(b)** to **(f)**. At \(\alpha_{0}=2.7\) a.u. the circle radius is just enough to form an optimum overlap at the hexagon center, giving the deepest point in the laser dressed potential to be the middle-of-the ring, which is expected to localize electrons.
At \(\alpha_{0}=\frac{r_{CC}}{2}\), the circles just touch each other, leading to deeper potentials at the middle of the bonds. For \(\alpha_{0}=r_{CC}\), all of the six circles intersect right in the middle of the regular hexagon, giving a deep light-induced vortex at the ring-center. The nature of this potential suggests a very interesting electronic structure for benzene in the presence of a strong CPL, with the possibility of electrons being seated at this ring-center.
The nature of these unexpected stable states achieved in the presence of the laser is described in the subsequent sections, using a comprehensive analysis of molecular orbitals. It is shown that a pair of electrons populates a bound state of an \(s\)-type character in the middle of the ring, as a result of a laser-induced hybridization. If the electrons' propensity for staying towards the middle of the ring is high in the presence of the laser, then this has consequences for proton stabilization at the ring-center. Interestingly, the \(\pi\)-cloud of benzene is unaffected by the laser polarization in the \(xy\)-plane, since the \(\pi\)-density is negligible in the plane of the ring. This ensures that the aromatic character of benzene is preserved. However, the circular distortions in the \(\sigma\)-framework lead to a delocalization and yield a new aromatic character, imparting a double aromaticity to the laser-dressed benzene. This is analyzed in terms of electron localization functions and electrostatic potentials. Finally, our work gives an experimental prescription in terms of laser parameters from calculations of electron dynamics, in the lab frame and explains how these states can be realized and stabilized by a laser pulse. Signatures of harmonic generation spectra from these peculiar electronic states are characterized by a high cut-off in the harmonics because of the high-frequency stabilization.
## 2 Results and Discussion
The first set of results are striking, and evident on an inspection of single-particle charge densities, \(\rho_{KH}\left(\vec{r}\right)\) for different values of \(\alpha_{0}=1.35,1.5\) and \(2.7\) a.u. These have been calculated by solving the multi-electron time-independent Schrodinger equation in the effective potential, \(V_{0}^{KH}\left(\vec{r}\right)\), given in Equation 6. The methodology adopted for this, together with the basis set used have been discussed in the section on Computational Methodology. Effective potentials plotted in three-dimensions (3-D) in terms of isosurfaces are given in Figure 3**(a)**, **(b)** and **(c)** in the upper panel, together with different orientations of the charge densities in the lower panel. The electron densities in the effective potential give a notion of structure in the presence of the CPL.
In contrast to the field-free electron density given in Figure 1**(c)**, the laser-dressed densities (\(\rho_{0}^{KH}\)) in do not possess nuclei-centered maxima, neither do they satisfy the Kato's cusp condition at nuclear positions. It is obvious why this is so, from the nature of the effective potentials. Once the toroidal densities start overlapping at \(\alpha_{0}\geq\frac{r_{CC}}{2}\), where \(r_{CC}\) denotes the carbon-carbon distance, maxima appear in the middle of the C-C bonds, indicating a favorable bonding behavior. The hydrogens are smeared out and do not contribute much to the laser-dressed densities. Hence, instead of a nuclei-centric structure, a bond-centric structure is evident in terms of electron densities, when \(\alpha_{0}=1.35\) a.u. and \(\alpha_{0}=1.5\) a.u.
On further increasing \(\alpha_{0}\), the density gets pushed towards the middle of the ring and outwards from the regular hexagon framework. This is achieved by the increasing radius of the circular currents induced by an increasing intensity of the CPL. When \(\alpha_{0}=r_{CC}\), the next topological change occurs, which is visualized in terms of lower-valued densities at the atomic positions as well as beyond the inter-nuclear axis. This is because the nuclei-centric tori now intersect, or have common grounds/volume at nuclear positions, as can be seen in Figure 2**(f)**. A sideways horizontal view of the charge density profile given in Figure 3**(c)**(lower panel) now indicates a bulge in the ring-center, which is evident and in contrast to the other horizontal profiles given in the lower panels of Figure 2**(a)** and **(b)**. Another feature to be noted is the appearance of _bonding_ maxima in the immediate periphery of the ring, which is attributed to a double aromaticity, discussed later in the manuscript as a property evident from the electronic structure. A numerical integration of the densities inside the ring revealed an increase in the number of electrons, moving away from the nuclear framework to the middle of the ring. A further evidence for this will be discussed in terms of a molecular orbital analysis in the KH oscillating framework, for the time-independent effective potential.
To summarize the results from the charge densities, there can be three different topological structures arising as a function of an increasing \(\alpha_{0}\): (i) when \(\alpha_{0}<\frac{r_{CC}}{2}\), the overlap between the tori of electron densities is the least and the density is yet to lose its nuclei-centric structure, (ii) when \(\alpha_{0}\geq\frac{r_{CC}}{2}\), when the overlap is substantial between the and then diminishing, pushing densities towards the middle of the ring and then outwards and finally (iii) when \(\alpha_{0}=r_{CC}\), which corresponds to an effective potential producing a topological phase of maximum density at the ring-center.
The laser-induced _structural effects_, as seen through the electron densities, can be ra
Isosurface plots for the time-averaged zeroth-order KH potentials (Red) and Electron densities (Blue) for varying quiver distances. The potential and densities are localized at the bond centers for \(\alpha_{0}\) = 1.35 a.u., and 1.50 a.u. and localized at the ring-center for \(\alpha_{0}\) = 2.7 a.u.
What happens when \(\alpha_{0}>r_{CC}\). Nothing much, since a further increase in \(\alpha_{0}\) makes the potential shallower. When \(\alpha_{0}>r_{CC}\), the deep vortex in the middle disappears. There would always be exactly two intersections for each pair of circles and the total number of intersections therefore would be 2 \(\times\)\({}^{6}C_{2}(\) = 30\()\), shallow points. The electron density would just keep moving away from the geometry with shallow maxima at the intersections, which would amount to undulations in an almost homogeneous electron density.
With the laser-dressed electron density yielding structural information, it is useful to
Isosurface Plots of Molecular Electrostatic Potential (MESP ; Purple and Light-pink color scheme) and Electron Localization Function (ELF ; Blue color scheme) for benzene with CPL-field of varying quiver distances. Lower panel : Interaction Energy plots for with-field and field-free \(C_{6}H_{6}\)-\(H^{+}\), \(\Delta V_{int}\) is a minimum at the ring-center, in contrast with the barrier seen in the field-free case.
The response of the laser-dressed density to external perturbations is quantified and understood here, through properties derived from the electron density _viz._, the electrostatic potential, electron localization function and energy of interaction with a proton. The electrostatic potential here is defined as:
\[V\left(\vec{r}\right)=\sum_{A}\frac{Z_{A}}{\left|\vec{r}-\vec{R}_{A}\right|}- \int\frac{\rho_{0}^{KH}\left(\vec{r}\,^{\prime}\right)}{\left|\vec{r}-\vec{r} \,^{\prime}\right|}d^{3}r^{\prime}.\]
The negative regions of the electrostatic potential of field-free densities indicate regions susceptible to electrophilic attack, with negative-valued isosurfaces enclosing negative-valued minima for lone pairs, \(\pi\)-clouds and aromatic \(\pi\)-clouds. This is seen from Figure 4**(a)** where the electrostatic potential for field-free benzene shows the presence of the aromatic \(\pi\)-cloud, with the light-pink isosurface enclosing six negative-valued minima arranged in a hexagon, on either side of the ring. The effect of the CPL with increasing \(\alpha_{0}\) shows the appearance of new negative regions at bond peripherals outside the ring and a negative-valued minimum inside the ring (Figure 4**(b)**). This is reminiscent of electron density getting pushed towards the middle of the ring and outwards. The outward negative-valued minima signify the doubly aromatic character developed in the system by the circularly polarized laser and the middle-of-the ring negative-valued minimum, indicates that a bare proton could bind to this site of a localized electron pair. It is also seen that in the presence of the CPL the negative regions of the electrostatic potential have become more negative in comparison with the field-free electrostatic potential. Bench-stable double aromatic compounds have been synthesized and isolated, where in addition to the \(\pi\)-aromaticity, a \(\sigma\)-aromaticity is realized in a hexaheteroatom-substituted benzene dication, and the \(\sigma\) electrons follow Huckel's \(4n+2\) rule, being circularly delocalized in the plane of the benzene ring. With the laser-parameters used in this work, a neutral unsubstituted benzene per se is shown to contain double-aromaticity.
The interaction energies of a proton with the with-field benzene (\(\Delta V_{int}\)) were plotted as a function of the proton distance from the ring-center. The energy is a minimum at the ring-center for the laser-dressed system, as against the energy in the field-free case where the proton sees a barrier (**(e)**). This is expected since the the MESP is a minimum (a negative and positive minimum for \(\alpha_{0}\) values 1.35 a.u. and 2.7 a.u., respectively) at the ring-center, which reveals the nucleophilicity of the with-field benzene, a requisite for electrophilic aromatic substitution reactions.
Further information is afforded by an analysis of the electron localization function (ELF) for the laser dressed densities. For a single determinant wavefunction built from Hartree Fock orbitals \(\psi_{i}\), the ELF on a 3-D grid is given as,\(ELF=\frac{1}{1+(\frac{D}{D_{h}})^{2}}\), where, \(D=\frac{1}{2}\sum_{i}\lvert\nabla\psi_{i}\rvert^{2}-\frac{1}{8}\frac{\lvert \nabla\rho\rvert^{2}}{\rho}\ ;\ \ D_{h}=\frac{3}{10}(3\pi^{2})^{\frac{5}{3}}\rho^{ \frac{5}{3}}\). The plots in the middle panel of show the isosurfaces for ELF and the maxima in ELF are in agreement with the proposal that there is a middle-of-the-ring electron localization. A clear signature of aromaticity is seen for \(\alpha_{0}=1.5\) a.u. and \(\alpha_{0}=2.7\) a.u. Outer ring maxima correspond to the doubly aromatic character that is developed. The reason for the doubly aromatic character is rationalized with respect to the molecular orbitals (MOs) and energy levels with and without the field.
Yet another evidence for the middle-of-the-ring localization is given by the Molecular orbitals in a mean-field picture of solving the KH time-independent Schrodinger equation. The MO's and MO energies as function of the parametric variation of \(\alpha_{0}\), the laser parameter give reasons for the middle-of-the-ring electron localization. The mixing of the orbitals, due to the increasing \(\alpha_{0}\) is evident in A Walsh-kind of diagram as a function of \(\alpha_{0}\), instead of the nuclear geometry, is portrayed in Figures 5**(c)** and **(d)**. What is clearly evident are the following:
* The first six MO's having a total of 12 electrons are linear combinations of the \(1s\)-state on each carbon atom. For the field-free benzene, the nodal structure of these emanate from the interaction due to the nuclear geometry. In the case of the laser-dressed benzene, for \(\alpha_{0}=1.35,1.50,2.7\) a.u. the circular electric field mixes these \(s\)-states, while preserving the nodal structure from the benzene geometry. For the ground state MO at \(\alpha_{0}=1.35\) a.u., clearly the first MO density is being pushed into the ring and outwards, giving a circular dichotomy. The nodeless first MO at \(\alpha_{0}\) is clearly an \(s\)-state with maximum density in the middle of the ring, evidencing a pair of electrons sitting at the ring-center. This state is clearly due to all the six tori centered on each carbon, intersecting to form a deep vortex in the middle which supports this state.
* The next feature of interest are the \(\pi\)-orbitals of benzene, accounting for the 6 \(\pi\)-electrons, which are almost unaffected by the circular-polarization.
A molecular orbital analysis for field-free and with-field benzene. (a) MOs are plotted with increasing energy. MOs 1-6 can be treated as core \(\sigma\)-orbitals which do not gain a node when the laser is switched on. MOs 17, 20, 21 are \(\pi\)-orbitals unaffected by the laser. The remaining MOs seem to have undergone the most change in the presence of a laser, with an increase in the number of nodes by one (except for 14 and 19) (b) Hartree-fock Energy plot as a function of quiver distance of the CPL. (C) MO energies as a function of the quiver distance.
The aromaticity of benzene is therefore preserved in the presence of the CPL. The nodal structure from benzene is preserved for these orbitals.
## (iii)
The third feature is the distorted \(\sigma\)-framework of benzene, which is induced by the CPL, resulting in a \(\sigma\)-aromaticity on top of the unaffected \(\pi\)-aromaticity, giving a totally doubly aromatic character. These orbitals account for 30 electrons which follow the \(4n+2\) count, where \(n=7\). Here the interpretation of the rule is to be modified, with six contributions from the carbon atoms and an extra contribution from the \(\sigma\)- electrons displaced by the circularly polarized light which amounts to a \(n=7\) count. These distorted \(\sigma\)-orbitals have an extra feature of CPL-induced circular nodal structure together with the nodal planes stemming from the benzene structure. The reason for this feature is explained in Figure 6, and it comes as a consequence of the circular dichotomy introduced by the CPL. The orbitals and thus the electronic structure therefore constitute a real laser-dressed state.
Orbitals 1 through 6 register no change in the number of nodes in going from the field-free to with-field realm. However, MO 7 sees a gain of one node. The same is true for MOs 8-13 and MOs 15-18. MOs 14 and 19 do not gain a node upon switching on the laser. This nodal picture becomes lucid with a particle-in-a-ring analysis, presented in Figures 7. For instance, the field-dressed MO 7 can be visualized as a superimposition of a particle-in-a-ring first-excited state over a field-free MO 7. This approach is extended to higher MOs and the constituent particle-in-a-ring states are depicted in the figures. The MOs which do not acquire a new nodal plane upon switching on the laser can be explained using a totally-symmetric particle-in-a-ring ground-state superimposed over the field-free MO. An intriguing feature is a flip in the direction of constituent ring states, from clockwise to counter-clockwise, for MOs 8, 11 and 18 when one goes from \(\alpha_{0}\) = 1.35 a.u. to \(\alpha_{0}\) = 2.7 a.u.
The Hartree-Fock and MO energies are plotted for varying quiver distances and shown in Figure 5(**b**) and (**c**), respectively with an enlarged view of the latter in Figure 5**(d)**. The core orbital (MO 1) contributes to a stabilized state owing to a visible dip in energy around \(\alpha_{0}\) = 2.45 a.u. Even upon switching on the laser, the core MOs, 1-6, preserve their character and stay well below the higher valence MOs. A few of the low-lying and higher valence MOs dip and rise in energy around \(\alpha_{0}\) = 1.35 a.u. However, these valence MOs return to more uniformly-spaced energy levels at \(\alpha_{0}\) = 2.7 a.u., and contribute to a laser-dressed state of benzene which is bound.
A simple way to understand the aforementioned nodal structure in the MOs forming the \(\sigma\)-framework is to think of the laser-dressed orbitals as a direct product of particle-on-a-ring
Particle-in-a-ring (PIAR) analysis for MOs of laser-dressed benzene. One sees a flip in the constituent PIAR states for MOs 8, 11 and 18, from clockwise at \(\alpha_{0}\) = 1.35 a.u. to _anti_-clockwise at \(\alpha_{0}\) = 2.7 a.u. The ground state of a PIAR is totally symmetric while the higher excited states are doubly-degenerate. The first excited state is a degenerate single-node state, imposing an angular momentum in clockwise or _anti_-clockwise directions on the field-free state upon switching on the laser.
This depiction is given in for the MOs forming the \(\sigma\) framework. It is as if there is a particle on a ring associated with each of the carbon atoms. When the rings start interacting with increasing \(\alpha_{0}\), there are flips in the signs according to the mixing and the energetic stabilization. The degenerate excited states for particle-on-a-ring can be thought of as a classical _clockwise_ and _anti-clockwise_ circulation. This amounts to a local angular momentum imparted by the CPL. However, at \(\alpha_{0}=2.7\), the net local angular momentum at the benzene ring-center is zero because of a localized \(s\)-state. The local angular momentum at each carbon atom is non-zero in some cases but the total vector addition will sum upto zero.
With all these exotic features including a middle-of-the-ring electron pair and double aromaticity, it is important to understand how these states can be realized in an experiment and what are their implications and applications. The nature of proton conduction when the benzene ring-center is electron-rich has already been discussed and this has implications for control over electrophilic substitution reactions. The double aromaticity also plays an important role here. However, a direct application is seen in terms of the harmonic generation spectra of these states for a given set of laser parameters. Firstly, we discuss an estimate of the laser parameters for intensity, frequency, risetime and pulse width of the circular polarized laser required for the formation of these states. The KH-Schrodinger equation in the oscillating frame of reference only gives the \(\alpha_{0}\) parameter (see Equation 6), which will be a measure of the intensity for a fixed frequency. To simulate experimental conditions, the solutions of a TDSE yielding the electronic dynamics in the lab-frame are presented in Figure 7, for \(\alpha_{0}=1.35,1.5\) and \(2.7\) a.u. The intensity and frequency have been varied for a fixed \(\alpha_{0}\). It is seen that for high values of \(\epsilon_{0}=0.886\) a.u. (\(2.754\)\(\times\)\(10^{16}\) W cm\({}^{-2}\)) and \(\omega=0.81\) a.u. (\(22.04\) eV), for \(\alpha_{0}=1.35\) a.u., about \(60\%\) of the benzene molecules survive the laser pulse and remain in this state as long as the peak CW region of the pulse is on. This is evident from the norm of the ground state wave function for all the cases.
Yet another factor that is analyzed is how fast should the rise time of the pulse be, so that these states have minimal ionization. This is portrayed in Figure 7**i**. The second panel in gives the harmonic generation spectra due to the formation of these states. It is seen that the harmonic generation spectra are highly intense and with a large cut-off going upto 30 in the case of \(\alpha_{0}=1.35\) and \(\alpha_{0}=1.5\) a.u. Electronic structure-wise, both of these belong to category (ii) discussed at the beginning of this section. What seems striking is that when one has a middle-of-the ring localization for \(\alpha_{0}=2.7\) a.u. the cut-off in the harmonic order dramatically increases to 55 as seen in Figure 7**(f)**, where the intensity and frequency of the wave-function are very similar. The same is true for the harmonic generation spectra due to the formation of these states. The same is true for the harmonic generation spectra due to the formation of these states.
This is evidently due to the new bound state of the paired up electrons in the middle of the ring which are bound by the CPL.
## 4 Concluding remarks
Our work has discussed the formation of exotic states for benzene in the presence of circularly polarized light, with plane of light polarization coinciding with ring plane in benzene. Through electronic dynamics at a fixed nuclear geometry, we have shown that such states can be experimentally obtained using a high intensity XUV-laser. Under high frequency conditions, as high as 16-22 eV at least 60% of the benzene molecules survive the laser pulse. The laser modifies the electronic structure of benzene such that the exotic states formed through a laser-induced hybridization have very interesting properties which include _two_ electrons sitting in the middle of the benzene ring and a double aromaticity.
The laser parameters obtained fall in line with current experimental capabilities. Previous works had shown ionization and Coulomb explosion of benzene when the intensity was of the order of \(10^{16}\) W/cm\({}^{2}\) at much lower frequencies of 1 eV to 8 eV. The presence of the high frequency stabilization given in our work, enables us to put forth a chemically modified benzene through laser-induced hybridization of its orbitals, that has properties very different from that of the usual benzene. It has been shown here through numerical calculations of properties and energetics that the laser-dressed benzene shown can stabilize a bare proton in the middle of the ring. With a benzene-proton complex forming a precursor for the Wheland ion intermediate, the new state predicted in the presence of the laser can imply temporal attosecond control over the dynamics of electrophilic substitution reactions.
Of late, improbable reactions being catalyzed by very strong static electric fields have been a subject of interest. A strong DC field is difficult to produce in a laboratory, however a time-dependent attosecond pulse constituting an oscillating electric field is nowadays state of the art. Moreover, the oscillating electric field has control over the ionization dynamics as well.
The experimental manifestation and effects of the xotic state have direct applications in terms of higher harmonics produced. The laser-dressed high frequency state in the case of benzene, with minimal ionization can produce a very large number of harmonics going upto 55 in comparison with the low intensity ionizing state where the harmonics last upto about 30. This again is due to the exotic set of bound states formed in the high-frequency stabilization.
Very recently light-induced charge density waves have been experimentally reported in a layered solid, LaTe\({}_{3}\). Again graphene and its laser-dressed states have had a recent experimental manifestation. A natural extension of our work related to benzene is to apply the same for graphene. In the case of graphene too, with circularly polarized light in the plane of the layer, instead of the light breaking the symmetries, _time crystals_, with structural and temporally periodic properties would certainly form a novel state of matter, the light-dressed state.
## 4 Computational methodology
Time-independent zeroth-order KH calculations in the effective potential \(V_{0}^{KH}\)of benzene have been done using GAMESS-US package with a COEMD-3 basis and an extra set of \(d\)-type even-tempered basis functions on the carbon atoms. Since, the KH-transformation only modifies the nuclear-electron attraction term, the GAMESS-US has been modified for the KH calculations. The integration for the zeroth order terms has been done using Gauss-legendre method with 501 number of grid points. Time-independent calculations have been done using 534 Cartesian basis functions at the Hartree-Fock (HF) level of theory, which is a sufficient description for the mean-field orbitals and the density.
Time-dependent Schrodinger equation for the many-electron system under the dipole approximation in the lab frame and length gauge will be:
\[i\frac{\partial}{\partial t}\Psi\left(\{\vec{r}_{j}\},t\right)=\left[-\frac{1}{2} \sum_{j}^{n}\nabla_{j}^{2}-\sum_{A}^{M}\sum_{j}^{n}\frac{Z_{A}}{\left|\vec{r}_{ j}-\vec{R}_{A}\right|}+\sum_{j>k}^{n}\frac{1}{\left|\vec{r}_{j}-\vec{r}_{k} \right|}-\sum_{j=1}^{n}\vec{r}_{j}.\vec{\epsilon}(t)\right]\Psi\left(\{\vec{r} _{j}\},t\right)\]
This has been used for the electronic dynamics and the harmonic generation spectra presented in the work. Time-dependent calculations have been done with 534 Cartesian basis functions which describe 4201 configuration state functions (CSF) using \((t,t^{\prime})\)-method. The MO-transformed Configuration interaction singles (CIS) Hamiltonian, dipole and complex absorbing potential (CAP) matrix have been extracted out from GAMESS-US. Our calculation therefore only includes singly excited states. The number of states taken is large and sufficient enough to describe the dynamics in the presence of the high frequencies that have been presented. The box type CAP matrix has been used for the ionization. The CAP has been put outside the classical turning points at \(x=19\) a.u., \(y=19\) a.u. and \(z=13\) a.u. The CAP integrals have been checked to be non-vanishing. The \((t,t^{\prime})\) method was developed by Peskin and Moiseyev. Here, a new fictitious time co-ordinate, denoted as \(t^{\prime}\), is introduced in an extended Hilbert Space as \(i\hbar\left[\frac{\partial}{\partial t}+\frac{\partial}{\partial t^{\prime}} \right]\tilde{\Psi}(\vec{r},t,t^{\prime})=\hat{H}(\vec{r},t^{\prime})\tilde{ \Psi}(\vec{r},t,t^{\prime})\). On the contour \(t=t^{\prime}\), this equation becomes equal to the physical time-dependent Schrodinger equation. It can be rearranged in the form:
\[i\frac{\partial}{\partial t}\tilde{\Psi}(\vec{r},t,t^{\prime})=\hat{\mathcal{H }}_{\mathcal{F}}(\vec{r},t^{\prime})\tilde{\Psi}(\vec{r},t,t^{\prime});\ \hat{\mathcal{H}}_{\mathcal{F}}(\vec{r},t^{\prime})=\left[\hat{H}(\vec{r},t^{ \prime})-i\hbar\frac{\partial}{\partial t^{\prime}}\right]\]
Here, a \(\hat{\mathcal{H}}_{\mathcal{F}}\) resembles a Floquet type operator in the \(t^{\prime}\) coordinate. The evolution equation of the wave function now takes the form:
\[\tilde{\Psi}(\vec{r},t,t^{\prime})=e^{-i\hat{\mathcal{H}}_{\mathcal{F}}(\vec{r },t^{\prime})(t-t_{0})}\tilde{\Psi}(\vec{r},t_{0},t^{\prime}).\]
The physical solution \(\Psi(\vec{r},t)\) can be extracted from the full solution \(\tilde{\Psi}(\vec{r},t,t^{\prime})\) at \(t=t^{\prime}\) and all the physical properties can then be calculated. The time-dependent Hamiltonian hasbeen solved for 65-optical cycles with a time step, \(dt\) of 0.01 a.u. High harmonic generation (HHG) spectra have been calculated using time-dependent induced dipole moments. The Fourier transform of the double derivative of the induced dipole moment gives the Harmonic generation spectra.
|
10.48550/arXiv.1911.06976
|
Bonding in light-induced vortices: benzene in a high-frequency circular polarized laser
|
Prashant Raj, Mishu Paul, Mythreyi R., Balanarayan Pananghat
| 5,025
|
10.48550_arXiv.2102.02427
|
## 1 Introduction
Quantum dynamics play a significant role in many chemical physics and biochemical physics problems. Frequently studied problems of this kind include exciton and electron transfer processes that are involved in photosynthetic systems, electron transfer, DNA, and photovoltaic systems. In these problems, the environments (baths), for example, proteins and solvents, play a central role; these baths are complex and strongly coupled to a molecular system of interest at finite temperatures. Recent theoretical works have demonstrated that such systems and baths are quantum mechanically entangled (b-thetanglement) and an understanding of these baths is essential to properly elucidate the quantum dynamics displayed by the system. For example, it has been shown that the optimal condition for excitation energy transfer in light-harvesting complexes is realized under non-Markovian system-bath interactions in a strong coupling regime, in which the noise correlation time of the bath is comparable to the time scale of the system dynamics. To conduct high-accuracy simulations with reduced computational costs, some approaches have utilized machine learning methods to develop models that reproduce open quantum dynamics, analyze two-dimensional spectroscopy images, and estimate chemical properties for classical molecular dynamics.
Although an irreversibility of the system dynamics results from quantum thermal activation and dissipation caused by the surrounding environment, it is difficult to conduct a quantum molecular dynamics simulation that exhibits such a characteristic feature arising from macroscopic degrees of freedom. Thus, we introduce a system-bath model in which the dynamics of excitons or electrons are described by a system Hamiltonian, while the other degrees of freedom that arise from environmental molecules are described by a harmonic-oscillator bath (HOB). The HOB, whose distribution takes a Gaussian form, exhibits wide applicability in simulating bath effects, despite its simplicity; this is because the influence of the environment can, in many cases, be approximated by a Gaussian process due to the cumulative effect of the large number of environmental interactions. In such a situation, the ordinary central limit theorem is applicable, and hence, the Gaussian distribution function is appropriate. The distinctive features of the HOB model are determined by the spectral distribution function (SDF) of the coupling strength between the system and the bath oscillators for various frequency values. By choosing the appropriate form of the SDF, the properties of the bath can be adjusted to represent various environments consisting of, for example, solid-state materials and protein molecules. Because the SDF can be different for different forms of a system Hamiltonian and system-bath coupling, it is difficult to find an optimized Hamiltonian associated with an optimized SDF, in particular for a bath describing a fluctuation in site-site interaction energy.
In a previous study, we employed a machine learning approach to construct a system-bath model for the intermolecular and intramolecular modes of molecular liquids using atomic trajectories obtained from molecular dynamics (MD) simulations. In this study, we extend the previous approach to investigate an exciton or electron transfer problem that is characterized by electronic states embedded in the molecular environment using quantum mechanics/molecular mechanics (QM/MM) calculations to determine the atomic coordinates of molecules. In particular, we focus on the exciton transfer process of the photosynthesis antenna system to investigate how natural systems can realize such highly efficient yields, presumably by manipulating quantum mechanical processes. As a demonstration, we consider a molecular dimer made of two dipole-coupled dye monomers as a model system that is often studied experimentally and theoretically. Then, we construct a model Hamiltonian of an indocarbocyanine dimer compound. The accuracy of this model is examined by calculating linear and two-dimensional electronic spectra.
This paper is organized as follows. In Section 2, we introduce a model that can be used for either exciton or electron transfer and is coupled to a harmonic heat bath. We then describe the machine learning approach that we use to determine the system parameters, the system-bath interactions, and the SDFs on the basis of QM/MM simulations. In Section 3, we present results for an indocarbocyanine dimer model constructed from the analysis of QM/MM trajectories. Linear absorption and two-dimensional spectra are calculated from analytical linear and nonlinear response functional expressions. Section 6 is devoted to concluding remarks.
## 2 Theory
### Hamiltonian
We consider the situations in which an exciton or electron transfer system interacts with molecular environments that give rise to dissipation and fluctuation in the system.
\[\hat{H}_{S}=\sum_{j}\hbar\omega_{j}|j\rangle\langle j|+\sum_{j\neq k}\hbar \Delta_{jk}|j\rangle\langle k|, \tag{1}\]
The interaction energy between the \(j\)th and \(k\)th states is described by \(\hbar\Delta_{jk}\). In our model, each state is coupled to a different molecular environment (labeledas \(a\)) that is treated as \(N_{a}\) harmonic oscillators.
\[H_{tot}= H_{\rm S}-\sum_{a}\sum_{l=1}^{N_{a}}\alpha_{l}^{a}\hat{V}^{a}\hat{x}_{l }^{a}\] \[+\sum_{a}\sum_{l=1}^{N_{a}}\left[\frac{(\hat{p}_{l}^{a})^{2}}{2m_{ l}^{a}}+\frac{1}{2}m_{l}^{a}(\omega_{l}^{a})^{2}(\hat{x}_{l}^{a})^{2}\right], \tag{2}\]
The system part of the system-bath interaction is expressed as
\[\hat{V}^{a}=\sum_{j,k}V_{jk}^{a}|j\rangle\langle k|, \tag{3}\]
The \(a\)th heat bath can be characterized by the spectral distribution function (SDF), defined as
\[J_{a}(\omega)\equiv\sum_{l=1}^{N_{a}}\frac{\hbar(\alpha_{l}^{a})^{2}}{2m_{l}^ {a}\omega_{l}^{a}}\delta(\omega-\omega_{l}^{a}), \tag{4}\]
Various environments, for example, those consisting of nanostructured materials, solvents, and protein molecules, can be modeled by adjusting the form of the SDF. For the heat bath to act as an unlimited heat source possessing an infinite heat capacity, the number of heat-bath oscillators \(N_{a}\) is effectively made infinitely large by replacing \(J_{a}(\omega)\) with a continuous distribution. The above model has been frequently used in the analysis of photosynthetic systems, electron transfer, DNA, and solar battery systems.
### Learning data: QM/MM simulations
We next consider the pigments in a molecular system, whose electric excitation or exciton states are described by Eq.. The electric states of the pigments depend on the configurations of the surrounding atoms at time \(t\). The time evolution of the excited statesof the system and environmental molecules are described by QM/MM simulations. Because our goal in constructing a system-bath model is to perform a full quantum simulation of the entire system, we should use quantum molecular dynamics (MD) simulations to provide data on the basis of all atomic coordinates. In practice, however, it is impossible to consider large environmental degrees of freedom accurately from a quantum mechanical perspective. Fortunately, we expect that we already have reasonable SDFs for quantum simulation, even though we evaluated them using the classical MD simulation. Such evaluations were conducted utilizing an ensemble of molecular trajectories that exhibit a Gaussian distribution in which the difference between the quantum and classical trajectories is expected to be minor. Further, the dynamics of harmonic oscillators are identical in both the classical and harmonic cases because both the classical and quantum Liouvillian for the \(l\)th oscillator in the \(a\)th bath are expressed as \(\hat{L}_{l}^{a}=-(p_{l}^{a}/m_{l}^{a})(\partial/\partial x_{l}^{a})-m(\omega_{ l}^{a})^{2})(\partial/\partial p_{l}^{a})\). We thus use the classical MD simulation technique to acquire the atomic coordinates of the pigments and the molecular environment. We then conduct quantum chemistry calculations to obtain the desired electronic states, typically the highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO) states of the pigments as a function of time. The excited energy of the \(j\)th pigment is denoted by \(\epsilon_{jj}(t)\), and the interaction energy between the \(j\)th and \(k\)th pigments that includes the bath-induced fluctuation is denoted by \(\epsilon_{jk}(t)\); these values can be obtained using any kind of numerical program for quantum chemistry calculations. If the main system is too large to enable evaluation of all electronic states, we evaluate the site energy \(\epsilon_{jj}(t)\) and the interaction energy \(\epsilon_{jk}(t)\) separately. From the calculated \(\epsilon_{jj}(t)\) and \(\epsilon_{jk}(t)\), we evaluate the system-bath coupling strength in \(\hat{V}_{jk}^{a}\) and its SDF, in addition to the excitation energy \(\hbar\omega_{j}\) and the interaction energy \(\Delta_{jk}\), based on the machine learning approach.
While the SDFs evaluated based on the MD simulations are temperature dependent, the SDFs for the HOB are temperature independent; we therefore eliminate the temperature dependence of optimized parameters, assuming that the sampled MD trajectories exhibit canonical ensembles at finite temperatures.
### Machine Learning
For \(n\) exciton or electronic excitation sites, we express the simulated data in terms of \(\epsilon_{jk}(t)\) when describing the excited and site-site interaction energies of interest obtained from the QM/MM simulation.
\[H(t)=\sum_{j,k=1}^{n}\epsilon_{jk}(t)|j\rangle\langle k|. \tag{5}\]
We then attempt to reproduce the trajectories of \(\epsilon_{jk}(t)\) for the total Hamiltonian, Eq., with Eqs. and. Although the system-bath model considers an infinite number of degrees of freedom, here, we employ a finite number of bath oscillators to estimate the SDFs. Then, the sampling used for machine leering training is considered the average of the classical bath oscillators for a certain selection of the system and system-bath parameters.
\[\epsilon_{jj}(t)=\hbar\omega_{j}-\delta_{jj}(t) \tag{6}\]
and
\[\epsilon_{jk}(t)=\hbar\Delta_{jk}-\delta_{jk}(t), \tag{7}\]
respectively, where \(\delta_{jk}(t)\) is expressed in terms of the linear function of the bath coordinates as
\[\delta_{jk}(t)=\sum_{a}\alpha_{jk}^{a}x_{jk}^{a}(t). \tag{8}\]
Here, the \(a\)th bath coordinate for the \(jk\) site is described as a function of time as
\[x_{jk}^{a}(t)=A_{jk}^{a}\sin\left(\phi_{jk}^{a}+\omega_{jk}^{a}t\right), \tag{9}\]where \(A_{jk}^{a}\) and \(\phi_{jk}^{a}\) are the amplitude and phase of the \(a\)th bath oscillator for the \(jk\) site, respectively. The phase \(\phi_{jk}^{a}\) is randomly chosen to avoid recursive oscillator motion. Although we can consider such correlated modes separately by introducing additional baths, here, we assume that the influences of the individual bath modes are all independent and that the correlations between the fluctuations among different modes can be ignored.
From Eqs.
\[\delta_{jk}(t)=\sum_{a}c_{jk}^{a}\sin\left(\phi_{jk}^{a}+\omega_{jk}^{a}t\right), \tag{10}\]
where
\[c_{jk}^{a}=\alpha_{jk}^{a}A_{jk}^{a} \tag{11}\]
In the machine learning context, the bath parameters and the system-bath interactions are expressed as a set of latent variables, defined as
\[\theta=(\left\{\omega_{j}\right\},\left\{\Delta_{jk}\right\},\left\{c_{jk}^{a} \right\}), \tag{12}\]
The trajectories of \(\epsilon_{jj}(t)\) and \(\epsilon_{jk}(t)\), obtained from the QM/MM calculations, are described as the vibrational motions of the pigment molecule and the surrounding molecules.
\[P(\lambda_{i}\mid\theta)=\int\prod_{k,j,a}d\phi_{jk}^{a}P(\lambda_{i}\mid \theta;\phi_{jk}^{a})P(\phi_{jk}^{a}), \tag{13}\]
Here, \(P(\phi_{jk}^{a})\) is the uniform distribution of \([0,2\pi)\) and
\[P(\lambda_{i}|\theta;\phi_{jk}^{a})\propto\exp\left[-\sigma\left(\lambda_{i}-E_{i} \right)^{2}\right], \tag{14}\]
Our goal in employing a machine learning method is to choose the optimal parameter set in \(E_{i}(\theta;\phi_{jk}^{a})\) that maximizes the probability distribution for given data \(\lambda_{i}\).
\[L=\sum_{i}(\lambda_{i}-E_{i})^{2}. \tag{15}\]
To find the maximum value of \(L\), we employ the Adam gradient method for optimization of the parameter set as
\[\theta\leftarrow\theta+\gamma\frac{\partial L}{\partial\theta}, \tag{16}\]
In this way, we obtain the \(J_{jk}\) element of the SDF for the \(jk\) site.
\[\left\langle A_{jk}^{a}\right\rangle=\frac{1}{\sqrt{\pi\beta m_{jk}^{a}\left( \omega_{jk}^{a}\right)^{2}}}. \tag{17}\]
Integrating Eqs. and into Eq., we obtain
\[J_{jk}(\omega)=\sum_{a=1}^{N_{a}}\frac{1}{2}\pi\beta\hbar\omega_{jk}^{a}(c_{jk }^{a})^{2}\delta(\omega-\omega_{jk}^{a}). \tag{18}\]
Because \(J_{jk}(\omega)\) rapidly changes over time in accordance with the structural changes in the pigment molecules and environments, we evaluate \(J_{jk}(\omega)\) by averaging the different sample trajectories. From a mathematical perspective, \(c_{jk}\) is the frequency domain expression of the time domain data, and \(J_{jk}(\omega)\) can be obtained by averaging the power spectra \(c_{jk}^{2}\) using the Wiener-Khinchin theorem.
It should be noted that the absolute intensity of \(J_{jk}(\omega)\) cannot be determined in the framework of the present study because, for simplicity, we do not evaluate the dipole moment of this complex material; we evaluate the intensity of \(J_{jk}(\omega)\) from the width of the experimentally obtained linear absorption spectrum.
## 3 Numerical demonstration
### Indocarbocyanine dimer
We now demonstrate our numerical approach for a dimer of identical indocarbocyanine molecules. displays the structure of the pigment molecule. The ground and excited states of each pigment are expressed as \(\left|0\right\rangle_{j}\) and \(\left|1\right\rangle_{j}\) for \(j=1\) and \(2\), respectively.
The molecular structure of the indocarbocyanine dimer. Two pigments are connected by methylene chains. The gray/blue/white atoms represent carbon/nitrogen/hydrogen, respectively. The red square represents pigment 1, whereas the blue square represents pigment 2.
The system Hamiltonian is then expressed as
\[\hat{H}= \omega_{0}\left(|1\rangle_{11}\langle 1|+|1\rangle_{22}\langle 1|\right)\] \[+\Delta\left(|0\rangle_{12}\langle 1|+|1\rangle_{12}\langle 0| \right), \tag{19}\]
By diagonalizing \(H\), we obtained the eigenvalues \(\omega_{k}\) for the \(k=+\) and \(-\) eigenstates of \(|1+\rangle=(|1\rangle_{1}|0\rangle_{2}+|0\rangle_{1}|1\rangle_{2})/\sqrt{2}\) and \(|1-\rangle=(|1\rangle_{1}|0\rangle_{2}-|0\rangle_{1}|1\rangle_{2})/\sqrt{2}\), respectively, as
\[\omega_{\pm}=\omega_{0}\pm\Delta. \tag{20}\]
The excitation energy and interaction energy fluctuations as functions of time, arising from intramolecular motions of the pigment and intermolecular motions of surrounding molecules, are expressed as \(\delta\omega_{\pm}(t)\) and \(\delta\Delta(t)\), respectively. These functions are evaluated based on the quantum chemistry calculations for given atomic trajectories of the entire molecular system determined by MD simulations.
In our model, because each exciton state is delocalized and the effects of the environmental modes are site specific, we employ an individual heat bath expressed as the sum of site-specific oscillators to describe the energy fluctuation at each exciton site. The distribution of the exciton-oscillator coupling strength is then evaluated based on the machine learning approach. Although it is possible to introduce a global heat bath to induce low-frequency environmental modes that are coupled to multiple exciton states, we find that such effects are not significant in the present case.
\[\delta\omega_{\pm}(t)=\sum_{m}^{1,2}w_{\pm,m}(t)\sum_{a}c_{\omega_{0}m}^{a} \sin(\omega^{a}t+\phi_{\omega_{0}m}^{a}), \tag{21}\]\[\delta\Delta(t)=w_{\Delta}(t)\sum_{a}c_{\Delta}^{a}\sin(\omega^{a}t+\phi_{\Delta}^{a }), \tag{22}\]
We introduce the localization weight functions \(w_{\pm,m}(t)\) and \(w_{\Delta}(t)\), as obtained from the diagonalization of the pigment-based Hamiltonian, expressed in Eq., to describe the pigment-specific environment effects in the delocalized exciton state representation. These localization weight functions are evaluated based on the electronic states of the pigment \(m=1\) and \(2\) established by the atomic orbitals (AO) obtained from quantum chemistry calculations.
Thus, the targeting eigenenergies to be described by the system-bath model, \(\lambda_{\pm}(t;\theta)\), are expressed as
\[\lambda_{\pm}(t;\theta)=\omega_{0}+\delta\omega_{\pm}(t)\pm\left(\Delta+\delta \Delta(t)\right), \tag{23}\]
As learning data, we compute the exciton energy \(E_{\pm}(t)\), the molecular orbital (MO) coefficients for each exciton state, and wavefunctions (atomic orbital (AO) coefficients for each MO) from quantum chemistry calculations for the given atomic coordinates as a function of time. Additionally, the movements of all atoms in the system are evaluated from the classical MD simulation. Using these data, we optimize the set of parameters \(\theta\). To evaluate the weight function \(w_{k,m}(t)\), we calculate the exciton and hole populations \(p_{k,m}^{ex}(t)\) and \(p_{k,m}^{h}(t)\) that are obtained as the summation of the absolute square of the AO coefficients, which are evaluated from the AO coefficients involved in the MO in pigment \(m\) for excited state \(k\). The weight function is then evaluated as \(w_{k,m}(t)=p_{k,m}^{ex}(t)p_{k,m}^{h}(t)\) and \(w_{\Delta}(t)=\sum_{k=\pm}\left(p_{k,1}^{h}(t)p_{k,2}^{ex}(t)+p_{k,2}^{h}(t)p_ {k,1}^{ex}(t)\right)\). As these definitions indicate, the exciton states are localized when \(w_{\pm,m}\) is close to \(1\), whereas the exciton states are distributed among the pigments when \(w_{\Delta}\) is close to \(1\).
To optimize the system and bath parameter set, we minimize the loss function
\[L =\sum_{n}\sum_{t}L^{n}(t)\] \[=\sum_{n}\sum_{t}\left[(\lambda_{-}(t;\theta^{n})-E_{-}^{n}(t))^{2} +(\lambda_{+}(t;\theta^{n})-E_{+}^{n}(t))^{2}\right], \tag{24}\]
Using the MLE method, we optimize \(c_{\omega_{0}m}^{a}\) and \(c_{\Delta}^{a}\) for each time series as a sample set. To apply the machine learning algorithm, the time series of the tuple \((E_{-}^{n}(t),E_{+}^{n}(t),w_{k,m}^{n}(t))\) are regarded as the input feature variables. In the indocarbocyanine case, the two pigments are symmetric, and the bath SDFs for each pigment are considered to be identical. Therefore, we use the averaged value \(c_{\omega_{0}}^{a}=(c_{\omega_{0}1}^{a}+c_{\omega_{0}2}^{a})/2\). We then evaluate \(J_{jj}(\omega)(j=1,2)\) and \(J_{12}(\omega)\), namely, \(J(\omega)\) for \(\omega_{0}\) and \(\Delta\), from \(c_{\omega_{0}}^{a}\) and \(c_{\Delta}^{a}\), respectively, using Eq..
### Fourier-based approach versus machine learning approach
A commonly used approach for evaluating the SDFs of \(\epsilon_{ij}(t)\) utilizes the Fourier transformation of the autocorrelation function expressed as \(\mathcal{F}\left[\langle\delta\epsilon_{ij}\delta\epsilon_{ij}(t)\rangle\right]\), where \(\delta\epsilon_{ij}(t)\equiv\epsilon_{ij}(t)-\langle\epsilon_{ij}\rangle\).
\[C_{ij}(t)=\frac{1}{N}\sum_{n}\left\langle\delta\epsilon_{ij}^{n}\delta \epsilon_{ij}^{n}(t)\right\rangle, \tag{25}\]
We then obtain the SDF as
\[J_{ij}(\omega)=\mathcal{F}\left[C_{ij}(t)\right]. \tag{26}\]Alternatively, using Wiener-Khinchin's theorem for stationary random processes, we can obtain the SDF as an average of power spectrum \(P_{ij}^{n}(\omega)=\left|\mathcal{F}\left[\epsilon_{ij}(t)\right]\right|^{2}\) as
\[J_{ij}(\omega)=\frac{1}{N}\sum_{n}P_{ij}^{n}(\omega). \tag{27}\]
Although this Fourier-based approach is simple and straightforward, for the system-bath Hamiltonian, the obtained SDFs are not necessarily the optimal choice for describing the QM/MM data because the exciton and interaction energies are mutually dependent on each other; thus, \(J_{ij}(\omega)\) and \(J_{ik}(\omega)\) cannot be evaluated separately. In the machine learning approach, however, it is possible to optimize not only \(J_{ij}(\omega)\) and \(J_{ik}(\omega)\) but also \(\omega_{0}\) and \(\Delta\) without assuming explicit relationships between the SDFs and the system parameters. Moreover, if necessary, we can introduce additional conditions for optimization of the SDFs and system parameters because we employed \(w_{k,m}(t)\) to account for the effects of the indocarbocyanine dimer exciton localization.
### Calculation Details
## Step 1: Classical MD
We prepared a system consisting of an indocarbocyanine dimer molecule with 1024 methanol molecules as the solvent. The classical MD simulations were carried out with the GROMACS software package. The conditions for preparation MD simulations were set as 1 atm and 300 K with an NPT ensemble. The equilibrium MD run was carried out for 20 ps in an NVT ensemble followed by a sampling MD run for 5 ps in an NVE ensemble. These equilibrium MD runs and sampling MD runs were repeated 100 times. The entire MD simulation was performed with a time step of 0.1 fs.
## Step 2: Data Preparation using Quantum Chemistry Calculations
To obtain the sample trajectories of the excitation energies, we conducted ZINDO calculations and natural transition orbital analysis for a 1 fs period in one sample using the ORCA software package. We then obtained 100 \((E_{-}(t),E_{+}(t),w(t))\) samples that were 5 ps in length.
## Step 3: Parameter Optimization for the Machine Learning Approach
We arrange the data with 5 ps lengths obtained from step 2 according to the starting time in each of 175 steps. We then extracted 604 trajectories containing 1000 data points in an interval of 4 fs. These sampling data were used as the input feature values in the machine learning calculations. To perform learning calculations, we developed Python codes using the TensorFlow library. The training was performed with the learning rate \(\alpha=1\times 10^{-4}\) for the first 200 steps and then the rate was reduced to \(\alpha=1\times 10^{-5}\) for the next 200 steps. The number of epochs was chosen to avoid the overfitting problem arising from the MLE that occurs with a gradient method. In the present case, this effect appears in the very law frequency region below 10 cm\({}^{-1}\) of \(J(\omega)\) (see Appendix B). Because such slow dynamics of the environment are not important in the present exciton transfer problem, we avoided this effect by simply choosing a shorter epoch known as the early-stopping technique. To minimize the loss function, we employed the Adam algorithm. The bath oscillator number \(N\) is 600. The frequency of the \(a\)th bath oscillator \(\omega^{a}\) is \(a\Delta\omega\) for \(a=1,2,\cdots,N\), where \(\Delta\omega\) is approximately 8.3391cm\({}^{-1}\).
The initial values of the target optimization variables for the SDF amplitudes were set as \(c^{a}_{\omega_{0}m}=1\times 10^{-5}\) and \(c^{a}_{\Delta}=1\times 10^{-5}\), and the exciton and interaction energies were set as \(\omega_{0}=(\langle E_{+}\rangle+\langle E_{-}\rangle)/2\) and \(\Delta=(\langle E_{+}\rangle-\langle E_{-}\rangle)/2\). The initial phases \(\phi^{a}_{b}\) were randomized 5 times for each series of samples. The loss functions were averaged over each set of 64 samples as a minibatch, while the parameters were optimized for every minibatch. For the 604 samples, each epoch contained 9 iterations.
## Step 4: Calculations of Optical Spectra
We assumed that the dipole operator for the indocarbocyanine dimer was given by \(\hat{\mu}_{1}+\hat{\mu}_{2}=\mu(|0\rangle_{11}\langle 1|+|1\rangle_{11}\langle 0|+|0 \rangle_{22}\langle 1|+|1\rangle_{22}\langle 0|)\), which created a transition between the ground state \(|00\rangle\) and the excitation states, \(|1+\rangle\), and \(|11\rangle\), while optical transitions from these states to the state \(|1-\rangle\) were forbidden. Thus, the optical transitions in the present system were modeled by a three-level system with eigenenergies of 0, \(\Omega_{+}\), and \(2\omega_{0}\). This allowed us to apply analytical expressions of the linear and nonlinear response functions, as presented in appendix A. We then calculated the linear absorption and two-dimensional (2D) electronic spectroscopy signals using line-shape functions.
### Results and Discussion
Representative examples of the prepared dataset are plotted in The abrupt change in the exciton energies in Fig. 2(b) occurs due to the exciton transfer between pigments 1 and 2 that takes place in the time period of 10-100 fs. As illustrated in Fig. 2(c), the difference in the exciton energies \(E_{+}-E_{-}\) exhibits minima in accordance with respect to the exciton transfer processes. As depicted by the red circles in Fig. 2(c), although such minimal points are significantly narrower and deeper than the minimal point caused by energetic fluctuation, it is difficult to separate the effects of exciton transfer from the energy fluctuation due to environmental motions. By introducing the localization weight functions \(w_{\pm,m}(t)\) and \(w_{\Delta}(t)\) in Eqs. and to eliminate the effects of the nonenvironmental origin involved in the learning trajectories, we can stabilize and enhance the efficiency of the machine learning process.
In Fig. 3, we depict the learning curve of the loss function, as defined in Eq.. Upon gathering random samplings of \(\phi\), the loss function converged monotonically to a certain positive value, which demonstrated the efficiency of the present algorithm. The initial parameter values of the excitation energy and the interaction energy were set as \(\omega_{0}\) = 17736 cm\({}^{-1}\) and \(\Delta\) = 1004 cm\({}^{-1}\), whereas the optimized values of the excitation energySamples of the data used in learning calculations for (a) the excitation energy \(E_{k}\) for \(k=\pm\), (b) the weight functions \(w_{\Delta}\) (green dashed curve), \(w_{k,m}(t)\) for pigments \(m=1\) (blue) and \(m=2\) (orange) for \(k=+\) (solid line) and \(k=-\) (dotted line), respectively. The panel (c) is plotted the differences between the energy levels \(E_{\pm}\) to illustrate the relationship between the energies and the weight functions.
The SDFs of the indocarbocyanine dimer in the methanol environment for the exciton energy \(J_{11}(\omega)\) (\(=J_{22}(\omega)\)) (blue) and the interaction energy \(J_{12}(\omega)\) (orange) obtained with the machine learning approach.
The learning curve of the loss function for the indocarbocyanine dimer model. The vertical line at epoch 200 indicates the epoch where the learning rate changed, and the vertical line at epoch 400 indicates early stopping.
Energy-level diagram for a dimer system that undergoes random fluctuations in the excited energy and coupling strength described by \(\delta\omega(t)\) and \(\delta\Delta(t)\), respectively. For the description of pure dephasing, only the difference between the energies involved in the optical excitation is important: the frequency fluctuation between \(|00\rangle\) and \(|1+\rangle\) is given by \(\delta\omega(t)+\delta\Delta(t)\), whereas that between \(|1+\rangle\) and \(|11\rangle\) is given by \(\delta\omega(t)-\delta\Delta(t)\). For perfectly uncorrelated fluctuations, we consider \(\delta\omega(t)\) and \(\delta\Delta(t)\) independently. The dashed line represents the energy level of the forbidden state \(|1-\rangle\).
Here, to avoid overfitting problems, we employed the early-stopping technique (see Appendix B).
In Fig. 4, we display the results of SDFs for the excitation energy \(J_{11}(\omega)\) (= \(J_{22}(\omega)\)) and the interaction energy \(J_{12}(\omega)\). Various intermolecular modes below 2000 cm\({}^{-1}\) are observed as prominent sharp peaks near 450, 570, 1185, 1393, 1541, 1791, 1842, and 1923 cm\({}^{-1}\). In the region above 2000 cm\({}^{-1}\), only two tiny peaks are observed at approximately 3000 cm\({}^{-1}\) and 3850 cm\({}^{-1}\). The normal mode analysis (B3LYP/def-SV(P)) indicates that these peaks under 3300 cm\({}^{-1}\) arise from the intramolecular modes of the indocarbocyanine dimer, whereas the peak at 3850 cm\({}^{-1}\) arises from a molecular vibration of the solvent methanol molecules. We found that each sharp peak can be fitted by the Brownian spectral distribution, whereas the broadened background peak in the range from 0 to 2000 cm\({}^{-1}\) corresponds to the intramolecular modes fitted by the Drude-Lorentz distribution. The intensities of the peaks in \(J_{12}(\omega)\) are considerably weaker than those in \(J_{11}(\omega)\): only the peaks near 456, 562, 1840 and 1920 cm\({}^{-1}\) are identified. As we expected, the intermolecular peak positions are governed by the classical MD simulation, whereas the heights of these peaks are predominately governed by the quantum chemistry calculation.
To verify the descriptions of the obtained SDFs and system parameters, we computed the linear absorption and two-dimensional electronic spectra (2DES), for the cases in which the experimentally obtained spectra were available. In general, these spectra should be calculated in the framework of open quantum dynamics that considers the complex interactions between the exciton sites. However, for demonstration purposes here, we employ the analytical expressions for response functions, ignoring the transitions to the state that are usually forbidden. The details of these calculations are presented in Appendix A.
The linear absorption spectrum calculated from Eqs. (A.1 ) and A.2 is presented in Here, the calculated peak is fitted by the Gaussian function \(\lambda\exp\left[-\left((\omega-\omega_{c})/\gamma\right)^{2}\right]\), where the amplitude, central frequency, and width are \(\lambda=351\), \(\omega_{c}=18583\)cm\({}^{-1}\), and \(\gamma=464\)cm\({}^{-1}\)respectively. Note that we could not determine the absolute SDF intensities because, for simplicity, we did not calculate the amplitude of the dipole operator. Here, we chose to use the intensity of \(J_{11}(\omega)\) to fit the experimentally obtained signal. As presented in Fig. 6, we observe a single broadened absorption peak at \(\omega_{0}+\Delta\) corresponding to the transition between \(|00\rangle\) and \(|1+\rangle\), while the transition between \(|00\rangle\) and the state \(|1-\rangle\) is forbidden (see Fig. 5). Although the experimentally observed linear absorption spectrum exhibits a \(0-1\) phonon sideband peak near \(\omega=19500\)cm\({}^{-1}\), here, we observe this phenomenon only as an asymmetry of the Gaussian peak in the high-frequency region.
The 2D correlation electronic spectra calculated using the analytical expressions of the response function (Eqs. (A.5 ) -A.7 ) are presented in At \(t_{2}=0\) fs, only one peak stretched near the \(\omega_{1}=\omega_{3}\) line, arising from the \(|00\rangle\rightarrow|1+\rangle\) transition, is observed. At \(t_{2}=10\), 25, and 40 fs, the peak is elongated in the low-frequency \(\omega_{1}\) direction due to a shift in the eigenenergy caused by the heat-bath-induced exciton-exciton interaction described by \(J_{12}(\omega)\). Because the system-bath interaction we considered here is non-Markovian and its effects appear only after a period longer than the inverse correlation time of noise, we do not observe such heat-bath effects for a small \(t_{2}\).
Linear absorption spectrum of an indocarbocyanine dimer, as calculated with Eqs. (A.1 )-A.2 and the line-shape function Eq. (A.3 ) for the system parameters and SDFs obtained with the machine learning approach. The dotted line is the fitted Gaussian peak centered at 18583cm\({}^{-1}\), indicating that the calculated peak is asymmetric due to the \(0-1\) phonon transition near 19500cm\({}^{-1}\).
2DES of an indocarbocyanine dimer, as calculated with Eqs. (A.5 ) -A.7 and the line-shape function Eq. (A.3 ) for the system parameters and SDFs obtained with the machine learning approach. The waiting time \(t_{2}\) for each signal is displayed at the top left of each panel. The peak intensity of the signal was normalized for each \(t_{2}\). The waiting time \(t_{2}\) was chosen to illustrate the maximal/minimal points of the oscillating feature of the peak elongation (see text).
As \(t_{2}\) increases, the intensities of these two peaks oscillate as a result of the population transitions among \(|10\rangle\), \(|1+\rangle\), and \(|11\rangle\) caused by \(\Delta\) and \(J_{12}(\omega)\). This phenomenon was also observed experimentally. The appearance of this oscillatory feature at a finite period in \(t_{2}\) indicates the importance of the off-diagonal heat bath, whose modeling is not easy in the framework of the existing approach.
While the off-diagonal peak still exhibits oscillatory motion at \(t_{2}\geq 100\) fs, the peak profile gradually elongates in the \(\omega_{1}=\omega_{3}\) direction due to the inhomogeneous broadening that arises from the diagonal bath modulation described by \(J_{11}(\omega)\) and \(J_{22}(\omega)\).
## 4 Conclusion
We introduced a machine learning approach for constructing a model that can be used to analyze the dynamics of exciton or electron transfer processes in a complex environment on the basis of considering the energy eigenstates evaluated from QM/MM simulations as functions of time. The key feature of the present study is the system-bath model, in which the primary exciton/electron dynamics are described by a system Hamiltonian expressed in terms of discretized energy states, while the other degrees of freedom are described by harmonic heat baths that are characterized by SDFs. An optimized system-bath Hamiltonian obtained from the machine learning approach allows us to conduct time-irreversible quantum simulations that are not feasible with a full quantum MD simulation approach.
Here, we demonstrated the above features by calculating linear and nonlinear optical spectra for the indocarbocyanine dimer system in a methanol environment in which the quantum entanglement between the system and bath plays a central role. The calculated results can be used to explain the experimental results reasonably well; we found that the heat bath plays a key role in describing the exciton transfer process for the exciton-exciton interaction in this system. Although here we ignore the transitions to the state that are usually forbidden due to an applicability of the analytical expression, if necessary, we can explicitly consider such transitions using the HEOM formalism.
Finally, we briefly discuss possible extensions of this study. As shown in a previous paper, the machine learning approach can be applied to a system described by reaction coordinates, which is useful for investigating chemical reaction processes characterized by potential energy surfaces. By combining the previous and present approaches, we can further investigate systems described by not only electronic states but also molecular configuration space, for example, photoisomerization, molecular motor, and nonadiabatic transition problems, with frameworks based on the system-bath model. In this way, we may construct a system-bath model for entire photosynthesis reaction processes consisting of photoexcitation, exciton transfer, electron transfer, and proton transfer processes, including conversion processes, such as exciton-coupled electron transfer and electron coupled proton transfer processes.
Further theoretical and computational efforts must be put forth that include providing learning data based on accurate and large quantum simulations, improving learning algorithms, and developing an accurate and efficient open quantum dynamics theory to treat a complex system-bath model. We leave such additional endeavors to future studies in accordance with recent progress in theoretical techniques.
The authors are thankful to Professor Yuki Kurashige for helpful discussions concerning the QM/MM simulations for providing an indocarbocyanine dimer system. Financial support from HPC Systems Inc. is acknowledged.
One-dimensional and two-dimensional spectra
Linear and nonlinear optical spectra can be expressed in the Fourier transformation of the response functions. In the present dimer case, we can analytically express the response functions in terms of a line-shape function including the contribution from an exciton-exciton interaction.
The linear absorption spectrum is given by
\[S(\omega)=\int_{0}^{\infty}dte^{i\omega t}R^{}(t)-c.c.,\] (A.1)
For a coupled dimer system, the analytical expression for the response function for Eq. (A.1 ) is expressed as
\[R^{}(t_{1})= \frac{i\mu}{\hbar}\exp[i\Omega_{+}t_{1}-g_{-}^{11}(t_{1})-g_{-}^{ 12}(t_{1})]\] \[-c.c.\] (A.2)
where the line-shape function, \(g_{\pm}^{a}(t)\), for the SDF, \(J_{a}(\omega)\), with \(a=11\) and \(12\) is given by
\[g_{\pm}^{a}(t) \equiv\int_{0}^{t}d^{\prime}t\int_{0}^{t^{\prime}}dt\int\frac{d \omega}{2\pi}\] \[\times J_{a}(\omega)\left[\coth\left(\frac{\beta\hbar\omega}{2} \right)\cos(\omega t)\pm i\sin(\omega t)\right],\] (A.3)
For the coupled dimer system, the third-order response function is
\[R^{}(t_{3},t_{2},t_{1})=\langle[\mu(t_{3}),[\mu(t_{2})\,[\mu(t_{1}),\mu ]]]\rangle\] (A.4)Figure A.0: Double-sided Feynman diagrams for the third-order response functions \(R^{}(t_{3},t_{2},t_{1})\). In each diagram, time increase from the bottom to the top, and \(t_{i}\) represents the time intervals for the \(i\)th sequence between successive laser–system interactions. The left line represents the time evolution of the ket, whereas the right line represents that of the bra. The black, red and blue lines indicate that the system is in the \(|00\rangle\), \(|1+\rangle\), and \(|11\rangle\) states, respectively. For these systems, the complex conjugate paths, which can be obtained by interchanging the ket and bra diagrams, are not shown here.
and can also be evaluated in the analytical form as
\[R^{}(t_{3},t_{2},t_{1})\equiv\frac{i\,\mu^{4}}{h^{3}}\sum_{\alpha=1}^{8}\exp[Q _{\alpha}(\mathrm{t})]-c.c.\] (A.5)
where
\[Q_{1}(t)= -i\Omega_{+}t_{1}-i\Omega_{+}t_{3}\] (A.6) \[-f_{1}^{1}(t_{1},t_{2},t_{3})-f_{1}^{2}(t_{1},t_{2},t_{3}),\] \[Q_{2}(t)= i\Omega_{+}t_{1}-i\Omega_{+}t_{3}\] \[-f_{2}^{1}(t_{1},t_{2},t_{3})-f_{2}^{2}(t_{1},t_{2},t_{3}),\] \[Q_{3}(t)= i\Omega_{+}t_{1}-i\Omega_{+}t_{3}\] \[-f_{3}^{1}(t_{1},t_{2},t_{3})-f_{3}^{2}(t_{1},t_{2},t_{3}),\] \[Q_{4}(t)= -i\Omega_{+}t_{1}-i\Omega_{+}t_{3}\] \[-f_{4}^{1}(t_{1},t_{2},t_{3})-f_{4}^{2}(t_{1},t_{2},t_{3}),\] \[Q_{5}(t)= -i\Omega_{+}t_{1}+i\Omega_{-}t_{3}\] \[-f_{5}^{1}(t_{1},t_{2},t_{3})-f_{5}^{2}(t_{1},t_{2},t_{3}),\] \[Q_{6}(t)= i\Omega_{+}t_{1}+i\Omega_{-}t_{3}\] \[-f_{6}^{1}(t_{1},t_{2},t_{3})-f_{6}^{2}(t_{1},t_{2},t_{3}),\] \[Q_{7}(t)= i\Omega_{+}t_{1}+2i\omega_{0}t_{2}+i\Omega_{-}t_{3}\] \[-f_{7}^{1}(t_{1},t_{2},t_{3})-f_{7}^{2}(t_{1},t_{2},t_{3}),\] \[Q_{8}(t)= -i\Omega_{+}t_{1}-2i\omega_{0}t_{2}-i\Omega_{+}t_{3}\] \[-f_{8}^{1}(t_{1},t_{2},t_{3})-f_{8}^{2}(t_{1},t_{2},t_{3})\]with
\[\begin{split}& f_{1}^{a}(t_{1},t_{2},t_{3})=-g_{-}^{a}(t_{1})-g_{+}^{ a}(t_{3})\\ &\quad-[g_{+}^{a}(t_{2})-g_{+}^{a}(t_{23})-g_{-}^{a}(t_{12})+g_{-}^{a}(t_{123})],\\ & f_{2}^{a}(t_{1},t_{2},t_{3})=-g_{+}^{a}(t_{1})-g_{+}^{a}(t_{3}) \\ &\quad+[g_{-}^{a}(t_{2})-g_{-}^{a}(t_{23})-g_{+}^{a}(t_{12})+g_{+} ^{a}(t_{123})],\\ & f_{3}^{a}(t_{1},t_{2},t_{3})=-g_{+}^{a}(t_{1})-g_{-}^{a}(t_{3} )\\ &\quad+[g_{+}^{a}(t_{2})-g_{+}^{a}(t_{23})-g_{+}^{a}(t_{12})+g_{+ }^{a}(t_{123})],\\ & f_{4}^{a}(t_{1},t_{2},t_{3})=-g_{-}^{a}(t_{1})-g_{-}^{a}(t_{3} )\\ &\quad-[g_{-}^{a}(t_{2})-g_{-}^{a}(t_{23})-g_{-}^{a}(t_{12})+g_{- }^{a}(t_{123})],\\ & f_{5}^{a}(t_{1},t_{2},t_{3})=-g_{-}^{a}(t_{1})-g_{+}^{a}(t_{3} )\\ &\quad-[g_{+}^{a}(t_{2})-g_{+}^{a}(t_{23})-g_{-}^{a}(t_{12})+g_{- }^{a}(t_{123})],\\ & f_{6}^{a}(t_{1},t_{2},t_{3})=-g_{+}^{a}(t_{1})-g_{+}^{a}(t_{3} )\\ &\quad+[g_{-}^{a}(t_{2})-g_{-}^{a}(t_{23})-g_{+}^{a}(t_{12})+g_{+ }^{a}(t_{123})],\\ & f_{7}^{a}(t_{1},t_{2},t_{3})=-g_{+}^{a}(t_{1})-g_{-}^{a}(t_{3} )\\ &\quad+[g_{+}^{a}(t_{2})-g_{+}^{a}(t_{23})-g_{+}^{a}(t_{12})+g_{+ }^{a}(t_{123})],\\ & f_{8}^{a}(t_{1},t_{2},t_{3})=-g_{-}^{a}(t_{1})-g_{-}^{a}(t_{3} )\\ &\quad-[g_{-}^{a}(t_{2})-g_{-}^{a}(t_{23})-g_{-}^{a}(t_{12})+g_{- }^{a}(t_{123})].\end{split}\] (A.7)
Here, \(t_{12}\equiv t_{1}+t_{2}\), \(t_{23}\equiv t_{2}+t_{3}\), and \(t_{123}\equiv t_{1}+t_{2}+t_{3}\). Because the fluctuationis in the \(|1+\rangle\) and \(|11\rangle\) states are described by \(J_{11}(\omega)+J_{12}(\omega)\) and \(J_{11}(\omega)+J_{22}(\omega)=2J_{11}(\omega)\), respectively, the line-shape function \(g_{\pm}^{a}(t)\) in Eq. (A.7 ) is now expressed as \(g_{\pm}^{1}(t)=g_{\pm}^{11}(t)+g_{\pm}^{12}(t)\), and \(g_{\pm}^{2}(t)=2g_{\pm}^{11}(t)\). By using third-order diagrams, the pump-probe spectrum and photon echo spectra are, for example, calculated from the \(Q_{1}(t)\), \(Q_{4}(t)\), and \(Q_{5}(t)\) elements, and the \(Q_{2}(t)\), \(Q_{3}(t)\), and \(Q_{6}(t)\) elements, respectively.
Although the change in the exciton population can be explored by pump-probe spec troscopy, If we wish to investigate not only population dynamics but also system-bath coherence, two-dimensional electronic correlation spectroscopy is a better choice.
\[I^{\rm(corr)}(\omega_{3},t_{2},\omega_{1})\] \[=I^{\rm(NR)}(\omega_{3},t_{2},\omega_{1})+I^{\rm(R)}(\omega_{3},t _{2},\omega_{1}),\] (A.8)
where the non-rephasing and rephasing parts of the signal are defined by
\[I^{\rm(NR)}(\omega_{3},t_{2},\omega_{1})=\] \[\quad\quad{\rm Im}\int_{0}^{\infty}{\rm d}t_{3}\int_{0}^{\infty}{ \rm d}t_{1}e^{i\omega_{3}t_{3}}e^{i\omega_{1}t_{1}}R^{}(t_{3},t_{2},t_{1}),\] (A.9)
and
\[I^{\rm(R)}(\omega_{3},t_{2},\omega_{1})=\] \[\quad\quad{\rm Im}\int_{0}^{\infty}{\rm d}t_{3}\int_{0}^{\infty}{ \rm d}t_{1}e^{i\omega_{3}t_{3}}e^{-i\omega_{1}t_{1}}R^{}(t_{3},t_{2},t_{1}).\] (A.10)
## Appendix B Overfitting problem of MLE
To illustrate the overfitting problem effect, we present the optimized results evaluated in the early-stopping case (epoch 400) and overfitting case (epoch 1200). The learning curve of the loss function for these two cases is presented in As illustrated in Fig. B.0, the results are similar and overlap everywhere except the low frequency region, \(0\sim 400\)cm\({}^{-1}\). The optimized system parameters in the early-stopping case are \(\omega_{0}\) = 17794 cm\({}^{-1}\) and \(\Delta\) = 963 cm\({}^{-1}\), whereas those in the overfitting case are \(\omega_{0}\) = 17795 cm\({}^{-1}\) and \(\Delta\) = 960 cm\({}^{-1}\). These results indicate that the learning process works very well even in epoch 400. We then found that the accuracy of the obtained SDF, particularly in the region below 400 cm\({}^{-1}\), decreases for the case of epoch 1200, because the fluctuation of the loss function in a larger epoch period suppresses the convergence of the SDF in the low frequency region, as illustrated in This phenomenon is known as the overfitting (or overtraining) problem of the MLE for the gradient method. We may avoid this problem by regularizing the model, for example, by adopting an L2 regularization, or by using the Bayesian inference method to account for the physical knowledge as a prior probability. Nevertheless, here we use an early-stopping method to simply reduce the numerical cost, because the low-frequency region \(100\leq\omega\leq 400\)cm\({}^{-1}\) is no longer significant in the ultrafast dynamics of the exciton transfer problem, whereas the region below 10 cm\({}^{-1}\) may alter the signal profile significantly due to the quantum thermal factor, \(\coth{(\beta\hbar\omega/2)}\), in the line-shape function defined as Eq. (A.3 ).
Figure B.0: SDF \(J_{11}(\omega)\) (\(=J_{22}(\omega)\)) obtained in the early-stopping case (blue) and overfitting case (orange).
|
10.48550/arXiv.2102.02427
|
Modeling and simulating the excited-state dynamics of a system with condensed phases: A machine learning approach
|
Seiji Ueno, Yoshitaka Tanimura
| 4,610
|
10.48550_arXiv.2103.09972
|
## Abstract
An accurate description of electron correlation is one of the most challenging problems in quantum chemistry. The exact electron correlation can be obtained by means of full configuration interaction (FCI). A simple strategy for approximating FCI at a reduced computational cost is selected CI (SCI), which diagonalizes the Hamiltonian within only the chosen configuration space. Recovery of the contributions of the remaining configurations is possible with second-order perturbation theory. Here, we apply adaptive sampling configuration interaction (ASCI) combined with molecular orbital optimizations (ASCI-SCF) corrected with second-order perturbation theory (ASCI-SCF-PT2) for geometry optimization by implementing the analytical nuclear gradient algorithm for ASCI-PT2 with the \(Z\)-vector (Lagrangian) formalism. We demonstrate that for phenalenyl radicals and anthracene, optimized geometries and the number of unpaired electrons can be obtained at nearly the CASSCF accuracy by incorporating PT2 corrections and extrapolating them. We demonstrate the current algorithm's utility for optimizing the equilibrium geometries and electronic structures of 6-ring-fused polycyclic aromatic hydrocarbons and 4-periacene.
## 1 Introduction
In quantum chemistry, appropriate treatment of the correlations between electronic motions (electron correlation) is crucial to obtain accurate energies. The perfect descriptions of electron correlation can be obtained via full configuration interaction (FCI), which diagonalizes the electronic Hamiltonian within the space of all possible electronic configurations (determinants). Unfortunately, the number of electronic configurations grows exponentially with the size of the orbitals. It is practically impossible to compute more than 22 electrons in 22 orbitals [i.e., (22\(e\),22_o_)], even with a carefully tailored algorithm on modern massive parallel computational devices.
Since the dawn of quantum chemistry, researchers have investigated many different approximate FCI methods. The density-matrix renormalization group (DMRG) method has been shown to successfully approximate the FCI wave function in the form of matrix-product states (MPS). The DMRG method is considered a _de facto_ standard when FCI is not possible. The variational two-electron reduced density matrix (v2RDM) method, which employs the 2-particle reduced density as a variational parameter, can treat the electron correlation to an impressive degree of accuracy. The FCI quantum Monte Carlo (FCIQMC) method, based on quantum Monte Carlo propagation, has been very successful in terms of accuracy and robustness, even for highly correlated systems such as solids. There are also incremental approximations to FCI theory [17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 140, 132, 133, 134, 135, 136, 137, 138, 139, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 199, 198, 199, 199, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 199, 199, 199, 198, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 1999, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 1999driven CIPSI, heat-bath CI (HCI), adaptive CI (ACI), iterative CI with selection, and adaptive sampling CI (ASCI) have been developed in the last two decades. The ASCI method is a cost-effective and deterministic way to sample determinants without introducing any randomness during the procedure. Recently, analogous methods were also employed to treat vibrational problems. For more details, the readers can refer to Eriksen's recent comprehensive perspective on FCI solvers.
One can embed a highly correlated system (a set of orbitals) into a mean-field system. We denote these orbitals as "active," and this embedding results in the complete active space self-consistent field (CASSCF) method, which was also known as fully optimized reactive space (FORS). The FCI calculation performed in the active space is complete active space CI (CASCI). The computational demand for CASCI can be prohibitive when the active space is large. A useful approximation is to impose a restriction on the orbitals' occupancies, which results in the restricted active space (RAS) and generalized active space (GAS) concepts. Of course, the approximate FCI methods can be used in place of CASCI. While the CIPSI method was combined with multiconfigurational SCF in 1986, approximate CASSCF methods were developed for DMRG, v2RDM, FCIQMC, and HCI theories in the recent two decades. For ASCI, the ASCI-SCF method was developed and compared with CASSCF.
The SCI methods include only some of the configurations in the variational Hamiltonian. The rest of the energy can be corrected for by applying perturbation theory, as done by the CIPSI method. This correction can be utilized in the SCI method embedded in the mean field as well. Usually, the Epstein-Nesbet partitioning of the electronic Hamiltonian is exploited for this purpose. The HCISCF method includes a PT2 correction for the orbital optimizations to obtain orbitals at the near-CASSCF level. In contrast, only variational SCI energies are applied in the orbital optimizations in the ASCI-SCF methods, and the PT2 contributions are evaluated later. In both cases, obtaining energies comparable to those of CASSCF is possible.
Geometry optimizations and dynamics simulations are among the most important applications of quantum chemistry calculations. In terms of computational cost and accuracy, having an analytical nuclear gradient is advantageous for such simulations. Recently, we developed analytical gradient methods for variational ASCI-SCF and applied this method to optimize the molecular geometries of periacenes and polyacenes. The analytical gradient used in Ref. did not include the PT2 correction to the analytical gradient. Therefore, improvement in the quality of the optimized geometry by incorporating the PT2 correction is warranted.
In this work, we formulate an analytical gradient method for perturbatively corrected ASCI-SCF, which we call ASCI-SCF-PT2, by implementing the response equation (or the so-called \(Z\)-vector equation), as the ASCI-SCF-PT2 energy is not variational with respect to the orbital rotations and the CI coefficients. We find that for phenalenyl radicals and anthracene, the quality of the ASCI-SCF-PT2 optimized geometries is comparable to that of CASSCF-generated geometries, while the improvements in energies are minor. Moreover, by extrapolating the ASCI-SCF-PT2 nuclear gradient and number of unpaired electrons, near-exact CASSCF-level geometries and radical indices are obtained. We demonstrate the algorithm's utility for optimizing the ground-state geometries of 6-ring polycyclic aromatic hydrocarbons (PAHs) and 4-periacene.
## 2 Theory
This section first reviews the ASCI algorithm and the second-order perturbation theory correction to the ASCI energy. Then, we introduce the analytical nuclear gradient theory for the ASCI-SCF-PT2 energy. The indices \(I\), \(J\), \(K\), \(\ldots\), \(T\), \(U\), \(V\), \(\ldots\), \(i\), \(j\), \(k\), \(\ldots\), \(r\), \(s\), \(t\), \(\ldots\), and \(x\), \(y\), \(z\), \(\ldots\), denote electronic configurations in the SCI space, electronic configurations out of the SCI space, doubly occupied (closed) orbitals, active orbitals, and general orbitals, respectively.
## Adaptive Sampling CI Algorithm.
In the ASCI method, the wave function is
\[\mid\Psi\rangle=\sum_{I}^{S_{D}}c_{I}\mid I\rangle\,, \tag{1}\]
First, the Hartree-Fock (HF) ground-state determinant, \(\mid 0\rangle\), is employed as an initial guess.
\[A_{T}=\frac{\langle 0\mid\hat{H}\mid T\rangle}{E_{\text{HF}}-H_{TT}}. \tag{2}\]
Here, \(E_{\text{HF}}\) is the HF energy, and \(H_{TT}\) is the diagonal element of the Hamiltonian for determinant \(T\). \(N_{\text{tdet}}\) "target" determinants are selected based on the perturbative amplitudes. Then, the Hamiltonian is constructed and diagonalized in the target space. The \(N_{\text{cdet}}\) "core" determinants are selected based on the eigenvectors (CI coefficients).
\[A_{T}=\frac{\sum_{I}^{\text{core}}H_{TI}c_{I}}{E-H_{TT}}\,, \tag{3}\]
This procedure is iteratively repeated until convergence is reached in terms of the ground-state energy. In our implementation, the default threshold is 1 \(\mu E_{\text{h}}\). This target determinant selection strategy can also be applied for computing excited states, while one would utilize a state-by-state bootstrap method. We embed the ASCI orbital space as the active space to establish the ASCI-SCF method. The second-order augmented Hessian (AH) orbital optimization scheme was used in our implementation, while we updated the list of the target determinants every 10 macroiterations (\(N_{\text{macro}}=10\)).
## Second-Order Perturbation Theory
The Epstein-Nesbet perturbation theory (ENPT) corrects the ASCI energy to account for the contributions from the determinants outside the SCI space.
\[\hat{H}^{}=\hat{P}\hat{H}\hat{P}+\sum_{T}\hat{Q}_{T}H_{\tau T}\hat{Q}_{T}\,, \tag{4}\]
In variational perturbation theory, PT2 is reformulated as a minimization of the Hylleraas functional
\[E_{\text{PT2}}=2\langle\Psi^{}\mid\hat{H}\mid\Psi^{}\rangle+\langle\Psi^ {}\mid\hat{H}^{}-E^{}\mid\Psi^{}\rangle\,, \tag{5}\]
Here, we parameterize the first-order correction to the wavefunction as
\[\mid\Psi^{}\rangle=\sum_{T}A_{T}\mid T\rangle\,. \tag{6}\]
Solving the equation
\[\frac{\partial E_{\text{PT2}}}{\partial A_{T}}=0 \tag{7}\]
yields the amplitude \(A_{T}\) as
\[A_{T}=\frac{\langle\Psi^{}\mid\hat{H}\mid T\rangle}{E-H_{\text{\tiny TT}}}= \sum_{I}\frac{c_{I}\langle I\mid\hat{H}\mid T\rangle}{E-H_{\text{\tiny TT}}}\,. \tag{8}\]
The final second-order energy with the ENPT2 is \[\begin{split} E_{\text{PT2}}&=\sum_{T}\frac{|\bra{\Psi^{(0 )}}\ket{\hat{H}}T\rangle|^{2}}{E-H_{TT}}\\ &=\sum_{T}A_{r}\langle T\ket{\hat{H}}\ket{\Psi^{}}=\sum_{T}\sum_ {I}A_{T}c_{I}H_{TI}\end{split} \tag{9}\]
One should evaluate the perturbative correction by looping over the determinants \(T\). We have used the algorithm based on triplet constraints, which groups \(T\) based on their three highest-occupied alpha orbitals. Compared to the naive algorithm without such grouping, this process dramatically reduces the memory requirement and cost for sorting \(T\). In our implementation, the ASCI-SCF energy is corrected with the ENPT to arrive at the ASCI-SCF-PT2 energy.
### Analytical Gradient Theory: ASCI Lagrangian.
Instead of directly differentiating ASCI-SCF-PT2 energy, we can derive the analytical gradient more elegantly with the Lagrangian formalism.
\[\begin{split}\mathcal{L}&=E_{\text{ASCI-PT2}}+\sum_{ xy}Z_{xy}\Bigg{(}\frac{\partial E^{}}{\partial U_{xy}}-\frac{\partial E^{}}{ \partial U_{yx}}\Bigg{)}-\frac{1}{2}\sum_{xy}X_{xy}\left(S_{xy}-\delta_{xy} \right)\\ &\quad+\Bigg{[}\sum_{I}z_{I}\Bigg{(}\frac{\partial E^{}}{ \partial c_{I}}-E^{}c_{I}\Bigg{)}-\frac{1}{2}x\Bigg{(}\sum_{U}c_{I}c_{J} \langle I\ket{J}-1\Bigg{)}\Bigg{]},\end{split} \tag{10}\]
We treat the convergence criteria for ASCI-PT2 as constraints. The matrices \(\mathbf{Z}\) and \(\mathbf{X}\) are antisymmetric and symmetric, respectively, from the form of these constraints. Also, \(\mathbf{z}\) is set orthogonal to \(\mathbf{c}\).
\[\mathbf{C}=\mathbf{C}^{}\mathbf{U}=\mathbf{C}^{}\exp(\mathbf{\kappa})\,, \tag{11}\]
This Lagrangian is stationary when \[\frac{\partial\mathcal{L}}{\partial U_{xy}}=0\,, \tag{12}\]
\[\frac{\partial\mathcal{L}}{\partial c_{I}}=0\,. \tag{13}\]
When we define
\[Y_{xy}=\frac{\partial E_{\text{ASCI-PT2}}}{\partial U_{xy}}\,, \tag{14}\]
\[y_{I}=\frac{\partial E_{\text{ASCI-PT2}}}{\partial c_{I}}, \tag{15}\]
The \(Z\)-vector equation is then basically the same as that for CASSCF but with only limited dimensions in CI space. The explicit expression for the \(Z\)-vector equation, which is a coupled form of Eqs.
\[\begin{pmatrix}\frac{\text{d}^{2}E}{\text{d}\text{d}\text{d}\text{d}\text{x}}& \frac{\text{d}^{2}E}{\text{d}\text{d}\text{d}\text{d}\text{e}}\\ \frac{\text{d}^{2}E}{\text{d}\text{d}\text{d}\text{d}\text{x}}&\frac{\text{d} ^{2}E}{\text{d}\text{d}\text{e}\text{d}\text{e}}\end{pmatrix}\!\!\!\begin{pmatrix} \mathbf{Z}\\ \mathbf{z}\end{pmatrix}\!=\!-\!\!\begin{pmatrix}\mathbf{Y}-\mathbf{Y}^{\dagger }\\ \mathbf{y}\end{pmatrix}\!. \tag{16}\]
Here, we have exploited the symmetricity of \(\mathbf{X}\) to remove it from the equation. One can solve this equation using coupled or uncoupled schemes. We have confirmed that both techniques resulted in the same gradients.
## Source Terms for ASCI-PT2.
The source terms \(\mathbf{Y}\) and \(\mathbf{y}\) should be evaluated to solve the \(Z\)-vector equation. The Hylleraas functional is stationary with respect to variations in \(A_{T}\). The Hylleraas functional in terms of \(\mathbf{A}\), \(\mathbf{c}\), and \(\mathbf{H}\) is \[\begin{split} E_{\text{ASCI-PT2}}&=E^{}+E_{\text{PT2}} \\ &=E^{}+2\sum_{T}A_{T}c_{t}H_{T}+\sum_{T}A_{T}^{2}(H_{TT}-\sum_{U}c_{ t}c_{t}H_{U}).\end{split} \tag{17}\]
Here, the zeroth-order energy (ASCI-SCF energy) is
\[E^{}=\sum_{U}H_{U}c_{t}c_{J}=\sum_{xy}h_{xy}d_{xy}^{}+\sum_{xy\in w}(xy \mid zw)D_{xy,zw}^{} \tag{18}\]
The mathematical derivations can be made simpler by defining the density matrices.
\[\mathbf{d}=\mathbf{d}^{}+\mathbf{d}^{}+\mathbf{d}^{}\,, \tag{19}\]
\[\mathbf{D}=\mathbf{D}^{}+\mathbf{D}^{}+\mathbf{D}^{}\,, \tag{20}\]
The total ASCI-PT2 energy is
\[E_{\text{ASCI-PT2}}=\sum_{rs}h_{rs}d_{rs}+\sum_{rstu}(rs\mid tu)D_{rs,u}\,. \tag{21}\]
The density matrices are then
\[d_{rs}^{}=\sum_{T}A_{T}c_{t}\Bigg{[}\frac{\partial H_{T}}{\partial h_{rs}}+ \frac{\partial H_{T}}{\partial h_{rs}}\Bigg{]}, \tag{22}\]
\[D_{rs,u}^{}=\sum_{T}A_{T}c_{t}\Bigg{[}\frac{\partial H_{T}}{\partial(rs\mid tu )}+\frac{\partial H_{T}}{\partial(rs\mid tu)}\Bigg{]}, \tag{23}\]
\[d_{rs}^{}=\sum_{T}A_{T}^{2}\frac{\partial H_{TT}}{\partial h_{rs}}-d_{rs}^{}N\,, \tag{24}\]
\[D_{rs,u}^{}=\sum_{T}A_{T}^{2}\frac{\partial H_{TT}}{\partial(rs\mid u)}-D_{rs,u}^{}N\,, \tag{25}\]
We symmetrize the first-order density matrices. Should there be closed (doubly occupied) orbitals, the core Fock integrals replace the one-electron integrals. With these density matrices, the orbital gradient \(\mathbf{Y}\) is obtained in the same way that the orbital gradients in the CASSCF calculation are evaluated. The only difference is that the zeroth-order density matrix is replaced with the total density matrix.
\[\begin{split} y_{I}&=2\Bigg{[}\sum_{T}A_{T}H_{Tt}+ \big{(}1-N\big{)}\sum_{J}H_{U}c_{J}\Bigg{]}\\ &=2\Bigg{[}\sum_{T}A_{T}H_{Tt}+\big{(}1-N\big{)}c_{I}E^{} \Bigg{]},\end{split} \tag{26}\]
#### 3.2.2 Final Gradient Evaluation and Extrapolation.
After the \(Z\)-vector equation is solved, the final gradient is evaluated as
\[\frac{\mathrm{d}E_{\mathrm{ASCPT2}}}{\mathrm{d}\mathbf{R}}=\frac{\partial \mathcal{L}}{\partial\mathbf{R}}=\sum_{\mu\nu}\frac{\mathrm{d}h_{\mu\nu}}{ \mathrm{d}\mathbf{R}}d_{\mu\nu}^{\mathrm{eff}}+\sum_{\mu\nu\alpha\sigma}\frac {\mathrm{d}(\mu\nu|\lambda\sigma)}{\mathrm{d}\mathbf{R}}D_{\mu\nu\lambda\sigma }^{\mathrm{eff}}+\sum_{\mu\nu}\frac{\mathrm{d}S_{\mu\nu}}{\mathrm{d}\mathbf{R }}X_{\mu\nu}\,, \tag{27}\]
One then evaluates the effective (relaxed) densities \(\mathbf{d}^{\mathrm{eff}}\) and \(\mathbf{D}^{\mathrm{eff}}\) in the AO basis as
\[\mathbf{d}^{\mathrm{eff}}=\mathbf{C}\Big{[}\mathbf{d}+\overline{\mathbf{d}}+ \mathbf{Z}\mathbf{d}^{}+\mathbf{d}^{}\mathbf{Z}^{*}\Big{]}\mathbf{C}^{ \dagger}\,, \tag{28}\]
\[D_{\mu\nu\lambda\sigma}^{\mathrm{eff}}=\sum_{z\nu}D_{\nu\sigma}^{\mathrm{eff}, z\nu}C_{\mu z}C_{\lambda\nu}\,, \tag{29}\]
\[\mathbf{D}^{\mathrm{eff},z\nu}=\mathbf{C}\Bigg{[}\mathbf{D}^{z\nu}+\frac{1}{2} \overline{\mathbf{Q}}^{z\nu}+\mathbf{Z}\mathbf{Q}^{z\nu}+\mathbf{Q}^{z\nu} \mathbf{Z}^{*}\Bigg{]}\mathbf{C}^{\dagger}\,, \tag{30}\]
where
\[\overline{\mathbf{d}}=\frac{1}{2}\sum_{I}z_{I}\left\langle I\left|\hat{E}_{z \nu}+\hat{E}_{z\nu}\right|\Psi^{}\right\rangle, \tag{31}\]
\[\overline{\mathbf{Q}}^{z\nu}_{xy}=\frac{1}{2}\sum_{I}z_{I}\left\langle I\left| \hat{E}_{z\nu,y\nu}+\hat{E}_{z\nu,y\nu}\right|\Psi^{}\right\rangle, \tag{32}\]\(D_{xy}^{zw}=D_{xz,yw}\), and \(Q_{xy}^{zw}=D_{xz,yw}^{}\). These expressions are the same as those in Refs. and.
One can extrapolate the PT2-corrected selected CI energy with respect to the PT2 correction to obtain the near-exact CASSCF energy. Simple linear regressions are usually sufficient for this purpose, and the \(R^{2}\) values for the fitting exceed 0.99 in almost all cases we have tested. We found that it is also possible to extrapolate the ASCI-SCF-PT2 nuclear gradients at each nuclear coordinate with respect to the PT2 correction to the point where the PT2 correction is zero to achieve near-exact CASSCF-level gradients. To this end, we compute the ASCI-SCF-PT2 analytical gradients for each \(N_{\rm{tdet}}\) (usually four to five points) and perform linear regressions for \(3N_{\rm{atom}}\) components of nuclear gradients to obtain the final gradient. At each extrapolation point, we perform the ASCI-SCF orbital optimization, but the iterative update of the list of the determinant is performed only at the point with the largest \(N_{\rm{tdet}}\) (i.e., \(N_{\rm{macro}}=\infty\) for each supermacroiteration in the ASCI-SCF optimization with smaller \(N_{\rm{tdet}}\)), as this approximation does not affect the efficiency to a significant degree. The linear fitting cost is negligible, and the significant computational burden for this procedure is the performance of the ASCI-SCF-PT2 correction at each \(N_{\rm{tdet}}\). We will denote the extrapolated ASCI-SCF-PT2 gradient as the "ASCI-SCF-PT2+X" gradient for the sake of brevity.
## 3 Numerical Examples
We have linked the programs for evaluating ASCI-SCF-PT2 density matrices and CI derivatives and solving the ASCI-SCF Z-vector equation to the program package BAGEL. We computed all two-electron integrals with density fitting (DF) approximations. We used the cc-pVDZ basis set and its corresponding JKFIT basis for DF approximation unless otherwise mentioned.
## Comparison with CASSCF: Optimized geometries.
To assess the improvement offered by the PT2 corrections, we used phenalenyl radicals and anthracene as test molecules. In the ASCI-SCF-PT2+X calculations, we employed five data points with \(N_{\text{tdet}}=20000\), 40000, 60000, 80000, and 100000. The errors in the energies and optimized geometries, compared to the CASSCF (FCI) references, are shown in Table 1.
First, we can notice that the improvements in the energy are rather minor compared to the ASCI-SCF-PT2 energy at the ASCI-SCF geometries, particularly when \(N_{\text{tdet}}\) is large. The largest improvement is 0.168 m\(E_{\text{h}}\) (which is \(\sim\)0.4 kJ/mol) for triplet anthracene with \(N_{\text{tdet}}=1000\), while the improvement is below 0.001 m\(E_{\text{h}}\) for (13\(e\),13\(o\)) phenalenyl radicals with \(N_{\text{tdet}}=100000\). In other words, to obtain an accurate ASCI-SCF-PT2 energy, it is sufficient to use the ASCI-SCF gradient in geometry optimization and perform a perturbative correction at the optimized geometry, particularly when \(N_{\text{tdet}}\) is large. We note that this argument still holds with larger active spaces tested in this work-the differences between the ASCI-SCF-PT2 energies at the ASCI-SCF-PT2 and ASCI-SCF geometries were below 0.5 m\(E_{\text{h}}\) in all cases. When ASCI-SCF-PT2+X gradients were used in optimizations, the resulting errors were within the 0.2 m\(E_{\text{h}}\) range of the CASSCF value, which is a very reliable estimate considering possible extrapolation errors.
Structures of the molecules used for tests against CASSCF results.
\begin{table}
\begin{tabular}{l c c c c c c c c c c c} \hline & & & & & & & & & & \multicolumn{2}{c}{RMS distance from CASSCF optimized geometry (Å)} \\ & active space & \(N_{\rm data}\) & \(N_{\rm data}\) & \(N_{\rm CASSCF}\)a & \(N_{\rm total}/N_{\rm CASSCF}\) & \(N_{\rm total}/N_{\rm CASSCF}\) & \(N_{\rm total}/N_{\rm CASSCF}\) & ASCI-SCF-FT2 & ASCI-SCF-FT2 \(\times\)CPt & ASCI-SCF-FT2 & ASCI-SCF-FT2 \(\times\)CPt \\ & & & & & & & & \multicolumn{2}{c}{SCF opt} & \\ \hline \multirow{3}{*}{\begin{tabular}{c} Phenalenyl radical \\ (double) \\ \end{tabular} } & \multirow{3}{*}{(13\(e\),13_o_)} & 200 & 1000 & 2944656 & 0.03 & 23.111 & 7.423 & 7.381 & & 0.005 & 0.003 & \\ & & & & & & & & & & \\ & & & & & & & & & & \\ & & & & & & & & & & \\ & & & & & & & & & & \\ \hline \multirow{3}{*}{\begin{tabular}{c} Anthracene \\ (single) \\ \end{tabular} } & \multirow{3}{*}{(14\(e\),14_o_)} & 200 & 1000 & 11778624 & 0.01 & 24.892 & 8.368 & 8.333 & & 0.007 & 0.003 & \\ & & & & & & & & & & \\ & & & & & & & & & & \\ & & & & & & & & & & \\ & & & & & & & & & & \\ \hline \multirow{3}{*}{
\begin{tabular}{c} Anthracene \\ (triplet) \\ \end{tabular} } & \multirow{3}{*}{(14\(e\),14_o_)} & 200 & 10000 & 11778624 & 0.08 & 6.741 & 2.292 & 2.271 & -0.134 & 0.002 & 0.001 & 0.000 \\ & & & & & & & & & & \\ & & & & & & & & & & \\ & & & & & & & & & & \\ & & & & & & & & & & \\ \hline \end{tabular}
\end{table}
Table 1: The errors in the energy (in m_E_h) and geometries (Å, root-mean-squared distance) between the ASCI-SCF and ASCI-SCF-PT2 optimized energy and geometries with respect to the CASSCF reference values.
Errors in bond lengths with respect to the CASSCF values in the phenalenyl radical and anthracene for the geometries optimized with ASCI-SCF (dashed), ASCI-SCF-PT2 (thin solid) with \(N_{\text{tdet}}=1000\) (green), 10000 (red), 100000 (blue), and ASCI-SCF-PT2+X (thick black).
The ASCI-SCF-PT2 geometries are two times closer to the CASSCF geometries than the ASCI-SCF geometries in terms of root-mean-squared distances. The geometries obtained with extrapolated gradients are within a distance of 0.0003 A from the CASSCF geometries. The deviations in the bond lengths of edge C-C bonds in phenalenyl radicals and anthracene from the CASSCF values are shown in In both systems, the errors are significantly reduced by increasing \(N_{\text{tdet}}\). Without PT2 corrections, the maximum error decreases from 0.008 A to 0.001 A (anthracene) and from 0.004 A to 0.000 4 A (phenalenyl radical) when \(N_{\text{tdet}}\) increases from 1000 to 10000. By considering the PT2 corrections, the absolute errors decrease by almost half. Including ~1% of complete electronic configurations in the ASCI-SCF-PT2 optimizations results in bond length deviations below 0.000 5 A. The geometries with the ASCI-SCF-PT2+X gradients exhibit errors below 0.000 2 A. Overall, these test results show that the ASCI-SCF-PT2 geometry optimizations can yield reasonable estimates of CASSCF-optimized geometries without the computational burden required for FCI diagonalization. Additionally, with linear extrapolation (ASCI-SCF-PT2+X), near-exact CASSCF-level geometries can be obtained.
## Comparison with CASSCF: Number of Unpaired Electrons.
The PT2 energy correction naturally improves the wavefunctions to the first order. In terms of RDMs, these corrections are given by Eqs. 19 and 20. Moreover, in analytical gradient theory, we make the Lagrangian (Eq. 10) stationary with respect to the orbital rotations by solving the \(Z\)-vector equation (Eqs. 12 and 13). In the sense of invariance, the Lagrangian is physically equivalent to the ASCI-SCF energy. The effective (relaxed) densities (Eq. 27), the RDMs for the Lagrangian, can be employed to analyze the electronic structures with a perturbative correction. This is a well-known strategy for examining electronic distributions in correlated methods.
\begin{table}
\begin{tabular}{c c c c c c c c} \hline & & \multicolumn{2}{c}{phenyl} & \multicolumn{2}{c}{anthracene} & \multicolumn{2}{c}{anthracene} \\ & & \multicolumn{2}{c}{radical (doublet)} & \multicolumn{2}{c}{(singlet)} & \multicolumn{2}{c}{(triplet)} \\ \hline \(N_{\rm det}\) & Density & \(N_{\rm U}\) & error & \(N_{\rm U}\) & error & \(N_{\rm U}\) & error \\ \hline \multirow{4}{*}{100000} & Effective & 1.351 & 0.008 & 0.427 & 0.013 & 2.357 & 0.014 \\ & ASCI-PT2 & 1.338 & 0.021 & 0.408 & 0.033 & 2.339 & 0.032 \\ & ASCI-SCF & 1.335 & 0.024 & 0.404 & 0.036 & 2.335 & 0.035 \\ & Effective & 1.315 & 0.044 & 0.379 & 0.062 & 2.304 & 0.067 \\ & ASCI-PT2 & 1.274 & 0.085 & 0.330 & 0.110 & 2.260 & 0.111 \\ & ASCI-SCF & 1.259 & 0.100 & 0.320 & 0.121 & 2.247 & 0.124 \\ & Effective & 1.259 & 0.099 & 0.312 & 0.128 & 2.237 & 0.134 \\ & ASCI-PT2 & 1.197 & 0.162 & 0.235 & 0.206 & 2.173 & 0.198 \\ & ASCI-SCF & 1.169 & 0.190 & 0.205 & 0.235 & 2.147 & 0.224 \\ \hline \multicolumn{2}{c}{CASSCF} & \multicolumn{2}{c}{1.359} & \multicolumn{2}{c}{0.441} & \multicolumn{2}{c}{2.371} \\ \hline Extrapolated & 1.357 & \multicolumn{2}{c}{0.442} & \multicolumn{2}{c}{2.370} \\ \hline \end{tabular}
\end{table}
Table 2: The numbers of unpaired electrons, \(N_{\rm U}\), evaluated with effective (relaxed), unrelaxed, ASCI-SCF-PT2, and ASCI-SCF densities and their errors with respect to the CASSCF results.
All \(N_{\rm U}\) values are evaluated using relaxed densities.
Results of linear extrapolation of the number of unpaired electrons (\(N_{\rm U}\)) with respect to the PT2 correction energy in phenalenyl radicals (doublet) and anthracene (singlet and doublet).
As an example, we evaluated the number of unpaired electrons with ASCI-SCF density and ASCI-SCF-PT2 densities (both unrelaxed and relaxed) in phenalenyl radicals and anthracene. We used the nonlinear model of Head-Gordon.
\[N_{\rm U}(i)=n_{i}^{2}(2-n_{i})^{2} \tag{33}\]
By back-transforming a matrix with diagonal elements of \(N_{\rm U}(i)\) into the atomic orbital basis, these unpaired electrons can be assigned to each atom as well.
The numbers of unpaired electrons, compared to the CASSCF values, are shown in Table 2. As expected, the errors were reduced by increasing \(N_{\rm det}\): With the ASCI-SCF density, the errors were reduced by 0.16 to 0.20 when we increased \(N_{\rm det}\) from 1000 to 100000. The unrelaxed ASCI-SCF-PT2 densities increased \(N_{\rm U}\), but to a minor degree. Inclusion of the Z-vector contributions in densities resulted in \(N_{\rm U}\) values closer to the CASSCF values. In the largest calculations, \(N_{\rm U}\) differed from the reference value by only \(\sim\)0.01. The \(N_{\rm U}\) values evaluated with ASCI-SCF and ASCI-SCF-PT2 were always smaller than the CASSCF values in all tested cases, which implies that the FCI space outside the ASCI space contributes to an increase in the unpaired electrons (which is very physical and natural).
We performed linear extrapolation of \(N_{\rm U}\) computed with the ASCI-SCF-PT2 relaxed density at the geometries optimized with \(N_{\rm det}=100000\). We have included five data points (\(N_{\rm det}\) = 20000, 40000, 60000, and 80000) in these extrapolations. The results of such extrapolations are shown in The \(N_{\rm U}\) values depend almost linearly on the PT2 correction energies, and the differences between the resulting estimates and CASSCF values were below 0.002 for all tested cases. We note that the \(N_{\rm U}\) value from the ASCI-SCF-PT2 unrelaxed densities does not have this property. This finding shows that despite the additional nonlinearity introduced by the orbital optimizations in ASCI-SCF, extrapolation also works for \(N_{\text{U}}\). This also suggests that this strategy could be used to compute reliable estimates of various electronic structure measures at the CASSCF level.
The extrapolations of both analytical gradients and \(N_{\text{U}}\) were quite successful. We hypothesize that the density matrices in the atomic orbital basis can be the target function for extrapolation. Unfortunately, it is not easy to satisfy the physical properties of the density matrices with linear extrapolation, and thus, we leave it as a target for future investigations.
### Demonstration for 6-ring-fused PAHs.
Next, we demonstrate the ASCI-SCF-PT2 analytical gradient method for the group of 6-ring-fused PAHs shown in Among these molecules, uthrene and triangulene are predicted to have triplet ground states, according to Ovchinnikov's rule, DFT, MR-CISD, and MR-AQCC calculations. Here, we perform ASCI-SCF-PT2 geometry optimizations for these molecules by employing full \(\pi\)-valence active space, which results in an active space ranging from (22\(e\),22\(o\)) for triangulene to (26\(e\), 26\(o\)) for hexacene and fulminene. We used \(N_{\text{tdet}}=500000\) and \(N_{\text{cdet}}=1000\) for the ASCI-SCF and ASCI-SCF-PT2 optimizations and \(N_{\text{tdet}}=100000\), 200000, 300000, 400000, and 500000 for the extrapolation to evaluate the ASCI-SCF-PT2+X gradient.
##
Molecular structures of 6-ring-fused PAHs.
Natural orbital occupancies in 6-ring-fused PAHs. Note the offsets of the NO indices in zethrene, uthrene, and triangulene for consistency in the indices of the HOMO and LUMO.
\begin{table}
\begin{tabular}{c c c c|c c c c} \hline & \multicolumn{2}{c|}{Singlet – triplet gaps (\(\Delta E_{\rm ST}=E_{\rm T}-E_{\rm S}\)) (kcal/mol)} & \multicolumn{4}{c}{\(N_{\rm U}\)} \\ \cline{2-7} & \multicolumn{2}{c|}{\(N_{\rm det}=500000\)} & \multicolumn{2}{c|}{ASCI-SCF-PT2} & \multicolumn{2}{c}{\(N_{\rm det}=500000\)} & \multicolumn{2}{c}{Extrapolated} \\ \cline{2-7} & ASCI-SCF & ASCI-SCF-PT2 & PT2+X & singlet & triplet & singlet & triplet \\ \hline hexacene & 19.30 & 18.96 & 19.34 & 0.73 & 2.49 & 1.01 & 2.65 \\ fulminene & 58.43 & 60.01 & 59.39 & 0.43 & 2.50 & 0.59 & 2.67 \\ zethrene & 28.82 & 25.94 & 22.19 & 0.57 & 2.44 & 0.75 & 2.59 \\ uthrene & –10.69 & –10.87 & –9.06 & 2.38 & 2.44 & 2.52 & 2.56 \\ triangulene & –12.36 & –13.48 & –13.48 & 2.41 & 2.46 & 2.48 & 2.56 \\ \hline \end{tabular}
\end{table}
Table 3: The adiabatic singlet–triplet gaps and the numbers of unpaired electrons (\(N_{\rm U}\)) in 6-ring PAHs, computed with ASCI-SCF, ASCI-SCF-PT2, and extrapolated ASCI-SCF-PT2.
The resulting singlet-triplet gaps and \(N_{\rm U}\) values for singlet and triplet states are shown in Table 3. As expected, the ground states are triplet states in uthrene and triangulene and singlet states in the other systems. Based on the extrapolated energies, the singlet-triplet gap (\(\Delta E_{\rm ST}\)) is largest in fulminene (\(\sim\)60 kcal/mol), and the gaps are similar in hexacene and zethrene. The \(\Delta E_{\rm ST}\) values of hexacene were calculated with various means of approximate CASSCF methods, and they range from 17.1 (v2RDM-SCF/cc-pVDZ) to 21.4 (DMRG/6-31+G**) kcal/mol. The ASCI-SCF values of 19.0 to 19.3 kcal/mol are within this range. The numbers computed with methods that incorporate dynamical correlations (11.2, 15.0, 16.8 kcal/mol with ACI-DSRG-PT2/cc-pVDZ, GAS-PDFT/6-31+G**, DMRG-PDFT/6-31+G**, respectively) are closer to the experimental value (\(\sim\)12 kcal/mol). These results suggest that the errors in calculated values with approximate CASSCF methods can be attributed to missing dynamical correlations. In both uthene and triangulene, the triplet state is \(\sim\)10 kcal/mol more stable than the singlet state. The number of unpaired electrons in the triplet states is approximately 2.5 for all the tested systems. The \(N_{\rm U}\) values for singlet states of uthrene and trianglene are greater than 2.0, which means that these molecules have diradical singlet states. These states are naturally less stable than the diradical triplet states according to Hund's rule. The experimental results also support the diradical ground state (singlet or triplet) in triangulene. The others have closed-shell singlet ground states. The natural orbital occupancies in the singlet states, computed with the relaxed densities, are shown in and reconfirm that the uthrene and triangulene singlet states are biradical, as confirmed with the MRCISD or CASSCF calculations using limited active space.
Atomic distributions of unpaired electrons in uthrene and triangulene with \(N_{\rm tdet}=500000\).
Optimized ground-state structures (singlet for hexacene, fulminene, zethrene; triplet for uthrene and triangulene) of 6-ring-fused PAHs. For utherene, the side view is also displayed. The molecular graphics are generated using the software IboView.
We evaluated the number of unpaired electrons on each carbon atom in uthrene and triangulene by back-transformations of the unpaired electron densities into the atomic orbital basis, followed by Mulliken analysis. The results with relaxed densities at \(N_{\text{tdet}}=500000\) are displayed in The unpaired electrons are mainly distributed alternatively along the edge of these molecules. The atoms with large unpaired electron densities belong to the same group, as defined by Ovchinnikov's rule. In the triplet and singlet states, these electrons are placed symmetrically and nonsymmetrically, respectively.
As predicted with DFT calculations, the optimized structure of uthrene exhibits geometries slightly twisted by \(\sim\)20 deg, while all the other molecules are planar. The singlet and triplet geometries in triangulene are isosceles-like and regular-triangle-like, respectively, due to their symmetric electronic structures. In the triplet state, the optimized geometry's symmetry is slightly broken unless the CASCI (perfect) wave function is used, as we did not exploit any geometrical symmetry (\(D_{3h}\)). In other words, the wave function and energy accurately approximate the CASSCF functions when the optimized triplet geometry is similar to a regular triangle. The edge bond lengths and the deviations from the symmetric geometries in triangulene are shown in When the geometries were optimized with ASCI-SCF, ASCI-SCF-PT2, and ASCI-SCF-PT2+X, the maximum deviations from the symmetric geometry were 0.0012 A, 0.0007 A, and 0.0005 A, respectively. The optimized conformations are slightly isosceles-like, and the triangular sides differ by 0.0016 A, 0.0016 A, and 0.0002 A. Again, these results imply that minimizing the extrapolated ASCI-SCF-PT2 energy can yield reasonable approximations to the CASSCF molecular conformations.
Optimized triplet geometry in triangulene. (a) C–C bond length alternation (BLA) pattern along the edge bonds and (b) their deviations from the symmetric geometry, with ASCI-SCF-PT2+X (bold solid line), ASCI-SCF-PT2 (thin solid line), and ASCI-SCF (dashed line). We defined the symmetric geometry by taking averages over the three triangular sides. The side lengths of the triangle are also shown.
C–C BLA pattern along the zigzag \(\pi\)-bonds in n-periacenes under the ASCI-SCF (dashed), ASCI-SCF-PT2 (solid thin), and ASCI-SCF-PT2+X (solid bold) singlet (black) and triplet (red) optimized geometries for 4-periacene. The size of the target determinant space was \(10^{6}\) for ASCI-SCF and ASCI-SCF-PT2 optimizations.
\begin{table}
\begin{tabular}{c c c c c c c c c c} \hline \hline & \multicolumn{3}{c}{ASCI-SCF geometry} & \multicolumn{3}{c}{ASCI-SCF-PT2 geometry} & \multicolumn{3}{c}{ASCI-SCF-PT2+X geometry} \\ \hline & \(N_{\text{dist}}=1000000\) & \multicolumn{3}{c}{\(N_{\text{dist}}=1000000\)} & \multicolumn{3}{c}{extrap.} & \multicolumn{3}{c}{\(N_{\text{dist}}=1000000\)} & \multicolumn{3}{c}{extrap.} & \multicolumn{3}{c}{extrap.} \\ & ASCI-SCF & ASCI-SCF-PT2 & & ASCI-SCF & ASCI-SCF-PT2 & & ASCI-SCF & ASCI-SCF-PT2 & \\ \hline singlet (\(E_{\text{d}}\)) & -1373.16579 & -1373.19884 & -1373.24458 & -1373.16550 & -1373.19943 & -1373.24681 & -1373.16000 & -1373.19710 & -1373.25225 \\ triplet (\(E_{\text{d}}\)) & -1373.15784 & -1373.19306 & -1373.23363 & -1373.15779 & -1373.19337 & -1373.23588 & -1373.15547 & -1373.19403 & -1373.23815 \\ \hline \(\Delta E_{\text{ST}}\) (kcal/mol) & 4.99 & 3.63 & 6.87 & 4.84 & 3.81 & 6.86 & 2.84 & 1.92 & 8.85 \\ \hline \hline \end{tabular}
\end{table}
Table 4: The ASCI-SCF, ASCI-SCF-PT2, and extrapolated ASCI-SCF-PT2 singlet, triplet, and singlet–triplet gap energies at the ASCI-SCF, ASCI-SCF-PT2, and ASCI-SCF-PT2+X geometries of 4-periacene.
## Case of 4-Periacene.
In our previous study using the ASCI-SCF analytical gradient, we found that for 4-periacene, the ASCI-SCF-optimized singlet geometry is not the true minimum for the singlet state. The extrapolated energy of the singlet-optimized geometry was higher than that of the triplet-optimized geometry. This suggested that the CASSCF-level minimum is not the ASCI-SCF singlet geometry but rather near the ASCI-SCF triplet geometry. To find the near-CASSCF-level geometries of 4-periacene, we optimized the singlet and triplet molecular geometries with ASCI-SCF, ASCI-SCF-PT2 (both with \(N_{\text{tdet}}=10^{6}\)), and ASCI-SCF-PT2+X. The (36\(e\),36\(o\)) active space with full \(\pi\)-valence active space was employed. Four data points (\(N_{\text{tdet}}=4\times 10^{5}\), 6\(\times 10^{5}\), 8\(\times 10^{5}\), and 10\({}^{6}\)) were utilized for extrapolation.
The total and singlet-triplet gap energies of the ASCI-SCF, ASCI-SCF-PT2, and ASCI-SCF-PT2+X geometries are shown in Table 4. We note that the total energies presented here are slightly different from those in Ref., as we have extrapolated the ASCI-SCF energies instead of the ASCI energies (i.e., the extrapolation points included the orbital relaxation effects at the points with smaller \(N_{\text{tdet}}\)). The ASCI-SCF-PT2+X optimization relaxed the energy by 7.7 m\(E_{\text{h}}\) (~5 kcal/mol) in the singlet state, and of course, the optimized energy (~1373.25225 \(E_{\text{h}}\)) was lower than the singlet extrapolated energy in the triplet ASCI-SCF-PT2 geometry (~1373.24464 \(E_{\text{h}}\)). This makes the resulting extrapolated ASCI-SCF-PT2 \(\Delta E_{\text{ST}}\) 8.9 kcal/mol. This value is smaller than the v2RDM-SCF/cc-pVDZ (13.4 kcal/mol) value but larger than the DMRG/STO-3G (5.3 kcal/mol) value. This value is also larger than the recent experimental value obtained via a SQUID experiment (2.5 kcal/mol). Most likely, including dynamic correlation will improve the computed \(\Delta E_{\text{ST}}\) value like in the case of hexacene (see above). We note that the optimized active orbitals in the ASCI-SCF-PT2+X geometry are almost the same as those in the ASCI-SCF-PT2 geometry.
The C-C bond lengths along the longer edge in 4-periacene in various optimized geometries are shown in The maximum deviations between the C-C bond length in the ASCI-SCF-PT2+X and ASCI-SCF geometries were 0.019 and 0.008 A for the singlet and triplet states, respectively. The singlet and triplet ASCI-SCF-PT2+X geometries move closer to the triplet and singlet ASCI-SCF geometries, respectively, but do not coincide, as hypothesized in our previous work or as in the v2RDM-SCF/cc-pVDZ optimization. Overall, we resolved some open questions regarding the ASCI-SCF method's performances for describing 4-periacene in our earlier work using the ASCI-SCF-PT2+X gradient. We finally note that the possible contributions of the dynamical correlation (or the electron correlation out of the active space) will probably be more significant than the ASCI-SCF-PT2+X corrections. The quantitative investigations of such dynamical correlation effects will be an intriguing target of future studies.
\begin{table}
\begin{tabular}{c c c c c c c c c c c c} \hline \hline & & & \multicolumn{4}{c}{Parameters} & \multicolumn{4}{c}{Wall clock time (secs)} \\ \hline & Active space & \(N_{\text{tun}}\)\({}^{a}\) & \(\begin{array}{c}\text{Largest}\\ N_{\text{dub}}\end{array}\) & \multirow{2}{*}{Number of \(T^{c}\)} & \multicolumn{4}{c}{ASCI-SCF} & \multicolumn{4}{c}{ASCI-SCF-PT2} & \multicolumn{4}{c}{Total ASCI-SCF-} \\ & & & & \(t_{\text{fbl}}\) & \(t_{\text{fbl}}\) & \(t_{\text{fbl}}\) & \(t_{\text{com}}\) & \(t_{\text{dus}}\)\({}^{f}\) & \(t_{\text{dus}}\)\({}^{f}\) & \(t_{\text{dus}}\)\({}^{f}\) & PT2+\(X\)\({}^{f}\) \\ \hline Anthracene & (14e,14o) & 246 & 100000 & 8 162 971 & 11 & 3 & 14.7 & 18 & 81 & 371 & 2170 \\ Hexacene & (26e,26o) & 444 & 500000 & 1 433 657 291 & 60 & 20 & 442 & 355 & 2000 & 1850 & 17000 \\
4-periacene & (36e,36o) & 584 & 1000000 & 18 274 677 825 & 0 & 75 & 3276 & 2950 & 24000 & 3380 & 71400 \\ \hline \hline \end{tabular} \({}^{a}\) Number of basis functions. \({}^{b}\) Largest \(N_{\text{dub}}\) employed in ASCI-SCF-PT2+X calculations. \({}^{c}\) Number of PT2 configurations included in the calculations with the largest \(N_{\text{dub}}\) for timing benchmark. We imposed the cutoff of \(10^{-7}\) to generate these contributions. \({}^{d}\) time for single diagonalization and single RDM calculation. \({}^{e}\) time for orbital optimization in a macroiteration. \({}^{f}\) time for generating the list of PT2 configurations. \({}^{g}\) time for sorting the list of PT2 configuration. \({}^{h}\) time for computing ASCI-PT2 energy, RDMs, and CI derivatives. \({}^{i}\) time for solving the \(Z\)-vector equation. \({}^{f}\) time for a single ASCI-SCF-PT2+X geometry optimization step. One step includes four (hexacene, 4-periacene) or five (anthracene) ASCI-SCF-PT2 gradient computations.
\end{table}
Table 5: Wall clock time for single iteration (gradient evaluation) in ASCI-SCF-PT2+X geometry optimizations. These times were measured using 18 physical cores in Intel Xeon Gold 6240 CPU (2.60 GHz). The detailed descriptions of times are in the footnote.
## Computational Cost.
Finally, we will discuss the computational cost of the current algorithm. Table 5 shows the wall clock time for ASCI-SCF-PT2+X gradient calculations for three systems tested in this work (singlet anthracene, hexacene, and 4-periacene). On eighteen physical cores in Intel Xeon Gold 6240 CPU (2.60 GHz), an ASCI-SCF-PT2+X geometry optimization step for anthracene, hexacene, and 4-periacene took 0.6, 4.7, and 20 hours, respectively. Roughly, the wall time needed for a single optimization step increases by a factor of ~10 when the numbers of electrons and active orbitals are increased by ~10. Because we are "sampling" the electronic configurations in the CASCI space, rather than computing with a rigorous ansatz like in DMRG, it is somewhat tricky to write scaling with simple molecular parameters like \(N_{\rm act}\). Instead, we can write the computational cost using the empirical parameters, such as the number of PT2 determinants (\(T\)). The number of operations for evaluating ASCI-SCF-PT2 energy is \(\sum\limits_{T}\alpha_{T}\), where \(\alpha_{T}\) is the number of the configurations in variational wave function that interact with \(T\). If we assume that \(\alpha_{T}\) is similar for all \(T\), the computational cost will be proportional to the number of \(T\), roughly the case for the data in Table 5.
The CASSCF calculations are not practical except for anthracene. For anthracene, a single CASCI iteration took 10 seconds. Approximating the wall time on the same CPU using the scaling of the most expensive step (\(N_{\rm det}N_{\rm act}^{4}\) ) in the Knowles-Handy algorithm yields the estimated single CASCI iteration time of ~30 and ~108 years for hexacene and 4-periacene, respectively.
Let us compare the computational cost for our ASCI-SCF-PT2+X algorithm with the other methods in the literature. The DMRG-SCF methods are considered to be the most accurate and are used for molecular geometry optimizations. A single DMRG-SCF geometry optimization step of (20\(e\),22\(o\)) spiropyran with the 6-31G(d) basis set took ~10 hours with \(M=512\) using 16 cores in Intel Xeon E5-2690 CPU (2.90 GHz), while \(M\) denotes the bond dimension. A singleDMRG-SCF iteration in (10\(e\),34\(o\)) indole using the aug-cc-pVTZ basis set with \(M=1000\) and (12\(e\),28\(o\)) Cr\({}_{2}\) using the cc-pVDZ basis set with \(M=500\) required 11 hours and 2 hours, respectively, using 16 cores in Intel Xeon E5-2670 (2.60 GHz) and Intel Xeon E5-2667 (2.90 GHz) CPUs. The v2RDM-SCF geometry optimizations are reported to be quite efficient, as a single geometry optimization step in (26\(e\),26\(o\)) hexacene and (38\(e\),38\(o\)) nonacene with the cc-pVDZ basis set took ~10 mins and ~50 mins, respectively, using six cores in i7-6850K CPU (3.60 GHz). The reported selected CI methods (although without geometry optimizations) require lower computational cost than our algorithm. A CIPSI calculation of (24\(e\),76\(o\)) Cr\({}_{2}\) using the cc-pVDZ basis set with 2\(\times\)10\({}^{7}\) determinants took ~14 mins using 800 physical cores at 2.70 GHz. A semi-stochastic heat-bath CI calculation and ASCI-PT2 calculation of (14\(e\),26\(o\)) F\({}_{2}\) with the cc-pVDZ basis required 5 and 49 seconds on 20 and single physical core in Intel Xeon E5-2680v2 (2.80 GHz) and Intel Xeon E5-2620v5 CPU (2.10 GHz), respectively. Of course, we should note that these comparisons are indirect, as the calculations were performed for different systems on diverse computer architecture with various software.
The benchmark results imply that there is plenty of room for improvement in our program. The generation of contributions (\(T\)) to each triplet constraint is not implemented optimally (using the methods like in Ref.). The most expensive term is the two-particle RDM from the second-order term (Eq. 25), which requires \(O(N_{T}N_{\rm ele}^{2})\) operations, where \(N_{\rm ele}\) is the number of electrons in the active space. One can efficiently compute the Hamiltonian diagonal element for each \(T\) (\(H_{TT}\) in Eq. 8) with the energy of determinants in ASCI wave function. One could similarly apply such a strategy for evaluating RDMs. Finally, the routine for computing the ASCI-related \(\sigma\)-term for the \(Z\)-vector equation should be improved as well, particularly in terms of its parallelization.
## 4 Summary and Future Prospects
This work has developed the analytical gradient theory for ASCI-SCF corrected with second-order perturbation theory (ASCI-SCF-PT2). To this end, we have developed the Lagrangian formalism (Z-vector equation) for the ASCI-SCF reference function. Extrapolation of the ASCI-SCF-PT2 analytical gradient with respect to the PT2 correction (ASCI-SCF-PT2+X) yields an optimized geometry that closely resembles the CASSCF conformation. This development enables geometry optimizations and molecular dynamics simulations with large active spaces (that were not tractable with the CASSCF method) at almost the CASSCF level.
The effective (relaxed) density at the ASCI-SCF-PT2 level can be obtained with the Z-vector equation solutions. The quantities related to the electronic structure, such as the number of unpaired electrons, can be obtained from the relaxed density and linearly extrapolated to obtain the near-CASSCF values, as is the case for the ASCI-SCF-PT2 energy. The source codes for the ASCI-SCF-PT2 analytical gradient are distributed in the form of patches on open-source BAGEL version 2021.02.05 at [http://sites.google.com/view/cbnuqbc/codes](http://sites.google.com/view/cbnuqbc/codes) under the GPU-v3 license.
There remains plenty of room for improvement. Of course, the current computational algorithm can be made more efficient. Multireference methods are actively used in studying excited states, so extending the theory for treating excited states is promising. This will enable the simulations that were impractical with conventional CASSCF or CASPT2, such as those for highly degenerate transition metal complexes or conical intersection dynamics. For example, one can extend the CASSCF study on the pyracylene to the larger PAHs like those investigated in the previous and this work. The dynamical correlation plays an important role in geometry optimizations and dynamics, and the combinations of ASCI-SCF(-PT2) with the MRPT, MRCI,or MRCC methods indeed will constitute significant contributions. The \(Z\)-vector equation with a limited CI space can be applied for other approximate FCI methods, such as other SCI methods, RASSCF, or GASSCF, to evaluate nuclear analytical gradients of SA-RASSCF or dynamical correlation methods such as RASPT2. We will report the progress in this direction in due course.
|
10.48550/arXiv.2103.09972
|
Near-Exact CASSCF-Level Geometry Optimization with a Large Active Space using Adaptive Sampling Configuration Interaction Self-Consistent Field Corrected with Second-Order Perturbation Theory (ASCI-SCF-PT2)
|
Jae Woo Park
| 1,223
|
10.48550_arXiv.1904.10650
|
## I Introduction
The impediment to applying quantum statistical mechanics to condensed matter systems at terrestrial temperatures and densities is the horrendous scaling with system size entailed by conventional approaches. Although the field is not exactly moribund, it would be fair to say that the rate of progress has been disappointing, and that there appears no obvious way of overcoming this fundamental limitation of the existing methods.
The pressing need to approach the problem from a different direction has motivated the author to develop a new methodology that is based on a formally exact transformation that expresses the quantum partition function as an integral over classical phase space. The validity of the formulation has been verified analytically and numerically for the quantum ideal gas, and for non-interacting quantum harmonic oscillators.
Application of the algorithm to interacting systems indicate that, for a given statistical error, the algorithm scales sub-linearly with system size. This places it in a class of its own for the treatment of condensed matter quantum systems.
Validation of the algorithm for an interacting system has been provided by quantitative tests for interacting Lennard-Jones particles against benchmarks obtained with more conventional methods by Hernando and Vanicek. The latter results were exact, apart from the fact that they were numerical and 'only' 50 energy levels were used for the statistical averages. One should not understate the difficulty of obtaining numerically 50 energy levels for the Lennard-Jones system; it is undoubtedly a greater computational challenge than obtaining analytically the 20,000 energy levels used below for the present one-dimensional harmonic crystal. If nothing else the figure of 50 levels does underscore the insurmountable intractability of conventional quantum approaches.
In fact the original motivation for the present paper was to sort out a small discrepancy between the putative exact results of Hernando and Vanicek and the author's mean field, classical phase space results at the highest temperature studied. It was unclear whether the differences in the original comparison were due to the mean field approximation used in the phase space simulation, or else to the limited number of energy levels used in the exact results. Accordingly the author has undertaken to establish his own exact results for use as benchmarks, and it was found that 20,000 energy levels were necessary for reliable results at temperatures high enough to give a departure from the ground state.
The results of these tests are reported below. It is found that the mean field approximation as originally formulated is essentially exact at the highest temperatures, but the error can be on the order of 5-10% at intermediate and low temperatures, depending upon model parameters. Accordingly, this papers develops a systematic general improvement to the phase space algorithm that might be called the cluster mean field approximation. This is here implemented at the singlet level, which is the original version, and also at the pair level, which is new. It is found that the pair mean field approximation gives essentially exact results at intermediate temperatures. At low temperatures such that the system is predominantly in the ground state, the singlet and pair mean field approximations are found to be in error by on the order of 5%.
This paper relies upon earlier work, which will not be re-derived here. The reader is referred Ref. for the derivation of quantum statistical mechanics, to Ref. for the derivation of the phase space formulation (see also Ref. for some formal details concerning symmetrization of multiparticle states), and to Ref. for the exact phonon analysis of the present one-dimensional harmonic crystal.
## II Analysis and Model
### Phase Space Formulation
The details of the classical phase space formulation of quantum statistical mechanics can vary with the particular quantity being averaged.
\[\left\langle\hat{\mathcal{H}}\right\rangle^{\pm}_{N,V,T}\] \[= \frac{1}{h^{dN}N!Z^{\pm}}\int\mathrm{d}\mathbf{\Gamma}\;e^{-\beta \mathcal{H}(\mathbf{\Gamma})}W_{p}(\mathbf{\Gamma})\eta^{\pm}_{q}(\mathbf{ \Gamma})\mathcal{H}(\mathbf{\Gamma}),\]
where the partition function is
\[Z^{\pm}(N,V,T)\] \[= \frac{1}{h^{dN}N!}\int\mathrm{d}\mathbf{\Gamma}\;e^{-\beta \mathcal{H}(\mathbf{\Gamma})}W_{p}(\mathbf{\Gamma})\eta^{\pm}_{q}(\mathbf{ \Gamma}).\]
In these \(N\) is the number of particles, assumed identical and spinless, \(V\) is the volume, \(T\) is the temperature, \(\beta=1/k_{\mathrm{B}}T\) is the inverse temperature, \(h\) is Planck's constant, and \(k_{\mathrm{B}}\) is Boltzmann's constant. Also, \(\mathbf{\Gamma}=\{\mathbf{p},\mathbf{q}\}\) is a point in phase space, with the vector of particles' momenta being \(\mathbf{p}=\{\mathbf{p}_{1},\mathbf{p}_{2},\ldots,\mathbf{p}_{N}\}\), and that of the particles' position being \(\mathbf{q}=\{\mathbf{q}_{1},\mathbf{q}_{2},\ldots,\mathbf{q}_{N}\}\). Finally, \(\hat{\mathcal{H}}=\mathcal{H}(\hat{\mathbf{p}},\hat{\mathbf{q}})\) is the energy or Hamiltonian operator, and \(\mathcal{H}(\mathbf{p},\mathbf{q})\) is the classical Hamiltonian function.
The plus sign is for bosons and the minus sign is for fermions. Actually, since this has been formulated for spinless particles, these refer to the fully symmetrized and fully anti-symmetrized spatial part of the wave function (see Appendix C of Ref.). Throughout the words 'boson' and 'fermion' should be understood in this sense.
The unsymmetrized position and momentum eigenfunctions in the position representation are respectively
\[|\mathbf{q}\rangle=\delta(\mathbf{r}-\mathbf{q}),\;\mathrm{and}\;|\mathbf{p} \rangle=\frac{e^{-\mathbf{p}\cdot\mathbf{r}/i\hbar}}{V^{N/2}}. \tag{3}\]
The symmetrization function is defined as
\[\eta^{\pm}_{q}(\mathbf{p},\mathbf{q})\equiv\frac{1}{\langle\mathbf{p}| \mathbf{q}\rangle}\sum_{\hat{\mathbf{P}}}(\pm 1)^{p}\left\langle\hat{\mathbf{P}} \mathbf{p}|\mathbf{q}\rangle.\right. \tag{4}\]
Here the sum is over the \(N!\) permutation operators \(\hat{\mathbf{P}}\), whose parity is \(p\). The imaginary part of these is odd in momentum, \(\eta^{\pm}_{q}(\mathbf{p},\mathbf{q})^{*}=\eta^{\pm}_{q}(-\mathbf{p},\mathbf{ q})\).
The symmetrization function can be written as a series of loop products,
\[\eta^{\pm}_{q}(\mathbf{\Gamma}) = 1+\sum_{ij}\!\!^{\prime}\eta^{\pm}_{q;ij}+\sum_{ijk}\!\!^{ \prime}\eta^{\pm}_{q;ijk} \tag{5}\] \[\quad+\sum_{ijkl}\!\!^{\prime}\eta^{\pm}_{q;ij}\eta^{\pm}_ {q;kl}+\ldots\]
Here the superscript is the order of the loop, and the subscripts are the atoms involved in the loop. The prime signifies that the sum is over unique loops (ie. each configuration of particles in loops occurs once only) with each index different (ie. no particle may belong to more than one loop).
\[\eta^{\pm(l)}_{q;1\ldots l}=(\pm 1)^{l-1}e^{\mathbf{q}_{1l}\cdot\mathbf{p}_{l} /i\hbar}\prod_{j=1}^{l-1}e^{\mathbf{q}_{j+1,j}\cdot\mathbf{p}_{j}/i\hbar}, \tag{6}\]
This corrects a typographical error in Eq. (3.4) of Ref.
The commutation function, which is essentially the same as the function introduced by Wigner and analyzed by Kirkwood, is defined by
\[e^{-\beta\mathcal{H}(\mathbf{p},\mathbf{q})}W_{p}(\mathbf{p},\mathbf{q})= \frac{\langle\mathbf{q}|e^{-\beta\hat{\mathcal{H}}}|\mathbf{p}\rangle}{ \langle\mathbf{q}|\mathbf{p}\rangle}. \tag{7}\]
Again one has \(W_{p}(\mathbf{p},\mathbf{q})^{*}=W_{p}(-\mathbf{p},\mathbf{q})\). High temperature expansions for the commutation function have been given.
The commutation function in phase space can also be written as a series of energy eigenfunctions and eigenvalues, \(\hat{\mathcal{H}}|\mathbf{n}\rangle=E_{n}|\mathbf{n}\rangle\).
\[e^{-\beta\mathcal{H}(\mathbf{p},\mathbf{q})}W_{p}(\mathbf{p}, \mathbf{q}) = \frac{\langle\mathbf{q}|e^{-\beta\hat{\mathcal{H}}}|\mathbf{p} \rangle}{\langle\mathbf{q}|\mathbf{p}\rangle}\] \[= \frac{1}{\langle\mathbf{q}|\mathbf{p}\rangle}\sum_{\mathbf{n}}e^ {-\beta E_{n}}\left\langle\mathbf{q}|\mathbf{n}\right\rangle\langle\mathbf{n }|\mathbf{p}\rangle.\]
This exact expression forms the basis of the mean field approximation to the commutation function.
### Cluster Mean Field Approximation
In the classical phase space formulation of quantum statistical mechanics, the symmetrization function is relatively trivial to obtain and implement. The commutation function is more of a challenge, with the most successful approach using a mean field approximation that exploits the analytic form of the commutation function in the case of independent simple harmonic oscillators. This has previously been tested for the simulation of a Lennard-Jones system.
This section begins with a summary of the singlet mean field approximation to the commutation function. Then the cluster mean field approximation is given.
#### ii.2.1 Singlet Mean Field Approximation
In general, the particles of the sub-system interact via the potential energy, which is the sum of one-body, two-body, three-body terms, etc.,
\[U(\mathbf{q}) = \sum_{j=1}^{N}u^{}(\mathbf{q}_{j})+\sum_{j<k}^{N}u^{}( \mathbf{q}_{j},\mathbf{q}_{k}) \tag{9}\] \[\quad+\sum_{j<k<l}^{N}u^{}(\mathbf{q}_{j},\mathbf{q}_{k}, \mathbf{q}_{l})+\ldots\]
Distributing the energy equally, the energy of particle \(j\) can be defined as
\[U_{j}(\mathbf{q}_{j};\mathbf{q}) = u^{}(\mathbf{q}_{j})+\frac{1}{2}\sum_{k=1}^{N}\!\!{}^{(k\neq j )}\,u^{}(\mathbf{q}_{j},\mathbf{q}_{k})\]\[+\frac{1}{3}\sum_{k<l}^{N}(k,l\neq j)\,u^{}({\bf q}_{j},{\bf q}_{k},{\bf q }_{l})+. \tag{10}\]
The argument \(({\bf q}_{j};{\bf q})\) means that \({\bf q}_{j}\) is here separated out from \({\bf q}\).
The potential energy of particle \(j\) in configuration \({\bf q}\) may be expanded to second order about its local minimum at \(\overline{{\bf q}}_{j}({\bf q})\),
\[U_{j}({\bf q}_{j};{\bf q})=\overline{U}_{j}({\bf q})+\frac{1}{2}[{\bf q}_{j}- \overline{{\bf q}}_{j}][{\bf q}_{j}-\overline{{\bf q}}_{j}]:\underline{ \overline{U}}_{j}^{\prime\prime}, \tag{11}\]
The \(d\times d\) second derivative matrix for particle \(j\) at the minimum, \(\underline{\overline{U}}_{j}^{\prime\prime}=\left.\nabla_{j}\nabla_{j}U_{j}({ \bf q}_{j};{\bf q})\right|_{{\bf q}_{j}=\overline{{\bf q}}_{j}}\), is assumed positive definite.
For configurations \({\bf q}\) that have no local minimum in the potential, or that have too large a displacement \(|{\bf q}_{j}-\overline{{\bf q}}_{j}|\), the corresponding single particle commutation function can be set to unity, \(W_{j}({\bf\Gamma})=1\), (or, in the multi-dimensional case, the commutation function of the corresponding mode). This is justified by analytic results for the simple harmonic oscillator.
The positive definite second derivative matrix has \(d\) eigenvalues \(\lambda_{j\alpha}({\bf q})>0\), and orthonormal eigenvectors, \(\underline{\overline{U}}_{j}^{\prime\prime}{\bf X}_{j\alpha}=\lambda_{j\alpha }{\bf X}_{j\alpha}\), \(\alpha=x,y,\ldots,d\).
\[\omega_{j\alpha}({\bf q})=\sqrt{\lambda_{j\alpha}({\bf q})/m},\ \ \alpha=x,y, \ldots,d. \tag{12}\]
With this the potential energy is
\[U({\bf q}) = \sum_{j=1}^{N}\overline{U}_{j}+\frac{1}{2}\sum_{j=1}^{N}({\bf q} _{j}-\overline{{\bf q}}_{j})({\bf q}_{j}-\overline{{\bf q}}_{j}):\underline{ \overline{U}}_{j}^{\prime\prime} \tag{13}\] \[= \sum_{j=1}^{N}\overline{U}_{j}+\frac{1}{2}\sum_{j,\alpha}\hbar \omega_{j\alpha}Q_{j\alpha}^{2}.\]
Here \(Q_{j\alpha}\equiv\sqrt{m\omega_{j\alpha}/\hbar}\,Q_{j\alpha}^{\prime}\), and \({\bf Q}_{j}^{\prime}=\underline{\underline{X}}_{j}^{\rm T}[{\bf q}_{j}- \overline{{\bf q}}_{j}]\). (This corrects a typographical error in Eqs and in Ref..)
The mean field approximation combined with the second order expansion about the local minima maps each configuration \({\bf\Gamma}\) to a system of \(dN\) independent harmonic oscillators with frequencies \(\omega_{j\alpha}\) displacements \(Q_{j\alpha}\), and momenta \(P_{j\alpha}=\left\{\underline{\underline{X}}_{j}^{\rm T}{\bf p}_{j}\right\}_ {\alpha}/\sqrt{m\hbar\omega_{j\alpha}}\). (This corrects a typographical error in Ref..)
With this harmonic approximation for the potential energy, the effective Hamiltonian in a particular configuration can be written
\[{\cal H}^{\rm SHO}({\bf p},{\bf q}-\overline{{\bf q}})=\sum_{j=1}^{N}\overline {U}_{j}+\frac{1}{2}\sum_{j,\alpha}\hbar\omega_{j\alpha}\left[P_{j\alpha}^{2}+ Q_{j\alpha}^{2}\right]. \tag{14}\]
The commutation function for the interacting system for a particular configuration can be approximated as the product of commutation functions for effective non-interacting harmonic oscillators which have the local displacement as their argument.
\[W_{p}^{\rm mf}({\bf\Gamma}) \approx W_{p}^{\rm SHO}({\bf p},{\bf q}-\overline{{\bf q}}) \tag{15}\] \[= e^{\beta{\cal H}^{\rm SHO}({\bf p},{\bf q}-\overline{{\bf q}})} \frac{\langle{\bf q}-\overline{{\bf q}}|e^{-\beta\widehat{{\bf H}}^{\rm SHO}}| {\bf p}\rangle}{\langle{\bf q}-\overline{{\bf q}}|{\bf p}\rangle}\] \[= \prod_{j,\alpha}W_{p,j\alpha}^{\rm SHO}(P_{j\alpha},Q_{j\alpha}).\]
The harmonic oscillator commutation function for a single mode is
\[W_{p,j\alpha}^{\rm SHO}(P_{j\alpha},Q_{j\alpha})\] \[= \sqrt{2}e^{-iP_{j\alpha}Q_{j\alpha}}e^{\beta\hbar\omega_{j\alpha }[P_{j\alpha}^{2}+Q_{j\alpha}^{2}]/2}e^{-[P_{j\alpha}^{2}+Q_{j\alpha}^{2}]/2}\] \[\times\sum_{n_{j\alpha}=0}^{\infty}\frac{i^{n_{j\alpha}}e^{- \beta\hbar\omega_{j\alpha}(n_{j\alpha}+1/2)}}{2^{n_{j\alpha}}n_{j\alpha}!}{ \rm H}_{n_{j\alpha}}(P_{j\alpha}){\rm H}_{n_{j\alpha}}(Q_{j\alpha}).\]
The prefactor \(e^{-iP_{j\alpha}Q_{j\alpha}}\) corrects the prefactor \(e^{-ip_{j\alpha}q_{j\alpha}/\hbar}\) given in Eq. (5.10) of Ref.. Here \({\rm H}_{n}(z)\) is the Hermite polynomial of degree \(n\). The imaginary terms here are odd in momentum. As justified by analytic results for the simple harmonic oscillator, for configurations such that \(\underline{\overline{U}}_{j}^{\prime\prime}({\bf q})\) is not positive definite (ie. a particular eigenvalue is not positive, \(\lambda_{j\alpha}\leq 0\)), or that the displacement \(Q_{j\alpha}\) exceeds a predetermined cut-off, the corresponding commutation function can be set to unity, \(W_{p,j\alpha}^{\rm SHO}=1\).
For the averages, the momentum integrals can be performed analytically, both here and in combination with the symmetrization function. This considerably reduces computer time and substantially increases accuracy.
#### ii.1.2 Cluster Mean Field Approximation
Any configuration \({\bf q}\) can be decomposed into disjoint clusters labeled \(\alpha=1,2,\ldots\). Of the different criteria that can be used to define a cluster, perhaps the simplest is that two particles belong to the same cluster if, and only if, they are connected by at least one path of bonds. Two particles are bonded if their separation is less than a nominated length. Some clusters, perhaps the great majority, will consist of a single particle.
An even simpler definition can be made in one dimension. In this case define the pair cluster \(\alpha\), \(\alpha=1,2,\ldots,N/2\), as the nearest neighbors \(\{2\alpha-1,2\alpha\}\), irrespective of their actual separation. This criteria is used in the results presented below.
Using a separation-based criterion for the definition of a cluster is useful not only for the mean field approximation to the commutation function, but also for the calculation of the symmetrization function. Dependingon the chosen bond length, only permutations of particles within the same cluster need to be considered. (This idea is not used in the results presented below.)
The cluster energy is the internal energy plus the relevant proportion of the interaction energy with other clusters: half for pair interactions, one third for triplet interactions, etc.
\[U({\bf q}) = \sum_{j=1}^{N}u^{}({\bf q}_{j})+\sum_{j<k}u^{}({\bf q}_{j}, {\bf q}_{k})+\ldots \tag{17}\] \[= \sum_{\alpha}U_{\alpha}({\bf q}_{\alpha};{\bf q}),\]
where the energy of cluster \(\alpha\) is
\[U_{\alpha}({\bf q}_{\alpha};{\bf q}) = \sum_{j\in\alpha}u^{}({\bf q}_{j})+\sum_{j<k}\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!The gradient vanishes when
\[\overline{d}_{j}({\bf q})\ =\ \frac{\lambda}{2\overline{U}_{j}^{\prime\prime}}[d_{j+1 }+d_{j-1}]. \tag{23}\]
The second derivative is
\[\overline{U}_{j}^{\prime\prime}\equiv\nabla_{j}\nabla_{j}U_{j}(q_{j};{\bf q})= \kappa+\lambda+\frac{\lambda}{2}[\delta_{j1}+\delta_{jN}], \tag{24}\]
The potential energy of particle \(j\) in configuration \({\bf q}\) may be expanded to second order about its local minimum at \(\overline{q}_{j}({\bf q})\),
\[U_{j}(q_{j};{\bf q})\ =\ \overline{U}_{j}({\bf q})+\frac{\overline{U}_{j}^{ \prime\prime}}{2}[q_{j}-\overline{q}_{j}({\bf q})]^{2}. \tag{25}\]
This second order expansion for the potential is exact for the present harmonic crystal. Note that the most likely position of particle \(j\) for the current configuration, \(\overline{q}_{j}({\bf q})\), is not the same as its lattice position \(\dot{\overline{q}}_{j}\).
For each molecule define the frequency \(\omega_{j}=\sqrt{\overline{U}_{j}^{\prime\prime}/m}\). This is the same for all configurations \({\bf q}\).
\[U({\bf q}) = \sum_{j=1}^{N}\overline{U}_{j}+\frac{1}{2}\sum_{j=1}^{N}\overline {U}_{j}^{\prime\prime}[q_{j}-\overline{q}_{j}({\bf q})]^{2} \tag{26}\] \[= \sum_{j=1}^{N}\overline{U}_{j}+\frac{1}{2}\sum_{j}\hbar\omega_{j} Q_{j}^{2}.\]
Here \(Q_{j}\equiv\sqrt{m\omega_{j}/\hbar}\,[q_{j}-\overline{q}_{j}]\). As above, the momenta are \(P_{j}=p_{j}/\sqrt{m\hbar\omega_{j}}\).
This is the approximation in the singlet mean field approach: each frequency mode is a displacement of a single particle. One may now apply Eqs,, and for the singlet mean field commutation function.
#### ii.2.3 Pair Mean Field Approximation
For the one dimensional harmonic crystal, define a cluster pair \(\alpha\) as the nearest neighbors \(\{2\alpha-1,2\alpha\}\), \(\alpha=1,2,\ldots,\lfloor N/2\rfloor\). For simplicity, assume \(N\) even. (Contrariwise, include particle \(N\) as a singlet cluster.)
The energy of cluster \(\alpha\) is
\[U_{\alpha}({\bf q}_{\alpha};{\bf q})\] \[= \frac{\kappa}{2}\left\{d_{2\alpha-1}^{2}+d_{2\alpha}^{2}\right\} +\frac{\lambda}{2}\left\{[d_{2\alpha}-d_{2\alpha-1}]^{2}\right.\] \[\left.+\frac{1}{2}[d_{2\alpha-1}-d_{2\alpha-2}]^{2}+\frac{1}{2}[d _{2\alpha+1}-d_{2\alpha}]^{2}\right\}\] \[+\frac{\lambda}{4}[d_{2\alpha-1}-d_{2\alpha-2}]^{2}\delta_{2 \alpha-1,1}+\frac{\lambda}{4}[d_{2\alpha+1}-d_{2\alpha}]^{2}\delta_{2\alpha,N},\]
Note that the interaction with the wall particles, when present, has to be counted fully. Note also that \(d_{0}=d_{N+1}=0\).
The second derivative matrix is
\[\frac{\overline{U}_{\alpha}^{\prime\prime}}{d_{2\alpha}} = \left(\begin{array}{cc}\kappa+\frac{3\lambda}{2}+\frac{\lambda }{2}\delta_{2\alpha-1,1}&-\lambda\\ -\lambda&\kappa+\frac{3\lambda}{2}+\frac{\lambda}{2}\delta_{2\alpha,N}\end{array}\right) \tag{28}\] \[\equiv -\lambda\left(\begin{array}{cc}K_{\alpha}^{\prime}&1\\ 1&K_{\alpha}^{\prime\prime}\end{array}\right).\]
Using it gives the optimum cluster displacement as
\[\left(\begin{array}{c}\overline{d}_{2\alpha-1}\\ \overline{d}_{2\alpha}\end{array}\right)=\frac{-1/2}{K_{\alpha}^{\prime}K_{ \alpha}^{\prime\prime}-1}\left(\begin{array}{cc}K_{\alpha}^{\prime\prime}d_ {2\alpha-2}-d_{2\alpha+1}\\ -d_{2\alpha-2}+K_{\alpha}^{\prime}d_{2\alpha+1}\end{array}\right). \tag{29}\]
The eigenvalues of the second derivative matrix (without \(-\lambda\)) are
\[\mu_{\alpha}^{\pm}\ =\ \frac{1}{2}(K_{\alpha}^{\prime}+K_{\alpha}^{\prime\prime}) \pm\frac{1}{2}\sqrt{(K_{\alpha}^{\prime}-K_{\alpha}^{\prime\prime})^{2}+4}. \tag{30}\]
Since the \(K_{\alpha}\) are negative, so are the \(\mu_{\alpha}^{\pm}\).
Writing the eigenvectors as \({\bf u}_{\alpha}^{\pm}=c_{\alpha}^{\pm}(1,a_{\alpha}^{\pm})\), \(c_{\alpha}^{\pm}\equiv 1/\sqrt{1+a_{\alpha}^{\pm 2}}\), from the eigenvalue equation one obtains \(a_{\alpha}^{\pm}=\mu_{\alpha}^{\pm}-K_{\alpha}^{\prime}\). This is presumably equivalent to \(a_{\alpha}^{\pm}=1/[\mu_{\alpha}^{\mp}-K_{\alpha}^{\prime}]\).
The mode frequencies are \(\omega_{\alpha}^{\pm}=\sqrt{-\lambda\mu_{\alpha}^{\pm}/m}\), which are real.
\[{\bf Q}_{\alpha} = \underline{\omega}_{\alpha}\underline{X}_{\alpha}^{\rm T} \Delta{\bf d}_{\alpha}\] \[= \left(\begin{array}{c}\sqrt{m\omega_{\alpha}^{+}/\hbar}\,\{c_{ \alpha}^{+}\Delta d_{2\alpha-1}+c_{\alpha}^{+}a_{\alpha}^{+}\Delta d_{2 \alpha}\}\\ \sqrt{m\omega_{\alpha}^{-}/\hbar}\,\{c_{\alpha}^{-}\Delta d_{2\alpha-1}+c_{ \alpha}^{-}a_{\alpha}^{-}\Delta d_{2\alpha}\}\end{array}\right),\]
The contribution to the total energy from cluster \(\alpha\) in configuration \({\bf\Gamma}\) is
\[{\cal H}_{\alpha}({\bf\Gamma}) = \frac{1}{2m}\left[p_{2\alpha-1}^{2}+p_{2\alpha}^{2}\right] \tag{32}\] \[\ \ \ +\overline{U}_{\alpha}+\frac{1}{2}\overline{\underline{U}}_{ \alpha}^{\prime\prime}:[{\bf q}_{\alpha}-\overline{{\bf q}}_{\alpha}]\,[{\bf q}_{ \alpha}-\overline{{\bf q}}_{\alpha}]\] \[= \overline{U}_{\alpha}+\frac{\hbar\omega_{\alpha}^{+}}{2}\left\{P_{ \alpha,+}^{2}+Q_{\alpha,+}^{2}\right\}\] \[\ \ \ +\frac{\hbar\omega_{\alpha}^{-}}{2}\left\{P_{\alpha,-}^{2}+Q_{ \alpha,-}^{2}\right\}.\]
In the computational implementation of the algorithm, these can be used to replace directly the singlet mean field terms, \(P_{\alpha}^{+}\Rightarrow P_{2\alpha-1}\), \(P_{\alpha}^{-}\Rightarrow P_{2\alpha}\), and similarly for \(Q_{\alpha}^{\pm}\) and \(\omega_{\alpha}^{\pm}\).
#### ii.2.4 Symmetrization Function
Symmetrization consists of a sum over all particle permutations. However because of the highly oscillatory Fourier contributions to the loop symmetrization function, Eq., only permutations of closely separated particles actually contribute to the statistical average.
In view of this one can define the permutation length,
\[d_{\rm m}(\hat{\rm P})\equiv\sum_{j=1}^{N}|j-j^{\prime}|,\;\;j^{\prime}\equiv\{ \hat{\rm P}\}_{j}. \tag{33}\]
One can see that \(d_{\rm m}=0\) corresponds to the identity permutation, \(d_{\rm m}=2\) corresponds to a single nearest neighbor transposition (dimer), and \(d_{\rm m}=4\) corresponds to either two distinct nearest neighbor transpositions (double dimer), or else a single cyclic permutation of three consecutive particles (trimer). One expects that the contributions to the symmetrized wave function will decrease with increasing permutation length.
Hence one can set an upper limit on the length of the permutations that are included in the symmetrization function. The numerical results below show that by far the greatest contribution comes from the identity permutation alone, \(d_{\rm m}=0\). In some cases a measurable change occurs by including also nearest neighbor permutations, \(d_{\rm m}=2\). Measurable but smaller change occurs upon also including permutations of length \(d_{\rm m}=4\) (not shown below). For \(N\) particles, the number of permutation terms that contribute to the symmetrization function is 1 for \(d_{\rm m}=0\), \(1+(N-1)\) for \(d_{\rm m}\leq 2\), and \(N+(N-2)(N-3)/2+2(N-2)\) for \(d_{\rm m}\leq 4\).
### Simulation Algorithm
The simulation algorithm was as previously described. Briefly the Metropolis algorithm in position space was used with the usual classical Maxwell-Boltzmann weight. The various momentum integrals were performed analytically. Averages were evaluated by umbrella sampling using the commutation function and symmetrization function as weight. Since three versions of the commutation function (unity (ie. classical), singlet mean field, pair mean field), and three versions of the symmetrization function (\(d_{\rm m}=0\) (ie. classical), and \(d_{\rm m}\leq 2\), for bosons and for fermions) were tested, 9 different averages for each quantity were obtained simultaneously.
Typically enough configurations were generated to make the statistical error less than 1%, sometimes much less. In the simulations the time depends on how many Hermite polynomials are used for the commutation function (eight in the results reported below; tests with six and twelve showed no great effect), and the cut-off for the mode amplitude beyond which the commutation function was set to unity (\(Q^{\rm cut}=1\) in the results below; tests with \(Q^{\rm cut}=2\) showed no great effect).
Interestingly enough, for an accuracy of about 1%, the Monte Carlo algorithm was a factor of about \(2,000\) times more efficient (in terms of total computer time) than the quasi-analytic exact phonon method. The main bottleneck in the latter was the crude numerical quadrature method that was used to evaluate the symmetrization function and density profile, and this was exacerbated by the large number of energy levels that were required for accurate results at higher temperatures. For this particular comparison at \(\beta\hbar\omega_{\rm LJ}=2\), \(l^{\rm max}=20,000\) energy levels were necessary; grid parameters \(\Delta_{Q}=0.15\) and \(Q^{\rm max}=6\) gave a quadrature error of about \(.8\%\). The phonon method is more efficient in one respect, namely that it requires negligible computer time for each additional temperature point; the simulations give results for only one temperature at a time.
The Lennard-Jones frequency used to scale the results below is \(\omega_{\rm LJ}=3.28\times 10^{12}\,\)Hz, the mass is \(m=3.35\times 10^{-26}\,\)kg, the well-depth is \(\varepsilon=4.93\times 10^{-22}\,\)J, and the equilibrium separation is \(r_{\rm e}=3.13\times 10^{-10}\,\)m. These parameters are appropriate for neon.
## III Results
At low temperatures the phonon modes are in the ground state, \(E_{1}=\sum_{n=1}^{N}\hbar\omega_{n}/2\), and at high temperatures the system approaches the classical result, \(\langle{\cal H}\rangle_{\rm cl}=N/\beta\). It can be seen in the main part of the figure that both the singlet and pair mean field approximations are in relatively good agreement with these limiting results and with the exact calculations over the whole temperature regime shown.
1, symmetrization effects are negligible and only the \(d_{\rm m}=0\) calculations are shown.
There are two approximations in the exact calculations: the number of energy levels used and the domain and spacing of the grid used for the numerical quadrature. (I persist in calling these results 'exact' because they use explicit analytic expressions for the energy eigenvalues and eigenfunctions.) The quadrature affects only the average density profile without symmetrization, and also the average energy and the average density profile with symmetrization. Hence the exact calculations of the average energy in the absence of symmetrization effects, \(d_{\rm m}=0\), as in Fig. 1, are approximate only as regards to the number of energy levels that are retained.
The exact calculations in use 10,000 energy levels. These are adequate for low and intermediate temperatures, \(\beta\hbar\omega_{\rm LJ}\gtrsim 0.24\), judged in part by comparison with results using 5,000 energy levels. The exact data begins to underestimate the classical result for higher temperatures than this, and are not shown in One might speculate that the exact quantum result for the average energy of the harmonic crystal should approach the classical limit from above. Using fewer energy levels reduces the domain of inverse temperatures in which the exact calculations are reliable.
Both the singlet and pair mean field simulations are practically exact at higher temperatures, \(\beta\hbar\omega_{\rm LJ}\lesssim 0.5\) in this case. As the temperature is decreased, the singlet mean field energy lies between the exact energy and the classical energy. The pair mean field result lies between the singlet mean field energy and the exact energy. It can be seen that at low temperatures the classical energy is substantially less than the exact ground state energy, but the pair mean field energy is only slightly less than the exact ground state energy. It may be concluded that the mean field approach is better than a high temperature expansion in that it yields the dominant quantum correction to the classical result over the entire range of temperatures.
The inset of scales the average energy by the inverse temperature and focusses on the low temperature regime. For inverse temperatures \(\beta\hbar\omega_{\rm LJ}\gtrsim 3\), the exact results with 10,000 energy levels are practically indistinguishable from the ground state energy. At low temperatures, both mean field classical phase space approximations give a lower energy than the ground state. For example, in the case of Fig. 1, the ground state energy is \(E_{1}/\hbar\omega_{\rm LJ}=3.385\). At \(\beta\hbar\omega_{\rm LJ}=10\), the singlet mean field theory gives \(\langle\hat{\cal H}\rangle/\hbar\omega_{\rm LJ}=3.116\pm.002\), and the pair mean field theory gives \(3.305\pm.001\). At this temperature the classical result given by the simulation was \(0.4002\pm.0002\), which is rather close to the exact classical result of 0.4. Here and throughout, the statistical error quoted for the simulations is twice the standard error on the mean, which is the 96% confidence level.
One can conclude from the data that at low temperatures such that the system is close to the ground state, the mean field approximations remain viable. The pair mean field approach substantially reduces the error in the singlet mean field approach. In the absence of exact data, the difference between the pair and singlet mean field results would give a guide to the quantitative accuracy of the former.
It is worth mentioning that at each temperature the classical phase space simulations took about five minutes on a desktop personal computer to obtain the quoted accuracy. In comparison, it took about 2 days to obtain the exact results with these energy levels and quadrature grid, the latter being the time limiting part of the computations. (This is independent of how many temperature points are saved.)
There is no singlet potential, \(\kappa=0\), and the nearest neighbor spring constant has been decreased by a factor of 50, \(\lambda/m\omega_{\rm LJ}^{2}=0.02\). The lattice spacing and spring length is unchanged. The ground state energy is \(E_{1}=0.3757\), which is approximately one tenth that for the parameters of
In the main body of the results of both mean field approximations appear indistinguishable from the exact results. At \(\beta\hbar\omega_{\rm LJ}=3\), the exact energy is 1.3707 (for \(l_{\rm max}=20,000\); 1.3627 for 10,000), the singlet mean field gives \(1.3322\pm.0009\) and the pair mean field gives \(1.3443\pm.0008\). The exact classical result here is 4/3, compared to \(1.3329\pm.0007\) given by the simulations. These results are for \(d=0\), so no symmetrization effects are included. For inverse temperatures \(\beta\hbar\omega_{\rm LJ}\lesssim 1.8\), the mean field algorithms appear more reliable than the exact results with \(l_{\rm max}=20,000\).
The inset of shows the average energy for bosons less that for fermions, \([\langle\hat{\cal H}\rangle^{+}-\langle\hat{\cal H}\rangle^{-}]/\hbar\omega_{ \rm LJ}\), \(d_{\rm m}\leq 2\).
Average energy versus inverse temperature (\(N=4\), \(\Delta_{q}=r_{\rm e}\), \(\kappa=0\), \(\lambda/m\omega_{\rm LJ}^{2}=0.02\), \(d_{\rm m}=0\)). The solid curve is the exact result using \(l_{\rm max}=20,000\) energy levels. The triangles are the singlet and the circles are the pair mean field Monte Carlo simulations, respectively. The dotted curve is the classical result, \(\langle E\rangle_{\rm cl}=N/\beta\), and the dashed line is the ground state energy. **Inset.** The boson energy minus the fermion energy using \(d_{\rm m}\leq 2\). The crosses are Monte Carlo simulations with the classical commutation function, \(W=1\).
At lower temperatures, \(\beta\hbar\omega_{\rm LJ}\gtrsim 1.5\), the difference is positive, which means that the energy for bosons is greater than that for fermions. It can be seen that the peak difference, which occurs at \(\beta\hbar\omega_{\rm LJ}=1.9\), is about \(3\%\) of the actual energy (exact results). (The error in the numerical quadrature used for the exact result is \(2\%\) at this temperature for the energy with \(d_{\rm m}=0\). Hopefully for \(d_{\rm m}\leq 2\) this error is the same for bosons as for fermions and therefore cancels.) The classical phase space results may be described as qualitatively correct and perhaps semi-quantitative in accuracy. For \(\beta\hbar\omega_{\rm LJ}\gtrsim 2\), the singlet mean field approximation overestimates the energy difference, whereas the pair mean field approximation perhaps halves the error. The classical results, with commutation function \(W=1\), performs surprisingly well in this low temperature regime. At higher temperatures than this all four approaches indicate that the energy difference turns negative. The extent to which the singlet and pair mean field predictions agree with each other gives an indication of their reliability in this regime. Although the exact results are terminated at the estimated limit of reliability of the energy, one should note that the results in the inset of represent the difference between two relatively large terms, and the effects of any errors or approximation are accordingly magnified.
In Ref., the non-monotonic behavior of the energy difference was attributed to two competing effects: on the one hand the thermal wavelength increases with decreasing temperature, and on the other the particles become more confined to their lattice positions as the temperature decreases, which reduces the amount of overlapping wave function and non-zero symmetrization exchange.
The density peaks are rather broad, with the central two particles merging into a single peak. Interestingly enough, the density profile spills over beyond the wall particles at \(q_{0}=0\) and \(q_{5}=5r_{\rm e}\). There is good agreement between all four methods, with the classical phase space simulations being closer to the exact results at the shoulders of the density profile.
The inset of the figure shows the difference between the density of bosons and that for fermions with symmetrization effects accounted for by only nearest neighbor transpositions (dimers). The phase space simulations may again be described as quantitatively correct. Evidently the bulk of the symmetrization effects are captured using the classical commutation function, \(W=1\).
It can be seen that the exact energy approaches the classical energy from above as the temperature is increased. Without symmetrization effects (main part of figure) the mean field simulations lie between the exact and the classical results, with the pair mean field results lying closer to the exact results than the singlet mean field results across the temperature range shown.
Both mean field results lie below the ground state energy at low temperatures. For example, in this case the ground state energy is \(E_{1}/\hbar\omega_{\rm LJ}=2.6569\), and at \(\beta\hbar\omega_{\rm LJ}=10\), the singlet mean field gives \(\langle\hat{\cal H}\rangle/\hbar\omega_{\rm LJ}=2.336\pm.001\), and the pair mean field gives \(2.575\pm.001\). Both are substantially more accurate than the classical simulation result of \(0.4001\pm 0.0002\).
The inset to shows the energy for bosons less that for fermions, calculated by including only nearest neighbor transpositions, \(d_{\rm m}=2\). It can be seen that there is a pole in the exact results for fermions at \(\beta\hbar\omega_{\rm LJ}\approx 0.72\).
Density profile at \(\beta\hbar\omega_{\rm LJ}=2\) (\(N=4\), \(\Delta_{q}=r_{\rm e}\), \(\kappa=0\), \(\lambda/m\omega_{\rm LJ}^{2}=0.02\), \(d_{\rm m}=0\)). The solid curve is the exact result using \(l_{\rm max}=20,000\) energy levels, and the symbols are the classical phase space Monte Carlo simulations, with the crosses being the classical result, \(W=1\), the triangles using the singlet and the circles using the pair mean field commutation function. **Inset.** The boson density minus fermion density using \(d_{\rm m}\leq 2\).
(This pole disappears when two nearest neighbor dimer transpositions or a cyclic permutation of three consecutive particles, \(d_{\rm m}=4\), are included.) It can be seen that the classical, \(W=1\), singlet mean field, and pair mean field commutation functions all give this pole for \(d_{\rm m}\leq 2\) at about the same location. There is little to choose between the three at high temperatures; the apparent agreement of the mean field approximations with the exact results for \(\beta\hbar\omega_{\rm LJ}\leq 0.3\) should not be taken seriously because this is about the limit of reliability of the exact results with \(l_{\rm max}=20,000\) in this case. At intermediate and low temperatures, \(\beta\hbar\omega_{\rm LJ}\gtrsim 1\), the classical results lie closer to the exact results than do the mean field results. It is difficult to obtain the results for fermions accurately when the denominator passes through zero.
It can be seen that there is essentially a single density peak in the center of the system, and that the density spills beyond the wall particles at \(q_{0}=0\) and \(q_{5}=0.5\). There is little to choose between the three simulation algorithms, at least in the case of monomers (ie. no symmetrization effects, \(d_{\rm m}=0\)). Compared to the exact calculations all three simulation algorithms are slightly broader at the peak.
For the case of bosons, Fig. 5B, including nearest neighbor transpositions makes the profile slightly narrower and more sharply peaked. There is again good agreement between the three simulation algorithms, with the pair mean field approach slightly underestimating the height of the density peak. For the case of fermions, Fig. 5C, the exact calculations give a single peak, that is narrower and higher than for bosons. There is no dip in the center of this peak, as one might have expected from Fermi repulsion. (At the lower temperature of \(\beta\hbar\omega_{\rm LJ}=2\), the pair mean field profile shows a bifurcated peak, but the other methods show a single peak similar to here). The classical and the singlet mean field approaches underestimate the height of the peak, with the latter being closer to the exact results than the former. The pair mean field approach overestimates the height of the peak. All three simulation algorithms miss the broad base to the density profile at the walls that is given by the exact calculations. Arguably for reliable results for fermions at this density and temperature, one should go beyond single dimer transpositions.
## IV Conclusion
This paper has been concerned with ascertaining the accuracy of a quantum Monte Carlo algorithm for an interacting system. The algorithm is based on a formulation of quantum statistical mechanics in classical phase space and it uses a mean field approximation for the commutation function. For the tests reliable exact results were required as benchmarks, and these were obtained for a one-dimensional harmonic crystal for which the energy eigenvalues and eigenstates can be expressed analytically in closed form. The analytic nature of the exact results for this model allowed up to 20,000 energy levels to be used to establish the benchmark results for the tests. Previous tests of the mean field classical phase space formulation of quantum statistical mechanics used benchmarks established for a one-dimensional Lennard-Jones model with 50 energy eigenvalues obtained numerically. Here it was found that the number of energy levels has a significant effect on the statistical average at higher temperatures, and so the present benchmarks can be relied upon in this regime.
Two versions of the mean field approximation were tested: the singlet version, which has previously been published, and a pair version, which is new here. It was found that the singlet mean field approach was qualitatively correct over the whole temperature (and density and coupling) regime studied. It appeared to be exact in the high temperature limit, and it was generally better than 10% accurate in the low temperature regime in which the system was predominately in the quantum ground state. The pair mean field algorithm significantly improved the accuracy of the singlet algorithm in the intermediate and low temperature regime. In the absence of benchmark results, the difference between the singlet and pair mean field predictions can be used as a guide to the quantitative accuracy of the latter.
There appear to be at least two advantages to the present mean field treatment of the commutation function compared to evaluating it from high temperature expansions. First, the mean field expressions remain accurate across the entire temperature regime, including the ground state. Second, algebraically the mean field expressions are relatively simple, and computationally they are easy to implement and efficient to evaluate. In contrast the high temperature expansions rapidly become algebraically complex as higher order terms are included, and there are corresponding challenges in their computational implementation and numerical evaluation.
The present paper also explored wave function symmetrization effects. This was at the dimer level, which means the transposition of nearest neighbor particles. It was found, somewhat surprisingly, that combining classical Monte Carlo in classical phase space with the symmetrization function (ie. neglecting the commutation function) in some, but not all, cases gave as good results as those obtained retaining the mean field commutation function. At the highest density studied, some features of the symmetrized system were not captured entirely by the present phase space simulations, particularly in the case of fermions. This suggests that retaining further terms in the symmetrization loop expansion (eg. double dimer and trimer) may be necessary in some cases. It may also be worth reflecting on the underlying philosophy of the mean field approach in the presence of wave function symmetrization.
Finally, a rather interesting conclusion from the present and earlier results is that the classical component dominates, not just in the high temperature limit but even in the quantum ground state (for structure, not energy). Of course quantum effects are not entirely absent, and when present these are captured by the present mean field commutation function and also the symmetrization function, but it is clear that these truly are a perturbation on the classical prediction. This underscores the utility of treating real world condensed matter systems via quantum statistical mechanics formulated as an integral over classical phase space, rather than formulating it as a sum over quantum states, or by parameterizing the wave function.
|
10.48550/arXiv.1904.10650
|
Quantum Monte Carlo in Classical Phase Space. Mean Field and Exact Results for a One Dimensional Harmonic Crystal
|
Phil Attard
| 5,990
|
10.48550_arXiv.2407.06571
|
## I Introduction
Quantum chemistry has appeared as one of the main contenders in the race for quantum supremacy.
\[\hat{H}=\sum_{ij}^{N}h_{ij}\hat{F}_{j}^{i}+\sum_{ijkl}^{N}g_{ijkl}\hat{F}_{j}^{ i}\hat{F}_{l}^{k}, \tag{1}\]
Footnote 1: Representing the Hamiltonian using only excitation operators is usually referred to as chemists’ notation. This entails a modification to the one-electron tensor with respect to physicists’ notation, which uses normal-ordered operators of the form \(\hat{a}_{i\sigma}^{\dagger}\hat{a}_{i\sigma}\hat{a}_{i\sigma}\). Our notation is related to the electronic integrals by \(g_{ijkl}=\frac{1}{2}\int\int d\vec{r}_{1}d\vec{r}_{2}\,\frac{\phi_{i}^{\dagger} (\vec{r}_{1})\phi_{j}(\vec{r}_{1})\phi_{k}^{\dagger}(\vec{r}_{2})\phi_{l}(\vec {r}_{2})\phi_{l}(\vec{r}_{2})}{\left|\vec{r}_{1}-\vec{r}_{2}\right|}\) and \(h_{ij}=-\sum_{k}g_{ikkj}+\int d\vec{r}\phi_{i}^{\dagger}(\vec{r})\Big{(}-\frac{ \nabla^{2}}{2}-\sum_{n}\frac{Z_{n}}{\left|\vec{r}-\vec{R}_{n}\right|}\Big{)} \phi_{j}(\vec{r})\), with \(\phi_{i}(\vec{r})\) the one-particle electronic basis functions, and \(Z_{n}/\vec{R}_{n}\) the charge/position of nucleus \(n\).
Efficiently finding eigenvalues of the electronic structure Hamiltonian can be achieved on a quantum computer through algorithms such as quantum phase estimation (QPE) and its variants. These algorithms require the implementation of functions of the Hamiltonian, such as the time evolution operator \(e^{-i\hat{H}t}\), within the quantum circuit to incorporate information from \(\hat{H}\). Generally, there are two approaches for encoding \(\hat{H}\) on a quantum circuit: Trotter product formulas and LCU-based encodings. Product formula approaches involve decomposing \(\hat{H}\) into a sum of fast-forwardable fragments to implement the time-evolution operator. Despite its advantages and low qubit costs, this family of approaches does not provide the optimal \(\sim\mathcal{\bar{6}}(t)\) time scaling. On the other hand, LCU-based encodings construct an \(\hat{H}\) oracle circuit and enable the implementation of an arbitrary polynomial of \(\hat{H}\). The LCU encoding also gives rise to the optimization approach implementing a walk operator denoted as \(\hat{\Psi}\), which has eigenvalues \(e^{\pm i\text{arccos}(E_{n}/\lambda)}\), where \(E_{n}\) are the eigenvalues of \(\hat{H}\) (see Fig.1) and \(\lambda=\sum_{k}\left|u_{k}\right|\) is a 1-norm of a coefficient vector for an LCU decomposition of \(H\)
\[\hat{H}=\sum_{k=1}^{K}u_{k}\hat{U}_{k}. \tag{2}\]
Here, \(\hat{U}_{k}\) are unitary operators, \(u_{k}\) are complex coefficients, and \(K\) is the total number of unitaries. By estimating the phase of the walk operator, the eigenvalues of \(\hat{H}\) can be extracted from \(\arccos(E_{n}/\lambda)\). The accuracy associated with the energy estimation is upper-bounded by \(\epsilon\leq\lambda e_{W}\), where \(\epsilon_{W}\) is the walk operator phase estimation accuracy. Consequently, employing an optimal phase estimation algorithm with Heisenberg scaling \(\tilde{\mathcal{\bar{6}}}(1/\epsilon_{W})\), the algorithm's cost to achieve a target energy accuracy of \(\epsilon\) would require \(\tilde{\mathcal{\bar{6}}}(\lambda/\epsilon)\) calls to the walk operator. It is evident that the cost of the LCU-based energy estimation method depends on both the 1-norm \(\lambda\) and the implementation cost of the walk operator.
In this work, we consider lowering the circuit cost of the LCU-based Hamiltonian encoding by optimizing the choice of unitary operators. The circuit cost of LCU decomposition is proportional to its 1-norm \(\lambda\). By employing the LCU decomposition, we can construct a Hamiltonian oracle, which corresponds to a circuit that encodes the action of \(\hat{H}/\lambda\) on a quantum register, as illustrated in The spectral range \(\Delta E\equiv E_{\text{max}}-E_{\text{min}}\), where \(E_{\text{max(min)}}\) represents the maximum (minimum) eigenvalue of \(\hat{H}\), provides a lower bound for the 1-norm, given by \(\lambda\geq\Delta E/2\). Based on this consideration, it becomes evident that the eigenvalues of \(\hat{H}/\lambda\) after a constant shift will fall within the range of \([-1,1]^{2}\). Having the Hamiltonian's spectrum within this range is necessary to ensure that the block encoding defines a unitary transformation in the quantum register plus ancilla space. Numerous LCU decompositions have been proposed for the electronic structure Hamiltonian. In this work, we present a unified framework for expressing the electronic structure Hamiltonian as an LCU. In addition, by working using polynomials of Majorana operators, we expect these decompositions to be applicable to other kind of Hamiltonians, such as the second-quantized form of a vibrational Hamiltonian. This unified framework enables us to establish connections and to generalize existing decomposition methods.
To provide an appropriate context for comparison of different LCUs, let us consider the main contributions to their quantum resource cost. The LCU-based Hamiltonian oracle shown in comprises two primary components: the PREPARE circuit, which transforms the ancilla state \(\ket{0}_{a}\) into a linear superposition with the LCU coefficients, and the SELECT circuit, which applies the LCU unitaries in a controlled manner. In general, it is possible to move operations from SELECT into PREPARE and vice-versa, and for a given LCU there might be several different associated oracles that realize the block-encoding. The unitaries in SELECT correspond to multiplexed operators, and their quantum gate cost depends on that of the associated unitaries \(\hat{U}_{k}\) and a unary iteration over \(k\) indices. The unary iteration is analogous to a for-loop over indices \(k\) conditionally applying \(\hat{U}_{k}\), and it can be efficiently incorporated by using a binary tree algorithm. This incurs an overhead scaling of \(4(K-1)\) T-gates and \(\lceil\log_{2}(K)\rceil-1\) additional ancilla qubits. However, specific structures within the LCU decomposition can enable more efficient SELECT circuits. We will explore this aspect further after introducing the different LCU decomposition techniques.
The rest of this paper is organized as follows. Section II presents the unified Majorana tensor decomposition (MTD) framework for LCU decompositions and a new LCU decomposition. Then a list of recently proposed LCU decompositions expressed as MTDs is given in Sec. III. Finally, benchmarks for small systems and concluding remarks are offered in Sec. IV.
## II Majorana tensor decomposition framework
We will use the Majorana operators formalism because it provides a convenient bridge between fermionic and qubit operators.
\[\hat{\gamma}_{j\sigma,0} \equiv\hat{a}_{j\sigma}+\hat{a}_{j\sigma}^{\dagger} \tag{3}\] \[\hat{\gamma}_{j\sigma,1} \equiv i\left(\hat{a}_{j\sigma}^{\dagger}-\hat{a}_{j\sigma}\right), \tag{4}\]
with the algebraic relations
\[\{\hat{\gamma}_{p},\hat{\gamma}_{q}\} =2\delta_{pq}\hat{1} \tag{5}\] \[\hat{\gamma}_{p}^{\dagger} =\hat{\gamma}_{p}\] \[\hat{\gamma}_{p}^{2} =\hat{1}, \tag{7}\]
where we used complex indices \(p\) and \(q\) containing sub-indices \(\{j\sigma,m\}\): \(m\in\{0,1\}\), \(j\) is a spacial orbital in
LCU-based implementation of walk operator with eigenvalues \(e^{\pm i\arccos\frac{\hat{H}}{\lambda}}\), for \(\lambda=\sum_{k}|c_{k}|\) the 1-norm of the LCU. Note the Z gate is applied conditioned on the state of all qubits in register \(k\) to be zero.
The electronic structure Hamiltonian for closed-shell systems can be written in the Majorana representation as
\[\hat{H} =\left(\sum_{i}h_{ii}+\sum_{ij}g_{iijj}\right)\hat{1}\] \[+\frac{i}{2}\sum_{\sigma}\sum_{ij}\left(h_{ij}+2\sum_{k}g_{ijkk} \right)\hat{\gamma}_{i\sigma,0}\hat{\gamma}_{j\sigma,1}\] \[-\frac{1}{4}\sum_{\sigma\tau}\sum_{ijkl}g_{ijkl}\hat{\gamma}_{i \sigma,0}\hat{\gamma}_{j\sigma,1}\hat{\gamma}_{k\tau,0}\hat{\gamma}_{l\tau,1}. \tag{8}\] \[\equiv\hat{H}_{0}+\hat{H}_{1}+\hat{H}_{2}, \tag{9}\]
Since no terms with more than four Majorana operators appear in \(\hat{H}\), a general MTD representation of the Hamiltonian can be written as
\[\hat{H}=\sum_{\begin{subarray}{c}w_{1},...,w_{L}\\ m_{1},...,m_{L}\end{subarray}}\Omega_{\vec{w},\vec{m}}\prod_{\nu=1}^{L}\hat{V} _{\vec{m}}^{(\nu)\dagger}\hat{p}_{\vec{w}}^{(\nu)}\hat{V}_{\vec{m}}^{(\nu)}, \tag{10}\]
We also constrain the product \(\prod_{\nu=1}^{L}\hat{p}_{\vec{w}}^{(\nu)}\) to yield a Majorana polynomial of degree \(\leq 4\) for all \(w_{1},...,w_{L}\) appearing in the sum, which restricts \(L\leq 4\). For all decompositions shown in this work only spacial orbital rotations are used, which form a subgroup of \(\text{Spin}(4N)\)
\[\hat{U}_{\vec{\theta}} \equiv\exp\left[\sum_{i>j}\theta_{ij}(\hat{F}_{j}^{i}-\hat{F}_{i} ^{j})\right]\in\text{Spin(N)} \tag{11}\] \[=\exp\left[\sum_{i>j}\frac{\theta_{ij}}{2}\sum_{\sigma}(\hat{ \gamma}_{i\sigma,0}\hat{\gamma}_{j\sigma,0}+\hat{\gamma}_{i\sigma,1}\hat{ \gamma}_{j\sigma,1})\right]. \tag{12}\]
For example, in the linear case \(\hat{p}_{1}^{\nu}=\gamma_{1\sigma,x}\), where \(x=\{0,1\}\), all other linear in Majorana unitary operators can be obtained by conjugating with \(\hat{U}_{\vec{\theta}}\) these two polynomials. For higher polynomial degrees, there is a still relatively low number of distinct \(\hat{p}_{\vec{w}}^{(\nu)}\)'s that give non-overlapping sets of unitary operators upon \(\hat{U}_{\vec{\theta}}\) conjugation.
Note that even though \(\hat{H}\) could be in principle decomposed with unitaries that are Majorana polynomials of degree \(>4\), these contributions would have to cancel to recover \(\hat{H}\). Since there is an exponential number of Majorana polynomials with degree larger than 4, keeping track of these operators and enforcing their cancellation is a computationally daunting task. By enforcing conservation of the Majorana polynomial degree and starting with polynomials of degree \(\leq 4\), we do not have to consider an exponentially large space, making the space of unitaries for the LCU more computationally tractable.
### Mtd-\(L^{4}\) decomposition
Here we start with a novel LCU decomposition, MTD-\(L^{4}\), that has not been suggested before. The idea behind the MTD-\(L^{4}\) decomposition is representing unitaries as products of four unitaries that are first degree polynomials of Majoranas, thus having \(L=4\). A way of writing this decomposition that conserves the spin-symmetry in Eq.
\[\hat{q}_{\vec{m}\sigma,x}^{(\nu)}=\hat{U}_{\vec{m}}^{(\nu)\dagger}\hat{\gamma} _{1\sigma,x}\hat{U}_{\vec{m}}^{(\nu)} \tag{14}\]
Here, we skip the \(\hat{H}_{1}\) part since there is a known optimal treatment of this part, and it will be discussed with other fermionic LCU decomposition techniques in Sec. III.2. Note that real orbital rotations \(\hat{U}\)'s [Eq.] do not introduce Majoranas with different \(x\) or \(\sigma\) indices:
\[\hat{U}^{\dagger}\hat{\gamma}_{i\sigma,x}\hat{U}=\sum_{j}U_{ij}\hat{\gamma}_{ j\sigma,x}. \tag{15}\]
Thus, the MTD-\(L^{4}\) decomposition expresses the four-Majorana tensor, with associated rank = 4, as a linear combination of outer products of four rank-1 vectors, with a trivial index \(\vec{w}=1\). Generally, the orbital rotations can be implemented as products of Givens rotations with \(\mathcal{O}(N^{2})\) operations for rotating \(N\) orbitals. However, the orbital rotations shown here only act on a single spacial orbital, thus, only \(\mathcal{O}(N)\) operations are necessary to implement these \(\hat{U}\)'s using Givens rotations. The corresponding angles of the Givens rotations can be found once the coefficients in Eq. have been specified; a detailed discussion of how to find these coefficients is shown in Sec. III.1.2.
Next, we show different ways of obtaining the MTD-\(L^{4}\) decomposition using common tensor decomposition techniques.
#### ii.1.1 MPS-based MTD-\(L^{4}\)
Matrix product state (MPS) formalism applied to 4-index tensor \(g_{pqrs}\) can be seen as iterative application of the singular value decomposition (SVD) (see Fig.).
\[A_{uv}=\sum_{\alpha}U_{u\alpha}S_{\alpha}W_{v\alpha}. \tag{16}\]
Treating the \(g_{ijkl}\) tensor as a matrix with indices \(i\) and \(jkl\) allows to start the decomposition as follows
\[g_{ijkl} =\sum_{u=1}^{N}U_{iu}^{}S_{u}^{}W_{jklu}^{} \tag{17}\] \[=\sum_{u=1}^{N}\sum_{v=1}^{N^{2}}S_{u}^{}S_{v}^{}U_{iu}^{(1 )}U_{jv}^{(2,u)}W_{klv}^{}\] \[=\sum_{u,w=1}^{N}\sum_{v=1}^{N^{2}}S_{u}^{}S_{v}^{}S_{w}^{( 3)}U_{iu}^{}U_{jv}^{(2,u)}U_{kw}^{(3,v)}W_{lw}^{}. \tag{19}\]
From this, we can write
\[\hat{H}_{2} =\sum_{ijkl}\sum_{uvw}\Omega_{uvw}U_{iu}^{}U_{jv}^{(2,u)}U_{kw} ^{(3,v)}W_{lw}^{}\sum_{\sigma\tau}\hat{\gamma}_{i\sigma,0}\hat{\gamma}_{j \sigma,1}\hat{\gamma}_{k\tau,0}\hat{\gamma}_{l\tau,1} \tag{20}\] \[=\sum_{uvw}\Omega_{uvw}\sum_{\sigma\tau}\hat{q}_{u,\sigma}^{} \hat{q}_{uv,\sigma}^{}\hat{q}_{vw,\tau}^{}\hat{q}_{w,\tau}^{}, \tag{21}\]
where we use the notation for the rotated polynomials
\[\hat{q}_{u,\sigma}^{} \equiv\sum_{i}U_{iu}^{}\hat{\gamma}_{i\sigma,0}\equiv\hat{U}_{ 1,u}^{\dagger}\hat{\gamma}_{1\sigma,0}\hat{U}_{1,u} \tag{22}\] \[\hat{q}_{uv,\sigma}^{} \equiv\sum_{j}U_{jv}^{(2,u)}\hat{\gamma}_{j\sigma,1}\equiv\hat{U} _{2,uv}^{\dagger}\hat{\gamma}_{1\sigma,1}\hat{U}_{2,uv}\] \[\hat{q}_{vw,\tau}^{} \equiv\sum_{k}U_{kw}^{(3,v)}\hat{\gamma}_{k\tau,0}\equiv\hat{U}_{ 3,vw}^{\dagger}\hat{\gamma}_{1\tau,0}\hat{U}_{3,vw}\] \[\hat{q}_{w,\tau}^{} \equiv\sum_{l}W_{lw}^{}\hat{\gamma}_{l\tau,1}\equiv\hat{U}_{4, w}^{\dagger}\hat{\gamma}_{1\tau,1}\hat{U}_{4,w} \tag{25}\]
Since in this case the SVD yields real unitary matrices \(U\) and \(W\), it follows that for all of these polynomials \(\hat{q}=\sum_{p}c_{p}\hat{\gamma}_{p}\), \(c_{p}\in\mathbb{R}\) and \(\sum_{p}|c_{p}|^{2}=1\), thus proving that these are real orbital rotations \(\in\mathrm{Spin}(N)\). The iterative SVD procedure for the \(g_{ijkl}\) tensor can be done efficiently on a classical computer, producing all necessary quantities for obtaining the LCU and specifying the corresponding unitaries. We will refer to this decomposition as the MPS decomposition. Note that the MPS decomposition can be done also by first factorizing the middle index with \(N^{2}\) components. This decomposition would be extremely similar to the double factorization (DF) factorization of the two-electron tensor, although with a different LCU representation of the Hamiltonian.
#### ii.2.2 CP4-based MTD-\(L^{4}\)
Alternatively the \(g_{ijkl}\) tensor can be decomposed as a linear combination of outer products of rank-1 vectors based on the CP4 tensor decomposition
\[g_{ijkl}=\sum_{m}\Omega_{m}v_{i}^{(1,m)}v_{j}^{(2,m)}v_{k}^{(3,m)}v_{l}^{(4,m)}, \tag{26}\]
The associated MPS representation would correspond to a linear combination of MPSs with bond dimension 1, and is shown in Fig.4. Note that the normalization can be done without loss of generality since all normalization constants can be absorbed in the \(\Omega_{m}\) coefficient. Obtaining the CP4 decomposition of the \(g_{ijkl}\) tensor also defines an MTD-\(L^{4}\) decomposition as shown in Eq., with the polynomials
\[\hat{q}_{m,\sigma}^{} \equiv\sum_{i}v_{i}^{(1,m)}\hat{\gamma}_{i\sigma,0}=\hat{U}_{m}^{\dagger}\hat{\gamma}_{1\sigma,0}\hat{U}_{m}^{} \tag{27}\] \[\hat{q}_{m,\sigma}^{} \equiv\sum_{j}v_{j}^{(2,m)}\hat{\gamma}_{j\sigma,1}=\hat{U}_{m}^{\dagger}\hat{\gamma}_{1\sigma,1}\hat{U}_{m}^{}\] \[\hat{q}_{m,\tau}^{} \equiv\sum_{k}v_{k}^{(3,m)}\hat{\gamma}_{k\tau,0}=\hat{U}_{m}^{(3 )\dagger}\hat{\gamma}_{1\tau,0}\hat{U}_{m}^{}\] \[\hat{q}_{m,\tau}^{} \equiv\sum_{l}v_{l}^{(4,m)}\hat{\gamma}_{l\tau,1}=\hat{U}_{m}^{ \dagger}\hat{\gamma}_{1\tau,1}\hat{U}_{m}^{}, \tag{30}\]
#### ii.2.3 SVD-based MTD-\(L^{4}\)
There is yet another alternative scheme of iterative SVD for performing MTD-\(L^{4}\).
\[g_{ijkl} =\sum_{\alpha_{1}=1}^{N}U_{i\alpha_{1}}^{}S_{\alpha_{1}}^{ }V_{jkl}^{(1,\alpha_{1})} \tag{31}\] \[=\sum_{\alpha_{1}=1}^{N}U_{i\alpha_{1}}^{}S_{\alpha_{1}}^{ }\sum_{\alpha_{2}=1}^{N}U_{j\alpha_{2}}^{(2,\alpha_{1})}S_{\alpha_{2}}^{(2, \alpha_{1})}V_{kl}^{(2,\alpha_{1}\alpha_{2})}\] \[=\sum_{\alpha_{1}=1}^{N}U_{i\alpha_{1}}^{}S_{\alpha_{1}}^{ }\sum_{\alpha_{2}=1}^{N}U_{j\alpha_{2}}^{}S_{\alpha_{2}}^{(2,\alpha_{1})}\] \[\quad\times\sum_{\alpha_{3}=1}^{N}U_{k\alpha_{3}}^{}S_{\alpha_ {3}}^{(3,\alpha_{1}\alpha_{2})}V_{l}^{(3,\alpha_{1}\alpha_{2}\alpha_{3})} \tag{33}\]
This decomposition gives rise to an MTD for which \(\vec{m}\cong(\alpha_{1},\alpha_{2},\alpha_{3})\) with \(N^{3}\) elements.
\[\hat{q}_{\alpha_{1},\sigma}^{} =\sum_{i}U_{i\alpha_{1}}^{}\hat{\gamma}_{i\sigma,0} \tag{34}\] \[\hat{q}_{\alpha_{1}\alpha_{2},\sigma}^{} =\sum_{j}U_{j\alpha_{2}}^{(2,\alpha_{1})}\hat{\gamma}_{j\sigma,1}\] \[\hat{q}_{\alpha_{1}\alpha_{2}\alpha_{3},\tau}^{} =\sum_{k}U_{k\alpha_{3}}^{(3,\alpha_{1}\alpha_{2})}\hat{\gamma}_ {k\tau,0}\] \[\hat{q}_{\alpha_{1}\alpha_{2}\alpha_{3},\tau}^{} =\sum_{l}V_{l}^{(3,\alpha_{1}\alpha_{2}\alpha_{3})}\hat{\gamma}_ {l\tau,1}\] \[\Omega_{\alpha_{1}\alpha_{2}\alpha_{3}} =S_{\alpha_{1}}^{}S_{\alpha_{2}}^{(2,\alpha_{1})}S_{\alpha_{3 }}^{(3,\alpha_{1}\alpha_{2})}. \tag{38}\]
Note that if we "flatten" the \(N\times N\times N\) vector \((\alpha_{1},\alpha_{2},\alpha_{3})\) into a one-dimensional array with \(N^{3}\) elements, we recover an MTD which is identical to the MTD-CP4, the only difference being how the tensor was decomposed. We refer to this family of decompositions as MTD-\(L^{4}\). However, the additional structure in some of these decompositions (e.g., MPS) make it amenable to more efficient LCU oracles, as shown in the circuits in the Appendix A.
MPS representation of CP4 factorization.
MPS diagram representation of iterative factorization of two-electron tensor with nested SVDs. Numbers in red show bond dimension for the matrices in the MPS.
## III Current LCUs as MTDs
Here we show how current LCUs can be written as MTDs, these mappings are summarized in Table 3.4.
### Qubit-based decompositions
#### iii.1.1 Pauli product decompositions
Also known as the sparse LCU, the simplest LCU decomposition can be obtained by mapping the electronic structure Hamiltonian [Eq.] into qubit operators through the use of some fermion-to-qubit mapping (e.g., Jordan-Wigner or Bravyi-Kitaev):
\[\hat{H}=\sum_{q}d_{q}\hat{P}_{q}, \tag{39}\]
To see the correspondence between electron integrals in the fermionic Hamiltonian Eq. and coefficients of Pauli products (\(d_{q}\)), it is convenient to write the fermionic Hamiltonian using the reflection operators \(\hat{Q}_{kl\sigma}\equiv i\hat{\gamma}_{k\sigma,0}\hat{\gamma}_{l\sigma,1}\):
\[\hat{H}=\hat{H}_{0}+\frac{1}{2}\sum_{ij}\tilde{h}_{ij}\sum_{\sigma}\hat{Q}_{ ij\sigma}+\frac{1}{4}\sum_{ijkl}g_{ijkl}\sum_{\sigma\tau}\hat{Q}_{ij\sigma} \hat{Q}_{kl\tau}, \tag{40}\]
This expression is an LCU corresponding to the sparse LCU with \(\tilde{h}_{ij}/2\) and \(g_{ijkl}/4\) equal to \(d_{q}\) in.
The associated MTD representation [Eq.] has \(L=1\), and can be obtained by a trivial \(\vec{m}=1\) index with a single element and no rotation, i.e. \(\hat{V}_{1}=\hat{1}\). There will be a single \(\vec{w}\to w\) coefficient, thus making the coefficients tensor \(\Omega_{w}\), with the corresponding \(\Omega_{w}/\hat{p}_{w}\)'s running over each of the coefficients/operators appearing in Eq..
A subtle point is that the two-electron component of \(\hat{H}_{c}\) can be expressed by separating \(\sigma=\tau\) and \(\sigma\neq\tau\) components
\[\hat{H}_{2} =\left(\frac{1}{2}\sum_{ij}g_{ijji}\right)\hat{1}\] \[+\frac{1}{4}\sum_{\sigma\neq\tau}\sum_{ijkl}g_{ijkl}\hat{Q}_{ij \sigma}\hat{Q}_{kl\tau}\] \[+\frac{1}{2}\sum_{\sigma}\sum_{i>k,l>j}(g_{ijkl}-g_{ilkj})\hat{Q} _{ij\sigma}\hat{Q}_{kl\sigma}. \tag{41}\]
This form of \(\hat{H}_{2}\) gives lower 1-norms than that in Eq., but it requires separating the spin components and as a consequence twice as many coefficients to be loaded by the PREPARE circuit, making it generally more expensive. Thus, we will always work with the Hamiltonian form shown in Eq., making the two-electron tensor representable using spacial orbitals. In the case where this spin-symmetry is broken, as is the case when obtaining a Hamiltonian from unrestricted Hartree-Fock calculations and generally for any open-shell system, only non-zero coefficients of the two-electron tensor in spin-orbitals need to be loaded. This can be done by using a contiguous register, as shown in Appendix A for the sparse LCU.
#### iii.1.2 Anti-commuting groupings
The 1-norm of the Pauli decomposition can be reduced by grouping Pauli products into larger unitaries.
\[\hat{H}=\sum_{n=1}^{N_{\rm AC}}a_{n}\hat{A}_{n}, \tag{42}\]
where \(N_{\rm AC}\) represents the total number of groups, and
\[a_{n} \equiv\sqrt{\sum_{q\in K_{n}}|d_{q}|^{2}}, \tag{43}\] \[\hat{A}_{n} \equiv\frac{1}{a_{n}}\sum_{q\in K_{n}}d_{q}\hat{P}_{q}, \tag{44}\]
Within the MTD framework [Eq.], this decomposition has \(L=1\), where \(\vec{m}\) takes on a single element such that \(\hat{V}_{1}=\hat{1}\). The \(\vec{w}\) index is identical to the \(n\) index in Eq., with \(\Omega_{n}=a_{n}\) and \(\hat{p}_{n}=\hat{A}_{n}\). Referring to Eq., we observe that the coefficients \(d_{q}\) associated with products of two Majoranas are purely imaginary, whereas those arising from four Majoranas are real. Considering that products of two Majoranas are anti-Hermitian and products of four Majoranas are Hermitian, it follows that \(\hat{A}_{n}^{\dagger}=\hat{A}_{n}\), indicating that \(\hat{A}_{n}\) operators are reflection operators.
\[\phi_{q}^{(n)}\equiv\frac{1}{2}\arcsin\frac{d_{q}}{\sqrt{\sum\limits_{\begin{subarray} {c}i\in K_{n}\\ i\leq q\end{subarray}}d_{i}^{2}}}. \tag{46}\]
Note that this decomposition is the only one presented here in which unitaries can combine both two-Majorana and four-Majorana terms. The application of controlled \(\hat{A}_{n}\) unitaries, as shown in Eq., would require \(2|K_{n}|\) controlled exponentials of Pauli operators, and as a result high T-gate cost. To bring this cost down, we introduce an alternative approach that reduces this number to only one controlled unitary. The key idea is that for each group \(K_{n}\), the anti-commutativity of the \(|K_{n}|\) Pauli products generates the Clifford algebra isomorphic to that of the same number of Majorana operators.
\[\theta_{1} =\frac{1}{2}\arccos c_{1},\] \[\theta_{2} =\frac{1}{2}\arccos\left(\frac{c_{2}}{\sin 2\theta_{1}}\right),\] \[\theta_{3} =\frac{1}{2}\arccos\left(\frac{c_{3}}{\prod_{j\leq 2}\sin 2 \theta_{j}}\right),\] \[\vdots\] \[\theta_{N-2} =\frac{1}{2}\arccos\left(\frac{c_{N-2}}{\prod_{j\leq(N-3)}\sin 2 \theta_{j}}\right),\] \[\theta_{N-1} =\frac{\text{sign}(c_{N})}{2}\arccos\left(\frac{c_{N-1}}{\prod_{j \leq N-2}\sin 2\theta_{j}}\right). \tag{48}\]
Given the identical algebraic structure, the same Givens-based diagonalization can be used to express any unitary consisting of mutually anti-commuting Paulis \(\hat{A}\equiv\sum_{j=1}^{J}a_{j}\hat{P}_{j}\):
\[\hat{A}=\left(\prod_{j=J-1}^{1}e^{\theta_{j}\hat{P}_{j}\hat{P}_{j+1}}\right) \hat{P}_{1}\left(\prod_{j=1}^{J-1}e^{-\theta_{j}\hat{P}_{j}\hat{P}_{j+1}} \right). \tag{49}\]
By using this expression, the controlled application of each \(\hat{A}_{n}\) only requires a single controlled operation, as shown in the SELECT circuit for this LCU in Appendix A.
#### iii.1.3 Orbital-optimized qubit operators
1-norm and the number of unitaries for qubit-based decompositions can be reduced further by working with an orbital basis that differs from the canonical molecular orbitals obtained from solving the Hartree-Fock equations. In this study, we explored two distinct orbital rotation schemes: the Foster-Boys orbitals and optimized orbitals obtained by directly minimizing the 1-norm of the LCU. In both cases, the Hamiltonian is transformed as \(\hat{U}^{\dagger}\hat{H}\hat{U}\), where \(\hat{U}\) represents the chosen orbital rotation. This amounts to choosing the Majorana-conserving rotation \(\hat{V}=\hat{U}\) instead to \(\hat{V}=1\) in the MTD expression [Eq.]. For the systems studied in Ref., the 1-norm scaling for the Foster-Boys orbitals closely approaches that of the optimal rotations, while requiring \(\mathcal{O}(N^{3})\) operations to find \(\hat{U}\), as opposed to the \(\mathcal{O}(N^{5})\) operations needed for finding the 1-norm minimizing rotation.
### Fermionic LCU decompositions
Next, we illustrate how fermionic-based LCU decompositions can be expressed as MTD. Unitaries in these LCUs are usually constructed as orbital rotations acting on reflections \(\hat{Q}_{ii\sigma}\equiv i\hat{\gamma}_{i\sigma,0}\hat{\gamma}_{i\sigma,1}=2 \hat{n}_{i\sigma}-\hat{1}\), where \(\hat{n}_{i\sigma}\equiv\hat{a}_{i\sigma}^{\dagger}\hat{a}_{i\sigma}\) represents the occupation number operator. The fermion-to-qubit mappings transform \(Q_{ii\sigma}\) operator to the qubit operator \(\hat{z}_{i\sigma}\).
Some methods, particularly those that factorize the \(g_{ijkl}\) tensor, separate the one-electron and two-electron operators, which amounts to having separate decompositions for the two-Majorana \(\left(\hat{H}_{1}\right)\) and four-Majorana \(\left(\hat{H}_{2}\right)\) parts in Eq., with the decomposition for \(\hat{H}_{1}\) being capped at \(L=2\) and polynomials \(\hat{p}\)'s with degree \(\leq 2\).
\[\hat{H}_{1}=\hat{U}^{\dagger}\left(\sum_{i}\mu_{i}\sum_{\sigma}\hat{Q}_{ii \sigma}\right)\hat{U}. \tag{50}\]
Note that the most efficient encodings for many fermionic-based decompositions, such as DF and tensor hypercontraction (THC), merge the one- and two-electron coefficients, using a quantum register to indicate whether a given term is coming from one- or two-electron operators. In the case of \(\hat{H}_{2}\), the coefficient invariance in the Hamiltonian with respect to spin enables us to decompose the two-electron tensor \(g_{ijkl}\) using the spacial-orbital indices, effectively working in the algebra defined by the \(\hat{F}_{j}^{i}\) operators. Apart from computational efficiency, operating within this spin-symmetric algebra facilitates compilation of the LCU unitaries while requiring less data to be loaded on the quantum computer, as demonstrated in Ref..
#### iii.1.1 Single factorization
Although originally deduced with fermionic operators, the single factorization (SF) technique can be written with Majoranas as:
\[\hat{H} =c_{\rm SF}\hat{1}+\frac{i}{2}\sum_{\sigma}\sum_{ij}\tilde{h}_{ij} \hat{\gamma}_{i\sigma,0}\hat{\gamma}_{j\sigma,1}\] \[\quad-\frac{1}{4}\sum_{\ell=1}^{N^{2}}\left(\sum_{\sigma}\sum_{ij }W_{ij}^{(\ell)}\hat{\gamma}_{i\sigma,0}\hat{\gamma}_{j\sigma,1}\right)^{2}, \tag{51}\]
We note that the \(\ell\) index is usually truncated at some \(M\ll N^{2}\). Equation illustrates how this decomposition can be cast as an MTD that separates \(\hat{H}_{1}\) and \(\hat{H}_{2}\), with \(L=2\) [Eq.]. The associated sum over \(\vec{m}\) has trivial rotations \(\hat{V}=\hat{1}\) and \(M\) elements. The Majorana polynomials \(\hat{p}\)'s correspond to the Majorana products in Eq.. Additionally, the complete-square structure of the two-electron elements allows them to be expressed as squares of \(\hat{p}_{\ell}\equiv\sum_{\sigma}\sum_{ij}W_{ij}^{(\ell)}\hat{\gamma}_{i\sigma, 0}\hat{\gamma}_{j\sigma,1}\) operators. Complete-squares can be encoded through the use of oblivious amplitude amplification, which implements through qubitization the second-degree Chebyshev polynomial of the normalized \(\hat{p}_{\ell}\), i.e. \(2\left(\hat{p}_{\ell}/N_{\ell}^{(\rm SF)}\right)^{2}-\hat{1}\), where the normalization constant \(N_{\ell}^{(\rm SF)}\equiv 2\sum_{ij}|W_{ij}^{(\ell)}|\). Finally, we note that the original procedure did not diagonalize the one-electron term.
\[\hat{H} =\tilde{c}_{\rm SF}\hat{1}+i\sum_{jk\sigma}\Lambda_{jk}\hat{ \gamma}_{j\sigma,0}\hat{\gamma}_{k\sigma,1}\] \[\quad+\sum_{\ell}\Omega_{\ell}\left(2\left(\frac{\hat{p}_{\ell} }{N_{\ell}^{(\rm SF)}}\right)^{2}-\hat{1}\right), \tag{52}\]
#### iii.1.2 Double factorization
The DF decomposition can be viewed as an extension of SF with the distinction that two-Majorana polynomials are first diagonalized to achieve an optimal 1-norm. In the DF approach, the one-electron term is diagonalized as shown in Eq.. This diagonalization is also used to implement the two-Majorana polynomials \(\hat{p}_{\ell}\). This approach has \(L=1\) for the MTD [Eq.].
\[\hat{p}_{\ell} \equiv i\sum_{ij}W_{ij}^{(\ell)}\sum_{\sigma}\hat{\gamma}_{i\sigma, 0}\hat{\gamma}_{j\sigma,1}\] \[=i\hat{U}_{\ell}^{\dagger}\left(\sum_{i}\mu_{i}^{(\ell)}\sum_{ \sigma}\hat{\gamma}_{i\sigma,0}\hat{\gamma}_{i\sigma,1}\right)\hat{U}_{\ell} \tag{53}\]
the DF LCU is written in the fermionic representation as
\[\hat{H} =c_{\rm DF}\hat{1}+\sum_{i}\mu_{i}^{}\sum_{\sigma}\hat{U}_{0} ^{\dagger}\hat{Q}_{ii\sigma}\hat{U}_{0}\] \[\quad+\sum_{\ell}\left(\sum_{i}\mu_{i}^{(\ell)}\sum_{\sigma}\hat{ U}_{\ell}^{\dagger}\hat{Q}_{ii\sigma}\hat{U}_{\ell}\right)^{2}, \tag{54}\]
The \(\mu^{(l)}\) vectors are obtained by diagonalizing the \(W^{(\ell)}\) matrices appearing in the SF approach. This can be recast in the MTD form:
\[\hat{H} =c_{\rm DF}\hat{1}+i\sum_{j}\mu_{j}^{}\sum_{\sigma}\hat{U}_{ \uparrow}^{\dagger}\hat{\gamma}_{j\sigma,0}\hat{\gamma}_{j\sigma,1}\hat{U}_{0}\] \[\quad+\sum_{\ell}\Omega_{\ell}\hat{U}_{\ell}^{\dagger}\left(2 \left(\frac{\hat{p}_{\ell}}{N_{\ell}^{(\rm DF)}}\right)^{2}-\hat{1}\right)\hat{ U}_{\ell}, \tag{55}\]
For the MTD, \(\vec{m}\) runs over \(M\leq N^{2}\) indices. However, \(\vec{m}\) is equal to polynomial subscript \(\ell\) (\(\hat{p}_{\ell}\)), which would correspond to an implicit Kronecker delta between \(\vec{m}\) and \(\vec{w}\) indices (here written as \(\ell\)) in the \(\Omega\) tensor, with \(\hat{V}_{\ell}=\hat{U}_{\ell}\in\mathrm{Spin}(N)\).
#### iii.1.3 Cartan sub-algebra
The Cartan sub-algebra (CSA) decomposition generalizes DF to include two-electron terms that are not parts of a complete-square. Thus, it corresponds to a full-rank factorization, as opposed to DF which is known as a low-rank factorization.
\[\hat{H} =c_{\mathrm{CSA}}\hat{1}+\sum_{i}\mu_{i}\sum_{\sigma}\hat{U}_{0}^ {\dagger}\hat{Q}_{i\sigma}\hat{U}_{0}\] \[\quad+\sum_{\ell}\sum_{ij}\lambda_{ij}^{(\ell)}\sum_{\sigma\tau} \hat{U}_{\ell}^{\dagger}\hat{Q}_{ii\sigma}\hat{Q}_{jj\tau}\hat{U}_{\ell}, \tag{56}\]
The corresponding MTD form is
\[\hat{H} =c_{\mathrm{CSA}}\hat{1}+i\sum_{j}\mu_{j}^{}\sum_{\sigma}\hat{ U}_{0}^{\dagger}\hat{\gamma}_{j\sigma,0}\hat{\gamma}_{j\sigma,1}\hat{U}_{0}\] \[\quad-\sum_{\ell,ij}\lambda_{ij}^{(\ell)}\sum_{\sigma\tau}\hat{ U}_{\ell}^{\dagger}\hat{\gamma}_{i\sigma,0}\hat{\gamma}_{i\sigma,1}\hat{\gamma}_ {j\tau,0}\hat{\gamma}_{j\tau,1}\hat{U}_{\ell}, \tag{57}\]
The associated \(\Omega\) tensor has an implicit Kronecker delta in the \(m_{1}\) and \(m_{2}\) indices. Despite being more flexible than DF, the non-linear optimization procedure required to obtain the Cartan sub-algebra decomposition becomes computationally expensive for large systems.
#### iii.1.4 Tensor Hypercontraction
The THC decomposition expresses the Hamiltonian as
\[\hat{H} =c_{\mathrm{THC}}\hat{1}+i\sum_{j}\mu_{j}\sum_{\sigma}\hat{U}_{0} ^{\dagger}\hat{Q}_{jj\sigma}\hat{U}_{0}\] \[\quad+\sum_{\mu,\nu=1}^{K}\zeta_{\mu\nu}\sum_{\sigma}\hat{U}_{\mu }^{\dagger}\hat{Q}_{11\sigma}\hat{U}_{\mu}\sum_{\tau}\hat{U}_{\nu}^{\dagger} \hat{Q}_{11\tau}\hat{U}_{\nu}, \tag{58}\]
Since the \(\hat{U}\)'s orbital rotations are only acting on orbital 1, their implementation requires \(\mathcal{C}(N)\) Givens rotations, as opposed to the \(\mathcal{C}(N^{2})\) required for rotating all orbitals.
\[\hat{U}=\prod_{i>j=1}^{N}e^{\theta_{ij}\hat{F}_{j}^{i}}, \tag{59}\]
Since \(\hat{Q}_{11\sigma}\) only has a component in orbital 1 we can simplify its transformation by leaving only \(\hat{F}_{j}^{i}\) generators where \(i\) or \(j\) is 1. To demonstrate validity of this simplification, we use a freedom to choose an order of exponential functions in the \(\hat{U}\) product. Such a choice can always be made because generators \(\hat{F}_{j}^{i}\) form a Lie algebra. The convenient order is to group all Givens rotations not acting on orbital 1 on the left-hand side of \(\hat{U}\),
\[\hat{U}^{\dagger}\hat{Q}_{11\sigma}\hat{U} = e^{\theta_{12}\hat{F}_{1}^{\perp}}e^{\theta_{13}\hat{F}_{3}^{1}}...e^{\theta_{1N}\hat{F}_{N}^{1}}\left(\prod_{i>j>1}e^{\theta_{ij}\hat{F}_{j}^ {i}}\right)^{\dagger}\hat{Q}_{11\sigma} \tag{60}\] \[\times \left(\prod_{i>j>1}e^{\theta_{ij}\hat{F}_{j}^{i}}\right)e^{ \theta_{1N}\hat{F}_{N}^{1}}...e^{\theta_{12}\hat{F}_{2}^{\perp}}\] \[= e^{\theta_{12}\hat{F}_{2}^{\perp}}e^{\theta_{13}\hat{F}_{3}^{1}}... e^{\theta_{1N}\hat{F}_{N}^{1}}\hat{Q}_{11\sigma}\] \[\times e^{\theta_{1N}\hat{F}_{N}^{\perp}}...e^{\theta_{12}\hat{F}_{2}^{ \perp}},\]
Evidently, \(N-1\) Givens rotations are enough for arbitrary orbital rotation of \(\hat{Q}_{11\sigma}\).
The \(\alpha\)-dependence of the \(\alpha\)-
Results and discussion
The most time-consuming part of the fault-tolerant QPE algorithm is performing T-gates and their error correction. Considering that the number of applications of the Hamiltonian LCU oracle is proportional to the 1-norm of the used LCU, as a quantum resource cost we use a product of the LCU 1-norm by the number of T-gates required for an implementation of the Hamiltonian oracle. The latter corresponds to the number of T-gates in one SELECT and two PREPARE operations. We will refer to the quantum resource cost simply as hardness. Additionally, we note that optimizing the compilation strategy for the SELECT circuit is imperative for minimizing the hardness. Compilation of these circuits typically exploit structures present in the used LCU _ansatz_, as seen in Ref. for DF and THC approaches which make use of the spin-symmetry that is present (i.e. working with \(\hat{F}_{j}^{i}\) operators). Whether better circuits can be found for the decompositions that have been introduced in this work is an open question.
All our fermionic LCU decompositions are performed on the two-body part of the Hamiltonian, which we define as the part of the Hamiltonian quartic in Majorana operators. Since we know that the remaining one-body part of the Hamiltonian can be optimally treated using DF, all 1-norms corresponding to fermionic LCU decompositions reported in this paper are, therefore, sum of the spectral range of the one-body part and the 1-norm for the fermionic LCU decomposition of the two-body part. We have obtained the THC decompositions using a non-linear optimization performed with the Levenberg-Marquardt algorithm, details of our THC implementation are in Appendix B. The Toffoli gate counts for our THC decompositions are obtained using Openfermion and the reported T-gate counts are 4 times the Toffoli counts.
To assess efficiency of various LCU decompositions we present quantum resource estimates for small molecules (Table 2) and 1D hydrogen chains (and Table 3), computational details are in Appendix B. Most of LCU methods can be partitioned into four categories based on the small molecule results: 1) low T-gate count and medium 1-norms, Pauli and Pauli-OO; 2) medium 1-norms and medium T-gate count, \(L^{4}\)-SVD and THC; 3) high T-gate count and low 1-norms, AC, OO-AC, and DF; 4) medium T-gate count and high 1-norms, \(L^{4}\)-CP4 and \(L^{4}\)-MPS.
Overall, the orbital-optimized Pauli LCU has the lowest hardness in virtue of the small number of T-gates required for its implementation. The PREPARE circuit for this decomposition can be optimized to exploit the 8-fold symmetry of the two-electron tensor as shown in Ref., greatly reducing the number of coefficients that need to be loaded. This, along with the simple structure of its SELECT circuit make it extremely efficient. For the orbital-optimized schemes in small molecules (Table 2) we used the full orbital optimization which minimizes the total 1-norm since the Foster-Boys localization scheme did not improve the results. In contrast, for 1D hydrogen chains, the orbital rotations that minimize the 1-norm yield results similar to those obtained using the Foster-Boys orbital localization scheme, which is in line with findings of Ref.. Therefore, for computational efficiency, we used the Foster-Boys localization scheme for the LCUs of hydrogen chains. It is worth noting that fermionic-based decompositions, which inherently include orbital rotations as part of the LCU, are unaffected by the orbital basis used to represent the electronic tensor.
Regarding the scaling of hardness with the system size (Table 3), the only decompositions which presented a better scaling than Pauli-OO are \(L^{4}\)-CP4 and THC. The THC decomposition offers the most optimal scaling with increasing system size \(N\), hardness scaling of \(O(N^{2.19})\) is obtained empirically from results of Table 3. Thus, THC can be advantageous over Pauli-OO for systems with \(N>23\). We note that extremely large 1-norms were obtained for the \(L^{4}\)-CP4 tensor decomposition for the small molecules (Table 2). We attribute this to an implementation that was used for the CP4 decomposition, and having a more robust implementation will be necessary for larger systems, as discussed in Ref.. At the same time, \(L^{4}\)-SVD decomposition gave competitive 1-norms for these molecules. On the other hand, the SVD-based 1-norms for the hydrogen chains were larger than their CP4 counterparts, which makes the SVD decomposition prohibitively expensive for larger chains. Overall, these results establish the \(L^{4}\) LCU decomposition as a promising decomposition in terms of hardness; finding the corresponding fragments in a robust and efficient way is work in progress. We believe that this decomposition could be useful for hybrid LCU approaches by obtaining a few large MTD-\(L^{4}\) fragments before decomposing the rest of the Hamiltonian with some other method. However, exploring these kind of hybrid LCUs is beyond the scope of this work.
The anti-commuting (AC) decomposition, when using the orbital optimized Hamiltonians, yielded very low 1-norms and requires the fewest qubits. However, the large number of rotations required for its SELECT circuit increases the T-gate count and makes its hardness one of the worst out of all the studied LCUs. As such, we believe that an implementation of the AC decomposition which loads the Givens rotation angles in the quantum computer could lower significantly the cost compared to the current implementation that we proposed where angles are "hard coded" into the circuits.
The results for the DF method correspond to the implementation discussed in the Appendix A. The "optimized DF" results in and Table 3 correspond to the DF implementation discussed in Ref. where each Toffoli gate has been considered to require 4 T-gates. We note that this decomposition also presents one of the best hardness scalings. However, the number of qubits required for the optimized DF is an order of magnitude larger than that required by most other approaches. Only the THC decomposition involves similar numbers of ancilla qubits as the Optimized DF approach.
_The classical pre-processing cost:_ The Pauli products decomposition of a Hamiltonian has \(\mathcal{E}(N^{4})\) operators. However, orbital localization schemes can reduce the scaling to \(\mathcal{E}(N^{2})\) for large molecules; the cost for optimized orbital localization schemes typically scales as \(\mathcal{E}(N^{3})\). The sorted insertion algorithm, used in AC, has a sorting cost of \(\mathcal{E}(M^{2})\), where \(M\) is the number of elements in the list being sorted. This results in a decomposition cost of \(\mathcal{E}(N^{8})\) for AC and \(\mathcal{E}(N^{4})\) when grouping is done on localized orbitals. For the fermionic methods, the cost for DF is dominated by the diagonalization of the reshaped \(N^{2}\times N^{2}\) tensor \(g_{ijkl}\), which scales as \(6(N^{6})\).
\begin{table}
\begin{tabular}{|c|c|c|c|c|c|} \hline Molecule & H\({}_{2}\) & LiH & BeH\({}_{2}\) & H\({}_{2}\)O & NH\({}_{3}\) \\ \hline \(\Delta E/2\) & 0.815 & 4.93 & 9.99 & 41.9 & 33.8 \\ \hline Pauli & 1.58 & 13.0 & 22.8 & 71.9 & 69.2 \\ \hline OO-Pauli & 1.58 & 12.4 & 21.9 & 61.0 & 54.5 \\ \hline AC & 1.49 & 10.2 & 18.0 & 57.2 & 48.9 \\ \hline OO-AC & 1.49 & 10.2 & 17.9 & 55.7 & 46.8 \\ \hline DF & 1.37 & 9.34 & 16.4 & 53.7 & 44.7 \\ \hline THC & 1.8 & 11.18 & 19.59 & 58.55 & 49.26 \\ \hline \(L^{4}\)-SVD & 2.19 & 11.8 & 22.0 & 61.1 & 53.3 \\ \hline \(L^{4}\)-CP4 & 5.65 & 34.5 & 61.1 & 125.0 & 121.0 \\ \hline \(L^{4}\)-MPS & 4.63 & 130.08 & 279.27 & 474.1 & 612.45 \\ \hline \end{tabular}
\end{table}
Table 2: 1-norms for electronic structure Hamiltonians in the STO-3G basis, the numbers of T-gates required to implement the associated Hamiltonian oracle circuit are given in parenthesis. \(\Delta E/2\) corresponds to a lower bound for the 1-norm. Pauli: Pauli products as unitaries; AC: anti-commuting grouping technique; OO-prefix indicates that orbitals were optimized to minimize the 1-norm; DF: double factorization; THC: tensor hypercontraction; \(L^{4}\)-SVD, \(L^{4}\)-CP4, and \(L^{4}\)-MPS: MTD-\(L^{4}\) decompositions introduced in Sec. II.1.1.
\begin{table}
\begin{tabular}{|c|c|c|c|} \hline LCU method & Fitted quantity & \(\alpha\) & \(\beta\) \\ \hline \multirow{2}{*}{Pauli} & \(\lambda\times\#\)T-gates & 0.01 & 5.59 \\ & \# qubits & 1.49 & 0.44 \\ \hline \multirow{2}{*}{AC} & \(\lambda\times\#\)T-gates & 1.85 & 5.53 \\ & \# qubits & 1.35 & 0.45 \\ \hline \multirow{2}{*}{OO-Pauli} & \(\lambda\times\#\)T-gates & 2.36 & **2.67** \\ & \# qubits & 1.54 & 0.41 \\ \hline \multirow{2}{*}{OO-AC} & \(\lambda\times\#\)T-gates & 4.04 & 2.82 \\ & \# qubits & 1.39 & 0.41 \\ \hline \multirow{2}{*}{\(L^{4}\)-SVD} & \(\lambda\times\#\)T-gates & 2.10 & 3.83 \\ & \# qubits & 1.86 & 0.23 \\ \hline \multirow{2}{*}{\(L^{4}\)-CP4} & \(\lambda\times\#\)T-gates & 3.17 & **2.36** \\ & \# qubits & 1.85 & 0.23 \\ \hline \multirow{2}{*}{\(L^{4}\)-MPS} & \(\lambda\times\#\)T-gates & 2.55 & 4.20 \\ & \# qubits & 2.12 & 0.14 \\ \hline \multirow{2}{*}{DF} & \(\lambda\times\#\)T-gates & 1.80 & 4.14 \\ & \# qubits & 1.91 & 0.24 \\ \hline \multirow{2}{*}{Optimized DF} & \(\lambda\times\#\)T-gates & 2.58 & 2.70 \\ & \# qubits & 0.66 & 1.75 \\ \hline \multirow{2}{*}{THC} & \(\lambda\times\#\)T-gates & 3.02 & **2.19** \\ & \# qubits & 1.62 & 0.91 \\ \hline \end{tabular}
\end{table}
Table 3: Linear fit coefficients for Fig. 5, \(\log_{10}y=\alpha+\beta\log_{10}x\).
For our THC implementation, the computation of an approximate Hessian matrix for the update step in the
Hardness (top) and the number of qubits (bottom) scaling for different LCUs for hydrogen chains with \(N\) atoms with a spacing \(r=1.4\) Å. Hardness is defined as the product of the 1-norm \(\lambda\) with the number of T-gates for one SELECT and two PREPARE oracles. Lines correspond to a least-square fit, \(\log_{10}y=\alpha+\beta\log_{10}x\). Coefficients \(\alpha\) and \(\beta\) are reported in Table III. Orbital optimizations were done using the Foster-Boys localization.
Levenberg-Marquardt algorithm dominates the computational cost. Assuming that the rank of the THC decomposition, \(K\) [Eq.] scales linearly with system size, the computational cost of our THC implementation scales as \(O(N^{8})\). This is in contrast to the \(O(N^{6})\) scaling of Openfermion's L-BFGS-B THC implementation. Our implementation of THC offers greater robustness with the choice of initial parameters in exchange for the higher computational scaling. For \(L^{4}\)-MPS, the cost for each step in the MPS factorization (see Fig. 3) goes as \(\mathcal{E}(N^{5})\), thus being the overall scaling cost of this decomposition. The \(L^{4}\)-SVD decomposition quickly becomes prohibitively expensive to do as the system size grows, while the \(L^{4}\)-CP4 decomposition requires \(\mathcal{E}(N^{3}W)\), where \(W\) is the number of terms in the decomposition. Among the algorithms presented in this work, orbital-localized AC has the best theoretical scaling, followed by \(L^{4}\)-MPS. However, in practice, the fermionic-based decompositions (DF and \(L^{4}\)-MPS) tend to have lower prefactors due to their efficient utilization of linear algebra routines. For example, in the case of the linear H\({}_{30}\) chain, orbital-localized AC took approximately 5 minutes, while DF required around 0.1 seconds and \(L^{4}\)-MPS required around 0.3 seconds. These calculations were done using a single core of a 2.2 GHz 6-Core Intel i7 processor.
## V Conclusion
In summary, this work has introduced the MTD framework, which provides a unified approach to LCU decompositions of the electronic structure Hamiltonian. This framework not only presents a unifying perspective but also leads to new decompositions like MTD \(L^{4}\), the latter expresses the electronic structure tensor as a minimal-rank representation. MTD establishes a connection between popular tensor decompositions and LCUs. A study over small systems revealed the LCU based on Pauli products with an orbital optimization to be the most competitive. However, the MTD \(L^{4}\)-CP4 and THC decompositions had better scalings with the system size, suggesting that for larger systems these decompositions could be more efficient to implement.
|
10.48550/arXiv.2407.06571
|
Majorana Tensor Decomposition: A unifying framework for decompositions of fermionic Hamiltonians to Linear Combination of Unitaries
|
Ignacio Loaiza, Aritra Sankar Brahmachari, Artur F. Izmaylov
| 2,402
|
10.48550_arXiv.2208.01257
|
### Hard Sphere Mapping: General Framework
An analytical expression of \(D_{0}\) for the given CG system is obtained through three steps. illustrates a schematic diagram of the hard sphere mapping procedure.
We first map the CG system to the corresponding hard sphere system. Since dimensionless static or dynamic properties of literal hard spheres are completely described by their packing fraction \(\eta\), the first step is to determine values of \(\eta\) that can reproduce the important properties of the reference CG system. Conventionally, one assumes the number density is known, and this involves determining the EHSD, \(\sigma\). Then, \(\eta\) is readily obtained from the EHSD by the following relationship:
\[\eta=\frac{\pi}{6}\sigma^{3}\rho=\frac{\pi}{6}\sigma^{3}\,\frac{N_{CG}}{V}\,, \tag{8}\]
However, determining \(\sigma\) may not be the only way to construct a dynamically consistent effective hard sphere fluid. In Section III, we introduce a simpler approach that allows for directly mapping \(\eta\).
Once the hard sphere packing fraction is determined from a mapping, the diffusion coefficient of the specific hard sphere system can be straightforwardly calculated using the hard sphere kinetic theory. Since we are dealing with the mapped hard sphere system, we first revisit the Dzugutov scaling scheme. Unlike generally complex FG systems, the Dzugutov scaling is valid only for the hard sphere systems. Thus, our ultimate goal is to analytically derive the \(D_{0}\) term for the original Rosenfeld scaling scheme by introducing what we perceive as a dynamically consistent hard sphere reference as a proxy, where its fundamental quantities can be analytically derived using the Dzugutov scaling.
Schematic diagram describing the overall procedure of this work: Map a given CG system to an effective hard sphere system. This step gives an EHSD or packing fraction of the CG model. From the EHSD or packing fraction, derive an analytical expression for the diffusion coefficient of the hard sphere system. Since the hard sphere diffusion process can be correctly described by the Dzugutov scaling if the packing fraction is not too high, this step is done on the basis of the Dzugutov relationship. From the hard sphere diffusion coefficient expression, infer the D\({}_{0}\) term in the Rosenfeld scaling of the original CG system.
1) in order to analytically derive the \(D_{0}\) expression in the hard sphere systems (Step 2 in Fig. 1) and transfer \(D_{0}\) back to the original CG system (Step 3 in Fig. 1).
To calculate the static and dynamic properties of a hard sphere fluid, one must carefully choose the proper EOS.
\[\mathbb{Z}(\eta)=1+2^{2}\eta+b_{3}\eta^{2}+b_{4}\eta^{3}+\cdots=1+4\eta+\sum_{ k^{\prime}=3}^{\infty}b_{k\prime}\eta^{k\prime-1}, \tag{11}\]
The particular choice of the analytic form for the EOS is vital for applying the perturbation theory and calculating various thermodynamic quantities of interest, even for systems exhibiting non-negligible attractive interactions. While classical liquid state theory asserts that such a hard sphere description cannot be directly applied at the microscopic level to associated liquids characterized by strong specific attractions (e.g., water), the goal of this paper is to demonstrate how this minimalist description can still work rather well for molecular CG models based on the mapping formulated for the problem of supercooled molecular liquid relaxation. A detailed discussion of this practical extension will be given in Section II-E.
## C.
Classical perturbation theories of atomic or molecular (at the interaction site level) liquids treat pair interactions by separating the short-range repulsions as a hard sphere reference and the long-range interactions as a perturbation (often attractions), assuming that the structure (and sometimes the dynamics) of the liquid is determined by repulsive interactions. With this in mind, for a system with soft-core repulsive interactions as relevant to our CG model, an accurate estimation of the EHSD is needed to calculate the static and dynamic properties of the corresponding hard sphere model. In this light, various free energy perturbation theories have been developed based on different criteria to determine the EHSD. To name a few, BH and WCA theories are two frequently used approaches.
Barker and Henderson proposed the first perturbation theory for simple Lennard-Jones systems. In brief, the short-range repulsive potential [positive part of the overall potential \(U(R)\)] was mapped to a hard sphere reference, and attractive long-range terms were treated as a perturbation. \({}^{65}\) At leading order, the simplest BH EHSD was then defined as
\[\sigma_{BH}=\int_{0}^{R_{0}}[1-\exp(-\beta U(R))]\cdot dR, \tag{12}\]
In this work, we use the many-body CG PMF from our prior work to determine the BH EHSD, which is different from the original work, where it corresponds to the Lennard-Jones interaction \(U_{Lf}(R)\).
### **D. Weeks-Chandler-Andersen Criterion and its Drawbacks**
The well-known WCA perturbation theory starts from an alternative, physically motivated separation method based on the sign of forces (positive or negative) rather than interaction energies. In doing so, the WCA perturbation theory provides a smoothly varying first-order perturbative treatment of free energies. The WCA EHSD criterion is derived by approximating the cavity correlation function of the reference system \(y_{\text{HS}}(R)\coloneqq g(R)\exp(\beta U(R))\) as that of the hard sphere system, i.e., equating the long wavelength responses in Fourier space between the hard sphere and the purely repulsive force from reference system.
\[S(k)=\frac{1}{N}\langle\rho(\vec{k})\rho(-\vec{k})\rangle, \tag{13}\]
Alternatively, using the radial distribution function (RDF), Eq.
\[S(k)=1+4\pi\rho\int_{0}^{\infty}dR\ R^{2}\frac{\sin\ (kR)}{kR}[g(R)-1]. \tag{14}\]
In the \(k=0\) limit, the structure factor reduces to
\[S(k=0)=1+4\pi\rho\int_{0}^{\infty}\!\!dr\ R^{2}[g(R)-1]. \tag{15}\]
Therefore, based on the WCA separation of the pair potential, the EHSD of the effective hard sphere fluid that best captures the behavior of the purely repulsive fluid is determined by matching the \(S(k=0)\) (dimensionless compressibility) quantities:
\[\int_{0}^{\infty}\!\!y_{HS}(R;\sigma_{WCA})\{\exp[-\beta U_{\text{WCA}}(R)]- \exp[-\beta U_{\text{HS}}(R;\sigma_{\text{WCA}})]\}R^{2}\cdot dR=0. \tag{16}\]
In practice, this procedure is conducted iteratively until \(\sigma_{WCA}\) converges. It has been shown that the WCA theory is superior to the BH description for atomic liquids, especially at the high densities of present interest.
However, the WCA theory suffers from three drawbacks pertinent to this study. First, a determination of \(\sigma_{WCA}\) requires solving Eq. iteratively, making it less efficient. Also, in our cases, for both the BH and WCA theories, the EHSD is estimated by many-dimensional CG PMFs projected on pairwise basis sets. Furthermore, it is worthwhile noting that the original BH and WCA approaches were designed for systems with harsh repulsive interactions, and extending these approaches to systems with soft repulsions, in which attractions vary rapidly in space and can change structure, is of unknown reliability. Given the complex interaction profile of CG PMFs of molecular liquids, we envisage that the EHSD scheme based on \(U(R)\) might be less accurate when applied to CG systems. This suggests that such a method should focus on determining the effective packing fraction for which the density and hard sphere diameter are not separable. Lastly, both the BH and WCA criteria for the EHSD are derived from the perspective of describing equilibrium thermodynamic properties (e.g., free energy), and thus it is not _a priori_ clear how good they are to capture CG dynamics. In this regard, a mapping method that matches a quantity believed to be strongly correlated with dynamics is desirable. We note that the dynamical reconstruction approach also employs a reference hard sphere model to understand the CG diffusion of polymers. However, in Refs., the EHSD is determined via the long-time limit of Rouse dynamics of polymer melts, and hence the underlying physical implementation of the hard sphere description are different in both cases.
### Alternative Hard Sphere Mapping: Fluctuation Matching
To reconcile the aforementioned issues, we employ a hard sphere mapping method that has been highly developed and applied by Schweizer and coworkers. It is dynamically motivated and has been shown to be useful for mapping real molecular and polymer liquids with both repulsive and attractive interactions to an effective hard sphere description for use in a dynamical theory of activated relaxation. The central idea is to require the dimensionless compressibility, \(S(k=0)\), of Eq. be exactly equal to its value for the real thermal liquid at all temperatures and pressures. Note that the right-hand side of Eq.
\[S(k=0)=\rho k_{B}T\kappa_{T}. \tag{17}\]
This quantity also enters WCA theory, but there it is employed to only match the dimensionless amplitude of long wavelength density fluctuations of the purely repulsive reference fluid and its effective hard sphere analog for atomic liquids. The criterion of matching of \(S(k=0)\) of attractive thermal liquids and a reference hard sphere fluid yields an effective hard sphere packing fraction that is chemistry and thermodynamic state dependent, and there is no need to explicitly know the effective hard sphere diameter. Under isobaric conditions, attractions can strongly modify density as the liquid is cooled, and this important effect enters the mapping procedure. The idea to use such a \(S(k=0)\) mapping approach in a dynamical context was motivated by the physical argument that collective density fluctuations is the key slow variable that quantifies caging effects in dense liquids. When this mapping is combined with the activated dynamics theories of Refs., an understanding of the dependence of the structural relaxation time of chemically complex molecular and polymer liquids over many orders of magnitudes, extending from the weakly to strongly supercooled regimes down to the kinetic glass transition, has been achieved. We believe this provides a firm foundation for our present work in the normal liquid regime where the dynamical problem is, in fact, simpler than in the supercooled regime since caging effects are far weaker.
Here, we extend the above ideas to CG models of molecular liquids in the normal state regime. _A priori_, the accuracy of this approach is unknown given the inter-particle potentials are softer and can have strong attractions, especially for water. The use of this mapping assumes these features do not change the usefulness of the basic idea that the effect of dynamic correlations and caging remains strongly correlated with collective density fluctuation dynamics. For hard spheres, since the isothermal compressibility is given by the derivative of the EOS from the compressibility route, one only needs the effective packing fraction, \(\eta\), i.e., \(S(k\to 0)_{\text{HS}}=f(\eta)\).
In the present context, we call this approach "fluctuation matching", which is defined mathematically as equating the dimensionless long wavelength density fluctuation amplitude of the CG and hard sphere systems:
\[S(k=0)_{\text{CG}}=S(k=0)_{\text{HS}}. \tag{18}\]
We emphasize this mapping via \(S(k=0)\) directly focuses on \(\eta\), and hence avoids the issue of determining the effective hard sphere diameter with \(\eta\) carrying the chemistry, temperature, and pressure dependences. This distinction enhances the power of a hard sphere mapping in terms of transferability, and employing only \(\eta\) rather than separating it into EHSD is consistent with Rosenfeld's original work. In practice, for a spherical particle system (i.e., a single site CG model), the right-hand side of Eq. reduces to a direct connection between the dimensionless compressibility and the RDF of any spherical particle fluid:
\[f(\eta)=1+4\pi\rho\int_{0}^{\infty}\!dr\,R^{2}[g(R)-1]. \tag{19}\]
Of course, _a priori_, one might expect that this proposed approach will face a significant challenge in the application to molecular CG systems that exhibit liquid dynamics influenced by strong- and short-range attractive interactions (associated liquids). While recent work in this direction has provided a deeper level of understanding of the connections between \(S(k=0)\) and local structure, being able to match dimensionless compressibilities is not guaranteed to be an accurate strategy in a dynamical context for associated normal state liquids. Moreover, the adoption of a hard sphere reference system and the above mapping effectively assumes that although attractions do modify the magnitude, temperature and pressure dependence of the effective packing fraction, and hence structure, for dynamics it is assumed repulsive forces dominate. References 72-77, where the analog of \(S(k=0)_{\text{CG}}\) is the experimental dimensionless compressibility under isobaric conditions, have shown success for supercooled (largely non-associated) molecular and polymer liquids and provides support for this simplification. Concerning practical implementation, as one sees from Eq. the key quantity \(f(\eta)\) is dependent on \(\kappa_{T}\) which is determined from the EOS. In later sections, we will determine the full form of Eq. using several hard sphere fluid EOSs along with the FG simulations. This will be a main focus of the rest of this paper.
## F.
## 1.
The previous section demonstrated that CG systems can be mapped onto effective hard sphere fluids by utilizing either the BH or fluctuation matching. With this in mind, we now derive an analytical expression of \(D_{0}^{\text{CG}}\) for the Rosenfeld scaling in terms of effective hard sphere packing fraction.
Our starting point is the alternative Dzugutov scaling relationship that accurately holds for the hard sphere fluid as well as for the Rosenfeld scaling: \(D_{Z}^{*}=D\cdot(\sigma^{-2}\Gamma_{E}^{-1})=D_{Z}^{0}\exp(s_{ex}^{})\). For the hard sphere system, we determine \(s_{ex}^{}\) from the excess entropy of the hard sphere fluid, \(s_{ex}^{HS}\), that can be computed analytically in many cases.
\[D=\sigma^{2}\Gamma_{E}D_{Z}^{0}\exp(s_{ex}^{\text{HS}}). \tag{20}\]
An advantage of the hard sphere system is that the hard sphere diffusion coefficient in Eq. can be alternatively expressed based on Enskog kinetic theory. By considering the positional correlations in a fluid of spheres, Enskog theory has been shown to provide a reasonably good description of dynamics up to moderate liquid-like densities. We adopt the Enskog expression for the hard sphere diffusion coefficient:
\[D=D^{HS}=\frac{3}{8\pi}\sqrt{\frac{\pi k_{B}T}{m}}\frac{1}{\rho \sigma^{2}g(\sigma)}.\]
Equating Eq. with Eq.
\[D_{Z}^{0}=\frac{3}{32\pi}\frac{\exp\left(-S_{ex}^{HS}\right)}{\rho^{2}\sigma^{ 6}g^{2}(\sigma)}.\]
An explicit form of Eq.
\[g(\sigma)=\frac{\mathbb{Z}-1}{4\eta}, \tag{23}\]
Furthermore, the excess entropy of the hard sphere system can be obtained from
\[s_{ex}^{HS}=-\int_{0}^{\eta}\frac{\mathbb{Z}-1}{\eta^{\prime}}d\eta^{\prime}.\]
Since \(\mathbb{Z}\) is a function of the packing fraction \(\eta\), \(D_{Z}^{0}\) is entirely determined from \(\eta\) as follows:\[D_{Z}^{0}=\frac{\pi}{24}\cdot\frac{1}{(\mathbb{Z}-1)^{2}}\exp\left(\int_{0}^{1} \frac{\mathbb{Z}-1}{\eta^{\prime}}d\eta^{\prime}\right). \tag{25}\]
## 2.
We now derive an analytic expression for \(D_{0}\) using the analytically solvable Percus-Yevick (PY) integral equation theory.\({}^{96}\) However, PY theory is thermodynamic inconsistent,\({}^{97}\) meaning that thermodynamic quantities derived from identical RDFs by different formally exact routes are not identical.
\[\mathbb{Z}_{\text{PY}}^{c}=\frac{(1+\eta+\eta^{2})}{(1-\eta)^{3}}, \tag{26}\]
which agrees with the Scaled Particle Theory.\({}^{98-100}\) However, an alternative \(\mathbb{Z}\) expression derived by Wertheim and Thiele from the virial route is\({}^{101-103}\)
\[\mathbb{Z}_{\text{PY}}^{v}=\frac{(1+2\eta+3\eta^{2})}{(1-\eta)^{2}}. \tag{27}\]
In this work, we chose the compressibility route \(\mathbb{Z}_{\text{PY}}^{c}\), since overall \(\mathbb{Z}_{\text{PY}}^{c}\) provides better description of virial coefficients\({}^{104}\) and compressibility factors\({}^{105}\) than \(\mathbb{Z}_{\text{PY}}^{v}\), and because the fluctuation matching mapping strategy focuses on the dimensionless compressibility. In the Appendix we derive the relevant static and dynamic quantities using the virial route.
Using the EOS as Eq., the excess entropy predicted from the PY EOS is:
\[s_{ex}^{HS}=-\int_{0}^{\eta}\frac{\eta^{\prime 2}-2\eta+4}{(1-\eta^{ \prime})^{3}}d\eta^{\prime}=\ln(1-\eta)-\frac{3}{2}\bigg{[}\frac{1}{(1-\eta)^{ 2}}-1\bigg{]}, \tag{28}\]
and the pair correlation function at contact \(g(\sigma)\) is
\[g(\sigma)=\frac{\eta^{2}-2\eta+4}{4(1-\eta)^{3}}. \tag{29}\]
## 3. Final Expression for \(D_{0}\)
Combining Eqs. and, we arrive at the \(D_{Z}^{0}\) expression based on the PY EOS (compressibility route), which we denote as \(D_{Z,\text{PY}}^{0}\):
\[D_{Z,\text{PY}}^{0}=\frac{\pi}{24}\cdot\frac{(1-\eta)^{5}}{\eta^{2}(\eta^{2}- 2\eta+4)^{2}}\exp\bigg{[}\frac{3(2\eta-\eta^{2})}{2(1-\eta)^{2}}\bigg{]}. \tag{30}\]
Therefore, the diffusion coefficient of the hard sphere using the Dzugutov scaling \(D=\sigma^{2}\Gamma_{E}D_{Z}^{0}\exp(S_{ex})\) is \[D=4\pi\sigma^{4}\rho\,g(\sigma)\left(\frac{k_{B}T}{\pi m}\right)^{\frac{1}{2}} \times\frac{\pi}{24}\cdot\frac{(1-\eta)^{5}}{\eta^{2}(\eta^{2}-2\eta+4)^{2}} \exp\left[\frac{3(2\eta-\eta^{2})}{2(1-\eta)^{2}}\right]\exp(s_{ex}). \tag{31}\]
Now, writing \(\frac{\pi}{24}\sigma^{4}\rho=\frac{\sigma}{4}\cdot\left(\frac{\pi}{6}\sigma^{3 }\rho\right)=\frac{\sigma}{4}\cdot\eta\), we obtain
\[D=\sigma^{2}\Gamma_{E}D_{Z}^{0}\exp(s_{ex})=\frac{\sigma}{4}\bigg{(}\frac{\pi k _{B}T}{m}\bigg{)}^{\frac{1}{2}}\times\frac{(1-\eta)^{2}}{\eta(\eta^{2}-2\eta+4 )}\exp\left[\frac{3(2\eta-\eta^{2})}{2(1-\eta)^{2}}\right]\exp(s_{ex}). \tag{32}\]
We now seek to transfer \(D\) (Step 3 in Fig. 1) back to the \(D_{0}\) of the CG system via the Rosenfeld scaling. From the "macroscopic" rescaling in Eq., we arrive at
\[D^{*}=D\cdot\frac{\rho^{\frac{1}{3}}}{\left(\frac{k_{B}T}{m}\right)^{\frac{1}{ 2}}}=\frac{6^{\frac{1}{3}}}{4}\pi^{\frac{1}{6}}\eta^{\frac{1}{3}}\frac{(1-\eta )^{2}}{\eta(\eta^{2}-2\eta+4)}\exp\left[\frac{3(2\eta-\eta^{2})}{2(1-\eta)^{2} }\right]\exp(s_{ex}). \tag{33}\]
An empirical aspect of the Rosenfeld scaling originates from an exponent \(\alpha\), which hinders the derivation of a general and analytic expression. However, since the excess entropy of the hard sphere fluid should be always larger (more positive) than the molecular CG system due to missing degrees of freedom, the origin of \(\alpha\) arises from \(\exp(\alpha^{\text{CG}}s_{ex}^{\text{CG}})\,/\,\exp(s_{ex}^{HS})\approx 1\). We confirmed that this _back-mapping approximation_ holds for the molecular system studied in this work. Detailed analysis and the underlying physical principles of this approximation will be pursued in a future article.
Finally, the Rosenfeld scaling is applied to Eq., giving:
\[D_{0}^{\text{PY}}\approx\frac{\frac{6^{1}}{3}}{4}\pi^{\frac{1}{6}}\eta^{\frac {1}{3}}\frac{(1-\eta)^{2}}{\eta(\eta^{2}-2\eta+4)}\exp\left[\frac{3(2\eta-\eta ^{2})}{2(1-\eta)^{2}}\right]. \tag{34}\]
Equation again highlights our central idea that the diffusion coefficient in a single-site CG model is solely governed by the packing fraction of the effective hard sphere system.
To summarize, in this subsection, we derived an approximate expression for \(D_{0}\) for molecular CG systems. While analytically deriving such an expression is not practically feasible for the real CG systems, we overcame this bottleneck by representing the fluid of CG particles as effective hard spheres. Since the dynamics of hard spheres can be (approximately) described analytically by both kinetic theory and the alternative scaling relationship by Dzugutov, we showed that for a given equation of state, \(D_{0}\) can be expressed as a function of the packing fraction if one could faithfully map the molecular CG system to its corresponding hard sphere system.
We note that the present approach is somewhat similar to the previous work of Bretonnet, where the semi-empirical relationship between the hard sphere diffusion coefficient and the Dzugutov scaling was discussed. Furthermore, Bretonnet derived an analytical expression using only the PY EOS as determined from the virial route, \(\mathbb{Z}_{\rm PY}^{p}\), for Dzugutov scaling. Since the Dzugutov relationship is limited to hard spheres, the novelty of the current work is to extend the Enskog theory to an effective CG system under a more general scaling relationship with an improved hard sphere EOS [Eq.]. Results for other choices of the EOS are discussed in the Results section below.
## G.
In this paper series, we utilize our recently developed CG model for water: Bottom-Up Many-Body Water (BUMPer). In Paper I, we have demonstrated the excess entropy scaling relationship of the BUMPer model parametrized by the SPC/E, SPC/Fw, TlP4P/2005, and TlP4P/lce force fields follows Eq.. In principle, any other CG model for water, e.g., monatomic Water (mW), can also be applied for this framework. While the strong orientational preferences of water may require higher-order interactions beyond the two-body interaction, the BUMPer model is designed to reproduce many-body correlations using only pairwise basis sets. This is because the force-matched three-body interactions are in the form of the Stillinger-Weber three-body interaction and can be effectively integrated into the short-range pairwise interactions of BUMPer. In this regard, BUMPer is computationally inexpensive compared to atomistic simulations, or even compared to other CG models with explicit three-body interactions. Despite having only pairwise interactions, our recent findings also demonstrated that BUMPer can accurately capture both pairwise and many-body density correlations. The explicitly integrated three-body contributions in BUMPer are further corroborated by its ability to capture nucleation at the ice/water interface. Moreover, a pairwise decomposable form of the BUMPer interaction allows for the excess entropy scaling to be valid, unlike other models with explicit many-body interactions.
One of the central assumptions made in this work originates from the classical perturbation theory asserting that, for a dense liquid, repulsive forces dominate the liquid structure. However, it is known that this idea cannot generally be applied for highly polar liquids, ionic solutions, and water. This is because strong attractive interactions, such as hydrogen bonding for water at the atomistic resolution, rapidly vary as an interparticle distance changes, and thus they play a non-negligible role in the structure. Nevertheless, we note that the current scope of this work is to apply a minimalist hard sphere mapping to water at the _CG_ resolution. In contrast to the atomistic resolution, hydrogen bonding interactions do _not_ directly appear at the current CG level, as they are implicitly folded into the CG model. Thus, we believe it is sensible to explore the possible usefulness of an effective hard sphere mapping scheme for the CG water system without any strong interactions to account for hydrogen bonding as in the sticky hard sphere model.
Below we analyze dynamical properties (diffusion coefficient) using the BUMPer model parameterized from four different atomistic force fields at 300 K conditions: SPC/E, SPC/Fw, TlP4P/2005, and TlP4P/ice. The procedure to compute the dynamical properties is discussed in Paper I, and the performance of the BUMPer models in terms of structural correlations is extensively analyzed in the original BUMPer paper series.
## III.
A primary goal of this section is to apply the aforementioned theories for molecular CG systems in order to derive the \(D_{0}\) values. The organization of this section follows that of In the first step, we discuss two different approaches for obtaining the effective hard sphere diameter or packing fraction for the given CG systems. In the second step, we utilize Eqs.- to estimate the \(D_{0}\) values for various equations of state. The accuracy of the present approach is evaluated by comparing it to the actual \(D_{0}\) values obtained from the scaling relationship.
## A.
The BH EHSD is directly determined from the many-body CG PMFs. Therefore, compared to the fluctuation matching approach, the BH EHSD is not affected by EOS choices. illustrates the many-body CG PMF used in this work from Paper I and Refs..
The four BUMPer interactions studied in this work are parameterized from FG simulations using SPC-type (Fig. 2(a)) and TIP4P-type (Fig. 2(c)) force fields. Regardless of FG force fields, the resultant BUMPer interactions have relatively similar profiles, represented by two characteristic length scales corresponding to hydrogen bonding and van der Waals interactions.
Many-body CG PMFs and corresponding effective hard sphere diameters via the BH perturbation theory. Effective CG pair potentials of the BUMPer CG models that are parameterized from (a) SPC-based force fields (SPC/E: blue, SPC/Fw: sky-blue) and (c) TIP4P-based force fields (TIP4P/2005: red, TIP4P/Ice: ivory). From each CG pair potential, we calculate the BH diameter by integrating the integrand \(1-\exp(-\beta U(R))\) over the restricted range of separations shown in (b) and (d).
\(dR\) where the BH integrands are depicted in Fig. 2(b) and 2(d), respectively. Here, we chose \(R_{0}\) as the shortest pair distance with a zero potential value, resulting in EHSD values of 2.549 A for SPC/Fw, 2.527 A for SPC/E, 2.563 A for TIP4P/2005, and 2.552 A for TIP4P/Ice.
### B. Step 1: Fluctuation Matching 1. Percus-Yevick EOS
By equating Eq. to Eq., the fluctuation matching approach obtains \(\eta\) by solving the following equation
\[f(\eta)|_{\text{HS}}=1+4\pi\rho\int_{0}^{\infty}dr\;R^{2}[g(R)-1]\Bigg{|}_{ \text{CG}}. \tag{35}\]
Note that the right-hand side of Eq. is evaluated from the CG simulation to fit \(f(\eta)\) from the hard sphere reference system.
\[\mathbb{Z}_{\text{PY}}=\frac{(1+\eta+\eta^{2})}{(1-\eta)^{3}}. \tag{36}\]
Then, the isothermal compressibility \(\kappa_{T}\) of the PY EOS can be obtained using the chain rule
\[\left(\frac{\partial P}{\partial V}\right)=\left(\frac{\partial P}{\partial \eta}\right)\cdot\left(\frac{\partial\eta}{\partial V}\right)=-\left(\frac{ \partial P}{\partial\eta}\right)\cdot\frac{\eta}{V}. \tag{37}\]
The last term \(\eta/V\) is from the definition of the packing fraction: \(\eta=\frac{\pi}{6}\sigma^{3}\frac{N}{V}\). Further simplification is possible using the compressibility factor \(\mathbb{Z}_{\text{PY}}\)
\[\left(\frac{\partial P}{\partial V}\right)=-\frac{\partial}{ \partial\eta}\left(\mathbb{Z}_{\text{PY}}\cdot\frac{N\beta}{V}\right)\cdot \frac{\eta}{V}=-\frac{6}{\pi\sigma^{3}}\cdot\frac{\partial}{\partial\eta} \left(\eta\mathbb{Z}_{\text{PY}}\right)\cdot\frac{\eta}{V}. \tag{38}\]
Substituting Eq. into Eq.
\[\left(\frac{\partial P}{\partial V}\right)=-\frac{\rho}{\beta V}\cdot\frac{(1 +2\eta)^{2}}{(1-\eta)^{4}}. \tag{39}\]
Finally, we arrive at the well-known dimensionless compressibility relationship of hard spheres using the PY EOS
\[S(k=0)^{\text{PY}}_{\text{HS}}=\rho k_{B}T\cdot\left(-\frac{1}{V}\cdot\frac{ \partial V}{\partial P}\right)=\frac{(1-\eta)^{4}}{(1+2\eta)^{2}}. \tag{40}\]
Analytically, Equation has four solutions, but only one solution \(\eta_{\text{PY}}\) satisfies \(0\leq\eta_{\text{PY}}\leq 1\)\[\eta_{\rm PY}=\sqrt{S(k\to 0)^{\rm PY}_{\rm HS}}-\sqrt{S(k\to 0)^{\rm PY}_{\rm HS}}+3\sqrt{S(k\to 0)^{\rm PY}_{\rm HS}}+1. \tag{41}\]
While Eqs.
\[\sigma_{\rm PY}=\left\{\frac{6}{\pi\rho}\cdot\left[\sqrt{S(k\to 0)^{\rm PY}_{\rm HS}}-\sqrt{S(k\to 0)^{\rm PY}_{\rm HS}+3\sqrt{S(k\to 0)^{\rm PY}_{\rm HS}}}+1\right]\right\}^{\frac{1}{3}}. \tag{42}\]
In terms of the EHSD, the fluctuation matching approach provides an EHSD value from a single calculation without any iterations in contrast to the WCA approach.
## 2 Carnahan-Starling EOS
An elegant phenomenological attempt to formulate an accurate, yet empirical, hard sphere EOS was proposed by Carnahan and Starling in order to solve the thermodynamic inconsistency problem of the PY EOS obtained from the virial and compressibility routes:
\[\mathbb{Z}_{\rm CS}=\frac{1+\eta+\eta^{2}-\eta^{3}}{(1-\eta)^{3}}. \tag{43}\]
We note that Eq. can be derived by fitting the integer part of the virial coefficient \(b_{k}\) as \(k^{2}+k-2\) up to the first six coefficients or equivalently obtained by a simple interpolation form:
\[\mathbb{Z}_{\rm CS}(\eta)=\frac{1}{3}\mathbb{Z}^{\nu}_{\rm PY}(\eta)+\frac{2}{ 3}\mathbb{Z}^{\rm c}_{\rm PY}(\eta). \tag{44}\]
The Carnahan-Starling (CS) EOS is almost exact over the stable liquid regime and even significantly into the metastable regime. Thus, we now choose the CS EOS to apply fluctuation matching. From Eq., the partial derivative \(\partial P/\partial\eta\) is
\[\left(\frac{\partial P}{\partial\eta}\right)=\frac{\partial}{\partial\eta} \left[\frac{6k_{B}T}{\pi\sigma^{3}}\cdot\frac{\eta+\eta^{2}+\eta^{3}-\eta^{4}} {(1-\eta)^{3}}\right]=\frac{6k_{B}T}{\pi\sigma^{3}}\cdot\left(\frac{\eta^{4}-4 \eta^{3}+4\eta^{2}+4\eta+1}{(1-\eta)^{4}}\right). \tag{45}\]
Utilizing the chain rule and Eq., one has
\[\left(\frac{\partial P}{\partial V}\right)=-\frac{\rho}{V}k_{B}T\cdot\frac{ \eta^{4}-4\eta^{3}+4\eta^{2}+4\eta+1}{(1-\eta)^{4}}. \tag{46}\]
Finally, the compressibility factor for the CS EOS is then given as
\[S(k=0)^{\rm CS}_{\rm HS}=\frac{(1-\eta)^{4}}{\eta^{4}-4\eta^{3}+4\eta^{2}+4 \eta+1}.\]
## 3 Carnahan-Starling-Kolafa EOS
Even though the CS EOS generally provides an excellent approximate EOS, it can be improved at very high densities while retaining its simple form corresponding to the Carnahan-Starling-Kolafa (CSK) form:
\[\mathbb{Z}_{\text{CSK}}=\frac{1+\eta+\eta^{2}-\frac{2}{3}(\eta^{3}+\eta^{4})}{( 1-\eta)^{3}}. \tag{48}\]
Compared to Eq., Equation correctly captures higher-order contributions in density effects. This slight adjustment has been shown to further enhance the evaluation of both the virial coefficients and the compressibility factors.
Other than the aforementioned EOSs, we note that there are numerous empirically or semi-empirically designed EOSs for hard disks, hard spheres, or even hard hyperspheres (e.g., Table 3.11 in Ref. and Table 1 in Ref.). However, most of them have complicated polynomial forms that rely on numerical fitting procedures in contrast to the PY and or CS EOSs, which we focus on in this paper. Nevertheless, any of these complicated EOSs could be employed in the proposed framework.
\[TP=\frac{6k_{B}T}{\pi\sigma^{3}}.\frac{\eta+\eta^{2}+\eta^{3}-\frac{2}{3}(1+ \eta)\cdot\eta^{4}}{(1-\eta)^{3}}. \tag{49}\]
By repeating the procedures employed to obtain Eqs.-, we arrive at the following fluctuation matching equation for the CSK EOS:
\[S(k=0)^{\text{CSK}}_{\text{HS}}=\frac{3(1-\eta)^{4}}{4\eta^{5}-8 \eta^{4}-8\eta^{3}+12\eta^{2}+12\eta+3}.\]
## 4 CG Fluctuation: Finite Size Effect
In order to solve Eq. in practice, we evaluate the structure factor at zero wave vector \(S(k=0)\) using \(S(k)_{\text{CG}}=1+4\pi\rho\int_{0}^{\infty}(g(R)-1)\frac{\sin{(kR)}}{kR}R^{2}dR\). From the mapped CG trajectory, we applied the Fast Fourier Transformation of \(g(R)\) using the LiquidLib suite. The \(g(R)\) functions were sampled with a bin size of 0.02 A. Alternatively, \(S(k=0)\) can be computed using direct numerical integration. In this way, in order to obtain an accurate and numerically stable value, the finite size effect of \(S(k=0)_{\text{CG}}\) must be considered by integrating up to the finite distance \(R_{\text{cut}}\):
\[S(k=0,R_{\text{cut}})=1+4\pi\int_{0}^{R_{\text{cut}}}dR^{\prime} \cdot R^{\prime 2}[g(R^{\prime})-1].\]Differences between the estimated \(S(k=0,R_{\rm cut})\) and actual \(S(k=0)\) values can be corrected via the scheme proposed by Salacuse et al.:
\[S(k,R_{\rm cut})\approx S_{N}(k,R_{\rm cut})+\frac{S(k=0)}{N}\frac{4}{3}\pi{R_{ \rm cut}}^{3}\cdot\left[\frac{3}{(kR_{\rm cut})^{3}}\cdot(\sin kR_{\rm cut}-kR _{\rm cut}\cos kR_{\rm cut})\right], \tag{52}\]
We used the maximum possible value for \(R_{\rm cut}\) as half of the system box length.
\[S(k=0)\approx\frac{S_{N}(k=0,R_{\rm cut})}{1-\frac{1}{N}\frac{4}{3}\pi\rho{R_ {\rm cut}}^{3}}. \tag{53}\]
The computed \(S(k=0)_{\rm CG}\) values from both methods are in close agreement and for water are: 0.1033 for SPC/E, 0.1051 for SPC/Fw, 0.08814 for TIP4P/2005, and 0.09871 for TIP4P/Ice. Importantly, these values at 300 K and 1 atm conditions are within the reported ranges of 0.06-0.08 from experiments and computer simulations. We attribute the slightly overestimated \(S(k=0)\) values in CG systems to minor differences in the \(g(R)\) during the coarse-graining process. Another source of error might be due to the finite size effect in estimating \(S(k=0)\). Since thermodynamic conditions studied here are within the normal regime and not near a critical point, the effect of an overly long correlation length should not be pronounced. As expected, we checked that doubling the system size gives a \(S(k=0)\) value that differs by a very small amount of 0.007 for the SPC/Fw at 300 K condition. A comprehensive computational analysis to determine the \(S(k=0)\) values under various force field and temperature conditions will be systematically pursued in future studies.
## 5 Fluctuation Matching: Results
We now solve the fluctuation matching equation, Eq., for each EOS discussed above by equating the dimensionless compressibility expression to the corrected \(S(k=0)\) values from our computer simulations. Table 1 lists the computed packing fractions for the CG water systems at 300 K. Interestingly, the BH and fluctuation matching approaches yield similar \(\eta\) values for different choices of force fields and EOSs. Given the different principles underlying these mapping approaches, this agreement indicates that both methods provide nearly similar hard sphere systems with \(\Delta\eta\approx 0.01\) for the same FG system. We also note that the \(\eta\) values for both approaches are well below \(\eta_{f}=0.494\), the volume fraction of freezing, validating the underlying assumption to treat CG particles as equilibrium hard sphere liquids not in the metastable or supercooled regime. The relatively low absolute values of the effective packing fraction near 0.3 may seem surprising for a dense liquid. We believe this is likely a consequence of the relatively large value of \(S(k=0)\) due to the high molecular number density and atypical compressibility due to hydrogen bonding of water. For nonpolar or weakly polar molecular liquids, one expects the mapping would deliver packing fractions substantially larger than 0.3. This will be tested in future work.
\begin{table}
\begin{tabular}{l l} \multicolumn{2}{l}{**Table 1:** Effective hard sphere packing fractions (\(\eta\)) of the hard sphere system mapped from the CG water system using the BH scheme and the fluctuation matching approach with the selected EOSs.} \\ \end{tabular}
\end{table}
Table 1: Effective hard sphere packing fractions (\(\eta\)) of the hard sphere system mapped from the CG water system using the BH scheme and the fluctuation matching approach with the selected EOSs.
## C. Step 2: Estimating the Hard Sphere Diffusion Coefficient 1.
We now determine an analytical form of the hard sphere diffusion coefficient using different EOSs. For the simplest PY EOS, we arrived at the following expression in Section II-F
\[D_{0}^{\text{PY}}\approx\frac{6^{\frac{1}{3}}}{4}\pi^{\frac{1}{6}} \eta^{\frac{1}{3}}\frac{(1-\eta)^{2}}{\eta(\eta^{2}-2\eta+4)}\exp\left[\frac{3 (2\eta-\eta^{2})}{2(1-\eta)^{2}}\right].\]
## 2.
The CS EOS has a more accurate contact value \(g(R=\sigma)\):
\[g(\sigma)=\frac{\mathbb{Z}-1}{4\eta}=\frac{1-\eta/2}{(1-\eta)^{3}}\,,\]
and excess entropy:
\[S_{ex}=-\int_{0}^{\eta}\frac{\mathbb{Z}-1}{\eta^{\prime}}\,d\eta^{\prime}=- \eta\,\frac{(4-3\eta)}{(1-\eta)^{2}}\,. \tag{56}\]
Substituting Eq. and in Eq. yields an expression for the entropy-free diffusion coefficient from the Dzugutov scaling, \(D_{Z,CS}^{0}\)
\[D_{Z,CS}^{0}=\frac{\pi}{96}\frac{(1-\eta)^{6}}{\eta^{2}(2-\eta)^{2}}\exp\left[ \frac{(4\eta-3\eta^{2})}{(1-\eta)^{2}}\right]. \tag{57}\]
In Step 3, the hard sphere system (\(D_{Z,CS}^{0}\)) is mapped back to the original CG system to assess the entropy-free diffusion coefficient \(D_{0,CS}^{HS}\) under the Rosenfeld scaling:
\[D_{0,CS}^{HS}\approx\frac{\pi^{\frac{1}{6}}}{48}\cdot 6^{\frac{4}{3}}\cdot \frac{(1-\eta)^{3}}{\eta^{\frac{2}{3}}(2-\eta)}\exp\left[\frac{(4\eta-3\eta^{2} )}{(1-\eta)^{2}}\right].\]Compared to Eq., differences in \(D_{0}^{HS}\) originate from changes in \(g(\sigma)\) that affect collision rates and excess entropy terms from the scaling relationship.
## 3.
Similarly, \(D_{0}^{HS}\) from the CSK EOS can be derived, and several structural and thermodynamic properties are required in order to derive \(D_{0}^{HS}\).The contact value of the pair correlation and excess entropy are given by
\[g(\sigma)= \frac{\mathbb{Z}-1}{4\eta}=\frac{1-\frac{\eta}{2}+\frac{\eta^{2}}{12}- \frac{\eta^{3}}{6}}{(1-\eta)^{3}}, \tag{59}\] \[S_{ex}= -\int_{0}^{\eta}\frac{\mathbb{Z}-1}{\eta^{\prime}}\,d\eta^{\prime} =-\int_{0}^{\eta}\frac{4-2\eta^{\prime}+\frac{\eta^{\prime}{}^{2}}{3}-\frac{2} {3}\eta^{\prime}{}^{3}}{(1-\eta^{\prime})^{3}}\,d\eta^{\prime}=-\frac{5}{3} \ln(1-\eta)-\eta\,\frac{4\eta^{2}-33\eta+34}{6(1-\eta)^{2}}. \tag{60}\]
Equation is further confirmed by the results in Ref.. Then, substituting Eq. into Eq.
\[D_{Z}^{0} =\frac{\pi}{24}\cdot\frac{(1-\eta)^{\frac{23}{3}}}{\eta^{2}\left( 4-2\eta+\frac{1}{3}\eta^{2}-\frac{2}{3}\eta^{3}\right)}\exp\left[\eta\frac{4 \eta^{2}-33\eta+34}{6(1-\eta)^{2}}\right].\]
Finally, simple manipulations yield an analytic expression for \(D_{0,CSK}^{HS}\):
\[D_{0,CSK}^{HS} \approx\frac{\pi^{\frac{1}{6}}}{96}\cdot 6^{\frac{4}{3}}\cdot\frac{(1-\eta)^{\frac{14}{3}}}{\eta^{\frac{2}{3}} \left(1-\frac{\eta}{2}+\frac{\eta^{2}}{12}-\frac{\eta^{3}}{6}\right)}\exp\left[ \eta\frac{4\eta^{2}-33\eta+34}{6(1-\eta)^{2}}\right].\]
## D.
We now evaluate the computed entropy-free diffusion coefficient \(D_{0}^{\text{HS}}\) of CG water from the dynamically consistent hard sphere system. Table 2 lists the predicted \(\ln(D_{0}^{\text{HS}})\) values from both the BH approach and fluctuation matching scheme based on three EOSs (PY, CS, and CSK). We find that both approaches predict \(D_{0}^{\text{HS}}\) values close to the CG reference value of \(D_{0}^{\text{CG}}=0.7047\), confirming the validity of our assumptions. This finding seems rather remarkable in the sense that: adoption of a simple hard sphere model can still be effective for estimating the accelerated CG dynamics, and the matching dimensionless compressibility idea is useful for molecular liquids. Overall, \(D_{0}^{\text{HS}}\) values estimated by the BH approach incur an error of 11.1%, whereas fluctuation matching yields values that incur errors within 17%. The modest differences between the two approaches for water may not be representative of the general performance of the two methods for other less complex and non-associated molecular liquids, especially in the supercooled regime, where we expect the effective packing fractions will be significantly larger than \(\eta\approx 0.3\) obtained for water. This issue will be explored in a future article.
We also assess the accuracy of the specific EOS chosen in terms of reproducing \(D_{0}\) values. While we find that the relative performance of the EOS in reproducing correct CG dynamics generally follows the accuracy of the EOS itself (BH approach), the relative enhancement is minor (within errors of 2%). This low sensitivity can be understood by the relatively low effective packing fractions of the mapped water system where choice of EOS does not result in major variations. We also note that in this case, the back-mapping approximation is a reasonable assumption, giving \(\exp(\alpha^{\text{CG}}s_{ex}^{\text{CG}})\) / \(\exp(s_{ex}^{HS})\) as 1.1.
In turn, the results listed in Table 2 confirm the high-fidelity nature of our approach for estimating the \(D_{0}\) value by treating CG dynamics with an effective hard sphere model.
\begin{table}
\begin{tabular}{l c c c c c c c} \hline \hline & & \multicolumn{3}{c}{**Barker-Henderson**} & \multicolumn{3}{c}{**Fluctuation Matching**} \\ \cline{3-8} FG & Reference & \multirow{2}{*}{
\begin{tabular}{c} Reference \\ value \\ \end{tabular} } & Percus- & Carnahan- & Carnahan- & Percus- & Carnahan- & Carnahan- \\ Force field & & Yevick & Starling & Starling & Yevick & Starling & Starling \\ \hline SPC/E & & 0.771 & 0.759 & 0.758 & 0.802 & 0.803 & 0.802 \\ SPC/Fw & & 0.789 & 0.777 & 0.776 & 0.796 & 0.797 & 0.795 \\ TIP4P/2005 & & 0.802 & 0.790 & 0.789 & 0.951 & 0.973 & 0.972 \\ TIP4P/Ice & & 0.792 & 0.780 & 0.780 & 0.855 & 0.862 & 0.860 \\ \hline \hline \end{tabular}
\end{table}
Table 2: Effective “entropy-free” diffusion coefficients \(D_{0}^{\text{HS}}\) from the Rosenfeld scaling is predicted from the effective hard sphere systems mapped from the CG water system using the BH and the fluctuation matching approaches with the selected EOSs.
By comparing these \(D_{\rm HS}\) values to the actual numerical CG diffusion coefficients \(D_{\rm CG}\), we find, as expected, that both approaches capture well the accelerated CG diffusion coefficients. Interestingly, after incorporating the excess entropy terms, fluctuation matching shows an almost identical level of description (average errors of 15.1%) compared to the BH approach (16.0%). Remarkably, we emphasize that the trend of acceleration factor ascribed to different atomistic force fields is qualitatively captured in our approach. Here, the effective dynamic acceleration factor is estimated by the ratio of \(D_{\rm HS}\) to \(D_{\rm FG}\). Given the acceleration factors of the CG water models (\(D_{\rm CG}/D_{\rm FG}\) is 6.0 for SPC/E, 3.4 for SPC/Fw, 5.5 for TIP4P/2005, and 7.8 for TIP4P/Ice),
Recovered CG diffusion coefficient using the dynamically consistent hard sphere models \(D_{\rm HS}\) for four different water force fields at 300 K: (a) SPC/E, (b) SPC/Fw, (d) TIP4P/2005, and (e) TIP4P/Ice. Hard sphere packing fractions or effective diameters are estimated via two different approaches: Barker-Henderson (BH) and fluctuation matching (FM). Then, the corresponding diffusion coefficients are obtained by employing different choices of EOS: Percus-Yevick (blue solid), Carnahan-Starling (half-filled), and Carnahan-Starling-Kolafa (double-filled). References values from FG (red solid) and CG (green solid) systems are included as a comparison.
### Estimation of CG Diffusion Coefficient at Different Temperatures
We emphasize that the main advantage of fluctuation matching is that we can estimate the overall diffusion coefficient of CG systems _a priori,_ solely based on information from the FG systems. This is because if the hard sphere EOS is chosen, the fluctuation matching equation _only_ requires \(S(k\to 0)\) to match the long wavelength density fluctuations, not the detailed intermolecular potentials. Especially, this feature would be advantageous for some bottom-up CG approaches that aim to reproduce important structural correlations. Among various bottom-up CG methodologies, using pairwise basis sets, the Iterative Boltzmann Inversion (IBI), Inverse Monte Carlo (IMC), and Relative Entropy Minimization (REM) approaches can match the two-body correlations, i.e., RDF. Also, the Multi-Scale Coarse-Graining (MS-CG) approach satisfies the Yvon-Born-Green equation in liquid physics, indicating that MS-CG models aim to reproduce up to three-body correlations. Notably, BUMPer is developed upon the MS-CG principle, where the many-body projection theory allows for recapitulating the pairwise correlation as well.
Therefore, for bottom-up CG models designed for capturing structural correlations, i.e., RDF, one can approximate the \(S(k\to 0)_{\text{CG}}\) without any CG simulation. In other words, with bottom-up approaches, we can approximate the FG RDF as the CG RDF in Eq. to predict \(D_{0}\) values for CG systems at different thermodynamic conditions, especially variable temperature at atmospheric pressure. This estimation is _not possible_ in the conventional BH approach, in which the effective CG interactions need to be first parameterized in order to employ Eq..
Based on the predicted \(D_{0}\) from the hard sphere description, a complete determination of \(D_{\text{CG}}\) is also possible by considering the effective excess entropy contribution remaining at the CG resolution. In Paper I, we elucidated the differences in excess entropy between FG and CG systems corresponding to the mapping entropy, i.e., the missing contributions to the configurational entropies beneath the CG resolution. For single-site CG models, this mapping entropy is the rotational and vibrational entropies from the FG resolution.
\[s_{\text{trn}}^{ex}|_{\text{CG}}\approx s_{\text{trn}}^{\text{FG}}-s_{\text{ trn}}^{\text{(id)}}=s_{\text{trn}}^{\text{FG}}-\Big{(}1.6787+\frac{3}{2}\ln T \Big{)}. \tag{63}\]
for further derivation and analysis. Equation assumes complete entropy representability between the FG and CG systems. However, we adopt an approximate notation instead of an equality since slight differences between \(s_{\text{trn}}^{ex}|_{\text{FG}}\) and \(s_{\text{trn}}^{ex}|_{\text{CG}}\) are expected due to the pairwise approximation introduced in the CG model, slight differences between \(s_{\text{trn}}^{ex}|_{\text{FG}}\) and \(s_{\text{trn}}^{ex}|_{\text{CG}}\) are expected. With this in mind, we compute the estimated CG diffusion coefficients over a range of temperatures from 280 K to 360 K at 20 K intervals by utilizing FG information only. For different temperature conditions, we used the excess entropies and FG diffusion coefficients reported in Paper I.
As depicted in Fig. 4, it is immediately evident that the hard sphere estimation using our approaches successfully recapitulates the CG diffusion coefficients of water over a wide range of temperatures and for different atomistic force fields. For example, in the case of SPC/E, the reference CG diffusion coefficients obtained are: 1.22\(\times\)10\({}^{-4}\), 1.54\(\times\)10\({}^{-4}\), 1.77\(\times\)10\({}^{-4}\), and 1.96\(\times\)10\({}^{-4}\) cm\({}^{2}\cdot\)s\({}^{-1}\), as temperature increases from 280, 320, 340, and 360 K, respectively. The effect of temperature on the CG diffusion coefficient is well reproduced in the hard sphere model as 9.33\(\times\)10\({}^{-5}\), 1.68\(\times\)10\({}^{-4}\), 2.05\(\times\)10\({}^{-4}\), and 2.55\(\times\)10\({}^{-4}\) cm\({}^{2}\cdot\)s\({}^{-1}\), respectively. In turn, the hard sphere description provides diffusion coefficients for water similar to the CG model values regardless of the EOS adopted.
CG diffusion coefficients (plotted on a logarithmic scale) predicted from FG information for five different temperatures: (a) 280 K, (b) 320 K, (d) 340 K, and (e) 360 K. The \(D_{\rm HS}\) values are computed by combining the entropy representability relationship from FG information (\(\rm\,s_{\rm\,true}^{ex}|_{\rm FG}\)) and dynamically consistent hard sphere models (\(D_{\rm\,0}^{\rm HS}\)) using three different EOSs: Percus-Yevick (blue solid), Carnahan-Starling (half-filled), and Carnahan-Starling-Kolafa (double-filled).
We attribute the success of a hard sphere description to the accurate prediction of CG excess entropies from FG models (entropy representability) and to the effective hard sphere nature of normal liquid dynamics, yielding accurate \(D_{0}\) values for different thermodynamic conditions. Although the short-range structural ordering (e.g., RDF intensities at the first peaks) of associated liquids can be significantly affected by temperature, slight changes appear in the dimensionless compressibility, resulting in modest changes of the \(\eta\) values. For example, in the case of SPC/E, as \(S(k=0)\) varies from 0.10 to 0.088, the resultant \(\eta\) values range from 0.28 to 0.31. In other words, from 280 K to 360 K at 1 atm pressure (ambient liquid water condition), we observe that \(D_{0}\) does not change significantly. Such a weak temperature dependence is consistent with the underlying hypothesis of the excess entropy scaling approach that assumes \(D_{0}\) is invariant under changes in temperature and density. Table 3 further supports our claim that the computed \(D_{0}^{\text{HS}}\) is not very sensitive to temperature. As discussed earlier, the strongly associated nature of water results in rather large values of \(S(k=0)\) and a different temperature dependence compared to nonpolar and weakly polar non-associated liquids.
Recently, Bernhardt and Van der Vegt suggested that there may be a single unified excess entropy scaling relationship that encompasses several CG liquid systems. Nevertheless, they also found that the correlation of the putative unified scaling law is not strong and does not agree with the scaling constant of the Rosenfeld relationship. Our work resolves these inconsistencies, showing that \(D_{0}\) is quite dependent on the effective packing fraction of different CG molecules, and, thus, a single scaling relationship for many CG systems is not generically valid.
\begin{table}
\begin{tabular}{l c c c c c c c} \hline \hline (a) SPC/E & & & & (b) SPC/Fw & & & \\ \hline Temp & \(D_{0,PY}^{HS}\) & \(D_{0,CS}^{HS}\) & \(D_{0,CSK}^{HS}\) & Temp & \(D_{0,PY}^{HS}\) & \(D_{0,CS}^{HS}\) & \(D_{0,CSK}^{HS}\) \\ \hline
280 K & 0.778 & 0.796 & 0.794 & 280 K & 0.774 & 0.792 & 0.791 \\
320 K & 0.827 & 0.849 & 0.848 & 320 K & 0.790 & 0.809 & 0.808 \\
340 K & 0.826 & 0.849 & 0.847 & 340 K & 0.815 & 0.837 & 0.835 \\
360 K & 0.847 & 0.872 & 0.871 & 360 K & 0.844 & 0.868 & 0.867 \\ \hline (c) TIP4P/2005 & & & & (d) TIP4P/Ice & & & \\ \hline Temp & \(D_{0,PY}^{HS}\) & \(D_{0,CS}^{HS}\) & \(D_{0,CSK}^{HS}\) & Temp & \(D_{0,PY}^{HS}\) & \(D_{0,CS}^{HS}\) & \(D_{0,CSK}^{HS}\) \\ \hline
280 K & 0.815 & 0.837 & 0.836 & 280 K & 0.799 & 0.820 & 0.818 \\
320 K & 0.883 & 0.912 & 0.910 & 320 K & 0.863 & 0.889 & 0.888 \\
340 K & 0.906 & 0.936 & 0.935 & 340 K & 0.892 & 0.921 & 0.919 \\
360 K & 0.930 & 0.963 & 0.962 & 360 K & 0.918 & 0.950 & 0.948 \\ \hline \end{tabular}
\end{table}
Table 3: Reduced CG diffusion coefficients \(D_{0}^{\text{HS}}\) of water predicted from the fluctuation matching approach using the FG compressibilities and the Percus-Yevick, Carnahan-Starling, and Carnahan-Starling-Kolafa EOSs. \(D_{0}^{\text{HS}}\) is predicted at temperatures ranging from 280 K to 360 K using four FG force fields: (a) SPC/Fw, (b) SPC/E, (c) TIP4P/2005, and (d) TIP4P/ice.
Altogether, our work sheds light on understanding accelerated CG diffusion by focusing on two points: leveraging the entropy representability relationship between FG and CG systems and constructing effective hard sphere systems that approximate CG dynamics. Given the empirical nature of excess entropy scaling relationships, we introduced another layer of coarsening to map CG systems to hard sphere systems where dynamical properties can be formulated analytically. For conditions that the CG particles can be represented as effective hard spheres, we believe that our findings open up a new avenue for bottom-up CG modeling by interpreting CG dynamics from classical perturbation theory and fluctuation matching. As this paper is the first study of such combined efforts, we anticipate several potential directions that can be taken to further analyze the faster CG dynamics for complex molecular systems. One natural extension would be to elucidate the role of resolution in CG dynamics. Since the continuous effort to establish a theoretical link between the CG dynamics and CG resolution has mainly remained _ad hoc_, the proposed methodology is expected to systematically bridge the choice of CG mapping and the resultant dynamics.
## IV.
In this paper, we developed a new mapping scheme for molecular liquids in the normal regime from a target coarse-grained (CG) system to an effective hard sphere system as part of our ongoing effort to understand accelerated CG dynamics in terms of excess entropy scaling, which was introduced in Paper I of this series. Even though CG interactions are intrinsically many-body potentials of mean force, the treatment of single-site CG systems as hard spheres is substantiated by perturbation theories of liquids where the repulsive intermolecular interaction primarily determines the structural and dynamical properties. Such simplified models, in which the equation of state (EOS) is a function only of packing fraction, allow for employing analytical theories to predict the diffusion coefficient in CG systems.
In determining the packing fraction, conventional perturbation theories may have a limitation to correctly reproduce dynamics as they mainly aim to determine the effective hard sphere diameter for the purpose of predicting equilibrium structural correlations of a reference fluid that interacts only via harsh repulsive forces from an effective hard sphere model. To construct an effective hard sphere model suitable for describing dynamics in the presence of chemical complexity and attractive interactions, we adopt the idea of Mirigian and Schweizer successfully employed for supercooled liquid activated relaxation of molecular and polymeric liquids that the long wavelength amplitude of density fluctuations or dimensionless compressibility is a physically appropriate quantity for constructing a mapping under isobaric conditions in a CG framework. This so-called "fluctuation matching" idea requires the chemistry and thermodynamic state dependent dimensionless compressibilities of the molecular CG and hard sphere systems are exactly equal. The fluctuation matching method directly determines effective hard sphere packing fractions in a manner consistent with the original excess entropy scaling relationship. Along with the conventional BH criterion, we employ fluctuation matching for CG one-site water systems to construct effective hard sphere systems based on adopting different hard sphere fluid EOSs.
Analytical formulations of the entropy-free diffusion coefficient are derived by applying two excess entropy scaling schemes sequentially in an elementary kinetic theory framework for dense liquids. For the mapped hard sphere system, we determine the dynamic properties using the well-known Enskog theory. As the Dzugutov scaling is derived for the hard sphere system, we first apply the Dzugutov scaling to obtain the entropy-free diffusion coefficients \(D_{Z}^{0}\), and then map it back to the Rosenfeld scaling \(D_{0}\) for the original CG system. This scheme allows for estimating the entropy-free diffusion coefficient for the underlying CG system. Given a wide range of applications of excess entropy scaling, this paper opens a promising direction for future research to better characterize the accelerated dynamics of complex molecules at the reduced level, e.g., confined or active matter systems, especially when comprehensive analysis of collective behaviors at longer timescales is computationally challenging.
The key finding from this work is that the estimated entropy-free diffusion coefficient \(D_{0}^{\text{HS}}\) is in remarkable agreement with the actual values from the excess entropy scaling of the CG system \(D_{0}^{\text{CG}}\), as well as in agreement with the full diffusion coefficient \(D^{\text{CG}}\). Notably, and nontrivially, we show that fluctuation matching can be faithfully be applied to molecular liquids with (specific) attractive interactions at the CG level and at temperatures in the normal liquid regime. We believe our work demonstrates the efficacy of the hard sphere mapping to single-site CG systems since the less important degrees of freedom are effectively integrated out, and the resultant CG interactions are spherically symmetric. By employing such a minimalist model, we claim that the acceleration in the CG dynamics can be understood from the FG point of view. We further corroborate our claim by successfully predicting temperature-dependent CG diffusion coefficients _a priori_ using only FG information combined with fluctuation matching.
Finally, our findings lead to a systematic rationalization of the acceleration factor, \(D_{CG}/D_{FG}\). Under the excess entropy scaling formalism, this acceleration can be understood from two contributing factors. While the first term \(\exp(\alpha s_{ex}^{\text{CG}})/\exp(\alpha s_{ex}^{\text{FG}})\) can be understood from the entropy representability relationship, the second term \(D_{0}^{\text{CG}}/D_{0}^{\text{FG}}\) is not relatively clear due to \(D_{0}^{\text{FG}}\). Returning to the issue of the correspondence between FG and CG scaling relationships demonstrated in Paper I, a relevant follow-up question based on the present article is how can we physically understand \(D_{0}^{\text{FG}}\)? Unlike the dynamics at the CG level, dynamics at the full atomistic resolution entail various motions other than pure translation, even at the single-site level of description. However, it might be natural to envision that these other motions, as well as the translational motions, are encoded in the center-of-mass diffusion behavior. In the following paper of this series, we will approach this problem by decoupling rotational motions from translational motion and assessing how rotational diffusion is correlated to the overall translational diffusion. Another possible direction would be to combine fluctuation matching for \(S(k\to 0)_{\text{CG}}\) with non-hard sphere EOSs that can capture the FG dynamics. Using idealized or semi-empirical EOSs, it is possible that rotational and vibrational contributions deviated from hard sphere translations could be addressed with additional variables. Altogether, these proposed directions may lead to a unified framework for unraveling the fundamental differences underlying FG and CG dynamics.
This material is based upon work supported by the National Science Foundation (NSF Grant CHE-2102677). Simulations were performed using computing resources provided by the University of Chicago Research Computing Center (RCC). J.J. acknowledges the Harper Dissertation Fellowship from the University of Chicago and insightful discussions with Professor Jeppe C. Dyre and Professor Eok Kyun Lee. K.S.S acknowledges the support of the University of Chicagoand the Pritzker School of Molecular Engineering during his sabbatical stay where this work was initiated.
|
10.48550/arXiv.2208.01257
|
Understanding Dynamics in Coarse-Grained Models: II. Coarse-Grained Diffusion Modeled Using Hard Sphere Theory
|
Jaehyeok Jin, Kenneth S. Schweizer, Gregory A. Voth
| 3,473
|
10.48550_arXiv.1212.5062
|
###### Abstract
Based on the operatorial formulation of the perturbation theory, the exciton-phonon problem is revisited for investigating exciton-mediated energy flow in a finite-size lattice. Within this method, the exciton-phonon entanglement is taken into account through a dual dressing mechanism so that exciton and phonons are treated on an equal footing. In a marked contrast with what happens in an infinite lattice, it is shown that the dynamics of the exciton density is governed by several time scales. The density evolves coherently in the short-time limit whereas a relaxation mechanism occurs over intermediated time scales. Consequently, in the long-time limit, the density converges toward a nearly uniform distributed equilibrium distribution. Such a behavior results from quantum decoherence that originates in the fact that the phonons evolve differently depending on the path followed by the exciton to tunnel along the lattice. Although the relaxation rate increases with the temperature and with the coupling, it decreases with the lattice size, suggesting that the decoherence is inherent to the confinement.
pacs: 05.60.Gg,71.35.-y,71.38.-k,63.22.+m
## I Introduction
In molecular lattices, understanding how excitons carry energy from one region to another is a key step for explaining many phenomena. Example among many are amide-I excitons in \(\alpha\)-helices that favor the transduction of the chemical energy into mechanical work, vibrons in adsorbed nanostructures that enhance surface reactions or promote quantum information transfer, and Frenkel excitons in light-harvesting complexes that convert solar energy into chemical energy. A molecular lattice exhibits regularly distributed atomic subunits along which the energy of an electronic transition, or of a high-frequency vibrational mode, delocalizes owing to dipole-dipole interaction. This gives rise to a narrow-band exciton whose eigenstates are superimpositions of local states. When the dynamics is governed by the exciton Hamiltonian, such superimpositions are coherent since a phase relation is kept between the local states. This favors a wavelike motion of the exciton that propagates coherently.
Unfortunately, the exciton does not evolve freely but it interacts with the vibrations of the host medium usually responsible for dephasing. Therefore, the fundamental question arises whether the energy delocalizes coherently or incoherently. For instance, recent experiments revealed that the coherent nature of the exciton may explain the remarkable efficiency of light-harvesting complexes. Similarly, in a peptide helix, it has been shown that when a localized amide-I mode is selectively excited, the energy transport rate drastically increases suggesting that a coherent energy flow takes place between amide-I modes.
From a phenomenological point of view, dephasing limited coherent motion was described within stochastic models. The main idea is that the lattice vibrations induce random fluctuations of the site energies that behave as independent stochastic variables. Consequently, the phase of each local state randomly fluctuates so that the coherent nature of the exciton gradually disappears. A wavelike motion takes place in the short-time limit whereas a diffusion like motion occurs in the long-time limit. However, it turns out that the previous scenario is not so simple when a microscopic description is used to mimic the influence of the fluctuating surrounding. To proceed, most theories are based on either the Frohlich model or the Holstein model that provide a general description of an exciton coupled with acoustic or optic phonons, respectively (see for instance Refs.). Depending on the model parameters, the exciton properties exhibit different facets ranging from quantum to classical, from weak coupling to strong coupling, from adiabatic to nonadiabatic and from large to small polarons.
In that context, a quite different result from that derived from stochastic approaches was obtained recently for a narrow-band exciton coupled with acoustic phonons in an infinite lattice. Indeed, within the nonadiabatic weak-coupling limit, that is, when the exciton moves slower than the phonons, it has been shown that the exciton propagates freely as if it was insensitive to the phonons. This feature was established by using a time-convolutionless generalized master equation (TCLGME) for describing the evolution of the exciton reduced density matrix (RDM). In an infinite lattice, the phonons behave as a reservoir insensitive to the exciton so that the Born approximation applies. The influence of the phonons is encoded in a time-dependentrelaxation operator. Up to second order, it involves exciton-phonon coupling correlation functions whose dynamics is governed by free phonons. Consequently, because acoustic phonons exhibit spatial correlations over an infinite length scale, the relaxation operator describes a fast dephasing-rephasing mechanism which prevents the exciton diffusivity. Note that similar results have been obtained for a donor-acceptor pair coupled with a solvent that exhibits spatial correlations.
In the present paper, the exciton-phonon problem is revisited for describing energy redistribution in a finite-size lattice, a question that is rarely addressed. However, characterizing size effects is of fundamental importance because in biology and in nanoscience, relevant structures have reduced dimensionality. In a confined environment, we are faced with a major problem because the phonons no longer behave as a reservoir. Indeed, as observed in the TCL-GME that governs the evolution of the excitonic coherences, that is, the off-diagonal RDM elements between the vacuum and the one-exciton states, the confinement favors quantum recurrences and strong memory effects in the exciton-phonon coupling correlation functions. Therefore, the relaxation operator is an almost periodic function so that the TCL-GME reduces to a linear system of differential equations with almost periodic coefficients. In that case, parametric resonances give rise to an exponential growth of the RDM indicating that the TCL-GME method breaks down. The Born approximation fails to capture the dynamics because the correlations between the exciton and the phonons are no longer negligible during the propagation.
To overcome these problems inherent in the non-Markovian dynamics, different strategies can be used, such as the correlated projector method, the time-dependent projection-operator approach, the effective-mode representation and the quantum jumps approach. Here, we use the operatorial formulation of the perturbation theory (PT) recently introduced for studying the dynamics of the excitonic coherences. In the nonadiabatic weak-coupling limit, this method is a powerful tool for describing the spectral properties of the exciton-phonon system over a broad energy scale. It is particularly suitable for characterizing the coherence dynamics and is more accurate than TCL-GME. Within PT, the system dynamics is governed by an effective Hamiltonian that explicitly accounts for exciton-phonon entanglement. Exciton and phonons are thus treated on an equal footing to go beyond the Born approximation.
The paper is organized as follows. In Sec. II, the exciton-phonon Hamiltonian is described. Then, the results provided by PT are summarized to derive the effective exciton-phonon Hamiltonian. Finally, the exciton RDM is defined as the key observable required for characterizing the energy transfer. Its expression is established within PT. In Sec. III, the simulation of the RDM dynamics is carried out numerically to extract the time evolution of the exciton density. The results are finally discussed in Sec. IV.
## II Theoretical background
### Finite-size exciton-phonon system
In a finite-size lattice containing \(N\) sites \(x=1,...,N\), the exciton-phonon Hamiltonian is defined as (see for instance Refs.)
\[H=H_{A}+H_{B}+\Delta H. \tag{1}\]
In Eq., the first term is the exciton Hamiltonian \(H_{A}\) that acts in the one-exciton subspace \(\mathcal{E}_{A}\). It refers to the dynamics of \(N\) coupled two-level systems with Bohr frequency \(\omega_{0}\). Note that a two-level system represents for instance the first two vibrational states of a localized high-frequency intramolecular vibration. The Hamiltonian \(H_{A}\) is expressed in terms of the bare hopping constant \(\Phi\) between nearest neighbor sites as (note that the convention \(\hbar=1\) will be used throughout this paper)
\[H_{A}=\sum_{x=1}^{N}\omega_{0}|x\rangle\langle x|+\sum_{x=1}^{N-1}\Phi(|x+1 \rangle\langle x|+|x\rangle\langle x+1|), \tag{2}\]
Owing to the confinement, the exciton eigenstates are stationary waves with quantized wave vectors \(K_{k}=k\pi/L\), with \(k=1,..,N\) and \(L=N+1\), as
\[|k\rangle=\sum_{x=1}^{N}\sqrt{\frac{2}{L}}\sin(K_{k}x)|x\rangle. \tag{3}\]
The corresponding eigenenergies \(\omega_{k}=\omega_{0}+2\Phi\cos(K_{k})\) form a symmetric ladder of \(N\) discrete energy levels that belong to a band centered on \(\omega_{0}\) and whose width is approximately \(4\Phi\). In the eigenbasis, the exciton Hamiltonian is thus defined as : \(H_{A}=\sum_{k=1}^{N}\omega_{k}|k\rangle\langle k|\).
In Eq., the second term is the phonon Hamiltonian \(H_{B}=\sum_{p=1}^{N}\Omega_{p}a_{p}^{\dagger}a_{p}\) that acts in the Hilbert space \(\mathcal{E}_{B}\). It describes the external motions of the lattice sites that behave as point masses \(M\) connected via force constants \(W\). The phonons correspond to \(N\) stationary normal modes with quantized wave vectors \(q_{p}=p\pi/L\), with \(p=1,..,N\). The corresponding frequencies are \(\Omega_{p}=\Omega_{c}\sin(q_{p}/2)\) with \(\Omega_{c}=\sqrt{4W/M}\). The dynamics is described using standard phonon operators \(a_{p}^{\dagger}\) and \(a_{p}\) and the eigenstates are the well-known number states denoted \(|n_{p}\rangle\equiv|n_{1},...,n_{N}\rangle\).
The last term in Eq. stands for the exciton-phonon interaction \(\Delta H=\sum_{p=1}^{N}M_{p}(a_{p}^{\dagger}+a_{p})\) that acts in \(\mathcal{E}_{A}\otimes\mathcal{E}_{B}\). It is defined in terms of the coupling matrix \(M_{p}\) that measures the strength of the interaction between the exciton and the \(p\)th phonon mode. The interaction yields a stochastic modulation of each two-level system Bohr frequency by the lattice vibrations.
\[M_{pkk^{\prime}}=\eta_{p}(\delta_{p,k-k^{\prime}}+\delta_{p,k^{\prime}-k}-\delta_{ p,k+k^{\prime}}-\delta_{p,2L-k-k^{\prime}}), \tag{4}\]
Equation shows that \(\Delta H\) favors exciton scattering from state \(|k\rangle\) to state \(|k^{\prime}\rangle\) via the exchange of a phonon \(p\). The allowed transitions are specified by the selection rules \(M_{pkk^{\prime}}\neq 0\) that generalize the concept of momentum conservation in a finite-size lattice.
The Hamiltonian \(H\) will be used for characterizing exciton-mediated energy redistribution. To proceed, we shall focus our attention to the nonadiabatic (\(4\Phi\ll\Omega_{c}\)) weak-coupling (\(E_{B}\ll\Phi\)) limit, a common situation for vibrational excitons in molecular lattices. In that case, \(\Delta H\) being a small perturbation, the unperturbed states \(|k,n_{p}\rangle\) refer to a free exciton accompanied by free phonons. Although \(\Delta H\) favors transitions between unperturbed states, the key point is that there is no resonance between coupled unperturbed states so that PT can be used to solve \(H\).
### Perturbation theory
The operatorial formulation of PT is based on the introduction of a unitary transformation \(U\) that diagonalizes the Hamiltonian \(\hat{H}=UHU^{\dagger}\) in the unperturbed basis. It is written as \(U=\exp(S)\), where \(S\) is an anti-Hermitian generator that is non-diagonal in the unperturbed basis. Within PT, \(S\) is expanded as a Taylor series with respect to \(\Delta H\) so that the diagonalization is achieved at a given order (see appendix A).
\[\hat{H}=\sum_{k=1}^{N}\left[(\omega_{k}+\delta\omega_{k})|k\rangle\langle k|+ \hat{H}_{B}^{(k)}\otimes|k\rangle\langle k|\right], \tag{5}\]
where \(\hat{H}_{B}^{(k)}\) is the Hamiltonian that governs the phonon dynamics when the exciton lies in the state \(|k\rangle\), as
\[\hat{H}_{B}^{(k)}=\sum_{p=1}^{N}(\Omega_{p}+\delta\Omega_{pk})a_{p}^{\dagger}a _{p}. \tag{6}\]
Equations and show that in a state \(|k\rangle\), the energy of an exciton \(\hat{\omega}_{k}=\omega_{k}+\delta\omega_{k}\) is renormalized due to its coupling with the phonons. Similarly, the frequency \(\hat{\Omega}_{pk}=\Omega_{p}+\delta\Omega_{pk}\) of phonon \(p\) accompanied by an exciton in the state \(|k\rangle\) is renormalized by an amount \(\delta\Omega_{pk}\).
\[\delta\omega_{k} = \sum_{p=1}^{N}\sum_{k^{\prime}=1}^{N}\frac{M_{pkk^{\prime}}^{2}}{ \omega_{k}-\omega_{k^{\prime}}-\Omega_{p}}\] \[\delta\Omega_{pk} = \sum_{k^{\prime}=1}^{N}\frac{2M_{pkk^{\prime}}^{2}(\omega_{k}- \omega_{k^{\prime}})}{(\omega_{k}-\omega_{k^{\prime}})^{2}-\Omega_{p}^{2}}. \tag{7}\]
The operator \(\hat{H}\) defines the effective exciton-phonon Hamiltonian. Being diagonal in the unperturbed basis, its eigenvalues \(\epsilon_{k,n_{p}}=\hat{\omega}_{k}+\sum_{p=1}^{N}n_{p}\hat{\Omega}_{pk}\) stand for the system eigenenergies up to second order in \(\Delta H\). The corresponding eigenstates are defined as \(|\Psi_{k,n_{p}}\rangle=U^{\dagger}|k,n_{p}\rangle\). Consequently, \(U\) provides a new point of view in which \(\hat{H}\) no longer describe independent excitations but refers to entangled exciton-phonon states. The entanglement results from a dual dressing effect because a state \(|k\rangle\) defines an exciton dressed by a virtual phonon cloud whereas a number state \(|n_{p}\rangle\) describes phonons clothed by virtual excitonic transitions. The behavior of the energy corrections has been studied in great details.
\[\delta\omega_{k}\approx-E_{B}(1-2/L)-\frac{16\Phi E_{B}}{3\pi\Omega_{c}}\cos( \frac{k\pi}{L}). \tag{8}\]
By contrast, the phonon frequencies are either redshifted or blueshifted, depending on the state occupied by the exciton that accompanies the phonons.
\[\delta\Omega_{pk} \approx -\frac{16\Phi E_{B}}{L\Omega_{c}}\sin(\frac{p\pi}{2L})\cos^{2}( \frac{p\pi}{2L}) \tag{9}\] \[\times \left(\cos(\frac{k\pi}{L})+\frac{\delta_{p,k}-\delta p,L-k}{2} \right).\]
To conclude, let us mention that for a given temperature and a fixed coupling strength, we have introduced a critical length \(L^{*}\approx 0.1\Omega_{c}^{2}/(E_{B}k_{B}T)\) (\(k_{B}\) is the Boltzmann constant) so that for \(L<L^{*}\), PT correctly describes the exciton-phonon dynamics. By contrast, PT breaks down for \(L>L^{*}\) owing to the occurrence of quasi-resonance between unperturbed states.
### Transport properties
According to the standard theory of open quantum systems, a complete understanding of the exciton dynamics is obtained from the knowledge of the RDM \(\sigma(t)=Tr_{B}[\exp(-iHt)\rho\exp(iHt)]\), where \(Tr_{B}\) stands for a partial trace over the phonon degrees of freedom. The time evolution corresponds to an Heisenberg representation with respect to \(H\) and the initial conditions are specified by the exciton-phonon density matrix \(\rho\).
To define \(\rho\), let us mention that without any perturbation the lattice is assumed to be in thermal equilibrium at temperature \(T\). Assuming \(\omega_{0}\gg k_{B}T\), each two-level system is in its ground state. This is no longer the case for the phonons that form a thermal bath described by the Boltzmann distribution \(\rho_{B}=\exp(-\beta H_{B})/Z_{B}\), where \(Z_{B}\) is the phonon partition function (\(\beta=1/k_{B}T\)). Consequently, to study exciton-mediated energy flow, one assumes that the lattice interacts with an external source which brings the system in a state out of equilibrium.
This source is resonantly coupled with the two-level systems, only, and it does not affect the remaining degrees of freedom. Moreover, it is supposed to act during a very short time scale so that the exciton is prepared adiabatically without any change in the quantum state of the phonons. To simplify the discussion, we consider a situation in which an exciton is created on the site \(x_{0}\) at time \(t=0\) so that the initial density matrix is \(\rho=\rho_{A}\otimes\rho_{B}\) where \(\rho_{A}=|x_{0}\rangle\langle x_{0}|\). Note that for amide-I exciton in biopolymers, such an excitation may result from the energy released by the hydrolysis of ATP or from charge neutralization upon electron capture by a protonated helix.
In that context, in the local basis, the exciton RDM is expressed as
\[\sigma(x,x^{\prime},t)=Tr_{B}[\rho_{B}\langle x_{0}|e^{iHt}|x^{\prime}\rangle \langle x|e^{-iHt}|x_{0}\rangle]. \tag{10}\]
Its diagonal elements \(P(x,t)=\sigma(x,x,t)\) yield the exciton density that measures the probability to observe the exciton on site \(x\) at time \(t\) given that it was created on site \(x_{0}\) at time \(t=0\). The exciton density is the central object of the present study. It provides information about the way the excitonic energy flows along the lattice after its initial implementation. Moreover, its knowledge allows us to characterize the influence of the phonons on the process of energy redistribution. The off-diagonal elements of the RDM measure the ability of the exciton to develop or to maintain coherent superimpositions of local states. They provide additional information for determining whether the energy redistribution is coherent or incoherent.
For describing the time evolution of the RDM, we do not derive a GME. Instead, we apply PT that directly provides an expression of the RDM. This procedure can be summarized as follows (see appendix B). We first introduce \(U\) and diagonalize \(H\) in Eq.. Second, we use the fact that \(\hat{H}\) is the sum of independent contributions, each contribution describing the exciton-phonon system when the exciton occupies a specific state \(|k\rangle\) (Eq.). Then, we define the following Heisenberg representation \(U_{k}(t)=e^{i\hat{H}_{B}^{(k)}t}Ue^{-i\hat{H}_{B}^{(k)}t}\).
\[\rho_{B}^{(kk^{\prime})}(t)=\frac{\exp\left[-\beta H_{B}-it(\hat{H}_{B}^{(k)}- \hat{H}_{B}^{(k^{\prime})})\right]}{Z_{B}^{(kk^{\prime})}(t)}, \tag{11}\]
where
\[Z_{B}^{(kk^{\prime})}(t)=Tr_{B}e^{-\beta H_{B}-it(H_{B}^{(k)}-H_{B}^{(k^{\prime })})}. \tag{12}\]
Strictly speaking, \(\rho_{B}^{(kk^{\prime})}(t)\) is not a density matrix because it yields complex values for the phonon population. However, it is isomorphic to \(\rho_{B}\) with the correspondence \(\beta\Omega_{p}\rightarrow\beta\Omega_{p}+i(\delta\Omega_{pk}-\delta\Omega_{pk^ {\prime}})t\). It thus yields averages equivalent to thermal averages and facilitates the calculations.
\[\sigma(x,x^{\prime},t)=\sum_{k=1}^{N}\sum_{K^{\prime}=1}^{N}\frac{Z_{B}^{(kk^ {\prime})}(t)}{Z_{B}}e^{-i(\hat{c}_{k}-\hat{c}_{k^{\prime}})t}Tr_{B}\left[\rho _{B}^{(kk^{\prime})}(t)\langle x_{0}|U_{k^{\prime}}^{\dagger}(-t)|k^{\prime} \rangle\langle k^{\prime}|U|x^{\prime}\rangle\langle x|U^{\dagger}|k\rangle \langle k|U_{k}(-t)|x_{0}\rangle\right]. \tag{13}\]
The final step, quite fastidious, consists in expanding \(U\) as a Taylor series with respect to \(\Delta H\). By bringing together the various terms, one finally obtains the second order expression of \(\sigma(x,x^{\prime},t)\) that will be used in the following of the text. Note that we felt it was unnecessary to give the expression of the RDM that exhibits about thirty different contributions.
To conclude, let us mention that in the weak-coupling limit, the exciton-phonon coupling mainly induces a renormalization of the system energies without significantly modifying the quantum states. The exciton-phonon eigenstates basically correspond to the unperturbed states so that the transformation \(U\) behaves as the unit operator in Eq..
\[\sigma_{0}(x,x^{\prime},t) = \sum_{k=1}^{N}\sum_{k^{\prime}=1}^{N}\frac{Z_{B}^{(kk^{\prime})} (t)}{Z_{B}}e^{-i(\hat{c}_{k}-\hat{c}_{k^{\prime}})t} \tag{14}\] \[\times \langle x_{0}|k^{\prime}\rangle\langle k^{\prime}|x^{\prime} \rangle\langle x|k\rangle\langle k|x_{0}\rangle.\]
## III Numerical results
In this section, Eq. is used for studying the excitonic energy flow in a finite-size lattice. Despite its general nature, the previous formalism will be applied for investigating energy redistribution in a lattice of H-bonded peptide units, a system for which the parameters are well-known (Note that this formalism can be used for describing many situations in which a narrow-band exciton interacts with the phonons of the lattice such as Frenkel exciton in molecular aggregates or vibrational exciton in adsorbed nanostructures). In such a lattice, peptide units linked by H bonds are regularly distributed. Each site contains an amide-I mode (C=O vibration) that gives rise to a vibrational exciton. This exciton interacts with the H bond vibrations that form a bath of acoustic phonons at temperature \(T\). To study the amide-I dynamics, the following parameters are used: \(\omega_{0}=1660\) cm\({}^{-1}\), \(W=15\) Nm\({}^{-1}\), \(M=1.8\times 10^{-25}\) kg, \(\Omega_{c}=96.86\) cm\({}^{-1}\) and \(\Phi=7.8\) cm\({}^{-1}\). To avoid PT breakdown, the coupling strength will vary around \(\chi=10\) pN. Of course, the model displayed in Sec. II is too simple to accurately describe vibrational energy flow in a real biopolymer whose dynamics exhibits a tremendous complexity due to the large number of degrees of freedom. In particular, this model is unable to account for the finite amide-I lifetime because it conserves the number of amide-I exciton. Nevertheless, it involves ingredients that play a key role in interpreting specific experiments such as pump-probe spectroscopy in \(\alpha\)-helices and Electron Capture Dissociation in finite-size polypeptides. Consequently, we do not claim that the model is relevant to explain in details the vibrational dynamics in a protein. Nevertheless, its interest lies in the fact that it provides a simple approach to promote the idea that the confinement modifies the exciton energy redistribution.
### Time evolution of the exciton density
As illustrated in for \(N=9\), \(T=300\) K, \(\chi=10\) pN and \(x_{0}=5\), the evolution of the exciton density is governed by several time scales whose duration depends on the model parameters. In the short-time limit (\(t<20-30\) ps), shows that \(P(x,t)\) flows almost coherently along the lattice. Note that a purely coherent motion is illustrated by the thick lines in Fig. 1a, as discussed in the Sec. IV. The exciton behaves as a confined wave that experiences reflections on the lattice sides. Indeed, after the initial excitation, two wave packets propagate on each side of the central site \(x_{0}\). At \(t\approx 0.4\) ps, the wave packets have left the excited region and more than 50% of the initial energy occur on the neigboring sites \(x_{0}\pm 1\). At \(t\approx 2.1\) ps, the wave packets reach the lattice sides where the exciton density reaches 0.36. Then, the wave packets are reflected and they propagate back to the central site. Consequently, 71% of the initial energy recur on the central site at \(t\approx 7.1\) ps. Such a mechanism takes place almost periodically so that quantum recurrences occur. Therefore, \(P(x_{0},t)\) shows a series of peaks. The first peaks take place at \(t=11.5\) ps, \(t=18.6\) ps and \(t=23.0\) ps, the corresponding amplitudes being equal to 0.81, 0.74 and 0.53.
Over a longer time scale, the coherent nature of the exciton tends to disappear. The recurrences in \(P(x_{0},t)\) become less and less pronounced. For instance, when the exciton moves coherently, \(P(x_{0},t)\) exhibits two strong recurrences at \(t=47.4\) ps and \(t=105.9\) ps, the corresponding amplitudes being 0.98 and 0.97 (the second recurrence is not drawn). By contrast, when the exciton-phonon coupling is taken into account, \(P(x_{0},t)\) remains smaller than 0.46 in the neighborhood of the first recurrence and it does not exceed 0.27 in the neighborhood of the second recurrence. In fact, \(P(x,t)\) exhibits several spectral components.
Time evolution of the exciton density \(P(x,t)\) for \(N=9\), \(T=300\) K, \(\chi=10\) pN and \(x_{0}=5\). (a) Short-time evolution and (b) time evolution over a longer time scale. Note that thick lines in refer to a coherent energy transfer that arises when the relaxation induced by the phonons is neglected (see the Sec. IV).
However, as time increases, specific spectral components disappear so that, after approximately 100 ps, \(P(x,t)\) reduces to a damped sine function. These features clearly characterize a relaxation mechanism that affects the energy redistribution over an intermediate time scale.
Finally, in the long-time limit, all the spectral components vanish. The density converges toward a stationary distribution that specifies a new equilibrium density \(P_{eq}(x)\). With the parameters used in Fig.1, \(P_{eq}(x)\) is not uniformly distributed. At \(t=500\) ps, we obtain \(P(x_{0},t)\approx 0.2\) whereas \(P(x,t)\approx 0.1\), \(\forall x\neq x_{0}\), suggesting that the exciton keeps the memory of its initial state despite its interaction with the phonons. Note that for \(\chi=10\) pN, we have verified that the zero order approximation of the RDM (Eq.) yields a quite good estimate of the time evolution of the density.
The rate at which the exciton reaches the equilibrium depends on both the coupling strength and the temperature. This feature is illustrated in Fig.2 that displays the time evolution of the density on the excited site \(P(x_{0},t)\) for different parameter values. As shown in Fig. 2a, the coupling \(\chi\) enhances the relaxation rate. The larger \(\chi\) is, the faster the \(P_{eq}(x_{0})\) is reached. To estimate the relaxation time \(\tau_{R}\), we assume that \(P(x_{0},t)\) scales as a rapidly varying signal dressed by a smooth envelope function that describes an exponential decay toward equilibrium. In doing so, one successively obtains \(\tau_{R}\approx 100\) ps, \(\tau_{R}\approx 40\) ps and \(\tau_{R}\approx 20\) ps for \(\chi=8\) pN, \(\chi=12\) pN and \(\chi=16\) pN, indicating that \(\tau_{R}\propto 1/\chi^{2}\). Similarly, as displayed in Fig. 2b, the temperature also enhances the relaxation rate. One thus obtains \(\tau_{R}\approx 190\) ps, \(\tau_{R}\approx 90\) ps and \(\tau_{R}\approx 60\) ps for \(T=100\) pN, \(T=200\) pN and \(T=300\) pN. Such a behavior suggests that the relaxation time is inversely proportional to \(T\). Note that the zero order RDM (Eq.) yields results almost identical to those displayed in
### Equilibrium distribution
The influence of the initial position of the exciton on the equilibrium distribution is illustrated in for \(N=9\), \(\chi=10\) pN and \(T=300\) K. Although \(P_{eq}(x)\) is invariant under mirror symmetry, that is, \(P_{eq}(x)=P_{eq}(L-x)\), two distinct situations occur. When \(x_{0}=L/2\), \(P_{eq}(x)\) is partially localized on the excited site. It is almost uniformly distributed over the sites \(x\neq x_{0}\) (\(P_{eq}(x)\approx 0.1\)\(\forall x\neq x_{0}\)) whereas it is approximately two times larger on the excited site (\(P_{eq}(x_{0})\approx 0.17\)). In fact, when \(\chi\) nearly vanishes or at very low temperature, \(P_{eq}(x)\) exhibits an universal behavior. It depends only on the lattice size and its scales as \(P_{eq}(x)\approx(1+\delta_{xx_{0}})/L\). By contrast, when \(x_{0}\) is not located at the center of the lattice, \(P_{eq}(x)\) partially localizes over the two sites \(x_{0}\) and \(L-x_{0}\). It remains almost uniform over the sites \(x\neq x_{0}\) and \(x\neq L-x_{0}\) (\(P_{eq}(x)\approx 0.1\)).
Influence of the initial position of the exciton \(x_{0}\) on the equilibrium distribution \(P_{eq}(x)\) for \(N=9\), \(\chi=10\) pN and \(T=300\) K.
Time evolution of the exciton density \(P(x_{0},t)\) for \(N=9\) and \(x_{0}=5\). (a) Influence of the exciton-phonon coupling strength for \(T=300\) K and (b) influence of the temperature for \(\chi=10\) pN.
One thus obtains \(P_{eq}(x_{0})=P_{eq}(L-x_{0})=0.146\), 0.142, 0.139 and 0.138 for \(x_{0}=1\), 2, 3 and 4, respectively. Note that when \(\chi\) nearly vanishes or at very low temperature, we have observed the following universal behavior: \(P_{eq}(x_{0})=P_{eq}(L-x_{0})\approx 3/2L\) and \(P_{eq}(x)\approx 1/L\) for \(x\neq x_{0}\) and \(x\neq L-x_{0}\).
The equilibrium distribution is also quite sensitive to the coupling and to the temperature, as shown in for \(N=9\) and \(x_{0}=L/2\). When \(\chi\) increases, \(P_{eq}(x_{0})\) decreases whereas \(P_{eq}(x)\) increases \(\forall x\neq x_{0}\). Whatever the \(x\) values, \(P_{eq}(x)\) scales linearly with \(E_{B}\). Therefore, when \(\chi\approx 16\) pN (\(E_{B}\approx 0.86\) cm\({}^{-1}\)), the equilibrium density becomes almost uniform in the neighborhood of the excited site. One thus obtains \(P_{eq}(x_{0})\approx P_{eq}(x_{0}\pm 1)\approx 0.12\). Then, when \(\chi\approx 18\) pN (\(E_{B}\approx 1.08\) cm\({}^{-1}\)) a hole occurs in the exciton density at the center of the lattice, that is, \(P_{eq}(x_{0})<P_{eq}(x)\)\(\forall x\neq x_{0}\). This hole continues to deepen when \(\chi\) increases again. Note that, as discussed in Sec. II.B, PT is valid provided that \(\chi\) remains smaller than a critical value. With the parameters used in Fig. 4a, this critical value is approximately \(\chi\approx 20\) pN (\(E_{B}\approx 1.4\) cm\({}^{-1}\)). Above this value, unphysical results were obtained, that is, \(P_{eq}(x_{0})<0\) when \(\chi>26\) pN. As shown in Fig. 4b, the influence of the temperature is quite similar to that of the coupling. Indeed, as \(T\) increases, \(P_{eq}(x_{0})\) decreases whereas \(P_{eq}(x)\) increases \(\forall x\neq x_{0}\). Note that the density evolves linearly with the temperature. The key point is that over a broad temperature range, the temperature slightly affects the equilibrium distribution. For instance, \(P_{eq}(x_{0})\) decreases from 0.2 to 0.17 when the temperature increases from 0 to 300 K. Consequently, the exciton density does not become uniform in the neighborhood of the excited site and it does not exhibit a hole at the center of the lattice. Finally, a numerical fit of the data has revealed that \(P_{eq}(x_{0})\approx 2/L-3\eta(B)E_{B}k_{B}T/\Omega_{c}^{2}\), where \(\eta(B)\) is a slowly varying function of the adiabaticity \(B=2\Phi/\Omega_{c}\) quite close to unity.
### Linear Entropy and relaxation rate
Let us now focus our attention on the relaxation process. As shown in Figs. 1 and 2, it is not easy to define a relaxation time because the exciton density exhibits several spectral components that decay according to different relaxation rates. To overcome this difficulty, a single relaxation time must be extract from a more "global observable". To proceed, we consider the linear entropy defined as \(S(t)=1-Tr_{A}[\sigma(t)^{2}]\).
Parameter dependence of the equilibrium distribution for \(N=9\) and \(x_{0}=5\). (a) Influence of the exciton-phonon coupling for \(T=300\) K and (b) influence of the temperature for \(\chi=10\) pN.
Time evolution of the linear entropy for \(N=9\), \(x_{0}=5\) and \(T=300\) K. The entropy \(S_{0}(t)\) (dashed lines) is defined in terms of the RDM \(\sigma_{0}(t)\) (Eq.) whereas the entropy \(S_{2}(t)\) (solid lines) is obtained from the RDM \(\sigma(t)\) (Eq.).
It accounts for the arrow of time and it describes the relaxation mechanism as a consequence of an information transfer "from the exciton toward the phonon bath" in the form of exciton-phonon correlations.
The behavior of the entropy is shown in for \(N=9\), \(x_{0}=5\), \(T=300\) K and for two \(\chi\) values. Two kinds of calculations have been carried out by introducing the zero order entropy \(S_{0}(t)\) (dashed lines), defined in terms of the RDM \(\sigma_{0}(t)\) (Eq.), and the second order entropy \(S_{2}(t)\) (solid lines), defined in terms of the RDM \(\sigma(t)\) (Eq.). The entropy \(S_{0}(t)\) is a monotonic function that increases with time. Initially equal to zero, it first scales as \(t^{2}\) in the very short-time limit. Then, it increases as time increases by following an exponential function that rises to a maximum. In the long-time limit, it converges to a constant value \(S_{\infty}\), quite close to unity, that depends only on \(L\) and \(x_{0}\). When the entangled nature of the exciton-phonon eigenstates is taken into account, the time evolution of the linear entropy slightly changes. As previously, \(S_{2}(t)\) increases from zero to reach a constant value. In the intermediate-time limit and in the long-time limit, it still behaves as an exponential function that rises to a maximum. However, two main differences occur. First, in the very short-time limit, \(S_{2}(t)\) rapidly increases over a few tenths of ps. Then, the entropy no longer defines a monotonic function. Instead, it behaves as an increasing function that supports high-frequency small-amplitude oscillations.
Such a behavior allows us to introduce a phenomenological measure of the relaxation rate. To proceed, we works with the zero order entropy whose evaluation is must easier. We thus assume that it scales as \(S_{0}(t)\approx S_{\infty}(1-\exp(-2\Gamma_{R}t))\), the factor 2 in the exponential accounting for the fact that \(S_{0}(t)\) involves the square of the RDM. Of course, this expression works quite well in the intermediate-time limit and in the long-time limit, but it fails in reproducing the short-time behavior of the entropy. So defined, the parameter \(\Gamma_{R}\) represents an effective relaxation rate that provides information on the time \(\tau_{R}=1/\Gamma_{R}\) needed to the exciton to reach the equilibrium. Note that, in the weak-coupling limit, it turns out that there is no much difference between the relaxation rates provided by the two entropies.
The behavior of the relaxation rate \(\Gamma_{R}\) is illustrated in for \(x_{0}=1\). As shown in Fig. 6a, \(\Gamma_{R}\) increases linearly with \(E_{B}\). For instance, at \(T=300\) K, \(\Gamma_{R}\) ranges from \(0.068\) cm\({}^{-1}\) (\(\tau_{R}=77.85\) ps) to \(0.153\) cm\({}^{-1}\) (\(\tau_{R}=34.59\) ps) when \(\chi\) varies from \(10\) pN (\(E_{B}=0.33\) cm\({}^{-1}\)) to \(15\) pN (\(E_{B}=0.75\) cm\({}^{-1}\)). Similarly, reveals that \(\Gamma_{R}\) increases almost linearly with the temperature over a broad temperature range. For instance, for \(\chi=12\) pN (\(E_{B}=0.48\) cm\({}^{-1}\)), \(\Gamma_{R}\) extends from \(0.015\) cm\({}^{-1}\) (\(\tau_{R}=361.90\) ps) to \(0.098\) cm\({}^{-1}\) (\(\tau_{R}=54.10\) ps) when \(T\) varies from \(50\) K to \(300\) K.
Behavior of the relaxation rate \(\Gamma_{R}\). (a) Influence of the exciton-phonon coupling for \(N=9\), \(x_{0}=1\) and for two values of the temperature. (b) Influence of the temperature for \(N=9\), \(x_{0}=1\) and for two values of the coupling. (c) Influence of the lattice size for \(x_{0}=1\), \(\chi=10\) pN and \(T=300\) K.
Similarly, \(\Gamma_{R}\) slightly depends on \(x_{0}\). A rather fast relaxation occurs for \(x_{0}=L/2\) whereas a quite slow relaxation takes place when \(x_{0}=2\) or \(x_{0}=L-2\). For instance, for \(N=9\), \(\chi=10\) pN and \(T=300\) K, we obtained \(\Gamma_{R}=0.087\) cm\({}^{-1}\) (\(\tau_{R}=60.50\) ps) for \(x_{0}=L/2\) and \(\Gamma_{R}=0.065\) cm\({}^{-1}\) (\(\tau_{R}=81.50\) ps) for \(x_{0}=2\). Finally, the most surprising effect arises from the dependence of the rate with respect to the lattice size \(L\). Indeed, as illustrated in Fig. 6c, \(\Gamma_{R}\) decreases with \(L\). It approximately scales as \(\Gamma_{R}\propto 1/L\) so that, for \(\chi=10\) pN and \(T=300\) K, it decreases from \(0.068\) cm\({}^{-1}\) (\(\tau_{R}=77.85\) ps) to \(0.039\) cm\({}^{-1}\) (\(\tau_{R}=135.97\) ps) when \(L\) varies from 10 to 20. This result suggests that the relaxation mechanism is an intrinsic property of the finite-size lattice. The relaxation is thus enhanced by the confinement so that the shorter the lattice is, the larger is the relaxation rate. To conclude, let us mention that the previous data were exploited to extract an analytical expression of the relaxation rate.
\[\Gamma_{R}\approx\alpha(x_{0})\frac{BE_{B}k_{B}T}{\Omega_{c}L}, \tag{15}\]
## IV Discussion
As shown in the previous section, our numerical results contrast sharply with what happens in an infinite lattice with translational invariance. Indeed, in this latter situation, spatial correlations in the phonon bath prevent the exciton diffusivity. The exciton develops a wavelike motion so that it propagates coherently along the lattice as if it was insensitive to the phonons. In a confined lattice, a fully different behavior takes place. Although the exciton delocalizes coherently in the short-time limit, its interaction with the phonons favors a relaxation mechanism when time elapses. Consequently, the coherent nature of the exciton density gradually disappears. The density finally converges toward an almost uniform equilibrium distribution that slightly keeps the memory of the initial exciton position. Of course, in agreement with what one expects from a physical point of view, the relaxation rate is enhanced by the coupling strength and by the temperature. However, more surprisingly, we have observed that an increase of the lattice size softens the influence of the phonon bath resulting in a slowdown in the relaxation process. In other words, the relaxation previously evidenced is an intrinsic property of a confined lattice.
### Approximate expression of the exciton density
To understand the numerical results, let us discuss the physics which derives from the expression of the exciton density. According to Eq.
\[P(x,t)=\sum_{n_{p},m_{p}}\left\langle n_{p}|\rho_{B}|n_{p}\right\rangle\left| \left\langle m_{p},x|e^{-iHt}|x_{0},n_{p}\right\rangle\right|^{2}. \tag{16}\]
So defined, \(P(x,t)\) characterizes a generalized quantum probability. It measures the probability to observe the exciton-phonon system in the factorized state \(|x,m_{p}\rangle\) at time \(t\) provided that it was in the state \(|x_{0},n_{p}\rangle\) at time \(t=0\). Because we are concerned with the exciton dynamics, a sum over all possible phonon states at time \(t\) is performed and an average over the initial phonon state is realized.
To evaluate \(P(x,t)\), one can take advantage of the fact that in the nonadiabatic weak-coupling limit, the exciton-phonon coupling mainly induces a renormalization of the system energies without significantly modifying the quantum states. As a result, the Hamiltonian \(H\) in Eq. can be replaced by the effective Hamiltonian \(\hat{H}\) (Eq.) so that the density becomes
\[P(x,t)=\left\langle\left|\sum_{k=1}^{N}e^{-i\hat{\omega}_{k}t}e^{-i\sum_{p}n_{p }\hat{\Omega}_{pk}t}\langle x|k\rangle\langle k|x_{0}\rangle\right|^{2}\right\rangle \tag{17}\]
In accordance with the laws of quantum mechanics, before the average, the exciton density reduces to the square modulus of the probability amplitude that the exciton tunnels from \(x_{0}\) to \(x\) during time \(t\), the phonons evolving freely from their initial state \(|n_{p}\rangle\). This amplitude is the sum over the elementary probability amplitudes associated to the different paths that the exciton can follow to tunnel. A given path defines a transition through a stationary wave \(|k\rangle\) and it exhibits three contributions. First, it involves the weight of the localized states \(|x\rangle\) and \(|x_{0}\rangle\) in the stationary wave \(|k\rangle\). Then, it depends on a phase factor that accounts for the free evolution of the exciton. Finally, it involves a second phase factor that refers to the quantum evolution of the phonons dressed by the exciton in the state \(|k\rangle\).
\[P(x,t) \approx \sum_{k=1}^{N}\sum_{k^{\prime}=1}^{N}F(kk^{\prime},t)e^{-i(\hat{ \omega}_{k}-\hat{\omega}_{k^{\prime}})t} \tag{18}\] \[\times \langle x_{0}|k^{\prime}\rangle\langle k^{\prime}|x^{\prime} \rangle\langle x|k\rangle\langle k|x_{0}\rangle,\]
where \(F(kk^{\prime},t)\) stands for the so-called decoherence factor defined as
\[F(kk^{\prime},t)=\left\langle e^{-i\sum_{p}n_{p}(\hat{\Omega}_{pk}-\hat{ \Omega}_{pk^{\prime}})t}\right\rangle\equiv\frac{Z_{B}^{(kk^{\prime})}(t)}{Z _{B}}. \tag{19}\]
At this step, it is straightforward to show that Eq. corresponds to the diagonal element of the zero order RDM \(\sigma_{0}(t)\) displayed in Eq.. Provided that \(k\neq k^{\prime}\), \(F(kk^{\prime},t)\) defines a decaying function that tends to zero in the long-time limit. It represents the way the phonon bath destroys the coherence of an excitonic state defined as the superimposition between two stationary waves \(|k\rangle\) and \(|k^{\prime}\rangle\).
\[F(kk^{\prime},t)=\prod_{p=1}^{N}\frac{1-e^{-\beta\Omega_{p}}}{1-e^{-\beta \Omega_{p}-i(\delta\Omega_{pk}-\delta\Omega_{pk^{\prime}})t}}. \tag{20}\]
According to Eq., the time evolution of the exciton density \(P(x,t)\) results from the quantum interferences between the different paths that the exciton can follow to tunnel from \(x_{0}\) to \(x\). Owing to these interferences, \(P(x,t)\) is expressed as the sum of different spectral components, a specific component being associated to a phase factor involving an excitonic Bohr frequency. In addition, each component is modulated by a decoherence factor that results form the fact that the phonons evolve differently depending on the path followed by the exciton. More precisely, when the exciton follows two different paths, the phonons develop two different quantum evolutions so that an additional phase factor occurs. Strictly speaking, this phase factor is not responsible for decoherence. However, because the phonons are initially described by a statistical mixture of number states, an average is required. This average yields a sum over phase factors that interfere with the other giving rise to quantum decoherence.
As shown previously, the decoherence factor behaves as a gaussian function whose time evolution can be extracted from the short-time limit of Eq..
\[F(kk^{\prime},t)\approx e^{-i\sum_{p}\bar{n}_{p}(\hat{\Omega}_{pk}-\hat{\Omega }_{pk^{\prime}})t}e^{-\Gamma_{kk^{\prime}}^{2}t^{2}}, \tag{21}\]
The so-called decoherence rate \(\Gamma_{kk^{\prime}}\), inversely proportional to the corresponding relaxation time \(\tau_{kk^{\prime}}=1/\Gamma_{kk^{\prime}}\), is defined as
\[\Gamma_{kk^{\prime}}=\sqrt{\frac{1}{2}\sum_{p=1}^{N}\Delta\bar{n}_{p}^{2}( \hat{\Omega}_{pk}-\hat{\Omega}_{pk^{\prime}})^{2}}, \tag{22}\]
### Coherent dynamics in the short-time limit
By combining Eqs. and, we are able to formally explain the behavior of the exciton density as follows.
\[P(x,t)\approx\left|\sum_{k=1}^{N}e^{-i\tilde{\omega}_{k}t}\langle x|k\rangle \langle k|x_{0}\rangle\right|^{2}, \tag{23}\]
where \(\tilde{\omega}_{k}\) defines the effective energy of the exciton in state \(|k\rangle\), as
\[\tilde{\omega}_{k}=\omega_{k}+\delta\omega_{k}+\sum_{p=1}^{N}\bar{n}_{p}\delta \Omega_{pk}. \tag{24}\]
As illustrated by the solid lines in Fig. 1a, Eq. shows that the exciton moves freely as if it was insensitive to the phonon bath. A coherent energy transfer takes place, this transfer being characterized by quantum recurrences that result from the confinement of the lattice. Nevertheless, although they preserve the coherent nature of the exciton propagation, the phonons affect the dynamical parameters that govern this wavelike motion through the concept of effective energy. Indeed, as shown in Eq., it is as if the exciton was characterized by temperature dependent eigenvalues \(\tilde{\omega}_{k}\) that exhibit two corrections. The first correction, encoded in the parameters \(\delta\omega_{k}\), accounts for the fact that during its propagation the exciton is dressed by a virtual cloud of phonons. The second correction, \(\sum_{p}\bar{n}_{p}\delta\Omega_{pk}\), results form the phase of the decoherence factors.
In that context, a moment's reflection will convince the reader that \(\tilde{\omega}_{k}\) reduces to the small polaron energy in a state \(|k\rangle\) obtained by using a finite temperature mean field theory (see for instance Ref.). Within this theory, the exciton-phonon coupling is partially removed by performing the so-called Lang-Firsov transformation. This transformation provides a new point of view in which the elementary excitation is no longer an exciton but a small polaron, that is, a composite particle that describes an exciton dressed by a lattice distortion. Then, a mean field approach is invoked to treat the remaining polaron-phonon interaction, i.e. an average is realized over the phonon degrees of freedom according to the phonon density matrix \(\rho_{B}\).
\[\omega_{k}^{(pol)}=\omega_{0}-\epsilon_{B}+2\Phi e^{-s(T)}\cos(\frac{k\pi}{L}), \tag{25}\]
The band-narrowing factor reduces to \(s=8E_{B}/3\pi\Omega_{c}\) at zero temperature where it scales as \(s(T)=4E_{B}k_{B}T/\Omega_{c}^{2}\) at high temperature. In that context, by inserting Eqs. and into Eq., it is straightforward to show that the effective energy \(\tilde{\omega}_{k}\) reduces to the polaron energy \(\omega_{k}^{(pol)}\) in the weak-coupling limit, that is, for \(s(T)\ll 1\). This result is quite interesting because it establishes the link between the present approach, based on a perturbative treatment of the exciton-phonon entanglement, and the mean field procedure usually introduced in the small polaron theory. Within PT,the temperature dependence of the effective energy originates in the modification of the quantum evolution of the phonons owing to exciton-phonon correlations.
### Relaxation mechanism
Over a time scale longer than the shortest relaxation time but shorter than the longest relaxation time, the decaying behavior of the decoherence factors can no longer be disregarded. Therefore, a relaxation mechanism takes place resulting in the decay of the different spectral components of the exciton density. Each component disappears according to a specific decoherence rate. As shown in Eq., the temperature dependence of \(\Gamma_{kk^{\prime}}\) is encoded in the fluctuations of the phonon numbers. Because \(\Delta\bar{n}_{p}\) approximately scales as \(k_{B}T/\Omega_{p}\), \(\Gamma_{kk^{\prime}}\) increases linearly with the temperature. Note that this dependence differs from the standard expression of the relaxation rate that characterizes the dephasing of an open system coupled with a reservoir of harmonic oscillators. Indeed, in most situations the temperature dependence of the rate originates in its dependence with respect to the average phonon number \(\bar{n}_{p}\). At high temperature, both approaches yield a similar temperature dependence since \(\Delta\bar{n}_{p}\approx\bar{n}_{p}\approx k_{B}T/\Omega_{p}\). This is no longer the case at low temperature since \(\Delta\bar{n}_{p}/\bar{n}_{p}=\exp(\beta\Omega_{p}/2)\). Moreover, \(\Gamma_{kk^{\prime}}\) is proportional to \(E_{B}\) through its dependence with respect to the phonon energy corrections (see Eq.). From these dependences, we thus recover that both the temperature and the coupling strength enhance the relaxation mechanism, as observed in In that context, it is obvious that the duration of the relaxation mechanism will be approximately given by the longest relaxation time \(\tau_{R}=Max[\tau_{kk^{\prime}}]\), that is, the invert of the smallest decoherence rate \(\Gamma_{R}=Min[\Gamma_{kk^{\prime}}]\). From Eq., it turns out that the smallest decoherence rate is \(\Gamma_{12}\). Therefore, by inserting Eq. into Eq., one obtains a typical expression for the relaxation rate, as
\[\Gamma_{R}=\frac{4BE_{B}k_{B}T}{\Omega_{c}L}\sqrt{1+\frac{27\pi^{4}}{16L^{3}}}. \tag{26}\]
Equation, quite similar to the expression of the rate extracted form numerical data (see Eq.), clearly evidences the role of both the exciton-phonon coupling and the temperature. In addition, it shows that the relaxation is an intrinsic property of a confined lattice, the rate \(\Gamma_{R}\) decreasing with the lattice size \(L\). Consequently, the larger is the lattice size, the smaller is the relaxation rate. Such an effect originates in the size dependence of the energy correction of the phonons that scales as \(1/L\), as shown in Eq..
### Equilibrium distribution
Finally, in the long-time limit, that is, over a time scale longer than the longest relaxation time, all the spectral components of \(P(x,t)\) have disappeared owing to quantum decoherence.
\[P_{eq}(x)\approx\sum_{k=1}^{N}\left|\langle x|k\rangle\langle k|x_{0}\rangle \right|^{2}. \tag{27}\]
Equation represents a time independent "classical" probability, the word classical meaning that there is no longer quantum interferences. It thus reduces to the sum of the probabilities associated to the realization of each path that the exciton can follow to tunnel from \(x_{0}\) to \(x\). Consequently, \(P_{eq}(x)\) no longer depends on the temperature nor on the coupling strength. It involves only the lattice size and the initial position of the exciton.
\[P_{eq}(x)=\left\{\begin{array}{ll}2/L&\mbox{if $x=x_{0}$}\\ 1/L&\mbox{if $x\neq x_{0}$}.\end{array}\right. \tag{28}\]
Otherwise, for \(x_{0}\neq L/2\), the equilibrium distribution is defined as
\[P_{eq}(x)=\left\{\begin{array}{ll}1.5/L&\mbox{if $x=x_{0}$ and $x=L-x_{0}$}\\ 1/L&\mbox{otherwise}.\end{array}\right. \tag{29}\]
In a quite good agreement with the results displayed in Fig. 3, \(P_{eq}(x)\) is partially localized on the sites \(x_{0}\) and \(L-x_{0}\) whereas it is almost uniformly distributed over the other sites. It is symmetric with respect to the center of the lattice. At equilibrium, the average position of the exciton is equal to \(\bar{x}=L/2\) and the standard deviation reduces to \(\Delta\bar{x}^{2}=(L^{2}+2)/12+x_{0}(x_{0}/L-1)\).
In fact, such a behavior can be explained as follows.
\[\sigma(k,k^{\prime},t)\approx F(kk^{\prime},t)e^{-i(\tilde{\omega}_{k}-\tilde {\omega}_{k^{\prime}})t}\langle x_{0}|k^{\prime}\rangle\langle k|x_{0}\rangle. \tag{30}\]
Because the decoherence factor \(F(kk^{\prime},t)\) decays with time provided that \(k\neq k^{\prime}\), the coherences of the RDM gradually disappear. By contrast, the populations are time independent and they remain unchanged. Consequently, the equilibrium corresponds to a statistical mixture of stationary waves \(|k\rangle\). The corresponding probabilities, that is, \(|\langle x_{0}|k\rangle|^{2}\), define the probabilities to observe the exciton in states \(|k\rangle\) given that it occupies the state \(|x_{0}\rangle\). From the expression of the stationary waves (Eq.), one easily understands the characteristic of the equilibrium distribution. Note that within the previous scenario, we do not recover that \(P_{eq}(x)\) depends on both the temperature and the coupling strength (see Fig. 4). In fact, the influence of these parameters results from the entangled nature of the exciton-phonon eigenstates, entanglement that is neglected within the zero order approximation of the RDM.
### Linear entropy
To conclude, let us discuss the behavior of the linear entropy. Because the entropy measures the missing information about the state of the exciton, its time evolution can be interpreted as follows. At time \(t=0\), the exciton is in a pure state \(|x_{0}\rangle\) so that the entropy vanishes. Then, when time elapses, quantum decoherence takes place. The coherences of the RDM gradually disappear so that the exciton is finally described by a statistical mixture of stationary waves. Consequently, this relaxation mechanism is accompanied by an increase of the linear entropy (see Fig. 5).
\[S_{\infty}=\left\{\begin{array}{ll}1-1.5/L&\mbox{if $x_{0}\neq L/2$}\\ 1-2/L&\mbox{if $x_{0}=L/2$}.\end{array}\right. \tag{31}\]
However, does the missing information correspond to an information transfer "from the exciton toward the phonon bath" in the form of exciton-phonon correlations? I don't think so. In fact, at zero order, the missing information on the exciton state is due to our ignorance of the phonon states. Indeed, because the phonons are in thermal equilibrium, they are described by a statistical mixture of number states. Therefore, the increase of the entropy originates in the quantum decoherence that results from the average over the phonon degrees of freedom. The missing information on the initial phonon state is thus converted into missing information on the exciton state. Such a conversion is unidirectional so that the entropy behaves as a monotonic function. By contrast, at second order, has revealed that the linear entropy exhibits high-frequency small-amplitude oscillations. These oscillations account for the fact that the process of information transfer between the exciton and the phonons is partially bidirectional. Basically, a certain amount of information is transferred from the exciton to the phonon bath. Then, owing to the entangled nature of the exciton-phonon eigenstates, a part of this information is restored to the exciton... and so on. Nevertheless, owing to quantum decoherence, this information exchange gradually disappears and the linear entropy overall increases as time increases.
## V Conclusion
In the present paper, the exciton-phonon problem was revisited for investigating exciton-mediated energy redistribution in a finite-size lattice. To proceed, the dynamics was described using the operatorial formulation of PT. Within this method, exciton-phonon entanglement is taken into account through a dual dressing effect, the exciton being dressed by a virtual phonon cloud whereas the phonons are clothed by virtual excitonic transitions. In the nonadiabatic weak-coupling limit, exciton and phonons were treated on an equal footing so that we were able to overcome the problem caused by the fact that the phonons no longer behave as a reservoir.
In that context, special attention has been paid to describing the exciton density that provides information about the way the excitonic energy flows along the lattice. In a marked contrast with what happens in an infinite lattice, we have shown that the dynamics of the density is governed by several time scales. In the short-time limit, the density propagates coherently as if the exciton was insensitive to the phonons. Then, over an intermediate time scale, the exciton-phonon interaction favors quantum decoherence so that the coherent nature of the density gradually disappears. Finally, in the long-time limit, the density converges toward an almost uniform equilibrium distribution.
From a physical point of view, the exciton density measures the probability for the exciton to tunnel between lattice sites, the phonons evolving freely. Therefore, its behavior results from the quantum interferences between the paths that the exciton can follow to tunnel. The key point is that when the exciton follows two different paths, the phonons develop two different quantum evolutions. The density thus depends on phase factors that reduce to time decaying functions after performing an average over the phonon degrees of freedom. Based on the characterization of the linear entropy, we have shown that the corresponding decay rate is enhanced by the exciton-phonon coupling strength and by the temperature. More surprisingly, our study revealed that an increase of the lattice size softens the influence of the phonons resulting in a slowdown in the decoherence process. Quantum decoherence is thus an intrinsic property of a confined lattice.
To conclude, we would like to draw the reader's attention to the fact that the present formalism must be restricted to the study of finite-size lattices. Indeed, we have shown that quantum decoherence tends to disappear as the lattice size increases. From these observations, we could extrapolate to conclude that the exciton moves coherently in an infinite lattice. In doing so, we thus recover the results established in Ref.. Unfortunately, this is a misleading appearance because the origin of the decoherence is twofold. First, it results from the modification of the phonon energies induced by the exciton. Consequently, the phonon quantum states evolve differently depending on the state occupied by the exciton. Dynamical phases arise in the exciton RDM so that pure dephasing occurs when the average over the phonon degrees of freedom is performed (see Eq.). Then, as illustrated by the presence of the unitary transformation \(U\) in Eq., the decoherence also originates in the entangled nature of the exciton-phonon eigenstates. In a confined lattice, the entangled nature of the eigenstates is negligible so that the first origin of the decoherence dominates, as detailed in the present paper. By contrast, in an infinite lattice, the opposite situation occurs because the phonon energy corrections vanish (see Eq.). Fortuitously, in the weak-coupling limit, it turns out thatthe exciton-phonon entanglement does not favor an incoherent (diffusion-like) motion of the exciton so that a coherent (wave-like) regime remains. But this is not always the case. For instance, when one considers the dynamics of the RDM elements that measure the coherence between the vacuum and the one-exciton states, we have shown that quantum decoherence takes place in finite lattices, owing to exciton-induced phonon energy shift, and in infinite lattices, owing to the entanglement of the exciton-phonon eigenstates. In that context, forthcoming works will be devoted to establish a quite general approach to treat the exciton-phonon system whatever the lattice size. In doing so, we aim at determining the nature of the decoherence process when the lattice size evolves continuously from a small value to infinity.
## Appendix A Generator of the unitary transformation
As detailed previously, the generator of the unitary transformation up to second order in \(\Delta H\) is given by the following equations
\[[H_{A}+H_{B},S_{1}]=\Delta H,\] \[[H_{A}+H_{B},S_{2}]=\frac{1}{2}[S_{1},\Delta H]^{nd}, \tag{10}\]
From the expression of the exciton-phonon Hamiltonian Eq., one easily obtains
\[S_{1} = \sum_{p}Z_{p}a_{p}^{\dagger}-Z_{p}^{\dagger}a_{p},\] \[S_{2} = \sum_{pp^{\prime}}E_{pp^{\prime}}a_{p^{\prime}}^{\dagger}a_{p}^{ \dagger}-E_{pp^{\prime}}^{\dagger}a_{p}a_{p^{\prime}} \tag{11}\] \[+ \sum_{pp^{\prime}}D_{pp^{\prime}}a_{p}^{\dagger}a_{p^{\prime}}-D _{pp^{\prime}}^{\dagger}a_{p^{\prime}}^{\dagger}a_{p}+\sum_{p}C_{p}.\]
In the exciton eigenbasis, the various operators that enter the definition of the generator are defined as
\[\langle k|Z_{p}|k^{\prime}\rangle = \frac{\langle k|M_{p}|k^{\prime}\rangle}{\omega_{k}-\omega_{k^{ \prime}}+\Omega_{p}}\] \[\langle k|E_{pp^{\prime}}|k^{\prime}\rangle = \frac{\langle k|B_{pp^{\prime}}|k^{\prime}\rangle}{\omega_{k}- \omega_{k^{\prime}}+\Omega_{p}+\Omega_{p^{\prime}}}\] \[\langle k|D_{pp^{\prime}}|k^{\prime}\rangle = \frac{\langle k|\bar{B}_{pp^{\prime}}|k^{\prime}\rangle}{\omega_{ k}-\omega_{k^{\prime}}+\Omega_{p}-\Omega_{p^{\prime}}}\] \[\langle k|C_{p}|k^{\prime}\rangle = \frac{\langle k|A_{p}^{nd}|k^{\prime}\rangle}{\omega_{k}-\omega_{ k^{\prime}}}, \tag{12}\]
where
\[B_{pp^{\prime}} = \frac{1}{2}[Z_{p},M_{p^{\prime}}]\] \[A_{p} = -\frac{1}{2}(Z_{p}^{\dagger}M_{p}+M_{p}Z_{p})\] \[\bar{B}_{pp^{\prime}} = B_{pp^{\prime}}^{\dagger}\delta_{pp^{\prime}}+B_{pp^{\prime}}(1 -\delta_{pp^{\prime}}). \tag{13}\]
## Appendix B General expression of the exciton RDM
By inserting the unitary transformation \(U\), the coherence Eq.
\[\sigma(x,x^{\prime},t) =\] \[Tr_{B}\left[\rho_{B}\langle x_{0}|U^{\dagger}e^{i\hat{H}t}U|x^{ \prime}\rangle\langle x|U^{\dagger}e^{-i\hat{H}t}U|x_{0}\rangle\right].\]
Since \(\hat{H}\) is a sum of independent contributions (see Eq.), the coherence is expressed as
\[\sigma(x,x^{\prime},t) = \sum_{k=1}^{N}\sum_{k^{\prime}=1}^{N}\exp\left[-i(\hat{\omega}_{ k}-\hat{\omega}_{k^{\prime}})t\right] \tag{14}\] \[Tr_{B}\left[\rho_{B}\langle x_{0}|U^{\dagger}|k^{\prime}\rangle e ^{i\hat{H}_{B}^{(k^{\prime})}t}\langle k^{\prime}|U|x^{\prime}\rangle\right.\] \[\left.\langle x|U^{\dagger}|k\rangle e^{-i\hat{H}_{B}^{(k)}t} \langle k|U|x_{0}\rangle\right].\]
Because the operator \(\exp(-i\hat{H}_{B}^{(k)}t)\) defines a unitary evolution and because \([H_{B},\hat{H}_{B}^{(k)}]=0\), the partial trace in Eq.
\[Tr_{B} \left[\rho_{B}e^{-i(\hat{H}_{B}^{(k)}-\hat{H}_{B}^{(k^{\prime})} )t}\langle x_{0}|e^{-i\hat{H}_{B}^{(k^{\prime})}t}U^{\dagger}e^{+i\hat{H}_{B}^ {(k^{\prime})}t}|k^{\prime}\rangle\right.\] \[\left.\langle k^{\prime}|U|x^{\prime}\rangle\langle x|U^{\dagger}| k\rangle\langle k|e^{-i\hat{H}_{B}^{(k)}t}Ue^{+i\hat{H}_{B}^{(k)}t}|x_{0}\rangle \right].\]
At this step, let first define \(U_{k}(t)=e^{+i\hat{H}_{B}^{(k)}t}Ue^{-i\hat{H}_{B}^{(k)}t}\) as the Heisenberg representation of the unitary transformation \(U\) with respect to the Hamiltonian \(\hat{H}_{B}^{(k)}\).
\[\rho_{B}e^{-i(\hat{H}_{B}^{(k)}-\hat{H}_{B}^{(k^{\prime})})t}=\frac{Z_{B}^{(kk^ {\prime})}(t)}{Z_{B}}\rho_{B}^{(kk^{\prime})}(t),\]
with
\[\rho_{B}^{(kk^{\prime})}(t) = e^{-\beta H_{B}-it(\hat{H}_{B}^{(k)}-\hat{H}_{B}^{(k^{\prime})} )}/Z_{B}^{(kk^{\prime})}(t)\] \[Z_{B}^{(kk^{\prime})}(t) = Tr_{B}\left[e^{-\beta H_{B}-it(\hat{H}_{B}^{(k)}-\hat{H}_{B}^{(k^ {\prime})})}\right]. \tag{15}\]
Finally, inserting Eq. into the previous expression of the partial trace and combining the results with Eq. yield the exciton RDM Eq..
|
10.48550/arXiv.1212.5062
|
Energy transfer in finite-size exciton-phonon systems : confinement-enhanced quantum decoherence
|
Vincent J. C. Pouthier
| 4,848
|
10.48550_arXiv.2008.08686
|
### Materials
Reduced \(\beta\)-NADH (SigmaAldrich) was used in experiments. Distilled water and methanol of 96% purity were used for preparing solutions at volume proportions ranging from pure water to 100% methanol. The NADH concentration in all cases was the same and equal to 1 mM. All solutions were prepared fresh daily and used at the temperature of 20\({}^{\circ}\)C.
### Experimental procedure
A femtosecond Ti:Sa oscillator (Mai Tai HP, Spectra Physics) tunable in the spectral range of 6901040 nm with a pulse duration of 100 fs and a repetition rate of 80.4 MHz was used as an excitation source. About 2% of the laser output at 720 nm was split out and sent to the synchronization channel consisted of an attenuator, a lens, and a fast silicon photodiode for synchronization of the fluorescence signals in the TCSPC module. The output laser beam was attenuated and then expanded by a telescope to the diameter of 4 mm. The laser beam polarization was controlled by a half-, or quarterwave plates. The linear polarization degree of the laser beam was better than 0.995, and the circular polarization degree was better than 0.95. The polarized laser beam was focused onto the center of a quartz cuvette containing NADH in water-methanol solution. Average power of the laser beam on the cuvette was kept at the level of about 100 mW.
The fluorescence was collected in the direction perpendicular to the laser beam propagation and formed by a lens into a quasi-parallel beam that was split by a Glan prism into two beams with orthogonal polarizations parallel to X and Y axes. Two orthogonal polarization components were detected by ultrafast avalanche photodiodes (ADP-050-CTC, MPD) with a HWHM of instrumental response function of about 56 ps. The IRF contour was determined experimentally by recording laser pulses scattered from a fiber sample. As found the IRFshape in the conditions of our experiments had a strong spectral dependence, therefore the spectral range of the detected NADH fluorescence was restricted by a narrow-band interference filter with the bandwidth of 460-490 nm installed after the cuvette in front of the Glan prizm. The fluorescence signals were analyzed by a TCSPS module (Picoharp 300, Picoquant). The fluorescence photons were counted for 350 s for each laser beam polarization with a time bin of 4 ps.
## 2.
According to the results of our recent research the fluorescence parameter values were independent of the excitation wavelength in the range 720 - 780 nm within the experimental error bars. Therefore, in this study the excitation wavelength was fixed at 720 nm. Typical fluorescence experimental signals in NADH under two-photon excitation in aqueous solutions for several combinations of polarization of the pump and fluorescence beams are shown in Fig.2.
The abbreviation **I** in Fig.2 with subscripts is the fluorescence intensity, where capital case subscript indices \(YY\) and \(LL\) label the polarization of pumping photons: \(Y\) and \(L\) denote linear polarization along axis Y and left handed circular polarization, respectively. The lower case subscripts \(x\) and \(y\) label the fluorescence beam polarizations. The initial laser beam intensity was the same for all four data sets in Fig.2, therefore different amplitudes and shapes of the four decay curves indicate the dependence of the two-photon excitation probability and fluorescence decay dynamics on the pump and fluorescence polarization indices.
### Experimental data fitting and estimation of experimental errors
Processing of experimental signals and determination of the fluorescence parameters based on eqs. and was carried out by a global fit procedure using a least square differential evolution fitting algorithm with convolution. Experimental data related to \(x\) and \(y\) fluorescence decay polarization components similar to that shown in Fig.2 were fitted using eqs. and separately for linearly and circularly polarized excitation light aimed to determine the global fluorescence parameters \(I_{l}\), \(r_{l}\), \(\tau_{1}\), \(\tau_{2}\), \(a_{1}\), \(a_{2}\), \(\tau_{r}\) and \(I_{c}\), \(r_{c}\), \(\tau_{1}\), \(\tau_{2}\), \(a_{1}\), \(a_{2}\), \(\tau_{r}\), respectively.
Fluorescence decay signals in NADH in aqueous solution under two-photon excitation at 720 nm.
Four different data sets relate to different combinations of the laser and fluorescence polarizations, see text for details. Symbols represent experimental data and solid curves are fits.
These parameters were assumed to be the same in both sets and their values were averaged to obtain final values.
Different sources of experimental data errors were carefully analyzed. According to the analysis made, a statistical method of error estimation dealing with the standard deviation by optimization of the cost function near the global minimum in the case of fitting with exponential functions usually resulted in significant underestimation of actual errors because the statistical errors obtained were typically much smaller than systematic experimental errors.
Although the fit quality obtained when processing the experimental data often looked perfect with the naked eye, the values of the fluorescence parameters fluctuated from one experiment to another much more than was given by the statistical analysis of each data set. To take proper account for systematic experimental errors the experiments were repeated several times at the same conditions to obtain several experimental data sets (usually six \(x\) and \(y\) pairs of fluorescence decay data), and the fluorescence parameters were calculated by fit for each set separately. Then the parameter mean values and standard deviations were calculated by averaging of all sets values with 95 % confidence using Student's \(t\)-value. The error bars determined in this way are used throughout this paper.
The role of the IRF(\(t\)) shape shown in Fig.2 on the determined fluorescence parameter values was carefully analyzed and found to be very important in the conditions of our experiment especially for determination of the rotational diffusion time \(\tau_{r}\) and weighting coefficients \(a_{1}\) and \(a_{2}\). Even small perturbations at the rising edge and the tail of IRF led to noticeable changes in the calculated values of the fluorescence parameters. Several IRF functions were used for experimental data analysis: experimentally determined IRF in a numerical form, smoothed experimental IRF, and analytical IRF form obtained by approximation of the experimental IRF with two Gaussian functions. The best results were obtained by using the experimental IRF smoothed without deformations of the rising edge and the tail. ThisIRF was used for analysis of the experimental data. The details of the fitting protocol and analysis of statistical and experimental errors are given in _SI_.
### Experimental data analysis
The analysis of the obtained fluorescence parameters was based on the general expression for fluorescence intensity \(I(t)\) after two-photon excitation derived in our previous publications and given in eq. in _SI_.
The expression for the fluorescence intensity in eq. was similar to that given earlier by McClain and Wan and Johnson however it was presented in a somehow more compact spherical tensor form allowing for complete separation of the light polarization part controlled by experimentalist from the molecular dynamics part described by the molecular parameters \(M_{K_{e}}(R,R^{\prime},t)\) (\(M\)-parameters).
The \(M\)-parameters are real values that can be extracted from experiment and contain all information on the symmetry and structure of the molecular electronic transitions. The set of \(M\)-parameters at the time of excitation \(t=0\) is equivalent to the set of McClain's Cartesian parameters \(\hat{Q}_{i}\). The relationship between the two sets of parameters is tabulated in ref. In the conditions of the collision-induced rotational diffusion the \(M\)-parameters can be expressed as:
\[M_{K_{e}}(R,R^{\prime},t)=-\sqrt{3}\sum_{q_{e},q^{\prime}_{e}}\left(\left[{\bf F }_{1}^{*}\otimes{\bf F}_{1}^{\prime}\right]_{K_{e}q_{e}}{\cal D}_{q_{e},q_{e}} ^{K_{e}}(t)\left[{\bf S}^{*}{}_{R^{\prime}}\otimes{\bf S}_{R}\right]_{K_{e}q^{ \prime}_{e}}^{*}\right), \tag{5}\]
in _SI_. If \(K_{e}=0\) this matrix element is equal to unity and if \(K_{e}=2\) in the condition of our experiment it is proportional to:
\[{\cal D}_{q_{e},q^{\prime}_{e}}^{2}(t)\sim\delta_{q_{e},0}\delta_{q^{\prime}_{ e},0}e^{-t/\tau_{r}}, \tag{6}\]where \(\tau_{r}\) is the rotational diffusion time.
The spherical components of the two-photon excitation tensor \({\bf S}_{R\gamma}\) in eq. with a rank \(R\) and its component \(\gamma\) are given by:
\[S_{R\gamma} = 2\sum_{q_{1},q_{2},n_{i}}C_{1q_{1}\ 1q_{2}}^{R\gamma}\frac{\langle n _{e}|\hat{d}_{q_{2}}|n_{i}\rangle\langle n_{i}|\hat{d}_{q_{1}}|n_{g}\rangle}{E_ {i}-E_{g}-h\nu}, \tag{7}\]
The terms \(E_{i}\) and \(E_{g}\) in eq. are the electronic energies of the intermediate and ground states, respectively. The body-frame spherical components of the one-photon emission dipole moment \({\bf F}_{1}\) in Eq. are given by:
\[F_{1q_{fl}}=\langle n_{f}|\hat{d}_{q_{fl}}|\widetilde{n}_{e}\rangle, \tag{8}\]
In the case of a one-color two-photon population of a nondegenerate excited electron state that is relevant to our experimental conditions, the rank \(K_{e}\) in Eq. is limited to \(K_{e}=0,2\) and the total number of the \(M\)-parameters is four. There are two zeroth rank (\(K_{e}=0\)) \(M\)-parameters: \(M_{0}\) and \(M_{0}\) that can have only non-negative values and contribute to the isotropic part of the fluorescence intensity and two second rank M-parameters (\(K_{e}=2\)): \(M_{2}(0,2,t)\), and \(M_{2}(2,2,t)\) that govern the anisotropic, polarization-dependent part of the fluorescence intensity. The evolution of the \(M\)-parameters with \(K_{e}=2\) in time due to the rotational diffusion is expressed in Eq. while the zeroth-rank parameters do not depend on time.
All four parameters can be directly determined from experiment (see e.g.) by combining different laser and fluorescence polarizations. The \(M\)-parameters can be readily expressed in terms of the anisotropy in eq. and the parameter \(\Omega\) as follows:
\[r_{l}(t)=\frac{2}{\sqrt{35}}\frac{\sqrt{7}M_{2}(0,2,t)+M_{2}(2,2,t)}{\sqrt{5}M_{ 0}+2M_{0}}\,, \tag{9}\]
\[r_{c}(t)=-\frac{1}{\sqrt{35}}\frac{M_{2}(2,2,t)}{M_{0}}\,, \tag{10}\]
\[\Omega=\frac{I_{LL}}{I_{YY}}=\frac{3M_{0}}{\sqrt{5}M_{0}+2M_{0}}. \tag{11}\]
The M-parameters are very important because according to eqs.,, and they contain all information on the dynamics of the two-photon excitation in terms of the quantum mechanical transition amplitudes and phases. The M-parameters are more general than the components of the two-photon absorption tensor **S** in eq. because they are scalar quantities that do not depend on the particular excitation model and on the choice of the coordinate frame. The \(S_{R\gamma}\) values in eq. can be determined from experiment under certain assumptions, in particular one should know the direction of the fluorescence transition dipole moment \({\bf F}_{1}\) in the relaxed exited state.
## III.
The fluorescence decay parameters determined from experiment as a function of methanol concentration are presented in Figs. 3 - 7.
The fluorescence decay times \(\tau_{1}\) and \(\tau_{2}\) are shown in Fig.3. As can be seen in Fig.3 both fluorescence decay times increase slightly with methanol concentration at the concentrations below 40% and remain practically the same within experimental error bars at higher concentrations. In pure aqueous solution the fluorescence decay time \(\tau_{1}\) and \(\tau_{2}\) values in Fig.3are in agreement with those reported earlier in NADH under one- and two-photon excitation within experimental error bars. Also, both decay times in Fig.3 agree roughly with those reported for NADH in pure methanol, where however a three-exponential model was used.
The preexponential coefficient \(a_{2}=1-a_{1}\) (see eq.) is presented in Fig.4. As can be seen in this figure at methanol concentrations below 70% the coefficient is almost constant and equal to \(a_{2}=0.24\) within the experimental error bars, however at higher concentrations it rises up to 0.4.
At zero and 50% methanol concentrations the coefficient \(a_{2}\) values in are in agreement with those reported by Blacker et al., with our recent results, and with the earlier results of Krishnamoorthy et al. within experimental error bars. In NADH in pure methanol the \(a_{2}\) value in is in general agreement with the result reported by Krishnamoorthy et al. who however used a three-exponential decay fit with a minor (\(a_{3}=0.15\)) contribution from the third exponent with \(\tau_{3}=2.0\) ns. Ladokhin et al. used two, three, and four-exponential decay fits and also reported for NADH diluted in pure methanol the increase of a relative contribution from the longer \(\tau_{2}\) time.
Decay times \(\tau_{1}\) and \(\tau_{2}\) in NADH as function of methanol concentration.
differ both from the results shown in and from those reported by Krishnamoorthy et al.. We do not know the exact reason for this discrepancy, however according to our analysis the coefficients \(a_{1}\) and \(a_{2}\) are the most variable among all fluorescence parameters as they can noticeably change their values even at small variations of the experimental conditions and fitting procedure details.
The rotational diffusion time \(\tau_{r}\) is presented in The water-methanol solution viscosity is also shown in this figure by a red solid curve. One can see that the viscosity in is represented with a non-linear curve with a maximum at about 40% of methanol. As can be seen in Fig. 5, at methanol concentrations from zero to 40% the rotational diffusion time \(\tau_{r}\) raises proportionally to the solution viscosity, has a maximum at about 60% of methanol and then decreases. However at methanol concentrations over 40% the direct proportionality between the \(\tau_{r}\) and the viscosity is violated.
As shown in in aqueous solution \(\tau_{r}=180\pm 30\) ps that agrees well with our recent result and with the result reported earlier by Couprie et al..
Preexponential coefficient \(a_{2}\) in NADH (see eq.) as a function of methanol concentration.
1.6 times smaller than the values reported recently by Blacker et al. at several excitation wavelengths. We believe that this discrepancy could arise because Blacker et al. probably did not take into account IRF in their fitting procedure.
The initial anisotropies \(r_{l}\) and \(r_{c}\) are presented in as function of methanol concentration. As can be seen in both anisotropies practically did not depend on methanol concentration. The anisotropy values are in perfect agreement with previously reported results.
The fluorescence parameters determined: \(a_{2}\), \(\tau_{1}\), \(\tau_{2}\), \(r_{l}\), \(r_{c}\), \(\tau_{r}\), and \(\Omega\) are collected in Table 1. The numbers in parentheses in Table 1 are experimental errors.
Rotational diffusion time \(\tau_{r}\) in NADH as function of methanol concentration. Red solid curve represents solution viscosity.
The two-photon excitation parameter \(\Omega=I_{LL}/I_{YY}\) in NADH is presented in As can be seen the parameter practically does not depend on methanol concentration.
Anisotropies \(r_{l}=r_{YY}\) and \(r_{c}=r_{LL}\) under two-photon excitation in NADH as function of methanol concentration.
## IV Theoretical studies
_Ab initio_ calculations of several conformers of NADH and of NMNH in vacuum, and in water and methanol solutions in the ground and excited states have been carried out to extract
\begin{table}
\begin{tabular}{c c c c c c c} \hline MeOH & \(a_{2}=(1-a_{1})\) & \(\tau_{1}\), ps & \(\tau_{2}\), ps & \(r_{l}\) & \(r_{c}\) & \(\tau_{r}\), ps & \(\Omega\) \\ \hline
0 \% & 0.24 (0.05) & 260 & 610 & 0.47 (0.02) & -0.22 (0.02) & 180 & 0.77 (0.04) \\
19 \% & 0.24 (0.05) & 280 & 680 & 0.47 (0.02) & -0.25 (0.02) & 240 & 0.80 (0.04) \\
26 \% & 0.22 (0.05) & 290 & 690 & 0.46 (0.02) & -0.23 (0.02) & 260 & 0.78 (0.04) \\
33 \% & 0.22 (0.05) & 280 & 700 & 0.47 (0.02) & -0.23 (0.02) & 270 & 0.79 (0.04) \\
39 \% & 0.24 (0.05) & 300 & 700 & 0.47 (0.02) & -0.23 (0.02) & 280 & 0.79 (0.04) \\
41 \% & 0.24 (0.05) & 300 & 710 & 0.47 (0.01) & -0.23 (0.01) & 290 & 0.82 (0.04) \\
50 \% & 0.25 (0.05) & 310 & 720 & 0.47 (0.02) & -0.24 (0.02) & 300 & 0.81 (0.04) \\
56 \% & 0.26 (0.05) & 310 & 720 & 0.47 (0.02) & -0.23 (0.02) & 320 & 0.79 (0.04) \\
67 \% & 0.26 (0.05) & 320 & 750 & 0.47 (0.01) & -0.22 (0.01) & 310 & 0.80 (0.04) \\
78 \% & 0.32 (0.05) & 310 & 730 & 0.47 (0.02) & -0.23 (0.02) & 300 & 0.82 (0.04) \\
86 \% & 0.36 (0.05) & 300 & 720 & 0.46 (0.01) & -0.22 (0.01) & 270 & 0.82 (0.04) \\
94 \% & 0.37 (0.05) & 290 & 730 & 0.46 (0.02) & -0.22 (0.02) & 240 & 0.81 (0.04) \\
96 \% & 0.39 (0.05) & 310 & 740 & 0.45 (0.02) & -0.22 (0.02) & 217 & 0.82 (0.04) \\ \hline \end{tabular}
\end{table}
Table 1: Experimentally determined fluorescence parameters: NADH in water-methanol solutions.
Parameter \(\Omega\) in NADH as function of methanol concentration.
Earlier, _ab initio_ calculations of NADH conformers have been reported by Wu _et al._ at the MP2/6-31G*//6-31G* level and by Kumar _et al._ at the B3LYP/6-311++G** level. To the best of our knowledge, no calculations of NADH dissolved in water and methanol have been carried out till now.
In this paper electronic structure computations of twenty four NADH conformers, including _cis_ and _trans_ modifications dissolved in water and methanol have been performed _ab initio_ by means of the polarizable continuum model (PCM) at the B3LYP-D3BJ/6-31G* level with the GAUSSIAN package. The functionals were extended with the D3 version of Grimme's dispersion correction. One of the conformers (N8, see below) has been calculated at the B3LYP-D3BJ/6-31+G* level + (PCM) for comparing parameters.
A schematic of NADH is shown in Fig.8 where a number of distances and angles between most important atoms and molecular groups are indicated.
Optimized stable ground state geometries in terms of the interatomic distances and tor
Schematic of NADH with important interatomic distances and torsion angles indicated.
The optimized geometries of the first excited state in aqueous solution are presented in _SI_ in Tab. 7.
Vertical excitation energies of four lowest electron excited states, corresponding one-photon oscillator strengths, and ground state dispersion energies (\(E_{D}\)) for _cis_ and _trans_ forms of twelve NADH conformations in water and methanol solutions are presented in _SI_ in Tables 8 and 9, respectively. According to the vertical excitation energy values in Tables 8 and 9 only the first excited state could be two-photon excited in NADH at 720 nm in the conditions of our experiments. Data on the higher excited states are also presented in both tables and can be used for analysis of interactions between excited molecular moieties.
As can be seen in Tables 8 and 9 in _SI_ vertical excitation energies and oscillator strengths to the first excited state shown in the first columns in the tables vary somewhat from one NADH conformation to another. However they do not show a distinct correlation with the values of geometrical molecular characteristics in Tables 5 and 6. On the other hand the dispersion energy in the last columns in Tables 8 and 9 changes dramatically with respect to the interatomic distances \(R_{C6A-C2N}\), \(R_{Nn-Op}\) and the torsion angle \(\chi_{O5B-O5D}\) in Fig.8 that characterizes NA - AD and NA - Pyrophosphate (PP) interactions. \(E_{D}\) as function of the distance \(R_{C6A-C2N}\) between AD and NA rings is shown in separately for _cis_ and _trans_ conformations of NADH in water and methanol. The computed values of \(E_{D}\) are shown in with black squares that are connected with blue lines for the sake of clarity.
At short and medium distances at \(R_{C6A-C2N}<8\) A the conformations can be attributed to the folded group where interaction between AD and NA dominates. This effect is known to be due to \(\pi\)-stacking interaction. At long NA-AD distances when \(R_{C6A-C2N}\geq 8\) A the conformers can be attributed to the unfolded group where the interaction between AD and NA moieties is small, or negligible.
As can be seen in the energy \(E_{D}\) decreases in general with increasing of the distance between AD and NA rings. However as can be seen in the dispersion energy contains a pronounced non-monotonic region at the distances \(R_{C6A-C2N}\) shorter than about 7 A. According to the data presented in Tables 5 and 6 in _SI_, this non-monotonic behavior correlates with sharp changes of the interatomic distance \(R_{C1B-C1D}\) and the torsion angle \(\chi_{O5B-O5D}\) shown in Fig.8. The nonmonotonic behavior can be attributed to the interaction between NA and Pyrophosphate (PP) moieties. This conclusion suggests that the widely used consideration of only folded and unfolded conformation groups is insufficient for understanding the excited state dynamics in NADH because interaction between the NA and PP moieties is sometimes also very important. An important role of NA - PP interactions in NADH was discussed recently by Smith and Tanner.
The direction of the transition dipole moment (TDM) with respect to the NA plane is an important characteristic for excited state symmetry analysis and for verifying the validity of the models developed.
Dispersion energy in NADH _cis_ and _trans_ conformations dissolved in water and methanol as function of the distance \(R_{C6A-C2N}\) in Fig.8
The results are presented in _SI_ in Tables 10 - 12.
## 5 V.
### Excited state lifetime heterogeneity in NADH
As was already mentioned in the Introduction, several authors suggested a model implying a strong relationship between two decay times observed in NADH (see Fig. 3) and the relative concentration of the folded and unfolded conformations. However this model failed to explain relatively stable values of the preexponential factors in NADH in solutions and existence of two decay times in mononucleotides NMNH (\(\beta\)-NA mononucleotide) and MNH (1-methylNA) that lack the AD moiety.
Our experimental data shown in Figs. 4 and 12 also do not support this model. As shown in the preexponential factor \(a_{2}\) that refers to the longer lifetime in Table 1 is practically constant and equal to \(a_{2}=0.24\) at methanol concentrations below 70% and rises up to 0.4 at higher concentrations. At the same time the relative ratio of the folded conformation \(N_{fol}\) is known to decrease dramatically with methanol concentration and approaches zero value at about 70-80% MeOH (see also Fig.12 below).
These results supported the explanation of the lifetimes heterogeneity by inherent photoprocesses occurring in the NA chromophore of NADH. We in general agree with the recent suggestion by Blacker et al. who drew attention on the profound role of _cis_ and _trans_ configurations of the NA ring for explanation of the lifetime heterogeneity. However as shown below the nature of the particular lifetime values is still controversial and remains a subject of discussions. As reported by Scott et al. the fluorescence quantum yield of NADH in water is 0.019 which means that the measured fluorescence decay times reflex mostly the non-radiative decay of NADH excited states. Particular routes of this decay in NADH are still unknown. These can be either interactions of excited NADH molecules with surrounding solvent molecules, or intramolecular processes like internal conversion via conical intersections to the ground electronic state, or intersystem crossing to the triplet manifold.
The dependence of the decay times on methanol concentration shown in Fig.3 and on water-methanol solution viscosity in Fig.5 do not support the hydrodynamic origin of the non-radiative relaxation decay mechanism suggested by Blacker et al. According to the hydrodynamics theory the decay time must be proportional to the solution viscosity: \(\tau=k_{nr}^{-1}\sim\eta\), where \(k_{nr}\) is a non-radiative decay rate. However the decay times \(\tau_{1}\) and \(\tau_{2}\) in Fig.3 do not follow the change of the solution viscosity as a function of methanol concentration shown in Fig.5.
We believe that the heterogeneity in the measured decay times is due to different charge distributions in the _cis_ and _trans_ configurations of the NA ring that results in different electrostatic field distributions in these two configurations. The influence of external electric field on the fluorescence decay time was recently investigated by Nakabayashi _et al._ who demonstrated that application of an external electric field increases the probability of non-radiative decay in the NADH excited state and therefore shortens the measured fluorescence decay time. The increase of non-radiative decay rate was attributed to the \(\pi\pi^{*}\) character of the electronic transition of the nicotinamide moiety. Moreover, Nakabayashi _et al._ studied the fluorescence mean decay time in NADH as a function of the solution polarity and reported its significant shortening upon the increase of polarity.
The schematics of the _cis_ and _trans_ nuclear configurations of the NA ring of NADH conformation N11 (see Tables 5 and 7 in _SI_) in aqueous solution are shown in Fig.10. Earlier Wu _et al._ reported the results of _ab initio_ calculations of the ground state NADH in vacuum and concluded that the _cis_ configuration is more planar and stabilized by electrostatic interaction between the negatively charged amide oxygen atom and the positively charged carbonyl atom C2, while the _trans_ configuration is somewhat non-planar and adopts a boat conformation due to the hydrogen bonding between the same oxygen atom and one of the two out-of-plane hydrogen atoms. These geometry changes characterise the ring puckeringeffect.
The results of our calculations of the geometry and charge distributions in the NADH ground and first excited states in aqueous solution presented in Fig.10 and in Tables 5 and 7 in _SI_ are in general agreement with the conclusions made by Wu _et al_ however contain new important information. Bonding and dihedral angle values that describe the deformations of the NA moiety in the conformation 11 of NADH in the ground and first excited states in aqueous solution are collected in Tables 13 and 14 in _SI_.
In brief, the conclusions on the nuclear geometry and the ring puckering in NA moiety in NADH are as follows:
1.
Schematic of the NA ring of NADH conformation N11 in water.
A and B are \(cis\) and \(trans\) conformations, respectively, in the ground electronic state.
C and D are \(cis\) and \(trans\) conformations, respectively, in the relaxed first excited electronic state.
The coordinate frame in the figure bottom refers to the TDM components shown in Table 10 in _SI_.
As can be seen in Table 5 in _SI_ in the _trans_ configuration the effect is much more pronounced, and the torsion angle \(\varphi_{C2N-Nn}\) in some conformations exceeds \(30^{\circ}\). However, in the first approximation the NA moiety can be treated as planar (see the last subsection of this section), because the deviations from planarity are in general not large.
2. The slope of the C3-C7 bond that is characterized by bonding angles in Table 13 in _SI_ differs in the _cis_ and _trans_ configurations by several degrees due to the attraction of the negatively charged amide oxygen atom to the carbonyl atom C2 and an out-of-plane hydrogen atom, respectively.
3. The NA ring deformation (puckering) that is characterized by dihedral angles in Table 14 in _SI_ differs from zero in both _cis_ and _trans_ configurations. The puckering is noticeably more pronounced in the excited state _trans_ configurations.
According to the results of our calculations shown in Tables 8 and 9 in _SI_ vertical excitation energies from the ground to the first excited state do not differ significantly from each other for the _cis_ and _trans_ configurations in various NADH conformations (except in only a few of them). At the same time, as can be seen in Fig.10 and confirmed by the torsion angle \(\varphi_{C2N-Nn}\) values in Table 5 in _SI_ the charge distributions of _cis_ and _trans_ configurations differ from each other significantly. These charge distributions result in different electric field configurations for _cis_ and _trans_ forms that likely leads to the decay time heterogeneity in NADH observed in experiment.
The angle \(\gamma\) between the direction of fluorescence TDM in _cis_ and _trans_ configurations of the conformation 11 and the NA ring plane is shown in fifth column in Table 12 in _SI_. As can be seen in Table 12 the fluorescence TDM is not parallel to the NA plane. Difference between the angles \(\gamma\) related to the _cis_ and _trans_ configurations exceeds \(10^{\circ}\) indicating different fluorescence conditions in _cis_ and _trans_ configurations and supporting the above suggestion.
The internal relaxation mechanism mentioned above that describes the decay time het erogeneity in NADH is assumed to be the major one as it explains the relative stability of the observed fluorescence lifetimes in different conformers and in various solutions. At the same time a slight but noticeable rise of the decay times \(\tau_{1}\) and \(\tau_{2}\) in Fig.3 with the increase of methanol concentration can be treated as a minor effect of the slowing of non-radiative relaxation due to solution polarity decrease.
The preexponential coefficients \(a_{2}\) and \(a_{1}=1-a_{2}\) in eq. describe the relative population of the the _cis_ and _trans_ potential energy wells. These populations reflect the balance between the incoming and outgoing population fluxes in each well. The rotational barrier for _cis_ to _trans_ interconversion was calculated by Wu _et al_. and found to be about 7 kcal/mol that is an order of magnitude higher than vibrational energy at room temperature. Our calculations of NMNH given in the next section also result in comparable rotational barrier heights. The outgoing fluxes is known from experiment being coincide with the decay time values, while sophisticated molecular dynamics calculations are needed for determination of the incoming fluxes.
For qualitative analysis we used a model developed in our recent publication. According to this model the initial excitation laser pulse promotes the molecule from the ground state to the highly excited vibration energy levels of the electronic excited state. The energy of these vibrational states is higher than the rotational barrier separating the _cis_ and _trans_ potential energy wells. The following vibrational relaxation occurring in the picosecond time domain leads to the population of the potential wells depending on the details of the molecular structure and of interactions with solvent molecules. As can be seen in Fig.4 at the methanol concentrations lower than 60% the coefficient \(a_{2}\) was a constant, \(a_{2}\approx 0.25\) within the experimental error bars, and at higher methanol concentrations it grew up to \(a_{2}\approx 0.4\).
However, the quantitative interpretation of the preexponential coefficient dependence in is still a challenging task as it needs several hard and still unresolved problems to be addressed. These are: (i) determination of the ground state population distribution of NADH conformations including their \(cis\) and \(trans\) configurations in various solutions, (ii) determination of the two-photon excitation probabilities for each of these conformations from the ground electronic state to the high-laying vibrational states of the first electronic excited state, and (iii) population of the excited state \(cis\) and \(trans\) potential wells during picosecond vibrational relaxation.
### Lifetime heterogeneity in NMNH
For clarifying the role of the amide group rotation by the dihedral angle \(\varphi\) in NADH in Fig.1A the theoretical study of a more simple NA-containing molecule NMNH has been performed. Full geometry optimization of a number of NMNH conformers in vacuum in the ground and first excited states has been carried out at the B3LYP-D3BJ/6-31G* level. The schematic of NMNH is presented in Fig.1B. An important feature of NMNH is that it lacks the AD moiety, however several authors reported the observation of two, or even three fluorescence lifetimes in NMNH in solutions. Potential energy surfaces (PES) of several NMNH conformations were computed for better understanding of the nature of this effect.
The relaxed PES scans of the amide group torsion angle \(\phi=\varphi+\pi\) in the NMNH ground electronic state in vacuum are shown in Fig.11.
Due to strong hydrogen bonding between the amide oxygen atom and the nearest hydrogen atom from the phosphate group the amide group rotation by the angle \(\varphi\) in Fig.1B is significantly restricted and results in simultaneous change of equilibrium positions of all other nuclei. Therefore the \(cis\) and \(trans\) configurations in NADH cannot be treated independently being closely interrelated with other molecular conformations. Moreover, the calculated rotational potential curves depend on the initial value of the torsion angle \(\varphi\) and have strong interaction between each other at certain angles \(\varphi\). This feature results in a large number of rotational energy curves having one, or several local minima each. Only two of these potential curves are shown in Fig.11 for simplicity. The red curve in Fig.11 is a relaxed PES scan of the torsion angle calculated from the potential minima at \(\phi=-4^{\circ}\)related to \(cis\) molecular geometry. The blue curve in Fig.11 is a relaxed PES scan calculated from the potential minima at \(\phi=156^{\circ}\) related to \(trans\) molecular geometry.
As can be seem in Fig.11\(cis\) and \(trans\) molecular geometries in NMNH are characterised by deep minima separated by a relatively high potential barrier. This founding suggests that the double-minimum potential energy curves in Fig.11 related to \(cis\) and \(trans\) molecular geometries are responsible for two fluorescence lifetimes observed in NMNH in water.
Relaxed PES scans of the amide group torsion angle \(\phi\) in the NMNH ground electronic state in vacuum.
The angle \(\phi=\varphi+\pi\), where \(\varphi\) is the dihedral angle C2N–C3N–C7N–Nn in Fig.1B.
Blue and red potential curves represent two selected NMNH conformations from many others possible.
### Rotational diffusion and the determination of relative conformation concentrations
The rotation diffusion time \(\tau_{r}\) in Fig.5 can be described by the generalized Stokes-Einstein-Debye expression
\[\tau_{r}=fC\frac{\eta V_{M}}{kT}, \tag{12}\]
It is known as the boundary condition parameter (\(C\) = 1 for "stick", \(C\) = 0 for "slip" boundary condition). In general, the factors \(f\) and \(C\) in eq. are strongly interrelated in their effect on solvent-solute friction. The shape factor \(f\) is usually taken in the form \(f=f_{stick}\) that is a well-specified hydrodynamical frictional coefficient for stick boundary conditions that depends only on the shape of the rotating molecule and can be calculated analytically for any symmetric top molecule. The boundary condition parameter \(C\) depends on relative size of the solute compared to the solvent and is accounted for various effects of solvent-solute interaction: hydrodynamic friction, solute molecule shape, complexes formation, and free volume of the solvent. When the dimensions of a solute molecule are sufficiently large compare with a solvent molecule the parameter \(C\) can be presented as \(C=f_{slip}/f_{stick}\), where \(f_{slip}\) is a hydrodynamical frictional coefficient for slip conditions. The values of \(f_{slip}\) related to the limiting case of pure slip conditions were tabulated for any prolate and oblate top molecule by Hu and Zwanzig, however very important intermediate cases between the slip and stick boundary conditions can hardly be treated theoretically. When the dimensions of the solute molecules are much larger than the solvent molecules the parameter \(C\) usually approaches unity that is related to the stick boundary conditions.
Equation is known to be valid at relatively low solvent viscosities of approximately \(\eta\leq 0.1\) Pa\(\cdot\)s. As can be seen in eq. the rotation diffusion time \(\tau_{r}\) is proportional to the solute van der Waals molecular volume \(V_{M}\). The dry volume of NADH was calculated using Edwards increment method and found to be 0.488 nm. A complementary hydration volume \(V_{hyd}\) can be added to this volume for taking account of the hydration effect.
The shape factors \(f_{stick}\) were calculated for several NADH conformations in Table 5 in _SI_ at the stick boundary conditions using eq. in ref. They are presented in Table 2. The calculation was carried out assuming that NADH can be approximated as a prolate ellipsoid with the ratio of the major and minor axes \(R_{max}/R_{min}\).
The conformation No 1 in the first raw in Table 2 is the most compact folded one and the conformation No 12 in the fourth raw is the most extended unfolded one while the conformations No 10 and 11 belong to the intermediate group.
In the conditions of our experiments the signals in Fig.2 were satisfactory approximated by a three-exponential expression in eqs.- containing two isotropic decay times \(\tau_{1}\) and \(\tau_{2}\) in eq. and one rotational diffusion time \(\tau_{r}\) in eq.. A single-exponential anisotropic distribution in eq. in fact represented an unresolved multiexponential anisotropic signal containing contributions from many NADH conformations (see Tables 5 and 6 in \(SI\)), however the relative populations of these conformations are still unknown. Therefore the interpretation of the dependence of the rotational diffusion time \(\tau_{r}\) on methanol concentration in Fig.5 was done by considering a simplified model presenting \(\tau_{r}\) in the form of a sum of
\begin{table}
\begin{tabular}{c c c} \hline Conf.
1 & 1.19 & 1.06 \\ \hline
10 & 1.68 & 1.29 \\ \hline
11 & 2.18 & 1.63 \\ \hline
12 & 2.96 & 2.30 \\ \hline \end{tabular}
\end{table}
Table 2: Axes ratios and the shape factors \(f_{stick}\) for several NADH conformations in Table 5 in _SI_.
contributions from the folded and unfolded conformations (the details are given in _SI_):
\[\frac{1}{\tau_{r}}\approx\frac{1-(1-N_{fol})\chi(\eta)}{\tau_{fol}}+\frac{N_{un} \chi(\eta)}{\tau_{un}}, \tag{13}\]
The term \(\chi(\eta)\) in eq. is the quantum yield of laser induced fluorescence from the first excited state of NADH as a function of viscosity \(\eta\) in the water-methanol solution. It is defined by the expression:
\[\chi(\eta)=\frac{I_{fl}(M)}{I_{fl}(\eta)}, \tag{14}\]
As known\(\chi(\eta)\) decreases monotonously from about 1.6 to 1 with increase of methanol concentration. For analysis, the rotational diffusion times \(\tau_{fol}\) and \(\tau_{un}\) in eq. were calculated from the Stokes-Einstein-Debye formula in eq., where the shape factors \(f_{stick}(fol)\) and \(f_{stick}(un)\) were taken from the first and fourth rows of Table 2, respectively, and the values \(C_{fol}\), \(C_{un}\), \(V_{fol}\), \(V_{un}\) were used as fitting parameters. The concentration \(N_{fol}\) determined from eq. using the values of \(\tau_{r}\) and \(\eta\) in Fig.5 and the values of \(\chi(\eta)\) from ref. at several sets of the parameters \(C_{fol}\), \(C_{un}\), \(V_{fol}\), and \(V_{un}\) that are relevant for our experimental conditions and consistent with the results of previous studies are given in Fig.12.
As can be seen in Fig.12 the relative concentration of the folded conformation \(N_{fol}\) in NADH in pure water was about 0.4 and then decreased down to zero with increasing methanol concentration due to denaturation. This conclusion agrees qualitatively with the results of all other early studies. The distributions of conformations in NADH in water and water-methanol solutions were earlier studied on the basis of the two-states folded-unfolded model by several groups by means of various experimental methods: NMR, energy transfer, and fluorescence lifetime measurements. The advantage of the method used in this paper is that it is based on the study of the anisotropy of excited molecular axes distribution and refers directly to a solute molecule shape. The method does not need a sophisticated model for data extraction from experiment as NMR measurements do. Moreover it allows to avoid a typical problem of optical methods dealing with a necessity to account for non-radiative transitions.
\(N_{fol}\) value in pure water at \(20^{\circ}\) shown in Fig.12 was found to be about \(0.4\pm 0.1\) at all values of the fitting parameters used. This value agrees within experimental error bars with the results reported by Oppenheimer et al., (0.36) and by Heiner et al., (\(0.26\pm 0.06\)). However, our result is somehow larger than that reported by McDonald (0.24) and smaller than that reported by Hull et al. (0.55).
Note that the quantity \(N_{f}\) determined from NMR and energy transfer experiments refs. is not exactly the same as the quantity \(N_{fol}\) determined in this paper. The former deals mostly with the ratio of stacked and unstacked conformations and refers to the extent of chemical interaction between NA and AD rings, while the latter refers mostly to the geometrical properties of NADH conformations. Therefore, some disagreement between the results obtained by the former and the latter methods is possible and should not surprise.
### Determination of the components of the two-photon excitation tensor S.
As can be seen in Figs. 6, 7, and in Table 1 the excitation anisotropy parameters \(r_{l}\), \(r_{c}\), and the parameter \(\Omega\) determined from experiments were found to be almost independent of methanol concentration. Therefore, we assumed that these parameters did not depend on the solution composition and used for analysis their mean values calculated from Table 1 and denoted as \(\bar{r}_{l}\)\(\bar{r}_{c}\), and \(\bar{\Omega}\). As can be seen in Table 1 the parameter mean values are equal to \(\bar{r}_{l}=0.47\)\(\bar{r}_{c}=-0.23\), and \(\bar{\Omega}=0.8\), ane the relationship between the anisotropy parameters follows almost perfectly the formula \(r_{l}=-2r_{c}\). The same relationship was also found in our recent study of two-photon excited fluorescence in indol. According to the model developed in our paper different signs and values of the anisotropies \(r_{l}\) and \(r_{c}\) were due to different types of alignment produced by linearly polarized (LP) and circularly polarized (CP) photons in the molecular axes distribution. In fact, in the case of one-photon excitation the formula \(r_{l}=-2r_{c}\) is exact and should be valid for any molecules. However for two-photon excitation the formula implies a certain relationship between the components of the two-photon excitation tensor \(\mathbf{S}\).
The determination of the components of the two-photon excitation tensor \(\mathbf{S}\) from the experimental data was carried out on the basis of eqs.,. The mean values of the fluorescence parameters \(\bar{r}_{l}\), \(\bar{r}_{c}\), \(\bar{\Omega}\) determined from experiment and the normalized molecular parameters \(M_{K}(R,R^{\prime})\) calculated according to eqs.- are collected in Table 3.
As can be seen in eq. the molecular parameters \(M_{K}(R,R^{\prime})\) depend not only on the two-photon excitation spherical tensor components \(S_{2\gamma}\) in eq., but also on the fluorescence TDM components \(F_{1q_{fl}}\) in eq.. However, by choosing a _fluorescence_ body frame, where axis Z\({}_{fl}\) is parallel to the fluorescence transition dipole moment direction two from three fluorescence TDM components \(F_{1\pm 1}\) in eq. were canceled out and only one component \(F_{10}=F_{z}\) remained non-zero. In this case all M-parameters in eq. were proportional to the term \(|F_{z}|^{2}\) that could be taken outside from the sum and the particular \(M\)-parameter expressions could be written in term of the two-photon excitation tensor components in the form of eqs.- in ref..
The inspection of the calculation data in Tables 10-12 in _SI_ reveals that Z-component of TDM in NADH is an order of magnitude, or more larger than X and Y components. Moreover, as shown in Table 12 the direction of the TDM calculated in the relaxed excited state molecular frame is almost parallel to the direction of the one-photon TDM calculated in the ground state molecular frame. One can conclude that both molecular frames in Tables 10-12 are close to the _fluorescence_ TDM frame described above. Therefore, within these subsection the fluorescence TDM in NADH was assumed to be approximately parallel to the NA ring.
The Cartesian components of the two-photon excitation tensor \(\mathbf{S}\) calculated from the \(M\)-parameters values in Table 3 by solving the systems of equations- and- in our paper are shown in Table 4.
\begin{table}
\begin{tabular}{c c c c c} \hline \(S_{zz}\) & \(S_{xx}\)+\(S_{yy}\) & \(S_{xx}\)–\(S_{yy}\) & \(|S_{xz}^{2}\)+\(S_{yz}^{2}|^{1/2}\) \\ \hline \(T_{T}\mathbf{S}\) & \(T_{T}\mathbf{S}\) & \(T_{T}\mathbf{S}\) & \(T_{T}\mathbf{S}\) \\ \hline
0.95\(\pm\)0.05 & 0.05\(\pm\)0.05 & -0.0\(\pm\)0.04 & 0.44\(\pm\)0.11 \\ \hline \end{tabular}
\end{table}
Table 4: Two-photon excitation parameters.
\begin{table}
\begin{tabular}{c c c c c c} \hline \(\bar{r}_{l}\) & \(\bar{r}_{c}\) & \(\bar{\Omega}\) & \(\frac{M_{0}}{M_{0}}\) & \(\frac{M_{2}}{M_{0}}\) & \(\frac{M_{2}}{M_{0}}\) \\ \hline
0.47 & -0.23 & 0.8 & 1.28\(\pm\)0.06 & 1.86\(\pm\)0.08 & 1.72\(\pm\)0.09 \\ \hline \end{tabular}
\end{table}
Table 3: Molecular parameters determined from experiment.
The normalization factor \(Tr{\bf S}\) in Table 4 is equal to \(Tr{\bf S}=S_{xx}+S_{yy}+S_{zz}\). As can be seen in Table 4 the longitudinal diagonal Cartesian component \(S_{zz}\) of the two-photon excitation tensor dominates over the transversal diagonal components \(S_{xx}+S_{yy}\) indicating that the two-photon excitation is parallel to the direction of the fluorescence TDM. As can be shown from eqs.- in ref. this component is described only by the normalized value of the molecular parameter \(M_{2}\) according to the expression:
\[\frac{S_{zz}}{Tr{\bf S}}=\frac{1}{3}\left[1+\frac{M_{2}}{M_{0}}\right]. \tag{15}\]
As can also be seen in Table 4 another excitation channel described by the off-diagonal two-photon tensor components \(S_{xz}\) and \(S_{yz}\) contributed to the experimental signals along with the diagonal two-photon excitation channel mentioned above. These off-diagonal tensor components have in general complex values and only the combination \(|S_{xz}^{2}+S_{yz}^{2}|\) could be determined from our experiment.
Note that the two-photon tensor components considered in this section describe the molecular excitation by an ultrashort laser pulse at the time \(t=0\). In general they contain contributions from many NADH conformations shown in Tables 5 and 6 in _SI_. The separation of these contributions within a given experimental accuracy is a challenging problem.
## 4 Conclusions
The decay of polarized fluorescence in NADH dissolved in water-methanol solutions under two-photon excitation at 720 nm by femtosecond laser pulses was studied experimentally and theoretically. A number of fluorescence parameters have been determined from experiment using the global fit procedure and then compared with the results reported by other authors. Interpretation of the experimental results obtained were supported by intensive _ab initio_ calculations of the structure of NADH and NMNH in various solutions. The analysis of the results obtained has led to a new explanation of the heterogeneity in the measured decay times in NADH and in NMNH. The explanation ia based on the influence of the internal molecular electric field on the non-radiative decay rates of the excited molecular states. We suggest that different charge distributions in the _cis_ and _trans_ configurations of the nicotinamide ring result in different electrostatic field distributions that lead to the decay time heterogeneity. A slight but noticeable rise of the decay times \(\tau_{1}\) and \(\tau_{2}\) with methanol concentration was observed that was treated as a minor effect of non-radiative relaxation slowing due to solution polarity decrease. As shown, the consideration of only folded and unfolded conformation groups is insufficient for understanding the excited state dynamics in NADH because interaction between the NA and phosphate moieties were sometimes also very important. Determination of the rotational diffusion time in water-methanol solutions as a function of methanol concentration allowed for determination of relative concentrations of the folded and unfolded NADH conformations. Two-photon excitation tensor components have been determined from the experimental data. The analysis of the results obtained suggests the existence of two excitation channels with comparable intensities: the diagonal longitudinal channel that dominated by the two-photon tensor component \(S_{zz}\), where axis Z is in NA ring plane and off-diagonal channel dominated by the tensor components \(|S_{xz}^{2}+S_{yz}^{2}|^{1/2}\).
|
10.48550/arXiv.2008.08686
|
Two-Photon Excited Fluorescence Dynamics in NADH in water-methanol solutions: the Role of Conformational States
|
Ioanna A. Gorbunova, Maxim E. Sasin, Jesus Rubayo-Soneira, Andrey G. Smolin, Oleg S. Vasyutinskii
| 6,255
|
10.48550_arXiv.1907.04082
|
###### Abstract
Increasing the efficiency of materials design and discovery remains a significant challenge, especially given the prohibitively large size of chemical compound space. The use of a chemically transferable coarse-grained model enables different molecular fragments to map to the same bead type, while also reducing computational expense. These properties further increase screening efficiency, as many compounds are screened through the use of a single coarse-grained simulation, effectively reducing the size of chemical compound space. Here, we propose new criteria for the rational design of coarse-grained models that allows for the optimization of their chemical transferability and evaluate the Martini model within this framework. We further investigate the scope of this chemical transferability by parameterizing three Martini-like models, in which the number of bead types ranges from five to sixteen for the different force fields. We then implement a Bayesian approach to determining which chemical groups are more likely to be present on fragments corresponding to specific bead types for each model. We demonstrate that a level of performance and accuracy comparable to Martini can be obtained by using a force field with fewer bead types. However, the advantage of including more bead types is a reduction of uncertainty with respect to back-mapping these bead types to specific chemistries. Just as reducing the size of the coarse-grained particles leads to a finer mapping of conformational space, increasing the number of bead types yields a finer mapping of chemical compound space. Finally, we note that, due to the relatively large size of the chemical fragments that map to a single marini bead, a clear resolution limit arises when using \(\Delta G_{\mathrm{W\to OI}}\) as the only descriptor when coarse-graining chemical compound space.
## I Introduction
Molecular design is a cornerstone of materials science, requiring a fundamental understanding of the relationships between molecular structure and the resulting properties. Traditionally, these structure-property relationships only arise after multiple rounds of screening and discovery of new materials. These cases are examples of direct molecular design, in which the space of all chemical compounds, known as the chemical compound space (CCS), is explored to determine the most suitable chemistry for the target application. Direct molecular design can be interpreted as identifying a hypersurface in the high-dimensional CCS onto a lower dimensional space defined by certain key molecular descriptors that strongly correlate with the desired property. In contrast, inverse molecular design, in which a structure-property relationship is used to infer a suitable chemical structure from a desired property, remains the holy grail of materials science. The main obstacle to achieving this goal is the inability to quickly establish structure-property relationships that can span broad regions of CCS. This is an exceedingly difficult task, given that the size of CCS was estimated to be \(10^{60}\) for drug-like molecules less than 500 Da. Experimentally, this process is inhibited due to both the material and time cost associated with synthesizing and testing a large variety of chemistries that are necessary to infer a relation that is both robust and accurate enough to enable inverse molecular design.
Computationally, recent advancements in processing power and in machine learning have enabled several efficient methods for estimating the electronic properties of a large variety of materials.
A cartoon schematic showing the projection of CCS onto the hydrophobicity descriptor \(\Delta G_{\mathrm{W\to OI}}\), allowing for the creation of top-down chemically-transferable coarse-grained models with a) five, b) nine, c) twelve, and d) sixteen bead types. The number of bead types included in these models defines the degree to which CCS is partitioned on the \(\Delta G_{\mathrm{W\to OI}}\) axis. By varying the number of bead types in each model, we obtain greater insight as to the range of chemistries spanned by a single bead type.
However, there has been relatively little success in applying computational high-throughput screening methods to determine the stability of chemical compounds in soft matter systems for which thermal fluctuations play a critical role. Force field based methods, such as molecular dynamics simulations, are typically used to account for the immense number of configurations that result from thermal fluctuations in these systems. Unfortunately, due to the extensive computational resources required, a high-throughput scheme based on atomistic molecular dynamics simulations is currently unfeasible for spanning the large regions of CCS needed to obtain broadly applicable structure-property relationships.
Coarse-grained molecular dynamics simulations provide a means to significantly reduce the computational expense relative to fully atomistic simulations while still capturing the relevant physical properties. Coarse-grained representations of molecules result from mapping groups of atoms to coarse-grained "pseudo-atoms" or beads. The governing interactions between beads are determined such that the desired properties of the atomistic system are retained. This usually corresponds to a smoothing of the underlying free-energy landscape, allowing for more efficient sampling. Conventionally, coarse-graining is applied to a single molecule with the goal of efficiently sampling a specific system of interest. The coarse-grained potentials are obtained via one of several possible methods (e.g., iterative Boltzmann inversion, force-matching). However these methods are computationally expensive, requiring an initial atomistic simulation that sufficiently explores the underlying free energy landscape of the system of interest. Therefore, adapting coarse-grained molecular dynamics simulations to high-throughput screening of chemical compounds requires flexible yet reliable mapping and force field parameterization methods that do not rely on results from higher resolution simulations for each compound screened.
The coarse-grained Martini force field has become widely used to simulate biological systems as it provides a robust set of transferable force field parameters by constructing biomolecules from a small set of bead types. The Martini model is a top-down model, which maps an atomistic compound or molecular fragment to a coarse-grained site based on its partitioning between aqueous and hydrophobic environments. In the context of molecular design, the main advantage that Martini provides is its chemical transferability. While the force field was explicitly parameterized for a set of specific molecules, a single Martini bead can represent several different chemistries that share similar oil/water partitioning characteristics. In this context, the main feature captured by the Martini model is hydrophobicity, which can act as a key driving force in the physics of soft-matter systems. Rather than running a single atomistic simulation that yields a single data point in CCS, a Martini coarse-grained molecular dynamics simulation provides a representative point in CCS, corresponding to the average behavior of all the chemistries that lay in the region surrounding that point. Thus, high-throughput coarse-grained (HTCG) simulations that use chemically-transferable force fields, such as Martini, are advantageous because they span vast regions of CCS to quickly infer the structure-property relationships and chemical descriptors that can be used to enable inverse molecular design at any resolution. Menichetti et al. recently demonstrated this by running Martini HTCG simulations to construct a structure-property relationship describing the thermodynamics of the insertion of a small organic molecule into a biological membrane across CCS. In doing so, they discovered a linear relationship between the bulk partitioning behavior of the solute and its potential of mean force. They were then able to identify a structure-property hyper surface to obtain membrane permeabilities for these solute molecules. Using the Generated DataBase (GDB), a systematically computer-generated set of organic drug-like compounds, as a proxy for CCS, we then related the regions of this surface to regions of CCS that were dominated by specific chemical moieties, enabling inverse molecular design of small molecules given a desired permeability. The question remains: how representative of CCS is the Martini force field? Given that Martini was designed to reproduce the partitioning behavior of certain solvents as well as the properties of lipid-bilayer membranes, is there a way to accurately parameterize a transferable coarse-grained force field with the goal of optimizing its coverage of CCS? In the context of high-throughput coarse-grained simulations that use Martini, creating a structure-property relationship that enables inverse design requires an understanding of the chemistry that is representative of a specific bead type. The metric used in assigning specific chemical fragments to Martini bead types is the water/octanol partition free energy (\(\Delta G_{\mathrm{W\to OI}}\)). Therefore, an intuition for which chemistry maps to a given bead type can only be obtained by understanding how \(\Delta G_{\mathrm{W\to OI}}\) varies as a function of chemistry. Given that the number of heavy (non-hydrogen) atoms that usually map to a Martini bead ranges from three to five, we can think of each bead as representing a small carbon scaffold perturbed to some degree by either replacing carbons with other heavy atom types (e.g., oxygen, nitrogen, or fluorine) or by replacing single bonds with double or triple bonds. We define a functional group as being one or a localized combination of these types of perturbations.
In this work, we quantify the information loss that occurs when a top-down coarse-grained model, like Martini, is used to reduce the resolution of CCS. Additionally, we parameterize three sets of coarse-grained force fields in the Martini framework. In this context, we use the terms "force field" and "model" interchangeably, defined as a set of parameters which describe the interactions between a fixed number of coarse-grained representations called bead types. Each force field developed in this work consists of five, nine, and sixteen neutral bead types, as well as two extra types to account for hydrogen bond donors and acceptors. We observe that Martini does not provide the most efficient reduction of CCS. We show that the nine-bead force field reduces CCS to the same degree as Martini despite having three fewer bead types, and that further increasing the number of bead types yields negligible improvements in the performance of the model. The models are validated by performing coarse-grained simulations to calculate the water/octanol partition free energies of approximately 500 compounds for which experimental data is available. Finally, we demonstrate that the main advantage of a force field with a large number of bead types is the reduction of uncertainty when back-mapping these coarse-grained representations to real chemical functional groups. Just as decreasing the resolution of the CG mapping reduces the resolution of the potential energy landscape, a reduction in the number of bead types of a chemically transferable CG force field allows for an increased degeneracy of chemical fragments that map to a single bead type, illustrated in Ideally, a well-designed chemically transferable CG force field would contain some number of bead types that can be intuitively back-mapped to single chemical functional groups. However, the size of a single functional group is small relative to the size of a Martini bead, such that many functional groups could be identified within a fragment mapping to a single Martini bead. Here, we demonstrate that this mismatch between the size of a Martini bead and a single functional group requires additional constraints in order to identify the unique chemistry that maps to each bead type. Incorporating these constraints into a Bayesian formalism yields probabilities of specific chemistries mapping to a given bead type, further promoting inverse molecular design. However, even these additional constraints allow for the same functional groups to be present in multiple bead types, indicating a natural resolution limit when using \(\Delta G_{\text{W}\to\text{OI}}\) as the sole basis for a chemically-transferable, top-down coarse-grained model.
## Methods
### The Auto-Martini Algorithm
This work relies on the auto-martini algorithm initially developed by Bereau and Kremer. The algorithm first determines an optimal mapping for an organic small molecule. The mapping provides the number of coarse-grained beads used to represent the molecule as well as their placement. A mapping cost function is minimized for each molecule so as to optimize both the number and placement of beads used in its coarse-grained representation. The assignment of coarse-grained potentials to each bead (bead-typing) occurs by assigning an existing Martini bead type that has the closest matched water/octanol partition free energy (\(\Delta G_{\text{W}\to\text{OI}}\)) with that of the molecular fragment encapsulated by the bead. The partition coefficients of these fragments are obtained by using alogps, a neural network algorithm that predicts these values given the chemical structure of the fragment. In this work, we use an updated version of the auto-martini algorithm that has three significant changes from the previous version. The first change is an increased energetic penalty for "lonely" atoms (i.e., atoms that fall outside the Van der Waals radius of the placed coarse-grained beads). The second change is a reduction of the multiplicative factor used when assigning bead types to rings for both five and six-membered rings. Finally, the cutoff value for the \(\Delta G_{\text{W}\to\text{OI}}\) for the assignment of donor and acceptor fragments to their corresponding bead types was modified such that the CG and atomistic population distributions more closely matched. All of these changes increased the algorithm's accuracy, which is quantified in the supporting information. Using the refined auto-martini algorithm, approximately 3.5 million molecules with ten heavy atoms or less that make up the GDB were mapped to coarse-grained representations for four different force fields. The molecules contain carbon, nitrogen, oxygen, fluorine, and hydrogen atoms only. Of these 3.5 million compounds, approximately 340,000 were successfully mapped to both coarse-grained unimers (1 bead representations) and dimers (2 bead representations) for all of the force fields described in this work. The majority of the remaining compounds were mapped to coarse-grained representations with a higher number of beads, and a small fraction of compounds were unable to be successfully mapped by the algorithm. Histograms comparing the distributions of \(\Delta G_{\text{W}\to\text{OI}}\) for each set of atomistic compounds mapping to CG unimers and dimers and their CG counterparts were constructed using the numpy histogram function, with the number of bins equal to 1000 and 1050 for unimers and dimers, respectively. These histograms are shown in Fig. 2a-d for Martini, while the other histograms can be found in the SI.
### The Jensen-Shannon Divergence
In this work, the main tool used to quantify information loss when going from atomistic to coarse-grained resolution is the relative entropy in the form of a Jensen-Shannon divergence (JSD). The relative entropy framework has been previously established as a useful tool for evaluating the quality of coarse-grained models. The JSD is a variation of the well-known Kullback-Leibler divergence used to calculate the relative entropy between two distributions. It offers two advantages over the Kullback-Leibler divergence in that it is symmetric and always has a finite value.
\[D_{JS}=\frac{1}{2}D_{KL}\left(P_{CG}||P_{avg}\right)+\frac{1}{2}D_{KL}\left(P _{AA}||P_{avg}\right), \tag{1}\]
\[\text{where }D_{KL}(A||B)=\sum_{i=1}^{N}a_{i}\ln\left(\frac{a_{i}}{b_{i}} \right),\]
\[\text{and }P_{avg}=\frac{1}{2}(P_{CG}+P_{AA}).\]
Here, we use the JSD to evaluate how well the distribution of the water/octanol partition free energies for the coarse-grained molecules (\(P_{CG}\)) match the corresponding distribution at the atomistic resolution (\(P_{AA}\)). A value of 0 indicates that the two distributions are the same. The use of the average distribution (\(P_{avg}\)) conveniently prevents divisions by zero when comparing histograms like those shown in Fig. 2a-d.
### Basin Hopping and Minimization Schemes
In this work, we use multiple methods to optimize the coarse-grained partition free energies to best match the atomistic distribution of free energies. The first such method is the basin-hopping method, which is a variation of Metropolis-Hastings Monte Carlo. The algorithm proceeds in the following steps. Given a set of initial coordinates and objective function, the initial coordinates are first randomly perturbed and subsequently minimized. The results of the minimization are either accepted or rejected based on a predefined Metropolis criterion. These two steps form a single iteration of the algorithm, and a large number of iterations may be required to find the desired minima. Here, we use the JSD as our objective function and a set of possible water/octanol partition free energies for each coarse-grained bead type as our initial coordinates. Each move then corresponds to shifting the values of \(\Delta G_{\mathrm{W\to QI}}\) for each coarse-grained bead type in a given force field. The optimizations were performed in order to define the desired \(\Delta G_{\mathrm{W\to QI}}\) values for the five-bead-type force field, using the basinhopping function provided by scipy with a Broyden-Fletcher-Goldfarb-Shanno local minimizer, a Metropolis temperature parameter of 0.008, and a step size of 0.024 kcal/mol. For the reference atomistic distribution, we applied the alogps neural network to predict \(\Delta G_{\mathrm{W\to QI}}\) for molecules in the GDB, restricting the maximum number of heavy atoms per molecule to eight. However, finding the optimal set of \(\Delta G_{\mathrm{W\to QI}}\) values for the sixteen-bead-type force field using this approach proved to be computationally unfeasible, as the dimensionality of the problem scales with \(M^{N}\), where \(N\) is the total number of bead types in the force field and \(M\) is the range of \(\Delta G_{\mathrm{W\to QI}}\) values spanned by the Martini bead types divided by the step size. To parameterize the sixteen-bead-type force field, we used the scipy minimize function with the modified Powell method, starting with an initial set of eighteen bead types that were evenly distributed along the \(\Delta G_{\mathrm{W\to QI}}\) axis. The results of the minimization indicated two sets of two bead types that were within 0.1 kcal/mol of each other, and so each pair was combined into a single bead type, resulting in sixteen bead types total.
### Clustering the GDB
In addition to optimization of the JSD, a new set of coarse grained water/octanol partition free energies was also proposed by clustering the GDB, leading to the 9-bead-type force field. Specifically, all molecules with eight heavy atoms or less that were known to map to single bead representations using the auto-martini algorithm were grouped based on the number and type of hetero-atom substitutions present in the molecule (i.e., the number of times that a C was replaced with N, O, or F). The resulting atomistic molecular populations as well as the mean and standard deviation of their water/octanol partition free energies are shown in Detailed information on each of the distributions (beyond what is provided in Fig. 3) is available in the SI. The desired water/octanol partition free energies are determined by clustering the points on this graph, starting from the highest populated points and accepting anything that was within plus or minus 0.5 kcal/mol of these points. For example, the first point with the highest population in is chosen as a starting point for the first bead type. All points that fall within 0.5 kcal/mol are assigned to this bead type and the \(\Delta G_{\mathrm{W\to QI}}\) is determined by taking a population-weighted average of all of these points. The next bead type is determined by selecting the highest point on that is not already assigned to a bead type and repeating the process.
### Functional Group Analysis
A statistical analysis of the functional groups found in the molecular fragments mapping to single beads is necessary in order to obtain a more detailed picture as to which chemistries are representative of specific bead types. The enumeration of functional groups was achieved through the use of the checkmol software developed by Haider. This software uses the 3D coordinates of each atom and the corresponding atom labels in a given molecule to identify common chemical functional groups. A full list of the functional groups identified can be found in the SI. Using checkmol, we determine the degeneracy of specific functional group pairs with respect to single bead types for the set of molecular fragments that mapped to a single bead. This amounts to counting the number of fragments containing a specific functional group pair and mapping to a single bead type. This population is then normalized with respect to the total number of fragments containing that same functional group pair across all bead types. It is useful to frame this statistical analysis in terms of conditional probabilities, as this yields specific information relevant for molecular design applications. For example, the aforementioned counting and normalization is equivalent to calculating the likelihood of assigning a bead type (\(T\)) given a specific functional group pair (\(F\)), defined as \(P(T|F)\). We use the fragment population distributions for each bead type and each functional group pair to obtain probabilities \(P(T)\) of a bead type and \(P(F)\) of a functional group pair.
\[P(F|T)=\frac{P(T|F)P(F)}{P(T)}. \tag{2}\]
The results are shown as a series of heat maps for each force field in
### Parameterization of New Bead Types
The new force fields share most of the parameters defined by the Martini force field. For the intra-molecular interactions, bonded, angle, and dihedral force constants remain the same as those prescribed by Martini. The non-bonded interactions only contain one deviation from those in Martini. We linearly interpolate across the interaction matrix defined in Martini, utilizing the distance between the established Martini \(\Delta G_{\text{W}\to 0\text{I}}\) and the desired \(\Delta G_{\text{W}\to 0\text{I}}\) for the interpolation. The partition free energies of each bead type was then confirmed by running coarse-grained molecular dynamics simulations of single beads of each new bead type. These results are included in the SI, and show that this method yields an accurate force field without relying on an iterative scheme. Linear interpolation is chosen as it is clear that there is no underlying functional form or smooth landscape that can be derived from this parameter space (see SI for details). The new bead types are named as T\(i\) types, with \(i\) ranging from 1 to \(N\) where \(N\) is the total number of bead types in the force field. The numbering is also ordered by polarity. For example, the T1 bead type for all new force fields is the most polar type. Conversely, the T5, T9, and T16 bead types are the most apolar bead types in the five, nine, and sixteen-bead-type force fields, respectively. The full list of bead types for each force field, their force field parameters, and their corresponding \(\Delta G_{\text{W}\to 0\text{I}}\) values is available in the SI.
### Coarse-Grained Simulations
Coarse-grained molecular dynamics simulations were performed in GROMACS version 4.6.6 using the standard Martini force field parameters as well as the new force field parameters derived in this work. A time step of \(\delta t=0,03\)\(\tau\) was used for all simulations, where \(\tau\) is the natural time unit for the propagation of the model defined in terms of the units of energy \(\mathcal{E}\), mass \(\mathcal{M}\) and length \(\mathcal{L}\) as \(\tau=\mathcal{L}\sqrt{\mathcal{M}/\mathcal{E}}\). The simulations were run in an \(NPT\) ensemble with a Langevin thermostat and Andersen barostat to keep the temperature and pressure at 300 K and 1 bar, respectively. The corresponding coupling constants were \(\tau_{T}=\tau\) and \(\tau_{P}=12\tau\).
Water/octanol partition free energies were obtained by simulating approximately 500 coarse-grained molecules in octanol and water. Approximately 250 octanol molecules and 350 Martini water molecules were simulated for their respective systems, with the appropriate number of antifreeze particles. The free energies were computed using the Bennett acceptance ratio method in which the coarse-grained solute was incrementally decoupled from the solvent via the coupling parameter, \(\lambda\). Twenty-one simulations were run for each molecule at evenly spaced \(\lambda\) values ranging from 0 to 1, with each simulation run for 200,000 time steps. Finally, the partition free energies were calculated using the relation \(\Delta G_{\text{W}\to 0\text{I}}=\Delta G_{\text{W}}-\Delta G_{0\text{I}}\).
## Results
### Quantifying information loss of CG models with varying number of bead types
The updated auto-martini algorithm was used to first map and subsequently assign bead types to 3.5 million molecules of the GDB containing ten or fewer heavy atoms using the Martini force field as well as the other three force fields parameterized by interpolating the Martini interaction matrix. shows a comparison of the atomistic and coarse-grained \(\Delta G_{\text{W}\to 0\text{I}}\) distributions for molecules mapping to Martini miners (Fig. 2a,b) and dimers (Fig. 2c,d). The corresponding histograms for the other three force fields, as well as a histogram constructed using the Martini force field but with the original auto-martini algorithm can be found in the SI. The width of the coarse-grained bars reflects the range of \(\Delta G_{\text{W}\to 0\text{I}}\) values within which a molecule must fall in order to be assigned that bead type, or, in the dimer case, a combination of bead types. The height of the bars is set such that the area covered by each bar is equal to the total number of molecules that were assigned that coarse-grained representation. We then calculate the JSD between the coarse-grained and atomistic histograms for each force field to quantify the information loss as a function of the number of bead types present in each force field. Increasing the number of bead types reduces the information loss when going from atomistic to CG resolution, though this reduction becomes insignificant after reaching nine bead types. The JSD comparing the unimer histograms (red curve in Fig. 2e) changes negligibly when increasing the number of bead types from nine to sixteen, with only a small increase for the Martini case (12 bead types). This is expected due to the fact that the atomistic histogram of GDB molecules mapping to a single bead is a simple, unimodal distribution with a peak at \(\Delta G_{\text{W}\to 0\text{I}}=0\). Since all of the force fields have at least one amphiphilic bead type with a \(\Delta G_{\text{W}\to 0\text{I}}\) close to 0, they all capture this defining feature of the histogram, and, comparatively, further information gains are negligible. However, the JSDs calculated from the dimer histograms (blue curve in Fig. 2e) show a variety of interesting features. Both the nine and sixteen-bead-type force fields maintain roughly the same JSD, suggesting that the combinatorial explosion that results from doubling the molecular weight is captured by these force fields. The slight increase seen in the unimer JSD for Martini is noticeably amplified for the dimer case, indicating that careful placement of bead types on the \(\Delta G_{\text{W}\to 0\text{I}}\) axis is necessary to maximize chemical transferability. Surprisingly, the greatest deviation in the JSD going from the unimer to dimer histogram comes from the five-bead-type force field, dropping well below the values for the higher bead type force fields. The reason for this can be seen in Fig. S3b, which shows that the distribution of atomistic compounds mapping to dimers in the five-bead-type force field is significantly different from its analogs for the other force fields. This indicates that a significant number of molecules that would map to dimers when using one of the other force fields are mapped to trimers or tetramers using the 5-bead-type force field.
### Relating chemistry to bead types
As an alternative to purely numerical methods for determining the optimal \(\Delta G_{\text{W}\to 01}\) values for the bead types of a CG force field that best partitions CCS, we cluster the GDB itself and use the weighted average of \(\Delta G_{\text{W}\to 01}\) for each cluster. shows the two descriptors upon which we project and subsequently cluster the GDB. Each point in represents the set of molecules in the GDB that have a specific number and type of heavy atom substitutions (i.e., N, O, or F). The points are placed on the \(\Delta G_{\text{W}\to 01}\) axis according to the average of their \(\Delta G_{\text{W}\to 01}\) distribution. The error bars represent the standard deviation of the \(\Delta G_{\text{W}\to 01}\) in each distribution. One of the corresponding distributions is shown in Interestingly, all of the distributions with populations of over 1000 molecules are unimodal. The points are clustered hierarchically with respect to population and average as shown in The highest-populated points are all chosen as cluster centers as long as they are separated by at least 0.5 kcal/mol, which is an arbitrarily chosen length-scale for the clustering to ensure a reasonable number of bead types in the final force field. After the points are clustered, the desired \(\Delta G_{\text{W}\to 01}\) of each bead type is determined by taking the population-weighted average of all the points in a cluster. This intuitively provides a basic understanding of the chemistry that maps to a specific bead type. For example, a T4 bead is more likely to back-map to a molecule with one N and one O substitution compared to two N substitutions because of the difference in the GDB populations of each molecule type.
It is important to characterize the degree to which unique chemistries are captured by the bead types of each force field. Using the GDB as a proxy for CCS enables a quantitative understanding of the chemical transferability of each bead type through the calculation of conditional probabilities. shows a series of heat maps corresponding to each of the four force fields investigated in this work. These heat maps are constructed by counting all fragments containing only five heavy atoms and assigned to a specific bead type, such that two functional groups are detected by the checkerloh software package. The fragment population distributions are then used to calculate the Bayesian likelihood \(P(T|F)\) and posterior \(P(F|T)\) for each bead type/functional pair combination in every force field. The numbers on the horizontal axis for each heat map denote specific pairs of functional groups found in the chemical fragments that are assigned to a bead type, while the color corresponds to either the likelihood or posterior probabilities. We see the localization of functional group pairs to specific bead types mainly because of the constraint of only including fragments with five heavy atoms. This constraint limits the combinatorics of hetero-atom and bond substitutions that result in functional group pairs. Despite the addition of these constraints, a large number of functional group pairs are still split across multiple bead types. The corresponding heat maps constructed using four-heavy-atom-fragments only are included in the SI and show far less degeneracy of functional group pairs across bead types compared to these heat maps, although the general trends observed are the same.
Histograms of 343,700 small molecules extracted from GDB that map onto one-bead or two-bead coarse-grained Martini representations. (a),(c) Coarse-grained and (b),(d) atomistic populations as a function of water/octanol partitioning free energy. The width of the bars in (a),(c) corresponds to the range of atomistic water/octanol partitioning free energies that can map to that coarse-grained representation. (e) Jensen-Shannon divergence calculated for the histograms corresponding to those in (a)-(d) for all force fields described in this work.
4, displaying the average number of functional group pairs per bead type, as well as the number of likelihood and posterior values above cutoff values of 0.99 and 0.2, respectively. As the number of bead types increases, both the average number of functional group pairs per bead type and the number of likelihood values greater than 0.99 decrease, indicating that fewer bead types in a force field increases the coverage of CCS for each bead type. The opposite trend is observed for the number of posteriors greater than 0.2, indicating that more bead types result in higher chemical specificity for each bead type.
### CG force field validation
While we have demonstrated that the careful placement of bead types on the \(\Delta G_{\mathrm{W\to O1}}\) axis leads to more chemical transferability, the force fields themselves must be validated. Because \(\Delta G_{\mathrm{W\to O1}}\) was used as the target property for the interpolation of the Martini interaction matrix, we must ensure that this property is indeed captured by the resulting models and determine to what extent the accuracy of these models changes as the number of bead types increases. shows correlation plots comparing \(\Delta G_{\mathrm{W\to O1}}\) values computed from coarse-grained MD simulations with experimental values for approximately 500 ring-less molecules obtained from the National Cancer Institute database. The comparison is made for all four of the models examined in this work. The number of compounds varies for each model, as the auto-martini algorithm was able to successfully find mappings for more molecules in the database when using a model with a higher number of bead types, ranging from 479 compounds mapped when using the five-bead-type model to 505 when using the sixteen-bead-type model. The full set of compounds as well as their corresponding coarse-grained representations is provided in the SI. The vertical series of points prominently seen in are a consequence of the increased degeneracy of CCS for the 5-bead-type model: they represent many compounds mapping to the same coarse-grained representation. As expected, the correlation becomes less discretized as the number of bead types increases. Examining Figs. 5e and 5f, we see corresponding gains and losses in the Pearson correlation coefficients and MAEs, respectively. Surprisingly, the gains in accuracy are very slight as a function of number of bead types--with the correlation coefficient only increasing by 0.01 and the MAE decreasing by 0.2 kcal/mol--despite tripling the number of bead types. Even with the five-bead-type model, we achieve an MAE of 0.8 kcal/mol, within the standard for chemical accuracy.
## Discussion
Given the immense size of CCS, the creation of reduced models that efficiently subdivide the space is necessary for
\begin{table}
\begin{tabular}{|c|c|c|c|} \hline \# of & Avg. \# of & \# of & \# of \\ Bead Types & Func. Group Pairs & Likelihoods \textgreater{} 0.99 & Posteriors \textgreater{} 0.20 \\ \hline
5 & 16.4 & 40 & 8 \\
9 & 10.1 & 33 & 17 \\
12 (Martini) & 7.4 & 35 & 20 \\
16 & 6.4 & 27 & 26 \\ \hline \end{tabular}
\end{table}
Table 1: For each force field, the number of bead types, the average number of functional group pairs per bead type, the total number of likelihood values over 0.99, and the total number of posterior values over 0.2.
(a): Population versus average values of the distributions of water/octanol partition free energies corresponding to GDB molecules with up to 8 heavy atoms and a specific number and type of hetero-atom substitutions. The error bars refer to the standard deviations of each distribution. The colored backgrounds denote how these average values are clustered to obtain new bead types that more efficiently divide CCS. (b): Example distribution corresponding to top-most point labeled in the graph on the left. The label refers to the number and type of hetero-atom substitution, in this case 1 nitrogen and 1 oxygen substitution.
Heat maps portraying the degeneracy of specific pairs of functional groups for a given bead type for force fields containing five (a), nine (b), twelve (c), or sixteen (d) bead types. The horizontal axes denote specific functional group pairs that exist in a chemical fragment with five heavy atoms only. The color corresponds to either the column-normalized or row-normalized probabilities. The column-normalized probabilities (left side) are equivalent to the Bayesian likelihood of a given functional group mapping to a specific bead type. The row-normalized heat maps (right side) show the Bayesian posterior probabilities of obtaining a specific functional group given a bead type.
Here we demonstrate the use of the water/octanol partition free energy as the parameter used to generate top-down chemically-transferable coarse-grained models of varying numbers of bead types. This choice of descriptor is inspired by the Martini force field, which prescribes the use of \(\Delta G_{\mathrm{W\to Ol}}\) when determining the bead type to be used to represent a molecular fragment. Here, we use the GDB as a proxy for CCS and apply the auto-martini algorithm to compare the populations of the GDB molecules and their corresponding CG representations for four different force fields with varying numbers of bead types. This effectively amounts to a discretization of CCS projected onto \(\Delta G_{\mathrm{W\to Ol}}\) at multiple resolutions. quantifies the level of information loss using the JSD as the resolution is varied, allowing us to determine how effectively each of these force fields, including Martini, represents CCS. The JSD decreases as the number of bead types increase. However, the information retention becomes negligibly greater, essentially plateauing after 9 bead types. Remarkably, despite the fact that the Martini force field was parameterized using a small number of chemical compounds (relative to the large distribution of compounds used to parameterize the other models in this work), it shows only a minuscule increase in the JSD. This is mainly due to the inefficient placement of the P3, P4, and P5 beads within close proximity to each other on the \(\Delta G_{\mathrm{W\to Ol}}\) axis. Unfortunately, this increase in the JSD is amplified when comparing the \(\Delta G_{\mathrm{W\to Ol}}\) distributions for dimer molecules, whereas for the 9 and 16 bead type models, the JSD seems to converge. The combinatorial explosion that results from doubling the size of molecules (i.e., going from unimer to dimer) is reflected in these histograms as a broadening of the total distribution, since more hydrophobic and hydrophilic values of \(\Delta G_{\mathrm{W\to Ol}}\) are possible as molecule size increases. shows that the 9 and 16 bead type force fields match this combinatorial explosion.
On the other hand, Figs. 5e and f clearly demonstrate that a high level of accuracy is already achieved with respect to \(\Delta G_{\mathrm{W\to Ol}}\) using the five-bead-type force field. What, then, is the benefit to using a model with more than five bead types? As we show from Figs. 3 and 4, the main advantage is in back-mapping the coarse-grained representations to their likely atomistic counterparts.
Correlation curves comparing the \(\Delta G_{\mathrm{W\to Ol}}\) calculated from coarse-grained MD simulations of approximately 500 molecules to their measured values from experiment. The results are shown for each of the force fields described in this work (a)-(d) as well as their respective Pearson correlation coefficients and mean absolute error (MAE) values (e)-(f).
The distributions that were clustered to make this force field provide a method for predicting the chemistries that are most representative of a bead type. Since the standard deviations of these distributions are so large, such that some span across three different bead types, this provides only a rough idea of the probable chemistry accessible to a bead type. Moreover, knowledge of the presence of one or two heavy-atom substitutions on a carbon scaffold of up to 8 heavy atoms is insufficient for back-mapping given the number of ways in which they can be arranged on that scaffold resulting in wildly different chemical properties. shows how different functional group pairs will map clearly to specific bead types when the scaffold size is reduced to five heavy atoms. This extra constraint enables a clearer understanding of the range of unique chemistries that are accessible to a specific bead type. Decreasing the size of the scaffold from five to four heavy atoms yields correspondingly narrower distributions of \(\Delta G_{\text{W}\to\text{Ol}}\), meaning that the same functional group pair can be found in fewer bead types.
Table 1 also demonstrates that the number of unique functional group pairs that map to a given bead type decreases as the number of bead types increases, to the point where, for Martini as well as the 16 bead type force field, there exist bead types that essentially back-map to a single functional group pair. Here, we see a clear parallel with structural coarse-graining methods: just as decreasing the size of the beads leads to a finer mapping of the configurational space, increasing the number of bead types leads to a finer mapping of CCS. The efficiency of a CG model can be optimized by tuning the mapping function and bead size of a CG model such that the accuracy of the model is balanced with respect to the computational cost of simulating a greater number of particles. By fixing the geometric mapping method and bead size, and only varying the number of bead types possible, we instead balance between the accuracy of representing specific chemical features and the cost of parameterization and validation of the inter-particle potentials. We circumvent this cost by interpolating the Martini interaction matrix to obtain accurate parameters for all of the force fields presented in this work. However, this cost will be significant for models requiring a more rigorous parameterization scheme relying on other molecular descriptors. Separate from this trade-off between accuracy and parameterization cost, a "back-mapping efficiency" can be defined as the average number of functional group pairs that map to a single bead type, indicating a larger region of CCS being captured by a single bead type. Unsurprisingly, Table 1 shows that the five-bead-type force field clearly has the highest back-mapping efficiency.
This statistical analysis of functional group pairs also suggests a Bayesian approach to computing the probability of a functional group pair, \(F\), given a bead type, \(T\), represented as \(P(F|T)\) in equation 2. \(P(F)\), the Bayesian prior, is the probability of finding the specific functional group pair in the set of molecular fragments (made up of five heavy atoms and containing two functional groups as defined by checkmol) that mapped to single beads, and \(P(T)\) is the probability of choosing the given bead type from that same data set. The likelihood, \(P(T|F)\), shown in the left side of prescribes the bead type or types to which a fragment could be assigned based on its chemistry--the equivalent of the Martini "bible" for assigning bead types. As shown in Table 1, the number of functional group pairs with likelihoods greater than 0.99 (essentially localized to a single bead type) decreases as the number of bead types increases. The Martini force field deviates slightly from this trend, with two more functional group pairs with high likelihoods as compared to the nine-bead-type force field. This may stem from the parameterization strategy used for Martini that relied on specific molecules and their functional groups rather than aiming to efficiently span chemical space by optimizing the JSD, as proposed in this work. The posterior probabilities, which provide a quantitative description of which chemistries are more representative of each bead type, increase as the number of bead types increases. This effect more easily facilitates the back-mapping of coarse-grained representations. These two quantities, the Bayesian likelihood and posterior, are essential for further exploring CCS covered by specific bead types and enabling both direct and inverse molecular design.
Interestingly, we immediately see a resolution limit with respect to the functional group pairs that map to specific bead types. Because there are certain length scales on the \(\Delta G_{\text{W}\to\text{Ol}}\) axis that correspond to the distribution of specific functional group pairs, increasing the number of bead types will naturally split these distributions, such that one functional group pair is represented in multiple bead types. shows that the majority of functional group pairs are encompassed either by a single bead type or one of its neighbors on the \(\Delta G_{\text{W}\to\text{Ol}}\) axis. Increasing the number of bead types causes these splits to become more exacerbated, spanning multiple bead types for an increased number of functional group pairs. This is the resolution limit of this type of top-down coarse-graining. The large bead sizes of these models leads to a high degree of variability in the chemistry, meaning that it is no longer obvious which functional group/functional group pair belongs to which bead type. The limit is most evident for the functional group pairs mapping to the T3 and T13 bead types in Fig. 4d, indicating that they are placed too close to their neighbors on the \(\Delta G_{\text{W}\to\text{Ol}}\) axis. These functional group pairs contain some combination of the following functional groups: alkene, alkyne, enamine, hydrazine, hydroxylamine, carboxylic acid derivatives, and fluorine substitution. The placement of these functional groups within a five-carbon scaffold will drastically shift the \(\Delta G_{\text{W}\to\text{Ol}}\) beyond the range of the next nearest bead type on the \(\Delta G_{\text{W}\to\text{Ol}}\) axis and highlights the limitations of only using this single descriptor for the projection of CCS. However, determining a suitable orthogonal descriptor and then parameterizing a chemically transferable CG force field to achieve a more direct relation with CCS is outside the scope of this work, and will be addressed subsequently.
## Conclusion
In this work, we use the JSD to quantify the information loss in chemically transferable top-down coarse-grained models with varying numbers of bead types, with the GDB as our proxy for chemical compound space (CCS). We find that Martini, while not designed to efficiently reduce CCS, only performs slightly worse than the other force-fields explicitly designed to minimize the JSD. All force fields yield roughly the same level of accuracy with respect to \(\Delta G_{\text{W}\to\text{OI}}\), but vary greatly in their coverage of CCS. We used a Bayesian approach to calculate the probabilities of back-mapping given bead-types to fragments containing specific chemical substitutions. Here, we found it necessary to constrain the size of chemical fragments to five heavy atoms and the presence of two functional groups in order to clearly differentiate between the chemical moieties mapping to each bead type. The results of this Bayesian analysis indicate that increasing the number of bead types decreases the range of accessible chemistry while increasing the corresponding posterior probabilities for each chemistry. However, there is a resolution limit when using this approach, as it does not take into account the specific positions of hetero-atom and bond substitutions within a fragment, causing different bead types to appear representative of the same chemistry. Martini, as well as other chemically-transferable coarse-grained models, can be used to quickly build structure-property relationships that span broad regions of CCS. Here we highlight the powerful combination of this method with Bayesian inference, providing an informed mapping of a coarse structure-property relationship to a higher resolution in chemical compound space and further enabling inverse molecular design.
|
10.48550/arXiv.1907.04082
|
Resolution limit of data-driven coarse-grained models spanning chemical space
|
Kiran H. Kanekal, Tristan Bereau
| 3,220
|
10.48550_arXiv.2108.02461
|
###### Abstract
Implicit solvation is an effective, highly coarse-grained approach in atomic-scale simulations to account for a surrounding liquid electrolyte on the level of a continuous polarizable medium. Originating in molecular chemistry with finite solutes, implicit solvation techniques are now increasingly used in the context of first-principles modeling of electrochemistry and electrocatalysis at extended (often metallic) electrodes. The prevalent ansatz to model the latter electrodes and the reactive surface chemistry at them through slabs in periodic boundary condition supercells brings its specific challenges. Foremost this concerns the difficulty to describe the entire double layer forming at the electrified solid-liquid interface (SLI) within supercell sizes tractableby commonly employed density-functional theory (DFT). We review liquid solvation methodology from this specific application angle, highlighting in particular its use in the widespread _ab initio_ thermodynamics approach to surface catalysis. Notably, implicit solvation can be employed to mimic a polarization of the electrode's electronic density under the applied potential and the concomitant capacitive charging of the entire double layer beyond the limitations of the employed DFT supercell. Most critical for continuing advances of this effective methodology for the SLI context is the lack of pertinent (experimental or high-level theoretical) reference data needed for parametrization.
###### Contents
* 1 Introduction
* 2 Fundamentals of implicit solvation
* 2.1 Coarse-graining the electrolyte
* 2.2 Separation of the grand potential energy functional
* 2.3 Electrostatics of solvation
* 2.3.1 Potential energy and polarization models
* 2.3.2 Dielectric function
* 2.4 Non-electrostatics of solvation
* 2.4.1 Cavitation grand potential, \(\Omega_{\mathrm{is}}^{\mathrm{cav}}\)
* 2.4.2 Exchange repulsion, \(G_{\mathrm{is}}^{\mathrm{rep}}\)
* 2.4.3 Dispersion interactions, \(G_{\mathrm{is}}^{\mathrm{dis}}\)
* 2.4.4 Thermal motion, \(G_{\mathrm{is}}^{\mathrm{tm}}\)
* 2.5 Electrolyte models
* 2.5.1 Poisson-Boltzmann theory
* 2.5.2 Finite ion size corrections
###### Contents
* 1 Introduction
* 2 The \(\mathrm{d}\sigma\)-model model
* 2.1 The \(\mathrm{d}\sigma\)-model model
* 2.2 The \(\mathrm{d}\sigma\)-model model
* 2.3 The \(\mathrm{d}\sigma\)-model model
* 2.4 The \(\mathrm{d}\sigma\)-model model
* 2.5 The \(\mathrm{d}\sigma\)-model model
* 2.6 Parametrization
* 2.7 Implicit solvation implementations in DFT program packages
* 3 Implicit solvation models applied to electrified SLIs
* 3.1 _Ab initio_ thermodynamics framework
* 3.2 Potential of zero charge
* 3.3 Computational hydrogen electrode
* 3.4 Surface charging and interfacial capacitance
* 3.5 Constant potential vs.
* 4 Conclusions and outlook
* 5 Acknowledgements
Introduction
Electrocatalysis, i.e. potential-driven chemistry at electrified interfaces, is one of the pillars of a future sustainable energy landscape, providing a green storage of renewable energy and its conversion to valuable chemicals. The concomitant increased global interest in electrochemical processes at extended surfaces and interfaces has triggered unprecedented academic and industrial research efforts to optimize catalyst materials and electrochemical cell designs for maximal efficiency, sustainability, and durability. In this development, predictive-quality computational simulations have played a key role, augmenting experimental results with atomic-scale mechanistic insights and increasingly supporting catalyst discovery and optimization.
Given the fact that electrochemical reactions depend on the movement of charges, the respective computer simulations are by necessity based on a quantum mechanical description of the involved materials. Yet, while first-principles quantum chemistry provides a conceptually exact toolkit to simulate chemical reactions, current (super-)computers can even with most efficient semi-local density-functional theory (DFT) only simulate a limited amount of atoms and at time scales where chemical reactions cannot be statistically resolved. Fortunately, energy conversion processes can often be considered as a path through thermodynamically equilibrated, meta-stable states, separated by kinetic barriers which are often in a direct, linear relation with free energy differences between those states. Furthermore, chemical reactions frequently occur at defined locations, the so-called active sites, and have a quite localized impact on their surrounding. As a consequence, and as shown in Fig. 1, to a good approximation one can in many cases carve out from the full constant-particle thermodynamic system a smaller grand-canonical sub-system which is in equilibrium with bulk reservoirs of species.
In this _ab initio_ thermodynamics approach to surface catalysis, this sub-system in form of a model of the active site and any adsorbed reaction intermediates can then conveniently be computed as a slab within a periodic boundary condition supercell, and a grand-canonicalthermodynamic framework is used to connect the obtained first-principles energetics with the reservoirs through defined chemical potentials for the catalyst atoms and the reactants. In thermal heterogeneous catalysis, where this approach was pioneered and is widely used, the surrounding reactant environment and its corresponding reservoirs are generally well approximated by neutral ideal gases. Concomitantly, also the finite supercell is charge neutral and there is no necessity to explicitly include in the first-principles supercell calculations.
_Ab initio_ thermodynamics approach to electrified solid-liquid interfaces as occurring in electrocatalysis. The electrode is here negatively charged and this surface charge is compensated by the built up of counter charge in the electrolyte. The formed electric double layer (DL) can be pictured as a localized capacitor at the interface of electrode and a rather rigid layer of ions (inner DL or Helmholtz layer) and a long-range contribution (outer or diffuse DL). As in particular the spatial extent of the diffuse DL challenges efficient first-principles calculations, the _ab initio_ thermodynamics approach considers a grand-canonical ensemble in which a finite supercell computed e.g. with DFT is in equilibrium with appropriate reservoirs for the catalyst atoms, solvent species and electrons. Since the supercell does then generally not comprise the entire DL, it misses part of the compensating charge and does not necessarily have to be overall charge neutral.
Instead, the actual DFT calculations are simply performed for a slab in perfect vacuum. Unfortunately, the situation is significantly more complex in surface electrocatalysis, where the solid catalyst exchanges electrons with the reactants and is in contact with a liquid electrolyte forming a solid-liquid interface (SLI). As further detailed below, this enforces the consideration of charged reservoirs (electrons, protons, or ionic species in the electrolyte) with which the then no longer necessarily overall charge-neutral supercell is in electrochemical equilibrium, cf. Furthermore, this exchange of charge species with the respective reservoirs and potentially ongoing surface reactions are driven by applied electrostatic potentials, which directly interact with the solvent structure near the surface. Apart from the specifically adsorbed reaction intermediates there is thus now in principle also the need to describe the liquid electrolyte species within the finite volume between the periodically repeating slabs in the supercell.
It is from the objective of reducing this complexity and recovering the efficiency of _ab initio_ thermodynamics as known from thermal surface catalysis, where much of the renewed interest in implicit solvation schemes in this field comes from. Corresponding methodologies form in general a long-standing coarse-grained approach to describe a solvent environment on the level of a dielectric continuum. While they thus have their own history (in particular for molecular systems), their application to extended SLIs and the context of _ab initio_ thermodynamics has its specific challenges and merits. It is from this particular angle that we here review such methodologies and discuss their recent application to the surface electrocatalysis context, especially at metal electrodes and for liquid, mostly aqueous electrolytes. We refer to excellent and comprehensive reviews for full theoretical and technical details and the more traditional uses of implicit solvation methods for molecular systems, and content ourselves here with a focused exposition of the general concepts. Instead, we elaborate more on the specific demands, benefits and persisting issues when applying such methods to electrified interfaces.
To set the stage for such a discussion, also summarizes some key properties and specificities of the electrified SLI. Central to this is the separation of (ionic and electronic) charges that results from the interaction of the metallic electrode with the surrounding electrolyte under an applied potential. A potential-dependent amount of net charge \(\rho\) is thus localized on the electrode surface and counter charges in the form of dissolved ions are redistributed to a certain depth into the electrolyte to compensate for this net charge. Additionally, rotational, translational and even vibrational degrees of freedom in particular of polar electrolyte molecules (like water in aqueous electrolytes) will be affected within this formed, so-called electric double layer (DL). As a consequence of the concomitant screening, the electrostatic potential \(\phi\) drops over the width of the DL. At least in aqueous electrolytes, this drop generally occurs over two regions: the inner or Helmholtz layer (iDL), where \(\phi\) drops linearly, and the outer or diffuse layer, where it drops non-linearly. The capacitance \(C\) arising from the charging of the DL is correspondingly also commonly separated into an inner and an outer contribution. While this was originally made without a direct reference to the actual atomic-scale nature of the DL, the different dielectric property of the iDL is now related to a crowding of counter ions directly at the charged electrode. This leads to the formation of a compact layer with almost rigid water molecules and thus a small dielectric permittivity. In contrast, depending on the applied potential and electrolyte concentration, the more diffuse re-distribution of ions in the outer DL can extend over hundreds of Angstroms into the electrolyte, cf.
From this simplified capacitor picture, it becomes clear that the true amount of net surface charge on the electrode at a given applied potential is a sensitive function of the entire DL. Adsorption energies and therefore reaction pathways in turn often depend sensitively on this surface charge and the potential drop in the DL, e.g. via electrostatic interactions of dipolar adsorbates with the electric field. Already this aspect alone thus reveals that electrochemical activity in the SLI is generally not merely a function of the electrode aka catalyst material. Instead, it is equally influenced by the electrolyte and the concomitant DL. Additional aspects of this influence concern also more classic solvation effects like steric or bonding interactions with electrolyte species in the inner DL (in aqueous electrolytes e.g. prominently hydrogen bonds). Capturing these multifaceted influences and in particular their net effect on reaction energetics is correspondingly a pivotal ingredient of predictive-quality computational simulations and theoretical analyses of catalysis at electrified interfaces. At the same time and as further discussed in Section 2.1 below, the outer DL's large extent plus the ions' very slow, typically nanosecond time scale diffusion render any atomic-scale first-principles calculations including an explicit and dynamical account of the full DL still prohibitively expensive.
Implicit solvation schemes are at the opposite end and promise an unsurpassed computational efficiency in simulating the SLI. In their original molecular form, these schemes define a solvation cavity in which the solute is embedded and surrounded by a dielectric continuum representing the solvent's dielectric response. On top of that, the contribution of ions to the overall electrostatic response can be modeled. In the application to SLIs, such implicit solvation schemes thus foremost allow to appropriately describe the capacitive charging of the DL beyond the confines of the finite supercell--though requiring the integration into an _ab initio_ thermodynamics framework to appropriately account for the flow of particles between the subsystem and the reservoirs (cf. Fig. 1) as detailed in Section 3.1. Next to effectively describing the counter charge, implicit solvation models obviously also aim to capture plain solvation effects. Yet, with the solvent represented by a continuum this is, of course, only on a highly effective, parametrized level, in particular in the present state-of-the-art that also includes the inner DL into the implicit description. As further discussed in Section 2.6, this situation is aggravated even more by the scarcity of reliable experimental SLI data to fit the empirical parameters to. The prevalent approach to instead more or less uncritically resort to established parameters from (unbiased) molecular systems represents one of the aforementioned persisting issues in the field. It is these open challenges that are specific to the application of implicit solvation schemes to the context of SLIs that we also want to openly voice in this review, while simultaneously surveying the impressive insights that can be achieved with this at first sight admittedly rather crude approach.
## 2 Fundamentals of implicit solvation
### Coarse-graining the electrolyte
Since the beginning of computational chemistry, the simulation of solid-liquid interfaces has been of particular interest to scientists. To facilitate such investigations, theorists have since developed various methods particularly to coarse grain the highly dynamic and thus complex liquid phase. Indeed, the oldest such methods go back to Kirkwood and Onsager, and were introduced even a few years before the invention of the electronic computer. The goal back then was essentially the same as for the here discussed contemporary SLI electrocatalysis context, namely to reduce the physical complexity of the liquid phase in such a way as to keep the essential physics of the problem intact. Practically, these theories are derived for the description of thermodynamic equilibrium states by averaging over the configurational phase space. Of course, this can be achieved at varying degrees of coarseness which we will briefly survey in the following.
The starting point of our discussion is a fully _ab initio_, quantum mechanical treatment of the liquid phase, including all electronic and nuclear degrees of freedom (DOFs). Given the mobility of the molecules in the liquid phase, the evaluation of equilibrium states requires some sort of averaging or sampling over the nuclear DOFs, most often achieved in the form of _ab initio_ molecular dynamics (AIMD). Unfortunately, even at this fully explicit level there is still some debate which first-principles electronic structure theory is actually best suited for the task. Specifically for the description of pure water, easily the most important of solvents, there are a number of well documented failures of semi-local DFT, which in terms of its computational efficiency would be the present-day method of choice to describe larger supercells and achieve longest possible simulation times. Instead, the use of hybrid DFT with advanced dispersion corrections, or the strongly constrained and appropriately normed (SCAN) meta-GGA functional is often recommended, potentially even including nuclear quantum effects. This best practice becomes challenged in the SLI context though--not only because of potentially exploding computational costs, but also because the same functional now has to describe the (metallic) solid and the liquid phase with their very different physical characteristics on the same footing. For this specific task, the use of generalized-gradient functionals, in particular the revised version of the variational approach is not a viable option for the variational approach.
**Hierarchy of coarse-graining approaches for the liquid phase in the context of electrocatalysis at SLIs.** The sketch depicts an aqueous electrolyte with salt ions (blue spheres) and dissolved CO\({}_{2}\) (red and black molecules) at a crystalline surface. Starting from a fully explicit quantum mechanical description (far left) one can conceptually coarse-grain away electronic DOFs to arrive at a force field or interatomic potential description (center left). From that one can gradually remove nuclear solvent DOFs to represent solvent molecules e.g. only through their spatial distributions or correlation functions like in RISM-type models (center right). Finally, replacing even this with simply a polarizable continuum one arrives at fully implicit models (far right). Note that in the derivation and parametrization of each coarse-grained level one does not necessarily need to follow each step and can e.g. directly parametrize an implicit model from fully explicit data.
However, one clearly has to stress that this consensus derives more from reasonably-appearing averaged properties and functionalities computed at this level of theory, rather than from detailed experimental validation of the predicted atomic-scale structure of the electric DL.
Although AIMD simulations provide valuable insights about SLIs, they can usually only sample a few or even a single basin of the system's potential energy surface (PES) during presently computationally tractable trajectories on the picosecond time scale. Proper thermodynamic averages would instead require nanoseconds of simulations or longer, especially if the DL contains slowly equilibrating components such as ions or strongly physisorbed water. Furthermore, the simulation cell sizes feasible even over restricted picosecond time scales can barely, if at all cover the up to \(\sim 100\,\)A extent of the outer DL, cf.
The first in the corresponding hierarchy of approaches focuses on eliminating the electronic DOFs. This results in a classical description of pair-wise or many-body interactions between point-like nuclei in the form of an effective force field or interatomic potential to model the high-dimensional PES. While this is an extensive field of its own with a plethora of most advanced force fields for (bulk) water, electrolytes or materials, the crux is again in requiring them to describe the SLI within the same simulation cell. Much fewer parametrizations exist for this task, in particular for the interactions of (organic) electrolyte species with the (inorganic and heavy) elements like Pt or Cu that form the metallic electrodes. On top of that, most traditional force fields can not reliably describe bond forming or breaking events and can thus not cover the reactive surface chemistry that is central to catalysis at electrified interfaces. While there are thus only few examples where fully classical simulations were used to study the structure of SLIs, there are currently two interesting developments to overcome these limitations. To one end, modern reactive force fields that can account for bond dissociation start being applied in SLI simulations even under applied potential. To the other end, machine-learned interatomic potentials are a most promising new possibility to establish a computationally efficient surrogate to direct first-principles calculations. By construction, their reliability and range of applicability is determined by the training data fed into them. If this data contains appropriate information on the SLI and its reactive events, dynamical simulations based on such a potential would produce the same insight as direct AIMD, just orders of magnitude faster. Precisely the development of corresponding data-efficient training protocols (that would not require prohibitive amounts of first-principles training data) is presently the focus of strong research efforts worldwide. As this research is ongoing, present applications of machine-learned potentials to the SLI context are still restricted to some first case studies though.
An important general aspect in switching to more coarse-grained descriptions is that different levels may suitably be chosen for different spatial regions of the overall simulation cell. In the SLI context, a widespread realization of such concurrent multiscale modeling is a quantum mechanical/molecular mechanical (QM/MM) approach, in which the solid electrode and the chemical reactions thereon are kept on a quantum chemical level, while a force field or interatomic potential is employed for the liquid electrolyte. This offers significant speed-ups as much of the electrolyte sampling is done classically, while in particular the reactive surface chemistry is still described at a first-principles level. Note that the (spatial) distinction of what is described at the more coarse-grained level can be chosen flexibly, with the limitation that approaches that allow to continuously morph say a classically described atom into a quantum mechanically described one during an ongoing dynamical simulation are still in their infancy. Typically, which atoms (or molecules) are described at which level is therefore defined at the onset of a simulation, and this is kept fixed regardless of where the actual dynamical motion drives the atom or molecule to. A classical description of all electrolyte species apart from (specifically) adsorbed reaction intermediates offers thereby obviously highest computational efficiency, but is by construction unable to cover situations where the liquid phase participates actively in the reactions, e.g. as a proton donor. Furthermore, it also requires in principle specific (interface-sensitive) parametrizations to account for the overall effect of the classical solvent species on the surface reactions. Both of these limitations can instead be mitigated by including (parts of) the inner DL into the quantum mechanical part of the simulation, yet at concomitantly increased computational costs.
Central to the value of such simulations is in any case the correct depiction of the interaction or embedding energy of solid and liquid phase, the solvation energy. In QM/MM models, the Coulomb contribution to the solvation energy is commonly described by the interaction of the QM charge distribution with fixed, fitted electrostatic charges of the classically described liquid molecules. In addition to this, non-Coulomb contributions, including Pauli repulsion, dispersion and induction forces, have to be carefully parametrized. Electronic induction of the solid phase by the liquid phase charge distribution is treated by self-consistently re-iterating the liquid distribution and electron density. Polarization of the liquid phase is instead most often only included through movement and reorientation of solvent molecules and ions. In certain situations an additional electronic polarization, i.e. changes of the partial charges of atomic sites of the solvent molecules, has been shown to be relevant and can in principle be included using polarizable force fields. The description of the other, non-Coulomb interactions is still a topic of ongoing research. Commonly they are simply represented by pairwise interactions with parameters obtained from high-level quantum chemical calculations or by fitting to thermodynamic or dielectric properties of the (bulk) solvent. Nevertheless, properly parametrized force fields have actually been shown to sometimes even surpass full AIMD simulations in accuracy concerning structural and dynamic properties of the solvent. Their still atomistic approach to representing the liquid phase also has advantages over the more coarse-grained models discussed in the following, in that they can more readily describe localized effects and directed interactions such as hydrogen bonds to surface adsorbates.
While a QM/MM description of the SLI greatly speeds up simulations by simplifying the computational treatment of the liquid DOFs, it still does not relieve the need to sufficiently sample the phase space of each solvent molecule. Combined with the need to still determine the QM polarization response to each new MM charge configuration, even such simplified models might not be computationally tractable. Recognizing the explicit sampling of the solvent dynamics as the bottleneck, a further coarse-graining step aims therefore at effectively averaging out the movement of solvent molecules and ions, and at replacing them instead with their respective spatial equilibrium distributions, cf. A prominent representative of this ansatz is the reference interaction site model (RISM), which evaluates the equilibrium radial correlation functions of each pair of species in the system through an analytical integral equation, known as the Ornstein-Zernike equation. Within given approximations, the equilibrium structure of the fluid around any form of solute is then fully encoded through these radial distribution functions and without further need for a costly dynamical sampling. In most variants of RISM, such as the popular 3D-RISM, the central pair correlation functions are evaluated on a three-dimensional grid centered on the solute to yield the spatial distribution functions of each solvent site species. These distribution functions can be integrated over space and summed up to yield an excess chemical potential of solvation due to the solute-solvent interaction and solvent reorganization in the presence of the solute. In RISM theory, it is this excess chemical potential that connects the coarse-grained solvent with the explicitly treated solute. Its functional derivative with respect to the electron density yields an effective potential that can directly be included into the solute's Hamiltonian. This potential includes all the interactions used in the determination of \(g_{ij}(\mathbf{r})\) such as electrostatics and, most commonly, Lennard-Jones type terms encompassing dispersion and exchange interactions. Given the implicit dependency of the solvent excess chemical potential on the electron density, the solvent response is then iterated together with the quantum-mechanical DFT-described part of the system to reach self-consistency.
Going beyond purely molecular solvents, RISM-like models recently have seen very successful use in the simulations of various electrochemical processes.
Inherent to effective models, both explicit classical and RISM-based descriptions of the liquid phase depend on a series of element-specific parameters that define interatomic interactions and have to be carefully chosen for each system of interest. This requirement is generally not a significant burden for detailed studies of individual systems, in particular if these are prototypical cases for which then typically a plethora of high-level or experimental data is available that can be used for the parametrization. It becomes critical though, if fast estimates are needed, for instance to assess the catalytic activity of a large variety of electrode materials, morphologies, active sites or electrolyte components, or if unknown and complex electrochemical reactions are studied for which no reference data is available. For such cases and for potential further increases in efficiency, an even higher level of coarse graining of the liquid phase becomes appealing, in which all solvent DOFs are altogether merely described via a polarizable continuum, cf.
Following the concurrent multiscale modeling philosophy of QM/MM or QM/RISM, such implicit solvation schemes are in the SLI context predominantly employed to describe the equilibrium solvent response on a (metallic) electrode computed at a first-principles level of theory. Again, flexibility exists whether to replace the entire electrolyte in a so-called fully implicit approach, or to retain an explicit quantum or molecular mechanical description of (parts of) the inner DL, with latter models referred to as hybrid explicit/implicit models. Reduced to a continuum, the implicitly treated electrolyte is then just a polarizable medium with a dielectric permittivity. While an isotropic, constant tensor in the bulk of the electrolyte, this permittivity can in principle vary closer to the symmetry-breaking SLI. Additionally, it needs to be artificially reduced to vacuum permittivity inside the explicitly treated region of the simulation cell so as to not introduce spurious polarizability on top of the one intrinsically provided by the quantum or molecular mechanical description of the corresponding atoms or molecules. This region of vacuum permittivity inside the overall simulation cell is commonly referred to as solvation cavity, a word coined within the traditional field of implicit solvation of finite moleculear solutes. As discussed in detail in Section 2.3, different classes of implicit solvation schemes are categorized by the functional form employed to describe these spatial variations of the dielectric permittivity tensor. This form determines the electrostatic solvent response and could in principle be chosen to be non-local to reach similar levels of accuracy as RISM models. However, the use of such functional forms would unavoidably require the introduction of a multitude of system-specific parameters, thereby nullifying the original motivation for this effective methodology.
For planar electrodes (typically described by crystalline slabs with low-index surfaces in the corresponding first-principles supercell calculations), it is therefore common to only consider a local and stationary dielectric tensor with components that vary exclusively as a function of the vertical distance \(z\) to the surface. In fact, typically even the tensorial nature of the permittivity is neglected and a simple functional form for the scalar permittivity \(\epsilon(z)\) is employed. As this omits all structure in the liquid and especially any kind of directed interactions with the surface, such effects are instead considered by additional effective non-Coulomb energy functionals as discussed in more detail in Section 2.4 below. This particular strategy then allows to employ a minimum number of parameters for the dielectric modeling function and these non-Coulomb energy corrections as further discussed in Section 2.6.
We also discuss prevalent fitting strategies for these parameters in Section 2.6, but note already here that the simplicity of this prevalent approach does not only reflect the objective of creating a computationally most effective, transferable solvation approach. To some extent and as mentioned before it is also dictated by the present scarcity of interface-sensitive experimental or high-level theoretical reference data that does not warrant a more detailed (physical) modeling with a concomitantly increased number of parameters. This aspect notably also concerns the powerful possibility of extending implicit solvation schemes from pure liquids to electrolytes by additionally modeling the ionic charge distribution as discussed more in Section 2.5. Most of these models rely on the traditional diffuse DL theory, providing a functional form between ion distributions and electrostatic potential as developed by Gouy, Chapman and Debye in the beginning of the last century. Since this original approach, many corrections regarding e.g. non-mean-field ionic correlation effects, steric size corrections or ion-surface interactions have been made. While physically clearly motivated, each of these corrections necessarily gives rise to further parameters. Even though it is in particular this capability to account for the ionic counter charges that is presently predominantly exploited for the SLI context, it is thus again a specific issue of this application field in how much these more advanced electrolyte models can be parametrized or are in fact really necessary for the specific counter charge modeling aspect.
### Separation of the grand potential energy functional
As apparent from the discussion in the last section, different levels of theory ranging from high-level quantum chemistry to force fields or interatomic potentials may generally also be chosen for the description of the solid electrode (and an explicitly treated part of the inner DL). In the remainder of this review we will nevertheless focus on the use of DFT for this task, as this is the predominantly taken approach in implicit solvation works on SLIs and electrocatalysis at metallic electrodes to day. With minor modifications, many of the concepts and discussions are readily adapted to the other levels of theory though.
As described in the introduction around Fig. 1, in the SLI context, the employed DFT supercell at volume \(V\) generally only represents a grand-canonical sub-system, which is connected to bulk reservoirs of species that represent the rest of the (macroscopic) system. For the electrochemical environment, these would naturally include an electrochemical potential \(\tilde{\mu}_{\rm el}\) for the electrons, electrochemical potentials \(\tilde{\mu}_{{\rm ion},i}\) for different ionic electrolyte species \(i\), and chemical potentials \(\mu_{{\rm solv},j}\) for different neutral solvent species \(j\). In Chapter 3 we will detail how these potentials are set for the SLI context, but for the time being they are simply given constants. For such given constants, the true equilibrium structure and composition of the electrified interface would result from an exhaustive grand-canonical sampling and thermodynamic averaging of all nuclei and electronic DOFs inside the supercell--with nuclei DOFs here and henceforth denoting the detailed geometric structure and chemical composition of the system and electronic DOFs referring to those of the DFT-part of the system. In the coarse-grained solvation modeling reviewed here, this typically infeasible task is separated into two stages. First, solvation effects are evaluated for an individual interface configuration characterized by say a given electrode geometry and chemical composition with specifically adsorbed reaction intermediates at its active sites. The chemical composition \(N_{\alpha}\) of chemical species \(\alpha\) in this explicitly and DFT-described part of the system is thus fixed, and under an unanimous Born-Oppenheimer approximation the thermodynamic sampling and averaging is restricted to the remaining (canonical) electronic and (grand-canonical) nuclei DOFs of the electrolyte. In other words, one thus evaluates the thermodynamic stability of the electronic ground-state configuration for the given static nuclei charge density \(\rho_{\rm nuc,QM}=\rho_{\rm nuc,QM}(\mathbf{r})\) and in contact with a fully equilibrated electrolyte. In a subsequent step detailed in Section 3.1, an _ab initio_ thermodynamics framework is then employed to compare the stability of different such explicit interface configurations and compositions, and the one exhibiting the highest stability is identified as the closest approximant to the true grand-canonical equilibrium SLI structure within the tested space of configurations.
In this and the remaining sections of this chapter we will concentrate on the first of these two stages. In this stage, there is thus one defined chemical composition \(N_{\alpha}\) of the DFT-described part of the system, and in this respect this stage then encompasses the more traditional use of implicit solvation schemes in the molecular DFT context with finite solutes.
\[\Omega^{N_{\alpha}}[\rho_{\rm el,QM},\rho_{\rm is}]=\left.F_{\rm QM}\left[ \rho_{\rm el,QM}\right]\right|_{\rho_{\rm nuc,QM}}\;+\;\Omega_{\rm is}[\rho_{ \rm el,QM},\rho_{\rm is}]\quad, \tag{1a}\]which is minimized by the equilibrated charge density distribution \(\rho_{\rm is}^{\circ}\) and the ground-state electron density \(\rho_{\rm el,QM}^{\circ}\). Here, \(F_{\rm QM}\) is the free energy functional of the pure quantum system and \(\Omega_{\rm is}\) is the grand potential of the surrounding electrolyte. For simplicity of notation, we drop in the following the subscript "QM" (e.g. \(F_{\rm QM}\to F\)), and consistently denote all properties related to the electrolyte with the subscript "is" (for implicit solvent). Within the employed Born-Oppenheimer approximation, we also henceforth refrain from explicitly stating the only parametric dependence of \(F\) on the nuclei charge density \(\rho_{\rm nuc}\).
\[F[\rho_{\rm el}]=\overbrace{\underbrace{{\cal T}_{\rm el}^{\rm S}[\rho_{\rm el }]+{\cal V}^{\rm mf}[\rho_{\rm el}]+E^{\rm xc}[\rho_{\rm el}]}_{U[\rho_{\rm el} ]}+{\cal T}_{\rm nuc}[\rho_{\rm el}]}^{E^{\rm KS}[\rho_{\rm el}]}-TS[\rho_{\rm el }]\quad. \tag{1b}\]
Here, \({\cal T}_{\rm el}^{\rm S}\) is the kinetic energy functional of the non-interacting electrons and \({\cal T}_{\rm nuc}\) represents the kinetic energy functional of the nuclei (usually evaluated only as a post-correction at \(\rho_{\rm el}^{\circ}\)). The Coulomb energy functional \({\cal V}^{\rm mf}\) contains both nuclei-nuclei interactions described explicitly and electronic interactions described on the mean-field level, while additional electronic interactions are accounted for through the DFT exchange-correlation functional \(E^{\rm xc}\). \(E^{\rm KS}\) is generally referred to as the KS energy functional, and finally, \(TS\) represents entropic corrections at the given temperature \(T\). As indicated, all terms in \(F\) with the exception of the last one are often summarized under the header of the internal energy functional \(U\).
Importantly, \(F[\rho_{\rm el}]\) with all its terms is exactly the functional also underlying regular DFT calculations and does thus not depend on the electrolyte distribution \(\rho_{\rm is}\). We correspondingly refer to a multitude of excellent accounts on KS DFT for further details on this functional. All electrolyte-induced changes of the ground-state electron density arise instead from the optimization of the grand potential \(\Omega^{N_{\alpha}}[\rho_{\rm el},\rho_{\rm is}]\) in eq. (1a) and not \(F[\rho_{\rm el}]\) alone. In contrast, \(\Omega_{\rm is}[\rho_{\rm el},\rho_{\rm is}]\) as the second part of this grand potential refers to the electrolyte in its equilibrium distribution, and this does depend on the detailed charge distribution of the DFT-described solute and thus its electron density \(\rho_{\rm el}\). Conceptually, in order to determine \(\Omega_{\rm is}[\rho_{\rm el},\rho_{\rm is}]\) all electrolyte DOFs would therefore have to be sampled in the presence of a given \(\rho_{\rm el}\), and then the interdependence of electrolyte and DFT system charge densities would require an iterative cycle or generally a numerical optimizer to minimize the functional \(\Omega^{N_{\alpha}}[\rho_{\rm el},\rho_{\rm is}]\) with respect to the electron density at a corresponding equilibrium charge density distribution of the electrolyte. This has e.g. been realized in electrostatic QM/MM embedding, where molecular dynamics simulations are used to sample the equilibrium distributions of the electrolyte DOFs. Similarly QM/RISM simulations have been employed, in which a more coarse-grained model is used to derive the electrolyte equilibrium distribution corresponding to a given electron density of the QM system as already mentioned in Section 2.1.
The great advantage of implicit solvation schemes over these less coarse-grained approaches is that there a model solvation grand potential \(\Omega_{\rm is}[\rho_{\rm el}]\) is derived solely as an explicit functional of the electron density \(\rho_{\rm el}\). This leads to a dramatic reduction of computational effort, as then the evaluation of the resulting closed form of \(\Omega^{N_{\alpha}}[\rho_{\rm el}]\) can be achieved for a given \(\rho_{\rm el}\) in one go. In fact, corresponding schemes are often directly integrated into the DFT program packages by simply adding routines that evaluate and add the \(\Omega_{\rm is}[\rho_{\rm el}]\) contribution within the regular KS DFT minimization procedure. For this, it seems at first natural to separate the model grand potential functional into formal terms analogous to the quantum free energy functional \(F[\rho_{\rm el}]\),
\[\Omega_{\rm is}[\rho_{\rm el}]=\underbrace{{\cal T}_{\rm is}[\rho_{\rm el}]+{ \cal V}_{\rm is}[\rho_{\rm el}]}_{U_{\rm is}[\rho_{\rm el}]}-TS_{\rm is}[\rho_{ \rm el}]-\sum_{i}\tilde{\mu}_{\rm ion,i}\langle N_{\rm is,ion,i}\rangle[\rho_{ \rm el}]-\sum_{j}\mu_{\rm solv,j}\langle N_{\rm is,solv,j}\rangle[\rho_{\rm el }]\quad, \tag{1c}\]
Note that as a grand potential, \(\Omega_{\rm is}\) formally also contains contributions due to the electrochemical potentials of the ionic (\(\tilde{\mu}_{\rm ion,i}\)) and chemical potentials of the neutral solvent species (\(\mu_{\rm solv,j}\)). The inclusion of these terms--and especially their average particle numbers \(\langle N_{\rm is,ion,i}\rangle\) and \(\langle N_{\rm is,solv,j}\rangle\) in the implicit electrolyte--does at first seem counter-intuitive given that all explicit solvent and ion degrees of freedom have been coarse-grained out. Yet, as we will see below, even implicit ionic and solvent molecule concentrations in the simulation box depend on the electrochemical environment (e.g. the applied potential). Therefore the exchange of both kinds of particles with the extended electrolyte (as represented by the (electro)chemical potentials) needs to be accounted for, at least approximately.
In general, and as further elaborated on in Section 3.1, one is actually rarely interested in the absolute grand potential of eq. (1a). Instead it is differences in free energies, and thus differences in grand potentials at their respective optimal electronic densities, that are the main descriptors of chemical reactions. Similarly, comparisons with experiment--which are generally used for model parametrization--are also most easily done on the level of solvation free energies, which in turn are differences between the grand potential at optimized densities in solvent and in vacuum. For this purpose and considering the strong approximations to be made anyway, the fine separation into the various formal terms in eq. (1c) is not ideal.
\[\Omega_{\rm is}[\rho_{\rm el}]={\cal V}_{\rm is}^{\rm mf}[\rho_{\rm el}]+\Omega _{\rm is}^{\rm non-el}[\rho_{\rm el}]+\Omega_{\rm is}^{\rm ion}[\rho_{\rm el}]\quad. \tag{2}\]
Here, \({\cal V}_{\rm is}^{\rm mf}\) is the mean-field contribution due to the electrostatic response of the continuous polarizable medium describing the pure liquid.
\[\Omega_{\rm is}^{\rm non-el}[\rho_{\rm el}]\;=\;\Omega_{\rm is}^{\rm cav}[\rho _{\rm el}]+G_{\rm is}^{\rm rep}[\rho_{\rm el}]+G_{\rm is}^{\rm dis}[\rho_{\rm el }]+G_{\rm is}^{\rm tm}[\rho_{\rm el}]\quad, \tag{3}\]
describes all additional effects introduced by ions in the electrolyte. Even though in practice often further lumped together, cf. Section 2.6, we here distinguish four non-electrostatic contributions. \(\Omega_{\rm is}^{\rm cav}[\rho_{\rm el}]\) denotes the grand potential cost of forming a cavity in the solvent for the solute to be placed in. Making this space for the solute necessarily changes the particle numbers of solvent molecules of the implicitly-described liquid in the supercell and thus involves particle exchange with the reservoirs with a concomitant dependence on the chemical potentials of the solvent components. We accordingly denote this term here as a grand potential, even though most available literature refers to it as a cavitation free energy functional. \(G_{\rm is}^{\rm rep}[\rho_{\rm el}]\) commonly represents the contribution due to exchange or Pauli repulsion interactions, effectively also including an entropic contribution due to the resulting changes to the potential energy surface (PES). The third term, \(G_{\rm is}^{\rm dis}[\rho_{\rm el}]\), similarly represents dispersion or van der Waals interactions. Finally, \(G_{\rm is}^{\rm tm}[\rho_{\rm el}]\) is the free energy functional accounting for changes in the thermal motion of the solute. Note that all of the non-electrostatic terms and \(\Omega_{\rm is}^{\rm ion}[\rho_{\rm el}]\) thus contain potential, kinetic and entropic contributions. Nevertheless, each of these terms has been proven to be computationally accessible and in the following sections, we will now further elaborate on these various contributions to \(\Omega_{\rm is}[\rho_{\rm el}]\), starting first with a pure solvent and the discussion of the dominant electrostatic \(\mathcal{V}_{\rm is}^{\rm mf}[\rho_{\rm el}]\) term in Section 2.3 and the non-electrostatic \(\Omega_{\rm is}^{\rm non-el}[\rho_{\rm el}]\) in Section 2.4. In Section 2.5, ions are then added on top of that to arrive at full implicit electrolyte models that also include a model \(\Omega_{\rm is}^{\rm ion}[\rho_{\rm el}]\) grand potential. The general objective in all of these sections is to derive (closed) expressions for these functionals of the electron density, which then allows to (straightforwardly) add these contributions into the KS DFT minimization process. As noted before, the true free energy is then formally given by the grand potential \(\Omega^{N_{\alpha}}[\rho_{\rm el}^{\circ}]\) evaluated at the resulting optimized density \(\rho_{\rm el}^{\circ}\), cf. eq. (1a). However, it is important to note that, to this end, the practical implementations acknowledge the aforementioned fact that predominantly only grand potential energy differences are required. In such differencesof say systems \(A\) and \(B\),
\[\Delta\Omega(A,B) = \Omega^{A}[\rho_{\rm el}^{\circ}(A)]-\Omega^{B}[\rho_{\rm el}^{ \circ}(B)]\quad, \tag{4}\]
From this perspective, no efforts are therefore made to account for such contributions in the derived functional expressions in the first place. While formally describing the absolute solvation grand potential of eq., we thus emphasize that in practice many of the expressions discussed in the next sections only work for free energy differences. In fact, not least for reasons of computational efficiency the practical implementations often also consider only some terms within \(\Omega_{\rm is}[\rho_{\rm el}]\) in the functional minimization. One justification for this is an assumed negligible impact of the omitted terms on the final optimized electron density. Another pragmatic one is that any error incurred through the omission is effectively compensated in the fitting of the model parameters to reference data. A prominent example for this is to only consider the electrostatic \({\cal V}_{\rm is}^{\rm mf}[\rho_{\rm el}]\) in the minimization, and evaluate all non-electrostatic free energy contributions only as a post-correction on the basis of the resulting electron density that was thus exclusively optimized with respect to the dominant mean-field polarization effect of the surrounding liquid.
### Electrostatics of solvation
#### 2.3.1 Potential energy and polarization models
The mean-field electrostatic \({\cal V}_{\rm is}^{\rm mf}\) is the contribution to the solvation grand potential most intuitively associated with the response of a solvent to a solute. Considering it jointly with the Coulomb energy functional \({\cal V}^{\rm mf}\) in the minimization of the grand potential energy functional of eq. (1a) accounts for the polarization response of the continuum solvent to the net charge distribution of the solute \(\rho\) (resulting from the electron \(\rho_{\rm el}\) and nuclei \(\rho_{\rm nuc}\) charge densitiesof the DFT part of the system) and vice versa. To derive this contribution, we consider the static displacement field \(\mathbf{D}\) which arises from the collection of these explicit charges in the system and is screened by the surrounding medium. \(\mathbf{D}\) is given by the generalized Poisson equation (GPE)
\[\nabla\mathbf{D}=(\rho_{\rm el}+\rho_{\rm nuc})\quad. \tag{5}\]
The displacement field is related to the electric field \(\mathbf{E}\) of the explicit charge distribution via the polarization vector \(\mathbf{P}\), representing permanent and induced dipoles in the system. The functional form of \(\mathbf{P}\) is generally quite complicated, but depends on the relative and generally non-local dielectric permittivity tensor \(\mathbf{\varepsilon}=\mathbf{\varepsilon}_{\rm tot}/\varepsilon_{0}\) (with the absolute \(\mathbf{\varepsilon}_{\rm tot}\) and vacuum permittivity \(\varepsilon_{0}\)),
\[\mathbf{D}=\varepsilon_{0}\mathbf{E}+\mathbf{P}[\mathbf{E},\mathbf{\varepsilon}]\quad. \tag{6}\]
Technically, this makes \(\mathbf{D}\) a functional of the electric field \(\mathbf{E}\), which itself is an implicit functional of the net charge density (\(\rho_{\rm el}+\rho_{\rm nuc}\)) via eq.. For reasons of legibility we dropped these dependencies though. In this definition of \(\mathbf{D}\), the permittivity tensor is assumed to be static--i.e. time independent, but may still vary in space, e.g. to account for the symmetry breaking through a finite solute or an extended interface. Note that eq. also omits higher-order multipolar terms that might arise in the medium. For water, this approximation is generally well justified because the solvent molecules' electric field is dominated by its dipole moment. Higher-order terms can, however, be important in non-aqueous solvents with sizeable higher-order multipole moments, but to our knowledge no DFT program package yet supports an implicit solvent parametrization including such higher-order terms.
The GPE of eqs. and provides a direct relation of electric field and charge density which is generally valid, with and without a polarizable medium.
\[\left(\mathcal{V}^{\mathrm{mf}}+\mathcal{V}_{\mathrm{is}}^{\mathrm{mf}}\right) \left[\rho_{\mathrm{el}}\right]=\frac{1}{2}\int\left(\rho_{\mathrm{el}}+\rho_{ \mathrm{nuc}}\right)\phi=\frac{1}{2}\int\mathbf{E}\mathbf{D}\mathrm{d}\mathbf{r}\quad, \tag{7}\]
into eq., using the divergence theorem, neglecting the surface terms and finally substituting \(\mathbf{E}=-\nabla\phi\), with \(\phi\) the electrostatic potential.
The assumption of a static, i.e. frequency independent, dielectric permittivity implies that the solvent adapts instantaneously to the electron and nuclei charge distribution of the solute. While this is generally a good approximation for the solvent response on thermodynamic equilibrium and potentially even for transition states of chemical reactions, it will over-screen fast molecular dynamics, such as vibrations or charge-transfer processes. On top of that, the simulation of electronically excited states has been shown to generally also necessitate a frequency-dependent dielectric response. For most other cases, however, the static, dipolar response model is a good starting point for further approximations. As compiled in Fig. 3, these lead to three main categories of dielectric models, namely non-linear, non-local and anisotropic ones.
**Categorization of different electrostatic solvation models**. From the general starting point of a static non-linear, non-local and anisotropic model (top) several approximations can be made to ultimately arrive at the linear, local and isotropic polarization model most commonly applied in present-day DFT codes.
Notwithstanding, mostly a Taylor expansion of \(\mathbf{P}\) as a function of \(\mathbf{E}\) around \(\mathbf{E}=0\) can be truncated after the linear term (linear-response approximation), i.e.
\[\mathbf{P}\approx 0+\varepsilon_{0}\left(\mathbf{\varepsilon}-\mathbf{I}\right)\mathbf{E}\quad, \tag{8}\]
Next, non-locality in the solvent response is important, whenever solvent molecule correlations occur, e.g. close to charged solutes. The spherically averaged liquid susceptibility (SaLSA) model represents one example that accounts for non-locality. SaLSA has been also coarse-grained into a computationally more feasible local version (charge-asymmetric nonlocally determined local-electric, CANDLE), with the dielectric permittivity being derived from the non-local response. Non-locality may also be employed to account for an effective size of solvent molecules, since the electric field at a certain position then affects the solvent density in a finite solvent radius around it. This can be relevant to prevent solvent from penetrating into small pockets formed by the solute, see the discussion on the dielectric function below. Finally, anisotropic properties of the dielectric permittivity are, of course, generally important in systems with reduced symmetry. This is notably the case at electrified SLIs where even at a planar interface the dielectric tensor would at least feature two independent dielectric tensor components, parallel \(\varepsilon_{||}\) and vertical \(\varepsilon_{\perp}\) to the surface.
While non-linearity, non-locality and anisotropy could thus well be of relevance for SLIs, most implicit solvation models that have been implemented into DFT program packages to date neglect all three of them and are based on the most simple case of a linear, local and isootropic dielectric model \(\varepsilon(\mathbf{r})\).
\[\nabla\mathbf{D}=\varepsilon_{0}\nabla\left[\varepsilon(\mathbf{r})\mathbf{E}\right]=- \varepsilon_{0}\nabla\left[\varepsilon(\mathbf{r})\nabla\phi\right]=\left(\rho_{ \mathrm{el}}+\rho_{\mathrm{nuc}}\right) \tag{9}\]
can be further simplified. Usingeq., it then features separately the electrostatic energy functional contributions of the DFT part and of the implicit solvent
\[\left(\mathcal{V}^{\mathrm{mf}}+\mathcal{V}_{\mathrm{is}}^{\mathrm{mf}}\right) \left[\rho_{\mathrm{el}}\right]=\frac{1}{2}\int\left(\rho_{\mathrm{el}}+\rho_{ \mathrm{nuc}}\right)\phi=\underbrace{\frac{\varepsilon_{0}}{2}\int\mathbf{E}^{2} \mathrm{d}\mathbf{r}}_{\mathcal{V}_{\mathrm{mf}}^{\mathrm{mf}}}+\underbrace{\frac{1 }{2}\int\varepsilon_{0}\left(\varepsilon(\mathbf{r})-1\right)\mathbf{E}^{2}\mathrm{d} \mathbf{r}}_{\mathcal{V}_{\mathrm{is}}^{\mathrm{mf}}}\quad, \tag{10}\]
Since the latter GPE cannot be solved analytically for most dielectric functions, a closed form is typically not attainable and a numerical solution is required. Common methods for this include fixed point iterations or the conjugate gradient technique employing the analytically known Green's function of the Poisson equation in vacuum. Alternatively, for certain functional forms of the dielectric function multi-center multipole expansions have been shown valuable, or mappings onto a finite grid and solution via standard finite difference or finite element techniques. Regardless of this technical realization, the conceptual changes to a DFT code to incorporate the Coulomb electrostatic contribution at this level of dielectric model are nevertheless minimal. In fact, while the entire self-consistency cycle around the KS equations is untouched, the only change is that the electrostatic potential does no longer satisfy the normal Poisson equation, but is instead given by the GPE of eq..
#### 2.3.2 Dielectric function
For the linear, local and isotropic case, the dielectric permittivity \(\varepsilon(\mathbf{r})\) may generally still vary in space. As already introduced in Section 2.1, in present-day implicit solvation schemes this is primarily reduced to modeling a transition from the bulk solvent permittivity \(\varepsilon_{0}\varepsilon_{\infty}\) (with the relative solvent permittivity \(\varepsilon_{\infty}\)) deep inside the electrolyte to the vacuum permittivity \(\varepsilon_{0}\) inside the DFT-described part of the supercell. The optimum location and form of this transition is generally system specific. Optimum refers hereby to the best possible reproduction of the true solvation effects within the confines of the chosen dielectric continuum model, and--in particular in the widespread approach to even include the inner DL fully into the implicit model--system specific includes an actual dependency on the electrode structure and chemical composition. In principle, this optimum location and form for a specific system could be determined from high-level explicit simulations. However, this would negate the original motivation to use an implicit solvent model for its efficiency gain and to e.g. screen a large number of different SLIs. Implicit solvation schemes rely therefore typically on a sufficiently simple functional form of \(\varepsilon(\mathbf{r})\) which includes as much system-relevant physics as possible while maintaining an optimum transferability. Obviously, this implies a trade-off between a more physically accurate description for particular systems (then typically involving a larger number of parameters that need to be determined) and a more simplified model with as little parameters as possible to describe qualitative trends over a wide range of systems.
Favoring higher transferability, the dielectric transition is often approximated by a mere switching function between bulk solvent and vacuum, resulting in the formation of a solvation cavity. The location of the dielectric transition thereby has to be expressed in an appropriate molecular descriptor that is readily available in any DFT calculation.
**Illustration of different types of dielectric transition between solute and solvent**. For the example of an adsorbed CO\({}_{2}\) molecule at a single-crystal surface, a) shows the solvation cavity resulting from the superposition of atom-centered spheres based on eq., while b) shows the solvation cavity as defined by an isosurface of the electron density.
4, traditional implicit solvation techniques as dominantly used in molecular chemistry often rely on defining a solvation cavity by summing up atom-centered shape functions \(s(\mathbf{r})\), so that
\[\varepsilon(s(\mathbf{r}))=(\varepsilon_{\infty}-1)\left\{\prod_{\alpha}s_{\mathbf{r}}(| \mathbf{r}-\mathbf{R}_{\alpha}|,\mathbf{\mathcal{P}}_{\alpha})\right\}+1\quad. \tag{11}\]
Here \(s_{\mathbf{r}}\) is a shape function going from 0 in the solute region to 1 in the bulk electrolyte, \(\{\mathbf{R}_{\alpha}\}\) are the positions of the nuclei, and \(\mathbf{\mathcal{P}}_{\alpha}=(\{r_{\alpha}\},\dots)\) is a vector of parameters, containing e.g. the exclusion radius \(r_{\alpha}\) for each atom and the transition smoothness of the shape function. The simplest shape function is just a single Heaviside function \(s_{\mathbf{r}}=\theta(|\mathbf{r}-\mathbf{R}_{\alpha}|-r_{\alpha})\), with the atomic radii \(r_{\alpha}\) as the only parameters. These radii are usually either taken as tabulated van der Waals radii for each chemical element or fitted to reproduce some experimental data as discussed in more detail in Section 2.6. One advantage of using such a sharp step function, also sometimes referred to as apparent surface charge approach, is the efficiency with which the GPE can be solved using boundary element methods. Yet, in most cases this comes at the expense of additional approximate corrections for errors due to parts of the QM charge density lying beyond the transition. Corrections for this outlying-charge error are correspondingly integral parts of well-known implicit solvation approaches like the polarizable continuum model (PCM), the solvent model (SMx), or the conductor-like screening model (COSMO) that rely on such sharp step functions. As an alternative, recently also smoothed step functions were proposed and adapted specifically for SLI simulations (soft-sphere continuum solvation - SSCS model), then, however, requiring additional parameters for the functional form of this transition.
In general, defining the cavity based on atom-centered shape functions has the advantage of easily being able to implement dielectric regions, e.g. at dielectric interfaces, by assigning different values to the local dielectric permittivity. Additionally, solvation radii can be assigned separately to each atom based on their chemical environment. This allows for great flexibility in the definition of the dielectric function and, potentially, a more accurate prediction of solvation energies. Unfortunately, such a treatment also results in a larger parameter space, risking overfitting with the generally rather small available training sets as further discussed in Section 2.6. Furthermore, the reliance on atom-centered shapes may lead to the formation of encapsulated solvent pockets in lower-density parts of the solute. In particular in the context of extended metallic electrodes, filling such pockets with solvent unlikely reflects the correct physics.
Both of these limitations may be overcome in a different, equally popular approach. It recognizes that the presence of electron density--readily available in a DFT calculation--naturally separates explicitly treated regions from the rest of the supercell. The solvation cavity can thus be defined by an iso-surface of the electron density. Regions of lower \(\rho_{\rm el}\) than the chosen iso-value are then classified as the solvent, while regions of higher \(\rho_{\rm el}\) obviously represent the DFT-treated part of the system.
\[\varepsilon(\rho_{\rm el}(\mathbf{r}))=(\varepsilon_{\infty}-1)s_{\rho_{\rm el}} \left(\rho_{\rm el}(\mathbf{r}),(\rho_{\rm el,min},\rho_{\rm el,max})\right)+1\quad, \tag{12}\]
This kind of parametrization has for instance been employed in the self-consistent continuum solvation (SCCS) model by Andreussi _et al._. Equivalently, also the iso-value itself could be used as parameter, with the transition width then as corresponding second parameter. Various smooth shape functions have been proposed in the literature, resulting, however, in quite similar predictive accuracy of molecular solvation energies. While this suggests the actual shape to be less influential for the model performance, some functions like the one proposed in the SCCS model are constructed to have an exactly zero gradient outside the transition region, which is beneficial for the numerical solution. The advantage of the electron density based approach in general is that the solvation cavity adapts self-consistently to the electron density and exhibits thus a more physically reasonable and smooth shape. Froma technical standpoint, though, dielectric functions based on the electron density are slightly more involved to implement due to additional Pulay forces arising there.
Both, atom-centered shape function and electron density based approaches are generally challenged in the description of solutes at different charge states. In the molecular context, different parameter sets defining the solvation cavity are often required for anions on the one hand, and cations and neutral molecules on the other. To overcome this limitation, Sundararaman _et al._ proposed an extended form of the dielectric function, that in addition to defining the transition region via the electron density allowed for a correction based on the locally averaged outward electric field. This field has inverse signs for cation- and anion-like regions, and thus provides the model with the fundamental capability to shift the dielectric transition region accordingly without the need to invoke different parameters. A similar approach has recently been followed by Truscott and Andreussi, who utilized the SSCS atom-centered shape function model and allowed the atomic spheres to relax their radius depending on the value of the electric field flux through their surface.
### Non-electrostatics of solvation
The interaction between solute and solvent is not solely restricted to the electrostatic mean-field treatment described in the last section, even though especially for the study of electrified interfaces changes in the electrostatic potential can be expected to be dominant. Nevertheless, it is often minute changes to free energy profiles of reactions at these interfaces that can result in crucial changes of the catalytic activity or in particular of catalytic selectivities--and for such minute changes the additional beyond mean-field and non-electrostatic interactions could prove decisive. In this section we discuss the corresponding terms in the solvation grand potential, cf. eq., the physical background for them and how they are commonly treated. As will become apparent, this treatment is generally highly effective and thus incurs in principle multiple additional parameters. Not least from a parametrization point of view, but also for reasons of computational efficiency and to exploit potential error can cellation, modern implementations in DFT packages therefore rarely calculate these terms individually. Instead, some or all of these terms are instead lumped together into empirical functions with a minimum number of parameters. Highly successful examples for this are the SMx family of methods or the SCCS approach. As it is important to understand the physical backgrounds of these terms to appreciate the origin of the added free parameters and the lumping strategies, we will nevertheless discuss each term in more detail in the following. The parametrization done in practice is then covered in Section 2.6, while a more complete overview of non-electrostatic treatments in other (not necessarily implicit) solvation models can for example be found in the recent review by Schwarz and Sundararaman or the exhaustive review by Tomasi, Menucci, and Cammi.
#### 2.4.1 Cavitation grand potential, \(\Omega_{\mathrm{is}}^{\mathrm{cav}}\)
The placement of a solute, be it a single molecule, a cluster, or an extended electrode surface always leads to the displacement of solvent molecules to form the solvation cavity. The work necessary for this displacement is commonly referred to as the cavity formation energy. It can, in principle, be calculated from explicit solvent simulations, e.g. employing Monte Carlo or molecular dynamics, or information-theoretic maximum-entropy simulations. Yet, such a costly treatment is obviously not a desirable basis for the development of a simple cavitation grand potential functional within the context of implicit solvation models.
Instead, such development relies to a large extent on scaled particle theory, which essentially employs a hard-sphere representation of solvent and solute. In this case, the formed cavity is simply the excluded volume around a solute given in terms of the hard spheres of solute and solvent molecules. For such a simplified model, \(\Omega_{\mathrm{is}}^{\mathrm{cav}}[\rho_{\mathrm{el}}]\) can then be established analytically to yield an explicit expression that depends only on molecular parameters of solute and solvent. One example is the solution of Pierotti, which is e.g.
\[\Omega_{\rm is}^{\rm cav}=k_{B}T\left\{-\ln(\zeta)+\xi\left(\frac{r_{\rm hs}}{r_{ \rm hs,solv}}\right)+\left[\xi+\frac{\xi^{2}}{2}\left(\frac{r_{\rm hs}}{r_{\rm hs,solv}}\right)^{2}\right]\right\}\quad. \tag{13}\]
Here, \(k_{B}\) is Boltzmann's constant, and both \(\zeta=\zeta(r_{\rm hs,solv})\) and \(\xi=\xi(r_{\rm hs,solv})\) are unit-less auxiliary functions of \(r_{\rm hs}\) and the solvent hard-sphere radius \(r_{\rm hs,solv}\). Note that this formulation only accounts for a single sphere type each for all solute and for all solvent species, and thus does not necessarily reflect the actual shape of the cavity very well. As a remedy, extensions to multiple different radii have e.g. been proposed by Claverie _et al._ Nevertheless, the accuracy of such scaled particle theory based approaches still rests fully on the choice of solute and solvent radii. Many approaches have correspondingly been taken to fit such radii to various experimental properties and at various experimental conditions (thereby implicitly including the grand-canonical dependence of the cavity formation on the electrochemical environment). For a comprehensive discussion of all these approaches we refer the reader to the excellent review by Tomasi and co-workers. Here we only note, that typically the cavity used to establish the expression for \(\Omega_{\rm is}^{\rm cav}[\rho_{\rm el}]\) does not resemble the solvation cavity used in the mean-field electrostatic \(\mathcal{V}_{\rm is}^{\rm mf}[\rho_{\rm el}]\). Given the effective nature of implicit solvation models, this is not _per se_ a problem. It does, however, potentially add more and unnecessary parameters.
A different approach, based on the seminal work of Uhlig, instead tries to link \(\Omega_{\rm is}^{\rm cav}[\rho_{\rm el}]\) to the solvent's macroscopic surface tension, thereby eliminating the need to define species-specific parameters altogether. Where this original formulation assumed a spherical cavity of size \(r_{\rm cav}\) around the entire solute and independence of solvent parameters beyond the surface tension, more recent formulations account for geometric properties and density of the solvent, or for deviations from the spherical shape.
\[\Omega_{\rm is}^{\rm cav}=4\pi r_{\rm cav}^{2}\overline{\gamma}\left(1-\frac{2 \delta}{r_{\rm cav}}\right) \tag{14}\]
In principle, a direct connection between cavitation energy and surface tension seems obvious, considering that a cavity is essentially an internal interface between solvent and vacuum. Yet, it is not at all clear, that such a relation also has to hold on the microscopic level where cavities are not significantly bigger than solvent molecules, or at least that \(\overline{\gamma}\) is in any sense connected to the macroscopic surface tension. Yet, a number of works have shown the Tolman equation, eq., to hold and \(\overline{\gamma}\) to be near indistinguishable from the macroscopic surface tension.
The fact remains, though, that also this approach needs parameters describing the shape of the cavity on top of those already used in the mean-field electrostatic model. This can be avoided by recognizing that the term \(4\pi r_{\rm cav}^{2}\left(1-\frac{2\delta}{r_{\rm cav}}\right)\) in eq. essentially just describes the surface area of the cavity, per definition of the surface tension as free energy per area.
\[\Omega_{\rm is}^{\rm cav}=\gamma A_{\rm cav}\quad, \tag{15}\]
In electrostatic models where the cavity is defined through a step function in the dielectric permittivity, such an area can be calculated quite straightforwardly through some form of tessellation of the surface. In models that rely on a continuous dielectric function with a smoothed transition, the surface area of the cavity seems less obvious. To this end, Scherlis _et al._ employed the concept of a quantum surface. Introduced by Cococcioni and co-workers and refined by Andreussi _etal._, this is essentially a continuous integral over the points in space, which are part of the finite transition region of the shape function \(s_{\rho_{\rm el}}\), where \(\nabla s_{\rho_{\rm el}}\neq 0\).
\[A_{\rm cav}=\int{\rm d}\mathbf{r}\nabla s_{\rho_{\rm el}}\approx\int{\rm d}\mathbf{r} \left\{\left(s_{\rho_{\rm el}}\left[\rho_{\rm el}(\mathbf{r})-\frac{\Delta}{2} \right]-s_{\rho_{\rm el}}\left[\rho_{\rm el}(\mathbf{r})+\frac{\Delta}{2}\right] \right)\times\frac{|\nabla\rho_{\rm el}(\mathbf{r})|}{\Delta}\right\}\quad. \tag{16}\]
This describes a thin film between two density iso-surfaces with a thickness \(\Delta\). The exact value of \(\Delta\) thereby proved to be unimportant as long as it is large enough to avoid numerical noise due to the real-space integration grid of the specific DFT code, and small enough to still follow the contours of the cavity.
On the plus side, based on eqs. and, \(\Omega_{\rm is}^{\rm cav}[\rho_{\rm el}]\) may then straightforwardly be determined without adding any free parameters beyond those already necessary for the electrostatic part--if indeed the macroscopic surface tension \(\gamma\) is employed. As discussed in Section 2.6, \(\gamma\) may also be seen as an empirical parameter, in which case at least still only one additional parameter would be required. This more effective view is also more consistent with a downside of the cavity definition through the quantum surface concept of eq.. Since the latter depends on \(\rho_{\rm el}\) and its gradient, additional terms arise when explicitly including a corresponding cavitation functional term in the KS DFT minimimization. For this reason, the free energy contribution due to a \(\Omega_{\rm is}^{\rm cav}[\rho_{\rm el}]\) based on eqs. and is typically only considered as a post-correction for an electrostatically optimized electron density as already discussed at the end of Section 2.2.
Finally, a conceptually related approach to this is the weighted-density cavity formation model by Sundararaman and co-workers. There, instead of a cavity composed of overlapping spheres, one formulates a solvent-center cavity, where the tails of the electron density are expanded by the van der Waals radius of the solvent molecules to gain a more physical representation of the solvent accessible area of a solute. Based on this approach one can then derive an expression for \(\Omega_{\rm is}^{\rm cav}[\rho_{\rm el}]\) that fulfills known physical limits for very small cavities or on the opposite end for droplets of solvent in vacuum.
#### 2.4.2 Exchange repulsion, \(G_{\rm is}^{\rm rep}\)
While \(\Omega_{\rm is}^{\rm cav}[\rho_{\rm el}]\) represents the thermodynamic cost of creating a cavity in the solvent for the solute to fit in, it does not include actual interactions between solute and solvent that are lost in the coarse graining of the solvent DOFs. The free energy functional \(G_{\rm is}^{\rm rep}[\rho_{\rm el}]\) is supposed to account for repulsive such interactions, predominantly arising from Pauli exchange. While there is a whole hierarchy of methods developed to treat this term, modern implicit solvation models generally employ only either of two routes, a more quantum-mechanically inspired one and a more empirical one. Recognizing that exchange repulsion originates fundamentally from the overlap of the electron densities of solute and solvent,\(G_{\rm is}^{\rm rep}[\rho_{\rm el}]\) is in the former approximated from the explicitly available electron density lying outside of the cavity or in the latter through a Lennard-Jones based metric of how close the various solute atoms could approach the cavity.
In the former more quantum-mechanically inspired ansatz, the exchange repulsion functional is specifically given as an overlap integral over the explicit DFT electron density outside the cavity with a model solvent electron density approximated as a simple Gaussian with a width \(\xi_{\rm G}\),
\[G_{\rm is}^{\rm rep}=\frac{4\pi}{\xi_{\rm G}}n_{\rm solv}^{\rm val}c_{\rm solv }\int_{>{\rm cavity}}{\rm d}\mathbf{r}\;\rho_{\rm el}(\mathbf{r})\quad. \tag{17}\]
Here, \(c_{\rm solv}\) is the constant solvent concentration and \(n_{\rm solv}^{\rm val}\) the number of valence electrons of the solvent species. The advantage of this approach is that the functional expression can be straightforwardly inserted into the KS DFT Hamiltonian. To this end, the integral over all external space of eq. is transformed into a 2D integral over the cavity surface \(A_{\rm cav}\), which is numerically solved via tesselation. The price for this simplicity is a parameter \(\xi_{\rm G}\) which largely lacks any physical motivation and with which the repulsion free energy contribution resulting from this model functional can be scaled to any desired value.
A corresponding tesselation is also the basis for the second, more empirical scheme, which essentially approximates a possible electron density overlap of solute and solvent by how close individual solute atoms come to the cavity surface.
\[G_{\text{is}}^{\text{rep}}=n_{\text{solv}}^{\text{val}}c_{\text{solv}}\sum_{j \in\text{solv}}N_{j}\sum_{\alpha\in\text{solute}}\sum_{k\in\text{cavity}}A_{ \text{cav},k}\,\mathbf{X}_{\alpha,j}(\mathbf{r}_{\alpha k})\cdot\mathbf{n}_{k}\quad.\] (18a) Next to the sum over surface tesserae, the other two sums range over all explicitly treated atoms \[\alpha\] in the solute and all chemically unique atomic species \[j\] in the solvent. \[N_{j}\] denotes the number of times the species \[j\] is contained in a solvent molecule, and the auxiliary distance vector \[\mathbf{X}_{\alpha,j}(\mathbf{r}_{\alpha k})=-\frac{1}{9}\frac{d_{\alpha j}^{} }{|\mathbf{r}_{\alpha k}|^{12}}\mathbf{r}_{\alpha k}\] (18b) encodes a Lennard-Jones type repulsive interaction between solute atom \[\alpha\] and solvent molecules, with the latter represented by the cavity units and thus at a distance \[\mathbf{r}_{\alpha k}\cdot\mathbf{n}_{k}\) apart. In the form of the Lennard-Jones \[d_{\alpha j}^{}\] for each pair of solute and solvent species, this approach adds multiple additional parameters, which need to be determined, e.g. via fitting to experimental reference values. On the other hand, the computational overhead of this approach is negligible given that most of the other contributions to the solvation free energy demand such a surface tesselation anyway.
Importantly, both methods reduce in fact again to integrals over the surface area of the cavity. This observation inspired Andreussi and co-workers to simplify the calculation of the repulsion energy even further. Making again use of Cococcioni and co-workers' quantum surface concept, they simply formulated \(G_{\text{is}}^{\text{rep}}[\rho_{\text{el}}]\) (actually only in sum together with \(G_{\text{is}}^{\text{dis}}[\rho_{\text{el}}]\) as discussed below) as linearly dependent on the electrostatic cavity surface area \(A_{\rm cav}\) and potentially also its volume \(V_{\rm cav}\),
\[G_{\rm is}^{\rm rep}+G_{\rm is}^{\rm dis}=\alpha A_{\rm cav}+\beta V_{\rm cav}\quad. \tag{19}\]
The advantage of this approach over eqs. and (18a) is its unparalleled computational efficiency (when again only evaluating it as a post-correction) and the fact that it adds only two adjustable parameters, as further discussed in Section 2.6.
#### 2.4.3 Dispersion interactions, \(G_{\rm is}^{\rm dis}\)
Similar to \(G_{\rm is}^{\rm rep}[\rho_{\rm el}]\) and indeed often treated in a very similar fashion or grouped together with it, \(G_{\rm is}^{\rm dis}[\rho_{\rm el}]\) is supposed to account for another type of intermolecular interaction between solute and solvent molecules that is lost in the coarse graining process, namely attractive dispersion. With the relevance of solute-solvent dispersion on solute structure and energetics well documented, a great number of methods have been devised to derive approximate expressions for \(G_{\rm is}^{\rm dis}[\rho_{\rm el}]\). Again, these approaches can be roughly categorized into more quantum mechanically inspired and more empirical approaches. Of the former, a popular approach, implemented e.g. in the PCM model, is based on the theory of McWeeny. Without going into too much detail--see for example ref. for a full description--and similarly to the quantum mechanical treatment of the repulsion energy, also this approach can be boiled down to an integral over the cavity surface, yet this time over the electrostatic potential and the normal component of the electrostatic field. Both are represented in the basis functions of the underlying DFT method, which, at least in localized basis function codes, tend to be not very dense near the cavity surface. Therefore, the accuracy of the quantum mechanical calculation of \(G_{\rm is}^{\rm dis}[\rho_{\rm el}]\) tends to strongly depend on the chosen basis set. Properties of the solvent and solute enter this approach in the form of a multiplicative factor that depends among others on the first ionization energy of the solvent or average electronic transition energies. In particular also the complex integrals involved in the calculation, render the overall computational cost of this approach significantly higher than that of the other non-electrostatic contributions.
For this reason, a lot of implementations opt for a more empirical approach instead. An ansatz analogous to eq. (18a) leads then to
\[G_{\mathrm{is}}^{\mathrm{dis}}=n_{\mathrm{solv}}^{\mathrm{val}}c_{\mathrm{solv}} \sum_{j\in\mathrm{solv}}N_{j}\sum_{\alpha\in\mathrm{solute}}\sum_{k\in\mathrm{ cavity}}A_{\mathrm{cav},k}\,\mathbf{X}^{\prime}{}_{\alpha,j}(\mathbf{r}_{\alpha k}) \cdot\mathbf{n}_{k}\quad,\] (20a) only now with an auxiliary distance vector that encodes a London type attractive dispersion, \[\mathbf{X}^{\prime}{}_{\alpha,j}(\mathbf{r}_{\alpha k})=-\frac{1}{3}\frac{d_{\alpha j }^{}}{|\mathbf{r}_{\alpha k}|^{12}}\mathbf{r}_{\alpha k}\quad. \tag{20b}\]
Obviously, this approach thus incurs again a set of parameters (\(d_{\alpha j}^{}\)) which need to be determined.
Finally and also in exact analogy to exchange repulsion, each of these approaches to estimating \(G_{\mathrm{is}}^{\mathrm{dis}}[\rho_{\mathrm{el}}]\) boils numerically down to a tessellation of the cavity surface. We note that instead of the here described geometric surface tessellation, one could in principle also integrate over any suitable cavity shape function, such as the aforementioned weighted density solvent-center cavity. In any case, based on the observation that \(G_{\mathrm{is}}^{\mathrm{dis}}\) is just an integral over the cavity surface, Still and co-workers proposed a simple description as a function of solvent accessible area, or indeed, the surface area of the solvation cavity. As noted above, this idea was later expanded upon in the work of Andreussi _et al._ where the dispersion functional is then described together with the exchange repulsion functional through eq..
#### 2.4.4 Thermal motion, \(G_{\mathrm{is}}^{\mathrm{tm}}\)
As discussed above, solvation of any solute generally alters that solute's PES. Foremost, one pictures this in form of an altered equilibrium structure of the solute compared to the vacuum one, such that e.g. hydrophobic groups avoid exposure to the solvent, zwitterionic structures are stabilized by polar solvents or the internal hydrogen-bond network is rearranged.
However, the altered PES could in principle also lead to changes in the vibrational modes of the solute that would correspondingly need to be accounted for through another free energy functional term \(G_{\rm is}^{\rm tm}[\rho_{\rm el}]\). For molecular solutes, this would then additionally cover changes of the solute's rotational and translational entropy. The latter do not play a role at extended SLIs, and solvent-induced changes to the vibrational modes of an adsorbate are likely small compared to those arising from the adsorption itself or from ongoing chemical reactions. Therefore, to our knowledge no implicit solvation implementations for the SLI context have hitherto explicitly considered a \(G_{\rm is}^{\rm tm}[\rho_{\rm el}]\) term.
### Electrolyte models
The theories introduced in the last two sections yield expressions for the mean-field electrostatic \({\cal V}_{\rm is}^{\rm mf}[\rho_{\rm el}]\) and non-electrostatic \(\Omega_{\rm is}^{\rm non-el}[\rho_{\rm el}]\) terms in the model grand solvation potential, cf. eq.. These expressions are already sufficient to establish implicit solvation models for pure liquids. However, real electrochemistry or electrocatalysis almost invariably works with electrolytes with finite salt concentrations. Indeed, the presence of salt can actually even be substantial for the chemical reactions and the way they proceed. As already discussed, at SLIs ions act as counter charges to compensate the surface charge of the electrode. They are thus potentially strongly enriched particularly in the inner DL close to the electrode, and their presence may not least crucially impact the stabilities of reaction intermediates. In this section, we therefore continue with the extension of implicit solvation models to electrolytes and notably Poisson-Boltzmann (PB) theory, which forms the unanimous basis for most of these extensions to day. Practically, this proceeds again by deriving computationally tractable or closed expressions for the contributions grouped into the ion grand potential term \(\Omega_{\rm is}^{\rm ion}[\rho_{\rm el}]\) of eq..
#### 2.5.1 Poisson-Boltzmann theory
In a dilute electrolyte solution, one may reasonably assume that the solvent dielectric response is not (significantly) modified by the small ion concentrations, and interactions between the generally quite distant ions can be well described on a mean-field level. In the DFT supercell, one realization of such a dilute electrolyte could be to simply place a small number of point-like ions at fixed and not too close positions to each other inside the implicit solvent part. For a corresponding static ion charge distribution \(\rho_{\rm ion}=\rho_{\rm ion}(\mathbf{r})\), as well as under the mentioned assumption of unmodified solvent dielectric response and only mean-field ion-ion interactions, then the only term that we would have to consider in \(\Omega_{\rm is}^{\rm ion}[\rho_{\rm el}]\) is a straightforward potential energy functional \(\mathcal{V}_{\rm is}^{\rm mf,ion}\).
\[\left(\mathcal{V}^{\rm mf}+\mathcal{V}_{\rm is}^{\rm mf}+\mathcal{V}_{\rm is}^{ \rm mf,ion}\right)[\rho_{\rm el}]=\frac{1}{2}\int\left(\rho_{\rm el}+\rho_{ \rm nuc}+\rho_{\rm ion}\right)\phi\quad. \tag{21}\]
As visualized in such static ion distributions are indeed employed in so-called planar counter charge (PCC) models that are specifically developed for the description of planar SLIs and that we will further motivate in Section 2.5.6. However, for the general objective of deriving functional expressions for an electrolyte that is equilibrated in its response to the given electrode or solute with its (DFT) net charge density, it makes no sense to manually ascribe fixed ion positions. Indeed, the equilibrated ion density should be a result of the theory, and not an input. Therefore, this equilibrated density will have to adapt to the electrostatic potential, to which the ions though actually contribute themselves. This already shows that in such a case self-consistency between \(\rho_{\rm ion}\) and \(\phi\) in eq. has to be reached. Most commonly, a corresponding self-consistent description of the ion distribution is achieved within the famous PB theory, which treats the ions as a gas that interacts only via mean-field electrostatic interactions within the continuum dielectric. Referring to dedicated accounts on PB theory for full derivations and a full appraisal, we here only compile the**Schematic representation of various electrolyte models currently used for the description of SLIs. Planar counter charge (PCC) models place rigid ions in a Helmholtz layer like arrangement, while Poisson Boltzmann (PB) models determine the ionic distribution self-consistently in the total electrostatic potential. Various important modifications of PB theory are highlighted and discussed in the text.**
For simplicity, we furthermore focus here and in the remainder of this electrolyte section also on an electrolyte with a cationic concentration \(c_{+}=c_{+}(\mathbf{r})\) due to only one cation species of mass \(m_{+}\) and charge \(+z\), and an anionic concentration \(c_{-}=c_{-}(\mathbf{r})\) due to only one anion species of mass \(m_{-}\) and charge \(-z\).
\[\Omega_{\mathrm{is}}^{\mathrm{ion}}[\rho_{\mathrm{el}}]=\mathcal{T}_{\mathrm{ is}}^{\mathrm{PB,ion}}[\rho_{\mathrm{el}}]+\mathcal{V}_{\mathrm{is}}^{\mathrm{mf,ion}}[ \rho_{\mathrm{el}}]-TS_{\mathrm{is}}^{\mathrm{PB,ion}}[\rho_{\mathrm{el}}]\quad,\] (22a) now includes kinetic and, applying the famous Sackur-Tetrode equation, also entropic contributions. For the \[+z/-z\] electrolyte they read \[\mathcal{T}_{\mathrm{is}}^{\mathrm{PB,ion}} =\frac{3}{2}k_{\mathrm{B}}\int\left\{c_{+}+c_{-}\right\}\mathrm{ d}\mathbf{r}\] (22b) \[S_{\mathrm{is}}^{\mathrm{PB,ion}} =-k_{\mathrm{B}}\int\biggl{\{}c_{+}\left[\ln\left(c_{+}\lambda_{ +}^{3}\right)-\frac{5}{2}\right]+c_{-}\left[\ln\left(c_{-}\lambda_{-}^{3} \right)-\frac{5}{2}\right]\biggr{\}}\mathrm{d}\mathbf{r}\quad,\] (22c) with \[\lambda_{\pm}=\frac{h}{\sqrt{2\pi m_{\pm}k_{\mathrm{B}}T}}\] the thermal wave length. The potential energy functional still holds as before in eq. ( 21 ), of course, now with the ion charge density given as \[\rho_{\mathrm{ion}}=z(c_{+}-c_{-})\].
The starting point to obtain the self-consistent ion concentrations and electrostatic potential to evaluate these functional expressions is as before the GPE. Within the prevalent isotropic, linear and local dielectric model, and under the already mentioned assumption that the dielectric response of the solvent is not changed by the ion density, the electrolyte charge distribution can straightforwardly be added to the GPE of eq., simply by extending the source terms on the right hand side
\[-\varepsilon_{0}\nabla\left[\varepsilon(\mathbf{r})\nabla\phi\right]=\left(\rho_{ \mathrm{el}}+\rho_{\mathrm{nuc}}+\rho_{\mathrm{ion}}\right)\quad. \tag{23}\]
PB theory then additionally makes the assumption that in the equilibrated electrolyte the ions are Boltzmann-distributed in the electrostatic potential
\[c_{\pm}=c_{\infty,\text{ion}}\exp\left(\mp\frac{z\phi}{k_{\text{B}}T}\right)\quad, \tag{24}\]
Within the mean-field PB gas ansatz, \(c_{\infty,\text{ion}}\) is in turn readily related to the bulk ion electrochemical potential via \(\tilde{\mu}_{\text{ion},\pm}=k_{\text{B}}T\ln\left(\lambda_{\pm}^{3}c_{\infty, \text{ion}}\right)\). Inserting eq. into the ion-including GPE of eq., leads finally to the famous PB equation (PBE) itself
\[\varepsilon_{0}\nabla\left[\varepsilon\nabla\phi\right]=-\left(\rho_{\text{el} }+\rho_{\text{nuc}}+2zc_{\infty,\text{ion}}\sinh\left(-\frac{z\phi[\rho_{\text {el}}(\mathbf{r})]}{k_{\text{B}}T}\right)\right)\quad, \tag{25}\]
In practice, this PBE can thus be implemented into the KS-DFT minimization in a way completely analogous to the GPE of the ion-free case, and then be solved at each electron density optimization step. However, due to the complicated non-linear nature of the PBE, and the associated computational cost of solving it, it is popular to instead solve a simplified linearized version of it. This linearized Poisson-Boltzmann equation (LPBE) can be obtained by truncating a Taylor expansion of the sinh-term in eq.
\[\varepsilon_{0}\nabla\left[\varepsilon\nabla\phi\right]=-\left(\rho_{\text{el} }+\rho_{\text{nuc}}-\frac{2z^{2}c_{\infty,\text{ion}}}{k_{\text{B}}T}\phi[\rho _{\text{el}}(\mathbf{r})]\right)\quad. \tag{26}\]
The corresponding LPBE grand potential functional terms can then be derived analogously from a Taylor expansion of eq..
As mentioned at the beginning of this sub-section, the assumptions underlying PB theory restrict its formal range of applicability to dilute electrolytes. Close to electrified SLIs, however, high ion concentrations may accumulate even for electrolytes that are indeed dilute in the bulk. This motivates corrections to PB theory that account for then increasedion-ion and ion-solvent correlations. Despite the non-locality of these interactions, a series of local ion density approximation models have been proposed to keep the simplicity of the PB model intact. They are summarized in above and will be briefly outlined in the following.
#### 2.5.2 Finite ion size corrections
In the original formulation of PB theory, ions are point-like. This means, that for stronger fields local ion concentrations could in principle reach unphysically high values. An immediate fix to this problem is to simply give ions a finite size, which then leads to size-modified PB (MPB) theory. While MPB can be derived in various ways, the most physically intuitive derivation is based on a lattice model with a uniform cell size \(a\) for (solvated) ions and solvent molecules, where each lattice site can at max hold only one particle, cf. This way, the lattice mimics short-range ion-ion repulsion and by construction does not allow unphysically high local ion concentrations. For this model, an ion grand potential functional can be developed using the configurational partition function of solvent molecules and ions, and then applying a mean-field approximation.
\[\mathcal{T}_{\text{is}}^{\text{MPB,ion}} =0 \tag{27}\] \[S_{\text{is}}^{\text{MPB,ion}} =-k_{\text{B}}\int\biggl{\{}c_{+}\left[\ln\left(c_{+}a^{3} \right)-1\right]+c_{-}\left[\ln\left(c_{-}a^{3}\right)-1\right]\] \[\quad+\left(\frac{1}{a^{3}}-c_{+}-c_{-}\right)\ln\left(1-c_{+}a^ {3}-c_{-}a^{3}\right)+c_{+}+c_{-}\biggr{\}}\text{d}\mathbf{r}\quad. \tag{28}\]The ion concentrations now effectively follow a Fermi-Dirac-like statistics due to the maximum occupancy of the lattice cells
\[c_{\pm}=c_{\infty,\text{ion}}\frac{\exp\left(\mp\frac{z\phi}{k_{\text{B}}T}\right) }{1-2c_{\infty,\text{ion}}a^{3}+2c_{\infty,\text{ion}}a^{3}\cosh\left(\frac{z \phi}{k_{\text{B}}T}\right)}\quad, \tag{29}\]
In direct analogy to the unmodified PB case, inserting these concentrations into the GPE of eq. leads then to the so-called MPB equation, which in turn gives the desired functional relation between electrostatic potential and electron density and which can be solved within the KS DFT minimization as before.
The MPB ion concentration profiles converge to the PB profiles for \(a\to 0\), if the same bulk electrochemical potential reference is used. One can easily see that this then implies \(a=\lambda_{+}=\lambda_{-}\), i.e. within PB theory the thermal wavelength plays the same role as the ion size in MPB theory. The two theories thus have a common algebraic origin, but a different physical one. Indeed, in contrast to PB theory and as seen in eq., the MPB model lacks an ion kinetic energy functional. Yet, since it equally lacks a corresponding entropy contribution from the ionic motion and with the two terms canceling each other in PB theory, the same functional form is nevertheless recovered in both theories for the small ion size limit. Note also that the original MPB theory, and the corresponding equations above, were developed for equally sized cations and anions. It has since been extended to asymmetric electrolytes, e.g. by extending the statistical lattice model with sublattices, by introducing potential-dependent ion sizes, or by other means. There are also efforts to go beyond the lattice approximation, deriving functional expressions from experimental equation of state data or from equations defining atomic or molecular interactions such as closure relations to the Ornstein-Zernicke equation, cf. Section 2.1. Notwithstanding, the resulting energy functionals are generally still based on a local approximation for the ion density. Such local approaches to ion-ion interactions offer generally a simple correction to PB theory for those situations where ions are crowded, e.g. due to strong electric fields. However, in case of strong variations of the ion concentration profiles the description of \(\Omega_{\rm is}^{\rm ion}[\rho_{\rm el}]\) as a local functional of ion concentrations may break down altogether. Such cases may then necessitate a more involved non-local treatment.
#### 2.5.3 Ion-induced solvent structuring
One important physical effect of the ions completely omitted so far is simply the fact that in regions with high ion concentrations few solvent molecules may reside, and if they do they are likely highly structured around the ions. In such situations the dielectric continuum approximation for the solvent likely breaks down and ion interactions become much more specific than the the hitherto included mean-field electrostatics.
Accounting for this effect, Burak and Andelman derived a corrective short-range ionic interaction potential contribution to the mean-field electrostatic potential from Monte Carlo simulations. By truncating the virial expansion of the PB partition function after second order, they were then able to derive a simple analytic expression for the free energy. Although this direct expansion approach is of great interest for the development of improved PB-based theories, it is less practical due to the required knowledge of the system-dependent fluctuating short-range potential. An approach that is in this respect more in the spirit of effective parametrized continuum models has been put forward by Bohinc, Shrestha and May. There, they represented the additional short-range forces by a parametrized Yukawa potential, arriving at a simple correction to PB theory. Next to this, several other approaches have been developed in the past for which we refer the interested reader to an extensive review on this topic.
All of the above corrections share the fact that the resulting corrections are non-local in the sense that solvent structure at a point in space is also influenced by the ion concentration in its vicinity (e.g. via the aforementioned short-range potentials). A much simpler and local variant to correct for ion-induced solvent structuring is the dielectric decrement approach,cf. From simulations and various experimental works, it has been found that the isotropic dielectric permittivity of water varies linearly with the salt concentration at small to medium (1.5 M) salt concentrations,
\[\varepsilon_{\infty}(c_{\infty,\text{ion}})=\varepsilon_{\infty}(c_{\infty,\text {ion}}=0)+\beta c_{\infty,\text{ion}}\quad, \tag{30}\]
The equation can be also written as a function of the local salt concentration, but importantly anionic and cationic contributions cannot be separated without making further assumptions, due to a lack of experimental data. In any case, based on this evidence for water one could simply consider \(\beta\) as a further variable parameter and modify the dielectric function employed for the mean-field electrostatics, cf. Section 2.3.2, to additionally depend (linearly) on the local ion concentration. In the SLI context with aqueous electrolytes, this effective dielectric decrement approach also enjoys recent popularity to model a stronger (sometimes ice-like) water structuring in the inner DL through an accordingly reduced dielectric permittivity in that region. In summary, there are thus a number of methods of varying degrees of complexity that allow to mimic an ion-induced local structuring of the solvent around a solute. Thereby, they extend the validity of PB or MPB approaches to higher ion concentrations, yet often at the price of additional parameters that need to be determined.
#### 2.5.4 Coulombic ion correlations
In the case of small ion concentrations with thus effectively large ion separations and in strongly screening solvents like water, the mean-field interaction between the dissolved ions assumed in PB theory is generally a good approximation. However, it may quickly break down for solvents with smaller dielectric permittivity, for higher ion concentrations (such as in ionic liquids), or for multi-valent ions with stronger Coulomb forces. In these cases, the electrostatic force is much more non-local and fluctuating, leading for example to the effect of overscreening at charged interfaces. Overscreening in the context of electrified SLIs refers to the presence of higher amounts of counter charge in the electrolyte close to the electrode than needed to compensate the surface charge, followed by a smaller net charge of opposite sign to satisfy overall electroneutrality.
An account for the corresponding ion correlations requires in general field theoretical approaches, using loop expansions, which lead to substantially more complicated expressions than in PB theory. A promising, more approximate approach by Bazant _et al._ instead leads to a simple correction of the mean-field electrostatic potential energy, cf. eq., in form of one added term
\[\mathcal{V}_{\mathrm{is}}^{\mathrm{non-mf,ion}}[\rho_{\mathrm{el}},\phi]=- \int\frac{\varepsilon_{0}\varepsilon(\mathbf{r})}{2}l_{c}^{2}(\mathbf{\nabla}^{2} \phi)^{2}\mathrm{d}\mathbf{r}\quad, \tag{31}\]
This demonstrates nicely that the electrostatic energy is lowered due to overscreening by enhancing the curvature of \(\phi\). The theory was shown to give overscreened ion distribution profiles in close agreement with molecular dynamics simulations resulting in realistic estimates of the potential-dependent capacitance compared to experimental reference data.
#### 2.5.5 Ion-solute interaction and Stern layer formation
PB and MPB theories as well as their extensions are usually derived without the actual presence of the solute. Therefore, the only coupling between solute and ions is the hitherto discussed mean-field electrostatic coupling. Just as highlighted in Section 2.4 for the pure liquid, this neglects additional interactions between the ions and the solute that were for the solvent summarized in the \(\Omega_{\mathrm{is}}^{\mathrm{non-el}}[\rho_{\mathrm{el}}]\) term in eq.. A prominent such non-electrostatic correction for the ions would be an additional repulsive contribution which prevents ions to approach the solute too closely. In protic solvents, the formation of a corresponding ion-freesolvent region called Stern layer is for instance a consequence of the large size of the hydrated cations.
Most straightforwardly, this kind of physics can be implemented by an additional repulsion potential added on top of the mean-field potential. Alternatively, the repulsion potential can simply be expressed as an exclusion function \(\alpha_{\rm ion,\pm}=e^{-\frac{\phi_{\pm}^{\rm rep}}{k_{\rm B}T}}\) for cations and anions, respectively. The exclusion function prevents the ions from approaching the solute to a certain distance, similar to the dielectric shape function controlling the solvent's place of closest approach (see ref. for a full derivation following the statistical lattice approach).
**Creation of an ion-free Stern layer around a molecular solute.** Compared is the solvation environment around the center of mass (COM) of naphthalene in a 2.18M NaCl solution as obtained from explicit molecular dynamics simulations (dashed lines) and with a Stern-layer corrected implicit MPB model (solid lines). Data in red represents the spherically-averaged radial distribution function (RDF) of the oxygen atoms in the explicit water solvent (\(g_{\rm H_{2}O}\)) and the corresponding spherically-averaged dielectric function \(\varepsilon\) for the implicit model. Data in black are the spherically-averaged RDF for the ions and the corresponding ion-exclusion function \(\alpha_{\rm ion}\). Both the onset of the solute solvation shell and the radial Stern layer shift of the ionic distribution are rather well reproduced. To better grasp the involved scales, two dotted vertical lines illustrate the radial distance to the molecule COM as shown in the top view in the inset. Adapted with permission from ref., American Institute of Physics (AIP).
leads to a modified set of ion concentration functions, which for the case of a +z/-z electrolyte read
\[c_{\pm}=c_{\infty,\text{ion}}\alpha_{\text{ion},\pm}[\rho_{\text{el}}]\frac{\exp \left(\mp\frac{z\phi}{k_{\text{B}}T}\right)}{1-2c_{\infty,\text{ion}}a^{3}+2c_{ \infty,\text{ion}}a^{3}\alpha_{\text{ion},\pm}[\rho_{\text{el}}]\cosh\left( \frac{z\phi}{k_{\text{B}}T}\right)}\quad. \tag{32}\]
For the sake of convenience the functional form for \(\alpha_{\text{ion},\pm}\) can be chosen identical to the dielectric shape function \(s_{\rho_{\text{el}}}\), to vary in between 0 in the ion-free region and 1 in the ion-contained electrolyte region, yet with a different cutoff parameter that can be individually tuned. This simplified model was shown to be able to account for short ranged ion-solute interactions, at the price of the additional cutoff and shape parameters. As detailed in Section 2.6, a careful tuning of these parameters to reproduce molecular experimental reference data yields a plot like in Fig. 6, which nicely illustrates the achieved creation of an ion-free Stern layer close to the solute with an extent and location that agrees well with the results of molecular dynamics simulations with explicit solvent. Below, we will refer to a MPB model that additionally provides such a Stern-layer functionality as S-MPB model.
#### 2.5.6 Planar counter charge models
All the ion models presented so far are based on diffuse layer theory, essentially assuming mobile gas like ions that migrate and equilibrate in a mean-field potential. While these more physical approaches are most valuable for a general treatment of solvation, the aforementioned, much simpler PCC approach to place rigid ions into the supercell, cf. Fig. 5, is of particular interest and convenience for the context of planar SLIs. As further discussed in Section 3.4 below, its primary purpose is to introduce a counter charge distribution that exactly compensates a net charge of the electrode to achieve an overall charge-neutral supercell. As the name says, this ionic counter charge distribution is simply modeled as a smoothed out Gaussian charge plane. The advantage of this method is that the ionic charges can be freely shifted in space, thereby providing some flexibility, e.g. in modeling asymmetric DFT supercells with only one slab side exposed to the electrolyte. Physically, the PCC model resembles if at all the situation in the inner DL, where the ions are assumed to be highly crowded. In dominantly studied aqueous electrolytes, the obvious crudeness of this approach is fortunately to some extent remedied by the strong screening capabilities of the polar solvent, which renders the DL potential drop less sensitive to the exact location of the ions.
### 2.6 Parametrization
As apparent from the presentation so far, implicit solvation methodologies come invariably with a set of (in principle system-dependent) parameters, and it is these parameters that crucially determine the accuracy of this highly effective approach to solvation. To recap the previous sections, parameters arise generally in the functional expressions accounting for electrostatic, non-electrostatic and ionic contributions. In the standard linear, local and isotropic dielectric formulation of the electrostatic contribution, parameters are needed to define the location of the solvation cavity, or additionally the dielectric transition region. As discussed in Section 2.3.2, these can be atomic radii in the case of spatially parametrized dielectric functions, or iso-values of the electron density in the density-dependent case. The non-electrostatic energy functional also gives rise to a varying number of parameters that depends strongly on the models of choice. If the models separately account for cavitation, dispersion or repulsion, quite a large number of parameters can quickly arise. In contrast, the simplified SCCS model of Andreussi _et al._ lumps all of this into just two parameters that scale the solvation cavity volume and surface. Lastly, the ionic energy functional comes with its own number of parameters to express deviations from the PB theoretical description. These can either be the introduction of a finite ion size parameter, a parameter to describe the Stern layer and thus solute-ion interactions, or parameters to describe ion-solvent interactions in form of a dielectric decrement.
Depending on the model complexity and the way it considers these various contributions,a largely different total number of parameters can result. This number can range from just four in the minimal SCCS model to 64 parameters in the popular SMD model. These parameters may then be determined to give an optimum account of maybe only a single solvent/solute combination, maybe a particular solvent (e.g. water in the case of the SCCS model), or aiming at maximum transferability for a whole range of solvents (as in the SMD case). At the same time, due to the fitting procedure, there is always the possibility of some degree of error cancellation, when for instance the non-electrostatic contributions compensate for some of the shortcomings of the DFT functional itself. A key question is thus, to which degree the use of parameters can improve the physics and transferability of the implicit solvation model. This question relates directly to the size, quality and information content of the available training data to which parameters can be fitted.
In principle, training data should be selected that is as close as possible to the intended application, in this case SLIs, or if possible even electrified SLIs. Unfortunately and as further discussed in Chapter 3, experimental reference data is very rarely available for these systems, and if it is, it is often not suited for the parametrization of a microscopic solvation model. The main reason for this is that at SLIs various other effects overlap with pure solvation contributions as we will see in the next chapter below. In contrast, molecular solvation data is much more widely available, at least for water as a solvent. Therefore, many implicit solvation studies on SLIs have adapted parameters that have originally been derived from a parametrization to such molecular data. This is not only critical from the viewpoint of the largely different chemistries, involving solutes composed of light organic elements in one case and extended electrodes composed of heavy transition metals in the other. There are also practical problems that arise not least from the different dimensionality of the problem. For instance and as outlined in Section 2.4.2 above, non-electrostatic contributions can be expressed as a function of the volume of the solvation cavity, which in turn is not defined for extended interfaces. Such issues have led to the formulation of functional expressions that only involve quantities compatible with SLIs. In the mentioned example, this is a non-electrostatic model that only considers the cavity surface and not the volume. Nevertheless, also such models are then re-parametrized using the existing molecular databases. The performance of the cavity-surface model was there found to be similar to the original one including the cavity volume. While this suggests that a parametrization of SLI-compatible models is possible, it still does not tell how well the molecular parameters will transfer to the SLI context and we will come back to this issue in the next chapter.
If one accepts that the parametrization is done with molecular data, the next obvious questions are which and how much of such data is available, how diverse the database is in terms of a wide range of molecular chemistries and whether the tabulated quantities are in fact really suited for the parametrization at hand. In the long, independent history of molecular solvation modeling, these questions have been satisfactorily addressed through the built-up of databases of primarily experimental solvation free energies. As apparent from Table 1, these databases are indeed sizable and partly contain data for a wide variety of solvents. Out of these, the Minnesota solvation (MNSol) database was among the first to provide experimental solvation energies of a wide range of over 600 neutral organic molecules in over 100 different solvents.
\begin{table}
\begin{tabular}{c|c|c|c|c} Database & \# Solvents & \# Solutes & \# Hydration & \# NAQ solv. energies \\ FreeSolv (v0.51) & 1 (W) & 643N & 643 & 0 \\ Wang & 1 (W) & 668N & 668 & 0 \\ Rizzo-DGHYD & 1 (W) & 538N/52C & 603 & 0 \\ Kelly & 1 (W) & 106C & 106 & 0 \\ MNSol (v2.0) & 106 (W,NAQ) & 662N & 389 & 2648 \\ (v1.0) & 145 (NAQ) & 658N & 0 & 5952 \\ CompSol & 732 (W,NAQ,IL,T) & 863N & 397 (581*) & 3786 (13386*) \\ \end{tabular}
\end{table}
Table 1: **Databases containing experimentally measured solvation energies of molecular solutes at room temperature.** C = charged, N = neutral, W = water, NAQ = no-aqueous solvents, IL = ionic liquids, T = temperature dependence. The numbers were extracted from the respective databases directly. The Wang database is to a large part constructed from the FreeSolv database. *Only solvation energies evaluated at \(25\pm 2^{\circ}\)C have been considered, while the number in brackets refers to the complete number of solvent-solute combinations for which at least one temperature data is available.
The FreeSolv database is currently the largest collection of neutral molecule solvation free energies in water (then called hydration energies). It consists of over 600 entries and should thus allow a meaningful parametrization even of more complex models.
Within the implicit solvation framework defined in this review, a molecular solvation free energy is calculated as
\[\Delta G_{\rm solv}=\Omega_{\varepsilon_{\infty}={\rm solv}}^{N_{\alpha}}[\rho_ {\rm el}^{\circ}]-\Omega_{\varepsilon_{\infty}=1}^{N_{\alpha}}[\rho_{\rm el}^ {\circ}]\quad, \tag{33}\]
Note that it is awkward to see a free energy on the left hand side of the equation, and a difference of grand potential energies on the right hand side. We here simply attest to the fact that (measurable) solvation free energies are generally seen as a property of the full (macroscopic) system and not of the grand-canonical sub-system technically employed in the calculations. Starting our survey of model performance with the ubiquitous solvent water, implicit solvation models trained by these molecular databases can typically predict hydration energies with a mean absolute error (MAE) between 0.6 kcal/mol (large parameter space models like SMD or SM8) up to 1.2 kcal/mol (small parameter space models like the SCCS model). In general, though, these numbers are difficult to compare, since rarely the same set of training and test molecules have been used, see e.g. ref. 219 for a notable exception. Nevertheless, it generally seems that standard implicit solvation models have had a hard time to decrease the accuracy below about 0.5 kcal/mol, which could thus somehow mark what can realistically be expected at such high level of coarse graining. Recent reports of ground-breaking 0.14 kcal/mol MAEs with new machine-learned implicit solvation models have thus also to be seen in light of the actual accuracy of the underlying experimental data. Different solvation databases have been found to have an error of up to 0.25 kcal/mol relative to each other, which agrees with the experimental error that is estimated for solvation energies of neutral solutes based on measured partition coefficients. Too highly parametrized models could therefore run the risk of overfitting of experimental errors. A connected problem is the occurrence of solutes in the training set which are reactive in solution. The optimized SCCS model was, for example, found to perform well for most molecular components, apart from carbonic acids and amines. These are precisely those compounds which are mostly present in solution in their dissociated form at associated vastly different solvation energy. This highlights the importance of a careful curation of the reference databases.
Next to the simulation of aqueous solvation, also non-aqueous solvents are of high importance for electrochemistry, such as e.g. for lithium ion batteries or the electrocatalytic reduction of CO\({}_{2}\).
**Number of solvation energy entries (“training set size”) per non-aqueous solvent in the three largest corresponding experimental databases. The solvents are sorted according to their largest training set size in all of the three databases. Training set sizes below 50 are prone to significant overfitting errors, in particular if the solutes are not homogeneously distributed over the chemical space.**
correspondingly compares the MNSol, and COMPSOL databases regarding this amount of available solvation energies for each non-aqueous solvent. It is apparent that for most of these solvents the databases actually contain less than 50 solvation free energy entries. As pointed out by Hille _et al._, such small training set sizes for the implicit solvation model can result in significant overfitting of the data. This would likely reduce the transferability even for those implicit solvation models that have only a few fitting parameters. Unfortunately, the situation is even worsened by an often low chemical diversity of the organic solutes contained in these small test sets, which may further lead to bias in the achieved parametrization. These issues provide a motivation especially for smallest parameter space implicit solvation models that rather trade quantitative accuracy with a somewhat robust extrapolation outside of the small training regime. As we will elaborate further in Chapter 3 below, this objective fits actually very well with the realization that the primary value of implicit solvation modeling at SLIs is presently more the provision of a counter-charge model than the actual account of solvation effects. Within this perspective, a recent reformulation was able to reduce the parameter space of the SCCS model to a single non-electrostatic parameter, while still resulting in a reasonably accurate prediction of solvation free energies for most solvents. This one parameter can furthermore be estimated from the solvent bulk dielectric permittivity, enabling the prediction of solvation free energies for arbitrary solvents with known permittivity.
The problem of small training set sizes becomes even more critical when transitioning to solvation free energies of charged solutes. Estimating the solvation energy of ions is a key challenge of high priority, as charged systems appear constantly as reactants or reaction intermediates in electrochemistry, but also e.g. in bio or organic chemistry. The solvation energy of dissociated acids for example is important for the estimation of the acid dissociation constant, or the solvation energy of charged redox species for the calculation of redox potentials. Due to their large electrostatic stabilization, charged solutes exhibit -solvation energies that are an order of magnitude larger than the ones of neutral solutes. However, the experimental measurement of single molecule ionic solvation energies requires thermodynamic cycles and the knowledge of the absolute solvation free energy of an arbitrary reference ion, usually a proton. Especially, the latter has been found to be prone to errors up to 2 kcal/mol. Nevertheless, various attempts have been made to parameterize implicit solvation models to ionic solvation data, as e.g. the Rizzo-DGHYD database containing 52 solvation energies of cations and anions. In using this database for parametrizing the SCCS model, Dupont _et al._ found substantially different cavity parameters for anions, cations and neutral molecules, which one would generally avoid for the modeling of SLIs with a varying charge state depending on the applied potential. This issue may be due to two drawbacks of the original SCCS approach. First, relying on the electron density to define the dielectric function, cf. Section 2.3.2, means that anions show significantly larger cavities compared to cations and neutral molecules at comparable density iso-values. Furthermore, SCCS does not account for explicit solvent-solute correlation, which becomes significant in the case of high fields near localized charges. As noted above, electric field corrected dielectric functions have shown promise to go beyond this limitation.
Next, in an attempt to increase the training data for implicit solvation models, the CompSol database has recently been published adding further quantities beyond the traditional solvation free energies at standard state. These are mainly temperature-dependent solvation free energies and solvation energies in ionic liquids. In terms of solvation energy entries, this database is now by far the largest. In order to make full use of it though, the implicit solvation model actually has to be able to somehow account for the additional physics in this data. Indeed, temperature-dependent solvation data could provide an additional constraint on the functional form of electrostatic, cavity or dispersion energy contributions, all of which could in principle depend on the temperature. While some implicit solvation models incorporating temperature effects have been communicated, these have not found their way into widespread use to date.
As a final point, we note that all of the databases discussed above focus on solvation free energies for vanishing ionic concentrations in the solvent. They are thus not suited for the determination of the ionic parameters appearing in electrolyte models that go beyond the plain PB approach. Ionic parameters suffer therefore presently from the highest scarcity of reference data. One remedy is to realize that finite salt concentrations are known to alter the solvation free energy of neutral solutes in aqueous solution nearly linearly.
\[\Delta G_{\rm solv}(c_{\infty,{\rm ion}})-\Delta G_{\rm solv}=k_{\rm s} \frac{k_{\rm B}T}{\log_{10}({\rm e})}c_{\infty,{\rm ion}}\quad, \tag{34}\]
From tabulated Setchenow coefficients, one can estimate that a 1 M ion concentration in the electrolyte decreases solvation free energies by 0.1-0.3 kcal/mol, where the dominating effect is the energy penalty to create the ionic cavity. Such apparently small changes to the solvation free energies can still have critical consequences, at least if one thinks of biochemistry where they are known to induce protein folding. Accounting for these changes may also be key in fitting implicit models to experimental solvation data with possibly finite salt concentrations. In this respect, tabulated Setchenow coefficients actually represent a direct way to determine the ionic parameters of the implicit solvation model. Ringe _et al._ used such a database of experimentally tabulated Setchenow coefficients to optimize the electron density cutoff that controls the ionic cutoff function \(\alpha_{\rm ion}^{\pm}\) using an SCCS/S-MPB model. This density cutoff was found to be correlated with the hydration number of the ions in the solution, showing that the parametrized model was able to predict ion-specific hydration effects for neutral molecules.
To recapitulate, most contemporary implicit solvation models applied to the simulation of SLIs tend to use parameters derived from molecular solvation databases as a basis. While the actual transferability of these parameters to the SLI context is still unclear, there is generally at least sufficient molecular experimental reference data available to achieve a coherent parametrization for water as a predominant solvent also in interfacial electrocatalysis applications. The situation worsens quickly for non-aqueous solvents and is critical for ionic parameters. While this sets a perspective for the complexity of implicit solvation modeling one can aspire to, it clearly shows that even in the context of molecular solvation there is still room for improvement through the establishment of larger and chemically more diverse reference data bases including entries beyond standard state solvation free energies.
### Implicit solvation implementations in DFT program packages
As we have shown in the previous sections, a wide variety of implicit solvation methodologies exist and it is their recent implementation into DFT program packages that can also deal with extended surfaces (e.g. through the use of periodic boundary condition supercells) that has enabled such kind of modeling for the SLI context at all.
Table 2 shows a compilation of the implicit solvation methods and features that have been implemented into various state-of-the-art, periodic and non-periodic DFT program packages at the time of submission of this article. It clearly shows a tendency of predominant use of local, linear and isotropic dielectric models. All-electron DFT packages traditionally use sharp apparent surface charge (ASC) models, cf. Section 2.3.2, instead of smooth dielectric models. This is partly due their logarithmic integration grid structure to resolve localized core basis functions, which makes the solution of the GPE or PBE over the whole computational domain numerically difficult. ASC models have been implemented also in connection with the LPB equation to simulate simplified SLIs. Among all these realizations of implicit solvation models, FHI-aims has been the first all-electron DFT package implementing a smooth dielectric response model (SCCS), extended by an advanced Stern-layer and ionic size (lattice model) corrected PB (S-MPB) ion representation. It also introduced an efficient Newton solver linearizing the S-MPB equation, which has recently also been adapted by other DFT packages.
solvation at vacuum-water or liquid-liquid interfaces) via ASC methods. Q-Chem has also recently implemented the smooth dielectric S-MPB model to support modeling of electrolytes.
The so far discussed program packages are ideal for the simulation of electrochemistry of finite-size nanoparticles. The implicit solvation simulation of extended metallic electrodes, realized by supercells with periodic boundary conditions, has been made available in several ASC schemes, implemented in CRYSTAL, GAUSSIAN and Dmol. However, this domain is clearly dominated by pseudo-potential and then mostly plane-wave based DFT program packages, with their efficient Fourier-transform algorithms for periodic systems. Here, QUANTUM ESPRESSO and JDFTx are presently the clear leaders with most advanced implementations of solvation and ion models. In addition, QUANTUM ESPRESSO provides most sophisticated correction schemes for removing periodic boundary condition in the normal direction of the surface slab to avoid artificial slab-slab interactions. VASP, arguably the most popular DFT code in the theoretical electrochemistry community, provides so far only basic implicit solvation functionality, but at least also a LPB solver which provides counter charges and then allows simulations of charged interfaces. As pointed out in recent works, the use of VASP requires special care though due to the not self-explanatory shifting of the electrostatic potential and also problems with the dipolar slab correction which is supposed to correct slab-slab interactions across periodic boundaries. Other packages, such as GPAW, ONETEP and BIGDFT have recently also reported the required implementations for S-MPB based implicit solvation models and should thus also be valid options for future modeling of electrified interfaces.
Implicit solvation models applied to electrified SLIs
### _Ab initio_ thermodynamics framework
Having established all methodological ingredients to implicit solvation schemes in Chapter 2, we now proceed to discuss their application in the context of electrified SLIs, and in particular at metal electrodes. Already in the introduction we had motivated that SLI applications presently focus predominantly on thermodynamic quantities, but that the actual DFT supercell typically only represents a grand-canonical sub-system in equilibrium with the general and electrochemical environment. In order to evaluate the true thermodynamics in SLI applications, it is therefore generally not sufficient to consider the hitherto discussed grand potential functional \(\Omega^{N_{\alpha}}[\rho_{\rm el}]=F[\rho_{\rm el}]+\Omega_{\rm is}[\rho_{\rm el}]\), cf. eq. (1a). This grand potential accounts for the exchange of all implicitly treated solvent particles and ions with their reservoirs and does therefore already depend on the electrochemical environment. However, this is only for one fixed chemical composition \(N_{\alpha}\) of the explicitly, and thus DFT-described part of the system.
\[\tilde{\Omega}[\rho_{\rm el}]=\Omega^{\langle N_{\alpha}\rangle}[\rho_{\rm el} ]-\sum_{\alpha}\tilde{\mu}_{\alpha}\langle N_{\alpha}\rangle\quad, \tag{35}\]
which additionally accounts for the possible exchange of all explicitly treated chemical species \(\alpha\) of charge \(q_{\alpha}\) with their corresponding reservoirs described through their electrochemical potentials
\[\tilde{\mu}_{\alpha}=\mu_{\alpha}+q_{\alpha}\phi\quad. \tag{36}\]
Full minimization of this total grand potential functional would then yield the average particle number \(\langle N_{\alpha}\rangle\) of each explicitly described species at equilibrium.
Depending on the application at hand, it is typically convenient to distinguish sub groups among these explicit chemical species. Frequently, one considers substrate atoms of the (metal) electrode with chemical potentials \(\mu_{\rm sub}\), neutral solvent species \(j\) with chemical potentials \(\mu_{\rm solv,j}\) (e.g. water molecules in aqueous solvents), ions \(i\) of the electrolyte with electrochemical potential \(\tilde{\mu}_{\rm ions,i}\) and electrons with electrochemical potential \(\tilde{\mu}_{\rm el}\), with the latter indeed also just another chemical species in the thermodynamic sense. To this end, the electrocatalysis context and the use of an implicit solvation model dictate utmost care and add severe challenges in establishing a fully consistent set of corresponding electrochemical potentials. For one, the same chemical species might exist in both explicit and implicit parts of the system. This applies notably to mixed explicit/implicit models, where the inner DL is (partly) included in the DFT-treated part of the system. A common example are ice-like rigid water layers to approximate the Helmholtz layer structure at metal electrodes in aqueous solutions. For such systems, inconsistencies between the \(\mu_{\rm solv,j}\) or \(\tilde{\mu}_{\rm ion,i}\) employed in the explicit minimization of eq. and the one of the implicit electrolyte model could connect an erroneous free energy gain to the exchange of in principle equivalent explicit particles with implicitly described ones, or vice versa. If one indeed performed the full formal minimization of eq., this would then for instance spuriously favor to either describe the entire solvent in the DFT supercell explicitly or implicitly. A further challenge comes from surface chemical reactions, which can again not only convert explicitly and implicitly described species into another, but which in the electrocatalysis context in fact often involve the interconversion of species commonly assigned to different sub-groups. A prominent example would be protonation reaction steps, where a (charged) proton from the electrolyte and an electron end up forming part of a (neutral) reaction intermediate specifically adsorbed at the electrode.
The ongoing struggle to achieve such consistent sets of electrochemical potentials, in particular within the confines of present-day implicit solvation models, is one central reason, why contemporary works dodge the formal full minimization of the total grand potential of eq.. A second crucial one concerns the intractability of the concomitant configurational sampling and thermodynamic averaging. Such sampling obviously would have to include all possible structures and chemical compositions of the SLI, a task that in particular due to the possibility of strong _operando_ changes of working (electro)catalysts is generally as unfeasible as it is in thermal surface catalysis. In fact, electrocatalysis adds an additional level of complexity by the existence of charged species \(\alpha\) (ions or electrons). If the SLI model contained in the DFT supercell does not account for the contribution of the diffuse DL to the full compensating counter charge, then as highlighted in Chapter 1 this would imply the necessity to extend the sampling also over different overall charge states of the DFT supercell. However, in order to achieve an appropriate description of the extended interface and the metallic band structure of the electrode, DFT implementations predominantly employ periodic boundary conditions. This, for technical reasons, generally restricts such calculations to overall charge-neutral supercells and would thus prevent any such sampling of different supercell charge states. As already alluded to at several occasions, it is specifically the versatility with which implicit electrolyte models allow to include ionic counter charges into the DFT supercell that addresses this problem and we will discuss this in more detail in Section 3.4. Even if the grand-canonical sampling involved in the formal minimization of eq. can then be restricted to overall charge-neutral supercells, it is still generally impractical to perform this sampling simultaneously for the number of electrons and explicit particle species. This has to do with the predominantly canonical ansatz for the electron DOFs of major DFT packages (JDFTx and very recently also ONETEP forming rare exceptions). Rather than adjusting the electron number as to grand canonically equilibrate with an imposed electron electrochemical potential (in electrochemistry given by the applied electrode potential), this ansatz imposes a fixed electron number \(N_{\mathrm{el}}\)--with the \(\tilde{\mu}_{\mathrm{el}}\) to which this refers to then an outcome of the calculation. While not least the coupling to a potentiostat can still allow to indirectly determine the electron number that matches an applied electrode potential, cf. Section 3.5, this is in general technically better not mixed with a simultaneous adaption of the chemical composition \(N_{\alpha}\) of the DFT calculation.
For these multiple reasons, the prevalent approach in first-principles based SLI works with implicit solvation is to use an _ab initio_ thermodynamics framework as also widespread in thermal surface catalysis research. Instead of the full minimization of the total grand potential functional of eq., such a framework considers a total grand potential functional at a fixed chemical composition \(N_{\alpha}\) of the DFT-described part,
\[\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}]=\Omega^{N_{\alpha}}[\rho_{\rm el}]- \sum_{\alpha}\tilde{\mu}_{\alpha}N_{\alpha}\quad. \tag{37}\]
This \(\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}]\) is then evaluated and minimized individually for different trial chemical compositions \(N_{\alpha}\) and geometric structures of the SLI, where we recall that the latter is included through the Born-Oppenheimer parametric dependence of the involved quantum-mechanical free energy functional \(F[\rho_{\rm el}]\) on the positions \(\{\mathbf{R}_{\alpha}\}\) of the explicitly treated species, cf. eq. (1a). Subsequently, comparison of the resulting free energies for the individual candidate structures and compositions allows to conclude on their relative thermodynamic stability. In fact, the true equilibrium SLI structure and composition will yield a minimum such free energy, and with the exception of the neglected fluctuations around this equilibrium (contained in \(\langle N_{\alpha}\rangle\) in eq.) this free energy will be the same as the one one would also obtain from the full formal minimization of \(\tilde{\Omega}[\rho_{\rm el}]\).
This _ab initio_ thermodynamics approach is highly convenient. Not least, as there are no additional force terms beyond those already arising in the consideration of \(\Omega^{N_{\alpha}}[\rho_{\rm el}]\). This allows to straightforwardly perform geometry optimizations or structural sampling through molecular dynamics, if only the employed DFT code has an implemented implicit solvent model to evaluate \(\Omega^{N_{\alpha}}[\rho_{\rm el}]\) and associated force terms. This ease is treacherous though, as the obtained relaxed structure and sampled ensemble is, of course, restricted to the once fixed chemical composition. Likely even more consequential, all chemical sampling is now outsourced to the trial \(N_{\alpha}\) explicitly tested. In other words, while the full formal minimization of \(\tilde{\Omega}[\rho_{\rm el}]\) will yield the true equilibrium SLI structure and composition by construction, any evaluation of \(\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}]\) will only allow to conclude that among all compositions \(N_{\alpha}\) (and corresponding structures \(\{\mathbf{R}_{\alpha}\}\)) explicitly tested, the one that yields the minimum free energy is the closest approximant to the true SLI structure and composition within the configurational space spanned by the trial structures and compositions.
This kind of "poor man's sampling" is not only critical because of the human bias possibly introduced in the selection of trial structures and compositions. This is a problem that is generic to the described _ab initio_ thermodynamics framework and we will not further discuss it here. More specific to the electrified SLIs and implicit solvation context is instead that also the fixed-composition total grand potential functional \(\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}]\) still depends on the bulk electrochemical potentials. Inconsistencies in these references will therefore also in general sensitively affect the relative stabilities of trial structures and compositions, and the corresponding conclusion on the closest equilibrium approximant. However, evaluation of \(\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}]\) in mindfully chosen configurational sub-spaces, e.g. in the simplest case just different structures of the same chemical composition, can ease or even entirely lift these dependencies. Furthermore, with the focus typically on free energy differences as discussed in Section 2.2, further error cancellation might be exploited by taking these differences already for the individual trial candidates, rather than only after performing the corresponding two full minimizations. As we will see in Section 3.5, this is prominently exploited in performing so-called constant-charge rather than constant-potential calculations. In this respect, the pragmatic focus on the fixed-composition total grand potential \(\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}]\) can provide highly useful insight and can circumvent issues that within the context of present-day implicit solvation models and DFT calculations would render a formal full minimization of \(\tilde{\Omega}[\rho_{\rm el}]\) largely useless--even if it could be achieved practically.
This tight integration of implicit solvation models into the _ab initio_ thermodynamics framework has advantages and disadvantages. On the negative side, it is often difficult to judge how well an employed implicit model really describes solvation effects at the electrified interface, as the computed thermodynamic quantities may also be affected by specificitiesof the _ab initio_ thermodynamics ansatz. This has, not least, bearings on the parametrization issues of present-day implicit solvation schemes, as many of the quantities commonly measured in contemporary SLI electrochemistry can simply not be used to assess, advance or re-parametrize existing implicit solvation models. On the positive side, implicit solvation capabilities like the representation of counter charges within the DFT supercell are actually central to overcome some of the _ab initio_ thermodynamics limitations--and this may in fact turn out to be even more relevant conceptually than the originally intended (and likely quite crude) account of the solvation effects _per se_. In the following sections we will further illustrate these various aspects. We will begin with specific thermodynamic quantities that are least affected by the _ab initio_ thermodynamics framework (and its limitations) and thus most sensitive to the implicit solvation modeling itself, and then gradually move over to quantities where the two approaches get increasingly intertwined.
### Potential of zero charge
The potential of zero charge (PZC) and thermodynamic quantities evaluated at the PZC are a natural starting point for this survey. The PZC is generally defined as the applied electrode potential at which there is no excess charge at the metal electrode. Within the scope of this review, this implies that all electrolyte counter charges in the DL vanish, cf. The DFT supercell is thus charge neutral and any corresponding restrictions in sampling the optimum charge state of \(\Omega^{N_{\alpha}}[\rho_{\rm el}]\) do not apply. If we furthermore concentrate on the PZC of the pristine metal electrode and assume for the moment that the known structure and composition of the latter is not changed e.g. by any specific adsorption of electrolyte species (as most likely fulfilled at unreactive coinage metals), then there is neither any chemical composition sampling issue, nor do we have to worry about any of the (electro)chemical potentials of explicit ions, solvent or electrode species. When aiming for a comparison with experimental values, we still have to define an appropriate reference for the electron electrochemical potential and in this respect even this simple application example providesalready a manifestation of the subtle issues related with the definition of a consistent set of electrochemical potentials and their references when using implicit solvation models.
Let us henceforth generally denote electrode potentials on the absolute scale (i.e. versus vacuum reference \(\Phi_{\text{E,vac}}=0\)) with \(\Phi_{\text{E}}\) and correspondingly an experimentally measured PZC on this scale with \(\Phi_{\text{E,PZC}}\). Then we can exploit that \(e\Phi_{\text{E,PZC}}\) with \(e\) the (positive) elementary charge corresponds identically to the work function of the metal immersed in solution. In a periodic, canonical gas-surface DFT supercell calculation with a slab model representing the surface, the work function vs. vacuum is conveniently computed as \(e(\phi_{\text{F}}-\phi_{\text{vac}})\). Here, \(\phi_{\text{F}}\) is the electron Fermi level, which in the canonical calculation equals the DFT-internal electron chemical potential and is as mentioned before an outcome of the DFT calculation once self-consistency is achieved. \(\phi_{\text{vac}}\) is the DFT-internal vacuum potential, which one approximately obtains as the position of the average electrostatic potential in the middle of the vacuum region between the (periodically repeating) slabs. If this vacuum region in the supercell is now filled with implicit solvent, one would think that the same difference would directly yield the PZC. Unfortunately, this is not the case, as this difference only accounts for bringing the electron to the bulk of the implicit solvent. What is thus missing to be able to align to the experimental absolute scale, is the potential difference between implicit solvent and vacuum.
The corresponding potential drop at say a water-vacuum interface can in principle be determined from higher-level explicit simulations. However, this drop differs substantially from the required implicit water-vacuum drop, as the average electrostatic inner potential of bulk water that dominates the prior drop vanishes in implicit models. Alternatively, one might argue that due to the vanishing polarization at implicit-vacuum interfaces, the missing potential difference might actually be small. While this seems indeed supported by a recent study, it is still not a firm basis for a quantitative alignment. Similar alignment issues arise equally for other experimental referencing scales like the predominantly employed standard hydrogen electrode (SHE) for aqueous environments. By and large, this then presently prevents the desirable direct comparison of computed and measured PZC values for different electrodes as an accuracy test of the implicit solvent model (and specifically its parametrization).
Rather than actually assessing the performance of the implicit solvation description, the comparison to measured PZCs is therefore instead employed to empirically fit the unknown implicit solvent-vacuum potential drop or to re-parametrize the implicit solvent model to effectively match an experimental PZC, e.g. for Pt electrodes. To this end, it has to be emphasized though that in such procedures experimental data needs often to be re-referenced from a reference electrode scale to the absolute scale, e.g. using the absolute SHE potential, which in itself introduces quite some uncertainties on the experimental numbers. Notwithstanding, since the implicit solvent-vacuum potential drop is solvent-specific, but electrode independent, useful insight into the performance of the implicit solvation model can still be obtained from relative PZCs, i.e. PZC trends over different electrodes.
A corresponding comparison with experimental data in water is shown in and
**Relative trend of PZC values as obtained in experiments and implicit solvent calculations.** Experimental PZCs for the low-index surfaces of Ag, Cu, Au and Pt (gray line) are on the absolute scale and taken as averages of literature data compiled in the SI of ref.. Calculated PZCs are arbitrarily aligned to the experimental PZC of Au and taken from refs. (red) and (blue).
As also apparent from Fig. 8, experimental absolute PZCs are consistently smaller than the corresponding vacuum work functions. In the computed PZC values were arbitrarily aligned to the experimental PZC of Au to illustrate the trend behavior, so that from there no conclusion can be drawn in how much implicit solvation models can reproduce this reduction. However, if the implicit water-vacuum drop is indeed small, then one can indeed show that they would effectively yield this reduction, albeit likely only on a quantitatively smaller scale. This is actually surprising, since the reduction originates in reality mainly from finite charge transfer from water molecules in the inner DL and a concomitant polarization within the first \(\sim 4\,\)A away from the metal surface. Outliers to the captured trend in the PZCs, e.g. the larger offset for Pt visible in Fig. 8, have been ascribed to an increased interfacial charge transfer that can no longer be mimicked by the implicit solvation model. However, for more reactive surfaces one also has to keep in mind that experimental PZCs can not least be influenced by specifically adsorbed electrolyte ions. If the calculations were repeated for a corresponding chemical composition \(N_{\alpha}\) including such ions, the agreement might thus improve. However, in general, this uncertainty is nothing but a consequence of the "poor man's sampling" in _ab initio_ thermodynamics, which requires the explicit testing of different such chemical compositions \(N_{\alpha}\), rather than yielding the true equilibrium one as an outcome of the theory.
### Computational hydrogen electrode
Many of the thermodynamic quantities that are of central interest in electrocatalysis concern energetics. Take the surface free energy as a measure of the stability of the catalyst surface, or adsorption free energies as central to the surface thermochemistry (or within Bronsted-Evans-Polanyi relationships even indicative of the reaction kinetics). Computing or evaluating such energetic quantities within the _ab initio_ thermodynamics framework de scribed in Section 3.1 almost invariably involves comparing the relative stability of interface configurations with different chemical composition \(N_{\alpha}\). Next to the electron electrochemical potential referencing issues discussed in the preceding section, this then also puts the electrochemical potentials of particle species \(\alpha\) like explicitly described ions, solvent molecules or electrode constituents on the agenda. In addition, the energetic quantities are typically not only required at the PZC, which could then bring up first trouble with the charge restriction of prevalent periodic boundary DFT supercell implementations.
Let us exemplify this general problem in this section for the simple case of hydrogen adsorption (formally better proton electrosorption) in an aqueous environment. A pertinent thermodynamic quantity to compute for this case is the adsorption free energy.
\[\Delta G_{\rm H}^{\rm ads,}N_{\alpha}(\Phi_{\rm E})=\tilde{\Omega}^{N_{\alpha}, \rm H}[\rho_{\rm el,H}^{\circ}(\Phi_{\rm E})]-\tilde{\Omega}^{N_{\alpha}}[\rho _{\rm el}^{\circ}(\Phi_{\rm E})]\quad. \tag{38}\]
Here, \(N_{\alpha}\) summarizes the entire chemical composition of the electrode, which we assume to be unchanged upon adsorption apart from the additional proton (H nucleus). For simplicity of notation, we consider here only one proton per supercell, even though one would in the implicit solvation context practically prefer symmetric slab setups with adsorption of one proton per surface and thus two protons per supercell. Obviously, within the employed periodic boundary conditions adsorption of this proton per supercell corresponds as always effectively to some finite coverage, but this does not matter for our present argument. Note also that similarly as in Section 2.6 we again attest to the fact that (measurable) adsorption free energies are generally seen as a property of the full (macroscopic) system and not of the grand-canonical sub-system technically employed in the calculations, which is why we denote them as free energies even though they are computed here as a difference of grand potential energies.
Each fixed-composition total grand potential energy in eq. is evaluated at its optimized equilibrium electron density. Due to the presence of the additional H nucleus, \(\rho_{\rm el,H}^{\circ}\) will not only differ in its detailed spatial form from \(\rho_{\rm el}^{\circ}\), but will generally also integrate up to a total number of electrons that differs by \(l\)--the so-called electrosorption valency. With the electrode chemical composition unchanged, the dependence on all electrode chemical potentials \(\mu_{\rm sub}\) cancels in the difference of eq.. What remains are the electrochemical potentials of the bulk reservoirs from where the proton and the additional electron density dragged to it upon adsorption came from. At the electrified interface in the aqueous environment this correspond to \(\tilde{\mu}_{\rm H^{+}}\) of the solvated proton and the electron electrochemical potential \(\tilde{\mu}_{\rm el}\) as determined by the applied potential. Using eq. we can thus rewrite eq.
\[\Delta G_{\rm H}^{\rm ads,N_{\alpha}}(\Phi_{\rm E})=\left(\Omega^{N_{\alpha},{ \rm H}}[\rho_{\rm el,H}^{\circ}(\Phi_{\rm E})]-\Omega^{N_{\alpha}}[\rho_{\rm el }^{\circ}(\Phi_{\rm E})]\right)-(\tilde{\mu}_{\rm H^{+}}+l\tilde{\mu}_{\rm el} )\quad\text{.} \tag{39}\]
This equation now clearly carves out the entire wealth of practical problems that have to be dealt with. Under an applied potential \(\Phi_{\rm E}\) away from the PZC, the electrode will generally be charged with a corresponding balancing counter charge built up in the electrolyte. Some of this counter charge will be located in the diffuse DL, which is likely outside of a practically feasible DFT supercell as discussed in Chapter 1. Unless this is suitably taken care of by an implicit electrolyte model as discussed in the next section, this would imply the computation of charged supercells. Furthermore, to evaluate eq. we also need to determine the two electrochemical potentials. While we have already seen the difficulties to align \(\tilde{\mu}_{\rm el}=-e\Phi_{\rm E}\) on the absolute scale with the DFT-internal Fermi level in an implicit solvation calculation in the last section, also the explicit computation of the electrochemical potential of a solvated proton \(\tilde{\mu}_{\rm H^{+}}\) is a tough endeavor.
Intriguingly, all of these problems vanish completely with just one single and ingenious approximation. If we assume that the optimized electron density of a given interface configuration \(N_{\alpha}\) at any applied potential \(\Phi_{\rm E}\) remains the same as the one at its PZC, then there is no electrolyte counter charge as discussed in the previous section and the DFT supercell is always charge neutral. For \(\rho_{\rm el,H}^{\circ}\) this implies that the adsorbed protonic charge is exactly compensated by one additional electron, i.e. the H adsorption (proton electrosorption) process is a so-called proton-coupled electron transfer (PCET).
\[\Delta G_{\rm H,CHE}^{\rm ads,N_{\alpha}}(\Phi_{\rm E})=\left(\Omega^{N_{ \alpha},{\rm H}}[\rho_{\rm el,H}^{\circ}(\Phi_{\rm E,H,PZC})]-\Omega^{N_{ \alpha}}[\rho_{\rm el}^{\circ}(\Phi_{\rm E,PZC})]\right)-(\tilde{\mu}_{\rm H^ {+}}+\tilde{\mu}_{\rm el})\quad. \tag{40}\]
Also, the electrochemical potential calculation and alignment problem is naturally resolved, as the remaining integer sum of the two potentials in eq.
\[(\tilde{\mu}_{\rm H^{+}}+\tilde{\mu}_{\rm el})\;=\;\tfrac{1}{2}\mu_{\rm H_{2} }-e\Phi_{\rm E}^{\rm SHE}-k_{\rm B}T\ln{\rm pH}\quad. \tag{41}\]
Here, \(\mu_{\rm H_{2}}\) is the chemical potential of hydrogen gas at standard state, which is straightforward to compute, and pH is the pH value of the aqueous electrolyte. Note that the SHE scale is the predominantly employed scale in experiments anyway, which thus does allow to directly compare with experiments (as long as they are not affected by mass transport effects). There is correspondingly no need anymore to align the DFT-internal Fermi level to the applied potential. In fact, as the difference of grand potential energies in eq. is now potential-independent, it suffices to compute it once, and the entire dependence of the adsorption free energy on the applied potential is then just analytically given by eq..
\[(\tilde{\mu}_{\rm H^{+}}+\tilde{\mu}_{\rm el})\;=\;\tfrac{1}{2}\mu_{\rm H_{2} }-e\Phi_{\rm E}^{\rm RHE}\quad\, \tag{42}\]i.e. all pH dependencies are in the CHE just trivially Nernstian.
The original intention to exploit the SHE definition to circumvent the electrochemical potential referencing issues was coined computational hydrogen electrode (CHE) by Rossmeisl, Norskov and coworkers. Nowadays, CHE is instead essentially equated with the somewhat stronger approximation to employ PZC optimized densities as introduced in the example above. This kind of CHE approach underlies the by far dominant part of contemporary first-principles based work on electrified interfaces and electrocatalysis at them. In fact, it is fair to say that without the computational simplicity enabled by the CHE, theoretical electrocatalysis would not be where it is today. The CHE philosophy is readily generalized to other electrodes (computational sulfur electrode, computational Li electrode,...) and employed for the computation of other thermodynamic quantities.
**Surface phase diagram of Pt in water as determined within the CHE approach.** Shown are computed potential-dependent surface free energies of bare Pt and various H, OH and O coverages on it. Within the _ab initio_ thermodynamics framework, surface terminations with lowest surface free energy are declared as most stable one at the corresponding potential. This yields the indicated gradual transition from H-covered over bare surface to OH- and O-covered terminations with increasingly positive potential. Reproduced with permission from ref. 378.
the cost to create a surface with a certain structure and composition, as well as thermodynamic reaction barriers or concomitant thermodynamic overpotentials. Using the prior applied-potential dependent surface free energies to compare the stability of a range of candidate surface structures and composition, one can readily establish surface phase diagrams, which--if the electrochemical potential dependence is resolved into a potential and pH dependence--are also known as Pourbaix diagrams. illustrates this with corresponding work from McCrum _et al._ for the Pt surface in a water environment. Such kind of CHE surface phase diagrams are nowadays widely used to draw first conclusions on the actual surface structure and composition of electrodes under true operating conditions and we refer to excellent reviews on this topic for a more detailed overview of the uses and merits of this kind of most popular CHE application.
If employed within the sketched CHE approach, the task of an implicit solvation model is to account for the solvation response at the PZC of the considered surface configuration. This is conceptually analogous to what was discussed in the previous section for the pristine electrode, yet with two notable, opposing differences. On the one hand, in particular for larger, more protruding, polar or hydrogen-bond affine adsorbates one can in principle expect larger solvation corrections even at the PZC [395, 396, 397, 47, 43, 43, 47, 43, 43, 43, 44, 45, 46, 47, 48, 498, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 199, 199, 190, 191, 193, 195, 196, 197, 198, 199, 199, 199, 190, 192, 194, 195, 196, 197, 198, 199, 199, 199, 198, 199parametrizations. More detailed analysis of the contributions points out that an arguably most important correction to existing schemes would be to account for so-called competitive solvent adsorption in the cavitation grand potential. This reflects the fact that in a thermodynamically consistent treatment of solvation at SLIs, there should be an energy cost associated with the need to first remove solvent from the pristine electrode to create space for the adsorbate. This would generally be a substrate-dependent cost, in contrast to the existing, substrate agnostic cavitation grand potential formulations discussed in Section 2.4.1.
We think it could mainly be this missing appropriate account of competitive solvent adsorption that stands behind the (partly) dramatic discrepancies between implicit solvation results and benchmark AIMD simulations in explicit water environments. This view would be supported by the strong correlations with the OH/H\({}_{2}\)O adsorption properties of the substrate. Note, however, that competitive solvent adsorption is also not appropriately considered in a wide range of simple explicit solvation strategies, while it is generally questionable anyway whether the limited trajectories obtained in the dynamic simulations can really faithfully mimic thermodynamic equilibrium. On the implicit solvation side, there are some hints that more substrate-specific models such as the SCSS model using soft-sphere atomic cavities might constitute a way forward while not compromising other observables. Nevertheless, while all of this surely indicates the need to further advance implicit solvation models (or to rather move over to mixed explicit/implicit solvation models for SLIs), the fact remains that in a CHE free energy difference like in eq. for the adsorption free energy, solvation corrections evaluated at the PZCs tend to be small. One can correspondingly find multiple practitioner works in the literature, where the CHE is applied and in fact no solvation treatment is included at all, i.e. the underlying DFT calculations are actually performed for slabs in vacuum. In historical perspective, the advent of implicit solvation methodology in major periodic boundary conditions DFT packages came after the CHE approximation was firmly established and widely employed by the theoretical electrocatalystscommunity. The new functionality was then often employed within the prevalent CHE, rather than realizing that it could actually constitute a powerful avenue beyond it.
### Surface charging and interfacial capacitance
It is indeed important to realize that the typically small solvation corrections within the CHE are an invariable outcome of the PZC assumption. To assess this assumption, let us recall the physical processes actually occurring at a pristine electrode on gradual application of a potential that brings us away from its PZC. Without loss of generality, let this be a potential positive from the PZC, which will thus withdraw electrons from the electrode and lead to the formation of a positive net surface charge on the electrode surface. In order to screen the resulting electric field, a compensating counter charge will build up in the electrolyte part of the DL. Initially, this is just a capacitive charging process of the electric DL as introduced in Chapter 1. This means that the concomitant changes to the electrode electron density might induce some atomic relaxation or even stronger rearrangements in the electrode material. Also, the molecular and ionic distributions within the electrolyte will obviously change when building up the counter charge. However, at the initially small potentials there is formally no exchange of (charged) matter between these two constituents of the DL. In a fully implicit solvation model, this would thus mean that the chemical composition \(N_{\alpha}\) of the DFT-part of the system does not change.
Upon further increase of the potential away from the PZC, the polarization of the DL might eventually become so large, that such an exchange will occur, specifically in form of a so-called interfacial (or Faradaic) charge transfer. Here, it is now generally the transfer of ions to or from the electrolyte with a concomitant change of their charge state that reduces the electric field. For the considered positive potential and an aqueous electrolyte, this could for instance be the specific adsorption of anions, or depending on the pH, the formation of hydroxyl groups at the metal electrode. Even in a fully implicit solvation model, these new surface species would be explicitly modeled and we would correspondingly arrive at a new chemical composition \(N_{\alpha^{\prime}}\). This new electrode configuration then has its own new PZC, typically at a more positive potential as the original pristine surface (cf. normal vs anomalous work function change). When now further increasing the potential, we are again moving away from this PZC and the sequence of capacitive charging and interfacial charge transfer upon exceeding DL polarization continues.
What the CHE does is to approximate this sequence with a series of pure charge-transfer processes through unpolarized electrode configurations. As it only considers the electron density at the PZC of each electrode configuration, it is agnostic to capacitive charging. The increasing applied potential enters the fixed-composition total grand potential only through the changed electron electrochemical potential term as in the hydrogen adsorption eq. above. Within the _ab initio_ thermodynamics framework, any change of relative stability of different electrode configurations can thus only be captured, if the concomitant change of \(N_{\alpha}\) includes a change in the involved number of electrons \(N_{\text{el}}\), as is the case for an interfacial charge transfer. By construction, the CHE approximation can therefore for instance not account for potential-induced geometric changes or stronger reconstructions of the electrode that leave the chemical composition \(N_{\alpha}\) unchanged.
In order to overcome these limitations of the CHE it is therefore imperative to include some form of surface charging into the modeling. Remember that one of the motivations for historically introducing the CHE was the charge-neutrality restriction of prevalent periodic boundary condition DFT implementations. This restriction is elegantly addressed by the PZC assumption as the electron density then automatically integrates up to exactly match the total nuclei charge of the DFT part of the supercell. Yet, even in these codes there is in practice nothing that prevents us from adding more or less electrons into the DFT calculation to mimic surface charging. What the codes would only do (more or less unnoticed), is to introduce a homogeneous background charge that exactly compensates the net charge that would result from this varied electron number. In principle, this straightforward, so-called jellium approach can still be and is in fact largely used to model a potential-dependentelectron density. Obviously though, a homogeneous background charge--if applied without further corrections--is unlikely a good representation of the DL counter charge and there are several studies that highlight the nonphysical surface charging behavior obtained within this approach.
It is especially to this problem of surface charging where implicit solvation methodology adds significant flexibility. As discussed in Section 2.5 and summarized in Fig. 5, current electrolyte models offer a wide spectrum of introducing counter ions into the supercell, thereby allowing to establish overall charge neutrality without the need for a jellium background. This spectrum ranges from the simple PCC Helmholtz-layer models to the self-consistent ion distributions of PB or MPB theory, where, importantly, the ion distributions described in the latter theories include the diffuse DL. Here, it is worthwhile to emphasize the elegance with which these implicit approaches solve the problem of the wide extension of the diffuse DL part highlighted in Chapter 1. Even if this extension largely exceeds the actual dimension of the supercell employed in the practical DFT calculation, this still plays no role as it only enters the generalized Poisson equation solver of the code, cf. Section 2.5. Corresponding solvers can be implemented with free boundary conditions in \(z\)-direction, i.e. vertical to the slab surface, and are then completely independent of the finite supercell size used in the other parts of the KS DFT minimization. Whatever the specific electrolyte model used, its ionic counter charges thus flexibly allow to satisfy the supercell charge-neutrality condition despite a varying net surface charge on the explicitly DFT-described electrode. The implementation of corresponding implicit electrolyte models in a range of major DFT packages as summarized in Table 2 marked therefore a big conceptual step forward for the first-principles based modeling of electrified SLIs. What remains to be seen though, is what this actually brings practically for the modeling of surface charging and the truly intended computation of fixed-composition total grand potential energies \(\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}^{\circ}(\Phi_{\rm E})]\) with potential-dependent optimized electron densities \(\rho_{\rm el}^{\circ}(\Phi_{\rm E})\), cf. eq..
\[\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}^{\circ}(\Phi_{\rm E})]=\tilde{\Omega} ^{N_{\alpha}}[\rho_{\rm el}^{\circ}(\Phi_{\rm E,PZC})]-\frac{AC_{\rm PZC}}{2} \left(\Phi_{\rm E}-\Phi_{\rm E,PZC}\right)^{2}+{\cal O}(\left(\Phi_{\rm E}-\Phi_ {\rm E,PZC}\right)^{3})\quad, \tag{43}\]
Consistent with the above discussed physics of capacitive charging when moving away from the PZC, it is thus \(C_{\rm PZC}\) that naturally appears in the thermodynamics as a central quantity. This analysis correspondingly suggests that next to the PZC discussed in Section 3.2, implicit electrolyte models should especially be able to appropriately describe the capacitance \(C_{\rm PZC}\) at the PZC to achieve a sound potential-dependence of grand potential energies \(\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}^{\circ}(\Phi_{\rm E})]\) away from the PZC.
**Dependence of the interfacial capacitance on implicit solvation model parameters.** Shown is the variation of the fixed-composition total grand potential energy \(\tilde{\Omega}^{N_{\alpha}}\) around the PZC for a model Li electrode in implicit ethylene carbonate (EC) solvent (\(\varepsilon(\exp)=89.9\)). The parabolic variation nicely reflects the Taylor expansion of eq. and allows to fit the interfacial capacitance. (Left panel) Variation as a function of the bulk permittivity employed in the implicit solvation model. (Right panel) Variation as a function of the threshold charge density (called \(n_{\rm c}\)) employed to define the solvation cavity. Adapted from ref..
Their data for Li nicely portrays the inverted parabolas of the Taylor expansion, eq., and the sensitive dependence of the extracted capacitance \(C_{\mathrm{PZC}}\) on central parameters of the implicit solvation model, namely the bulk permittivity value and the iso-surface value of the electron density defining the solvation cavity, cf. Section 2.3. Note in particular the small capacitance values obtained in vacuum that can only be increased when considering the polarization response of the surrounding liquid through the implicit solvation model.
Further away from the PZC, higher-order terms in the Taylor expansion of eq. will start to play a role, which will then feature variations of the interfacial capacitance with the applied potential. As a result, quite some work has been dedicated to construct implicit models that can reproduce experimentally observed capacitance variations with applied potential as well as electrolyte composition. As already introduced in Chapter 1, the total interfacial capacitance can be seen as arising from two capacitors in series, the inner DL and the outer DL. For high electrolyte concentrations or for potentials far away from the PZC, this total capacitance will correspondingly be dominated by the capacitance of the inner DL, where the highest potential drop occurs. Consistent with this picture, the modeling of the diffuse DL is often found to play a minor role to describe the interfacial capacitance in this limit. Instead, it is the appropriate parametrization of the solvation cavity boundary that critically determines the overall accuracy, and it is within this understanding that refined models that include nonlinear electrolyte and dielectric response e.g. in form of a dielectric decrement as discussed in Section 2.5.3 are currently being pursued.
### 3.5 Constant potential vs.
As discussed in Section 3.3, the capability to compute applied-potential dependent fixed-composition total grand potential energies \(\tilde{\Omega}^{N_{\alpha}}[\rho_{\mathrm{el}}^{\circ}(\Phi_{\mathrm{E}})]\) is a key prerequisite to access thermodynamic energetic quantities like surface free energies or adsorption free energies, see e.g. eq. for the discussed example of hydrogen adsorption. To this end, the flexibility with which implicit electrolyte models allow to consider finite surface charges in the periodic DFT supercell calculations provides primarily an opportunity to go beyond the CHE approximation. Within the prevalent canonical DFT implementations that work with a prescribed number of electrons \(N_{\rm el}\), the amount of surface charge that corresponds to a particular applied potential \(\Phi_{\rm E}\) can e.g. straightforwardly be obtained from the condition that the DFT-internal electron chemical potential has to equal the external electron electrochemical potential at equilibrium. As discussed before in Section 3.2, the prior is given by the Fermi level position \(\phi_{\rm F}(N_{\rm el})\) with respect to the DFT-internal vacuum reference \(\phi_{\rm vac}\), while the latter is imposed by the applied potential.
\[e(\phi_{\rm F}(N_{\rm el})-\phi_{\rm vac})=\tilde{\mu}_{\rm el}=-e\Phi_{\rm E} \tag{44}\]
For this electron number, the desired fixed-composition total grand potential energy is then evaluated as
\[\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}^{\circ}(\Phi_{\rm E})]=\Omega^{N_{ \alpha}}[\rho_{\rm el}^{\circ}(N_{\rm el}(\Phi_{\rm E}))]+e\Phi_{\rm E}N_{\rm el }(\Phi_{\rm E})\quad, \tag{45}\]
Equivalent results can be achieved via the use of a potentiostat or in grand canonical DFT via an extended Hamiltonian. Note that, if it is of interest to understand the interface energetics for a range of applied potentials, it can be computationally more efficient to simply compute \(\Omega^{N_{\alpha}}[\rho_{\rm el}^{\circ}(N_{\rm el})]\) as well as \(\Phi_{\rm E}(N_{\rm el})\) (\(\leftrightarrow\phi_{\rm F}(N_{\rm el})\)) for a set of electron numbers \(\{N_{\rm el}\}\) from which the (continuous) \(\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}^{\circ}(\Phi_{\rm E})]\) can then be obtained in a straightforward way e.g. via interpolation.
In principle, correspondingly determined \(\tilde{\Omega}^{N_{\alpha}}[\rho_{\rm el}^{\circ}(\Phi_{\rm E})]\) still suffer from the difficulty of referencing to the DFT-internal vacuum reference \(\phi_{\rm vac}\) in implicit solvation calculations as discussed in Section 3.2. However, fortunately only differences of such fixed-composition to tal grand potential energies are typically required for targeted quantities like an adsorption free energy. In such differences, one can consistently reference to the available bulk implicit solvent potential, which then only implies a residual constant shift of the potential dependence of a quantity like \(\Delta G^{\rm ads,}N_{\alpha}(\Phi_{\rm E})\) with respect to an experimental scale. Accepting such constant uncertainty, empirical values are then also conveniently taken for additionally required electrochemical potentials of those particle species that vary in these differences, like the \(\tilde{\mu}_{\rm H^{+}}\) of a solvated proton in the hydrogen adsorption example.
On the other hand, the necessity to evaluate differences also creates challenges, in particular with respect to the implicit solvation modeling. To understand this, let us recap the equation with which the hydrogen adsorption free energy of Section 3.3 would be determined in this constant-potential approach
\[\Delta G^{\rm ads,}_{\rm H}(\Phi_{\rm E})=\tilde{\Omega}^{N_{\alpha},{\rm H}}[ \rho^{\circ}_{\rm el,H}(\Phi_{\rm E})]-\tilde{\Omega}^{N_{\alpha}}[\rho^{\circ }_{\rm el}(\Phi_{\rm E})]\quad. \tag{46}\]
Both fixed-composition total grand potential energies are obviously here evaluated at the same applied potential \(\Phi_{\rm E}\). However, as discussed in the last section, the PZCs of the H-covered and the clean surface are generally different. This means that the evaluation of the two grand potential energies proceeds at a different relative potential with respect to the respective PZC. It could for example be that the applied potential is actually quite close to the PZC of the clean surface, but rather far away from the H-covered one. In the last section we had seen that the accuracy of implicit electrolyte models typically depends on this relative difference from the PZC. A model might be better suited to describe the potential region around the PZC, while another one is tailored just for the inner DL-dominated region far away from the PZC. In the difference of eq., the model can instead be required to describe quite different relative potential regions and this could introduce quite some error.
Aiming for better error cancelation, we can revisit the CHE and emphasize an aspect of it that is actually often overlooked. As discussed, the prevalent form of the CHE assumes that the optimized electron density of a given interface configuration \(N_{\alpha}\) at any applied potential \(\Phi_{\rm E}\) always remains the same as at its PZC, cf. Section 3.3. Intriguingly, this implies that in the difference required for the adsorption free energy, the two fixed-composition grand potential energies are actually evaluated at different potentials. In the hydrogen adsorption case and the corresponding eq., this would namely be at the PZC of the hydrogen-covered surface \(\Phi_{\rm E,H,PZC}\) and at the PZC of the clean surface \(\Phi_{\rm E,PZC}\). It is now tempting to transfer this aspect to the surface charging case. If both terms in eq. were not evaluated at the same potential, but at the same amount of surface charging, then both fixed-composition total grand potential energies are determined at an approximately equal relative potential with respect to their PZC and one can hope for maximum error cancellation in the implicit model. Of course, on the other hand, error is introduced because at least one of the two terms is not computed at the correct applied potential--but possibly this incurred error is smaller than the error cancellation achieved.
This is essentially the philosophy of so-called constant-charge calculations, which can in principle be carried out with explicit or implicit charging. In such calculations the amount of surface charge according to an applied electrode potential is determined e.g. from experimental or simulated charge-potential relations. The resulting decoupling of quantum chemistry and surface charging thus allows to control the accuracy of both scales roughly independently. While clearly inspired by the CHE, it is worthwhile to note that this approach is also closely related to traditional constant field calculations as interfacial fields are naturally proportional to surface charge, and it is correspondingly not surprising that latter calculations were also taken into consideration in the early developments of the CHE approximation.
Both the constant-potential and constant-charge approach enjoy present popularity, with the former also often denoted as fully grand-canonical (FGC) approach. While they give quantitatively different results in practical supercell sizes, it is gratifying that in the thermodynamic limit of low-coverage adsorption, i.e. one adsorbate in a laterally infinitely extended supercell, both approaches will eventually coincide. In this limit, the PZC of the clean surface will only be infinitesimally changed upon adsorption. At any applied potential, both the single-adsorbate covered and clean surface will thus be equally charged anyway, and one is in both cases also at an equal absolute and relative potential with respect to the joint PZC. For this limit, one can analogously to the constant-potential case also consider a constant-charge appropriate Taylor series expansion of the adsorption free energy, i.e. now in terms of the excess charge. Here it has been found that in many cases, the first order term corresponding to the PZC change is dominating. In contrast, higher-order terms depending on the capacitance and thus electrolyte description are indeed less relevant, supporting the error cancelation motivation of this approach. For the constant-potential case, a dipole-field-type first-order correction term to the CHE adsorption free energy can even analytically be derived.
\[\Delta G_{\mathrm{H}(\theta\to 0)}^{\mathrm{ads},N_{\alpha}}(\Phi_{\mathrm{E}})- \Delta G_{\mathrm{H}(\theta\to 0),\mathrm{CHE}}^{\mathrm{ads},N_{\alpha}}(\Phi_{ \mathrm{E}})\approx e\;(l-1)\;(\Phi_{\mathrm{E}}-\Phi_{\mathrm{E,PZC}})\quad, \tag{47}\]
In the CHE, \(l_{\mathrm{CHE}}=1\), corresponding to a PCET process, while in general \(l\neq 1\).
Compared to the CHE, changes in the potential-dependence of adsorption free energies as obtained in emerging constant-potential or constant-charge calculations seem indeed generally largely be understandable in terms of dipole-field interaction, even for larger molecules, such as e.g. CO\({}_{2}\) reduction intermediates. Changes in this dipole-field interaction with applied potential can then lead to a wide range of conceptual physics that was outside of the realm of CHE theory. This includes a potential-induced switching of the most stable adsorption site or altered adsorbate geometries or adsorption motives, including e.g. interfacial water. Recent corresponding results have for instance also helped to clarify the impact of different cationic species on the interfacial capacitance, and how this can influence in an indirect way via the variation in the dipole-field interaction a variety of electrochemical observables such as the stability of adsorbed CO\({}_{2}\) and the Stark shifts of CO. In addition, the now available possibility to appropriately account for effects of the applied potential beyond the CHE has enabled significant progress in the simulation of electrochemical reaction barriers and bridged the gap to works that employ explicit charging strategies. Further works reported potential-induced surface reconstructions or lifting of those as well as non-Nernstian dependencies for surface coverages, nanoparticle shapes, Pourbaix diagrams, and last but not least (thermodynamic) cyclic voltammograms, where peak positions and shapes can indeed be extremely sensitive to the electrochemical conditions. These developments are all quite recent and we expect significant further progress in our understanding of interfacial electrocatalysis to emerge from such constant-potential or constant-charge calculations.
**Theoretical surface Pourbaix diagram of Cu in implicit water considering H and CO adsorbates.** The diagram obtained within the CHE approximation (left panel) shows only a trivial Nernstian pH-dependence, which vanishes on the here employed RHE scale. In contrast, non-trivial pH-dependencies are obtained with constant-potential aka FGC calculations (right panel). Figure created from data published in ref..
eq.. In the Pourbaix diagram for Cu shown on the RHE potential scale, there is correspondingly no further pH dependence within the CHE, but significant structure when computed with the constant potential approach.
## 4 Conclusions and outlook
Predictive-quality first-principles calculations based on DFT have undoubtedly become a cornerstone in modern materials, catalysis and energy research. In the specific context of catalysis at electrified interfaces, this development is largely connected with the ingenious computational hydrogen electrode (CHE) approach of Rossmeisl, Norskov and coworkers. It is difficult to understate the impact that this single approach has made on the design of electrocatalysts or the unraveling of electrochemical reaction mechanisms. By the very nature of its approximation, the CHE puts the predominant emphasis on the electrode site. Over the last decade or so, first-principles electrocatalysts research at solid-liquid interfaces (SLIs) was correspondingly dominated by finding optimum catalyst materials that lie at the top of reaction volcanos or gaining mechanistic understanding in terms of surface chemical bonds, yet without much caring for the electrolyte side of the SLI.
It is only within the last few years that an ever increasing understanding of electrified SLIs starts to trigger a return to this foundational pillar of electrochemistry, namely the influence of the electrolyte at the SLI. Unfortunately, it is also only when one starts to devise strategies of how to actually do so within the realm of present-day DFT and supercomputer capabilities, that one really starts to appreciate the ingenuity of the CHE approximation and the simplicity of the calculations it enables. Any real consideration of the extended double layer (DL) with its intricate long-range electrostatics and inherent dynamics quickly leads to excessive computational costs. In this respect, implicit solvation methodology forms a unique compromise. As we have surveyed in this review, consideration of corresponding methodology within the _ab initio_ thermodynamics framework commonly employed in surface catalysis anyway, immediately gives rise to multiple avenues beyond the CHE. At the same time, the computational cost of corresponding constant potential or constant charge calculations stays not too different from the one of the CHE approach.
While thus highly promising, this approach is not without its own challenges. The implementation of implicit solvation functionality into a series of powerful DFT software packages that can describe extended SLIs typically within the frame of periodic boundary condition supercells has been the enabler for this new field and a great community effort. However, in these implementations the methodological framework, historically developed to assess solvation effects on molecular solutes, has largely been left unaltered. To one end, this concerns the usage of parametrizations derived from molecular experimental reference data. To the other end, functional expressions for the effectively treated explicit electrode - implicit electrolyte interactions have if at all only marginally been modified, for instance if they contained quantities like a cavity volume that is not accessible at an extended SLI. As we have seen in the course of this review, the primary advance brought about by implicit solvation for the SLI context is more the flexibility with which one can represent the counter charges in the DL, rather than the actual account of solvation effects. For this purpose, the present state-of-the-art may largely be sufficient--and in addition to the already obtained massive insight into catalysis at electrified interfaces, we expect truly grand-canonical results (like the discussed constant potential or constant charge calculations) on the basis of existing implicit electrolyte models to continue carving out important electrochemistry that was not accessible within the CHE framework.
However, this can only be a first step. At present, the community is at a crossroad. One route is to focus efforts towards mixed explicit/implicit solvation approaches. The other is to refine the implicit solvation technology itself. For both case, what will centrally be required is a re-thinking of the functional expressions, in particular of the non-electrostatic terms, and reference data that is more pertinent for extended SLIs. Regarding the latter, we have seen throughout the review, that many experimentally accessible quantities commonly measured in electrochemistry are not ideal for this task, as their computation intricately mixes solvation effects with the specificities of the employed _ab initio_ thermodynamics ansatz. In this respect, more, systematic and accurate measurements of PZCs and interfacial capacitances for well-defined model electrodes would certainly be helpful. In our view, also contact angles could be another highly useful class of quantities. Without any such data, it is largely impossible to develop highly parametrized and thus potentially more accurate implicit solvation methods without running into overfitting. From this perspective, the actual developments of first-principles machine-learned interatomic potentials are probably most exciting. Explicit AIMD data has long been used to validate and improve implicit solvation methodology. If machine-learned interatomic potentials allow to generate comparably accurate, but orders of magnitude longer trajectories and in larger simulation cells, then this will be an invaluable asset that might even ultimately enable to validate and refine implicit solvation schemes for application outside the domain of _ab initio_ thermodynamics, notably the modeling of kinetic reaction barriers.
|
10.48550/arXiv.2108.02461
|
Implicit Solvation Methods for Catalysis at Electrified Interfaces
|
Stefan Ringe, Nicolas G. Hörmann, Harald Oberhofer, Karsten Reuter
| 1,103
|
10.48550_arXiv.1302.0771
|
###### Abstract
Memristors have been compared to neurons (usually specifically the synapses) since 1976 but no experimental evidence has been offered for support for this position. Here we highlight that memristors naturally form fast-response, highly reproducible and repeatable current spikes which can be used in voltage-driven neuromorphic architecture. Ease of fitting current spikes with memristor theories both suggests that the spikes are part of the memristive effect and provides modeling capability for the design of neuromorphic circuits.
Keywords:memristors, d.c., current transients, mem-con theory
# Observation, Characterization and Modeling of Memristor Current Spikes
Ella Gale 1
1.
Frenchay Campus, Coldharbour Lane, Bristol, BS16 5SR, UK
Ben de Lacy Costello
1.
Frenchay Campus, Coldharbour Lane, Bristol, BS16 5SR, UK
Andrew Adamatzky
1.
Frenchay Campus, Coldharbour Lane, Bristol, BS16 5SR, UK
Footnote 1: e-mail:
November 6, 2021
## 1 Introduction
Neuromorphic computing is the concept of using computer components to mimic biological neural architectures, primarily the mammalian brain. Although an area of current and active research, we do not know exactly how the brain works, however it is believed that the brain is a neural net. Signals travel along neurons via voltage spikes known as action potentials which are caused by the movement of ions across the neuron's cell membrane, and the signals pass between neurons via chemical neurotransmitters (the gap crossed between neurons is the synapse). The interaction of these spikes is thought to be a cause of brain waves, thought, learning and cognition. The long-term potentiation of neurons is related to a change in structure of the synaptic cleft, which is thought to result from the Spike Time Dependent Plasticity (STDP) of these synapses and result in Hebbian (associative) learning.
The memristor is the \(4^{\mathrm{th}}\) fundamental circuit element as predicted by Leon Chua. First reported experimentally using this terminology in 2008, memristors have been an object of scientific study for at least 200 years. Memristor theory was first demonstrated in a model of the action of nerve axon membranes in 1976, which was proposed as an alternative to the Hodgkin-Huxley circuit model) and this has led to the suggestion that they would be appropriate components for a computer built using a neuromorphic architecture. Several simulations of neural nets containing memristors have been performed (see for example). Recently, it was reported that circuits combining two memristors with two capacitors could produce self-initiating repeating phenomena similar in form to brain waves.
Perhaps it is not merely the case that memristor models fit neuron behavior, but that neurons themselves are memristive. Thus, we would expect that advances in the study of memristors would explain neurological phenomena (as happened with computer science and STDP). A circuit theoretic analysis of an updated version of Hodgkin-Huxley's model of the neuron has been undertaken. The Hodgkin-Huxley model is often used to explain the transmission of voltage spikes along the neuron. However, this model predicts huge inductances which are not experimentally observed in biology and it has been demonstrated that updating the Hodgkin-Huxley model with memristors avoids this requirement. A recentpaper suggested that memristance could explain the STDP in neural synapses. The authors used memristor equations to adjust simulated spikes, found a similarity to experimentally measured biological synapse action and concluded that a memristive mechanism was behind the biological STDP phenomenon.
In this paper we will show experimentally that memristors spike naturally and do not require a spiking input to cause them to spike in a manner qualitatively similar to neurons. We shall attempt to quantify the spikes. We will then demonstrate that these spikes are also present in theoretical models of memristors and discuss the cause of them. We think that utilizing these naturally-occurring spikes will be the most fruitful way to create neuromorphic memristor architectures.
## 2 Properties of Memristor Spikes
Memristors come in two flavours, charge-controlled (left) and flux-controlled (right) as shown below in Equation 2 where \(q\) is the charge, \(\varphi\), is the magnetic flux, \(M\) is the memristance and \(W\) is the memductance (inverse memristance)
\[V(t)=M(q(t))I(t),\;I(t)=W(\varphi(t))V(t)\,.\]
For a charge-controlled memristor we would input a current, \(I\), and measure the voltage, \(V\). Biological neurons may be described as charge-controlled because it is the movement of ions that causes the change in voltage giving rise to a voltage spike. Our memristors are flux-controlled and a change in voltage causes a spike in the current. Thus, creation of a neuromorphic computer with memristors will be using the complimentary effect to the one utilized by nature, in that memristors have voltage-change-caused current spikes and neurons have current-change-caused voltage spikes. That both types of spikes have a similar form arises from the similarity in the underlying electromagnetics, in that circuits can considered as being constructed with either a voltage source or a current source.
Current spikes recorded from a memristor subjected to the voltage square wave in The spike heights are highly repeatable and qualitatively resemble neuronal spikes.
Our memristors are flexible sol-gel titanium dioxide gel layers sandwiched between aluminium electrodes and they show a distinctive large spike that occurs when the voltage is changed. The experiments reported here were carried out with a Keithley 2400 sourcemeter sourcing voltage. There are no spikes in the voltage profiles, (see Figure 2) and no current spikes are seen when the same experiment is done across a resistor. It has been suggested that these spikes are capacitance; however the timescale is too long. The spikes have been reported by other groups in their memristors (see for example,), however they are usually overlooked or attributed to artefacts arising from the experimental set-up or not reported at all (many researchers only report the \(I-V\) curves to demonstrate that they have a memristor). However, the current spike is an equilibrating process that is responsible for the frequency dependence of the \(I-V\) curves. In each voltage step had 40 timesteps (\(\approx\) 3.3s) to equilibrate. If the voltage is scanned quicker than this, the current has not equilibrated and thus current is higher than the equilbration current. Thus, a faster switching time increases the hysteresis of the \(I-V\) loop. This effect increases with frequency until it reaches the limit where the voltage frequency is too fast for the memristor to relax at all and the \(I-V\) curve just traces out the maximal spike currents for each voltage.
These current spikes can be seen whenever a voltage change occurs across the memristor. Unlike some neuronal spikes, the voltage does not need to spike. The current spikes are highly reproducible. For the experiment shown in (10 pairs of positive to negative switches), the standard deviation was 0.0729% of the mean for the negative voltages (where \(n=10\)) and 0.1192% of the positive voltages (where \(n=9\), due to incomplete recording of the first spike)). For the repeated spikes in (3 repeats each of both positive and negative ramps, as shown in figure 4) the largest difference between the spike current repeats was only 3.06\(\times 10^{-9}\)A and only 2.33\(\times 10^{-10}\)A for the equilibrated current - both taken from the positive side as it has a larger hysteresis than the negative side.
The direction of the current spikes is related to the change in voltage, not its sign, so a change from a positive voltage to zero (turning the voltage source off) gives a negative spike and vice versa for a negative voltage to zero. The spike current still flows for a short while after the voltage source has been turned off. This lag is a general thing and has been recorded in several different devices. In different devices the spikes are the same shape and seem to be following similar dynamics. The spike current is proportional to the equilibrated current.
Voltage square wave that the memristor measured in was subjected to.
## 3 A Mathematical Description of the Spikes
As shown in figure 6, there is a steady-state current, \(i_{\infty}\), a spike current \(i_{0}\) and a transition between the two which is a time-dependent transient \(i(t)\).
Voltage ramps for 5 sets of positive voltage ramps-negative voltage ramps were run, to give the spike response in
The spikes for 5 successive runs up and then down the voltage staircase shown in The runs are coloured and overlap. The spikes are highly reproducible on successive runsAn example of a voltage step as applied to a memristor.
An example Spike. Red dashed line: \(\tau_{50}\); orange dotted line \(\tau_{90}\); green dot-dashed \(\tau_{95}\); blue dotted \(\tau_{99}\). Horizontal purple dot-dashed line is \(i_{\infty}\) and the spike height is \(i_{0}\).
We do know that \(i_{0}\) is related to \(i_{\infty}\). Until a thorough experimental study is undertaken, we shall assume that \(i(t)\) is not dependent on \(i_{0}\) as this is what the experimental evidence seems to suggest.
Thus, the time-dependent current response, \(I(t)\) is assumed to be of the form:
\[I(t)=i_{\infty}+i(t)\]
The current response to the voltage is thus:
\[\Delta I=\frac{V}{R(T)}\]
The time taken to get to \(i(t)=0\) the equilibration lifetime which we shall call \(\tau\), and this lifetime is the short-term memory of the memristor and relates to its dynamical properties; from longer time spike studies with our devices, we know that \(\tau\) is approximately 3.3s. We shall define the concept of the equilibration frequency as the 'frequency' associated with changing a descretised triangular voltage waveform such that each voltage step \(n\) lasts for \(\tau\) seconds.
We know that
\[q_{e}=\int I(t)dt.\]
thus, the total measured charge in a memristor spike is
\[\Delta q_{\rm spike}=\int_{t=0}^{\tau}=i(t)dt+i_{\infty}\tau.\]
This number includes all the charge carrying species in the system. Knowledge of this number may help us elucidate the mechanism of the spikes. For our example system shown in figures 6, we have an \(i_{0}\) of \(1.37\times 10^{-8}\)A, an \(i_{\infty}\) of \(2.40\times 10^{-10}\)A, with the \(\tau_{50}\) of 0.56s and an \(\tau_{90}\) of 0.84s, which shows how quick the fall off is (and \(\tau_{95}\) of 1.13s and \(\tau_{99}\) of 2.34s, as drawn in figure 6).
The resistance profile for the memristor subjected to the voltage in Note that the ‘zero’ resistance is due to zero measured resistance as no voltage is applied, not a true zero resistance.
## 4 The Mem-Con Theory as Applied to Memristor Spikes
The mem-con model of memristance is a recently announced theoretical model that relates real world \(q\) and \(\varphi\) to Chua's constitutive equations and has been successful in modeling our memristors. The mem-con theory has the concept of a memory property, the physical or chemical attribute of the device that holds the memory of the device. In titanium dioxide (and many others) it is related to the number of the oxygen vacancies. The presence of oxygen vacancies allows the creation of a doped form of titanium dioxide TiO\({}_{2-x}\) which is more conducting than the undoped (TiO\({}_{2}\)) form. The mem-con theory requires that we calculate the memristance from the point of view of the memory property, i.e. the ions.
Theoretically, the voltage step is a discontinuous function and the voltage changes from voltage A, \(V_{A}\) to voltage B, \(V_{B}\) in an infinitesimal, i.e. \(\Delta V=\frac{V_{B}-V_{A}}{t},t\rightarrow\delta t\). Experimentally this is not the case of course, but the response timescale of the memristor is long enough that we needn't worry about this approximation.
Thus to elucidate what happens to the memristor during a current spike, and how the final current \(i_{\infty}\) is determined, we take differences of the mem-con theory. We shall assume our device is a TiO\({}_{2}\) memristor, with oxygen vacancies acting as the memory property.
As a reaction to the voltage step, we get a current spike, \(\Delta i\), which can be expressed as a volume current within the device as \(\Delta J\) as given by:
\[\Delta\vec{J}=\{\frac{\Delta q_{v}\mu_{v}\vec{L}}{vol},0,0\}\]
The change in the magnetic field at point \(p\), \(\Delta\vec{B}(p)\) would then be:
\[\Delta\vec{B}(p)=\frac{\mu_{0}}{4\pi}\int\frac{\Delta J\vec{j}\vec{\times}\vec {r}}{r^{2}}d\tau \tag{1}\]
Thus, to get a measure of the effect of the spikes, we need to solve this integral over a time-interval covering from the start of the spike to the tail-off of the memristor's response. The voltage input is non-integrable, but we can integrate from the start of the step, which we shall take as \(t(n)\) where \(n\) is the number of the voltage step, which is zero for this case if it is understood that this is not the zero at the start of an experiment with many steps (i.e. we are considering a case as in figure 6) to when the memristor has responded, which we shall take as \(T\). Dependent on the situation \(T\) can be one of many values, for a staircase we would presumably want \(T=t(n+1)\) where \(t(n+1)\) is the time that the voltage step is input. For a response to a single step function we could take the integral out to \(\infty\) (which is what we shall do here). For experimental purposes we might be more interested in integrating to \(\tau\) or \(\tau_{90}\).
Solving the integral gives:
\[\Delta\vec{B}(p)=\frac{\mu_{0}}{4\pi}L\mu_{\nu}\Delta q\{0,-xzP_{y},xyP_{z}\}\]
with
\[P_{y} = \frac{F}{2\left(\Delta w^{2}+E^{2}+F^{2}\right)^{\frac{3}{2}}}\] \[-\frac{1}{2\Delta wEF}\frac{\left(\Delta wE\left(F^{2}\left(E^{2} +F^{2}\right)^{2}+a+b\right)\right)}{c}\] \[+F\arctan\left(\frac{\Delta wE}{F\sqrt{\Delta w^{2}+E^{2}+F^{2}}} \right),\] and \[P_{z} = \frac{E}{2\left(\Delta w^{2}+E^{2}+F^{2}\right)^{\frac{3}{2}}}\] \[-\frac{1}{2\Delta wEF}\frac{\left(\Delta wF\left(E^{2}\left(E^{2} +F^{2}\right)^{2}+a+b\right)\right)}{c}\] \[+E\arctan\left(\frac{\Delta wF}{E\sqrt{\Delta w^{2}+E^{2}+F^{2}}} \right)\,,\]
where
\[a=\Delta w^{4}\left(2E^{2}+F^{2}\right)\]
\[b=\Delta w^{2}\left(2E^{4}+5E^{2}F^{2}+2F^{4}\right)\]
\[c=\left(\Delta w^{2}+F^{2}\right)\left(E^{2}+F^{2}\right)\left(\Delta w^{2}+E^ {2}+\,F^{2}\right)^{\frac{3}{2}}.\]
Where the effect on the magnetic field is due to both the influx of charge and the resulting movement of the boundary between doped and undoped TiO\({}_{2}\).
To calculate the change in magnetic flux through a surface associated with this field, \(\varphi\), we need to take the surface integral
\[\Delta\varphi=\int\Delta\vec{B}\cdot d\vec{A}\]
As it is a surface integral, to calculate the magnetic flux we need to pick a surface to evaluate over. It makes sense to choose a surface that correlates to one of the surfaces of the device. Picking the surface just above the device (\(0<x<D\), \(0<y<E\), \(z=F\)), we use the surface normal area infinitesimal, \(d\vec{A}\), which is given by \(d\vec{A}=\{0,0,ij\}\). As is standard in electromagnetism, we integrate over the entire area. The limits of the surface are taken to be the dimensions of the device.
Thus we derive the general form of the magnetic flux passing through a surface \(i\)-\(j\): where, because \(\varphi\) is entirely dependent on \(q\), which is time-varying, we can include the time varying effects by taking the differentials thus
\[\delta\varphi=\frac{\mu_{0}}{4\pi}L\mu_{v}ijP_{k}\delta q_{v}\, \tag{2}\]
And, as in mem-con theory, by using Chua's constitutive relation for the memristor, we can then arrive at the change in the Chua memristance as experienced by the ions:\[\Delta M_{q}\left(\Delta q_{v}\left(t\right)\right)=UX\mu_{v}\Delta P_{k}\left( \Delta q_{v}\left(t\right)\right)\text{,} \tag{3}\]
Equation 3 can be considered as three separate parts:
1. \(U\), the universal constants: \(\frac{\mu_{0}}{4\pi}\).
2. \(X\), the experimental constants: \(DEL\).
3. the material variable: \(\mu_{v}P_{k}\) (called \(\beta\) is ref [!!Mem-Con]), this includes the physical dimensions of the doped part of the device and the drift speed of the dopants.
Writing out the differences explicitly of equation 3 we end up with:
\[M(B)=M(A)+UX\mu_{v}[P_{k}(q_{B})-P_{k}(q_{A})],\]
The Chua memristance is written for the vacancy charge, so to put it into the standard format for the electronic current we need to scale it thus:
\[R_{M}=C_{M}M,\]
### Conservation function
The memory part of the function only describes the effect of the memristance change on the doped part of the memristor.
\[w(t)=\frac{\mu_{v}Lq(t)}{EFv_{d}}.\]
Thus, the difference in conservation function, \(\Delta R_{\mathrm{con}}\), written as a difference equation is:
\[R_{\mathrm{con}}(B)=R_{\mathrm{con}}(A)+\frac{\left(D-[w(B)-w(A)]\right)\rho_{ \mathrm{Off}}}{EF}\]
The mem-con model describes a memristor by being the sum of the memory and conservation functions (both written for the electrons) and this then gives us the following expression for the change in time-varying resistance, \(R(t)\), as measured after a change from \(V_{A}\to V_{B}\) as:
\[\Delta R(t) = c_{m}M(A)+R_{\mathrm{con}}(A)+\frac{\rho_{\mathrm{off}}D}{EF}\] \[+c_{M}UX\mu_{v}[p_{k}(q_{B}(t))-p_{k}(q_{A}(t_{T}))]\] \[-\frac{L\rho_{\mathrm{off}}\mu_{v}[q_{B}(t)-q_{A}(t_{T})]}{E^{2}F ^{2}v_{d}},\]
This equation has two parts:1. \(S\), the time-invariant part, which is: \(c_{m}M(A)+R_{\rm con}(A)+\frac{\rho_{\rm off}D}{EF}\)
2. \(Y\), the time variant part: \(c_{M}UX\mu_{v}[p_{k}(q_{B}(t))-p_{k}(q_{A}(t_{T}))]\) \(-\frac{L\rho_{\rm off}\mu_{v}[q_{B}(t)-q_{A}(t_{T})]}{E^{2}F^{2}v_{d}}\),
In the above equation 4 highlights a few subtleties of the model. \(p_{k}\) and \(q\) are time-dependent and thus change after the voltage step from \(V_{A}\to V_{B}\). If we ask the question of what the difference will be between the equilibrated current at \(V_{A}\) and that at \(V_{B}\), \(\Delta R_{A_{\infty}\to B_{\infty}}\) equation 4 collapses to:
\[\Delta R = c_{m}M(A)+R_{\rm con}(A)+\frac{\rho_{\rm off}D}{EF}\] \[+c_{M}UX\mu_{v}[p_{k}(q_{B}(\tau))-p_{k}(q_{A}(\tau))]\] \[-\frac{L\rho_{\rm off}\mu_{v}[q_{B}(\tau)-q_{A}(\tau)]}{E^{2}F^{2 }v_{d}}\,,\]
What if there was previous step in which the device did not equilibrate to \(i_{\infty}\)? This would happen if the voltage was changed quicker than \(\tau\), i.e. \(T\) where \(T<\tau\). The \(q_{A}(t_{T})\) is not \(q_{A}(\tau)\) and thus needs to be shifted by its value as a proportion of \(\tau\). As an example, if we sped the voltage ramps up to 90% of the equilibration frequency, \(q_{A}\) would be \(q_{A}(\tau_{90})\) and the length of a time step would be \(\tau_{90}\). At first glance it might appear that this would merely modulate the starting point for \(q_{B}(t)\), which, at times under \(t<\tau\), this would be time dependent. But there is the interaction between \(q_{B}(t)\) and \(q_{A}(t_{T})\), the memristor hasn't finished responding to \(V_{A}\) and that response should be mixed in with \(V_{B}\), further complicating predictive efforts.
## 5 Modeling Memristor Spikes
The mem-con model consists of sum of two components: the memory function, \(M_{e}\), and Conservation function, \(R_{c}\). The memory function has a fitting parameter \(c_{m}\) within the model to account for the conversion between the material's resistance as for an oxygen vacancy and as for an electron. The conservation function has the fitting parameter \(c_{c}\) which accounts for the resistivity of the undoped material, \(\rho_{\rm off}\), which may not be the same as the bulk titanium dioxide. \(R_{\rm on}\) is the final fitting parameter and relates to the resistivity of the doped material, which is the memristor in the equilibrated state and any resistance in the wires.
\[I(t)=\frac{V}{R_{\rm on}}-\frac{V}{c_{c}R_{c}(t)-c_{m}M_{e}(t)}\;.\]
As figures 8 and 9 shows, the mem-con model fits these spikes quite well and much better than an exponential fit. For the positive spike, \(c_{M}-3.83\times 10^{6}\), \(c_{c}=1.76\times 10^{6}\) and \(V/R_{\rm on}=2.97\times 10^{-9}\), with a summed square of residuals of 1.61\(\times 10^{-17}\). For the negative spike, \(c_{M}-1.06\times 10^{6}\), \(c_{c}=1.86\times 10^{-6}\) and \(V/R_{\rm on}=-3.16\times 10^{-9}\), with a summed square of residuals of 1.63\(\times 10^{-17}\). For the exponential fit, \(I(t)=Ae^{\lambda t}\), and \(A=3.96\), \(\lambda=-19.5\) with a summed square of residuals of 2.43\(\times 10^{-15}\). The exponential fit could be fit to either the short time spike or the long time tail but not both, the short term spike fit goes erroneously to zero and the long-term spike fit grossly over-estimates the size of the spike. Furthermore, there is no experimental justification for using an exponential fit, unlike the mem-con fit. This model can be utilized to perform simulations of memristor spiking networks to test out possible neuromorphic architectures.
## 6 What is the Mechanism?
The memory property of these memristors is the oxygen ions, usually viewed as positive holes in a semi-conducting material. We suspect that the motion of these ions is behind both the spikes and the memristance as we postulate that the two are the same phenomena. The current that flows at \(t=0\)s may be the ionic current, which would have a greater inertia, and thus takes longer to stop compared to the electrons, which may explain the cause of the devices hysteresis. This current flow can also explain the open-loop memristors (suggested by Pershin and di Ventra to explain experimental results such as which are similar to ones seen in our labs and others'). The spike shape would then be the result of the equilibrating of the ionic current to a change in voltage. We expect that the timescale and dynamics of the spikes will relate to the frequency effects seen in memristors. However, there is much further experimental work to be done to prove this mechanism.
## 7 Comparison between memristors and neurons
Chua's definitions of his two types of memristors, flux and charge controlled, was given above. The mem-con model has the concept of a two-level system where we have two charge carriers, \(q\), our memory property and \(e^{-}\) the electronic current which is what is measured in an experiment. Level 0 is the relationship between the vacancy charge, \(q\) and vacancy flux, \(\varphi\).
A longer-term spike response fit by the mem-con theory. The mem-con theory fits the experimental data well, the best result fitting the data with an exponential is added as a comparison. Blue dots: experimental data, red line: mem-con fit, green line: exponential fit to the spike.
The circuit measurables are the voltage, \(V\) and the total current \(I\) where \(I=\frac{dq}{dt}+I_{e^{-}}\).
For our memristors, driven by a voltage, the right hand side of summarizes the operation. There is a change in voltage, which acts on the electrons and the vacancies, causing a change in the number of charge carriers (\(\Delta e^{-}\) and \(\Delta q\). The change in \(q\) causes a change in the magnetic flux associated with \(q\) and thus a change in the Chua memristance. This, due to the conservation of space, causes a change in the amount of material described by the conservation function \(R_{c}\), which then changes the total resistance \(\Delta R\). This change in resistance will draw more current, \(e^{-}\) and thus the change in the number of electrons is influenced by both the change in voltage and the change in resistance that change has caused. The change in total current is due to both the electrons and the vacancies.
A neuron is the opposite way round, see the left hand side side of The cell is always pumping ions back and forth, so we have a change current due to an influx of charge carrier. This causes a change in magnetic flux and affects the total resistance (the values of the memory and conservation functions for this system have not yet been worked out). This change in resistance causes a voltage spike. Thus, similarities can been seen between neuronal voltage spikes and memristive current spikes, in that they are the opposite way round with respect to the circuit measurables, in that the memristor as operated here is a current response to a voltage-sourced circuit, and the neuron is a voltage response to a current-sourced circuit. Essentially the shape of the circuit variable, i.e. that which is being measured, is qualitatively similar.
## 8 Towards Neuromorphic Computing
It has been suggested since 1976 that neurons are memristive, but experimental evidence for neuron-like spiking in memristors had not been collated or analyzed in this way before. If this spiking behavior is an integral result of memristance then it is evidence for the suggestion that neurons may be memristive in action and further understanding of memristor theory may further the neurological understanding.
A longer-term negative spike, demonstrating that the negative spikes are fit equally well by the mem-con theory. Blue dots: experimental data, red line: mem-con fit.
This work shows that to make neuromorphic computers that compute with spikes memristors are an obvious choice for this task as they spike naturally. Interruption to the equilibrating current curves as shown in figure, by, for example, changing voltage, would potentiate the connection by modifying the memristance and could thus be used to do STDP with memristors without requiring CMOS neurons to generate the spikes.
## 9 Conclusion
Memristors, when subject to a change in voltage, undergo a current spike. This spike has been shown to be reproducible and repeatable. The mem-con theory have been shown to fit the time-dependent current behaviour with only two fitting parameters (which come from the missing material values in the theory) suggesting that this \(I-t\) spike behaviour is an aspect of memristance. Rewriting the mem-con theory as a difference equation allows the formulation of a predictive equation to related the equilibrated currents at different (and successive) voltages. Application of the equilibration lifetime (\(\tau\)) to this equation highlights where the time-responsive interactions might arise in a memristor switched faster than the equilibration frequency.
|
10.48550/arXiv.1302.0771
|
Observation, Characterization and Modeling of Memristor Current Spikes
|
Ella Gale, Ben de Lacy Costello, Andrew Adamatzky
| 2,254
|
10.48550_arXiv.2211.15812
|
## 1 Introduction
Electromagnetic field-assisted processes in the form of microwave heating span a wide range of research and industrial areas, such as low-temperature sintering, low-temperature de-crystallization, enhancement of reaction rates, and catalytic effects in organic and inorganic synthesis. While the source of the heating mechanism is well understood, certain effects such as enhancement of reaction rates, increased yield, and creation of distinct transition pathways, which cannot be achieved through the conventional heating mechanism, have led to the term non-thermal or EMF specific effects. The term non-thermal effect is loosely described as any reaction or process that cannot be achieved through the conventional heating technique, which has gained a reputation for being something mysterious. There has been an increasing debate among the scientific community regarding the non-thermal effects being just a manifestation of the thermal effect in the form of non-uniform localized heating, hotspots, and/or inaccuracy in temperature measurement. Exhaustive reviews have been published on both the thermal and non-thermal effects. While there are satisfactory explanations for the observed phenomena in each case, modeling these effects to get an intuitive understanding of the process remains challenging.
The microwave energy is typically too low to cause electronic transitions in materials, and unlike the laser, microwave absorption has a more subtle effect on the system. The microwave field interacts with the dipole of the molecule, increasing its energy. The thermal effect in microwave absorption can be explained due to energy lost to the system due to the relaxation of these excited dipole moments. By tracking the internal vibrational energy distribution of the system, we can get information regarding the equilibrium properties and temperature of the system. We can model microwave absorption as multiphoton absorption in a collisional environment to implement this idea. Intramolecular vibrational redistribution and vibrational energy transfer can be used to model energy dissipation among vibrational levels. As we will see, collisional effects play a vital role in the redistribution of energy, which gives rise to the heating of the sample as an observable quantity.
Ma J., in their paper, used internal energy distribution to explain non-thermal microwave effects using a two-channel process where he hypothesized that microwave activates a rapid channel to enhance the reaction rate demonstrating the distribution of internal states deviate from equilibrium under a strong microwave field. The non-thermal microwave effect in certain applications has been explained by the increase in pre-exponential factor in the Arrhenius Equation, which directly influences the molecular collisions, Decrease in the activation energy, which directly increases the reaction rate, Localized microscopic high temperatures Intermediate transition states, and polar specific effects in the medium.
This paper introduces a model using the master equation formulation to demonstrate the effect of the microwave as a multiphoton absorption in a collisional environment.
## 2 Methodology
### Master Equation
The temporal evolution of the population of molecules under thermal dissociation, including the multiphoton absorption, can be described by the energy-grained master equation:\[\begin{split}\frac{dN_{i}}{dt}=\frac{I(t)}{\hbar\omega}[\sigma_{i,i-1 }N_{i-1}+\frac{\rho_{i}}{\rho_{i+1}}\sigma_{i+1,i}N_{i+1}-(\sigma_{i+1,i}+\frac {\rho_{i}}{\rho_{i+1}}\sigma_{i,i-1})N_{i}]\\ +\omega\sum_{j}P_{i,i+1}N_{i}-\omega\sum_{j}P_{i,i-1}N_{i}-\sum_{ m}k_{m}N_{i}\end{split} \tag{1}\]
I(t) is the intensity (W/cm\({}^{2}\)) of the microwave radiation. \(P_{i}\) is the transfer probability of the molecule from the ith energy level. \(\omega\) is the collisional frequency. \(k_{i}\) is the unimolecular rate constant of the ith channel. \(\sigma_{i+1,i}\) is the microscopic cross-section for the absorption of microwave energy from level i to i+1. \(\rho_{i}\) is the density of states at the ith energy level.
As it is difficult to find a solution for the Master equation given in Eq, a stochastic algorithm is used to solve it, as provided by Gillespie. The algorithm determines the time step and reaction that would occur in an evenly distributed system accounting for random fluctuations. To implement the stochastic process, first, a cumulative reaction term \(k_{0}\) is calculated as the sum of all the reactions \(k_{v}\) occurring in the system, according to Eq. Using random numbers given by \(r_{1}\) and \(r_{2}\), the time step \(\tau\) is added to the simulation time, and the reaction at that time step is chosen, according to Eq and, respectively.
\[k_{0}=\sum_{v=1}^{M}k_{v} \tag{2}\]
\[\tau=(\frac{1}{k_{0}}ln\frac{1}{r_{1}}) \tag{3}\]
\[\sum_{v-1}^{\mu-1}k_{v}<r_{2}k_{0}\leq\sum_{v=1}^{\mu}k_{v} \tag{4}\]
To put it all together, given the initial conditions, the program calculates the reaction rate for the different processes: microwave absorption/emission, elastic/inelastic collision, and dissociation. Each molecule then starts with the initial internal energy. Based on the reaction chosen by Eq and the time step by Eq, the appropriate energy is updated to the molecule's internal energy, and the program progresses. This cycle repeats until the time exceeds the set limit or the molecule dissociates. The flowchart in demonstrates this process.
### 2.2.
The electromagnetic wave interacts with a molecule by distorting the electron cloud in the direction of the applied field. As this distortion relaxes, energy is gained in the system due to intermolecular interactions leading to heating in the system. Thismechanism is known as dielectric heating, and it primarily depends on a cluster of atoms or molecules, and it is not a quantum mechanical phenomenon.
The property of the material that determines the coupling of electromagnetic waves is the relative permittivity. A material with permanent or induced dipoles will store charges if placed between two electrodes in a circuit, and the permittivity of the material defines this charge storing ability. The permittivity is defined as complex variable given in Eq.
\[\epsilon^{*}=\epsilon^{\prime}-j\epsilon" \tag{5}\]
Where \(\epsilon^{\prime}\) is the real part of the permittivity which is related to the charge storage and \(\epsilon"\) is the imaginary part which indicates the loss term. As dielectric heating is known to heat the bulk of the sample, it is directly related to the loss tangent of the system. The loss tangent or the energy dissipation factor is given by Eq.
\[tan\delta=\frac{\epsilon^{\prime}}{\epsilon"} \tag{6}\]
The loss tangent can be experimentally calculated for different temperatures. Dielectric heating can also be viewed as the efficiency of conversion of EMF energy into heat, and it is related to both the dielectric and thermal properties of the system. The relationship is given by Eq.
The flowchart for solving the master equation using Gillespie’s algorithm. The system is initialized to the set parameters of temperature (Temp), Number of Trajectories (N\({}_{Traj}\)), time limit (T\({}_{LIM}\)), Initial energy (E\({}_{BEGIN}\)) and size of energy grain(E\({}_{grid}\) reaction rates and density of states are initially calculated, and the stochastic process begins. Energy (E) is updated at each time step (\(\tau\)) and the simulation for each molecule runs if the set time has exceeded (T\(>\)T\({}_{LIM}\)), or the molecule dissociates (k\({}_{dissociation}\)).
\[P=\sigma|E|^{2}=(\omega\epsilon_{0}\epsilon^{\prime}tan\delta)|E|^{2} \tag{7}\]
Where P represents the power dissipated per unit volume in a material, \(\sigma\) is the conductivity, E is the electric field and \(\omega\) is the angular frequency. The power dissipated in the system is equivalent to the multiphoton jump due to the EMF absorption. In particular, we consider a bulk system with a continuum in the density of states. Therefore, modeling the EMF absorption as a multiphoton absorption is justified as the energy is assumed to be absorbed by all the molecules equally, i.e., we can relate this to the power absorbed by the whole system (not a single molecule). In the next step, the energy gained by the multiphoton absorption is dissipated quickly through collisional effects trying to maintain the equilibrium in the system. The thermal effect is seen as an increase in temperature, while the non-thermal effect is observed as a deviation from this equilibrium.
The EMF intensity plays a crucial role in the dynamics of the system. At higher microwave intensity, voltage breakdown occurs leading to arc discharges and highly ionized gas plasma. Very high values of microwave power(\(>\)10\({}^{6}\)W/cm\({}^{2}\)) can lead to a higher number of molecules in the excited states with insufficient time for redistribution of energy which can lead to wrong interpretation of the result.
### Molecular Collisions
The collision frequency \(\omega\) is a function of temperature, pressure, and atomic potential. The molecules are modelled as Lennard-Jones sphere and interact via the Lennard-Jones potential between molecules A and bath gas M.
\[\omega=\pi\sigma_{LJ}[M]\sqrt{\frac{8RT}{\pi\mu_{A-M}}}\Omega_{A-M}^{*} \tag{8}\]
Where \(\Omega_{A-M}^{*}\) is the collision integral given in Eq, [M] the number density of the system, \(\sigma_{LJ}M\) the collision diameter, T(K) the temperature, and \(\mu\) the atomic mass. The collision integral in this model is given by where \(\epsilon_{A-M}\) is the depth of the Lennard Jones potential well.
\[\Omega\approx[0.636+0.567\log\frac{kT}{\epsilon_{A-M}}] \tag{9}\]
A bi-exponential model gives the transfer probability function used in this model as in Eq..
\[P_{ij}=(1-a)e^{\frac{(E_{i}-E_{j})}{\alpha}}+be^{\frac{(E_{i}-E_{j})}{\beta}} \tag{10}\]
Where a,b,\(\alpha\), and \(\beta\) are coefficients for collisional energy transfer. The up and down collision are related by the balance equation given by:\[\frac{P_{i,j}}{P_{j,i}}=\frac{\rho_{i}}{\rho_{j}}e^{(\frac{E_{i}-E_{j}}{k_{b}T})} \tag{11}\]
The energy transfer, collision, and dissociation processes are all connected to the density of states. Many methods exist to calculate the density of states based on the size and accuracy of the molecule. Statistical tools and computer programs are available to generate the density of states of a molecule based on the vibrational and rotational frequencies.
The system parameters, such as the energy distribution, photons absorbed /emitted, and trajectories reacted, are stored to evaluate the time evolution of the system. The time bin is fixed and stores the average value in that interval. This gives an advantage to the memory required for each simulation. Considering the system to be linear, the average values of an ensemble of molecules can describe the system's behavior. The size of the energy grain must be chosen to simulate the molecule with sufficient accuracy. Finer energy grain will result in more memory space and would be required in reactions with multiple channels.
### Unimolecular Dissociation
A critical quantity to evaluate the reaction rate is the dissociation rate constant. The Rice-Ramsperger-Kassel-Marcus (RRKM) model is widely used to study the unimolecular kinetics and is based on the fundamental assumption that a micro-canonical ensemble of states exists initially for the activated molecule and is maintained as it decomposes. It is given by:
\[k(E,J)=\frac{W(E,J)}{h\rho(E,J)} \tag{12}\]
Where \(W(E,J)\) is the transition state's sum of states. To include the EMF effect, we can assume that the system reaction to sudden changes is slow, and the average energy shifts as EMF energy is absorbed. Due to the non-equilibrium distribution of the system, the vibrational temperature of the system would be different. This can be considered a change in its vibrational temperature and be included empirically in the RRKM equation, according to Eq.
\[k_{MW}=Ae^{\frac{-E_{crit}}{\beta(E)T}} \tag{13}\]
Here A and \(\beta\) values are determined experimentally by fitting the plot of the observed values.
## 3 Results
EMReact is written in MATLAB, and the source code is available and can be downloaded from the provided link. The code, without the EMF inclusion, is written based on the MULTIWELL suite developed by Barker et. al.. The program utilizes MATLAB functionalities and offers graphical visualization and prompt inputs for a user-oriented approach.
For this work, we have chosen two systems which that are known to exhibit thermal as well as non-thermal effects. Dimethyl Sulfoxide which is a solvent and good EMF absorber, and Benzyl Chloride which is a precursor to several microwaves-assisted reactions. The density of states is calculated for an energy grid of 5 cm\({}^{-1}\) employing the Wang-Landau method implemented in the Desnum package of the MULTIWELL suite, as shown in The density of states, as shown, is relatively smooth even at lower energy which leads, leading to little error during interpolation for the selected energy grid. The vibrational frequencies used to calculate the density of states for the two systems are calculated and experimentally determined by Chen and Naganathappa.
### Dimethyl Sulfoxide
Dimethyl Sulfoxide (DMSO) is an aprotic polar solvent that has been used extensively in microwave-assisted processes. DMSO can couple to the microwave field increasing its temperature due to dielectric losses which makes it one of the reasons it is widely used. The parameters used for the simulations are given in Table 1.
The internal energy distribution at different microwave intensities for a time limit of 5 ms is given in Figure 3(_a_). Without microwave radiation (MW=0W/cm\({}^{2}\)) the internal energy follows a Boltzmann distribution given by \(Ae^{(-(E-E_{1})/k_{B}T)}\). The starting internal energy set at 1000 cm\({}^{-1}\) reaches equilibrium due to the vibrational-vibrational energy transfer by the collisional process in 500 \(\mu\)s. This is shown by the average energy plot shown in Figure 3(_b_). The average energy \(<E>\) of the internal energy distribution depends on the temperature of the system. As microwave radiation is added to the system, the average energy plot shifts based on the microwave intensity.
At the microwave power of 100 W/cm\({}^{2}\), the energy distribution is modified such that the trailing edge of the distribution has a lower slope than the leading edge indicating a non-uniform distribution.
\begin{table}
\begin{tabular}{c c} \hline \hline \(\sigma_{LJ}\)=3.66, & \(\epsilon_{A-M}=168.18\) \\ \(tan_{S}\)=0.8524, & \(\epsilon\)=13, \(\omega\)=2.4 GHz \\ N\({}_{TRAJ}=3000\), E\({}_{BEGIN}\)=1000 cm\({}^{-1}\), T\({}_{TLIM}\)=5 ms, Temp=300 K \\ \end{tabular}
\end{table}
Table 1: Simulation Parameters for DMSO
Density of states for DMSO and Benzyl Chloride
The thermal effect in the form of shifting of the distribution due to temperature increase is not observed due to the time limit of the simulation. The average energy plot for 100 W/cm\({}^{2}\) is less than the average energy at no microwave radiation due to the spread of the energy state in the trailing edge of the energy distribution.
At higher microwave power at 1 kW/cm\({}^{2}\), the thermal and non-thermal effect is evident through the shift in the energy distribution and the slight spread in the population at high energy states. The higher energy states are more susceptible to microwave absorption due to the continuum of the density of states compared to the lower states. The average energy plot is also shown to be shifted higher than the equilibrium distribution.
An important factor in many microwave experiments is the effect of stirring the system under radiation. It has been observed that the microwave effect is reduced under stirring. The effect of stirring can be incorporated into the system by increasing the collisional frequency. The vibrational-vibrational energy transfer is responsible for redistributing the energy in the system leading to equilibrium at a fixed temperature. Increasing the collisional frequency increases this redistribution and effectively cancels the microwave effect. As shown in Figure 4, the internal energy distribution for a system under 2 kW/cm\({}^{2}\) of microwave power represents a thermal equilibrium distribution after stirring.
## 4 Benzyl Chloride
Microwave-assisted synthesis of phosphonium salt through the nucleophilic substitution of benzyl chloride with triphenylphosphine has been reported and has shown to exhibit non-thermal effects at 373 K. Benzyl chloride is a relatively neutral to microwave absorption with a low loss tangent. The simulation parameters are shown in Table 2.
The internal energy distribution at a time limit of 5 ms is shown in Figure 5(_a_) at different MW intensities. The corresponding average energy plot with respect to time is shown in Figure 5(_b_). At low microwave intensity, the shift in the equilibrium is not observed. As Benzyl Chloride is not a good absorber of MW, the temperature changes through the shift in the equilibrium distribution under microwave radiation are not prominent for the time limit of 5 ms. The internal distribution at a microwave power of 100 W/cm\({}^{2}\) and 1 kW/cm\({}^{2}\) does not show many changes compared to the distribution with no Microwave Power.
b: Average energy \(<E>\) with respect to time
The non-thermal effect, seen as a nonequilibrium distribution, is caused by some states pushing to higher energies while the collisional reaction constantly tries to re-distribute this change, resulting in a longer energy trail in the higher energy side with a shift in the peak of the distribution.
Based on the simulation parameters and the competing microwave absorption/emission, there is a microwave intensity beyond which the non-thermal effect will be more evident, which we call the characteristic intensity. The characteristic intensity can be defined as the intensity beyond which the microwave absorption/emission rate dominates over the collision rate. It is expected that microwave induces a nonequilibrium distribution at this level and exhibits a non-thermal effect. shows the microwave absorption/emission rate of the two reactants studied, along with the collisional frequency. The rate of microwave absorption/emission increases linearly with microwave power. It is determined by the permittivity and the density of states of the system; hence, it is a material-dependent value. The microwave characteristic intensity for DMSO is found to be 2 kW/cm\({}^{2}\), and for Benzyl Chloride is 15 kW/cm\({}^{2}\).
### Dissociation Reactions
In studying the enhancement of reaction rate under the microwave, we have implemented unimolecular dissociation simulations with CO\({}_{2}\) and H\({}_{2}\)O\({}_{2}\) molecules in an argon bath under microwave radiation. The microwave effect on CO\({}_{2}\) and H\({}_{2}\)O\({}_{2}\) has been studied and reported for applications in lasers, carbon monoxide conversion, microwave plasma/catalyst and waste-water treatment. The dissociation rate constant for CO\({}_{2}\) is given in the literature. Figure 7(_a_) presents the plot of trajectories reacted with increasing microwave intensity. Following the trend in the figure, the reaction rate increases with both temperature and microwave intensity. At low MW intensities, the reaction rate is too low for temperatures below 1200 K. However, when the MW intensity increases above 10\({}^{4}\) W/cm\({}^{2}\), the reaction rate becomes significant even at 1000 K. At high temperatures, the reaction rate saturates at lower MW intensity. For example, at 1500 K, the rate slows down at about 5\(\times\)10\({}^{4}\) W/cm\({}^{2}\), while at 1000 K, the rate change happens at about 2\(\times\)10\({}^{5}\) W/cm\({}^{2}\). The main reason for this difference is that at higher temperatures, the collision rate dominates and eventually spreads the distribution over the MW absorption and emission; hence, the effect of MW becomes less significant.
For unimolecular dissociation of H\({}_{2}\)O\({}_{2}\), the trajectories reacted with respect to microwave intensity are shown in Figure 7(_b_).
Interaction rate and collisional frequency versus microwave intensity for reactants (DMSO: red, Benzyl chloride: blue) along with the average collisional frequency(black).
b: Reaction Rates for H\({}_{2}\)O\({}_{2}\) for increasing microwave intensities at different temperatures.
The difference in the plot for these two molecular systems can be attributed to the difference in the density of states and dissociation rate constant. At higher temperatures, the collisional effects in the H\({}_{2}\)O\({}_{2}\) molecule play an essential role in the redistribution of energy attained from multi-photon absorption. Reducing the reaction rate at a higher temperature may seem counter-intuitive, but the similar argument of energy dissipation due to collision is the reason for this effect. The collisional parameters at higher temperatures dominate and due to the increase in the accessible density of states at high temperatures leads to faster energy dissipation from photon absorption and a reduction in the reaction rate.
## 5 Conclusion
The discussed model gives a realistic description of the temporal evolution of the population due to multiphoton absorption in the presence of collision. Solving the master equation using a Monte Carlo approach can be used to solve several problems related to EMF-assisted reactions. Through the examples of DMSO and Benzyl Chloride, an idea of how we can differentiate between the thermal and non-thermal effects under EMF radiation is discussed.
Educational insight into stirring leading to the reduction of the EMF effect is explored as a collisional effect. The thermal effect under EMF radiation can be seen as a shift in the internal energy distribution due to the dielectric loss leading to an increase in the temperature of the system. The non-thermal effect was shown as a deviation from the equilibrium distribution with an unequal number of molecules in the high and low energy states.
|
10.48550/arXiv.2211.15812
|
EMReact: A Tool for Modelling Electromagnetic Field Induced Effects in Chemical Reactions by Solving the Discrete Stochastic Master Equation
|
Kelvin Dsouza, Daryoosh Vashaee
| 2,910
|
10.48550_arXiv.2404.14021
|
## I Introduction
Open quantum systems are ubiquitous in nature and have versatile applications across various domains such as loss of coherence in quantum information, quantum memory, quantum transport, proton tunnelling in DNA and energy transfer in photosynthetic systems. Being a multi-body problem, the exact characterization of open quantum systems is not feasible owing to exponential growth in Hilbert space dimension and a large number of environment degrees of freedom. However, the problem becomes more tractable by tracing out environment degrees of freedom \(\mathbf{Tr}_{\mathrm{E}}(\,\cdot\,)\) or treating the environment and/or system within the classical phase space. To investigate open quantum systems, numerous approaches have been developed so far, spanning from entirely classical to fully quantum methods. While each of these approaches has been successful in its own right, they are hampered by many limitations, such as the inability to account for quantum effects, or demanding significant computational resources arising from the need of employing a very small descretization step due to stability constraints. Furthermore, the comprehensive integration of environmental effects, especially in highly non-Markovian scenarios, contributes significantly to the computational overhead.
Neural networks (NNs) present an efficient approach to learn complex spatio-temporal dynamics in high-dimensional space. NNs and other machine learning (ML) methods have proven to be proficient at predicting future time evolution of quantum states as a function of historical dynamics. In addition, NNs can directly predict the future quantum states as a function of time and/or simulation parameters.
However, a crucial aspect of quantum simulations is adherence to fundamental physical principles. In simulating open quantum systems, it is essential for an approach to uphold the core physical principle of conserving the trace (the sum of probabilities for all possible states) of the reduced density matrix (RDM \(\hat{\rho}_{\mathrm{S}}\)), which should always be equal to 1, i.e., \(\mathbf{Tr}_{\mathrm{S}}(\hat{\mathbf{\rho}}_{\mathrm{S}})=1\), where \(\mathbf{Tr}_{\mathrm{S}}\) represents trace over system degrees of freedom.
Despite the appeal of NNs, existing research on ML-based simulations of quantum dissipative dynamics has largely ignored trace conservation. To the best of our knowledge, only one study has mentioned, albeit in the context of a relatively simple system (spin-boson), that ML models, given sufficient data, were able to implicitly learn trace conservation to a reasonable degree. However, we cannot expect that it always holds, especially in much more complex situations and when it is difficult to obtain ample amount of data for implicit learning of the trace conservation. In general, the ML models can implicitly learn physical laws from the data but if left unchecked (unconstrained) or applied for situations too far from the training data, they can also spectacularly fail.
Physics-Informed Neural Networks (PINNs), introduced in 2017, present a promising solution to this problem. By incorporating physical constraints directly into the neural network architecture, PINNs ensure that the model's predictions adhere to underlying physical laws. This approach has been successfully applied across various fields, including fluid dynamics, seismic inversions in 2D acoustic media, chemical simulations, quantum dynamics, and electronic structure calculations.
In this paper, we explore whether NNs inherently conserve trace and demonstrate that unconstrained models can lead to unphysical results due to trace violations. To address this, we develop physics-informed neural networks that significantly reduce trace conservation violations. However, we find that even with the integration of physical knowledge, physics-informed NNs alone are not sufficient. To ensure correct physical behavior, we introduce an uncertainty-aware hard constraint (U-aware HC) approach that enforces perfect trace conservation by design.
The subsequent sections of this paper are structured as follows. In the "Theory and Methodology" section, we establish the foundational theory of open quantum systems and detail the various NN models employed in our study, including physics-agnostic and unconstrained NNs. We highlight the trace violations by these models and introduce physics-informed NNs (PINNs). Additionally, we discuss the associated loss functions used for training these models and introduce the U-aware HC constraint for rigorous enforcement of physical constraints. Following that, in the "Results and Discussion" section, we present our findings, comparing the performance of our PINN approach and HC constraint against existing models. We discuss the effectiveness of these approaches in enforcing physical laws and achieving accurate simulations. Finally, in the "Concluding Remarks" section, we summarize our key findings, explore the broader implications of our study, and outline potential future research directions.
## II Theory and Methodology
Let us consider an open quantum system (S) with \(n\)-number of states coupled to an outside environment (E). The dynamics of the composite system (S+E) is governed by the Liouville-von Neumann equation (\(\hbar=1\))
\[\dot{\mathbf{\rho}}(t)=-i[\mathbf{H},\mathbf{\rho}(t)], \tag{1}\]
As the composite system is an isolated systems, the dynamics is unitary. Assuming the initial state of the system and environment is uncoupled (i.e., \(\mathbf{\rho}=\mathbf{\rho}_{\mathrm{S}}\otimes\mathbf{\rho}_{\mathrm{E}}\)), the non-unitary reduced dynamics of the system can be extracted by taking a partial trace over environment degrees of freedom, i.e.,
\[\tilde{\mathbf{\rho}}_{\mathrm{S}}(t)=\mathbf{Tr}_{\mathrm{E}}\left(\mathbf{U}(t,0 )\mathbf{\rho}\mathbf{U}^{\dagger}(t,0)\right), \tag{2}\]
While most real-world systems technically qualify as "open" due to their environment, the immense complexity arising from all possible environmental interactions (known as the curse of dimensionality) makes exact theoretical solutions impractical. In the following, we present a brief theory of two broadly studied pedagogical systems: the two-state SB model and the Fenna-Matthews-Olson Complex (FMO) complex.
_SB model_: The SB model describes the temporal evolution of a qubit system (two-state system) interacting with an environmental bath comprising multiple independent harmonic oscillators. The system's total Hamiltonian, expressed in the basis of the excited (\(\ket{e}\)) and ground (\(\ket{g}\)) states, is given by:
\[\mathbf{H}=\epsilon\mathbf{\sigma}_{z}+\Delta\mathbf{\sigma}_{x}+\sum_{k}\omega_{k} \mathbf{b}_{k}^{\dagger}\mathbf{b}_{k}+\mathbf{\sigma}_{z}\sum_{k}c_{k}(\mathbf{b }_{k}^{\dagger}+\mathbf{b}_{k}), \tag{3}\]
The environment's creation and annihilation operators for the \(k\)th mode are \(\mathbf{b}_{k}^{\dagger}\) and \(\mathbf{b}_{k}\), respectively, with \(\omega_{k}\) being the mode's frequency. The coupling strength between the system and the \(k\)th environmental mode is denoted by \(c_{k}\). The environmental influence on the system is characterized by an Ohmic spectral density function with a Drude-Lorentz cutoff:
\[J(\omega)=2\lambda\frac{\gamma\omega}{\omega^{2}+\gamma^{2}}, \tag{4}\]
_FMO Complex_: The FMO complex, a trimer in green sulfur bacteria, plays a crucial role in photosynthesis. Each monomer in the complex contains chlorophyll molecules that act as energy transfer sites, typically numbering seven or eight. The energy transfer within an FMO monomer is described by the Frenkel exciton model Hamiltonian:
\[\mathrm{H}= \sum_{n=1}^{N}\ket{n}\epsilon_{n}\bra{n}+\sum_{n\neq m}^{N}\ket{ n}J_{nm}\bra{m}\] \[+\sum_{n=1}^{N}\sum_{k=1}\left(\frac{1}{2}\mathrm{P}_{k,n}^{2}+ \frac{1}{2}\omega_{k,n}^{2}\mathrm{Q}_{k,n}^{2}\right)\mathbf{I}\] \[-\sum_{n=1}^{N}\sum_{k=1}\ket{n}c_{k,n}\mathrm{Q}_{k,n}\bra{n}+ \sum_{n=1}^{N}\ket{n}\lambda_{n}\bra{n}, \tag{5}\]
The environmental contribution is represented by \(P_{k,n}\) and \(Q_{k,n}\), the momentum and coordinate of the \(k\)th mode interacting with site \(n\), with \(\omega_{k,n}\) as the mode's frequency. The identity matrix \(\mathbf{I}\) ensures dimensional consistency. \(c_{k,n}\) is the coupling strength between the \(k\)th mode and site \(n\), and \(\lambda_{n}\) is the reorganization energy for site \(n\).
For our analysis, we utilize the same Ohmic spectral density function with a Drude-Lorentz cutoff as in Eq., assuming a uniform spectral density across all sites.
### NN-accelerated quantum dissipative dynamics
Within NN framework, learning the time evolution of \(n\)-dimensional RDM can be defined as learning a mapping function \(\Psi:\mathbb{R}^{n\times k}\mapsto\mathbb{R}^{n\times r}\) which takes a vector of descriptive variables, \(\mathbf{x}\in\mathbb{R}^{n\times k}\), and maps it to the corresponding target RDM, \(\mathbf{y}\in\mathbb{R}^{n\times r}\). NN approaches for this task can be categorized into two main types: recursive and non-recursive, depending on how they handle the descriptor and inference.
_Recursive methods_: In recursive NN methodologies, a mapping function denoted as \(\Psi_{\text{rec}}\), is employed to predict future RDMs based on their past history. This approach mimics traditional quantum dynamics, where the evolution at any given time explicitly depends on the current state and implicitly on the past states. Mathematically, a recursive method can be described as:
\[\Psi_{\text{Rec}}: \{\mathbb{R}^{n\times r}\}\rightarrow\mathbb{R}^{r}\quad\text{ such that}\] \[\Psi_{\text{rec}}\left[\ldots,\tilde{\mathbf{\rho}}_{\text{S}}(t_{m-1 }),\tilde{\mathbf{\rho}}_{\text{S}}(t_{m})\right]=\tilde{\mathbf{\rho}}_{\text{S}}(t_ {m+1}), \tag{6}\]
Recursive methods make predictions iteratively. The predicted RDM (\(\tilde{\mathbf{\rho}}\)s) at time t is added to the history, and the oldest one is discarded to maintain a fixed-size memory. This updated history becomes the new input for the next prediction.
_Non-recursive methods_: Non-recursive methods, as seen in 28,29 learn the mapping function \(\Psi\) as a function of simulation parameters and/or temporal information. The time-dependent non-recursive method as used in 28, establishes a mapping function (\(\Psi_{\text{A}\text{IQD}}\)) between RDM and simulation parameters including time \(t\). Mathematically:
\[\Psi_{\text{A}\text{IQD}}: \mathbb{R}^{p}\rightarrow\mathbb{R}^{r}\] \[\text{such that}\quad\Psi_{\text{A}\text{IQD}}(t,\mathbf{p})= \tilde{\mathbf{\rho}}_{\text{S}}(t). \tag{7}\]
This approach allows for parallel computation of all time steps since the prediction for each step does not depend on the output of the previous step. In time-independent non-recursive methodology, the mapping function \(\Psi_{\text{OSTL}}\) predicts the entire trajectory of the RDM for a set of time steps \(t_{1}\) to \(t_{k}\) in one go:
\[\Psi_{\text{OSTL}}: \mathbb{R}^{p}\rightarrow\mathbb{R}^{k\times r}\quad\text{such that}\] \[\Psi_{\text{OSTL}}(\mathbf{p})=[\tilde{\mathbf{\rho}}_{\text{s}}(t_{1 }),\tilde{\mathbf{\rho}}_{\text{s}}(t_{2}),...,\tilde{\mathbf{\rho}}_{\text{s}}(t_{k})]. \tag{8}\]
### Limitations of existing NNs for open quantum systems: purely data-driven approaches
In the NN framework, establishing the mapping function \(\Psi\) between the descriptor \(x\) and its target dynamics can be approached in two ways: purely data-driven or based on known physical laws and constraints. Unfortunately, current research, including our own work, on machine learning (ML)-based simulations of open quantum systems, relies exclusively on data-driven approximations of the mapping function \(\Psi\). As a result, these models often fail to capture underlying physical laws, leading to non-physical RDMs that violate trace conservation.
Here, we classify purely data-driven NNs into two categories: "physics-agnostic NNs" and "unconstrained NNs". Physics-agnostic NNs are models that are not exposed to the complete data and thus remain unaware of the underlying physical laws and constraints. Unconstrained NNs, in contrast, are exposed to the entire data but do not incorporate physical constraints in their loss functions.
To emphasize on the issue of trace-violation by these data-driven NNs, we show their performance in with two examples: the relaxation dynamics within the SB model and the excitation energy transfer (EET) in the 7-site Fenna-Matthews-Olson (FMO) complex. As shown, these data-driven models fail to conserve the trace in both processes. In each case, we utilize convolutional neural networks (CNNs) and OSTL-based recursive dynamics propagation (Rec-OSTL)
\[\Psi_{\text{Rec-OSTL}}: \{\mathbb{R}^{n\times r+p}\}\rightarrow\mathbb{R}^{k\times r} \quad\text{such that}\] \[\Psi_{\text{Rec-OSTL}}\left[\ldots,\tilde{\mathbf{\rho}}_{\text{S}}(t _{m-1}),\tilde{\mathbf{\rho}}_{\text{S}}(t_{m}),\mathbf{p}\right]=\] \[[\tilde{\mathbf{\rho}}_{\text{s}}(t_{1}),\tilde{\mathbf{\rho}}_{\text{s} }(t_{2}),...,\tilde{\mathbf{\rho}}_{\text{s}}(t_{k})]\,. \tag{9}\]
We use MLQD package and train the models on data from the QD3SET-1 database (see Results and Discussion section for details). The training approach mirrors state-of-the-art methods reported previously.
In essence, for the physics-agnostic scenario (and C), we train individual CNNs for each diagonal RDM element, employing a loss function that gauges the deviation of NN-predicted values \(\tilde{\tilde{\rho}}_{\text{S},nn}\) from their reference counterparts \(\tilde{\rho}_{\text{S},nn}\):
\[\mathcal{L}_{nn}=\sum_{m=1}^{M}\left(\tilde{\rho}_{\text{S},nn,m}-\tilde{\rho} _{\text{S},nn,m}\right)^{2}, \tag{10}\]
As these models are not exposed to the dynamics of all states, they lack knowledge of trace conservation. Weshow that a much better solution is the unconstrained NN--a single, multi-output CNN designed to learn all RDM elements, incorporating a loss function that aggregates errors across all states (sites) (and D):
\[\mathcal{L}_{\text{multi}}=\sum_{n=1}^{N}\mathcal{L}_{nn}. \tag{11}\]
However, despite being exposed to the dynamics of all states, this solution still exhibits minor but noticeable trace violations. It is important to note that trace violations can be reduced to some extent with additional training, as demonstrated in Fig. S1 of the Supporting Information. However, further improvement becomes limited as the model approaches the point of overfitting. Additionally, our observations indicate that the improvement in trace conservation with increasing memory time \(t_{m}\) is somewhat unpredictable and does not follow a consistent trend. Despite this, there was a noticeable improvement in the accuracy of the dynamics predictions, as shown in Table S1.
### Our proposed solution: PINNs and beyond
In the preceding subsection, we explored the shortcomings of the state-of-the-art purely data-driven NNs for simulating open quantum systems where they often struggle to enforce fundamental physical laws like trace conservation. This leads to inaccurate and non-physical results. To address this limitation, we first explore PINNs which integrate physical constraints into the loss function inspired by similar ideas in the literature.
\[\mathcal{L}_{\text{Tr}}=\sum_{m=1}^{M}\left(1-\sum_{n=1}^{N}\tilde{\rho}_{ \text{S},nn,m}\right)^{2}. \tag{13}\]
In these equations, we can tune \(\mathcal{L}_{nn}\) and the deviations from trace conservation by weight factors \(\alpha\) and \(\eta\), respectively. Here we use \(\alpha=2.0\) and \(\eta=1.0\). Note that the unconstrained NN with the loss defined by Eq. is a special case of the PINN with \(\alpha=1.0\) and \(\eta=0\).
While PINNs significantly improve trace conservation compared to purely data-driven NNs, they can still exhibit minor violations (as we'll demonstrate later). This is because the physical constraints incorporated within the PINNs loss function are typically considered "soft." In simpler terms, PINNs are nudged towards satisfying the constraints during training, but they aren't strictly enforced.
To overcome the limitations of PINNs, we propose a novel approach that enforces trace conservation by design. This approach utilizes an U-aware HC (uncertainty-aware hard-coded) constraint, guaranteeing strict adherence to physical laws during simulations. Unlike PINNs, the U-aware HC constraint operates outside of the loss function. This allows for a more direct and rigorous enforcement of the trace conservation law, rectifying potential violations during the simulation process.
The key idea is as follows: After making predictions with machine learning models, there will inevitably be a deviation from perfect trace conservation. We can calculate this residual deviation for each time step as:
\[\Delta\text{Tr}(t)=1-\sum_{n=1}^{N}\tilde{\rho}_{\text{S},nn}(t). \tag{14}\]
We can redistribute the residual deviations between each state as:
Trace violations in quantum dissipative dynamics using machine learning. Panels A and C in their respective order illustrate trace violations in a physics-agnostic scenario for a symmetric SB model and the 7-site FMO complex, where a CNN is trained for each state (site). Panels B and D demonstrate the improvement achieved with the unconstrained multi-output CNN for the same two systems. In all cases, an initial dynamics of length 0.2 (in the respective time units), with ideal trace conservation, serves as the seed for model predictions, derived from reference calculations. For the symmetric SB model, results are shown for an unseen dynamics with a characteristic frequency \(\gamma/\Delta=9.0\), system-bath coupling \(\lambda/\Delta=0.6\), and inverse temperature \(\beta\Delta=1.0\). For the FMO complex, the initial excitation is considered on site-1, with parameters \(\gamma=400\) cm\({}^{-1}\), \(\lambda=40\) cm\({}^{-1}\), and temperature \(T=90\) K. Further details on training and prediction can be found in the Results and Discussion section.
\[\tilde{\rho}_{\mathrm{S},nn}^{\mathrm{HC}}(t)=\tilde{\rho}_{\mathrm{S},nn}(t)+w_{n }(t)\Delta\mathrm{Tr}(t). \tag{15}\]
Here, we need to make such a choice for state-specific weighting factors \(w_{n}\) that the trace is one. Also, it should be statistically motivated. Different states might be predicted with different uncertainty and for certain predictions we want smaller corrections (smaller weighting factors). Hence, we also need state-specific uncertainty quantification (UQ) of NN predictions. Similar problems were also faced in the prediction of partial atomic charges predicted by statistical models which do not necessarily add up to integer values: the suggested solution also was to redistribute the deviation from the correct total charge over atoms based on the UQ calculated as the disagreement between the models in ensemble. This shows how very different research field can inspire the solutions in the unrelated field.
Here we introduce a new approach for UQ. We train an additional, auxiliary multi-output CNN with the same loss function as the main PINN but we shift the reference values by a prior \(p_{2}\) (we assume that the main PINN model is trained with prior \(p_{1}=0\)). In other words, we train the CNN on \(\tilde{\mathbf{\rho}}_{\mathrm{S}}+p_{2}\mathbf{J}\) (\(\mathbf{J}\) is a unit matrix with all elements 1) with the predictions given by:
\[\tilde{\rho}_{\mathrm{S},nn}^{\mathrm{aux}}(t)=\tilde{\rho}_{\mathrm{S},nn}^{ \mathrm{aux-NN}}(t)-p_{2}\mathbf{J}. \tag{16}\]
The UQ metric is given then as the absolute deviation of the \(\tilde{\rho}_{\mathrm{S},nn}^{\mathrm{aux}}(t)\) from the main model predictions:
\[D_{nn}(t)=\big{|}\tilde{\rho}_{\mathrm{S},nn}^{\mathrm{aux}}(t)-\tilde{\rho}_ {\mathrm{S},nn}(t)\big{|}. \tag{17}\]
The state-specific weighting factors \(w_{n}\) we suggest to obtain as the normalized distances:
\[w_{n}(t)=\frac{D_{nn}(t)}{\sum_{n=1}^{N}D_{nn}(t)}. \tag{18}\]
The implementation of Eq. with the weighting factors defined with the Eq. ensures that \(\mathbf{Tr}_{\mathrm{S}}\left(\tilde{\rho}_{\mathrm{S}}^{\mathrm{HC}}(t) \right)=1\). It's crucial to distinguish our proposed U-aware HC constraint-based approach from the conventional trace normalization technique, \(\tilde{\rho}_{\mathrm{S}}/\mathbf{Tr}_{\mathrm{S}}\left(\tilde{\rho}_{\mathrm{S }}\right)\), commonly employed in non-trace conserving traditional methods.
Here's why our proposed U-aware HC constraint approach stands out:
* **Generality:** The U-aware HC constraint approach is purely machine learning-based approach and not limited to trace conservation. It can be tailored to enforce various physical constraints across diverse domains within machine learning studies. For example, it could be used to ensure the preservation of total charge in simulations of molecular systems, especially when learning individual charges for each atom.
* **Uncertainty-Aware Correction:** The U-aware HC constraint approach goes beyond simple normalization by incorporating an UQ metric (Eq.) along with a weighting factor (Eq.). This allows for targeted corrections. States (or sites) with greater uncertainty (deviations) receive larger corrections, while those with smaller deviations receive smaller adjustments. This ensures a refined correction process tailored to the level of uncertainty observed.
## III Results and Discussion
In this section, we evaluate the effectiveness of PINNs and our proposed U-aware HC constraint in enforcing trace conservation during simulations. We compare their performance against state-of-the-art purely data-driven neural networks commonly used in open quantum system simulations. For comprehensive assessment, we utilize two distinct processes as benchmarks: relaxation dynamics within the spin-boson (SB) model and the energy transfer process (EET) within the 7-site FMO complex.
For the SB model, we acquire high-quality training data from the publicly available QD3SET-1 database. This comprehensive database provides pre-computed dynamics using the hierarchical equations of motion (HEOM) approach. The specific training dataset, denoted by \(\mathcal{D}_{\mathrm{sb}}\), consists of 1,000 trajectories simulated across a four-dimensional parameter space encompassing system-bath coupling strength, bath reorganization energy, bath relaxation rate, and inverse temperature (represented by \(\epsilon/\Delta\), \(\lambda/\Delta\), \(\gamma/\Delta\), and \(\beta\Delta\), respectively). " In similar manner, training data for 7-site FMO complex was also extracted from QD3SET-1 database. This dataset encompasses 1,000 training instances, capturing the dynamics for both possible initial excitation sites (site-1 and site-6) within the complex. In the considered data set, the dynamics is propagated for a range of simulation parameters chosen from a three-dimensional space \(\mathcal{D}_{\mathrm{fmo}}=(\lambda,\gamma,T)\). The method used for propagation is the trace conserving local thermalizing Lindblad master equation (LTLME) with the system Hamiltonian parameterized by Adolphs and Renger.
\[H_{\mathrm{S}}=\] \[\begin{pmatrix}200&-87.7&5.5&-5.9&6.7&-13.7&-9.9\\ -87.7&320&30.8&8.2&0.7&11.8&4.3\\ 5.5&30.8&0&-53.5&-2.2&-9.6&6.0\\ -5.9&8.2&-53.5&110&-70.7&-17.0&-63.6\\ 6.7&0.7&-2.2&-70.7&270&81.1&-1.3\\ -13.7&11.8&-9.6&-17.0&81.1&420&39.7\\ -9.9&4.3&6.0&-63.3&-1.3&39.7&230\end{pmatrix}, \tag{19}\]
For the training process, we adopted OSTL-based recursive dynamics propagation (Eq.) where the RDM\(\tilde{\rho}_{\rm s}(t)\) at each time step transforms into a 1D vector with dimension \(M\) = number of sites + (2 \(\times\) number of the upper off-diagonal terms). As in RDM \(\tilde{\rho}_{\rm S_{\rm am}}(t)=\tilde{\rho}_{\rm S_{\rm am}}^{\rm S}(t)\,(n\neq m)\), only the upper off-diagonal terms are learned. In addition, the real and imaginary parts of each off-diagonal term are separated. More details can be found in Ref.. The target is multi-time step dynamics which is in the same shape as the input. Here we predict the dynamics of 20 time-steps in one shot and which is then fed to the model recursively for the prediction of the next 20 time-steps dynamics. In all cases, we trained a CNN model, implemented in the MLQD package. and the uncertainty-aware HC constraint is integrated with priors set as \((p_{1},p_{2})=(0,0.1)\).
For the training process, we employed an OSTL-based recursive dynamics propagation approach (Eq.), where the reduced density matrix (RDM) \(\tilde{\rho}_{\rm s}^{(i)}(t)\) at each time step is transformed into a vector with dimension \(M\), representing the number of sites plus twice the number of upper off-diagonal terms. Since \(\tilde{\rho}_{\rm s_{\rm am}}(t)=\tilde{\rho}_{\rm s_{\rm am}}^{*}(t)\) for \(n\neq m\), only the upper off-diagonal terms are learned, with the real and imaginary parts treated separately. For further details, please refer to Ref.. The target is multi-time step dynamics, maintaining the same shape as the input. We predict the dynamics for 20 time steps in one shot, and these predictions are then recursively fed into the model to forecast the next 20 time steps. Additionally, in our calculations, the uncertainty-aware HC constraint was integrated with priors set at \((p_{1},p_{2})=(0,0.1)\), and the MLQD package was used for all computations.
To improve training efficiency, we utilized farthest point sampling to select a subset of training trajectories. For both the symmetric SB model (\(\epsilon/\Delta=0\)) and FMO complex with initial excitation on site-1, 400 trajectories were chosen for training, with the remaining used for testing.
In our study, we trained CNN models with identical architectures across all four scenarios. The models used for dynamics propagation yielded nearly identical validation losses, with approximately \(1.2\times 10^{-5}\) in the SB case and \(1.1\times 10^{-7}\) in the FMO complex. Introducing trace constraints and adding a prior do impact computational efficiency. Including a trace constraint in the loss function increases its complexity, and the addition of a prior makes the model more challenging to fit, potentially leading to longer training times.
For example, in our experiments, the unconstrained NN model for FMO complex reached a validation loss of \(1.01\times 10^{-7}\) at epoch 194. In contrast, the PINN model with the same architecture achieved a similar loss of \(1.62\times 10^{-7}\) at epoch 785, and the auxiliary model in the case of PINN with U-aware HC attained loss of \(2.22\times 10^{-7}\) at epoch 1142. On our machine (GeForce RTX Nvidia 4090 GPU), each epoch took approximately 1 second, resulting in total training times of 194 seconds, 785 seconds, and 1142 seconds, respectively. While the addition of constraints and priors increases computational time, the overall increase is not significant given the advanced computational resources available today.
We revisit the same cases as presented in for purely data-driven NNs. As expected, the PINNs (Figs. 2 A and C) shows a significant improvement in trace conservation compared to purely data-driven neural networks. However, as previously discussed, PINNs rely on "soft constraints" within the loss function, which can lead to minor deviations from perfect trace conservation.
Perfect trace conservation is achieved via the U-aware HC constraint, as demonstrated in Figs. 2 B and D. By explicitly incorporating this constraint within the PINNs framework, we maintain perfect trace conservation throughout the simulations for both the SB model and the FMO complex. This finding underscores the benefit of enforcing strict physical constraints by design, rather than solely relying on the model's ability to learn physical principles indirectly.
Additionally, we present the corresponding population dynamics for all four cases in Fig. S2 and Fig.
Trace conservation in NN-based simulations using PINNs and the uncertainty-aware HC constraint approach. This figure replicates (data-driven NN) for SB model and FMO complex, demonstrating improved conservation with PINNs (Panels A and C) and perfect trace conservation achieved by combining U-aware HC constraint with PINNs (Panels B and D). In the case of SB model, an initial period of \(t_{m}\Delta=2.0\) serves as a seed for the model’s predictions and results are presented for a test trajectory with characteristic frequency \(\gamma/\Delta=9.0\), system-bath coupling \(\lambda/\Delta=0.6\), and inverse temperature \(\beta\Delta=1.0\). For the FMO complex, the initial excitation is considered on site-1, with parameters \(\gamma=400\) cm\({}^{-1}\), \(\lambda=40\) cm\({}^{-1}\), and temperature \(T=90\) K.
To evaluate the accuracy of each model in dynamics propagation, we provide the MAE averaged over all time steps for each state (site) in Table 1. From the MAE comparison, we observe that all models have tiny errors for populations, so the trace conservation did not have much impact on the quality of the dynamics in the studied cases. However, the trace conservation might have a big impact in the cases where ML struggles to learn and predict dynamics with such an accuracy. As described above, the additional computational cost for enforcing the trace conservation is not that high either, which does not justify the use of the non-conserving approaches in case they break down and have even worse behavior than in In any case, using trace-conserving approaches can be considered as a good prophylactic against unphysical behavior.
.................................................................
## IV Concluding remarks
This work addresses the critical issue of trace conservation in NN-based simulations of open quantum systems. While NN models are adept at capturing complex dynamics, they often struggle to maintain fundamental physical principles such as trace conservation. Our investigation reveals three key findings.
First, purely data-driven NN models, including physics-agnostic and unconstrained NNs, can effectively capture correlations between state-specific populations. However, they lack explicit enforcement of physical laws, leading to potential violations of trace conservation.
Second, PINNs offer a significant improvement by incorporating physical knowledge into the loss function. This method penalizes deviations from physical constraints, enhancing the accuracy of simulations. Despite this advancement, PINNs still rely on "soft constraints," which can result in minor violations of physical constraints like trace conservation.
Finally, U-aware HC constraint approach addresses the limitations of PINNs by enforcing trace conservation by design rather than solely through the loss function. The U-aware HC constraint utilizes uncertainty quantification techniques to redistribute residual errors and correct potential trace violations, ensuring physically consistent simulations throughout.
It is important to note that while we did not explicitly enforce a positivity constraint in our case-since all diagonal elements remained strictly positive-such a constraint could be incorporated if necessary.
To conclude, our findings underscore the importance of integrating well-defined physical constraints into NN models. The methods developed in this study are broadly applicable and can be adapted to enforce other essential constraints in various domains. For instance, in molecular simulations where individual atomic charges are learned, our different-prior approach for uncertainty quantification as well as an approach for redistributing residual error in atomic charges could be used as an alternative to existing, related approaches for ensuring total charge conservation. By extending these techniques, we can improve the fidelity and reliability of NN-based simulations across a wide range of scientific and engineering applications.
|
10.48550/arXiv.2404.14021
|
Physics-Informed Neural Networks and Beyond: Enforcing Physical Constraints in Quantum Dissipative Dynamics
|
Arif Ullah, Yu Huang, Ming Yang, Pavlo O. Dral
| 5,866
|
10.48550_arXiv.1010.1093
|
## Abstract. The effect of optical radiation on the rate of aggregation of nanoscopic particles is studied in metal aerosols. It has been shown that under light exposure, polydisperse metal aerosols can aggregate up to two orders faster due to the size dependent photoelectron effect from nanoparticles. Different size nanoparticles undergo mutual heteropolar charging when exchanging photoelectrons through the interparticle medium to result in an increased rate of aggregation. It is shown that long-range electrostatic attractive forces drive the particles into closer distances where the short-range Van-der-Waals forces become dominating. Attention is drawn to the fact that this effect may occur in various types of dispersed systems as well as in natural heteroaerosols.
## 1.
Studying the ability of aerosols containing metallic or dielectric dispersed phases as well as heteroaerosols of natural disperse media to aggregate under light exposure is of great practical interest. Photostimulated coagulation of disperse systems widely occurs in nature and affects other processes. In particular, aggregation of superdispersed particles of prevailing atmospheric aerosols can cause significant changes in their optical extinction spectra. A correct computation of this effect enables one to assess the influence of absorption of solar radiation by dust particles in the air on the thermal balance in the atmosphere. Finding a solution to this problem may be related to problems of global climatology.
Understanding the mechanisms of photoaggregation is useful for fabrication and storage of nanomaterials with aggregative instability. It is also useful when dealing with different technological processes that can be accompanied by formation of polydisperse metal nanoparticles and by their undesirable photoaggregation, which happens for example in coating chambers, or in vacuum atomizer cells with metal vapors used for nonlinear optical frequency conversion of pulsed laser radiation.
Analysis of the influence of optical radiation on the rate of aggregation of heterogeneous aerosols in combination with the accompanying factors, such as external corpuscular fluxes under close to real conditions (dusty planet atmospheres, interstellar medium, industrial air pollutants), can reveal general mechanisms underlying the aggregation kinetics and evolution of a disperse system.
It has to be admitted that the experimental facts on the effect of light on coagulation of aerosols are not given due consideration and they are usually ignored because of the lack of theoretical models for their explanation.
In our paper we study the role of optical radiation in coagulation kinetics of metal aerosols and seek to answer the question: How can light accelerate the aggregation of this type of disperse systems? We propose one of the possible approches to explanation of photostimulated aggregation of metal aerosols. Similarly to photoaggregation in metal hydrosols, our model is based on the photoelectric effect. Under certain conditions, the photoelectric effect can stimulate mutual heteropolar charging of aerosol particles, provided the particles are polydisperse. Under these conditions, long-range electrostatic attractive forces between particles initiate their approach into the range of action of the short-range Van-der-Waals forces and cause the particles to coagulate. In addition, electrostatic interactions of the particles randomize their motion, which stimulates collisions.
Optical radiation can induce electrostatic interaction of electroconductive particles, subject to the following.
Firstly, the metal sol must be polydisperse. In real nanoparticle sols, there are always particles of various sizes. Characteristics of a particle, such as the Fermi energy, photoelectric work function, and quantum yield, are controlled by the size of the particle. Under electromagnetic irradiation, smaller particles of such an ensemble will emit more photoelectrons from unit surface area than larger particles due to higher photoelectron emission constants of smaller particles.
Secondly, the interparticle medium must be transparent for photoelectrons, i.e. this medium should not impede transfer of photoelectrons from one particle to another. Due to the exchange of photoelectrons, such a system comes to equilibrium once the particles of a smaller size become charged positively while the particles of a larger size acquire the negative charge (it was studied by Nagaev in and in his earlier papers cited therein. This is the basic idea of the mutual heteropolar charging effect. In the process of electron exchange the system tends to equalize Fermi energies in different particles. If the characteristics of particles were independent of their sizes, the total capture of photoelectrons and their emission would be controlled only by the surface of particles, and mutual charging would be impossible.
Thirdly, the interparticle medium (electrically neutral on the whole) must have a fairly low concentration of electric charges so as not to allow them to condense into electric double layers on the surface of the particle thereby compensating the metal kernel's own charge. The most favorable case would be when there are no charges in the interparticle medium other than the charges of photoemissive origin.
The aim of the present paper is to study the conditions for accelerated aggregation of aerosols with electroconductive nanoparticles under light exposure and to fill the gaps in the theoretical concepts as well as to prove a feasibility of the effect of mutual heteropolar charging in polydisperse ensembles of metal nanoparticles and to stimulate experimental studies of these effects in aerosols.
The paper is organized as follows. In Section 2, we describe the theoretical model, which includes the conditions for size-dependent photoelectron emission from nanoparticles, capture of photoelectrons and charging of particles, and also the basic equations of the molecular dynamics method used in studying the kinetics of aggregation under interparticle electrostatic and van-der-Waals interactions. In Section 3, we present the results of computation of the coagulation kinetics employing the model described in the previous Section and analyze the influence of different characteristics of the disperse system and applied radiation on the kinetics.
## 2.
## 2.1.
In this Section, we deal with conditions for photoelectron emission from nanoparticles.
\[W_{e}(r_{i})=W_{e}-\Delta W_{e}(r_{i})-\Delta W_{F}(r_{i})+|\,e\,|\,(Z_{i}+|\, e\,|)/(4\pi\varepsilon_{o}\varepsilon r_{i}), \tag{2}\]where \(W_{e}\) is the work function of the bulk, the terms \(\Delta W_{e}(r_{i})=(5/8)\cdot e^{2}/(4\pi\varepsilon_{0}\varepsilon r_{i})\) and \(|\,e\,|\,(Z_{i}+|\,e\,|)/(4\pi\varepsilon_{0}\varepsilon r_{i})\) are size dependent additions to the work function of a charged spherical particle, \(\varepsilon\) is the dielectric permeability of the interparticle medium (\(\varepsilon=1\) for gas media), \(\varepsilon_{0}\) is the permittivity of free space, \(e\) is the electron charge.
\[\Delta W_{F}(r_{i})=W_{F}(r_{i})-W_{\infty}\approx\pi W_{\infty}S_{i}/(4k_{F}V _{i}). \tag{3}\]
The importance of the size-dependent addition to the Fermi energy was emphasized in the paper.
The photoemission constant \(c(r_{i})\) for silver has been measured experimentally in in the UV spectral range (in the range of photon energies near threshold). A strong size dependence of \(c(r_{i})\) in the range \(2r_{i}=4-6\) nm and an increase in the constant by a factor of 100 (compared to the bulk) was reported in (see Figure 1).
The conventional values for the work function of \(Ag\) bulk (4.25-4.3 eV) were borrowed from. It should be emphasized that for metals, the Fowler law in its form is valid near the photoemission threshold of the photon energy only within the range \((h\omega-W_{e}(r_{i}))\leq(1.5+2)\) eV. Beyond this range, dependence can deviate from the square-law \((h\omega-W_{e}(r_{i}))^{2}\Rightarrow(h\omega-W_{e}(r_{i}))^{m}\); the index \(m\neq 2\).
In, the quantum yield \(Y_{w}(h\omega)\) dependence (in arbitrary units) was studied experimentally for nanoparticles in a wider spectral range (up to 10-11 eV) (see for a study on \(Ag\) nanoparticles with a 5 nm radius).
\[F(h\omega)=Y_{w}^{0}(h\omega)/Y^{0}(h\omega)\,.\]
The dependence of the photoemission constant on the particle radius (in the range 2–3 nm) for silver at 230 nm (5.4 eV) (Schmidt-Ott et al 1980) (for larger radii the curve is a result of approximation).
Here \(Y_{\nu}^{0}(h\omega)\) is the experimentally determined quantum yield (in relative units) in a wide range of photon energies covering the range of non-Fowler behavior
\(Y^{0}(h\omega)\) is the spectral dependence of the quantum yield (in relative units) as found from expression. In terms of the above definition, \(F(h\omega_{min})\sim\)1 in the spectral range of validity of the Fowler law and deviates from 1 at higher photon energies.
We have thus harmonized the experimental and calculated quantum yield data obtained in different papers by bringing them to a common system of relative units in a wide spectral range. Taking into consideration these results, square-law expression, which is valid in a narrow spectral range, can be rewritten with the same index \(m\)=2 as follows:
\[Y_{c}(h\omega,r_{i})=F(h\omega)c(r_{i})[h\omega-W_{e}(r_{i})]^{2}. \tag{4}\]
Expression in this form is valid in the range up to 10 eV.
Within the range of the particle radii 2-5 nm, we assume the spectral term \(F(h\omega)\) (in relative units) to be independent of the particle size because the \(Y(h\omega)\) curves for the bulk and for a nanoparticle are of the same shape. Expressions and similarly describe photoelectron emission at a small detuning of the photon energy from the red threshold, but at \((h\omega-W_{e}(r_{i}))\geq 3-3.5\) eV the effect may be significant, depending on both the type of metal and the surface contamination.
An example of a comparative diagram of the experimental and calculated absolute yields is shown in for the particle radius 5 nm.
Curve 3 for contaminated particles abundant in the experiment (see) was found to be the most realistic for further computations in our model. Similar curves can be calculated for other sizes of particles used in our study by taking into account the size dependent factors in Expression.
Spectral dependences of the absolute quantum yield for a silver nanoparticle with the radius \(r_{i}\)=5 nm: 1 – derived from expression; 2 – experimental dependence for clean particles, 3 – experimental dependence for contaminated nanoparticles (see comments in the text). The dependence \(Y_{c}(h\omega)\) calculated with expression taking into consideration the experimental data reproduces the curve.
## 2.2.
## 2.2.1.
To investigate the aggregation process, we apply the method of molecular dynamics. Initially, all particles are randomly spaced in a rectangular cell with the sides about \(10^{2}\)-\(10^{3}\) times the particle size. The walls of the cell can be assumed to be reflecting; that is considered as the boundary conditions. To simulate an infinite medium, we can change the boundary conditions to allow a particle irreversibly escape the cell while an identical particle with antiparallel momentum is simultaneously injected into the cell from the opposite side. As our computations showed, the use of both models provides similar results. Because of a simpler realization of the first model all computations in our work have been performed with the use of the reflecting wall cell.
Initial velocities are assigned to particles according to the Maxwell distribution. Next, mutual potential forces between the particles are introduced and discrete-time simulation of the Newtonian dynamics is performed (Section 2.3).
Emission of a given monovelocity \((V_{e})_{k}\) fraction of photoelectrons ceases when the charge of the \(j\)-th particle reaches \(Z_{j}\) and the photon energy is not sufficient to exceed the additional electrostatic energy of the particle after the emission of electrons with the total charge \(Z_{j}\).
Velocities of the emitted photoelectrons should meet the following requirement \((V_{e}(h\omega))_{k}=\sqrt{2[\hbar\omega-W_{e}(r_{j})]/m_{e}}\) for the \(k\)-th electron and monochromatic irradiation. Estimates show that electron velocities at typical electron concentrations never become maxwellian, - neither during the time-step, nor at the initial stage of aggregation. In such situations, the probabilistic method is the most suitable tool for description of emission and capture of photoelectrons by particles.
To describe emission of the \(k\)-th electron by the \(j\)-th particle during the time-step \(\Delta t\), a random process was realized. The requirement for emission from the \(j\)-th particle is \(f\leq(P_{em})_{j}\), where \(0<f<1\) is a random number chosen from the continuous uniform distribution,\((P_{em})_{j}\) is the probability of emission of an electron for a time step \(\Delta t\). If \(f\leq(P_{em})_{j}\), emission occurs.
\[(P_{em})_{j}=I_{0}\cdot Y_{e}\cdot(S_{p})_{j}\cdot\Delta t. \tag{5}\]
Restrictions on the parameters in this formula (primarily on the time-step) are set by the condition: \(0\leq(P_{em})_{j}\leq 1\). \(I_{0}\) is the number of photons incident onto unit area of the particle surface \((S_{p})_{j}\) during the time step. The emission dependence on the particle charge is included in the term \(Y_{e}(W(r_{j}))\). When an electron is emitted from a particle, one elementary negative charge (\(e\)) is subtracted from the current value the charge of \(j\)-th particle (\(Z_{j}\)).
As shown in our paper on the study of kinetics of aggregation during _half-coagulation time_ (\(t_{1/2}\)), single-particle and two-particle sub-aggregates prevail while the share of three-particle aggregates is negligible (according to the definition (see e.g.) \(t_{1/2}\) is the time of coagulation within which the number of particles in the ensemble including sub-aggregates becomes half the initial number).
Thus, consideration of the emission and electron capture by particles is complicated by the presence of a relatively large number of two-particle sub-aggregates and by ohmic contacts of particles in aggregates. In this case, we are to solve a separate complex problem of emission and capture of electrons by a pair of particles with the charge (\(Z_{i}\)+\(Z_{j}\)) and non-isotropic electric potential. It is obvious, that we have to apply a simplified approach. Estimates show that in the case of a pair of particles the potential of a two-particle sub-aggregate with the particle radii and \(r_{j}\) and charges (\(Z_{i}\)+\(Z_{j}\)) is close to the potential of an equivalent spherical particle with the radius \(r_{c}^{(i)}=\sqrt{r_{i}^{2}+r_{j}^{2}}\) and the same charge (\(Z_{i}\)+\(Z_{j}\)). After coagulation of \(i\)-th and \(j\)-th heteropolar particles the total charge (\(Z_{i}\)+\(Z_{j}\)) is close to minimum. According to, the difference in the potential of two monopolarly charged spheres and the potential of such an equivalent sphere with radius \(r_{c}^{(i)}\) and the same charge lies within 15%. This justifies our using a simplified method of calculation of emission and capture of electrons by two-particle sub-aggregates by way of substituting them with an equivalent spherical particle with the radius \(r_{c}^{(i)}\)and charge (\(Z_{i}\)+\(Z_{j}\)). This condition has been realized in our model.
Consideration of later stages of aggregation with a precise analysis of the potential of more complex multi-particle sub-aggregates in the calculation of emission and electron capture and, consequently, the effect of mutual charging on the aggregation is a separate problem and it is beyond the scope of our work.
#### 2.2.2 Electron capture
Emitted electrons are randomly distributed in the interparticle space of a cell and are captured by particles due to collisions. Capture of photoelectrons depends on the sign of the particle charge. The maximum number of photoelectrons in the interparticle medium in a monovelocity fraction, which can be adsorbed by a particle, is limited because particles with negative charges repulse electrons.
Capture of the \(k\)-th electron by the \(j\)-th particle during the time-step \(\Delta t\) is also realized as a random process by means of a random number (\(0<f<1\)) chosen from the continuous uniform distribution \(\rho(x)\).
\[P_{j}^{(k)}=P^{(k)}\left(x_{j-1}^{(k)}\leq f\leq x_{j}^{(k)}\right)=\int_{ \begin{subarray}{c}x_{j-1}^{(k)}\\ x_{j-1}^{(k)}\end{subarray}}^{x_{j-1}^{(k)}}\rho(x)dx=x_{j}^{(k)}-x_{j-1}^{(k)}\,.\]
In other words, the condition for capture of the \(k\)-th electron by the \(j\)-th particle during time-step \(\Delta t\) is
\[\sum_{l=1}^{l=j}P_{l}^{(k)}\geq f\geq\sum_{l=1}^{l=j-1}P_{l}^{(k)}\,, \tag{6}\]
If condition is satisfied, the act of capture occurs.
To check this condition, the probability of capture of the \(k\)-th electron by the \(j\)-th particle for the time step \(\Delta t\) is calculated from the expression
\[P_{j}^{(k)}=[\sigma_{j}^{(k)}((V_{e})_{k},\varphi_{j})\cdot(V_{e})_{k}\,/\, \upsilon_{c}\,]\Delta t\,. \tag{7}\]
\(\upsilon_{c}\) is the cell volume.
The cross-section of capture of electrons by the \(j\)-th particle is
\[\sigma_{j}^{(k)}=\begin{cases}\pi r_{j}^{2}\left(1+\dfrac{2e\varphi_{j}}{m_{e} (V_{e})_{k}^{2}}\right),&\dfrac{2e\varphi_{j}}{m_{e}(V_{e})_{k}^{2}}>-1\\ 0,&\dfrac{2e\varphi_{j}}{m_{e}(V_{e})_{k}^{2}}<-1\end{cases}\,,\]
Restrictions on the parameters in Formula are also imposed by the condition \(0\leq P_{j}^{(k)}\leq 1\).
If an electron is captured, one elementary negative charge (\(e\)) is added to the current charge of the \(j\)-th particle (\(Z_{j}\)). The exchange of electrons emitted by different size particles continues until the charges of the particles prevent the emission and capture of electrons.
### Kinetics of sol aggregation under mutual heteropolar charging
Basic equations of the molecular dynamics method used for studying aggregation kinetics under interparticle electrostatic interactions are described in. The resultant force acting on the \(i\)-th particle is subject to electrostatic influence of the remaining particles. This influence randomizes the motion of particles, giving rise to the appearance electrostatically linked particles with fast orbital rotation around the center of mass, which additionally intensifies the collision of particles. Coming into contact results in an irreversible coagulation of the particles. Irreversible coagulation of particles occurs when they come close to each other (approximately defined as \(0.1(r_{i}+r_{j})/2\)) within the range of action of van-der-Waals forces. This provides for the sticking probability of particles in our model to be equal to 1. Brownian dynamics models taking into consideration stochastic forces and short-range repulsion of particles with adsorption layers are described in, where we analyze coagulation of particles in viscous media in hydrosols with the sticking probability of particles less than 1. However the use of this model for aerosols is inappropriate because in aerosols the Langevin equations for ensembles of particles are transformed into simple Newtonian equations.
In the absence of collisions, the coordinates and velocity of the \(i\)-th particle change during one iteration step according to the equations:
\[\frac{\mathrm{d}\mathbf{r}_{i}}{\mathrm{d}t}=\mathbf{v}_{i},\,\mathrm{m}_{i} \,\frac{\mathrm{d}\mathbf{v}_{i}}{\mathrm{d}t}=\mathbf{F}_{i},\,\mathbf{F}_{i} =\mathbf{grad}(\mathrm{U}_{\mathrm{tot}})_{i}, \tag{8}\]
The one-step Runge-Kutta method is used to solve the equations of motion of the particles. The force is calculated at each time moment \(t\) from the total interparticle interaction potential for all the pairs of particles which include an \(i\)-th particle (for the given size distribution function of particles). The pair-interaction energy \(\mathrm{U}_{\mathrm{tot}}\) involves van-der-Waals and electrostatic interactions:
\[U_{\mathrm{v}}(r_{ij})=-\frac{A_{H}}{6}\Bigg{(}\frac{2i_{i}r_{j}}{r_{ij}^{2}- \left(r_{i}+r_{j}\right)^{2}}+\frac{2r_{i}r_{j}}{r_{ij}^{2}-\left(r_{i}-r_{j} \right)^{2}}+\ln\frac{r_{ij}^{2}-\left(r_{i}+r_{j}\right)^{2}}{r_{ij}^{2}- \left(r_{i}-r_{j}\right)^{2}}\Bigg{)}, \tag{9}\]
where, \(A_{H}\) is the Hamaker constant (\(A_{H}\approx 2\cdot 10^{-19}\) J for silver); \(r_{i}\) and \(r_{j}\), are the radii of particles, and \(r_{ij}\) is the interparticle distance;
\[U_{\mathrm{el}}(r_{ij})=\frac{Z_{i}Z_{j}}{4\pi\varepsilon_{0}\varepsilon F_{ij }}; \tag{10}\]
\(Z_{i},Z_{j}\) are the charges of interacting particles.
Distances between all pairs of particles are calculated during each iteration step. When the distance \(r_{ij}\) between two particles \(i\) and \(j\) becomes equal to or less than the sphere diameter (when particles collide), these two particles are considered to be rigidly attached to each other, and the sub-aggregates so formed continue moving as a rigid body for the rest of the aggregation process under the influence of forces acting on each individual particle. Note that the particle sticking condition is checked after an elementary move. This means that technically they can overlap (the distance between the particle centers may become smaller than the sum of radii), but the time step is chosen to be sufficiently small (and velocity dependent) so that the depth of such overlapping is small compared to the sum of radii, and when overlapping happens, the interparticle distance is slightly increased to the point of exact touching.
Elementary rotation of a sub-aggregate (with more than one particle) is taken into account with the use of the rotation matrices
\[\mathrm{R}_{\mathrm{x}}^{\mathrm{\alpha}}=\begin{pmatrix}1&0&0\\ 0&\cos\alpha&-\sin\alpha\\ 0&\sin\alpha&\cos\alpha\end{pmatrix};\ \ \mathrm{R}_{\mathrm{y}}^{\mathrm{\beta}}= \begin{pmatrix}\cos\beta&0&\sin\beta\\ 0&1&0\\ -\sin\beta&0&\cos\beta\end{pmatrix};\ \ \mathrm{R}_{\mathrm{z}}^{\mathrm{\gamma}}= \begin{pmatrix}\cos\gamma&-\sin\gamma&0\\ \sin\gamma&\cos\gamma&0\\ 0&0&1\end{pmatrix}. \tag{11}\]
These transformations are applied step by step to all particles in a given sub-aggregate. New coordinates resultant from the rotation are calculated as \(\mathbf{r}^{\prime}{}_{i}=\mathrm{R}_{\mathrm{x}}^{\mathrm{\alpha}}\mathrm{R} _{\mathrm{y}}^{\mathrm{\beta}}\mathrm{R}_{\mathrm{z}}^{\mathrm{\gamma}}\mathrm{r }_{i}\), where \(\alpha=\omega_{\mathrm{x}}\mathrm{dt}\), \(\beta=\omega_{\mathrm{y}}\mathrm{dt}\), \(\gamma=\omega_{\mathrm{z}}\mathrm{dt}\), and \(\mathbf{\omega}\) is the angular velocity of an aggregate, which satisfies \(\widetilde{\mathrm{\mathbf{I}}\mathrm{\omega}}=\mathbf{J}\), where \(\mathbf{J}\) is the angular moment and \(\widetilde{I}\) is the tensor of inertia of the aggregate being rotated. The angular velocity is calculated from the moment of inertia by inverting the tensor \(\widetilde{I}\), namely, \(\mathbf{\omega}=\widetilde{I}^{-1}\mathbf{J}\). Linear velocities of individual particles in an aggregate change as the result of rotation according to \(\mathbf{v}^{\prime}{}_{i}=\mathbf{v}_{\mathrm{c}}+\mathbf{\omega}\times\left( \mathbf{r}_{i}-\mathbf{r}_{\mathrm{c}}\right)\) where \(\mathbf{r}_{\mathrm{c}}\) is the center of mass of the aggregate. In turn, the angular momentum is updated according to \(\mathbf{J}=\mathbf{J}+\mathbf{M}\mathrm{\mathbf{d}t}\) where \(\mathbf{M}\) is the total torque acting on the aggregate being rotated. At the starting instance, \(\mathbf{M}=\)0. For each subsequent iteration step, \(\mathbf{M}\) is found from the known forces that act on each individual particle in the sub-aggregate: \(\mathbf{M}=\mathbf{M}+d\mathbf{M}\); \(d\mathbf{M}=\Sigma_{\mathrm{F}}\mathbf{{}_{i}}\times\left(\mathbf{r}_{i}- \mathbf{r}_{\mathrm{c}}\right)\).
As mentioned above, particles are distributed in a cubic cell with a given volume concentration. Simulating particle coagulation in real time (aggregation time is on the order of tens of seconds and longer) with the time step \(\Delta t=10^{-8}\) s and preserving all parameters of the system would require a significant amount of computation time.
To speed up computations, an accelerated sol aggregation under mutual charging was imitated. The real aerosols studied in the experiment had very low particle concentrations (on the order of \(10^{10}-10^{12}m^{-3}\)) and hence the rate of spontaneous aggregation was very slow. In order to achieve acceleration within shorter times (less than \(10^{-4}\) s), the concentration of the dispersed phase in the model has been increased to \(10^{15}\) cm\({}^{-3}\) (see validation of this method in Section 3.2).
Another problem of computation is posed by the photoelectron flux from the particles being rather small at low light intensities close to the experimental value (less than 1 W/cm\({}^{2}\)). Mutual charging of nanoparticles in a sol with a given concentration ceases when the rate of emission of photoelectrons and the inverse time of redistribution of photoelectrons among particles are below or comparable with the inverse half-time (\(t_{1/2}^{-1}\)) of particle coagulation.
These conditions impose the lower limit on the intensity of radiation that can initiate aggregation. Appearance of oppositely charged particles even in low-concentration aerosols essentially accelerates their aggregation in comparison with a spontaneous process. In computations, in order to compensate for insufficient photoelectron emission, the radiation intensity (or the density of the photon flux \(I\)) is increased by 3-4 orders of magnitude and more, which allows significant mutual charging and accelerated photoaggregation to be achieved within shorter computation times.
To speed up computations further, we used a time dependent dynamic step. The time step depends on the position of particles and is \(10^{-9}\) maximum, averaging at \(10^{-11}\) s. When particles are far from each other, the time step is maximal, as the particles approach each other (even if in just one pair) the step shortens.
## 3 Results and Discussion
In this Section we present the results of computation of the coagulation kinetics employing the model described in Section 2 and analyze the influence of different characteristics of the disperse system and radiation on the kinetics.
Maxwellian velocities of particles in random directions are found depending on the aerosol temperature. As mentioned above, a polydisperse ensemble comes to equilibrium because photoelectron emission depends on the particle size. Electrons are randomly redistributed among particles causing mutual heteropolar charging of the particles. As mentioned in Section 2.3, to verify this effect, we have calculated electric charge distribution histograms for 100 silver particles ranging in size from 4 to 6 nm with the Poisson distribution (see figure 3).
When the above described conditions are realized, we observe a significant exchange of electrons in a polydisperse ensemble of metal nanoparticles. Histograms of charges in Ag nanoparticles of a polydisperse ensemble are shown in for half-coagulation time.
The histogram of distribution of an electric charge on 100 \(Ag\) particles of a polydisperse ensemble for the Poisson distribution of particle sizes after UV irradiation. (a) corresponds to the energy of photons 6 eV and (b) corresponds to the photon energy 8 eV. The radiation intensity is \(I\)=10\({}^{6}\) Wm\({}^{2}\); the particle concentration is \(n\)=10\({}^{21}\) m\({}^{-3}\); half-coagulation time \(t_{\gamma_{2}}\) is approximately equal to 10\({}^{-5}\) s. The ratio of the maximal and minimal radii of particles for the Poisson size distribution function is \(r_{min}\)/\(r_{max}\)=2/3, \(r_{min}\)=2 nm.
Calculations of charge distribution histograms for different size particles (the value of the charge in particles of a given size) under mutual charging of the particles are made for both single particles and multi-particle sub-aggregates treated as equivalent spheres. As mentioned above, upon coagulation of heteropolyardy charged particles their charges undergo redistribution due to the ohmic contact that can be interpreted as the appearance of equivalent particles with the radius \(r_{c}^{(i)}=\sqrt{r_{i}^{2}+r_{j}^{2}}\) and total charge (\(Z_{i}\)+\(Z_{j}\)).
### Coagulation kinetics
The particles were irradiated by UV light with the photon energy 6 eV. The same histogram for the photon energy 8 eV is shown in The range larger than 6 nm corresponds to multi-particle sub-aggregates. These figures show that the mutual charging effect really takes place. A strong dispersion of the charge values is explained by a random nature of the process.
The photon energies for our simulations have been chosen arbitrary but with some justifications. The energy 6 eV corresponds to the spectral range of validity of the Fowler law \(\hbar\omega-W_{e}(r_{i})\)=1.7 eV and it is close to the photon energy in the experiments by. The value 8 eV is outside this range but within the range of transparency of solids and yet the experimental realization does not require any sophisticated means of registration of photostimulated aggregation of aerosols under such conditions.
The difference in 6 eV (a) and 8 eV (b) histograms lies in that in Case (b) there are particles with larger charges and a larger dispersion of particle charges is observed because following expression the higher the energy of photons, the larger the particle charge. But even under such conditions we register smaller particles to be charged positively and the larger ones to be charged negatively.
Predominance of the fraction of negatively charged particles is due to small particles emitting more electrons (see the size-dependence of the photoemission constant in Figure 1) and larger ones capturing more electrons.
In the process of aggregation, emission and capture of electrons are realized both for single particles and particles in sub-aggregates. Changes in the histograms are observed for large particles due to the capture of electrons by multi-particle sub-aggregates. Some of them have a charge close to zero.
Dependence of the rate of photostimulated aggregation of a polydisperse ensemble of 100 silver nanoparticles on the energy of photons under monochromatic irradiation. \(I\)=10\({}^{6}\) Wm\({}^{-2}\), the particle concentration is \(n\)=10\({}^{21}\) m\({}^{-3}\), \(r_{min}\)/\(r_{max}\)=2/3, \(r_{min}\)=2 nm. \(t_{1/2}^{\max}\)=4.5\({}^{\star}\)10\({}^{-5}\) s, \(t_{1/2}^{\min}\)=7.3\({}^{\star}\)10\({}^{-7}\) s.
The rate of photostimulated aggregation of a polydisperse ensemble of 100 silver nanoparticles with the radius 2-3 nm is shown in as a function of the photon energy of monochromatic UV radiation.
Saturation at high photon energies results from the electrostatic limitation on the maximum charge of particles in acts of emission and capture of electrons, which in turn limits the interparticle Coulomb attractive forces.
Hereinafter the acceleration factor can be found from the ratio \(t_{1/2}^{\max}/t_{1/2}^{\min}\) of maximal and minimal coagulation times given in the figure captions.
Dependence of the rate of photoaggregation of an ensemble of \(Ag\) nanoparticles on the the particle size distribution function is shown in for the bimodal and Poisson size distribution functions and 6 eV photon energy of monochromatic radiation complying with the condition \((r_{\max}+r_{\min})/2=3\).
The dependence exhibits a tendency to saturation due to saturation of the dependence of the photoemission constant on the particle size: the larger the particle, the smaller the photoemission constant.
Both dependences in look similar: the coagulation time is the longest at equal particle sizes and rapidly shortens with the growing difference in the particle sizes when the mutual charging effect is maximal (points with \(t_{1/2}^{\max}\) for cases (a) and (b) coincide).
This dependence demonstrates saturation at higher intensities at the given wavelength. Saturation at high intensities is explained by the electrostatic limitation on the maximal charge of particles.
Dependence of the rate of photostimulated aggregation of a polydisperse ensemble of 100 silver nanoparticles on the particle size distribution function: (a) Poisson distribution and (b) bimodal distribution. The photon energy is 6 eV. (\(I\)=\(10^{6}\) Wm\({}^{-2}\), \(n\)=\(10^{21}\) m\({}^{-3}\); the average radius in a pair \((r_{\max}\), \(r_{\min}\)) is kept constant at 3 nm; \(t_{1/2}^{\max}\)=\(4.5\,\mbox{\raisebox{-2.15pt}{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{ \mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox \mbox{ \mboxmbox{ }}}}}}}}}}}}}}}}}}}}\)\)\(10^{5}\) s, \(t_{1/2}^{\min}\)=\(6.7\,\mbox{\raisebox{-2.15pt}{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{ \mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{ \mboxmboxmbox{ \mboxmboxmboxmboxmbox{ \mboxmboxmboxmboxmbox { \mboxmboxmbox{ { \mboxmboxmbox { \mboxmbox { \mboxmbox { \mboxmboxmbox { \mboxmbox { \mboxmbox { \mbox { \mboxmbox { \mboxmbox { \mboxmbox { \mboxmbox { \mboxmbox { \mboxmbox { \mbox {\mboxmbox { \mboxmbox { \mboxmbox { \mboxmbox {\mboxmbox { \mboxmbox { \mboxmbox { \mboxmbox { \mboxmbox { \mboxmbox { \mboxmbox{ \mboxmboxmbox { \mboxmbox { \mboxmbox{ \mboxmboxmbox{ \mbox { \mboxmboxmbox { \mbox { \mboxmboxmbox { \mboxmbox { \mboxmbox{ \mboxmboxmbox { \mboxmbox { \mboxmbox { \mboxmbox{ \mboxmboxmbox { \mboxmboxmbox{ \mboxmbox { \mboxmbox { \mboxmbox { \mboxmbox { \mboxmboxmbox{ \mboxmbox { \mboxmbox{ \mboxmboxmbox { \mboxmboxmbox { \mboxmbox { \mboxmbox{ \mboxmboxmbox { \mboxmbox { \mboxmbox{ \mboxmboxmbox { \mboxmbox{ \mboxmbox { \mboxmboxmbox { \mboxmbox{ \mboxmboxmbox { \mboxmbox { \mboxmboxmbox { \mboxmbox{ \mboxmbox { \mboxmboxmbox{ \mboxmbox { \mboxmboxmbox{ \mboxmbox { \mboxmboxmbox { \mboxmbox{ \mboxmbox { \mboxmboxmbox { \mboxmbox{ \mboxmboxmbox{ \mboxmbox{ \mboxmboxmbox { \mboxmbox{ \mboxmbox {\mboxmboxmbox{ \mboxmbox {\mboxmboxmbox {\mboxmboxmbox{\mboxmbox {\mboxmboxmbox{\mboxmbox \mbox{\mboxmbox{\mbox \mbox{\mboxmbox{ \mboxmbox{\mbox \mbox{ \mbox{\mboxmboxmbox{ \mboxmboxmbox{ \mbox{\mboxmbox{ \mboxmbox{\mboxmbox \mbox{\mboxmbox{ \mbox{\mboxmboxmbox }}}}}}}}}}}\)\)\)\(\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{ \mbox{\mbox{\mbox{\mbox{{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox{\mbox \mboxmboxmbox{ \mboxmbox \mbox{ \mboxmbox { \mboxmbox { \mboxmbox { \mboxmboxmbox { \mboxmboxmbox { \mboxmboxmbox {\mboxmbox \mbox{ \mbox{\mbox \mbox{ \mbox \mbox{ \mbox{\mbox \mbox{ \mbox \mbox{ \mbox{ \mboxmbox \mbox{ \mbox{ \mbox \mbox{ \mbox{\mboxmbox \mbox{ \mbox{\mbox \mbox{ \mboxmbox{ \mboxmbox \mbox{\mbox{ \mboxmbox{\mbox \mbox{\mbox \mbox{\mbox{ \mbox \mbox{\mbox \mbox{\mboxmbox \mbox{\mbox \mbox{\mbox{\mbox \mbox{\mbox \mbox{\mbox \mbox{\mbox \mbox{\mbox{\mbox \mbox{\mbox \mbox{\mbox {\mbox \mbox{\mbox \mbox{ \mbox \mbox{\mbox \mbox{\mbox \mbox{\mbox \mbox{\mbox \mbox{\mbox \mbox{\mbox \mbox{\mbox \mbox \mbox{\mbox \mbox{\mbox \mbox{\mbox\mbox \mbox{\mbox\mbox \mbox{\mbox \mbox{\mbox\mbox \mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox \mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox\mbox{\mbox \mbox{\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox\mbox{\mbox \mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox \mbox\mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox \mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox \mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox\mbox\ \mbox{\mbox{\mbox\mbox\mbox{\mbox\mbox\mbox \mbox{\mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox \mbox{\mbox\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox \mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox\ \mbox{\mbox{\mbox\mbox\mbox{\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\ \mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox\ \ \ \mbox{\mbox{\mbox\mbox{\mbox\mbox\mbox\ \ \mbox{\mbox\mbox{\mbox\mbox{\mbox\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \ \mbox{\mbox{\mbox\mbox\mbox\ \mbox{\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\ \mbox\ \mbox{\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox\ \ \mbox{\mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \ \mbox{\mbox{\mbox\mbox\mbox\ \mbox{\mbox\mbox{\mbox\ \mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\ \mbox\mbox{\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox{\mbox\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox\mbox{\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox\mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\ \mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox{\mbox\mbox{\mbox\mbox\mbox{\mbox\mbox{\mbox\mbox\mbox\
At too high radiation intensities, some electrons are not captured by particles and remain in the interparticle medium. Curve (a) corresponds to all electrons having been captured, (b) and (c) to 50% and 20% captured electrons, respectively. As we can see from these dependences, the lower the percentage of captured electrons, the slower the aggregation processes. This is due to a larger fraction of positively charged particles.
### Limitations of the model
Kinetics of photostimulated aggregation depending on the particle concentration is shown in
Dependence of the half-coagulation time of a polydisperse ensemble of silver nanoparticles irradiated by monochromatic light on the radiation intensity for different fractions of captured electrons (a – 1, b – 0.5, and c – 0.2). The photon energy is 6 eV, the particle concentration \(n\)=\(10^{21}\) m\({}^{-3}\), the minimal radius of a particle is 2 nm, and the particle size distribution function is Poissonian (\(r_{min}/r_{max}\)=\(2/3\), \(r_{min}\)= 2 nm). \(t_{1/2}^{\max}\) =\(4.0\,\mathbf{\cdot}\,10^{-5}\,\mbox{s}\), \(t_{1/2}^{\min}\) =\(9.5\,\mathbf{\cdot}\,10^{-7}\,\mbox{s}\).
Dependence of the half-coagulation time of a polydisperse ensemble of silver nanoparticles on the particle concentration. The particles are irradiated by monochromatic light with the photon energy 6eV, \(I\)=\(10^{6}\) Wm\({}^{-2}\). \(r_{min}/r_{max}\)=\(2/3\), \(r_{min}\)= 2 nm, \(t_{1/2}^{\max}\)=\(1.5\,\mathbf{\cdot}\,10^{-6}\) s, \(t_{1/2}^{\min}\) =\(2\,\mathbf{\cdot}\,10^{-7}\,\mbox{s}\).
The revealed tendency agrees with the theoretical models derived from the Smoluchowski theory. In concordance within this theory, the dependence of the sol coagulation rate on the particle concentration (\(t_{1/2}\)\(\sim\)\(n^{-1}\)) (see, e.g.,) valid for arbitrary particle concentration in random systems is given by the equation
\[t_{1/2}=\frac{6\eta r}{4\rho nk_{B}T}\,. \tag{12}\]
In the case of a rarefied interparticle gas medium and nanoscopic particles an approximate estimate can be provided by the expression
\[t_{1/2}\approx\frac{6\eta r}{4\rho nk_{B}T\bigg{(}1+A\frac{\lambda_{0}}{r} \bigg{)}}\,, \tag{13}\]
For the van-der-Waals forces responsible for spontaneous aggregation the ratio \((\rho/r)_{v}\)\(\approx\)2.3.
\[\frac{Z_{i}Z_{j}}{4\pi\varepsilon_{0}\varepsilon\rho}=(3/2)k_{B}T\,. \tag{14}\]
Here \(Z_{i}\) and \(Z_{j}\) are charges of particles. So for the particle charges \(Z_{i}\) and \(Z_{j}\)being \(\pm\) 1e, \(\pm\) 2e, and \(\pm\) 3e, we have the ratio \((\rho/r)_{a}\)\(\approx\) 38, 152 and 342. In this case, a rough estimate of the acceleration factor of aggregation in comparison with a spontaneous process for the same particle radii \(r\), with the unit sticking probability of particles can be found as the ratio \([(\rho/r)_{a}]/[(\rho/r)_{v}]\) or \([(t_{1/2})_{a}]/[(t_{1/2})_{v}]\). For charges \(\pm\) 1e, \(\pm\) 2e, \(\pm\) 3e it yields an acceleration factor of 16, 66, and 148 that is close to the computation results obtained by the molecular dynamics method.
According to expressions and, the acceleration factor is independent of the particle concentration. This means our method can be applied to investigate systems with much lower particle concentrations typical of experimental conditions (10\({}^{10}\)-10\({}^{12}\) m-3). Note that in vacuum interparticle media, randomization of charged particles is caused by their electrostatic interactions.
The above results have been obtained for the case when the mean free path of emitted photoelectrons (\(I_{e}\)) exceeds both the mean interparticle distances (\(I_{p}\)) and the cell size. In this case, photoelectrons fill the cell as a continuous homogenous background. Obviously, if an electron reaches a neighboring particle within the time between collisions of particles, the effect of mutual charging may occur. If the mean free path decreases, the effect is less manifest.
In real sols, a dense interparticle medium reduces the mean free path of photoelectrons because of scattering on molecules. For example, in water the mean free path of photoelectrons with energy 4 eV does not exceed 5 nm.
In simulating this process, one of the simplest ways to describe the slowdown of electron propagation in the interparticle medium is to introduce the factor (V\({}_{m}\)) into the real photoelectron velocity \(V_{e}\) and to substitute \(V_{e}\) by V\({}_{m}\)\(\times\)\(V_{e}\) for emitted photoelectrons.
Dependence of the half-coagulation time on the slowing factor of photoelectrons. The particle concentration is 10\({}^{21}\) m\({}^{3}\), \(t_{\%}\) = 4.6\({}^{\bullet}\)10\({}^{\circ}\)5 s for the Poisson size distribution function, \(r_{min}\)\(r_{max}\)=2/3, \(r_{min}\).2 nm. \(I\)=10\({}^{6}\) Wm\({}^{\text{-2}}\) and the photon energy is 6 eV. \(t_{1/2}^{\text{max}}\)=2.5\({}^{\bullet}\)10\({}^{\text{-5}}\) s, \(t_{1/2}^{\text{min}}\)=1.6\({}^{\bullet}\)10\({}^{\text{-6}}\) s.
It can be seen that the half-coagulation time increases with decreasing V\({}_{\text{m}}\). This is due to recapture of the emitted electron by the same particle. The cross-section of the electron capture by the \(j\)-th particle is \(\sigma_{j}^{(i)}=\pi r_{j}^{2}\bigg{(}1+\dfrac{2e\varphi_{j}}{m_{e}(V_{e})_{i} ^{2}}\bigg{)}\). Following this expression, \(\sigma_{j}^{(i)}\)increases as \(V_{e}\) reduces (if \(\varphi_{j}>0\)). Irradiated particles tend to minimize their charges. In the limit it corresponds to the rate of spontaneous aggregation.
At high radiation intensities some fraction of electrons is not captured by particles and remains in the interparticle medium, which is another contributing factor to the slowing of the aggregation process.
Dependence of the half-coagulation time on the fraction (p\({}_{f}\)) of photoelectrons captured by particles. The ratio of radii of Ag particles \(r_{min}\)\(r_{max}\)=2/3, the particle concentration is 10\({}^{21}\) m\({}^{3}\), \(t_{\%}\) = 4.6\({}^{\bullet}\)10\({}^{\text{-5}}\) s for the Poisson size distribution function, \(r_{min}\)\(r_{max}\)=2/3, \(r_{min}\)=2 nm., \(I\)=10\({}^{6}\) Wm\({}^{\text{-2}}\) and photon energy 6 eV. \(t_{1/2}^{\text{max}}\rightarrow\infty\), \(t_{1/2}^{\text{min}}\)=10\({}^{\text{-6}}\) s.
Computations show that as the number of captured electrons reduces, the half-coagulation time indefinitely grows because of domination of particles with the positive charge.
## Conclusion
We can conclude that the proposed model can be applied for studying light accelerated coagulation of real disperse systems with various concentrations of particles as well as rarefied gaseous or vacuum interparticle media. We hope this model could provide a useful tool to predict the effect of heteropolar mutual charging of nanoparticles due to the size dependent photoelectric effect in polydisperse metal sols and what is even more important in heterosols with an interparticle medium transparent for photoelectrons. It has been shown that under considered conditions the mutual charging effect can accelerate photoaggregation of Ag aerosols up to 60 times and more in comparison with the rate of their spontaneous aggregation without light.
An accelerated aggregation of silver aerosols with nanoparticles of various sizes was observed in the experiments in a \(N_{2}\) flow. The nanoparticles were produced by electrode sputtering in a spark discharge between \(Ag\) electrodes accompanied by generation of plasma. This experimentally established fact of an anomalously rapid aggregation of metal aerosols remains unexplained so far. The paper has been the first to draw attention to a possibility of explaining the results of by mutual charging of the particles. In our opinion, one of possible explanations for the accelerated aggregation of aerosols in that paper should be sought in the experimental conditions. In that experiment, the concentrations of particles with different sizes were determined by the method of photoelectron emission from the obtained nanoparticles under UV irradiation. Note also that the aerosol used in the experiment was not strictly monodisperse. The contribution of these two factors to the aggregation kinetics can provide mutual charging of the particles.
Optimal conditions for manifestation of the mutual charging effect primarily involve enrichment of the radiation spectrum by high-energy photons (over 10 eV), which eases restrictions on the values of particle charges, enhances the Coulomb interparticle attraction forces and further accelerate aggregation. This concerns observation of the effect in aerosols under natural conditions, under which the photoemission characteristics of particles may be associated not only with the size dependence, but also with the chemical composition of particles. At high intensities of pulsed laser fields, an accelerated aggregation of aerosols may be induced by multi-photon electron photoemission. Another important factor is the presence of particles in metal aerosols with a diameter smaller than 10 nm because the photoemission constant of such particles increases rapidly with the decreasing particle size.
Our paper demonstrates how fine quantum size effects in metal nanoparticles can influence the aggregation kinetics, and we would like to draw attention to the possibility of manifestation of this effect.
|
10.48550/arXiv.1010.1093
|
Photostimulated Aggregation of Metal Aerosols
|
Sergei V. Karpov, Ivan L. Isaev
| 1,379
|
10.48550_arXiv.1603.03285
|
#### ii.1.1 Gaussian charge distribution models
For **Model 1**, the nonbonded interaction \(V_{M1}\) consists of a spherical Gaussian function for the electrostatic interaction and an LJ term representing dispersion and Pauli repulsion.
\[V_{M1}(\vec{r})=\sum_{nonbonded}\Big{[}q_{i}q_{j}\int\frac{|\psi_{i}|^{2}|\psi _{j}|^{2}}{R_{ij}}dr+\epsilon(\frac{\sigma_{j}^{12}}{R_{ij}^{12}}-2\frac{ \sigma_{ij}^{6}}{R_{ij}^{6}})\Big{]}, \tag{3}\]
The integrals over the Gaussian distributions required for the electrostatic interactions between two atoms \(i\) and \(j\) can be evaluated as a function of the distance \(R_{ij}\) using the error functions since
\[\int\frac{|\psi_{i}|^{2}|\psi_{j}|^{2}}{R_{ij}}dr=\frac{1}{R_{ij}}Erf\frac{ \sqrt{2}R_{ij}}{\sqrt{\omega_{i}^{2}+\omega_{j}^{2}}}. \tag{4}\]
In analogy, the interaction between a point charge and a Gaussian distribution can be expressed as \[\int\frac{|\psi_{j}|^{2}}{R_{ij}}dr=\frac{1}{R_{ij}}Erf\frac{\sqrt{2}R_{ij}}{\omega_ {j}}, \tag{5}\]
Based on these expressions, the integrals required for the calculation of the interaction potential can be implemented in a computationally inexpensive way using numerically efficient methods to evaluate the error function, as used in the eFF implementation.
For **Model 2**, the nonbonded interaction of the atoms \(V_{M2\_at}\) has the same terms as \(V_{M1}\), whereas for the electron potential sites \(V_{M2\_elp}\), the LJ potential is replaced by a repulsive potential for simplicity in order to reduce the interaction to the minimal requirements for an electron potential.
\[V_{M2\_at}(\vec{r})=\sum_{nonbonded}\Big{[}q_{i}q_{j}\int\frac{|\psi_{i}|^{2}| \psi_{j}|^{2}}{R_{ij}}dr+\varepsilon(\frac{\sigma_{j}^{12}}{R_{ij}^{12}}-2 \frac{\sigma_{ij}^{6}}{R_{ij}^{6}})\Big{]}, \tag{6}\]
\[V_{M2\_elp}(\vec{r})=\sum_{nonbonded}\Big{[}q_{i}q_{j}\int\frac{|\psi_{i}|^{2 }|\psi_{j}|^{2}}{R_{ij}}dr+\varepsilon(\frac{\sigma_{ij}}{R_{ij}})^{6}\Big{]}, \tag{7}\]
For **Model 3**, the nonbonded interaction of the atoms \(V_{M3\_at}\) is described by point charges instead of Gaussian distributions.
\[V_{M3\_at}(\vec{r})=\sum_{at\_at}\Big{[}\frac{q_{i}q_{j}}{R_{ij}}+\varepsilon (\frac{\sigma_{j}^{12}}{R_{ij}^{12}}-2\frac{\sigma_{ij}^{6}}{R_{ij}^{6}})\Big{]}, \tag{8}\]
For the interaction between electron potential sites \(V_{M3\_elp\_elp}\), the potential function is identical to \(V_{M2\_elp}\) (Equation 7). For the interaction between atom cores and electron potentials \(V_{M3\_at\_elp}\), interactions between Gaussian distributions and point charges have to be evaluated.
\[V_{M3\_at\_elp}(\vec{r})=\sum_{at\_elec}\Big{[}q_{i}q_{j}\int\frac{|\psi_{j}|^ {2}}{R_{ij}}dr+\varepsilon(\frac{\sigma_{ij}}{R_{ij}})^{6}\Big{]}, \tag{9}\]
Point charge models
For each of the three interaction models, a point charge model is evaluated in addition.
\[V_{elec}=q_{i}q_{j}\int\frac{|\psi_{i}|^{2}|\psi_{j}|^{2}}{R_{ij}}dr \tag{10}\]
are replaced on all interactions sites by
\[V_{elec}=\frac{q_{i}q_{j}}{R_{ij}} \tag{11}\]
In the case of Model 1, the interaction potential is then identical to conventional force field functions.
### Parametrization of interaction models
As explained above, the number of adjustable parameters is identical for all three models and there is only one adjustable parameters for the electrostatic potential per molecule. This is achieved as follows: The positions of the additional interaction sites are determined based on the molecular structure and the Slater radii \(S^{24}\) of the atoms. Bond electron potentials are simply placed at the geometric center of each bond. The position of the lone pair potential of atom \(i\) is determined by constraining it to a generic distance of \(\frac{3}{4}S_{i}\) from the atom and minimizing the interaction energy between all electron potentials in the molecule given in Equation 7. For the electrostatics the parameter space is reduced by either considering charge transfer between atoms due to electronegativity differences or by using the valence electron structure of the molecule. For the \(\sigma\) and \(\omega\) parameters, additional interaction sites in principle require additional parameters. For \(\omega\) this is avoided here by using only one adjustable parameters for all atoms and electron potential sites based on initial parameter values for atom radii and generic values for electron potentials. For \(\sigma\) additional parametrization of the electron potential sites is avoided by using a generic value of \(\sigma_{elp}=1.0\) A on all electron potential sites.
The parametrization is carried out separately for the electrostatic potential surface (EPS) and intermolecular interactions since the optimal parameters are not identical in both cases. This is on one hand due to polarization effects, on the other hand due to error compensation. For the EPS partial charges need to be determined for all models and the width of the Gaussian distribution \(\omega\) needs to be determined for the Gaussian charge distribution models. For the intermolecular interactions the width of the \(\sigma\) and \(\epsilon\) parameters of the LJ potential and the repulsive potential for the electron potential sites needs to be determined in addition. The \(\omega\) and \(\sigma\) parameters are in principle related as both of them describe the width of the electron density distribution. However, since they are parametrized differently and used in different evaluations two different symbols are used for clarity.
#### iii.1.1 Charges on atoms and electron potentials
For **Models 1 and 2** partial charges are determined based on Pauling electronegativities\(P\). For each pair of atoms \(i\) and \(j\) in a bond, an electronegativity difference coefficient \(\delta_{bond}=1.0-\frac{P_{i}}{P_{i}+P_{j}}\) is determined.
\[q_{i}=\zeta\sum_{n=1}^{Nb}{(0.5-\delta_{n})Zb_{n}}, \tag{12}\]
For Model 1 these charges are used directly, whereas for Model 2 the atom charge population is distributed equally among all sites assigned to the atom. Lone pair electron potential sites are assigned to their corresponding atom, whereas bond electron potential sites are assigned to the atom in the bond with the smaller electronegativity. In the case of equal electronegativity the sites are assigned to both atoms equally.
## Models 3
The initial bond charges \(z_{i}\) for this model are given by the bond order, i.e. \(z_{i}=-2.0\) for single bonds, \(z_{i}=-4.0\) for double bonds and \(z_{i}=-6.0\) for triple bonds. In the case of delocalized bonds the charge is distributed equally amongst all sites sharing it, e.g. for all benzene C-C bonds \(z_{i}=-3.0\). For lone pair potentials \(z_{i}=-2.0\). For each atom core the basic charge \(z_{i}\) is determined as the charge of the nucleus minus the number of valence electrons. The initial charge \(z_{i}\) is however only the correct atom core charge in cases where no charge transfer takes place between atoms and electrons are localized entirely on the on the electron potential sites.
\[q_{i}=\zeta z_{i}-\sum_{n=1}^{Nb}(0.5-\delta_{n})Zb_{n}, \tag{13}\]
In contrast to Models 1 and 2, the charge transfer between atoms is not affected by the \(\zeta\)-parameter in Model 3. The effect of a small \(\zeta\)-value is equivalent to having parts of the valence electron density localized on the atom core, i.e. the charge separation is smaller.
#### ii.1.2 Width of Gaussian distributions and LJ potential parameters
## Width \(\omega\) of Gaussian distributions:
As initial value for \(\omega\) the Slater radius \(S^{24}\) is used for the atoms and a generic initial value of \(\omega\)=1.0 A for all electron potential sites. Based on these initial values, the final width \(\omega_{i}\) is determined using again a single scaling parameter \(\upsilon\) per molecule.
## Radius \(\sigma\) and \(\varepsilon\) for LJ and repulsive potentials:
On the atoms, empirical values derived from X-ray diffraction data are used for \(\sigma\) except in the case of hydrogen where this value is generally too large and therefore replaced by the Slater radius of \(\sigma\)=0.25 A. For \(\varepsilon\) a single adjustable parameter is again used which is determined by fitting to the intermolecular interaction energies.
#### ii.1.3 Adjustable parameter fitting
As the number of adjustable parameters is kept small, fitting parameters to either the EPS or intermolecular interaction energies is simple. For fitting the point charge models to the EPS there is only one parameter \(\zeta_{m}\) to be determined for each molecule \(m\). This is done by calculating the average potential energy difference \(\bar{\Delta}_{Epot}\) between the _ab initio_ EPS outside the Slater radii of each atom on a 3-dimensional grid of 20 A side length centered about the geometric center of the molecule, with a distance of 1.0 A between grid points. The molecular parameter \(\zeta_{m}\) is determined as \(\zeta_{m}=ArgMin(\bar{\Delta}_{Epot}(\zeta))\) in the range \(\zeta=[0.0,4.0]\).
For fitting the Gaussian charge distribution models to the EPS two parameters, \(\zeta_{m}\) and \(\upsilon_{m}\) need to be determined. This is done in two steps: first \(\upsilon_{m}=ArgMin(\bar{\Delta}_{Epot}(\upsilon)|\zeta_{m})\) is determined in the range \(\upsilon=[0.0,8.0]\) for \(\zeta_{m}=\{\zeta_{mPC},\frac{5}{4}\zeta_{mPC}\}\), with \(\zeta_{mPC}\) being the \(\zeta\)-parameter determined for the point charge model. In the second step \(\zeta_{m}\) is determined as \(\zeta_{m}=ArgMin(\bar{\Delta}_{Epot}(\zeta)|\upsilon_{m})\) in the range \(\zeta=[0.0,4.0]\). The parameters for fitting all models to the EPS are given in in Table 1.
\begin{table}
\begin{tabular}{c||c|c|c||c|c|c|c|c} & \multicolumn{4}{c||}{point charge models} & \multicolumn{4}{c}{Gaussian distribution models} \\ \hline & Model 1 & Model 2 & Model 3 & Model 1 & Model 2 & Model 3 \\ \hline & \(\zeta\) & \(\zeta\) & \(\zeta\) & \(\zeta\) & \(\upsilon\) & \(\zeta\) & \(\upsilon\) & \(\zeta\) & \(\upsilon\) \\ \hline \hline H\({}_{2}\)O & 1.7 & 1.6 & 0.15 & 1.7 & 0.9 & 1.6 & 1.0 & 0.15 & 1.2 \\ \hline NH\({}_{3}\) & 1.9 & 1.3 & 0.15 & 1.9 & 1.0 & 1.3 & 1.1 & 0.15 & 1.4 \\ \hline CH\({}_{4}\) & 1.9 & 3.4 & 0.1 & 2.1 & 1.0 & 3.7 & 0.9 & 0.15 & 1.6 \\ \hline H\({}_{2}\)S & 1.1 & 1.0 & 0.01 & 1.1 & 1.6 & 1.0 & 2.1 & 0.015 & 3.5 \\ \hline PH\({}_{3}\)\({}^{a}\) & 0.9 & 0.8 & 0.0 & 1.0 & 0.9 & 0.8 & 2.2 & 0.01 & 4.4 \\ \hline CO\({}_{2}\) & 2.2 & 2.1 & 0.1 & 2.2 & 0.7 & 2.1 & 0.9 & 0.15 & 0.6 \\ \hline ethanol & 0.9 & 0.9 & 0.05 & 0.9 & 1.1 & 0.9 & 1.1 & 0.05 & 1.8 \\ \hline CH\({}_{5}\)N & 1.2 & 0.9 & 0.15 & 1.2 & 1.4 & 0.9 & 1.6 & 0.15 & 1.4 \\ \hline OCH\({}_{2}\) & 1.7 & 1.7 & 0.25 & 1.7 & 0.7 & 1.7 & 0.9 & 0.25 & 1.2 \\ \hline benzene & 1.7 & 2.1 & 0.1 & 1.7 & 1.3 & 2.2 & 1.5 & 0.1 & 1.7 \\ \hline pyrrole & 0.0 & 0.0 & 0.3 & 0.0 & 0.0 & 0.0 & 0.0 & 0.3 & 1.3 \\ \hline thiophene & 1.4 & 0.9 & 0.15 & 1.4 & 1.2 & 0.9 & 0.9 & 0.15 & 1.7 \\ \hline glycine & 1.4 & 0.5 & 0.0 & 1.5 & 1.4 & 0.5 & 0.9 & 0.01 & 3.6 \\ \hline \end{tabular}
\end{table}
Table 1: Molecular parameters fitted to the EPS for all three interaction models in their point charge and Gaussian charge distribution version. Parameter values of 0.0 indicate that \(\bar{\Delta}_{Epot}\) is smallest if the corresponding energy term is omitted. \({}^{a}\) For PH\({}_{3}\) Mulliken electronegativities are used instead of Pauling electronegativities due to the small Pauling electronegativity difference between the atoms.
For fitting the point charge models to intermolecular interaction energies there are also two parameters to determine, \(\zeta_{m}\) and \(\epsilon_{m}\). This is done by calculating the average potential energy difference \(\tilde{\Delta}_{E\_inter}\) between a set of 26 _ab initio_ intermolecular interaction energies and the interaction energy calculated from each model. (For details to the _ab initio_ calculations see next section). In the first step \(\epsilon_{m}\) is determined as \(\epsilon_{m}=ArgMin(\tilde{\Delta}_{E\_inter}(\epsilon)|\zeta_{m})\) in the range \(\epsilon=[0.0,2.0]\) for \(\zeta_{m}=\{0.1,1.0\}\). In the second step \(\zeta_{m}=ArgMin(\tilde{\Delta}_{E\_inter}(\zeta)|\epsilon_{m})\) is determined in the range \(\zeta=[0.0,4.0]\). The parameters for fitting all models to interaction energies are given in in Table 2.
### Electronic structure calculations and distributed multipole moments
## EPS calculations:
Electrostatic potential surfaces have been calculated for all test molecule using ORCA. As a reference for parameter fitting, density functional theory was used with the B3LYP functional and an aug-cc-pVTZ basis set.. The EPS was evaluated outside the Slater radii of the molecule on a 3-dimensional grid of 20 A side length centered about the geometric center of the molecule, with a distance of 1.0 A between grid points. For the evaluations of the accuracy, a reference EPS was calculated at the MP2/aug-cc-pVTZ level of theory.
\begin{table}
\begin{tabular}{c||c|c|c|c|c|c} & Model 1 & Model 2 & Model 3 \\ \hline & \(\zeta\) & \(\epsilon\) & \(\zeta\) & \(\epsilon\) & \(\zeta\) & \(\epsilon\) \\ \hline \hline H\({}_{2}\)O & 1.75 & 0.51 & 1.0 & 0.13 & 0.0085 & 0.09 \\ \hline NH\({}_{3}\) & 1.0 & 0.13 & 1.1 & 0.31 & 0.009 & 0.29 \\ \hline CH\({}_{4}\) & 0.0 & 0.57 & 0.1 & 0.3 & 0.0 & 0.66 \\ \hline H\({}_{2}\)S & 0.0 & 0.71 & 0.7 & 0.16 & 0.0045 & 0.16 \\ \hline PH\({}_{3}\)\({}^{a}\) & 1.2 & 1.23 & 0.45 & 1.5 & 0.0 & 1.38 \\ \hline CO\({}_{2}\) & 1.55 & 0.16 & 1.55 & 0.03 & 0.0045 & 0.03 \\ \hline CH\({}_{5}\)N & 0.1 & 0.83 & 0.1 & 0.85 & 0.0005 & 0.55 \\ \hline OCH\({}_{2}\) & 1.2 & 0.71 & 1.0 & 0.12 & 0.0006 & 0.0 \\ \hline \end{tabular}
\end{table}
Table 2: Molecular parameters fitted to the intermolecular interaction energies for all three interaction models. Parameter values of 0.0 indicate that \(\tilde{\Delta}_{E\_inter}\) is smallest if the corresponding energy term is omitted.
This choice of gridpoints is due to the fact that the evaluation is carried out for point segments within selected distance ranges of any atom as explained in the Section 'Results'. For comparison Mulliken charges and distributed multipole moments were calculated based on B3LYP/aug-cc-pVTZ calculations carried out with GAUSSIAN. Distributed multipole moments were obtained based on the electron density distributions using GDMA.
#### 3.2.2 Intermolecular interaction energies:
Intermolecular interaction energies were calculated at the MP2/aug-cc-pVTZ level of theory for 26 dimer geometries of the eight molecules selected for the evaluation using ORCA. Counterpoise corrections were used to account for the basis set superposition error. For each of the selected molecules, the dimer geometry was first optimized. Starting from the optimized geometry, 25 new dimer geometries were generated by random translation and rotation of the two monomers. In order to avoid very unfavorable geometries which are not of interest for the evaluation as they are unlikely to be observed, geometries with atom distances \(<1.0\) A were rejected.
## 4 Results
### Comparison of electrostatic potential surfaces
As a first step for all evaluations, the coordinates of the electron potential sites need to be determined. The coordinates of the bond electron potentials are calculated based on the atom coordinates, while lone pair electron potential coordinates are obtained by minimizing the interaction energy of all electron potentials in the molecule using the nonbonded interaction potentials given in Section 2.1.
The differences between the EPS of the three models, _ab initio_ calculations, distributed multipoles and Mulliken charges are first illustrated for three molecules, NH\({}_{3}\), CH\({}_{4}\) and formaldehyde (OCH\({}_{2}\)). The EPS of all models is calculated on a two-dimensional grid and is shown in Figures 3, 4 and 5. Comparison of the potential energy surfaces shows that that the EPS of Model 3 is qualitatively the most similar to the _ab initio_ EPS calculated at the MP2/aug-cc-pVTZ level of theory. The largest differences to _ab initio_ are observed for Mulliken charges. Models 1 and 2 have similar potential surfaces, with Model 2 appearing as slightly better than Model 1 overall. For distributed multipoles it can clearly be seen that the EPS is becoming more accurate as the distance to the atoms becomes larger. This is due to the fact that the distributed multipole expansion converges at a given radius from the multipole sites which varies for different molecules. In addition to the convergences of the multipole expansion with distance there is a convergence of the expansion as a function of the highest multipole rank to consider. For this and the following evaluations, multipole expansions are truncated at rank 2 (quadrupole). This rank has been found to provide a good trade-off between convergence and the computational effort to calculate multipolar energies and interactions.
Molecules with optimized electron potential coordinates (yellow spheres). First row: H\({}_{2}\)O, NH\({}_{3}\), methane benzene. Second row: H\({}_{2}\)S, PH\({}_{3}\), CO\({}_{2}\), pyrrole. Third row: ethanol, methylamine, formaldehyde, thiophene.
The comparison of the EPS above is obviously just qualitative, but it illustrates the differences between the _ab inito_ EPS and the different models. Most importantly it shows that there are systematic differences between the three models that are observed for different molecules. Furthermore the accuracy of the different models varies as a function of the distance to the atoms. Models which are more similar to _ab initio_ in the short range may perform badly at longer distances and vice versa, therefore for the quantitative evaluation of the EPS accuracy in the next section, the distance dependence will be considered.
Electrostatic potential surfaces for NH\({}_{3}\). Upper row: comparison to MP2/aug-cc-pVTZ, DMA and Mulliken charges. Lower row: EPS of the three valence electron potential force field (VePff) models. White areas: energies outside the range of \(\pm\)40 kcal/mol.
Electrostatic potential surfaces for CH\({}_{4}\). Upper row: comparison to MP2/aug-cc-pVTZ, DMA and Mulliken charges. Lower row: EPS of the three valence electron potential force field (VePff) models. White areas: energies outside the range of \(\pm\)40 kcal/mol.
Electrostatic potential surfaces for OCH\({}_{2}\). Upper row: comparison to MP2/aug-cc-pVTZ, DMA and Mulliken charges. Lower row: EPS of the three valence electron potential force field (VePff) models. White areas: energies outside the range of \(\pm\)40 kcal/mol.
### Evaluation of electrostatic potential accuracy
In order to quantitatively compare the EPS accuracy, the differences between the electrostatic potential of different models and the _ab initio_ EPS calculated at the MP2/aug-cc-pVTZ level of theory are evaluated on a 3-dimensional grid centered at the geometric center of the molecules. The evaluation is carried out for segments of points within increasing distance ranges of any atom in the molecule. This distance dependent evaluation allows to compare not only the overall accuracy of each model, but also to assess the distance range in which the errors occur. As the electrostatic potential is higher at short distances, the errors are in general also higher in the near range. In the results are compared between the three models in their point charge version and in their Gaussian distribution version. In the point charge version of the three models is compared to Mulliken charges and distributed multipoles. For clarity additional representation of this comparison for each model separately are shown in the Supplementary material (SI Figures 1-3).
All interaction models including DMA and Mulliken charges have been parametrized at the B3LYP/aug-cc-pVTZ level of theory, therefore the evaluations here also assess the transferability between a computationally less expensive method used for parametrization and a computationally more expensive and more accurate reference. The molecules are divided into two groups of six molecules where the upper group of six molecules in both Figures shows molecules with only two atom types, whereas the lower group contains more than two atom types. This distinction is important due to the fact that for molecules with only two atom types, the parameter space of the electrostatic parameter \(\zeta\) contains a value corresponding to Mulliken charges. Therefore if these molecules are more accurate than Mulliken charges this is mainly due probing a larger parameter space, as well as due to the transferability between the DFT EPS used for fitting and the MP2 reference calculation. For the molecules with more than two atom types in contrast, the parameter space does not necessarily contain Mulliken charges, therefore higher accuracy than Mulliken charges implies a real advantage of this model and parametrization.
Comparison of the evaluations on all 12 molecules shows that in general Mulliken charges are the least accurate method, DMA is most accurate in the longer range, whereas in the short range its equally or less accurate than the three VePff models. Models 1 and 2 perform very similar,therefore the green lines representing Model 1 are mostly covered by the the red lines representing Model 2. Model 3 performs better than Models 1 and 2 except for CO\({}_{2}\) and and formaldehyde. The Gaussian distribution models are systematically more accurate than the point charge models, which are more accurate than the point charge models.
Accuracy evaluation for Model 1 (green), Model 2 (red) and Model 3 (blue) in their point charge (solid lines) and Gaussian distribution version (dashed lines). The differences between each model and the MP2/aug-cc-pVTZ EPS are evaluated for segments of increasing distance to the atoms. The upper six molecules contain only two atom types, the lower six molecules more than two atom types.
Due to this result, the Gaussian charge distribution models will not be evaluated for intermolecular interaction energies in the next section as the differences between different models are found here to be
Comparison of the accuracy of point charge Model 1 (green), Model 2 (red) and Model 3 (blue) to DMA (black) and Mulliken charges (brown). The differences between each model and the MP2/aug-cc-pVTZ EPS are evaluated for segments of increasing distance to the atoms. The upper six molecules contain only two atom types, the lower six molecules more than two atom types.
Furthermore intermolecular interaction energies are dominated by repulsive interaction potentials in the short range, so most likely the accuracy of the repulsive energy terms is more important than the additional accuracy due to the use of Gaussian functions.
### Evaluation of intermolecular interaction energies
The electrostatic potential evaluations are suitable to assess the general properties of the different models. For the use of these models in atomistic force fields it is however more important to evaluate the accuracy of the intermolecular interaction energies. In principle the parameters obtained for the EPS could also be used for the interaction energies, however, in practice it has been found that it is better to parameterize force fields based on interaction energies and to jointly fit all parameters to the intermolecular energy. Therefore the \(\zeta\)- and \(\epsilon\)-parameters have been fitted to the intermolecular interaction energies of eight example molecules (Table 2), including both, molecules with only two atom types and molecules with more than two atom types. The accuracy of the three models compared to _ab initio_ is first illustrated for NH\({}_{3}\), CH\({}_{4}\) and formaldehyde as example molecules. The interaction energies and a subset of the dimer geometries are shown in Figures 8, 9 and 10.
Comparison of the three molecules shows a clearly visible advantage for Model 2 in the case of NH\({}_{3}\) and formaldehyde, whereas for methane Model 3 is the most accurate. The average energy differences \(\bar{\Delta}_{E\_inter}\) between the _ab initio_ interaction energies for all eight molecules are shown in Table 3. For all cases the most accurate model is either Model 2 or Model 3, i.e.one of the two models containing electron potential sites. This demonstrates the advantage of using electron potentials despite the fact that a number of generic parameters are used and the number of parameters to be fitted is the same as for Model 1. Model 2 seems to perform better for polar molecules, i.e. for molecules with larger electronegativity differences, whereas Model 3 performs better for apolar molecules. A comparison of the \(\zeta\)-parameters obtained from the EPS in Table 1 to the parameters obtained from the intermolecular interaction energies in Table 2 shows that there is no systematic difference between the two sets of parameters for Models 1 and 2, whereas for Model 3 the parameters obtained from the interaction energies are systematically smaller. This indicatesthat Model 3 captures a systematic difference between the electrostatic potential of monomers and dimers, most likely due to polarization. Potentially this result could be interesting for describing polarization effects systematically. For Model 1 and 2, the differences in the parameters are not systematic and therefore most likely arising due to a combination of polarization effects and error compensation.
NH\({}_{3}\) interaction energies fitted and compared to MP2/aug-cc-pVTZ calculations (black): Model 1 (green squares), Model 2 (red circles), Model 3 (blue triangles). Example conformations are shown with the corresponding conformation number.
CH\({}_{4}\) interaction energies fitted and compared to MP2/aug-cc-pVTZ calculations (black): Model 1 (green squares), Model 2 (red circles), Model 3 (blue triangles). Example conformations are shown with the corresponding conformation number.
\begin{table}
\begin{tabular}{l||c|c|c} & \multicolumn{3}{c}{\(\bar{\Delta}_{E,\mu iter}\) kcal/mol} \\ \hline & Model 1 & Model 2 & Model 3 \\ \hline \hline H\({}_{2}\)O & 1.0198 & 0.9897 & 1.3504 \\ \hline NH\({}_{3}\) & 1.8860 & 0.3326 & 0.8152 \\ \hline CH\({}_{4}\) & 0.0107 & 0.0320 & 0.0078 \\ \hline H\({}_{2}\)S & 1.5788 & 0.6627 & 0.6166 \\ \hline PH\({}_{3}\) & 0.0141 & 0.0178 & 0.0116 \\ \hline CO\({}_{2}\) & 0.3651 & 0.2853 & 0.3317 \\ \hline CH\({}_{5}\)N & 0.0047 & 0.0046 & 0.0254 \\ \hline OCH\({}_{2}\) & 0.3358 & 0.1060 & 0.3631 \\ \hline \end{tabular}
\end{table}
Table 3: Average energy differences \(\bar{\Delta}_{E,\mu iter}\) between the _ab initio_ interaction energies calculated at the MP2/aug-cc-pVTZ level of theory and each model in its point charge version.
Formaldehyde (OCH\({}_{2}\)) interaction energies fitted and compared to MP2/aug-cc-pVTZ calculations (black): Model 1 (green squares), Model 2 (red circles), Model 3 (blue triangles). Example conformations are shown with the corresponding conformation number.
Discussion and Conclusions
The possibility of constructing valence electron based potentials for the nonbonded interactions in atomistic force fields has been explored in this paper. Three charge distribution models using simple potential functions and only one adjustable parameter for the electrostatic potential of a molecule have been introduced and compared. It was shown that even with this simple parametrization which uses empirical constants and generic parameter values for the additional sites, the electrostatic potential is more or equally accurate than with population derived charges for all three models. For half of the evaluated molecules the EPS is equally or more accurate than distributed multipole moments for at least one of the models. The comparison of point charge interaction potentials to spherical Gaussian distributions showed that the accuracy of the electrostatic potential can be further increased by using Gaussian distributions. However, in most cases the improvement due to Gaussian functions is small and only relevant in the distance range close to the atoms. Since in this distance range the interactions are dominated by Pauli repulsion it is more important to assess the overall accuracy of the intermolecular interactions composed of electrostatics, LJ-potentials and repulsive potentials on the electron potential sites. The evaluation of the intermolecular interaction energies compared to _ab initio_ revealed a systematic advantage of having electron potential sites. For the charge distribution models it was found that Model 2 which uses a polar charge distribution scheme is more accurate for molecules with larger electronegativity differences i.e. polar molecules, whereas Model 3 is more accurate for molecule with small electronegativity differences.
As all the molecules used for the evaluations here were small, it was possible to use molecular parameters for fitting the different molecular parameters. In order to generalize this approach to larger molecules it would most likely be necessary to use fitting parameters for groups of atoms rather than entire molecules. As the accuracy of the different models seems to depend on the polarity of the molecules, it would probably be suitable to parametrize larger molecules by partitioning them into functional groups. Overall the methods presented here offers a new concept for introducing and parametrizing additional interaction sites to improve the accuracy of intermolecular interactions in atomistic force fields. The concept of charge distribution models provides a pathway to use more accurate potential functions without increasing the parameter space and therefore the parametrization effort.
|
10.48550/arXiv.1603.03285
|
Exploring the properties of valence electron based potential functions for the nonbonded interactions in atomistic force fields
|
Nuria Plattner
| 2,251
|
10.48550_arXiv.1803.04536
|
###### Abstract
Within the framework of linear-scaling Kohn-Sham density functional theory, a robust method for maintaining compact localized orbitals close to the ground state is coupled with nuclear dynamics. This allows to obviate the commonly employed optimization of the one-electron density matrix and thus create an efficient orbital-only molecular dynamics method for weakly-interacting systems. An application to liquid water demonstrates that the low computational overhead of the method makes it well-suited for routine simulations whereas its linear-scaling complexity allows to extend first-principle dynamical studies of molecular systems to previously inaccessible length scales.
Since the unification of molecular dynamics and density functional theory (DFT), _ab initio_ molecular dynamics (AIMD) has become an important tool to study processes in molecules and materials. Unfortunately, the computational cost of the conventional Kohn-Sham (KS) DFT grows cubically with the number of atoms, which severely limits the _length scales_ accessible by AIMD. To address this issue, substantial efforts have been directed to the development of linear-scaling (LS) DFT.
In all LS DFT methods, the _delocalized_ eigenstates of the effective KS Hamiltonian must be replaced with an alternative set of _local_ electronic descriptors. Most LS methods explore the natural locality of the one-electron density matrix (DM). However, the DM DFT becomes advantageous only for impractically large systems when accurate multifunction basis sets are used. This issue is rectified in optimal-basis DM methods that contract large basis sets into a small number of new localized functions and then optimize the DM in the contracted basis. Despite becoming the most popular approach to LS DFT, the efficiency of these methods is hampered by the costly optimization of both the contracted orbitals and the DM. From this point of view, a direct variation of molecular orbitals that are strictly localized within predefined regions is preferable because LS can be achieved with significantly fewer variables. Advantages of the orbitals-only LS DFT are especially pronounced in accurate calculations that require many basis functions per atom. Unfortunately, the development of promising orbital-based LS methods has been all but abandoned because of the inherently difficult optimization of localized orbitals.
Thus, despite impressive progress of the LS description of the electronic and atomic structure of large static systems, the high computational overhead of existing LS methods restrict their use in dynamical simulations to very short _time scales_, systems of low dimensions, and low-quality minimal basis sets. On typical length and time scales required in practical and accurate AIMD simulations, LS DFT still cannot compete with the straightforward low-cost cubically-scaling KS DFT.
In this work, we present an AIMD method that overcomes difficulties of orbital-only local DFT to achieve LS with extremely low computational overhead. To demonstrate advantages of the new method we applied it here to systems of weakly-interacting molecules. However, the same approach is readily applicable to systems of strongly-interacting fragments that do not form strong covalent bonds such as ionic materials--salts, liquids, and semiconductors. A generalization of the method to all finite-gap systems, including covalently bonded atoms, will be reported later.
The new AIMD method utilizes a recently developed LS DFT based on absolutely localized molecular orbitals (ALMOs). Unlike delocalized KS orbitals, each ALMO has its own _localization center_ and a predefined _localization radius_\(R_{c}\) that typically includes nearby atoms or molecules. In the current implementation, a localization center is defined as a set of all Gaussian atomic orbitals of one molecule. However, the approach can use other local and nonlocal basis sets. The key feature of ALMO DFT is that its one-electron wavefunctions are constructed in a two-stage self-consistent-field (SCF) procedure to circumvent the problem of the sluggish variational optimization emphasized above. In the first stage, ALMOs are constrained to their localization centers whereas, in the second stage, ALMOs are relaxed to allow delocalization onto the neighbor molecules within their localization radius \(R_{c}\). To achieve a robust optimization in the problematic second stage, it is important to keep the delocalization component of the trial wavefunction orthogonal to the fixed orbitals obtained in the first stage. For mathematical details, see the ALMO SCF method in Ref..
ALMO constraints imposed by \(R_{c}\) prohibit electron density transfer between distant molecules, but retain all other types of interaction such as long-range electrostatic, exchange, polarization, and--if the exchange-correlation (XC) functional includes them--dispersion interactions. Since the importance of electron transfer decays exponentially with distance in finite-gap materials, the ALMO approximation is expected to provide a natural and accurate representation of the electronic structure of molecular systems. Because of the greatlyreduced number of electronic descriptors and the robust optimization, the computational complexity of ALMO DFT grows linearly with the number of molecules while its computational overhead remains very low. These features make ALMO DFT a promising method for accurate AIMD simulations of large molecular systems.
The challenge of adopting ALMO DFT for dynamical simulations arises from the slightly nonvariational character of the localized orbitals. While ALMOs are variationally optimized in both SCF stages, the occupied subspace defined in the first stage must remain fixed during the second stage to ensure convergence. In addition, electron transfer effects can suddenly become inactive in the course of a dynamical simulation when a neighboring molecule crosses the localization threshold \(R_{c}\). Furthermore, the variational optimization in any AIMD method is never complete in practice and interrupted once the maximum norm of the gradient of the energy with respect to the electronic descriptors drops below small but nevertheless finite convergence threshold \(\epsilon_{\rm SCF}\). These errors do not affect the accuracy of static ALMO DFT calculations, geometry optimization, and Monte-Carlo simulations. Unfortunately they tend to accumulate in molecular dynamics trajectories leading to non-physical sampling and eventual failure. Traditional strategies to cope with these problems are computationally expensive and include computing the nonvariational contribution to the forces via a variational coupled-perturbed procedure, increasing \(R_{c}\), and decreasing \(\epsilon_{\rm SCF}\).
In this work, we propose another approach that obviates the need in a coupled-perturbed solver, relaxes tight constraints on \(R_{c}\) and \(\epsilon_{\rm SCF}\), and thus enables us to maintain stable dynamics and to keep the algorithmic complexity and cost of simulations low. In our approach, the forces on atoms are calculated _approximately_ after the two-stage ALMO SCF using a straightforward procedure that computes only the Hellmann-Feynman and Pulay components and neglects the computationally intense nonvariational component of the forces. The difference between these approximate ALMO forces and the _reference_ forces that could be obtained from perfectly converged fully-delocalized KS orbitals is \(\delta f_{i\alpha}(t)\):
\[f_{i\alpha}^{\rm KS}(t)=f_{i\alpha}^{\rm ALMO}(t)+\delta f_{i\alpha}(t), \tag{1}\]
\(\delta f_{i\alpha}(t)\) comprises all neglected terms that originate from a finite localization radius \(R_{c}\) and incomplete SCF optimization. \(\delta f_{i\alpha}(t)\) can be reduced to zero _systematically_ by increasing \(R_{c}\) and decreasing \(\epsilon_{\rm SCF}\).
Our approach to compensate for the missing \(\delta f_{i\alpha}(t)\) term is inspired by the methodology introduced into AIMD by Krajewski _et al._, formalized by Kuhne _et al._ and rationalized by Dai _et al._ before becoming informally known as the second generation Car-Parrinello molecular dynamics. Adopting the principle of Refs., ALMO AIMD is chosen to be governed by the Langevin equation of motion that can be written in terms of the unknown reference forces
\[m_{i}\ddot{r}_{i\alpha}=f_{i\alpha}^{\rm KS}(t)-\gamma m_{i}\dot{r}_{i\alpha}+R _{i\alpha}^{\gamma}(t), \tag{2}\]
where \(m_{i}\) is the mass of atom \(i\), \(r_{i\alpha}\) is its position along dimension \(\alpha\), \(\gamma\) is the Langevin scaling factor, and \(R_{i\alpha}^{\gamma}(t)\) is the stochastic force represented by a zero-mean white Gaussian noise
\[\langle R_{i\alpha}^{\gamma}(t)\rangle =0, \tag{3}\] \[\langle R_{i\alpha}^{\gamma}(t)R_{j\beta}^{\gamma}(t^{\prime})\rangle =2k_{B}T\gamma m_{i}\delta_{ij}\delta_{\alpha\beta}\delta(t-t^{ \prime}). \tag{4}\]
The last relation means that, for any value of \(\gamma\), the damping and stochastic terms are in perfect balance and trajectories generated with Eq. will sample the canonical ensemble at a specified temperature \(T\). In the limit \(\gamma\to 0\), the Newton equation is recovered and the microcanonical ensemble is sampled.
The main assumption of ALMO AIMD is that the error in the ALMO forces is well approximated by Gaussian noise \(R_{i\alpha}^{\Delta}(t)\):
\[\delta f_{i\alpha}(t)=R_{i\alpha}^{\Delta}(t) \tag{5}\]
that obeys
\[\langle R_{i\alpha}^{\Delta}(t)\rangle =0, \tag{6}\] \[\langle R_{i\alpha}^{\Delta}(t)R_{j\beta}^{\Delta}(t^{\prime})\rangle =2k_{B}T\Delta m_{i}\delta_{ij}\delta_{\alpha\beta}\delta(t-t^{ \prime}). \tag{7}\]
This assumption, shown to be well justified, allows us to rewrite the Langevin equation using the ALMO forces
\[m_{i}\ddot{r}_{i\alpha}=f_{i\alpha}^{\rm ALMO}(t)-\gamma m_{i}\dot{r}_{i\alpha} +R_{i\alpha}^{\gamma+\Delta}(t), \tag{8}\]
The only missing piece in the modified Langevin equation is the value of \(\Delta\), which describes the strength of the newly introduced stochastic term. This term compensates for imperfections in ALMO forces and must be adjusted to re-balance the damping and stochastic components in ALMO AIMD.
In principle, \(\Delta\) can be calculated using the integral of Eq. averaged over atoms with different \(m_{i}\)
\[\Delta=(2k_{B}Tm_{i})^{-1}\int_{-\infty}^{\infty}\frac{1}{3}\langle\delta\vec{ f}_{i}\cdot\delta\vec{f}_{i}(\tau)\rangle d\tau \tag{9}\]
\(R_{c}\rightarrow\infty\) and \(\epsilon_{\rm SCF}\to 0\)) for a short representative AIMD trajectory. In practice, we found (see results below) that this approach is not particularly accurate because the \(\delta_{ij}\delta_{\alpha\beta}\) assumption in Eq. does not strictly hold. Nevertheless, the ACF integral can provide a reasonable starting value of \(\Delta\). This value can be further fine-tuned in a series of short trial-and-error ALMO AIMD runs until the average kinetic energy corresponds to the requested temperature \(\langle\frac{1}{2}m_{i}\dot{\mathbf{r}}_{i}^{2}\rangle=\frac{3}{2}k_{B}T\).
The inherently stochastic approach presented here does not aim to produce fully time-reversible dynamics for atomic nuclei. Nevertheless, it is capable to reproduce correct dynamical properties of a system as long as \(\gamma\) is set to a small value and partially optimized ALMOs remain close to the ground state resulting in \(\Delta\ll\gamma\).
ALMO AIMD was implemented in CP2K, an open source materials modeling package. Accuracy and efficiency of ALMO AIMD was tested using liquid water as an example. This system is challenging because intermolecular electron delocalization is a critical component of hydrogen bonding and must be described correctly to reproduce static and dynamical properties of liquid water. A periodic cell containing 125 molecules was simulated for \(30\,\mathrm{ps}\) at \(T=298\,\mathrm{K}\) and a constant density of \(1.01\,\mathrm{g\cdot cm^{-3}}\). Ricci-Ciccotti algorithm was used to integrate the Langevin equation. We found that \(\gamma=10^{-3}\,\mathrm{fs^{-1}}\) is large enough to thermostat the system efficiently and small enough not to significantly affect dynamical properties of liquid water. In the dual Gaussian and plane-wave scheme implemented in CP2K, the TZV2P basis set was used to represent molecular orbitals, and a plane-wave cutoff of \(320\,\mathrm{Ry}\) used to represent electron density. The XC energy was approximated using the dispersion-corrected PBE functional. Separable norm-conserving pseudopotentials were used and the Brillouin zone was sampled at the \(\Gamma\)-point. The predictor of the Kolafa scheme was adopted to localized orbitals to generate a highly accurate initial ALMOs in both SCF stages, which can be brought close to the ground state with just a few SCF steps of the robust two-stage optimization procedure.
The _reference_ forces were calculated with fully delocalized electrons using the tightly converged, \(\epsilon_{\mathrm{SCF}}=10^{-6}\) a.u., orbital transformation (OT) method. In ALMO AIMD, the element-specific cutoff radius for electron delocalization \(R_{c}\) was set to 1.6 in units of the elements' van der Waals radii (vdWR). This localization radius includes approximately two coordination shells of an average water molecule and was shown to reproduce the reference radial distribution function (RDF) perfectly in Monte-Carlo simulations. To check the ability of the \(R^{\Delta}(t)\) term to compensate for imperfections in ALMO forces, we varied \(\epsilon_{\mathrm{SCF}}\) between tight \(10^{-6}\) a.u. and loose \(10^{-2}\) a.u.
Even with \(\epsilon_{\mathrm{SCF}}=10^{-2}\), the simulation is stable with the correct average temperature and perfect Maxwell-Boltzmann distribution. \(\Delta\) was initially estimated at \(2\times 10^{-5}\,\mathrm{fs^{-1}}\) using Eq. and then refined heuristically to \(6\times 10^{-5}\,\mathrm{fs^{-1}}\). We found that it is easier to optimize \(\Delta\) when \(\gamma\) is set to zero because of reduced noise in the trial runs. Analysis of \(\delta\vec{f}_{i}(t)\) shows that the error indeed resembles Gaussian white noise. The mean of the error is zero (black line in Figure 2a). Its ACF decays rapidly so that the errors can be considered uncorrelated on time scale of \(50\,\mathrm{fs}\). Thus the main assumption behind our approach to ALMO AIMD is justified for liquid water. We established that the main source of error in forces for this system is the loose convergence criterion and not the finite \(R_{c}\): fully converged ALMO SCF calculations remove the oscillating component of \(\delta f\). We also verified that the ALMO forces converge to the reference forces in the limit \(R_{c}\to\infty\) (Figure S1 in the Supplemental Material).
To test the accuracy of ALMO AIMD we used the trajectory analyzer TRAVIS to compute the infrared (IR) spectrum, RDF, and diffusion coefficient of liquid water from both the ALMO trajectory (\(\epsilon_{\mathrm{SCF}}=10^{-2}\) a.u. and \(R_{c}=1.6\) vdWR) and from the reference trajectory. The diffusion coefficients \(D_{\mathrm{OT}}=1.7\times 10^{-10}\,\mathrm{m^{2}\cdot s^{-1}}\) and \(D_{\mathrm{ALMO}}=1.8\times 10^{-10}\,\mathrm{m^{2}\cdot s^{-1}}\) and RDFs are in good agreement.
Calculated properties of water using ALMO AIMD with \(\epsilon_{\mathrm{SCF}}=10^{-2}\) a.u. and \(R_{c}=1.6\) vdWR (red line) and fully converged OT reference (black line). (a) RDF, (b) kinetic energy distribution (the gray curve shows the theoretical Maxwell-Boltzmann distribution), (c) IR spectrum.
These stringent tests show that despite noticeable errors in the ALMO forces, the compensating \(R^{\Delta}(t)\) term in the modified Langevin equation makes it possible to recover atomic dynamics properly. We would like to note that ALMO AIMD could not be stabilized with \(\Delta=0\). Neither were we able to find any values of \(\Delta\) that stabilize trajectories generated using perturbative versions of ALMO DFT.
To demonstrate the computational efficiency of ALMO AIMD, we compared the average wall-time per MD step for a variety of methods in It is important to emphasize that the comparison is performed for a three-dimensional condensed phase system described with an accurate triple-\(\zeta\) basis set with polarized functions--a particularly challenging case for DM-based LS methods. ALMO AIMD shows clear LS behavior for all values of \(\epsilon_{\text{SCF}}\), even for medium-size systems. While the second generation Car-Parrinello method decreases the computational overhead of the cubically-scaling AIMD for small systems, ALMO AIMD exploits the modified Langevin concept to substantially reduce the simulation cost for systems of all sizes. The crossover point between ALMO AIMD and cubically scaling methods lies in the region of 256 molecules--length scales routinely accessible with AIMD today.
Weak scaling benchmarks for very large systems show that localized orbitals are naturally suited for parallel execution: LS is retained for a wide range of systems and compute cores. We were able to successfully simulate systems as large as \(\sim 10^{5}\) atoms within reasonable wall-clock time using only moderate number of compute cores--an impressive feat for AIMD considering that accurate molecular orbitals and the idempotent DM are computed on each step. The horizontal line in is shown as a rough guide to time and length scales accessible in a fixed wall-clock time given various computational resources. It indicates that ALMO AIMD can extend the range of routine simulations to \(\sim 10^{4}\) atoms on modern HPC platforms.
To summarize, we demonstrated--for the first time--that compact localized orbitals can be utilized to perform accurate and efficient LS AIMD without concomitant optimization of the DM. High efficiency of the presented method is achieved without sacrificing accuracy with a combination of two techniques: on-the-fly calculation of approximate forces without lengthy self-consistent optimization of localized orbitals and integration of a modified Langevin equation of motion that is fine-tuned to retain stable dynamics even with imperfect forces. By obviating the optimization of the DM, the method remains remarkably efficient even with large localized basis sets. Using liquid water as an example, we showed that the new approach enables simulations of molecular systems on previously inaccessible length scales. The developed method will have a significant impact on modeling of complex molecular systems (e.g. interfaces or nuclei) making completely new phenomena accessible to AIMD. Generalization of the methodology to systems of strongly interacting atoms (e.g. covalent crystals) is underway.
|
10.48550/arXiv.1803.04536
|
Low-cost orbital-based linear-scaling \emph{ab initio} molecular dynamics for weakly-interacting systems
|
Hayden Scheiber, Yifei Shi, Rustam Z. Khaliullin
| 1,174
|
10.48550_arXiv.2005.03938
|
###### Abstract
Pseudoscalar or pseudovector cosmic fields, that serve as a source of parity (\(\mathcal{P}\)) violation, are invoked in different models for cold dark matter or in the standard model extension that allows for Lorentz invariance violation. A direct detection of the timelike-component of such fields requires a direct measurement of \(\mathcal{P}\)-odd potentials or their evolution over time. Herein, advantageous properties of chiral molecules, in which \(\mathcal{P}\)-odd potentials lead to resonance frequency differences between enantiomers, for direct detection of such \(\mathcal{P}\)-odd cosmic fields are demonstrated. Scaling behavior of electronic structure enhancements of such interactions with respect to nuclear charge number and the fine-structure constant is derived analytically. This allows a simple estimate of the effect sizes for arbitrary molecules. The analytical derivation is supported by quasi-relativistic numerical calculations in the molecules H\({}_{2}\)X\({}_{2}\) and H\({}_{2}\)XO with X = O, S, Se, Te, Po. Parity violating effects due to cosmic fields on the C-F stretching mode in CHBrClF are compared to electroweak parity violation and influences of non-separable anharmonic vibrational corrections are discussed. On this basis it was estimated from a twenty year old experiment with CHBrClF that bounds on Lorentz invariance violation as characterized by the parameter \(|\aleph_{0}^{6}|\) can be pushed down to the order of \(10^{-17}\,\mathrm{GeV}\) in modern experiments with suitably selected molecular system, which will be an improvement of the current best limits by at least two orders of magnitude. This serves to highlight the particular opportunities that precision spectroscopy of chiral molecules provides in the search for new physics beyond the standard model.
## I Introduction
In our recent work the virtues and prospects of chiral molecules as direct sensors for pseudovector and pseudoscalar cosmic fields were demonstrated. In the present paper we derive scaling laws for interactions of electrons with these fields, presented in and provide support from numerical calculations. Furthermore, the methods applied for derivation of limits on cosmic field interactions from experiments with chiral molecules are presented in a more detailed manner and accompanied by comparison to other computational methods.
One of the biggest puzzles of modern physics is the nature and composition of dark matter (DM) (see e.g.). Many different models for dark matter exist, considering objects that range from macroscopic to microscopic and from being hot (ultra-relativistic) to cold (non-relativistic). Among these DM theories cold DM (CDM) theory serves to provide a simple explanation for many cosmological observations. However, the constituents of CDM are unknown and can in principle fall in the range from macroscopic objects such as black holes to new fundamental particles like weakly interacting massive particles (WIMPs), axions, sterile neutrinos or dark photons (see e.g. Refs.).
Despite its merits, the model of CDM has several drawbacks. A possible solution of some of these provide fuzzy CDM models. Fuzzy CDM is supposed to consist of ultra light particles with masses of \(m_{\phi}\sim 1\times 10^{-22}\,\mathrm{eV}/c^{2}\). This model makes searches for ultralight CDM oscillating with frequencies on the order of \(1\,\mathrm{\SIUnitSymbolMicro Hz}\) particularly interesting.
CDM can consist of various types of weakly interacting particles (an overview can be found e.g. in Ref.). Among those pseudoscalar and pseudovector fields are of special interest as they are a source of parity violation.
_Pseudoscalar_ CDM particles behave like axions, which were originally proposed to solve the strong \(\mathcal{CP}\)-problem of quantum chromodynamics (QCD). The search for CDM particles can be restricted to a comparatively small parameter space assessable to the QCD axion (see e.g.) or can involve a wide range for axionic particles that are not bound to solve the strong \(\mathcal{CP}\)-problem. The latter are often referred to as axion-like particles (ALPs). _Pseudovector_ fields are important for models such as dark photons and also appear as sources of local Lorentz invariance violation in the Standard Model Extension (SME) by Kostelecky and coworkers.
In the last decade many new proposals for new experiments and improved bounds on pseudoscalar CDM appeared, employing atomic spectroscopy (see e.g.). Among those, strict limits on static \(\mathcal{P}\)-odd fields were set from direct detection of parity violation with modern atomic precision spectroscopy. In these experiments the dominating effect for parity violation stems from the electroweak \(Z^{0}\)-mediated electron-nuclear interaction.
Such \(\mathcal{P}\)-odd effects are strongly enhanced in chiral molecules as well (for recent reviews on molecular parity violation see). The chiral arrangement of the nuclei in the molecule leads to helicity of the electron cloud (see e.g. Ref.). Additional \(\mathcal{P}\)-odd effects can thenbe measured as energy difference between enantiomers of chiral molecules or as resonance frequency differences between the two non-identical mirror-image molecules. As frequency shifts can be measured very accurately, this appears to be a particularly promising tool to search for \(\mathcal{P}\)-odd cosmic fields.
In the following we analyse in detail the effects that emerge from \(\mathcal{P}\)-odd cosmic fields in chiral molecules. We derive scaling laws with respect to nuclear charge and the fine structure constant and compare to what is known from parity violation due to electroweak interactions. From our analysis we demonstrate advantages of the use of chiral molecules to search for \(\mathcal{P}\)-odd cosmic fields. We perform quasi-relativistic calculations at different levels of theory and estimate the effect sizes in the vibrational spectra of the chiral methane derivate CHBrClF. Thereby, the computational difficulties are highlighted. From a twenty year old experiment with this molecule we estimate the sensitivity on cosmic parity violation and discuss the scope for improvement on these limits in modern experiments with chiral molecules and by improvement of present theoretical methods.
## II Theory
### Parity non-conserving interactions of electrons with cosmic fields
\(\mathcal{P}\)-odd interactions of electrons with pseudoscalar and pseudovector cosmic fields were discussed in detail in Ref.. A light pseudoscalar cosmic field obeys the Klein-Gordon equation. Assuming it to be non-relativistic, i.e. \(\hbar\omega_{\phi}\approx m_{\phi}c^{2}\) with \(m_{\phi}\) being the CDM particle mass and \(c\) being the speed of light in vacuum, we can write
\[\phi(\vec{r},t)=\phi_{0}\cos\left(\omega_{\phi}t-\frac{\vec{r}\cdot\vec{p}_{ \phi}}{\hbar}+\varphi\right), \tag{1}\]
As the relative velocity of the ALP field is suppressed by \(10^{-3}\) with respect to the speed of light (see Refs. for details), for terrestrial experiments we can assume \(\frac{\vec{r}\cdot\vec{p}_{\phi}}{\hbar}\) to be constant and choose \(\varphi\) such that eq. can be written as \(\phi(\vec{r},t)=\phi_{0}\cos(\omega_{\phi}t)\) (see also Ref.).
The interaction of the electronic field \(\psi_{\rm e}\) with such pseudoscalar fields \(\phi\) can be described by (see e.g.)
\[\mathcal{L}^{\phi}_{\rm ps}=g_{\phi{\rm e}{\rm e}}(\hbar c\,\partial_{\mu}\phi) \bar{\psi}_{\rm e}\mathbf{\gamma}^{\mu}\mathbf{\gamma}^{5}\psi_{\rm e}\,, \tag{2}\]
Herein the Dirac matrices are defined as
\[\mathbf{\gamma}^{0}=\begin{pmatrix}\mathbf{1}_{2\times 2}&\mathbf{0}_{2\times 2} \\ \mathbf{0}_{2\times 2}&-\mathbf{1}_{2\times 2}\end{pmatrix},\qquad\mathbf{\gamma}^{k}= \begin{pmatrix}\mathbf{0}_{2\times 2}&\mathbf{\sigma}^{k}\\ -\mathbf{\sigma}^{k}&\mathbf{0}_{2\times 2}\end{pmatrix}, \tag{3}\]
\(\mathbf{\gamma}^{5}=\mathbf{\imath}\mathbf{\gamma}^{0}\mathbf{\gamma}^{1}\mathbf{\gamma}^{2}\mathbf{ \gamma}^{3}\), where \(\imath=\sqrt{-1}\) is the imaginary unit, \(\partial_{\mu}=\frac{0}{\partial x^{2}}\) is the first derivative with respect to the four-vector \(x^{\mu}=(ct,x,y,z)\) and Einstein's sum convention is used. Additionally a direct pseudoscalar coupling between the electrons and the pseudoscalar cosmic field can be considered (see e.g. Ref.):
\[\mathcal{L}^{\phi}_{\rm dps}=-\imath\tilde{g}_{\phi{\rm e}{\rm e}}m_{\rm e}c^{ 2}\phi\bar{\psi}_{\rm e}\mathbf{\gamma}^{5}\psi_{\rm e}\,, \tag{4}\]
Whereas this interaction can lead to parity violating couplings when considering transition matrix elements of atomic or molecular excitations, it does not contribute to parity violating expectation values, which give dominant contributions to frequency differences in spectra of chiral molecules. Thus these interactions are not discussed any further in the following.
The time-derivative of the pseudoscalar field leads to the \(\mathcal{P}\)-odd single-electron Hamiltonian
\[\hat{h}_{\rm ps}=g_{\phi{\rm e}{\rm e}}\sqrt{2(hc)^{3}\rho_{\rm CDM}}\sin( \omega_{\phi}t)\mathbf{\gamma}^{5}, \tag{5}\]
We use lowercase letters (\(\hat{h}\)) for single-electron operators and uppercase letters (\(\hat{H}\)) for multi-electron operators. These are in the case of \(\hat{H}_{\rm ps}\) (as well as \(\hat{H}_{\rm pv}\), \(\hat{H}_{\rm ew}\) given below) simple sums over all electrons of the system, e.g. \(\hat{H}_{\rm ps}=\sum_{\imath}\hat{h}_{\rm ps}(i)\)
Electronic interactions with pseudovector cosmic fields can be described by the Lagrangian
\[\mathcal{L}^{b}_{\rm pv}=-b_{\mu}\bar{\psi}_{\rm e}\mathbf{\gamma}^{\mu}\mathbf{\gamma} ^{5}\psi_{\rm e}, \tag{6}\]
in the local Lorentz invariance violating Standard Model Extension (SME) (for details see Refs.).
The parity non-conserving interaction Hamiltonian for the temporal component is
\[\hat{h}_{\rm pv}=b_{0}(t)\mathbf{\gamma}^{5}, \tag{7}\]
Here \(b_{0}^{\rm e}\) is the interaction strength of the timelike-component of the pseudovector field with the electrons.
In spectra of chiral molecules the interactions discussed above lead to shifts (static fields) or oscillations (dynamic fields) of frequency shifts due to the nuclear spin-independent electroweak interactions, the main contribution to which is in closed-shell molecules expected to arise from the electron-nuclei weak neutral-current interaction Hamiltonian (see e.g.):
\[\hat{h}_{\rm ew}=\frac{G_{\rm F}}{2\sqrt{2}}\sum_{A=1}^{N_{\rm nuc}}Q_{\rm W,A} \rho_{A}(\vec{r})\mathbf{\gamma}^{5}\,, \tag{8}\]where \(G_{\rm F}=2.222\,49\times 10^{-14}\,E_{\rm h}\,{a_{0}}^{3}\) is Fermi's weak coupling constant, \(Q_{\rm W,A}\) and \(\rho_{A}\) are the weak charge and normalized charge density of nucleus \(A\), respectively. The total number of nuclei is \(N_{\rm nuc}\). Contributions from \(\mathcal{P}\)-odd nuclear-spin dependent terms when combined with \(\mathcal{P}\)-even hyperfine coupling are estimated to give only minor contributions in closed-shell molecules. Similar considerations hold for the contribution from neutral-current interaction terms between electrons.
It shall be noted that in chiral molecules weakly interacting dark matter candidates, such as WIMPs, or cosmic neutrinos can also lead to shifts or oscillations of the \(\mathcal{P}\)-odd potential as was discussed by Bargueno _et.al._. These interactions as well as those of electrons with pseudoscalar and pseudovector fields discussed above are proportional to \(\left\langle\mathbf{\gamma}^{5}\right\rangle\). In the following we will discuss in general the chiral operator \(\mathbf{\gamma}^{5}\), which leads to parity non-conservation and compare to known properties of operator.
### Molecular expectation value of \(\mathbf{\gamma}^{5}\)
The time-independent Dirac-Coulomb equation for the electronic system of the molecule reads
\[\hat{H}_{\rm DC}\Psi_{I}=E_{I}\Psi_{I}, \tag{9}\]
with \(\Psi_{I}\) and \(E_{I}\) being the \(I\)th eigenfunction and eigenvalue of the Dirac-Coulomb Hamiltonian being given by
\[\begin{split}\hat{H}_{\rm DC}&=\sum_{i}^{N_{\rm nuc }}\left[c\mathbf{\gamma}^{0}\mathbf{\tilde{\gamma}}\cdot\hat{\vec{p}}_{i}+\left(\mathbf{ \gamma}^{0}-\mathbf{1}\right)m_{\rm e}c^{2}\right.\\ &\left.+V_{\rm nuc}(\vec{r}_{i})+\frac{1}{2}\sum_{j\neq i}^{N_{ \rm nuc}}k_{\rm res}\frac{e^{2}}{|\vec{r}_{i}-\vec{r}_{j}|}\right]\,,\end{split} \tag{10}\]
Here \(e\) is the elementary electric charge, \(k_{\rm res}\) is in SI units \(\frac{1}{4\pi\varepsilon_{0}}\) with \(\epsilon_{0}\) being the electric constant and \(V_{\rm nuc}\) being the potential the nuclei in the molecule produce.
In the Dirac-Hartree-Fock-Coulomb (DHFC) approach, the multi-electron states \(\Psi_{I}\) are approximated by a Slater determinant build from an orthonormal set of single-electron bi-spinors \(\psi_{i}\) with orbital energy \(\epsilon_{i}\).
\[\psi_{i}(\vec{r})=\begin{pmatrix}\varphi_{i}(\vec{r})\\ \chi_{i}(\vec{r})\end{pmatrix} \tag{11}\]
can be found via
\[\chi_{i}(\vec{r})=c\left(2m_{\rm e}c^{2}-\hat{V}+\varepsilon_{i}\right)^{-1} \mathbf{\tilde{\sigma}}\cdot\hat{\vec{p}}\,\varphi_{i}(\vec{r}), \tag{12}\]
For the remaining part of this section we will use atomic units, in which \(\hbar\), \(|e|\) and \(m_{\rm e}\) have the numerical value of 1. Then, the term in parentheses in eq.
\[c\left(2c^{2}-\hat{V}+\varepsilon_{i}\right)^{-1}=\frac{\alpha}{2}\sum_{k=0}^{ \infty}\left[\frac{\alpha}{2}\left(\hat{V}-\varepsilon_{i}\right)\right]^{k}\,. \tag{13}\]
Truncation after first order yields the Pauli approximation:
\[\chi_{i}(\vec{r})=\left[\frac{\alpha}{2}+\frac{\alpha^{3}}{4}\left(\hat{V}- \varepsilon_{i}\right)\right]\mathbf{\tilde{\sigma}}\cdot\hat{\vec{p}}\,\varphi_ {i}(\vec{r}). \tag{14}\]
In a molecule, the expectation value of \(\mathbf{\gamma}^{5}\) for a single Slater determinant is determined by a summation over contributions from all occupied molecular orbitals \(i\):
\[\left\langle\psi_{i}\,\middle|\,\mathbf{\gamma}^{5}\,\middle|\,\psi_{i}\right\rangle =\left\langle\varphi_{i}\,\middle|\,\chi_{i}\right\rangle+\left\langle\chi_{i }\,\middle|\,\varphi_{i}\right\rangle \tag{15}\]
Insertion of the first term of the expansion in eq. gives the first order contribution to \(\mathbf{\gamma}^{5}\):
\[\left\langle\psi_{i}\,\middle|\,\mathbf{\gamma}^{5}\,\middle|\,\psi_{i}\right\rangle \approx\alpha\left\langle\varphi_{i}\,\middle|\,\mathbf{\tilde{\sigma}}\cdot\hat{ \vec{p}}\,\middle|\,\varphi_{i}\right\rangle \tag{16}\]
This obviously vanishes if the overall electron density of the molecule is non-helical, but can, in the static case and when remaining in first order with respect to \(\mathcal{P}\)-odd operators, only be non-zero for a chiral molecule, in which the electron density can have non-vanishing helicity.
In order to determine scaling laws with respect to the nuclear charge number \(Z\) and the fine-structure constant \(\alpha\), eq. itself is not immediately useful. This is why we follow Ref. and write the operator \(\mathbf{\gamma}^{5}\) for electron \(i\) as a commutator:
\[\mathbf{\gamma}^{5}_{i} =\frac{\imath}{c}\left[\hat{H}_{\rm DC},\mathbf{\tilde{\Sigma}}_{i} \cdot\vec{r}_{i}\right]_{-}+2\begin{pmatrix}\mathbf{0}&\hat{\mathbf{k}}_{i}\\ \hat{\mathbf{k}}_{i}&\mathbf{0}\end{pmatrix}, \tag{17}\] \[\hat{\mathbf{k}}_{i} =-(\mathbf{1}_{i}+\mathbf{\tilde{\sigma}}_{i}\cdot\hat{\vec{l}}_{i})\,, \qquad\mathbf{\tilde{\Sigma}}_{i}=\begin{pmatrix}\mathbf{\tilde{\sigma}}_{i}&\mathbf{0}\\ \mathbf{0}&\mathbf{\tilde{\sigma}}_{i}\end{pmatrix}\,. \tag{18}\]
Eigenvalues of the operator \(\hat{\mathbf{K}}=\sum_{i}\hat{\mathbf{k}}_{i}\) in atomic systems correspond to the relativistic quantum numbers \(\varkappa=(\ell-j)(2j+1)\), where \(\ell\) and \(j\) are the orbital and total angular momentum quantum numbers, respectively.
As long as we are interested in expectation values of the operator \(\mathbf{\gamma}^{5}\) on the molecular DHFC-orbitals \(\psi_{i}\), the commutator part in eq. turns to zero. DHFC molecular orbital matrix elements of the second term in eq.
The non-relativistic limit of \(\left\langle\mathbf{\gamma}^{5}\right\rangle\) vanishes as can be shown by insertion of the first term of the expansion in eq.:
\[\left\langle\psi_{i}\,\middle|\,\mathbf{\gamma}^{5}\,\middle|\,\psi_{i}\right\rangle \approx\alpha\left\langle\varphi_{i}\,\middle|\,\left\{\hat{\mathbf{k}},\mathbf{ \tilde{\sigma}}\cdot\hat{\vec{p}}\right\}_{+}\,\middle|\,\varphi_{i}\right\rangle =0\,, \tag{20}\]where we use the fact that operator \(\hat{\mathbf{k}}\) anti-commutes with \(\vec{\mathbf{\sigma}}\cdot\hat{\vec{p}}\):
\[\left\{\hat{\mathbf{k}},\vec{\mathbf{\sigma}}\cdot\hat{\vec{p}}\right\}_{+}=0. \tag{21}\]
The terms of order \(\alpha^{3}\) give:
\[\left\langle\psi_{i}\left|\,\mathbf{\gamma}^{5}\,\right|\psi_{i}\right\rangle\approx \frac{\alpha^{3}}{2}\left\langle\varphi_{i}\left|\,(\vec{\mathbf{\sigma}}\cdot\hat{ \vec{p}})\hat{V}\hat{\mathbf{k}}+\hat{\mathbf{k}}\hat{V}(\vec{\mathbf{\sigma}}\cdot \hat{\vec{p}})\,\right|\varphi_{i}\right\rangle\,, \tag{22}\]
Equation can be rewritten as:
\[\left\langle\psi_{i}\left|\,\mathbf{\gamma}^{5}\,\right|\psi_{i}\right\rangle \approx\frac{\alpha^{3}}{2}\left\langle\varphi_{i}\left|\,\left[\vec{\mathbf{ \sigma}}\cdot\hat{\vec{p}},\hat{V}(\vec{r})\right]_{-}\hat{\mathbf{k}}+\hat{V }(\vec{r})\left\{\hat{\mathbf{k}},\vec{\mathbf{\sigma}}\cdot\hat{\vec{p}}\right\} _{+}+\left[\hat{\mathbf{k}},\hat{V}(\vec{r})\right]_{-}\vec{\mathbf{\sigma}}\cdot \hat{\vec{p}}\right|\varphi_{i}\right\rangle\\ =\frac{\alpha^{3}}{2}\left\langle\varphi_{i}\left|\,\left[\vec{\bm {\sigma}}\cdot\hat{\vec{p}},\hat{V}(\vec{r})\right]_{-}\hat{\mathbf{k}}+\left[ \hat{\mathbf{k}},\hat{V}(\vec{r})\right]_{-}\vec{\mathbf{\sigma}}\cdot\hat{\vec{p} }\right|\varphi_{i}\right\rangle, \tag{23}\]
In general, the molecular potential energy operator \(\hat{V}\) does not commute with both operators \(\hat{\mathbf{k}}\) and \((\vec{\mathbf{\sigma}}\cdot\hat{\vec{p}})\). However, its spherically symmetric part \(\hat{V}_{\mathrm{s}}(|\vec{r}|)\) commutes with the operator \(\hat{\mathbf{k}}\). Therefore, for the spherically symmetric potential the last term in eq. turns to zero. Let us separate the contribution of \(\hat{V}_{\mathrm{s}}(|\vec{r}|)\):
\[\left\langle\psi_{i}\left|\,\mathbf{\gamma}^{5}\,\right|\psi_{i}\right\rangle =\left\langle\psi_{i}\left|\,\mathbf{\gamma}^{5}\,\right|\psi_{i} \right\rangle_{\mathrm{s}}+\left\langle\psi_{i}\left|\,\mathbf{\gamma}^{5}\, \right|\psi_{i}\right\rangle_{\mathrm{a}}\,, \tag{24}\] \[\left\langle\psi_{i}\left|\,\mathbf{\gamma}^{5}\,\right|\psi_{i} \right\rangle_{\mathrm{s}} =\frac{\alpha^{3}}{2}\left\langle\varphi_{i}\left|\,\to\!\!\left( \vec{\mathbf{\sigma}}\cdot\vec{r}\right)\!\frac{\hat{V}_{\mathrm{s}}^{\prime}(| \vec{r}|)}{|\vec{r}|}\hat{\mathbf{k}}\,\right|\varphi_{i}\right\rangle \tag{25}\]
Note that \(\hat{V}_{\mathrm{s}}^{\prime}(|\vec{r}|)/|\vec{r}|\) commutes with both operators \(\vec{\mathbf{\sigma}}\cdot\vec{r}\) and \(\hat{\mathbf{k}}\). By analogy with we can assume that \(\left\{\vec{\mathbf{\sigma}}\cdot\vec{r},\hat{\mathbf{k}}\right\}_{+}=0\). Thus, we can write:
\[\imath(\vec{\mathbf{\sigma}}\cdot\vec{r})\hat{\mathbf{k}}=\frac{\imath}{2}\left[ \vec{\mathbf{\sigma}}\cdot\vec{r},\hat{\mathbf{k}}\right]_{-}\,, \tag{26}\]
which proves that the operator in is hermitian, and allows to rewrite this expression as:
\[\left\langle\psi_{i}\left|\,\mathbf{\gamma}^{5}\,\right|\psi_{i}\right\rangle_{ \mathrm{s}} =\frac{\alpha^{3}}{4}\left\langle\varphi_{i}\left|\,\vec{\mathbf{ \sigma}}\cdot\vec{v}_{\mathcal{T},s}\,\right|\varphi_{i}\right\rangle, \tag{27}\] \[\vec{v}_{\mathcal{T},\mathrm{s}} =\frac{\hat{V}_{\mathrm{s}}^{\prime}(|\vec{r}|)}{|\vec{r}|}\left( |\vec{r}|^{2}\hat{\vec{p}}-\vec{r}(\hat{\vec{p}}\cdot\vec{r})\right). \tag{28}\]
We see that expectation value has the form of a scalar product of the spin with an electronic orbital \(\mathcal{T}\)-odd vector \(\vec{v}_{\mathcal{T},\mathrm{s}}\). Molecular matrix elements of \(\vec{\mathbf{\sigma}}\cdot\vec{v}_{\mathcal{T}}\) turn to zero in the non-relativistic approximation for two reasons: (i) for a singlet state an expectation value of the spin is zero; (ii) matrix elements of orbital \(\mathcal{T}\)-odd vectors are imaginary, so their expectation values are zero. In order to get a non-zero expectation value of such operators one needs to include spin-orbit interactions \(\hat{H}_{\mathrm{so}}\), which mix singlet and triplet molecular states and have imaginary matrix elements. Therefore, the energy shift \(\delta E_{\gamma^{5},\mathrm{s}}\) of the molecular (ground) singlet state due to the interaction \(\vec{\mathbf{\sigma}}\cdot\vec{v}_{\mathcal{T},\mathrm{s}}\) appears in double perturbation theory as:
\[\delta E_{\gamma^{5},\mathrm{s}}=\frac{\alpha^{3}}{2}\frac{\mathfrak{Re} \left\{\left\langle\Psi_{\mathrm{s}}\left|\,\vec{\mathbf{\sigma}}\cdot\vec{v}_{ \mathcal{T},\mathrm{s}}\,\right|\Psi_{\mathrm{t}}\right\rangle\left\langle\Psi_ {\mathrm{t}}\left|\,\hat{H}_{\mathrm{so}}\,\right|\Psi_{\mathrm{s}}\right\rangle \right\}}{E_{\mathrm{s}}-E_{\mathrm{t}}}\,. \tag{29}\]
Equation allows to estimate the scaling law for \(\delta E_{\gamma^{5},\mathrm{s}}\) with the nuclear charge \(Z\) and the fine structure constant \(\alpha\). The matrix element of the spin-orbit interaction \(\left\langle\psi_{\mathrm{t}}\left|\,\hat{H}_{\mathrm{so}}\,\right|\psi_{ \mathrm{s}}\right\rangle\) scales as \(\alpha^{2}Z^{2}\). The \(Z\) scaling of the matrix element of the operator \(\vec{v}_{\mathcal{T},\mathrm{s}}\) depends on the distances where the integral is accumulated. Taking into account that this operator appears in third order in \(\alpha\), we can assume that the integral is accumulated at short distances near the nucleus, where relativistic corrections are larger. At such distances the potential of the nucleus is practically unscreened, \(\hat{V}_{\mathrm{s}}\sim Z/r\). Furthermore, at these distances the electron moves \(Z\) times faster, so \(\hat{\vec{p}}\sim Z\). Therefore, we can assume that \(\int v_{\mathcal{T},\mathrm{s}}d^{3}r\sim Z^{2}\). Then the overall scaling is:
\[\delta E_{\gamma^{5},\mathrm{s}}\sim\alpha^{5}Z^{4}\,. \tag{30}\]
The last expression does not take into account "the single center theorem", which implies that electron helicity in molecules is suppressed in the vicinity of a single heavy nucleus and one has to take two matrix elements of expression at two different heavy centers. Therefore, the final scaling should be:
\[\delta E_{\gamma^{5},\mathrm{s}}\sim\alpha^{5}Z_{A}^{2}Z_{B}^{2}\,, \tag{31}\]where \(A\) and \(B\) are typically taken as the two heaviest atoms in the molecule.
Now let us analyze the second term in eq.. In this case both terms from eq. can contribute. For the first term we can use the same arguments as above, but the asymmetric part of the molecular potential at short distances is much weaker, so this term will add small corrections to eq.. Thus, we will focus on the second term, which was zero for the symmetric potential.
We assume again that the matrix element is accumulated at short distances, where the molecular potential can be expanded in spherical harmonics. The second term of this expansion can be written as \((\vec{a}\cdot\vec{r})\hat{V}_{\rm a}(|\vec{r}|)\), where \(\vec{a}\) is some constant polar vector. In this approximation we get:
\[\left[\hat{\bf k},\hat{V}(|\vec{r}|)\right]_{-}=-\imath(\vec{\mathbf{\sigma}} \cdot(\vec{r}\times\vec{a}))\hat{V}_{\rm a}(|\vec{r}|)\,, \tag{32}\]
Substituting this into the second term in eq. we find that:
\[\left\langle\psi_{i}\,\right|\mathbf{\gamma}^{5}\left|\,\psi_{i}\right\rangle_{ \rm a}\\ \approx\frac{\alpha^{3}}{2}\left\langle\varphi_{i}\,\right|- \imath(\vec{\mathbf{\sigma}}\cdot\vec{r}\times\vec{a})\hat{V}_{\rm a}(|\vec{r}|)( \vec{\mathbf{\sigma}}\cdot\hat{\vec{p}})\left|\,\varphi_{i}\right\rangle. \tag{33}\]
Simplifying this further and neglecting the term, which is similar to, we get:
\[\left\langle\psi_{i}\,\right|\mathbf{\gamma}^{5}\left|\,\psi_{i} \right\rangle_{\rm a} \approx\frac{\alpha^{3}}{2}\left\langle\varphi_{i}\,|\,\vec{a} \cdot\vec{v}_{\rm a}\,|\,\varphi_{i}\right\rangle, \tag{34}\] \[\vec{v}_{\rm a} =2\hat{V}_{\rm a}(|\vec{r}|)\,\vec{r}\times\vec{\nabla}\,. \tag{35}\]
The orbital pseudovector \(\vec{v}_{\rm a}\) is \(\mathcal{T}\)-even. The expected scaling with \(\alpha\) is given by eq.. Scaling with \(Z\) for operators and should be similar, so we assume:
\[\delta E_{\mathbf{\gamma}^{5},{\rm a}}\sim\alpha^{3}Z^{2}\,. \tag{36}\]
Combining the two terms in eq. together suggests an estimate for a molecule with two heavy atoms \(A\) and \(B\):
\[\delta E_{\mathbf{\gamma}^{5}}\approx c_{1}\alpha^{5}Z_{A}^{2}Z_{B}^{2}+c_{2} \alpha^{3}Z_{A}^{2}+c_{3}\alpha^{3}Z_{B}^{2}\,. \tag{37}\]
The first term is formed on both heavy centers, while the other two terms are formed independently in the vicinity of each heavy nucleus. The chiral structure of the molecule is weakly felt locally, so we can expect that \(|c_{2,3}|\ll|c_{1}|\).
In the following we discuss the implications in molecular systems of the equation derived above for \(\left\langle\mathbf{\gamma}^{5}\right\rangle\) and compare to results from numerical computations. Hereby, we focus on scaling with respect to the nuclear charge number and the fine structure constant. Furthermore, we compare to energy shifts due to nuclear spin-independent electroweak neutral-current interactions.
## III Computational details
Quasi-relativistic two-component calculations of H\({}_{2}\)X\({}_{2}\) and H\({}_{2}\)XO with X = O, S, Se, Te, Po and CHBrClF are performed within the zeroth order regular approximation (ZORA) at the level of complex generalized Hartree-Fock (cGHF) or Kohn-Sham (cGKS) with a modified version of the quantum chemistry program package Turbomole.
For calculations of H\({}_{2}\)X\({}_{2}\) and H\({}_{2}\)XO compounds a basis set of 25 s, 25 p, 14 d and 11 f uncontracted Gaussian functions with the exponential coefficients \(\alpha_{i}\) composed as an even-tempered series by \(\alpha_{i}=a\cdot b^{N-i};\ i=1,\ldots,N\) with \(a=0.02\,a_{0}{}^{-2}\), \(b=(5/2\times 10^{10})^{1/25}\approx 2.606\) and \(N=26\) was used for X = O, S, Se, Te, Po. The largest exponent coefficients of the s, p, d and f subsets are \(5\times 10^{8}\ a_{0}^{-2}\), \(1.91890027\times 10^{8}\ a_{0}^{-2}\), \(13300.758\ a_{0}^{-2}\) and \(751.8368350\ a_{0}^{-2}\), respectively. A similar but slightly smaller basis set (three f functions less) has proven successful in calculations of parity violating energy shifts in H\({}_{2}\)Po\({}_{2}\). The H atom was represented with the s,p-subset of a decontracted correlation-consistent basis of quadruple-\(\zeta\) quality.
Structure parameters of H\({}_{2}\)X\({}_{2}\) were chosen as in Refs.. For H\({}_{2}\)XO compounds the equilibrium bond-length of the O-X bond, for X = S, Se, Te, Po was obtained by full structure optimization at the level of GHF-ZORA. As convergence criteria an energy change of less than \(10^{-5}\)\(E_{\rm h}\) was used. Bond angles H-O-X and bond distances H-O of H\({}_{2}\)XO were assumed to be equal to H\({}_{2}\)O\({}_{2}\) and bond angles H-X-O and distances H-X were assumed to be equal to H\({}_{2}\)X\({}_{2}\). Employed structure parameters are summarized in Table 1.
Structure parameters, harmonic vibrational wave numbers and normal coordinates, of CHBrClF, as well as electronic densities and vibrational wave functions along the C-F stretching mode were employed as described in Ref.. Electronic densities along other normal coordinates were calculated on the level of ZORA-cGHF and ZORA-cGKS with the same basis set employed in Ref.. Properties were calculated on the levels of ZORA-cGHF and ZORA-cGKS. Used density functionals are the local density approximation (LDA) and the Lee, Yang and Parr correlation functional (LYP) with a generalized gradient exchange functional by Becke (BLYP) or the hybrid Becke three parameter exchange functional (B3LYP).
The ZORA-model potential \(\tilde{V}(\vec{r})\) as proposed by van Wullen was employed with additional damping.
For calculations of two-component wave functions and properties a finite nucleus was used, described by a normalized spherical Gaussian nuclear density distribution \(\rho_{\rm nuc,}A(\vec{r})=\frac{\zeta_{A}^{3/2}}{\pi^{3/2}}{\rm e}^{-\zeta_{A}| \vec{r}-\vec{r}_{A}|^{2}}\), where \(\zeta_{A}=\frac{3}{2r_{\rm nuc,}A}\) and the root mean square radius \(r_{\rm nuc,}A\) of nucleus \(A\) was used as suggested by Visscher and Dyall. The mass numbers \(A\) were chosen to correspond to the isotopes \({}^{1}\)H, \({}^{12}\)C, \({}^{16}\)O, \({}^{19}\)F, \({}^{32}\)S, \({}^{35}\)Cl, \({}^{79}\)Br, \({}^{80}\)Se, \({}^{130}\)Te, \({}^{209}\)Po. The weak nuclear charges \(Q_{\mathrm{W},A}\) of the various isotopes with charge number \(Z_{A}\) and neutron number \(N_{A}\) were included as \(Q_{\mathrm{W},A}\approx(1-4\sin^{2}\theta_{\mathrm{W}})Z_{A}-N_{A}\), where we have used \(\sin^{2}\theta_{\mathrm{W}}=0.2319\) as the numerical value of the Weinberg parameter.
All relativistic expectation values of \(\mathbf{\gamma}^{5}\) and \(\hat{H}_{\mathrm{ew}}\) were calculated with our ZORA property toolbox approach described in Ref..
## IV Results
### Scaling laws for \(\left<\mathbf{\gamma}^{5}\right>\) in molecules
In order to confirm results of section II.2 we performed quasi-relativistic numerical calculations at the level of ZORA of (\(P\))-enantiomers of H\({}_{2}\)X\({}_{2}\) compounds with an dihedral angle of 45\({}^{\circ}\), varying X = O, S, Se, Te, Po. These compounds are established as a common test system for electroweak parity violation and its scaling behavior with respect to nuclear charge. In the above scaling law a factor of \(\alpha^{2}Z_{B}^{2}\) emerges from spin-orbit coupling. This factor is in good approximation equal to \(\alpha^{2}\) in main group element containing molecules with only one heavy center (see e.g. Refs.). Therefore, for a variation of one heavy X atom while holding the other one fixed as oxygen atom (H\({}_{2}\)XO) we would expect roughly a scaling of \(\sim\alpha^{3}Z_{A}^{2}\) (corresponding to the second term in eq.) as the spin-orbit coupling contribution (corresponding to the first term in eq.) is suppressed by a factor of \(\alpha^{2}\).
The numerical results are summarized in Table 2 and Table 3. shows a double logarithmic plot and a linear fit for the determination of the \(Z\)-scaling law in ZORA-cGHF calculations. From numerical calculations of H\({}_{2}\)X\({}_{2}\) compounds we find a \(Z\)-scaling with \(Z^{4.4}\), which agrees well with the analytical prediction. Furthermore for H\({}_{2}\)XO compounds we find a scaling of \(Z^{2.1}\), which is in perfect agreement with the expectations above and shows the missing spin-orbit coupling contribution as the nuclear charge of oxygen is close to 1.
In order to test the predicted \(\alpha\)-dependence the speed of light was varied in the quasi-relativistic calculations of wave functions and properties for H\({}_{2}\)PoO and H\({}_{2}\)Po\({}_{2}\). The results show the expected scaling of \(\alpha^{5.4}\approx\alpha^{5}\) for H\({}_{2}\)Po\({}_{2}\) and a scaling of \(\alpha^{3.6}\) for H\({}_{2}\)PoO showing the weak influence of spin-orbit coupling in compounds with only one heavy nucleus. The results are in perfect agreement with the analytical analysis.
### Comparison to electroweak electron-nucleon interactions
Similar considerations, as detailed in the previous section, are known to hold also for parity non-conserving nuclear spin-independent electroweak interactions described by Hamiltonian in chiral molecules. The main difference of this Hamiltonian to the ones discussed in the theory section is that \(\hat{H}_{\mathrm{ew}}\) evaluates the expectation value of \(\mathbf{\gamma}^{5}\) at positions inside the nuclei only. To further compare \(\hat{H}_{\mathrm{ew}}\) with \(\mathbf{\gamma}^{5}\) we evaluated the dependence of the expectation value of both operators on the dihedral angle in H\({}_{2}\)X\({}_{2}\) for X = O and Po, and found similar behavior (see and for the explicit data see the Supplement). It shall be noted, that the sign of \(\hat{H}_{\mathrm{ew}}\) is inverted in comparison to \(\mathbf{\gamma}^{5}\) as \(\hat{H}_{\mathrm{ew}}\) contains in addition the weak charge for which \(Q_{\mathrm{W}}\approx-N<0\).
In a recent work, similar calculations on \(\mathbf{\gamma}^{5}\) in H\({}_{2}\)X\({}_{2}\) compounds were performed and similar results were obtained. However, unfortunately, in Ref. insufficient basis sets for oxygen were employed resulting in qualitatively wrong results for the dihedral angle dependence in H\({}_{2}\)O\({}_{2}\).
The similar dependence on the molecular structure together with the steep scaling with nuclear charge indicates that contributions at the nuclear centers dominate also the expectation value of \(\mathbf{\gamma}^{5}\) and, thus, imply that molecular experiments that aim to test parity violation due to weak interactions can also be used for searches of parity violating cosmic fields with a comparable sensitivity. This aspect will be discussed in the following in detail.
### Limits on cosmic fields from experiments with chiral molecules
#### iv.3.1 Test system and choice of methods
The expected sensitivity of experiments with chiral molecules to \(\mathcal{P}\)-odd cosmic fields characterized by \(b_{0}^{e}\) is estimated from an experiment with CHBrClF performed by Daussy _et. al._, in which a hyperfine component of the \(40_{7,34}\gets 40_{8,33}\) transition (\(J^{\prime}_{K^{\prime}_{4},K^{\prime}_{5}}\gets J^{\prime\prime}_{K^{ \prime\prime}_{4},K^{\prime\prime}_{5}}\)) of the C-F stretching fundamental in enantiomerically enriched samples of the mirror images \(R\)-CHBrClF and \(S\)-CHBrClF was studied.
Our interest is in a possible splitting of the vibrational resonance frequency between enantiomers that is caused by cosmic fields interacting through \(\left<\mathbf{\gamma}^{5}\right>\). For this purpose frequency shifts in the vibrational spectrum due to electronic interactions via \(\mathbf{\gamma}^{5}\) have to be evaluated. This test system, CHBrClF, was excessively studied by theory and experiment and is supposed to be reasonably well understood with respect to electroweak parity violation.
However, the influence from non-separable anharmonic effects (multimode effects) on electroweak parity violation in CHBrClF is largely unexplored. Quack and Stohner studied the deuterated isotopomer CDBrClF with respect to multimode contributions in a four-dimensional, anharmonically treated subspace involving the C-F stretch, C-D stretch and the two C-D bending modes to find an increase of the parity-violating frequency splitting in the C-F stretch fundamental \(\nu_{4}\) by al most a factor of two -- depending on the specific model, they obtained up to about 75 % relative deviation with respect to the separable anharmonic adiabatic approximation. Although not directly comparable due to the different isotope, this at least suggests that pronounced multimode effects can also exist for \(\left\langle\mathbf{\gamma}^{5}\right\rangle\).
We have reported major findings and implications for future experiments in a separate letter, but provide herein more details on the computational challenges and subsequent analysis.
We estimate the influence of multimode effects within a perturbative treatment by calculation of derivatives of the property of interest with respect to all normal coordinates. One-dimensional and two-dimensional vibrational corrections to a property \(O\) for a single dimensionless reduced normal coordinate \(q_{r}\) are in leading order given by:
\[O_{q_{r}}^{\text{1D}} \approx\frac{1}{2}\left(v_{r}+\frac{1}{2}\right)\left(\frac{ \partial^{2}O_{0}}{\partial q_{r}^{2}}-\frac{\phi_{rrr}}{\tilde{\nu}_{r}} \frac{\partial O_{0}}{\partial q_{r}}\right) \tag{38}\] \[O_{q_{r}}^{\text{2D}} \approx-\frac{1}{2}\left(v_{r}+\frac{1}{2}\right)\sum_{s\neq r} \frac{\phi_{rrs}}{\tilde{\nu}_{s}}\frac{\partial O_{0}}{\partial q_{s}}, \tag{39}\]
Properties are evaluated along the dimensionless reduced normal coordinate \(q_{r}\) and fitted to a polynomial of degree 4:
\[\left\langle\psi_{\text{e}}\left|\,\hat{H}_{\text{ew}}\left|\, \psi_{\text{e}}\right\rangle_{r}\approx\sum_{k=0}^{4}c_{\text{ew},r,k}q_{r}^{k}\right. \tag{40}\] \[\left\langle\psi_{\text{e}}\left|\,\mathbf{\gamma}^{5}\left|\,\psi_{ \text{e}}\right\rangle_{r}\approx\sum_{k=0}^{4}c_{\mathbf{\gamma}^{5},r,k}q_{i}^{k }\,.\right. \tag{41}\]
In the dependence of \(\left\langle\mathbf{\gamma}^{5}\right\rangle\) and \(\left\langle\hat{H}_{\text{ew}}\right\rangle\) on the normal coordinates for the different methods in the region \(q_{r}=-3,\ldots,3\) (for the explicit data see the Supplement). Within this region the probability density of the first two vibrational states in the mode \(q_{4}\) is sufficiently decayed (see of Ref.), as can also be expected by considering classical turning points of a harmonic approximation to the parity-conserving potential, which are located at \(\left|q_{4}\right|=1\) for the ground vibrational state of a harmonic oscillator and at \(\left|q_{4}\right|=\sqrt{3}\) in the first vibrationally excited state. The resulting fit parameters \(c_{\mathbf{\gamma}^{5},r,k}\) alongside the explicit values for the one-dimensional cuts through the hypersurface for all normal coordinates \(q_{r}\) are reported in the Supplement.
The derivatives of the properties with respect to the normal coordinate \(q_{r}\) are given by
\[\frac{\partial\left\langle\psi_{\text{e}}\left|\,\mathbf{\gamma}^{5} \left|\,\psi_{\text{e}}\right\rangle_{r}\right.}{\partial q_{r}}=c_{\mathbf{\gamma }^{5},r,1} \tag{42}\] \[\frac{\partial^{2}\left\langle\psi_{\text{e}}\left|\,\mathbf{\gamma} ^{5}\left|\,\psi_{\text{e}}\right\rangle_{r}\right.}{\partial q_{r}^{2}}=2c_{ \mathbf{\gamma}^{5},r,2}\,, \tag{43}\]
Resulting first and second derivatives from the fit in are listed in Table 5 and Table 6. From these we see that the C-F stretching mode has a weak influence on \(\left\langle\mathbf{\gamma}^{5}\right\rangle\) in comparison to the other modes and, thus, is not an optimal choice for an experiment. In particular along the deformation normal coordinates \(q_{9}\) (Br-Cl), \(q_{8}\) (Br-F), \(q_{3}\) (H) and \(q_{2}\) (H) the first derivatives of \(\left\langle\mathbf{\gamma}^{5}\right\rangle\) are considerably larger in magnitude than for \(q_{4}\). The second derivatives with respect to the C-F stretching coordinate are smaller in absolute value than those first derivatives mentioned, by about an order of magnitude (see Table 5 and Table 6). We may assume that anharmonic constants can be roughly of the order \(\phi_{rrr}\sim\mathcal{O}(0.1\tilde{\nu}_{r})\) and \(\phi_{rrs}\sim\mathcal{O}(0.01\tilde{\nu}_{s})\) or even larger (see e.g. Ref. for some cubic force constants in CDBrClF). In total, two-dimensional effects on the C-F stretching mode for \(\left\langle\mathbf{\gamma}^{5}\right\rangle\) can be on the same order as one-dimensional vibrational effects. Thus not only the effect of parity violating interactions on the C-F stretching mode is very weak, but also the theoretical description is limited by the need of an excellent description of all modes, which is exceedingly difficult.
It is important to note, that the use of a different vibrational mode (such as Br-F (\(v_{8}\)) or H (\(v_{3}\)) deformation) in CHBrClF can result in vibrational frequency splittings that are larger by about an order of magnitude and may reduce error bars considerably. This has to be analyzed in more detail, however, using anharmonic vibrational force fields.
Due to the resulting large error bars for vibrational corrections for the C-F stretching mode we do not provide a final value for the enhancement of \(b_{0}^{*}\) in the C-F stretching but rather give an order of magnitude estimate.
For this purpose, within the separable anharmonic adiabatic approximation as described in Ref., where we follow for this specific application Ref. closely, the vibrationally averaged expectation value for the C-F stretching mode is evaluated from a series expansion in the vibrational moments \(\left\langle v\left|\,q^{k}\left|\,v\right.\right\rangle\right.\), where \(v\) represents the vibrational quantum number of the \(v\)th vibrational state. The vibrational wave functions and corresponding moments were received in Ref. from a discrete variable representation on an equidistant grid. The moments were reported in the supplementary material to Ref. and are reused for calculating interactions of CHBrClF with cosmic fields.
In order to estimate electron correlation effects, for the C-F stretching mode the vibrationally averaged expectation values where evaluated at the DFT and HF level, the former with different flavors of density functionals. The results of these methods are compared in Table 4.
In previous studies on electroweak parity-violating vibrational frequency splittings in CHBrClF with density functional approaches much reduced variations between the methods were found for the C-F stretching fundamental as can be expected by the nearly parallel curves shown in In Ref. we have observed a spread of about 20 % from the mean value for the four methods used also in the present work. The variation amongst the various density functionals (B3LYP, BLYP and LDA) was below 5 %. In Ref. it was found that B3LYP, BLYP and LDA estimates deviate by 6 % or less from the values predicted on the second order many-body perturbation theory level (MP2), with the latter method giving also absolute values at the equilibrium structure that agree well with the corresponding CCSD(T) estimates. Hartree-Fock based predictions, in contrast, displayed larger deviations from those of the mentioned density functional calculations. Similar trends are observed in the present work (see Table 4), but with more pronounced variations for the structure dependence of \(\left\langle\mathbf{\gamma}^{5}\right\rangle\) as compared to \(\left\langle\hat{H}_{\mathrm{ew}}\right\rangle\): Vibrational splittings vary by about 50 % from the mean value of all four methods, with variations amongst the density functionals being on the order of 25 % or less from their mean. Assuming again that the density functionals outperform the Hartree-Fock approach for this property and give again similar results as MP2, we are lead to a rough error estimate of about 30 % for the density functionals. Of the different functionals, we give herein tentative preference to the B3LYP results as i) the absolute values at the equilibrium structures for electroweak parity violation were for B3LYP closer to the MP2 and CCSD(T) values, ii) the atomic contributions studied in Refs., which are differently weighted by \(\left\langle\mathbf{\gamma}^{5}\right\rangle\) as compared to \(\left\langle\hat{H}_{\mathrm{ew}}\right\rangle\), were found to be more consistent with MP2 and CCSD(T) values and iii) the vibrational splitting on the B3LYP level is smaller than for the other functionals, which results in more conservative sensitivity estimates.
#### iv.2.2 Sensitivity to static cosmic fields
The expectation values of \(\mathbf{\gamma}^{5}\) and splittings between enantiomers are given in Table 4. As discussed above, we expect multimode effects of the same size as single-mode effects and at the present stage are not able to set upper bounds on \(b_{0}^{e}\) from the CHBrClF experiment. In Ref. we rather estimated the sensitivity of this experiment. Assuming B3LYP to give the best performance (see discussion above) \(\Delta_{(R,S)}\left\langle\mathbf{\gamma}^{5}\right\rangle\) is on the order of \(10^{-10}\) (\(\mathcal{O}(10^{-10})\)).
The sensitivity of the CHBrClF experiment, performed by Daussy et al. in 1999, to \(b_{0}^{e}\) was in Ref. estimated from the experimental upper bound of the parity violating frequency splitting in the C-F stretching fundamental \(\left|\Delta\nu\right|=12.7\,\mathrm{Hz}\) as:
\[\left|b_{0}^{e}\right|\lesssim\left|\frac{12.7\,\mathrm{Hz}}{\mathcal{O}(10^{ -10})}h\right|\sim\mathcal{O}(10^{-12}\,\mathrm{GeV}) \tag{44}\]
In comparison to the actual best direct limits on \(b_{0}^{e}\) from modern atomic experiments, that are \(2\times 10^{-14}\,\mathrm{GeV}\) from Cs and \(7\times 10^{-15}\,\mathrm{GeV}\) from Dy, the 1999 CHBrClF experiment is less sensitive by about two orders of magnitude. However, it is as sensitive as atomic experiments with Tl and Yb (\(\left|b_{0}^{e}\right|<2\times 10^{-12}\,\mathrm{GeV}\), see Ref.).
As emphasized in the discussion of multimode effects the sensitivity of future experiments can be increased by an order of magnitude, when choosing favorable vibrational transitions. As we pointed out in Ref., it was emphasized in Refs. that the sensitivity of the experiment discussed above is improvable by at least two orders of magnitude by experimental refinement. A choice of a more favorable molecule is expected to lead to further enhancement by two orders of magnitude. Thus it was estimated in Ref. that in future \(\mathcal{P}\)-violation experiments with chiral molecules the limits from the 1999 experiment can be improved down to \(10^{-17}\,\mathrm{GeV}\), i.e. an improvement of the actual best limit by at least two orders of magnitude. This makes experiments with chiral molecules highly powerful tools to search for Lorentz invariance violation beyond the Standard Model of particle physics.
The accuracy of the estimate for cosmic field effects in CHBrClF, which was in this work indirectly inferred by comparison to previous studies on electroweak parity violation, can in principle be benchmarked by future explicit calculations with systematically improvable electron correlation methods and the presently neglected multi-mode contributions can be accounted for by explicit calculation of anharmonicity constants. As the main purpose of the present studies was to explore the general potential of chiral molecules to act as sensitive probes for new physics, more accurate theoretical estimates specifically for CHBrClF do not seem to be pressing until new experiments with higher accuracy are performed. Given the pronounced scaling with nuclear charge that was shown analytically and confirmed numerically in this paper, the main focus will likely be shifted to accurate estimates for chiral compounds with heavier elements. Furthermore, our study showed that care has to be taken by choice of the vibrational mode, which on the one hand can directly influence the sensitivity by an order of magnitude and on the other hand can be crucial for accurate theoretical predictions, which are essential to provide limits on cosmic fields from experiments.
## V Conclusion and outlook
In this paper we have shown that interactions of electrons with the timelike-component of pseudovector cosmic fields are strongly pronounced in chiral molecules. Due to the \(\mathcal{P}\)-odd contributions of the nuclear potential, that electrons experience in a chiral molecule, these interactions lead to \(\mathcal{P}\)-odd resonance frequency splittings between enantiomers, similar to those from electroweak parity-violating interactions. We could show analytically and numerically that these interactions are strongly enhanced in heavy element containing molecules and are dominated from contributions that stem from the region near the nucleus. It was demonstrated that \(\mathcal{P}\)-odd interactions of electrons with cosmic fields show similar behavior to interactions due to electroweak coupling of electrons and nucleons in chiral molecules. Thus, knowledge from electroweak quantum chemistry can be employed to find promising candidate molecules to limit \(\mathcal{P}\)-odd electronic coupling to cosmic fields. However, care has to be taken as our calculations revealed a stronger dependence of \(\mathbf{\gamma}^{5}\) on molecular structure.
We calculated matrix elements of \(\mathcal{P}\)-odd cosmic field interactions in CHBrClF with quasi-relativistic ab initio methods, including vibrational corrections, and compared the results of different DFT functionals. Our calculations of \(\mathcal{P}\)-odd effects along the different normal coordinates of CHBrClF revealed an important role of non-separable anharmonic effects and showed that the C-F stretching mode in particular is from this perspective not ideally suited for a measurement of \(\mathcal{P}\)-violation due to cosmic fields. Effects on some other modes are expected to be larger by an order of magnitude. These findings underline the importance to select not only a favorable molecule, but also to carefully choose the vibrational transition. However, from our calculations the sensitivity of a 20 year old experiment with CHBrClF to \(|b_{0}^{e}|\) was estimated to be \(\mathcal{O}(10^{-12}\,\mathrm{GeV})\). This sensitivity is inferior by two orders to the actual best direct measurements drawn from modern atomic parity violation experiments, but was considered to be improvable to the order of \(\mathcal{O}(10^{-17}\,\mathrm{GeV})\) or better for static pseudovector fields, which would be an improvement of the actually best limit on \(b_{0}^{e}\) by at least two orders of magnitude. This demonstrates the specific virtue that studies on chiral molecules provides in the search for new physics beyond the standard model.
|
10.48550/arXiv.2005.03938
|
Parity nonconserving interactions of electrons in chiral molecules with cosmic fields
|
Konstantin Gaul, Mikhail G. Kozlov, Timur A. Isaev, Robert Berger
| 2,835
|
10.48550_arXiv.1203.6010
|
###### Abstract
The problem of the one dimensional electro-diffusion of ions in a strong binary electrolyte is considered. In such a system the solute dissociates completely into two species of ions with unlike charges. The mathematical description consists of a diffusion equation for each species augmented by transport due to a self consistent electrostatic field determined by the Poisson equation. This mathematical framework also describes other important problems in physics such as electron and hole diffusion across semi-conductor junctions and the diffusion of ions in plasmas. If concentrations do not vary appreciably over distances of the order of the Debye length, the Poisson equation can be replaced by the condition of local charge neutrality first introduced by Planck. It can then be shown that both species diffuse at the same rate with a common diffusivity that is intermediate between that of the slow and fast species (ambipolar diffusion). Here we derive a more general theory by exploiting the ratio of Debye length to a characteristic length scale as a small asymptotic parameter. It is shown that the concentration of either species may be described by a nonlinear integro-differential equation which replaces the classical linear equation for ambipolar diffusion but reduces to it in the appropriate limit. Through numerical integration of the full set of equations it is shown that this nonlinear equation provides a better approximation to the exact solution than the linear equation it replaces.
pacs: 82.45.Gj,82.45.-h,82.70.y,47.57.eb,47.57.jd,72.20.Dp The one dimensional electro-diffusion equations describing the evolution of ionic concentrations \(n(x,t)\), \(p(x,t)\) in a fully dissociated binary electrolyte with a self consistent electrostatic potential \(\phi(x,t)\) may be put in the form
\[\partial_{t}n-\partial_{x}(n\partial_{x}\phi) = \partial_{xx}n \tag{1}\] \[u^{-1}\partial_{t}p-z\partial_{x}(p\partial_{x}\phi) = \partial_{xx}p\] \[\partial_{xx}\phi = -n-zp \tag{3}\]
We will take n as the species with lesser mobility, so that \(u\geq 1\). Since the two species are oppositely charged, \(z\) is negative. The above equations have been rendered dimensionless by scaling the concentrations by the concentration of n at infinitely far points (\(n_{*}\)) so that \(n(\infty,t)=1\), the co-ordinate \(x\) by a length scale \(\lambda_{*}\) related to Debye length, the time by a diffusion time \(\lambda_{*}^{2}/D_{n}\) and the potential \(\phi\) by the thermal energy (\(k_{B}T\))/(\(ez_{n}\)). In the above, \(D_{n}\) and \(D_{p}\) are the diffusivities of the two species which are related to the mobilities \(u_{n},u_{p}\) through the Einstein relation \(D_{n}/u_{n}=D_{p}/u_{p}=k_{B}T\), where \(k_{B}\) is the Boltzmann constant and \(T\) the absolute temperature. The length scale \(\lambda_{*}\) referred to above is defined by the expression \(\lambda_{*}^{2}=(\epsilon_{0}k{k_{B}T})/(z_{n}^{2}e^{2}n_{*})\) in terms of the electronic charge (\(e\)), permittivity of vacuum (\(\epsilon_{0}\)) and dielectric constant (\(\kappa\)). It is related to the Debye length in a homogeneous charge neutral electrolyte where the less mobile species has a dimensionless concentration of \(n\) as \(\lambda_{D}(n)=\lambda_{*}/\sqrt{n(1-z)}\).
Eq.- and its modifications also describe other transport problems of interest in various areas of physics. For example, if \(\phi(x,t)=-E_{0}x+\phi^{\prime}\) where \(E_{0}\) is an external field and \(\phi^{\prime}\) is the perturbation in the potential and if a term \(-N(x)\) representing the density of free charges is added to the right hand side of Eq., then these equations become the Van Roosbroeck equations describing the migration of charge carriers in solid state physics, where \(n\), \(p\) and \(N\) represent the concentrations of electrons, holes and dopants. If \(N(x)\) is interpreted as the density of ion exchange sites within a membrane, then the same set of equations also describes various filtration processes such as electrodialysis. In plasma physics one often encounters situations where the electrons and positive ions may be regarded as two interpenetrating charged fluids of different diffusivities coupled by a self-consistent electric field. It is then described mathematically by Eq.-. The generalization of these equations to three or more species in the presence of an applied electric field \(E_{0}\) describe the separation of charged macromolecules in capillary electrophoresis and other separation processes based on electric charge.
One of the simplest experimental realizations of Eq.- is the "Liquid Junction" problem, where a barrier initially separates two semi-infinite regions of a binary electrolyte (such as sodium chloride in water) of different concentrations, and subsequently, at time \(t=0\), the barrier is removed. Since generally the positive and negative components would diffuse at different rates a polarization vector and consequently a measurable "Liquid Junction Potential" (LJP) appears across the interface. Planck derived an expression for the LJP on the basis of the local charge neutrality assumption, an idea that has since become a central paradigm in many problems involving electro-diffusion. The physical basis of the approximation is that the electrostatic force thatwould arise if a charge separation was realized is so strong that in practice it precludes any departure from electro-neutrality anywhere in space.
Mathematically, the local electro-neutrality assumption amounts to neglecting the term \(\partial_{xx}\phi\) on the left hand side of Eq., so that it reduces to \(p=-n/z\). This relation can now be used to eliminate the terms involving \(\phi\) in Eq. and.
\[\partial_{t}n=D\partial_{xx}n \tag{4}\]
Since \(z<0\), clearly \(u\geq D\geq 1\). The coulomb attraction between the two species enhances the rate of spreading of the slower species and reduces that of the faster species so that both diffuse at an equal, ambipolar (that is, the same for either charge) rate. In the analysis outlined above, though the term in \(\phi\) is dropped in Eq., it must be retained in Eq.-, a situation that appeared self contradictory and led to some controversy until Hickman Hickman; Hickman and Heck provided a proper interpretation within the framework of an asymptotic theory based on expansion in the small parameter \(k\lambda_{D}\), where \(k\) is a characteristic wave number and \(\lambda_{D}\) is the Debye length. Since the Debye length \(\lambda_{D}\) is normally a very small quantity (about 3 nm in a 0.01 M solution of a common salt like sodium or potassium chloride) in most applications the charge neutrality assumption is very accurate. However, it could be violated in many modern applications such as in nanochannels where one or more geometric dimensions may be of the order of nanometers. Such departures from charge neutrality may give rise to new effects. It would therefore seem worthwhile to first study these effects in the context of the simple model system represented by Eq.-.
In this paper we use the method of multiple scales (Heck, 1988) to reduce Eq.- to a nonlinear one dimensional system in the limit of long (compared to the Debye length) length scales and slow (compared to the diffusion time over a Debye length) time scales. We show that at the lowest order the equation for ambipolar diffusion is recovered. If the asymptotic theory is continued to the next order, a nonlinear integro-differential equation for the concentration \(n\) emerges. Numerical solutions of this reduced equation is seen to agree better with that of the full system of Eq.- and lead to effects not captured in the lowest order theory based on local charge neutrality.
We would like to study the behavior of Eq.- at long length and time scales, under the boundary conditions that \(n,p,\phi\) all approach constant values and respect local charge neutrality as \(x\rightarrow\pm\infty\). Thus, we are considering passive diffusion in the absence of imposed electric fields and currents. Following the usual procedure (Heck, 1988), we introduce slow variables \(\xi=\sqrt{\epsilon}\,x\) and \(\tau=\epsilon\,t\) and suppose that \(n,p,\phi\) depend solely on \(\xi\) and \(\tau\). Then Eq.- reduce to
\[\partial_{\tau}n-\partial_{\xi}(n\partial_{\xi}\phi) = \partial_{\xi\xi}n \tag{5}\] \[u^{-1}\partial_{\tau}p-z\partial_{\xi}(p\partial_{\xi}\phi) = \partial_{\xi\xi}p\] \[\epsilon\partial_{\xi\xi}\phi = -n-zp \tag{7}\]
Expanding all dependent variables in asymptotic series in \(\epsilon\), such as \(n=n_{0}+\epsilon n_{1}+\epsilon^{2}n_{2}+\cdots\), substituting in Eq.- and equating the coefficients of \(\epsilon\) on both sides, we get a series of equations starting with the lowest order ones:
\[\partial_{\tau}n_{0}-\partial_{\xi}(n_{0}\partial_{\xi}\phi_{0}) = \partial_{\xi\xi}n_{0} \tag{8}\] \[u^{-1}\partial_{\tau}p_{0}-z\partial_{\xi}(p_{0}\partial_{\xi} \phi_{0}) = \partial_{\xi\xi}p_{0}\] \[-n_{0}-zp_{0} = 0 \tag{10}\]
If the last equation is used to eliminate the second terms from the first two, we get
\[\partial_{\tau}n_{0} = D\partial_{\xi\xi}n_{0} \tag{11}\] \[p_{0} = -n_{0}/z \tag{12}\]
with \(D=u(1-z)/(1-uz)\), representing ambipolar diffusion and if the time derivative terms are eliminated instead and the result integrated, we get
\[\phi_{0}=\frac{u-1}{1-uz}\ln n_{0}. \tag{13}\]
If the above equation is evaluated at \(x=-\infty\), we get Planck's formula (Planck, 1977) for the potential drop across a liquid junction.
\[\partial_{\tau}n_{1}-\partial_{\xi}(n_{1}\partial_{\xi}\phi_{0}) -\partial_{\xi}(n_{0}\partial_{\xi}\phi_{1}) = \partial_{\xi\xi}n_{1} \tag{14}\] \[u^{-1}\partial_{\tau}p_{1}-z\partial_{\xi}(p_{1}\partial_{\xi} \phi_{0})-z\partial_{\xi}(p_{0}\partial_{\xi}\phi_{1}) = \partial_{\xi\xi}p_{1}\] \[\partial_{\xi\xi}\phi_{0}=-n_{1}-zp_{1} \tag{16}\]
If we add Eq. and and use Eq., the term in \(\phi_{1}\) is eliminated. We then obtain after substituting \(\phi_{0}\) from Eq.
\[\alpha = -\frac{uz(u-1)^{2}}{(1-uz)^{3}} \tag{19}\] \[\beta = \frac{u(u-1)(2-z-uz)}{2(1-uz)^{3}} \tag{20}\]
If we add Eq. to \(\epsilon\) times Eq.
\[\partial_{t}n=D\partial_{xx}n+\alpha\partial_{xxxx}(\ln n)-\beta\partial_{xx}( \partial_{x}\ln n)^{2} \tag{21}\]
In arriving at Eq. we have replaced the terms \(\alpha\partial_{\xi\xi\xi\xi}(\ln n_{0})\)by \(\alpha\partial_{\xi\xi\xi\xi}(\ln n)\) and \(-\beta\partial_{\xi\xi}(\partial_{\xi}\ln n_{0})^{2}\) by \(-\beta\partial_{\xi\xi}(\partial_{\xi}\ln n)^{2}\), which is justified because doing so introduces an error in Eq. that is higher than order \(\epsilon^{2}\). Similarly, combining Eq. with Eq.
\[p=-\frac{n}{z}-\frac{u-1}{z(1-uz)}\partial_{xx}(\ln n) \tag{22}\]
The last term in Eq. is due to the departure from charge neutrality.
Let us now consider some of the consequences of Eq. and. First, we consider weak perturbations from the constant values at infinity: \(n=1+n^{\prime}\) and \(p=-1/z+p^{\prime}\), where \(|n^{\prime}|,|p^{\prime}|\ll 1\). This gives a hyper-diffusion equation for \(n^{\prime}\):
\[\partial_{t}n^{\prime}=D\partial_{xx}n^{\prime}+\alpha\partial_{xxxx}n^{ \prime}. \tag{23}\]
Therefore, solutions of the form \(\exp[at+ikx]\) have growth rates \(a=-k^{2}D+k^{4}\alpha\). Since \(\alpha>0\), high wavenumber modes \(k>k_{\rm max}=\sqrt{(D/\alpha)}\) are unstable. We will show that this instability is entirely spurious. Indeed, Eq. is not even valid for modes \(k>k_{max}\sim 1\) since these solutions violate the premise \(k\ll 1\). The instability arises as a consequence of truncating the asymptotic series, as can be seen by considering the linearized version of Eq.- with \(u^{-1}=0\):
\[\partial_{t}n^{\prime}-\partial_{xx}\phi = \partial_{xx}n^{\prime}, \tag{24}\] \[\partial_{xx}\phi = \partial_{xx}p^{\prime},\] \[\partial_{xx}\phi = -n^{\prime}-zp^{\prime}. \tag{26}\]
If Eq. is integrated and the result substituted in Eq. an exact expression for \(\phi\) may be found in terms of \(n^{\prime}\):
\[\phi = -\frac{1}{z}(1+z^{-1}\partial_{xx})^{-1}n^{\prime}, \tag{27}\] \[= -\int_{-\infty}^{+\infty}G(|x-y|;\sqrt{-z})n^{\prime}(y)\,dy,\] \[= -\frac{1}{z}\left[1-z^{-1}\partial_{xx}+z^{-2}\partial_{xxxx}+ \cdots\right]n^{\prime} \tag{29}\]
If the exact inversion, that is Eq., is substituted in Eq.
\[a=-k^{2}\left[1+\frac{1}{k^{2}-z}\right]\sim-k^{2}\left(1-\frac{1}{z}\right)+ \frac{k^{4}}{z^{2}}+\cdots \tag{30}\]
The exact formula for \(a\) indicated by the \(=\) sign above shows that there is no high wavenumber instability if the exact inversion, Eq., is employed. The second part, indicated by the \(\sim\) sign, shows the result of the approximate inversion, Eq., based on the low wavenumber approximation. In this case there is a high wavenumber instability if the asymptotic series is truncated after an even number of terms. That is, the root cause of the instability is the oscillatory approach to the limit of the series indicated on the right hand side of Eq..
The above analysis suggests a simple way of modifying Eq. without violating its asymptotic validity, but at the same time eliminating the spurious high wavenumber instability. We recognize that the term \(\alpha\partial_{xxxx}(\ln n)=\partial_{xx}(\alpha\partial_{xx}\ln n)\) in Eq. most likely resulted from a truncated expansion of the Helmholtz operator: \(\partial_{xx}[1-\alpha\partial_{xx}]^{-1}\sim\partial_{xx}(1+\alpha\partial_{ xx}+\cdots)=\partial_{xx}+\alpha\partial_{xxxx}+\cdots\) and simply replace \(\alpha\partial_{xx}\) in Eq. by \((1-\alpha\partial_{xx})^{-1}-1\). Thus, we get the nonlinear integro-differential equation:
\[\partial_{t}n=\partial_{xx}{\cal F}[n] \tag{31}\]
where \({\cal F}\) is the functional:
\[{\cal F}[n] = Dn-\beta[\partial_{x}(\ln n)]^{2} \tag{32}\] \[+ \frac{1}{2\sqrt{\alpha}}\int_{-\infty}^{+\infty}e^{-|x-y|/\sqrt{ \alpha}}\ln\left\{\frac{n(y)}{n(x)}\right\}\,dy.\]
Eq. is asymptotically equivalent to Eq. since they differ only by terms of order higher than \(\epsilon^{2}\) but it is free from the spurious high wavenumber instability. By virtue of its construction it also has the property that when \(u^{-1}=0\), the linearized version of Eq. is the exact solution of Eq.-. Thus, when \(u\) is large and amplitudes are small, the solutions of Eq. match closely the true solution, irrespective of the validity of the assumption of long length scales and slow time scales exploited in the asymptotic theory. The formal condition for the validity of Eq. is \(\lambda_{D}\bar{n}^{-1}\partial_{\bar{x}}\bar{n}\ll 1\), where the bars on top of the variables indicate that they are in dimensional form. In dimensionless notation, the condition of validity reduces to \(|\partial_{x}n|/[n^{3/2}\sqrt{1-z}]\ll 1\). It
The quantity \(\Delta=(2D)^{-1}d\sigma^{2}/dt-1\) (lower quadrants) where \(\sigma^{2}\) is the variance of the distribution \(n(x,t)\), and the excess kurtosis \(\gamma\) (upper quadrants) determined using the models (II) [dashed] and (III) [solid]. For model (I), \(\gamma=\Delta=0\) at all \(t\). The initial state is \(n(x,0)=1+A\exp(-x^{2}/2)\) and parameters are \(A=1.0\), \(u=6\), \(z=-1\).
as well as the lower order effective diffusivity approximation fails in the limit of very low concentrations (\(n\to 0\)). This is because the Debye length becomes infinitely large in this limit invalidating our premise of long length scales compared to the local value of the Debye length. Eq. together with Eq. to calculate the concentration of the faster diffusing species are our principal results.
A numerical method employing a fourth order finite difference algorithm for spatial derivatives coupled with a fourth order Runge-Kutta time stepping scheme with fixed grid and step size was used to solve a time dependent electrodiffusion problem. An initial condition \(n(x,0)=1+A\exp(-x^{2}/2)\) was used with \(A=1.0\) and the other parameters were set as \(u=6\) and \(z=-1\). The problem was solved at three levels of approximation:
(I) Using the ambipolar diffusion equation, Eq..
(II) Using the nonlinear equation, Eq..
(III) Using the full electro-diffusion model Eq.-.
It is clear that with (I) the variance \(\sigma^{2}\) would increase linearly with time, but not so in the case of (II). Further, a distribution that is initially Gaussian remains so at future times under (I) but not under (II). Therefore, the excess kurtosis \(\gamma\) is expected to remain zero at all times in the case of (I) but not in the case of (II). In the quantities \(\Delta\equiv(2D)^{-1}d\sigma^{2}/dt-1\) and \(\gamma\) are plotted as a function of time (\(t\)) for (II) and (III). In case of (I), \(\Delta=\gamma=0\) at all \(t\) and this case is not shown for clarity. It is seen that at sufficiently long times (I), (II) and (III) all approach a common asymptotic limit. However, at shorter times, (II) is in better accord with (III) than (I) is.
The qualitative nature of the time variation of \(\gamma\) and \(\Delta\) may also be understood on the basis of Eq.. To see this, we use Eq. to expand the integral operator and put Eq.
\[\partial_{t}n=D\partial_{xx}n+\alpha\partial_{xxxx}(\ln n)-\beta\partial_{xx}( \partial_{x}\ln n)^{2}+\cdots \tag{33}\]
If we further restrict ourselves to small amplitudes, \(n=1+n^{\prime}\), with \(|n^{\prime}|\ll 1\), then \(\ln n=n^{\prime}-\frac{1}{2}(n^{\prime})^{2}+\cdots\), so that the Eq.
\[\partial_{t}n^{\prime}=D\partial_{xx}n^{\prime}+\alpha\partial_{xxxx}n^{\prime }-\frac{\alpha}{2}\partial_{xxxx}(n^{\prime})^{2}-\beta\partial_{xx}(\partial _{x}n^{\prime})^{2}+\cdots \tag{34}\]
Here the \(\cdots\) now include the nonlinear terms of higher order. If we multiply Eq. by \(x^{k}\) and integrate, we can generate ordinary differential equations for the moments \(m_{k}=\int_{-\infty}^{\infty}x^{k}n^{\prime}(x)\ dx\), starting with \(dm_{0}/dt=dm_{1}/dt=0\). The neglected sixth and higher derivative terms do not contribute for \(k\leq 4\). Without loss of generality, we can assume \(m_{1}=0\) so that here \(\sigma^{2}=m_{2}/m_{0}\) and \(\gamma=m_{4}m_{0}/m_{2}^{2}-3\).
\[\frac{d\sigma^{2}}{dt}=2D-\frac{2\beta}{m_{0}}\int_{-\infty}^{+\infty}(\partial _{x}n^{\prime})^{2}\ dx \tag{35}\]
and
\[\frac{d\gamma}{dt}=-\frac{4\gamma D}{\sigma^{2}}+\frac{12\alpha}{\sigma^{4}}+\cdots \tag{36}\]
Now, if \(\alpha=0\), then either \(\gamma=0\) (if it is zero initially) or else \(\gamma\) monotonically decays to zero. When deviations from the ambipolar diffusion limit is considered (\(\alpha>0\)), we find that \(\gamma\) initially increases (when \(\sigma\) is small enough for the second term to dominate), but as \(\sigma\) becomes larger, the first term eventually dominates resulting in a decrease in \(\gamma\) towards zero. Further, Eq. shows, that since \(\beta>0\) the rate of increase of the variance is slightly less than \(2D\) but the deficit in the growth rate becomes progressively smaller at large times.
In summary, a prototypical problem in electrodiffusion was considered in the limit where the Debye-length was non-zero but nevertheless small compared to a characteristic scale of spatial variation. It was shown that in this limit the concentration can be described by a one dimensional nonlinear integro-differential equation which reduces to the linear diffusion equation describing ambipolar diffusion if all but the leading order terms in the ratio of Debye-length to a characteristic spatial scale are neglected. Numerical solution shows that the nonlinear integro-differential equation provides a better approximation to the true solution. The approach described here could be useful for other problems in electrodiffusion where the Debye length is small but may not be considered negligible in relation to other length scales.
_Acknowledgement:_ Support from the NIH under grant R01EB007596 is gratefully acknowledged.
|
10.48550/arXiv.1203.6010
|
A nonlinear equation for ionic diffusion in a strong binary electrolyte
|
Sandip Ghosal, Zhen Chen
| 549
|
10.48550_arXiv.1905.10635
|
###### Abstract
We investigate how different chemical environment influences magnetic properties of terbium(III) (Tb)-based single-molecule magnets (SMMs), using first-principles relativistic multireference methods. Recent experiments showed that Tb-based SMMs can have exceptionally large magnetic anisotropy and that they can be used for experimental realization of quantum information applications, with a judicious choice of chemical environment. Here, we perform complete active space self-consistent field (CASSCF) calculations including relativistic spin-orbit interaction (SOI) for representative Tb-based SMMs such as TbPc\({}_{2}\) and TbPcNc in three charge states. We calculate low-energy electronic structure from which we compute the Tb crystal-field parameters and construct an effective pseudospin Hamiltonian. Our calculations show that ligand type and fine points of molecular geometry do not affect the zero-field splitting, while the latter varies weakly with oxidation number. On the other hand, higher-energy levels have a strong dependence on all these characteristics. For neutral TbPc\({}_{2}\) and TbPcNc molecules, the Tb magnetic moment and the ligand spin are parallel to each other and the coupling strength between them does _not_ depend much on ligand type and details of atomic structure. However, ligand distortion and molecular symmetry play a crucial role in transverse crystal-field parameters which lead to tunnel splitting. The tunnel splitting induces quantum tunneling of magnetization by itself or by combining with other processes. Our results provide insight into mechanisms of magnetization relaxation in the representative Tb-based SMMs.
## 1 Introduction
Single-molecule magnets (SMMs) are magnetic molecules typically composed of one or several transition metal or lanthanide ions surrounded by ligands. The key feature of SMMs is existence of inherent magnetic anisotropy due to interplay between the ligand crystal field (CF) and spin-orbit interaction (SOI). This feature allows one to explore interesting phenomena and potential applications of SMMs for magnetic information storage, spintronics and quantum information processing. Recently, it was reported that monometallic lanthanide-based SMMs can have an effective energy barrier of over 1000 cm\({}^{-1}\) with magnetic hysteresis above liquid nitrogen temperature. For reviews of lanthanide-based SMMs, see Refs.. Furthermore, since the first proposal of the implementation of Grover's algorithm into the prototype SMM Mn\({}_{12}\), quantum bits (qubits) and quantum gates based on SMMs have been experimentally realized. One promising SMM candidate is a double-decker TbPc\({}_{2}\) family (Pc = phthalocyanine), where Rabi oscillations and Grover's algorithm were experimentally implemented by tuning the nuclear spin states of the Tb ion with AC electric fields within single-molecule transistor set-ups.
SMM TbPc\({}_{2}\) consists of a Tb\({}^{3+}\) ion sandwiched between two Pc ligands, as shown in Three different charge states were experimentally realized for TbPc\({}_{2}\) molecules such as anionic [TbPc\({}_{2}\)]\({}^{-}\) (Fig. 1(b)) neutral [TbPc\({}_{2}\)]\({}^{0}\), (Fig. 1(a)) and cationic [TbPc\({}_{2}\)]\({}^{+}\). In the latter case, however, X-ray crystallography data have not been reported yet. Recently, [TbPcNc]\({}^{0,+}\) molecules, where one of the Pc ligand rings was replaced by a larger Nc (naphthalocyaninato) ligand, were experimentally studied. The synthesized TbPc\({}_{2}\) (TbPcNc) molecules have only approximate \(D_{4d}\) (\(C_{4v}\)) symmetry. The degree of symmetry deviation varies with crystal packing, diamagnetic dilution molecules, or solvent molecules used in synthesis processes.
Both TbPc\({}_{2}\) and TbPcNc molecules in different charge states were experimentally shown to exhibit SMM behavior. For the TbPcNc molecule, the measured effective energy barrier is in the range of 340-580 cm\({}^{-1}\) depending on oxidation number, while for the TbPc\({}_{2}\) molecule, the measured barrier is in the range of 230-640 cm\({}^{-1}\) depending on oxidation number. There are no theoretical studies of the origin of this wide range of the energy barrier for TbPcNc or TbPc\({}_{2}\) molecules. Observed magnetization relaxation in SMMs may arise from combined contributions of different relaxation mechanisms such as quantum tunneling of magnetization (QTM), Raman process, Orbach process, hyperfine coupling, and/or intermolecular interaction. In extracting the experimental barrier, it is often assumed that there are no intermolecular interactions and hyperfine interactions. The experimental barrier may also depend on magnetization relaxation mechanisms considered in the fitting of experimental data. Considering the complexity of the relaxation mechanisms and the assumption and ambiguity in the fitting process, it may be difficult to unambiguously determine the magnetization relaxation mechanisms solely from measurements. Furthermore, depending on experimental set-ups, additional environmental factors qualitatively affect the magnetic properties of the SMMs.
In TbPc\({}_{2}\), the Tb\({}^{3+}\) ion has the 4f electronic configuration. According to Hund's rules, its ground multiplet corresponds to spin, orbital and total angular momentum quantum numbers of \(S=3\), \(L=3\) and \(J=6\), respectively. The unquenched orbital angular momentum gives rise to strong SOI. CF of Pc ligands in conjunction with the SOI splits the Tb \(J\)-multiplets, leading to large magnetic anisotropy. CF parameters are often referred to as magnetic anisotropy parameters. As shown in Fig. 2, for low-energy \(J\)-multiplets, different \(J\)-multiplets are well separated from one another. In this case, assuming uniaxial magnetic anisotropy, total angular momentum projected onto the magnetic easy axis (\(M_{J}\)) remains a good quantum number. For a given \(J\)-multiplet, states with the same magnitude of \(M_{J}\) are degenerate, whereas states with different \(|M_{J}|\) values are split.
Side and top views of experimental atomic structure of several TbPc\({}_{2}\)-type molecules. (a) Neutral TbPc\({}_{2}\) with experimental geometry from Ref. (**M1**). (b) Anionic TbPc\({}_{2}\) with experimental geometry from Ref. (**M5**). (c) Neutral TbPcNc with experimental geometry from Ref. (**M3**). Red, blue, orange, and gray spheres represent Tb, N, C, and H atoms, respectively.
Scalar relativistic \(4f\) (\(S=3\), \(L=3\)) energy level for an isolated Tb\({}^{3+}\) ion (left) splitted by SOI into \(J\)-multiplet structure (center). For TbPc\({}_{2}\) molecules the \(J\)-multiplets are further split by the CF of the Pc ligands (right). The energy levels in the boxed region correspond to the \(J=6\) multiplet and are shown in for charged TbPc\({}_{2}\)-type molecules. The energy levels in the left and center panels are calculated for an isolated Tb\({}^{3+}\) ion while those on the right panel are obtained for **M5** molecule (the relative energy scales are aligned such that the **M5** lowest energy level is 3500 cm\({}^{-1}\) below the \(S=3\), \(L=3\) state).
Note that this barrier can differ from experimental effective energy barrier. The energy difference between the ground-state and the first-excited doublets for a given \(J\)-multiplet is referred to as zero-field splitting (ZFS).
For cationic or anionic TbPc\({}_{2}\) or TbPcNc SMMs, the Tb \(J=6\) multiplet structure constitutes the entire low-energy spectrum. In this case, the transverse CF may split the two-fold degeneracy of nonzero \(|M_{J}|\) levels by mixing states with different \(M_{J}\). This phenomenon is referred to as tunnel splitting. For neutral TbPc\({}_{2}\) SMMs, one unpaired electron (with the spin \(s=1/2\)) is delocalized (or shared) within the two Pc ligands. This ligand spin interacts with the Tb magnetic moment by exchange coupling (\(J_{\rm ex}\)) and doubles the number of low-energy levels. In addition, this extra electron makes neutral TbPc\({}_{2}\) a Kramers system, and irrespective of symmetry of the ligand field, it ensures at least two-fold degeneracy of all electronic levels enforced by time-reversal symmetry. Low-energy levels within the ground multiplet, ZFS, tunnel splitting, \(J_{\rm ex}\), and separation between the ground and first-excited multiplets are important energy scales that control the magnetic properties of TbPc\({}_{2}\) and TbPcNc SMMs. They play an important role in elucidation of magnetization relaxation mechanisms.
Despite the great interests and the experimental efforts and ambiguity, there are quite few computational studies of TbPc\({}_{2}\) molecule or its derivatives. In Ref. a neutral TbPc\({}_{2}\) molecule was investigated using density-functional theory (DFT) calculations with and without an on-site Coulomb repulsion \(U\) term in the absence of SOI. However, SOI on Tb ion is much stronger than the ligand CF and, therefore, it is imperative to include SOI for an even qualitative description of TbPc\({}_{2}\). Equally importantly, lanthanides atoms are known to have nearly degenerate electronic configurations demanding multireference treatments. The addition of the \(U\) term alone does not suffice to describe the electronic structure and magnetic properties of the TbPc\({}_{2}\) molecule even qualitatively. Some multireference calculations of the TbPc\({}_{2}\) molecule using complete active space self-consistent field (CASSCF)method including SOI within restricted active space state-interaction (RASSI) have also been reported. In particular, in Ref. a neutral TbPc\({}_{2}\) molecule was studied using a DFT-optimized structure when the molecule is adsorbed on a Ni substrate. In Ref., ZFS and MAB of anionic TbPc\({}_{2}\) were calculated with a particular experimental geometry. There are no multireference calculations of the magnetic properties of the TbPcNc molecule. Therefore, there is still a lack of multireference _ab initio_ studies of magnetic energy scales of TbPc\({}_{2}\)-type SMMs as a function of charge state, type of ligand, and details of molecular geometry.
Here we investigate the effects of ligand and oxidation on the electronic and magnetic properties of the TbPc\({}_{2}\) and TbPcNc molecules in a gas phase, by using the CASSCF multireference method including SOI within RASSI. Using experimental geometries, we study the electronic levels characteristics and analyze the dependence of important magnetic energy scales on oxidation, ligand type and details of molecular structure. Furthermore, we construct an effective pseudospin Hamiltonian that includes Tb CF parameters and the Zeeman interaction (as well as exchange coupling between the Tb magnetic moment and the ligand spin for neutral molecules). The paper is structured as follows. Geometries of interest and methods used in this study are described in Sec. 2 and Sec. 3, respectively. Results of the TbPc\({}_{2}\) molecule in charged forms are followed by those for neutral forms in Sec. 4. We make conclusions in Sec. 5.
## 2 Geometries of Study
We perform calculations for TbPc\({}_{2}\)-type molecules with different charge states and ligand types for which experimental atomic structures are available. The use of experimental molecular geometries is preferred over theoretically optimized geometries since the latter, due to prohibitive computational costs, cannot include counter ions and solvent or dilution molecules which can be quite sizable. Indeed, the type of solvent molecules or diamagnetic dilution molecules as well as crystal packing, can significantly affect the ligand geometry of TbPc\({}_{2}\)-type molecules. For example, two [TbPc\({}_{2}\)]\({}^{-}\) molecules in Refs. and (referred to as **M4** and **M5** later) have significantly different geometries despite the same oxidation and ligand type (Table 1).
We consider the following six molecules (Table 1): (**M1**) neutral TbPc\({}_{2}\) with experimental geometry from Ref., (**M2**) neutral TbPc\({}_{2}\) with experimental geometry from Ref., (**M3**) neutral TbPcNc with experimental geometry from Ref., (**M4**) anionic TbPc\({}_{2}\) with experimental geometry from Ref., (**M5**) anionic TbPc\({}_{2}\) with experimental geometry from Ref., (**M6**) cationic TbPcNc with experimental geometry from Ref.. As is a common practice, for all experimental geometries we correct the carbon-hydrogen bond length to 1.09 A since this distance cannot be reliably extracted from X-ray measurements. Solvent molecules, diamagnetic dilution molecules, or counter ions are not included in our calculations.
The structure can be viewed in terms of the Tb ion sandwiched between two approximately flat Pc or Nc ligand planes. The ligands are rotated with respect to each other by roughly 45\({}^{\circ}\) angle. Each ligand has four roughly identical branches that form \(\approx\)90\({}^{\circ}\) angle with each other. Each branch starts with a nitrogen atom that is the nearest neighbor of the Tb ion and is denoted as N\({}_{\rm nn}\). Away from the Tb ion, N\({}_{\rm nn}\) is connected to two intermediate carbon atoms which, in turn, are connected to a benzene-like carbon ring. Two carbon atoms from the ring which are the farthest from the Tb ion are denoted as C\({}_{\rm far}\). For Nc, C\({}_{\rm far}\) atoms are further connected to an additional carbon ring. The neighboring branches are linked by bridging nitrogen atoms which form bonds with the intermediate carbon atoms.
Schematic molecular structure illustrating definitions of the structural parameters in Table 1. The highlighted region in (a) is zoomed in (b). The Tb ion is at the origin.
In order to characterize the experimental geometries we introduce several structural parameters (see Table 1 and Fig. 3). The analysis of these parameters allows us to quantify deviations of the molecular structure from the ideal \(D_{4d}\) (or \(C_{4v}\) for TbPcNc) symmetry induced by counter cations/anions, solvent molecules, diamagnetic dilution molecules, or crystal packing. For **M1**, **M3**, and **M6** molecules, both \(\theta_{\rm N_{nn}TbN_{nn}}\) and \(\theta_{\rm C_{far}TbC_{far}}\) angles are close to 45\({}^{\circ}\) which indicates that the four-fold symmetry is approximately preserved. To a somewhat lesser degree, this is also true for **M2** and **M4** molecules. This is consistent with a small standard deviation of the \(D_{\rm TbN_{nn}}\) and \(D_{\rm TbC_{far}}\) bond lengths for these molecules. On the other hand, for **M5** molecule, we have strong deviations from the four-fold symmetry as reflected in the fact that both the angles and the \(\theta_{\rm N_{nn}TbN_{nn}}\) and \(\theta_{\rm C_{far}TbC_{far}}\) as well as the \(D_{\rm TbN_{nn}}\) and \(D_{\rm TbC_{far}}\) bond lengths vary significantly. For all considered geometries, we observe curving of the carbon parts of the ligand planes away from each other (this effect can be also seen in Fig. 1). The curving can be different for different ligand branches and it is most pronounced for the **M5** molecule. The strong deviations of the **M5** geometry from the ideal TbPc\({}_{2}\) structure is likely a result of bulky diamagnetic dilution molecules used in the synthesis process.
## 3 Methods
The multireference calculations are performed using the Molcas quantum chemistry code (version 8.2). Scalar relativistic effects are included based on the Douglas-Kroll-Hess Hamiltonian using relativistically contracted atomic natural orbital (ANO-RCC) basis sets. In particular, polarized valence triple-\(\zeta\) quality (ANO-RCC-VTZP) is used for the Tb ion, polarized valence double-\(\zeta\) quality (ANO-RCC-VDZP) is used for the nitrogen and carbon atoms, and valence double-\(\zeta\) quality (ANO-RCC-VDZ) is used for the hydrogen atoms. Such a choice of the basis set is made to maintain a high accuracy and to not exceed computational capabilities. More details on the basis set dependence are discussed in Tables S4 and S5 in Supporting Information.
First, in the absence of SOI, for a given spin multiplicity, the spin-free eigenstates are obtained using state-averaged CASSCF method. The valence electronic configurations of Tb\({}^{+3}\) ion consists of eight electrons at \(4f\) orbitals which must be included in the active space. It would be desirable to include all ligand \(\pi/\pi^{*}\)-type orbitals as well. This is, however, computationally prohibitive and, therefore, only several near-in-energy ligand orbitals are included in the active space. In order to identify such ligand orbitals, we consider molecules in the cationic state and perform CASSCF calculations with eight electrons and seven Tb \(4f\)-type orbitals in the active space. We find that HOMO (L1) is always energetically separated (\(\sim\)0.1 a.u.) from other occupied ligand orbitals. For unoccupied orbitals we find that LUMO-1 (L2) as well as nearly degenerate LUMO-2 (L3) and LUMO-3 (L4) are separated from higher unoccupied states. Therefore, we include L1, L2, L3, and L4 ligand orbitals in the active space. Altogether, as illustrated in Fig. 4, we consider eleven active orbitals with ten, eleven, and twelve active electrons for cationic, neutral, and anionic molecules, respectively. The effect of the choice of the active space on the final results is discussed in Supporting Information (Tables S7 and S8).
In the case of charged molecules we consider only the \(S=3\) configuration since other spin configurations such as \(S=2\), \(S=1\), and \(S=0\) are much higher in energy and their inclusion changes the energy levels of the ground \(J\)-multiplet only by about a few cm\({}^{-1}\). See Table S6 in Supporting Information for details. For neutral molecules, depending on whether the ligand spin (\(s=1/2\)) is parallel or antiparallel to the Tb spin \(S=3\), we have two possible values of the total spin of the molecule: \(S_{\rm tot}=7/2\) or \(S_{\rm tot}=5/2\) (see Fig. 4). We consider both \(S_{\rm tot}\) values since they lie close in energy. For a given spin configuration, we evaluate seven lowest spin-free states (roots) that correspond to different configurations of eight electrons in Tb \(4f\)-type orbitals with Tb spin of \(S=3\). These seven spin-free states are used in the state-averaged procedure.
In the next step, we include SOI, within the atomic mean-field approximation, in Schematic illustration of the active space used in the CASSCF calculations for different charge states. L1, L2, L3, and L4 denote four ligand orbitals that lie close to the HOMO-LUMO gap. The figure shows nominal occupation of the orbitals.
With SOI, all possible \(J\)-multiplets from the addition of \(\mathbf{L}\) and \(\mathbf{S}\) are generated as illustrated in For the calculation of the CF parameters, we use the methodology implemented in the SINGLE_ANISO module of the Molcas code.
The technique from Ref. that we use for the charged molecules cannot be directly applied to the neutral molecules. Indeed, for lanthanides this method finds CF parameters and \(g\)-tensor elements for a given \(J\)-multiplet using lowest \((2J+1)\)_ab initio_ eigenvalues and the corresponding eigenfunctions. For neutral molecules, however, the low-energy levels do not only originate solely from the ground \(J\)-multiplet but also they involve the unpaired ligand electron spin-flip states. Both types of excitations are entangled and there is no obvious way how to extract multiplet levels and their wave functions. In fact, the entanglement is essential for formation of Kramers doublet. In order to circumvent this problem, we take the experimental geometries of the neutral molecules and consider them in the cationic state for the calculation of the Tb CF parameters. In this way we remove the unpaired ligand electron so that the calculated low-energy spectrum can be put into correspondence with the Tb \(J=6\) multiplet and the CF parameters as well as the \(g\)-tensor elements can be calculated. Assuming that the unpaired ligand electron has a small contribution to the Tb CF and \(g\)-tensor, these parameters as well as the exchange coupling constant can be used in effective pseudospin Hamiltonian for the neutral molecules.
## 4 Results and Discussion
In each subsection, we present the calculated magnetic energy levels obtained from CASSCF-RASSI-SOI method and construct the effective pseudospin Hamiltonian with the calculated CF parameters and \(g\) tensor. We then compare with relevant theoretical results and experimental data when they are available.
### Charged molecules
For charged TbPc\({}_{2}\)-type molecules (**M4**, **M5** and **M6**), we find that among eleven active molecular orbitals, seven are Tb \(4f\)-type. These orbitals are similar for all considered molecules and are shown in Fig. S1 in Supporting Information. The occupation of each \(4f\)-type orbital is approximately 1.14 which corresponds to about eight electrons occupying these orbitals (note that this is a result of state-averaged calculations). The remaining four active orbitals (L1-L4) are ligand orbitals and are shown in Figs. S5-S7 in Supporting Information. For cationic **M6** molecule, L1 is almost doubly occupied while L2-L4 are almost empty. All ligand orbitals arise mostly from inner C atoms and they have in-plane symmetry due to \(C_{4}\) symmetry of the molecule (Table 1). There is, however, asymmetry between orbitals in the top Pc and the bottom Nc ligands. For anionic **M4** molecule, the occupation of nominally filled L1 and L2 orbitals is somewhat lower than two while the nominally empty L3 and L4 orbitals have a significant occupation (Fig. S5). The effect is even stronger for the **M5** molecule (Fig. S6) and it suggests sizable correlations. The orbitals have significant in-plane asymmetry as well as asymmetry between the two Pc planes.
\begin{table}
\begin{tabular}{l c c c c c c} \hline & **M1** & **M2** & **M3** & **M4** & **M5** & **M6** \\ \hline \(E_{\rm TS}^{1,2\,a}\) & n/a & n/a & n/a & 0.000 & 0.007 & 0.000 \\ \(E_{\rm TS}^{3\,4\,b}\) & n/a & n/a & n/a & 0.015 & 0.090 & 0.000 \\ \(E_{\rm TS}^{5,6\,c}\) & n/a & n/a & n/a & 0.530 & 7.969 & 0.499 \\ \(\Delta E_{\rm ex}^{\,d}\) & 8.2 & 6.6 & 5.1 & n/a & n/a & n/a \\ \(J_{\rm ex}^{\,e}\) & 0.8 & 0.6 & 0.6 & n/a & n/a & n/a \\ \(E_{\rm ZFS}^{\,f}\) & 308 & 305 & 308 & 289 & 292 & 298 \\ \(E_{\rm MAB}^{\,g}\) & 658 & 639 & 622 & 582 & 734 & 592 \\ \(\Delta E_{J}^{\,h}\) & 2083 & 2068 & 2068 & 2036 & 2045 & 2049 \\ \hline \end{tabular} \({}^{a}\) Tunnel splitting for the ground-state quasi-doublet; \({}^{b}\) Tunnel splitting for the first excited quasi-doublet; \({}^{c}\) Tunnel splitting for the second excited quasi-doublet; \({}^{d}\) Energy difference between states with parallel and antiparallel orientation of the Tb angular momentum and the ligand spin; \({}^{e}\) Exchange coupling between the Tb angular momentum and the ligand spin; \({}^{f}\) ZFS; \({}^{g}\) MAB; \({}^{h}\) Separation between the ground and first-excited \(J\) multiplets.
\end{table}
Table 2: Calculated Magnetic Energy Scales for Different TbPc\({}_{2}\)-Type SMMs in Units of cm\({}^{-1}\)Calculated low-energy (\(J=6\)) spectrum for different charged TbPc\({}_{2}\)-type molecules. For **M1**, **M2**, and **M3** molecules calculations are done in the cationic state with neutral-state geometries.
For the charged molecules, the low-energy spectrum is largely determined by Tb atomic levels which are split in the ligand CF. As listed in Table 2, the separation between the ground and the first-excited \(J\)-multiplet, \(\Delta E_{J}\), is almost a factor of three larger than the MAB of the ground \(J\)-multiplet, \(E_{\rm MAB}\). Thus, we can project the lowest \(J\) multiplet onto a multiplet of effective pseudospin \(\tilde{J}=6\). shows the thirteen lowest calculated energy levels corresponding to the \(J=6\) ground multiplet for all considered charged molecules (numeric data together with spin-free energies are provided in Tables S1 and S2 in Supporting Information).
\[\hat{H}=\sum_{k=2,{\rm even}}\sum_{q=-k}^{k}B_{k}^{q}\hat{O}_{k}^{q}(\hat{\bf J })-\mu_{B}{\bf B}\cdot{\bf g}\cdot\hat{\bf J}, \tag{1}\]
Here \(\hat{\bf J}\) is the Tb pseudospin operator, and \(\hat{O}_{k}^{-q}(\hat{\bf J})=[\hat{O}_{k}^{q}(\hat{\bf J})]^{\star}\), where \(\star\) means complex conjugate. For example, second-rank Stevens operators are \(\hat{O}_{2}^{0}=(3J_{z}^{2}-J(J+1)I)\), \(\hat{O}_{2}^{1}=(J_{z}(J_{+}+J_{-})+(J_{+}+J_{-})J_{z})/4\) and \(\hat{O}_{2}^{2}=(J_{+}^{2}+J_{-}^{2})/2\), where \(J_{\pm}\) are raising and lowering operators and \(I\) is a 2\(\times\)2 identity matrix. Higher-rank Stevens operators are listed in Table S9 in Supporting Information. Here \(\hat{O}_{k}^{q=0}\) terms represent uniaxial or diagonal contributions, while \(\hat{O}_{k}^{q\neq 0}\) terms are transverse or off-diagonal contributions. Time-reversal symmetry enforces only even integer \(k\) in the summation. Molecular symmetry dictates allowed nonzero \(B_{k}^{q}\) values. When the second-order uniaxial term, \(\hat{O}_{2}^{0}\), is dominant, ZFS and MAB are approximated to be \(3|B_{2}^{0}|(2J-1)\) and \(3|B_{2}^{0}|J^{2}\), respectively, for integer \(J\). The second term in Eq. is the Zeeman interaction with the \(g\)-tensor, \({\bf g}\).
We evaluate the elements of the \(g\)-tensor and \(B_{k}^{q}\) values using the thirteen _ab initio_ energy levels and the corresponding eigenfunctions. The results are shown in Tables 3 and 4 (higher order CF parameters are shown in Table S10 in Supporting Information). Weuse the coordinate system along the principal axes of the \(g\)-tensor. In this coordinate system, the \(z\) axis is roughly perpendicular to the ligand planes. Note that the principal values of the \(g\)-tensor are close to the ideal Lande \(g\)-factor value of 3/2. The calculated \(|B_{2}^{0}|\) value is the largest for **M5** and the smallest for **M4**. Due to the absence of any symmetry, however, **M4** and **M5** molecules have significant all-order off-diagonal CF parameters (Table 4). In particular, for **M5**, the \(B_{2}^{2}\) value is about 63% of the \(|B_{2}^{0}|\) value, while the \(|B_{4}^{4}|\) value is similar to the \(|B_{4}^{0}|\) value. On the other hand, for **M6**, only off-diagonal terms with \(q=\pm 4\) are significant, which is dictated by the molecular \(C_{4}\) symmetry. Including the full set of CF parameters up to sixth order (\(k=6\)) and the diagonal eighth order term (\(B_{8}^{0}\)), diagonalization of Eq. reproduces the _ab initio_ energy levels up to 0.5 cm\({}^{-1}\). This indicates that Eq. is a proper Hamiltonian for the low-energy spectrum of charged TbPc2-type SMMs.
The lower part of the spectrum (six lowest levels for **M4** and **M6** and four lowest levels for **M5**) is composed of approximate doublets, as shown in The former three doublets correspond to states \(M_{J}=\pm 6\), \(\pm 5\), and \(\pm 4\), while the latter two doublets are states \(M_{J}=\pm 6\) and \(\pm 5\), respectively. For further level characteristics, see Table S14-S16 in Supporting Information. The transverse CF, however, mixes states with different \(|M_{J}|\) values and breaks the degeneracies of the doublets leading to tunnel splitting. It is important to understand tunneling splitting of the levels within the ground multiplet since magnetization can be relaxed via phonon-assisted tunneling.
\begin{table}
\begin{tabular}{c c c c c c c} \hline & **M1\({}^{b}\)** & **M2\({}^{b}\)** & **M3\({}^{b}\)** & **M4** & **M5** & **M6** \\ \hline \(g_{x}\) & 1.505 & 1.505 & 1.504 & 1.505 & 1.510 & 1.504 \\ \(g_{y}\) & 1.504 & 1.503 & 1.504 & 1.502 & 1.498 & 1.504 \\ \(g_{z}\) & 1.476 & 1.477 & 1.477 & 1.480 & 1.477 & 1.478 \\ \hline \end{tabular} \({}^{a}\) The coordinate system is along the principal axes of the \(g\)-tensor with the \(z\) axis being roughly perpendicular to the ligand plane; \({}^{b}\) The \(g\)-tensor was calculated assuming a cationic charge state (see the text).
\end{table}
Table 3: Calculated Principal Values\({}^{a}\) of the \(g\)-Tensor for Different TbPc\({}_{2}\)-Type SMMswith opposite \(M_{J}\). For high-energy levels with small \(|M_{J}|\), the transverse CF becomes more important and tunnel splitting is more pronounced. In fact, for higher part of the spectrum, \(|M_{J}|\) ceases to be a good quantum number and the doublet structure disappears. The tunnel splitting and \(|M_{J}|\) mixing for **M5** are significantly larger than those for **M4** and **M6**. This is because the substantial distortions and curving of the ligand planes for **M5** lead to significant transverse CF (Table 4). The geometrical distortions are shown in the larger spread of \(D_{\rm TbN_{nn}}\) and \(D_{\rm TbC_{far}}\) values and of \(\theta_{\rm N_{nn}TbN_{nn}}\) and \(\theta_{\rm C_{far}TbC_{far}}\) values and in the larger range of the \(Z_{\rm C^{up}}\) and \(Z_{\rm C^{down}}\) values in Table 1, compared to the other molecules of interest. For **M5**, the tunnel splitting values are of the order of \(10^{-2}\), \(10^{-1}\), and \(10\) cm\({}^{-1}\) for the ground state and the first- and second-excited states, respectively (Table 2). For **M6**, the \(C_{4}\) symmetry allows mixing of \(M_{J}\) levels with \(\Delta M_{J}\)=\(\pm 4\) only. This explains the significant tunnel splitting only for a pair of states 5 and 6 and a pair of states 9 and 13 (Table 2 and Table S2 in Supporting Information).
The calculated ZFS and MAB values are shown in Table 2.
\begin{table}
\begin{tabular}{c c c c} \hline \(B_{k}^{q}\) & **M4** & **M5** & **M6** \\ \hline \(B_{2}^{-2}\) & -0.12802616 & -0.02332695 & 0.00000240 \\ \(B_{2}^{-1}\) & -0.03939183 & -0.00727754 & 0.00001437 \\ \(B_{2}^{0}\) & -5.05068413 & -5.52638125 & -5.33499223 \\ \(B_{2}^{1}\) & -0.02663218 & -0.00634158 & -0.00001624 \\ \(B_{2}^{2}\) & 0.80463902 & 3.45788034 & 0.00003865 \\ \hline \(B_{4}^{-4}\) & -0.00332496 & -0.00213682 & 0.00610215 \\ \(B_{4}^{-3}\) & 0.00793656 & -0.00424813 & -0.00000026 \\ \(B_{4}^{-2}\) & 0.00410350 & 0.00096348 & -0.00000007 \\ \(B_{4}^{-1}\) & -0.00375441 & 0.00049609 & 0.00000289 \\ \(B_{4}^{0}\) & -0.01406960 & -0.01300444 & -0.01412822 \\ \(B_{4}^{1}\) & -0.00514061 & 0.00707655 & -0.00000337 \\ \(B_{4}^{2}\) & -0.00151449 & -0.02369506 & -0.00000004 \\ \(B_{4}^{3}\) & -0.00030393 & 0.00188537 & 0.00000259 \\ \(B_{4}^{4}\) & -0.00079989 & 0.01117566 & 0.00540901 \\ \hline \end{tabular} \({}^{a}\) The coordinate system is along the magnetic axes that diagonalize the \(g\)-tensor (\(z\) axis roughly perpendicular to the ligand planes); \({}^{b}\) Higher order \(B_{k}^{q}\) are provided in Table S10 in Supporting Information.
\end{table}
Table 4: CF Parameters\({}^{a,b}\) in cm\({}^{-1}\) for Tb Ion for M4, M5 and M6 SMMs.
As seen from Fig. 5, ZFS for **M6** is similar to ZFS for other cationic TbPc\({}_{2}\) or TbPcNc molecules. Therefore, we conclude that ZFS does not depend on ligand type and geometry details, and it only shows a weak dependence on oxidation number. This reflects the fact that ZFS characterizes the lower part of the energy spectrum where the transverse CF has a small effect.
Our calculated results for **M5** molecule can be directly compared with similar calculations from Ref. where the same experimental geometry was used. ZFS and MAB reported in Ref. are 308 cm\({}^{-1}\) and 809 cm\({}^{-1}\), respectively. While ZFS is reasonably close to our value (292 cm\({}^{-1}\)), the MAB differs more significantly from our value (770 cm\({}^{-1}\)). We check that both the choice of the active space and the effect of higher spin-free states are not responsible for the significant difference (see Supporting Information). The most likely reason is the basis set difference between our study and Ref..
On the experimental side, the Tb CF parameters for [TbPc\({}_{2}\)]\({}^{-}\)TBA\({}^{+}\) estimated by Ishikawa _et al._ are widely used in the community. These CF parameters were obtained by fitting both the experimental nuclear magnetic resonance (NMR) shift and magnetic susceptibility data to a ligand-field model, assuming perfect C\({}_{4}\) symmetry. The estimated CF parameters are listed in Thiele _et al._ In this estimate, three uniaxial terms \(B_{2}^{0}\), \(B_{4}^{0}\), and \(B_{6}^{0}\) as well as only one transverse term \(B_{4}^{4}\) were considered. The estimated values are \(B_{2}^{0}\)=\(-\)4.18193, \(B_{4}^{0}\)=\(-\)0.02790, \(B_{4}^{4}\) =0.00122, and \(B_{6}^{0}\)=\(-\)0.00004 cm\({}^{-1}\). Since X-ray crystallography data was not reported and perfect C\({}_{4}\) symmetry was assumed in Ishikawa _et al._, our calculated CF parameters cannot be directly compared to their fitted values. Nonetheless, it is worthwhile providing some remarks. Note that **M4** molecule has the same counter cation without bulky dilution molecules as in Ishikawa _et al._ Thus, we discuss major differences between our calculated CF parameters for **M4** molecule with those by Ishikawa _et_al._ For **M4** molecule, the \(|B_{2}^{0}|\) value is somewhat greater than the fitted value, while the \(|B_{4}^{0}|\), \(|B_{4}^{4}|\), and \(|B_{6}^{0}|\) values are comparable to the fitted values. See Table 6 and Table S10 in Supporting Information. The major difference is that we find large low-order transverse CF parameters such as \(|B_{2}^{q\neq 0,2}|\), \(|B_{4}^{q\neq 0,4}|\) and \(|B_{6}^{q\neq 0}|\), whereas the literature considered only one transverse CF parameter \(B_{4}^{4}\). As a result, our result predicts a much larger tunnel splitting and qualitatively different level characteristics above the first-excited states.
Table 5 shows the experimental effective barrier for several TbPc\({}_{2}\)-like molecules. The measured barrier for **M6** was reported, considering Raman and Orbach processes, while the barrier for a diluted crystal of neutral TbPcNc (**M3**) molecules was obtained, considering Orbach and quantum tunneling processes. It is interesting to compare the measured barrier for **M5** with and without bulky diamagnetic dilution molecules. These experimental values are much larger than the earlier reported values from Refs. and, for anionic TbPc\({}_{2}\) molecular crystals. Comparison with the experimental data indicates that for **M6** molecule, the experimental barrier is close to states 11 and 12, and for **M5** molecule, the range of the experimental barrier falls on states 10 and 11. See Table S2 in Supporting Information.
### Neutral molecules
Neutral TbPc\({}_{2}\)-type molecules (**M1**, **M2** and **M3**) have much richer electronic structure than their charged counterparts due to presence of an extra unpaired electron that is expected to reside in the ligands. As in the case of the charged molecules, seven of the eleven active orbitals are Tb 4\(f\)-like (Fig. S1 in Supporting Information).
\begin{table}
\begin{tabular}{l c c c c c c c} \hline Molecule & TbPc\({}_{2}\)\({}^{a}\) & TbPc\({}_{2}\)\({}^{b}\) & TbPc\({}_{2}\)\({}^{b}\) & TbL\({}_{2}\)\({}^{c,d}\) & TbL\({}_{2}\)\({}^{c,d}\) & TbPcNc\({}^{e}\) & TbPcNc\({}^{e}\) \\ \hline Charge & -1 & -1 & -1 & -1 & +1 & +1 & 0 \\ Dilution & Yes & No & Yes & No & No & No & No \\ MAB & 260 & 584 & 641 & 594 & 550 & 584 & 342 \\ Relevance & n/a & **M5** & **M5** & n/a & n/a & **M6** & **M3** \\ \hline \end{tabular} \({}^{a}\) Ref.; \({}^{b}\) Ref.; \({}^{c}\) L=Pc(OEt)\({}_{8}\); \({}^{d}\) Ref.; \({}^{e}\) Ref.;
\end{table}
Table 5: Measured effective magnetic anisotropy barrier (in cm\({}^{-1}\)) for TbPc\({}_{2}\)-like single-molecule magnetsGraphical representation of SOMO (L2) obtained from scalar relativistic CASSCF calculations for neutral **M1** molecule (top) and for a flattened **M1** molecule (bottom). Red and blue regions represent opposite phases of the orbital wave function. The figure is prepared using the Luscus software.
The remaining four active orbitals (L1-L4) are ligand orbitals and are shown in Figs. S2-S4 in Supporting Information. For all neutral molecules L1 is almost doubly occupied, L2 is singly occupied (SOMO) while L3 and L4 have small occupation numbers. For **M3** molecule, the SOMO (L2) has top-down plane asymmetry due to different ligand type but shows in-plane symmetry. Interestingly, for **M1** and **M2** molecules, despite the same ligand type, the SOMO is highly asymmetric with large weight on the bottom Pc plane only (Fig. S2 and S3). This is due to the asymmetric curvature within the two Pc planes, as listed in Table 1. Compare the \(Z_{\rm C^{up}}\) and \(Z_{\rm C^{down}}\) values in Table 1. In order to check on this, we enforce mirror symmetry about the \(xy\) plane (\(\sigma_{h}\)) in **M1** molecule and compute the ligand orbitals in the active space. For this flattened **M1** molecule, we find that top-bottom plane symmetry is more or less restored in the SOMO. Compare Fig. 6(a) with (b).
The spin of the ligand electron can be either parallel (\(S_{\rm tot}=7/2\)) or antiparallel (\(S_{\rm tot}=5/2\)) to the Tb spin. The calculated spin density for both cases is shown in for **M1** and **M3** molecules. The Tb spin density is localized in the vicinity of Tb ionic core.
Spin density from scalar relativistic CASSCF calculations for **M1** molecule for \(S_{\rm tot}=7/2\) (a) and \(S_{\rm tot}=5/2\) (b) and for **M3** molecule for \(S_{\rm tot}=7/2\) (c) and \(S_{\rm tot}=5/2\) (c). Red and blue colors denote spin-up and spin-down density, respectively. The figure is prepared using the Luscus software.
Pc plane. This is consistent with the SOMO (Fig. 6(a)) that is primarily delocalized on these carbon atoms. The same behavior is observed for **M2** molecule (not shown). For **M3**, consistently with SOMO, the majority of the ligand spin density is symmetrically shared by the inner carbon atoms from the Nc ligand. Table S1 in Supporting Information shows the calculated seven lowest spin-free states for both values of \(S_{\rm tot}\) for different neutral molecules. For all systems, the parallel configuration of the Tb and ligand spin (\(S_{\rm tot}=7/2\)) has a lower energy than the antiparallel configuration (\(S_{\rm tot}=5/2\)).
## Table 6: CF Parameters\({}^{a,b}\) in cm\({}^{-1}\) for Tb Ion for M1, M2 and M3 SMMs
\begin{tabular}{c c c c} \hline \(B_{k}^{q}\) & **M1\({}^{c}\)** & **M2\({}^{c}\)** & **M3\({}^{c}\)** \\ \hline \(B_{2}^{-2}\) & 0.00000001 & -0.04824674 & -0.00000478 \\ \(B_{2}^{-1}\) & -0.00000003 & 0.03000187 & 0.00000119 \\ \(B_{2}^{0}\) & -5.93432163 & -5.68334414 & -5.65157592 \\ \(B_{2}^{1}\) & -0.00444594 & -0.04186354 & -0.00001973 \\ \(B_{2}^{2}\) & 0.32444131 & 0.44453996 & 0.00019387 \\ \hline \(B_{4}^{-4}\) & 0.00000000 & -0.01188166 & -0.00302990 \\ \(B_{4}^{-3}\) & 0.00000000 & -0.00089983 & 0.00000000 \\ \(B_{4}^{-2}\) & 0.00000000 & 0.00162901 & 0.00000014 \\ \(B_{4}^{-1}\) & 0.00000000 & -0.00034396 & 0.00000135 \\ \(B_{4}^{0}\) & -0.01336535 & -0.01369266 & -0.01380654 \\ \(B_{4}^{1}\) & -0.00043523 & -0.00806695 & -0.00000275 \\ \(B_{4}^{2}\) & -0.00143680 & -0.00160371 & -0.00000081 \\ \(B_{4}^{3}\) & 0.00160576 & -0.00288249 & -0.00000487 \\ \(B_{4}^{4}\) & -0.00027466 & -0.01198852 & -0.00720075 \\ \hline \end{tabular}
\({}^{a}\) The coordinate system is along the magnetic axes that diagonalize the \(g\)-tensor (\(z\) axis roughly perpendicular to the ligand planes); \({}^{b}\) Higher order \(B_{k}^{q}\) are provided in Table S10 in Supporting Information; \({}^{c}\) The CF parameters were calculated assuming a cationic charge state (see the text).
Since neutral TbPc\({}_{2}\) and TbPcNc molecules have an odd number of electrons, the Kramers theorem dictates that all electronic levels are at least doubly degenerate. The low symmetry of the considered molecules prevents from appearance of higher degeneracy. The low-energy part of the calculated spectra for **M1**, **M2**, and **M3** molecules consists of a group of thirteen Kramers doublets (and Table S3 in Supporting Information).
The thirteen lowest Kramers doublets result from Tb \(J=6\) multiplet that is coupled with the unpaired ligand electron spin. The next electronic level belongs to the first-excited(a) Calculated low-energy spectrum for **M1**, **M2**, and **M3** neutral molecules. Each line represents a Kramers doublet. The red rectangle denotes the low-energy part of the spectrum that is schematically illustrated in panel (b).
Therefore, for analysis of the magnetic properties we can focus only on the thirteen lowest doublets. Such an energy spectrum can be described by the following generalization of Eq. in order to include the ligand spin:
\[\hat{H}=\sum_{kq}B_{k}^{q}\hat{O}_{k}^{q}(\hat{\bf J})+J_{\rm ex}\hat{\bf s} \cdot\hat{\bf J}-\mu_{B}{\bf B}\cdot\left({\bf g}\cdot\hat{\bf J}+g_{e}\hat{ \bf s}\right), \tag{2}\]
In addition, the Zeeman term is generalized to describe interaction of the ligand spin with magnetic field (\(g_{e}\approx 2\) is a free-electron \(g\)-factor). Note that in the presence of strong SOI, anisotropic and antisymmetric exchange couplings may play an important role and rigorous treatment of exchange interaction requires more general formalism. We expect, however, that the simple isotropic Heisenberg form in Eq. is a good approximation for our molecules since the exchange interaction is very small. The anisotropic and antisymetric exchanges (as well as higher order spin interactions) are, in general, smaller than the isotropic exchange and, therefore should not play a significant role for TbPc\({}_{2}\)-type molecules. As discussed below, we find that Hamiltonian provides a good representation of the low-energy spectrum of the neutral molecules.
First-principles evaluations of the Tb \(B_{k}^{q}\) coefficients and the \(g\)-tensor elements are not straightforward for the neutral molecules. As discussed in Sec. 3, we obtain the Tb \(B_{k}^{q}\) coefficients and the \(g\)-tensor by using the calculated low-energy spectra (Fig. 5(a)-(c) and Table S2 in Supporting Information) of the neutral molecular geometries with one electron removed. Here we assume that the contribution of the unpaired ligand electron to the Tb CF and \(g\)-tensor is negligible. Later we show that this is indeed a valid assumption. The calculated elements of the \(g\)-tensor and the Tb CF parameters are shown in Tables 3 and 6 (see Table S10 in Supporting Information for a full set of \(B_{k}^{q}\)). Note that as in the case of **M4**, **M5** and **M6** molecules, we use a coordinate system along the principal axes of the \(g\)-tensor for which the \(z\) axis (easy axis) points approximately in the direction perpendicular to the ligand planes. We find all the principal values of the \(g\)-tensor being close to the ideal Lande \(g\)-factor value of 3/2 (Table 3). The calculated \(|B_{2}^{0}|\) value for **M1** is a bit larger than that for **M2** and **M3**, which is reflected in the largest ZFS among the three neutral molecules. The calculated \(|B_{2}^{0}|\) value for **M1**, **M2**, and **M3** is consistently larger than that for the charged molecules. Compared to the charged TbPc\({}_{2}\) molecules, the degree of the structural distortion is much less for the neutral TbPc\({}_{2}\) molecules. See \(D_{\rm TbC_{far}}\), \(Z_{\rm C^{up}}\), and \(Z_{\rm C^{down}}\) values in Table 1. This explains overall smaller transverse CF parameters in the neutral TbPc\({}_{2}\) molecules than in the charged TbPc\({}_{2}\) molecules.
In order to evaluate the exchange coupling \(J_{\rm ex}\), we fit the _ab-initio_ energies of the lowest thirteen doublets to eigenvalues of Eq. with a fitting parameter \(J_{\rm ex}\). For all molecules, we obtain a high-quality fit. This indicates that the method we use for evaluation of CF parameters is reliable and that Eq. provides a reasonable description of the low-energy spectrum of neutral TbPc\({}_{2}\)-type molecules. We find a small ferromagnetic exchange coupling with \(J_{\rm ex}\approx 0.6-0.8\) cm\({}^{-1}\) (Table 2). An increase of the active space or the atomic basis sets does not affect this number significantly (see Tables S4 and S7 in Supporting Information). The insensitivity of \(J_{\rm ex}\) to ligand type and synthesis process indicates that the exchange coupling does not depend on details of molecular geometry much as long as structural changes are moderate. This conclusion is only applied to molecules in crystals or films on substrates rather than within single-molecule transistors where the molecules experience much stronger structural changes.
Using the pseudospin Hamiltonian, Eq., we can determine the characteristics of energy levels shown in For all considered molecules, four lowest Kramers doublets have a well-defined \(|M_{J}|\) value. For each pair, the lower (higher) energy level has the ligand spin parallel (antiparallel) to the Tb angular momentum. Within each Kramers doublet, it is convenient to choose such a basis set in which each Kramers partner state is characterized by \(M_{J}\) and \(m_{s}\) quantum numbers. Since Kramers partner states are related by time reversal symmetry, they have opposite values of \(M_{J}\) and \(m_{s}\). If we focus on the Kramers partner state with positive \(M_{J}\), we find (\(M_{J}=6\), \(m_{s}=1/2\)) for the ground state level (level 1), (\(M_{J}=6\), \(m_{s}=-1/2\)) for the first-excited level (level 2), (\(M_{J}=5\), \(m_{s}=1/2\)) for the second-excited level (level 3), (\(M_{J}=5\), \(m_{s}=-1/2\)) for the third-excited level (level 4). For the most symmetric **M3** molecule, level 5 consists of a majority of (\(M_{J}=4\), \(m_{s}=1/2\)) and small contributions from (\(M_{J}=-4\), \(m_{s}=1/2\)) and (\(M_{J}=0\), \(m_{s}=1/2\)), while level 6 consists of a majority of (\(M_{J}=4\), \(m_{s}=-1/2\)) with small contributions from (\(M_{J}=-4\), \(m_{s}=-1/2\)), (\(M_{J}=-4\), \(m_{s}=1/2\)), (\(M_{J}=0\), \(m_{s}=-1/2\)) and (\(M_{J}=3\), \(m_{s}=1/2\)).
For the neutral molecules, we define ZFS as an energy difference between \(|M_{J}|=6\) and \(|M_{J}|=5\) levels with the same direction of ligand spin with respect to the Tb total angular momentum (values for parallel and antiparallel configurations are very close). As seen in Table 2 and Fig. 8, ZFS is similar for **M1**, **M2** and **M3** molecules. This indicates that ZFS is not sensitive to the ligand type and geometry details, similarly to \(J_{\rm ex}\). However, the ZFS for the neutral molecules is somewhat larger than the values for the anionic and cationic molecules. Higher-energy part of the spectrum shows some dependence on ligand type and geometry details but we have less sensitivity than in the case of charged molecules.
We now compare our calculated CASSCF-RASSI-SO results to other theoretical calculations and experimental data. Regarding \(J_{\rm ex}\), a similar value to our value has been obtained from CASSCF-RASSI-SO calculations for a neutral TbPc\({}_{2}\) molecule adsorbed on a Ni substrate using a DFT-optimized atomic structure. Compared to the experimental \(J_{\rm ex}\) from electron paramagnetic resonance spectra for a crystal of **M1** molecules, the sign of our calculated \(J_{\rm ex}\) agrees with experiment but the magnitude is a bit larger than the experimental value. Experiments on TbPc\({}_{2}\)-based single-molecule transistors have shown both antiferromagnetic and ferromagnetic couplings between the ligand spin and the Tb spin. These seemingly conflicting experimental results are not surprising, considering the small energy scale of \(J_{\rm ex}\) and possible large configurational changes of the ligand planes of the TbPc\({}_{2}\) molecule bridged between gold electrodes in single-molecule transistor set-ups. Experimental data for the effective energy barrier is rare for neutral TbPc\({}_{2}\) molecules. There exists an experimental report on a neutral TbPcNc molecular crystal diluted with YPcNc. In this case, the experimental barrier is 342 cm\({}^{-1}\) (Table 5). The measured value seems to be close to our calculated ZFS for **M3** molecule. See Tables 2 and 6.
## 5 Conclusions
We investigate electronic structure and magnetic properties of six different TbPc\({}_{2}\) and TbPcNc molecules in different charge states using relativistic multireference methods and pseudospin Hamiltonian technique. For the charged molecules, we evaluate CF parameters and \(g\)-tensor elements by projecting the low-energy spectrum onto effective \(J=6\) pseudospin. For the neutral molecules, we consider exchange coupling of the Tb magnetic moment with the ligand spin and extract Tb CF parameters and \(g\)-tensor elements by separating the Tb ground multiplet from the unpaired ligand electron spin-flip states using artificial oxidation. The key findings are as follows:
* For the neutral molecules, the exchange coupling constant between the Tb magnetic moment and the ligand spin does not depend much on ligand type and geometry details. This result is valid as long as molecular structures are more or less controlled such as in crystals or layers on substrates.
* Geometry details and ligand type do not affect ZFS.
* ZFS weakly depends on oxidation number. The neutral molecules have somewhat higher ZFS than the charged molecules.
* The higher-energy levels and associated tunnel splittings strongly depend on ligand type, oxidation number, and geometry details.
* Comparison to experimental effective barrier suggests that in some cases higher-energy levels rather than just ZFS may contribute to the magnetization relaxation through phonon-assisted tunneling.
These results provide insights in separating the effects intrinsic to individual molecules from extrinsic effects on magnetization relaxation and in interpretation of reported experimental data and stimulating new experiments on TbPc\({}_{2}\)-type molecules.
This work was funded by the Department of Energy (DOE) Basic Energy Sciences (BES) grant No DE-SC0018326. Computational support by Virginia Tech ARC and San Diego Supercomputer Center (SDSC) under DMR060009N. We also thank Dr. Benjamin Pritchard for helpful discussion and insight.
The Supporting Information is available free of charge: Molecular orbitals, spin-free energies, numeric data for energy levels, basis set and active space dependence, definitions of extended Stevens operators, full set of crystal field parameters, and character of electronic levels.
|
10.48550/arXiv.1905.10635
|
Multireference Ab Initio Studies of Magnetic Properties of Terbium-Based Single-Molecule Magnets
|
Ryan Pederson, Aleksander L. Wysocki, Nicholas Mayhall, Kyungwha Park
| 3,692
|
10.48550_arXiv.2107.12481
|
## 1 **Estrutura**
A estrutura da melanina natural ainda e objeto de discussao na literatura. Contudo, sabe-se que a melanina e composta por unidades de 5,6-dihidroxi-indol em diferentes estados redox, Figura 1. Essas unidades formam principalmente estruturas covalentemente ligadas (com ate 12) em uma mistura heterogenea de tres subestruturas planas empilhadas, atraves de ligacoes \(\pi\)-\(\pi\), com 3,4 A de espacamento e dimensoes aproximadas de 10 A de altura e 20 A de comprimento.\({}^{24-26}\) Vale destacar que, embora as medidas atuais de microscopia sugerema formacao predominante deste tipo de estrutura, ha a formacao de outras estruturas polimericas.\({}^{27}\) E interessante destacar que a presenca de especies polaronicas de nitrogenio, como Ndef\({}^{+}\), nao e comumente apresentada na estrutura da melanina. No entanto, existem fortes evidencias experimentais da sua presenca obtidas por espectroscopia de fotoeletrons excitados por raios X, espalhamento Raman de superficie e espectroscopia infravermelha de que elas existem.\({}^{28-33}\) Tal estrutura esta relacionada a subprodutos formados durante o processo sintee.\({}^{34}\)
A combinacao de estruturas carboxiladas e nao-carboxiladas e uma importante caracteristica das melaninas, uma vez que diferentes proporcoes dos monomeros DHI e DHICA podem influenciar diretamente a estrutura final da macromolecula. Sabe-se que as melaninas ricas em DHICA sao formadas por estruturas menores e mais lineares,\({}^{2,35}\) enquanto que as ricas em DHI tem estrutura maiores e com tendencia globular. Esta diferenca tem relacao com o numero e com a topologia das posicoes das ligacoes cruzadas entre os monomeros. Enquanto DHICA predominantemente polimeriza-se pelas posicoes C-4 e C-7, DHI pode formar adutos pelas posicoes C-2, C-3, C-4 e C-7.
A biossintese de melanina, conhecido como processo de melanogenese, ocorre de maneira generalizada em todos os organismos a partir de precursores de tirosina. Nos seres humanos, a melanogenese se da atraves da hidroxialquilacao da tirosina na presenca de oxigenio para formacao da L-DOPA e/ou oxidacao da L-DOPA para dopaquinona sendo ambos os processos catalisados pela enzima tirosinase (TYR). Entretanto, a dopaquinona e um intermediario bastante reativo que se rearranga e cicliza para gerar o dopacromo. O oligomero de melanina final consiste em varios tipos de monomeros que exibem diferentes estados de oxidacao, derivados de DHI (Figura 1).
O mecanismo de formacao da melanina comecou a ser estudado a partir da decada de 1920, com os trabalhos de Raper e Mason. Segundo Raper-Mason, o dopacromo seria espontaneamente descarboxilado formando os monomeros DHI, que soferiam polimerizacoes oxidativas sucessivas para formacao de outros compostos instaveis e finalmente fomacao do polimero de melanina. Entretanto, este mecanismo nao explica a presenca da grande quantidade de grupos carboxilicos, como o presente no DHICA, na melanina natural. Portanto, considera-se que durante a formacao da melanina duas enzimas derivadas da tirosinase participam do processo cataltico: proteina-1 relacionada a tirosinase (Tyrp1) e proteina-2 relacionada a tirosinase (Tyrp2), tambem chamadas de dopacromo tautomerase (DCT). Foi proposto que, na sintese _in vivo_, a enzima Tyrp2 catalisa a tautomerizacao da dopacromo em DHICA e a Tyrp1 catalisa a oxidacao da DHICA, de tal forma a promover sua polimerizacao, explicando a proporcao de DHICA/DHI maior que 50% na melanina natural. Ainda, durante a sintese da melanina, peroxido de hidrogenio (H\({}_{2}\)O\({}_{2}\)) e formado e se acumula no meio reacional. O mecanismo de formacao do H\({}_{2}\)O\({}_{2}\) e sua funcao durante a melanogenese ainda nao sao completamente compreendidos, mas estudos indicam que tanto H\({}_{2}\)O\({}_{2}\) quanto o anion radical superoxido (O\({}_{2}\)-) modulam a attividade da tirosinase. Enquanto H\({}_{2}\)O\({}_{2}\) e um inibidor competitivo da tirosinase, O\({}_{2}\)- ativa essas enzimas 40 vezes. A tironinase tambem e inibida por tioredoxina e por ditiois. Assim, estas regulacoes podem ser responsaveis pela alteracao na quantidade de monomeros de DHI e DHICA no produto final. Por isso, as condicoes na qual a sintese ocorre sao de extrema importancia, visto que pequenas modificacoes na reacao podem afetar a estrutura polimericae consequentemente as propriedades do material. Por exemplo, enquanto o processo sintetico tradicional de melanina proporciona uma macroestrutura com predominancia de DHI, com proporcao de DHICA proxima de 10%, na sepia-melanina, um tipo natural de melanina obtida a partir de moluscos marinhos da classe _Cephalopoda_, ocorre uma razao DHICA/DHI maior que 50% (Esquema 1).
Devido a insolubilidade e a dificuldade de extracao das melaninas naturais, diversos estudos foram realizados para desenvolver a sintese da melanina _in vitro_. Muitas rotas sinteticas foram propostas para a obtencao de melanina a partir de diferentes precursors, tais como L-DOPA, DHI, DHICA, triosina e dopamina. Embora a dopamina venha ganhando bastante atencao devido a facilidade de processamento e a disponibilidade comercial,, tradicionalmente, a melanina e obtida a partir da oxidacao da L-DOPA, pormeio de reacao enzimatica ou pela auto-oxidacao em solucao aquosa alcalina com o borbulhamento de oxigenio molecular. Ambos os processos sinteticos nao sao finance controlados, resultando em materiais polimericos quimicamente heterogeneos e com diferentes proporcoes DHICA/DHI. Alem disso, assim como na melanina natural, a melanina obtida sinteticamente e insoluvel em agua ou em solventes organicos, o que dificulta a producao de filmes finos.
Bronze-Uhle e colaboradores desenvolveram uma rota sintetica mais rapida do que a tradicional tambem utilizando o borbulhamento de oxigenio molecular. A formacao do produto acontece atraves da polimerizacao oxidativa da L-DOPA sob pressao de oxigenio em meio alcalino. A nova metodologia resultou em uma melanina sintetica homogenea com maior teor de grupos carboxilados, mais mantendo semelhanca com a melanina natural. Neste processo, a oxidacao da L-DOPA na fase aquosa e meio alcalino e promovida com o NH4OH, para desprotonar as hidroximas, levando a ciclizacao e a formacao de dopacromo, que segue uma serie de etapas oxidativas para formar DHI e DHICA. Nesta rota sintetica, a descarboxilacao e evitada, o que resultara em uma maior proporcao de monomeros DHICA.
A obtencao de um derivado de melanina soluvel foi desenvolvida pelo grupo de pesquisa do Professor Carlos Graeff em 2004. Nesta rota, a sintee e controlada o que propicia a obtencao de um material homogeneo, com maior estabilidade termica e solubilidade em solventes organicos como dimetilsulfoxido (DMSO), \(N\),\(N\)-dimetilformamida (DMF) e \(N\)-metilpirrolidona (NMP), possibilitando a producao de filmes finos por _spin coating_ com melhor aderencia nos substratos.
Analogamente a sintese em meio aquoso, esta rota sintetica e baseada na oxidacao da L-DOPA, porem, o agente oxidante e o peroxido de benzoila e o DMSO e utilizado como solvente para substituir a agua. Neste ambiente quimico, o peroxido de benzoila alem de oxidar a L-DOPA para formacao dos monomeros derivados de DHI e DHICA, conforme o modelo de Raper-Mason, tambem oxida o DMSO para formacao do anidrido metanossulfonico. Este subproduto e responsavel pela protecao das hidroximas fenclicas dos monomeros da DHI e DHICA com grupos sulfonados (-\(SO_{2}CH_{3}\)).
Apesar do potencial tecnologico deste material, o tempo sintetico e lento, mais de 50 dias. Assim, uma alteracao sintetica baseada neste mecanismo foi realizada a partir do aquecimento do meio reacional. Observou-se durante o aquecimento da solucao a 100 \({}^{\circ}\)C quea sintese e ainda mais controlada e o tempo reacional diminui para 8 dias. Alem disto, este aquecimento ocasiona uma diminuicao dos grupos carboxilicos nos produtos intermediarios e, consequentemente, no polimero final.\({}^{47}\) Um resumo do mecanismo proposto para a sintese da melanina em DMSO (DMSO-Melanina) e apresentado no Esquema 2. A partir do mecanismo sintetico, foi proposto que a estrutura da DMSO-Melanina e um conjunto de oligomeros com diferentes grupos protetores ligados as hidroxilas fenolicas e ao nitrogenio do anel indol, Esquema 2.
## Esquema 2. Etapas reacionais propostas para a sintese da melanina sulfonada in vitro.
A preparacao da melanina sintetica e um campo com muitos desafios a serem compreendidos e amadurecidos.\({}^{67}\) A melanina e um importante biomaterial e as adaptacoesem sua rota sintetica podem levar a novos produtos multifuncionais com caracteristicas necessarias para as aplicacoes desejadas.
## Propriedades fisico-quinicas
## Propriedades opticas
Diferente de muitos pigmentos naturais, a melanina apresenta uma intensa absorcao otica de banda larga, sem nenhum pico proeminente de transicao eletronica, que diminui de intensidade do ultravioleta ao visivel e ao infravermelho proximo, o que da origem a sua coloracao escura, comportando-se de maneira equivalente a um semicondutor amorfo.\({}^{27}\)
Atualmente, este comportamento optico e explicado atraves de dois modelos de desordens: desordem quimica e desordem fisica. No primeiro caso, a absorcao optica decorre da sobreposicao dos diversos monomeros que compoe sua estrutura.\({}^{68}\) Este modelo foi reforcado com a demonstracao de que os diferentes estados redox contribuem para um espectro amplo de absorcao\({}^{69}\) e por mudancas no espectro de emissao com diferentes energias de excitacao.\({}^{70}\) Este modelo, contudo, nao consegue explicar porque a absorcao e maior na regiao do ultravioleta. Para isto, o modelo da desordem fisica propoe que tal comportamento seja consequencia de folhas de unidades oticas oligomericas de diferentes tamanhos e empilhadas e aleatoriamente.\({}^{71}\)
## Propriedades paramagneticas
Outra propriedade que chama bastante atencao nas melaninas e a presenca de um sinal paramagnetico facilmente detectado em medidas de ressonancia paramagnetica eletronica (EPR) em onda continua.\({}^{72}\) Atualmente, acredita-se que o sinal paramagnetico possa ser uma combinacao de radicais intrinsecos e extrinsecos. Os radicais intrinsecos, recentemente denominados radicais centralos em carbono (CCR, do ingles _carbono-centered radicals_),\({}^{73}\) sao formados durante o crescimento dos oligomeros e agregados da melanina; enquanto os radicais extrinsecos, denominados radicais livres de semiquinona (SFR, do ingles _semiquinone-free radicals_),\({}^{73}\) sao associados aos diferentes estados redox dos monomeros da melanina.\({}^{72,74-76}\) Contudo, mesmo apos decadas de estudo, ainda nao ha consenso a respeito da origem estrutural deste sinal, principalmente com relacao ao CCR.
Considerando que o sinal de EPR pode ser modificado por uma serie de fatores como agentes oxidantes e redutores, luz, pH, temperature, ions metalicos e hidratacao e seguindo o equilibrio de comproporcionamento (Esquema 3), acreditava-se que os radicais do tipo semiquinona dessem origem a este sinal. Recentemente, entretanto, atraves de calculos eletronicos da teoria do funcional de densidade, propos-se que o monomero anionic do tipo indolquinona tambem apresenta caracteristicas espectrais semelhantes ao da semiquinona. Este mesmo trabalho, sugere que as especies CCR sao decorrentes dos monomeros de hidroquinona e de defeitos no nitrogenio neutros (Ndef). Simulacoes espectrais dos sinais experimentais reforcaram tais atribuicoes.
\(\begin{array}{c}\includegraphics[width=142.
appresenta um comportamento de _switch_ reversivel, compativel com o de um semicondutor organico amorfo.\({}^{98,99}\)
Entre os diversos trabalhos que se sucederam, um dos resultados mais importantes foi a caracterizacao do efeito da hidratacao nas propriedades eletricas da melanina, uma vez que a condutividade eletrica varia de \(10^{-13}\) Scm-1, em uma amostra completamente seca, a \(10^{-5}\) Scm-1, quando a melanina esta completamente hidratada. Alem disso, um estudo de coulometria com controle de hidratacao mostrou que, quando em estado timido, os portadores de carga da melanina eram 65% protons e 35% eletrons.\({}^{100,99}\)\({}^{100,101}\)
Outro trabalho bastante relevante foi publicado em 2006 quando a natureza puramente eletronica do transporte de carga da melanina comecou a ser posta a prova. Os autores propuseram que o tipo de carga dominante na melanina era de protons,\({}^{102}\) que modulariam a densidade de eletrons presentes no sistema, semelhante ao efeito do pH na condutividade de polianilina. Este modelo foi, alguns anos mais tarde, aprimorado. Mostert _et al._ propuseram que as propriedades eletronicas da melanina nao eram de origem semicondutora, mas sim que a adsorcao de agua daria origem a ions (protons, H+) e eletrons, seguindo o modelo de comprocorcionamento (Esquema 3), de maneira similar ao efeito do pH nas propriedades paramagneticas da melanina.\({}^{103,104}\) Baseado neste modelo, a melanina e um sistema de condutividade mista ionica-eletronica, seguindo o fenomeno de auto dopagem quimica induzida por agua.\({}^{104}\) Diversos experimentos posteriores fortaleceram este ponto de vista para as propriedades eletricas da melanina.\({}^{65,80,83,84,105,106}\) No entanto, este nao e o fim da linha.
Recentemente, o mesmo grupo de pesquisa questionou o modelo de condutividade mista. Atraves de medidas de ressonancia paramagnetica eletronica induzida por luz e espectroscopia de impedancia e dieletrica, sugeriu-se que a melanina deva ser um condutor exclusivamente ionico/protonico, isto e, sem componente eletronica.\({}^{81,107}\) Porem, outro grupo de pesquisa, liderado pela Professora Santato, revisitou o modelo classico de semicondutividade e forneceu uma nova perspectiva para transporte protonico em biomateriais. Neste caso, o tipo de portador de carga nao e determinado pelas propriedades semicondutoras, mas sim por seus sistemas estendidos de bandas de energia.\({}^{108}\) Neste novo modelo de semicondutividade, a melanina teria ambos os tipos de portadores de carga (ionico e eletronico), mas que mudariam de uma conducao predominantemente eletronica para uma conducao predominantemente ionica/protonica, dependendo do grau de hidratacao.\({}^{108}\) Estenovo modelo de condutividade poderia explicar tanto o comportamento de switch observado em amostras secas,\({}^{109}\) quanto o comportamento ionico em amostras timidas.\({}^{107}\)
## Biocompatibilidade
Testes _in vitro_ de biocompatibilidade mostraram que a melanina nao impede a adesao e proliferacao de celulas de fibroblastos gengivais humanos,\({}^{110}\) celulas-tronco embrionarias\({}^{111}\) e celulas de Schwann.\({}^{112}\) Interessantemente, o derivado sulfonado de melanina mostrou uma maior viabilidade celular quando comparada a melanina padrao.\({}^{110}\)
Testes _in vivo_, por outro lado, demonstraram que filmes de melanina quando implantados em nervos ciaticos possuem baixa propriedades mecanicas e sao capazes de provocar uma resposta inflamatorio similar a implantes de silicone.\({}^{112}\) Vale mencionar que estes mesmos filmes foram totalmente reabsorvidos apos oito semanas.\({}^{112}\)
## ASPECTOS BIOQUIMICOS
## Melanogenese
Como ja mencionado acima, melaninas sao pigmentos polimericos naturais encontrados em uma variedade enorme de seres vivos, desde bacterias ate animais superiores.\({}^{113,114}\) Com composicao heterogenea, e responsavel por caracteristicas visuais como a pigmentacao da pele, cor dos cabelos e olhos, mas apresenta outras funcoes, como atividade termoreguladora em vertebrados inferiores e acao fotoprotetiva em peles de maniferos.\({}^{115,116}\) As melaninas sao sintetizadas em celulas epiteliais especializadas, denominadas de melanocitos, provenientes dos melanoblastos (celulas derivadas da crista neural embrionaria), que se diferenciam em diversos tipos celulares e dentre elas, as celulas pigmentarias.\({}^{117}\)
No interior dos melanocitos, localizam-se as organelas responsaveis pela sinteese e o armazenamento da melanina, os melanosomos, que produzem o pigmento durante a sua propria maturacao. Existem 4 estagios de amadurecimento dos melanosomos: o estagio I, tambem chamado de pre-melanosomo, a organela possui morfologia esferica e matriz amorfa. No estagio II, passa para um formato eliptico, uma matriz fibrilar organizada, porem ainda sem catalisar a sintee do pigmento. Evoluindo para o estagio III, o melanosomo iniciao processo de sintese da melanina, que se deposita nas fibras internas. Na fase final, estagio IV, a melanina ja se deposita sobre todo o interior do melanossomo, que perde sua atividade enzimatica, para ser transportado para as celulas epiteliais vizinhas (queratinocitos) ou para as fibras capilares. Variacoes na distribuicao dos melanossomos e fatores como sua composicao, numero e tamanho, sao responsaveis por diferencas observaveis de pigmentacao.
Como ja mencionado detalhadamente acima, a sintese da melanina tem inicio a partir do aminoacido tirosina, o qual na presenca da enzima _tirosinase_ e oxigenio molecular e na ausencia de cisteina, sofre uma etapa de hidroxilacao, gerando L-DOPA, e subsequente oxidacao para DOPAquinona (DQ). A DQ e uma molecula extremamente reativa, que produz outros intermediarios como a ciclodopa, o dopacromo o DHI e o DHICA, que atraves de ligacoes cruzadas formam oligomeros e posteriormente o polimero. Na presenca de cisteina/glutationa, a DQ reage formando majoritariamente a 5-S-cisteinildopa e glutationildopa, que apos subsequente etapa de oxidacao, formam a feomelanina. As concentracoes de cisteina e de glutationa sao determinantes na producao de um tipo de melanina em detrimento de outro. Baixas concentracoes, levarao preferencialmente a formacao de eumelanina, enquanto altas concentracoes conduzirao a formacao da feomelanina. Como ja mencionado, H\({}_{2}\)O\({}_{2}\) e O\({}^{2^{*}}\) sao moduladores da atividade da tirosinase, e portanto, enzimas antioxidants como a _catalase_, _glutationa peroxidase_, _glutationa reductase_ e _superoxido dismutase_ sao cruciais na regulacao deste processo. A proporcao de grupos tiois conjugados a DQ e diretamente relacionada ao equilbrio homeostatico intracelular e, portanto, o ciclo das pentoses e a regulacao do sistema antioxidante das celulas tambem estao intimamente ligados ao processo de sintese das melaninas. O balanco hormonal tambem assume um papel importante, pois no interior dos melanocitos e queratinocitos ocorre a producao da proteina pro-opiomelanocortina (POMC), precursorsa de varios compostos ativos, como o hormonio estimulante de alfa melanocitos (\(\alpha\)-MSH) e o hormonio adrenocorticotfico (ACTH). Estes peptideos, ligam-se ao receptor de melanocortina 1 (MC1-R) presentes na membrana celular dos melanocitos, sendo capazes de regular processos de proliferacao dos mesmos e estimular a melanogenese.
## Propriedades Fotoquimicas e antioxidantes
Alem destes fatores endogenos, a exposicao a radiacao solar participa de forma ativa no processo de melanogaster. Gracas a sua estrutura abundante em duplas ligacoes, a melanina absorve tanto radiacao UV (200 a 400 nm) quanto luz visvel (400 e 700 nm), sendo que a absorcao luminosa e maxima na regiao do UVB (290-320 nm). De fato, a radiacao solar estimula a produccao de proteinas como a POMC, que induzem o processo hormonal citalo acima. A faixa de absorcao maxima da melanina coincide com a absorcao de outras moleculas importantes presente da pele de humanos, como as bases nitrogenadas pirimidinas constituintes do DNA, que ao absorverem radiacao UVB podem formar fotoprodtos carcinogenicos. Sendo assim, as melaninas exercem um papel importante de protecao contra a mutag enese induzida principalmente por radiacao UVB e por isso sao reconhecidas como moleculas fotoprotetoras. No entanto, a melanina tambem gera especies excitadas ao interagir com radiacao UVA (320-400nm) e luz visivel, gerando especies triplets e oxigenio singlete (\({}^{1}\)O\({}_{2}\)), que podem reagir com biomoleculas, gerando efeitos nocivos e ate mesmo mutag enicos. Allem de gerar diversas especies reativas por excitacao eletronica, as melaninas tambem reagem com essas especies reativas, funcionando com agente antioxidante de sacrificio. \({}^{1}\)O\({}_{2}\), por exemplo, adiciona-se aos aneis indolicos presentes nas melaninas, formando hidroperoxidos e causando o foto-branqueamento das melaninas.
A oxidacao e um processo inerente ao metabolismo celular, e consequentemente, a formacao de radicais e outros compostos reativos ocorre naturalmente em nosso organismo, podendo ser potencializada ou tambem desencadeada por fatores externos. Alguns dos compostos gerados sao genericamente chamados de especies reativas de oxigenio (EROs), que incluem compostos radicalares como o radical hidroxila (HO'), superoxido (O\({}_{2}\)^), e peroxida (ROO') e nao radicalares, como o peroxido de hidrogenio (H\({}_{2}\)O\({}_{2}\)). Estas especies podem reagir com moleculas biologicamente ativas, alterando o regime de homeostasisa celular e podendo gerar patologias, que podem acometer diversas regioes do corpo humano (pele, olhos, mucosa). O excesso destes compostos oxidantes e anulado atraves de compostos antioxidantes, que podem sequestra-los, transformando-os em compostos menos reativos. A acao de um antioxidante pode ocorrer atraves de mecanismos fisicos e/ouquimicos. A melanina atua tambem desta forma, absorvendo a luz e neutralizando especies radicalares prejudiciais.\({}^{132}\)
Apesar de a melanina ser capaz de sequestrar radiciais livres, ela tambem e capaz de gera-los apos um processo de fotoexcitacao por luz UV ou visivel.\({}^{135}\) Em geral, a extensao de suas propriedades antioxidantes ou pro-oxidantes serao governadas, de acordo com o tipo de pigmento produzido. Ja se constatou que a feomelanina pode estimular a producao de radiciais livres ao passo que a eumelanina possui maior capacidade de consumi-los e portanto, ter um efeito predominantemente protetivo.\({}^{132}\) Sendo assim, quanto maior a razao entre eumelanina e feomelanina, menos pronunciada sera a manifestacao das especies reativas e, portanto, maior o efeito antioxidante. Estudos em culturas de celulas epiteliais variando fenotipos de I a VI (o nivel de pigmentacao aumenta de forma crescente de I a VI), mostram que celulas contendo maiores concentracoes de eumelanina sao mais resistenes aos danos fotoquimicos e apresentam maior taxa de sobrevida pos fotossensibilizacao com radiacao UV, em especial UVB.\({}^{136}\) A capacidade da eumelanina de neutralizar os radiciais livres depende do numero de grupos passiveis de softerem oxidacao/reducao presentes nas unidades, bem como, de seus potenciais redox, e tambem da facilidade de acesso destes grupos aos radiciais.\({}^{137}\)
Durante o processo de polimerizacao das unidades constitutivas da melanina, ocorre a formacao de intermediarios que possuem uma alta taxa de transferencia de eletrons e de hidrogenio, o que as tornam moleculas susceptiveis a absorcao da radiacao visivel e UV. Apos a irradiacao dessas entidades, ocorre a formacao de radiciais superoxido, e estes podem reagir com os intermediarios da rota sintetica, podendo por exemplo, reduzir quinonas a semiquinonas, e ao mesmo tempo oxidar hidroquinonas a semiquinonas ou quinonas.\({}^{138}\) Como este processo diminui a concentracao de radiciais superoxido, por consequencia tambem reduz a concentracao de radiciais hidroxila, que e reativo e consequentemente mais deletero. Outra rota de protecao se da pela habilidade da melanina em quelar metais, como fions Fe\({}^{2+}\), o que suprime a reacao de Fenton, principal geradora deste ultimo radical.\({}^{138,139}\) Estas evidencias corroboram com a hipotese de a melanina ser uma molecula capaz de reduzir os niveis de radiciais livres presentes no ambiente celular.
No entanto, com o envelhecimento celular, a exposicao constante de celulas pigmentarias a um ambiente com elevadas concentracoes de oxigenio e constante exposicaoa radiacao solar podem depreciar essa atuacao, levando ate mesmo a causar o efeito inverso, elevando as especies oxidantes.\({}^{140}\) Essa mudanca de atividade pode estar relacionada as modificacoes estruturais sofridas em decorrencia da fotodegradacao de suas unidades.\({}^{141}\) Alem disso, um dos fatores de maior influenza pode ser atribuido a acumulo de lipofuscina, um pigmento associado ao envelhecimento celular, que se deposita no citoplasma de celulas danilicadas e que produz oxigenio singlete por fotossensibilizacao na regiao espectral do visivel.\({}^{142,143}\) A resistencia ao envelhecimento e a perda das propriedades antioxidantes da melanina tambem esta diretamente relacionada a proporcao entre eumelanina/feomelanina presente.\({}^{132}\)
Para enriquecer as teorias acerca destes fenomenos, muitos estudos tem utilizado seus analogos sinteticos para avaliacao de suas propriedades, como habilidade sequestrante de radicais, paramagnetismo, taxas de troca de ions e estrutura.\({}^{137}\) O estudo por meio de modelos parece vantajoso, pois sua sintese e controlada e costumam conter estruturas de elevada solubilidade. Por exemplo, a DOPA e polidopamina (PDA) sao polimeros utilizados como modelos de estudo,\({}^{144}\) assim como DHI e DHICA sao extensivamente estudados como antioxidantes, gracas as suas ligacoes O-H fenolicas e N-H indolicas, que sao capazes de sequestrar radicais livres.\({}^{145}\) Estes compostos, portanto, auxiliam no estudo do comportamento da melanina observado _in vivo_.
## Melanina e protecao solar
A luz solar apresenta um espectro continuo de radiacao eletromagnetica, que e arbitrariamente dividida em tres grandes faixas: UV, visivel e infravermelho. A radiacao UVC e altamente reativa, mas e completamente absorvida pela camada de ozonio, que tambem absorve parte da radiacao UVB.\({}^{146-148}\) A interacao da radiacao solar com a pele sofre influencia das caracteristicas absorvedoras de luz e fotoquimicas dos cromoforos presentes na pele, que tambem afeta a profundidade de penetracao da radiacao na pele.\({}^{132,149,150}\) Sabese que a exposicao a luz solar pode causar efeitos agudos, como eritema ou bronzeamento, porem a exposicao constante e prolongada pode provocar efeitos cronicos como melasma, fotoenvelhecimento, podendo levar inclusive a formacao de tumores. Porem, e importante ressaltar que a radiacao solar nao traz apenas maleficios. Os raios UVB sao os principais responsaveis pelo inicio das reacoes intracelulares que levam a formacao e posterior ativacaoda vitamina D, e este composto tem uma infinidade de attividades biologicas, podendo inclusive inibir a proliferacao de celulas tumorais, atraves do estimulo o aumento do numero de celulas apoptoticas.\({}^{149,151,152}\)
Durante o processo de melanogenese, os dendritos presentes nos melanocitos transferem os granulos de pigmento para cerca de 30 a 40 queratinocitos,\({}^{128}\) realizando seu deposito ao redor do nucleo destas celulas. Isso nao ocorre por acaso, mas devido a capacidade da melanina em proteger o DNA nuclear da radiacao UVB.\({}^{123}\) A eumelanina e particularmente responsavel por este fenomeno, gracas ao seu baixo rendimento quantico de geracao de especies excitadas, ou seja, a razao molecular de formacao de transientes reativas por fotons absorvidos e muito baixa.\({}^{152}\) Isso leva a concluir que a eumelanina dissipa a maior parte da energia que recebe em calor, uma caracteristica de moleculas fotoprotetivas.\({}^{27,70}\) Este fato, aliado ao poder sequestrante de radicais livres, mostra que a melanina assume um papel importante e vital na protecao do organismo contra radiacao. Ja a feomelanina, se mostra um forte fotossensibilizador, uma vez que e capaz de gerar maiores concentracoes de oxigenio singlete apos exposicao a radiacao UVA e a luz visivel, do que a eumelanina, sob as mesmas condicoes.\({}^{131,152,153}\) Por isso e importante ressaltar que o comportamento da melanina pode ser variavel, dependendo da razao entre as quantidades de cada pigmento.\({}^{152-157}\)
Segundo Meredith & Sarna,\({}^{27}\) podemos classificar a reacoes fotoinduzidas da melanina em dois subtipos, os processos anaerobios e aerobios. O primeiro, pode ser iniciado por radiacao UV-visivel, em comprimentos de onda acima de 300 nm. Nestes processos, a radiacao ira deslocar o equilibrio redox existente entre as proprias subunidades formadoras da melanina, que se encontram em estado de equilibrio em suas formas oxidadas e reduzidas, causando transferencia continua de eletrons, que com a dose extra de energia, poderao sofrer deslocamentos e na presencia de outros compostos, podem formar novos radicais. No segundo caso, em presenca de oxigenio, a radiacao UV-visivel, em frequencias maiores, pode originar as especies reativas ja mencionadas, como \({}^{1}\)O\({}_{2}\), o H\({}_{2}\)O\({}_{2}\) e \({}^{O_{2}\thicksim}\), responsaveis pela fotodegradacao da melanina.\({}^{27,155,157}\)
Pathak et al\({}^{154}\) realizaram ensaios histoquimicos, confirmando a presenca de melanina recem-formada, assim como presenca de attividade recente da enzima _tirosinase_ em amostras de celulas de pele humana irradiadas por radiacao UV e luz visivel, demonstrando que este estimulo implica em aumento do processo de melanogenese. Estes dados abriramcaminho para as propostas de mecanismos de formacao e estimulo da melanina apos processos de fotossensibilizacao. Hoje sabemos que a radiacao UV e a luz visivel aumentam a proliferacao e diferenciacao das celulas melanocticas, ao promover a producao dos hormonios precursores da melanogenese nos queratinocitos, como endotelina 1 (estimulator da \(a\)-MSH) e a proteina POMC, e estes por sua vez atuam de forma paracrina nos melanocitos com os quais possuem contato. Ela tambem atua diretamente nas celulas pigmentarias ao estimular tanto a POMC, quanto seu receptor, e tambem promovendo enzimas como a _tirosinase_ e _Tyrp1_, dentre outros fatores.
Apesar de ainda existirem controversias, estudos recentes vem mostrando uma forma de participacao da melanina nas mutacoes provocadas pela exposicao a radiacao UV, e consequentemente no aparecimento de patologias como o cancer. Ao softerem excesso de exposicao ao sol, todo o balanco redox no ambiente celular (homeostase) sofre uma desregulacao, que resulta no aumento de especies oxidativas, como o anion radical superoxido, ja mencionado anteriormente e o radical oxido nitrico (NO\({}^{\bullet}\)). A juncao destes dois compostos pode dar origem a uma terceira entidade, o peroxinitrito (ONOO'). Este anion e um potente oxidante, naturalmente produzido em nosso organismo, e capaz de degradar a melanina. Uma vez presentes no citoplasma, os fragmentos resultantes da degradacao de eumelanina ou feomelanina, tornam-se mais soluveis e sao difundidos ate o nucleo celular, onde sofferao troca de energia e irao formar, em uma serie de reacoes, especies eletronicamente excitadas, capazes de formarem compostos como os dimeros de ciclobutano de pirimidina (CPDs, do ingles _cyclobutane pyrimidine dimers_). A geracao de oxigenio singlete tambem aumenta substancialmente em melaninas degradas, gerando lesoes oxidativas em DNA nuclear. Estes compostos sao atualmente tidos como causadores de mutagenicidade, e logo, podendo sofrer influencia da presenca de melanina no seu processo de formacao.
Diante destas evidencias, e possivel perceber que o papel fotoprotetor da melanina nao exclui o fato de que concomitantemente, este biopolimero seja pivo em diversas outras reacoes, gerando fotoprodutos, radicais livres, intensificando processos de oxidacao, decomposicao ou ate mesmo de mutacoes e morte celular. Portanto, independentemente do nivel de melanina na pele de um individuo, deve-se sempre evitar o excesso de exposicao ao sol, mesmo com o uso dos filtros solares, uma vez que os efeitos decorrentes da fotossensibilizacao da melanina cocorrem tambem na regiao espectral do visivel.\({}^{147,161}\) A acao antioxidante parece ser uma maneira de proteger os melanocitos de danos induzidos por exposicao a luz visivel.
## Aplicações Tecnologicas
Tendo em vista as diversas propriedades descritas acima, a melanina como uma classe de materiais foi considerada para uma vasta gama de aplicacoes tecnologicas.
Em relacao a aplicacoes biomedicas, a liberalcao controlada de particulas de melanina em ferimentos na pele mostrou-se capaz de reduzir a inflamacao do ferimento.\({}^{162}\) Alem disso, nanoparticulas de melanina funcionalizada com PDA\({}^{163}\) ou ainda o recobrimento de matrizes ceramicas com melanina\({}^{164}\) relevou o potencial do material para _drug-delivery_. Em camundongos, tambem apresentou efeitos de protecao do figado contra o estresse induzido pelo alcool\({}^{165}\) ou ainda, em dietas ricas em gordura, prevencao e controle de hiperlipidemia.\({}^{166}\) Melanina tem sido apontada como um biomarcador para melanomas,\({}^{167}\) e tambem como um agente terapeutico fototetermico capaz de matar celulas tumorais por hipertermia quando irradiado com radiacao na regiao do infravermelho proximo.\({}^{168-170}\)
Os diferentes estados de oxidacao da melanina permitem que ela seja capaz de ligar com farmacos e metais. Esta caracteristica possibilitou a sua utilizacao em terapia guiada por imagem,\({}^{171,172}\) agentes de contraste fotoacustico e de ressonancia magnetica\({}^{173}\) e propriedades antibacterialas.\({}^{174}\) Esta mesma propriedade tambem pode ser utilizada para purificacao de aguas poluidas com corantes,\({}^{175}\) metais pesados como uranio\({}^{176}\) ou farmacos (como cloroquina).\({}^{177}\) Devido as suas propriedades opticas, a melanina tambem foi utilizada como agentes foto-estabilizadores de la\({}^{178}\) e plasticos.\({}^{179}\) Alem disso, tambem foi mostrado que ela e capaz de aumentar a protecao contra radiacao ultravioleta e estabilidade termica de polimeros.\({}^{180}\)
Do mesmo modo, a melanina e amplamente utilizada na eletronica organica. Nesta area, uma importante caracteristica e producao de filmes-finos homogeneos. Devido a sua natureza insoluvel, houve grande esforco para obtencao de filmes com qualidade para desenvolvimento de dispositivos nas ultimas duas decadas.\({}^{61,62,65,111}\) Todo o desenvolvimento do processamento de filmes de melanina possibilitou a sua utilizacao como eletrolito solidoem transistores eletroquimicos organicos capaz de transduzir corrente ionica em eletronica.\({}^{181}\) Ou ainda em baterias de sodio\({}^{182}\) e magnesio,\({}^{183}\) capacitores,\({}^{184}\) supercapacitores operando em eletrolitos aquososos\({}^{185,186}\) ou em estado solido.\({}^{187}\) Aplicacoes fotovoltaicas como em celulas solares sensibilizadas por corantes foram estudas com melaninas naturais\({}^{188,189}\) e melaninas eletrodepositadas.\({}^{190}\) Alem disso, a melanina tambem foi utilizada como camada ativa para aplicacoes em sensores de pH.\({}^{191,192}\)
## CONCLUSAO E PERSPECTIVAS FUTUREAS
No presente trabalho apresentamos as melaninas e potenciais aplicacoes. As melaninas vem despertando um interesse crescente quanto ao seu papel biologico assim como em aplicacoes na eletronica e bioeletronica. Como consequencia ha um numero crescente de grupos de pesquisa estudando-as. Apesar disso, ha varios aspectos de sua estrutura e funcionalidade que ainda nao sao entendidos. Do ponto de vista das aplicacoes, um dos grandes desafios, que era a producao de filles finos, parece estar resolvido. Nesta revisao apresentamos varios exemplos envolvendo a combinacao de solvente com a funcionalizacao do polimero. No entanto, muitas questoes permanecem, entre elas: qual e a natureza do transporte de cargas, o transporte e ionico, eletronico ou uma mistura dos dois? Seria possivel dopar a melanina e aumentar sua condutividade? Qual a relacao do sinal de ressonancia paramagnetica com o transporte de cargas? Outra questao importante para a aplicacao em larga escala e um melhor controle e diminuica do tempo de sintese. Do ponto de vista biologico, discutimos os mecanismos de ativacao da melanogenese e as respectivas propriedades fotoquimicas e antioxidantes das melaninas. Nao restam dtividas de que as melaninas exercem um papel protetor e anti-oxidante importante, mas que tambem podem causar danos em tecidos biologicos. Embora a absorcao da radiacao UV e da Luz visivel pelas melaninas ofereca protecao contra danos no DNA das celulas epiteliais, as melaninas, em especial a feomelanina, tambem geram especies reativas, incluindo oxigenio singlete, causando danos oxidativos em DNA por fotossensibilizacao. Discutimos a consequencia desta dicotomia nas estrategias atuais de fotoprotecao. Ha ainda muitas incognitas nos mecanismos de transformacao maligna que levam melanocitos a se transformar em melanoma, mas ja ha informacoes suficiente que sugerem que as melaninas e a melanogenee podem ter um papel importante neste processo.
## Agradecimentos
Os autores agradecem o apoio da Fundacao de Amparo a Pesquisa do Estado de Sao Paulo (FAPESP, processos: 2013/07296-2, 2015/23000-1), Conselho Nacional de desenvolvimento Cientifico e Tecnologico (CNPq) e Coordenacao de Aperfeicoamento de Pessoal de Nivel Superior (CAPES). J.V.P e C.F.O.G tambem agradecem ao Prof. Jose H. D. da Silva (UNESP) por ter inspirado a elaboracao deste trabalho.
|
10.48550/arXiv.2107.12481
|
Melanina, um pigmento natural multifuncional
|
Joāo V. Paulin, Barbara Fornaciari, Bruna A. Bregadiolli, Mauricio S. Baptista, Carlos F. O. Graeff
| 1,657
|
10.48550_arXiv.2103.14119
|
###### Abstract
We present a new non-adiabatic ring polymer molecular dynamics (NRPMD) method based on the spin mapping formalism, which we refer to as the spin-mapping NRPMD (SM-NRPMD) approach. We derive the path-integral partition function expression using the spin coherent state basis for the electronic states and the ring polymer formalism for the nuclear degrees of freedom (DOFs). This partition function provides an efficient sampling of the quantum statistics. Using the basic property of the Stratonovich-Weyl transformation, we derive a Hamiltonian which we propose for the dynamical propagation of the coupled spin mapping variables and the nuclear ring polymer. The accuracy of the SM-NRPMD method is numerically demonstrated by computing nuclear position and population auto-correlation functions of non-adiabatic model systems. The results from SM-NRPMD agree very well with the numerically exact results. The main advantage of using the spin mapping variables over the harmonic oscillator mapping variables is numerically demonstrated, where the former provides nearly time-independent expectation values of physical observables for systems under thermal equilibrium, the latter can not preserve the initial quantum Boltzmann distribution. We also explicitly demonstrate that SM-NRPMD provides invariant dynamics upon various ways of partitioning the state-dependent and state-independent potentials.
## I Introduction
One of the central challenges in theoretical chemistry is to accurately simulate chemical reactions involving non-adiabatic processes and nuclear quantum effects. These reactions, such as the electron transfer, proton-coupled electron transfer, or the scattering reactions involving non-adiabatic transitions among many electronic states and nuclear quantum effects, are commonly encountered from photo-catalysis, biochemistry and enzymatic reactions, to astrochemistry. Developing accurately yet numerically efficient approaches became a key focus in physical chemistry.
To this end, a large number of these approaches are developed, including the popular trajectory surface-hopping method (mixed quantum-classical approach), the Linearized semi-classical (LSC) path-integral approaches, partially linearized density matrix (PLDM) path-integral approaches, the mixed quantum-classical Liouville equation, and the symmetrical quasi-classical (SQC) approach, to name a few. Despite their successes, these approaches generally do not preserve quantum detailed balance or zero-point energy (ZPE) associated with the nuclear degrees of freedom (DOFs), and often suffer from numerical issues such as ZPE leakage.
Imaginary-time path-integral approaches, such as the ring polymer molecular dynamics (RPMD), resemble classical dynamics in an extended phase space and provide a convenient way to compute approximate quantum time-correlation functions. The classical evolution of RPMD preserves its initial quantum distribution captured by the ring polymer Hamiltonian, and it is free of the zero-point energy leaking problem. Despite its success in describing quantum effects in the condensed phase, RPMD is limited to one-electron non-adiabatic dynamics or nuclear quantization, as well as the lack of real-time electronic coherence effects.
Recently emerged state-dependent RPMD approaches, such as non-adiabatic RPMD (NRPMD), mapping variable RPMD (MV-RPMD), and coherent state RPMD (CS-RPMD) are promising to provide accurate non-adiabatic dynamics with an explicit description of electronic states, in addition to the reliable treatment of nuclear quantum effects through ring polymer quantization. The common ingredient of these approaches is the Meyer-Miller-Thoss-Stock (MMST) mapping formalism, which maps \(N\) electronic states onto \(N\) singly excited harmonic oscillators. The electronic non-adiabatic dynamics are hence mapped onto the phase space trajectories of the mapping oscillators, which evolve together with the nuclear ring polymer. Hence, these MMST-based RPMD approaches can be viewed as unified theories of the mapping oscillators and the ring polymer. These methods are shown to provide both accurate non-adiabatic dynamics as well as nuclear quantum effects. In particular, the NRPMD approach has been rigorously derived from the non-adiabatic Matsubara dynamics framework.
One potential limitation of these state-dependent RPMD approaches is rooted in the MMST mapping representation. It is well known that the MMST representation has a larger size of Hilbert space than the original electronic subspace, and requires projection back to that subspace to obtain accurate results. In addition, the total population along a single trajectory is not guaranteed to be unitary, hence breaking the dynamicalinvariance under different ways of partitioning the potentials into the state-dependent and state-independent components. Besides the widely used MMST representation, there exist other mapping formalisms based upon spin coherent states. In particular, a new spin mapping formalism based on the Stratonovich-Weyl transform was recently developed by Runeson and Richardson. In this spin-mapping approach, two electronic states are mapped onto two angles defining the spin coherent state on the Bloch sphere. One of the advantages of this approach, compared to the MMST formalism, is that the dimensionality of the spin coherent state basis is of the same size of the electronic states of the original system, hence it provides a more consistent mapping than the MMST approach and it does not require additional projections back to the electronic subspace. The spin mapping (SM) variables, being bounded on the Bloch sphere, also guarantees the total population along a single trajectory to be unitary. This further enforces the independence of the dynamics to the splitting between the state-dependent and state-independent parts of the Hamiltonian. It has been shown that in the LSC and the PLDM approaches, using spin-mapping approach provides more accurate non-adiabatic dynamics compared to the corresponding approaches when using the MMST formalism. These exciting theoretical developments of the spin mapping variables motivate us to develop the NRPMD approach with the spin mapping representation.
In this paper, we develop a new non-adiabatic RPMD method which we refer to as the spin mapping NRPMD (SM-NRPMD) approach, based on the recently developed spin mapping formalism. We first derive a partition function formalism based on the SM representation that allows one to efficiently sample the exact quantum statistics. We then derive a SM-NRPMD Hamiltonian for propagating dynamics. With the proposed SM-NRPMD approach, we compute the Kubo-transformed position and population auto-correlation functions with non-adiabatic model systems, and demonstrate that this approach is capable of accurately describing both the correct quantum statistics as well as the electronic Rabi oscillations. Compared to the MMST-based NRPMD approaches, SM-NRPMD seems to preserve the quantum detailed balance, resulting in a nearly time-independent expectation value of the nuclear position or population for the system under the thermal equilibrium. Finally, we demonstrate that the dynamics is invariant of the partitioning of the potential into the state-dependent and the state-independent components.
## II Basic theory of the spin mapping formalism
In this section, we review the spin-mapping for electronic states introduced by Runeson and Richardson. A comprehensive introduction of this material can be found in Ref..
The total Hamiltonian operator of the system is
\[\hat{H}=\frac{\hat{P}^{2}}{2m}\hat{\mathcal{I}}+U_{0}(\hat{R})\hat{\mathcal{I}} +\begin{pmatrix}V_{1}(\hat{R})&\Delta(\hat{R})\\ \Delta(\hat{R})&V_{2}(\hat{R})\end{pmatrix}, \tag{1}\]
The Hamiltonian can also be written in terms of the spin operator as
\[\hat{H}=H_{0}\hat{\mathcal{I}}+\frac{1}{\hbar}\mathbf{H}\cdot\hat{\mathbf{S}} =H_{0}\hat{\mathcal{I}}+\frac{1}{\hbar}(H_{x}\cdot\hat{S}_{x}+H_{y}\cdot\hat{S }_{y}+H_{z}\cdot\hat{S}_{z}), \tag{2}\]
where \(\hat{\mathcal{I}}\) is the \(2\times 2\) identity matrix, \(\hat{S}_{i}=\frac{\hbar}{2}\hat{\sigma}_{i}\) (for \(i\in\{x,y,z\}\)) is the quantum spin operator, with \(\hat{\sigma}_{i}\) being the Pauli matrices expressed as follows
\[\hat{\sigma}_{x}=\begin{pmatrix}0&1\\ 1&0\end{pmatrix},\ \hat{\sigma}_{y}=\begin{pmatrix}0&-i\\ i&0\end{pmatrix},\ \ \hat{\sigma}_{z}=\begin{pmatrix}1&0\\ 0&-1\end{pmatrix}. \tag{3}\]
Various components of the Hamiltonian in Eq. 2 are expressed as
\[H_{0} =\frac{\hat{P}^{2}}{2m}+U_{0}(\hat{R})+\frac{1}{2}(V_{1}(\hat{R}) +V_{2}(\hat{R})), \tag{4a}\] \[H_{x} =2\Re(\Delta(\hat{R})),\] (4b) \[H_{y} =2\Im(\Delta(\hat{R})),\] (4c) \[H_{z} =V_{1}(\hat{R})-V_{2}(\hat{R}), \tag{4d}\]
Note that for a molecular Hamiltonian, one often has \(\Im(\Delta(\hat{R}))=0\).
Following the original work on spin-mapping variables, we introduce the spin coherent state (SCS) basis
\[|\mathbf{u}\rangle=\cos\frac{\theta}{2}e^{-i\varphi/2}|1\rangle+\sin\frac{ \theta}{2}e^{i\varphi/2}|2\rangle, \tag{5}\]
The SCS vector is normalized \(\langle\mathbf{u}|\mathbf{u}\rangle=1\).
\[S_{i}(\mathbf{u})=\langle\mathbf{u}|\hat{S}_{i}|\mathbf{u}\rangle=\frac{\hbar }{2}u_{i},\ \ \ \ i\in\{x,y,z\}, \tag{6}\]
where \(u_{x}\), \(u_{y}\), and \(u_{z}\) are expressed as follows
\[u_{x} =\sin\theta\cos\varphi, \tag{7a}\] \[u_{y} =\sin\theta\sin\varphi,\] (7b) \[u_{z} =\cos\theta. \tag{7c}\]
We further introduce three different functions to define for the _Stratonovich-Weyl_ (SW) transformation of any operator in the SM representation and hence obtain expectation values. They are the Q-, P- and W-functions.
These functions depend on the _kernel_\(\hat{w}_{\rm s}\) and the _spin radius_\(r_{\rm s}\) as follows
\[\hat{w}_{\rm s}({\bf u}) =\frac{1}{2}\hat{\mathcal{I}}+r_{\rm s}{\bf u}\cdot\hat{\mathbf{\sigma }},\ \ \ \ {\rm s}\in\{{\rm Q},{\rm P},{\rm W}\}, \tag{8a}\] \[r_{\rm Q} =\frac{1}{2},\ \ r_{\rm P}=\frac{3}{2},\ \ r_{\rm W}=\frac{\sqrt{3}}{ 2}, \tag{8b}\]
The SCS projection operator is \(|{\bf u}\rangle\langle{\bf u}|=\cos^{2}\frac{\theta}{2}|1\rangle\langle 1|+ \cos\frac{\theta}{2}\sin\frac{\theta}{2}e^{-i\varphi}|1\rangle\langle 2|+ \cos\frac{\theta}{2}\sin\frac{\theta}{2}e^{i\varphi}|2\rangle\langle 1|+ \sin^{2}\frac{\theta}{2}|2\rangle\langle 2|\).
\[\hat{w}_{\rm Q}=|{\bf u}\rangle\langle{\bf u}|, \tag{9}\]
On the other hand, \(\hat{w}_{\rm P}\) and \(\hat{w}_{\rm W}\) do not have simple relations with \(|{\bf u}\rangle\langle{\bf u}|\).
### Spin Mapping of Diabatic Electronic States
The SW transform of an operator \(\hat{A}\) is defined as
\[A_{\rm s}({\bf u})\equiv[\hat{A}]_{\rm s}({\bf u})={\rm Tr}_{\rm e}[\hat{A} \hat{w}_{\rm s}], \tag{10}\]
Mapping an operator \(\hat{A}\) onto the spin Hilbert subspace corresponds to the following relation
\[\hat{A}\to A_{\rm s}({\bf u})={\rm Tr}_{\rm e}[\hat{A}\hat{w}_{\rm s}]. \tag{11}\]
Generalizing the theory to many states is also possible by using the generators of the SU(\(N\)) Lie algebra (when \(N=3\) it corresponds to the Gell-Mann matrices in the \(SU\)-symmetry theory of quarks).
For the \({\rm s}={\rm Q}\) special case, this mapping relation means that
\[A_{\rm Q}({\bf u})={\rm Tr}_{\rm e}[\hat{A}\hat{w}_{\rm Q}]={\rm Tr}_{\rm e}[ \hat{A}|{\bf u}\rangle\langle{\bf u}|]=\langle{\bf u}|\hat{A}|{\bf u}\rangle. \tag{12}\]
The Q-relation maps the spin operator \(\hat{S}_{i}\) with \([\hat{S}_{i}]_{\rm Q}=\frac{\hbar}{2}u_{i}\), which is its expectation value in the SCS through Eq. 6.
Using the spin-mapping defined in Eq. 11, it is easy to show that \([\hat{\mathcal{I}}]_{\rm s}({\bf u})=1\) (because \({\rm Tr}_{\rm e}\hat{\sigma_{i}}=0\) for all \(i\)), as well as
\[{\bf S}_{\rm s}({\bf u})\equiv[\hat{\bf S}]_{\rm s}({\bf u})={\rm Tr}_{\rm e} [\frac{\hbar}{2}\hat{\mathbf{\sigma}}(\frac{1}{2}\hat{\mathcal{I}}+r_{\rm s}{\bf u }\cdot\hat{\mathbf{\sigma}})]=\hbar r_{\rm s}{\bf u}. \tag{13}\]
The projection operators are transformed as
\[[|1\rangle\langle 1|]_{\rm s}({\bf u}) =[\frac{1}{2}\hat{\mathcal{I}}+\frac{1}{\hbar}\hat{S}_{z}]_{\rm s }({\bf u})=\frac{1}{2}+r_{\rm s}\cos\theta, \tag{14a}\] \[[|2\rangle\langle 2|]_{\rm s}({\bf u}) =[\frac{1}{2}\hat{\mathcal{I}}-\frac{1}{\hbar}\hat{S}_{z}]_{\rm s }({\bf u})=\frac{1}{2}-r_{\rm s}\cos\theta,\] (14b) \[[|1\rangle\langle 2|+|2\rangle\langle 1|]_{\rm s}({\bf u})=2[\frac{1}{ \hbar}\hat{S}_{x}]_{\rm s}({\bf u})=2r_{\rm s}\sin\theta\cos\varphi. \tag{14c}\]
The Hamiltonian in Eq. 2 is mapped as \(\hat{H}\to[\hat{H}]_{\rm s}({\bf u})\), with the following expression
\[H_{\rm s}({\bf u})\equiv[\hat{H}]_{\rm s}({\bf u})=H_{0}+r_{\rm s }{\bf H}\cdot{\bf u} \tag{15}\] \[=\frac{P^{2}}{2m}+U_{0}+(\frac{1}{2}+r_{\rm s}\cos\theta)\cdot V _{1}+(\frac{1}{2}-r_{\rm s}\cos\theta)\cdot V_{2}\] \[\quad+2r_{\rm s}\sin\theta\cos\varphi\cdot\Delta.\]
Note that \(H_{0}\) and \({\bf H}\) are in principle \(R\)-dependent. The SW mapping is closely related to the MMST mapping approach, and a brief discussion between these two formalisms is provided in Appendix A, whereas a thorough comparison can be found in Ref..
To obtain the equations of motion (EOM) governed by \(H_{\rm s}({\bf u})\) in Eq. 15 for the spin-mapping variables, we start with the following Heisenberg EOM for \(\hat{\bf S}\)
\[\frac{d}{dt}\hat{\bf S}=\frac{1}{i\hbar}[\hat{\bf S},\hat{H}]=\frac{1}{\hbar}{ \bf H}(\hat{R})\wedge\hat{\bf S}, \tag{16}\]
2) is \(\hat{\bf S}\)-independent, hence commutes with \(\hat{\bf S}\). Applying the SW transform (Eq. 10) on both sides of the above equation, we have
\[\frac{d}{dt}{\bf u}=\frac{1}{\hbar}{\bf H}(\hat{R})\wedge{\bf u}. \tag{17}\]
Note that the above equation is exact, regardless of the \(r_{\rm s}\)-dependence of \(\hat{\bf H}\). Of course, the EOM for the nuclear DOF is not yet explicitly expressed. When choosing the Wigner representation for the nuclei and using the quantum-classical Liouville equation (QCLE), Eq. 17 can also be rigorously derived. This equation can be solved by treating \({\bf u}\) as dynamical variables, or equivalently, \(\theta\) and \(\varphi\) as dynamical variables. Further analysis of this is provided in Appendix B.
### Properties of the Stratonovich-Weyl Transform
Here, we briefly summarize several basic properties of the Stratonovich-Weyl transform, which will be used to derive the quantum partition function and the spin-mapping NRPMD Hamiltonian in the next section.
\[{\rm Tr}_{\rm e}[\hat{A}] =\int{\rm d}{\bf u}\langle{\bf u}|\hat{A}|{\bf u}\rangle=\int{\rm d }{\bf u}A_{\rm Q}({\bf u}) \tag{18}\] \[=\frac{1}{2\pi}\int_{0}^{\pi}{\rm d}\theta\sin\theta\int_{0}^{2 \pi}{\rm d}\varphi A_{\rm Q}(\theta,\varphi),\]
Note that because \(\langle{\bf u}|\hat{A}\hat{B}|{\bf u}\rangle\neq\langle{\bf u}|\hat{A}|{\bf u} \rangle\langle{\bf u}|\hat{B}|{\bf u}\rangle\) (the uncertainty property), the Q-function cannot be used to directly compute the quantum mechanical trace of a product of operators, _i.e._, \({\rm Tr}_{\rm e}[\hat{A}\hat{B}]\neq\int{\rm d}{\bf u}A_{\rm Q}({\bf u})B_{\rm Q }({\bf u})\).
To solve this issue, one can use the P-function and the following property
\[\mathrm{Tr}_{\mathrm{e}}[\hat{A}\hat{B}]=\int\mathrm{d}\mathbf{u}A_{ \mathrm{Q}}(\mathbf{u})B_{\mathrm{P}}(\mathbf{u})=\int\mathrm{d}\mathbf{u}A_{ \mathrm{P}}(\mathbf{u})B_{\mathrm{Q}}(\mathbf{u}). \tag{19}\]
The W-function can also be used for this purpose
\[\mathrm{Tr}_{\mathrm{e}}[\hat{A}\hat{B}]=\int\mathrm{d}\mathbf{u}A_{\mathrm{W}}( \mathbf{u})B_{\mathrm{W}}(\mathbf{u}). \tag{20}\]
Summarizing the above properties, we have
\[\mathrm{Tr}_{\mathrm{e}}[\hat{A}\hat{B}]=\int\mathrm{d}\mathbf{u}A_{\mathrm{s}} (\mathbf{u})B_{\mathrm{\bar{s}}}(\mathbf{u}), \tag{21}\]
The proof of Eq. 21 is elementary and is provided in Appendix C.
Choosing \(\hat{B}=\hat{\mathcal{I}}\), Eq. 21 becomes
\[\mathrm{Tr}_{\mathrm{e}}[\hat{A}]=\int\mathrm{d}\mathbf{u}A_{\mathrm{s}}( \mathbf{u})[\hat{\mathcal{I}}]_{\mathrm{\bar{s}}}(\mathbf{u})=\int\mathrm{d} \mathbf{u}A_{\mathrm{s}}(\mathbf{u}), \tag{22}\]
This suggest that for the quantum mechanical trace of an operator \(\hat{A}\), one can freely choose \(\mathrm{s}\in\{\mathrm{Q},\mathrm{P},\mathrm{W}\}\), which all provide the identical answer, even though different kernel \(\hat{w}_{\mathrm{s}}\) and radius \(r_{\mathrm{s}}\) is used. Further using the definition of \(A_{\mathrm{s}}(\mathbf{u})\) (Eq. 11) into Eq. 22, we have
\[\mathrm{Tr}_{\mathrm{e}}[\hat{A}]=\int\mathrm{d}\mathrm{u}\mathrm{Tr}_{ \mathrm{e}}[\hat{A}\hat{w}_{\mathrm{s}}]=\mathrm{Tr}_{\mathrm{e}}\big{[}\hat{A }\cdot\int\mathrm{d}\mathbf{u}\hat{w}_{\mathrm{s}}\big{]}, \tag{23}\]
The above equality indicates the following resolution of identity
\[\mathds{1}_{\mathbf{u}} =\int\mathrm{d}\mathbf{u}\hat{w}_{\mathrm{s}} \tag{24}\] \[=\frac{1}{2\pi}\int_{0}^{\pi}\mathrm{d}\theta\sin\theta\int_{0}^ {2\pi}\mathrm{d}\varphi\Big{(}\frac{1}{2}\hat{\mathcal{I}}+r_{\mathrm{s}} \mathbf{u}\cdot\hat{\mathbf{\sigma}}\Big{)},\]
18) and \(\hat{w}_{\mathrm{s}}\) (Eq. 8a). This identity can also be easily verified through elementary integrals, which is provided in Appendix C.
When choosing \(\mathrm{s}=\mathrm{Q}\), the resolution of identity is
\[\mathds{1}_{\mathbf{u}}=\int\mathrm{d}\mathbf{u}\hat{w}_{\mathrm{Q}}=\int \mathrm{d}\mathbf{u}|\mathbf{u}\rangle\langle\mathbf{u}|, \tag{25}\]
9).
## III Quantum partition function with spin-mapping variables
### Spin Coherent State (SCS) Partition Function
The canonical partition function is expressed as \(\mathcal{Z}=\mathrm{Tr}_{\mathrm{n}}\mathrm{Tr}_{\mathrm{e}}[e^{-\beta\hat{H }}]\), where \(\mathrm{Tr}_{\mathrm{n}}\) and \(\mathrm{Tr}_{\mathrm{e}}\) represent the trace over the nuclear and electronic DOFs, respectively, and \(\beta=1/k_{\mathrm{B}}T\). The partition function can be exactly evaluated in the limit \(N\rightarrow\infty\) by the Trotter discretization, where \(N\) is the number of ring polymer beads.
We start from expressing the quantum partition function as follows
\[\mathcal{Z}=\mathrm{Tr}_{\mathrm{e}}\mathrm{Tr}_{\mathrm{n}}\Big{[}\big{(}e^{ -\beta_{N}(H_{0}\hat{\mathcal{I}}+\frac{1}{2}\mathbf{H}\cdot\mathbf{\bar{s}}) }\big{)}^{N}\Big{]}, \tag{26}\]
Inserting \(N\) copies of the identities in the nuclear subspace, \(\mathds{1}_{R}=\int\mathrm{d}R_{\alpha}|R_{\alpha}\rangle\langle R_{\alpha}|\) and \(\mathds{1}_{P}=\int\mathrm{d}P_{\alpha}|P_{\alpha}\rangle\langle P_{\alpha}|\), where \(\alpha\) is the label of the imaginary-time (bead) index, and using the standard path-integral techniques, we obtain
\[\mathcal{Z}= \frac{1}{(2\pi\hbar)^{N}}\lim_{N\rightarrow\infty}\int\mathrm{d} \{R_{\alpha}\}\int\mathrm{d}\{P_{\alpha}\}e^{-\beta_{N}\tilde{H}_{0}( \mathbf{R})}\] \[\times\mathrm{Tr}_{\mathrm{e}}\Big{[}\prod_{\alpha=1}^{N}e^{- \beta_{N}\frac{1}{2}\mathbf{H}_{\alpha}\cdot\mathbf{\bar{S}}}\Big{]}. \tag{27}\]
Here, we use the notation \(\int\mathrm{d}\{X_{\alpha}\}=\prod_{\alpha=1}^{N}\int\mathrm{d}X_{1}\cdots \mathrm{d}X_{N}\), \(\mathbf{R}\equiv\{R_{\alpha}\}\), and \(\mathbf{H}_{\alpha}=[H_{x}(R_{\alpha}),H_{y}(R_{\alpha}),H_{z}(R_{\alpha})]\) (see their definition in Eq. 4b-4d).
\[\tilde{H}_{0}(\mathbf{R})= \sum_{\alpha=1}^{N}\big{[}\frac{P_{\alpha}^{2}}{2m}+\frac{m}{2 \beta_{N}^{2}\hbar^{2}}(R_{\alpha}-R_{\alpha-1})^{2} \tag{28}\] \[+U_{0}(R_{\alpha})+\frac{1}{2}(V_{1}(R_{\alpha})+V_{2}(R_{\alpha} ))\big{]}.\]
To perform the electronic trace, we insert \(N\) copies of the following spin coherent state identities (by choosing \(\mathrm{s}=\mathrm{Q}\))
\[\mathds{1}_{\mathbf{u}}= \int\mathrm{d}\mathbf{u}_{\alpha}|\mathbf{u}_{\alpha}\rangle \langle\mathbf{u}_{\alpha}|=\int\mathrm{d}\mathbf{u}_{\alpha}\hat{w}_{\mathrm{Q}} \tag{29a}\] \[= \frac{1}{2\pi}\int_{0}^{\pi}\mathrm{d}\theta_{\alpha}\sin\theta_{ \alpha}\int_{0}^{2\pi}\mathrm{d}\varphi_{\alpha}|\mathbf{u}_{\alpha}\rangle \langle\mathbf{u}_{\alpha}|, \tag{29b}\]
and rearranging the terms (as well as neglecting a normalization constant), resulting in
\[\mathcal{Z}\propto \lim_{N\rightarrow\infty}\int\mathrm{d}\{R_{\alpha}\}\int\mathrm{ d}\{P_{\alpha}\}\int\mathrm{d}\{\{\mathbf{u}_{\alpha}\}e^{-\beta_{N}\tilde{H}_{0}(\mathbf{R})}\] \[\times\prod_{\alpha=1}^{N}\langle\mathbf{u}_{\alpha}|e^{-\beta _{N}\frac{1}{2}\mathbf{H}_{\alpha}\cdot\mathbf{\bar{S}}}|\mathbf{u}_{\alpha+1}\rangle. \tag{30}\]
The above partition function can also be equivalently expressed by inserting electronic projection operators \(\hat{\mathcal{P}}=\sum_{n}|n\rangle\langle n|\), leading to
\[\mathcal{Z}\propto \lim_{N\rightarrow\infty}\int\mathrm{d}\{R_{\alpha}\}\int\mathrm{ d}\{P_{\alpha}\}\int\mathrm{d}\{\{\mathbf{u}_{\alpha}\}e^{-\beta_{N}\tilde{H}_{0}( \mathbf{R})}\] \[\times\prod_{\alpha=1}^{N}\langle\mathbf{u}_{\alpha}|\sum_{n}|n \rangle\langle n|e^{-\beta_{N}\frac{1}{2}\mathbf{H}_{\alpha}\cdot\mathbf{\bar{S}}} \sum_{m}|m\rangle\langle m|\mathbf{u}_{\alpha+1}\rangle. \tag{31}\]Note that the size of the spin mapping Hilbert space \(\mathbf{I}_{\mathbf{u}}\) is the same as the original electronic subspace \(\hat{\mathcal{P}}=\sum_{n}|n\rangle\langle n|\). Hence with or without \(\hat{\mathcal{P}}\), the partition function is invariant. This is different than the mapping in harmonic oscillators based on the MMST formalism, where the mapping Hilbert space is larger than the original electronic subspace, and projection often leads to a better result.
We further express the matrix elements of spin coherent state projected by \(\hat{\mathcal{P}}\) as follows
\[\mathbf{C}(\mathbf{u}_{\alpha}) \equiv\langle\mathbf{u}_{\alpha}|\sum_{n}|n\rangle\langle n| \tag{32a}\] \[=\cos\frac{\theta_{\alpha}}{2}e^{i\frac{\pi\alpha}{2}}\langle 1|+ \sin\frac{\theta_{\alpha}}{2}e^{-i\frac{\pi\alpha}{2}}\langle 2|,\] \[\mathbf{D}(\mathbf{u}_{\alpha+1}) \equiv\sum_{m}|m\rangle\langle m|\mathbf{u}_{\alpha+1}\rangle\] (32b) \[=\cos\frac{\theta_{\alpha+1}}{2}e^{-i\frac{\pi\alpha+1}{2}}|1 \rangle+\sin\frac{\theta_{\alpha+1}}{2}e^{i\frac{\pi\alpha+1}{2}}|2\rangle.\]
Using these, we can write the special form of the Spin Coherent State (SCS) partition function (with \(\mathrm{s=Q}\) case) as follows
\[\mathcal{Z}\propto\lim_{N\rightarrow\infty}\int\mathrm{d}\{R_{\alpha}\}\int \mathrm{d}\{P_{\alpha}\}\int\mathrm{d}\{\mathbf{u}_{\alpha}\}\mathrm{Tr}_{ \mathrm{e}}[\mathbf{\Gamma}_{\mathrm{Q}}]\cdot e^{-\beta_{N}\hat{H}_{0}( \mathbf{R})}, \tag{33}\]
where the electronic trace has the following expression
\[\mathbf{\Gamma}_{\mathrm{Q}}=\prod_{\alpha=1}^{N}\sum_{n,m}C_{n}( \mathbf{u}_{\alpha})\mathcal{M}_{nm}(R_{\alpha})D_{m}(\mathbf{u}_{\alpha+1}), \tag{34a}\] \[\mathcal{M}_{nm}(R_{\alpha})=\langle n|e^{-\beta_{N}\frac{1}{k} \mathbf{H}_{\alpha}\cdot\hat{\mathbf{S}}}|m\rangle. \tag{34b}\]
This partition function is analogous to those used with MMST mapping variables, such as the mapping-variable RPMD partition function or the coherent state mapping (CSM) ring polymer partition function. In CSM partition function, a similar derivation procedure is conducted with the coherent-state representation of the MMST mapping oscillators.
The above procedure relies on inserting N copies of the identities \(\mathbf{I}_{\mathbf{u}}=\int\mathrm{d}\mathbf{u}_{\alpha}|\mathbf{u}_{\alpha} \rangle\langle\mathbf{u}_{\alpha}|\equiv\int\mathrm{d}\mathbf{u}_{\alpha}\hat {w}_{\mathrm{Q}}(\mathbf{u}_{\alpha})\) (where \(\alpha\) is the bead index). Of course, one can insert the general resolution of identity \(\mathbf{\hat{1}}_{\mathbf{u}}=\int\mathrm{d}\mathbf{u}\hat{w}_{\mathrm{s}}\) (Eq. 24) inside the \(\mathrm{Tr}_{\mathrm{e}}[...]\) of Eq. 27, then moving the \(\int\mathrm{d}\mathbf{u}_{\alpha}\) integral outside \(\mathrm{Tr}_{\mathrm{e}}\), resulting in
\[\mathcal{Z}\propto\lim_{N\rightarrow\infty}\int\mathrm{d}\{R_{\alpha}\}\int \mathrm{d}\{P_{\alpha}\}\int\mathrm{d}\{\mathbf{u}_{\alpha}\}\mathrm{Tr}_{ \mathrm{e}}[\mathbf{\Gamma}_{\mathrm{s}}]\cdot e^{-\beta_{N}\hat{H}_{0}( \mathbf{R})}, \tag{35}\]
where the expression of the electronic trace is
\[\mathbf{\Gamma}_{\mathrm{s}}=\prod_{\alpha=1}^{N}e^{-\beta_{N}\frac{1}{k} \mathbf{H}_{\alpha}\cdot\hat{\mathbf{S}}}\cdot\hat{w}_{\mathrm{s}}(\mathbf{u} _{\alpha}). \tag{36}\]
By Taylor expanding the Boltzmann operator and using the properties of the Pauli matrices, we can prove the following identity
\[e^{-\beta_{N}\frac{1}{k}\mathbf{H}_{\alpha}\cdot\hat{\mathbf{S}}}=\cosh\frac {\beta_{N}|\mathbf{H}_{\alpha}|}{2}\hat{\mathcal{I}}-\sinh\frac{\beta_{N}| \mathbf{H}_{\alpha}|}{2}\cdot\frac{2\mathbf{H}_{\alpha}\cdot\hat{\mathbf{S}}} {\hbar|\mathbf{H}_{\alpha}|}, \tag{37}\]
Plugging this identity back into Eq. 36 we obtain the general expression \(\mathbf{\Gamma}_{\mathrm{s}}\) as follows
\[\mathbf{\Gamma}_{\mathrm{s}}= \Big{[}\prod_{\alpha=1}^{N}\Big{(}\frac{1}{2}\cosh\frac{\beta_{N }|\mathbf{H}_{\alpha}|}{2}-r_{\mathrm{s}}\frac{\mathbf{H}_{\alpha}}{|\mathbf{H }_{\alpha}|}\cdot\mathbf{u}_{\alpha}\sinh\frac{\beta_{N}|\mathbf{H}_{\alpha}|}{ 2}\Big{)}\hat{\mathcal{I}}\] \[+\Big{(}r_{\mathrm{s}}\mathbf{u}_{\alpha}\cosh\frac{\beta_{N}| \mathbf{H}_{\alpha}|}{2} \tag{38}\] \[\qquad-\frac{1}{|\mathbf{H}_{\alpha}|}(\frac{\mathbf{H}_{\alpha}}{ 2}+ir_{\mathrm{s}}\mathbf{H}_{\alpha}\wedge\mathbf{u}_{\alpha})\sinh\frac{\beta _{N}|\mathbf{H}_{\alpha}|}{2}\Big{)}\cdot\hat{\mathbf{\sigma}}\Big{]}.\]
A detailed derivation of Eqs. 37 and 38 is provided in Appendix D. When \(\mathrm{s=Q}\), Eq. 38 is equivalent to the expression of \(\mathbf{\Gamma}_{\mathrm{Q}}\) in Eq. 33. The numerical advantage of Eq. 38 is that it replaces the \(\mathcal{M}_{nm}(R_{\alpha})\) matrix in Eq. 34b with an analytic expression in Eq. 37.
### Spin-Mapping (SM)-NRPMD Hamiltonian
The SCS partition function in Eq. 35 gives the exact quantum statistics for a non-adiabatic system. The effective Hamiltonian from the SCS partition function can be used to propagate the dynamics. However, it will not provide accurate electronic dynamics (such as electronic Rabi oscillation) due to the inter-bead coupling among the different electronic and nuclear DOFs inside \(\mathbf{\Gamma}_{\mathrm{s}}\).
Instead of proposing a reasonable Hamiltonian for dynamics propagation, here, we try to theoretically justify a Hamiltonian from an alternative expression of the partition function. To this end, we evaluate the electronic trace in Eq. 27 using the property \(\mathrm{Tr}_{\mathrm{e}}[\hat{A}]=\int\mathrm{d}\mathbf{u}A_{\mathrm{s}}( \mathbf{u})\) in Eq. 22, leading to
\[\mathrm{Tr}_{\mathrm{e}}\Big{[}\prod_{\alpha=1}^{N}e^{-\beta_{N} \frac{1}{k}\mathbf{H}_{\alpha}\cdot\hat{\mathbf{S}}}\Big{]}=\int\mathrm{d} \mathbf{u}_{1}\Big{[}\prod_{\alpha=1}^{N}e^{-\beta_{N}\frac{1}{k}\mathbf{H}_{ \alpha}\cdot\hat{\mathbf{S}}}\Big{]}_{\mathrm{s}}(\mathbf{u}_{1}), \tag{39}\]
We further separate \(\prod_{\alpha=1}^{N}e^{-\beta_{N}\frac{1}{k}\mathbf{H}_{\alpha}\cdot\hat{ \mathbf{S}}}\) into \(e^{-\beta_{N}\frac{1}{k}\mathbf{H}_{1}\cdot\hat{\mathbf{S}}}\prod_{\alpha=2}^{N}e ^{-\beta_{N}\frac{1}{k}\mathbf{H}_{\alpha}\cdot\hat{\mathbf{S}}}\), and use the property expressed in Eq. 21, leading to
\[\int\mathrm{d}\mathbf{u}_{1}\Big{[}e^{-\beta_{N}\frac{1}{k}\mathbf{ H}_{1}\cdot\hat{\mathbf{S}}}\prod_{\alpha=2}^{N}e^{-\beta_{N}\frac{1}{k}\mathbf{H}_{ \alpha}\cdot\hat{\mathbf{S}}}\Big{]}_{\mathrm{s}}(\mathbf{u}_{1}) \tag{40}\] \[=\int\mathrm{d}\mathbf{u}_{1}\Big{[}e^{-\beta_{N}\frac{1}{k} \mathbf{H}_{1}\cdot\hat{\mathbf{S}}}\Big{]}_{\mathrm{s}}(\mathbf{u}_{1})\cdot \Big{[}\prod_{\alpha=2}^{N}e^{-\beta_{N}\frac{1}{k}\mathbf{H}_{\alpha}\cdot \hat{\mathbf{S}}}\Big{]}_{\mathrm{\bar{s}}}(\mathbf{u}_{1}),\]
21.
To evaluate \([e^{-\beta_{N}\frac{1}{\hbar}\mathbf{H}_{1}\cdot\hat{\mathbf{S}}]_{\mathrm{s}}}( \mathbf{u}_{1})\), we Taylor expand the exponential and neglect the terms of order equals to or higher than \(\beta_{N}^{2}\) (which is exact under the limit \(N\rightarrow\infty\)), leading to
\[[1-\beta_{N}\frac{1}{\hbar}\mathbf{H}_{1}\cdot\hat{\mathbf{S}}+ \mathcal{O}(\beta_{N}^{2})]_{\mathrm{s}}(\mathbf{u}_{1}) \tag{41}\] \[=\exp[-\beta_{N}\cdot\frac{1}{\hbar}\mathbf{H}_{1}\cdot[\hat{ \mathbf{S}}]_{\mathrm{s}}(\mathbf{u}_{1})]=\exp[-\beta_{N}\cdot\frac{1}{\hbar}r _{\mathrm{s}}\mathbf{H}_{1}\cdot\mathbf{u}_{1}].\]
Plugging it back into Eq. 40, we have
\[\mathrm{Tr}_{\mathrm{e}}\Big{[}\prod_{\alpha=1}^{N}e^{-\beta_{N} \frac{1}{\hbar}\mathbf{H}_{\alpha}\cdot\hat{\mathbf{S}}}\Big{]} \tag{42}\] \[=\int\mathrm{d}\mathbf{u}_{1}e^{-\beta_{N}\cdot r_{\mathrm{s}} \mathbf{H}_{1}\cdot\mathbf{u}_{1}}\cdot\Big{[}\prod_{\alpha=2}^{N}e^{-\beta_{ N}\frac{1}{\hbar}\mathbf{H}_{\alpha}\cdot\hat{\mathbf{S}}}\Big{]}_{\mathrm{\bar{s}}}( \mathbf{u}_{1})\] \[=\int\mathrm{d}\mathbf{u}_{1}e^{-\beta_{N}\cdot r_{\mathrm{s}} \mathbf{H}_{1}\cdot\mathbf{u}_{1}}\cdot\mathrm{Tr}_{\mathrm{e}}\Big{[}\prod_{ \alpha=2}^{N}e^{-\beta_{N}\frac{1}{\hbar}\mathbf{H}_{\alpha}\cdot\hat{\mathbf{ S}}}\hat{w}_{\mathrm{\bar{s}}}(\mathbf{u}_{1})\Big{]}.\]
Further inserting the identity \(\int\mathrm{d}\mathbf{u}_{2}\hat{w}_{\mathrm{s}}(\mathbf{u}_{2})\) (see Eq. 24) inside the \(\mathrm{Tr}_{\mathrm{e}}\), we have
\[\mathrm{Tr}_{\mathrm{e}}\Big{[}\prod_{\alpha=1}^{N}e^{-\beta_{N} \frac{1}{\hbar}\mathbf{H}_{\alpha}\cdot\hat{\mathbf{S}}}\Big{]} \tag{43}\] \[=\int\mathrm{d}\mathbf{u}_{1}e^{-\beta_{N}\cdot r_{\mathrm{s}} \mathbf{H}_{1}\cdot\mathbf{u}_{1}}\cdot\mathrm{Tr}_{\mathrm{e}}\Big{[}\int \mathrm{d}\mathbf{u}_{2}\hat{w}_{\mathrm{s}}(\mathbf{u}_{2})e^{-\beta_{N}\frac {1}{\hbar}\mathbf{H}_{2}\cdot\hat{\mathbf{S}}}\] \[\times\prod_{\alpha=3}^{N}e^{-\beta_{N}\frac{1}{\hbar}\mathbf{H} _{\alpha}\cdot\hat{\mathbf{S}}}\hat{w}_{\mathrm{\bar{s}}}(\mathbf{u}_{1}) \Big{]}\] \[=\int\mathrm{d}\mathbf{u}_{1}e^{-\beta_{N}\cdot r_{\mathrm{s}} \mathbf{H}_{1}\cdot\mathbf{u}_{1}}\int\mathrm{d}\mathbf{u}_{2}e^{-\beta_{N} \cdot r_{\mathrm{s}}\mathbf{H}_{2}\cdot\mathbf{u}_{2}}\] \[\times\mathrm{Tr}_{\mathrm{e}}\Big{[}\prod_{\alpha=3}^{N}e^{- \beta_{N}\frac{1}{\hbar}\mathbf{H}_{\alpha}\cdot\hat{\mathbf{S}}}\hat{w}_{ \mathrm{\bar{s}}}(\mathbf{u}_{1})\hat{w}_{\mathrm{\bar{s}}}(\mathbf{u}_{2}) \Big{]},\]
11, and in the third equality, we have used the property in Eq. 40.
Repeating the above argument for all \(N\) beads, we obtain the following partition function
\[\mathcal{Z}\propto\lim_{N\rightarrow\infty}\int\mathrm{d}\{R_{ \alpha}\}\int\mathrm{d}\{P_{\alpha}\}\int\mathrm{d}\{\mathbf{u}_{\alpha}\} \Phi_{\mathrm{\bar{s}}}\cdot e^{-\beta_{N}\hat{H}_{\mathrm{s}}}, \tag{44}\]
where \(\Phi_{\mathrm{\bar{s}}}=\mathrm{Tr}_{\mathrm{e}}\left[\prod_{\alpha=1}^{N}\hat {w}_{\mathrm{\bar{s}}}(\mathbf{u}_{\alpha})\right]\), and the spin-mapping (SM)-NRPMD Hamiltonian is
\[\tilde{H}_{\mathrm{s}}=\tilde{H}_{0}(\mathbf{R})+\sum_{\alpha=1}^{N}r_{ \mathrm{s}}\mathbf{H}_{\alpha}\cdot\mathbf{u}_{\alpha}, \tag{45}\]
15 (with the additional ring polymer potential in Eq. 28). Based on our previous experience with the MMST version of NRPMD approach, we conjecture that \(\tilde{H}_{\mathrm{s}}\) should be the Hamiltonian for the NRPMD propagation when using the spin mapping variables. This is because the correct equations of motion for the MMST mapping variables can be derived based on the partition function through a similar procedure as above, which coincides with the Liouvillian derived from generalized Kubo-transformed TCF with Matsubara approximation and ring polymer approximation We also note that in principle, the partition function in Eq. 45 should generate the same result as the one in Eq. 35, under the limit \(N\rightarrow\infty\). However, with a finite \(N\), we find that the numerical convergence by using Eq. 45 is much slower compared to Eq. 35, likely due to the limit we took in Eq. 41 (which requires a large \(N\)). Hence, we emphasize that Eq. 45 is only used as a justification for the Stratonovich-Weyl NRPMD Hamiltonian in Eq. 45, and not used for sampling the quantum initial condition.
## IV Spin-Mapping (SM)-NRPMD time-correlation function
The Kubo-transform real-time correlation function for two operators \(\hat{A}\) and \(\hat{B}\) is expressed as
\[C_{AB}^{\mathrm{K}}(t)=\frac{1}{Z\beta}\int_{0}^{\beta}\mathrm{ d}\lambda\mathrm{Tr}\left[e^{-(\beta-\lambda)\hat{H}}\hat{A}e^{-\lambda\hat{H}}e^{i \hat{H}t/\hbar}\hat{B}e^{-i\hat{H}t/\hbar}\right]. \tag{46}\]
We propose that the above Kubo-transformed TCF (Eq.46) can be approximated as the following Spin-Mapping TCF
\[C_{AB}(t)= \frac{1}{\mathcal{Z}}\lim_{N\rightarrow\infty}\int\mathrm{d}\{R_ {\alpha}\}\int\mathrm{d}\{P_{\alpha}\}\int\mathrm{d}\{\mathbf{u}_{\alpha}\}\] \[\times\mathrm{Tr}_{\mathrm{e}}[\mathbf{\Gamma}_{\mathrm{s}}]e^{- \beta_{N}\tilde{H}_{0}}[A]_{N}[B]_{N}(t), \tag{47}\]
When operators \(\hat{A}\) and \(\hat{B}\) are related to the electronic DOF, the TCF is proposed as
\[C_{AB}(t)= \frac{1}{\mathcal{Z}}\lim_{N\rightarrow\infty}\int\mathrm{d}\{R_{ \alpha}\}\int\mathrm{d}\{P_{\alpha}\}\int\mathrm{d}\{\mathbf{u}_{\alpha}\}\] \[\times\mathrm{Tr}_{\mathrm{e}}[\mathbf{\Gamma}_{\mathrm{s}}\hat{A}]e ^{-\beta_{N}\tilde{H}_{0}}[B_{\mathrm{\bar{s}}}]_{N}(t), \tag{48}\]
21 in order to satisfy the requirement at \(t=0\) to compute the trace of two operators (i.e, \(e^{-\beta H}\hat{A}\) and \(\hat{B}\)).
\[\mathcal{P}_{nn}^{\rm s} =\mathrm{Tr}_{\rm e}[\mathbf{\Gamma}_{\rm s}|n\rangle\langle n|]= \frac{1}{N}\sum_{\mu=1}^{N}\Big{[}\prod_{\alpha^{\prime}=1}^{N-\mu}e^{-\beta_{N} \frac{1}{2}\mathbf{H}_{\alpha^{\prime}}\cdot\mathbf{\hat{\mathrm{s}}}}\cdot\hat{ \mathrm{w}}_{\rm s}(\mathbf{u}_{\alpha^{\prime}})\] \[\times|n\rangle\langle n|\prod_{\alpha^{\prime\prime}=N-\mu+1}^{N }e^{-\beta_{N}\frac{1}{2}\mathbf{H}_{\alpha^{\prime\prime}}\cdot\mathbf{\hat{ \mathrm{s}}}}\cdot\hat{\mathrm{w}}_{\rm s}(\mathbf{u}_{\alpha^{\prime\prime}}) \Big{]}, \tag{49}\]
The analytic expression can be evaluated in the same way as \(\mathbf{\Gamma}_{\rm s}\) in Eq. 38, leading to \(|n\rangle\langle n|\) inserted in-between the \(\alpha\) and the \(\alpha+1\) bead. Specifically, for \(\mathrm{s}=\mathrm{Q}\), using Eq. 34a we have
\[\mathcal{P}_{nn}^{\mathrm{Q}}= \frac{1}{N}\sum_{\alpha=1}^{N}\frac{\sum_{m}C_{n}(\mathbf{u}_{ \alpha})\mathcal{M}_{nm}(R_{\alpha})D_{m}(\mathbf{u}_{\alpha+1})}{\sum_{n,m}C _{n}(\mathbf{u}_{\alpha})\mathcal{M}_{nm}(R_{\alpha})D_{m}(\mathbf{u}_{\alpha +1})}. \tag{50}\]
The population estimator for the operator \(\hat{B}\) is obtained by
\[[B_{\rm s}]_{N}=\frac{1}{N}\sum_{\alpha=1}^{N}B_{\rm s}(\mathbf{u}_{\alpha}), \tag{51}\]
where \(B_{\rm s}(\mathbf{u}_{\alpha})\) is the SW transform of \(\hat{B}\), and when \(\hat{B}=|n\rangle\langle n|\), it is expressed as
\[B_{\rm s}(\mathbf{u}_{\alpha})=\mathrm{Tr}_{\rm e}\left[|n\rangle\langle n| \hat{w}_{\rm s}\right]=\begin{cases}1/2+r_{\rm s}\cos\theta_{\alpha},&n=1\\ 1/2-r_{\rm s}\cos\theta_{\alpha},&n=2\end{cases}.\]
The function \([B]_{N}(t)\) or \([B_{\rm s}]_{N}(t)\) is evaluated along the classical trajectory \(\{R_{\alpha}(t),\mathbf{u}_{\alpha}(t)\}\), and the dynamics is _proposed_ to be governed by
\[\tilde{H}_{\rm s}=\tilde{H}_{0}(\mathbf{R})+\sum_{\alpha=1}^{N}r_{\rm s} \mathbf{H}(R_{\alpha})\cdot\mathbf{u}_{\alpha} \tag{52}\]
45.
\[\dot{R}_{\alpha} =\frac{\partial\tilde{H}_{\rm s}}{\partial P_{\alpha}}=\frac{P_{ \alpha}}{m} \tag{53a}\] \[\dot{P}_{\alpha} =-\frac{\partial\tilde{H}_{\rm s}}{\partial R_{\alpha}}=-\frac{ \partial\tilde{H}_{0}}{\partial R_{\alpha}}-r_{\rm s}\frac{\partial\mathbf{H} (R_{\alpha})}{\partial R_{\alpha}}\cdot\mathbf{u}_{\alpha},\] (53b) \[\dot{\mathbf{u}}_{\alpha} =\frac{1}{\hbar}\mathbf{H}(R_{\alpha})\wedge\mathbf{u}_{\alpha}. \tag{53c}\]
In the original NRPMD method, the corresponding equation of motion was first proposed, then recently proved through the non-adiabatic Matsubara dynamics formalism. We envision that the above EOM (Eq. 53a-Eq. 53c) can also be proved in a similar way when using the spin mapping variables, and we will explore this in future studies.
## V Computational details
To test the performance of the derived SCS-partition function in Eq. 35, we adapt a widely used model system and compute the state-dependent nuclear probability distribution. The model Hamiltonian \(\hat{H}=\hat{P}^{2}/2m+\hat{V}\), with nuclear mass \(M=3600\) a.u., and the diabatic potential \(\hat{V}\) is defined as
\[V_{ij}=\begin{cases}\frac{1}{2}k_{i}(R-R_{i})^{2}+\epsilon_{i},&i=j\\ 5\times 10^{-5}e^{-0.4R^{2}},&i\neq j\end{cases}, \tag{54}\]
1. We refer to this model as Model 0. The physical temperature of the system is set to be \(T=8\) K.
The initial quantum distribution is sampled using the Metropolis-Hastings algorithm according to the following distribution function
\[\rho(\{R_{\alpha},\mathbf{u}_{\alpha}\})=|\mathrm{Tr}_{\rm e}[\mathbf{\Gamma}_ {\rm s}]|\cdot e^{-\beta_{N}\tilde{H}_{0}(\mathbf{R})}, \tag{55}\]
with a complex weighting factor of
\[\Xi_{\rm s}(\{R_{\alpha},\mathbf{u}_{\alpha}\})=\mathrm{Tr}_{\rm e}[\mathbf{ \Gamma}_{\rm s}]/|\mathrm{Tr}_{\rm e}[\mathbf{\Gamma}_{\rm s}]|. \tag{56}\]
The nuclear probability distribution is obtained by computing
\[\mathrm{P}(R_{0})= \frac{\mathrm{Tr}[e^{-\beta\hat{H}}\delta(\hat{R}-R_{0})]}{ \mathcal{Z}} \tag{57}\] \[= \frac{1}{\langle\Re(\Xi_{\rm s})\rangle}\cdot\langle\Re(\Xi_{\rm s })\cdot\delta(R-R_{0})\rangle,\]
55.
\[\mathrm{P}_{n}(R_{0}) =\frac{\mathrm{Tr}[e^{-\beta\hat{H}}|n\rangle\langle n|\delta( \hat{R}-R_{0})]}{\mathcal{Z}} \tag{58}\] \[=\frac{1}{\langle\Re(\Xi_{\rm s})\rangle}\cdot\langle\Re(\Xi_{ \rm s}\cdot\mathcal{P}_{nn}^{\rm s})\cdot\delta(R-R_{0})\rangle,\]
49. To compute \(\mathrm{P}(R_{0})\) and \(\mathrm{P}_{n}(R_{0})\), \(N=10\) beads were required to converge the results, using a total of \(2.4\times 10^{7}\) configurations sampled from the Monte-Carlo procedure for \(\mathrm{s}=\mathrm{Q}\). Exactly identical results can be obtain with the same bead-convergence for other choices of \(\mathrm{s}\), but the required number of configurations to achieve the same level of convergence is much higher. In particular for \(N=10\) beads, using \(\mathrm{s}=\mathrm{W}\) requires 24 times more trajectories, while using \(\mathrm{s}=\mathrm{P}\) requires almost 2000 times more trajectories.
To assess the accuracy of the SM-NRPMD approach, we compute time correlation functions and compare our results with numerically exact Kubo-transformed quantum TCF, as well as non-adiabatic RPMD approach based on the MMST formalism.
\[\hat{H}=\frac{\hat{P}^{2}}{2m}+\frac{1}{2}m\omega^{2}\hat{R}^{2}+\begin{pmatrix} \hat{R}+\epsilon&\Delta\\ \Delta&-\hat{R}-\epsilon\end{pmatrix}, \tag{59}\]
We choose \(m=\hbar=\omega=\beta=1\). The rest of the parameters are provided in Table 2, changing the non-adiabaticity of the system from adiabatic (model I with \(\beta\Delta=10\)) to highly non-adiabatic (model VII with \(\beta\Delta=0.1\)). The number of beads to generate the converged results are also provided in Table 2.
The position and population auto-correlation functions are computed as follows
\[\mathrm{C}_{RR}(t)= \frac{1}{\left\langle\Re(\Xi_{\mathrm{s}})\right\rangle}\cdot \left\langle\Re(\Xi_{\mathrm{s}})\cdot\bar{R}\cdot\bar{R}(t)\right\rangle, \tag{60}\] \[\mathrm{C}_{nn}(t)= \frac{1}{\left\langle\Re(\Xi_{\mathrm{s}})\right\rangle}\left\langle \Re(\Xi_{\mathrm{s}}\cdot\mathcal{P}_{nn}^{\mathrm{s}})\cdot[B_{\mathrm{s}} ]_{N}(t)\right\rangle. \tag{61}\]
For s = W, between \(10^{4}\) and \(10^{6}\) trajectories were run for 4-to-6 beads for the results presented hereafter, with a time-step of 0.01 a.u.
## VI Results and Discussion
54. These distributions agree perfectly with the numerically exact results obtained from the DVR calculations. The numerical convergence is achieved with only \(N=10\) beads. The SCS partition function in Eq. 35 only requires two independent variables \(\{\theta_{\alpha},\varphi_{\alpha}\}\) for each bead, which is consistent with the number of electronic states. The MMST based partition function, such as those used in NRPMD or MV-RPMD requires 4 independent variables. As the number of beads increases, the MMST-based approaches becomes numerically expensive. In addition, previous numerical investigations suggest that 16-32 beads are required to reach to the same level of convergence with the MMST-based path-integral approaches. This is likely due to a larger set of free variables needed to be sampled. Moreover, the general formalism of \(\mathbf{\Gamma}_{\mathrm{s}}\) in Eq. 38 does not explicitly require the evaluation of \(\mathcal{M}_{nm}(R_{\alpha})=\langle n|e^{-\beta_{N}\frac{\mathrm{i}}{2}\mathbf{ H}_{\alpha}\cdot\mathbf{\hat{S}}}|m\rangle\) matrix, avoiding explicit diagonalization of the \(2\times 2\) matrix at a given \(R_{\alpha}\).
A brief description of NPRMD and the mean-field RPMD approach are provided in Appendix E. The SM-NRPMD calculations have been done with the choice of \(\mathrm{s}=\bar{\mathrm{s}}=\mathrm{W}\) (sampling with \(\mathbf{\Gamma}_{\mathrm{s}}\equiv\mathbf{\Gamma}_{\mathrm{W}}\) and dynamics with \(\tilde{H}_{\bar{\mathrm{s}}}\equiv\tilde{H}_{\mathrm{W}}\)). In the adiabatic regime (\(\beta\Delta\gg 1\)) in panel a, all methods agree perfectly with the exact result as expected. In the intermediate regime \(\beta\Delta\approx 1\) in panel b, all RPMD based approaches captures the correct oscillation frequency of the TCF, but they give different amplitudes that deviate from the exact results, except the SM-NRPMD approach which provides an excellent agreement with the exact results. In the non-adiabatic regime though (\(\beta\Delta\gg 1\)) in panels c and d, MF-RPMD method can not provide the correct amplitude nor oscillation frequency for the TCF. On the other hand, both SM-NRPMD and NRPMD results are in agreement with the exact results at short times. We can notice that even in the most challenging highly non-adiabatic case, only 4 beads are required to converge results, and for the other models above 6 beads, the results are converged.
Nuclear probability distribution (black curve) \(\mathrm{P}(R_{0})\) of model 0 obtained from the SCS partition function, with \(N=10\) beads. The state specific distributions for state 1 (blue) and state 2 (red) are also shown. The results are compared with quantum exact calculations (filled circles).
We have also performed the SM-NRPMD simulations with (i) s = Q sampling and dynamics obeying \(\bar{\mathrm{s}}=\mathrm{P}\), and (ii) s = P sampling and dynamics obeying \(\bar{\mathrm{s}}=\mathrm{Q}\). Additional results and discussions are provided in Appendix E.
Accurately describing electronic Rabi oscillations are essential for non-adiabatic dynamics simulations. Both the SM-NRPMD and the NRPMD agree well with exact results in the adiabatic regime for model II presented in Fig. 3a, and provide reasonably good results for the model systems in the intermediate regimes presented in Fig. 3b-c. MV-RPMD on the other hand, cannot correctly capture the electronic oscillations in these population auto-correlation functions (results not shown), due to the contamination of the true electronic Rabi oscillations with the inter-beads couplings in the mapping ring polymer Hamiltonian.
These expectation values are computed with both SM-NRPMD (solid lines) and
Expectation values of the nuclear position operator (left panel) and the electronic population of state 1 (right panel) for model IV (intermediate regime). Results are obtained from SM-NRPMD (solid lines) and NRPMD (dashed lines) with s = W results for \(N=2\) (magenta), \(N=4\) (blue), and \(N=6\) (green) beads.
The Kubo-transformed nuclear position auto-correlation functions (left panels) and the electronic population auto-correlation functions (right panels) for (a) Model II, (b) Model III, and (c) Model V. The results are obtained from SM-NRPMD with s = W (black lines) and NRPMD (green dashed lines), compared to the numerically exact result (red dots).
Kubo-transformed nuclear position auto-correlation functions for (a) model I, (b) model V, (c) model VI, and (d) model VII. Results are obtained from SM-NRPMD with s = W (black lines), MF-RPMD (blue dashed lines) and NRPMD (green dashed lines), as well as numerically exact result (red dots).
NRPMD (dashed lines) and are compared to the exact value. Because the system is under thermal equilibrium, these values should be conserved along the dynamics. As we can see in Fig. 4, by increasing the number of beads from \(N=2\) (magenta), to \(N=4\) (blue), and \(N=6\) (green), SM-NRPMD (with \(\mathrm{s=W}\)) almost provides time-independent expectation values. The MMST based approach, such as the NRPMD (dashed lines) can not provide a constant expectation value with the same number of beads. We conjecture that at a large number of beads, SM-NRPMD (with \(\mathrm{s=W}\)) might preserve the initial quantum Boltzmann distribution. This conjecture is also corroborate by the numerical evidence that the initial distribution function \(\mathrm{Tr}_{\mathrm{e}}[\mathbf{\Gamma}_{\mathrm{s}}]\cdot e^{-\beta N\tilde{H}_{ 0}(\mathbf{R})}\) (inside Eq. 35) is conserved by the equations of motion in Eqs. 53a-53c at the single trajectory level with a large number of beads (\(N>32\)). To summarize, with a finite number of beads, the SM-NRPMD (with \(\mathrm{s=W}\)) largely conserves the initial quantum Boltzmann distribution, providing an almost time-independent expectation value for systems under thermal equilibrium. This is a significant improvement compared to the MMST based NRPMD dynamics.
Compared to the previous NRPMD approach with the MMST formalism, the SM-NRPMD approach provides an additional advantage that the dynamics is invariant with respect to the splitting between the state-independent potential \(U_{0}(\tilde{R})\) and the state-dependent potential. This is because the spin-mapping formalism explicitly enforces the total population to be \(1\), such that \([\hat{\mathcal{I}}]_{\mathrm{s}}(\mathbf{u})=1\). More explicitly, this can be seen in Eq. 14a-14b, leading to \(U_{0}=(\frac{1}{2}+r_{\mathrm{s}}\cos\theta)\cdot U_{0}+(\frac{1}{2}-r_{ \mathrm{s}}\cos\theta)\cdot U_{0}\). The MMST formalism, on the other hand, does not guarantee this property, and a brief discussion between these two mapping approaches is provided in Appendix A.
\[\hat{H}=\frac{\hat{P}^{2}}{2m}+\begin{pmatrix}\frac{1}{2}m\omega^{2}\hat{R}^{2 }+\hat{R}+\epsilon&\Delta\\ \Delta&\frac{1}{2}m\omega^{2}\hat{R}^{2}-\hat{R}-\epsilon\end{pmatrix}.\]
The results are obtained with the NRPMD using the MMST formalism (panels a-d) and with the SM-NRPMD approach using the spin-mapping formalism (panels e-h). When including \(U_{0}\) into the state-dependent potential, the NRPMD dynamics becomes unstable and completely breaks down at \(t\approx 3.5\) a.u., as some trajectories within
The influence of including the quadratic potential \(U_{0}(R)\) into the state-dependent Hamiltonian for model V. The left panels (a-d) present the results obtained from the NRPMD method using the MMST mapping formalism. The Kubo-transformed position auto-TCF (panel a) and population auto-TCF (panel b) are computed with NRPMD (black lines) and exact approach (red dots). The panels (c) and (d) present three representative trajectories (blue, magenta and green) along which \(\tilde{R}=\frac{1}{N}\sum_{\alpha=1}^{N}R_{\alpha}\) (panel c) and \(\tilde{\mathrm{P}}_{1}=\frac{1}{N}\sum_{\alpha=1}^{N}\frac{1}{2}([q_{\alpha}] ^{2}+[p_{\alpha}]^{2}-1)\) are computed (see Eq. 81 in Appendix E). The dynamics are propagated using the NRPMD approach with \(U_{0}(R)\) (dashed lines) and without \(U_{0}(R)\) (solid lines) inside the state-dependent potential. Right panels (e-h) present the results obtained from the SM-NRPMD method using the spin mapping with \(\mathrm{s=W}\). Panels (g) and (h) provides three representative trajectories along which \(\tilde{R}=\frac{1}{N}\sum_{\alpha=1}^{N}R_{\alpha}(t)\) (panel c) and \(\tilde{\mathrm{P}}_{1}=\frac{1}{N}\sum_{\alpha=1}^{N}\frac{1}{2}+r_{\mathrm{ W}}\cdot\cos\theta_{\alpha}\) are computed using SM-NRPMD.
Three representative nuclear position trajectories and population trajectories when including \(U_{0}\) into the state-dependent potential are shown with the dashed lines in Fig. 5c-d, compared to the case when treating the quadratic term \(U_{0}\) as a state-independent potential (solid lines). When individual trajectories have a total population deviated from 1 (as shown in Fig. 5d) in the MMST formalism, the total population also multiplies in front of \(U_{0}\), resulting in an incorrect force acting on the nuclear DOF, as well as unstable motions. Due to this, including \(U_{0}\) into the state-dependent Hamiltonian could be numerically challenging and eventually causes numerical instabilities. In addition, the results of the auto-correlation functions (before diverging) are different than those obtained in and Fig. 3c, indicating that different splitting of state-dependent and state-independent potential in the MMST formalism can lead to different numerical results when using approximate quantum dynamics approaches.
Fig. 5e-h present the same comparisons using the spin-mapping approach SM-NRPMD (with s = W). As expected, the dynamics is invariant under different ways of partitioning \(U_{0}(R)\). and present the Kubo-transformed TCF when including \(U_{0}(R)\) inside the state-dependent potential, providing identical results to those presented in and In fact, the dynamics is invariant at the single-trajectory level, as clearly indicated in Fig. 5g-h. This is guaranteed because the total population is always bounded by one in spin mapping, hence the quadratic potential is always \(\left(\frac{1}{2}+r_{\text{s}}\cos\theta\right)\cdot U_{0}+(\frac{1}{2}-r_{ \text{s}}\cos\theta)\cdot U_{0}=U_{0}\). This is another unique advantage of using the spin mapping formalism compared to the MMST mapping formalism, in addition to the better preservation of the initial quantum distribution demonstrated in Note that in the SM approach, a negative population is still possible in the case of the \(\tilde{H}_{\text{P}}\) and \(\tilde{H}_{\text{W}}\) (see Fig. 5h), but the population is not directly involved in the potential related to \(U_{0}\). In addition, the mapping dynamics of the spin variables \(\theta\) and \(\varphi\) are bounded on the Bloch sphere of radius \(r_{\text{s}}\), as opposed to un-bounded phase space variables (in the mapping oscillator phase space) in the MMST formalism (see Appendix A). Together, these advantages of the spin-mapping variables make it a more accurate and convenient mapping representation for developing non-adiabatic dynamics methods, and we extend it to the NRPMD dynamics in this work.
## VII Conclusion
In this paper, we present a new non-adiabatic RPMD method based on the recent development of spin mapping (SM) formalism. The basis of the spin mapping variables, the spin coherent states, is of the same dimensionality as the electronic Hilbert subspace of the original system. Hence, the SM approach is numerically advantageous compared to the original harmonic oscillator-based mapping approach. These include the total population for a single trajectory is always bounded by one, the dynamics is invariant under different ways of partitioning the state-independent and state-dependent potentials, and the further projections back to the electronic subspace is not necessary to compute the physical observables.
Using the spin mapping representation, we derive a general quantum partition function for the coupled electronic-nuclear system, which we refer to as the Spin Coherent State (SCS) Partition Function. We test the performance of the SCS partition function by computing state-dependent nuclear distribution in a two-level system coupled to a harmonic DOF. Our result suggest that the SCS partition function provides the exact quantum results using \(N=10\) beads, requiring fewer beads compared to the MMST-based quantum partition functions. Further, the SCS partition function provides an analytical expression of the matrix elements of the thermal Boltzmann operator (Eq. 37), facilitating the Monte-Carlo numerical simulations. Using various choices of \(r_{\text{s}}\) in the Stratonovich-Weyl transformation, we find that the \(\text{s}=\text{Q}\) requires the fewest MC configuration to converge, whereas \(\text{s}=\text{W}\) approach requires 10 times more than \(\text{s}=\text{Q}\) approach, and \(\text{s}=\text{P}\) requires \(10^{3}\) more configurations to converge for \(N=10\) beads (this ratio increases when increasing the number of beads). Compared to the MMST based approaches, the \(\text{s}=\text{W}\) approach requires a similar amount of configurations and fewer number of beads for convergence compared to the original NRPMD or CS-RPMD approach.
Using the property of the Stratonovich-Weyl transformation, we further derive the spin-mapping (SM)-NRPMD Hamiltonian, which can be viewed as the unified Hamiltonian of the spin-mapping Hamiltonian and the ring polymer Hamiltonian. Based on this Hamiltonian, we propose the SM-NRPMD dynamics, where the initial sampling is governed by the SCS partition function and the dynamics is governed by the SM-NRPMD Hamiltonian. Using the degrees of freedom of \(r_{\text{s}}\) and \(r_{\bar{\text{s}}}\), we find that by choosing \(\{\text{s}=\text{W},\bar{\text{s}}=\text{W}\}\), SM-NRPMD provides accurate Kubo-transformed nuclear-position autocorrelation function compared to the exact results for model systems that exhibit a broad range of parameters, from electronically adiabatic to the non-adiabatic regime. It can also provide the accurate population autocorrelation function with the correct electronic Rabi oscillation frequency. The accuracy of SM-NRPMD appears to be equivalent (with some slight improvements in certain cases) to those obtained from MMST based non-adiabatic RPMD methods, such as NRPMD or CS-RPMD, with a similar number of beads to converge the dynamics and a similar amount of trajectories required to converge the calculations.
From our numerical results, the SM-NRPMD seems to preserve the initial quantum Boltzmann distribution by providing a nearly time-independent expectation value of the nuclear position and electronic population. The MMST-based RPMD approaches, on the other hand, failed to generate time-independent expectation value of an observable for systems under thermal equilibrium. Moreover, the SM-NRPMD provides stable and invariant results regardless of how to partition the state-independent and state-dependent potentials, whereas the MMST-based NRPMD dynamics are highly sensitive to the specific choice of splitting the potentials.
To summarize, SM-NRPMD provides accurate electronic non-adiabatic dynamics with explicit nuclear quantization, with additional advantages compared to the original MMST based approaches including a normalized total population along a single trajectory, and the invariant dynamics under different ways of partition of potentials. Future directions include generalizing the current formalism to multi-electronic states, as well as rigorously derive SM-NRPMD formalism through the recent development of the non-adiabatic Matsubara framework.
|
10.48550/arXiv.2103.14119
|
Non-Adiabatic Ring Polymer Molecular Dynamics with Spin Mapping Variables
|
Duncan Bossion, Sutirtha N. Chowdhury, Pengfei Huo
| 1,569
|
10.48550_arXiv.1611.08901
|
###### Abstract
We introduce and investigate the escape problem for random walkers that may eventually die, decay, bleach, or lose activity during their diffusion towards an escape or reactive region on the boundary of a confining domain. In the case of a first-order kinetics (i.e., exponentially distributed lifetimes), we study the effect of the associated death rate onto the survival probability, the exit probability, and the mean first passage time. We derive the upper and lower bounds and some approximations for these quantities. We reveal three asymptotic regimes of small, intermediate and large death rates. General estimates and asymptotics are compared to several explicit solutions for simple domains, and to numerical simulations. These results allow one to account for stochastic photobleaching of fluorescent tracers in bio-imaging, degradation of mRNA molecules in genetic translation mechanisms, or high mortality rates of spermatozoa in the fertilization process. This is also a mathematical ground for optimizing storage containers and materials to reduce the risk of leakage of dangerous chemicals or nuclear wastes.
leakage, safety, diffusion, escape problem, first passage time, mixed boundary condition pacs: 02.50.-r, 05.40.-a, 02.70.Rr, 05.10.Gg
## I Introduction
The safe long-term storage of dangerous species is of paramount importance for chemical industries, nuclear waste containers and landfills, and military arsenals. In spite of numerous efforts to improve containers for dangerous chemicals or nuclear wastes, a complete isolation is not realistic because of eventual defects and material degradation in time. If the species remain active forever, their leakage is certain and is just a matter of time. In such situation, the quality of isolation can be characterized by the survival probability for diffusing species to remain inside a confining domain up to time \(t\). This is an example of the first passage time (FPT) problems that have attracted much attention during the last decade, with numerous chemical, biological and ecological applications ranging from diffusion in cellular microdomains to animal foraging strategies. Most analytical results were obtained for the mean first passage time (MFPT) through a small escape region, also known as the narrow escape problem.
In this paper, we introduce and discuss an important extension of the escape problem to "mortal" walkers. In fact, a finite lifetime of diffusing species is a typical situation for many biological, chemical and ecological processes: (i) an animal or a bacterium should remain alive while searching for food; (ii) in bio-imaging techniques, progressive extinction of the fluorescence signal can be either due to degradation of the tagged protein through an enzymatic reaction (i.e. finding the target) or due to bleaching (i.e. finite fluorescence lifetime); the latter mechanism should be taken into account for reliable interpretation of such measurements; (iii) in order to trigger translation, messenger RNA should not be degraded before reaching a ribosome; (iv) in spite of a very high mortality rate, the spermatozoa that search for a small egg in the uterus or in the Fallopian tubes, need to remain alive to complete the fertilization; (v) molecules should remain active or intact before reaching a reactive site on the surface of a catalyst; (vi) protective materials may trap, bind or deactivate dangerous species via a bulk reaction before they leak through defects in the boundary of the container; (vii) for a safe storage of nuclear wastes, the motion of radioactive nuclei should be slowed down enough to ensure their disintegration or at least to reduce the amount of released nuclei, etc. While some first passage problems have been recently extended to mortal walkers (see and references therein), the effect of a finite lifetime of a walker onto the escape through the boundary of two- and three-dimensional confining domains has not been investigated. Since the escape is not certain, because of a possible "death" of the walker, the contribution of long trajectories towards the escape region can be greatly reduced, thus strongly affecting the conventional results.
The paper is organized as follows. In Sec. II, we formulate the general escape problem for mortal walkers and we then discuss the most relevant case of a first-order bulk kinetics. We also introduce the exit probability of mortal walkers. In Sec. III, we quantify the impact of the finite lifetime of the walkers onto their survival probability, MFPT, and exit probability. For this purpose, we derive the upper and lower bounds of the Laplace-transformed survival probability and related quantities for arbitrarybounded domains, and analyze their asymptotic behavior at small, intermediate, and large death rates. To check the quality of these general estimates and asymptotics, we compare them to exact solutions that we obtain for concentric domains and for an escape region on the boundary of a disk. In Sec. IV, various extensions and applications of these results are discussed, in particular, the problem of leakage control and optimization. Section V summarizes and concludes the paper.
## II Notations and equations
In mathematical terms, the first passage time \(\tau\) is a random variable which is determined by the survival probability \(S(t;x_{0})=\mathbb{F}_{x_{0}}\{\tau>t\}\) that a particle started at a point \(x_{0}\in\Omega\) has not left a confining domain \(\Omega\subset\mathbb{R}^{d}\) through an escape region \(\Gamma\) on the boundary \(\partial\Omega\). The survival probability can be expressed through the diffusion propagator for the escape problem, \(G_{t}(x;x_{0})\), i.e., the probability density for a particle started at \(x_{0}\in\Omega\), to be at position \(x\) after time \(t\).
\[S(t;x_{0})=\int\limits_{\Omega}dx\,G_{t}(x;x_{0}). \tag{2}\]
The Laplace transform reduces the diffusion equation for the survival probability (which follows from Eq.) to a simpler Helmholtz equation,
\[\big{[}D\Delta-p\big{]}\tilde{S}(p;x_{0})=-1, \tag{3}\]
with the same mixed boundary condition, where tilde denotes the Laplace-transformed survival probability:
\[\tilde{S}(p;x_{0})=\int\limits_{0}^{\infty}dt\,e^{-tp}\,S(t,x_{0}). \tag{4}\]
Once Eq. is solved, the inverse Laplace transform of \(\tilde{S}(p;x_{0})\) yields the survival probability \(S(t;x_{0})\) in time domain.
\[\rho(t;x_{0})=-\frac{\partial}{\partial t}S(t;x_{0}), \tag{5}\]
yields the usual Poisson equation for the MFPT).
Next, we formulate the escape problem for mortal walkers. We assume that the lifetime \(\chi\) of a walker is independent of the search process. This situation corresponds to the case of a uniform bulk reaction or to some internal death mechanism of a walker, e.g., a life cycle of a bacterium or a radioactive decay. In this case, the survival probability of a mortal walker inside the domain \(\Omega\) up to time \(t\), \(S_{\mu}(t;x_{0})\), is simply the product of \(S(t;x_{0})\) and \(Q(t)=\mathbb{P}\{\chi>t\}\):
\[S_{\mu}(t;x_{0})=\mathbb{P}\{\min\{\tau,\chi\}>t\}=S(t;x_{0})\ Q(t). \tag{6}\]
While this expression is general, it remains formal as the survival probability \(S(t;x_{0})\) is not known analytically except for some elementary cases (see Appendix A).
A particular simplification occurs in the relevant case of an exponentially distributed lifetime, \(Q(t)=\exp(-\mu t)\), with \(\mu\) being the death rate (or \(1/\mu\) being the mean lifetime), for which the Laplace transform of Eq.
\[\tilde{S}_{\mu}(p;x_{0})=\tilde{S}(p+\mu;x_{0}). \tag{7}\]
This is equivalent to adding the term \(-\mu\tilde{S}\) into Eq. to describe a first-order bulk kinetics or radioactive decay.
(Color online) Schematic illustration of the escape problem. **(a)** A general domain \(\Omega\subset\mathbb{R}^{d}\) with an escape region \(\Gamma\) (dashed line) on its boundary \(\partial\Omega\). A particle is released at a starting point \(x_{0}\) and diffuses, with eventual reflections on \(\partial\Omega\backslash\Gamma\), until it reaches the escape region \(\Gamma\). **(b,c)** Two explicitly solvable cases of the escape problem: from a disk through an arc on its boundary **(b)**, see Appendix C, and from an annulus through its inner circle **(c)**, see Appendix A.
The MFPT for mortal walkers is then
\[\langle\tau_{\mu}\rangle=\tilde{S}_{\mu}(0;x_{0})=\tilde{S}(\mu;x_{0}), \tag{8}\]
Since the Laplace transform of a positive function monotonously decreases, the MFPT monotonously decreases with \(\mu\). In particular, the MFPT for mortal walkers is smaller than that for immortal ones. This is expected because long trajectories to the escape region are progressively eliminated as \(\mu\) increases.
Finally, we introduce the exit probability of mortal walkers, \(H_{\mu}(x_{0})\), i.e., the probability that a walker started from \(x_{0}\) leaves the confining domain before dying: \(H_{\mu}(x_{0})=\mathbb{P}\{\tau<\chi\}\).
\[\begin{split} H_{\mu}(x_{0})&=\int\limits_{0}^{ \infty}\mathbb{P}\{\chi>t\}\,\mathbb{P}\{\tau\in(t,t+dt)\}\\ &=\int\limits_{0}^{\infty}dt\,Q(t)\,\rho(t;x_{0}).\end{split} \tag{9}\]
For an exponentially distributed lifetime, the exit probability becomes
\[H_{\mu}(x_{0})=\int\limits_{0}^{\infty}dt\,e^{-\mu t}\,\rho(t;x_{0})=1-\mu \tilde{S}(\mu;x_{0})=1-\mu\langle\tau_{\mu}\rangle, \tag{10}\]
The exit probability satisfies
\[\begin{split} D\Delta H_{\mu}(x_{0})-\mu H_{\mu}(x_{0})& =0\quad(x_{0}\in\Omega),\\ H_{\mu}(x_{0})&=1\quad(x_{0}\in\Gamma),\\ \frac{\partial}{\partial n}H_{\mu}(x_{0})&=0\quad(x_ {0}\in\partial\Omega\backslash\Gamma).\end{split} \tag{11}\]
As expected, this probability is equal to 1 for immortal walkers (i.e., for the conventional escape problem with \(\mu=0\)) and monotonously decreases to 0 as \(\mu\) increases. For mortal walkers, the exit probability characterizes the storage safety of a container.
In many applications, the starting point \(x_{0}\) is not fixed but randomly distributed over the confining domain. In this case, one considers the global, or volume-averaged MFPT:
\[\overline{\langle\tau_{\mu}\rangle}=\frac{1}{|\Omega|}\int\limits_{\Omega}dx_ {0}\,\langle\tau_{\mu}\rangle\,, \tag{12}\]
In analogy, we define the global exit probability (GEP) to characterize the overall safety of the container:
\[\overline{H}_{\mu}=\frac{1}{|\Omega|}\int\limits_{\Omega}dx_{0}\,H_{\mu}(x_{0 }). \tag{13}\]
Integrating Eq. over the confining domain and using the Green formula and boundary conditions, one gets another representation of the GEP:
\[\overline{H}_{\mu}=\frac{D}{\mu|\Omega|}\int\limits_{\Gamma}dx_{0}\,\frac{ \partial H_{\mu}(x_{0})}{\partial n} \tag{14}\]
When the mean lifetime \(1/\mu\) is much larger that the time needed to find the escape region by immortal walkers, i.e., \(\mu\langle\tau_{0}\rangle\ll 1\), the Taylor expansion of Eq.
\[\langle\tau_{\mu}\rangle=\langle\tau_{0}\rangle-\frac{1}{2}\mu\langle\tau_{0} ^{2}\rangle+O(\mu^{2}) \tag{15}\]
and, from Eq.,
\[H_{\mu}(x_{0})=1-\mu\langle\tau_{0}\rangle+\frac{1}{2}\mu^{2}\langle\tau_{0}^ {2}\rangle+O(\mu^{3}). \tag{16}\]
The problem is reduced to the analysis of the moments of the FPT of immortal walkers
\[\langle\tau_{0}^{n}\rangle=(-1)^{n}\left(\frac{\partial^{n}}{\partial\mu^{n}} H_{\mu}(x_{0})\right)_{\mu=0}. \tag{17}\]
In particular, the exact formulas for the MFPT are known for several simple domains (see also Appendices A and C). When the escape region is small, the following asymptotic behavior was established (see the review and references therein):
\[\begin{split}\langle\tau_{0}\rangle\simeq\begin{cases}\frac{| \Omega|}{\pi D}\bigl{(}\ln(1/\epsilon)+O\bigr{)}&(d=2),\\ \frac{|\Omega|}{4D\ell\epsilon}+O(\ln\epsilon)&(d=3),\end{cases}\end{split} \tag{18}\]
In the context of chemical reactions, the escape region can be interpreted as an active catalytic site on an inert confining boundary of a reactor, the FPT \(\tau_{\mu}\) is the (random) reaction time (for a perfectly reactive catalyst), \(\langle\tau_{\mu}\rangle\) is the mean reaction time, and \(H_{\mu}(x_{0})\) is the target encounter probability. While we adopt the language of first-passage processes, the following results can be easily translated into the chemical or biochemical context.
In what follows, we investigate the Laplace-transformed survival probability, the MFPT, the exit probability and the GEP for mortal walkers in the limit of short lifetimes, \(\mu\langle\tau_{0}\rangle\gg 1\), which turns out to be more relevant for many applications, e.g., for developing safe containers. Note also that the large \(\mu\) limit captures the short-time asymptotic behavior of the survival probability and related quantities according to Tauberian theorems. In contrast to the long-time asymptotics, the short-time behavior of first passage distributions is generally not universal. Nevertheless, we will obtain a universal relation for the global exit probability within the short-time scale regime.
## III Main results
### Upper and lower bounds
First, we establish the upper and lower bounds for the Laplace-transformed survival probability. An elementary upper bound follows from the trivial inequality \(S(t,x_{0})\leq 1\) by applying the Laplace transform:
\[\tilde{S}(\mu;x_{0})\leq\frac{1}{\mu}. \tag{19}\]
The lower bound can be obtained from the continuity of Brownian motion that implies that the FPT to a subset \(\Gamma\) of the boundary \(\partial\Omega\) is greater than the FPT from the center of the ball of radius \(|x_{0}-\partial\Omega|\) to its boundary, where \(|x_{0}-\partial\Omega|\) is the distance from \(x_{0}\) to the boundary \(\partial\Omega\).
\[S(t;x_{0})\geq S_{B_{d}(|x_{0}-\partial\Omega|)}(t;0), \tag{20}\]
We claim a stronger inequality
\[S(t;x_{0})\geq S_{B_{d}(|x_{0}-\Gamma|)}(t;0), \tag{21}\]
Since the Laplace transform of positive functions does not affect the inequality, one also gets the explicit lower bound in the Laplace domain
\[\tilde{S}(\mu;x_{0})\geq\tilde{S}_{B_{d}(|x_{0}-\Gamma|)}(\mu;0). \tag{22}\]
The right-hand side is known explicitly:
\[\tilde{S}_{B_{d}(R)}(\mu;0)=\frac{1}{\mu}\big{(}1-U_{d}(R\sqrt{\mu/D})\big{)}, \tag{23}\]
with
\[U_{d}(x)=\frac{\big{(}x/2\big{)}^{\frac{d}{2}-1}}{\Gamma\big{(}\frac{d}{2} \big{)}\ I_{\frac{d}{2}-1}(x)}, \tag{24}\]
When \(x\gg 1\), one has
\[U_{d}(x)\simeq\frac{\sqrt{\pi}\,2^{\frac{3-d}{2}}}{\Gamma(d/2)}\,x^{\frac{d-1 }{2}}e^{-x}. \tag{25}\]
Combining the upper and lower bounds, one concludes that the MFPT, \(\langle\tau_{\mu}\rangle=\tilde{S}(\mu;x_{0})\), can be accurately approximated as \(1/\mu\) at large \(\mu\) for any domain and any escape region, the error of this approximation vanishing as a stretched-exponential function according to Eq.. As a consequence, the MFPT becomes fully controlled by the lifetime of the walker in this limit, as expected. Similarly, the global MFPT, \(\overline{\langle\tau_{\mu}\rangle}\), behaves as \(1/\mu\) at large \(\mu\), although it includes contribution from points which are close to the escape region.
One can see that the two bounds accurately approximate the MFPT at large \(\mu\). In turn, they do not control the behavior of the MFPT at small \(\mu\), in which case the lower bound tends to \(|x_{0}-\Gamma|^{2}/(2dD)\), and the upper bound diverges. In this limit, one can use the asymptotic relation.
In turn, the exit probability \(H_{\mu}(x_{0})\) becomes a non-trivial characteristics of the escape problem at large \(\mu\). According to Eqs., \(H_{\mu}(x_{0})\) is bounded as
\[H_{\mu}(x_{0})\leq U_{d}\big{(}|x_{0}-\Gamma|\sqrt{\mu/D}\big{)}. \tag{26}\]
(Color online) The MFPT \(\langle\tau_{\mu}\rangle\) **(a)** and the exit probability \(H_{\mu}(x_{0})\)**(b)**, as functions of \(\mu R^{2}/D\) and \(R\sqrt{\mu/D}\), respectively, for the disk of radius \(R\) with the escape arc \((\pi-\varepsilon,\pi+\varepsilon)\) on the boundary, for several \(\varepsilon\) (here \(x_{0}=0\)). Inset shows the MFPT on log-log scale. For the MFPT, the exact solution is compared to the upper and lower bounds. For the exit probability, the exact solution is compared to the upper bound (the lower bound being zero).
Since the upper bound is independent of the size of the escape region, it does not capture well the accessibility of this region. In fact, the exit probability is expected to decay faster with \(\mu\) for smaller escape regions. This is illustrated in that shows \(H_{\mu}(x_{0})\) for the disk with escape regions of various sizes. While the upper bound over-estimates the exit probability, it captures correctly its asymptotic decay as a stretched-exponential function. The quality of the upper bound is further discussed in the case of a disk in Appendix C.3. Note also that this upper bound is equal to the exit probability in the case of a ball with the escape region on the whole boundary and the starting point at the origin.
According to the inequality and numerical analysis, the exit probability \(H_{\mu}(x_{0})\) at large \(\mu\) exponentially decays with the distance from the starting point to the escape region. For a better control of this decay, one needs a lower bound for the exit probability. In Appendix D, we partly solve this problem for the specific case when the escape region is the whole boundary: \(\Gamma=\partial\Omega\). Finding an appropriate lower bound for the general escape problem (with \(\Gamma\neq\partial\Omega\)) remains an open problem.
### Global exit probability
When the starting point is uniformly distributed, some walkers start near the escape region, and the stretched-exponential decay with \(\mu\) is expected to be replaced by a slower power law. Determining an exact expression for the GEP appears to be a challenging problem because of mixed boundary conditions. Here, we derive an approximation for the GEP for intermediate death rates, which also turns out to be exact in the limit of small death rates. For small escape regions, this approximation shows satisfactory agreement with the results of numerical simulations over a broad range of \(\mu\).
The survival probability \(S(t;x_{0})\) in a bounded domain admits a spectral decomposition on appropriate Laplacian eigenfunctions and thus decreases exponentially at long times as \(S(t;x_{0})\sim A(x_{0})e^{-\lambda t}\), where \(\lambda\) is the smallest Laplacian eigenvalue, and \(A(x_{0})\) is a coefficient. When the escape region is small as compared to the size of the confining domain, one expects \(A(x_{0})\simeq 1\) and \(\overline{\langle\tau_{0}\rangle}=1/\lambda\), independently of the starting point \(x_{0}\) if \(x_{0}\) is far from the escape region.
\[\overline{H}_{\mu}=1-\mu\,\overline{S}(\mu)\approx\frac{1}{1+\mu\,\langle\tau_ {0}\rangle}. \tag{27}\]
In the limit \(\mu\to 0\), one retrieves \(\overline{H}_{0}=1\) for any escape region size, independently of its smallness, as expected. In turn, the exit probability vanishes for any \(\mu>0\) in the limit of shrinking escape region because \(\overline{\langle\tau_{0}\rangle}\to\infty\). Note also that this expression can be re-written as \(1/\overline{\langle\tau_{\mu}\rangle}=1/\overline{\langle\tau_{0}\rangle}+\mu\), i.e., the volume-averaged MFPT for mortal walkers is the harmonic mean of \(\overline{\langle\tau_{0}\rangle}\) and \(1/\mu\). We point out that Eq. is exact, to the first order, in the limit \(\mu\overline{\langle\tau_{0}\rangle}\ll 1\). This can be seen by comparing the asymptotic expansion of Eq. to the integral over \(x_{0}\) of the Taylor expansion.
We compare the approximation to two exact solutions: the concentric escape region (Appendix A) and the disk with an escape arc \((\pi-\varepsilon,\pi+\varepsilon)\) on the boundary (Appendix C). We present the results only for the latter case, for which shows the GEP as a function of the death rate. The approximation (shown by dashed lines) accurately captures the behavior of the GEP at small \(\mu\) and under-estimates its values at larger \(\mu\). In the narrow escape configuration, when \(\overline{\langle\tau_{0}\rangle}\gg R^{2}/D\), the approximation remains accurate even for \(\mu\sim 1/\overline{\langle\tau_{0}\rangle}\). However, Eq. fails at large death rates \(\mu\gg D/(\varepsilon R)^{2}\) that correspond to the short-time behavior of the FPT distribution.
The mono-exponential approximation of the survival probability fails at very short times. Similarly to, we show that in the large death rate limit (which corresponds to the few-encounter regime in), the notion of a kinetic rate in the traditional bulk sense does not hold, and the asymptotic regime of the GEP should be different. In this limit, the exit probability \(H_{\mu}(x_{0})\) rapidly decays from the escape region \(\Gamma\) towards the bulk. In the vicinity of each escape point, the exit probability along normal vector can be approximated by the one-dimensional solution of Eq. on the positive half-line: \(H_{\mu}(x_{0})=\exp(-x_{0}\sqrt{\mu/D})\). Substituting this approximation in Eq., one finds
\[\overline{H}_{\mu}\underset{\mu\rightarrow\infty}{\sim}\frac{|\Gamma|}{| \Omega|}\left(\mu/D\right)^{-\frac{1}{2}}. \tag{28}\]
(Color online) The global (or volume-averaged) exit probability \(\overline{H}_{\mu}\) (symbols) from Eq. as a function of \(\mu R^{2}/D\) for the disk of radius \(R\) with the escape arc \((\pi-\varepsilon,\pi+\varepsilon)\) on the boundary. Dashed lines show the approximation, with the volume-averaged MFPT \(\overline{\langle\tau_{0}\rangle}=\frac{R^{2}}{D}\left(1/8-\ln(\sin(\varepsilon /2))\right)\) while solid lines present the asymptotic relation.
The right-hand side can be interpreted as the probability that a uniformly distributed starting point lies within a thin layer of width \(\ell=(\mu/D)^{-1/2}\) near the escape region. The expression is confirmed on the exactly solvable cases presented in Appendices A and C. We also found a perfect agreement of Eq. with two numerical simulations in which the target is (i) a square concentric to the confining disk; and (ii) a sphere included within the boundary of a sphere, both two and three dimensions (results are not shown).
Comparing approximate relations, we define the transition death rate \(\mu_{c}\) between two asymptotic regimes:
\[\mu_{c}=\frac{(|\Omega|/|\Gamma|)^{2}}{D\overline{\langle\tau_{0}\rangle}^{2} }\,. \tag{29}\]
Substituting the asymptotic relation for the global MFPT, one finds that \(\mu_{c}\sim D/a^{2}\) in both two and three dimensions, where \(a\) is the size of the escape region.
### First passage position
While we mainly focused on the FPT to the escape region, the identification and further isolation of the most probable exit locations is also of practical importance. In this section, we show how the distribution of first passage positions (FPP) can be expressed through the MFPT of mortal walkers.
For this purpose, we recall that the probability flux density,
\[q(t,x;x_{0})=-D\frac{\partial}{\partial n}G_{t}(x;x_{0}), \tag{30}\]
In particular, the integral of this function over \(x\in\Gamma\) yields the marginal probability density \(\rho(t;x_{0})\) for the FPT,
\[\rho(t;x_{0})=\int\limits_{\Gamma}dx\,q(t,x;x_{0}) \tag{31}\]
(see Appendix E).
\[\omega_{\mu}(x;x_{0})=\int\limits_{0}^{\infty}dt\,q(t,x;x_{0})\,e^{-\mu t}, \tag{32}\]
Finally, if the starting point \(x_{0}\) is distributed uniformly, one gets for any \(x\in\partial\Omega\)
\[\overline{\omega}_{\mu}(x) =\frac{1}{|\Omega|}\int\limits_{\Omega}dx_{0}\,\omega_{\mu}(x;x_{ 0})=-\frac{D}{|\Omega|}\int\limits_{0}^{\infty}dt\,\frac{\partial}{\partial n }S(t;x)\,e^{-\mu t}\] \[=-\frac{D}{|\Omega|}\frac{\partial\tilde{S}(\mu;x)}{\partial n}=- \frac{D}{|\Omega|}\frac{\partial}{\partial n}\langle\tau_{\mu}\rangle=\frac{D} {\mu|\Omega|}\frac{\partial H_{\mu}(x)}{\partial n}. \tag{33}\]
In other words, the MFPT for mortal walkers also determines their escape positions.
We illustrate the result in the case of an escape from a ball of radius \(R\), for which the MFPT \(\langle\tau_{\mu}\rangle\) is given in Eq.. In the limit of small death rates, the distribution converges to \(\overline{\omega}_{\mu}\sim 1/|\Gamma|\), where \(|\Gamma|=|\partial\Omega|\) is the surface of the \((d-1)\)-sphere, which corresponds to the expected uniform measure over the sphere. At large death rates, we find that \(\overline{\omega}_{\mu}\sim\frac{1}{|\Omega|}(\mu/D)^{-1/2}\) in any dimensions. The latter relation is expected since it corresponds to the product of the uniform measure on the sphere and the GEP defined in Eq..
The expression is of particular interest in the study of the narrow escape problem. As discussed in for immortal walkers, the probability density \(\overline{\omega}_{0}(\theta)\) diverges at the boundaries of the escape region. In other words, immortal walkers tend to exit through the edges of the escape region. We show that this proclivity is hindered at large death rates \(\mu\), at which \(\overline{\omega}_{\mu}(\theta)\sim(\mu/D)^{-1/2}\) (see Fig. 4). In the next section, we discuss the implications of the asymptotic behavior at low and high death rates for the leakage control.
(Color online) FPP probability density \(\overline{\omega}_{\mu}(x)\) (given by Eq.), multiplied by \(\sqrt{\mu R^{2}/D}\), as a function of angular coordinate \(\theta\) of the passage position \(x\) on \(\Gamma\), for the disk of radius \(R\) with an escape arc \((\pi-\varepsilon,\pi+\varepsilon)\) and \(\varepsilon=0.2\) (see Fig. 1b). For a small death rate \(\mu R^{2}/D=0.1\), the density greatly increases on the edges of the escape region (at \(\theta=\varepsilon\)). For a larger death rate \(\mu R^{2}/D=100\), the density becomes flatter, illustrating the convergence to a uniform distribution as \(\mu\to\infty\). Note that only the half of the density, for \(\pi\leq\theta\leq\pi+\varepsilon\), is shown (the other half being symmetric).
Discussion
### Leakage control
In many real-life applications, one needs to fabricate efficient containers for a safe storage of dangerous species such as nuclear wastes or toxic chemicals. For this purpose, one can either improve the isolation of the container, or incorporate mechanisms to bind, transform, deactivate or disintegrate dangerous species in the bulk or on the boundary. In mathematical terms, the first strategy aims at shrinking escape regions to increase the MFPT \(\left\langle\tau_{0}\right\rangle\) for immortal walkers according to the asymptotic relation. Since a complete isolation is not realistic (e.g., due to a slow but permanent degradation of container materials), the exit probability of intact diffusing species (with \(\mu=0\)) would be equal to 1. In other words, whatever the isolation improvements are, the leakage is only a matter of time. It is therefore important to implement in parallel the second strategy that aims at reducing the mean lifetime \(1/\mu\). Note that this is a natural frame for nuclear wastes which disintegrate by a radioactive decay.
The asymptotic analysis in Sec. III has shown how the MFPT \(\left\langle\tau_{\mu}\right\rangle\) and the exit probability \(H_{\mu}(x_{0})\) are affected by the finite lifetime of the walkers. When the mean lifetime \(1/\mu\) is much larger than the time \(\left\langle\tau_{0}\right\rangle\) needed to find the escape region by immortal walkers, the decay mechanism weakly affects the escape process. In particular, the exit probability remains close to 1, indicating on a poor leakage protection. Improving either the isolation, or the decay mechanism, or both, one aims at switching to the opposite limit \(\mu\langle\tau_{0}\rangle\gg 1\), in which the exit probability can be significantly reduced.
In this regime, the MFPT \(\left\langle\tau_{\mu}\right\rangle\) was shown to decrease universally as \(1/\mu\), i.e., this quantity essentially reflects the mean lifetime. In turn, the exit probability \(H_{\mu}(x_{0})\) remains informative and was shown to decay as stretched-exponential, \(\exp(-|x_{0}-\Gamma|\sqrt{\mu/D})\), when the death rate \(\mu\) increases. While we could not provide a general form of the dependence of the exit probability \(H_{\mu}(x_{0})\) on the size of the escape region, we derived the asymptotic behavior of the exit probability for several specific domains (see Appendices A and C).
The leakage of many species uniformly distributed in a container can be characterized by the global exit probability \(\overline{H}_{\mu}\) which exhibits a much slower decay with \(\mu\), as suggested by approximations. The approximation can be interpreted as the result of competition of two first-order kinetics:
\[\text{Dead}\underset{\mu}{\longleftarrow}\text{Alive}\underset{k_{e}}{ \longrightarrow}\text{Exited} \tag{34}\]
In the stationary state, the law of mass action predicts that the proportion of dead \(n_{D}\) to exited \(n_{E}=1-n_{D}\) walkers reads \(n_{D}/n_{E}=\mu\,\overline{\left\langle\tau_{0}\right\rangle}\), from which \(n_{E}=1/\left(1+\mu\,\overline{\left\langle\tau_{0}\right\rangle}\right)\). Since the proportion \(n_{E}\) of exited walkers is precisely the exit probability \(\overline{H}_{\mu}\), one recovers the approximation. One can conclude that the GEP remains close to 1 at small \(\mu\) even for high quality isolation (small \(\varepsilon\)), whereas shrinking the escape region allows one to significantly improve the safety of a container when \(\mu R^{2}/D\gtrsim 1\). At very large death rates, the above kinetic argument fails, and the GEP is determined by the universal asymptotic relation.
Note also that active protecting mechanisms can be implemented not only in the bulk, but also on the container boundary, including the escape regions. As shown in, the introduction of an energetic or entropic barrier at the escape region significantly affects the MFPT \(\left\langle\tau_{0}\right\rangle\) for immortal walkers, making the escape process "barrier-limited" instead of "diffusion-limited". In particular, the conventional asymptotic behavior is replaced by a faster divergence as \(\left(\kappa/D\right)\,\epsilon^{1-d}\), where \(\kappa\) is the barrier reactivity or permeability. In other words, the presence of energetic or entropic barrier greatly improves the quality of isolation and allows one to reduce the exit probability. In the same vein, long-ranged repulsive interactions keep diffusing species expelled from the boundary and thus increase the MFPT of immortal walkers. In addition, our general result concerning the distribution of exit positions can be used to design protecting shields that would be adapted to the death rate of the toxic reactant. In particular, we showed that highly mortal walkers hit uniformly the escape region, in contrast to immortal walkers that mainly exit through the edges of the escape region (see discussion in Sec. III.3).
### Scaling argument for mRNA translation
As another application of our results, we propose a scaling argument to express the variability of the mRNA lifetime between E. Coli, Yeast and Human cells in terms of the cell volume and the number of ribosomes. In this biological context, the escape region is the surface of ribosomes, while the exit probability can be interpreted as the probability for an mRNA to encounter a ribosome before being degradated. Assuming that the asymptotic relation holds (i.e., that mRNA translation occurs in the high degradation rate regime), we get the scaling relation
\[\frac{Ns_{r}}{V}\,\sqrt{D/\mu}\sim\overline{H}_{\mu} \tag{35}\]
In addition, we assume that (i) the diffusion coefficient \(D\) is the same for different species and (ii) that an identical "success rate" \(\overline{H}_{\mu}\) in the reaction kinetics should be maintained for all cells to ensure an efficient translation mechanism. From Eq., the ratio \(N/(V\sqrt{\mu})\)should then be approximatively constant among different species. This scaling argument is confirmed by comparing E. Coli, Yeast, and Human cells (Table 1). Using the known data for one cell type, one should thus be able to estimate the degradation rate of mRNA in another cell type from its volume and the number of ribosomes. Similar arguments could hold for the transcription problem, involving the search of a specific DNA site by transcription factors. As discussed in, it appears that only the fastest \(0.01\%\) to \(1\%\) of transcription factors actually matter for the cellular response.
In summary, our results for the GEP provides a new interpretation for the cell size scaling problem, which has received a large attention within the recent years (see and references therein). By narrowing the spread of the first passage times, the degradation rate might be used to focus the temporal cellular response to external perturbations.
### Extensions
The escape problem for mortal walkers can be extended in various ways. First, the diffusion equation in Eqs. can be replaced by a more general backward Fokker-Planck (or Kolmogorov) equation to account for the effect of external forces or potentials. For instance, external forces can model the effect of chemotactic gradient which was proved to be relevant for the egg search problem by spermatozoa. Second, the Dirichlet boundary condition on the escape region can be replaced by a Robin condition to model the presence of a recognition step, partial reflections, stochastic gating, or microscopically heterogeneous distribution of exit channels. The survival probability and the distribution of FPTs to the whole partially absorbing boundary have been earlier studied for immortal walkers. Third, one can consider intermittent processes, with alternating phases of bulk and surface diffusion. Fourth, one can investigate the effect of heterogeneous materials or multiple layers of a container with different diffusion or trapping properties (this is a typical situation for nuclear waste storage). Fifth, the retarding effect of geometric or energetic traps can be included by considering the continuous time random walk (CTRW) with a fat-tailed waiting time distribution. The extension of these problems to mortal walkers consists in replacing the Laplace transform variable \(p\) by \(p+\mu\), as in Eq. for normal diffusion.
To illustrate this point, we consider an extension to CTRW for which the propagator obeys the fractional diffusion equation, while the Laplace-transformed survival probability satisfies the Helmholtz equation
\[\big{[}p^{1-\alpha}D_{\alpha}\Delta-p\big{]}\tilde{S}^{\alpha}(p;x_{0})=-1, \tag{36}\]
Changing the variable \(p\) to \(Dp^{\alpha}/D_{\alpha}\), the solution of this equation can be expressed through the earlier obtained \(\tilde{S}(p;x_{0})\):
\[\tilde{S}^{\alpha}(p;x_{0})=\frac{p^{\alpha-1}D}{D_{\alpha}}\,\tilde{S}(Dp^{ \alpha}/D_{\alpha};x_{0}). \tag{37}\]
From this relation, one immediately retrieves that the MFPT of immortal walkers is infinite: \(\langle\tau_{0,\alpha}\rangle=\tilde{S}^{\alpha}(0;x_{0})=\infty\) because the walkers can be trapped for long periods of time until they reach the escape region. In contrast, the MFPT for mortal walkers is finite:
\[\langle\tau_{\mu,\alpha}\rangle=\frac{\mu^{\alpha-1}D}{D_{\alpha}}\,\tilde{S} (D\mu^{\alpha}/D_{\alpha};x_{0})=\frac{\mu^{\alpha-1}D}{D_{\alpha}}\,\langle \tau_{D\mu^{\alpha}/D_{\alpha}}\rangle. \tag{38}\]
In fact, too long trajectories whose contribution led to divergence of the MFPT for CTRW, are eliminated because of a finite lifetime of the walker. In other words, the MFPT for mortal CTRW is related to the MFPT for mortal normal walkers with a modified death rate: \(\mu_{\alpha}=D\mu^{\alpha}/D_{\alpha}\). Finally, the exit probability for mortal CTRWs that can be defined in analogy with Eq.
\[H_{\mu,\alpha}(x_{0})=H_{\mu_{\alpha}}(x_{0}). \tag{40}\]
One can therefore apply the results from Sec. III to mortal CTRWs.
In all these extensions, the first-order bulk kinetics or, equivalently, an exponentially distributed lifetime of the walker, controls the duration of trajectories, assigning smaller weights to longer trajectories. We also mention the possibility of considering other lifetime distributions \(Q(t)\) beyond the exponential one. In this case, Eq. is still applicable but one needs to get the survival probability in time domain by inverse Laplace transform. Finally, if the death mechanism is coupled to diffusion (e.g., in the case of a space-dependent death rate), one has to treat the whole diffusion-reaction mixed boundary value problem.
\begin{table}
\begin{tabular}{|c|c|c|c|} \hline & E. Coli & Yeast & Human \\ \hline \(N\) & \(10^{4}\) & \(10^{5}\) & \(10^{6}\) \\ \(V\) (in \(\mu\)m\({}^{3}\)) & 2 & 40 & 2000 \\ \(1/\mu\) (in min) & 5 & 20 & 600 \\ \hline \(10^{-4}\)\(N/(V\sqrt{\mu})\) & 1.1 & 1.1 & 1.2 \\ (in \(\mu\)m\({}^{-3}\).min\({}^{1/2}\)) & & & \\ \hline \end{tabular}
\end{table}
Table 1: The number of ribosomes (\(N\)), the cell volume (\(V\)), and the mRNA degradation rate (\(\mu\)) for E. Coli, Yeast, and Human fibroblast cells.
Conclusion
We formulated the escape problem to mortal walkers and investigated how their finite lifetime drastically affects the survival and exit probabilities. The latter is the likelihood of escape or leakage from the confining container and can thus characterize its isolation quality. We focused on the most relevant case of a first-order bulk kinetics or, equivalently, an exponentially distributed lifetime, for which the problem is reduced to finding the Laplace-transformed survival probability for immortal walkers. We derived the upper and lower bounds for the MFPT \(\langle\tau_{\mu}\rangle\) and the exit probability \(H_{\mu}(x_{0})\) and analyzed their asymptotic behavior at small and large death (or reaction) rates. When the mean lifetime, \(1/\mu\), is much larger than the MFPT for immortal walkers, \(\langle\tau_{0}\rangle\), the exit probability remains close to 1, meaning a poor isolation. In this situation, the leakage of dangerous species is just a matter of time. For a safer protection, one needs both to improve the boundary isolation by shrinking escape regions (thus increasing \(\langle\tau_{0}\rangle\)), and to implement efficient trapping, binding or deactivation mechanisms (thus increasing \(\mu\)). Improving only one of these aspects is not sufficient to significantly reduce \(H_{\mu}(x_{0})\). For the volume-averaged, or global exit probability \(\overline{H}_{\mu}\) that quantifies the overall safety of a container, we obtained two approximations at intermediate and large death rates. The quality of the obtained analytical results for general confining domains was confirmed by comparison with several explicitly solvable cases, and with numerical simulations. We also introduced and investigated the distribution of the first passage positions as a mathematical ground for design and optimization of containers. The density \(\overline{\omega}_{\mu}(x)\) was shown to exhibit a transition from a singular function highly localized near edges of the escape region for immortal walkers, to asymptotically uniform density at large death rates. Various extensions and applications of the studied escape problem have been discussed including the leakage control of dangerous chemicals and scaling relation for mRNA translation mechanism.
|
10.48550/arXiv.1611.08901
|
The escape problem for mortal walkers
|
D. S. Grebenkov, J. -F. Rupprecht
| 4,139
|
10.48550_arXiv.2409.16905
|
###### Abstract
Mixtures involving nitrobenzene and hydrocarbons, or 1-alkanols and 1-nitroalkane, or nitrobenzene have been investigated on the basis of a whole set of thermophysical properties available in the literature. The properties considered are: excess molar functions (enthalpies, entropies, isobaric heat capacities, and volumes), vapour-liquid and liquid-liquid equilibria, permittivities or dynamic viscosities. In addition, the mixtures have been studied by means of the application of the DISQUAC, ERAS, and UNIFAC models, and using the formalism of the concentration-concentration structure factor. The corresponding interaction parameters in the framework of the DISQUAC and ERAS models are reported. In alkane mixtures, dipolar interactions between 1-nitroalkane molecules are weakened when the size of the polar compound increases, accordingly with the relative variation of their effective dipolar moment. Dipolar interactions are stronger in nitrobenzene solutions than in those containing the smaller 1-nitropropane, although both nitroalkanes have very similar effective dipole moment (aromaticity effect). Systems with 1-alkanols are characterized by dipolar interactions between like molecules which sharply increases when the alkanol size increases. Simultaneously, interactions between unlike molecules become weaker, as the OH group is then more sterically hindered. Interactions between unlike molecules are stronger in systems with nitromethane than in nitrobenzene solutions. The replacement of nitromethane by nitroethane in systems with a given 1-alkanol leads to strengthen those effects related with the alcohol self-association. Permittivity data and results on Kirkwood's correlation factors show that the addition of 1-alkanol to a nitroalkane leads to cooperative effects, which increase the dipolar polarization of the solution, in such way that the destruction of the existing structure in pure liquids is partially counterbalanced. This effect is less important when longer 1-alkanols are involved.
Keywords: nitroalkanes; 1-alkanols; thermophysical data; models; dipolar interactionsIntroduction
Nitroalkanes are aprotic solvents of high polarity as it is demonstrated by their large dipole moments (3.56 D (nitromethane); 3.60 D, (nitroethane); 4.0 (nitrobenzene)). They have many applications. For example, nitromethane is widely used in the manufacture of pharmaceuticals, pesticides or fibers. The industrial interest on the chemistry of nitrobenzene mixtures is due to this compound plays an essential role in the aniline production, and in the preparation of other substances as dyes, paint solvents or the analgesic paracetamol. Unfortunately, it is highly toxic and the hazardous effects to soil, groundwater and human health must be taken into account.
There is little evidence that nitroalkanes are self-associated in the pure state. However, as a consequence of their large \(\mu\) values, strong dipolar interactions exist between nitroalkane molecules, and binary mixtures formed by 1-alkanol (from 1-butanol) and nitromethane show liquid-liquid equilibrium (LLE) curves with upper critical solution temperatures (UCST) ranged between 291.1 K (1-butanol) and 352.6 K (1-pentadecanol). Similarly, the UCST of the 1-decanol + nitroethane system is 294.1 K. In addition, rather large positive values of molar excess Gibbs energies (\(G_{m}^{E}\)), and of enthalpies (\(H_{m}^{E}\)) are encountered for the methanol, or ethanol or 1-propanol, or 1-butanol + nitromethane mixtures. That is, 1-alkanol + 1-nitroalkane systems are characterized by positive deviations from the Raoult's law. Interestingly, non-random effects in the mentioned solutions have been investigated by measuring isobaric excess molar heat capacities (\(C_{pm}^{E}\)), a very useful magnitude to gain insights into the variation of the solution structure with concentration. In fact, it is well-known that mixtures of the type polar compound + alkane, at temperatures in the vicinity of the critical one, are characterized by W-shaped \(C_{pm}^{E}\) curves, where non-random effects appear at intermediate compositions.
Regarding to nitrobenzene systems, a large database exists containing LLE measurements for alkane solutions. Their critical temperatures vary from 291.9 K for the heptane system, up to 309.7 K for the hexadecane mixture. These data have been used for the determination of the critical exponents. Special attention has been also paid to the dielectric behaviour of these systems near the critical point.
The main purpose of the present work is to get a deeper understanding of the interactions and structure of nitrobenzene + hydrocarbon mixtures and of 1-alkanol + 1-nitroalkane, or + nitrobenzene systems. At this end, a whole set of experimental data available in the literature, \(H_{m}^{E}\), \(C_{pm}^{E}\), excess molar volumes (\(V_{m}^{E}\)), vapour-liquid equilibria (VLE), LLE, permittivities (\(\varepsilon_{r}\)), or dynamic viscosities (\(\eta\)), are analyzed. Of particular interest is the investigation of the aromaticity effect, by means of the study of nitrobenzene solutions. Inaddition, the selected systems are also treated in the framework of the DISQUAC and ERAS models, and the results are compared with those obtained from UNIFAC (Dortmund version) using interaction parameters from the literature. The systems are also investigated using the concentration-concentration structure factor (\(S_{\rm CC}\)) formalism, based on the Bhatia-Thorton partial structure factors. The \(S_{\rm CC}\) formalism is concerned with the study of fluctuations in the number of molecules regardless of the components, the fluctuations in the mole fraction and the cross fluctuations, and arises from the generalization of the Bhatia-Thorton partial structure factors to link the asymptotic behaviour of the ordering potential to the interchange energy parameters in the semi-phenomenological theories of thermodynamic properties of liquid solutions. Thus, we continue our detailed programme concerned with the research of 1-alkanol \(+\) strong polar compound mixtures. Within this programme, we have studied mixtures involving, e.g., sulfolane, tertiary amides, or nitriles.
## 2 Models
### 2.1 Disquac
The group contribution model DISQUAC is based on the rigid lattice theory developed by Guggenheim. Some of its more relevant features are now briefly summarized. (i) The geometrical parameters of the mixture compounds, total molecular volumes, _r_i, surfaces, _q_i, and the molecular surface fractions, \(\alpha_{\rm si}\), are calculated additively on the basis of the group volumes \(R_{\rm G}\) and surfaces \(Q_{\rm G}\) recommended by Bondi. At this end, the volume \(R_{\rm CH4}\) and surface \(Q_{\rm CH4}\) of methane are taken arbitrarily as equal to 1. The geometrical parameters for the groups used in this work can be found elsewhere. (ii) The partition function is factorized into two terms. The excess functions are the result of two contributions: a dispersive (DIS) term arising from the contribution from the dispersive forces; and a quasichemical (QUAC) term which comes from the anisotropy of the field forces created by the solution molecules. For \(G_{\rm m}^{\rm E}\), a combinatorial term, \(G_{\rm m}^{\rm E,COMB}\), represented by the Flory-Huggins equation must be also included.
\[G_{\rm m}^{\rm E} = G_{\rm m}^{\rm E,DIS}+G_{\rm m}^{\rm E,QUAC}+G_{\rm m}^{\rm E, COMB} \tag{1}\] \[H_{\rm m}^{\rm E} = H_{\rm m}^{\rm E,DIS}+H_{\rm m}^{\rm E,OUAC} \tag{2}\]
(iii) The interaction parameters change with the molecular structure of the mixture components; (iv) The coordination number is assumed to be the same for all the polar contacts (\(z=4\)). This is a very important shortcoming of the model, and is partially removed via the hypothesis of considering structure dependent interaction parameters.
The equations used to calculate the DIS and QUAC contributions to \(G_{\mathrm{m}}^{\mathrm{E}}\)and \(H_{\mathrm{m}}^{\mathrm{E}}\) are given elsewhere. The temperature dependence of the interaction parameters is expressed in terms of the DIS and QUAC interchange coefficients, \(C_{\mathrm{s}\mathrm{t}\mathrm{J}}^{\mathrm{DIS}};C_{\mathrm{s}\mathrm{t} \mathrm{J}}^{\mathrm{QUAC}}\) where \(\mathrm{s}\neq\mathrm{t}\) are two contact surfaces present in the mixture and \(l=1\) (Gibbs energy; \(C_{\mathrm{s}\mathrm{t}\mathrm{,}1}^{\mathrm{DIS}\mathrm{QUAC}}=g_{\mathrm{s} \mathrm{t}}^{\mathrm{DIS}\mathrm{QUAC}}(T_{\mathrm{o}})/RT_{\mathrm{o}})\); \(l=2\) (enthalpy, \(C_{\mathrm{s}\mathrm{t}\mathrm{,}2}^{\mathrm{DIS}\mathrm{QUAC}}=h_{\mathrm{s} \mathrm{t}}^{\mathrm{DIS}\mathrm{QUAC}}(T_{\mathrm{o}})/RT_{\mathrm{o}})\)), \(l=3\) (heat capacity, \(C_{\mathrm{s}\mathrm{t}\mathrm{,}3}^{\mathrm{DIS}\mathrm{QUAC}}=c_{\mathrm{p} \mathrm{t}}^{\mathrm{DIS}\mathrm{QUAC}}(T_{\mathrm{o}})/R\)). \(T_{\mathrm{o}}=298.15\) K is the scaling temperature and \(R\), the gas constant. The equations can be found elsewhere.
As in previous applications, DISQUAC calculations on LLE were conducted taking into account that the values of the mole fraction \(x_{1}\) of component 1 (\(x_{1}^{\cdot},x_{1}^{\cdot}\)) relating to the two phases in equilibrium are such that the functions \(G_{\mathrm{m}}^{\mathrm{M}},G_{\mathrm{m}}^{M^{\ast}}\) (\(G_{\mathrm{m}}^{\mathrm{M}}=G_{\mathrm{m}}^{\mathrm{E}}+G_{\mathrm{m}}^{ \mathrm{ideal}}\) ) have a common tangent.
## 2.2 ERAS
Some important features of the model are the following. (i) The excess functions are calculated as the sum of two terms. One is linked to hydrogen-bonding effects (the chemical contribution, \(X_{\mathrm{m,chem}}^{\mathrm{E}}\)), and the other is related to non-polar van der Waals' interactions including free volume effects (physical contribution, \(X_{\mathrm{m,phys}}^{\mathrm{E}}\)). Equations for \(X_{\mathrm{m}}^{\mathrm{E}}=H_{\mathrm{m}}^{\mathrm{E}}\), \(V_{\mathrm{m}}^{\mathrm{E}}\) are given elsewhere. (ii) It is assumed that only consecutive linear association occurs.
\[A_{\mathrm{m}}+B\longleftarrow\stackrel{{ K_{\mathrm{AB}}}}{{ \longleftarrow}}A_{\mathrm{m}}B \tag{4}\]
The association constants (\(K_{\mathrm{AB}}\)) of equation are also assumed to be independent of the chain length. Equations and are characterized by \(\Delta h_{1}^{\ast}\), the enthalpy of the reaction that corresponds to the hydrogen-bonding energy, and by the volume change (\(\Delta v_{i}^{*}\)) related to the formation of the linear chains. (iii) The \(X_{\rm m,phys}^{\rm E}\) term is derived from the Flory's equation of state, which is assumed to be valid not only for pure compounds but also for the mixture.
\[\frac{\vec{P}\vec{V}_{\rm i}}{\vec{T}_{\rm i}}=\frac{\vec{V}_{\rm i}^{1/3}}{ \vec{V}_{\rm i}^{1/3}-1}-\frac{1}{\vec{V}_{\rm i}}\,\vec{T}_{\rm i} \tag{5}\]
In equation, \(\vec{V}_{\rm i}=V_{\rm mi}\,/\,V_{\rm mi}^{*}\); \(\vec{P}_{\rm i}=P\,/\,P_{\rm i}^{*}\);\(\vec{T}_{\rm i}=T\,/\,T_{\rm i}^{*}\) are the reduced volume, pressure and temperature respectively. The pure component reduction parameters\(V_{\rm mi}^{*}\), \(P_{\rm i}^{*}\), \(T_{\rm i}^{*}\) are determined from \(P\)-\(V\)-\(T\) data (density, \(\alpha_{\rm p}\), isobaric thermal expansion coefficient, and isothermal compressibility, \(\kappa_{T}\)), and association parameters. The reduction parameters for the mixture \(P_{\rm M}^{*}\) and \(T_{\rm M}^{*}\) are calculated from mixing rules. The total relative molecular volumes and surfaces of the compounds were calculated additively using the Bondi's method.
### Modified UNIFAC (Dortmund version)
This version of UNIFAC differs from the original UNIFAC model by the combinatorial term and the temperature dependence of the interaction parameters. The equations used to calculate \(G_{\rm m}^{\rm E}\) and \(H_{\rm m}^{\rm E}\) are obtained from the fundamental equation for the activity coefficient \(\gamma_{\rm i}\) of component i:
\[\ln\gamma_{\rm i}=\ln\gamma_{\rm i}^{\it COMB}+\ln\gamma_{\rm i}^{\it RES} \tag{6}\]
Equations are available elsewhere. In Dortmund UNIFAC, two main groups, OH and CH\({}_{3}\)OH, are defined for predicting thermodynamic properties of mixtures with alkanols. The main group OH is subdivided in three subgroups: OH(p), OH(s) and OH(t) for the representation of primary, secondary and tertiary alkanols, respectively. The CH\({}_{3}\)OH group is a specific group for methanol solutions. In the case of nitroalkanes and nitrobenzene two main groups exist. The main group CNO2 is divided in three subgroups: CH\({}_{3}\)NO\({}_{2}\) for nitromethane; CH\({}_{2}\)NO\({}_{2}\) for the remainder 1-nitroalkanes, and CHNO\({}_{2}\) for 2-nitroalkanes. There is also a main group, ACNO\({}_{2}\), for nitrobenzene. The subgroups within the same main group have different geometrical parameters, and identical group energy-interaction parameters. It is remarkable that the geometrical parameters, the relative van der Waals volumes and the relative van der Waalssurfaces are not calculated from molecular parameters like in the original UNIFAC, but fitted together with the interaction parameters to the experimental values of the thermodynamic properties considered. The geometrical and interaction parameters were taken from literature and used without modifications. No interaction parameters are available for the methanol + nitrobenzene system.
### 2.4 The concentration-concentration structure factor
Mixture structure can be investigated by means of the \(S_{\rm CC}\) function:
\[S_{\rm CC}=\frac{RT}{(\partial^{2}G^{\rm M}/\partial x_{1}^{2})_{P,T}}= \frac{x_{1}x_{2}}{D} \tag{7}\]
with
\[D=\frac{x_{1}x_{2}}{RT}(\partial^{2}G^{\rm M}/\partial x_{1}^{2})_{P,T}=1+\frac {x_{1}x_{2}}{RT}\left(\frac{\partial^{2}G_{\rm m}^{\rm E}}{\partial x_{1}^{2} }\right)_{P,T} \tag{8}\]
\(D\) is a function closely related to thermodynamic stability. For ideal mixtures, \(G_{\rm m}^{\rm E,id}=0\) (excess Gibbs energy of the ideal mixture); \(D^{\rm id}=1\) and \(S_{\rm CC}=x_{1}x_{2}\). From stability conditions,\(S_{\rm CC}>0\). If a system is close to phase separation, \(S_{\rm CC}\) must be large and positive (\(\infty\), if the mixture presents a miscibility gap). If compound formation between components exists, \(S_{\rm CC}\) must be very low (0, in the limit). Therefore, \(S_{\rm CC}>x_{1}x_{2}\) (\(D<1\)) indicates that the dominant trend in the system is the homocoordination (separation of the components). The mixture is then less stable than the ideal. If \(0<S_{\rm CC}<x_{1}x_{2}=S_{\rm CC}^{\rm id}\), \((D>1)\), the fluctuations in the system have been removed, and the main feature of the solution is compound formation (heterocoordination). The system is then more stable than ideal. In summary, \(S_{\rm CC}\) is an useful magnitude to evaluate the non-randomness in the mixture. In this work, we have used DISQUAC to evaluate \(S_{\rm CC}\) for a number of mixtures.
## 3.
### 3.1 DISQUAC interaction parameters
In terms of DISQUAC, the studied systems are regarded as possessing the following types of surfaces: (i) type a, aliphatic (CH\({}_{3}\), CH\({}_{2}\), in \(n\)-alkanes, or toluene, or 1-nitroalkane, or 1-alkanols); (ii) type r (NO\({}_{2}\) in 1-nitroalkanes or nitrobenzene); (iii) type s (s = b, C\({}_{6}\)H\({}_{6}\), or C\({}_{6}\)H\({}_{5}\) in benzene, toluene or nitrobenzene; s = c-CH\({}_{2}\) in cyclohexane; s = h, OH in 1-alkanols).
The general procedure applied in the estimation of the interaction parameters have been explained in detail in earlier works. Final values of our fitted parameters are listed in Tables 1 and 2. Some important remarks are provided below.
#### 3.1.1 Nitrobenzene + benzene
This system is only characterized by the (b,r) contact, which is assumed to be represented by DIS interaction parameters. Such choice is supported by the low experimental \(H_{\rm m}^{\rm E}\)values of this system (261 J\(\cdot\)mol-1 at equimolar composition and 293.15 K). The \(C_{\rm br,1}^{\rm DIS}\) coefficient was obtained from data on activity coefficients at infinite dilution. Final parameters are given in Table 1.
#### 3.1.2 Nitrobenzene + alkane, or + toluene
Mixtures with alkane are built by three contacts: (a,b), (b,r) and (a,r). The interaction parameters for the (a,b) contacts are dispersive and are known from the research of alkylbenzene + alkane systems. The interaction parameters of the (b,r) contacts are already known and thus only those corresponding to the (a,r) contacts must be determined (Table 1). As in other many applications, we have used \(C_{\rm ar,1}^{\rm QUAC}\)=\(C_{\rm cr,1}^{\rm QUAC}\) (\(l\) = 1,2,3). That is, the QUAC coefficients for the (a,r) and (c,r) contacts are independent of the alkane. In addition, the \(C_{\rm ar,1}^{\rm DIS}\) coefficient is assumed to be dependent on the chain length of the \(n\)-alkane in order to get improved results for the coordinates of the critical points (see below).
The system involving toluene is characterized by the same contacts. Here, we have held the interaction parameters for the (a,b) and (a,r) contacts and determined newly those for the (b,r) contact (Table 1), assuming that it is dispersive.
#### 3.1.3 1-Alkanol + 1-nitroalkane
The contacts present in these solutions are: (a,h), (a,r) and (h,r). The interaction parameters for the (a,h) contacts are described by DIS and QUAC interaction parameters, previously determined from the study of 1-alkanol + \(n\)-alkane systems, and the \(C_{\rm ar,1}^{\rm DIS,QUAC}\)(l = 1,2,3) coefficients are also known from the corresponding treatment of 1-nitroalkane + alkane systems. Therefore, only the interaction parameters for the (h,r) have to be obtained (Table 2). Detailed calculations showed that the QUAC coefficients could be assumed to be independent of the 1-alkanol (from 1-propanol). Consequently, these parameters were fixed and the DIS ones determined. We remark that, for a similar reason to that given for nitrobenzene + \(n\)-alkane systems, the \(C_{\rm br,1}^{\rm DIS}\) coefficient is considered to be dependent on the alkanol size.
#### 3.1.4 1-Alkanols + nitrobenzene
We have now six contacts: (a,b), (a,h), (a,r), (b,h), (b,r) and (h,r). The contacts (b,h) are represented by DIS and QUAC interaction parameters, which are already known from the study of 1-alkanol + toluene mixtures. Thus, only the interaction parameters for the (h,r)contacts have to be determined (Table 2) as those for the remainder contacts are known. The procedure is similar to that just explained.
### Adjustment of ERAS parameters
Values of \(V_{\mathrm{mi}},V_{\mathrm{mi}}^{{}^{\ast}}\) and \(P_{\mathrm{i}}^{{}^{\ast}}\) of pure compounds at \(T=298.15\) K, needed for calculations, have been taken from the literature in the case of 1-alkanols, and are listed in Table S1 of supplementary material for 1-nitroalkanes or nitrobenzene. For the 1-alkanols, \(K_{\mathrm{Av}}\), \(\Delta h_{\mathrm{A}}^{{}^{\ast}}\) (=\(-\) 25.1 kJ*mol*1) and \(\Delta v_{\mathrm{A}}^{{}^{\ast}}\) (=\(-\) 5.6 cm3*mol*1) are known from \(H_{\mathrm{m}}^{\mathrm{E}}\) and \(V_{\mathrm{m}}^{\mathrm{E}}\) data for the corresponding mixtures with alkanes. These values have been used in many other applications. The binary parameters to be fitted against \(H_{\mathrm{m}}^{\mathrm{E}}\) and \(V_{\mathrm{m}}^{\mathrm{E}}\) data available in the literature for 1-alkanol + 1-nitroalkane or + nitrobenzene systems are then \(K_{\mathrm{AB}},\Delta h_{\mathrm{AB}}^{{}^{\ast}}\), \(\Delta v_{\mathrm{AB}}^{{}^{\ast}}\) and \(X_{\mathrm{AB}}\). They are collected in Table 3.
## 4 Theoretical results
Results from DISQUAC on phase equilibria, \(H_{\mathrm{m}}^{\mathrm{E}}\) and \(C_{\mathrm{pm}}^{\mathrm{E}}\) are shown in Tables 4-8 and in Figures 1-7 (see also Figure S1 of supplementary material).
\[\sigma_{\mathrm{r}}(P)=\{\frac{1}{N}\sum\left[\frac{P_{\mathrm{exp}}-P_{ \mathrm{calc}}}{P_{\mathrm{exp}}}\right]^{2}\}^{1/2} \tag{9}\]
\[dev(H_{\mathrm{m}}^{\mathrm{E}})=\{\frac{1}{N}\sum\left[\frac{H_{\mathrm{m,exp }}^{\mathrm{E}}-H_{\mathrm{m,calc}}^{\mathrm{E}}}{H_{\mathrm{m,exp}}^{\mathrm{ E}}(x_{1}=0.5)}\right]^{2}\}^{1/2} \tag{10}\]
ERAS results on \(H_{\mathrm{m}}^{\mathrm{E}}\) for 1-alkanol + nitromethane or + nitrobenzene systems are shown in Table 7 (Figures 5 and 6). Some ERAS calculations on \(V_{\mathrm{m}}^{\mathrm{E}}\) are collected in Table S2 (see Figure S2 of supplementary material). Figure S1 compares, as an example, ERAS results with experimental \(G_{\mathrm{m}}^{\mathrm{E}}\) values for the methanol + nitromethane mixture. Results from the application of the UNIFAC model on VLE, \(H_{\mathrm{m}}^{\mathrm{E}}\)and \(C_{\mathrm{pm}}^{\mathrm{E}}\) are collected in Tables 4, 7 and 8.
## 5 Discussion
Hereafter, we are referring to values of the thermodynamic excess functions at 298.15 K and equimolar composition. The number of C atoms in 1-alkanols and in 1-nitroalkanes are represented by \(n_{\text{OH}}\)and\(n_{\text{NO2}}\), respectively. The impact of polarity on bulk properties can be examined through the effective dipole moment, \(\overline{\mu}\), defined by:
\[\overline{\mu}=\left[\frac{\mu^{2}N_{A}}{4\pi\varepsilon_{0}V_{m}k_{B}T}\right] ^{1/2} \tag{11}\]
Here, \(\mu\) is the dipole moment; \(N_{A}\) the Avogadro's number, \(\varepsilon_{0}\) the permittivity of the vacuum, \(V_{m}\) the molar volume, and \(k_{B}\) the Boltzmann's constant. Values of \(\overline{\mu}\) for 1-nitroalkanes and nitrobenzene are collected in Table 9. As for a pure polar liquid, the potential energy related to dipole-dipole interactions is, in first approximation, proportional to (\(-\)\(\overline{\mu}^{4}/r^{6}\)) or more roughly to (\(-\)\(\overline{\mu}^{4}/V_{m}^{2}\)) (\(r\) is the distance between dipoles), dipolar interactions between nitroalkane molecules decrease in the order: nitromethane (\(\overline{\mu}=1.855\)) \(>\) nitroethane (\(\overline{\mu}=1.625\)) \(>\) 1-nitropropane (\(\overline{\mu}=1.453\)). For nitrobenzene, \(\overline{\mu}=1.510\). Dipolar interactions between 1-alkanol molecules are much weaker and also decrease with the increasing of molecular size: \(\overline{\mu}=1.023\) (\(n_{\text{OH}}=1\)) \(>\) 0.852 (\(n_{\text{OH}}=2\)) \(>\) 0.752 (\(n_{\text{OH}}=3\)) \(>\) 0.664 (\(n_{\text{OH}}=4\)) \(>\) 0.580 (\(n_{\text{OH}}=6\)).
## 5.1 Nitrobenzene + hydrocarbon mixtures
Firstly, we must remark that nitromethane or nitroethane + alkane systems show LLE curves characterized by UCSTs, which in the case of nitromethane solutions are very high: 387.2 K and 398.1 K for the systems with octane and decane, respectively. The critical temperatures of nitroethane + octane (314.5 K), or + decane (325.8 K) are lower and \(H_{m}^{\text{E}}/\text{J}\cdot\text{mol}^{-1}\) values of heptane mixtures decrease when \(n_{\text{NO2}}\) is increased: 1593 (\(n_{\text{NO2}}=3\)); 1390 (\(n_{\text{NO2}}=4\)); 1220 (\(n_{\text{NO2}}=5\)). Similarly, \(H_{m}^{\text{E}}(\text{C}_{6}\text{H}_{12})/\text{J}\cdot\text{mol}^{-1}=1690\) (\(n_{\text{NO2}}=2\)); 1549 (\(n_{\text{NO2}}=3\)); 1370 (\(n_{\text{NO2}}=4\)); 1225 (\(n_{\text{NO2}}=5\)). These experimental results are in agreement with the existence of strong dipolar interactions between nitroalkane molecules and reveal that the mentioned interactions become weaker when \(n_{\text{NO2}}\)increases, accordingly with the relative variation of \(\overline{\mu}\). Regarding nitrobenzene + alkane systems, UCST/K = 293.01 K (octane); 295.96 K (decane) and \(H_{m}^{\text{E}}(\text{cyclohexane}\), \(T=293.15\) K) = 1654 J\(\cdot\)mol\({}^{-1}\); 1447 (hexane). Such set of data suggests that dipolar interactions between nitrobenzene molecules are stronger than those between 1-nitropropane molecules, even when the distance between nitrobenzene molecules is larger (Table 9). This can be ascribed to the existence of intramolecular effectsbetween the phenyl and the NO\({}_{2}\) groups of nitrobenzene which lead to enhanced dipolar interactions. Intermolecular effects between the phenyl ring and the polar group of a given aromatic polar molecule is encountered in many systems. Thus, 1-alkanol (from ethanol) or 1-alkylamine (from ethylamine) + heptane mixtures are miscible at any composition at 298.15 K, while the UCST of the corresponding solutions with phenol or aniline are 327.3 K and 343.1 K, respectively. Similarly, UCST(decane)/K = 266.8 (2-propanone), 277.4 (acetophenone), or UCST(dodecane)/K = 284.7 (butanenitrile); 293.1 (benzonitrile).
It is interesting to conduct a short comparison between mixtures containing 1-nitroalkanes and alkanes or benzene. For example, \(H_{\mathrm{m}}^{\mathrm{E}}\) (benzene)/J*mol-1 = 790 (\(n_{\mathrm{NO2}}\) = 1); 63 (\(n_{\mathrm{NO2}}\) = 3); and 261 (nitrobenzene, _T_/K = 293.15). That is, when the \(n\)-alkane is replaced by benzene, mixtures become miscible and \(H_{\mathrm{m}}^{\mathrm{E}}\) values decrease, which can be ascribed to the new interactions between unlike molecules created upon mixing (intermolecular effects). This is a rather general trend, as the following values reveal: \(H_{\mathrm{m}}^{\mathrm{E}}\)(C\({}_{6}\)H\({}_{6}\))/ J*mol-1 = 751 (aniline); 860 (phenol, \(T\) = 313.15 K); 125 (acetophenone); 138 (2-propanone); - 66 (butanenitrile); 32 (benzonitrile). 1-Alkanols are a remarkable exception and \(H_{\mathrm{m}}^{\mathrm{E}}\)(1-alkanol (\(\neq\) methanol) + heptane) \(>H_{\mathrm{m}}^{\mathrm{E}}\)(1-alkanol + benzene). Thus for 1-hexanol mixtures, \(H_{\mathrm{m}}^{\mathrm{E}}\)/J*mol-1 = 527 (heptane); 1141 (benzene). Aromatic hydrocarbons are better breakers of the alcohol self-association than \(n\)-alkanes.
## 5.2 1-alkanol + 1-nitroalkane
We start remarking that the existence, for these solutions, of LLE curves with relatively high UCST values (see Introduction Section) show that dipolar interactions are very important. For the sake of comparison, we provide UCST values of hexane mixtures with tertiary amides together with dipole moments of these very polar substances. Thus, UCST/K = 337.7 (_N,N_-dimethylformamide; \(\mu\) = 3.68 D); 305.3 (_N,N_-dimethylacetamide; \(\mu\) = 3.71 D); 324.6 (_N_-methylpyrrolidone; \(\mu\) = 4.09 D. For the 1-hexanol + nitromethane system, UCST = 308.7 K. This value is much lower than the corresponding result for the nitromethane + hexane mixture (375.4 K). That is, the replacement of an \(n\)-alkane by an isomeric 1-alkanol leads to a decreased UCST value, which reveals that 1-alkanols are better breakers of the dipolar interactions between nitroalkane molecules. The existence of alkanol-nitroalkane interactions is supported by the fact that systems with \(n_{\mathrm{OH}}\) = 1-4 and \(n_{\mathrm{NO2}}\) = 1 are miscible at 298.15 K and any concentration (see below).
Next, we are going to evaluate the enthalpy of the H-bonds between 1-alkanols and nitroalkanes (termed as\(\Delta H_{\mathrm{OH-NO2}}\)). Neglecting structural effects, \(H_{\mathrm{m}}^{\mathrm{E}}\) can be considered as the result of three contributions. The positive ones, \(\Delta H_{\text{OH-OH}},\Delta H_{\text{NO2-NO2}}\), come, respectively, from the breaking of alkanol-alkanol and nitroalkane-nitroalkane interactions along the mixing process. The negative contribution, \(\Delta H_{\text{OH-NO2}}\), is due to the new OH--NO2 interactions created upon mixing. That is:
\[H_{\text{m}}^{\text{E}}=\Delta H_{\text{OH-OH}}+\Delta H_{\text{NO2-NO2}}+\Delta H _{\text{OH-NO2}} \tag{12}\]
An evaluation of \(\Delta H_{\text{OH-NO2}}\) can be conducted extending the equation to \(x_{\text{i}}\to 0\). Then, \(\Delta H_{\text{OH-OH}}\) and \(\Delta H_{\text{NO2-NO2}}\) can be replaced by \(H_{\text{m1}}^{\text{E},\infty}\) (partial excess molar enthalpy at infinite dilution of the first component) of 1-alkanol or nitroalkane + heptane systems.
\[\Delta H_{\text{OH-NO2}}=H_{\text{m1}}^{\text{E},\infty}(1-\text{alkanol}+\text {nitroalkane})\]
\[-H_{\text{m1}}^{\text{E},\infty}(1-\text{alkanol}+\text{heptane})-H_{\text{m1}}^{\text{E},\infty}(\text{nitroalkane}+\text{heptane}) \tag{13}\]
There are some shortcomings for this estimation of \(\Delta H_{\text{OH-NO2}}\) values. (i) Some \(H_{\text{m1}}^{\text{E},\infty}\) data used were calculated from \(H_{\text{m}}^{\text{E}}\) measurements over the entire mole fraction range. (ii) For 1-alkanol + _n_-alkane systems, it was assumed that \(H_{\text{m1}}^{\text{E},\infty}\) is independent of the alcohol, a common approach when applying association theories. We have used in this work, as in previous applications, \(H_{\text{m1}}^{\text{E},\infty}=23.2\) kJ'mol\({}^{-1}\). Nevertheless, it should be remarked that the values of \(\Delta H_{\text{OH-NO2}}\)collected in Table 10 are still meaningful as they were obtained following the same procedure that in other previous investigations, which allows to compare enthalpies of interaction between 1-alkanols and different organic solvents. Inspection of Table 10 shows that \(\Delta H_{\text{OH-NO2}}\) increases more or less smoothly with \(n_{\text{OH}}\), that is, interactions between unlike molecules become weaker, which may be ascribed to the OH group is more sterically hindered in longer 1-alkanols. On the other hand, the increased UCST values for nitromethane mixtures with longer 1-alkanols (\(n_{\text{OH}}\geq 6\)) suggests a sharp decrease of the number of interactions between unlike molecules, while interactions between like molecules become strongly dominant.
#### 5.2.1 _Molar excess enthalpies and entropies_
The 1-alkanol + nitromethane mixtures are characterized by: (i) large and positive \(H_{\rm m}^{\rm E}\) values (\(H_{\rm m}^{\rm E}/J\cdot mol^{-1}=1265\) (\(n_{\rm OH}=1\)); 1633 (\(n_{\rm OH}=2\)); 1911 (\(n_{\rm OH}=3\)); 2131 (\(n_{\rm OH}=4\)); 2781 (\(n_{\rm OH}=6\)), \(T=313.15\) K). (ii) Symmetrical \(H_{\rm m}^{\rm E}\) curves, which, at the middle of the concentration range, are more or less flattened as consequence of the proximity of the UCST. (iii) Positive \(TS_{\rm m}^{\rm E}\) (\(=H_{\rm m}^{\rm E}-G_{\rm m}^{\rm E}\)) values: 228 (\(n_{\rm OH}=1\)); 463 (\(n_{\rm OH}=2\)); 1459 (\(n_{\rm OH}=6\), \(T=313.15\) K) (values calculated using \(G_{\rm m}^{\rm E}/J\cdot mol^{-1}=1040\) (\(n_{\rm OH}=1\)); 1170 (\(n_{\rm OH}=2\)); 1322 (\(n_{\rm OH}=6\), DISQUAC value at \(T=313.15\) K)). These features support our previous conclusion on the relevance of dipolar interactions in the present systems. Note that several typical properties of 1-alkanol + alkane mixtures are: (i) low \(H_{\rm m}^{\rm E}\) values, as alkanes are poor breakers of the alkanols self-association; (ii) for the same reason, the \(H_{\rm m}^{\rm E}\) curves are skewed to low alcohol mole fractions; (iii) very negative \(TS_{\rm m}^{\rm E}\) values. For example, for the 1-propanol + hexane system, \(G_{\rm m}^{\rm E}=1295\), \(H_{\rm m}^{\rm E}=533\) and \(TS_{\rm m}^{\rm E}=-762\) (all values in J\(\cdot mol^{-1}\)).
The \(H_{\rm m}^{\rm E}(n_{\rm OH}\)) variation of 1-alkanol + nitromethane systems can be explained as follows. (i) Interactions between unlike molecules are weakened (lower \(\left|\Delta H_{\rm OH\cdot NO2}\right|\)values) when \(n_{\rm OH}\) increases. (ii) Dipolar interactions between like molecules become more important at this condition, as the \(TS_{\rm m}^{\rm E}(n_{\rm OH}\)) variation shows. Note the very high result for \(TS_{\rm m}^{\rm E}\) of the 1-hexanol solution at 313.15 K, a temperature only 4.45 K higher than the UCST. Dipolar interactions are weakened when nitromethane is replaced by nitroethane, as the UCST values of the corresponding 1-nonanol mixtures reveal: (328 K (\(n_{\rm NO2}=\)1); 283.6 K (\(n_{\rm NO2}=2\)).
#### 5.2.3 _Molar excess volumes_
It is well stated that \(V_{\rm m}^{\rm E}\) is the result of different contributions. Those which are positive arise from the breaking of interactions between like molecules. Interactions between unlike molecules and structural effects (changes in free volume, differences in size and shape between the system components, interstitial accommodation) contribute negatively to\(V_{\rm m}^{\rm E}\). Thus, the positive \(V_{\rm m}^{\rm E}\)/(cm\({}^{3}\cdot mol^{-1}\)) values of the systems 1-nitropropane + C\({}_{6}\)H\({}_{12}\) (0.690) or 1-propanol + heptane (0.271) indicate that the main contribution to \(V_{\rm m}^{\rm E}\) comes from the disruption of the dipolar interactions between 1-nitropropane molecules and from the breaking of the alcohol self-association, respectively. The lower \(V_{\rm m}^{\rm E}\)/(cm\({}^{3}\cdot mol^{-1}\)) value of 1-propanolnitronethane mixture (0.236) newly points out to the existence of interactions between unlike molecules in this kind of systems, which are also characterized by strong structural effects. In fact, \(H_{\rm m}^{\rm E}\)and \(V_{\rm m}^{\rm E}\)values of some solutions are of opposite sign (e.g., \(V_{\rm m}^{\rm E}\)(methanol + nitromethane) = - 0.152 cm\({}^{3}\cdot\)mol\({}^{\text{-1}}\)). This is a typical feature of mixtures where strong structural effects exist. In addition, the positive \(V_{\rm m}^{\rm E}\)values encountered for these solutions (0.342 cm\({}^{3}\cdot\)mol\({}^{\text{-1}}\) for 1-butanol + nitromethane, see also above) are rather low, if one takes into account the very large \(H_{\rm m}^{\rm E}\)values involved. On the other hand, for a given 1-nitroalkane, both \(H_{\rm m}^{\rm E}\)and \(V_{\rm m}^{\rm E}\)increase in line with \(n_{\rm OH}\). That is, the \(V_{\rm m}^{\rm E}\)(\(n_{\rm OH}\))variation is closely related to that of the corresponding interactional contribution to this excess function. Interestingly,\(V_{\rm m}^{\rm E}\)is lower for the methanol + nitromethane system (- 0.168 cm\({}^{3}\cdot\)mol\({}^{\text{-1}}\) than for the corresponding mixture with nitroethane (- 0.141 cm\({}^{3}\cdot\)mol\({}^{\text{-1}}\), values at 293.15 K). An inversion of this behaviour is observed for solutions with \(n_{\rm OH}\geq 2\). Thus, \(V_{\rm m}^{\rm E}\)(1-propanol)/cm\({}^{3}\cdot\)mol\({}^{\text{-1}}\): 0.213, (\(n_{\rm NO2}=1\)); 0.111 (\(n_{\rm NO2}=2\)). This suggests that effects linked to interactions between unlike molecules are more important in the methanol + nitromethane system, and that those related to the breaking of dipolar interactions between nitroalkane molecules are predominant in systems with \(n_{\rm OH}\geq 2\) and \(n_{\rm NO2}=1\). At the latter conditions, effects due to alcohol self-association become more relevant in solutions with \(n_{\rm NO2}=2\). Accordingly with this interpretation, the \(V_{\rm m}^{\rm E}\)curve of the 1-propanol + nitroethane systems is skewed to lower concentrations in the alcohol, while the corresponding curve for 1-propanol + nitromethane is nearly symmetrical.
#### 5.2.4 _Molar excess heat capacities at constant pressure_
It is important to keep in mind that \(C_{\rm p,m}^{\rm E}\)values of 1-alkanol + alkane mixtures are high and positive (11.7 J\(\cdot\)mol\({}^{\text{-1}}\cdot\)K\({}^{\text{-1}}\)for ethanol + heptane). In addition, \(C_{\rm p,m}^{\rm E}\)increases with the temperature and, at enough high values of this magnitude, decreases. For example, in the case of the 1-decanol + decane mixture, \(C_{\rm p,m}^{\rm E}\)(\(x_{\rm i}\)= 0.3925)/J\(\cdot\)mol\({}^{\text{-1}}\cdot\)K\({}^{\text{-1}}\)= 14.81 (298.15 K); 25.04 (348.15 K); 24.31 (368.15 K). Mixtures characterized by strong dipolar interactions show low positive \(C_{\rm p,m}^{\rm E}\)values. Thus, \(C_{\rm p,m}^{\rm E}\)/J\(\cdot\)mol\({}^{\text{-1}}\cdot\)K\({}^{\text{-1}}\)= 4.4 (1-propanol + 2,5,8-trioxanonane); 0.96 (ethanol + DMF). Therefore, the \(C_{\rm p,m}^{\rm E}\)value of the methanol + nitromethane mixture (9.3 J\(\cdot\)mol\({}^{\text{-1}}\cdot\)K\({}^{\text{-1}}\)) reveals that association/solvation effects are still important in this solution. The available \(C_{\rm p,m}^{\rm E}\)data for 1-propanol or 1-butanol + nitromethane mixtures are very large (16.6 and 20.6 J\(\cdot\)mol\({}^{\text{-1}}\cdot\)K\({}^{\text{-1}}\), respectively), and decrease when the temperature is increased (Table 8). These are typical features encountered in systems at temperatures not far from their UCST. In this framework, the observed \(C_{\rm p,m}^{\rm E}\)(_n_OH ) variation at 298.15 K in nitromethane solutions (Table 8) can be ascribed to an increase of non-random effects related to the proximity of the critical temperature. The \(C_{\rm p,m}^{\rm E}\) result of the ethanol + 1-nitropropane mixture (14.9 J*mol-1*K-1) seems to be higher than the corresponding value of the nitroethane system, which can be explained considering that dipolar interactions are much less relevant in 1-nitropropane mixtures where effects related to alcohol self-association become more relevant.
## 5.3 1-alkanol + nitrobenzene
The \(H_{\rm m}^{\rm E}\) values of these systems are also large and positive: \(H_{\rm m}^{\rm E}/\)J*mol-1 = 1109 (_n_OH = 1); 1430 (_n_OH = 2); 1946 (_n_OH = 4). Consequently, the main contribution to \(H_{\rm m}^{\rm E}\) arises from the breaking of interactions between like molecules. The \(H_{\rm m}^{\rm E}\)(_n_OH ) variation can be explained as above. i.e., in terms of a weakening of the alkanol-nitrobenzene interactions produced when _n_OH increases (Table 10), together with the corresponding increase of dipolar interactions between like molecules. The excess molar volumes are negative: - 0.191 (_n_OH = 3), - 0.117 (_n_OH = 4), - 0.075 (_n_OH = 5) (_T_ = 303.15 K), i.e., the contribution to \(V_{\rm m}^{\rm E}\) from structural effects is here dominant.
On the other hand, comparison of \(\Delta H_{\rm OH\mbox{-}N\mbox{\scriptsize 2}}\) values for systems with nitromethane or nitrobenzene (Table 10) suggests that interactions between unlike molecules are weaker in the latter solutions. However, the comparison between \(H_{\rm m}^{\rm E}\)values pertaining to different homologous series should be conducted with caution in order to state reliable conclusions. It is known that \(H_{\rm m}^{\rm E}\)is not only determined by interactional effects, but also by structural effects. In fact, it is more appropriated compare \(U_{\rm vm}^{\rm E}\)results (isochoric molar excess internal energy), which can be obtained from:
\[U_{\rm vm}^{\rm E}=H_{\rm m}^{\rm E}-\frac{\alpha_{\rm p}}{\kappa_{\rm T}}TV_{ \rm m}^{\rm E} \tag{14}\]
The determination of the eos contribution needs of accurate volumetric data. Here, the very different \(V_{\rm m}^{\rm E}\)values of 1-alkanol + nitromethane, or + nitrobenzene systems (see above)suggest that the eos term may be decisive when comparing \(H_{\rm m}^{\rm E}\)values pertaining to these series. We have roughly determined \(U_{\rm Vm}^{\rm E}\)/J\(\cdot\)mol\({}^{-1}\) values for 1-propanol + nitromethane, or + nitrobenzene systems, and for 1-butanol + nitromethane, or + nitrobenzene mixtures, assuming the ideal behavior to calculate \(\alpha_{\rm p}\) and \(\kappa_{T}\)values of the mixtures. The rather similar results obtained for solutions with a given 1-alkanol point out that the different \(H_{\rm m}^{\rm E}\) values are largely due to the different structural effects in the considered systems.
## 5.4_ _Dielectric constants and Kirkwood's correlation factor
Here, we analyze the permittivity \(\rm data,\varepsilon_{r}\), for 1-alkanol + nitromethane, or + nitrobenzene mixtures. The excess permittivities, \(\varepsilon_{\rm r}^{\rm E}\), are defined as:
\[\varepsilon_{\rm r}^{\rm E}=\varepsilon_{\rm r}-\phi_{\rm I}\varepsilon_{\rm r 1}-\phi_{\rm I}\varepsilon_{\rm r 2} \tag{16}\]
The \(\varepsilon_{\rm r}^{\rm E}\) values referred to below are, except when indicated, at 298.15 K and \(\phi_{\rm i}=0.5\). For the present systems \(\varepsilon_{\rm r}^{\rm E}\) values are very negative. Thus, \(\varepsilon_{\rm r}^{\rm E}\)(nitromethane; 293.15 K) = - 1.74 (\(n_{\rm OH}=1\)); \(-\) 2.55 (\(n_{\rm OH}=2\)); \(-\) 3.20 (\(n_{\rm OH}=3\)); and \(\varepsilon_{\rm r}^{\rm E}\)(nitrobenzene) = - 2.46 (\(n_{\rm OH}=2\)); \(-\) 4.01 (\(n_{\rm OH}=3\)); \(-\) 4.23 (\(n_{\rm OH}=4\)); \(-\) 2.67 (\(n_{\rm OH}=5\)); \(-\) 0.90 (\(n_{\rm OH}=7\)) (\(T=293.15\) K). These negative experimental results indicate that the predominant trend in the solutions is the breaking of the alcohol self-association, as well as the disruption of the dipolar interactions between nitroalkane molecules, as such effects lead to a decrease of the dipolar polarization of the mixture. The contribution to \(\varepsilon_{\rm r}^{\rm E}\) due to interactions between unlike molecules may be either positive or negative, and that depends on the chemical nature of the mixture compounds and of the size and shape of the multimers formed upon mixing by the two components. Negative contributions are encountered when the mentioned multimers form, eg., cyclic structures and have less effective dipole moments than those of the multimers built by the pure components. Interestingly, the \(\varepsilon_{\rm r}^{\rm E}\)(\(\phi_{\rm i}=0.5\)) value of the nitrobenzene + benzene mixture is also very negative (\(-\) 4.88). It is clear that benzene is a good breaker of the dipolar interactions between nitrobenzene molecules. The variation of \(\varepsilon_{\rm r}^{\rm E}\)(nitrobenzene) with \(n_{\rm OH}\) seems to be a rather general trend, as many mixtures behave similarly. For example, \(\varepsilon_{\rm r}^{\rm E}\)(hexane)= - 1.12 (\(n_{\rm OH}=4\)); \(-\) 2.43 (\(n_{\rm OH}=5\)); \(-\) 2.52 (\(n_{\rm OH}=7\)); \(-\) 1.62 (\(n_{\rm OH}=10\)); or \(\varepsilon_{\rm r}^{\rm E}\)(cyclohexylamine) = 2.22 (\(n_{\rm OH}=1\))\(-\) 0.27 (\(n_{\rm OH}=3\)); \(-\) 0.85 (\(n_{\rm OH}=4\)); \(-\) 0.91 (\(n_{\rm OH}=7\)): \(-\) 0.41 (\(n_{\rm OH}=10\)). This behaviour has been explained in terms of the lower and weaker self-association of longer 1-alkanols.
We have determined the Kirkwood's correlation factor, \(g_{\rm K}\), of systems containing 1-alkanols by means of the equation:
\[g_{\rm K}=\frac{9k_{\rm B}TV_{\rm m}\varepsilon_{0}(\varepsilon_{\rm r}- \varepsilon_{\rm r}^{\infty})(2\varepsilon_{\rm r}+\varepsilon_{\rm r}^{\infty })}{N_{\rm A}\mu^{2}\varepsilon_{\rm r}\left(\varepsilon_{\rm r}^{\infty}+2 \right)^{2}} \tag{17}\]
Details of the calculation procedure have been given previously. The equation needed for the determination of \(g_{\rm K}\) for the nitrobenzene + benzene system is different as benzene is a non-polar compound and can be found elsewhere. Physical properties of pure compounds and density data needed for calculations were taken from the literature. In absence of experimental measurements on refractive indices, this magnitude was considered as ideal. For 1-alkanol + nitrobenzene systems, results plotted in show that the mixture structure increases smoothly over a rather wide concentration range, and that this increase becomes even softer when \(n_{\rm OH}\)increases. Thus, the addition of 1-alkanol to nitroalkane leads to cooperative effects, which increase the total effective dipole moment of the solution, and that partially compensate the destruction of the existing structure in pure liquids. This effect becomes less relevant when longer 1-alkanols are involved. The same behaviour is encountered for 1-alkanol + nitromethane mixtures. Here, it is important to pay attention to the nitrobenzene + benzene system.
### Dynamic viscosities
The discussion is now conducted in terms of deviations of viscosity from the linear behaviour, defined as
\[\Delta\eta=\eta-x_{\rm I}\eta_{\rm I}+x_{\rm Z}\eta_{\rm Z} \tag{18}\]
Results given below are at equimolar composition. Firstly, we must remark that \(\Delta\eta\) values of the studied systems are negative. At 303.15 K, \(\Delta\eta\left(n_{\rm NO2}=1\right)\)/mPa* \(=\)\(-\)\(0.084\) (\(n_{\rm OH}=\)1); \(-\)\(0.211\) (\(n_{\rm OH}=\)2); \(-\)\(0.408\) (\(n_{\rm OH}=\)3); \(\Delta\eta\left(n_{\rm NO2}=\)2)/mPa* \(=\)\(-\)\(0.066\) (\(n_{\rm OH}=\)1); \(-\)\(0.193\) (\(n_{\rm OH}=\)2); \(-\)\(0.407\) (\(n_{\rm OH}=\)3); and \(\Delta\eta\) (nitrobenzene)/mPa* \(=\)\(-\)\(0.043\) (\(n_{\rm OH}=\)1); \(-\)\(0.172\) (\(n_{\rm OH}=\)2); \(-\)\(0.342\) (\(n_{\rm OH}=\)3); \(-\)\(0.494\) (\(n_{\rm OH}=\)4). These results can be explained in terms of a higher fluidization of the solution due to the breaking of alcohol self-association and of dipolar interactions between nitroalkane molecules. As viscosity is strongly dependent on the size and shape of the mixture components, we compare now the results of nitrobenzene solutions with those of 1-alkanol + toluene systems at 303.15 K. The latter mixtures show lower values. Thus, \(\Delta\eta\)/mPa\(\cdot\)s =\(-\)0.056 (\(n_{\rm OH}\) = 1); \(-\)0.228 (\(n_{\rm OH}\) = 2); \(-\)0.381 (\(n_{\rm OH}\) = 3); \(-\)0.557 (\(n_{\rm OH}\) = 4). From this comparison, it is possible to conclude that alkanol-nitrobenzene interactions lead to a lower fluidization of the corresponding mixture (higher \(\Delta\eta\) values). Of course, interactions between unlike molecules are also present in systems with nitromethane or nitroethane. This can be demonstrated by comparing (\(\Delta\eta\)/mPa\(\cdot\)s) data for mixtures with \(n_{\rm NO2}\) = 2 at 293.15 K (\(-\)0.239 (\(n_{\rm OH}\) = 2); \(-\)0.540 (\(n_{\rm OH}\) = 3)) with results for 1-alkanol + pentane systems at 298.15 K (\(-\)0.264 (\(n_{\rm OH}\) = 2); \(-\)0.599 (\(n_{\rm OH}\) = 3)). Positive contributions to \(\Delta\eta\), related to interactions between molecules, are also encountered in many others systems, as 1-alkanol + cyclohexylamine. In fact, if interactions between unlike molecules are enough important, \(\Delta\eta\) values may be positive. This is the case of 1-propanol, or 1-butanol + cyclohexylamine mixtures at 303.15 K (0.344 and 0.181 mPa\(\cdot\)s, respectively). The effect of the replacement of nitromethane by nitroethene in systems with a given 1-alkanol is not clear. It seems that \(\Delta\eta\) values are slightly lower for methanol or ethanol + nitromethane mixtures (see above). This might be due to this nitroalkane is a better breaker of the alcohol self-association. More experimental work is needed in this field. Finally, we note that \(\Delta\eta\)(nitroalkane) decreases when \(n_{\rm OH}\) increases, which is the same behaviour observed for 1-alkanol + alkane, or + toluene systems. Therefore, it can be ascribed not only to a weakening of the interactions between unlike molecules caused by the OH group become more sterically hindered when \(n_{\rm OH}\) increases, but also to the decreasing of the alcohol self-association at this condition.
## 5.6 \(S_{CC}\) results.
Inspection of Figures 9a and 9b allows state some interesting remarks. (i) We note that, for the studied systems, \(S_{CC}\) > 0.25. This means that homocoordination (i.e., interactions between like molecules) is the main feature for such mixtures. (ii) In the case of the nitrobenzene + heptane system, the \(S_{CC}\) curve shows a large maximum at 298.15 K, a temperature close to the corresponding UCST (291.9 K). As already mentioned, this is a typical behaviour shown by systems at temperatures in vicinity of the UCST, which are characterized by a strong homocoordination. (iii) The replacement of heptane by methanol leads to a large \(S_{CC}\) decrease. Homocoordination becomes weaker due to the new methanol-nitrobenzene interactions created upon mixing. This newly demonstrates the existence of interactions between unlike molecules in 1-alkanol + nitroalkane mixtures. (iv) For nitronethane solutions, the maximum of the \(S_{CC}\) curves increases in the sequence: methanol \(>\) ethanol \(>\) 1-propanol \(>\) 1-butanol. Interactions between like molecules become more relevant when \(n_{OH}\) increases, as then the system temperature is closer to the UCST. (v) For ethanol mixtures, \(S_{CC}\) changes in the order: nitromethane \(>\) nitroethane \(\geq\) 1-nitropropane. That is, interactions between like molecules are more relevant in nitromethane mixtures, and slightly more important in nitroethane systems than in those containing 1-nitropropane. (vi) Finally, the model consistently predicts the \(S_{CC}\) decrease of the 1-propanol \(+\) nitromethane system at 333.15 K, i.e., when the separation from the UCST increases. (Figures 9a and 9b).
## 5.7 Comparison between results from models and experimental data
DISQUAC results on VLE, LLE, \(H_{m}^{E}\), or \(C_{p,m}^{E}\) (Tables 4-8) show that the model can be applied rather successfully over a wide range of temperature. Indeed, deviations between experimental and theoretical \(H_{m}^{E}\) results for the systems methanol, ethanol, 1-propanol, or 1-butanol \(+\) nitromethane at 313.15 K are somewhat large. However, it must be underlined that the experimental data seem to be overestimated. Thus, the \(C_{p,m}^{E}\) value, roughly evaluated from data at 298.15 K and 313.15 K, for the methanol \(+\) nitromethane mixture is 24.3 J\(\cdot\)mol\({}^{\text{-}1\text{-}}\)K\({}^{\text{-}1}\), much higher than the value directly measured (9.3 J\(\cdot\)mol\({}^{\text{-}1\text{-}}\)K\({}^{\text{-}1}\)). As a general trend, one can state that the larger differences between experimental results and DISQUAC calculations arise when the system temperature is close to the UCST. At this condition, the experimental \(H_{m}^{E}\)and LLE curves become flattened, while the theoretical ones are more rounded. This is due to DISQUAC is a mean field theory and calculations are conducted assuming that the thermodynamic properties are analytical close to the critical temperature, when, really, they are expressed in terms of power laws. LLE curves are determined by means of DISQUAC assuming, erroneously, that \(G_{m}^{E}\) is an analytical function close to the critical point. The instability of a system is given by \((\partial^{2}G_{m}^{M}/\partial x_{1}^{2})_{P,T}\) and represented by the critical exponent \(\gamma>\)1 in the critical exponents theory. In the framework of this theory, mean field models (\(\gamma=1\)) provide LLE curves which are too high at the UCST and too low at the LCST (lower critical solution temperature). For this reason, the \(C_{\text{ar,1}}^{\text{DIS}}\) coefficients (s = a,h) must be ranged between certain limits in order to provide not very high calculated critical temperatures. This explains the difficulty in describing simultaneously VLE and LLE using the same interaction parameters and the slightly larger \(\sigma_{t}(P)\) values obtained in the case of 1-alkanol \(+\) nitromethane systems (Table 4). On the other hand, the critical exponent linked to the order parameter is, in mean field theories, \(\beta=0.5\), and the derived LLE curves are more rounded close to the UCST, as fluctuations of the order parameter due to the sharp increase of the correlation length are not considered. For the nitrobenzene + _n_-alkane mixtures, there is an additional difficulty when describing the UCST variation with the number of C atoms of the alkane, as the experimental UCST values show a minimum for the heptane system (Table 6). That is, the solubility of mixtures involving pentane or hexane is lower. _N_-methylpyrrolidone systems show a similar trend, which has been explained by the phase rich in alkane is approaching to it gas-liquid critical point. As already mentioned, \(C_{\mathrm{p,m}}^{\mathrm{E}}\) curves of systems of the type polar compound + non-polar compound are W-shaped. In the case of 1-alkanol + nitromethane systems, small negative values are only encountered at low alcohol concentrations and at the temperatures closer to the UCST. This concentration dependence of \(C_{\mathrm{p,m}}^{\mathrm{E}}\) is not properly described by the model. However, DISQUAC correctly predicts the decrease of \(C_{\mathrm{p,m}}^{\mathrm{E}}\) when the difference (_T_-UCST) is increased (Table 8).
Comparison between DISQUAC and UNIFAC shows that DISQUAC describes more accurately the thermodynamic properties under consideration. Particularly, the temperature dependence of the mentioned properties is better represented by DISQUAC. For example, UNIFAC predicts UCST(nitrobenzene)/K = 280 (decane), 300 (hexadecane), which are poorer results than those provided by DISQUAC (Table 6). See also the results for \(C_{\mathrm{p,m}}^{\mathrm{E}}\) (Table 8). DISQUAC also improves ERAS results on \(H_{\mathrm{m}}^{\mathrm{E}}\) (Table 7) and \(G_{\mathrm{m}}^{\mathrm{E}}\) (Figure S1, supplementary material). The most interesting result from the ERAS model is that deviations between experimental \(H_{\mathrm{m}}^{\mathrm{E}}\) values and theoretical results increase with \(n_{\mathrm{OH}}\) for nitromethane systems. That is, larger deviations are encountered when dipolar interactions become progressively more important. Particularly, we note that the theoretical \(H_{\mathrm{m}}^{\mathrm{E}}\) curve of the 1-hexanol system is skewed towards low mole fractions of the alkanol, which suggests that effects related to the alcohol self-association are overestimated by the model. For the nitrobenzene solutions examined, the corresponding deviations are lower, as dipolar interactions are here comparatively less relevant. Nevertheless, we remark that the same type of asymmetry is encountered for the \(H_{\mathrm{m}}^{\mathrm{E}}\) curve determined using ERAS for the 1-butanol mixture.
These so different results point out that the experimental values should be taken with caution, as we have demonstrated, along a number of works, that DISQUAC is a reliable tool to predict VLE and \(H_{\rm m}^{\rm E}\) of ternary mixtures using only binary interaction parameters. Regarding the mentioned systems, other theories do not provide better results. For example, the graph theory gives \(dev(H_{\rm m}^{\rm E})=\) 0.148 and 0.102 for the systems with heptane and cyclohexane, respectively. The Flory model provides, in the same order, 0.205 and 0.035 for this magnitude, and from UNIFAC one obtains \(dev(H_{\rm m}^{\rm E})=\) 0.127 (heptane) and 0.316 (cyclohexane).
Finally, we have paid attention to the nitrobenzene + indane mixture, as solid-liquid equilibrium (SLE) data are available for this system. Indane is important in the petrochemical industry and it is built by one aromatic ring and one aliphatic ring. Thus, there are three contacts in this solution: (b,c); (b,r) and (c,r). The experimental SLE data (\(N=23\)) are well represented by fitting only the \(C_{\sigma,1}^{\rm D\!S}\) coefficient (= - 1.55). Results (see Figure S3, supplementary material) were obtained using the needed physical constants of pure compounds available in the literature. DISQUAC provides 0.005 and the Ideal Solubility Model 0.028 for \(\sigma_{\rm r}(T)\) (relative standard deviation for the temperature, defined similarly to \(\sigma_{\rm r}(P)\) (equation. 9)).
(ii) The \(C_{\rm h,1}^{\rm QUAC}\) coefficients (l =1,2,3) are independent of the alcohol in systems with nitrobenzene. This behavior is also encountered in systems such as 1-alkanol + linear organic carbonate, or + \(n\)-alkanoate or + \(n\)-alkanone or + benzene, or + toluene.
(iii) The \(C_{\rm h,1}^{\rm QUAC}\) coefficients (l = 1,2) for methanol, ethanol or 1-propanol + nitromethane systems are somewhat different from those of the remainder 1-alkanol + 1-nitroalkane mixtures. This merely reflects the difficulty in describing thermodynamics properties of systems including first members of homologous series. Similar trends have been observed for 1-alkanol + \(N\),\(N\)-dialkylamide, + \(n\)-alkanenitrile, + aniline, or + cyclic ether mixtures.
### ERAS interaction parameters
The model provides rather reliable \(H_{\rm m}^{\rm E}\) results (Table 7) using a consistent set of parameters (Table 3). Inspection of Table 3 allows conclude: (i) interactions between unlike molecules become less relevant when \(n_{\rm OH}\) increases, as it is indicated by the corresponding \(K_{\rm AB}\) decrease; (ii) the contribution to the excess functions arising from physical interactions is larger when \(n_{\rm OH}\) increases. Table S2 (supplementary material) shows that the model gives correct \(V_{\rm m}^{\rm E}\) (\(x_{\rm i}=0.5\)) values. However, it fails when describing \(V_{\rm m}^{\rm E}\) (\(x_{\rm i}\)) (Figure S2,supplementary material). It means that the strong structural effects present in these solutions are not properly described by ERAS. In a previous study, better results on \(V_{\rm m}^{\rm E}\) for 1-alkanol + nitromethane systems were obtained assuming that the nitroalkane is weakly self-associated, which is not justified. However, no result on \(H_{\rm m}^{\rm E}\) was provided in that work.
## 6 Conclusions
From the existing database for the studied mixtures, it has been shown that, in alkane solutions, dipolar interactions between 1-nitroalkane molecules are weakened when \(n_{\rm NO2}\) increases. This is in agreement with the \(\overline{\mu}\)(\(n_{\rm NO2}\)) variation. The aromaticity effect leads to interactions between nitrobenzene molecules are stronger than those involving the smaller 1-nitropropane. Dipolar interactions between like molecules are prevalent in systems with 1-alkanols, and become more important when \(n_{\rm OH}\) increases. At this condition, interactions between unlike molecules become weaker, as the OH group is more sterically hindered. Interactions between unlike molecules are stronger in systems with nitromethane than in nitrobenzene solutions. The replacement of nitromethane by nitroethane in systems with a given 1-alkanol leads to effects related with the alcohol self-association are more important. Permittivity data and the application of the Kirkwood-Frohlich model for the \(g_{\rm K}\) determination show that the addition of 1-alkanol to a nitroalkane leads to cooperative effects, which increase the total effective dipole moment of the solution, in such way that the destruction of the existing structure in pure liquids is partially compensated. This effect is less important for larger \(n_{\rm OH}\) values.
## 7 List of symbols
\(C\):\(C_{p}\):\(C_{p}\):\(C_{p}\):\(\Delta H\):\(\Delta H\):\(\Delta H\):\(\Delta H\):\(\Delta H\):\(\Delta H_
\begin{tabular}{l l} \(n_{\rm OH}\) & number of C atoms in 1-alkanol \\ \(p\) & pressure \\ \(P_{\rm i}^{*}\) & reduction parameter for pressure of the component i \\ \(\overline{P_{\rm i}}\) & reduced pressure of component i \\ \(S\) & entropy \\ \(S_{\rm CC}\) & concentration-concentration structure factor (eq. 7) \\ \(T\) & temperature \\ \(U_{\rm V}\) & internal energy at constant volume \\ \(V\) & volume \\ \(V_{\rm mi}^{*}\) & reduction parameter for molar volume of the component i \\ \(\overline{V_{\rm i}}\) & reduced volume of component i \\ \(\Delta v_{\rm i}^{*}\) & self-association volume of component i \\ \(\Delta v_{\rm AB}^{*}\) & association volume of component A with component B \\ \(x\) & mole fraction in liquid phase \\ _Greek letters_ & \\ \(\alpha_{\rm P}\) & isobaric thermal expansion coefficient \\ \(\varepsilon_{r}\) & relative permittivity \\ \(\eta\) & dynamic viscosity \\ \(\kappa_{T}\) & isothermal compressibility \\ \(\mu\) & dipole moment \\ \(\overline{\mu}\) & effective dipole moment (eq. 11) \\ \(\sigma_{r}\) & relative standard deviation (eq. 9) \\ \(X_{12}\) & physical parameter in ERAS \\
## Superscripts
& \\ E & excess property \\
## Subscripts
& \\ i,j & compound in the mixture, (i, j =1,2) \\ m & molar property \\ s,t & type of contact surface in DISQUAC (s \(\neq\) t = a (CH\({}_{3}\); CH\({}_{2}\)); b (C\({}_{6}\)H\({}_{5}\)); c (c-CH\({}_{2}\)); h, (OH); r (NO2) \\ \end{tabular}
|
10.48550/arXiv.2409.16905
|
Thermodynamics of mixtures containing a very strongly polar compound. 12. Systems with nitrobenzene or 1-nitroalkane and hydrocarbons or 1-alkanols
|
Juan Antonio González, Fernando Hevia, Luis Felipe Sanz, Isaías García de la Fuente, Cristina Alonso Tristán
| 828
|
10.48550_arXiv.2310.08938
|
## 1 Introduction
Organic opto-electronic materials offer many advantages over their silicon counterparts, such as lower production cost, smaller weight, higher flexibility and easier disposability, but are hampered by low exciton mobility. Enhancing that mobility therefore has become a major optimization target and several solutions have been proposed, which include increasing the lifetime via triplet formation, ordering molecules to increase exciton delocalization, or coupling the excitons to the confined light modes of an optical resonator. Because the latter solution does not require chemical modifications of the molecules, which may compromise other properties, utilizing strong light-matter coupling could be a promising route towards improving the performance of organic opto-electronic devices.
Because the confinement of light into smaller volumes by an optical resonator increases the interaction with molecular transitions, the enhanced exciton mobility in the strong coupling regime has been attributed to hybridization of excitons and confined light modes into polaritons, which can form when the interaction strength exceeds the decay rates of both excitons and cavity modes. The hybrid states with contributions from cavity modes are bright and can hence be accessed optically. Because the cavity mode energy depends on the in-plane momentum, or wave-vector, \(k_{z}\), these states have dispersion and form the upper and lower polaritonic branches, as shown in Figure 1**f**. These branches are separated by the Rabi splitting, which is defined as the energy gap at the wave-vector for which the energy of the exciton and cavity dispersion are resonant. Most of the hybrid states, however, have negligible contribution from the cavity modes, and are hence dark. These dark states therefore also lack dispersion and form a quasi-degenerate manifold instead that is situated in between the two bright polaritonic branches.
Owing to their dispersion, the bright polaritonic states support ballistic motion of population at their group velocity (_i.e._, \(v_{\mathrm{g}}=\partial\omega(k_{z})/\partial k_{z}\), with \(\hbar\omega(k_{z})\) the energy of a polariton with in-plane momentum \(k_{z}\), Figure 1**f**). However, while in inorganic micro-cavities, such ballistic propagation wasindeed observed, transport in organic microcavities is a diffusion process because of rapid dephasing in disordered organic materials. Results from molecular dynamics (MD) simulations suggest that such dephasing is due to reversible exchange of population between the stationary dark states and propagating polaritonic states. Although polariton-mediated exciton transport is not ballistic in organic systems, polaritonic diffusion can still dramatically outperform the intrinsic exciton diffusivity of the material. However, despite several experimental realizations, and an emerging theoretical understanding of polariton propagation in organic microcavities, strong light-matter coupling has so far not been leveraged systematically for practical applications.
One of the obstacles on the path to polaritonic devices for enhanced exciton transfer is that polariton propagation requires laser excitation of either wavepackets of polaritonic states, or higher-energy electronic states of the molecules. Yet, for practical applications, such as light-harvesting, it will be essential that transport can also be initiated with low-intensity incoherent light sources. To address this specific challenge for a Fabry-Perot optical resonator, we propose to initiate polariton propagation in a strongly coupled molecule-cavity system with a suitable donor that, upon excitation at wavelengths for which the cavity mirrors are more transparent, undergoes a rapid photo-chemical reaction into an excited-state photo-product with an emission maximum that is resonant with both the cavity and acceptor dye molecules. As illustrated in Figures 1 and 2, such system could potentially be realized if we combine 10-hydroxybenzo[h]quinoline (HBQ) that upon excitation at 375 nm or 360 nm undergoes ultra-fast excited-state intra-molecular proton transfer (ESIPT) on a femtosecond timescale into a photo-product with a broad emission centered at 620 nm, with Methylene Blue (MeB) in an optical micro-cavity made of silver mirrors and resonant with the broad absorption peaks of MeB at 668 nm or 609 nm (Figure S3, in the Supplementary Material). Here, we demonstrate the feasibility of this concept, which is somewhat similar to radiative pumping of the LP of a strongly-coupled dye with the emission of a weakly coupled second dye, or to the photo-transformation of an uncoupled reactant into a coupled photo-product, by means of multi-scale molecular dynamics simulations.
## 2 Materials and Methods
In the simulations, the details of which can be found in the Supplementary Material (SM), a single HBQ molecule solvated in cyclohexane, was combined with 1023 hydrated MeB molecules. While the number of molecules in real cavities is estimated to be much higher (_i.e._, \(10^{5}-10^{8}\)), we could show in previous work that for modeling polariton propagation, including 1024 molecules in the simulation provides a reasonable compromise. The electronic ground (S\({}_{0}\)) and excited (S\({}_{1}\)) states of HBQ were modeled with Density Functional Theory (DFT) and time-dependent density functional theory (TDDFT) within the Tamm-Dancoff approximation (TDA), respectively, using the CAM-B3LYP functional in combination with the 3-21G basis set. The cyclohexane solvent molecules were modelled with the GROMOS 2016H66 force field. At this level of theory the vertical excitation energy of HBQ is \(h\nu^{\text{HBQ}}=4.06\) eV (305 nm), while the energy gap to the ground state is 2.58 eV (480 nm) in the S\({}_{1}\) minimum. Despite the overestimation of the S\({}_{1}\)-S\({}_{0}\) energy gap, our model provides potential energy surfaces (Figure 1**d**) that are in qualitative agreement with the more accurate description at the TPSSh/cc-pVDZ level of theory for this system (Figure S4 in SM). The S\({}_{0}\) and S\({}_{1}\) electronic states of MeB were modelled with DFT and TDDFT based on the Casida equations, respectively, using the B97 functional and the 3-21G basis set. The water molecules were described with the TIP3P model. Although at this level of theory the vertical excitation energy of MeB is \(h\nu^{\text{MeB}}=2.5\) eV, and thus significantly overestimated with respect to experiment, there is a fortuitous overlap with the emission of HBQ that we exploit in this work (Figure 1**e**). Thus, while MeB may not be the optimal choice for a practical realization, this dye should be suitable for demonstrating the feasibility of inducing polariton-mediated exciton transfer with a photochemical reaction in our simulations.
We modelled the optical resonator as a one-dimensional Fabry-Perot cavity with 239 discrete modes (Figure 1**f**) and a cavity vacuum field strength of 0.21 MVcm\({}^{-1}\) (0.00004 au). The fundamental mode at \(k_{z}=0\) was red-detuned by 323 meV with respect to the excitation maximum of MeB (2.5 eV at the TD-B97/3-21G level of theory).
Illustration of a Fabry-Pérot microcavity (panel **a**, not to scale) containing a 10-hydroxybenzo[h]quinoline donor molecule (HBQ, panel **b**) and 1023 Methylene Blue acceptor molecules (MeB, panel **c**). The first singlet excited states (S\({}_{1}\)) of the MeB molecules are coupled to the 239 modes of the cavity. Upon absorbing a photon at a frequency where the mirrors have become more transparent (\(\sim\) 4.0 eV at the TDA-CAMB3LYP/3-21G level of theory, Figure S3 in SM), HBQ undergoes ultra-fast intra-molecular proton transfer on the S\({}_{1}\) excited-state potential energy surface (panel **d**) into an excited-state photo-product that is resonant with both the absorption maximum of MeB and the cavity. Panel **e** shows the QM/MM absorption (magenta) and emission (cyan) spectra of HBQ and the absorption spectrum of MeB (red). The normalised angle-resolved absorption spectrum of the molecule-cavity system (panel **f**) shows the Rabi splitting of 282 meV between the lower polariton (LP) and upper polariton (UP) branches. The cavity dispersion is plotted as a white dashed line, while the excitation maxima of the MeB molecules (\(\sim\) 2.5 eV at the TD-B97/3-21G level of theory) and HBQ are plotted as straight red and magenta lines, respectively.
We computed Ehrenfest MD trajectories, in which the nuclei evolve classically on the mean-field potential energy surface of the total light-matter wave function (SM). To account for the finite lifetime of the cavity modes (here \(\tau_{\rm cav}\) = 15 fs, in line with experiments on metallic micro-cavities), the wave function was propagated along the classical trajectory under the influence of an effective non-Hermitian Hamiltonian (SM), in which the loss-terms were added to the cavity mode energies (_i.e._, \(\hbar\omega(k_{z})-1/2i\gamma_{\rm cav}\), with \(\hbar\omega(k_{z})\) the dispersion of the empty cavity, shown in dashed white lines in Figure 1**f**, and \(\gamma_{\rm cav}=1/\tau_{\rm cav}\) the decay rate of the cavity).
## 3 Results and Discussion
In we plot the progress of the ESIPT reaction, defined as the distance between the hydroxyl oxygen and the proton (**a**), the excitonic wavepacket \(|\Psi_{\rm exc}(z,t)|^{2}\) (**b**), the contributions of the molecular excitations to the total wave function, \(|\Psi(z,t)|^{2}\) (**c**,**d**), and the Mean Squared Displacement of the excitonic wavepacket (MSDexc, **e**). After photo-excitation into the highest-energy eigenstate of the molecule-cavity system (_i.e._, \(|\psi^{1263}\)), which is dominated by the S\({}_{1}\) electronic state of HBQ (_i.e._, \(|\beta_{\rm HBQ}^{1263}|^{2}>0.9995\), Figure 3**c**), with a photon of an energy above the polaritonic manifold, the proton transfers from the hydroxyl oxygen to the nitrogen atom (Figure 3**a**). Because this end to keto transformation is accompanied by a 1.4 eV redshift of the S\({}_{1}\)-S\({}_{0}\) energy gap, HBQ becomes resonant with the MeB molecules as well as the cavity modes, and enters the dark state manifold around 10 fs after excitation.
Due to displacements along vibrations parallel to the non-adiabatic coupling vector between this new dark state that is localized on HBQ, on the one hand, and the delocalized bright polaritonic states of the strongly coupled molecule-cavity system on the other hand, population is transferred from HBQ into the bright states. Because these bright polaritonic states have group velocity, the transferred population starts propagating, as shown in Figure 3**b**. However, since the population transfer is reversible, this propagation continues ballistically until the population transfers back into the dark state manifold.
Simplified Jablonski diagram of states involved in photo-chemically induced polariton propagation. The cavity modes and polaritons are schematically shown as continuous dispersions. After photo-excitation of HBQ at \(h\nu^{\rm HBQ}\), excited-state intra-molecular proton transfer (ESIPT) brings the excited state population that was initially localized on HBQ, into the dark state manifold. Reversible population exchanges between the stationary dark states and the propagating lower polaritonic (LP) bright states, cause the population to move away from the HBQ molecule and diffuse into the cavity. The path along which the population arrives in the LP states is illustrated by a blue arrow.
Panel **a** shows the distance between the oxygen and proton, which we use as the reaction coordinate for the excited-state proton transfer reaction (inset and see also Figure 1**c**) as a function of time after instantaneous excitation into the highest energy eigenstate of the molecule-cavity system, which is dominated by the S\({}_{1}\) state of HBQ (_i.e._, \(|\beta_{\text{HBQ}}|^{2}\approx 0.99\), panel **c**). Panel **b** depicts the probability density of the excitonic part of the wave function \(|\Psi_{\text{exc}}|^{2}\) as a function of the \(z\)-coordinate (horizontal axis) and time (vertical axis). Panels **c** and **d** show the contribution of the HBQ molecule (red) and the Methylene Blue molecules (grey) to the total wave function, as well as the population of the ground state (blue). Panel **e** shows the Mean Squared Displacement (_i.e._, \(\text{MSD}_{\text{exc}}=(z(t)-z)^{2}\)) of the excitonic wavepacket.
The observation of diffusion rather than ballistic motion, is in line with experiments on organic micro-cavities.
The initial structures for our simulations were sampled from equilibrium QM/MM trajectories at 300 K (SM) and therefore can capture the heterogeneity as indicated by the absorption line-widths of the molecules in Figure 1**e**. Because of such structural disorder, the ESIPT reaction rates span a distribution. To confirm that the proton transfer in HBQ initiates the polariton-mediated exciton transport process, we show in Figure S5 (SM) that for a system in which the ESIPT is delayed, also the transport starts at a later point in time, and that this time point coincides with the formation of the HBQ photo-product.
## 4 Conclusion
To summarize, the results of our MD simulations suggest that long-range polariton-mediated exciton transport can be induced with an excited-state proton transfer reaction. While the excitation scheme proposed here resembles the off-resonant laser excitation conditions employed in previous experiments of polariton transport, the absorption cross-section of HBQ should be high enough to initiate the propagation with incoherent light, in particular for a cavity with a thin silver top mirror, which is more than 50% transparent at the required wave length (Figure S3, SM).
|
10.48550/arXiv.2310.08938
|
Photochemical Initiation of Polariton Propagation
|
Ilia Sokolovskii, Gerrit Groenhof
| 179
|
10.48550_arXiv.1707.00856
|
## I Introduction
Despite the enormous success of density functional theory (DFT) in many fields of modern physics, its predictive power is impaired by the intrinsic errors associated with the approximations made for the exchange-correlation functional. Delocalization error is one of the dominant errors of mainstream density functional approximations (DFAs). It is responsible for many failures of DFT calculations. Its physical origin is the violation of the Perdew-Parr-Levy-Balduz (PPLB) condition, which requires the total energy of a system as a function of electron number, \(E(N)\), to be piecewise straight lines interpolating between integers. DFAs suffering from delocalization error yield \(E(N)\) underneath the piecewise linear segments for finite systems, and thus tend to give too low energy for delocalized electron distributions, and produce excessively delocalized electron distribution and sometimes qualitatively wrong density, as it falsely lowers the system energy. Moreover, the error in the total energy also transfers to the error in the energy derivatives with respect to the electron number, i.e., the chemical potentials. As a consequence, the frontier orbital energies as predicted by DFAs significantly deviate from the true ionization potentials or electron affinities. This also applies to infinite systems, where delocalization error is indicated by the unphysical narrowing of the band gap.
The manifestation of delocalization error is size-dependent. Figure 1(a) explores the evolution of error in a series of loosely bound \(\mathrm{He}_{M}\) clusters, where all the He atoms are chemically equivalent but separated by a large distance. For a small \(M\), for example \(M=1\), the energy of the highest occupied molecular orbital (\(\epsilon_{\mathrm{HOMO}}\)) resulting from the paradigmatic local density approximation (LDA) significantly overestimates the calculated negative vertical ionization potential (\(-I_{\mathrm{ve}}\)), leading to a large positive error bar (shown in red). In comparison, the deviation of \(-I_{\mathrm{ve}}\) from the experimental value (\(I_{\mathrm{exp}}\)) is negligibly small. As the cluster grows from the atomic limit (\(M=1\)) to the bulk limit (\(M=\infty\)), however, while \(\epsilon_{\mathrm{HOMO}}\) gradually approaches \(-I_{\mathrm{ve}}\), \(I_{\mathrm{ve}}\) deviates increasingly from the experimental value (\(I_{\mathrm{exp}}\)), resulting in a large negative error bar.
The above deviations can be represented and understood from the fractional charge perspective. Figure 1(b) shows how the calculated \(E(N)\) curve of a helium atom deviates from the PPLB condition. The discrepancy \(\Delta I=I_{\mathrm{ve}}-I_{\mathrm{exp}}\) for \(\mathrm{He}_{M}\) corresponds quantitatively to the deviation of total energy from linearity, \(\Delta E\), for a helium atom upon removal of \(\frac{1}{M}\) electron. In particular, \(\Delta E(\frac{1}{M})=\frac{1}{M}\Delta I(\mathrm{He}_{M})\). Moreover, \(\epsilon_{\mathrm{HOMO}}-I_{\mathrm{ve}}(\mathrm{He})=\frac{\partial E}{ \partial N}|^{-}-I_{\mathrm{ve}}(\mathrm{He})=-\frac{d\Delta E}{d\sigma} \Big{|}_{x=0^{+}}\) is the tangent slope error at integer \(N\) (note the positive \(x\) direction in Figure 1(b) is to the left), which agrees with \(-\Delta I(\mathrm{He}_{M})\) only in the infinite \(M\) limit.
\[-\frac{d\Delta E}{dx}\Big{|}_{x=0^{+}} = -\lim_{M\to\infty}\frac{\Delta E(\frac{1}{M})-\Delta E}{\frac{1}{M}} \tag{1}\] \[= -\lim_{M\to\infty}M\Delta E(\frac{1}{M})\] \[= -\lim_{M\to\infty}\Delta I(\mathrm{He}_{M}).\]
Therefore, the positive error bar in the orbital energy for \(M=1\) exactly agrees in absolute value with the negative error bar in the total energy for \(M=\infty\), demonstrating that the delocalization error for finite (small) and bulk systems are similar but are manifested in two different ways.
The situation is similar for many other DFAs, such as the popular hybrid B3LYP functional. As shown in Fig. 1, inclusion of Hartree-Fock (HF) exchange only slightly reduces the energy deviations, because the delocalization error is compensated partly by the localization error (referring to \(E(N)\) above the piecewise linear segment) associated with the HF exchange.
Here we make a comparison between the delocalization error with related concepts of self-interaction error (SIE) and many electron self-interaction error (MSIE). SIE was the first concept to describe systematic error in DFAs and refers to the failure of compensation between the Hatree and exchange-correlation functional, or the error in the electron interaction energy, with DFAs in _one-electron_ systems. The SIE concept also applies to fractional electron systems with less than one electron. It had been presumed to be the key reason for some systematic failure in DFAs and various SIE-free functionals have been constructed. However, in 2006 this presumption was discovered untrue: it was found that two well-developed functionals MCY2 and B05, while both are SIE-free by construction and have improvements over hybrid functionals in describing chemical reaction barriers, still exhibit similar behavior on the remaining difficulties such as incorrect dissociation limit for molecular ions and overestimation of molecular polarizability for polymers, which was once associated with SIE. This leads to the identification of a different source of the systematic errors of DFAs, namely, the essentially convex deviation from the exact linear condition of fractional number of electrons, and the introduction of the concept of many-electron self-interaction error (MSIE). However, multiple definitions of MSIE exist: the term many-electron self-interaction error has also been defined in Ref, yet it was used to describe both convex and concave behaviors of fractional charges, which leads to totally different behaviors of errors in DFAs.
The concepts of localization error (LE) and delocalization error (DE) were then introduced to capture the physical essence of the problem for systems with any number of electrons. Delocalization error, which exists in all commonly used DFAs including hybrids, describes the essentially convex deviation from the exact linear \(E(N)\) curve for fractional charges and highlights the associated unphysical delocalization of electrons, or unphysically low energies for delocalized electrons. Localization error, which exists in Hartree-Fock (HF) approximation, describes the concave deviation from the exact linear \(E(N)\) curve for fractional charges and highlights the associated unphysical localization of electrons. Moreover, the concept of DE essentially incorporates the concept of SIE. For one-electron systems, the SIE grows with increasing local fractional electron numbers and this is the reason for the failure in describing \(\mathrm{H}_{2}^{+}\) dissccioation, as first pointed out in Ref. This can also be characterized by DE. The only missing component of SIE in the description of DE is the integer energy error for one electron systems that are compact as in molecular equilibrium structures. Nevertheless, for such systems, the SIE of commonly used DFAs is much smaller compared to fractional systems. Neglecting the integer error, DE and SIE become identical for systems with one electron or less. In other word, functionals that are free of DE is essentially SIE-free, while SIE-free functionals can still suffer significantly from DE, just as other commonly used DFAs.
DE thus pinpoints the critical flaws in commonly used DFA and its reduction has been the driving force in many functional developments.
(a) Deviations between the calculated \(\epsilon_{\mathrm{HOMO}}\) and \(-I_{\mathrm{ve}}\) and between \(I_{\mathrm{ve}}\) and \(I_{\mathrm{exp}}\) for a series of \(\mathrm{He}_{M}\) clusters. In each cluster all the He atoms are chemically equivalent. The nearest neighboring atoms are separated by \(10\,\mathrm{\AA}\), and the \(I_{\mathrm{exp}}\) of \(\mathrm{He}_{M}\) is well approximated by \(I_{\mathrm{exp}}\) of a He atom. (b) Calculated total energy deviation from the linearity condition of a fractionally charged He atom as a function of the fractional charge \(\delta\). Here \(\Delta E(\mathrm{He}^{\delta+})=E(\mathrm{He}^{\delta+})-\delta E(\mathrm{He}^ {+})-(1-\delta)E(\mathrm{He})\), and the \(\delta\) values have been scaled in the figure for a direct comparison with (a).
The corresponding localization error for HF is shown in the supplemental material.
Enormous efforts have been devoted towards systematic removal of delocalization error. These include the development of global hybrid, local hybrid, doubly hybrid and range-separated functionals. Like the B3LYP, the performance of these DFAs relies on the cancellation of errors, and is often system-dependent. There are also attempts focusing on specific properties, such as the Koopmans-compliant functional, the generalized transition state method and related approaches, and others.
To achieve a universal elimination of delocalization error, it is crucial to treat fractional electron distribution explicitly and properly. Following this idea, in 2011 we have developed a global scaling correction (GSC) approach by imposing the PPLB condition globally on any given system. The GSC largely restores the energy linearity condition at any fractional electron number, and thus reduces substantially the discrepancy between \(\epsilon_{\rm HOMO}\) (\(\epsilon_{\rm LUMO}\)) and \(-I_{\rm ve}\) (\(-A_{\rm ve}\)) for systems of all sizes. Here, \(\epsilon_{\rm LUMO}\) is the energy of the lowest unoccupied molecular orbital, and \(A_{\rm ve}\) is the vertical electron affinity. Retrieving PPLB condition for electron addition within the framework of DFT guarantees the improvement in the prediction of \(-I_{\rm ve}\) using \(\epsilon_{\rm HOMO}\), as well as \(-A_{\rm ve}\) using \(\epsilon_{\rm LUMO}\). Despite the success of GSC, it does not offer any cure to the size-dependent discrepancy \(\Delta I\), because it gives zero correction to energies at integers. This implies that the parent functionals such as LDA is not _size-consistent_ for calculations of \(I_{\rm ve}\) (\(A_{\rm ve}\)); and any functional correction that cannot affect the energies at integers is not _size-consistent_ either. To address this issue, in 2015 we have developed a local scaling correction (LSC) scheme that enforces the PPLB condition on local subsystems. The LSC is capable of correcting the total energy and electron density of integer-\(N\) systems self-consistently, and thus leads to much improved description of dissociating molecules, transition-state species and charge-transfer systems. However, the LSC has difficulty in capturing an infinitesimal amount of fractional electron, and so it cannot improve the prediction of \(\epsilon_{\rm HOMO}\), \(\epsilon_{\rm LUMO}\) and thus fundamental gaps from the chemical potential differences.
It would be ideal to incorporate the merits of describing global fractions (as in GSC) and local fractions (as in LSC) in one framework while avoiding their difficulties. To this end, two major advancements are needed: a unified scheme for characterizing global and local fractional electron distributions; and an explicit and size-consistent correction to system energies at both fractional and integer electron numbers. In this paper, we develop a localized orbital scaling correction (LOSC) framework, which realizes these advancements. To the best of our knowledge, this is the first density functional approximation that alleviates delocalization error in all aspects without involving system dependent parameters. In the rest of the paper, we will first revisit the forms of GSC and LSC, and then proceed to formulate the LOSC functional. Finally we will close with some concluding remarks.
## II Motivation
The basic idea of GSC is to add a correction functional to linearize each of the nonlinear components of the KS functional in the presence of fractional electron in the global system. The nonlinear components include the Hartree functional and the exchange-correlation functional, with the former being a quadratic functional of the electron density \(\rho\) and the latter having a more complicated scaling relation with \(\rho\).
\[\Delta E^{\rm GSC}=\frac{1}{2}\kappa\left(n_{\!f}-n_{\!f}^{2}\right). \tag{2}\]
Here \(n_{\!f}=N-[N]\), and also equals the fractional electron occupation number on the KS frontier orbital, and \(\kappa\) is a functional of the frontier orbital \(\varphi_{f}(\mathbf{r})\), which itself is a functional of the KS reduced density matrix \(\rho_{s}\).
\[\frac{1}{2}\kappa=\frac{1}{2}\!\!\int\!\!\!\int\frac{\rho_{f}(\mathbf{r})\rho_ {f}(\mathbf{r}^{\prime})}{|\mathbf{r}-\mathbf{r}^{\prime}|}d\mathbf{r}d \mathbf{r}^{\prime}-\frac{C_{\rm x}}{3}\!\int\![\rho_{f}(\mathbf{r})]^{\frac{ 4}{3}}d\mathbf{r}, \tag{3}\]
One can verify that \(\frac{\partial\Delta E^{\rm GSC}}{\partial n_{\!f}}=\pm\frac{1}{2}\kappa\) at \(n_{\!f}\to 0\) or \(1\) gives a correction to the tangent slope at integers, while \(\frac{\partial^{2}\Delta E^{\rm GSC}}{\partial n_{\!f}^{2}}=-\kappa\) modifies the curvature of \(E(N)\) of the DFA.
The functional form of LSC can be viewed as a "local" version of Eq.,
\[\Delta E^{\rm LSC}\approx\frac{1}{2}\int\!\!\!\int d\mathbf{r}d\mathbf{r}^{ \prime}\,\tilde{n}_{\!f}(\mathbf{r})\left[1-\tilde{n}_{\!f}(\mathbf{r}^{ \prime})\right]\tilde{\kappa}(\mathbf{r},\mathbf{r}^{\prime}), \tag{4}\]
These local variables are introduced as functionals of the KS reduced density matrix to capture the local fractional information. Note despite the fact that the original form of LSC as in Ref slightly differs from Eq., with some range separation and extra non-local attraction, Eq. captures the main idea of LSC. Here for the simplicity of comparison, we stick with Eq. as the approximate form of LSC.
Let us consider two typical examples: H\({}_{2}^{+}\) at large internuclear distance \(R\); He\({}_{M}\) cluster as mentioned in In, at large \(R\), the one-electron H\({}_{2}^{+}\) molecule yields a delocalized electron density over the two separated protons, with each subsystem density integrating to half an electron. The LDA severely underestimates the energy of the stretched H\({}_{2}^{+}\) because of delocalization error. By design, GSC cannot capture the locally half electron information because it counts fractional electrons globally. In contrast, LSC yields \(\tilde{n}_{f}(\mathbf{r})\simeq 0.5\) near each proton by imposing a spatial screening on the density matrix, from which the local information is extracted, and then effectively performing the calculation of \(n_{f}(\mathbf{r})-n_{f}(\mathbf{r})^{m}\), with \(n_{f}(\mathbf{r})\) being the fractional component of the locally screened density matrix and \(m\) a large integer (fixed to be 10). It is by the vanishing nature of high powers of any fractional number that enables \(\tilde{n}_{f}(\mathbf{r})\) to approximate \(n_{f}(\mathbf{r})\), and thus capture the half electron information. Yet, this only allows us to distinguish a fractional number from integer 0 or 1; it cannot distinguish two integers because \(n_{f}(\mathbf{r})-n_{f}(\mathbf{r})^{m}=0\) identically for all such occupations. Moreover, distinguishing a tiny fractional \(\tilde{n}_{f}(\mathbf{r})\) from zero is also numerically difficult. This causes trouble in example. On the one hand, in the case of He\({}_{M}^{+}\), as \(M\rightarrow\infty\) the local fraction becomes vanishingly small, which poses numerical challenge for LSC to capture the tiny fraction. On the other hand, in the case of neutral He\({}_{M}\) where local fraction is absent, LSC cannot capture the right "local frontier orbital" to compute the local curvature for the frontier orbital energy correction.
The above analysis thus suggests that the key is to capture the "local frontier orbital" and its local occupation in order to solve all the problem together. Here we highlight that the two key words are "local" and "frontier". The extension from GSC to LSC captures "local" but overlooks "frontier", with the latter requiring energy information to enter in the local variable construction, rather than simply invoking the density matrix. This is a great challenge if our local extension is through the \(\mathbf{r}\) space, since then we have to devise a "local frontier energy" function to achieve our purpose, which is difficult in both conceptual and practical manner.
## IV Formulation of Losc
We now pursue an alternative way to realize the local extension through the orbital space, by invoking localized orbitals (LOs). Note that the KS density matrix is a sum over KS canonical orbital (CO) projections, \(\rho_{s}=\sum_{m}n_{m}|\varphi_{m}\rangle\langle\varphi_{m}|\), and the occupation numbers \(\{n_{m}\}\) are all integers (0 or 1) for integer systems. The COs are not the only choice for unraveling \(\rho_{s}\).
\[\rho_{s}=\sum_{ij}|\phi_{i}\rangle\langle\phi_{i}|\rho_{s}|\phi_{j}\rangle \langle\phi_{j}|=\sum_{ij}\lambda_{ij}|\phi_{i}\rangle\langle\phi_{j}|, \tag{5}\]
In particular, one can show that the diagonal elements satisfy \(0\leqslant\lambda_{ii}\leqslant 1\) and \(\sum_{i}\lambda_{ii}=N\), so that each \(\lambda_{ii}\) plays the role of an occupation number associated with \(\phi_{i}\). Moreover, our desired local fractions arise naturally through the fractional components of \(\{\lambda_{ii}\}\). This is the motivation of LOSC. Now the remaining task is to construct the LOs from \(\rho_{s}\) and to build a correction functional out of \(\{\phi_{i}\}\) and \(\lambda_{ij}\).
In the orbital space, the LOs should come from a unitary transformation of the COs through a localization procedure. In the traditional procedures, such as in the Boys and the Ruedenberg prescription, one minimizes a target function of only occupied COs, rendering the virtual COs irrelevant. These approaches, however, cannot serve our purpose. In the H\({}_{2}^{+}\), for example, only one occupied CO exists, which leads to identical orbital with itself no matter what localization scheme is implemented. To obtain the desired LO, we have to incorporate the virtual COs into the localization scheme to form collective unitary rotations with the occupied orbitals. Moreover, we note that at large \(R\), the HOMO and LUMO are near-degenerate, and energetically separated from other COs; the desired LOs should come from linear combinations (mixtures) of only HOMO and LUMO, and not from any other CO. Therefore, the localization procedure should be able to limit the mixing of these other orbitals with HOMO/LUMO.
This suggests that our desired localized orbitals should achieve _a compromised localization both in the physical space and in the energy space_. In contrast, traditional localized orbitals keep localization only in the physical space by construction, while traditional canonical orbitals maintain localization only in the energy space, being energy eigenstates of an one-particle Hamiltonian. Here to better describe our desired localized orbitals, we introduce a new concept and call them _orbitallets_, in analogy to wavelets, which achieve a compromised localization both in the physical space and in the momentum space.
There are many ways to obtain localization in both the physical and energy spaces. A simple way to achieve this is to modify the localization target function by adding a penalty function that enforces the localization of CO energies. Here to have localization in energy space for our LOs (orbitallets), we have to define an energy spectrum for them. In this paper, we define it to be the same as the CO spectrum. One can invoke other definitions, but our definition seems to be the most natural choice. Each LO with energy \(\epsilon_{i}^{\rm LO}\) results from a mixture of the COs whose energies are within a certain energy window of a fixed size centered at \(\epsilon_{i}^{\rm LO}\), and the mixing coefficients come from a unitary matrix \(U_{im}=\langle\phi_{i}|\varphi_{m}\rangle\). See for a schematic illustration. It is worth mentioning that localization involving both occupied and virtual orbitals have been used previously within a fixed energy window near the Fermi level for constructing maximally localized Wannier functions for systems with entangled energy bands. Defined in a different way and for a different purpose, our localized orbitals, orbitallets, have dynamic energy windows opened at each \(\epsilon_{i}^{\rm LO}\) and are designed to capture the local fractions in the system.
In our present implementation of LOSC, \(\{\phi_{i}({\bf r})\}\), the LOs, are generated by a restrained Boys localization procedure, which aims at minimizing the following spread function
\[F=\sum_{i}^{\infty}\Big{[}\langle\phi_{i}|{\bf r}^{2}|\phi_{i}\rangle-\langle \phi_{i}|{\bf r}|\phi_{i}\rangle^{2}\Big{]}+\sum_{im}^{\infty}w_{im}|U_{im}|^ {2}. \tag{6}\]
Here the first term on the right hand side of Eq. resembles the Boys spread function, except that the sum runs over all the orbitals. In the second term, a penalty function \(w_{im}=w(|\epsilon_{i}^{\rm LO}-\epsilon_{m}^{\rm CO}|)\) is introduced to suppress mixing of orbitals beyond an energy radius \(\epsilon_{0}\), in order to achieve energy localization. As a remark, without the penalty term, minimization of \(F\) with infinite number of orbitals (which span the entire Hilbert space of functions) will lead to a set of \(\delta\) functions centered at different positions and \(F=0\). Therefore, from pure mathematical perspective, the penalty function is also needed if the virtual orbitals participate in the localization. From the physical considerations, the penalty function \(w(x)\) should satisfy the following requirements: \(w=0\), which implies no penalty imposed against mixing between degenerate COs; \(w(x)\) is a monotonically increasing function; \(w(\infty)=\infty\), which forbids mixing between COs that are far apart in energy.
\[w(x)=R_{0}^{2}(\frac{x}{\epsilon_{0}})^{\gamma}[1-e^{-(\frac{x}{\epsilon_{0}}) ^{\eta}}], \tag{7}\]
\(R_{0}\) can be considered as related to the typical spread radius of small molecules, and is introduced here to factor out the length unit; \(\epsilon_{0}\) is the energy window as mentioned above; and \(\gamma\) and \(\eta\) are the other parameters to adjust the shape of the function, in particular, the asymptotic behavior as \(x\to 0\) and \(x\rightarrow\infty\). Note that \(\gamma\) has to be positive to satisfy condition. However, after some numerical experiment, we find that a positive \(\gamma\) imposes insufficient penalty for \(x<\epsilon_{0}\), leading to excessively artificial local fractions for compact molecules. Therefore, we modify the function by dividing it into a piecewise function, with \(\gamma=0\) when \(x<\epsilon_{0}\) and \(\gamma>0\) when \(x\geqslant\epsilon_{0}\),
\[w(x)=\left\{\begin{array}{cc}R_{0}^{2}[1-e^{-(\frac{x}{\epsilon_{0}})^{ \eta}}],&\mbox{if }x<\epsilon_{0},\\ R_{0}^{2}(\frac{x}{\epsilon_{0}})^{\gamma}[1-e^{-(\frac{x}{\epsilon_{0}})^{\eta }}],&\mbox{if }x\geqslant\epsilon_{0}.\end{array}\right. \tag{8}\]
The parameters are optimized for a balanced behavior on thermochemistry, reaction barrier heights and dissociation curves, resulting in \(R_{0}=2.7\)A, \(\epsilon_{0}=2.5\)eV, \(\gamma=2.0\) and \(\eta=3.0\), see the supplemental materials for more details. The penalty function in Eq. is continuous but not smooth at \(x=\epsilon_{0}\). Yet this artificial feature has little impact on our results and can be easily removed by rounding it out or by a smooth interpolation between the two constructing functions. In the present work, since we put more emphasis on the general idea, we will not need to refine the details and make more sophisticated forms.
In the minimal basis, the H\({}_{2}^{+}\) molecule has two KS orbitals. Figure 3(a) depicts the LO densities of a compact H\({}_{2}^{+}\). Both LOs resemble the original COs, and the diagonal elements of the local occupation matrix remain integers, \(\lambda_{11}=1\) and \(\lambda_{22}=0\). This is because the HOMO and LUMO are far apart in energy so that their mixing is suppressed, rendering LOs of the similar character as the COs. This is reasonable since there is hardly any fractional electron distribution at a small internuclear distance \(R\). In contrast, at a large \(R\) the two nearly degenerate COs fully mix into two spatially separated LOs. The two LOs locate symmetrically at the two nuclei with \(\lambda_{11}=\lambda_{22}=0.5\), which reveals the fact that each proton carries half an electron (see Figure 3(b)). In addition, the above behavior is almost independent of the basis set.
Here as a remark, for H\({}_{2}^{+}\) at a small \(R\) or a large \(R\), different choice of the localization target functions, for example Boys or Ruedenberg, does not make a difference
Schematic illustration of the relation between LOs and COs and the energy window. Here each energy window is centered at \(\epsilon_{i}^{\rm LO}\) with a fixed radius \(\epsilon_{0}\).
This, however, will make a difference in the intermediate \(R\) or in a more complicated molecule. Yet this is a minor effect and beyond the scope of the present paper. Here we choose to modify Boys localization because it can be easily implemented and applied to systems of all sizes, and more importantly the Boys LOs can be replaced by Wannier functions for periodic solids.
_Explicit functional form._ With the LOs (orbitallets) and \(\mathbf{\lambda}\)-matrix, an explicit form of LOSC is constructed, similar to Eqs. and:
\[\Delta E^{\text{LOSC}}=\sum_{ij}\frac{1}{2}\kappa_{ij}\lambda_{ij}\left(\delta _{ij}-\lambda_{ij}\right)=\frac{1}{2}\text{tr}(\mathbf{\kappa}\mathbf{\omega}), \tag{9}\]
Here the diagonal terms in the summation of Eq. is a natural extension of the energy correction formula of Eq. from one global fraction to all the local fractions. Note that for \(\lambda_{ii}=0\) or \(1\), their contributions to \(\Delta E^{\text{LOSC}}\) are zero and thus have no impact on the sum. The off-diagonal terms are introduced as non-local corrections to the unphysical interaction between the local fractions centered at different positions. They play similar role as the long-range attraction term in the LSC functional.
\[\frac{1}{2}\kappa_{ij}=\frac{1}{2}\!\iint\frac{\rho_{i}(\mathbf{r})\rho_{j}( \mathbf{r}^{\prime})}{|\mathbf{r}-\mathbf{r}^{\prime}|}d\mathbf{r}d\mathbf{r} ^{\prime}\!-\!\frac{\tau C_{\text{x}}}{3}\!\int[\rho_{i}(\mathbf{r})]^{\frac{ 2}{3}}[\rho_{j}(\mathbf{r})]^{\frac{2}{3}}d\mathbf{r}, \tag{10}\]
This is a straightforward extension of Eq. in a symmetric manner with respect to \(i\) and \(j\), but with a parameter \(\tau\). In Eq., \(\tau\) is set to \(1\), which is good for the orbital corrections. In particular, in the case of HOMO energy correction of a hydrogen atom, with the exchange only LDA functional one can show that Eq. exactly compensates the wrong slope under the frozen orbital assumption. However, \(\tau=1\) does not retrieve the right amount of correction for \(\text{H}^{1/2+}\). This is because the LDA exchange functional is not quadratic, so that the energy deviation from linearity in the \(E\) vs \(N\) curve of (exchange only) LDA is not strictly quadratic either, although it can be approximately treated as a parabola. As a consequence, under this parabolic assumption, one cannot simultaneously retrieve the right slope at integer and the energy at half integer. In order to do that, higher order corrections have to be introduced. In the present paper, we are biased towards the half integer energy within the parabolic correction under the frozen orbital analysis, and assign \(\tau\) a nonempirical value of \(\tau=6(1-2^{-1/3})\approx 1.2378\). In the frozen-orbital approximation, Eq. leads to the following correction to the CO energies:
\[\Delta\epsilon_{m}=\sum_{i}\kappa_{ii}\left(\frac{1}{2}-\lambda_{ ii}\right)|U_{im}|^{2}-\sum_{i\neq j}\kappa_{ij}\lambda_{ij}U_{im}U_{jm}^{*}. \tag{11}\]
As shown in Fig. 3(c), the dissociation energy curve of \(\text{H}_{2}^{+}\) is greatly improved for \(R>2\,\text{\AA}\) by using Eq.; while for more compact geometries (\(R<2\,\text{\AA}\)) the energy correction is almost zero. The latter is due to the specific form of the penalty function adopted in the restrained Boys localization of Eq. - it is designed to preserve the good accuracy of parent DFAs on thermochemistry for small- and medium-sized molecules. In particular, at a small \(R\), the localization plays trivial role in the sense that the unitary transformation \(\mathbf{U}\) is almost an identity matrix. This suggests that the minimizer of the restrained Boys target function is achieved almost at its boundary due to the small overall spread of COs and large penalty against their mixing. As \(R\) increases, the CO spread grows while the penalty term is alleviated. When \(R\) reaches a critical value (which is mainly dependent on the \(R_{0}\) parameter; smaller \(R_{0}\) will reduce the critical \(R\)), mixing between COs becomes favorable so that \(\mathbf{U}\) becomes different from identity. This causes a kink at the critical \(R\), an artifact due to the choice of our present localization scheme, and shall be addressed through a better construction scheme in future work. In addition, we note that for large \(R\), the LOSC-LDA energy is slightly above zero, which suggests that \(\text{H}^{1/2+}\) is over-corrected. This is because the LDA correlation energy correction has been neglected- we introduce a non-empirical \(\tau\) parameter only to account for the LDA
Distribution of LO densities along the bonding axis of \(\text{H}_{2}^{+}\) at the internuclear distance of (a) \(R=1\,\text{\AA}\) and (b) \(R=5\,\text{\AA}\). The two protons locate at \(x=0\) and \(x=R\). The data are extracted from LDA calculations. (c) Dissociation energy curve of \(\text{H}_{2}^{+}\) calculated by various DFAs. The energy of a hydrogen atom is set to zero.
Self-consistent implementation could only modestly relax the energy, without achieving the zero energy limit; in order to reach the zero energy limit, one has to go beyond the frozen orbital assumption in the design of curvature matrix and total energy correction formula.
One can also apply LOSC to other more accurate parent functionals. In this paper, we have achieved this by designing flexible forms of \(\kappa_{ij}\) on the basis of the curvature formula of LDA for many other types of DFAs, including the GGAs, the hybrids such as the B3LYP, the range-separated functionals such as the CAM-B3LYP, etc. These DFAs suffer from the delocalization error to different extents, while the LOSC gives similar corrected results; see for instance the H\({}_{2}^{+}\) dissociation curves calculated by LOSC-LDA and LOSC-B3LYP in Fig. 3(c). Moreover, the fact that LOSC-B3LYP energy at large \(R\) is almost perfect suggests that LOSC applied to a better parent functional leads to better results.
In a related effort, Anisimov and Kozhevnikov have developed a generalized transition state (GTS) method to improve the LDA calculation for band gaps of solids. Their suggested energy correction amounts to \(\Delta E^{\rm GTS}=\sum_{i}\frac{1}{2}\kappa_{ii}(\lambda_{ii}-\lambda_{ii}^{2})\), where each \(\kappa_{ii}\) is determined by a separate constrained LDA calculation. A similar scheme was recently constructed by Ma and Wang. In these works the LOs come from mixing of only occupied or virtual COs in the localization and their energy corrections, thus do not change the total energies for physical systems with integer number of electrons; hence these energy functionals are not _size consistent_ and can only correct orbital energies. These methods have only been implemented as post-DFT corrections rather than in a self-consistent manner, thus they rely on the qualitatively correct DFT densities to produce reasonable corrections. In contrast, the LOSC framework uses mixing of the occupied and virtual COs in the localization and offers an explicit form of Eq. for computing the \(\mathbf{\kappa}\)-matrix. It changes the DFA energies both at integer and fractional electron numbers. Moreover, Eq. involves off-diagonal \(\kappa_{ij}\) and \(\lambda_{ij}\) that are crucial because they dispel the unwanted interactions between LO pairs. In the case of H\({}_{2}^{+}\) dissociation, it is only with these off-diagonal terms that the correct asymptotic behavior as \(R\rightarrow\infty\) can be retrieved. Importantly, LOSC is a functional of the non-interacting density matrix and can be implemented self-consistently within the generalized Kohn-Sham approach.
_Self-consistent corrections to energy and electron density._ For practical calculations we have devised a self-consistent field (SCF) procedure, with which the LOSC approach improves \(E(N)\) and \(\rho(\mathbf{r})\) simultaneously.
\[\mathbf{\rho}_{s}^{\rm in}\stackrel{{\rm(I)}}{{\longrightarrow}}\mathbf{h }_{p}\stackrel{{\rm(II)}}{{\longrightarrow}}\{\varphi_{m}\} \stackrel{{\rm(III)}}{{\longrightarrow}}\{\phi_{i}\}\stackrel{{ \rm(IV)}}{{\longrightarrow}}\Delta\mathbf{h}\stackrel{{\rm(V)}}{{ \longrightarrow}}\mathbf{\rho}_{s}^{\rm out}. \tag{12}\]
In step (I) a projected KS (or generalized KS, GKS) Hamiltonian is constructed from the initial density matrix as \(\mathbf{h}_{p}=\mathbf{\rho}_{s}\mathbf{h}_{0}\mathbf{\rho}_{s}+(\mathbf{I}-\mathbf{\rho}_{s})\mathbf{h}_ {0}(\mathbf{I}-\mathbf{\rho}_{s})\), with \(\mathbf{h}_{0}=\frac{\delta E^{\rm DFA}}{\delta\mathbf{\rho}_{s}}\) being the KS/GKS Hamiltonian of the parent DFA. Here, the projections on \(\mathbf{h}_{0}\) using \(\mathbf{\rho}_{s}\) and \(\mathbf{I}-\mathbf{\rho}_{s}\) avoid overcorrecting the energies of compact molecules by rendering \(\lambda_{ii}\) close to integer 0 or 1.. In step (II) \(\mathbf{h}_{p}\) is diagonalized to generate the auxiliary COs. In step (III) the restrained Boys localization is carried out to obtain the LOs. In step (IV) the LOSC contribution to GKS Hamiltonian matrix is computed via \(\Delta\mathbf{h}=\frac{\delta\Delta E^{\rm LOSC}}{\delta\mathbf{\rho}_{s}}\). Finally in step (V) \(\mathbf{\rho}_{s}\) is updated by minimizing the total energy with the aid of approximate gradient \(\mathbf{h}_{0}+\Delta\mathbf{h}\). Steps (I)-(V) are iterated until the initial and final density matrices are equal. It is worth pointing out that although the LOSC involves orbitals, it remains a functional of the density matrix and its SCF implementation can be carried out within the GKS scheme (or the Hartree-Fock-Kohn-Sham scheme), because the Hamiltonian \(\mathbf{h}_{p}\) and the COs and LOs are all determined by \(\mathbf{\rho}_{s}\). Thus, \(\epsilon_{\rm HOMO}/\epsilon_{\rm LUMO}\) of the LOSC hamiltonian \(\mathbf{h}_{0}+\Delta\mathbf{h}\) is the chemical potential for electron removal/addition and the HOMO-LUMO gap is the derivative gap, which are theoretical predictions of the fundamental gap in both finite molecules and bulk.
The implementation of LOSC is very efficient, since the computation of pertinent quantities such as \(\mathbf{\lambda}\) and \(\mathbf{\kappa}\) are straightforward. The restrained Boys localization can be conducted efficiently using the Jacobi sweep approach. Consequently, the extra computational cost due to the LOSC procedure usually amounts to a small portion of the overall cost.
## III Results and Conclusion
In the following, we demonstrate that the LOSC approach generally alleviates the delocalization error associated with the mainstream DFAs, and thus curves many related problems in practical DFT calculations.
To start with, the size-dependent deviations between the calculated \(\epsilon_{\rm HOMO}\) and \(-I_{\rm ve}\) and between \(I_{\rm ve}\) and \(I_{\rm exp}\) are mostly eliminated by applying the LOSC, suggesting that LOSC achieves size-consistency; see Fig. 1(a). As indicated in Fig. 1(b), the LOSC largely straightens the \(E(N)\) curve between integers. Furthermore, the straightening of \(E(N)\) curve is achieved not only for electron removal (related to HOMO prediction), but also for electron addition (related to LUMO prediction); not only for compact systems, but also for dissociating molecules such as He\({}_{2}^{+}\). In the latter case, when the parent DFA predicts wrong integer energy, merely straightening the \(E(N)\) curve does not cure the problem. In such case, the LOSC not only straightens the curve, but also shifts the integer energy so as to point to the right slope, see for instance the \(E(N)\) curve connecting He\({}_{2}\) and He\({}_{2}^{+}\) at \(R=5\)A.
Besides the H\({}_{2}^{+}\) shown in Fig. 2(c), the LOSC also systematically improves the dissociation behavior of many other molecular cation species, such as the He\({}_{2}^{+}\), the water dimer cation, and the benzene dimer cation. This suggests that the LOSC does not deteriorate with system size.
The LOSC systematically improves the prediction of \(\epsilon_{\text{HOMO}}\), \(\epsilon_{\text{LUMO}}\) and thus the fundamental gaps for systems of all sizes, ranging from atoms and molecules to polymers such as polyacenes and trans-polyacetylenes; see In contrast to GSC which has no effect (and LSC which has numerical difficulty) on band gaps of bulk, the LOSC gives a promising improvement as it has a nonzero correction on polymers in the extrapolated infinite chain length limit.
In addition to \(\epsilon_{\text{HOMO}}\) and \(\epsilon_{\text{LUMO}}\), the LOSC also corrects the energies of other KS orbitals via Eq.. As depicted in Fig. 4(d), the orbital energies of pentacene predicted by the LOSC-LDA agree accurately with the peak positions in experimental photoemission spectrum. This feature will be further studied in our subsequent papers.
The y axes in (a)-(c) stand for \(\epsilon_{\text{HOMO}}\) (\(\epsilon_{\text{LUMO}}\)) in unit of eV, and in (d) stands for density of states (DOS) in atomic unit. In (a)-(c), \(\epsilon_{\text{HOMO}}\) (\(\epsilon_{\text{LUMO}}\)) are represented by the solid (hollow) symbols. The plot details are as follows. (a) Calculated \(\epsilon_{\text{HOMO}}\) (\(\epsilon_{\text{LUMO}}\)) of LDA and LOSC-LDA for 82 atoms and molecules in a modified G2–97 set, in comparison with the reference \(-I_{\text{ve}}\) (\(-A_{\text{ve}}\)). The mean absolute error (MAE) is 4.43 (3.76) eV with the LDA, and 0.50 (0.50) eV with the LOSC–LDA. (b) Calculated \(\epsilon_{\text{HOMO}}\) versus \(-I_{\text{exp}}\) (\(\epsilon_{\text{LUMO}}\) versus \(-A_{\text{exp}}\)) for trans-polyacetylene oligomers. The reference experimental values and the second order restrictive active space perturbation theory (RASPT2) calculated values are obtained from Refs.. (c) Calculated \(\epsilon_{\text{HOMO}}\) versus \(-I_{\text{exp}}\) (\(\epsilon_{\text{LUMO}}\) versus \(-A_{\text{exp}}\)) for polyacene oligomers. The green lines in (a) and (c) are guide to the eyes. The reference experimental data are given by Refs.. (d) DOS spectrum of pentacene in comparison with experimental photoemission spectrum. All peaks in the calculated DOS and the peak at -1.4eV in the experimental spectrum are broadened by Gaussian functions for clarity.
Finally, the LOSC also corrects the wrong electron density caused by the delocalization error. A typical example is a solvated Cl anion. As shown in Fig. 5, the PBE functional erroneously predicts that the excess electron delocalizes over several water molecules, while the LOSC-PBE yields the correct distribution that the excess electron mostly localizes on the Cl atom.
The examples shown demonstrate that the LOSC approach remedies a wide range of problems caused by the delocalization error of DFAs. The LOSC functional is very different from conventional density functional constructions, which are explicit analytical functionals of the density, the density gradients or the KS reduced density matrix. The LOSC framework utilizes many other information, such as the localized orbitals, local occupation matrix and parent DFA reference spectrum, which themselves are implicit functionals of \(\rho_{s}\). This opens up a lot more possibilities in the exploration of the exact functional within the functional space. As a first effort, the LOSC functional presented in this paper addresses the size-consistency problem and exhibits a systematic elimination of delocalization error in all aspects. With a more sophisticated localization procedure and a more complete form of \(\Delta E^{\text{LOSC}}\) addressing local fractional spins, it is possible to eliminate other intrinsic errors of the mainstream functionals, and systematically improve the density functional approximation to a greater extent.
Support from the National Science Foundation (CHE-1362927) (CL), the Ministry of Science and Technology of China (Grants No. 2016YFA0400900 and No. 2016YFA0200600) (XZ), the National Natural Science Foundation of China (Grants No. 21573202 and No. 21233007) (XZ), the Strategic Priority Research Program (B) of the Chinese Academy of Sciences (XDB01020000) (XZ), the Fundamental Research Funds for Chinese Central Universities (Grant No. 234000074) (XZ), and the Center for Computational Design of Functional Layered Materials (Award DE-SC0012575), an Energy Frontier Research Center funded by the US Department of Energy, Office of Science, Basic Energy Sciences. (NQS, WY) is appreciated. Discussions with Dadi Zhang, Prof. Jianfeng Lu and Dr. Tomasz Janowski were helpful.
|
10.48550/arXiv.1707.00856
|
Localized Orbital Scaling Correction for Systematic Elimination of Delocalization Error in Density Functional Approximations
|
Chen Li, Xiao Zheng, Neil Qiang Su, Weitao Yang
| 6,316
|
10.48550_arXiv.1511.04114
|
###### Abstract
We present models for a heteronuclear diatomic molecular ion in a linear Paul trap in a rigid-rotor approximation, one purely classical, the other where the center-of-mass motion is treated classically while rotational motion is quantized. We study the rotational dynamics and their influence on the motion of the center-of-mass, in the presence of the coupling between the permanent dipole moment of the ion and the trapping electric field. We show that the presence of the permanent dipole moment affects the trajectory of the ion, and that it departs from the Mathieu equation solution found for atomic ions. For the case of quantum rotations, we also evidence the effect of the above-mentioned coupling on the rotational states of the ion.
Laser-cooled or sympathetically-cooled atomic and molecular ions have been in the spotlight in recent years, due to their great potential in the study and development of many fields, such as chemical reactions, high-precision spectroscopy, atomic optics, and in many other fast developing fields related to quantum information and quantum computing, where the manipulation of the internal states of these ions has become possible.
One of the most applied techniques to trap these cold ions is known as the Paul trap, using time-dependent radio-frequency electric fields. Paul traps, specifically those with a linear configuration, consist of electrodes designed to surround an inner space where the trapping of a single charged particle, as well as the simultaneous trapping a string of charged particles, occurs. This makes them highly suitable to be used in experiments involving laser control of the ions, since the spacing between the electrodes as well as the spacing between the ions provide a good environment for this purpose. In this regard, large scale studies of trapped ions in large scales, in the form of Coulomb crystals, have become feasible, with promising developments for large-scale quantum simulations and quantum computations.
While there have been many experimental studies using molecular ions, on the theoretical side there have been few attempts to treat them differently than atomic ions, at least with respect to their interaction with the trapping electric field. In particular, the rotation of the molecular ions has been considered mostly with respect to a coupling with an external laser field.
In this paper we study the dynamical behavior of a rigid, heteronuclear diatomic molecular ion trapped inside a linear Paul trap. We investigate the coupling of the permanent dipole moment of the molecular ion with the trapping electric field and its effect on the rotational dynamics of the ion and also on the center-of-mass motion of the ion. We develop classical and semi-classical models of the molecular ion, where rotation is treated classically and quantum mechanically, respectively. We carry out simulations of the motion of the center of mass, including the effects of the interaction between the trapping electric field and the permanent dipole moment of the molecular ion, by solving the classical equations of motion. Classical rotation is treated using the "body vector method," while a basis set of spherical harmonics is used to describe the quantum rotational wave function.
This paper is organised as follows. First, we introduce our model for a rigid diatomic molecular ion trapped inside a linear Paul trap (Sec. II). The numerical methods for the classical and semi-classical simulations are described in Sec. III. The results obtained from both classical and semi-classical methods are presented in Sec. IV.1 and Sec. IV.2, respectively. Finally, concluding remarks are given in Sec. V.
## II Hamiltonian of a rigid diatomic molecular ion in a linear Paul trap
### Quantum-mechanical Hamiltonian
We consider a diatomic molecular ion inside a linear Paul trap, combining a quadrupole, time-dependent, radio-frequency electric potential and a static electric potential,
\[\Phi(t)=\Phi_{\mathrm{rf}}(t)+\Phi_{\mathrm{s}}\,, \tag{1}\]
in which, for an ion located at position \((X,Y,Z)\),
\[\Phi_{\mathrm{rf}}(t) =\frac{1}{2}\frac{V_{0}}{r_{0}^{2}}\left(X^{2}-Y^{2}\right)\cos \Omega t\,, \tag{2a}\] \[\Phi_{\mathrm{s}} =\frac{\kappa U_{0}}{z_{0}^{2}}\left[Z^{2}-\frac{1}{2}\left(X^{2 }+Y^{2}\right)\right]\,, \tag{2b}\]
The Hamiltonian of a free diatomic molecular ion in a _rigid rotor_ model, where the nuclear vibrations of the ion are ignored, has the form
\[\hat{H}=-\frac{\hbar^{2}}{2M}\nabla_{\mathbf{R}}^{2}+\frac{B\hat{J}^{2}}{\hbar ^{2}}\,, \tag{3}\]
The derivatives in the first term are with respect to the position of the center of the mass of the ion, \(\mathbf{R}\).
\[B=\frac{\hbar^{2}}{2I}, \tag{4}\]where \(I=m_{\rm red}r^{2}\) is the moment of inertia, with \(m_{\rm red}\) and \(r\) the reduced mass and internuclear distance, respectively. The rigid rotor approximation we are using implies that the latter is fixed, i.e., \(r={\rm const}\).
In addition to having a non-zero electric charge, we take the diatomic molecular ion to be heteronuclear, which means that it possesses a permanent dipole moment that will also interact with the electric field of the Paul trap.
\[\hat{H}=-\frac{\hbar^{2}}{2M}\nabla_{\bf R}^{2}+\frac{B\hat{J}^{2}}{\hbar^{2}} +Ze\Phi({\bf R},t)-\mathbf{\mu}\cdot{\bf E}({\bf R},t)\,, \tag{5}\]
### Classical Hamiltonian
Classically, we approximate the diatomic molecular ion inside a trapping potential as a dipolar rigid rotor, consisting of two masses \(m_{-}\) and \(m_{+}\) with partial charges \(\delta_{-}\) and \(\delta_{+}\), respectively, kept spatially separated at a constant distance.
\[H=K+V=\frac{1}{2}M\dot{{\bf R}}^{2}+\frac{1}{2}I\dot{\mathbf{\omega}} ^{2}+Ze\Phi({\bf R},t)-\mathbf{\mu}\cdot{\bf E}({\bf R},t)\,, \tag{6}\]
The corresponding classical equation of motion for the center of mass is
\[M\ddot{{\bf R}}+Ze\mathbf{\nabla}\Phi({\bf R},t)-\mathbf{\nabla }\left[\mathbf{\mu}\cdot{\bf E}({\bf R},t)\right]=0\,, \tag{7}\]
As shown in many previous works, the equations of motion of an atomic ion inside a Paul trap obey the Mathieu equation
\[\frac{{\rm d}^{2}{\bf R}}{{\rm d}\tau^{2}}+(a-2q\cos 2\tau){\bf R}=0\,. \tag{8}\]The Mathieu equation has stable solutions and hence bounded trajectories depending on whether its parameters (\(a\) and \(q\)) fall within a certain range - called stability regions - in the \(a-q\) plane or not. For an ion inside a Paul trap, \(a\) and \(q\) depend on the mass of the ion, the magnitude of the field, and the parameters of the trap. For an ion in a linear Paul trap, corresponding to the potential in Eqs.
\[a_{x}=a_{y} =-\frac{4Ze}{M\Omega^{2}}\frac{\kappa U_{0}}{z_{0}^{2}}, \tag{9a}\] \[a_{z} =\frac{8Ze}{M\Omega^{2}}\frac{\kappa U_{0}}{z_{0}^{2}},\] (9b) \[q_{x}=-q_{y} =-\frac{2Ze}{M\Omega^{2}}\frac{V_{0}}{r_{0}^{2}},\] (9c) \[q_{z} =0\,. \tag{9d}\]
Even though the equation of motion for a diatomic molecular ion, Eq., does not reduce to a Mathieu equation, taking the term \(\mathbf{\nabla}\left[\mathbf{\mu}\cdot\mathbf{E}(\mathbf{R},t)\right]\) as a small perturbation (we will investigate this below), one can still use the stability criteria of the Mathieu equation as a guideline for the stability of the trap for a molecular ion.
## III Numerical Methods
### Classical approach
The classical representation of the rigid rotor, Hamiltonian, has five degrees of freedom: the three Cartesian coordinates of the center of mass, along with two angles positioning the internuclear axis of the diatomic ion in space, see We consider three different coordinate systems: the laboratory-fixed system \((X,Y,Z)\), the COM (center-of-mass) system \((x,y,z)\), which is taken to be parallel to the laboratory-fixed frame, but centered on the center of mass of the molecular ion (see also Appendix A), and the body-fixed system \((x^{\prime},y^{\prime},z^{\prime})\), with its origin at the center of mass, but with axis \(z^{\prime}\) always co-linear with the internuclear axis (and hence with the dipole moment \(\mathbf{\mu}\)).
Expanding the dipole moment vector into its components in the COM frame \((x,y,z)\), we get
\[\mathbf{\mu}=\mu_{0}[\sin\theta\cos\varphi\mathbf{\hat{x}}+\sin\theta\sin\varphi \mathbf{\hat{y}}+\cos\theta\mathbf{\hat{z}}]\,, \tag{10}\]
1).
\[\mathbf{E}(\mathbf{R},t)=-\nabla\Phi(\mathbf{R},t)=\left[\frac{\kappa U_{0}}{z_ {0}^{2}}-\frac{V_{0}}{r_{0}^{2}}\cos\Omega t\right]X\mathbf{\hat{x}}+\left[ \frac{\kappa U_{0}}{z_{0}^{2}}+\frac{V_{0}}{r_{0}^{2}}\cos\Omega t\right]Y \mathbf{\hat{y}}-\frac{2\kappa U_{0}}{z_{0}^{2}}Z\mathbf{\hat{z}}\,, \tag{11}\]
and. The coupling term of Eq.
\[\begin{split}\mathbf{\mu}\cdot\mathbf{E}&=\left[\frac{ \kappa U_{0}\mu_{0}\sin\theta\cos\varphi}{z_{0}^{2}}-\frac{V_{0}\mu_{0}\sin \theta\cos\varphi}{r_{0}^{2}}\cos\Omega t\right]X\\ &\quad+\left[\frac{\kappa U_{0}\mu_{0}\sin\theta\sin\varphi}{z_{0} ^{2}}+\frac{V_{0}\mu_{0}\sin\theta\sin\varphi}{r_{0}^{2}}\cos\Omega t\right] Y\\ &\quad-\left[\frac{2\kappa U_{0}\mu_{0}}{z_{0}^{2}}\cos\theta \right]Z\,.\end{split} \tag{12}\]
It is clearly evident from this equation that the interaction between the dipole moment and the electric field has not only a spatial dependence, but also is dependent on the orientation
(Color online) The orientation of the permanent dipole moment \(\mathbf{\mu}\) of a diatomic molecule in space with respect to the center-of-mass (COM) frame \((x,y,z)\), which is indicated by angles \(\varphi\) and \(\theta\). The body-fixed frame \((x^{\prime},y^{\prime},z^{\prime})\) shows the principal axes (or symmetry axes) of this molecule. The orientation of the dipole moment is fixed along \(z^{\prime}\). The COM frame is aligned with the laboratory frame \((X,Y,Z)\), but has its origin at the center-of-mass of the molecule.
A direct numerical simulation of the equations of motion corresponding to the Hamiltonian is impractical, as these equations contain a term in \(1/\sin^{2}\theta\), and thus are singular at \(\theta=0\).
\[I=\begin{pmatrix}I_{x^{\prime}}&0&0\\ 0&I_{y^{\prime}}&0\\ 0&0&0\end{pmatrix} \tag{13}\]
Following the body vector method, we introduce unit vectors along these axes, namely \(\mathbf{\hat{x}}^{\prime}\), \(\mathbf{\hat{y}}^{\prime}\), \(\mathbf{\hat{z}}^{\prime}\).
\[\mathbf{\hat{u}}^{\prime}=\left(u^{\prime}_{x},u^{\prime}_{y},u^{\prime}_{z} \right). \tag{14}\]
Setting the internuclear axis (or \(\mathbf{\mu}\)) along the \(z^{\prime}\) axis, the unit vector \(\mathbf{\hat{z}}^{\prime}\) simply gives the direction of the dipole moment in space. These vector are similar to the variables described by Cheung to avoid singularities in the equations of motion for diatomic molecules.
\[\mathbf{\mu}=\mu_{0}\mathbf{\hat{z}}^{\prime}\,, \tag{15}\]
in which the components of \(\mathbf{\hat{z}}^{\prime}\) can be written in terms of \(\theta\) and \(\varphi\)
\[z^{\prime}_{x} =\sin\theta\cos\varphi\,,\] \[z^{\prime}_{y} =\sin\theta\sin\varphi\,, \tag{16}\] \[z^{\prime}_{z} =\cos\theta\,,\]
and
We can now simply rewrite Eq. using the new representation of the dipole moment, with the interaction term in the Hamiltonian not being explicitly dependent on the angles\((\theta,\varphi)\).
\[\hat{\mathbf{u}}^{\prime}=\mathbf{\omega}_{P}\times\hat{\mathbf{u}}^{\prime}\,, \tag{17}\]
To write the equations of motion, we need to calculate the torque in the lab-fixed frame, \(\mathbf{T}_{L}=\mathbf{\mu}\times\mathbf{E}\), which is imposed on the dipole moment by the external electric field. To calculate the time evolution of the dipole moment, Eq., we need the principal torque \(\mathbf{T}_{P}\), which is obtained from the relation
\[\mathbf{T}_{L}=\mathbf{A}^{T}\mathbf{T}_{P}\,, \tag{18}\]
where \(\mathbf{A}^{T}\) is the transpose of the rotation matrix whose columns are the principal axes unit vectors
\[\mathbf{A}^{T}=(\hat{\mathbf{x}}^{\prime},\hat{\mathbf{y}}^{\prime},\hat{ \mathbf{z}}^{\prime})\equiv\begin{pmatrix}x^{\prime}_{x}&y^{\prime}_{x}&z^{ \prime}_{x}\\ x^{\prime}_{y}&y^{\prime}_{y}&z^{\prime}_{y}\\ x^{\prime}_{z}&y^{\prime}_{z}&z^{\prime}_{z}\end{pmatrix}\,. \tag{19}\]
As there is no rotation along the internuclear axis, \(\omega_{Pz^{\prime}}=0\) and there is no torque in the \(z_{P}\) direction [see Eq.], \(T_{Pz^{\prime}}=0\), the equations of motion for rotation are written as
\[I_{x^{\prime}}\dot{\omega}_{Px^{\prime}} =T_{Px^{\prime}}+\omega_{Py^{\prime}}\omega_{Pz^{\prime}}(I_{y^{ \prime}}-I_{z^{\prime}})=T_{Px^{\prime}}\] \[I_{y^{\prime}}\dot{\omega}_{Py^{\prime}} =T_{Py^{\prime}}+\omega_{Pz^{\prime}}\omega_{Px^{\prime}}(I_{z^{ \prime}}-I_{x^{\prime}})=T_{Py^{\prime}} \tag{20}\] \[I_{z^{\prime}}\dot{\omega}_{Pz^{\prime}} =0\,.\]
Finally, the equations which describe the dynamical behavior of the ion inside the linear Paul trap, classically, are
\[\frac{d\mathbf{R}}{dt} =\frac{\mathbf{P}}{M}\,, \tag{21a}\] \[\frac{dP_{X}}{dt} =e\left[b_{1}-b_{2}\cos\Omega t\right]X+\mu_{0}\left[b_{1}-b_{2} \cos\Omega t\right]z^{\prime}_{x}\,,\] (21b) \[\frac{dP_{Y}}{dt} =e\left[b_{1}+b_{2}\cos\Omega t\right]Y+\mu_{0}\left[b_{1}+b_{2} \cos\Omega t\right]z^{\prime}_{y}\,,\] (21c) \[\frac{dP_{Z}}{dt} =-2eb_{1}Z-2\mu_{0}b_{1}z^{\prime}_{z}\,,\] (21d) \[\frac{d\hat{\mathbf{u}}^{\prime}}{dt} =\mathbf{\omega}_{P}\cdot\hat{\mathbf{u}}^{\prime}\,,\text{ where } \hat{\mathbf{u}}^{\prime}\in\{\hat{\mathbf{x}}^{\prime},\hat{\mathbf{z}}^{\prime}\} \tag{21e}\](since only two coordinates are necessary to locate the dipole in space, we obtain \(\mathbf{\hat{y}}^{\prime}\) from \(\mathbf{\hat{y}}^{\prime}=\mathbf{\hat{z}}^{\prime}\times\mathbf{\hat{x}}^{\prime}\)),
\[\frac{d\mathbf{\omega}_{P}}{dt}=\frac{1}{I}\mathbf{T}_{P}\ =\frac{1}{I}(\mathbf{A}^{ \mathrm{T}})^{-1}\mathbf{T}_{L}\,, \tag{21f}\]
with
\[T_{Lx} =\mu_{0}\left[-2b_{1}Zz_{y}^{\prime}-(b_{1}+b_{2}\cos\Omega t)Yz_ {z}^{\prime}\right]\,, \tag{22a}\] \[T_{Ly} =\mu_{0}\left[2b_{1}Zz_{x}^{\prime}+(b_{1}-b_{2}\cos\Omega t)Xz_ {z}^{\prime}\right]\,,\] (22b) \[T_{Lz} =\mu_{0}\left[(b_{1}+b_{2}\cos\Omega t)Yz_{x}^{\prime}-(b_{1}-b_{ 2}\cos\Omega t)Xz_{y}^{\prime}\right]\,, \tag{22c}\]
where we have defined
\[b_{1}\equiv\frac{\kappa U_{0}}{z_{0}^{2}}\,,\quad b_{2}\equiv\frac{V_{0}}{r_{0 }^{2}}\,. \tag{23}\]
### Semi-classical approach
A direct simulation using the full quantum Hamiltonian would be quite numerically intensive, owing to the five degrees of freedom, the large amplitude of the motion in the trap, and the impossibility of separating the problem along Cartesian coordinates (in contrast to the case for atomic ions) due to the coupling of the rotation with the trapping potential.
\[i\hbar\frac{d}{dt}\Psi^{\mathrm{rot}}(\theta,\varphi;t)=\left[\frac{B}{\hbar} \tilde{J}^{2}-\mathbf{\mu}\cdot\mathbf{E}(\mathbf{R},t)\right]\Psi^{\mathrm{rot}} (\theta,\varphi;t)\,, \tag{24}\]
below.
The wave function can be expanded on the basis of field-free rotor states (spherical harmonics) \(Y_{J,M}(\theta,\varphi)\) as
\[\Psi^{\mathrm{rot}}(\theta,\varphi;t)=\sum_{J,M}D_{J,M}(t)Y_{J,M}(\theta, \varphi)\,, \tag{25}\]
Solving the time-dependent Schrodinger equation using the expression of the rotational wave function in Eq.,
\[i\hbar\sum_{J,M}\dot{D}_{J,M}(t)=\sum_{J,M}D_{J,M}(t)BJ(J+1)-\sum_{J,M}\sum_{J^{ \prime},M^{\prime}}D_{J^{\prime},M^{\prime}}(t)\left\langle J,M|\mathbf{\mu}\cdot \mathbf{E}|J^{\prime},M^{\prime}\right\rangle\,, \tag{26}\]
results in a set of coupled equations for the time evolution of the coefficients,
\[\dot{D}_{J,M}(t)=-\frac{i}{\hbar}D_{J,M}(t)BJ(J+1)+\frac{i}{\hbar}\sum_{J^{ \prime},M^{\prime}}D_{J^{\prime},M^{\prime}}(t)\left\langle J,M|\mathbf{\mu}\cdot \mathbf{E}|J^{\prime},M^{\prime}\right\rangle\,, \tag{27}\]
Using Eq.
\[\mathbf{\mu}\cdot\mathbf{E}=\sqrt{\frac{2\pi}{3}}\mu_{0}\left[(E_{x}+iE_{y})Y_{1,- 1}+\sqrt{2}E_{z}Y_{1,0}+(-E_{x}+iE_{y})Y_{1,1}\right] \tag{28}\]
and hence
\[\left\langle J,M|\mathbf{\mu}\cdot\mathbf{E}|J^{\prime},M^{\prime}\right\rangle =\sqrt{\frac{2\pi}{3}}\mu_{0}\left[(E_{x}+iE_{y})\left\langle J,M |Y_{1,-1}|J^{\prime},M^{\prime}\right\rangle+\sqrt{2}E_{z}\left\langle J,M|Y_{1,0}|J^{\prime},M^{\prime}\right\rangle\right.\] \[\left.+\left(-E_{x}+iE_{y}\right)\left\langle J,M|Y_{1,1}|J^{ \prime},M^{\prime}\right\rangle\right]\,, \tag{29}\]
Using Clebsch-Gordan coefficients and their related 3j-symbols, we have
\[\left\langle J,M|Y_{1,n}|J^{\prime},M^{\prime}\right\rangle =\int_{0}^{\pi}\int_{0}^{2\pi}Y_{J,M}^{*}(\theta,\varphi)Y_{1,n}( \theta,\varphi)Y_{J^{\prime},M^{\prime}}(\theta,\varphi)\sin\theta d\theta d\varphi\] \[=(-1)^{n}\sqrt{\left[\frac{(2J+1)(2J^{\prime}+1)}{4\pi}\right] \begin{pmatrix}J&1&J^{\prime}\\ 0&0&0\end{pmatrix}\begin{pmatrix}J&1&J^{\prime}\\ -M&n&M^{\prime}\end{pmatrix}}\] \[=\sqrt{\left[\frac{(2J^{\prime}+1)}{4\pi(2J+1)}\right]}C_{J^{ \prime}010}^{J0}C_{J^{\prime}M^{\prime}1n}^{JM}\,, \tag{30}\]
The selection rules for non-zero values of the the Clebsch-Gordan coefficients in Eq.
\[\Delta J \equiv|J-J^{\prime}|=0,\pm 1\;\;(J=0\not\leftrightarrow 0) \tag{31a}\] \[\Delta M \equiv|M-M^{\prime}|=0,\pm 1\,. \tag{31b}\]
The classical equations for the motion of the center of mass were given in the previous section, Eqs. (21a)-(21d), and they involve the orientation of the dipole through the vector \(\hat{\mathbf{z}}^{\prime}\), Eq.. As \(\theta\) and \(\varphi\) are not, in the semi-classical approach, classical variables, we need to modify the above treatment for the center-of-mass motion.
\[z_{x}^{\prime} \equiv\left\langle\sin\theta\cos\varphi\right\rangle(t)=\sum_{J,M} \sum_{J^{\prime},M^{\prime}}D_{J,M}^{*}(t)D_{J^{\prime},M^{\prime}}(t)\left\langle J,M|\sin\theta\cos\varphi|J^{\prime},M^{\prime}\right\rangle\,, \tag{32a}\] \[z_{y}^{\prime} \equiv\left\langle\sin\theta\sin\varphi\right\rangle(t)=\sum_{J,M} \sum_{J^{\prime},M^{\prime}}D_{J,M}^{*}(t)D_{J^{\prime},M^{\prime}}(t)\left\langle J,M|\sin\theta\sin\varphi|J^{\prime},M^{\prime}\right\rangle\,,\] (32b) \[z_{z}^{\prime} \equiv\left\langle\cos\theta\right\rangle(t)=\sum_{J,M}\sum_{J^{ \prime},M^{\prime}}D_{J,M}^{*}(t)D_{J^{\prime},M^{\prime}}(t)\left\langle J,M |\cos\theta|J^{\prime},M^{\prime}\right\rangle\,, \tag{32c}\]
Finally, the time evolution of a rigid-rotor diatomic molecular ion trapped in a linear Paul trap, in the semi-classical model, is obtained from the coupled differential equations
\[\frac{d\mathbf{R}}{dt} =\frac{\mathbf{P}}{M}\,, \tag{33a}\] \[\frac{dP_{X}}{dt} =e\left[b_{1}-b_{2}\cos\Omega t\right]X+\mu_{0}\left[b_{1}-b_{2} \cos\Omega t\right]\left\langle\sin\theta\cos\varphi\right\rangle(t)\,,\] (33b) \[\frac{dP_{Y}}{dt} =e\left[b_{1}+b_{2}\cos\Omega t\right]Y+\mu_{0}\left[b_{1}+b_{2} \cos\Omega t\right]\left\langle\sin\theta\sin\varphi\right\rangle(t)\,,\] (33c) \[\frac{dP_{Z}}{dt} =-2eb_{1}Z-2\mu_{0}b_{1}\left\langle\cos\theta\right\rangle(t)\,,\] (33d) \[\frac{dD_{J,M}}{dt} =-\frac{i}{\hbar}D_{J,M}(t)BJ(J+1)+\frac{i}{\hbar}\sum_{J^{\prime },M^{\prime}}D_{J^{\prime},M^{\prime}}(t)\left\langle J,M|\mathbf{\mu}\cdot \mathbf{E}|J^{\prime},M^{\prime}\right\rangle\,, \tag{33e}\]
Note that this is different from the result that would be obtained by a simple classical approximation of the momentum term in Hamiltonian, as this would omit the effect of the coupling term \(\mathbf{\mu}\cdot\mathbf{E}\) on the center-of-mass motion.
### Numerical implementation
To solve the equations of motion, both for the classical and the semi-classical models, we use a 4th-order Runge-Kutta-Fehlberg integrator. The time step is taken to be equal to \(\Delta t=1\times 10^{-14}\,\mathrm{s}\). In the classical simulations, this choice of time step ensures that the unit vector \(\mathbf{\hat{u}}^{\prime}\) remains normalised at each time step.
## IV Results
In both classical and semi-classical simulations we have used MgH\({}^{+}\) ion as an example. The magnitude of the permanent dipole moment of the ion, \(\mu_{0}\), as well as its rotational constant, \(B\), the mass of each of the consisting atoms, \(m_{\rm H}\) and \(m_{\rm Mg}\), and the internuclear distance \(r\) are all listed in Tab. 1. We use one set of trap parameters for all simulations, except for the amplitudes of the static and radio-frequency electric potentials \(U_{0}\) and \(V_{0}\), that may vary in different numerical calculations. The value of the fixed parameters are given in Tab. 2 and are based on typical experimental realizations. With these parameters, we can calculate the stability regions of the Mathieu equation, using a point-charge model, without the interaction of the dipole moment with the trap (as discussed in Sec. II.2).
We ran our simulations for three different pairs of trapping potentials (\(U_{0},V_{0}\)) equal to (\(680\,{\rm V},1650\,{\rm V}\)), (\(2175\,{\rm V},2000\,{\rm V}\)) and (\(7080\,{\rm V},3000\,{\rm V}\)), resulting in values of \(a\) and \(q\), Eqs., which are located inside the second stability region of the Mathieu equation for
\begin{table}
\begin{tabular}{c c} \hline \hline Parameter & Value \\ \hline \(r_{0}\) & \(0.769\times 10^{-3}\,{\rm m}\) \\ \(z_{0}\) & \(1.25\times 10^{-3}\,{\rm m}\) \\ \(\kappa\) & \(0.31\) \\ \(\Omega\) & \(2\pi\times 8\times 10^{6}\,{\rm s}^{-1}\) \\ \hline \hline \end{tabular}
\end{table}
Table 2: Trap configuration parameters for a MgH\({}^{+}\) ion.
The I and II stability regions corresponding to a MgH\({}^{+}\) molecular ion in a linear Paul trap are shown in Fig. 2, as calculated from the characteristic values \(a_{0}\), \(b_{1}\), etc., of the Mathieu equation. Note that these stability regions hold along both \(x\) and \(y\), as they depend on \(a\) and \(|q|\), while the trapping along \(z\) is unconditionally stable.
We chose our parameters from the second stability region since it provides trapping potentials which are large enough to magnify the effect of the coupling between the dipole moment and the trapping field. However it needs to be mentioned that we also ran simulations for much smaller trapping potentials [for example \((1820\,\mathrm{V},500\,\mathrm{V})\)], in the first stability region.
(Color online) I and II stability regions of the Mathieu equation, corresponding to a MgH\({}^{+}\) molecular ion, in a point-charge model. Shaded areas correspond to stable trajectories, with the borders delimited by the characteristic values of the Mathieu equation.
In our simulations, the ion is located initially (\(t=0\)) at the center of the trap. However as we mentioned in the previous paragraph, choosing an initial position other than the center of the trap gives the same qualitative results. We also assign an initial linear momentum (in both methods) and an initial angular momentum to the ion (in the classical method), approximating an initial temperature. The latter is chosen based on experimental works on trapping MgH\({}^{+}\) ions. For example, MgH\({}^{+}\) ions have been cooled translationally down to temperature of \(\sim 10\,\)mK by Doppler laser cooling where ions form a Coulomb crystal, or temperatures of \(\sim 10\,\)K are obtained where they are cooled rotationally. In this work, we have assigned the same initial temperature to both classical rotational and translational motion of the ion, equivalent to \(20\,\)K.
### Classical trajectories of the center-of-mass motion of the ion
In the classical simulations, in addition to assigning an initial position and initial linear and angular momenta to the ion, we have also assigned an initial orientation of the ion inside the linear Paul trap with respect to the COM frame, whose \(z\)-axis is considered to be parallel to the elongated part of the trapping device, in correspondence with Eq. (see also Fig. 1). Consequently, we will consider three different initial orientations of the dipole moment (internuclear axis): i) along the \(z\)-axis (\(\theta=0^{\circ}\)); ii) in the \(x-y\) plane (\(\theta=90^{\circ},\varphi=45^{\circ}\)); iii) along an arbitrarily-chosen orientation in space (\(\theta=45^{\circ},\varphi=30^{\circ}\)). These three orientations are indicated by the angles \(\theta\) and \(\varphi\) with respect to the COM frame, see Let us note that the simulation requires two vectors, \(\mathbf{\hat{x}}^{\prime}\) and \(\mathbf{\hat{z}}^{\prime}\), see Eqs., while \(\theta\) and \(\varphi\) only fix \(\mathbf{\hat{z}}^{\prime}\). We therefore take an arbitrary initial orientation for \(\mathbf{\hat{x}}^{\prime}\), orthogonal to \(\mathbf{\hat{z}}^{\prime}\). We ran the classical simulations for a total time of \(5\,\)us and for three different pairs of trapping potentials mentioned above.
The trajectories of the center-of-mass motion of the ion for two cases, when there is an interaction between the dipole moment and the electric field (\(\mu\neq 0\)) and when this interaction is ignored (\(\mu=0\)), are shown for all three components of the motion in For these simulations, \((U_{0},V_{0})\) are equal to \((680\,\mathrm{V},1650\,\mathrm{V})\) and the diatomic ion is initially oriented in space with spherical angles \(\theta=45^{\circ}\) and \(\varphi=30^{\circ}\). As expected from Eqs., one sees that in the presence of a permanent dipole moment, the trajectory of the center of mass does not correspond to the Mathieu equation, which is obtained for \(\mu=0\). For such a strength of the trapping field and initial orientation, the deviation is significant enough to be observable in the \(x\) direction. However, in the \(y\) and \(z\) directions the difference between these trajectories is very small compared to the actual amplitudes of the motion. In order to magnify the deviation in the trajectories, especially in the \(y\) and \(z\) directions, we have plotted the difference between the trajectories with and without the dipole moment in Figs. 3(b), (d), and (f). In the \(z\) direction, this difference is small enough to be assumed equal to zero. This arises from the fact that the equation of motion in the \(z\) direction, Eq. (21d), lacks the term containing the radio-frequency field, resulting only in a fast-oscillating term (due to the time evolution of \(z_{z}^{\prime}\)) that averages out to zero. In all cases, the presence of the coupling between the dipole moment and the trapping field does not affect the stability of ion's trajectory in the trap, which remains bounded and even displays the same oscillation period as when there is no coupling.
Increasing the magnitude of the fields \((U_{0},V_{0})\) increases the magnitude of the interaction energy between the dipole moment and the electric field. Results for the previous field (\(U_{0}=680\,\mathrm{V},V_{0}=1650\,\mathrm{V}\)) are compared to stronger fields in Fig. 4, where we plot the difference in trajectories with and without a dipole moment, only for the \(X\) component of motion. To increase visibility, we have only plotted a part of the entire trajectory, in the time interval \([$2.5\,\mathrm{\SIUnitSymbolMicro s}$,$3.0\,\mathrm{\SIUnitSymbolMicro s}$]\). We see in that, as expected, the deviation of the trajectories increases with the magnitude of the trapping fields.
We have also investigated the role of the initial orientation of the ion, by changing the initial values of \(\theta\) and \(\varphi\), as mentioned above. The result is also shown in Fig. 4, where we see a strong dependence of the trajectory on this initial orientation. It is interesting to note that, nevertheless, the deviations in the trajectory show the same periodicity and have the same zero crossings. It appears that the orientation of the dipole significantly changes the trajectory only at the turning points of the latter. As expected, these turning points are also where the rotations are the most hindered by the dipole-field coupling, as evidenced by (Color online) Time evolution of the center-of-mass motion of a MgH\({}^{+}\) ion inside a linear Paul trap, with classical rotation: (a)–(b) \(X\) component of the motion, (c)–(d) \(Y\) component of the motion, and (e)–(f) \(Z\) component of the motion. In panels (a), (c) and (e), the dotted (black) curves correspond to the trajectories of the ion when \(\mu_{0}=0\), while the solid (green) curves show the trajectories when \(\mu_{0}\neq 0\). Panels (b), (d), and (f) present the difference between the trajectories, with and without the dipole moment. The magnitudes of the trapping potentials are \(U_{0}=680\,\mathrm{V},V_{0}=1650\,\mathrm{V}\).
(Color online) Time evolution of the \(X\) component of the center-of-mass motion, with classical rotation, for different initial angles and different trap potentials: (a) \(U_{0}=680\,\)V, \(V_{0}=1650\,\)V; (b) \(U_{0}=2175\,\)V, \(V_{0}=2000\,\)V; (c) \(U_{0}=7080\,\)V, \(V_{0}=3000\,\)V. The curves present the difference between the trajectories of motion for the two cases of \(\mu=0\) and \(\mu=\mu_{0}\).
the rotational energy
\[E_{\rm rot}=\frac{1}{2}I(\omega_{Px^{\prime}}^{2}+\omega_{Py^{\prime}}^{2})\,, \tag{34}\]
see
### Quantum rotation
In the semi-classical simulations, the initial state of the ion is chosen as for the classical simulations, except for rotations, where we take the initial quantum rotational state to correspond to a specific spherical harmonic \(|J,M\rangle\). In other words, all coefficients \(D_{J,M}\) in Eq. are equal to zero at \(t=0\), except for one which is set to 1. All the results presented here for the semi-classical model are for the highest magnitude of the trapping field used in the classical simulations, namely \(U_{0}=7080\,\)V, \(V_{0}=3000\,\)V.
(Color online) Time evolution of the \(X\) component of the center-of-mass motion, with classical rotation, for different trap potentials: (a) \(U_{0}=2175\,\)V, \(V_{0}=2000\,\)V; (b) \(U_{0}=7080\,\)V, \(V_{0}=3000\,\)V.
#### iii.1.1 Effect of the dipole moment on the trajectory
The results for the center-of-mass motion of the ion in the semi-classical model, starting from the initial rotational states \(|0,0\rangle\) and \(|1,0\rangle\), are shown in As previously, we also present the difference between the trajectories obtained with and without a dipole moment. Only results along the \(X\) and \(Y\) axes are presented, as there is no effect on the \(Z\) component of the trajectory. Comparing to the result for the classical model, Fig. 4, we see that the effect of the quantum rotation on the center-of-mass motion is smaller. In addition, the deviations in the trajectories are smaller when the ion is initially in an excited state (\(J=1\)) in comparison to the rotational ground state (\(J=0\)).
#### iii.1.2 Effect of the trapping field on the rotation of the ion
We now shift our attention from the center-of-mass motion to the effect of the external trapping electric field on the rotational dynamics of the molecular ion. In Fig. 8, we plot the
(Color online) Time evolution of the rotational energy of a classical MgH\({}^{+}\) molecular ion in a linear Paul trap. The time evolution of the trapping electric field, at the position \(\mathbf{R}\) of the ion, is also shown as a dotted (black) curve. The magnitudes of the trapping potentials are \(U_{0}=2175\,\mathrm{V},V_{0}=2000\,\mathrm{V}\).
We see that the trapping field causes a small perturbation in the quantum rotation of the diatomic ion, as is clearly evidenced by the local value of the field felt by the ion, which is also plotted in The trapping field is thus coupling the ground rotational state to excited states, but the ion appears to return adiabatically to the ground state as it revisits a low-field region.
Comparable results are obtained when starting from the excited states \(\left|1,0\right\rangle\) or \(\left|1,1\right\rangle\), as shown in Starting from the state \(\left|1,1\right\rangle\), one sees, Fig. 9(a), a strong exchange with the corresponding sub-level with opposite \(M=-1\), \(\left|1,-1\right\rangle\), with limited transfer to the state \(\left|1,0\right\rangle\) [note the difference in scale in Fig. 9(a) for the latter state].
(Color online) Time evolution of the center-of-mass motion of a MgH\({}^{+}\) ion inside a linear Paul trap, with quantum rotation, for: (a)–(b) \(X\) component of motion, (c)–(d) \(Y\) component of motion. Panels (a) and (c) show the trajectories of the motion when \(\mu=0\). Panels (b) and (d) present the difference between the trajectories of motion when \(\mu=0\) and when \(\mu\neq 0\). In the case when \(\mu\neq 0\), the ion is initially in a specific rotational state.
9(b).
By summing over the probabilities of finding the ion to be in one of the sub-levels above, i.e.,
\[\left|D_{J}(t)\right|^{2}\equiv\sum_{M=-J}^{J}\left|D_{J,M}(t)\right|^{2} \tag{35}\]
9(c) that the excited states are less perturbed by the external field than the ground state, especially when starting from \(\left|1,0\right\rangle\). Nevertheless, the states corresponding to \(J=1\) follow the same temporal changes as the local electric field, as was the case for the ground state.
Finally, we have also compared the time evolution of the first three rotational states of the ion corresponding to quantum numbers \(J=0\), \(J=1\) and \(J=2\), by studying their populations \(\left|D_{J}(t)\right|^{2}\), Eq., in As we see, the rotational ground state is the most affected state by the interaction of the electric field with the dipole, while on the other hand, the states corresponding to \(J=2\) are hardly affected by the trapping field.
(Color online) Time evolution of the quantum rotational ground state of a MgH\({}^{+}\) molecular ion with permanent dipole moment \(\mu_{0}\), in a linear Paul trap. The ion is initially in its rotational ground state. The time evolution of the trapping electric field, at the position \(\mathbf{R}\) of the ion, is also shown with a dotted (green) curve. The magnitudes of the trapping potentials are \(U_{0}=7080\,\mathrm{V},V_{0}=3000\,\mathrm{V}\).
(Color online) Time evolution of the first excited quantum rotational state of a MgH\({}^{+}\) molecular ion with permanent dipole moment \(\mu_{0}\), in a linear Paul trap. Panels (a)–(b) show the time evolution of sub-levels \(M=0,\pm 1\) when the ion is initially in either rotational states corresponding to \(J=1\), \(M=1\) [panel (a)] and \(J=1\), \(M=0\) [panel(b)]. Panel (c) compares the time evolution of state with \(J=1\), for the two initial conditions mentioned, along with the time evolution of the trapping electric field at the position \(\mathbf{R}\) of the ion. The magnitudes of the trapping potentials are \(U_{0}=7080\,\mathrm{V},V_{0}=3000\,\mathrm{V}\).
8 and 9, that the higher the energy of the rotational state, the less it is affected by the interaction between the trapping electric field and the dipole moment. It can be understood by considering that the lower the value of \(J\), the more the presence of the dipole-field interaction \(\mathbf{\mu}\cdot\mathbf{E}\) affects the rotational eigenstates, see Refs..
## V Conclusion
We have studied the dynamics of a rigid heteronuclear diatomic molecular ion trapped in a linear Paul trap by carrying out both classical and semi-classical numerical simulations. We have investigated the effect of the interaction between the permanent dipole moment of the ion and the time-varying trapping electric field on the motion of the center of mass of the ion, along with the effect of the field on the rotation of the ion.
Classically, we showed that the trajectories of the center of mass differ from those of
(Color online) Comparison of the time evolution of the quantum rotational ground state \(J=0\) (when the ion is initially in this state), first rotational excited state \(J=1\) (when the ion is initially in this state), and second rotational excited state \(J=2\) (when the ion is initially in this state) with each other and with the time evolution of the trapping electric field, at the position \(\mathbf{R}\) of the ion. The magnitudes of the trapping potentials are \(U_{0}=7080\,\mathrm{V},V_{0}=3000\,\mathrm{V}\).
The trajectories obtained therefore do not follow the Mathieu equation, as atomic ions do, but present nevertheless the same periodic oscillations. We observed that the deviation of the trajectories of the motion of the molecular ion is strongly dependent on its initial orientation (orientation of the dipole) inside the trap, for the same magnitude of the electric field. Similarly, a stronger trapping field lead to a greater divergence between the trajectories with and without a coupling between the field and the dipole moment of the ion.
Considering quantum rotation in a semi-classical model, we found that the time evolution of the rotational states of the ion follows closely the time evolution of the trapping electric field. Therefore, when the electric field is zero, the initial rotational state of the ion is fully populated and as soon as the electric field increases from zero, the probability of finding the ion in other rotational states becomes non-zero. We showed that higher rotational states of the ion are less affected by the interaction of the dipole moment with the trapping electric field. We also found that the trajectory of the ion is less affected by the coupling between the trapping field and the dipole moment when quantum rotation is considered, compared to classical rotation.
In all cases, we found that the deviations of the trajectories from the Mathieu equation did not lead to unbounded trajectories. In other words, the stability of the trap was not affected by the presence of the coupling between the dipole moment and the trapping field for the potential values \(U_{0}\) and \(V_{0}\) used in this study. Future work should explore whether there exists instances where this is not the case, maybe very close to the border between stable and unstable trajectories.
|
10.48550/arXiv.1511.04114
|
Rotational dynamics of a diatomic molecular ion in a Paul trap
|
A. Hashemloo, C. M. Dion
| 5,288
|
10.48550_arXiv.1106.0594
|
Biosorption of Ag(I)-Spirulina platensis for different pH
E.Gelagutashvili, E.Ginturi, N,Kuchava, N.Bagdavadze, A.Rcheulishvili
Iv.
E.
0186, 6, Tamarashvili St.,
Tbilisi, Georgia
Abstract
Biosorption of _Ag(I)-Spirulina platensis_ for different _pH_ were investigated using dialysis and Atomic-absorbtion analysis. It was shown, that the biosorption constant for Ag(I) _Spirulina platensis_ complex and the capacity depend on the change of _pH_. In particular, with the increase of _pH (pH=5.5. and pH=8.6_ cases), the biosorption constant increase and the capacity decreases. The nature of interaction is also changed. In case of neutral _pH_, the interaction Ag(I)-_S. platensis_ is of cooperative character and maximum metal biosorption by _S. platensis_ biomass was observed at _pH 7.0._
Introduction
Cyanobacteria _Spirulina platensis_ (also known as blue-green algae) is gaining more attention in the field medical science because of its nutraceutical and pharmaceutical importance. _S. platensis_ consists of several nutritions elements, which are important for health improvement. _S. platensis_ are most diverse group of photosynthetic prokaryotes. An aqueous extract of _S. platensis_ inhibited HIV-1 replication in human T-cell lines, peripherial blood mononuclear cells and Langerhans cells.
Various species of cyanobacteria and algae have been known to adsorb and take up heavy metal ions. Understanding the bioavailability of heavy metals is a advantageous for plant cultivation and phytoremediation. Quantification of metal-biomass interactions is fundamental to the evaluation of potential implementation strategies, hence sorption isotherms, as well as models use to characterize algae biosorption.
The role of silver for health promotion and disease prevention is generally accepted world wide.
The aim of this study was to examine silver bioavailability with _S. platensis_ for different pH.
Materials and Methods
The study of biosorption of Ag(I)-_S. platensis_ was carried out by the methods of dialysis and atomic absorption analysis. _S. platensis_ was used in different state:
1. In suspension, when it is dissolved in medium and its pH=8.6.
2. Dissolved in water, pH=5.5
3. Dissolved in phosphate buffer, pH=7.0.
The experiments of dialysis were carried out in 5ml cylindrical vessels made of organic glass. A cellophane membrane of 30\(\upmu\)m width (type - "Visking" manufacturer - "serva") was used as a partition. The duration of dialysis was 72 hours. The experiments were carried out at 20-23\({}^{\rm{o}}\)C temperature.
In all mentioned cases, the concentration of _Spirulina platensis_ was 400mg/100ml. The metal concentration ranged from 10\({}^{\rm{-3}}\) to 10\({}^{\rm{-5}}\). The metal concentration after the dialysis was measured by atom-absorption analysis at the wavelength of _l_=_328.1_ nm. Each value was determined as an average of three independent estimated values \(\pm\) the standard deviation. Several models for the apparent Ag dissociation constant were fit to the observed binding isotherm. The best fits was obtained for binding of silver to _S.platensis_ by using Freundlich and Hill models[3.4].
Results and Discussions
For A and B cases, cyanobacteria were dissolved in nutrient medium _pH=8.6_ and in water _pH=5.5_, respectively.
Biosorption isotherms for Ag(I)-_Spirulina platensis_ in the medium (A) at different \(pH\) and in water (B) by using the fitted Freundlich model
By means of Freundlich isotherms the biosorption constant (\(K\)) and the capacity (\(n\)) were determined for Ag(I)-\(S\). _platensis_. The data are shown in Table I.
Table 1.
\begin{tabular}{|l|l|l|l|l|l|} \hline & Biosorption & Biosorption & Hill coefficient & Standard & Correlation \\ & constant & capacity & & deviation & coefficient \\ \hline & _Kx10-4_ & \(n\) & \(n_{H}\) & \(SD\) & \(R\) \\ \hline _S.platensis_ & 9.4 & 1.67 & - & 1.18 & 0.92 \\ dissolved in medium & & & & & \\ pH=8.6 & & & & & \\ \hline _S.platensis_ & 13.0 & 5.27 & 4.34 & 0.045 & 0.97 \\ dissolved in phosphate buffer & & & & & \\ pH=7.0 & & & & & \\ \hline _S.platensis_ & 2.9 & 2.78 & - & 0.06 & 0.97 \\ dissolved in water & & & & & \\ pH=5.5 & & & & & \\ \hline \end{tabular} As it is seen from the Table, with the change of \(pH\) the biosorption parameters are changed. Namely, in case of high \(pH\) (\(pH\)=8.6), the biosorption constant _K=9.4 x 10-4_
Biosorption isotherms for Ag(I)-_Spirulina platensis_ for neutral \(pH\) (dissolved in phosphate buffer).
(\(C\) shows the hypothetical theoretical curve chosen by \(\chi^{2}\) criterion (\(\chi^{2}\)\(<\)0.002),
\(D\) shows the same parameters in Hill coordinates).
shows the biosorption isotherm for the case, when the cyanobacteria is dissolved in phosphate buffer _pH=7.0_. The dots on the figure show the experimental data, and the solid line (C case) is the hypothetical theoretical curve chosen by \(\chi^{2}\) criterion (\(\chi^{2}\) 0.002, R\({}^{2}\) 0.97) in _Y=C\({}_{\emph{b/n}}\) vs Logm_ coordinates, where \(n\), the number of active centers (capacity) is determined as the maximum value of \(C_{b}\), and \(m\) is the concentration of free metal ions. Each dot is the average of three independent values, and the standard deviation \(<9\%\) of average value.
The type of \(y\) versus _logm_ dependence is nonlinear- S-shape, means that there exists positive cooperation of interaction of metal ions bound with C-PC, i.e. binding of the first metal ion increases affinity of the site for the second one. For more argumentation, as an additional criterion of cooperativity, the Hill plot (dependence _LogY/I-Y vs Logm_) was used. The line, deviation of which is the Hill coefficient (_nH_). Taking this into account, the same data are presented in Hill coordinates, Fig. 2(D), where the parameters are the same, as in previous case, Fig. 2(C). By using Hill equality, the following values of \(K\) and _nH_ (Table 1) were obtained: _K=13.0 x 10\({}^{\emph{7}}\)_ and _nH=4.34_. As it is seen from the Table, nH=1, showing the incomplete cooperativity of the interaction between the silver ions and _S. platensis_.
In all cases, the correlation between the experimental and the theoretical data is obvious (_R_ is more than 0.9).
The mentioned model gives the possibility to characterize the interaction Ag(I)-_Spirulina platensis_. The result obtained from the cooperative interaction shows that both, the biosorption constant for Ag(I) _Spirulina platensis_ complex and the capacity depend on the change of _pH_. In particular, with the increase of _pH (pH=5.5. and pH=8.6_ cases), the biosorption constant increase and the capacity decreases. The nature of interaction is also changed. In case of neutral _pH_, the interaction Ag(I)-_S. platensis_ is of cooperative character and maximum metal biosorption by _S. platensis_ biomass was observed at _pH 7.0_.
Thus, in the case of neutral _pH_, in contrast to the previous case, it is possible to characterize the nature of interaction. In the interaction a number of functional groups take place having different location that lead to the creation of coordination centers of different type strongly differing in composition and structure. But, in the case of neutral _pH_ the biosoprtion constant is the highest.
References
1. Ayehunie S.,Blay A.,Baba T.W., Ruprecht R.M. Inhibition of HIV-1 replication by an aqueous extract of _Spirulina platensis_. J.Acquir. Immune. Defic. Syndr. Hum. Retrovirol 1998,18,7-12.
2. Kuyucak N., Volesky, B., Raton F.L. Biosorption of heavy metals. CRC Press, Boca Raton,p.173
3. Freundlich,H., Adsorption in solutions,Phys.Chem. 1906,57,384-410.
4. Hill A.V. The possible effects of the aggregation of the molecules of hemoglobin on its dissociation curves, J.Physiol.,1910, London,40,iv.
|
10.48550/arXiv.1106.0594
|
Biosorption of Ag(I)-Spirulina platensis for different pH
|
E. Gelagutashvili, E. Ginturi, N. Kuchava, N. Bagdavadze, A. Rcheulishvili
| 6,288
|
10.48550_arXiv.1304.5764
|
## I Introduction
Nonequilibrium states like glasses from supercooled liquids (SCLs) are abundant in Nature, whose entropy \(S\) can only be estimated by calorimetrically measured entropy \(S_{\rm expt}\), which can then be extrapolated to absolute zero. The extrapolated value \(S_{\rm R}\) at absolute zero is commonly known as the _residual entropy_ and is normally found to satisfy \(S_{\rm R}>0\). In practice, one considers the isobaric entropy \(S(T_{0})\) of the system as a function of the temperature \(T_{0}\) of the surrounding medium; see The existence of \(S_{\rm R}\) was first theoretically demonstrated by Pauling and Tolman; see also Tolman. In addition, the existence of the residual entropy has been demonstrated rigorously for a very general spin model by Chow and Wu. The residual entropy for glycerol was observed by Gibson and Giauque and for ice by Giauque and Ashley. Pauling provided the first numerical estimate for the residual entropy for ice, which was later improved by Nagle. Nagle's numerical estimate has been recently verified by simulation. The numerical simulation carried out by Bowles and Speedy for glassy dimers also supports the existence of a residual entropy. For a brief review of the history of the residual entropy, see. Thus, it appears that the support in favor of the residual entropy, see the curve Glass1 in see Fig. 2, is quite strong. Its existence also does not violate Nernst's postulate, as the latter is applicable only to true equilibrium states with a _non-degenerate_ ground state. Indeed, many exactly solved statistical mechanical models show a non-zero entropy at absolute zero. However, as of yet, no experiment can be performed at absolute zero to experimentally determine the residual entropy; in all cases, some sort of _extrapolation_ is required. This point should not be forgotten in the following whenever we speak of the residual entropy. Despite the above mentioned support for the reality of the residual entropy, it has become a highly debated issue in the literature as discussed by these authors. The reason for the debate is that the relationship among \(S(T_{0})\), \(S_{\rm expt}(T_{0})\) and the entropy \(S_{\rm SCL}(T_{0})\) of the corresponding stationary state is not well understood, and understanding this relationship is the main theme of this work.
In the following, we will speak of "the equilibrium" state associated with a nonequilibrium state as the stationary (time-independent) state. Depending on the context, the equilibrium state may represent a true equilibrium state such as a crystal or a stationary metastable state such as the supercooled liquid. A nonequilibrium state in this work will always be taken as time-dependent. Accordingly, \(S(T_{0})\) above should be correctly expressed as \(S(T_{0},t)\), and in some cases can be expressed as a function \(S(T(t))\) of the instantaneous temperature \(T(t)\) of the system; see In this work, we will not be concerned with \(T(t)\). Thus, we will simply use \(S(T_{0})\) for the nonequilibrium state, knowing well that the state continues to change with time.
For the purpose of clarity, we will consider supercooled liquids and associated nonequilibrium states (glasses) in the following, but the arguments are applicable to all nonequilibrium states. The supercooled liquid undergoes a glass transition over a transition range, see Fig. 2, over which the entropy falls rapidly with lowering temperature \(T_{0}\).
An isolated system \(\Sigma_{0}\) consisting of the system \(\Sigma\) in a surrounding medium \(\widetilde{\Sigma}\). The medium and the system are characterized by their fields \(T_{0},P_{0},...\) and \(T(t),P(t),...\), respectively, which are different when the two are out of equilibrium.
As the irreversibility due to the glass transition does not allow for an exact evaluation of the entropy, it has been suggested that the entropy decreases by an amount almost equal to \(S_{\rm R}\) within the glass transition region so that the glass (see Glass2 in Fig. 2, whose entropy lies below the supercooled liquid) would have a vanishing entropy at absolute zero. It has been shown by Goldstein that Glass2 results in a violation of the second law. It should be stressed that if there is ever any _conflict_ between the second law and any other law in physics such as the zeroth or the third law, it is the second law that is believed to hold in _all_ cases. One can also argue that to confine the glass into a _unique_ basin in the energy landscape requires _microscopic_ information; hence, the particular glass _cannot_ be considered in a macrostate. Oppenheim has also raised somewhat of a similar objection.
We have drawn the two entropy curves (Glass1 or Glass2) in that emerge out of the entropy curve for the equilibrated supercooled liquid for a given \(\tau_{\rm obs}\) in such a way that Glass1 has its entropy above (so that \(S_{\rm R}\geq 0\)) and Glass2 below (so that \(S_{\rm R}\equiv 0\)) that of the supercooled liquid. The entropy of Glass1 (Glass2) approaches that of the equilibrated supercooled liquid entropy from above (below) during isothermal (fixed temperature of the medium) relaxation; see the two downward vertical arrows for Glass1. It is the approach to equilibrium that distinguishes the two glasses, Glass1 and Glass2. Almost all experimental investigations leave open the possibility that Glass2 may materialize if the irreversibility is too large. Our work clarifies the situation.
It is abundantly clear from the above discussion that there is a need to look at the relationship between various entropies in As is customary, we treat the supercooled liquid as an equilibrium state, even though it not a true equilibrium state; see above. We proceed by following the strict second law inequality \(d_{\rm i}S>0\), see Eq., and use it to prove the following results applicable to _all_ nonequilibrium systems, regardless of how close or far they are from their equilibrium state:
1. Various entropies obey the following strict inequalities \[S(T_{0})>S_{\rm expt}(T_{0})>S_{\rm SCL}(T_{0})\quad\mbox{for $T_{0}<T_{0\rm g}$},\] so that the entropy variation in time has a unique direction as shown by the _downward arrows_ in Thus, \(S(T_{0})\) cannot drop below \(S_{\rm SCL}(T_{0})\) (such as Glass2 in Fig. 2) without violating of the second law.
2. The experimentally observed non-zero entropy at absolute zero in a vitrification process is a _strict lower bound of the residual entropy_ of any system: \[S_{\rm R}\equiv S>S_{\rm expt}>S_{\rm SCL}.\]
The Eq. is consistent with Glass1 but not with Glass2. All experiments on or exact/approximate computations for nonequilibrium systems _must_ obey the strict inequalities in Eqs. without exception. This is the meaning behind the usage of "... rigorous..." in the title. The actual values of the entropy are not relevant for the aim of this work, which is to find the relationship among different entropies under vitrification. Because of the possibility that the systems may be far away from equilibrium such as in a fast quench, where the irreversible contributions may not be neglected, our results go beyond the previous calorimetric evidence. The systems we are interested in include glasses and imperfect crystals as special cases. However, to be specific, we will only consider glasses below.
## II Entropy bounds during vitrification
The vitrification process we consider is carried out at some cooling rate as follows. The temperature of the medium is isobarically changed by some small but fixed \(\Delta T_{0}<0\) from the current value to the new value, and we wait for (not necessarily fixed) time \(\tau_{\rm obs}\) at the new temperature to make an instantaneous measurement on the system before changing the temperature again. At some temperature \(T_{0\rm g}\), see Fig. 2, the relaxation time \(\tau_{\rm relax}\), which continuously increases as the temperature is lowered, becomes equal to \(\tau_{\rm obs}\).
Schematic behavior of the entropy of the equilibrated, i.e. stationary supercooled liquid (solid curve) and two possible glasses (Glass1-dotted curve, Glass2-dashed curve) during vitrification. The transition region between \(T_{0\rm g}\) and \(T_{0\rm G}\) has been exaggerated to highlight the point that the glass transition is not a sharp point. For all temperatures \(T_{0}<T_{0\rm g}\), any nonequilibrium state undergoes isothermal structural relaxation in time towards the supercooled liquid. The entropy of the supercooled liquid is shown to extrapolate to zero per our assumption, but that of Glass1 to a non-zero value and of Glass2 to zero at absolute zero.
This solid is identified as a _glass_. The location of both temperatures depends on the rate of cooling, i.e. on \(\tau_{\rm obs}\). Over the glass transition region between \(T_{0\rm G}\) and \(T_{0\rm g}\) in Fig. 2, the system gradually turns from an equilibrium supercooled liquid at or above \(T_{0\rm g}\) into a glass at or below \(T_{0\rm G}\). We overlook the possibility of the supercooled liquid ending in a spinodal. It is commonly believed that \(S_{\rm SCL}\) will vanish at absolute zero (\(S_{\rm SCL}\equiv 0\)), as shown in the figure. However, it should be emphasized that the actual value of \(S_{\rm SCL}\) has no relevance for the theorems below.
We will only consider isobaric cooling (we will not explicitly exhibit the pressure in this section), which is the most important situation for glasses. The process is carried out along some path from an initial state A at temperature \(T_{0}\) in the supercooled liquid state which is still higher than \(T_{0\rm g}\) to the state A\({}_{0}\) at absolute zero. The state A\({}_{0}\) depends on the path A\(\rightarrow\)A\({}_{0}\), which is implicit in the following. The change \(dS\) between two neighboring points along such a path is \(dS=d_{\rm e}S+d_{\rm i}S\) in modern notation.
\[d_{\rm e}S(t)=d_{\rm e}Q(t)/T_{0}\equiv C_{P}dT_{0}/T_{0} \tag{3}\]
It also represents the _calorimetrically_ determined change in the entropy in any process.
\[d_{\rm i}S>0 \tag{4}\]
The equality in Eq. holds for a reversible process, which we will no longer consider unless stated otherwise. A discontinuous change in the entropy is ruled out from the continuity of the Gibbs free energy \(G\) and the enthalpy \(H\) in vitrification proved elsewhere.
## Theorem 1
## The experimentally observed (extrapolated) non-zero entropy at absolute zero in a vitrification process is a strict lower bound of the residual entropy of any system:
\[S_{R}\equiv S>S_{\rm expt}.\]
## Proof.
We have along A\(\rightarrow\)A\({}_{0}\)
\[S=S(T_{0})+\int\limits_{\rm A}^{{\rm A}_{0}}\!\!d_{\rm e}S+\int\limits_{\rm A }^{{\rm A}_{0}}\!\!d_{\rm i}S, \tag{5}\]
Since the second integral is always _positive_, and since the residual entropy \(S_{\rm R}\) is, by definition, the entropy \(S\) at absolute zero, we obtain the important result
\[S_{\rm R}\equiv S>S_{\rm expt}\equiv S(T_{0})+\int\limits_{T_{0}}^{0}C_{ P}dT_{0}/T_{0}. \tag{6}\]
This proves Theorem 1. The integral represents the calorimetric contribution.
The strict forward inequality above clearly establishes that the residual entropy at absolute zero must be strictly larger than \(S_{\rm expt}\) in any nonequilibrium process.
## Theorem 2
## The calorimetrically measured (extrapolated) entropy during processes that occur when \(\tau_{obs}<\tau_{\rm relax}(T_{0})\) for any \(T_{0}<T_{0g}\) is larger than the supercooled liquid entropy at absolutely zero
\[S_{\rm expt}>S_{\rm SCL}.\]
## Proof.
Let \(\dot{Q}_{\rm e}(t)\equiv d_{\rm e}Q(t)/dt\) be the rate of net heat loss by the system.
\[|d_{\rm e}Q| \equiv C_{P}\left|dT_{0}\right|=\int\limits_{0}^{\tau_{obs}}\left| \dot{Q}_{\rm e}\right|dt<\left|dQ\right|_{\rm eq}\left(T_{0}\right)\] \[\equiv\int\limits_{0}^{\tau_{\rm relax}(T_{0})}\left|\dot{Q} \right|dt,\qquad T_{0}<T_{0\rm g}\]
become supercooled liquid during cooling at \(T_{0}\). For \(T_{0}\geq T_{0\rm g}\), \(d_{\rm e}Q\equiv d_{\rm e}Q_{\rm eq}(T_{0})\equiv C_{P,\rm eq}dT_{0}\).
\[\int\limits_{T_{0}}^{0}C_{P}dT_{0}/T_{0}>\int\limits_{T_{0}}^{0}C_{P,\rm eq}dT_ {0}/T_{0}.\]
We thus conclude that
\[S_{\rm expt}>S_{\rm SCL}. \tag{7}\]
This proves Theorem 2.
The strict inequalities above are the result of glass being a nonequilibrium state. We have now verified the second statement in the Introduction.
The difference \(S_{\rm R}-\)\(S_{\rm expt}\) would be larger, more irreversible the process is. The quantity \(S_{\rm expt}\) can be determined calorimetrically by performing a cooling experiment. We take \(T_{0}\) to be the melting temperature \(T_{0\rm M}\), and uniquely determine the entropy of the supercooled liquid at \(T_{0\rm M}\) by adding the entropy of melting to the crystal entropy \(S_{\rm CR}(T_{0\rm M})\) at \(T_{0\rm M}\). The latter is obtained in a unique manner by integration along a reversible path from \(T_{0}=0\) to \(T_{0}=T_{0\rm M}\):
\[S_{\rm CR}(T_{0\rm M})=S_{\rm CR}+\int\limits_{0}^{T_{0\rm M}}C_{P,\rm CR} dT_{0}/T_{0},\]here, \(S_{\rm CR}\) is the entropy of the crystal at absolute zero, which is traditionally taken to be zero in accordance with the third law, and \(C_{P,\rm CR}(T_{0})\) is the isobaric heat capacity of the crystal. This then uniquely determines the entropy of the liquid to be used in the right hand side in Eq.. We will assume that \(S_{\rm CR}=0\). Thus, the experimental determination of \(S_{\rm expt}\) is required to give the _lower bound_ to the residual entropy in Eq.. Experiment evidence for a non-zero value of \(S_{\rm expt}\) is abundant as discussed by several authors; a textbook also discusses this issue. Goldstein gives a value of \(S_{\rm R}\simeq 15.1\) J/K mol for _o_-terphenyl from the value of its entropy at \(T_{0}=2\) K. We have given above a mathematical justification of \(S_{\rm expt}>0\) in Eq..
The inequality in Eq. takes into account any amount of irreversibility during vitrification; it is no longer limited to only small contributions of the order of 2% considered by several others, which makes our derivation very general.
By considering the state A\({}_{0}\) above to be a state A\({}_{0}\) of the glass in a medium at some arbitrary temperature \(T_{0}^{\prime}\) below \(T_{0\rm g}\), we can get a generalization of Eq.:
\[S(T_{0}^{\prime})>S_{\rm expt}(T_{0}^{\prime})\equiv S(T_{0})+\int\limits_{T_ {0}}^{T_{0}^{\prime}}C_{P}dT_{0}/T_{0}. \tag{8}\]
We again wish to remind the reader that all quantities depend on the path A\(\rightarrow\)A\({}_{0}\), which we have not exhibited. By replacing \(T_{0}\) by the melting temperature \(T_{0\rm M}\) and \(T_{0}^{\prime}\) by \(T_{0}\), and adding the entropy \(\widetilde{S}(T_{0\rm M})\) of the medium on both sides in the above inequality, and rearranging terms, we obtain (with \(S_{\rm L}(T_{0\rm M})=S_{\rm SCL}(T_{0\rm M})\) for the liquid)
\[S_{\rm L}(T_{0\rm M})+\widetilde{S}(T_{0\rm M})<S(T_{0})+\widetilde{S}(T_{0 \rm M})-\int\limits_{T_{0\rm M}}^{T_{0}}C_{P}dT_{0}/T_{0}, \tag{9}\]
This provides us with an independent derivation of the inequality given by Sethna and coworker.
It is also clear from the derivation of Eq.
\[S_{\rm expt}(T_{0})>S_{\rm SCL}(T_{0}), \tag{10}\]
Thus, \(S_{\rm expt}(T_{0})\) appears to have a form similar to that of Glass1 in but strictly lying below it. We have now verified the first statement in the Introduction.
While we have only demonstrated the forward inequalities, the excess \(S_{\rm R}-S_{\rm expt}\) can be computed in nonequilibrium thermodynamics, which provides a clear prescription for calculating the irreversible entropy generation. We do not do this here as we are only interested in general results, while the calculation of irreversible entropy generation will, of course, be system-dependent and will require detailed information. Gutzow and Schmelzer provide such a procedure with a single internal variable but under the assumption of equal temperature and pressure for the glass and the medium. However, while they comment that \(d_{i}S\geq 0\) whose evaluation requires system-dependent properties, their main interest is to only show that it is negligible compared to \(d_{\rm e}S\).
We have proved Theorems 1 and 2 by considering only the system without paying any attention to the medium. For Theorem 1, we require the second law, i.e. Eq.. This is also true of Eq.. The proof of Theorem 2 requires the constraint \(\tau_{\rm obs}<\tau_{\rm relax}(T_{0})\) for any \(T_{0}<T_{0\rm g}\), which leads to a nonequilibrium state. The same is also true of Eq..
## III Conclusions
We have considered the role of irreversible entropy generation during isobaric vitrification to rigorously justify the two statements in the Introduction. They are valid regardless of how far the system is out of equilibrium. Thus, our results are very general and are not restricted by the small amount of irreversibility that is normally considered in the literature. The first statement shows that the instantaneous entropy \(S(T_{0},t)\) must always be higher than \(S_{\rm expt}(T_{0})\), which in turn must always be higher than \(S_{\rm SCL}(T_{0})\) of the equilibrated supercooled entropy. The second statement shows that the extrapolation of the calorimetrically measured entropy to absolute zero forms a strict lower bound to the residual entropy \(S_{\rm R}\). As the former is usually positive, this proves that the residual entropy has to be at least as large as this value. From the first statement, it also follows that Glass2 is not realistic.
The statements follow from considering the thermodynamic entropy that appears in the second law, and their validity is not affected by which equivalent statistical definition of entropy one may wish to use for the thermodynamic entropy, an issue that has been investigated by us recently.
|
10.48550/arXiv.1304.5764
|
Some Rigorous Results Relating Nonequilibrium, Equilibrium, Calorimetrically Measured and Residual Entropies during Cooling
|
P. D. Gujrati
| 3,355
|
10.48550_arXiv.2410.24192
|
## 1 Introduction
With the increase in demand for high-accuracy first-principles simulations of the quantum dynamics of molecular systems at the attosecond time scale comes the need for theoretical frameworks that allow derivation of affordable, accurate, and systematically improvable computational tools. The cornerstone of approximate quantum dynamics has for more than half a century been the _time-dependent variational principle_ (TDVP) and the _McLachlan variational principle_, often collectively reffered to as the Dirac-Frenkel-McLachlan variational principle, even though they are not equivalent in all circumstances. However, the most popular wavefunction-based method in quantum chemistry is the coupled-cluster (CC) method, which is notable for being _not variational_ - it is _bivariational_.
In this article, we present a comprehensive study of _the time-dependent bivariational principle_ (TD-BIVP) which generalizes the TDVP, and forms the proper setting for the various forms of time-dependent CC theory that have been developed over the last decades; from time-dependent traditional CC theory, via equation-of-motion CC theory, to orbital-adaptive and orbital-optimized CC theory. Applications of time-dependent CC theory with a bivariational formulation covers applications as diverse as electronic-structure theory, the vibrational Schrodinger equation,and nuclear structure theory. For a recent review, see Ref.. Considering the importance of variational principles for the development of computational tools on one hand, and the growing importance of time-dependent CC theory on the other, establishing the theoretical framework of the TD-BIVP can therefore catalyze rapid progress in the development of sophisticated and relatively low-cost real-time propagation methods in several fields, and in particular in attochemistry.
The TD-BIVP was first mentioned in passing by Chernoff and Marsden, without coining the term, who devised a Lagrangian density (in the sense of field theory), and a corresponding action \(\mathcal{A}\), in which the system wavefunction and its complex conjugate were formally independent variables \(\psi\) and \(\tilde{\psi}\) forming canonically conjugate variables in an abstract phase space \(\mathbb{H}\). The Euler-Lagrange equations were the time-dependent Schrodinger equation and its dual, written as a pair of Hamilton's equations of motion on complex form. The principle was independently discovered by Arponen in his seminal treatise on coupled-cluster (CC) theory, where the CC amplitudes turn out to be canonical variables, preserving the form of Hamilton's equations of motion. In later publications by the trio of Arponen, Bishop and Pajanne on the extended CC method, the principle was occasionally invoked, emphasizing the canonical structure of these methods, including small oscillations around the ground state solution. A more formal symplectic geometry formulation of coupled-cluster theory was first considered by Arponen, who identified a real Hamiltonian system; see Sec. 2.1 in this article.
While the canonical transformation of time-dependent CC theory is an exact reformulation of quantum dynamics, approximations are invariably introduced in the form of the conventional hierarchy of CC with singles, doubles, triples, etc. Such approximations amount to choosing a submanifold \(\mathcal{M}\subset\mathbb{H}\) of phase space and restricting variations of the variables in the action \(\mathcal{A}\) to be in the tangent space of \(\mathcal{M}\), in a similar fashion as is done for the Dirac-Frenkel and McLachlan principles. Traditionally, the manifold \(\mathcal{M}\) has been assumed to be complex in the bivariational case, i.e., the local coordinates are complex numbers and the points on \(\mathcal{M}\) are complex differentiable with respect to the coordinates. The argument has been that since \(\mathcal{A}\) is complex-valued, the manifold must be complex in order to give well-defined equations of motion. In this article, we show that this can be refined, and a careful consideration of real manifolds yields two distinct time-dependent bivariational bivariational principles, \(\delta\operatorname{Re}\mathcal{A}=0\) and \(\delta\operatorname{Im}\mathcal{A}=0\), that, just like the Dirac-Frenkel and McLachlan principles, are equivalent when the tangent spaces of \(\mathcal{M}\) are complex vector spaces, and exact when \(\mathcal{M}=\mathbb{H}\), the full phase space. For general real manifolds, the variational principles are distinct. The principle \(\delta\operatorname{Re}\mathcal{A}=0\) is exemplified by the time-dependent orbital-optimized CC method by Sato and coworkers, while the second variational principle has not been put to use, to our knowledge.
The remainder of this article is structured as follows. In Section 2 we introduce the TD-BIVP, with emphasis on symplectic geometry. In Section 3 we consider the restriction of the bivariational dynamics to (symplectic) submanifolds, and develop the equations of motions in local coordinates. We consider both real and complex submanifolds. We also discuss the relation of the TD-BIVP to the Dirac-Frenkel and McLachlan variational principles. In Section 5, we formulate bivariational dynamics methods in the literature using the present abstract framework. Finally, in Section 6 we present our conclusion and future perspectives. An appendix provides additional details.
## 2 The time-dependent bivariational principle
The time-dependent bivariational principle is a stationary-action principle for the time-dependent Schrodinger equation and its dual.
\[\mathcal{A}=\int_{0}^{T}i\left\langle\tilde{\psi}|\dot{\psi}\right\rangle- \langle\tilde{\psi}|H|\psi\rangle\;\,dt. \tag{1}\]We assume for simplicity that the system Hamiltonian \(H\) is a bounded self-adjoint operator, in order to avoid unnecessary formal complexity. Here, \(\psi(t)\in{\cal H}\), a wavefunction in a complex (separable) Hilbert space, and where \(\tilde{\psi}(t)\in{\cal H}^{*}\), the complex-conjugate Hilbert space, or dual space. The notation \(\left\langle\cdot|\cdot\right\rangle\) is thus the dual pairing on \({\cal H}^{*}\times{\cal H}\). (Note that \(\tilde{\psi}\) is _not_ meant to be the complex conjugate of \(\psi\), but an independent variable that lives in dual space.) In physics terms, \({\cal H}\) is the space of "kets", while \({\cal H}^{*}\) is the space of "bras".
The action is stationary (\(\delta{\cal A}=0\)) under arbitrary smooth variations (vanishing at the endpoints) of \(\psi\) and \(\tilde{\psi}\) if and only if
\[i\dot{\psi}=H\psi,\quad-i\dot{\tilde{\psi}}=H^{t}\tilde{\psi}, \tag{2}\]
Here, \(H^{t}\tilde{\psi}=\left\langle\tilde{\psi}\right|H\) in bra-ket notation is called the Banach adjoint, or operator transpose. With the Hamiltonian function \({\cal E}(\tilde{\psi},\psi)=\left\langle\tilde{\psi}|H\psi\right\rangle\), Eq.
\[i\dot{\psi}=\frac{\partial{\cal E}}{\partial\tilde{\psi}},\quad\mbox{and} \quad-i\dot{\tilde{\psi}}=\frac{\partial{\cal E}}{\partial\psi}, \tag{3}\]
Thus, \(\mathbb{H}\equiv{\cal H}^{*}\oplus{\cal H}\) serves as phase space, with \((\tilde{\psi},\psi)\in\mathbb{H}\) forming a pair of (infinite-dimensional) momenta and coordinate vectors, respectively.
The action \({\cal A}\) implies two conserved quantities: First, the overlap is conserved, \(d\left\langle\tilde{\psi}|\psi\right\rangle/dt=0\), and second, the energy is conserved, \(d{\cal E}/dt=0\).
We make note of the small curiosity, also present for the Modified Hamilton's Principle for classical dynamics, that only if \((\psi(T),\tilde{\psi}(T))\) are _actually_ solutions of Eqs. with initial conditions \((\tilde{\psi},\psi)\), will \({\cal A}\) have a critical point. On the other hand, the variations do not "see" the boundary conditions, since they are supported in the interior of the time interval \([0,T]\). Thus, the action functional can be viewed as a semi-local integral formulation of the time-dependent Schrodinger equation, and we will omit the specification of the time boundaries when writing the action integrals.
In exact quantum mechanics and in the TDVP, the only allowed initial conditions satisfy \(\tilde{\psi}=\psi^{\dagger}/\|\psi\|^{2}\), and consequently \(\tilde{\psi}(t)=\psi(t)^{\dagger}/\|\psi(t)\|^{2}\) for all \(t\). (In general, \(\psi^{\dagger}\) is given by Riesz' representation theorem. If \({\cal H}\) is a space of square-integrable functions, \(\psi^{\dagger}\) can be taken to be the complex conjugate function.) However, the power of the bivariational principle is that it allows far more flexible approximation schemes than the TDVP, since we may allow \(\tilde{\psi}\) and \(\psi\) to have independent approximations, such as is the case in coupled-cluster theory.
### Complex Hamiltonian systems as real Hamiltonian systems
Complex Hamiltonian equations of motion may seem strange and very different from the standard real Hamiltonian systems of classical mechanics. In this section we demonstrate, however, that the complex Hamiltonian system is in fact just real Hamiltonian systems in disguise.
The real and imaginary parts of the action \({\cal A}\) must be simultaneously stationary if \(\delta{\cal A}=0\). Following the idea of Arponen _et al._, we show that the complex Hamiltonian system is in fact equivalent to a standard real-valued Hamiltonian system: Let \(\{\phi_{\mu}\}\subset{\cal H}\) and \(\{\tilde{\phi}^{\mu}\}\subset{\cal H}^{*}\) be biorthogonal bases, i.e. \(\langle\tilde{\phi}^{\mu},\phi_{\nu}\rangle=\delta^{\mu}_{\nu}\), and define real-valued vectors \(q_{i}\) and \(p_{i}\), \(i\in\{1,2\}\), such that
\[\psi=\sum_{\mu}\phi_{\mu}(q_{1}^{\mu}+ip_{2}^{\mu}),\quad\tilde{\psi}=\sum_{\mu }\tilde{\phi}^{\mu}(q_{2,\mu}-ip_{1,\mu}). \tag{4}\]
We obtain, up to a total time derivative,
\[\mbox{Re}\,{\cal A}=\int p_{1}\cdot\dot{q}_{1}+p_{2}\cdot\dot{q}_{2}-\mbox{Re} \,{\cal E}\;dt,\] (5a) and \[\mbox{Im}\,{\cal A}=\int q_{2}\cdot\dot{q}_{1}+p_{1}\cdot\dot{p}_{2}-\mbox{Im} \,{\cal E}\;dt. \tag{5b}\]We recognize the real part of \({\cal A}\) as the Lagrangian from the Modified Hamilton's Principle, and consequently,
\[\dot{q}_{i}=\frac{\partial\mathop{\rm Re}\nolimits{\cal E}}{\partial p_{i}}, \quad\mbox{and}\quad\dot{p}_{i}=-\frac{\partial\mathop{\rm Re}\nolimits{\cal E }}{\partial q_{i}}.\] (6a) For the imaginary part, set \[(Q_{1},Q_{2})=(q_{1},p_{2})\] and \[(P_{1},P_{2})=(q_{2},p_{1})\] to obtain another Hamiltonian system \[\dot{Q}_{i}=\frac{\partial\mathop{\rm Im}\nolimits{\cal E}}{\partial P_{i}}, \quad\mbox{and}\quad\dot{P}_{i}=-\frac{\partial\mathop{\rm Im}\nolimits{\cal E }}{\partial Q_{i}}. \tag{6b}\]
Since \({\cal A}\) is complex differentiable, Eqs. (6a) and (6b) can be seen to be related by the Cauchy-Riemann equations, and hence equivalent. We make the observation that the complex Hamiltonian system is equivalent to two _distinct_ standard real Hamiltonian systems. Hamiltonian systems with multiple distinct symplectic structures are called _bi-Hamiltonian systems_.
### Symplectic formulation of the bivariational principle
We now develop the theory of the bivariational principle using a more abstract approach. This will in the end produce very concise and informative expressions.
\[\omega(u,v)=\langle\tilde{\psi}_{u}|\psi_{v}\rangle-\langle\tilde{\psi}_{v}| \psi_{u}\rangle=\langle\!\langle u,Jv\rangle\!\rangle, \tag{7}\]
The notation \(\langle\!\langle u,Jv\rangle\!\rangle\) is the dual pairing of \(\mathbb{H}={\cal H}^{*}\oplus{\cal H}\) and \(\mathbb{H}^{*}={\cal H}\oplus{\cal H}^{*}\), and does _not_ denote the Hermitian inner product on \(\mathbb{H}\). In fact, we will never use an inner product on \(\mathbb{H}\). (The reader may note that in this particular dual pairing - natural from the context - the dual element is to the _right_, as opposed to the left as in \(\langle\cdot|\cdot\rangle\).) Consequently, \((\mathbb{H},\omega)\) is now a (linear) symplectic space.
For an operator \(A:{\cal H}\rightarrow{\cal H}\), let us introduce the "symmetrization" \(\hat{A}:\mathbb{H}\rightarrow\mathbb{H}^{*}\) as
\[\hat{A}(\tilde{\psi},\psi)=\frac{1}{2}(A\psi,A^{t}\tilde{\psi}), \tag{8}\]
We define the notation \({\cal E}(u)\equiv{\cal E}(u,u)=\langle\tilde{\psi}_{u}|H\psi_{u}\rangle\).
\[{\cal A}=\int_{0}^{T}\frac{i}{2}\omega(u,\dot{u})-{\cal E}(u)\;dt. \tag{9}\]
Let \(\delta u(t)\in\mathbb{H}\) be an arbitrary variation.
\[\begin{split}\delta{\cal A}&=\int i\omega(\delta u,\dot{u})-d{\cal E}(u)(\delta u)\;dt\\ &=\int i\langle\!\langle\delta u,J\dot{u}\rangle\!\rangle-2 \langle\!\langle\delta u,\hat{H}u\rangle\!\rangle\;dt.\end{split} \tag{10}\]
Here, \(d{\cal E}(u)\) is the Frechet derivative of \({\cal E}\), and \(\delta{\cal E}=d{\cal E}(u)(\delta u)\) is correspondingly the directional derivative of \({\cal E}\) in the direction \(\delta u\).
\[i\dot{u}=J^{-1}d{\cal E}(u), \tag{11}\]
which, due to the special form of \({\cal E}\) reduces to
\[i\dot{u}=2J^{-1}\hat{H}u, \tag{12}\]
## 3 Evolution on manifolds
In this section we describe the bivariational evolution on submanifolds of phase space. We first deal with complex manifolds, where complex differentiation can be used, and then with the more general real manifolds, which will allow us to identify two distinct time-dependent bivariational principles which reduce to the complex case under certain conditions.
### Complex manifolds
Approximate time evolution is obtained from the bivariational principle by introducing a smooth submanifold \(\mathcal{M}\subset\mathbb{H}\) and restricting the principle of stationary action to \(\mathcal{M}\); see for an illustration. We assume for simplicity that \(\mathcal{M}\) is a complex manifold of finite dimension \(n\).
Let \(u=(\tilde{\psi},\psi)\in\mathcal{M}\). Let \(\delta u(t)\in T_{u}\mathcal{M}\) be an arbitrary variation.
\[\begin{split}\delta\mathcal{A}&=\int i\omega(\delta u,u)-\langle\!\langle\delta u,d\mathcal{E}(u)\rangle\!\rangle\;dt\\ &=\int i\omega(\delta u,\dot{u})-\omega(\delta u,2J^{-1}\hat{H}u) \;dt\end{split} \tag{13}\]
and the corresponding Euler-Lagrange equation
\[i\dot{u}=P(u)J^{-1}d\mathcal{E}(u)=2P(u)J^{-1}\hat{H}u, \tag{14}\]
where \(P(u):\mathbb{H}\to T_{u}\mathcal{M}\) is the _symplectic projection_ at \(u\): For any \(v\in\mathbb{H}\), the projection \(P(u)v\in T_{u}\mathcal{M}\) is defined by the condition
\[\omega(\delta u,v)=\omega(\delta u,P(u)v)\;\forall\delta u\in T_{u}\mathcal{M}. \tag{15}\]
The symplectic projection \(P(u)\) is well-defined whenever the restriction \(\omega:T_{u}\mathcal{M}\times T_{u}\mathcal{M}\to\mathbb{C}\) of the symplectic form is nondegenerate, which is turn is the definition of \(\mathcal{M}\) being a symplectic submanifold of \(\mathbb{H}\).
### Real manifolds
Suppose now \(\mathcal{M}\) is a _real_ manifold of finite dimension \(m\).
\[\operatorname{Re}\mathcal{A} =-\int\frac{1}{2}\operatorname{Im}\omega(u,\dot{u})+ \operatorname{Re}\mathcal{E}(u)\;dt \tag{16a}\] \[\operatorname{Im}\mathcal{A} =\int\frac{1}{2}\operatorname{Re}\omega(u,\dot{u})-\operatorname{ Im}\mathcal{E}(u)\;dt, \tag{16b}\]
This suggests that we can, in principle, generate _different_ approximate time evolutions on \(\mathcal{M}\) from each real-valued action.
The forms \(\operatorname{Re}\omega\) and \(\operatorname{Im}\omega\) are antisymmetric nondegenerate bilinear forms on a _real_ phase space we will denote \(\mathbb{H}_{\mathbb{R}}\): There is a standard way to view a complex linear space \(\mathcal{V}\) as a real linear space \(\mathcal{V}_{\mathbb{R}}\), called _the realification of \(\mathcal{V}\)_, by restricting the field of multiplicative scalars to \(\mathbb{R}\). The set of vectors is the same. Multiplication \(u\mapsto iu\) still yields an element of \(\mathcal{V}_{\mathbb{R}}\), but now as a linear operator \(\mathbf{i}\) that satisfies \(\mathbf{i}^{2}=-\operatorname{Id}\). (Such an operator is called a complex structure on the vector space.) In particular, \(u\) and \(iu\) are linearly independent in \(\mathcal{V}_{\mathbb{R}}\), and the dimension of the space is therefore doubled. Moreover, suppose a nondegenerate bilinear form \(a:\mathcal{V}\times\mathcal{V}\to\mathbb{C}\) is given. Both \(\operatorname{Re}a\) and \(\operatorname{Im}a\) are now nondegenerate bilinear forms on \(\mathcal{V}_{\mathbb{R}}\). If \(a\) is symmetric/antisymmetric, then \(\operatorname{Re}a\) and \(\operatorname{Im}a\) are also symmetric/antisymmetric. We conclude that \((\mathbb{H}_{\mathbb{R}},\operatorname{Im}\omega)\) and \((\mathbb{H}_{\mathbb{R}},\operatorname{Re}\omega)\) are distinct symplectic linear manifolds.
Select a _real_ submanifold \(\mathcal{M}\subset\mathbb{H}_{\mathbb{R}}\), and find, in a similar manner as previously, that \(\delta\operatorname{Re}\mathcal{A}=0\) for all infinitesimal variations \(\delta u(t)\in T_{u}\mathcal{M}\) if and only if
\[\dot{u}=-P_{\operatorname{Im}}(u)J^{-1}d\operatorname{Re}\mathcal{E},\] (17a) where \[P_{\operatorname{Im}}(u)\] is the symplectic projection operator onto the _real_ tangent space \[T_{u}\mathcal{M}\] obtained from the symplectic form \[\operatorname{Im}\omega\] Similarly, \[\delta\operatorname{Im}\mathcal{A}=0\] for all infinitesimal variations
Illustration of infinite dimensional phase space, a submanifold \(\mathcal{M}\), and the symplectic projection that dictates time evolution on the manifold.
if and only if
\[\dot{u}=P_{\rm Re}(u)J^{-1}d\,{\rm Im}\,{\cal E}, \tag{17b}\]
The existence of the symplectic projection depends on the invertibility of its matrix in the tangent basis; see Section 3.4. Equivalently, we must require that \({\cal M}\) is a symplectic submanifold, where the symplectic form is non-degenerate by definition.
The equations of motion (17a) and (17b) are explicitly real equations of motion, and, in the case \({\cal M}=\mathbb{H}\), equivalent to the canonical equations of motion (6a) and (6b). For general submanifolds, however, the Cauchy-Riemann equations do not apply, and the two Euler-Lagrange equations are not equivalent, i.e., they generate distinct time evolutions on \({\cal M}\). On the other hand, it can happen that \({\cal M}\) is simply a re-expression of a complex manifold using real and imaginary parts of the complex coordinates. In that case, the Cauchy-Riemann equations again apply, and \({\rm Re}\,{\cal A}\) and \({\rm Im}\,{\cal A}\) are equivalent functionals, i.e., Eqs. (17a) and (17b) generate the same time evolution on \({\cal M}\).
### Manifold normalization
A word on normalization of bivariational approximation manifolds \({\cal M}\) is in place. Implicit in Eq. is the assumption that \(\langle\tilde{\psi},\psi\rangle=1\), and this normalization is preserved by the time evolution.
\[\rho=\frac{|\psi\rangle\,\langle\tilde{\psi}|}{\langle\tilde{\psi}|\psi\rangle}. \tag{18}\]
If the manifold \({\cal M}\) either satisfies \(\langle\tilde{\psi}|\psi\rangle=1\) everywhere, or if \({\cal M}\) contains phase and normalization scalings \((\alpha\tilde{\psi},\beta\psi)\in{\cal M}\) for any fixed \((\tilde{\psi},\psi)\in{\cal M}\), then the two principles of stationary action give the same solutions. In practice, such normalized or scale invariant manifolds are always easy to construct, given a manifold that does not initially satisfy the constraint.
### Euler-Lagrange equations in local coordinates
We express the complex and real Euler-Lagrange equations in terms of local coordinates, beginning with the complex case. Let \({\cal M}\subset\mathbb{H}\) be a complex submanifold of dimension \(n<+\infty\), for simplicity. Let \(u\in{\cal M}\subset\mathbb{H}\) be given in local coordinates by \(u=\Phi(z)\), with \(z\in\mathbb{C}^{n}\), and let \(t_{\mu}=(\tilde{v}_{\mu},v_{\mu})=\partial_{z^{\mu}}(\tilde{\psi},\psi)= \partial_{z^{\mu}}\Phi(z)\) define the coordinate basis vectors. By assumption, this is a linearly independent set. Any tangent vector \(\delta u\in T_{u}{\cal M}\) is now expanded as \(\delta u=\sum_{\mu}t_{\mu}\delta z^{\mu}\). The equations of motion are readily obtained by considering Eq.
\[i\Omega_{\mu\nu}\dot{z}^{\nu}=\frac{\partial{\cal E}(z)}{\partial z^{\mu}} \quad\forall\mu, \tag{19}\]
where
\[\frac{\partial{\cal E}(z)}{\partial z^{\mu}}=\langle\tilde{v}_{\mu}|H\psi \rangle+\langle\tilde{\psi}|Hv_{\mu}\rangle\,, \tag{20}\]
and where
\[\Omega_{\mu\nu}=\omega(t_{\mu},t_{\nu})=\langle\tilde{v}_{\mu}|v_{\nu}\rangle- \langle\tilde{v}_{\nu}|v_{\mu}\rangle \tag{21}\]
We used the Einstein summation convention. We note that the antisymmetry implies that (the complex dimension) \(\dim({\cal M})=n\) must be even. The matrix \(\Omega\) is invertible over the manifold if and only if \(({\cal M},\omega)\) is a symplectic submanifold.
Inverting \(\Omega\) in Eq., multiplying with \(t_{\nu}\) and summing, we obtain the following representation of the symplectic projection operator:
\[P(u)=t_{\nu}(\Omega^{-1})^{\nu\mu}\omega(t_{\mu},\;\cdot\;). \tag{22}\]
The calculation of the Euler-Lagrange equations for the real manifold case is very similar. We let \({\cal M}\subset\mathbb{H}_{\mathbb{R}}\) be a real submanifold of dimension \(n<+\infty\), and denote as before \((\tilde{v}_{\mu},v_{\mu})=(\partial_{x^{\mu}}\tilde{\psi},\partial_{x^{\mu}} \psi)\in T_{u}{\cal M}\;,\) the coordinate basis for tangent space at \(u=(\tilde{\psi},\psi)\in{\cal M}\), where now \(x=(x^{\mu})\) are real-valued local coordinates.
\[-\operatorname{Im}\Omega_{\mu\nu}\dot{x}^{\nu} =\operatorname{Re}\left(\frac{\partial\mathcal{E}(x)}{\partial x^{ \mu}}\right),\quad\text{and} \tag{23a}\] \[\operatorname{Re}\Omega_{\mu\nu}\dot{x}^{\nu} =\operatorname{Im}\left(\frac{\partial\mathcal{E}(x)}{\partial x ^{\mu}}\right). \tag{23b}\]
In both cases, the coefficient matrix is a real antisymmetric matrix. It follows that \(\dim(\mathcal{M})=n\) must be even both cases. Whenever the matrix inverses exist over the manifold, \((\mathcal{M},\operatorname{Re}\omega)\) and \((\mathcal{M},\operatorname{Im}\omega)\) are real symplectic manifolds, which must of even (real) dimension. By a similar argument as for Eq., we derive the following representations for the real symplectic projection operators:
\[P_{\operatorname{Im}}(u)=t_{\nu}((\operatorname{Im}\Omega)^{-1})^{\nu\mu}( \operatorname{Re}\omega)(t_{\mu},\,\cdot\,),\] (24a) and \[P_{\operatorname{Re}}(u)=t_{\nu}((\operatorname{Re}\Omega)^{-1})^{\nu\mu}( \operatorname{Im}\omega)(t_{\mu},\,\cdot\,). \tag{24b}\]
### Interpretation
Taking the real part of \(\mathcal{A}\) to define an action functional is not a novel idea, and dates back at least to Pedersen and Koch. The time-dependent optimized-orbital CC method of Sato and coworkers (see also Section 5) is defined in terms of taking the real part of a non-complex differentiable action. Indeed, for bivariational methods such as the coupled-cluster method, it is a well-known problem, or feature, that observables may attain non-zero imaginary values. It is customary to simply insist on taking the real part of the computed observable, and discard the (hopefully) small imaginary parts. This would be equivalent to using the action \(\operatorname{Re}\mathcal{A}\) and using the Hellmann-Feynman interpretation of expectation values. Indeed, for \(\operatorname{Re}\mathcal{A}\), the real part \(\operatorname{Re}\mathcal{E}\) of the energy expectation value is now the generator for the dynamics, and for a generic observable \(O\), the expectation value functional reads \(\operatorname{Re}\mathcal{E}_{O}(u)=\operatorname{Re}(\langle\!\langle u, \hat{O}u\rangle\!\rangle/\langle\!\langle u,\hat{I}u\rangle\!\rangle)\), where \(\hat{I}\) is the symmetrization of the identity operator.
On the other hand, we have also found that taking the _imaginary_ part \(\operatorname{Im}\mathcal{A}\) yields a distinct approximation when \(\mathcal{M}\) is real. Then, \(\operatorname{Im}\mathcal{E}\) is the generator for dynamics and hence conserved.
Using the imaginary part of the energy as generator for dynamics may seem odd. However, consider the following: For \(\operatorname{Re}\mathcal{A}\), the real part \(\operatorname{Re}\mathcal{E}\) is conserved in time, while we have no conservation law for \(\operatorname{Im}\mathcal{E}\). If the manifold \(\mathcal{M}\) is accurate, we can expect \(\operatorname{Im}\mathcal{E}\) to remain small. On the other hand, for \(\operatorname{Im}\mathcal{A}\), the imaginary part \(\operatorname{Im}\mathcal{E}\) is conserved - and if \(\mathcal{M}\) is accurate, it will be small - but it is now \(\operatorname{Re}\mathcal{E}\) that fluctuates, but should hopefully be _almost_ conserved. In this sense, the two principles are complementary, and reflect the ubiquitous compromise in bivariational theory, which is "non-Hermitian" in nature.
### Poisson brackets and conservation laws
### Poisson bracket
Let \(\mathcal{F}(u)\) and \(\mathcal{G}(u)\) be smooth scalar-valued functions of \(u\in\mathcal{M}\), for the moment assumed to be a complex submanifold of \(\mathbb{H}\).
\[\{\mathcal{F},\mathcal{G}\}:=-i(\nabla_{z}\mathcal{F})^{T}\Omega^{-1}\nabla_{ z}\mathcal{G}, \tag{25}\]
Due to the antisymmetry of \(\omega\),
\[\{\mathcal{F},\mathcal{G}\}=-\{\mathcal{G},\mathcal{F}\}, \tag{26}\]
and it is also readily shown that the bracket satisfies the Jacobi identity
\[\{\mathcal{F},\{\mathcal{G},\mathcal{H}\}\}+\{\mathcal{G},\{\mathcal{H}, \mathcal{F}\}\}+\{\mathcal{H},\{\mathcal{F},\mathcal{G}\}\}=0. \tag{27}\]
In the case where \(\mathcal{M}=\mathbb{H}\), the Poisson bracket takes the form
\[\{\mathcal{F},\mathcal{G}\}=i\langle\!\langle J^{-1}d\mathcal{F},d\mathcal{G }\rangle\!\rangle, \tag{28}\]
By using the representation we obtain a coordinate-free formula for the Poisson bracket on the complex manifold \(\mathcal{M}\)
\[\{\mathcal{F},\mathcal{G}\}=i\langle\!\langle P(u)J^{-1}d\mathcal{F},d\mathcal{G} \rangle\!\rangle. \tag{29}\]
The Poisson bracket \(\{\cdot,\mathcal{E}\}\) generates time evolution.
\[\frac{d\mathcal{F}}{dt}=\{\mathcal{F},\mathcal{E}\}. \tag{30}\]
In particular, the coordinates themselves are smooth over \(\mathcal{M}\), which gives
\[\frac{dz^{\mu}}{dt}=\{z^{\mu},\mathcal{E}\}. \tag{31}\]
### Ehrenfest's Theorem
Recall that for an operator \(A\) on \(\mathbb{H}\), we have defined its symmetrization \(\hat{A}\) on \(\mathbb{H}\to\mathbb{H}^{*}\) as \(\hat{A}(\tilde{\psi},\psi)=\frac{1}{2}(A\psi,A^{t}\tilde{\psi})\).
\[\langle\!\langle J^{-1}\hat{A}u,\hat{B}u\rangle\!\rangle=-\frac{1}{4}\langle \!\langle u,\widehat{[A,B]}u\rangle\!\rangle. \tag{32c}\]
We define the expectation value of \(A\) with respect to \((\tilde{\psi},\psi)\equiv u\in\mathbb{H}\) by
\[\langle\!\langle A\rangle\!\rangle_{u}:=\langle\!\langle u,\hat{A}u\rangle\! \rangle=\langle\tilde{\psi}|A\psi\rangle.\]
Let \(u(t)\) be a solution to Eq.. Then, using Eq. (32c) and the Poisson bracket with \(d\langle\!\langle A\rangle\!\rangle_{u}=2\hat{A}u\), one obtains the following Poisson bracket:
\[\{\langle\!\langle A\rangle\!\rangle_{u},\langle\!\langle B\rangle\!\rangle_{u }\}=-i\langle\!\langle[A,B]\rangle\!\rangle_{u}. \tag{33}\]
In particular, one obtains Ehrenfest's Theorem on bivariational form,
\[\frac{d}{dt}\langle\!\langle A\rangle\!\rangle_{u}=-i\langle\!\langle[A,H] \rangle\!\rangle_{u}. \tag{34}\]
### Conservation laws on complex manifolds
We next present the generalization of the Bir variational Ehrenfest theorem to manifolds. Let \(\mathcal{M}\subset\mathbb{H}\) be a complex manifold.
\[\forall u\in\mathcal{M}:J^{-1}\hat{A}u\in T_{u}\mathcal{M}.\]
Let \(u(t)=(\tilde{\psi}(t),\psi(t))\in\mathcal{M}\) be a solution to the bivariational principle on \(\mathcal{M}\), and hence a solution to the \(\mathcal{M}\)-projected Euler-Lagrange equation, Eq..
Using Eq., we see that
\[\frac{d}{dt}\langle\!\langle\hat{A}\rangle\!\rangle_{u}=\{\langle\!\langle A \rangle\!\rangle_{u},\mathcal{E}\}.\]
Applying Eq. (32b) to the Poisson bracket formula in Eq., and inserting \(d\langle\!\langle\hat{A}\rangle\!\rangle_{u}=2\hat{A}u\) and \(d\mathcal{E}=2\hat{H}u\), we get
\[\frac{d}{dt}\langle\!\langle\hat{A}\rangle\!\rangle_{u}=-4i\langle\!\langle J ^{-1}\hat{A}u,JP(u)J^{-1}\hat{H}u\rangle\!\rangle\]
If \(\hat{A}\) preserves \(\mathcal{M}\), then we can use the definition of \(P(u)\) in Eq. together with Eq. (32c) to obtain
\[\frac{d}{dt}\langle\!\langle\hat{A}\rangle\!\rangle_{u}=-i\langle\!\langle[A, H]\!\rangle\!\rangle_{u}\]
In general, \(\{\langle\!\langle A\rangle\!\rangle_{u},\langle\!\langle B\rangle\!\rangle_{u }\}=-i\langle\!\langle[A,B]\rangle\!\rangle_{u}\) if \(\hat{A}\) or \(\hat{B}\) preserves \(\mathcal{M}\). If \(\hat{A}\) does not preserve \(\mathcal{M}\), it is possible to express the deviation from the Ehrenfest theorem in terms of the distance between \(J^{-1}\hat{A}u\) and \(T_{u}\mathcal{M}\). One can thus obtain bounds for change in expectation value of observables.
\[\{\langle\!\langle A\rangle\!\rangle_{u},\langle\!\langle B\rangle\!\rangle_{u }\}=-i\langle\!\langle[A,B]\rangle\!\rangle_{u}\]
\[-4i\langle\!\langle u,\hat{B}Q(u)J^{-1}\hat{A}u\rangle\!\rangle,\]
Thus, for a preserved variable \(\langle\!\langle A\rangle\!\rangle\) evolving on a complex manifold, we obtain the bound of the error
\[\int\left|\frac{d}{dt}\langle\!\langle A\rangle\!\rangle\right|\!dt\leq 4\int| \langle\!\langle u,\hat{A}Q(u)J^{-1}\hat{H}u\rangle|dt\]
Such error bounds can be expressed in terms of the curvature of the manifold \(\mathcal{M}\),5 but we do not consider this further here.
## 4 Relation to unvariational theory
### TDVP vs.
The branching of the complex TD-BIVP into two distinct real bivariational principles is similar to the relationship between the time-dependent variational principle (TDVP) and the McLachlan variational principle for _univariational_ approximations of the Schrodinger equation.
\[\mathcal{S}=\int\left\langle\psi(t),\dot{\psi}(t)+iH\psi(t)\right\rangle\,dt \tag{35}\]
Using the fact that \(H\) is Hermitian it is straightforward to show that \(\delta\mathcal{S}=0\) if and only if \(i\dot{\psi}=H\psi\). Suppose now \(\mathcal{M}\subset\mathcal{H}\) is a _real_ submanifold.
\[\mathrm{Im}\langle\delta\psi,\dot{\psi}+iH\psi\rangle=0 \tag{36}\]
If we define the symplectic form \(\tilde{\omega}(\psi,\phi)=2i\,\mathrm{Im}\langle\psi,\phi\rangle\) on \(\mathcal{H}\), then, Eq.
\[\tilde{\omega}(\delta\psi,\dot{\psi}+iH\psi)=0\;\forall\psi\in T_{\psi} \mathcal{M}.\]
We may rephrase in terms of the symplectic projection \(P_{\tilde{\omega}}(\psi):\mathcal{H}\to T_{\psi}\mathcal{M}\), such that Eq.
\[\dot{\psi}=P_{\tilde{\omega}}(\psi)(-iH\psi). \tag{37}\]
The McLachlan variational principle states that \(\dot{\psi}=\Theta\), where \(\Theta\) is defined by
\[\Theta=\underset{\Theta\in T_{\psi}\mathcal{M}}{\mathrm{argmin}}\,\|iH\psi+ \Theta\|^{2}.\]
Differentiating \(\|iH\psi+\Theta\|^{2}\) with respect to \(\Theta\), equating to zero and inserting \(\dot{\psi}=\Theta\) gives
\[\mathrm{Re}\langle\delta\psi,\dot{\psi}+iH\psi\rangle=0, \tag{38}\]
The form \(\mathrm{Re}\left\langle\psi,\phi\right\rangle\) is the inner product on the realification \(\mathcal{H}_{\mathbb{R}}\), so that Eq.
\[\dot{\psi}=P_{\perp}(\psi)(-iH\psi), \tag{39}\]
Suppose now \(\mathcal{M}\) has the property that \(\delta\psi\in T_{\psi}\mathcal{M}\iff i\delta\psi\in T_{\psi}\mathcal{M}\). This happens if \(\mathcal{M}\) is actually a complex submanifold of \(\mathcal{H}\). Then Eq. and Eq. are equivalent, and the symplectic and orthogonal projections coincide, \(P_{\tilde{\omega}}(\psi)=P_{\perp}(\psi)\).
The TDVP and the McLachlan principles are distinct for real manifolds, but equivalent for complex manifolds, where the symplectic projection and orthogonal projections coincide. We contrast this with the bivariational case, where the projections \(P_{\mathrm{Re}}(u)\) and \(P_{\mathrm{Im}}(u)\) are distinct for real manifolds, and \(P_{\mathrm{Re}}(u)=P_{\mathrm{Im}}(u)=P(u)\) for complex manifolds. Notably, in all cases we deal with the symplectic projection.
### TD-BIVP contains the TDVP, but not the McLachlan principle
Consider the approximate univariate dynamics on a real or complex submanifold \(\mathcal{M}\subset\mathcal{H}\) generated by the TDVP (Eq.). We denote by \(\Phi_{t}:\mathcal{M}\rightarrow\mathcal{M}\) the TDVP flow, such that given an initial condition \(\psi\in\mathcal{M}\), we have \(\psi(t)=\Phi_{t}(\psi)\).
Let \(j:\mathcal{H}\rightarrow\mathbb{H}\) be the map \(j(\psi)=(\bar{\psi},\psi)\), where \(\bar{\psi}\) is the injection of \(\mathcal{H}\) into \(\mathcal{H}^{*}\). (It is not the complex conjugate, which does not exist for general complex Hilbert spaces, but rather a "renaming" of \(\psi\) so that it lies in \(\mathcal{H}^{*}\).) A straightforward calculation shows that
\[\omega(j(\psi),j(\phi))=\tilde{\omega}(\psi,\phi), \tag{40}\]
Moreover, let \(\mathcal{N}=j[\mathcal{M}]\). Wehave \(T_{j(\psi)}\mathcal{N}=j[T_{\psi}\mathcal{M}]\) for any \(\psi\in\mathcal{M}\). The manifold \(\mathcal{N}\) is complex if and only if \(\mathcal{M}\) is complex. The symplectic form \(\omega\) on \(\mathbb{H}\) is purely imaginary on \(j[\mathcal{M}]\), so we select the real symplectic form \(\operatorname{Im}\omega\) and thus the real part of the bivariational action, Eq. (16a), and generate a flow \(\Psi_{t}:\mathcal{N}\to\mathcal{N}\) such that \(u(t)=\Psi_{t}(u)\).
It is now a straightforward calculation to show that the following diagram commutes:
In other words, the TDVP is contained in the TD-BIVP on the form \(\operatorname{Re}\delta\mathcal{A}=0\), when promoting a real manifold \(\mathcal{M}\) to \(j[\mathcal{M}]\) as bivariational manifold. If \(\mathcal{M}\) is complex, we may use any bivariational action, real or complex.
We now turn to the McLachlan principle, which uses an orthogonal projection. For a complex submanifold \(\mathcal{M}\subset\mathcal{H}\), the McLachlan principle and the TDVP coincide. If, however, \(\mathcal{M}\) is a real manifold, \(P_{\perp}(\psi)\) is not a symplectic projection, indicating that the McLachlan principle is not contained in the BIVP in this case. Indeed, we instead find that it is a special case of the following bivariational version of the McLachlan principle:
\[\Delta=\langle\!\langle\delta u,\hat{I}(\dot{u}+2iJ^{-1}\hat{H}u)\rangle\! \rangle=0,\quad\forall\delta u\in T_{u}\mathcal{N}, \tag{41}\]
The principle does not seem to be derivable from a stationary action. The reason is that the first term \(\langle\!\langle\delta u,\hat{I}\dot{u}\rangle\!\rangle\) is a symmetric bilinear form, and the functional \(\int\langle\!\langle u,\hat{I}\dot{u}\rangle\!\rangle dt\) is therefore a total time derivative.
To derive Eq., we note that
\[\Delta \equiv\operatorname{Re}\left\langle\delta\psi,\dot{\psi}+iH\psi \right\rangle=\operatorname{Im}\left\langle\delta\psi,i\dot{\psi}-H\psi\right\rangle\] \[=\frac{1}{2i}\omega(j(\delta\psi),j(i\dot{\psi}-H\psi)). \tag{42}\]
For a point \(u=j(\psi)\in\mathcal{N}\), we now have the relations \(\dot{\psi}=j^{-1}\dot{u}\) and \(jH\psi=\hat{I}^{-1}\hat{H}u\), verified by direct computation. Moreover, \(jj^{-1}=\frac{1}{2}i\hat{I}^{-1}J\), also verified by direct computation.
\[\Delta=\frac{1}{2}\omega(\delta u,\hat{I}^{-1}(J\dot{u}+2i\hat{H}u)).\]
At this point, we insert the definition of the symplectic form, with some easy manipulations, to obtain Eq..
## 5 Unifying view of current time-dependent bivariational approaches
The abstract formalism in this article is applicable to many methods for real-time propagation encountered in the literature. In this section, we give a brief overview of what we consider to be some important examples in the language of the present work, all being varieties of coupled-cluster (CC) theory: Time-dependent traditional CC theory (TDCC), orbital-adaptive time-dependent CC (OATDCC) theory, orthogonal orbital-optimized time-dependent CC (TD-OCC) theory, and time-dependent equation-of-motion CC (TD-EOM-CC) theory.
Familiarity with CC theory is assumed, and this section will only serve as an overview. We refer to the original publications for full details and complete specifications. For a recent review of real-time propagation with CC theory, Ref..
### The traditional CC ansatz
The traditional CC method is the most popular wavefunction-based method for electronic-structure theory, with the CCSD(T) model often termed "the gold standard of quantum chemistry" due to its balance of cost and accuracy. Similarly, the CC method offers an attractive approach for approximating the solution to the vibrational Schrodinger equation. Although the physical nature of the degrees of freedom and of the Hamiltonian is very different in the electronic and vibrational cases, both cases benefit from fast convergence of the CC hierarchy, polynomial-scaling cost, and size extensivity.
The traditional CC method is usually formulated in finite-dimensional subspace of \(\mathcal{H}\), defined in terms of a finite orthonormal set of single-particle functions (a "basis set"). However, we here take a broader picture, and merely assume a _biorthogonal_ set of single-particle functions partitioned into occupied and unoccupied subsets, \(\mathcal{B}=\mathcal{B}_{\mathrm{occ}}\cup\mathcal{B}_{\mathrm{unocc}}\), and \(\tilde{\mathcal{B}}=\tilde{\mathcal{B}}_{\mathrm{occ}}\cup\tilde{\mathcal{B}}_ {\mathrm{unocc}}\).. This induces a biorthogonal many-particle basis set \(\{\phi_{\mu}\}\subset\mathcal{H}\) and \(\{\tilde{\phi}_{\mu}\}\subset\mathcal{H}^{*}\) with \((\tilde{\phi},\phi)=(\tilde{\phi}_{0},\phi_{0})\) being formal excitation references for bras and kets, respectively. The full configuration-interaction wavefunction and its dual are written \(\psi=C\phi\) and \(\tilde{\psi}=\tilde{C}^{t}\tilde{\phi}\), with \(C=\sum_{\mu}c_{\mu}X_{\mu}\in\mathcal{C}\) and \(\tilde{C}=\sum_{\mu}\tilde{c}_{\mu}\tilde{X}_{\mu}\in\tilde{\mathcal{C}}\) being cluster operators (the summations run over \(\mu\geq 0\)). Here, \(X_{\mu}\phi=\phi_{\mu}\) is an elementary excitation, and similarly \(\tilde{X}_{\mu}^{t}\tilde{\phi}=\tilde{\phi}_{\mu}\) is an elementary dual excitation, or de-excitation. Note that \(X_{0}\phi=\phi\) and \(\tilde{X}_{0}^{t}\tilde{\phi}=\tilde{\phi}\), i.e. \(X_{0}\) and \(\tilde{X}_{0}\) are simply identity operators or null excitations. The notation \(A^{t}\) is defined via the dual pairing \(\langle A^{t}\phi|\psi\rangle\equiv\langle\phi|A\psi\rangle\). Equivalently, using bra notation, \(\langle A^{t}\tilde{\phi}|=\langle\tilde{\phi}|A\).
The spaces \(\tilde{\mathcal{C}}\) and \(\mathcal{C}\) are nilpotent abelian algebras that depend on the subdivisions \(\mathcal{B}=\mathcal{B}_{\mathrm{occ}}\cup\mathcal{B}_{\mathrm{unocc}}\) and \(\tilde{\mathcal{B}}=\tilde{\mathcal{B}}_{\mathrm{occ}}\cup\tilde{\mathcal{B}} _{\mathrm{unocc}}\) of the single-particle basis set into occupied and unoccupied subsets. The spaces \(\tilde{\mathcal{C}}\) and \(\mathcal{C}\) are duals to each other, with the dual pairing being given by \(\langle\tilde{\phi}|\tilde{C}C\phi\rangle=\sum_{\mu}\tilde{c}_{\mu}c_{\mu}\).
In traditional CC theory, the phase-space point \((\tilde{\psi},\psi)\) is parameterized in terms of a pair of cluster operators \((\Lambda,T)\in\tilde{\mathcal{C}}\oplus\mathcal{C}\) as
\[\psi=e^{T}\phi,\quad\tilde{\psi}=e^{-T^{t}}\Lambda^{t}\tilde{\phi}, \tag{43}\]
Again, \(\mu=0\) is included in the summations. \(\tau_{0}\) plays the role of a phase/norm factor, while \(\lambda_{0}\) determines the intermediate normalization in the sense \(\langle\tilde{\psi}|\psi\rangle=\lambda_{0}\). Equation defines a map \(\Phi(\Lambda,T)=(\tilde{\psi},\psi)\) being a global coordinate chart for a smooth complex submanifold \(\mathcal{M}_{\mathrm{CC}}\subset\mathbb{H}\). In fact, this submanifold is covers _almost_ all possible points in \(\mathbb{H}\). The additional conditions are \(\langle\tilde{\psi}|\psi\rangle\neq 0\) and \(\langle\tilde{\phi}|\psi\rangle\neq 0\).
The energy functional in these coordinates is
\[\mathcal{E}(\Lambda,T)=\langle\tilde{\phi}|\Lambda e^{-T}He^{T}\phi\rangle\,, \tag{44}\]
the conventional CC Lagrangian (which is a Lagrangian in the sense of constrained optimization, and must not to be confused with the Lagrangian density encountered in the bivariational principle), and the action functional reads
\[\mathcal{A}_{\mathrm{CC}}=\int i\lambda\cdot\dot{\tau}-\mathcal{E}(\Lambda,T )\;dt. \tag{45}\]
(The integrand is the Lagrangian density in our language.) In particular, the functional form is preserved compared to Eq.. This means, that the coordinate transformation \((\tilde{\psi},\psi)=\Phi(\Lambda,T)\) is a canonical transformation in the sense of classical mechanics, and it follows that Hamilton's equations of motion (both the complex and real forms) are preserved as well.
The full CC case is not practical, and conventional truncation schemes \(\mathcal{T}\) of the cluster operators imply an approximate submanifold \(\mathcal{M}_{\mathrm{CC}}(\mathcal{T})\subset\mathcal{M}_{\mathrm{CC}}\). For electronic-structure theory, \(\mathcal{T}\) is typically the CCSD\(\cdots K\) scheme, where all excitations of up to \(K\) electrons are included. In the vibrational case, the analogous approach is usually denoted VCC\([K]\) and includes up to \(K\)-mode excitations. Since the coordinates are canonical, the induced symplectic form on \(\mathcal{M}_{\mathrm{CC}}(\mathcal{T})\) is trivially non-degenerate, and the submanifold is always symplectic.
Since the coordinates \((\Lambda,T)\) are canonical, the Poisson bracket takes on the simple form
\[\begin{split}\{\mathcal{F},\mathcal{G}\}&=i\langle \!\langle J^{-1}d\mathcal{F},d\mathcal{G}\rangle\!\rangle\\ &=i\left(\left\langle\frac{\partial\mathcal{G}}{\partial\tau} \Big{|}\frac{\partial\mathcal{F}}{\partial\lambda}\right\rangle-\left\langle \frac{\partial\mathcal{F}}{\partial\tau}\Big{|}\frac{\partial\mathcal{G}}{ \partial\lambda}\right\rangle\right).\end{split} \tag{46}\]
For expectation values, \(\mathcal{F}=\langle\tilde{\psi}|F\psi\rangle=\langle\tilde{\phi}|\Lambda e^{-T }Fe^{T}|\phi\rangle\), we have \(\partial\mathcal{F}/\partial\tau^{\mu}=\langle\tilde{\phi}|\Lambda[\tilde{F},X _{\mu}]|\phi\rangle\), where \(\tilde{F}=e^{-T}He^{T}\), and \(\partial\mathcal{F}/\partial\lambda^{\mu}=\langle\tilde{\phi}_{\mu}|\tilde{F}\phi\rangle\).
Evaluation of the Poisson bracket yields
\[\{\mathcal{F},\mathcal{G}\} =-i\left\langle\tilde{\psi}|[F,G]\psi\right\rangle\] \[+i\sum_{\mu}((\tilde{\phi}|\Lambda X_{\mu}\bar{G}|\phi)\left\langle \tilde{\phi}_{\mu}|\bar{F}|\phi\right\rangle\] \[\qquad\qquad-\left\langle\tilde{\phi}|\Lambda X_{\mu}\bar{F}|\phi \right\rangle\left\langle\tilde{\phi}_{\mu}|\bar{G}|\phi\right\rangle) \tag{47}\]
In particular, if we set \(\mathcal{F}=\mathcal{E}\), we obtain that \(\mathcal{G}\) is exactly conserved under dynamics only if the last two terms vanish.
### The OACC ansatz
The CC ansatz described above is defined in terms of a _static_ single-particle basis. In particular, the references \((\tilde{\phi},\phi)\) are static, which is a serious limitation in terms of describing, say, large oscillations in the wavefunction, or motion far away from the ground state. The traditional CC ansatz works well when the amplitudes are sufficiently small, i.e., when the reference describes a large part of the wave function. Conversely, if the wave function moves too far from the reference, the amplitudes grow and the ansatz tends to break down. The paradigmatic example of such a situation is ionization or dissociation, i.e. the removal of one or more particles from the system (typically by a laser pulse). However, much less violent phenomena can also initiate the breakdown of the CC ansatz, as exemplified in vibrational CC theory by the internal vibrational energy redistribution (IVR) in water.
This problem can be alleviated by introducing an _adaptive_ single-particle basis: Both the occupied and unoccupied single-particle bra and ket basis functions are time dependent. This in turn defines a pair of adaptive references _and_ excited determinants that move with the wave function, so to speak. In the present exposition, we assume that the span of the single-particle bases \(\mathcal{B}\) and \(\tilde{\mathcal{B}}\) are fixed. Equivalently, there is no _secondary space_, and the CC parameterizaton is allowed to correlate all available single-particle functions. This means that the OACC ansatz becomes formally equivalent to non-orthogonal orbital-optimized CC theory (NOCC).
Consider therefore a pair of uncorrelated states \((\tilde{\phi},\phi)\in\mathcal{H}^{*}\oplus\mathcal{H}\) satisfying \(\langle\tilde{\phi}|\phi\rangle=1\). The set of such binormalized reference determinants is a smooth submanifold \(\mathcal{U}\subset\mathbb{H}\). Every element \((\tilde{\phi},\phi)\in\mathcal{U}\) defines a unique CC manifold \(\mathcal{M}_{\text{CC}}(\tilde{\phi},\phi)\).
\[\mathcal{M}_{\text{OACC}}(\mathcal{T})=\bigcup_{(\tilde{\phi},\phi)\in \mathcal{U}}\mathcal{M}_{\text{CC}}(\tilde{\phi},\phi;\mathcal{T}). \tag{48}\]
It has been found that singles excitations must be removed from \(\mathcal{T}\) in order to produce a well-defined manifold.
The manifold \(\mathcal{M}_{\text{OACC}}(\mathcal{T})\) is a complex manifold due to the independence of the bra and ket single-particle functions.
The orbital-adaptive time-dependent coupled-cluster (OATDCC) ansatz was first introduced by Kvaal for describing electron dynamics. An analogous ansatz for the vibrational problem was proposed by Madsen et al. under the name time-dependent modal vibrational coupled-cluster (TDMVCC). The formulation by Kvaal is done in an abstract infinite-dimensional setting, but, for simplicity, we consider here a finite basis and use the exponential parameterization of the manifold \(\mathcal{U}\) of reference determinants.
\[\phi^{\prime}=e^{\kappa}\phi,\quad\tilde{\phi}^{\prime}=e^{-\kappa^{t}}\tilde {\phi}. \tag{49}\]
Here, \(\kappa\) is a generic (neither Hermitian nor anti-Hermitian) one-particle operator,
\[\kappa=\sum_{pq}\kappa_{pq}\tilde{c}_{p}^{\dagger}c_{q} \tag{50}\]
We have introduced the creation and annihilation operators associated with the biorthogonal single-particle basis of the traditional CC method. In fact, \(\exp(\kappa)\) and \(\exp(-\kappa^{t})\) are change of single-particle basis operators.
\[(\tilde{\psi}^{\prime},\psi^{\prime})=(e^{-\kappa^{\prime}}\tilde{\psi},e^{\kappa} \psi)\in{\cal M}_{\rm CC}(\tilde{\phi}^{\prime},\phi^{\prime};{\cal T}). \tag{51}\]
However, the parameterization contains redundancies that must be eliminated or fixed by a suitable gauge condition. The source of the redundancy is the invariance of the CC wavefunctions under mixings of occupied single-particle functions and unoccupied single-particle functions separately. Thus, one allowed gauge condition (at \(\kappa=0\)) is to only keep all elements in \(\kappa\) that mix occupied single-particle functions [those that comprise \((\tilde{\phi},\phi)\)] with unoccupied single-particle functions, and set all other elements to zero. The mathematical structure is that of a principal bundle.
It is instructive to consider the Lagrangian, i.e. the integrand of the action, Eq.:
\[\begin{split}{\cal L}_{\rm OACC}&=i\,\langle\tilde{ \psi}^{\prime}|\dot{\psi}^{\prime}\rangle-\langle\tilde{\psi}^{\prime}|H\psi^{ \prime}\rangle\\ &=i\,\langle\tilde{\psi}|\dot{\psi}\rangle-\langle\tilde{\psi}|( \bar{H}-G)\psi\rangle\,.\end{split} \tag{52}\]
Here, \(\bar{H}=e^{-\kappa}He^{\kappa}\) and \(G=ie^{-\kappa}(de^{\kappa}/dt)=-i(de^{-\kappa}/dt)e^{\kappa}\). \({\cal L}_{\rm OACC}\) is in fact identical to the integrand of the action in traditional CC theory, cf. Eq., provided we substitute \(H\leftarrow\bar{H}-G\), so the AOCC amplitude equations have the same form as the CC amplitude equations. Stationarity of \({\cal A}_{\rm OACC}\) leads to a set of linear equations for \(G\), which in turn determines \(\dot{\kappa}\). At the point \(\kappa=0\), the relation is particularly simple, namely \(G=i\dot{\kappa}\) (the \(\kappa\neq 0\) case has been treated in detail in Ref.). We remark that Refs. and did not use the exponential parameterization for the single-particle basis, but the linear equations are the same.
### Orthogonal orbital-optimized CC
The orthonormal orbital-optimized CC (OCC) ansatz is identical to the OACC ansatz except that the single-particle basis is restricted to being orthonormal. This induces an orthonormal many-particle basis and, in particular, \(\tilde{\phi}=\phi^{\dagger}\).
\[{\cal M}_{\rm OCC}=\bigcup_{(\phi^{\dagger},\phi)\in{\cal U}}{\cal M}_{\rm CC} (\phi^{\dagger},\phi). \tag{53}\]
Whereas \({\cal M}_{\rm OACC}\) was a complex manifold, the OCC manifold is a _real_ manifold, which can be seen by realizing that the "coordinates" being single-particle basis functions appear both as complex conjugates and as they are. Consequently, one must turn to one of the real action principles \(\delta\,{\rm Re}\,{\cal A}_{\rm OCC}=0\) or \(\delta\,{\rm Im}\,{\cal A}_{\rm OCC}=0\), which lead to distinct time evolutions. Sato et al. used \({\rm Re}\,{\cal A}\), which amounts to letting \({\rm Re}\,{\cal E}\) generate the time evolution. As the physical energy is a real quantity, this seems to be the natural choice.
The CC amplitudes \(\lambda\) and \(\tau\) appears in a complex differentiable manner in the OCC ansatz, so \({\rm Re}\,{\cal A}_{\rm OCC}\) leads to the same amplitude equations as \({\cal A}_{\rm OCC}\). The OCC and OACC amplitude equations are thus identical, and both are essentially identical to the traditional CC amplitude equations. The linear equations that determine the OCC basis set evolution are, however, symmetrized in OCC compared to the OACC equations. As an example, the density matrices that appear in the OACC working equations are Hermitianized in the OCC working equations (cf. Refs. and).
One may consider appealing to the action principle \(\delta\,{\rm Im}\,{\cal A}_{\rm OCC}\). This will again lead to the same amplitude equations as before, while the basis set equations will contain anti-Hermitianized density matrices etc. While it seems like an unconventional choice, it may be worth investigating in future research.
### Time-dependent EOM-CC
Equation-of-motion CC theory is perhaps the simplest example of a bivariational method that is not explicitly Hermitian. In EOM-CC, a ground-state calculation is first performed with traditional CC theory, producing a cluster operator \(T_{0}\in{\cal C}\) (and also a \(\Lambda_{0}\in\tilde{\cal C}\)) such that \(\psi_{0}=e^{T_{0}}\phi\) is an approximate ground state. Using the theory of linear response, excited-state energies are approximated by the eigenvalues of a "dressed", or effective Hamiltonian, \(A=Pe^{-T_{0}}He^{T_{0}}P\), where \(P\) is the orthogonal projector onto the configuration-interaction space, usually at the same level of truncation \(\mathcal{T}\) as the underlying CC calculation. Thus, a bivariate Rayleigh quotient is set up, \(\mathcal{E}(L,R)=\left\langle\tilde{\phi}|LAR\phi\right\rangle/\left\langle \tilde{\phi}|LR|\phi\right\rangle\), where \((L,R)\in\tilde{\mathcal{C}}\oplus\mathcal{C}\). The left and right eigenvectors of \(A\) are treated as approximations to excited states.
\[\mathcal{A}_{\text{EOM-CC}}=\int i\left\langle\tilde{\phi}|L\dot{R}|\phi \right\rangle-\left\langle\tilde{\phi}|LAR|\phi\right\rangle\;dt, \tag{54}\]
## 6 Conclusion
In this article, we studied the time-dependent bivariational principle, and employed a differential geometric point of view. We introduced an action principle \(\delta\mathcal{A}=0\), where the field variables are the wavefunction and its complex conjugate \((\tilde{\psi},\psi)\). Approximate propagation techniques of bivariational type are then obtained by restricting these to lie in a smooth submanifold, or ansatz space.
We demonstrated that taking the real and imaginary parts \(\operatorname{Re}\mathcal{A}\) and \(\operatorname{Im}\mathcal{A}\) resulted in two independent variational principles that both reproduce exact dynamics when no approximations in the wavefunctions are introduced. A distinction was further made of approximate methods depending on the ansatz space being parameterized with complex or real coordinates. When complex coordinates are used, all variational principles are equivalent. When real coordinates are used, the real and imaginary principles are not always equivalent. Comparison with the time-dependent (uni-)variational principle and the McLachlan variational principle were made.
The imaginary principle is valid, yet its physical meaning is presently unclear, since the generator for time evolution is not the real part of the energy, but instead the imaginary part. It is not from the outset "unphysical", since bivariational methods invariable introduce formally compled-valued energies and expectation values. The imaginary principle ensures that the imaginary part of the energy is conserved and thus guaranteed to remain small. We relegate the study of this principle to future investigations.
In analogy with classical mechanics, Poisson brackets were introduced that allow analogy with the transition from classical to quantum mechanics. In particular, time evolution of observables become Poisson brackets.
In the final section of the article, we formulated various methods for real-time propagation using the TD-BIVP. We formulated time-dependent traditional coupled-cluster theory, orbital-adaptice time-dependent coupled-cluster theory and orthogonal orbital-optimized coupled-cluster theory, where the single-particle functions are allowed to move during dynamics, and time-dependent equation-of-motion coupled cluster theory.
|
10.48550/arXiv.2410.24192
|
The time-dependent bivariational principle: Theoretical foundation for real-time propagation methods of coupled-cluster type
|
Simen Kvaal, Håkon Richard Fredheim, Mads Greisen Højlund, Thomas Bondo Pedersen
| 5,489
|
10.48550_arXiv.2405.17246
|
###### Abstract
As pioneering experiments have shown, strong vibrational coupling between molecular vibrations and light modes in an optical cavity can significantly alter molecular properties and even affect chemical reactivity. However, the current theoretical description is limited and far from complete. To explore the origin of this exciting observation, we investigate how the molecular structure changes under strong light-matter coupling using an _ab-initio_ method based on the cavity Born-Oppenheimer Hartree-Fock ansatz. By optimizing H\({}_{2}\)O and H\({}_{2}\)O\({}_{2}\) resonantly coupled to cavity modes, we study the importance of reorientation and geometric relaxation. In addition, we show that the inclusion of one or two cavity modes can change the observed results. On the basis of our findings, we derive a simple concept to estimate the effect of the cavity interaction on the molecular geometry using the molecular polarizability and the dipole moments.
## I Introduction
When molecules are placed in a non-classical photonic environment present in optical or nanoplasmonic cavities, it is possible to form strong light-matter-coupled hybrid states called polaritons. The control of the photonic environment allows to couple the cavity photon modes to vibrational or electronic transitions in molecules, called vibrational-strong coupling (VSC) or electronic-strong coupling (ESC), respectively. Both types of strong coupling can be an effective tool for modifying molecular properties and offer a possible novel approach to control chemical reactions using optical resonators. The experimental advances reported cover a wide range of applications, from manipulating the selectivity of organic reactions, changing the ionic conductivity of water to even influencing enzymatic activity. Driven by these experimental advances, considerable efforts have been made to develop theories that elucidate the mechanisms governing polaritonic chemistry. Even so, the current theoretical description is limited and far from complete. However, in recent years, a substantial number of studies using different theoretical approaches have proposed that a variety of additional reactions can be enhanced, inhibited, or controlled. In particular, the combination of electronic structure methods and quantum electrodynamics has significantly improved theoretical understanding and will hopefully help close the existing gaps between theory and experiment. Most of these studies model cavity-induced electronic structure changes using a single molecule coupled to a single-cavity mode in the strong-coupling limit. In almost every example, both a fixed orientation relative to the polarization axes of the cavity mode and fixed molecular geometries are assumed, despite the fact that the coupling strengths used are quite large.
In this work, we use the cavity Born-Oppenheimer Hartree-Fock (CBO-HF) ansatz together with analytical gradients to optimize H\({}_{2}\)O and H\({}_{2}\)O\({}_{2}\) resonantly coupled to one and two cavity photon modes. The CBO-HF ansatz is capable of describing the electronic ground state of single molecules, as well as an ensemble of molecules coupled to an optical cavity under VSCs conditions. By performing the optimizing the geometries in the laboratory frame together with the polarization vectors of the cavity, we are able to study both the orientation and the relaxation of the internal coordinates of the molecules induced by the interaction with the cavity photon modes. Furthermore, we calculate vibro-polaritonic IR spectra within the harmonic approximation, we allow for the verification of the structures found as real minima and to analyze in detail the formed polaritonic states. As will become clear in the remainder of this article, without a physical mechanism to fix the orientation, molecules inside an optical cavity will orient and change their geometry depending on the coupling strength. Furthermore, these observed effects will vary if one- or two-cavity orthogonal photon modes are included in the simulation. The main results of this work are indication that some theoretical studies in the literature may overestimate the cavity-induced effects on the ground-state chemistry. Finally, we establish a useful and straightforward connection between the molecular polarizability and dipole moment and the expected reorientation and relaxation of a molecule coupled to an optical cavity.
## II Results & Discussion
### Optimization of H\({}_{2}\)O embedded in a cavity
As a first example, we optimize a H\({}_{2}\)O molecule coupled to one photon mode of an optical cavity. The cavity frequency \(\omega_{m}\) is resonant with the bending mode, which has a field-free vibrational frequency of 1744 cm\({}^{-1}\), and the cavity polarization axis \(\mathbf{e}\) is aligned with the \(x\) axis of the laboratory frame. shows the optimized parameters for the coupled molecular cavity system as afunction of the coupling strength \(\lambda_{m}\) and the corresponding vibro-polaritonic IR spectra in the region of the H\({}_{2}\)O bending mode.
For the optimized single-mode cavity-H\({}_{2}\)O system, the magnitude of the dipole \(|\langle\mu\rangle|\) moment, shown in a), is only slightly reduced with increasing \(\lambda_{m}\). This rather small change is consistent with the observed small decrease in both the OH bond length (c)) and bond angle (d)) even at high coupling strengths. The only system parameter significantly affected by VSC is the relative orientation of the molecular dipole moment and the polarization axis of the cavity mode visualized as the angle \(\phi\) between them in b). For the arbitrary chosen initial configuration, \(\phi\) has a value of about 55\({}^{\circ}\) (dashed dotted line), which changes to exactly 90\({}^{\circ}\) after optimization, regardless of the coupling strength. The interaction with teh photon field leads to an orientation of the molecule so that not only the dipole moment but also the molecular plan is orthogonal to the polarization axis. The relevant parts of the vibro-polaritonic IR spectra for the optimized water-cavity system are shown in e) and f) for different coupling strengths. Given the orientation of the dipole moment perpendicular to the cavity polarization axis, one would expect no hybrid photonic states to be formed because of the lack of dipole-cavity interaction. This assumption holds for lower coupling strengths below 0.062 au, shown in e). Within the used broadening of 10 cm\({}^{-1}\), only a single peak is observed, almost unshifted to the field-free value (black dashed line). For \(\lambda_{m}=0.062\) au a small shoulder appears at a slightly lower frequency, the bright purple line in e). If the coupling is further increased, as shown in f), this shoulder becomes a separate peak with increasing intensities and shifts to lower frequencies with increasing \(\lambda_{m}\). As introduced in our early work, the normal mode component that describes the change in the classical photon displacement field is a measure of how much photon character the corresponding transition is. The main peak in all shown vibro-polaritonic IR spectra has no photonic character and can be described as a pure molecular transition corresponding to the bending mode. The smaller peak at higher coupling strengths is predominantly photonic, and we do not see the formation of a typical hybrid matter-photon lower polariton (LP) state and upper polariton (UP) state here. This can be explained by a rotational motion of the H\({}_{2}\)O molecule. Due to the confinement of the cavity, this motion is not a free rotation anymore. This motion creates a dipole moment parallel to the cavity polarization axis, which leads to a coupling with the photon mode. Due to this coupling, the photonic transition gains intensity and shifts to lower frequencies. More details and a thorough analysis of the photonic characters of all relevant transitions can be found in the Supporting Information.
Next, we discuss the optimization results for H\({}_{2}\)O coupled to two cavity modes of orthogonal polarization with the same frequency, effectively modeling a Fabry-Perot-like setup. In practice, we use the same cavity frequency and coupling strength for both modes and take the same values as for the single-mode case. The polarization axes are aligned with the \(x\) axis (\(\mathbf{e}_{1}\)) and the \(y\) axis (\(\mathbf{e}_{2}\)) of the laboratory frame.
When the optimized parameters are compared for the single-mode case and the two-mode case, some similarities but also distinct differences are observed. The change in magnitude of the dipole moment, shown in a), is nearly one order of magnitude smaller for the two-mode optimization. In contrast, the OH bond length (c)) and the bond angle (d)) are more strongly influenced by the coupling to two cavity modes. In particular, the change in the bond angle is significantly larger compared to the case of a single mode, although it increases only by 1.5\({}^{\circ}\) for the highest coupling strength. Again, the most significant changes due to optimization are the relative orientations of the molecular dipole moment with respect to the cavity-mode polarization axes. The corresponding angles \(\phi\) for \(\mathbf{e}_{1}\) and \(\mathbf{e}_{2}\) are shown in b). Starting from a configuration where both \(\mathbf{e}_{1}\) and \(\mathbf{e}_{2}\) are in the molecular plane but not aligned with the dipole moment, the optimization reorients the molecule independently of the coupling strength so that \(\mathbf{e}_{1}\) is parallel to the dipole moment, while \(\mathbf{e}_{2}\) is orthogonal to the dipole moment and the molecular plane.
Optimized parameters of a H\({}_{2}\)O molecule coupled to a single photon mode of an optical cavity as a function of the coupling strength \(\lambda_{m}\). a) Magnitude of the dipole moment \(|\mu|\) b) the angle \(\phi\) between the polarization axis of the cavity and the dipole moment, c) averaged OH bond length and d) bond angle. The dashed-dotted lines in a)-d) indicate the initial values. Relevant parts of the vibro-polaritonic IR spectra for different coupling strengths (color-coded) are shown in e) and f). The cavity frequency \(\omega_{m}\) is 1744 cm\({}^{-1}\) shown as a black dashed line in e) and f). The cavity coupling \(\lambda_{m}\) increases from 0.015 au to 0.123 au and the cavity polarization axis is \(\mathbf{e}=\).
Already, for the lowest coupling strength \(\lambda_{m}=0.015au\) in e), a splitting into a LP transition and a UP transition is observed. The corresponding normal modes clearly show a hybridization of the vibrational transition and the cavity photon mode with the polarization axis \(\mathbf{e}_{1}\). As \(\lambda_{m}\) increases, the Rabi splitting between the LP and UP transitions increases and becomes more asymmetric, consistent with our previous work. At the same time, a third weaker peak between the LP and UP transitions becomes visible. Similarly to the single-mode case, this peak is due to the coupling between a specific constricted rotational degree of freedom and the photon mode with the polarization axis \(\mathbf{e}_{2}\). Due to this coupling, the photonic transition gains intensity and shifts to low frequencies with increasing coupling strength. More information on the photonic characters of all relevant transitions can be found in the Supporting Information.
To gain more insight into the underlying driving force for the energetically favorable orientation, we discuss how VSC modifies the polarizability \(\alpha\) and the dipole self-energy (DSE) contribution. The principal components of the polarizability tensor \(\alpha\) and the change in the DSE contribution are shown in as a function of the coupling strength for both optimized H\({}_{2}\)O-cavity systems.
In the single-mode case shown in a) and the two-mode case shown in b) all three principal components of \(\alpha\) are reduced with increasing coupling strength or, in other words, the interaction with the photon mode contracts the electronic density. Therefore, it is logical that the components aligned with the cavity polarization axes (highlighted in bold) are the most affected. These components already have the smallest value in the field-free case, and the optimization reorients the molecule so that these components align with the cavity polarization axes. For H\({}_{2}\)O these are the components orthogonal to the molecular plane corresponding to the green lines in a) and b) and the one parallel to the molecular dipole moment corresponding to the orange lines in a) and b). As shown in c), the observed reorientation also minimizes the DSE contribution to the coupled cavity-molecule system.
Optimized parameters of a H\({}_{2}\)O molecule coupled to two orthogonal cavity photon modes as a function of the coupling strength \(\lambda_{m}\). a) Magnitude of the dipole moment \(|\mu|\) b) the two angle \(\phi\) between the polarization axes of the cavity and the dipole moment of the molecule, c) averaged OH bond length and d) the bond angle. The dashed-dotted lines indicate the initial values in a)-d). The relevant part of the vibro-polaritonic IR spectra for different coupling strengths (color-coded) is shown in e) and f). The cavity frequency \(\omega_{m}\) is 1744 cm\({}^{-1}\) shown as a black dashed line in e) and f). The cavity coupling \(\lambda_{m}\) increases from 0.015 au to 0.123 au and the cavity polarization axes are \(\mathbf{e}_{1}=\) and \(\mathbf{e}_{2}=\).
Principal components of the polarizability tensor \(\alpha\) as a function of the coupling strength \(\lambda_{m}\) for optimized H\({}_{2}\)O structures coupled to a) a single cavity mode and to b) two orthogonal cavity modes. The bold line indicates the component aligned with the cavity polarization axes, and the black dashed dotted lines represent the field-free principal components of the polarizability. c) Change in DSE contribution as a function of the coupling strength \(\lambda_{m}\) compared to the initial H\({}_{2}\)O geometry for the single-mode case (green) and the two-mode case (orange). The cavity frequency \(\omega_{m}\) is 1744 cm\({}^{-1}\).
In general, reorientation can be rationalized as an effective way to reduce molecular polarizability along the cavity polarization axes and thus minimizing the DSE contribution.
All results presented for the optimization of H\({}_{2}\)O resonantly coupled to a single or two orthogonal cavity photon modes show that, without restriction of the rotational degrees of freedom, the cavity interaction leads to a reorientation of the molecule. This rotational motion occurs for all coupling strengths and is more pronounced than changes in the internal coordinates of H\({}_{2}\)O. For higher coupling strengths, the rotational effects even become visible in the vibro-polaritonic IR spectra.
### Optimization of H\({}_{2}\)O\({}_{2}\) embedded in a cavity
The optimization of a H\({}_{2}\)O\({}_{2}\) molecule coupled to a single photon mode and two to orthogonal photon modes serves as the second example. The cavity is resonant with the asymmetric bending mode of \(1491\,\mathrm{cm}^{-1}\), and the cavity polarization axis \(\mathbf{e}\) in the case of a single mode is aligned with the molecular dipole moment, which corresponds to the \(z\) axis of the laboratory frame.
Regardless of the coupling strength used, the dipole moment of H\({}_{2}\)O\({}_{2}\) remains aligned with the polarization axis of the single cavity mode, as shown in b). However, its size decreases significantly with increasing coupling up to the situation where the dipole moment is zero at \(\lambda_{m}=0.150\,\mathrm{au}\) (see a)). This large change in the dipole moment with the increasing coupling strength is induced by an increase in the dihedral angle, leading to a planarization of H\({}_{2}\)O\({}_{2}\) (see orange line d)). For \(\lambda_{m}=0.150\,\mathrm{au}\) the H\({}_{2}\)O\({}_{2}\) molecule becomes completely planar in a _trans_ configuration, which results in a zero dipole moment. Note that for this _trans_ structure, the cavity polarization axis is orthogonal with respect to the molecular plane, very similar to the H\({}_{2}\)O case. In contrast, the bond lengths and the averaged bond angle shown in c) and d), respectively, remain nearly constant for increasing couplings strengths. The observed change in the molecular geometry due to the cavity interaction can be clearly seen in the corresponding vibro-polaritonic IR spectra, shown in e) and f) for different coupling strengths. Details about the photonic characters of all relevant transitions can be found in the Supporting Information. For smaller values of \(\lambda_{m}\), as depicted in e), there is a clear splitting into a LP transition and a UP transition, which increases with increasing coupling strength. If \(\lambda_{m}\) is further increased, the changes in dipole moment and dihedral angle are more pronounced, which changes the vibro-polaritonic IR spectra, see f). These spectra are characterized by a single peak that is a pure molecular transition but is red-shifted relative to the field-free asymmetric bending mode. For a coupling strength below \(0.13\,\mathrm{au}\) a small shoulder a slightly lower frequency is present. This signal has a photonic character and is visible due to the weak coupling of the photon mode with the twisting mode in H\({}_{2}\)O\({}_{2}\) at \(424\,\mathrm{cm}^{-1}\). This mode is the most intense transition in the field-free spectrum and is characterized by a change in the dihedral angle of H\({}_{2}\)O\({}_{2}\). Similarly to the rotational motion observed in H\({}_{2}\)O, this twisting mode induces a dipole moment parallel to the polarization axis of the cavity. However, as the coupling increases further and H\({}_{2}\)O\({}_{2}\) becomes more planar, this coupling weakens, and only the single molecular peak remains visible.
To optimize the H\({}_{2}\)O\({}_{2}\) molecule coupled to a two mode cavity, we add a second orthogonal cavity mode with the same frequency and coupling strength. The two polarization axes are aligned with the \(z\) axis (\(\mathbf{e}_{1}\)) and the \(x\) axis (\(\mathbf{e}_{2}\)) of the laboratory frame.
Optimization of H\({}_{2}\)O\({}_{2}\) coupled to a two-mode cavity setup leads to similar but significantly weaker changes than in the case of coupling to a single cavity mode.
Optimized parameters of a H\({}_{2}\)O\({}_{2}\) molecule coupled to a single photon mode of an optical cavity as a function of the coupling strength \(\lambda_{m}\). a) Magnitude of the dipole moment \(|\mu|\) b) the angle \(\phi\) between the polarization axis of the cavity and the dipole moment of the molecule, c) averaged OH bond length (purple) and OO bond length (orange) and d) the averaged bond angle (purple) as well as the dihedral angle (orange). The dashed-dotted lines in a)-d) indicate the initial values. Relevant parts of the vibro-polaritonic IR spectra for different coupling strengths (color-coded) are shown in e) and f). The cavity frequency \(\omega_{m}\) is \(1491\,\mathrm{cm}^{-1}\) shown as a black dashed line in e) and f). The cavity coupling \(\lambda_{m}\) increases from \(0.017\,\mathrm{au}\) to \(0.150\,\mathrm{au}\) and the cavity polarization axis is \(\mathbf{e}=\).
The other internal coordinates of H\({}_{2}\)O\({}_{2}\), shown in c) and d), are slightly more affected in the two-mode case, but still in a rather insignificant way. Similarly to the single-mode coupling, we do not observe any reorientation of the H\({}_{2}\)O\({}_{2}\) dipole moment with respect to the polarization axes for the two-mode case due to the chosen orientation of the initial geometry. The dipole moment is parallel to the polarization axis \(\mathbf{e}_{1}\) and therefore orthogonal to the other, as visualized in b). Next, we discuss the vibro-polaritonic IR spectra of the optimized H\({}_{2}\)O\({}_{2}\) molecule coupled to a two-mode cavity. e) shows spectra for smaller coupling strengths, which can be roughly divided into two groups. When \(\lambda_{m}\) is less than 0.05 au, a weak Rabi splitting is observed. The LP transition and the UP transition are formed by the asymmetric bending mode and the cavity photon mode with the polarization axis \(\mathbf{e}_{1}\). This "standard" pair of LP and UP is present in all spectra and the Rabi splitting increases with increasing coupling strength, see e) and f). But similar to the case of H\({}_{2}\)O coupled with two cavity modes, a third weaker signal is present first as a shoulder (\(\lambda_{m}=0.05\) au e)) later as a distinct peak as shown in f). This weaker "middle" signal is mostly photonic and is characterized by the cavity photon mode with the polarization axis \(\mathbf{e}_{2}\). In line with the single-mode case, this transition is due to a weak coupling to the twisting mode in H\({}_{2}\)O\({}_{2}\) for higher coupling strengths. A visualization of the photonic characters of all relevant transitions can be found in the Supporting Information.
As discussed for the optimization of H\({}_{2}\)O, we also want to understand the observed changes in H\({}_{2}\)O\({}_{2}\) due to the cavity interaction by analyzing the polarizability \(\alpha\) and the DSE contribution. The principal components of the polarizability tensor \(\alpha\) and the change in the DSE contribution are shown in as a function of the coupling strength for both optimized H\({}_{2}\)O\({}_{2}\)-cavity systems.
Consistent with the H\({}_{2}\)O results shown in Fig. 3, in both the single-mode case and the two-mode case, all three principal components of the polarizability \(\alpha\) de
Principal components of the polarizability \(\alpha\) as a function of the coupling strength \(\lambda_{m}\) for optimized H\({}_{2}\)O\({}_{2}\) structures coupled to a single cavity mode a) and to two cavity modes b). The bold line indicates the component aligned with the cavity polarization axes, and the black dashed dotted lines represent the field-free principal components of the polarizability. c) Change in DSE contribution as a function of the coupling strength \(\lambda_{m}\) compared to the initial H\({}_{2}\)O\({}_{2}\) geometry for the single-mode case (green) and the two-mode case (orange). The cavity frequency \(\omega_{m}\) is resonant with the bending mode (1491 cm\({}^{-1}\)).
Optimized parameters of a H\({}_{2}\)O\({}_{2}\) molecule coupled to two orthogonal cavity photon modes as a function of the coupling strength \(\lambda_{m}\). a) Magnitude of the dipole moment \(|\mu|\) b) the angle \(\phi\) between the cavity polarization axes and the dipole moment of the molecule, c) averaged OH bond length (purple) and OO bond length (orange) and d) the averaged bond angle (purple) as well as the dihedral angle (orange). The dashed-dotted lines in a)-d) indicate the initial values. Relevant parts of the vibro-polaritonic IR spectra for different coupling strengths (color-coded) are shown in e) and f). The cavity frequency \(\omega_{m}\) is 1491 cm\({}^{-1}\) shown as a black dashed line in e) and f). The cavity coupling \(\lambda_{m}\) increases from 0.017 au to 0.150 au and the cavity polarization axes are \(\mathbf{e}_{1}=\) and \(\mathbf{e}_{2}=\).
The components aligned with the cavity polarization axes, highlighted in bold, are more strongly reduced by the cavity interaction, while especially for the single-mode case a) the other two components remain almost unchanged. We do not see reorientation of H\({}_{2}\)O\({}_{2}\) for the cavity polarization axes studied in the single-mode and two-mode cases, since for the chosen molecular ordination the axes are already aligned with the smallest polarizability components \(\alpha^{I}\) and \(\alpha^{II}\). Consequently, the observed reduction in polarizability is achieved only by geometrical changes and the direct response of the electronic structure to the cavity photon field. These geometrical changes and the electronic response also minimizes the DSE contribution. The change in \(E_{DSE}\) is plotted as a function of \(\lambda_{m}\) in c). Interestingly, the energy change is much larger for the single-mode case (green line) compared to the two-mode case (orange line), although the change in \(\alpha\) is comparable. To explain this, we have to consider that the total DSE contribution can be separated into a one-electron and two-electron part. uniquly The behavior of the one-electron contribution, and usually larger part, can be estimated by the polarizability, while the two-electron contribution is defined by the dipole moment. Since in the single-mode case H\({}_{2}\)O\({}_{2}\) planarizes with increasing coupling strength, its dipole moment vanishes and the two-electron part of the DSE is zero. Consequently, in the single-mode case, the change in DSE reflects both the decrease in polarizability and the decrease in the dipole moment. The present result for the optimization of H\({}_{2}\)O and H\({}_{2}\)O\({}_{2}\) coupled to an optical cavity clearly shows that not only the dipole moment of the molecule is important, but also the polarizability is a decisive factor in determining the influence of the molecule-cavity interaction.
## III Conclusion
To conclude, we studied the importance of geometry relaxation and relative orientation in the context of molecules coupled to the photon modes of optical cavities. Both effects are currently neglected in most computational studies in the field of polaritonic chemistry. As two illustrative examples, we have optimized H\({}_{2}\)O and H\({}_{2}\)O\({}_{2}\) resonantly coupled to one or two cavity photon modes. For the case of H\({}_{2}\)O, the predominant effect observed during the optimization processes is a rotation that leads to a reorientation of the molecule with respect to the photon modes. Surprisingly, we could even observe an effect of the restricted rotational degrees of freedom on the vibro-polaritonic IR spectrum for larger coupling strengths. These results clearly show that rotational motion is no longer an unrestricted degree of freedom in a cavity molecular system that can be neglected. In contrast, for the optimization of H\({}_{2}\)O\({}_{2}\), no rotational motion occurred for the chosen initial conditions (orientation of the molecule and polarization axes of the cavity), and geometric relaxation is more important. Consequently, the interpretation of computational studies based on fixed molecular structures and fixed orientation should be viewed with caution. Comparing the optimizations of H\({}_{2}\)O\({}_{2}\) coupled to one- and two-cavity modes, we want to highlight that the results have some similarities, but also show significant differences. The more realistic two-mode case with two cavity modes predicts smaller changes in the molecular structure, while the single-mode case overemphasizes them, even leading to a planar structure of H\({}_{2}\)O\({}_{2}\).
In agreement with recent studies, we can confirm the minimization of the DSE contribution as the driving force for both orientation and geometrical relaxation. The DSE has been shown to be strictly necessary to obtain a finite polarization and a bounded solution. However, despite its importance, DSE is still a rather abstract idea, so we derived a more accessible concept to estimate the effect of cavity interaction on molecular geometry. We are able to explain the observed rotation of the cavity-coupled molecule by determining the main components of the polarizability tensor \(\alpha\). Without fixing its orientation, the molecule will reorient so that the cavity mode polarization axes align with the smallest components of the polarizability. With respect to geometric relaxation, we observed two effects: First, it aims to reduce the polarizability of the molecule and, second, it also reduces the dipole moment itself. The evaluation of both the dipole moment and the polarizability is likely to be a useful and simple tool to evaluate the possibility of influencing molecules and chemical reactions by strong light-matter interaction inside an optical cavity. Exploring the changes in orientation and geometries in molecular ensembles are a logical next step.
## IV Methods
### Theoretical background
Starting from the non-relativistic Pauli-Fierz Hamiltonian in the length gauge representation we apply the cavity Born-Oppenheimer approximation (CBOA) to simulate molecules interacting with the confined electromagnetic field in an optical cavity under VSC conditions. By making use of a generalized Born-Huang expansion the cavity modes are grouped with the nuclei and the resulting electronic subsystem can be solved in an _ab-initio_ manner. Atomic units (\(\hbar=4\pi\varepsilon_{0}=m_{e}=1\)) are used in the following unless otherwise noted, and bold symbols denote vectors.
\[\hat{H}_{CBO}=\hat{H}_{el}+\sum_{m}^{N_{M}}\frac{1}{2}\omega_{m}^{2}q_{m}^{2}- \omega_{m}q_{m}\left(\mathbf{\lambda}_{m}\cdot\mathbf{\hat{\mu}}\right)+\frac{1}{2} \left(\mathbf{\lambda}_{m}\cdot\mathbf{\hat{\mu}}\right)^{2}\,, \tag{1}\]where \(\hat{\mathbf{\mu}}\) represents the molecular dipole operator, which is defined by the operators of the \(N_{el}\) electron coordinates \(\mathbf{\hat{r}}\) and the classic coordinates \(\mathbf{R}\) of the \(N_{Nuc}\) nuclei. \(\hat{H}_{el}\) is the Hamiltonian for the field-free many-electron system. Each of the \(N_{m}\) cavity modes contributes three additional terms to the electronic Hamiltonian. The first term is a harmonic potential introduced by the photon displacement field, with the classic photon displacement coordinate \(q_{m}\) and \(\omega_{m}\) being the frequency of the cavity mode. The second term describes the dipole coupling between the molecular system and the photon displacement field, which is characterized by the coupling strength \(\mathbf{\lambda}_{m}\). The last term is the DSE operator, which is an energy contribution that describes the self-polarization of the molecule-cavity system. The cavity mode-specific coupling parameter \(\mathbf{\lambda}_{m}\) for a cavity with effective mode volume \(V_{m}\) is defined as follows:
\[\mathbf{\lambda}_{m}=\mathbf{e}_{m}\lambda_{m}=\mathbf{e}_{m}\sqrt{\frac{4\pi}{V_{m}}}\,. \tag{2}\]
The unit vector \(\mathbf{e}_{m}\) denotes the polarization axis of the cavity mode \(m\).
The many-electron problem described by Eq. 1 can be solved using the CBO-HF approach. The resulting energy \(E_{CBO}\) is a function of the nuclear coordinates \(\mathbf{R}\) and the photon displacement coordinates \(\mathbf{q}\), represented as a vector grouping all \(q_{m}\):
\[E_{CBO}= \big{\langle}E_{CBO}\big{\rangle}(\mathbf{R},\mathbf{q})=\big{\langle}E_{ cl}\big{\rangle}(\mathbf{R}) \tag{3}\] \[+\sum_{m}^{N_{M}}\langle E_{lin}^{(m)}\rangle(\mathbf{R},q_{m})+ \big{\langle}E_{des}^{(m)}\rangle(\mathbf{R})+E_{dis}^{(m)}(q_{m})\]
The first derivative of the energy \(E_{CBO}\) with respect to the nuclear and photon displacement coordinates can be calculated analytically and defines the CBO-HF gradient \(\mathbf{g}_{CBO}\) as a \((3N_{A}+N_{M})\) vector, where \(N_{A}\) is the number of atoms in the molecule.
\[\mathbf{g}=\mathbf{\nabla}E_{CBO} \tag{4}\]
Based on the analytic gradients, the CBO-Hessian matrix \(\mathbf{H}\) of size \((3N_{A}+N_{M})\,(3N_{A}+N_{M})\) is accessible via finite differences.
\[\mathbf{H}_{ij}=\mathbf{\nabla}_{i}\mathbf{g}_{j}\approx\frac{\mathbf{g}_{j}\left(x_{i}+ \Delta\right)-\mathbf{g}_{j}\left(x_{i}-\Delta\right)}{2\Delta} \tag{5}\]
The molecular polarizability tensor \(\alpha\) at the CBO-HF level is calculated as the first derivative of the CBO-HF dipole moment vector with respect to a small external field, also using finite differences. Both the analytic gradient and the numerical Hessian can be used to optimize molecules coupled to cavity photon modes. In this manuscript we use the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm:
\[\mathbf{x}_{n+1}=\mathbf{x}_{n}-\tilde{\mathbf{H}}^{-1}\mathbf{g}_{CBO}\,, \tag{6}\]
The approximate Hessian matrix \(\tilde{\mathbf{H}}\) at each point \(n\) is updated with the Hessian matrix at stage \(n-1\) according to:
\[\tilde{\mathbf{H}}_{n}=\tilde{\mathbf{H}}_{n-1}+\frac{\Delta\mathbf{g}\cdot\Delta\mathbf{g}^{T }}{\Delta\mathbf{g}^{T}\cdot\Delta\mathbf{x}}-\frac{\tilde{\mathbf{H}}_{n-1}\cdot\Delta \mathbf{x}\cdot\Delta\mathbf{x}^{T}\tilde{\mathbf{H}}_{n-1}}{\Delta\mathbf{x}^{T}\cdot\tilde{ \mathbf{H}}_{n-1}\cdot\Delta\mathbf{x}} \tag{7}\]
A detailed benchmark of the implemented BFGS algorithm against the Steepest Descent method and the Newton-Raphson method can be found in the Supporting Information.
### Computational details
The BFGS algorithm, the necessary analytical gradients and the numerical Hessian for the CBO-HF ansatz have been implemented in the Psi4NumPy environment, which is an extension of the PSI4 electronic structure package. All calculations were performed using the aug-cc-pVDZ basis set and all geometries were pre-optimized at the Hartree-Fock level of theory. In all CBO-HF calculations performed in this work, we consider a lossless cavity. The coupling strength \(\lambda_{m}\) is chosen between \(0.015\,\mathrm{au}\) and \(0.150\,\mathrm{au}\) to assess the medium and strong coupling situation in the case of a single molecule. We emphasize that parts of the light-matter coupling used here are significantly larger than what can presently be achieved in experiments. For example, \(\lambda_{m}=$0.1\,\mathrm{au}$\) corresponds to an effective mode volume of less than \(0.2\,\mathrm{nm}^{3}\), which is less than the typical mode volumes of approximately \(10.0\,\mathrm{nm}^{3}\) that can be achieved in plasmonic cavities.
The CBO-HF energy and gradients depend on the relative orientation of the molecule and the polarization axes of the cavity. For this reason, the internal coordinate system traditionally used to optimize molecular geometries is a poor choice because it neglects spatial orientation. Therefore, all geometry optimizations in this work were performed using Cartesian coordinates. For the first step of the BFGS optimization, the exact Hessian matrix was computed, and to improve convergence for the H\({}_{2}\)O optimization, the exact Hessian was recomputed after every 10th step. The maximum component and the root mean square deviation of the gradient and the displacement vector are used as convergence criteria. For the gradient, both must be less than \(10^{-5}\,\mathrm{au}\) and for the displacement vector, both must be less than \(10^{-4}\,\mathrm{au}\). The semiclassical harmonic approximation was used to determine the normal modes and frequencies of the optimized coupled cavity-molecular systems. All optimized structures discussed in this manuscript are real minima, since they have no imaginary frequencies.
|
10.48550/arXiv.2405.17246
|
Do Molecular Geometries Change Under Vibrational Strong Coupling?
|
Thomas Schnappinger, Markus Kowalewski
| 1,009
|
10.48550_arXiv.1608.03167
|
###### Abstract
The molecular electric dipole, quadrupole and octupole moments of a selected set of 21 spin-compensated molecules are determined employing the extended version of the Piris natural orbital functional 6 (PNOF6), using the triple-\(\zeta\) Gaussian basis set with polarization functions developed by Sadlej, at the experimental geometries. The performance of the PNOF6 is established by carrying out a statistical analysis of the mean absolute errors with respect to the experiment. The calculated PNOF6 electric moments agree satisfactorily with the corresponding experimental data, and are in good agreement with the values obtained by accurate _ab initio_ methods, namely, the coupled-cluster single and doubles (CCSD) and multi-reference single and double excitation configuration interaction (MRSD-CI) methods.
## I Introduction
The interpretation and understanding of intermolecular forces, particularly those relating to long-range electrostatic interactions, require knowledge of the electrostatic moments. The electric moments are essential to provide simple ways to figure out the electric field behaviour of complex molecules. These electrical properties provide also information about the molecular symmetry since the electric moments depend on the geometry and charge distribution of the molecule.
It has long recognized the role of electrostatic interactions in a wide range of biological phenomena. The electrostatic energy is frequently the ruling contribution to molecular interactions in large biological systems, hence it is extremely important to describe properly the electrostatic potentials around these molecules. In order to improve the current treatment of the electrostatics for biomolecular simulations, which are traditionally modeled using a set of atom-centered point charges, the knowledge of higher multipole moments is required to include the effects of non-spherical charge distributions on intermolecular electrostatic interactions.
In principle, one can experimentally find the components of the electric field at each point, but it turns into a formidable task for large molecular systems. There are several techniques to determine experimentally the dipole moments, but it is still very difficult to obtain precise experimental values of higher multipole moments such as quadrupole or octupole moments, independently of the experimental conditions. Theoretical calculations are therefore essential but challenging for quantum chemistry methods. The accurate calculation of these properties is highly dependent on the method employed, either regarding approximate density functionals or methods based on wavefunctions. Consequently, calculating the multipole moments is a way to assess any electronic structure method.
The natural orbital functional (NOF) theory has emerged in recent years as an alternative method to conventional _ab initio_ approaches and density functional theory (DFT). A series of functionals has been proposed by Piris and collaborators (PNOFi, i=\(\overline{1,6}\)) using a reconstruction of the two-particle reduced-density matrix (2-RDM) in terms of the one-particle RDM (1-RDM) by ensuring necessary N-representability positivity conditions on the 2-RDM. In this work, we employ the PNOF6, which has proved a better treatment of both dynamic and non-dynamic electron correlations than its predecessors.
The aim of the present paper is to apply PNOF6, in its extended version, to the determination of molecular dipole and quadrupole moments of selected spin-compensated molecules, namely, H\({}_{2}\), HF, BH, HCl, H\({}_{2}\)O, H\({}_{2}\)CO, C\({}_{2}\)H\({}_{2}\), C\({}_{2}\)H\({}_{4}\), C\({}_{2}\)H\({}_{6}\), C\({}_{6}\)He, CH\({}_{3}\)CCH, CH\({}_{3}\)F, HCCF, ClF, CO, CO\({}_{2}\), N\({}_{2}\), NH\({}_{3}\), and PH\({}_{3}\). Moreover, the octupole moment of CH\({}_{4}\), a molecule without dipole and quadrupole moments is also studied. The Gaussian basis set of Sadlej, which has been specially developed to compute accurately molecular electric properties, is employed to perform all calculations.
We compare the obtained PNOF6 results with the experimental values reported in the literature, as well as with the theoretically computed values of Bundgen et al. who used the multi-reference single and double excitation configuration interaction (MRSD-CI) method, and the coupled-cluster single and doubles (CCSD) values calculated by us. Recall that the CCSD values for one-electron properties differ from full-CI results in only 2% if no multiconfigurational character is observed, so they can be considered as benchmark calculations. To our knowledge, this is the first NOF study of higher multipole moments such as quadrupole and octupole moments.
This article is organized as follows. We start in section II with the basic concepts and notations related to PNOF6 and electric multipole moments. The section III is dedicated to present our results and those obtained by using CCSD and MRSD-CI methods. Here, we discuss the outcomes obtained for the dipole, quadrupole and octupole principal moments, in separate sections. The performance of PNOF6 is established by carrying out astatistical analysis of the mean absolute errors (MAE) with respect to the experimental marks.
## II Theory
### The NOF Theory
We briefly describe here the theoretical framework of our approach. A more detailed description of PNOF6 can be found in reference. We focus on the extended version of PNOF, which provides a more flexible description of the electron pairs in the NOF framework.
Recall that PNOF6 is an orbital-pairing approach, which is reflected in the sum rule for the occupation numbers, namely,
\[\sum_{p\in\Omega_{g}}n_{p}=1\,,\quad g=\overline{1,F} \tag{1}\]
This involves coupling each orbital \(g\), below the Fermi level (\(F\)), with \(N_{c}\) orbitals above it (\(p>F\)), so the orbital subspace \(\Omega_{g}\equiv\left\{g,p_{1},p_{2},\cdots,p_{N_{c}}\right\}.\) Taking into account the spin, each subspace contains an electron pair. Henceforth, we will denote PNOF6(\(N_{c}\)) the method we use, emphasizing the number \(N_{c}\) of usually weakly-occupied orbitals employed in the description of each electron pair.
The PNOF6(\(N_{c}\)) energy for a singlet state of an \(N\)-electron molecule can be cast as
\[E=\sum_{g=1}^{F}E_{g}+\sum_{f\neq g}^{F}\sum_{p\in\Omega_{f}}\sum_{q\in\Omega_ {g}}E_{pq}^{int} \tag{2}\]
The first term of the energy draws the system as independent \(F=N/2\) electron pairs described by the following NOF for two-electron systems,
\[E_{g}=\sum_{p\in\Omega_{g}}n_{p}\left(2\mathcal{H}_{pp}+\mathcal{J}_{pp}\right) +\sum_{p,q\in\Omega_{g},p\neq q}E_{pq}^{int} \tag{3}\]
It is worth noting that the interaction energy, the last term of equations and, is equal for electrons belonging to the same subspace \(\Omega_{g}\) or two different subspaces (\(\Omega_{g}\neq\Omega_{f}\)), therefore, the intrapair and interpair electron correlations are equally balanced in PNOF6(\(N_{c}\)).
The interaction energy \(E_{pq}^{int}\) is given by
\[E_{pq}^{int}=\left(n_{q}n_{p}-\Delta_{qp}\right)\left(2\mathcal{J}_{pq}- \mathcal{K}_{pq}\right)+\Pi_{qp}\mathcal{L}_{pq} \tag{4}\]
\(\mathcal{L}_{pq}=\langle pp|qq\rangle\) is the exchange and time-inversion integral, which reduces to \(\mathcal{K}_{pq}\) for real orbitals. \(\Delta\) and \(\Pi\) are the auxiliary matrices proposed in reference in order to reconstruct the 2-RDM in terms of the occupation numbers. The conservation of the total spin allows to determine the diagonal elements as \(\Delta_{pp}=n_{p}^{2}\) and \(\Pi_{pp}=n_{p}\), whereas known analytical necessary \(N\)-representability conditions provide bounds for the off-diagonal terms.
\[\begin{array}{c|c|c}\Delta_{qp}&\Pi_{qp}&Orbitals\\ \hline e^{-2S}h_{q}h_{p}&-e^{-S}\left(h_{q}h_{p}\right)^{\nicefrac{{1}}{{2}}}& q\leq F,p\leq F\\ \frac{\gamma_{q}\gamma_{p}}{S_{\gamma}}&-\Pi_{qp}^{\nicefrac{{1}}{{2}}}&q>F,p \leq F\\ e^{-2S}n_{q}n_{p}&e^{-S}\left(n_{q}n_{p}\right)^{\nicefrac{{1}}{{2}}}&q>F,p>F \end{array} \tag{5}\]
The other magnitudes are defined as
\[\begin{array}{c}\gamma_{p}=n_{p}h_{p}+\alpha_{p}^{2}-\alpha_{p}S_{\alpha}\\ \\ \alpha_{p}=\begin{cases}e^{-S}h_{p}\,,&p\leq F\\ e^{-S}n_{p}\,,&p>F\end{cases}\end{array} \tag{6}\]
\[\begin{array}{c}\Pi_{qp}^{\gamma}=\left(n_{q}h_{p}+\frac{\gamma_{q}\gamma_{p }}{S_{\gamma}}\right)^{\nicefrac{{1}}{{2}}}\left(h_{q}n_{p}+\frac{\gamma_{q} \gamma_{p}}{S_{\gamma}}\right)^{\nicefrac{{1}}{{2}}}\\ S=\sum_{q=F+1}^{F+FN_{c}}\!\!n_{q},\quad S_{\alpha}=\sum_{q=F+1}^{F+FN_{c}}\!\! \alpha_{q},\quad S_{\gamma}=\sum_{q=F+1}^{F+FN_{c}}\!\!\gamma_{q}\end{array} \tag{7}\]
It is noteworthy that the reconstruction of the 2-RDM, and therefore the functional, are independent of the orbital-pairing sum rules. These additional constraints are imposed to ensure that no fractional electron numbers appear when non-dynamic electron correlation effects become important. Additionally, this allows the constraint-free minimization of the PNOF6(\(N_{c}\)) energy with respect to the occupation numbers, which yields substantial savings of computational time.
At present, the procedure for the minimization of the energy with respect to both the occupation numbers and the natural orbitals is carried out by the iterative diagonalization method developed by Piris and Ugalde implemented in the DoNOF program package. The matrix element of the kinetic energy and nuclear attraction terms, as well as the electron repulsion integrals are inputs to our computational code. In the current implementation, we have used the GAMESS program for this task.
### Dipole, Quadrupole and Octupole Moments
The potential of the electric field at any point outside a distribution of charges is simply related to the electric multipole moments. As any distribution function, the essential features of the charge distribution can be characterized by its moments, thereby for an uncharged molecule the first (dipole), second (quadrupole) and third (octupole) electric moments are the most important terms in the multipole expansion, therefore, are usually sufficient to characterize its interaction with an external field.
\[\mu_{\alpha}=-\int\rho(\mathbf{r})r_{\alpha}dV+\sum_{i=1}^{NUC}Z_{i}R_{i\alpha} \tag{7}\]
\[\begin{split}\Theta_{\alpha\beta}&=-\frac{1}{2} \int\rho(\mathbf{r})(3r_{\alpha}r_{\beta}-\delta_{\alpha\beta}r^{2})dV\\ +\frac{1}{2}\sum_{i=1}^{NUC}Z_{i}(3R_{i\alpha}R_{i\beta}-\delta_ {\alpha\beta}R_{i}^{2})\end{split} \tag{8}\]
\[\begin{split}\Omega_{\alpha\beta\gamma}&=-\frac{5}{2 }\int\rho(\mathbf{r})\,r_{\alpha}r_{\beta}r_{\gamma}dV\\ +\frac{1}{2}\,\int\rho(\mathbf{r})r^{2}\left(r_{\alpha}\delta_{ \beta\gamma}+r_{\beta}\delta_{\alpha\gamma}+r_{\gamma}\delta_{\alpha\beta} \right)dV\\ +\frac{5}{2}\,\sum_{i=1}^{NUC}Z_{i}\,R_{i\alpha}R_{i\beta}R_{i\gamma }\\ -\frac{1}{2}\sum_{i=1}^{NUC}Z_{i}R_{i}^{2}\left(R_{i\alpha}\delta_ {\beta\gamma}+R_{i\beta}\delta_{\alpha\gamma}+R_{i\gamma}\delta_{\alpha\beta}\right) \end{split} \tag{9}\]
Note that the nuclear contribution is taken into account separately from the electronic contribution, which arises from the negative charge distribution over all the space. The formulas and define symmetric tensors in all subscripts. Moreover, equation defines a traceless tensor for the quadrupole moment, namely, \(\Theta_{xx}+\Theta_{yy}+\Theta_{zz}=0\), similarly, equation leads to \(\Omega_{xxx}+\Omega_{yyz}+\Omega_{zzz}=0\) for octupole tensor, and respective permutations between the subscripts \(x,y\) and \(z\).
## III Results and Discussion
In the following sections, we show the PNOF6(\(N_{c}\)) results obtained for the dipole, quadrupole, and octupole moments with respect to the center of mass.
The chosen basis set is known to be an important factor in the calculation of molecular electric properties. We used the Gaussian basis set of Sadlej, which is a correlation-consistent valence triple-\(\zeta\) basis set augmented with additional basis functions selected specifically for the correlated calculation of electric properties. Thus, it contains sufficient diffuse and polarization functions in order to give an accurate description of the outer-valence region. It has been shown that the Sadlej basis set has effectively the same accuracy as the aug-cc-pVTZ basis set.
Since the number \(N_{c}\) of usually weakly-occupied orbitals is related to the description of the electron pairs, we begin studying the H\({}_{2}\) molecule, where there are not interpair correlation effects. This molecule has zero dipole moment, hence the calculated quadrupole moment values for different \(N_{c}\) values are shown in table 1.
As expected, the best description of the electron pair is obtained when the number of usually weakly-occupied orbitals is maximum, in fact, the calculated quadrupole moment converges to the CCSD value, which is the full CI result for this molecule. In the present work, we will carry out all PNOF6(\(N_{c}\)) calculations by using the maximum \(N_{c}\) value allowed by the Sadlej basis set for each molecule. In our calculations, we have set to one the occupancies of the core orbitals. Consequently, the maximum possible value of \(N_{c}\) is given by the number of basis functions above the Fermi level, divided by the number of the considered strongly occupied orbitals.
For comparison, we have included the available experimental data, and the calculated Hartree-Fock (HF) and CCSD values using the GAMESS program. The experimental equilibrium geometries have been used to carry out all calculations. The performance of theoretically obtained results is established by carrying out a statistical analysis of the mean absolute errors (MAE) with respect to the experimental data. Atomic units (a.u.) are used throughout.
### Dipole Moment
In this work, the dipoles are aligned along the principal symmetry axis of the studied molecules, set on \(z\) direction. Table 1 shows the independent component \(\mu_{z}\) of the dipole moments obtained at the HF, PNOF6(\(N_{c}\)), and CCSD levels of theory.
Overall, the inclusion of electron correlation effects through, both PNOF6(\(N_{c}\)) and CCSD, improves significantly the performance of the HF method. PNOF6(\(N_{c}\)) and CCSD afford MAEs with respect to experimental data of 0.0309 a.u. and 0.0177 a.u., respectively. It is worth noting the agreement between PNOF6(\(N_{c}\)) and CCSD results, as well as with the experimental data. Note that the aug-cc-pVTZ basis set of Dunning was used for the BH molecule since there is no Sadlej-pVTZ basis set available for Boron. In this case, the PNOF6 result is very close to the Full-CI/aug-cc-pVTZ value obtained by Halkier et al., 0.5433 a.u., showing a result as good as the CCSD one.
The electronic structure and bonding situation of carbon monoxide is of special interest for modern electronic structure methods. The dipole moment of CO, extensively studied in Refs., is very small (0.0481 a.u.) and ends at the carbon atom, although carbon is less electronegative than oxygen. The result shown in Table 3 is representative, while HF gives the wrong direction for the CO dipole moment, PNOF6 corrects the sign, giving a result that is in excellent agreement with the experimental value. Remarkably, the result obtained at CCSD level is 34% away from the experimental value, so that it is necessary to include third order triplet excitations in the cluster theory in order to obtain a reasonable value, such as the one reported by Maroulis at the CCSD(T) level, 0.0492 a.u.. Accordingly, the relevant electron correlation for CO is well accounted by the PNOF6 method.
Regarding the values obtained for HF, H\({}_{2}\)O, H\({}_{2}\)CO, HCl, NH\({}_{3}\), and ClF, PNOF6(\(N_{c}\)) competes with Coupled Cluster, providing values that differ from experimental data in less than a 7%. In the case of HCCF and PH\({}_{3}\), PNOF6(\(N_{c}\)) seems to lack relevant dynamic electron correlation and thereby the obtained dipole moments are not as accurate as the CCSD ones. Conversely, our values are in excellent agreement with experimental data in the case of the methyls CH\({}_{3}\)F and CH\({}_{3}\)CCH, often attached to large organic molecules, giving dipole moments with errors of 0.4% and 2% respectively, with respect to experimental values.
A special case is ozone, which is a molecule with strong multiconfigurational character. The PNOF6 dipole moves into the right direction from the HF value, but overestimates the effects of the electron correlation. Taking into account the good CCSD result for O\({}_{3}\), which is not valid for higher electric moments, it seems that the dynamic electron correlation compensates for the lack of non-dynamical in this method, and could improve our numerical value of the dipole.
### Quadrupole Moment
Tables 2 and 3 list the molecular quadrupole moments obtained at the HF, CCSD, MRSD-CI and
\begin{table}
\begin{tabular}{l c c c c c} \hline \hline Molecule & HF & PNOF6 (\(N_{c}\)) & CCSD & EXP. \\ \hline HF & 0.7565 & 0.7223 & 7 & 0.6994 & 0.7089 \\ BH\({}^{*}\) & 0.6854 & 0.5395 & 38 & 0.5551 & 0.4997 \\ H\({}_{2}\)O & 0.7808 & 0.7458 & 9 & 0.7225 & 0.7268 \\ H\({}_{2}\)CO & 1.1134 & 0.9872 & 10 & 0.9084 & 0.9175 \\ HCl & 0.4746 & 0.4598 & 8 & 0.4416 & 0.4301 \\ HCCF & 0.3535 & 0.3189 & 9 & 0.2733 & 0.2872 \\ NH\({}_{3}\) & 0.6372 & 0.6153 & 12 & 0.5943 & 0.5789 \\ PH\({}_{3}\) & 0.2780 & 0.2755 & 13 & 0.2340 & 0.2258 \\ O\({}_{3}\) & 0.3033 & 0.1370 & 7 & 0.2276 & 0.2099 \\ ClF & 0.4453 & 0.3226 & 6 & 0.3451 & 0.3462 \\ CH\({}_{3}\)F & 0.7706 & 0.7283 & 10 & 0.6919 & 0.7312 \\ CH\({}_{3}\)CCH & 0.3203 & 0.3141 & 12 & 0.2866 & 0.3070 \\ CO & \(-\)0.0987 & 0.0414 & 9 & 0.0725 & 0.0481 \\ \hline MAE & 0.0843 & 0.0309 & & 0.0177 & \\ \hline \hline \end{tabular} \({}^{*}\)Calculations performed with the aug-cc-pVTZ basis set.
\end{table}
Table 2: \(\mu_{z}\) component of molecular dipole moments in atomic units (\(ea_{0}\)) computed with the Sadlej-pVTZ basis set at the experimental equilibrium geometries. \(N_{c}\) is the number of weakly-occupied orbitals employed in PNOF6(\(N_{c}\)) for each molecule.
\begin{table}
\begin{tabular}{l c c c c c} \hline \hline PNOF6 & PNOF6 & PNOF6 & PNOF6 & CCSD & EXP. \\ \hline
0.3697 & 0.4030 & 0.3965 & 0.3935 & 0.3935 & 0.39\(\pm\) 0.01 \\ \hline \hline \end{tabular}
\end{table}
Table 1: \(\Theta_{xz}\) component of H\({}_{2}\) quadrupole moment, in atomic units, obtained by employing PNOF6(\(N_{c}\)) and CCSD with the Sadlej-pVTZ basis set at the experimental equilibrium geometry, together with the experimental value.
PNOF6(\(N_{c}\)) levels of theory, along with the experimental values taken from Refs.. Inspection of these Tables shows that PNOF6(\(N_{c}\)) quadrupole moments agree satisfactorily with the experimental data, whereas the discrepancies are consistent with those observed using the CCSD and MRSD-CI methods in most cases.
In the case of linear molecules (H\({}_{2}\), HF, BH, HCl, HCCF, ClF, CO, C\({}_{2}\)H\({}_{2}\), CO\({}_{2}\) and N\({}_{2}\)), NH\({}_{3}\) and PH\({}_{3}\), belonging to the \(C_{3v}\) point symmetry group, the \(D_{6h}\) C\({}_{6}\)H\({}_{6}\) molecule, and the trigonal planar C\({}_{2}\)H\({}_{6}\), which has \(D_{3d}\) symmetry, the relation \(\Theta_{xx}=\Theta_{yy}=-\frac{1}{2}\Theta_{zz}\) holds for quadrupole moment tensor, so \(\Theta_{zz}\) alone is sufficient to determine completely the quadrupole moment. Setting the main axis of symmetry in the z direction of the coordinate system, the results for these molecules are reported in Table 2. From the latter, one can observe that PNOF6(\(N_{c}\)) yields a MAE of 0.15 a.u., hence considering the added complexity of the quadrupole moment, the performance of PNOF6(\(N_{c}\)) is within a reasonable accuracy.
Taking into account the experimental uncertainty, PNOF6(\(N_{c}\)) results agree with the experimental data for H\({}_{2}\), HCl, CO, N\({}_{2}\), PH\({}_{3}\), ClF, CH\({}_{3}\)F, C\({}_{2}\)H\({}_{6}\), and C\({}_{6}\)H\({}_{6}\). The value obtained for H\({}_{2}\) reproduces the experimental one with high precision. It is also worth noting the excellent agreement with the experiment obtained for the quadrupole moment of Benzene, which is of great interest for many fields of chemistry and biology. Indeed, the quadrupole moment of Benzene plays an important role in determining the crystal structures and molecular recognition in biological systems because it is the key to the intermolecular interactions between \(\pi\)-systems.
For HCCF, NH\({}_{3}\), C\({}_{2}\)H\({}_{2}\), and CO\({}_{2}\), the quadrupole moments fall out of the experimental error intervals, however, in the case of HCCF, C\({}_{2}\)H\({}_{2}\), and CO\({}_{2}\) the mean relative percentage error is below 11%, whereas the results obtained for NH\({}_{3}\) is only 0.05 a.u. away from the higher limit of the experimental uncertainty. For CH\({}_{3}\)CCH the PNOF6 result deviates from the experimental value in a 13%, ergo more dynamic correlation is clearly necessary to improve this result, an effect not observed for the dipole moment of this molecule.
For the Hydrogen fluoride, the HF result is the closest to the experimental value, however, the PNOF6 result is in outstanding agreement with the full-CI/aug-cc-pVTZ value of 1.6958 a.u.. For the Boron monohydride, the experimental quadrupole moment is not available, so we use the full-CI/aug-cc-pVTZ calculation reported by Halkier et al., 2.3293 a.u., in order to carry out the comparison. The agreement between PNOF6 and full-CI is good, according to the relative percentage error obtained below 1.7%.
Table 3 shows the \(\Theta_{zz}\) and \(\Theta_{xx}\) components obtained for H\({}_{2}\)O, H\({}_{2}\)CO, C\({}_{2}\)H\({}_{4}\), and O\({}_{3}\). In this work, we use the traceless quadrupole moment, hence two components are sufficient to determine completely this magnitude. On the other hand, MRSD-CI values are significantly better than CCSD calculations when many components of the quadrupole tensor are studied, thereby MRSD-CI is used as benchmark theoretical method in Table 3.
According to the results reported in Table 3, PNOF6(\(N_{c}\)) performs better than the MRSD-CI method for this selected set of molecules. For H\({}_{2}\)O and H\({}_{2}\)CO, the PNOF6(\(N_{c}\)) values fall into the experimental error interval, which is specially broad for H\({}_{2}\)CO. In the case of the C\({}_{2}\)H\({}_{4}\) molecule, the longitudinal component \(\Theta_{zz}\) obtained with PNOF6 is near the limit of the experimental error interval, as well as the \(\Theta_{xx}\) component. Finally, we have the results obtained for O\({}_{3}\), which is a stringent test for quadrupole calculations due to its two-configurational character. One can observe that Ozone is well described by PNOF6 comparing to the results obtained by using HF and MRSD-CI methods.
### Octupole Moment
The octupole moment is particularly interesting in the case of methane. The octupole moment is the first non-zero term in the multipole expansion of the electrostatic interaction for methane molecule, so it is crucial in order to describe properly its interactions with external fields. Actually, the octupole-octupole interaction is the main long-range orientation dependent interaction in methane. Moreover, the complex charge distribution of methane, which has long been studied in the literature, is mainly dependent on its octupole moment, thus, the octupole moment is essential to characterize the charge distribution of tetrahedral molecules.
For tetrahedral molecules the octupole moment is simply given by one component, namely \(\Omega=\Omega_{xyz}\). Employing PNOF6 with the Sadlej-pVTZ basis set at the experimental equilibrium geometry, the result obtained for CH\({}_{4}\) is \(\Omega_{xyz}=2.1142\,a.u.\), whereas the experimental mark reported in Ref. is \(\Omega_{xyz}=2.95\pm 0.17\,a.u.\) Although the PNOF6 result falls out of the experimental interval error, this value is reasonable taking into account the discrepancies between experimental marks obtained by different experimental techniques. Besides, comparing to theoretical calculations, the PNOF6 value is very close to the result obtained by using CCSD, \(\Omega_{xyz}=2.0595\,a.u.\). Consequently, we can conclude that PNOF6 describes properly the octupole moment of methane.
## IV Conclusions
The PNOF6 method, in its extended version, has been assessed by comparing the molecular electric moments with the experimental data as well as with CCSD and MRSD-CI theoretical values. The dipole, quadrupole and octupole moments for a selected set of well-characterized 21 molecules have been calculated at the experimental equilibrium geometries using the triple-\(\zeta\) Gaussian basis set with polarization functions developed by Sadlej. Our results show that PNOF6(\(N_{c}\)) is able to predict electric properties as accurate as high-level electronic structure methods such as CCSD or MRSD-CI, therefore the functional computes quite accurately the charge distribution of molecular systems. To our knowledge, this is the first NOF study of higher multipole moments such as quadrupole and octupole moments.
For PNOF6(\(N_{c}\)) dipole moments, the obtained MAE with respect to experimental data is \(0.0309\:a.u.\), being consistent with the theoretical benchmark calculations. Remarkable is the result obtained by PNOF6 for Carbon monoxide, for which, HF gives a wrong direction of the dipole and CCSD overestimates it severely, whereas PNOF6 corrects the sign, giving a result that is in excellent agreement with the experimental mark.
The high performance of PNOF6(\(N_{c}\)) in computing electric quadrupole moments has been shown by most of the studied molecules, for which the computed values fall into the experimental interval error. It has been shown that the method is capable of providing the different components of the quadrupole moment tensor. The PNOF6(\(N_{c}\)) MAE with respect to the experiment is \(0.1291\:a.u.\), which is very close to the corresponding MAEs of \(0.0902\:a.u.\) and \(0.1448\:a.u.\) obtained by using the well-established CCSD and MRSD-CI methods, respectively.
\begin{table}
\begin{tabular}{l c c c c c} \hline Molecule & HF & PNOF6 (\(N_{c}\)) & MRSD-CI & EXP. \\ \hline H\({}_{2}\)O (\(xx\)) & 1.7966 & 1.7808 & 9 & 1.8050 & \(1.86\pm 0.02\) \\ H\({}_{2}\)O (\(zz\)) & 0.0981 & 0.0869 & 9 & 0.0950 & \(0.10\pm 0.02\) \\ H\({}_{2}\)CO (\(xx\)) & 0.1019 & 0.0516 & 10 & 0.1100 & \(0.04\pm 0.12\) \\ H\({}_{2}\)CO (\(zz\)) & 0.0921 & 0.1255 & 10 & 0.2230 & \(0.20\pm 0.15\) \\ C\({}_{2}\)H\({}_{4}\) (\(xx\)) & 2.7819 & 2.5892 & 13 & 2.3700 & \(2.45\pm 0.12\) \\ C\({}_{2}\)H\({}_{4}\) (\(zz\)) & 1.4942 & 1.3266 & 13 & 1.1700 & \(1.49\pm 0.11\) \\ O\({}_{3}\) (\(xx\)) & 1.1175 & 1.2426 & 7 & 1.2830 & \(1.03\pm 0.12\) \\ O\({}_{3}\) (\(zz\)) & \(-0.2387\) & 0.3606 & 7 & 0.1680 & \(0.52\pm 0.08\) \\ \hline MAE & 0.1772 & 0.1066 & & 0.1448 & & \\ \end{tabular}
\end{table}
Table 4: \(\Theta_{zz}\) and \(\Theta_{xx}\) components of molecular quadrupole moments, in atomic units, computed using the Sadlej-pVTZ basis set at the experimental equilibrium geometries. \(N_{c}\) is the number of weakly-occupied orbitals employed in PNOF6(\(N_{c}\)) for each molecule.
\begin{table}
\begin{tabular}{l c c c c c} \hline Molecule & HF & PNOF6 (\(N_{c}\)) & CCSD & EXP. \\ \hline H\({}_{2}\) & 0.4381 & 0.3935 & 17 & 0.3935 & \(0.39\pm 0.01\) \\ HF & 1.7422 & 1.6939 & 7 & 1.7156 & \(1.75\pm 0.02\) \\ BH\({}^{*}\) & 2.6772 & 2.3706 & 38 & 2.3388 & \(2.3293^{\dagger}\) \\ HCl & 2.8572 & 2.7753 & 8 & 2.7233 & \(2.78\pm 0.09\) \\ HCCF & 3.3530 & 3.2482 & 9 & 2.9335 & \(2.94\pm 0.10\) \\ CO & 1.5366 & 1.4562 & 9 & 1.4889 & \(1.44\pm 0.30\) \\ N\({}_{2}\) & 0.9397 & 1.0530 & 9 & 1.1712 & \(1.09\pm 0.07\) \\ NH\({}_{3}\) & 2.1258 & 2.1080 & 12 & 2.1661 & \(2.45\pm 0.30\) \\ PH\({}_{3}\) & 1.7217 & 1.6507 & 13 & 1.5695 & \(1.56\pm 0.70\) \\ ClF & 0.9413 & 1.1122 & 6 & 1.0514 & \(1.14\pm 0.05\) \\ CH\({}_{3}\)F & 0.3482 & 0.3269 & 10 & 0.3002 & \(0.30\pm 0.02\) \\ C\({}_{2}\)H\({}_{2}\) & 5.3655 & 5.1531 & 12 & 4.6850 & \(4.71\pm 0.14\) \\ C\({}_{2}\)H\({}_{6}\) & 0.6329 & 0.6275 & 13 & 0.6234 & \(0.59\pm 0.07\) \\ C\({}_{6}\)H\({}_{6}\) & 6.6418 & 6.3571 & 12 & 5.6653 & \(6.30\pm 0.27\) \\ CH\({}_{3}\)CCH & 4.2913 & 4.1146 & 12 & 3.6939 & \(3.58\pm 0.01\) \\ CO\({}_{2}\) & 3.8087 & 3.6012 & 8 & 3.1966 & \(3.19\pm 0.13\) \\ \hline MAE & 0.2646 & 0.1517 & 0.0902 & & \\ \end{tabular}
\end{table}
Table 3: \(\Theta_{zz}\) component of the quadrupole moments, in atomic units, computed with the Sadlej-pVTZ basis set at the experimental equilibrium geometries for molecules with linear, \(C_{3v}\), \(D_{6h}\), and \(D_{3d}\) symmetry. \(N_{c}\) is the number of weakly-occupied orbitals employed in PNOF6(\(N_{c}\)) for each molecule.
Finally, the study of the octupole moment was focused here on methane, due to its important role in the description of the long-range electrostatic interactions for this molecule. The PNOF6 result is in excellent agreement with the value provided by the CCSD method.
|
10.48550/arXiv.1608.03167
|
Molecular electric moments calculated by using natural orbital functional theory
|
Ion Mitxelena, Mario Piris
| 3,878
|
10.48550_arXiv.1703.07712
|
## B. P.
## Gatchina, St.
###### Abstract
We show that the theoretical predictions on high energy behavior of the photoionization cross section of fullerenes depend crucially on the form of the function \(V(r)\) which approximates the fullerene field. The shape of the high energy cross section is obtained without solving the wave equation. The cross section energy dependence is determined by the analytical properties of the function \(V(r)\).
## 1 Introduction
In this paper we calculate the high energy nonrelativistic asymptotics for the photoionization cross section of the valence electrons of fullerenes \(C_{N}\). We consider the fullerenes which can be treated approximately as having the spherical shape. The photon carries the energy \(\omega\) which is much larger than the ionization potential \(I\). We find the leading term of the cross section expansion in terms of \(1/\omega\). We keep the photon energy to be much smaller than the electron rest energy \(mc^{2}\). Here we consider only the ionization of \(s\) states. We employ the relativistic system of units in which \(\hbar=1\); \(c=1\) and the squared electron charge \(e^{2}=\alpha=1/137\).
The actual potential experienced by the fullerene valence electrons is a multicentered screened Coulomb potential produced by the \(C^{+4}\) carbon ions of the fullerene. The ionized electron approaches one of these centers transferring large momentum. The asymptotic of the photoionization cross section \(\sigma(\omega)\) in the screened Coulomb field is the same as in the unscreened one, i.e. \(\sigma\sim\omega^{-7/2}\). Thus we expect the observed asymptotic also to be \(\sigma\sim\omega^{-7/2}\).
However one usually uses a model central potential \(V(r)\) for description of the field created by the fullerene. For a model potential the asymptotics may be a different one. We consider spherical fullerene with the radius \(R\) and the width of the layer \(\Delta\ll R\). The general properties of the potential \(V(r)\) are well known-see, e.g.. It is located mostly inside the fullerene layer \(R-\Delta/2\leq r\leq R+\Delta/2\) being negligibly small outside.
In the simplest (or even the oversimplified) version it is just the well potential which is constant inside the layer and vanishes outside.
\[V(r)\ =\ -V_{0}\theta(r-R_{1})\Big{(}1-\theta(r-R_{2})\Big{)};\quad V_{0}\ >\ 0. \tag{1}\]Recall that \(\theta(x)=1\) for \(x\geq 0\) while \(\theta(x)=0\) for \(x<0\).
\[V(r)\ =\ -U_{0}\delta(r-R), \tag{2}\]
Here, as well as in Eqs. and \(U_{0}>0\) are the dimensionless constants.
The nowadays calculations are often based on the jellium model (see, e.g.,). In this approach the charge of the positive core consisting of nuclei and the internal electrons is assumed to be distributed uniformly in the fullerene layer.
\[V_{1}(r)\ =\ const=-U_{0}\frac{3}{2}\frac{R_{2}^{2}-R_{1}^{2}}{R_{2}^{3}-R_{1}^{ 3}};\quad V_{2}(r)=-\frac{U_{0}}{2(R_{2}^{3}-R_{1}^{3})}\Big{(}3R_{2}^{2}-r^{2 }(1+\frac{2R_{1}^{3}}{r^{3}})\Big{)}; \tag{3}\]
\[V_{3}(r)=-\frac{U_{0}}{r}.\]
Sometimes model potentials are determined by analytical functions of \(r\) with a sharp peak at \(r=R\).
\[V(r)=-\frac{U_{0}}{\pi}\frac{a}{(r-R)^{2}+a^{2}}. \tag{4}\]
It describes the Dirac bubble potential at \(a\to 0\).
\[V(r)\ =\ -\frac{V_{0}}{\pi}\exp\frac{-(r-R)^{2}}{a^{2}}, \tag{5}\]
Strictly speaking our analysis is true for the negative ion \(C_{N}^{-}\). However, since there are many valence electrons in the fullerene shell, we expect it to be true for photoionization of the neutral fullerene \(C_{N}\) as well.
As it stands now, the asymptotics for the photoionization cross section is known only for the Dirac bubble potential. Here we demonstrate that the energy behavior of the asymptotic cross section is strongly model dependent. It is determined by the analytical properties of the potential \(V(r)\). In Sec. 2 we obtain the general equation for the asymptotics of the photoionization cross section. In Sec.3 we calculate the asymptotics for the potentials mentioned in Introduction. We analyze the results in Sec.4.
## 2 Asymptotics of the cross section
The photoionization cross section can be presented as (see Eq.(56.3) of or Eq.(5.76) of)
\[d\sigma=n_{e}\frac{mp}{(2\pi)^{2}}|F|^{2}d\Omega. \tag{6}\]
Here \(m\) is the electron mass, \(p=|{\bf p}|\), while \({\bf p}\) is the photoelectron momentum, \(\Omega\) is the solid angle of the photoelectron, and \(n_{e}\) is the number of electrons in the ionized state. The normalization factor of the photon wave function \(n(\omega)=\sqrt{4\pi}/\sqrt{2\omega}\) is included in the photoionization amplitude \(F\). Averaging over polarizations of the incoming photon is assumed to be carried out.
We consider the photon energy \(\omega\) which is much larger than the ionization potential \(I\), i.e. \(\omega\gg I\). Limiting ourselves by the condition \(\omega\ll m\) we can treat the photoelectron in nonrelativistic approximation. The kinetic energy of the photoelectron is \(\varepsilon=\omega-I=p^{2}/2m\). The electron momentum \(p\) is much larger than the characteristic momentum \(\mu=(2mI)^{1/2}\) of the bound state (\(p\gg\mu\)). At \(\omega\gg I\) the photoionization requires large momentum \({\bf q}={\bf k}-{\bf p}\) to be transferred to the recoil fullerene. Here \({\bf k}\) is the photon momentum, and \(k=|{\bf k}|=\omega\). One can see that \(k\ll p\) if \(I\ll\omega\ll m\) and thus we can put \(|{\bf q}|=q=p\).
If the electron-photon interaction is written in the velocity form, momentum \(q\) is transferred in the initial state in ionization of \(s\) states,. Thus the photoelectron can be described by plane wave. Interaction of the photoelectron with the ionized fullerene provides the contributions of the relative order \(O(1/p)\) to the amplitude. Hence they contribute to the cross section beyond the asymptotics. The photoionization amplitude can be written as \(F=n(\omega)\int d^{3}r\psi_{\bf p}^{*}({\bf r})\gamma\psi(r)\) with \(\psi_{\bf p}\) and \(\psi\) the wave functions of the photoelectron and the bound electron correspondingly; \(\gamma=-i\sqrt{\alpha}{\bf e}\cdot\nabla/m\) is the operator of interaction between the photon and electron.
\[F=\sqrt{\alpha}n(\omega)\int\frac{d^{3}f}{(2\pi)^{3}}\psi_{\bf p}({\bf f}) \frac{{\bf e}\cdot{\bf f}}{m}\psi({\bf f}-{\bf k}).\]
Since the photoelectron is described by the plane wave, i.e. \(\psi_{\bf p}({\bf f})=(2\pi)^{3}\delta({\bf f}-{\bf p})\), the amplitude of photoionization can be written as
\[F=N(\omega)\frac{{\bf e}\cdot{\bf p}}{m}\psi(p);\quad N(\omega)=\Big{(}\frac{4 \pi\alpha}{2\omega}\Big{)}^{1/2}. \tag{7}\]
We replaced \(q\) by \(p\) in the argument of the Fourier transform of the wave function of the fullerene electron. The latter is \(\psi(p)=\int d^{3}r\psi(r)e^{-i{\bf p}\cdot{\bf r}}\).
Now we present the wave function \(\psi(p)\) in terms of the Fourier transform of the potential
\[V(p)=\int d^{3}rV(r)e^{-i{\bf p}\cdot{\bf r}}=\frac{4\pi}{p}\int_{0}^{\infty} drrV(r)\sin pr. \tag{8}\]
The function \(\psi(p)\) can be expressed by the Lippmann-Schwinger equation
\[\psi=\psi_{0}+G(\varepsilon_{B})V\psi, \tag{9}\]
The matrix element of the propagator is
\[\langle{\bf f}_{1}|G(\varepsilon_{B})|{\bf f}_{2}\rangle=g(\varepsilon_{B},f_ {1})\delta({\bf f}_{1}-{\bf f}_{2});\quad g(\varepsilon_{B},f_{1})=\frac{1}{ \varepsilon_{B}-f_{1}^{2}/2m}.\]
For a bound state \(\psi_{0}=0\), and thus Eq.
\[J(p)=\int\frac{d^{3}f}{(2\pi)^{3}}\langle{\bf p}|V|{\bf f}\rangle\langle{\bf f }|\psi\rangle=\int\frac{d^{3}f}{(2\pi)^{3}}V({\bf p}-{\bf f})\psi({\bf f}). \tag{10}\]Putting also \(g(\varepsilon_{B},p)=-2m/p^{2}\) we obtain
\[\psi(p)=-\frac{2m}{p^{2}}J(p)=-\frac{J(p)}{\omega}. \tag{11}\]
The integral \(J(p)\) is saturated at \(f\sim\mu\ll p\). Thus its dependence on \(p\) is determined by that of \(V(p)\).
\[J(p)=\int d^{3}r\psi(r)V(r)e^{-i{\bf p}\cdot{\bf r}}=\frac{4\pi}{p}\int_{0}^{ \infty}dr\chi(r)V(r)\sin{(pr)};\quad\chi(r)=r\psi(r), \tag{12}\]
We shall demonstrate that the potential can take the form \(V(p)=V_{1}(p)+V_{2}(p)\) with the two terms corresponding to two fullerene characteristics \(R_{1}\) and \(R_{2}\).
\[J(p)=V_{1}(p)\kappa_{1}+V_{2}(p)\kappa_{2}, \tag{13}\]
Thus Eq.
\[\psi(p)=-\frac{1}{\omega}\Big{(}V_{1}(p)\kappa_{1}+V_{2}(p)\kappa_{2}\Big{)}, \tag{14}\]
can be presented as
\[F=-\frac{N(\omega)}{\omega}\frac{{\bf e}\cdot{\bf p}}{m}\sum_{i=1,2}V_{i}(p) \kappa_{i}. \tag{15}\]
Hence the asymptotics of the photoionization cross section can be expressed through the potential \(V(p)\):
\[\sigma(\omega)=\frac{4\alpha}{3}\frac{p}{\omega^{2}}n_{e}|\sum_{i}V_{i}(p) \kappa_{i}|^{2};\quad p=\sqrt{2m\omega}. \tag{16}\]
If the potential is determined by analytical function of \(R\) and \(\Delta\), we have just \(J(p)=V(p)\kappa\). It was shown in that in the simplest case \(\kappa=\psi(r=0)\).
Thus one can find the asymptotic energy dependence of the photoionization cross section without solving the wave equation.
## 3 Asymptotics for the model potentials
Start with the well potential defined by Eq.. Employing Eq.
\[\psi(p)=-\frac{4\pi V_{0}R}{\omega p^{2}}\lambda;\quad\lambda=\cos{(pR_{2})} \psi(R_{2})-\cos{(pR_{1})}\psi(R_{1}). \tag{17}\]
Here we neglected the terms of the order \(\Delta/R\ll 1\). The parameter \(\lambda\) can be expressed in terms of the fullerene characteristics \(R\) and \(\Delta\). Note that the function \(\psi(r)\) varies noticeably inside the fullerene layer \(\Delta\). Hence the difference \(\psi(R_{2})-\psi(R_{1})\) is not small.
\[\sigma=\frac{2^{6}\alpha\pi^{2}}{3}\frac{V_{0}^{2}R^{2}}{\omega^{2}p^{3}}n_{e} \lambda^{2}. \tag{18}\]
It drops as \(\omega^{-7/2}\).
For the Dirac bubble potential given by Eq. we obtain employing Eq.
\[\psi(p)=\frac{4\pi U_{0}R}{\omega p}\sin{(pR)}\psi(R), \tag{19}\]
Thus the cross section is
\[\sigma=\frac{2^{6}\alpha\pi^{2}}{3}\frac{U_{0}^{2}R^{2}}{\omega^{2}p}\sin^{2}{ (pR)}n_{e}\psi^{2}(R);\quad p^{2}=2m\omega, \tag{20}\]
This is just the result obtained in by solving the wave equation.
Both Dirac bubble and well potentials have singularities on the real axis. However, the wave functions exhibit different behavior at the singular points. This leads to different high energy behavior in these cases. The wave function corresponding to the Dirac bubble potential is continuous at \(r=R\) while its first derivative suffers a jump at this point.
\[\chi^{}(r)=2mV(r)\chi(r)-2m\varepsilon_{B}\chi(r), \tag{21}\]
over the small interval \(R^{-}\leq r\leq R^{+}\) with \(R^{\pm}=R\pm\delta,\delta\to 0\).
The wave function of the bound electron in momentum representation can be expressed in terms of these jumps.
\[\psi(p)=\frac{4\pi}{p}\int_{0}^{\infty}dr\sin{(pr)}\chi(r),\]
we find for the Dirac bubble potential
\[\psi(p)=\frac{4\pi}{p^{3}}R\sin{(pR)}[\psi^{\prime}(R^{-})-\psi^{\prime}(R^{+})] \tag{22}\]
\[-\frac{4\pi}{p^{3}}\Big{[}\int_{0}^{R^{-}}dr\sin{pr}\chi^{}(r)+\int_{R^{+}} ^{\infty}dr\sin{pr}\chi^{}(r)\Big{]}\]
Further integration by parts of the second term demonstrates that it is about \(1/p\) times the first one. Hence the leading contribution is provided by the first term and the Fourier transform of the function \(\psi(r)\) is determined by the jump of the first derivative. This leads to the \(\omega^{-5/2}\) law for the cross section.
In the case of the well potential one can present the function \(\psi(p)\) in similar form, but with two singular points \(r=R_{2,1}=R\pm\Delta/2\). The first derivatives \(\psi^{\prime}(r)\) are continuous at these points while the second derivatives experience jumps.
\[\psi(p)=-\frac{4\pi}{p^{4}}\sum_{i=1,2}R_{i}\cos{(pR_{i})}\delta\psi^{}(R_{ i})\]where \(\delta\psi^{}(R_{i})=\psi^{}(R_{i}+\delta)-\psi^{}(R_{i}-\delta)\) are the jumps of the second derivative \(\psi^{}(R_{i})\). Thus the asymptotic wave function in the field of the well potential is \(1/p\) times that in the Dirac bubble potential. This provides additional small factor of the order \(1/\omega\) in the photoionization cross section given by Eq.compared with that for the Dirac bubble potential presented by Eq..
The potential \(V(r)\) for the jellium model given by Eq. is continuous at the real axis as well as its derivative \(V^{\prime}(r)\). One can see that \(V_{1}(R_{1})=V_{2}(R_{1})\) and \(V_{2}(R_{2})=V_{3}(R_{2})\). Also \(V^{\prime}_{1}(R_{1})=V^{\prime}_{2}(R_{1})\), \(V^{\prime}_{2}(R_{2})=V^{\prime}_{3}(R_{2})\). The second derivative \(V^{}(r)\) experiences jumps at \(r=R_{1,2}\). Employing Eq.
\[\psi(p)=\frac{4\pi}{\omega p^{4}}\lambda;\quad\lambda=\sum_{i=1,2}R_{i}\cos{( pR_{i})}\psi(R_{i})\delta V^{}(R_{i});\quad\delta V^{}(R_{i})=V^{}(R_{i} ^{+})-V^{}(R_{i}^{-}). \tag{23}\]
Note that the second derivatives \(\psi^{}(r)\) are continuous at \(r=R_{1,2}\). This can be seen from the wave equation.
\[\sigma=\frac{64\alpha\pi^{2}}{3\omega^{2}p^{7}}n_{e}\lambda^{2}. \tag{24}\]
It drops as \(\omega^{-11/2}\).
Turn now to the Lorentz bubble potential determined by Eq. with \(a\ll R\). Employing Eq.
\[V(p)=V_{A}(p)+V_{B}(p);\quad V_{A,B}=-4\frac{a}{p}X_{A,B},\]
with
\[X_{A}(p)=U_{0}(-\frac{\partial}{\partial p})\cos{(pR)}\int_{-\infty}^{\infty} dx\cos{(px)}u(x); \tag{25}\]
\[X_{B}(p)=U_{0}\frac{\partial}{\partial p}\int_{R}^{\infty}dx\cos{((p(x-R))u(x )}\quad u(x)=\frac{1}{x^{2}+a^{2}}.\]
Note that \(X_{A}(p)\) is determined by \(r\) close to \(R\), i.e. by \(|R-r|\sim a\).
\[X_{B}(p)=-U_{0}\int_{0}^{\infty}dr^{\prime}r^{\prime}\sin{(pr^{\prime})}u(R+r ^{\prime}).\]
It is dominated by small \(r^{\prime}\sim 1/p\).
Since the potential \(V_{A}\) is determined by the space region where the electron density reaches it largest values, we expect it to be the most important one. Neglecting for a while the contribution of \(V_{B}\) we find in the lowest order of expansion in powers of \(a/R\)
\[X_{A}(p)=\pi\frac{U_{0}R}{a}e^{-pa}\sin{pR};\quad V(p)=-4\pi\frac{U_{0}R}{p}e^ {-pa}\sin{pR}. \tag{26}\]
We calculate the function \(\psi(p)\) by employing Eqs. and. The potential \(V(v)=-4\pi U_{0}Re^{-va}\sin{(vR)}/v\), with \(v=|{\bf p}-{\bf f}|\) can not be expanded in powers of \(f\). Note, however that it is sufficient to put \(v=p-{\bf p}\cdot{\bf f}/p\) for calculation of the asymptotics of the wave function.
\[\psi(p)=4\frac{\pi U_{0}R}{\omega}\int\frac{d^{3}f}{(2\pi)^{3}}\frac{e^{-va}}{v} [\sin{(pR)}\cos{(ftR)}-\cos{(pR)}\sin{(ftR)}]\psi(f);\quad t={\bf p}\cdot{\bf f }/pf. \tag{27}\]
Due to the factor in the square brackets the integral over \(f\) on the right hand side of Eq. is saturated by small \(|ft|\sim 1/R\ll 1/a\). This enables to put \(v=p\) and \(e^{-(p-ft)a}=e^{-pa}\) in the integrand.
\[\psi(p)=4\frac{\pi U_{0}R}{\omega p}e^{-pa}\sin{(pR)}A;\quad A=\frac{4\pi}{R} \int_{0}^{\infty}\frac{dff}{(2\pi)^{3}}\sin{(fR)}\psi(f). \tag{28}\]
Noting that \(A=\psi(R)\), we obtain
\[\psi(p)=4\frac{\pi U_{0}R}{\omega p}e^{-pa}\sin{(pR)}\psi(R). \tag{29}\]
The cross section
\[\sigma_{A}=\frac{64}{3}\frac{\alpha\pi^{2}U_{0}^{2}R^{2}}{\omega^{2}p}e^{-2pa} \sin^{2}{(pR)}n_{e}\psi^{2}(R), \tag{30}\]
by the exponential factor \(e^{-2pa}\). Note that the latter is caused by the poles \(r=R\pm ia\) of the potential in the complex plane.
The lowest term of the asymptotic expansion of the function \(X_{B}(p)\) provides the potential
\[V_{B}(p)=\frac{16U_{0}a}{p(pR)^{3}}.\]
From the formal point of view, the corresponding cross section
\[\sigma_{B}=\frac{2^{10}\alpha}{3}\frac{U_{0}^{2}a^{2}}{R^{6}}\frac{\psi^{2}(r= 0)}{\omega^{2}p^{7}}, \tag{31}\]
However, e.g., at characteristic values \(R=6\)a.u, \(a=\Delta=1\) a.u. the cross section determined by Eq. becomes comparable with that determined by Eq. only at the photoelectron energies \(\varepsilon\geq 5\) keV. At these energies both cross sections \(\sigma_{A}\) and \(\sigma_{B}\) are more than \(10^{8}\) times smaller than the cross section \(\sigma_{A}\) at \(\varepsilon=100\) eV. They have no chances to be observed. Thus only the cross section \(\sigma_{A}\) is of physical interest.
Similar analysis can be carried out for the Gaussian-type potential given by Eq.. Now \(V(p)=V_{A}(p)+V_{B}(p)\) with \(V_{A,B}(p)=-4V_{0}X_{A,B}(p)/p\). The functions \(X_{A}(p)\) and \(X_{B}(p)\) are given by Eq. with \(u(x)=e^{-x^{2}/a^{2}}\) and \(U_{0}=1\).
\[\sigma=\frac{64\alpha\pi}{3}\frac{V_{0}^{2}R^{2}a^{2}}{\omega^{2}p}e^{-p^{2}a^ {2}/2}\sin^{2}{(pR)}n_{e}\psi^{2}(R), \tag{32}\]
The Gaussian drop changes to the power behavior \(\sigma(\omega)\sim\omega^{-11/2}\) at the photoelectron energies of about 3 keV. Here the cross section is too small to be detected.
Summary
We found that the high energy behavior of the photoionization cross section of fullerenes depends on the form of the function \(V(r)\) which is chosen for approximation of the fullerene field. We expressed the asymptotic cross section in terms of the Fourier transform \(V(p)\) of the potential \(V(r)\) without solving the wave equation.
The shape of the function \(V(p)\) at large \(p\) is known to be determined by the analytical properties of the function \(V(r)\). Thus the latter determine the shape of the asymptotic cross section as well.
The three potentials presented by Eqs.-have singularities on the real axis. In each case the cross section exhibits a power drop. The cross section decreases as \(\omega^{-5/2}\) in the Dirac bubble potential which turnes to infinity at \(r=R\). It behaves as \(\omega^{-7/2}\) in the well potential with the finite jumps of \(V(r)\). The potential of jellium model is more smooth, and only the second derivatives \(V^{}(r)\) experience jumps. In this model the cross section drops as \(\omega^{-11/2}\).
The Lorentz bubble potential given by Eq. has poles in the complex plane. In this case the observable cross section exhibits exponential drop \(e^{-2pa}\). The Gaussian type potential determined by Eq. with the essential singularity in the complex plane provides Gaussian drop of the photoionization cross section. In both cases these fast drops change to a slower power drop \(\omega^{-11/2}\). However this takes place at the energies of several keV where the cross sections become unobservably small.
|
10.48550/arXiv.1703.07712
|
High energy photoionization of fullerenes
|
E. G. Drukarev, A. I. Mikhailov
| 625
|
10.48550_arXiv.1611.04886
|
## 4 The SPDERG approach to boundary layer problems
The SPDERG approach and its connections with the renormalization group, with attention to the context of boundary layer problems, have been extensively reviewed in, and a preliminary attempt to apply it to MM kinetics within the tQSSA framework is presented in that work, too. Moreover, a recent review of the general method, in a context that is instead different from the boundary layer one, is presented in.
In \(A)\) and \(C)\) we present the behaviour of the concentrations of the substrate \(s(t)\), whereas in \(B)\) and \(D)\) we present the ones of the complex \(c(t)\) for the \(a\) and \(b\) sets of ICVs given in, respectively. Hence, in \(A)\) and \(B)\) we are in the case with \(\varepsilon=\varepsilon^{a}=0.1\), whereas in \(C)\) and \(D)\) we are in the one with \(\varepsilon=\varepsilon^{b}=0.5\). We plot both the numerical solutions of Eqs. already shown in the previous figures, and the analytical solutions computed from the 1st order PE UAs (by using a standard numerical approximation for the Lambert function) as given in. We finally plot our corresponding rough evaluations of the two different time scales involved, too, with \(\tau_{s}\) describing the substrate decay time and \(\tau_{c}\) the complex saturation time. Notice that the time is in logarithmic scale.
Indeed, in those works, the method was already shown successful in some of these last cases.
As MM kinetics, these problems are generally characterized by a boundary layer of a given small thickness (that is here \(O(\varepsilon)\), but that could also be, for instance, \(O(\sqrt{\varepsilon})\)), in which the solution is rapidly varying. Correspondingly, in order to predict the system's dynamics, one needs to solve singularly perturbed ODEs and one usually resorts to a standard PE method.
To start to sketch the general SPDERG procedure in these cases, one starts from the singular 2nd order ODE for \(y(t)\) and makes the transformation \(t\to\tau=t/\varepsilon\), thereby studying the ODE to be obeyed by the inner solution \(Y(\tau)\). The approach basically consists in focusing on the large \(\tau\) behaviour of this solution, by outlining the secular terms, with the aim of proposing first of all the correct renormalization of the integration constants, that allows to eliminate them.
At this point, we anticipate that in the end one usually gets a physically meaningful solution, that also contains the leading terms of the outer expansion, and that is therefore to be interpreted as the UA within this approach, to be compared with the one of the standard PE methods. As we are going to verify in detail in the present application, this approximation is expected to work even better than the corresponding UA solution of the PE at the previous order in the matching region. Moreover, as we are going to discuss, it appears possible to make an analogy between the basic mechanism of the SPDERG approach to boundary layer problems and the one that allows to impose the correct MCs in the standard methods.
More precisely, one studies the solution \(Y(\tau)\) with given boundary conditions at a given time \(\tau_{0}\) and _renormalizes_ the \(Y_{div}(\tau)\) part of this solutions that contains the Cauchy data, the other leading order terms and the terms that do not tend to zero or to a constant value at large times. This is obtained by renormalizing the corresponding _bare_ integration constants (_i.e._, the Cauchy data of the problem, though these integration constants can also more generally _contain_ them in some given form), let us say \(A_{0}(\tau_{0})\) and \(B_{0}(\tau_{0})\), as \(A_{0}(\tau_{0})\to Z_{1}A(\lambda)\), \(B_{0}(\tau_{0})\to Z_{2}B(\lambda)\). In fact, the renormalization constants \(Z_{1}\) and \(Z_{2}\) depend both on \(\tau_{0}\) and \(\lambda\), their basic role being to change, in the part to be renormalized of the function, the initial dependence of the Cauchy data on \(\tau_{0}\) in a dependence on the arbitrary time \(\lambda\).
In detail, they are assumed to have the expansion (with \(a_{0}=b_{0}=1\)):
\[Z_{1}=\sum_{n=0}^{\infty}a_{n}(\tau_{0},\lambda)\varepsilon^{n};\ \ \ \ Z_{2}= \sum_{n=0}^{\infty}b_{n}(\tau_{0},\lambda)\varepsilon^{n}. \tag{26}\]
Since one can always take \((\tau-\tau_{0})=(\tau-\lambda)+(\lambda-\tau_{0})\), the secular terms in \((\lambda-\tau_{0})\) are correspondingly absorbable in appropriate redefinitions of \(A_{0}\) and \(B_{0}\), that have to be made by correctly choosing the coefficients \(\{a_{n}\},\{b_{n}\}\). As it will become clear in the application that we are going to present, one can usually safely expect to be able to absorb in the coefficients of the renormalization constants also other possible secular terms, for instance those corresponding to higher powers of \((\tau-\tau_{0})\). In particular, we will see in the present 2nd order calculations that, when writing \((\tau-\tau_{0})^{2}=(\tau-\lambda)^{2}+(\lambda-\tau_{0})^{2}+2(\tau-\lambda)( \lambda-\tau_{0})\), only the second term is to be absorbed, whereas the last one is cancelled by the previously chosen 1st order renormalization coefficients, in the same way as in a case that is considered in (see Section B in that work).
Summarizing, the basic hypothesis, that is analogous to the scaling one in the renormalization group theory, is that the bare quantities need to be renormalized in such a way that the solution turns out to be independent of the arbitrary time \(\lambda\), as it has reasonably to be. Hence, \(\lambda\) plays a key role in the whole approach. Intuitively, one can therefore think to \(\lambda\) as the equivalent of the _unknown time_ at which the matching needs to be obeyed, and from this point of view there is an analogy with the imposition of the MCs in the standard PE methods. Nevertheless, though this analogy can be useful, as we will show, it is not to be taken too rigorously, particularly in the present peculiar case in which the 1st order matching requires two term conditions, and one expects that the correct matching could require even more terms at higher orders.
Going back to the recalling of the main details of the approach in the case of boundary layer problems, the appropriate choice of the coefficients of the renormalization constants, and the appropriate redefinitions of \(A_{0}\) and \(B_{0}\), turns out in the replacement of these bare quantities with \(A(\lambda)\) and \(B(\lambda)\), and in the corresponding change of variable \(\tau_{0}\to\lambda\) in \(Y_{div}(\tau)\). Then, on this \(\lambda\)-depending \(Y_{div}\), one imposes the _scaling_ condition \(dY_{div}/d\lambda=0\), that allows to obtain the 1st order ODEs to be obeyed by \(A(\lambda)\) and \(B(\lambda)\). Finally, the change of variable \(\lambda\to\tau\) in the renormalized \(Y_{div}\), in which the integration constants are replaced by the solutions of these equations, and the imposition of the original boundary conditions of the problem, allow to get the physically meaningful result.
## 5 Results and discussion: i) First order contribution
Here we use, for the first time to our knowledge, the SPDERG approach to study MM kinetics beyond the sQSSA, whose known PE UAs we already recalled in Sections 2 and 3. In order to make the presentation as clear as possible, we stress from the beginning that we introduce a procedure that is slightly different from the one considered in. In fact, we make explicit the dependence from the ICVs in the solutions, thus outlining that, though in principle one can renormalize even general integration constants that contain them, the bare quantities that are basically to be renormalized are just the ICVs of the problem.
We notice first of all that one can rewrite the system for the inner solutions in the form of a 2nd order ODE (see in particular for the first study of MM kinetics in terms of a 2nd order ODE) to be obeyed by the adimensional substrate concentration:
\[\ddot{\tilde{s}}^{in}(\tau)-\frac{\left[\dot{\tilde{s}}^{in}(\tau)\right]^{2}} {\tilde{s}^{in}(\tau)+m}+\left[\tilde{s}^{in}(\tau)+M\right]\dot{\tilde{s}}^{ in}(\tau)+\varepsilon\left[\frac{m}{\tilde{s}^{in}(\tau)+m}\dot{\tilde{s}}^{ in}(\tau)+(M-m)\dot{\tilde{s}}^{in}(\tau)\right]=0, \tag{27}\]
Analogously, one can write the same system in the form of a 2nd order ODE to be obeyed by the adimensional complex concentration:
\[\left[1-\tilde{c}^{in}(\tau)\right]\ddot{\tilde{c}}^{in}(\tau)+\left[\dot{ \tilde{c}}^{in}(\tau)\right]^{2}+M\tilde{c}^{in}(\tau)\dot{\tilde{c}}^{in}( \tau)+\varepsilon\left[1-\tilde{c}^{in}(\tau)\right]^{2}\left[\dot{\tilde{c}} ^{in}(\tau)+(M-m)\dot{c}^{in}(\tau)\right]=0, \tag{28}\]
Actually, the physically meaningful solutions of these 2nd order ODEs are identical to the solutions of the original system. In fact, at the 1st order in \(\varepsilon\), the equations are satisfied by \(\tilde{s}^{in}(\tau)=\tilde{s}^{in}_{0}(\tau)+\varepsilon\tilde{s}^{in}_{1}(\tau)\) and \(\tilde{c}^{in}(\tau)=\tilde{c}^{in}_{0}(\tau)+\varepsilon\tilde{c}^{in}_{1}(\tau)\), with \(\tilde{s}^{in}_{0}(\tau)\), \(\tilde{c}^{in}_{0}(\tau)\) given by, and \(\tilde{s}^{in}_{1}(\tau)\), \(\tilde{c}^{in}_{1}(\tau)\) given by, respectively.
Therefore, instead of attempting to apply the approach to these 2nd order ODEs, we study the SPDERG (_i.e., renormalization group_, that we label \(rg\)) adimensional substrate and complex concentrations, \(\tilde{s}^{rg}(\tau)\) and \(\tilde{c}^{rg}(\tau)\), that are solutions of the system. The difference, with respect to the \(\tilde{s}^{in}(\tau)\) and \(\tilde{c}^{in}(\tau)\) that we previously considered within the PE method, is that here the ICVs are given at a time \(\tau_{0}\), to be considered in principle different from zero, as \(\tilde{s}^{rg}(\tau_{0})=\tilde{s}^{*}\) and \(\tilde{c}^{r}{}^{g}(\tau_{0})=\tilde{c}^{*}\), respectively. Indeed, these \(\tilde{s}^{*}\) and \(\tilde{c}^{*}\) values are the bare quantities to be renormalized, whereas the original ICVs of the problem, _i.e._, \(\tilde{s}=1\) and \(\tilde{c}=0\), will be taken into account after the renormalization procedure.
For the sake of clarity, at the cost of being somehow repetitive, at the 1st order in \(\varepsilon\), we search once again solutions in the form:
\[\left\{\begin{array}{l}\tilde{s}^{rg}(\tau)=\tilde{s}^{reg}_{0}(\tau)+ \varepsilon\tilde{s}^{rg}_{1}(\tau)\\ \tilde{c}^{rg}(\tau)=\tilde{c}^{reg}_{0}(\tau)+\varepsilon\tilde{c}^{reg}_{1} (\tau).\end{array}\right. \tag{29}\]
In these formulas, \(\tilde{s}^{rg}_{0}(\tau)\) and \(\tilde{c}^{rg}_{0}(\tau)\) solve as usual the 0th order system, but the ICVs are given at \(\tau=\tau_{0}\), and their values are \(\tilde{s}^{reg}_{0}(\tau_{0})=\tilde{s}^{*}_{0}\) and \(\tilde{c}^{rg}_{0}(\tau_{0})=\tilde{c}^{*}_{0}\). These solutions are:
\[\left\{\begin{array}{lcl}\tilde{s}^{rg}_{0}(\tau)&=&\tilde{s}^{*}_{0}\\ \tilde{c}^{rg}_{0}(\tau)&=&\tilde{c}^{*}_{0}e^{-}(\tilde{s}^{*}_{0}+M)(\tau- \tau_{0})+\frac{\tilde{s}^{*}_{0}}{\tilde{s}^{*}_{0}+M}\left[1-e^{-}(\tilde{ s}^{*}_{0}+M)(\tau-\tau_{0})\right].\end{array}\right. \tag{30}\]
On the other hand, \(\tilde{s}^{rg}_{1}(\tau)\) and \(\tilde{c}^{rg}_{1}(\tau)\) need to solve the 1st order system (that is slightly different from Eqs. because of the presence of \(\tilde{s}^{*}_{0}\neq 1\)):
\[\left\{\begin{array}{lcl}\dot{\tilde{s}}^{rg}_{1}(\tau)&=&(\tilde{s}^{*}_{0} +m)\tilde{c}^{rg}_{0}(\tau)-\tilde{s}^{*}_{0}\\ \dot{\tilde{c}}^{rg}_{1}(\tau)&=&-(\tilde{s}^{*}_{0}+M)\tilde{c}^{rg}_{1}(\tau )-\left[\tilde{c}^{rg}_{0}(\tau)-1\right]\tilde{s}^{rg}_{1}(\tau),\end{array}\right. \tag{31}\]
We obtain, in the case of the substrate:
\[\tilde{s}^{rg}_{1}(\tau) = \tilde{s}^{*}_{1}-(M-m)\frac{\tilde{s}^{*}_{0}}{\tilde{s}^{*}_{0} +M}(\tau-\tau_{0})+\]\[- \frac{\tilde{s}_{0}^{rg}+m}{(\tilde{s}_{0}^{*}+M)^{2}}\left[\tilde{s}_ {0}^{*}-\tilde{c}_{0}^{*}(\tilde{s}_{0}^{*}+M)\right]\left[1-e^{-(\tilde{s}_{0}^ {*}+M)(\tau-\tau_{0})}\right], \tag{32}\]
Let us start by studying \(\tilde{s}^{rg}(\tau)\). Following, we look at \(\tilde{s}_{div}^{rg}(\tau)\), that contains just the ICVs (_i.e._, the terms that give \(\tilde{s}^{rg}(\tau_{0})=\tilde{s}_{0}^{*}+\varepsilon\tilde{s}_{1}^{*}\)), the other possible leading terms at this order (absent in the present case), and the secular terms (here the one proportional to \((\tau-\tau_{0})\) in \(\tilde{s}_{1}^{rg}(\tau)\)). Therefore, we write:
\[\tilde{s}^{rg}(\tau)=\tilde{s}_{div}^{rg}(\tau)+\varepsilon{\cal R}_{1}^{*}( \tau)+O(\varepsilon^{2}), \tag{33}\]
One has:
\[\tilde{s}_{div}^{rg}(\tau)=\tilde{s}_{0}^{*}+\varepsilon\tilde{s}_{1}^{*}- \varepsilon(M-m)\frac{\tilde{s}_{0}^{*}}{\tilde{s}_{0}^{*}+M}(\tau-\tau_{0}). \tag{34}\]
As previously recalled, the SPDERG approach, at this point, is based on writing \((\tau-\tau_{0})=(\tau-\lambda)+(\lambda-\tau_{0})\), by correspondingly assuming, in the present case, that the secular term proportional to \((\lambda-\tau_{0})\) can be absorbed in an appropriate redefinition of the bare constants. In detail, by labelling \(\tilde{s}_{0}^{rg}\) and \(\tilde{s}_{1}^{rg}\) the contributions to the renormalized ICV at the 0th and at the 1st order, respectively, we put \(\tilde{s}_{0}^{*}=Z_{s_{0}}\tilde{s}_{0}^{rg*}(\lambda)\) and \(\tilde{s}_{1}^{*}=Z_{s_{1}}\tilde{s}_{1}^{rg*}(\lambda)\). The renormalization constants, \(Z_{0}^{s}\) and \(Z_{1}^{s}\), are assumed to have the same \(\varepsilon\)-expansions as the ones given in. This expansion, at the 1st order in \(\varepsilon\), that we considered here, implies \(\tilde{s}_{0}^{*}=[1+\varepsilon z_{s_{0},1}(\tau_{0},\lambda)]\tilde{s}_{0}^ {rg*}(\lambda)\) and \(\tilde{s}_{1}^{*}=\tilde{s}_{1}^{rg*}(\lambda)\). Thus, the present secular term can be absorbed by choosing:
\[z_{s_{0},1}(\tau_{0},\lambda)=\frac{(M-m)}{\tilde{s}_{0}^{rg*}+M}(\lambda- \tau_{0}). \tag{35}\]
Therefore, we end up to study:
\[\tilde{s}_{div}^{rg}(\tau,\lambda)=\tilde{s}_{0}^{rg*}(\lambda)+\varepsilon \tilde{s}_{1}^{rg*}(\lambda)-\varepsilon(M-m)\frac{\tilde{s}_{0}^{rg*}(\lambda )}{\tilde{s}_{0}^{rg*}(\lambda)+M}(\tau-\lambda). \tag{36}\]
Hence, by imposing the scaling condition, \(ds_{div}^{rg}(\tau,\lambda)/d\lambda=0\), we get the 1st order ODEs to be obeyed by the renormalized quantities.
In detail, at the 1st order in \(\varepsilon\), we obtain:
\[\frac{ds_{0}^{rg*}(\lambda)}{d\lambda}+\varepsilon\frac{ds_{1}^{rg*}(\lambda) }{d\lambda}-\varepsilon(M-m)\frac{\tilde{s}_{0}^{rg*}(\lambda)}{\tilde{s}_{0} ^{rg*}(\lambda)+M}=0, \tag{37}\]
and, correspondingly:
\[\left\{\begin{array}{rcl}\frac{dS_{0}^{rg*}(\lambda)}{d\lambda}&=&- \varepsilon(M-m)\frac{\tilde{s}_{0}^{rg*}(\lambda)}{\tilde{s}_{0}^{rg*}( \lambda)+M}\\ \frac{dS_{1}^{rg*}(\lambda)}{d\lambda}&=&0.\end{array}\right. \tag{38}\]
Consistently, when making the further transformation\(\lambda\rightarrow\tau=t/\varepsilon\), the first of these equations is just the ODE to be obeyed by the 0th order outer adimensional substrate concentration, that we encountered in Eqs.. Its solution is therefore \(\tilde{s}_{0}^{rg*}(t)=\tilde{s}_{0}^{out}(t)\), that is given in by means of the Lambert function, and that also satisfies the ICV \(\tilde{s}_{0}^{rg*}=1\).
Clearly, this result already suggests that, also in the present particular demanding case of MM kinetics, the SPDERG approach proposed in turns out to be able to reproduce the leading order terms of the PE UAs.
On the other hand, the result \(d\tilde{s}_{1}^{rg*}(\lambda)/d\lambda=0\) is here to be interpreted as the verification that the \(\tilde{s}_{1}^{rg*}\) term can be neglected at this order. In fact, we find \(\tilde{s}_{1}^{rg*}=const=0\) by imposing the ICV (more correctly the ICV, that is fixed here to the value \(\tilde{s}=1\), should be imposed on the solution at the end, but this is not influential in the present case, since one generally has \({\cal R}=0\) for the appropriately calculated contribution of the parts not to be renormalized of the inner solutions).
Finally, we get the renormalized result:
\[\tilde{s}_{1}^{rg,u}(t)=\tilde{s}_{0}^{out}(t)+\varepsilon{\cal R}_{1}^{*}(t /\varepsilon)+O(\varepsilon). \tag{39}\]Notice that, when we were considering \(\tilde{s}^{rg}(\tau)\) as the inner solution, in, before applying the SPDERG approach, we knew that it was correct up to order \(O(\varepsilon^{2})\). Here, instead, the renormalized \(\tilde{s}_{1}^{rg,u}(t)\) (that for this reason we label \(\tau\!g,u\), by moreover making explicit that it is a 1st order approximation) is to be interpreted as the SPDERG UA to the solution of the original problem, clearly bearing in mind that, from this point of view, it is correct only up to order \(O(\varepsilon)\). Actually, it is only expected to contain the leading order terms of the outer solution, as one can check that is indeed the case, by comparing it with.
Notice moreover that the terms in \({\cal R}_{1}^{s}\), that did not need to be renormalized, are to be evaluated in \(\tau_{0}=0\), by using the correct ICVs for the bare constants (_i.e._, \(\tilde{s}_{0}^{*}=1\), \(\tilde{s}_{1}^{*}=0\), \(\tilde{c}_{0}^{*}=0\)). Thus, the SPDERG 1st order UA to the solution for the adimensional substrate concentration does instead contain the 1st order terms of the PE UA to the correct solution that comes from \(\tilde{s}_{1}^{in}(t/\varepsilon)\). On the other hand, correspondingly, the usual asymptotically vanishing solution behaviour that one finds when applying the recalled PE method to MM kinetics (_i.e._, \(\lim_{t\to\infty}\tilde{s}(t)=0\)), is not verified here (one has \(\lim_{t\to\infty}\tilde{s}_{1}^{rg,u}(t)=O(\varepsilon)\)). We will better discuss this failure of the present application of the SPDERG approach in the following, by proposing a way to bypass it, too.
Let us now study the complex. First of all, despite of the presence of a large number of terms in (the formula given in Appendix A), that describes the behaviour of the 1st order inner adimensional complex concentration within the SPDERG approach, \(\tilde{c}_{1}^{rg}(\tau)\), the part of the function to be renormalized up to the 1st order in \(\varepsilon\) is as manageable as in the case of the substrate.
In detail, we write the same kind of expression as in:
\[\tilde{c}^{rg}(\tau)=\tilde{c}_{div}^{rg}(\tau)+\varepsilon\mathcal{R}_{1}^{c} (\tau)+O(\varepsilon^{2}), \tag{40}\]
Correspondingly, we find:
\[\tilde{c}_{div}^{rg}(\tau)=(\tilde{c}_{0}^{*}+\varepsilon\tilde{c}_{1}^{*})e^ {-}(\tilde{s}_{0}^{*}+M)(\tau-\tau_{0})+\frac{\tilde{s}_{0}^{*}}{\tilde{s}_{0 }^{*}+M}-\varepsilon\frac{M(M-m)}{\left(\tilde{s}_{0}^{*}+M\right)^{3}}\tilde{ s}_{0}^{*}(\tau-\tau_{0}). \tag{41}\]
In fact, here we considered explicitly the terms that give the original ICV, (_i.e._, \(\tilde{c}_{div}^{rg}=\tilde{c}_{0}^{*}+\varepsilon\tilde{c}_{1}^{*}\)). Nevertheless, both the 0th order contribution and the 1st order one to it are exponentially suppressed in the large time limit. Therefore, one can hypothesize (see an analogous case in, Section C) that also \(\tilde{c}_{0}^{rg*}(\lambda)\), when renormalizing, should obey an ODE of the kind \(d\tilde{c}_{0}^{rg*}(\lambda)/d\lambda=0\). Hence, one expects to find \(\tilde{c}_{0}^{rg*}(\lambda)=const=0\) when imposing the ICV (fixed here to \(\tilde{c}=0\)) at the end of the renormalization procedure. Moreover, one also expects \(\tilde{c}_{1}^{rg*}(\lambda)=const=0\), a result that is even more predictable at this order, since we already verified that \(\tilde{s}_{1}^{rg*}\) gives no contribution in the case of the substrate. Correspondingly, we assume that these terms can be neglected in the present application of the SPDERG approach to MM kinetics.
Therefore, the function to be studied is further reduced to:
\[\tilde{c}_{div}^{rg}(\tau)\left|{}_{\tilde{c}_{1}^{\sigma=0}\atop\tilde{c}_{1 }^{\sigma=0}}\right.=\frac{\tilde{s}_{0}^{*}}{\tilde{s}_{0}^{*}+M}-\varepsilon (M-m)\frac{M\tilde{s}_{0}^{*}}{\left(\tilde{s}_{0}^{*}+M\right)^{3}}(\tau- \tau_{0}). \tag{42}\]
As already recalled, in the context of the standard method, the MC that is verified by the adimensional substrate concentration does consistently also satisfy the matching of the inner and outer adimensional complex concentrations. For this reason, on the basis of the previously sketched analogy, we find not particularly surprising that, when making the transformations \(\left\{\tau_{0}\to\lambda,\tilde{s}_{0}^{*}\to\tilde{s}_{0}^{rg*}(\lambda)\right\}\), the scaling condition \(d\tilde{c}_{div}^{rg}(\tau,\lambda)/d\lambda=0\) gives again the same ODE to be obeyed by \(\tilde{s}_{0}^{rg*}(\lambda)\) as the one that we just found in the study of \(\tilde{s}_{div}^{rg}(\tau,\lambda)\) (_i.e._, Eqs.). Hence (with the transformation \(\lambda\to\tau=t/\varepsilon\)), we get again \(\tilde{s}_{0}^{rg*}(t)=\tilde{s}_{0}^{out}(t)\). One can check in particular that, since \(d[\tilde{s}_{0}^{*}/(\tilde{s}_{0}^{*}+M)]/d\tilde{s}_{0}^{*}=M/(\tilde{s}_{0} ^{*}+M)^{2}\), when renormalizing the bare constant \(\tilde{s}_{0}^{*}\) by using the value of the coefficient \(z_{s_{0},1}\) that is already fixed by, the contribution of the first term to the order \(\varepsilon\) is exactly equal to \(\varepsilon(\lambda-\tau_{0})M(M-m)\tilde{s}_{0}^{rg*}/(\tilde{s}_{0}^{rg*}+M)^ {2}\), and therefore it suitably absorbs the secular term that is present in this case.
By recalling that \(\tilde{c}_{0}^{out}(t)=\tilde{s}_{0}^{out}(t)/(\tilde{s}_{0}^{out}(t)+M)\) from Eqs., the obtained SPDERG 1st order UA to the correct solution is:
\[\tilde{c}_{1}^{rg,u}(t)=\tilde{c}_{0}^{out}(t)+\varepsilon\mathcal{R}_{1}^{c} (t/\varepsilon)+O(\varepsilon), \tag{43}\]
Writing explicitly the contribution of the different terms, we obtain the following SPDERG 1st order UAs to the correct solutions for the time behaviour of the adimensional substrate concentration and of the adimensional complex concentration in MM kinetics beyond the sQSSA, respectively (with \(\omega\) the Lambert function):
\[\left\{\begin{array}{lcl}\hat{s}_{1}^{rg,u}(t)&=&M\omega(e^{-(M-m)t/M+1/M}/M)+ \\ &-&\varepsilon\frac{1+m}{(1+M)^{2}}\left[1-e^{-(1+M)t/\varepsilon}\right]+O( \varepsilon)\\ \hat{c}_{1}^{rg,u}(t)&=&\frac{\omega(e^{-(M-m)t/M+1/M}/M)}{\omega(e^{-(M-m)t/M+1 /M}/M)+1}-\frac{1}{1+M}e^{-(1+M)t/\varepsilon}+\\ &-&\varepsilon\frac{M(1+2m-M)}{(1+M)^{4}}\left[1-e^{-(1+M)t/\varepsilon}\right]+ \\ &-&\varepsilon\left[\frac{(1-M)(1+m)}{(1+M)^{3}}(t/\varepsilon)+\frac{M-m}{(1+M )^{2}}\frac{(t/\varepsilon)^{2}}{2}\right]e^{-(1+M)t/\varepsilon}+\\ &+&\varepsilon\frac{(1+m)}{(1+M)^{4}}e^{-(1+M)t/\varepsilon}\left[1-e^{-(1+M)t /\varepsilon}\right]+O(\varepsilon).\end{array}\right. \tag{44}\]
We plot in [Fig. 5] our corresponding results for the two considered sets of ICVs, as usual in comparison with the numerical solutions of the original problem (the same curves as in [Fig. 1]).
The figure shows that, as expected, these results are more refined than the PE 0th order UA ([Fig. 2A], [Fig. 3A] and [Fig. 2C], [Fig. 3B] for the results for the \(a\) and \(b\) sets of ICVs, respectively], but not as much refined as the PE 1st order ones ([Fig. 4]).
In \(A)\) and \(C)\) we present the behaviour of the concentrations of the substrate \(s(t)\), whereas in \(B)\) and \(D)\) we present the ones of the complex \(c(t)\) for the \(a\) and \(b\) sets of ICVs given in, respectively. Hence, in \(A)\) and \(B)\) we are in the case with \(\varepsilon=\varepsilon^{a}=0.1\), whereas in \(C)\) and \(D)\) we are in the one with \(\varepsilon=\varepsilon^{b}=0.5\). We plot both the numerical solutions of Eqs. already shown in the previous figures, and the analytical solutions computed from the SPDERG 1st order UAs (with a standard numerical approximation for the Lambert function), as given in. We plot moreover the (physically meaningless) limits for \(t\to\infty\) of the analytical solutions: \(s_{0}^{\otimes}\hat{s}_{1,\infty}^{rg,u}\simeq-0.622\), \(e_{0}^{\otimes}\hat{c}_{1,\infty}^{rg,u}\simeq-0.0122\), and \(s_{0}^{\otimes}\hat{s}_{1,\infty}^{rg,u}\simeq-3.11\), \(e_{0}^{\otimes}\hat{c}_{1,\infty}^{rg,u}\simeq-0.305\), respectively. We finally plot our corresponding rough evaluations of the two different time scales involved, too, with \(\tau_{s}\) describing the substrate decay time and \(\tau_{c}\) the complex saturation time. Notice that the time is in logarithmic scale.
Actually, as expected, both the substrate and the complex approach (physically meaningless) \(O(\varepsilon)\) values for \(t\to\infty\). In fact, from, one has \(\lim_{t\to\infty}\tilde{s}_{1}^{rg,u}(t)=\tilde{s}_{1,\infty}^{rg,u}=- \varepsilon(1+m)/(1+M)^{2}\) and \(\lim_{t\to\infty}\tilde{c}_{1}^{rg,u}(t)=\tilde{c}_{1,\infty}^{rg,u}=- \varepsilon M(1+2m-M)/(1+M)^{4}\) (the corresponding asymptotic constants for the two considered sets of ICVs are also plotted in the figures). Both this unphysical outcome and the lack of the 1st order contribution to the outer solution are particularly evident in the result for the substrate in the case of the \(b\) set of ICVs, corresponding to the larger \(\varepsilon=\varepsilon^{b}=0.5\) ([Fig. 5C]).
On the other hand, noticeably, the figure definitely shows that the 1st order SPDERG approach works better than the 0th order PE method in capturing the qualitative features of the correct solutions. In detail, one can observe the presence of the three inflection points in the curves for the substrate both in [Fig. 5A] and in [Fig. 5C]. From the quantitative point of view, also the correct complex maximum values turn out to be better predicted than in the corresponding [Fig. 3], further confirming that the approach is particularly successful in the proximity of the matching region.
Therefore, on the basis of these initial results, that appear as a whole to support the SPDERG usefulness, both for further testing its correctness and for possibly obtaining better approximations to the correct solutions than the ones given by the PE 1st order UAs, we also consider the 2nd order in \(\varepsilon\).
## 6 Results and discussion: ii) Second order contribution
In order to simplify the calculations, here we only look at the case in which the complex ICVs, in \(\tau=\tau_{0}\), are fixed at \(\tilde{c}_{0}^{*}=\tilde{c}_{1}^{*}=\tilde{c}_{2}^{*}=0\) from the beginning, \(\forall\tau_{0}\). This could appear unrealistic, but it is not expected to influence the result, since we already checked that there is no need for renormalizing these bare constants at the 1st order, and the same thing should reasonably apply to the 2nd order, too.
From a different point of view, this choice is not expected to influence the results since at the end one is interested in taking as initial time \(\tau_{0}=0\), and at the 2nd order, as at the 1st one, the renormalized part of \(\tilde{c}^{rg}\) should be obtainable by simply using the renormalized ICV for the substrate (after appropriately redefining the corresponding bare quantities for removing the secular terms). We will anyway verify in the following that indeed one gets the same ODE to be obeyed by \(\tilde{s}_{1}^{rg*}\) both from the study of the substrate and from the study of the complex, as it can be once again better understood within the analogy with the matching in the PE, where there is freedom for fixing only one of the two conditions, whereas the other turns out to be consistently also satisfied.
We take moreover \(\tilde{s}_{2}^{*g}=0\) from the beginning, since we expect that it can be anyway neglected within this approach at the 2nd order, in the same way as \(\tilde{s}_{1}^{*}\) turned out to be negligible at the 1st order (notice that we should instead allow for \(\tilde{s}_{2}^{*}\neq 0\) if we were to use these solutions to calculate the 3rd order contribution, too).
Clearly, one looks for 2nd order solutions in the form:
\[\left\{\begin{array}{l}\tilde{s}^{rg}(\tau)=\tilde{s}_{0}^{rg}(\tau)+ \varepsilon\tilde{s}_{1}^{rg}(\tau)+\varepsilon^{2}\tilde{s}_{2}^{rg}(\tau)\\ \tilde{c}^{rg}(\tau)=\tilde{c}_{0}^{rg}(\tau)+\varepsilon\tilde{c}_{1}^{rg}( \tau)+\varepsilon^{2}\tilde{c}_{2}^{rg}(\tau).\end{array}\right. \tag{45}\]
Here, \(\tilde{s}_{0}^{rg}(\tau)\) and \(\tilde{c}_{0}^{rg}(\tau)\) are still given by, with \(\tilde{c}_{0}^{*}=0\) in the case of the complex. Moreover, \(\tilde{s}_{1}^{rg}(\tau)\) and \(\tilde{c}_{1}^{rg}(\tau)\) are still given by and, respectively, with \(\tilde{c}_{1}^{*}=0\) in the case of the complex, whereas the condition \(\tilde{c}_{0}^{*}=0\) does also apply to the solution for the substrate.
Correspondingly, we end up to study the system:
\[\left\{\begin{array}{lcl}\tilde{s}_{2}^{rg}(\tau)&=&(\tilde{s}_{0}^{*}+m) \tilde{c}_{1}^{rg}(\tau)+\left[\tilde{c}_{0}^{rg}(\tau)-1\right]\tilde{s}_{1}^ {rg}(\tau)\\ \tilde{c}_{2}^{rg}(\tau)&=&-(\tilde{s}_{0}^{*}+M)\tilde{c}_{2}^{rg}(\tau)- \left[\tilde{c}_{0}^{rg}(\tau)-1\right]\tilde{s}_{2}^{rg}(\tau)-\tilde{s}_{1} ^{rg}(\tau)\tilde{c}_{1}^{rg}(\tau),\end{array}\right. \tag{46}\]
The obtained formulas are anyway cumbersome, as one could expect: in fact, they contain both (implicitly) the 1st order contributions to the outer solutions and (explicitly) the 2nd order contributions to the inner ones. As anticipated, these last 2nd order contributions are calculated in the present work for the first time to our knowledge.
By carrying on the study as in the previous section, we start from the part of the substrate solution to be renormalized. Interestingly, beyond the 2nd order contribution that originates from the first three terms in the solution for \(\tilde{s}_{2}^{rg}(\tau)\) given by (that are the secular ones, whose coefficients are reported in), the correct function needs also to contain the constant term that appears in \(s_{1}^{rg}(\tau)\), since this is now a leading order term. In fact, it depends on one of the bare constants \((\tilde{s}_{0}^{*})\), and it does not tend to zero for \((\tau-\tau_{0})\rightarrow\infty\).
Hence, one finds:
\[\tilde{s}_{div}^{rg}(\tau) = \tilde{s}_{0}^{*}+\varepsilon\tilde{s}_{1}^{*}-\varepsilon\frac{ \tilde{s}_{0}^{*}+m}{(\tilde{s}_{0}^{*}+M)^{2}}\tilde{s}_{0}^{*}-\varepsilon \frac{(M-m)\tilde{s}_{0}^{*}}{\tilde{s}_{0}^{*}+M}(\tau-\tau_{0})+ \tag{47}\] \[+ \varepsilon^{2}\frac{2M(M-m)\tilde{s}_{0}^{*}}{(\tilde{s}_{0}^{*} +M)^{4}}\left(\tilde{s}_{0}^{*}+m\right)(\tau-\tau_{0})-\varepsilon^{2}\frac{M (M-m)\tilde{s}_{1}^{*}}{(\tilde{s}_{0}^{*}+M)^{2}}(\tau-\tau_{0})+\] \[+ \varepsilon^{2}\frac{M(M-m)^{2}\tilde{s}_{0}^{*}}{2\left(\tilde{s} _{0}^{*}+M\right)^{3}}(\tau-\tau_{0})^{2}.\]
We then replace the bare constants with the renormalized ones, _i.e._, \(\tilde{s}_{0}^{*}=Z_{s_{0}}^{s}(\tau_{0},\lambda)\tilde{s}_{0}^{rg*}(\lambda)\) and \(\tilde{s}_{1}^{*}=Z_{s_{1}}^{s}(\tau_{0},\lambda)\tilde{s}_{1}^{rg*}(\lambda)\), with, at the 2nd order in \(\varepsilon\):
\[\left\{\begin{array}{rcl}Z_{s_{0}}(\tau_{0},\lambda)&=&1+ \varepsilon z_{s_{0},1}(\tau_{0},\lambda)+\varepsilon^{2}z_{s_{0},2}(\tau_{0}, \lambda),\\ Z_{s_{1}}(\tau_{0},\lambda)&=&1+\varepsilon z_{s_{1},1}(\tau_{0},\lambda).\end{array}\right. \tag{48}\]
Correspondingly, we write \((\tau-\tau_{0})=(\tau-\lambda)+(\lambda-\tau_{0})\) in and we absorb the secular terms in \((\lambda-\tau_{0})\) by appropriately choosing the renormalization constant coefficients. In detail, \(z_{s_{0},1}(\tau_{0},\lambda)\) is already fixed by. Nevertheless, the contributions proportional to this coefficient to the third term and to the fourth one in are to be taken carefully into account, since they turn out to be \(O(\varepsilon^{2})\).
In particular, in the case of the third term, when renormalizing \(\tilde{s}_{0}^{*}\), one has:
\[-\varepsilon\frac{(\tilde{s}_{0}^{*}+m)\tilde{s}_{0}^{*}}{(\tilde{s}_{0}^{*}+ M)^{2}}=-\varepsilon\frac{(\tilde{s}_{0}^{rg*}+m)\tilde{s}_{0}^{rg*}}{( \tilde{s}_{0}^{rg*}+M)^{2}}-\varepsilon^{2}\left\{\frac{d}{d\tilde{s}_{0}^{*} }\left[\frac{(\tilde{s}_{0}^{*}+m)\tilde{s}_{0}^{*}}{(\tilde{s}_{0}^{*}+M)^{2 }}\right]\left|{}_{\tilde{s}_{0}^{*}-\tilde{s}_{0}^{rg*}}\right.\right\}z_{s_ {0},1}(\tau_{0},\lambda)\tilde{s}_{0}^{rg*}, \tag{49}\]
with:
\[-\varepsilon^{2}\left\{\frac{d}{d\tilde{s}_{0}^{*}}\left[\frac{( \tilde{s}_{0}^{*}+m)\tilde{s}_{0}^{*}}{(\tilde{s}_{0}^{*}+M)^{2}}\right]\left| {}_{\tilde{s}_{0}^{*}-\tilde{s}_{0}^{rg*}}\right.\right\}z_{s_{0},1}(\tau_{0}, \lambda)s_{0}^{rg*}= \tag{50}\] \[= -\varepsilon^{2}\left[\frac{(M-m)\tilde{s}_{0}^{rg*}}{(\tilde{s} _{0}^{rg*}+M)^{3}}+\frac{M(\tilde{s}_{0}^{rg*}+m)}{(\tilde{s}_{0}^{rg*}+M)^{3 }}\right]z_{s_{0},1}(\tau_{0},\lambda)\tilde{s}_{0}^{rg*}=\] \[= -\varepsilon^{2}\left[\frac{(M-m)^{2}(\tilde{s}_{0}^{rg*})^{2}}{ (\tilde{s}_{0}^{rg*}+M)^{4}}+\frac{M(M-m)\tilde{s}_{0}^{rg*}}{(\tilde{s}_{0}^ {rg*}+M)^{4}}(\tilde{s}_{0}^{rg*}+m)\right](\lambda-\tau_{0}).\]
Clearly, both of these two terms do in fact contribute, and they have to be taken into account in the definition of \(z_{s_{0},2}(\tau_{0},\lambda)\). Let us note moreover that \(d[\tilde{s}_{0}^{*}/(\tilde{s}_{0}^{*}+M)]/d\tilde{s}_{0}^{*}=M/(\tilde{s}_{0}^ {*}+M)^{2}\). Thus, one can check that the 2nd order contribution proportional to the same coefficient, \(z_{s_{0},1}(\tau_{0},\lambda)\) (that originates from the fourth term in \(\tilde{s}_{div}^{rg}\), in the same way as we just explained in detail with the formulas and), partially absorbs the term proportional to \((\tau-\lambda)(\lambda-\tau_{0})\) (that originates from the last term), and partially contributes to \(z_{s_{0},2}(\tau_{0},\lambda)\).
Thus, by taking:
\[\left\{\begin{array}{rcl}z_{s_{0},2}(\tau_{0},\lambda)&=& \left[\frac{(M-m)^{2}\tilde{s}_{0}^{rg*}}{(\tilde{s}_{0}^{rg*}+M)^{4}}-\frac{ M(M-m)}{(\tilde{s}_{0}^{rg*}+M)^{4}}(\tilde{s}_{0}^{rg*}+m)\right](\lambda- \tau_{0})+\\ &+&\frac{M(M-m)^{2}}{2(\tilde{s}_{0}^{rg*}+M)^{3}}(\lambda-\tau_{0})^{2}\\ z_{s_{1},1}(\tau_{0},\lambda)&=&\frac{M(M-m)}{(\tilde{s}_{0}^{rg*}+M)^{2}}( \lambda-\tau_{0}),\end{array}\right. \tag{51}\]
Noticeably, moreover, though we wrote explicitly the first two terms in \(z_{s_{0},2}\) to make evident the different origins of the contributions, their sum can be clearly simplified in such a way that the coefficient of the term proportional to \((\lambda-\tau_{0})\) is simply equal to \(-m(M-m)/(\tilde{s}_{0}^{rg*}+M)^{3}\).
Therefore, we continue our analysis by studying directly the derivative with respect to \(\lambda\) of \(\hat{s}^{rg}_{div}(\tau,\lambda)\), and we obtain, at the 2nd order:
\[\frac{d\hat{s}^{rg}_{div}(\tau,\lambda)}{d\lambda} = \frac{d\hat{s}^{rg*}_{0}(\lambda)}{d\lambda}+\varepsilon\frac{d \hat{s}^{rg*}_{1}(\lambda)}{d\lambda}+\varepsilon\frac{(M-m)\hat{s}^{rg*}_{0} (\lambda)}{\hat{s}^{rg*}_{0}(\lambda)+M}+\varepsilon^{2}\frac{(M-m)^{2}\left[ \hat{s}^{rg*}_{0}(\lambda)\right]^{2}}{\left[\hat{s}^{rg*}_{0}(\lambda)+M \right]^{4}}+ \tag{52}\] \[+ \varepsilon^{2}\frac{M(M-m)\hat{s}^{rg*}_{0}(\lambda)}{\left[\hat{ s}^{rg*}_{0}(\lambda)+M\right]^{4}}\left[\hat{s}^{rg*}_{0}(\lambda)+m\right]+ \varepsilon^{2}\frac{M(M-m)\hat{s}^{rg*}_{0}(\lambda)}{\left[\hat{s}^{rg*}_{0} (\lambda)+M\right]^{3}}(\tau-\lambda)+\] \[- \varepsilon^{2}\frac{2M(M-m)\hat{s}^{rg*}_{0}(\lambda)}{\left[\hat {s}^{rg*}_{0}(\lambda)+M\right]^{4}}\left[\hat{s}^{rg*}_{0}(\lambda)+m\right] +\varepsilon^{2}\frac{M(M-m)\hat{s}^{rg*}_{1}(\lambda)}{\left[\hat{s}^{rg*}_{ 0}(\lambda)+M\right]^{2}}+\] \[- \varepsilon^{2}\frac{M(M-m)^{2}\hat{s}^{rg*}_{0}(\lambda)}{\left[ \hat{s}^{rg*}_{0}(\lambda)+M\right]^{3}}(\tau-\lambda).\]
In fact, one already makes use of the known 1st order result on \(d\hat{s}^{rg*}_{0}/d\lambda\), given in Eqs., in the derivation of this equation.
Here, for the sake of clarity, we also wrote explicitly both of the 2nd order terms proportional to \((\tau-\lambda)\), that obviously cancel each other. In fact, this appears a quite consistent result of the present approach, since they have completely different origins. Indeed, the first of these terms originates from part of the derivative with respect to \(\lambda\) of the 1st order term proportional to \((\tau-\lambda)\) in \(\hat{s}^{rg}_{div}(\tau,\lambda)\) (_i.e._, from the term that is equal to \(-\varepsilon(M-m)\{d[\hat{s}^{rg*}_{0}/(\hat{s}^{rg*}_{0}+M)]/d\lambda\}(\tau -\lambda)\)). The second of these terms originates instead from the derivative with respect to \(\lambda\) of the last term in \(\hat{s}^{rg}_{div}(\tau,\lambda)\), that is proportional to \((\tau-\lambda)^{2}\).
Moreover, interestingly, the second term proportional to \(\varepsilon^{2}\) (part of the contribution that originates from the constant term in \(\hat{s}^{rg}_{1}(\tau)\)), letting aside a factor 2, is the same as the one that is found when deriving the first of the terms proportional to \(\varepsilon^{2}(\tau-\lambda)\) in with respect to \(\lambda\) (_i.e._, the term in \(\hat{s}^{rg}_{div}(\tau,\lambda)\) that corresponds to the fourth term proportional to \(\varepsilon^{2}\) here), but that has opposite sign.
In fact, both of these results are obtained in a similar manner to the one that we described in detail previously, in the context of the derivation of the appropriate renormalization constants (from Eqs., one has \((\lambda-\tau_{0})d\hat{s}^{rg*}_{0}/d\lambda=-\varepsilon z^{*}_{s_{0},1}( \tau_{0},\lambda)\hat{s}^{rg*}_{0}\)).
Finally, by imposing the scaling condition \(d\hat{s}^{rg}_{div}(\tau,\lambda)/d\lambda=0\), and by grouping the terms in \(\varepsilon\) and in \(\varepsilon^{2}\), we derive the two ODEs to be obeyed by \(\hat{s}^{rg*}_{0}(\lambda)\) and \(\hat{s}^{rg*}_{1}(\lambda)\), respectively:
\[\left\{\begin{array}{lcl}\frac{d\hat{s}^{rg*}_{0}(\lambda)}{ d\lambda}&=&-\varepsilon\frac{(M-m)\hat{s}^{rg*}_{0}(\lambda)}{\hat{s}^{rg*}_{0}( \lambda)+M}\\ \frac{d\hat{s}^{rg*}_{1}(\lambda)}{d\lambda}&=&\varepsilon\left\{-\frac{(M-m )^{2}\left[\hat{s}^{rg*}_{0}(\lambda)\right]^{2}}{\left[\hat{s}^{rg*}_{0}( \lambda)+M\right]^{4}}+\frac{M(M-m)\hat{s}^{rg*}_{0}(\lambda)}{\left[\hat{s}^{ rg*}_{0}(\lambda)+M\right]^{4}}\left[\hat{s}^{rg*}_{0}(\lambda)+m\right]+\right.\\ &&\left.-&\frac{M(M-m)\hat{s}^{rg*}_{1}(\lambda)}{\left[\hat{s}^{rg*}_{0}( \lambda)+M\right]^{2}}\right\}=\varepsilon\left\{\frac{m(M-m)\hat{s}^{rg*}_{ 0}(\lambda)}{\left[\hat{s}^{rg*}_{0}(\lambda)+M\right]^{3}}-\frac{M(M-m)\hat{ s}^{rg*}_{1}(\lambda)}{\left[\hat{s}^{rg*}_{0}(\lambda)+M\right]^{2}}\right\}. \end{array}\right. \tag{53}\]
The first of these ODEs is just the already known 1st order result for \(d\hat{s}^{rg*}_{0}(\lambda)/d\lambda\) reported in Eqs.. Actually, at the 2nd order, the interesting ODE is the one to be obeyed by \(\hat{s}^{rg*}_{1}(\lambda)\).
Let us remark that, when making the final transformation \(\lambda\rightarrow\tau=t/\varepsilon\), and when recalling that \(\hat{s}^{rg*}_{0}(t)=\hat{s}^{out}_{0}(t)\) (given in), the ODE to be obeyed by \(\hat{s}^{rg*}_{1}(t)\) turns out to be different from the one for the 1st order outer substrate within the PE method reported in Eqs.. Here, we wrote the equation both with and without the simplification due to the sum of the first two terms just in order to make evident the difference, since it is the first term in the not simplified expression that was absent there. Indeed, to get in particular the same coefficient as in the ODE in Eqs. for the second term in the present not simplified expression, it is essential to correctly take into account the constant in \(\hat{s}^{rg}_{1}(\tau)\) in the part to be renormalized of the function (at the 2nd order). In fact, as outlined, the contribution of the only term proportional to \(\varepsilon^{2}\) would give an incorrect (twice larger) coefficient.
Interestingly, the present ODE turns out to be simpler than the one encountered in the PE method. One can check that the solution is given by:
\[\hat{s}^{rg*}_{1}(t)=\frac{m\hat{s}^{out}_{0}(t)}{M\left[\hat{s}^{out}_{0}(t)+ M\right]}\log\left[\frac{\hat{s}^{out}_{0}(t)+M}{(1+M)\hat{s}^{out}_{0}(t)} \right], \tag{54}\]with the choice \(\tilde{s}_{1}^{rg*}=0\), that is reasonable in this context, since it allows to get correctly \(\tilde{s}_{2}^{rg,u}=1\). It is also important to stress that the solution also satisfies the asymptotic condition \(\lim_{t\to\infty}\tilde{s}_{1}^{rg*}(t)=0\), though this in fact implies \(\lim_{t\to\infty}\tilde{s}_{2}^{rg,u}(t)=O(\varepsilon^{2})\).
On the other hand, consistently, when adding the 1st order term that originates from the replacement \(\tilde{s}_{0}^{*}\to\tilde{s}_{0}^{rg,*}(\lambda)\) in the 1st order constant term in the part to be renormalized of the function, the total 1st order _outer_ contribution to the SPDERG 2nd order UA to the correct solution for the substrate is just:
\[\tilde{s}_{1}^{rg,out}(t)=-\frac{[\tilde{s}_{0}^{out}(t)+m]\tilde{s}_{0}^{out} (t)}{[\tilde{s}_{0}^{out}(t)+M]^{2}}+\tilde{s}_{1}^{rg*}(t)=\tilde{s}_{1}^{ out}(t). \tag{55}\]
Thus, one recovers the 1st order outer contribution to the PE UA reported in. As we will discuss more in detail in the following, this was not a result to be taken for granted. Obviously, the result implies that the total 1st order outer contribution satisfies the MC, too (_i.e._, in particular that \(\tilde{s}_{1}^{rg,out}=-(1+m)/(1+M)^{2}\)).
Hence, we recall that \({\cal R}_{2}^{s}(\tau)\) is to be evaluated from the remaining part of the solution in \(\tilde{s}_{0}^{*}=1\), \(\tilde{s}_{1}^{*}=0\) and \(\tau_{0}=0\). It is given by (see and, in Appendix B):
\[{\cal R}_{2}^{s}(\tau) = D_{s_{2}^{rg}}+\left[F_{s_{2}^{rg}}+H_{s_{2}^{rg}}\tau+ J_{s_{2}^{rg}}\tau^{2}\right]e^{-(1+M)\tau}+K_{s_{2}^{rg}}e^{-2}(1+M)\tau= \tag{56}\] \[= -\frac{1}{2(1+M)^{5}}[2M^{2}(2m+1)-M(6m^{2}+5m+3)-m^{2}+m]+\] \[+ \frac{1}{(1+M)^{5}}[M^{2}(2m+1)-M(3m^{2}+3m+2)-m^{2}-m-1]e^{-(1+M )\tau}+\] \[- \left\{\frac{1}{(1+M)^{4}}[M^{2}+M(m^{2}+1)-2m-1]\tau-\frac{(M-m )}{2(1+M)^{3}}(m+1)\tau^{2}\right\}e^{-(1+M)\tau}+\] \[+ \frac{(m+1)}{2(1+M)^{5}}(M+m+2)e^{-2}(1+M)\tau,\]
Correspondingly, we get the complete result for the SPDERG 2nd order UA to the adimensional substrate solution:
\[\tilde{s}_{2}^{rg,u}(t)=\tilde{s}_{1}^{rg,u}(t)+\varepsilon\left\{-\frac{ \tilde{s}_{0}^{out}(t)+m}{[\tilde{s}_{0}^{out}(t)+M]^{2}}\tilde{s}_{0}^{out}(t )+\frac{1+m}{(1+M)^{2}}+\tilde{s}_{1}^{rg*}(t)\right\}+\varepsilon^{2}{\cal R }_{2}^{s}(t/\varepsilon)+O(\varepsilon^{2}), \tag{57}\]
We wrote the solution in the present form to make more evident the analogies and the differences with the 1st order PE result. In particular, the constant term that is equal to \(-(1+m)/(1+M)^{2}\) originates here from the constant term depending on \(\tilde{s}_{0}^{*}\) that appeared (calculated in \(\tilde{s}_{0}^{*}=1\)) in \(\tilde{s}_{1}^{rg,u}(t)\), the one that has been now included in \(\tilde{s}_{d_{2}^{rg}}^{rg}\) at the 2nd order. Moreover, one can notice that the result implies, as expected, \(\lim_{t\to\infty}\tilde{s}_{2}^{rg,u}(t)=\tilde{s}_{2,\infty}^{rg}=\varepsilon ^{2}D_{s_{2}^{rg}}=O(\varepsilon^{2})\), where \(D_{s_{2}^{rg}}\) is the first of the terms in, reported in Appendix B, too.
For the complex, the 2nd order terms to be added to the part of the solution to be renormalized are the three ones that appear in \(\tilde{c}_{2,div}^{rg}(\tau)\), that is reported in, in Appendix A. Moreover, we have again to take into account also the constant term in \(\tilde{c}_{1}^{rg}(\tau)\), _i.e._, the one corresponding to the coefficient \(B_{\tilde{c}_{1}^{rg}}(\tilde{s}_{0}^{*},\tilde{s}_{1}^{*},\tilde{c}_{0}^{*})\) that is reported in, in Appendix A, and that needs to be calculated in \(\tilde{c}_{0}^{*}=0\). We obtain:
\[\tilde{c}_{div}^{rg}(\tau) = \frac{\tilde{s}_{0}^{*}}{\tilde{s}_{0}^{*}+M}-\varepsilon\frac{M (\tilde{s}_{0}^{*}-M+2m)}{(\tilde{s}_{0}^{*}+M)^{4}}\tilde{s}_{0}^{*}+ \varepsilon\frac{M\tilde{s}_{1}^{*}}{(\tilde{s}_{0}^{*}+M)^{2}}-\varepsilon \frac{M(M-m)\tilde{s}_{0}^{*}}{(\tilde{s}_{0}^{*}+M)^{3}}(\tau-\tau_{0})+ \tag{58}\] \[+ \varepsilon^{2}A_{C_{2,div}^{rg}}(\tilde{s}_{0}^{*})(\tau-\tau_{0 })+\varepsilon^{2}B_{\tilde{c}_{2,div}^{rg}}(\tilde{s}_{0}^{*},\tilde{s}_{1}^{ *})(\tau-\tau_{0})+\varepsilon^{2}C_{\tilde{c}_{2,div}^{rg}}(\tilde{s}_{0}^{* })(\tau-\tau_{0})^{2},\]
Though the proof involves more demanding calculations in the present case, one can check that here, in the same way as we showed in detail in the case of the substrate, with the same renormalization coefficients of the ICVs (\(z_{s_{0},1}\) already fixed in, and \(z_{s_{0},2}\), \(z_{s_{1},1}\) already fixed in, respectively), one obtains a function \(\tilde{c}_{div}^{rg}(\tau,\lambda)\) exactly corresponding to, with \(\tilde{s}_{0}^{*}\to\tilde{s}_{0}^{rg*}(\lambda)\), \(\tilde{s}_{1}^{*}\to\tilde{s}_{1}^{rg*}(\lambda)\) and \(\tau_{0}\to\lambda\). This was in fact the expected result from the point of view of the analogy with the matching in the PE.
Moreover, as furthermore expected within the same context, we verify that, when imposing the scaling condition \(\hat{c}^{rg}_{div}(\tau,\lambda)/d\lambda=0\), one recovers once again the two ODEs in, to be obeyed by \(\tilde{s}_{0}^{rg*}(\lambda)\) and \(\tilde{s}_{1}^{rg*}(\lambda)\). Thus, one obtains in particular the same result on \(\tilde{s}_{1}^{rg*}(\lambda)\).
In detail, also in the case of the complex, one needs to use the known 1st order result on \(d\tilde{s}_{0}^{rg*}(\lambda)/d\lambda\) in the derivation of the 2nd order one on \(d\tilde{s}_{1}^{rg*}(\lambda)/d\lambda\). Finally, also in this case the result on the derivative with respect to \(\lambda\) is obtained in a very similar way to the proof of the correspondence between \(\hat{c}^{rg}_{div}(\tau,\lambda)\) and the original \(\hat{c}^{rg}_{div}(\tau)\) given by.
Clearly, in the present case, there is a definitely larger number of relevant terms that anyway either cancel each other or contribute in the correct way to the final result. Therefore, the verification of the expectations we made on the basis of the recalled analogy, that is reported in Appendix C, appears to give further consistency to the whole approach.
Correspondingly, we obtain a 1st order _outer_ contribution to the SPDERG 2nd order UA to the correct solutions for the adimensional complex concentration that turns out to be exactly equal to the 1st order PE outer contribution given in. It is obtained here from an algebraic relation that could appear different from the one in Eqs., but that is in fact equivalent. This becomes evident when writing \(\tilde{s}_{1}^{rg*}(t)\) in terms of \(\tilde{s}_{0}^{out}(t)\) and \(\tilde{s}_{1}^{out}(t)\) by means of:
\[\tilde{c}_{1}^{rg,out}(t)=-\frac{M(\tilde{s}_{0}^{rg*}(t)-M+2m)}{\left[\tilde{ s}_{0}^{rg*}(t)+M\right]^{4}}\tilde{s}_{0}^{rg*}(t)+\frac{M\tilde{s}_{1}^{rg*}(t)} {\left[\tilde{s}_{0}^{rg*}(t)+M\right]^{2}}=\tilde{c}_{1}^{out}(t). \tag{59}\]
The equality can also be easily checked by reminding that \(\tilde{s}_{0}^{rg*}(t)=\tilde{s}_{0}^{out}(t)\), and by using the known result for \(\tilde{s}_{1}^{rg*}(t)\) reported in.
The SPDERG 2nd order UA to the correct solution for the complex is then obtainable by taking into account also the remaining part of the 2nd order inner solution, and is given by:
\[\tilde{c}_{2}^{rg,u}(t) = \tilde{c}_{1}^{rg,u}(t)+\varepsilon\left\{-\frac{M(\tilde{s}_{0} ^{rg*}(t)-M+2m)}{\left[\tilde{s}_{0}^{rg*}(t)+M\right]^{4}}\tilde{s}_{0}^{rg*} (t)+\frac{M\tilde{s}_{1}^{rg*}(t)}{\left[\tilde{s}_{0}^{rg*}(t)+M\right]^{2}} +\frac{M(1-M+2m)}{(1+M)^{4}}\right\}+ \tag{60}\] \[+ \varepsilon^{2}\mathcal{R}_{2}^{c}(t/\varepsilon)+O(\varepsilon^ {2}).\]
Here, we outline once again similarities and differences with the PE result. In fact, because of, one can equivalently write the terms in curly brackets as \(\tilde{c}_{1}^{out}(t)\) minus the constant terms that in the standard method appear twice.
In detail, the various terms that appear in are reported: \(\tilde{c}_{1}^{rg,u}(t)\) in; \(\tilde{s}_{0}^{rg*}(t)=\tilde{s}_{0}^{out}(t)\) in; \(\tilde{s}_{1}^{rg*}(t)\) in; and \(\mathcal{R}_{2}^{c}(\tau)\), correctly evaluated in \(\tau_{0}=0\), \(\tilde{s}_{0}^{*}=1\) and \(\tilde{s}_{1}^{*}=0\) in, with the coefficients given in, in Appendix B.
This UA to the correct solution verifies the ICV \(\tilde{c}_{2}^{rg,u}=0\), whereas one finds \(\lim_{t\to\infty}\tilde{c}_{2}^{rg,u}(t)=\tilde{c}_{2,\infty}^{rg,u}=\varepsilon ^{2}A_{\mathcal{R}_{2}^{c}}=O(\varepsilon^{2})\), with \(A_{\mathcal{R}_{2}^{c}}\) the constant term in \(\mathcal{R}_{2}^{c}(\tau)\), whose detailed dependence on \(\tilde{s}_{0}^{*}\) and \(\tilde{s}_{1}^{*}\) is reported in, and that is calculated in \(\tilde{s}_{0}^{*}=1\), \(\tilde{s}_{1}^{*}=0\) in, in Appendix B.
We plot in [Fig. 6] our numerical results on the SPDERG 2nd order UAs to the solutions for the substrate and the complex, respectively, for the two considered sets of ICVs. The plots are as usual in comparison with the numerical solutions of the original problem (the same curves as in [Fig. 1]).
As it is in fact explicable on the basis that they contain the 2nd order terms of the inner solutions, the approximations work better than the 1st order PE ones in a region that encompasses the matching one. Indeed, the results here are nearly indistinguishable, within our numerical precision, from the correct ones on a definitely larger time window. This is true also in the particularly demanding case of the substrate in [Fig. 6C], and the outcome is clearly different from the one observed in the same case at the 1st order, when applying the standard method, that is reported in [Fig. 4C].
Nevertheless, one can still note a minor discrepancy at large times, that is at least partially to be related to the failure of the approximations in reproducing the physically correct asymptotically vanishing solutions. Actually, on the basis of the results that we already obtained in the present study, this failure appears easily correctable in a reasonable way. In the following, we just consider the SPDERG 2nd order UAs that can be proposed that satisfy the asymptotic conditions \(\lim_{t\to\infty}s(t)=\lim_{t\to\infty}c(t)=0\), too.
## 7 Results and discussion: iii) Second order with proposed refinement
Within the framework of the already obtained results, by assuming that the found solution behaviours could be iterated, one can hypothesize that, in the SPDERG approach (at least in its present kind of application to a boundary layer problem, in which we directly renormalized the bare initial constant values), the constant term at a given order will contribute to the outer component of the solution at the following order. Actually, this appears to us the SPDERG approach ingredient that is equivalent to take into account both the MCs and the removing of one of the constants that appear twice in the standard PE method, though the observation needs to be better formalized for investigating its possible generalizations.
In fact, in the present case, when passing from the 2nd to the 3rd order, these iteratively expected solution behaviours should be obtainable by means of the substitutions \(\tilde{s}_{0}^{*}\rightarrow\tilde{s}_{0}^{\prime\prime*}(t)\) and \(\tilde{s}_{1}^{*}\rightarrow\tilde{s}_{1}^{\prime\prime*}(t)\) in the constant terms \(D_{s_{2}^{\prime\prime}}(\tilde{s}_{0}^{*})+E_{s_{2}^{\prime\prime}}(\tilde{s} _{0}^{*},\tilde{s}_{1}^{*})\) and \(A_{\mathcal{R}_{2}^{*}}(\tilde{s}_{0}^{*},\tilde{s}_{1}^{*})\). These terms appear in the 2nd order inner solution for the substrate and the complex, respectively, that are given in and, in Appendix B. We notice that, despite of the substitutions, they will remain terms of 2nd order, in agreement with the general consideration that, at the \(n\)-th order, within the SPDERG approach, one in fact obtains the \((n-1)\)-th order outer components of the corresponding PE UAs, to be interpreted as their leading order terms.
Therefore, we can finally consider as refined SPDERG 2nd order UAs to the correct solutions the functions \(\tilde{s}_{2,r}^{\sigma,u}(t)\) and \(\tilde{c}_{2,r}^{\sigma,u}(t)\), that satisfy by construction the physically meaningful asymptotic
In \(A)\) and \(C)\) we present the behaviour of the concentrations of the substrate \(s(t)\), whereas in \(B)\) and \(D)\) we present the ones of the complex \(c(t)\) for the \(a\) and \(b\) sets of ICVs given in, respectively. Hence, in \(A)\) and \(B)\) we are in the case with \(\varepsilon=\varepsilon^{a}=0.1\), whereas in \(C)\) and \(D)\) we are in the one with \(\varepsilon=\varepsilon^{b}=0.5\). We plot both the numerical solutions of Eqs. already shown in the previous figures, and the analytical solutions computed from the SPDERG 2nd order UAs (with a standard numerical approximation for the Lambert function), as given in and, respectively. We plot moreover the (physically meaningless) asymptotic limits of these analytical solutions: \(s_{0}^{\prime}\tilde{s}_{2,\infty}^{\sigma,u}\simeq 0.0012\), \(\varepsilon_{0}^{a}\tilde{c}_{2,\infty}^{\sigma,u}\simeq-0.00023\), and \(s_{0}^{b}\tilde{s}_{2,\infty}^{\sigma,u}\simeq 0.3\), \(s_{0}^{b}\tilde{c}_{2,\infty}^{\sigma,u}\simeq-0.029\), for the \(a\) and \(b\) ICV sets, respectively. We finally plot our corresponding rough evaluations of the two different time scales involved, too, with \(\tau_{s}\) describing the substrate decay time and \(\tau_{c}\) the complex saturation time. Notice that the time is in logarithmic scale.
conditions \(\lim_{t\to\infty}\hat{s}_{2,r}^{rg,u}(t)=\lim_{t\to\infty}\hat{c}_{2,r}^{rg,u}(t)=0\), given by:
\[\left\{\begin{array}{rcl}\hat{s}_{2,r}^{rg,u}(t)&=&\hat{s}_{2}^{rg,u}(t)+ \varepsilon^{2}\left\{D_{s_{2}^{rg}}[\hat{s}_{0}^{rg*}(t)]+E_{s_{2}^{rg}}[\hat{ s}_{0}^{rg*}(t),\hat{s}_{1}^{rg*}(t)]-D_{s_{2}^{rg}}\right\}+O(\varepsilon^{2}) \\ \hat{c}_{2,r}^{rg,u}(t)&=&\hat{c}_{2}^{rg,u}(t)+\varepsilon^{2}\left\{A_{ \mathcal{R}_{2}^{c}}[\hat{s}_{0}^{rg*}(t),\hat{s}_{1}^{rg*}(t)]-A_{\mathcal{R }_{2}^{c}}\right\}+O(\varepsilon^{2}),\end{array}\right. \tag{61}\]
We remind that \(E_{s_{2}^{rg}}=0\), whereas \(\hat{s}_{2}^{rg,u}(t)\) and \(\hat{c}_{2}^{rg,u}(t)\) are reported in and, respectively, \(\hat{s}_{0}^{rg*}(t)=\hat{s}_{0}^{out}(t)\) in, and \(\hat{s}_{1}^{rg*}(t)\) in.
It is not to be taken for granted that these approximations could work better than the previously considered ones, since they anyway lack of a part of the 2nd order outer contribution. On the other hand, these appear to us the most refined SPDERG 2nd order UAs that one can propose by exploiting as much as possible the obtained results.
We present the corresponding numerical solutions for the two different considered ICV sets in [Fig. 7]. The plots are as usual in comparison with the numerical solutions of the original problem (the same curves as in [Fig. 1]).
In fact, one could already observe in the previous [Fig. 6A], [Fig. 6B] that the SPDERG 2nd order UAs appeared indistinguishable from the correct solutions in the case of the \(a\) set of ICVs. Indeed, this set corresponds to the relatively small \(\varepsilon=\varepsilon^{a}=0.1\), and the PE 1st order UAs appeared indistinguishable from the correct solutions for this ICV set ([Fig. 4A], [Fig. 4B]), too. In this case, we limit ourselves to underline that the present proposed 2nd order approximations, that are presented in [Fig. 7A] for the substrate concentration and in [Fig. 7B] for the complex one, are moreover correctly asymptotically vanishing. In fact, as previously discussed, in the not refined 2nd order approximations the limits for \(t\to\infty\) were not zero, though they gave a practically negligible contribution to the plotted curves.
In \(A)\) and \(C)\) we present the behaviour of the concentrations of the substrate \(s(t)\), whereas in \(B)\) and \(D)\) we present the ones of the complex \(c(t)\) for the \(a\) and \(b\) sets of ICVs given in, respectively. Hence, in \(A)\) and \(B)\) we are in the case with \(\varepsilon=\varepsilon^{a}=0.1\), whereas in \(C)\) and \(D)\) we are in the one with \(\varepsilon=\varepsilon^{b}=0.5\). We plot both the numerical solutions of Eqs. already shown in the previous figures, and the analytical solutions computed from the refined SPDERG 2nd order UAs (with a standard numerical approximation for the Lambert function), as given in. We finally plot our corresponding rough evaluations of the two different time scales involved, too, with \(\tau_{s}\) describing the substrate decay time and \(\tau_{c}\) the complex saturation time. Notice that the time is in logarithmic scale.
On the other hand, when looking at [Fig. 7C] and [Fig. 7D], that present the behaviour of the substrate and complex concentration, respectively, for the \(b\) ICV case, _i.e._, for the larger value of the expansion parameter \(\varepsilon=\varepsilon^{b}=0.5\), the plots appear not enough detailed for understanding up to which point the present approximated solutions are better than the ones without the refinement. Both for this reason and for roughly quantifying the differences between the PE 1st order UAs and the present SPDERG 2nd order results, we are led to a more careful study.
## 8 Results and discussion: iv) A conclusive comparison
We present in [Fig. 8] the detailed time depending behaviours of the substrate concentration, \(s(t)\), and of the complex one, \(c(t)\), for the more demanding ICV case with \(\varepsilon=\varepsilon^{b}=0.5\), in the two relevant parts of the time window, by comparing the different _best_ approximations we considered: i) the PE 1st order UAs (as given in, already presented in [Fig. 4C], [Fig. 4D]); ii) the SPDERG 2nd order UAs (as given in,, already presented in [Fig. 6C], [Fig. 6D]); iii) the refined SPDERG 2nd order UAs (as given in, already presented in [Fig. 7C], [Fig. 7D]).
Here, we neglect first of all the initial time interval, up to \(t=0.03s\) for the substrate and to \(t=0.08s\) for the complex, respectively. Indeed, in this interval, the different results are indistinguishable, within our numerical precision, both each other and with the correct numerical solutions. Instead, we consider in detail, as usual in logarithmic time scale, the central intervals, _i.e._, the ones that encompass the matching region. These intervals are \(t\in[0.03,3.5]s\) for the substrate and \(t\in[0.08,2.5]s\) for the complex, respectively. Finally, we present in non logarithmic time scale the relevant large time window, [Fig. 8B] and [Fig. 8D], for the substrate and the complex, respectively. In fact, this window ranges up to \(t\sim 12s\) for the substrate and up to \(t\sim 14s\) for the complex, since at larger times (as it is clear from [Fig. 8B] in particular) the solutions have already reached, within our numerical precision, their asymptotic values (that are different from zero in the case of the non refined SPDERG).
Interestingly, the curves shown in [Fig. 8A] not only make once more evident that, in the case of the substrate, the SPDERG 2nd order UAs we considered works definitely better than the PE 1st order ones, both with and without the refinement, but it also allows to naked-eye evaluate the range in which this happens. In fact, this range corresponds to \(t\sim 0.05\div 2s\), hence it covers about the 15% of the whole relevant time window in the case without refinement, whereas it corresponds to \(t\sim 0.05\div 3s\) (with the larger time being of the same order of \(\tau_{s}^{b}\)), and hence it covers about the 25% of the whole relevant time window, in the case with the refinement. On the other hand, for \(t\stackrel{{>}}{{\sim}}3s\) ([Fig. 8B]), the refined SPDERG 2nd order UA appears to tend to zero slightly too rapidly, with respect both to the correct solution and to the PE 1st order result, though it makes a smaller error than the same SPDERG approximation without the refinement.
In the case of the complex, as already anticipated, the curves shown in [Fig. 8C] make clear that the PE 1st order UA slightly over-evaluates the value of the maximum of this quantity. In fact, it appears to over-evaluate the dynamical behaviour of this quantity in the whole range \(t\sim 0.25\div 0.7s\), that is a time interval roughly centered around the maximum abscissa. On the other hand, both of the SPDERG 2nd order UAs that we considered are successful in correctly capturing the maximum value of the complex numerical solution. In particular, the refined SPDERG approximation turns out to be the one more in agreement with the correct result up to the larger time \(t\sim 2s\), _i.e._, in about the 15% of the whole relevant time window. It is moreover both as much correct as the 1st order result of the standard method (though going to zero slightly more rapidly than the correct solution) and definitely better than the same approximation without refinement at large times. In particular, this is true for \(t\sim 9\div 14s\), that covers about a remaining 35% of the whole relevant time window. Nevertheless, the refined SPDERG approximation turns out to be the one that makes the largest error, by over-evaluating the correct complex concentration behaviour, in the remaining part of the relevant time window (in detail, for \(t\sim 2.5\div 6.5s\)).
Thus, the refined SPDERG 2nd order UAs we proposed successfully capture the correct dynamical behaviour in a large part of the relevant time window, despite of the considered case being very demanding. First of all, this further confirms both the correctness and the utility of the SPDERG approach in general. Moreover, these findings support the correctness of the present proposed way for obtaining asymptotically vanishing solutions, that exploits as much as possible the results, too. At least within our kind of SPDERG application procedure, the proposed refined SPDERG UAs appear easily generalizable to other similar cases. At the same time, the present analysis makes clear that the remaining part, to be analytically calculated, of the 2nd order outer contributions would be important for an approximation to MM kinetics, beyond the sQSSA, that could work in the whole relevant time window, whatever the kinetic constants are, for values of the expansion parameter as large as \(\varepsilon=e_{0}/s_{0}\sim 0.5\).
## 9 Conclusions
In the present work, we start by recalling the standard PE method in the case of MM kinetics, beyond the sQSSA, _i.e._, the method that is generally used to deal with the problem. Against this background, we are able to successfully apply to this case the alternative SPDERG approach to boundary layer problems, by clarifying similarities and differences.
The procedure that we choose for applying the approach makes use of a basic observation in, that the bare quantities to be effectively renormalized are the ICVs of the problem (in fact, here, the substrate one). By starting from the 1st order ODEs to be obeyed by the inner solutions, with ICVs given at a generic time, and by performing the calculations up to the 2nd order, besides generally outlining the contribution of the single terms in the functions to the behaviours captured by the obtained approximations, we are able to show that one gets exactly the same outer component, as in the standard PE 1st order UA, for the substrate (thus, as expected, also for the complex).
First of all, this result was not to be taken for granted.
In \(A)\) (respectively \(C\))) we present the behaviour of the concentration of the substrate \(s(t)\) (respectively the complex \(c(t)\)) in the central part of the time window, in logarithmic time scale, whereas in \(B)\) (respectively \(D\))) we present the large time behaviour of these two quantities. The results are for the \(b\) set of ICVs given in, _i.e._, with \(\varepsilon=\varepsilon^{b}=0.5\). We plot the numerical solutions of Eqs., already shown in the previous figures, and we compare with these correct behaviours: i) the PE 1st order UAs (as given in, already presented in [Fig. 4C], [Fig. 4D]); ii) the SPDERG 2nd order UAs (as given in,, already presented in [Fig. 6C], [Fig. 6D]); iii) the refined SPDERG 2nd order UAs (as given in, already presented in [Fig. 7C], [Fig. 7D]). All these UAs to the correct solutions are computed by means of the same standard numerical approximation for the Lambert function. When they belong to the relevant time window, we finally plot our corresponding rough evaluations of the two different time scales involved, too, with \(r_{s}^{b}\) describing the substrate decay time and \(\tau_{e}^{b}\) the complex saturation time.
Therefore, we interpret the outcome of the present study also as a further confirmation, from a very different point of view, of the correctness of the known PE results up to 1st order.
At the same time, within our application, the SPDERG approach turns out to make possible to correctly manage both the secular terms and the constant ones, without the imposition of MCs (that in the present case, already at the 1st order, need in fact to be two term conditions, that involve the first derivatives of the outer solutions, too).
Actually, one can see an analogy between the present approach and the MCs of the standard method, since the key ingredient here, _i.e._, the fact that the renormalized part of the solution has to be independent of the arbitrary time \(\lambda\), appears similar to the imposition of the MCs at an unknown time. Then, one notices that, when imposing the MCs in the standard PE method, the one for the substrate needs to apply to the complex, too. Thus, within this context, it appears expectable that, as we verified, the ODE to be obeyed by the renormalized substrate ICV is the only one to be obtained. In detail, we verified that the one found from the study of the substrate is the same as the one found in the case of the complex.
On the one hand, within the SPDERG approach, it is necessary to perform the calculations up to the 2nd order for recovering the 1st order outer contribution of the standard PE method. On the other hand, this allowed us to present the 2nd order contribution of the inner solutions for the first time to our knowledge. Moreover, this also allowed us to hypothesize, by assuming that the observed solution behaviours could be iterated, that the constant terms at a given order, not to be renormalized, play the role of a part of the outer component at the following order. Indeed, this appears equivalent to both the imposition of the first term of the MC and the need for cancelling one of the constant terms that appear twice in the UA, within the PE. In fact, it is reasonable to expect that this is a general characteristic of the SPDERG approach to boundary layer problems, at least in the present kind of application procedure, in which one directly renormalizes the bare ICVs. In particular, this observation allowed us to propose refined SPDERG 2nd order UAs, that contain the parts of the 2nd order outer components that are predictable on the basis of our hypothesis, and that are thus asymptotically vanishing, too.
The conclusive comparison among the best different approximations that we considered, in the more demanding case that we studied (the one with the larger value of the expansion parameter, \(\varepsilon=\varepsilon^{b}=0.5\)), shows that the time region in which the SPDERG approach at the 2nd order works better than the PE method at the 1st order, definitely encompasses the matching region. In fact, the obtained SPDERG approximations are nearly indistinguishable, within our numerical precision, from the correct solutions of the problem in about the first 15% of the relevant time window. In detail, the extension of this region is slightly different for the substrate and for the complex, and it is in both of the cases larger when considering the refined UAs. Moreover, in the case of the complex, this last refined approximation works at least as well as the PE 1st order UAs in a part of the relevant time window that also includes the large times.
Actually, we studied particular demanding cases of MM kinetics. In fact, with the present kinetic constant choice, in logarithmic time scale, the curve for the substrate displays three inflection points instead of a single one already for values of the expansion parameter as small as \(\varepsilon=0.1\). This observation appears related to an ICV for the sQSSA to the complex behaviour that, apart from being as usual larger than zero, is even larger than the maximum reached by the correct solution for this quantity. Indeed, these qualitative observations would deserve a more careful study, that could allow a better understanding of the parameter's dependence of the relevant time scales in MM kinetics.
From a different point of view, an interesting advantage of the present procedure for applying the SPDERG approach is that, at the 2nd order, one needs to solve a simpler ODE for the 1st order outer component than the corresponding one in the standard method (just because the other part of the outer contribution is already known). This could turn out to be particularly useful when attempting to apply the same procedure within the different framework of the tQSSA since, despite of this framework being more largely applicable from the experimental point of view, the outer solution is not known explicitly at the 1st order and at the 0th order (in any event, it is not known in terms of the Lambert function).
Finally, we notice that, despite of their being quite technical and cumbersome, there is no particular difficulty in the present calculations. The performed detailed verifications, apart from being possibly better formalized in the future, can be in any event taken for granted in other similar cases, too. Thus,the present results definitely support further applications of the SPDERG approach to boundary layer problems.
|
10.48550/arXiv.1611.04886
|
An alternative approach to Michaelis-Menten kinetics that is based on the Renormalization Group: Comparison with the perturbation expansion beyond the sQSSA
|
Barbara Coluzzi, Alberto Maria Bersani, Enrico Bersani
| 3,691
|
10.48550_arXiv.1709.07648
|
## 1 Introduction
Nuclear quantum effects, such as the zero-point energy and its interplay with the anharmonic character of the hydrogen bond (HB), affect a large number of water's properties ranging from its microscopic structure and dynamics to its thermodynamic and chemical behavior.
Recently, new experimental and simulation techniques have been used to probe the quantum state of hydrogen nuclei in water and water systems by examining the hydrogen nuclear momentum distribution, \(n(p)\), and the hydrogen nuclear mean kinetic energy, \(\langle E_{K}\rangle\). These physical quantities are influenced by quantum effects and can be uniquely accessed via high energy neutron scattering using the deep inelastic neutron scattering (DINS) technique. The DINS refers to a specific regime of inelastic neutron scattering in which the incident neutron energy is well above the binding energies of the scattering atoms. This condition is experimentally achieved at high energy \(\hbar\omega\) (\(\geq\) 1 eV) and momentum \(\hbar q\) transfers (\(\geq\) 25 A\({}^{-1}\)). It is a specific regime where the neutron-scattering process is theoretically described within the framework of the impulse approximation (IA), which is exact in the limit of infinite momentum transfer, \(\hbar q\).
There are several reports on DINS experiments and theoretical studies of \(n(p)\) lineshapes and \(\langle E_{K}\rangle\) values of hydrogen and light nuclei in water systems and in a variety of other materials. In particular, \(n(p)\) and \(\langle E_{K}\rangle\) observables are routinely used to fingerprint changes in the hydrogen bond network in water and water systems. Most recent reviews on the experimental studies and the use of eV neutron spectroscopy to investigate the properties of light nuclei in water and complex materials can be found in Ref. and Ref., respectively. DINS measurements of these observables are benchmarked with the results of electronic density functionals used in path integral molecular dynamics (PIMD) for the description of hydrogen bonded systems in _ab initio_ numerical simulations. The most recent examples are DINS and PIMD studies in ice and water, ice, supercritical water, and supercooled water. In these cases, simulations and DINS experimental results provide new information on the three-dimensional effective potential energy surface experienced by the hydrogen nucleus. In this context, a relevant parameter is the mean force (MF) function, \(f(x)\), which provides an insight into the forces in a molecular system and is used to describe the average force acting on an atomic particle by keeping all other particles in the system fixed. Indeed, as pointed out by Feynman, many problems of the molecular structure are essentially concerned with forces, such as the stiffness of chemical bonds and geometrical arrangements due to repulsions and attractions between atoms.
The mean force is expressed in terms of the spherical end-to-end distribution, \(\tilde{n}(x)\), i.e., the Fourier transform of \(n(p)\). For example, in the DINS and simulation study on hexagonal ice in Ref., it has been reported how the \(f(x)\) function can be derived from the experimental data and howthe accuracy required to unambiguously resolve and extract the effective hydrogen nuclear potential can be evaluated. It is also shown how the derivation depends on the signal-to-noise ratio in the DINS count rate. The latter is a consequence of data uncertainties and error propagation derived from the sequence of experimental correction routines and data analysis procedures.
In this study, we extend the formalism of mean force as a direct, model-independent, non-parametric approach to probe the experimental sensitivity to anharmonicity in the hydrogen nuclear effective potential, in order to separate the effects of anharmonicity from those of molecular anisotropy. This is applied first to synthetic, spherically averaged, momentum distribution data from model systems representing the local hydrogen environment in harmonic anisotropic potentials, and then, as a first approximation, by adding anharmonicity along the bond direction by using a simple Morse potential. Finally, the formalism is applied to experimental data on amorphous and polycrystalline ices, showing that the local environment of the hydrogen nucleus in amorphous ices is characterized by a more pronounced anharmonicity of the hydrogen nuclear effective potential, in comparison to that in ice 1\(h\).
## 2 Response Function from DINS experiments
The IA assumes that the scatterers recoil freely from the collision with neutrons, the inter-particle interaction in the final state being negligible. The regime can be regarded as a special case of the incoherent approximation, where, in the case of high-energy collisions, a short-time expansion (\(t\to 0\)) of the atomic position operator, \(\mathbf{R}(t)\), is applied to the position operator of any scatterer of mass \(M\) and momentum \(\mathbf{p}\), i.e., \(\mathbf{R}(t)=\mathbf{R}+\frac{t}{M}\mathbf{p}\). By applying the momentum- and energy-conservation laws, it can be shown that the energy distribution of the scattered neutrons is directly related to the distribution of particle momenta parallel to the wave vector transfer \(q\), and the resulting (incoherent) dynamic structure factor yields:
\[S(\mathbf{q},\omega)=\hbar\int n(\mathbf{p})\delta\left(\hbar\omega-\hbar \omega_{r}-\frac{\hbar\mathbf{q}\cdot\mathbf{p}}{M}\right)d\mathbf{p} \tag{1}\]
Using the West scaling formalism, the two dynamic variables \(\omega\) and \(\mathbf{q}\) can be coupled by introducing the West variable \(y=\frac{1}{\hbar}\mathbf{p}\cdot\mathbf{\hat{q}}=\frac{M}{\hbar^{2}q}(\hbar \omega-\hbar\omega_{r})\), so that Eq. 1 can be re-written as
\[S(\mathbf{q},\omega)=\frac{M}{\hbar q}J(y,\mathbf{\hat{q}}), \tag{2}\]
where \(J(y,\mathbf{\hat{q}})\) is the response function, or Neutron Compton Profile (NCP), within the IA framework:
\[J(y,\mathbf{\hat{q}})=\hbar\int n(\mathbf{p})\delta\left(\hbar y-\mathbf{\hat{ q}}\cdot\mathbf{p}\right)d\mathbf{p}. \tag{3}\]
\(J(y,\mathbf{\hat{q}})\) represents the probability that the atomic nucleus has a momentum parallel to \(\mathbf{\hat{q}}\) of magnitude between \(\hbar y\) and \(\hbar(y+dy)\).
For isotropic samples, the momentum distribution only depends on \(|\mathbf{p}|\), and the \(\mathbf{\hat{q}}\) direction becomes immaterial. Thus, the NCP is expressed by \(2\pi\hbar\int_{|y|}^{\infty}pn(p)dp\), and the expression for \(n(p)\) yields:
\[n(p)=-\frac{1}{2\pi\hbar^{3}}\left[\frac{dJ(y)}{dy}\right]_{\hbar y=p}. \tag{4}\]
The IA is exact only in the limit of infinite wave vector transfer. At finite values of \(\mathbf{q}\), deviations occur, which are caused by the localization of the scatterer in its final state due to surrounding atoms, and are termed final state effects (FSEs). This causes a broadening of \(J(y)\) that resembles an instrumental-resolution effect. In the presence of FSEs, the \(J(y)\) function shows an additional dependence on \(q\), which in the isotropic case is expressed as a \(\frac{1}{q}\) power series:
\[J(y,q)=J(y)-\frac{A_{3}}{q}\frac{d^{3}J(y)}{dy^{3}}+.... \tag{5}\]
Currently, DINS measurements are carried out at the VESUVIO beamline at the ISIS pulsed neutron and muon source (Rutherford Appleton Laboratory, Chilton, Didcot, UK) using the time of flight technique. VESUVIO is the only instrument designed to exploit the DINS technique at high energy- and momentum transfers.
A DINS experiment on a condensed-water sample yields an experimental NCP, \(F_{l}(y,q)\), from each \(l\)-\(th\) detector, for the hydrogen (or oxygen) nuclei. For each detector element \(l\), the experimental NCP, due to finite \(q\) values in the neutron scattering process, retains the \(q\) dependence and is related to the DINS count rate via the expression:
\[F_{l}(y,q)=\frac{BM}{E_{0}\ I(E_{0})}q\,C_{l}(t) \tag{6}\]
DINS data sets of all samples are \(y\)-scaled according to Eq..
In a DINS experiment the asymptotic IA profile, strictly valid in the limit of infinite \(q\) (asymptotic regime), is broadened for each individual \(l\)-\(th\) detector by finite \(q\) corrections terms,\(\Delta J_{l}(y,q)\), known as final state effects (FSEs) and by the instrumental resolution function, \(R_{l}(y,q)\) :
\[F_{l}(y,q)=[J(y)+\Delta J_{l}(y,q)]\otimes R_{l}(y,q). \tag{7}\]
This equation is used to describe the experimental NCP of Eq. for each individual _l-th_ detector, \(F_{l}(y,q)\). Full details on DINS formalism, description of operation of the VESUVIO instrument and experimental set up, experimental corrections and data analysis are reported in References.
## 3 Potential of Mean Force
The interparticle potential in a condensed system can be expressed as a sum of pairwise terms which depend on the relative coordinates between particles. A coordinate \(R(\vec{\mathbf{q}})\) is used to indicate a hydrogen bond, a torsional angle, or linear combinations of similar quantities; \(\vec{\mathbf{q}}\) is the generalized vector coordinate, along which the free-energy profile can be determined.
The free-energy profile, referred to as the Potential of Mean Force (PMF), is defined as the potential energy arising from the average force acting between two fixed particles, with the average taken over the ensemble of configurational states for the remaining \(N-2\) particles.
Through the use of \(R(\vec{\mathbf{q}})\), one can define the system in a hypersurface within the phase space, allowing one to derive the free energy, \(F_{R}(R^{\prime})\), the partition function, \(Z_{R}(R^{\prime})\), and the end-to-end reaction-coordinate distribution function, \(P_{R}(R^{\prime})\). These functions yield:
\[Z_{R}(R^{\prime})=\frac{1}{h^{3N}N!}\int\int e^{-\frac{\mathcal{H}(\vec{ \mathbf{q}})}{k_{B}T}}\delta(R^{\prime}-R(\vec{\mathbf{q}}))d\vec{\mathbf{p}} d\vec{\mathbf{q}}; \tag{8}\]
\[P_{R}(R^{\prime})=\frac{Z_{R}(R^{\prime})}{Z}=\frac{\int\int e^{-\frac{ \mathcal{H}(\vec{\mathbf{q}})}{k_{B}T}}\delta(R^{\prime}-R(\vec{\mathbf{q}}))d \vec{\mathbf{p}}d\vec{\mathbf{q}}}{\int\int e^{-\frac{\mathcal{H}(\vec{ \mathbf{q}})}{k_{B}T}}d\vec{\mathbf{p}}d\vec{\mathbf{q}}}; \tag{9}\]
\[F_{R}(R^{\prime})=-k_{B}TlnP_{R}(R^{\prime})-k_{B}TlnZ, \tag{10}\]
Following Lin _et al._, one can express both the partition function and the momentum distribution, \(n(p)\), in terms of the one body density matrix \(\rho(\mathbf{r},\mathbf{r}^{\prime})=\left<\mathbf{r}|\mathbf{e}^{-\beta \mathcal{H}}|\mathbf{r}^{\prime}\right>\):
\[Z=\int d\mathbf{r}\rho(\mathbf{r},\mathbf{r}^{\prime}) \tag{11}\]
and
\[n(\mathbf{p})= \frac{1}{(2\pi\hbar)^{3}Z}\int d\mathbf{r}d\mathbf{r}^{\prime}e^ {\frac{i}{2}\mathbf{p}\cdot(\mathbf{r}-\mathbf{r}^{\prime})}\rho(\mathbf{r}, \mathbf{r}^{\prime})=\]
\[=\frac{1}{(2\pi\hbar)^{3}}\int d\mathbf{x}e^{\frac{i}{2}\mathbf{p}\cdot \mathbf{x}}\tilde{n}(\mathbf{x}) \tag{12}\]
with
\[\tilde{n}(\mathbf{x})=\frac{1}{Z}\int d\mathbf{r}d\mathbf{r}^{\prime}\delta( \mathbf{r}-\mathbf{r}^{\prime}-\mathbf{x})\rho(\mathbf{r},\mathbf{r}^{\prime }). \tag{13}\]
A viable computational strategy in the investigation of a condensed system is the statistical sampling using the Feynman path integral (PI) representation: \(\tilde{n}(\mathbf{x})\) is the end-to-end distribution derived by a sum over open paths whereas closed paths determine \(Z_{R}(R^{\prime})\).
In such a representation, the density matrix is expressed by
\[\rho(\mathbf{r},\mathbf{r}^{\prime})=\int_{\mathbf{r}=\mathbf{r},\mathbf{ r}(\beta\hbar)=\mathbf{r}^{\prime}}\mathcal{D}\mathbf{r}(\mathbf{r})e^{- \frac{1}{\hbar}\int_{0}^{\beta\hbar}d\tau\left(\frac{\mathbf{r}^{2}(\tau)}{2 }+V[\mathbf{r}(\tau)]\right)} \tag{14}\]
If the linear transformation \(\mathbf{r}(\tau)=\mathbf{\tilde{r}}(\tau)+y(\tau)\mathbf{x}\) is carried out in the path space, then one can express the NCP in terms of the distribution \(\tilde{n}(\mathbf{x})\). This action reshapes the open path \(\mathbf{r}(\tau)\) into the closed path \(\mathbf{\tilde{r}}(\tau)\), with the free particle contribution coming from the derivative of \(y(\tau)\). Thus, the end-to-end distribution is given by:
\[\tilde{n}(\mathbf{x})=\frac{\int_{\mathbf{r}-\mathbf{r}(\beta\hbar)= \mathbf{x}}\mathcal{D}\mathbf{r}(\tau)e^{-\frac{1}{\hbar}\int_{0}^{\beta \hbar}d\tau\left(\frac{\mathbf{r}^{2}(\tau)}{2}+V[\mathbf{r}(\tau)]\right)}}{ \int_{\mathbf{r}(\beta\hbar)=\mathbf{r}}\mathcal{D}\mathbf{r}(\tau)e^{- \frac{1}{\hbar}\int_{0}^{\beta\hbar}d\tau\left(\frac{\mathbf{r}^{2}(\tau)}{ 2}+V[\mathbf{r}(\tau)]\right)}}=\\ =e^{-\frac{\mathbf{n}\mathbf{x}^{2}}{2\beta\hbar^{2}}\int_{ \mathbf{r}(\beta\hbar)=\mathbf{r}}\mathcal{D}\mathbf{\tilde{r}}(\tau)e^{ -\frac{1}{\hbar}\int_{0}^{\beta\hbar}d\tau\left(\frac{\mathbf{r}^{2}(\tau)}{2 }+V[\mathbf{r}(\tau)]\right)}}\\ \int_{\mathbf{r}(\beta\hbar)=\mathbf{r}}\mathcal{D}\mathbf{r}( \tau)e^{-\frac{1}{\hbar}\int_{0}^{\beta\hbar}d\tau\left(\frac{\mathbf{r}^{2}( \tau)}{2}+V[\mathbf{r}(\tau)]\right)}. \tag{15}\]
The equations above allow us to express the NCP in terms of \(\tilde{n}(\mathbf{x})\):
\[J(y,\mathbf{\hat{q}})=\frac{1}{2\pi\hbar}\int dx_{\parallel}\tilde{n}(x_{ \parallel}\mathbf{\hat{q}})e^{\frac{i}{2}\mathbf{r}_{\parallel}y}, \tag{16}\]
By making use of the primitive approximation:
\[\tilde{n}(\mathbf{x})=e^{-\frac{\mathbf{n}\mathbf{x}^{2}}{2\beta\hbar^{2}}}e^{ -\beta U(\mathbf{x})}, \tag{17}\]
the expressions for \(U(x_{\parallel}\mathbf{\hat{q}})\) and the MF, \(f(x_{\parallel}\mathbf{\hat{q}})\), become:
\[U(x_{\parallel}\mathbf{\hat{q}})=-\frac{mx_{\parallel}^{2}}{2\beta^{2}\hbar^{ 2}}-\frac{1}{\beta}\ln\int dyJ(y,\mathbf{\hat{q}})e^{ik_{\parallel}y} \tag{18}\]
and
\[f(x_{\parallel}\mathbf{\hat{q}})=-\frac{mx_{\parallel}}{\beta^{2}\hbar^{2}}+ \frac{1}{\beta}\frac{\int_{0}^{\infty}y\sin(x_{\parallel}y)J(y,\mathbf{\hat{q}} )dy}{\int_{0}^{\infty}d\y\cos(x_{\parallel}y)J(y,\mathbf{\hat{q}})}, \tag{19}\]For finite temperature systems, Eq. 17, which is valid for time step \(\tau=\frac{\beta}{N}\to 0\), with \(N\) the number of virtual replicas, is of particular relevance.
Indeed, the sequence of Feynman-Trotter approximations to the thermal Feynman path integral for a general non-relativistic system characterized by a smooth, single-minimum interaction potential converges pointwise to the quantum thermal propagator at every non-zero temperature, but in the zero-temperature limit, for high-order elements of the sequence, an abrupt "collapse" from the quantum to the classical ground-state takes place. In other words, for all _finite_\(N\)-values the \(T\to 0\) limit is unphysical. This situation can be mitigated by either increasing \(N\) (which can prove computationally demanding) or implementing any alternative low-\(T\) formulation to Feynman's original one, such as coherent-state path integral (CSPI).
In any case, Eq. 18 is an application of Feynman mapping of the quantum system onto a set of replicas obeying to classical mechanics, where each particle in a string of replicas is referred to as a "bead", and adjacent beads interact via a harmonic potential of frequency \(\sim\sqrt{\frac{1}{\beta^{2}\hbar^{2}}}\).
It is worth noticing here that the operational temperature mainly impacts the MF curve in terms of the slope of its local tangent in \(x_{\parallel}=0\) A, via the dominant negative addend in Eq. 19.
## 4 Mean Force in anisotropic harmonic potentials
Previous DINS and simulation studies have described the momentum distribution of the hydrogen nuclei in water as a spherical average of a multivariate Gaussian according to
\[4\pi p^{2}n(p)=\Big{\langle}\frac{\delta(p-|\mathbf{p}|)}{\sqrt{8\pi^{3}} \sigma_{x}\sigma_{y}\sigma_{z}}\exp\left(-\frac{p_{x}^{2}}{2\sigma_{x}^{2}}- \frac{p_{y}^{2}}{2\sigma_{y}^{2}}-\frac{p_{z}^{2}}{2\sigma_{z}^{2}}\right) \Big{\rangle}, \tag{20}\]
The set of parameters, \(\sigma_{x,y,z}\), determines the anisotropy in the momentum distribution line shape, with:
\[\sigma_{i}^{2}=\frac{M\omega_{i}}{2\hbar}coth\left(\frac{\beta\hbar\omega_{i}} {2}\right), \tag{21}\]
\(\omega_{i}\) being an effective principal frequency.
The spherical average of \(n(p)\) in Eq. 20 is carried out over all possible molecular orientations, explicitly yielding:
\[n(p)=\frac{1}{4\pi}\frac{1}{(2\pi)^{\frac{3}{2}}}\frac{1}{\sigma_{x}\sigma_{y} \sigma_{z}}\int_{0}^{2\pi}d\phi\int_{0}^{\pi}\left[sin(\theta)e^{-\frac{1}{2} p^{2}S(\theta,\phi)}\right]d\theta. \tag{22}\]
and the corresponding expression for the NCP is:
\[J(y)=\frac{\hbar}{2}\frac{1}{(2\pi)^{\frac{3}{2}}}\frac{1}{\sigma_{x}\sigma_{ y}\sigma_{z}}\int_{0}^{2\pi}d\phi\int_{0}^{\pi}sin(\theta)\frac{1}{S(\theta, \phi)}e^{-\frac{1}{2}S(\theta,\phi)\hbar^{2}y^{2}}d\theta, \tag{23}\]
where
\[S(\theta,\phi)=\frac{sin^{2}(\theta)cos^{2}(\phi)}{\sigma_{x}^{2}}+\frac{sin^ {2}(\theta)sin^{2}(\phi)}{\sigma_{y}^{2}}+\frac{cos^{2}(\theta)}{\sigma_{z}^{ 2}}. \tag{24}\]
Eqs. 22 and 23 can be evaluated numerically.
For an isotropic system, the NCP is a univariate Gaussian, i.e., \(J(y)\)=\(\frac{1}{\sqrt{2\pi}\sigma}e^{-\frac{y^{2}}{2\sigma^{2}}}\) and the mean force in Eq. yields:
\[f(x_{\parallel}\mathbf{\hat{q}})=-\frac{mx_{\parallel}}{\beta^{2}\hbar^{2}}+ \frac{1}{\beta}\frac{\frac{\sqrt{\pi}}{4}(2\sigma^{2})^{\frac{3}{2}}x_{ \parallel}e^{-\frac{\sigma^{2}}{2}x_{\parallel}^{2}}}{\frac{\sqrt{\pi}}{2}(2 \sigma^{2})^{\frac{3}{2}}e^{-\frac{\sigma^{2}}{2}x_{\parallel}^{2}}}=\left(- \frac{m}{\beta^{2}\hbar^{2}}+\frac{\sigma^{2}}{\beta}\right)x_{\parallel} \tag{25}\]
This expression for the mean force shows a linear dependence on the coordinate \(x_{\parallel}\). Eq. is interpreted as Hooke's law governing the "nanospring" that connects two atoms along the direction of \(x_{\parallel}\), with \(k=\frac{m}{\beta^{2}\hbar^{2}}-\frac{\sigma^{2}}{\beta}\) being the related elastic constant. In the anisotropic case, a three-dimensional harmonic potential would similarly produce such a linear behavior, when directional distributions are considered.
Therefore, deviations of \(f(x_{\parallel}\mathbf{\hat{q}})\) from linearity provide evidence of underlying anharmonicity of the local potential. However, in experiments, only the spherically averaged NCP is accessible in liquids, amorphous solids, and polycrystalline solids. The effect of the spherical average is to introduce deviations from linearity on \(f(x_{\parallel}\mathbf{\hat{q}})\).
This is shown in Figs. 1 and 2, where we report the calculated hydrogen NCP and MF, respectively, of a model system with a multivariate Gaussian momentum distribution characterized by a robust anisotropy (with average variance \(\overline{\sigma}_{aniso}\) at \(T\)=100 K), together with the hydrogen NCP and MF of a model system with an isotropic momentum distribution at \(T\)=100 K, and \(\overline{\sigma}_{iso}=\overline{\sigma}_{aniso}\).
Here the deviation of the spherical force from linearity at finite \(x_{\parallel}\) results from the averaging process and is not a sign of anharmonicity.
We note that the effect of anisotropy is to introduce a concavity in the MF trend, but the slope of the resulting mean force is always lower than the slope of the corresponding isotropic reference system.
Thus, in the interpretation of the experimental Compton profiles, which result from the contribution of many particles, one must distinguish the case of an anisotropic harmonic potential energy surface from that of an anharmonic potential energy surface. In view of the above, a practical means to identify anharmonicity in the experimental data is to compare the slope of the mean force with that of the corresponding isotropic model, which can be easily derived from the spherically averaged standard deviations of the neutron Compton profiles.
### Evaluation of anharmonicity along the bond direction
In this subsection, we show how one can fingerprint anharmonicity through the inspection of the mean force.
Let us consider an anisotropic hydrogen-containing system at \(T\)=100 K. We describe the hydrogen \(n(p)\) as a spherical average of an anharmonic contribution along one direction, i.e., the \(z\) axis, and two harmonic components along \(x\) and \(y\), respectively.
Let us suppose that a simple Morse potential is acting along the O\(-\)H covalent bond (\(z\) axis), which, in the typical tetrahedral arrangement of molecules in condensed water, can be represented, to a first approximation, as collinear with a hydrogen bond.
Of course, this type of modeling neglects the effects caused by both anharmonic coupling between covalent- and HB vibrations (in turn dependent on deviations from collinearity) and any departure from Morse-like 1D anharmonicity (in favor of a double-well potential), which can occur in stronger intermolecular hydrogen bonds.
If the anharmonic contribution to the eigenfunction \(\phi(\mathbf{p})\) of the ground state in the momentum space is represented as coming from a Morse-oscillator motion, then the spherical average of the momentum distribution yields:
\[\begin{array}{l}n(p)=\frac{1}{16\pi^{3}}\frac{2^{k-1}}{\alpha_{z}\sigma_{z }\sigma_{z}\hbar}\frac{2\lambda_{z}-1}{1(2\lambda_{z})}\int_{0}^{2\pi}d\phi~{} \int_{0}^{\pi}d\theta\times\\ \\ \times\left[sin(\theta)e^{-\frac{p^{2}}{2}\left(\frac{\sin^{2}(\theta)\omega_{ \mathrm{osc}}^{2}(\phi)}{\alpha_{z}^{2}}+\frac{\sin^{2}(\theta)\omega_{ \mathrm{osc}}^{2}(\phi)}{\alpha_{z}^{2}}\right)}\right]\Gamma\left(\lambda_{z}- \frac{1}{2}+i\frac{pcos(\theta)}{\alpha_{z}\hbar}\right)\left|^{2}\right]\end{array} \tag{26}\]
and
\[J(y)=2\pi\hbar\int_{\left|y\right|\right\}^{\infty}pn(p)dp=\]
\[=\frac{\hbar}{8\pi^{2}}\frac{2^{k-1}}{\alpha_{z}\sigma_{z}\sigma_{z}\hbar} \frac{2\lambda_{z}-1}{\Gamma(2\lambda_{z})}\int_{0}^{2\pi}d\phi\int_{0}^{\pi} sin(\theta)d\theta~{}\times\]
\[\times\int_{\left|y\right|}^{\infty}\left[pe^{-\frac{p^{2}}{2}\left(\frac{ \sin^{2}(\theta)\omega_{\mathrm{osc}}^{2}(\phi)}{\alpha_{z}^{2}}+\frac{\sin^{2 }(\theta)\omega_{\mathrm{osc}}^{2}(\phi)}{\alpha_{z}^{2}}\right)}~{}\left|\Gamma \left(\lambda_{z}-\frac{1}{2}+i\frac{pcos(\theta)}{\alpha_{z}\hbar}\right) \right|^{2}\right]~{}dp. \tag{27}\]
We recall here that, since \(D_{z}\) is the depth of the Morse potential minimum, \(\alpha_{z}\) is its curvature, and \(\omega_{0}=\sqrt{\frac{2D_{z}\alpha_{z}^{2}}{M}}\) is the harmonic fundamental frequency of the Morse oscillator, then \(D_{z}\alpha_{z}^{2}\) can be derived from the analysis of the experimental line shapes through the following expression:
\[\sigma_{z}=\left[\frac{M}{2\hbar}\sqrt{\frac{2D_{z}\alpha_{z}^{2}}{M}}coth\left( \frac{\beta\hbar}{2}\sqrt{\frac{2D_{z}\alpha_{z}^{2}}{M}}\right)\right]^{\frac{ 1}{2}}, \tag{28}\]
The evaluation of the anharmonic contributions to the Morse potential shown above can be carried out by considering the anharmonic constant, \(x_{amh}=\frac{h\omega_{0}}{4D_{z}}=\frac{h\omega_{z}}{\sqrt{8D_{z}M}}=\frac{1} {2\lambda_{z}}\). Using typical values of \(D_{z}\) and \(\lambda_{z}\) from the literature we obtain \(x_{amh}=0.032\). The resulting behavior of the mean force is expected to show small differences from the anisotropic harmonic case. More significant changes in the curvature of the mean force can be obtained by using \(x_{amh}=3.2\), although this choice of the anharmonicity constant implies an almost flat-bottom potential. The two Morse potentials with \(x_{amh}=0.032\) and \(x_{amh}=3.2\) are shown in
Numerical calculation of \(J(y)\) were carried out for the anisotropic harmonic model, and the anisotropic anharmonic model with the two anharmonic constants described above; the corresponding mean forces along the bond (\(z\)) direction, are shown in below.
As anticipated above, considering a simple Morse potential along an O\(-\)H covalent bond collinear with a hydrogen bond amounts to neglecting, to a first approximation, a non-negligible portion of anharmonic behavior.
Indeed, hydrogen bonding normally introduce competing quantum effects on the vibrational motion of the hydrogen nucleus in the plane of the water molecule, and perpendicular to it, due to its influence on the interplay between the O\(-\)H stretch and HB bending. These effects produce a _prominent_ modification of the absorption band of the stretching mode of the X\(-\)H donor group, in terms of a red shift and spectral broadening, which are a manifestation of the competition between the anharmonic quantum fluctuations of intramolecular covalent bond stretching and intermolecular hydrogen-bond bending.
In our first- approximation model, any large anharmonic effect may be modeled by an increase of the anharmonic constant \(x_{amh}\) along the bond direction, reflecting the enhancement of the non-parabolic character of the vibrational potential. This may lead, for example, to a very shallow model potential as seen in (blue curve). Such a potential is clearly not adequate in describing the quantum state of the hydrogen nucleus in condensed water phases.
These findings suggest that an improved description of the anharmonicity due to the intermolecular hydrogen bonding has to take into account the non-collinearity of the covalent-bond and hydrogen-bond directions, in the sense that each hydrogen atom should move under at least two non-collinear anharmonic potentials, one along the direction of the intramolecular O\(-\)H stretch and one along the intermolecular hydrogen bond.
Schematics of a single water molecule (right) and its tetrahedral arrangement in low-temperature condensed phases (left). O\(-\)H covalent bonds are in green, hydrogen bonds are in black.
Analogously, an upgrade of potential modeling to an asymmetric double-well might be an interesting test.
Notwithstanding all of the above considerations, one must notice that the model presented in this paper is already suitable, in the absence of biases due to the reduction of experimental data, to offer a practical, first-approximation tool to tell anharmonicity of the local potential apart from mere anisotropy. This is done by assessment of the experimental MF possibly exceeding the slope of the isotropic linear reference.
## 5 Application to experimental DINS data
### Amorphous ices
The determination of the hydrogen mean force from DINS data has been carried out in order to apply the models described in the previous sections to the interpretation of experimental data on amorphous and polycrystalline ice samples, with special attention to the identification of anharmonicities in the hydrogen nuclear effective potential.
The procedure adopted consists in the use of the single-detector NCPs, \(F_{l}(y,q)=[J_{\rm IA}(y)+\Delta J_{l}(y,q)]\otimes R_{l}(y,q)\) (see equation 7, and the corresponding final-state-effect contributions obtained from the previous DINS experiments of References). The single-detector NCPs were corrected for the final state effects and averaged over the set of individual detectors, yielding:
\[J_{R}(y)=\langle F_{l}(y,q)-\Delta J_{l}(y,q)\rangle_{l}. \tag{29}\]
Full details on the determination of \(F_{l}(y,q)\) and \(\Delta J_{l}(y,q)\) can be found in References. The detector-averaged profile, \(\hat{J}_{R}(y)\), contains the broadening due to the instrumental resolution. The latter is neglected to a first approximation in the present work, since its contribution is on the order of 15% of the full width at half maximum of the neutron Compton profiles. The hydrogen mean forces for the amorphous and polycrystalline samples are then determined by a numerical evaluation of Equation 19 applied to \(J_{R}(y)\) for each sample.
At first we derive the mean force using DINS data from amorphous ices at \(T\)= 80 K and standard pressure, namely, very-high-density (vHDA), unannealed high-density (uHDA), and low-density (LDA) amorphous ices, respectively. The results, plotted in Fig. 6, are compared to model profiles from a univariate (1D) Gaussian momentum distribution with a standard deviation equal to the spherically averaged standard deviation, \(\overline{\sigma}_{AI}\), found in Reference, that is, \(\sigma=\overline{\sigma}_{AI}\) (Eq. 25).
In view of the calculations and figuresLeft panel: hydrogen MF along \(\hat{\mathbf{q}}\) for vHDA (top, confidence region in yellow), uHDA (center, confidence region in magenta), and LDA (bottom, confidence region in blue), respectively, at \(T\)=80 K. The linear isotropic reference is obtained using a 1D Gaussian with \(\sigma=\overline{\sigma}_{AI}\) (black). The turquoise line appearing in the plot for vHDA represents the mean force calculated using the detector-averaged resolution profile as experimental data. AI densities signaled in the three plots are those reported in. Right panel: the same hydrogen MFs divided by the linear MF for the isotropic reference, in order to enhance MF curvature differences - with respect to the reference itself - as a function of density. The \(x_{||}\)-axis in each plot starts from 0.002 Å since the point in the origin has zero error. This is caused by error bars computed by error propagation from \(J(y)\) alone.
The right panel of clearly shows that, as density decreases (vHDA \(\longrightarrow\) uHDA \(\longrightarrow\) LDA) and within the experimental error, the local radius of curvature of the MF becomes smaller, or the \(x_{||}\)-range over which the MF is placed above its sample-specific isotropic reference becomes larger. This suggests that, in harmony with findings from inelastic and deep inelastic scattering measurements reported in Ref., the hydrogen local environment in amorphous ices is characterized by an anharmonic character of the local potential, which goes beyond mere anisotropy, and decreases as density increases. This behavior is consistent with the structural differences found in the various forms of amorphous ices: indeed, all molecules in amorphous ices are hydrogen-bonded to four approximately tetrahedrally arranged neighbors (the 'Walrafen pentamer'). HDA holds an additional molecule (at a similar distance) not directly hydrogen bonded to the central molecule and located in between its first and second coordination shell; vHDA then holds two interstitial molecules. We note that the structure of uHDA has recently been described as a "de-railed" state along the ice I to ice IV pathway. This picture is consistent with a longer average O\(-\)O distance between hydrogen-bonded molecules for vHDA (2.85 A) than for uHDA (2.82 A); an even shorter distance is found for LDA (2.77 A), in spite of a decrease in density and in the hydrogen mean kinetic energy.
In conclusion, as the hydrogen bonds weaken with increasing density, the effective potential becomes less and less anharmonic and its shape becomes similar to the shapes inferred from other spectroscopic techniques within the harmonic assumption. This is consistent with recent findings on LDA and HDA from 2D IR spectroscopy.
### Hexagonal ice I\(h\)
An early application of the MF formalism described in section 3 can be found in Ref., concerning DINS data from a polycrystalline sample of ice I\(h\) at 271 K and standard pressure.
In order to assess whether the anharmonic behavior reported above might be resulting from systematic experimental contributions, we report below the determination of the hydrogen mean force for a polycrystalline ice sample from a DINS measurement at 71 K and standard pressure.
According to the literature, polycrystalline ice I\(h\) is expected to be a quasi-harmonic system, especially at low temperature, with a small amount of anisotropy stemming from molecular orientations in the crystal.
The MF derived from experimental DINS data for this system seems to confirm the above picture: as reported in Fig. 7, bottom panel, within experimental uncertainties the mean force
Top: Numerically evaluated hydrogen MF along \(\hat{\mathbf{q}}\) for ice I\(h\) at 271 K and standard pressure (dark blue-green confidence region). In black we report the linear model obtained using a 1D Gaussian with \(\sigma=\overline{\sigma}_{Ih}\). Red full points represent the numerically evaluated MF of a spherically averaged harmonic anisotropic model system characterized by \(\sigma_{x}=3.7\) Å\({}^{-1}\), \(\sigma_{y}=4.3\) Å\({}^{-1}\), and \(\sigma_{x}=6.5\) Å\({}^{-1}\). Bottom: Numerically evaluated hydrogen MF along \(\hat{\mathbf{q}}\) for ice I\(h\) at 71 K and standard pressure (green confidence region). In black we report the linear model obtained using a 1D Gaussian with \(\sigma=\overline{\sigma}_{Ih}\).
## 6 Conclusions
This paper presents a new procedure to obtain information on the hydrogen-nucleus energy surface in water by directly expressing the mean force function, \(f(x)\), in terms of the neutron Compton profiles measured in DINS experiments, beyond what was introduced by Lin _et al._ in. The new formalism is illustrated and applied to experimental DINS data in a variety of low-temperature condensed phases of water. The calculations on model systems allow to obtain a practical tool to identify anharmonicity in the hydrogen-nucleus effective potential, and to distinguish the case of an anisotropic harmonic potential from that of an anharmonic potential, by simple inspection of the concavity and slope of the mean force.
By applying the above tools to the experimental data from DINS measurements, it is found that the shape of the mean force for amorphous and polycrystalline ice is primarily determined by the anisotropy of the underlying quasi-harmonic effective potential, and that data from amorphous ice show an additional curvature reflecting the more pronounced anharmonicity of the hydrogen-nucleus effective potential, with respect to that of ice I_h_.
The present work joins the stream of efforts to better understand the relation between the experimental momentum distribution and the atomic/molecular environment: a challenging task, though crucial to the study of condensed matter and worthy of development.
We plan to further refine the proposed approach to better formalize the role of molecular hydrogen bonding with the aim of quantitative determinations of the anharmonicities of atomic local potentials.
|
10.48550/arXiv.1709.07648
|
Hydrogen mean force and anharmonicity in polycrystalline and amorphous ice
|
Alexandra Parmentier, Carla Andreani, Giovanni Romanelli, Jacob J. Shephard, Christoph G. Salzmann, Roberto Senesi
| 321
|
10.48550_arXiv.1801.03713
|
## Abstract
Typically, a Lewis acid and a base react with each other and form classic acid-base adducts. The neutralization reaction is however prevented by the introduction of bulky substitutes and this interesting finding leads to a new concept called "frustrated Lewis pairs, FLPs". Since both reactivities of Lewis acids and bases are remained in the same systems, FLPs have been shown many important applications. One of them is hydrogen activation, which showed for the first time the use of a non-metal catalyst for that purpose. In this mini-review, we have summarized all important findings regarding the H\({}_{2}\) activation by FLPs. This includes pre-organisation of FLPs, reaction path for the activation, polarization of H-H bond and the factors affected the reactivity. In light of some recent developments, we aim to clarify the reaction mechanism for the H\({}_{2}\) actitation by FLPs, which has been under debate for decades since the first discovery of FLPs. We believe that this mini-review can be served as a guideline for the future fundamental studies and industrial applications.
|
10.48550/arXiv.1801.03713
|
Reaction mechanism of hydrogen activation by frustrated Lewis pairs
|
Lei Liu, Binit Lukose, Pablo Jaque, Bernd Ensing
| 3,802
|
10.48550_arXiv.0903.1786
|
###### Abstract
We calculate the inelastic scattering probabilities in the wide band limit of a local polaron model with quadratic coupling to bosons. The central object is a two-particle Green function which is calculated exactly using a purely algebraic approach. Compared with the usual linear interaction term a quadratic interaction term gives higher probabilities for inelastic scattering involving a large number of bosons. As an application we consider the problem hot electron mediated energy transfer at surfaces and use the delta self-consistent field extension of density functional theory to calculate and compare coupling parameters and probabilities for exciting different vibrational modes of CO adsorbed on a Cu surface.
pacs: 71.38.-k, 71.45.-d, 31.15.xr, 71.15.Qe, 82.20.Gk, 82.20.Kh
## I Introduction
The local polaron model describes a localized electronic state which is coupled to a boson field. The local state is then assumed to be hybridized with a continuum of delocalized states and are thus not an eigenstate of the electronic part of the Hamiltonian.
One of the first applications of the model was the coupling of plasmons to core holes and valence holes in metals. The boson field then represents the plasmons which can be excited by the introduction of a structureless core-hole or a valence-hole which may be hybridized with metallic states. Plasmon excitation spectra are typically measured using Electron Energy Loss Spectroscopy (EELS) where the energy loss of highly energetic electrons are measured after transmission through a metallic film. A similar application is that of deep-level spectroscopy where the boson field represents a phonon system which can be excited by the introduction of a core-hole. Hybridization of the core-hole is then introduced to capture the degeneracy of the core-hole with a continuum of states with no core-hole but a high energy Auger electron present. A somewhat different line of application is that of certain rare earth compounds which are known to give rise to mixed valence states. These states are characterized by an alternating valence in an otherwise periodic lattice which can result in unusual thermodynamic properties. The reason is that the difference in valence results in a difference in ionic radii and the extra valence electron thus have a strong coupling to the phonon system. The model designed to capture the effect consists of a localized \(f\)-state (the extra valence electron) coupled to a continuum of delocalized electrons and a phonon field coupled to the \(f\)-state. Although, there is orders of magnitude differences between typical plasmon and phonon energies the physics in the models are very similar and only the model parameters differ.
Finally the local polaron model has been applied to the problem of resonant tunneling in the context of electronic transport, and the very similar problem of Hot Electron Femtochemistry at Surfaces (HEFatS). The idea of HEFatS is that an adsorbate system on a metal surface can have unoccupied electronic states which obtain a broadening due to interaction with the metallic states. If a hot electron (an electron above the Fermi level) is generated in the metal it may interact with the unoccupied state and induce a chemical reaction on the surface. As an example we can think of a single molecule on a metal surface with one unoccupied electronic state well above the Fermi level. A hot electron with an energy that matches the unoccupied orbital has the possibility of tunnelling from the metal to the molecule resulting in a transient occupation of the orbital. If the molecule was initially in an equilibrium position the electron will assert a force on the internal molecular degrees of freedom and can excite vibrational modes of the molecule before it tunnels back into the conductor. The molecule may acquire enough energy in this process to undergo a chemical reaction or a desorption event. A clever method to produce hot electrons is based on a Metal-Insulator-Metal (MIM) heterostructure as suggested by Gadzuk. With an ideal MIM device it is possible to tune hot electrons to any desired resonance of an adsorbate system and the approach thereby suggests the highly attractive possibility of performing selective chemistry at surfaces. Such devices have been constructed and characterized and comprise a promising candidate for advanced HEFatS experiments.
Common to all these applications is that a linear bosonic coupling term has been assumed. It is by no means obvious that linear coupling captures the possibly complicated interaction of bosons and electrons, although it is probably often a good approximation. To examine the local polaron model beyond linear coupling we calculate the consequences of substituting the linear coupling term with a quadratic coupling term. In principle we should add the quadratic coupling on top of the linear, but this renders the model somewhat tedious to work with and the physics of quadratic coupling become hidden in complicated expressions. In contrast, having only quadratic coupling allow us to obtain inelastic scattering amplitudes very similar to those with linear coupling andthe comparison is very instructive. In terms of bosonic potentials, a linear coupling term corresponds to a shift in the potential minimum whereas a quadratic coupling term corresponds to a shift in the frequency of the potential.
The paper is organized as follows: In section II we present the local polaron model with a general coupling function and no bosonic dispersion. The electronic part is briefly reviewed and the wide band limit which is imposed in the remainder of the paper is defined. We then present the well known spectral function and inelastic scattering probabilities of the model with linear coupling and compare with a quadratically coupled model calculated in the present work. It is shown that for inelastic scattering involving a large number of bosons, quadratic coupling can give rise to much larger scattering probabilities. In section III we apply the theory to hot electron mediated excitation of the different modes of CO adsorbed on Cu. The model parameters are calculated using density functional theory and the delta self-consistent field method and we find that linear coupling dominates desorption probabilities for the normal modes along the molecular axis, but vanishes for the frustrated rotations where quadratic coupling has to be taken into account. In appendix A we derive a path integral representation of the Newns-Anderson retarded Green function and show that the special properties of the wide band limit allow us to decouple bosonic and electronic degrees of freedom. Appendix B and appendix C present the details of the calculations leading to the spectral functions and inelastic scattering probabilities associated with linear and quadratic coupling. In appendix D, we show how a linear transformation of creation operators makes it possible to obtain the exact Green functions including both linear and quadratic coupling.
## II Model
### The Newns-Anderson model with coupling to bosons
The general model we are concerned with is composed of a Newns-Anderson type Hamiltonian coupled to a dispersionless (single frequency) boson field through a single electronic state. A dispersionless boson field naturally corresponds to a single mode of oscillation in an adsorbate system, whereas we can think of the dispersionless model as describing an Einstein band if the boson field represents a phonon system. Thus it is a model of non-interacting metallic electrons \(|k\rangle\), a localized resonant state \(|a\rangle\), and a harmonic oscillator described by the coordinate \(x\) or equivalently the bosonic creation and annihilation operators \(a^{\dagger}\) and \(a\).
\[H= \sum_{k}\epsilon_{k}c_{k}^{\dagger}c_{k}+\sum_{k}\left(V_{ak}c_{a }^{\dagger}c_{k}+V_{ak}^{*}c_{k}^{\dagger}c_{a}\right)\] \[+\hbar\omega_{0}a^{\dagger}a+\varepsilon_{a}(x)c_{a}^{\dagger}c_ {a}, \tag{1}\]
The function \(\varepsilon_{a}(x)\) couples the resonant electron to the oscillator degrees of freedom. If one considers a hole coupled to the bosons instead of an electron, the order of \(c_{a}\) and \(c_{a}^{\dagger}\) should be exchanged.
It is natural to Taylor expand the coupling function in the vicinity of the ground state minimum \(x_{0}\). Including only the zeroth order term \(\varepsilon_{a}(x_{0})=\varepsilon_{0}\) results in the Newns-Anderson model. Since it is quadratic in the electronic creation and annihilation operators one could in principle formally diagonalize it.
\[c(t)=e^{iHt/\hbar}ce^{-iHt/\hbar}.\]
It is easily calculated in the energy domain using the Dyson equation and the result is
\[G_{R}^{0}(\omega)=\frac{1}{\hbar\omega-\varepsilon_{0}-\Sigma(\omega)+i\Gamma( \omega)/2}, \tag{3}\]
with
\[\Gamma(\omega)=2\pi\sum_{k}|V_{ak}|^{2}\delta(\hbar\omega-\epsilon_{k}), \tag{4}\]
and
\[\Sigma(\omega)=\int\frac{d\omega^{\prime}}{2\pi}\frac{\Gamma(\omega)}{\omega- \omega^{\prime}}. \tag{5}\]
Assuming the hopping matrix elements \(V_{ak}\) to be constant, \(\Gamma(\omega)\) becomes proportional to the metal density of states. If we furthermore assume the metal density of states to be wider than the resonance energy we can write \(\Gamma(\omega)=\Gamma(\varepsilon_{0})\) and \(\Sigma=0\). This is the wide band limit which will be imposed in the present paper. It allow us to separate electronic and bosonic degrees of freedom in the general case and we can calculate Green functions corresponding to linear and quadratic coupling exactly.
In the wide band limit the electronic retarded Green function is
\[G_{R}^{0}(t)=-i\theta(t)e^{-(i\varepsilon_{0}+\Gamma/2)t/\hbar}, \tag{6}\]
When boson coupling terms are included (first and second order Taylor expansions of \(\varepsilon_{a}(x)\) in) the spectral function changes and inelastic scattering on the resonance becomes possible.
Suppose the resonance is initially unoccupied and the oscillator is in the state \(n\).
\[R(n;\varepsilon^{\prime},\varepsilon)= \Gamma^{2}\int\frac{d\tau dsdt}{2\pi\hbar^{3}}e^{i(\varepsilon- \varepsilon^{\prime})\tau/\hbar+i\varepsilon^{\prime}t/\hbar-i\varepsilon s/\hbar}\] \[\times G(n;\tau,s,t), \tag{7}\]
where
\[G(n;\tau,s,t)=\theta(s)\theta(t)\langle n|c_{a}(\tau-s)c_{a}^{\dagger}(\tau)c_ {a}(t)c_{a}^{\dagger}|n\rangle.\]
The probability of transferring a given amount of energy to the bosons can thus be obtained by integrating the inelastic scattering matrix over the relevant values of \(\varepsilon^{\prime}\). In this paper we focus on inelastic scattering by electrons, since this is the relevant quantity in the context of HEFatS and EELS. However, the two-particle Green function also appear in the calculation of optical transition amplitudes of an adsorbed molecule and knowing \(G(n;\tau,s,t)\) allows one to calculate a variety of observable quantities.
Finally, we note that the lifetime of the electron is independent of the boson coupling in the wide band limit.
\[p_{a}(n;t)=\sum_{m=0}^{\infty}|\langle m,a;t|n,a;0\rangle|^{2}=e^{-\Gamma t/ \hbar}, \tag{8}\]
### Coupling function and adiabatic potentials
Consider the state \(|x,a\rangle\) with the oscillator at \(x\) and an electron occupying the resonance. The expectation value of the Hamiltonian on such a state will depend on the value of \(x\) due to the coupling \(\varepsilon_{a}(x)\) and if we could calculate the electronic energy for all values of \(x\) we would obtain an excited state potential \(V_{1}(x)=\langle x,a|H|x,a\rangle\). Doing the same for the state with no electron in the resonance \(|x,0\rangle\) would result in a different potential \(V_{0}(x)\) and the coupling function should then be given by \(\varepsilon_{a}(x)=V_{1}(x)-V_{0}(x)\) which is illustrated in In the model we have implicitly assumed that the potential \(V_{0}(x)\) is quadratic, but in general it could have any form. The potentials \(V_{1}(x)\) and \(V_{0}(x)\) are called Born-Oppenheimer surfaces and are obtained by moving the oscillator adiabatically in the electronic environment.
### Linear coupling
We now Taylor expand the coupling function \(\varepsilon_{a}(x)\) to first order and express the boson coordinate in terms of creation and annihilation operators.
\[\lambda_{1}=\frac{l}{\sqrt{2}}\frac{\partial}{\partial x}V_{1}\Big{|}_{x=x_{0 }},\qquad l=\sqrt{\frac{\hbar}{m\omega_{0}}}, \tag{10}\]
This model corresponds to the potentials and coupling function shown in
A shifted excited state corresponds to a linear coupling function \(\varepsilon_{a}(x)\sim x\). The strength of the coupling is proportional to the derivative of the excited state at \(x_{0}\).
A general example of adiabatic potentials \(V_{1}(x)\) and \(V_{0}(x)\) and the coupling function \(\varepsilon_{a}(a)=V_{1}(x)-V_{0}(x)\). The vertical distance between the two potentials at the ground state minimum is \(\varepsilon_{0}\).
As shown in appendix A, the bosonic degrees of freedom decouple from electronic degrees of freedom in the wide band limit and the retarded Green function thus becomes a product of an electronic part given by and an bosonic part. Since the interacting term is linear in the oscillator coordinate the oscillator part of the Hamiltonian can be diagonalized by "completing the square" Or, equivalently, performing a canonical transformation which shifts the boson coordinate an amount proportional to \(c_{a}^{\dagger}c_{a}\):
\[H\to e^{iP}He^{-iP},\qquad P=-i\frac{\lambda_{1}}{\hbar\omega_{0}}c_{a}^{ \dagger}c_{a}(a^{\dagger}-a). \tag{11}\]
The retarded Green function can then be calculated exactly for the \(n\)'th excited state giving
\[G_{R}^{}(n;t)= -i\theta(t)e^{(-i\omega_{0}-\Gamma/2)t/\hbar}e^{-g_{1}(1-i\omega_ {0}t-e^{-i\omega_{0}t})}\] \[\times L_{n}\big{[}g_{1}|1-e^{i\omega_{0}t}|^{2}\big{]}, \tag{12}\]
In this paper \(g_{n}\) denotes a dimensionless effective coupling constant and \(G^{(n)}\) denotes the _exact_ Green function corresponding to a coupling term \(\varepsilon_{a}(x)\sim x^{n}\) and _not_ the contribution from an \(n\)'th order perturbative calculation as is sometimes custom.
\[A_{0}^{}(\omega) =\Gamma e^{-g_{1}} \tag{13}\] \[\times\sum_{m=0}^{\infty}\frac{g_{1}^{m}}{m!}\frac{1}{(\hbar \omega-\varepsilon_{0}+(g_{1}-m)\hbar\omega_{0})^{2}+(\Gamma/2)^{2}}.\]
The spectral function is thus a sum of Lorentzians of width \(\Gamma\) and an internal spacing of \(\omega_{0}\) and the amplitude of the \(m\)'th peak follows a Poisson distribution. It should be noted that the peaks do not represent excited states of the oscillator. It is the spectral function of the resonant electron with the oscillator in the ground state and the different peaks show that the coupling term mixes the eigenstates of the isolated oscillator. The real part of the self energy is always negative and given by \(-\hbar\omega_{0}g_{1}\) and all physical observables are invariant to \(\lambda_{1}\rightarrow-\lambda_{1}\) since linear coupling corresponds to a shifted harmonic oscillator and the direction of the shift is irrelevant.
The two-particle Green function and inelastic scattering matrix can also be calculated exactly and the probability that an incoming electron scatters on the resonance and excites the oscillator from the ground state to the \(n\)'th excited state is
\[P_{n}^{}(\varepsilon)=\Gamma^{2}e^{-2g_{1}}\frac{g_{1}^{n}}{n!}|F_{n}^{ }(\varepsilon)|^{2}, \tag{14}\]
with
\[F_{n}^{}(\varepsilon)= \sum_{k=0}^{n}(-1)^{j}\binom{n}{j}\] \[\times \sum_{l=0}^{\infty}\frac{g_{1}^{l}}{l!}\frac{1}{\varepsilon- \varepsilon_{0}+(g_{1}-j-l)\hbar\omega_{0}+i\Gamma/2}.\]
The probability of exciting the \(n\)'th vibrational state thus essentially conserves the Poisson distribution, but the Lorentzians are replaced by the interference factor \(|F_{n}(\varepsilon)|^{2}\). The results and can also be obtained using a disentangling theorem as shown in appendix B.
### Quadratic coupling
We will now consider an quadratic excited state potential energy surface \(V_{1}(x)\) which has a minimum that coincides with the ground state minimum, but has a different harmonic evolution. The potentials and coupling function corresponding to this is shown in Alternatively we could regard this model as a second order Taylor expansion of the phonon coupling function \(\varepsilon_{a}(x)\) when the first order contribution vanishes.
\[H_{I}=\lambda_{2}c_{a}^{\dagger}c_{a}(a^{\dagger}+a)^{2}, \tag{15}\]
with
\[\lambda_{2}=\frac{\hbar}{2m\omega_{0}}\frac{1}{2}\frac{\partial^{2}(V_{1}-V_ {0})}{\partial x^{2}}\Big{|}_{x=x_{0}}. \tag{16}\]
In this work we will only consider bound excited state potentials of the form \(V_{1}(x)=m\omega_{1}^{2}(x-x^{0})^{2}/2\) and we can then write
\[\lambda_{2}=\frac{(\hbar\omega_{1})^{2}-(\hbar\omega_{0})^{2}}{4\hbar\omega_{0 }},\qquad\omega_{1}=\omega_{0}\sqrt{1+4\lambda_{2}/\hbar\omega_{0}}. \tag{17}\]
In the wide band limit the electronic and bosonic degrees of freedom decouple and the boson propagator can be evaluated using a generalization of the Baker-Campbell-Hausdorff formula. As a result the retarded Green function and spectral function can be calculated exactly.
A frequency shifted excited state gives rise to a quadratic coupling function: \(\varepsilon_{a}(x)\sim x^{2}\).
The derivation is shown in appendix C and gives for the ground state:
\[A_{0}^{}(\omega)=\Gamma\sqrt{1-g_{2}} \tag{18}\] \[\times\sum_{m=0}^{\infty}\frac{b_{m}g_{2}^{m}}{(\hbar\omega- \varepsilon_{0}+\hbar(\omega_{0}-\omega_{1})/2-2m\hbar\omega_{1})^{2}+(\Gamma/ 2)^{2}},\]
where
\[g_{2}=\Big{(}\frac{\omega_{0}-\omega_{1}}{\omega_{0}+\omega_{1}}\Big{)}^{2}, \quad b_{m}=\frac{1}{m!}\frac{\partial^{m}}{\partial x^{m}}(1-x)^{-1/2}\Big{|} _{x=0}.\]
This result is valid for \(\omega_{1}>0\) which implies that \(g_{2}<1\). Again, the spectral function is a sum of Lorentzians, but with the \(m\)'th peak damped by a factor of \(b_{m}g_{2}^{m}\) instead of a Poisson distribution. The real part of the self energy is now given by half the frequency shift. The internal spacing between the peaks is \(2\omega_{1}\) and we see that the quadratic coupling only mixes the oscillator ground state with the even excited states of \(V_{1}(x)\). This is due to the mirror symmetry of \(\varepsilon_{a}(x)\) which implies that only oscillator states with equal parity mix. In the quadratic case, the effective dimensionless coupling \(g_{2}\) is not simply given by \(\lambda_{2}^{2}/(\hbar\omega_{0})^{2}\) as may have been anticipated, and thus the \(m\)'th term in does not correspond to a \(m\)'th order perturbative calculation of \(G_{R}^{}(t)\) in \(\lambda_{2}\). The calculation leading to the exact result is very different from the perturbative approach and we have checked that the second order Taylor expansion of indeed gives the result obtained from second order perturbation theory.
Allthough the spectral function shows a series of peaks spaced by \(2\omega_{1}\) it is only possible to excite an integer number of \(\omega_{0}\) through inelastic scattering. The reason is of course that the boson field is completely decoupled from the electronic states in the asymptotics of a scattering event and will thus be observed in an free oscillator eigenstate. Again, the symmetry of the quadratic coupling means that transitions involving an uneven number of bosons are forbidden.
\[F_{n}^{}(\varepsilon)=\sum_{j=0}^{n}(-1)^{j}\binom{n}{j}\sum_ {k=0}^{\infty}\sum_{l=0}^{\infty}\frac{b_{l}g_{2}^{k+l}(n+k-1)!}{k!(n-1)!}\] \[\times\frac{1}{\varepsilon-\varepsilon_{0}+(\hbar\omega_{0}- \hbar\omega_{1})/2-2(j+k+l)\hbar\omega_{1}+i\Gamma/2}.\]
The structure is very similar to the case of linear coupling. With linear coupling the probability for an electron to create \(n\) bosons are proportional to the \(n\)'th order Taylor expansion of \(e^{g_{1}}\) and normalized by \(e^{-2g_{1}}\) whereas in the quadratic case the probability to create \(2n\) bosons are proportional to the \(n\)'th order Taylor expansion of \((1-g_{2})^{-1/2}\) and normalized by \((1-g_{2})\). In the context of EELS and plasmon excitations, one would now observe a series of peaks spaced by \(2\omega_{0}\). If the plasma frequency is not known the spacing itself cannot give clues to whether linear or quadratic coupling governs the transitions, but one could use the relative amplitude between peaks since these follow a Poisson distribution if linear coupling dominates and the distribution \(b_{n}g_{2}^{n}\) if quadratic coupling dominates. If both linear and quadratic coupling is present one would observe a coupling dependent enhancement of every second peak.
In a model with linear coupling, the probability of exciting \(2n\) vibrational quanta is proportional to \(g_{1}^{2n}\) whereas it is proportional to \(g_{2}^{n}\) in a quadratic coupled model. This implies that if \(g_{2}>g_{1}^{2}\) a quadratic coupling term will give rise to larger inelastic scattering probabilities than a linear term. Even with \(g_{2}<g_{1}^{2}\) a quadratic coupling term may have stronger effect for large \(n\) since the expansion coefficients of \((1-x)^{-1/2}\) decay slower than those of \(e^{x}\). This is illustrated in figure 4, where the probability of transferring \(n\) vibrational quanta to the ground state is shown for linear and quadratic coupling.
Probabilities of making the transition \(0\to n\) through resonant inelastic scattering with linear and quadratic coupling. The parameters are \(\Gamma/\hbar\omega_{0}=0.5\), \(g_{1}=0.2\), and \(\omega_{1}=0.75\omega_{0}\) (\(g_{2}=0.02\)). Even though \(g_{2}<g_{1}^{2}\) the quadratic coupling becomes dominating for large \(n\) due to the slowly decaying expansion coefficients. One should also note the spacing between peaks which is \(\omega_{0}\) for linear coupling and \(2\omega_{1}\) for quadratic coupling. The centers of the probability distributions are approximately shifted by \(n\omega_{0}/2\) for the linear coupling and \(n\omega_{1}/2\) for the quadratic coupling relative to the bare resonance energy \(\varepsilon_{0}\), since this is where the binomial coefficients in and have their maxima.
## III Application to Hot Electron Mediated Desorption
As an example of a system where the dynamics can be approximated by a local polaron model, we consider the problem of hot electron mediated energy transfer on a metal surface. Such an energy transfer can lead to desorption of adsorbed molecules or induce chemical reactions which cannot proceed by thermal heating.. The conceptual picture of the process is the following: Hot electrons are generated in the metal by means of an MIM device or a femtosecond laser. The hot electrons may then interact with a chemisorped molecule by tunneling from the metal to an unoccupied molecular state and excite vibrational states in the molecule. If enough energy is transferred to the molecule either by a single or multiple scattering events, the molecule may eventually desorp or break an internal chemical bond. As a particular example we will calculate transition probabilities for CO adsorbed on Cu.
To calculate inelastic scattering probabilities within the local polaron model we need to obtain the coupling function \(\varepsilon_{a}(x)\). As described in section II.2, we can fix the molecule at different positions and calculate the potential energy surfaces \(V_{1}(x)\) and \(V_{0}(x)\) at each point and \(\varepsilon_{a}(x)=V_{1}(x)-V_{0}(x)\). The model does not directly contain Coulomb interactions between electrons, but these are included in the calculation of \(\varepsilon_{a}(x)\) which thus becomes an effective coupling that is supposed to contain all the electronic interaction associated with the excited state of the molecule.
The potential energies \(V_{1}(x)\) and \(V_{0}(x)\) has been obtained using the code gpaw which is a real-space Density Functional Theory (DFT) code that uses the projector augmented wave method. In all our calculations we used the Revised Perdew-Burke-Ernzerhof (RPBE) exchange-correlation functional since this has been designed to perform well for molecules adsorbed on surfaces, and has been shown to perform better than the original PBE functional for adsorbed molecules.
We set up a Cu surface consisting of three atomic layers with the top layer being relaxed. 10 A of vacuum has then been introduced above the slab and 0.50 monolayer of adsorbate molecules relaxed at top sites which is the preferred adsorption site. Both molecules adsorb with their molecular axis perpendicular to the surface with O pointing away from the surface. We then did a normal mode analysis and mapped out the three ground state potential energy functions \(V_{0}(x_{i})\) corresponding to the two normal modes that involve the perpendicular degrees of freedom and a frustrated rotation. The perpendicular modes roughly correspond to an internal stretch \(d=x_{O}-x_{C(N)}\) and center of mass \(z=(m_{O}x_{O}+m_{C(N)}x_{C(N)})/(m_{O}+m_{C(N)})\). We do not include the three remaining molecular modes since one is another frustrated rotation with identical properties to the one considered, and the two frustrated translations are only weakly coupled to the resonant electron and are not expected to play a significant role in the femtochemistry. In all calculations we use a p(2x2) cell, sample 12 irreducible k-points in the surface plane, and use a grid spacing of 0.2 A.
To find the excited state potential energies \(V_{1}(x_{i})\) corresponding to the three normal modes of interest, we have used the method of linear expansion \(\Delta\)SCF which has been published in a previous work and implemented in gpaw. In the previous publication we have tested the method against inverse photo-emission spectroscopy, and found that it performed well for molecules chemisorped on surfaces. In each step of the self consistency cycle an electron is removed from the Fermi level, the density of an excited state is added to the total density, and the band energy of this state is added to the total energy. To get the band energy right we need to expand the excited state on the Kohn-Sham orbitals found in each iteration. The method is thus a generalization of the usual \(\Delta\)SCF where occupation numbers are changed. Instead of changing occupation numbers we occupy an orbital which is not an eigenstate of the Kohn-Sham Hamiltonian but a superposition of eigenstates in such a way that the state is as close as possible to the original molecular state. In the present case the excited state is the anti-bonding \(2\pi\) orbital of CO. In we show the ground and excited state potential energy surfaces corresponding to the frustrated rotation along with \(\varepsilon_{a}(x)\). It is clear that the excited state potential is not exactly a quadratic potential and the parameter \(\lambda_{2}\) which we need to calculate transition probabilities will depend on how we fit this potential to a quadratic form. However the width of the Gaussian ground state vibrational wavefunction corresponds to \(x=0.08\) A and for low lying excitations we can thus use this region of the potential which is rather flat.
Potential energy surfaces along the frustrated rotation mode of CO adsorbed on a Cu surface. The coordinate \(x\) is a generalized coordinate representing the deviation from equilibrium. \(x=0.4\) corresponds to a \(24^{\circ}\) angular deviation from the perpendicular position.
For both perpendicular modes we find that \(\varepsilon_{a}(x_{i})\sim x_{i}\) and quadratic coupling can thus be neglected. In contrast, due to symmetry the excited state potential energy of frustrated rotation is invariant to \(x_{i}\rightarrow-x_{i}\) and the linear coupling term thus vanishes. We have calculated the excitation energy to \(\varepsilon_{a}(x_{0})=2.8\ eV\) and the resonance width is estimated from the Kohn-Sham projected density of states to \(\Gamma\approx 1.0\ eV\). In table 1, we display the calculated parameters corresponding to the three modes.
We note that when calculating transition probabilities we should include all modes in the model, because even if the modes are not coupled directly they have an indirect coupling since they all interact with the resonance. It is possible to obtain expressions for the scattering matrix including more than one mode, but these are rather complicated to handle and for weakly coupled systems the physics can usually be extracted from three one-mode models. In we show the calculated probabilities for a hot electron to excite the different modes of CO adsorbed on Cu. The internal stretch and and frustrated rotation show transition probabilities on the same order of magnitude whereas the center of mass vibrations are very unlikely to get excited. This is in accord with calculations of the electronic friction coefficients of this system which is very closely related to the coupling function \(\varepsilon_{a}(x)\). The frequency of internal vibration is five times larger than both the center of mass and frustrated rotation frequencies and as previously shown the stretch mode will completely dominate the total energy transfer. Thus, in a simple model where hot electron mediated desorption is reduced to calculating the probability of transferring the chemisorption energy to the adsorbate, the internal mode governs the desorption probability. Nevertheless, our estimate of \(\Gamma\) is based on the Kohn-Sham density of states which may give a poor description of the electronic spectral function \(A^{0}(\omega)\). If \(\Gamma\) is significantly smaller than our estimate, the quadratically coupled frustrated rotation will play an important role in hot electron mediated desorption for this system.
## IV Summary and discussion
We have calculated the spectral function and inelastic scattering amplitudes in a local polaron model with quadratic coupling to bosons. The probability of exciting \(n\) bosons is found to be damped by a distribution function given by the \(n\)'th Taylor expansion of \(1/\sqrt{1-g_{2}}\) which decays much slower than the Poisson distribution appearing in a linearly coupled model. Hence for comparable values of linear and quadratic coupling constants, a quadratic term will dominate inelastic scattering probabilities involving a large number of bosonic excitations.
As an application we have considered the problem of hot electron mediated vibrational excitations of molecules adsorbed on metal surfaces. The coupling constants were calculated from the excitation energy along the molecular normal modes using delta self-consistent field DFT. It was found that quadratic coupling is important for exciting the frustrated rotations since this mode does not couple linearly due to symmetry.
A major approximation in the model is the quadratic assumption for the ground state potential. In our numerical example with HEFatS it is clear from that the potentials is not exactly quadratic. For the center of mass mode the situation is even worse and a Morse potential is much better suited to describe this mode. The anharmonic deviations are likely to have a significant effect on high lying excited states, but renders the model much more complicated. In fact, since the coupling to the internal stretch mode seems to govern the rate of energy transfer, one has to assume that the energy is readily redistributed to other degrees of freedom and anharmonic coupling is thus expected to play a vital role in the actual desorption process.
The wide band limit has been essential in the deriva
\begin{table}
\begin{tabular}{c|c|c|c} Mode & \(\hbar\omega\) & \(\lambda_{1}\) & \(\lambda_{2}\) \\ \hline Frustrated rotation & 0.037 & 0 & -0.009 \\ Center of mass & 0.043 & -0.006 & \(\sim 0\) \\ Internal stretch & 0.248 & -0.170 & \(\sim 0\) \\ \end{tabular}
\end{table}
Table 1: Parameters for CO adsorbed on Cu. All number are \(eV\). Note that while the quadratic coupling for the two perpendicular modes are very small and thus neglectable, the linear coupling of frustrated rotation vanishes exactly due to symmetry.
Probabilities of exciting two and four quanta of vibrations to the center of mass, internal stretch and frustrated rotation modes CO adsorbed on Cu.
However, it would be very interesting to do perturbation theory with the general retarded Green function to examine the effect of energy dependence in the electronic self energy.
|
10.48550/arXiv.0903.1786
|
Inelastic scattering in a local polaron model with quadratic coupling to bosons
|
Thomas Olsen
| 1,354
|
10.48550_arXiv.0902.3559
|
## 1 Introduction
Although the Density Functional Theory (DFT) based methods of modeling electronic structure of molecules and solids widely proliferate during last decades, the problem of consistent description of transition metal and rare earth compounds with open \(d\)- and \(f\)-shells, respectively, remains a still unresolved, challenging problem in this framework. One of the main reasons for this failure of the DFT is that the multiplet spin/orbital momentum states are generally not easily described within the DFT paradigm. The source of that intimate "unfriendlieness" of the DFT to the multiplet states lays in the "oversymmetry" of the fundamental quantity pertaining in the realm of DFT: the one-electron density. As it has been demonstrated many times, states of different total spin and/or spatial symmetry may produce equal one-electron densities. The complication arising from this is the impossibility to distinguish the nature of the ground state on the basis of the total density only: although only one of say two functions represents the ground state, i.e. the _exact_ energies of the involved states may be different, they turn out to be degenerate in the DFT context. In other terms, if the same densities are fed to the "universal" density functional implied by the DFT, it is going to produce the same value of the electronic energy for states whose _exact_ energies are _different_. Of course the latter remark may be opposed by noting that the "universal" functional is going to output the ground state energy only, but in this case it is not clear how other importantpieces of information concerning the nature of this ground state (e.g. its spin multiplicity) can be extracted from such an answer.
This situation certainly requires some clarification which is addressed in the present paper. In order to do so we give below a brief description of relevant elements of the electronic structure theory (Section 2). Then we consider an archetypical example of the problems encountered by the DFT while trying to reproduce correct spin properties of many-electron systems (Section 3). Then we propose a general scheme allowing to include states of definite spin in the DFT theory (Section 4). This, however, does not solve the problem of the multiple states in the open \(d\)- and \(f\)-shells of the transition and rare earth ions. For this end we explore in Section 6 the possibility to circumvent these problems with use of the short/long range separation of Coulomb interaction between electrons and propose in Section 5 some conceivable state-dependent definition of exchange-correlation functionals capable to reproduce the energies of nontrivially correlated many-electronic states in the \(d\)-shell of the Fe\({}^{2+}\) ions.
## 2 Theoretical background
### Electronic distribution.
The main idea of the DFT is to reduce the description of entire electronic structure to a single quantity: the one-electron density - the diagonal part of the one-electron density matrix. The possibility of such a reduction is proven by the Hohenberg-Kohn theorems which state an existence of a universal one-to-one correspondence between one-electron external potential and the one-electron density in that sense that not only the one-electron potential acting upon a given number of electrons uniquely defines the ground state of such a system _i.e._ its wave function and thus the one-electron density, but also that for each given density integrating to a given number of electrons \(N\) a one-electron potential yielding that given density can be uniquely defined from the density. The "density only" formulation of the electronic structure problem, even if it is practically achieved, leaves unanswered an important question of the nature of the ground state thus obtained _e.g._ about its total spin (or other symmetry features).
Incidentally, the symmetry properties of quantum states, like total spin, are easier expressed in terms of wave functions (see Ref.) so it would be practical to consider tentative relation between the wave function and density only pictures of the electronic structure. The required relation can be established with use of the reduced one- and two-electron density _matrices_ as much simpler objects than the wave functions, providing equivalent description of electronic structure. The reduced density matrices respectively depend on two \((x,x^{\prime})\) and four \((x_{1}x_{2},x^{\prime}_{1}x^{\prime}_{2})\) coordinates:
\[\begin{array}{lcl}\rho^{}(x;x^{\prime})&=&N\int\Psi^{*}(x,x_{2},\ldots x_{N })\times\\ &\times&\Psi(x^{\prime},x_{2},\ldots,x_{N})dx_{2}\ldots dx_{N},\\ \rho^{}(x_{1}x_{2};x^{\prime}_{1}x^{\prime}_{2})&=&\frac{N(N-1)}{2}\int\Psi^ {*}(x_{1},x_{2},x_{3},\ldots x_{N})\times\\ &\times&\Psi(x^{\prime}_{1},x^{\prime}_{2},x_{3},\ldots,x_{N})dx_{3}\ldots dx_{ N},\end{array} \tag{1}\]
The transition to the description in terms of reduced density matrices is itself a significant simplification (although being absolutely exact). The one-electron density implied by the DFT theory appears then as a result of further reduction of eq.:
\[\rho({\bf r})=\sum_{s}\rho^{}({\bf r}s;{\bf r}s) \tag{2}\]
Thus according to the DFT paradigm the one-electron density which depends on one spatial radius-vector must be able to serve as an equivalent substitute to the \(N\)-electronic wave function dependent on \(N\) radius vectors and \(N\) more spin projections of all electrons involved. The obvious loss of information which takes place by going from eq. to eq. - we remind that going from the wave function \(\Psi(x_{1},x_{2},x_{3},\ldots x_{N})\) to the reduced density matrices by eq. does not produce any loss of at least important information - must be compensated by the "universal" and "exact" density functional which is in general unknown.
### Electronic energy.
Leaving aside the "ideal" DFT using the unknown "universal" and "exact" functional of the density eq. and turning to pragmatic methods pertaining to the DFT realm needs some approximate expressions for the energy presented as a functional of the density eq. only. In the wave function and in the equivalent reduced density matrix formulations the energy has the form:
\[E = \langle\hat{T}_{e}\rangle+\langle\hat{V}_{ne}(\{{\bf R}\})\rangle+ \langle\hat{V}_{ee}\rangle+V_{nn}(\{{\bf R}\}). \tag{3}\]
In the coordinate representation the above averages acquire familiar appearance:
\[\begin{array}{l}\left\langle\hat{T}_{e}\right\rangle=-\frac{1}{2}\sum_{s} \int\limits_{{\bf r}={\bf r}^{\prime}}\Delta^{\prime}\rho^{}({\bf r}s;{\bf r }^{\prime}s)d{\bf r}\\ \left\langle\hat{V}_{ne}(\{{\bf R}\})\right\rangle=\sum_{i}Z_{i} \int\frac{\rho({\bf r})d{\bf r}}{|{\bf R}_{i}-{\bf r}|}\\ \left\langle\hat{V}_{ee}\right\rangle=\frac{1}{2}\sum_{ss^{\prime}} \int\int\frac{\rho^{}({\bf r}s,{\bf r}^{\prime}s^{\prime};{\bf r}s,{\bf r} ^{\prime}s^{\prime})}{|{\bf r}-{\bf r}^{\prime}|}d{\bf r}d{\bf r}^{\prime}\\ V_{nn}(\{{\bf R}\})=\frac{1}{2}\sum_{i\neq j}\frac{Z_{i}Z_{j}}{ |{\bf R}_{i}-{\bf R}_{j}|};\mbox{where}\\ \Delta^{\prime}=\frac{\partial^{2}}{\partial x^{\prime 2}}+\frac{\partial^{2}}{ \partial y^{\prime 2}}+\frac{\partial^{2}}{\partial z^{\prime 2}}\end{array} \tag{4}\]
are assumed to be specific for a given geometry \(\{{\bf R}\}\) and for an electronic state described by the \(N\)-electronic wave function \(\Psi=\Psi\left(x_{1},...,x_{N}\right)\) employed to define the density matrices eq.. The first row in eq. is the kinetic energy of electrons, the second row is the energy of Coulomb attraction of electrons to nuclei, the third row is the energy of interelectronic repulsion; the fourth one is the energy of Coulomb repulsion of the nuclei, which does not depend on the electronic density/wavefunction.
In the above expressions eqs, only the average of the nuclear potential \(\hat{V}_{ne}\) is _exactly_ a functional of the required form: that of the one-electron density eq.. All other terms in eqs, require further consideration. It applies similarly to the remaining one- and two-electron contributions to the energy. As for the one-electron term, the kinetic energy requires knowledge of the one-electron density _matrix_ eq. rather than its diagonal part eq. although effectively in a narrow range of spatial separations \({\bf r}-{\bf r}^{\prime}\) which must be sufficient to determine the second derivative. The attempts to avoid this bottleneck and to obtain pragmatic DFT methods brought Kohn and Sham to their famous orbital construct which allowed them to express the kinetic energy in terms of some single-determinant wave function yielding by definition the required (exact) one-electron density. Then the kinetic energy is calculated as one of the system of non-interacting electrons described by a single determinant built of KS orbitals.
### Electronic density and electronic energy decompositions.
While the Kohn-Sham construct offers an efficient technique to handle the difficult kinetic energy problem and provide a very good first approximation to it, the representation of the electron-electron interaction energy in terms of the one-electron density (and possibly further parameters derived from the KS determinant) remains the central problem on modern density functional theory. Generally, calculating the Coulomb electron-electron energy (3-rd row of eq.) requires knowledge of the two-electron density matrix. According to it decomposes:
\[\rho^{}(x_{1},x_{2};x_{1}^{\prime},x_{2}^{\prime})=\frac{1}{2}\left|\begin{array}[ ]{cc}\rho^{}(x_{1};x_{1}^{\prime})&\rho^{}(x_{2};x_{1}^{\prime})\\ \rho^{}(x_{1};x_{2}^{\prime})&\rho^{}(x_{2};x_{2}^{\prime})\end{array} \right|-\chi(x_{1},x_{2};x_{1}^{\prime},x_{2}^{\prime}), \tag{5}\]
The second term in eq. - the cumulant of the two-particle density matrix - is responsible for deviation of electrons' behavior from the independent electron model, _i.e._ for their Coulomb correlations. The Coulomb interaction of electrons eq. can be decomposed to contributions associated to the terms of the above two-particle density matrix decomposition eq.:
\[\langle V_{ee}\rangle = E_{H}+\overline{E}_{xc}; \tag{6}\] \[\overline{E}_{xc} = \overline{E}_{x}+\overline{E}_{c}.\]
by singling out first the "classical" part of the Coulomb interaction energy - the Hartree energy:
\[E_{H}=\frac{1}{2}\sum_{ss^{\prime}}\int\frac{\rho^{}(\mathbf{r}s,\mathbf{r} s)\rho^{}(\mathbf{r}^{\prime}s^{\prime},\mathbf{r}^{\prime}s^{\prime})}{| \mathbf{r}-\mathbf{r}^{\prime}|}d\mathbf{r}d\mathbf{r}^{\prime}=\frac{1}{2}\int \frac{\rho(\mathbf{r})\rho(\mathbf{r}^{\prime})}{|\mathbf{r}-\mathbf{r}^{ \prime}|}d\mathbf{r}d\mathbf{r}^{\prime} \tag{7}\]and then the exchange and correlation energies:
\[\overline{E}_{x} = -\frac{1}{2}\sum_{s}\int\frac{\rho^{}({\bf r}s,{\bf r}^{\prime}s) \rho^{}({\bf r}^{\prime}s,{\bf r}s)}{|{\bf r}-{\bf r}^{\prime}|}d{\bf r}d{ \bf r}^{\prime} \tag{8}\] \[\overline{E}_{c} = -\frac{1}{2}\sum_{ss^{\prime}}\int\frac{\chi({\bf r}s,{\bf r}^{ \prime}s^{\prime};{\bf r}s,{\bf r}^{\prime}s^{\prime})}{|{\bf r}-{\bf r}^{ \prime}|}d{\bf r}d{\bf r}^{\prime} \tag{9}\]
While the definition of the Hartree-energy is unique, and constitutes together with the nuclear-electron repulsion energy the part of the total energy that can be written straightforwardly as a simple analytic functional of the one-particle density, the exchange and correlation energies are defined in quantum chemistry and in DFT in different ways. As far as the exchange energy is concerned, one should remark that the one-particle density matrix, appearing in eq. is supposed to be exact. This quantity is not available even in exact KS theory, where we have at best the one-particle density matrix associated to the single determinant constructed from the exact KS orbitals. By consequence, the exact exchange energy in DFT is in general not equal to \(\overline{E}_{x}\).
It must be observed that the usual definition of the correlation energy in quantum chemistry, proposed by Lowdin in Ref. differs from that given in Eq., which follows rather the suggestion due to Kutzelnigg and Mukherjee. This latter definition has the conceptual advantage that it uses the quantities entering eqs. - irrespective to any approximate method of calculation of the electronic energy. Some authors, (cf. _e.g._ Refs.) argue that pragmatic DFT methods can be considered as approximations to the two-electron density matrix cumulant.
This situation is as well slightly more complicated in conventional Kohn-Sham theory, where the correlation energy involves also the difference of the exact and KS kinetic energies. However, this kinetic energy contribution can be assimilated to a potential energy term the virtue of the adiabatic connection procedure, which allows one to write the total correlation energy as an average electron-electron interaction over the adiabatic connection path.
Symmetry non-sensitivity of the density-only methods
### Archetypical example of existing problems
In order to better understand the problems which appear in the DFT realm when trying to describe the correct total spin of a many electronic state we consider the simplest system of two electrons occupying spatial orbitals \(|a\rangle\) and \(|b\rangle\) (which can be understood as notation for one-dimensional irreducible representations of a point group) and forming corresponding singlet and triplet states \({}^{1}B\) and \({}^{3}B\). The relevant wave functions in the coordinate representation are given by:
\[\begin{array}{l}\Psi_{{}^{1}B}(x_{1},x_{2})=\frac{1}{2}\left(a({\bf r}_{1})b ({\bf r}_{2})+b({\bf r}_{1})a({\bf r}_{2})\right)\left(\alpha(s_{1})\beta(s_{ 2})-\beta(s_{1})\alpha(s_{2})\right),\\ \Psi_{{}^{3}B}(x_{1},x_{2})=\frac{1}{2}\left(a({\bf r}_{1})b({\bf r}_{2})-b({ \bf r}_{1})a({\bf r}_{2})\right)\left(\alpha(s_{1})\beta(s_{2})+\beta(s_{1}) \alpha(s_{2})\right),\end{array} \tag{10}\]
Following the definitions of the one-electron density matrices eq. the states eq. immediately yield _exactly_ the same one-electron density _matrix_:
\[\rho_{{}^{2S+1}B}^{}(x,x^{\prime})=\frac{1}{2}\left(\alpha^{*}(s)\alpha(s^{ \prime})+\beta^{*}(s)\beta(s^{\prime})\right)\left(a^{*}({\bf r})a({\bf r}^{ \prime})+b^{*}({\bf r})b({\bf r}^{\prime})\right) \tag{11}\]
This result is well known for decades and appears even in textbooks. Obviously the density eq. which is required by the DFT is as well the same for the two spin states.
The _exact_ two electron density matrices calculated according to their definition eq. from the wave functions eq. _are_, however, different:
\[\begin{array}{l}\rho_{{}^{1,3}B}^{}(x_{1}x_{2},x_{1}^{\prime}x_{2}^{\prime })=\\ \frac{1}{4}\left(\alpha^{*}(s_{1})\beta^{*}(s_{2})\mp\beta^{*}(s_{1})\alpha^{*} (s_{2})\right)\left(\alpha(s_{1}^{\prime})\beta(s_{2}^{\prime})\mp\beta(s_{1}^ {\prime})\alpha(s_{2}^{\prime})\right)\times\\ \left(a^{*}({\bf r}_{1})b^{*}({\bf r}_{2})\pm b^{*}({\bf r}_{1})a^{*}({\bf r}_{2 })\right)\left(a({\bf r}_{1}^{\prime})b({\bf r}_{2}^{\prime})\pm b({\bf r}_{1} ^{\prime})a({\bf r}_{2}^{\prime})\right)\end{array}\]
Comparing the above expression with the decomposition eq. one easily sees that only the _cumulant_ of the two-electron density matrix can be responsible for the distinguishing of the two-electron density matrices for the singlet and triplet states.
The "oversymmetry" of the density (and even of the first order density matrix) with respect to the total spin exemplified by eq. is not accidental, but is a consequence of a very general result (see Ref. and references therein). Even a higher symmetry can be proven. In its modern formulation (Theorem 1 of Ref.) it reads: 'The electron density of an arbitrary \(N\)-electron system, characterized by the \(N\)-electron wave function corresponding to the total spin \(S\) and constructed on some orthonormal orbital set, does not depend upon the total spin \(S\) of the state and always preserves the same form as it is for a single-determinant wave function'. The proof given in Ref. relies not upon the spin properties themselves rather on the manifestation of permutation symmetry of the exact wave function in terms of the total spin. We address this issue later in Section 4.2.1.
### Methods proposed to treat coinciding densities
#### 3.2.1 Multiplet sum method
The first attempt to get around this problem of coinciding densities in the DFT context dates back to the work Ref.. The analysis of problems performed there is precisely repeated in the above two-electron two-orbital model. The prescription Ref. concerning the way out reads as follows: to evaluate correctly the energy of the singlet and triplet states \({}^{1}B\) and \({}^{3}B\) in terms of the quantities which can be obtained with use of single determinant wave functions. To do so one has to address the single determinant function \(|a\alpha b\beta|\) which is not a pure spin state, but in fact is a linear combination of two above spin states:
\[|a\alpha b\beta|=\frac{1}{\sqrt{2}}\left(\left|{}^{1}B,S_{z}=0\right\rangle+ \left|{}^{3}B,S_{z}=0\right\rangle\right) \tag{12}\]
Averaging the Hamiltonian over the linear combination eq. of the pure spin states immediately yields:
\[\frac{1}{2}E(^{1}B)+\frac{1}{2}E(^{3}B) \tag{13}\]
The energy of the triplet state entering the above combination can be independently extracted from another single determinant wave function: \(|a\alpha b\alpha|\) corresponding to the component of the triplet with the spin projection \(+1\). Thus one can express the energy of the (non-single-determinant) singlet state linearly combining the averages of the Hamiltonian over the single determinant states of which, however, one belongs to the spin projection \(+1\). Obviously the above move was only possible because the off-diagonal matrix element of the Hamiltonian between the singlet and triplet contributions to the determinant of interest vanishes due to the spin symmetry. The different expressions for \(E(^{1}B)\) and \(E(^{3}B)\) thus obtained are then treated as required _distinct_ energy functionals to be used to calculate the energy respectively for the singlet and triplet states possessing the same one-electron density. It is instructive to check (and in this simple case it can be done by direct evaluation) where the difference between the energy expressions comes from. Inserting the one-electron density matrix eq. which is the same for both spin states in the definitions of the Hartree and exchange energies eqs., yields for the both spin states equal Hartree and exchange contributions:
\[\begin{array}{ll}{\rm Hartree}&\frac{1}{2}\left[(aa|aa)+(aa|bb)+(bb|aa)+(bb |bb)\right];\\ {\rm exchange}&\frac{1}{2}\left[(aa|aa)+(ab|ba)+(ba|ab)+(bb|bb)\right].\end{array} \tag{14}\]
One can see that (i) the self interaction terms in the Hartree contribution are precisely cancelled by the corresponding terms in the exchange part; (ii) at the same time, obviously, there is no other source where the difference between the spin state energies could come from except the cumulant of the two-electron density matrix and thus the correlation energy as defined by eq. is responsible for the difference in the resulting expressions:
\[\begin{array}{ll}E(^{1}B)&=(aa|bb)+(ab|ba)\\ E(^{3}B)&=(aa|bb)-(ab|ba)\end{array} \tag{15}\]
On the other hand one may notice that the classification of the energy contributions as exchange or correlation ones by eqs., is in some way arbitrary as well. Indeed, for the above model the energy of the triplet state with the spin projection \(+1\) is exactly the sum of the Hartree and exchange contributions since the latter state is represented by a single determinant wave function for which the cumulant precisely vanishes. However, the equal energy for the triplet state with the zero spin projection breaks down differently: into the Hartree, exchange, and correlation contributions, where the Hartree contribution is the same as in the case of the spin projection \(+1\), but only the sum of the exchange and correlation contributions is the same for different values of \(S_{z}\).
The above way leading to the energy expressions for different spin states is not completely satisfactory (and not clearly generalizable): although formally the results can be treated as functionals of the density the difference of the two energy expressions is obtained by a kind of trick. Referring to the triplet component with \(S_{z}=+1\) in the derivation of the multiplet energy looks out as an alien element (in fact the energy is uniquely determined by the spatial multipliers in the wave functions eq. - without any reference to the spin components at all). This strange element of the derivation appeared in order to compensate somehow the element of the general theory which is missing in the DFT - the cumulant of the two-electron density matrix. Despite this criticism the result of the derivation is very transparent: it reduces to deriving according to McWeeny's notice in Ref. 'of a particular type of energy expression--irrespective of the nature of the wavefunction', namely one - linear in the Coulomb and exchange two-electron integrals over the involved orbitals.
Further development of this approach is based on the _assumption_ that it is _always_ (or at least for unspecifically wide class of cases) possible to express the energy of a pure spin multiplet state allowable for a given number of electrons/orbitals as a linear combination:
\[E(n\Gamma S)=\sum_{i}w_{i}^{n\Gamma S}E_{i} \tag{16}\]
The one-electron density (matrices) corresponding to these determinants are different and the whole scheme becomes workable provided the set of the coefficients (weights) \(w_{i}^{\Gamma S}\) exists and they are uniquely defined by the spin and symmetry quantum numbers \(\Gamma\) and \(S\) and other quantum numbers \(n\) serving to distinguish potentially existing states with equal \(\Gamma\) and \(S\). Apparently only a restricted number of examples of such functional forms is known. The reason is quite simple and the above consideration allows to single out the range of cases where the derivation analogous to that of eq. can be performed. It applies if the total spin allows to completely distinguish the electronic states. In this case the energy of the state of the highest available spin \(S_{\rm max}\) can be expressed through a single determinant function (the cumulant is vanishing) with the highest available projection of the total spin. Then this result can be used to evaluate the energy of the state with \(S_{\rm max}-1\), _etc._ The recipe immediately fails as soon as multiple states of the same total spin appear in the system. This is however the everyday life, so in what follows we switch to considering further possibilities of constructing the energy functionals useful in this situation.
#### 3.2.2 Restricted open shell KS (ROKS) method
The situation with reproducing total spin dependence of the energy as it appears in the DFT context is by no means unique: the same problem arises in the Hartree-Fock-Roothaan (HFR) context since the latter lacks any adequate representation of the cumulant of the two-electron density matrix as well. Within the "extended" HFR context some ways out have been proposed. Incidentally, the method of Ref. is precisely the Slater multiplet sum method Ref. which migrated from the HFR to the DFT context. Another option is the ROHF (restricted open shell Hartree - Fock) method whose respective migration resulted in a range of the ROKS (restricted open shell Kohn-Sham procedures) being the DFT counterpart of the former. Despite different appearance they have many common features (and we do not address here the methods based on the statistical - ensemble - averaging).
The derivation of the ROHF (or equivalently 'old MC SCF'- see below) bases on the general expression of the form:
\[E(n\Gamma S)=\sum_{ij}C_{i}^{n\Gamma S}C_{j}^{n\Gamma S}H_{ij} \tag{17}\]
In this case no alien states of wrong spin projection may appear. On the other hand the contribution of the off-diagonal elements \(H_{ij}\) to the energy may be nontrivial (in contrast with eq.). The knowledge of the expansion coefficients \(C_{i}^{n\Gamma S}\) in general requires diagonalization of the Hamiltonian matrix making the expansion coefficients and thus the energy itself some sophisticated irrational function of the Hamiltonian matrix elements including two-electron integrals. It was Roothaan who first noticed that certain states \(\Psi_{n\Gamma S}\) of atoms and linear molecules, even those requiring many-determinant (multi-reference, multi-configurational) wave functions, yield energy expressions which are linear with respect to two-electron integrals \((ii|jj)\) and \((ij|ji)\) (respectively Coulomb and exchange ones). It is only possible if the wave function expansion coefficients \(C_{i}^{n\Gamma S}\) in eq. can be determined on the symmetry grounds _i.e._ without nontrivial diagonalization. In this case there is no need that the off-diagonal elements \(H_{ij}\) which are linear expressions in the two-electron integrals and thus give a linear contribution to the energy functional disappear as required by eq.. Only the possibility to have the expansion coefficients \(C_{i}^{n\Gamma S}\) independent on the specific values of the Hamiltonian matrix elements is the true prerequisite for obtaining the expressions for the energy of the required (linear) form. Nevertheless the number of cases when the described procedure was possible in fact reduces to the \(p^{n}\) states of atoms and \(\pi^{n}\) and \(\delta^{n}\) states of linear molecules. Similarly the ROKS sheme proposed in Ref. and representing a migration of the Roothaan's reasoning to the DFT context allowed to obtain the functional forms for the same set of states: \(p^{n}\), \(\pi^{n}\), and \(\delta^{n}\). Thus the forecast of the year 1960 due to Roothaan: 'It is a relatively simple matter to extend the open-shell theory just presented in such a way that other important classes of atomic states can be accommodated, as for instance, the \(d^{N}\) configurations for the transition elements. We postpone such generalizations for the present, and include whatever new treatments may be necessary with the actual applications planned for the future' never became true and the \(d^{N}\) states generally cannot be squeezed in the ROHF/ROKS scheme.
Under other angle of view, validity of the Roothaan or similar schemes means that the cumulant of the two-electron density matrix can be in some particular case recovered by symmetry based manipulations. In the cases considered by Roothaan himself and recently used in the DFT context in Ref. the possibility of obtaining closed expressions for the energy functional in terms of two-electron integrals over orbitals involved is stipulated by additional symmetry of the system (in the chemical context it goes about additional symmetry group \(G\) with irreducible representations \(\Gamma\), where \(G=SO\) for an atom \(G=SO\) for a linear molecule, and may be some point group for other molecules) which allows to figure out the expansion coefficients \(C_{i}^{n\Gamma S}\). It is clear that for an overwhelming majority of cases it is impossible to find any nontrivial symmetry group \(G\neq C_{1}\) which predefines a restricted character of any Roothaan-like treatment. It is thus our next purpose is to expore other possibilities of designing energy functionals distinguishing the states of different spin multiplicities in the DFT context.
Spin and unitary symmetry of the electronic wave function
As we mentioned previously any reference to the spin projections throughout the derivation of the energy expressions for the two-electron two-orbital model looks out as an alien element. The ultimate reason for that is that the non-relativistic Hamiltonian does not depend on spin variables at all and the energy itself as well as the differences in its form for different spin states originates solely from the spatial multiplier of the many-electronic wave function (spatial function). The idea to restrict the entire consideration by those spatial functions persists almost from the beginning of the quantum chemistry and is known as "spin-free quantum chemistry". It can be given different formulations of which we use one based on the use of the unitary group (see Ref.). We briefly remind its basic facts in the Appendix.
### Unitary symmetry of the spatial multiplier.
The construct using the permutational symmetry of the spatial part of the wave function had been used for developing the so called generalized Hartree-Fock procedure which had numerous descendants (see _e.g._). They basically performed the task of presenting the energy in the HFR-like form: linear with respect to Coulomb and exchange integrals over the orbitals involved with the coefficients dependent on the permutational symmetry of the spatial part of the wave function and thus on the total spin. The permutational symmetry, however, addresses the many-electron wave functions in the coordinate representation which is of restricted use in quantum chemistry. By contrast the wave functions actually used are those in the representation of the occupation numbers of the orbitals involved. For that reason it is more practical to switch to labeling of the many-electron functions by irreducible representations of the unitary group which are closely related to those of the \(S_{N}\) group. The corresponding construct is described in the Appendix.
### Physical quantities in terms of unitary group.
Going to the representation of the \(U(M)\) group has that advantage that it allows to easily write down the energy of many electron states. This is done as follows: for each Young pattern \(\Upsilon\) one can construct the set of generators \({\bf E}_{ij}^{\Upsilon}\) (\(ij=1\div M\)) of the group \(U(M)\) acting in the space of the irreducible representation \(\Upsilon=\Upsilon(M,N,S)\) whose matrix elements between the tableaux \(\upsilon\) and \(\upsilon^{\prime}\) can be calculated irrespective to the physical nature of the system described. The set of generators completely defines the action of the group \(U(M)\) in the irreducible subspace of its tensors of the rank \(N\) with the permutational/spin symmetry stipulated by the Young pattern \(\Upsilon\).
The diagonal generators \({\bf E}_{ii}^{\Upsilon}\) are diagonal in the basis of Young tableaux and their matrix elements are equal to the occupation number (\(n_{i}=2,1,0\)) of the \(i\)-th orbital in the Young tableau \(\Upsilon\upsilon:\)
\[\left\langle\Upsilon\upsilon\left|{\bf E}_{ii}^{\Upsilon}\right|\Upsilon \upsilon^{\prime}\right\rangle=\delta_{\upsilon\upsilon^{\prime}}\left\langle {\bf E}_{ii}^{\Upsilon}\right\rangle_{\Upsilon\upsilon}=\delta_{\upsilon \upsilon^{\prime}}n_{i}\]
By contrast off-diagonal generators \({\bf E}_{ij}^{\Upsilon}\) (raising ones if \(i>j\) and lowering ones if \(j>i\)) have non-vanishing matrix elements \(\left\langle\Upsilon\upsilon\left|{\bf E}_{ij}^{\Upsilon}\right|\Upsilon \upsilon^{\prime}\right\rangle\) if the tableau \(\upsilon^{\prime}\) contains at least one orbital symbol \(j\) whereas the tableau \(\upsilon\) contains one less orbital symbol \(j\) than \(\upsilon^{\prime}\) and one more orbital symbol \(i\) than it. From this it follows that the off-diagonal generators \({\bf E}_{ij}^{\Upsilon}(i\neq j)\) have no non-vanishing diagonal matrix elements.
#### 4.2.1 One-electron density in the unitary group formalism
The generators \({\bf E}_{ij}^{\Upsilon}\) are _by definition_ the components of the spatial one-electron density operator restricted to the subspace of the states belonging to the \(\Upsilon\) pattern (those having transformation properties of the corresponding rank \(N\) tensors with the permutational symmetry stipulated by the Young pattern \(\Upsilon\) or equivalently having the total spin prescribed by this pattern):
\[{\bf E}_{ij}^{\Upsilon}=\sum_{\sigma}P^{\Upsilon}\psi^{+}(i\sigma)\psi(j\sigma) P^{\Upsilon},\]
It is remarkable to note that the Young tableaux states have an important property similar to that of the Slater determinants: the one-electron density matrices generated from such states are diagonal in the basis of the orbitals involved in their construction.
With use of this construct one can easily check the validity of the Kaplan's Theorem 1. Indeed, for whatever Young tableau \(\Upsilon\upsilon\) the one electron density pertinent to the corresponding \(N\)-electron state reads:
\[\rho_{\Upsilon\upsilon}({\bf r},{\bf r}^{\prime}) = \sum_{s}\rho_{\Upsilon\upsilon}^{}({\bf r}s,{\bf r}^{\prime}s)= \sum_{s}\left\langle\Upsilon\upsilon\left|\psi^{+}({\bf r}s)\psi({\bf r}^{ \prime}s)\right|\Upsilon\upsilon\right\rangle\] \[= \sum_{\sigma}\sum_{s}\sigma^{*}(s)\sigma(s)\sum_{ij}\varphi_{i}^{* }({\bf r})\varphi_{j}({\bf r}^{\prime})\left\langle\Upsilon\upsilon\left|\psi^ {+}(i\sigma)\psi(j\sigma)\right|\Upsilon\upsilon\right\rangle=\] \[= \sum_{ij}\varphi_{i}^{*}({\bf r})\varphi_{j}({\bf r}^{\prime}) \left\langle\Upsilon\upsilon\left|{\bf E}_{ij}^{\Upsilon}\right|\Upsilon \upsilon\right\rangle=\] \[= \sum_{ij}\varphi_{i}^{*}({\bf r})\varphi_{j}({\bf r}^{\prime}) \delta_{ij}\left\langle\Upsilon\upsilon\left|{\bf E}_{ii}^{\Upsilon}\right| \Upsilon\upsilon\right\rangle\] \[= \sum_{i}n_{i}\varphi_{i}^{*}({\bf r})\varphi_{i}({\bf r}^{\prime})\]
Thus even the spatial density matrix (not only the density) is permutation/spin independent as stated in Ref..
#### 4.2.2 Energy in the unitary group formalism
Further development is based on the possibility to express the blocks of the Hamiltonian matrix pertaining to \(N\) electrons in \(M\) orbitals with total spin \(S\) through the generators \({\bf E}_{ij}^{\Upsilon}\), with \(\Upsilon=\Upsilon(M,N,S)\). The required representation reads Ref.:
\[{\bf H} = \bigoplus_{\Upsilon}{\bf H}^{\Upsilon} \tag{19}\] \[{\bf H}^{\Upsilon} = \sum_{ij}h_{ij}{\bf E}_{ij}^{\Upsilon}+\frac{1}{2}\sum_{ijkl}(ij| kl)\left({\bf E}_{ij}^{\Upsilon}{\bf E}_{kl}^{\Upsilon}-\delta_{jk}{\bf E}_{il}^{ \Upsilon}\right) \tag{20}\]
The matrix elements \(h_{ij}\) are the sums of the respective matrix elements of the kinetic energy \(\hat{T}_{e}\) of electrons and of the external Coulomb potential \(\hat{V}_{ne}\); the quantities \((ij|kl)\) - the two-electron matrix elements of the Coulomb interactions.
For each of the states \(|\Upsilon\upsilon\rangle\) represented by the Young tableau with the Young pattern \(\Upsilon\) and the filling \(\upsilon\) (this information suffice to define the spatial part of the \(N\)-electron wave function) the expectation value of the energy reads:
\[E(\Upsilon\upsilon)=\sum_{ij}h_{ij}\left\langle{\bf E}_{ij}^{\Upsilon}\right\rangle _{\Upsilon\upsilon}+\frac{1}{2}\sum_{ijkl}(ij|kl)\left\langle\left({\bf E}_{ij}^ {\Upsilon}{\bf E}_{kl}^{\Upsilon}-\delta_{jk}{\bf E}_{il}^{\Upsilon}\right) \right\rangle_{\Upsilon\upsilon} \tag{21}\]
For the one-electron contribution to the energy one gets:
\[\sum_{ij}h_{ij}\left\langle{\bf E}_{ij}^{\Upsilon}\right\rangle_{\Upsilon \upsilon}=\sum_{i}h_{ii}\left\langle{\bf E}_{ii}^{\Upsilon}\right\rangle_{ \Upsilon\upsilon}=\sum_{i}h_{ii}n_{i}\]
and the Coulomb interaction of electrons is expressed through the Coulomb and exchange integrals with respect to the orbitals involved in the construction of the states represented by the Young tableaux:
\[\begin{array}{ll}{\rm Hartree}&\frac{1}{2}\sum_{ij}\left(ii|jj\right) \left\langle{\bf E}_{ii}^{\Upsilon}{\bf E}_{jj}^{\Upsilon}\right\rangle_{ \Upsilon\upsilon}+\\ {\rm exchange+}&\frac{1}{2}\sum_{i\neq j}\left(ij|ji\right)\left\langle{\bf E}_ {ij}^{\Upsilon}{\bf E}_{ji}^{\Upsilon}-{\bf E}_{ii}^{\Upsilon}\right\rangle_{ \Upsilon\upsilon}-\sum_{i}\left(ii|ii\right)\left\langle{\bf E}_{ii}^{\Upsilon }\right\rangle_{\Upsilon\upsilon}\end{array} \tag{22}\]
The Young tableau states \(\Upsilon\upsilon\) are the eigenstates of the diagonal generators \({\bf E}_{ii}^{\Upsilon}\). For that reason the Hartree contribution to the energy can be rewritten:
\[{\rm Hartree}\ \ \frac{1}{2}\sum_{ij}\left(ii|jj\right)\left\langle{\bf E}_{ii }^{\Upsilon}\right\rangle_{\Upsilon\upsilon}\left\langle{\bf E}_{jj}^{\Upsilon }\right\rangle_{\Upsilon\upsilon}=\frac{1}{2}\sum_{ij}\left(ii|jj\right)n_{i}n _{j} \tag{23}\]
From this we see that the Hartree part of the Coulomb energy is uniquely defined by the occupation numbers of the spatial orbitals _i.e._ only by the spatial density in the representation of orbitals. We see that as in the other representations the Hartree term is contaminated by the self-interaction of electrons and that the principal effect for which the true exchange term is responsible in the HFR context - the avoiding of the self interaction - is guaranteed by the specific form of the coefficient at the integrals of the \((ii|ii)\) type which must be absorbed by the exchange contribution to the energy. The averages of the off-diagonal generators' products entering the expression eq. are not however uniquely defined either by the occupation numbers of the orbitals in the tableau \(\Upsilon\upsilon\) or by the total spin, prescribed by the pattern \(\Upsilon\). They depend also on the mutual positions of the orbital symbols in the tableau. This is precisely the result obtained many years ago in Ref. under the name of the spin-free self consistent field theory.
?From the ROHF (old MCSCF) point of view the result eqs., can be considered as a recipe of obtaining the coupling coefficients \(a_{ij}\) and \(b_{ij}\) at the Coulomb and exchange integrals in the ROHF expressions for the energy which incidentally acquire the \(\Upsilon\upsilon\) dependence:
\[a_{ij}^{\Upsilon\upsilon} = \left\langle{\bf E}_{ii}^{\Upsilon}{\bf E}_{jj}^{\Upsilon}-\delta _{ij}{\bf E}_{ii}^{\Upsilon}\right\rangle_{\Upsilon\upsilon} \tag{24}\] \[b_{ij}^{\Upsilon\upsilon} = \left\langle{\bf E}_{ij}^{\Upsilon}{\bf E}_{ji}^{\Upsilon}-{\bf E }_{ii}^{\Upsilon}\right\rangle_{\Upsilon\upsilon} \tag{25}\]
Turning back to expressions eq. one can say that constructing the spatial Young tableaux states provide the expansion coefficients \(C_{i}^{\Upsilon\upsilon}\) for the respective linear combinations of the \(N\)-electron Slater determinants, yielding the total spin specified by the Young pattern. These coefficients are derived by purely symmetry reasons and do not depend on the matrix elements of the Hamiltonian thus satisfying the requirement of "universality". On the other hand it is obvious that specifying the total spin only does not suffice to specify the electronic state. The procedure implied by eq. provides for each allowable set of \(M,N,S\) the whole bunch of energy expressions labeled by the rows \(\upsilon\) of the irreducible representation \(\Upsilon=\Upsilon(M,N,S)\).
### Multiplet sum method from the unitary perspective
The first usage of the above formalism is to repeat the success of the MSM in case of two electrons in two orbitals without addressing explicitly the foreign component of the triplet state with \(S_{z}=+1\). Indeed, the spatial parts of the multiplet states in eq. are equivalently represented as the Young tableaux states:
\[{}^{1}B\]
\[{}^{3}B\]
\[\left|\begin{array}{c}\framebox{a}\\ \framebox{b}\end{array}\right\rangle\]
Two electrons in two orbitals form only one spatial function for the spin triplet state, but in addition to one given above two more functions compatible with the spin singlet state:
\[\framebox{a}\]
\[\left|\begin{array}{c}\framebox{a}\\ \framebox{b}\end{array}\right\rangle,\framebox{b}\]
Three spatial functions compatible with the spin singlet state together form a basis of the three-dimensional irreducible representation of the group \(U\) corresponding to the total spin \(0\). The single spatial function compatible with the spin triplet state spans the one-dimensional irreducible representation of the group \(U\). The Young pattern label \(\Upsilon\) here can be replaced by indicating the total spin only. Then the generator \({\bf E}_{ab}^{S=1}=0\), but for \(S=0\) one has:
\[\left\langle\framebox{a}\framebox{a}\framebox{}\framebox{}{\bf E}_{ab}^{S=0} \framebox{a}\framebox{b}\right\rangle=\sqrt{2}=\left\langle\framebox{b} \framebox{b}\framebox{}\framebox{}{\bf E}_{ab}^{S=0}\framebox{a}\framebox{b} \right\rangle\]
These values suffice to perform the matrix multiplication of the generators \({\bf E}_{ab}^{S=0}{\bf E}_{ba}^{S=0}\) in the general expressions eq., so that we obtain for the contribution of the average interaction to the energy:
\[\begin{array}{ll}E(^{1}B)&=(aa|bb)+(ab|ba)\\ E(^{3}B)&=(aa|bb)-(ab|ba)\end{array}\]
We see that the archetypical result is reproduced within the Young tableaux technique without addressing the component of the spin multiplet with a foreign value of the spin projection. Also the self-interaction contamination is removed automatically.
### DFT implications
All above treatment was not in any way related to the DFT realm. The possibility of establishing such a relation can be based on the recognition of the fact that the symmetry (in particular the total spin) dependence must be extraneously introduced into DFT considerations analogously to the treatment by Filatov and Shaik Ref.. The unitary group formalism allows us to conclude that for a given set of consistent values of \(M,N,S\) one arrives to the family of functionals labelled by the rows \(\upsilon\) of the irreducible representation \(\Upsilon=\Upsilon(M,N,S)\) of the \(U(M)\) group. The spin symmetry features of these functionals are condensed in the \(\left\langle{\bf E}_{ij}^{\Upsilon}{\bf E}_{ji}^{\Upsilon}-{\bf E}_{ii}^{ \Upsilon}\right\rangle_{\Upsilon\upsilon}\) (or \(a_{ij}^{\Upsilon\upsilon}\), \(b_{ij}^{\Upsilon\upsilon}\)) coefficients given above.
The energy matrix elements reflecting specific features of the system can be easily figured out. The coefficients \(a_{ij}^{\Upsilon\upsilon}\) for the Coulomb integrals \((ii|jj)\) which define the Hartree part of the Coulomb energy are known, but they are of no practical use in the DFT context, where the Hartree part of the interaction is determined directly from the electron density. Relatively problematic (in the DFT context) is to decide where the energy matrix element to be combined with the coupling coefficients \(\left\langle{\bf E}_{ij}^{\Upsilon}{\bf E}_{ji}^{\Upsilon}-{\bf E}_{ii}^{ \Upsilon}\right\rangle_{\Upsilon\upsilon}\) (exchange)and \(-\left\langle{\bf E}_{ii}^{\Upsilon}\right\rangle_{\Upsilon v}\) (self-interaction) has to come from. This choice must be compatible with various theoretical settings. First of all we notice that if a hybrid functional is used which contains some fraction of the Hartree-Fock exchange the latter must be modified accordingly so that the corresponding \((ij|ji)\) integrals over the Kohn-Sham orbitals be included with the correct coefficients \(\left\langle{\bf E}_{ij}^{\Upsilon}{\bf E}_{ji}^{\Upsilon}-{\bf E}_{ii}^{ \Upsilon}\right\rangle_{\Upsilon v}\). The same applies to the integrals \((ii|ii)\) which together with coefficients \(-\left\langle{\bf E}_{ii}^{\Upsilon}\right\rangle_{\Upsilon v}\) will take care about some fraction of self-interaction.
Further concerns are related with the treatment of the nontrivial parts of the exchange-correlation functionals within the \(\Upsilon v\) numbering of the spin (permutation) states. This can be solved on the basis of certain consistency requirements. Among possible consistency requirements the most natural is that with the TDDFT. The TDDFT approximation is equivalent to constructing the corresponding time evolution of the many-electronic state in the basis of single electron excitations (particle-hole pairs) above the KS single determinant wave function. Leaving aside the question of the area of applicability of such an approach we notice that it requires an estimate of the two-electron integrals coupling between different singly excited determinants. The interaction appears as second functional derivative of the energy with respect to density (first functional derivative of the exchange-correlation functional). In the orbital representation these derivatives acquire the necessary form of two-electron integrals with the kernels determined by the form of the used exchange-correlation functional. On the other hand the \((ij|ji)\) integrals appear in ROHF and in unitary group formalism for \(\Upsilon v\) states as a consequence of configuration interaction between different Slater determinants implicitly entering in the expansion of the Young tableau state \(\Upsilon v\). Thus in order to ensure the compatibility of the corresponding components of the theory: the couplings between the elementary excitations in TD-DFT and between Slater determinants in expansions of \(\Upsilon v\) states must be the same. Thus they can be expressed through the integral kernels of the interaction (_e.g._ according to):
\[\begin{array}{rcl}(ij|ji)^{\rm xc}&=&\int\int\varphi_{i}^{*}({\bf r})\varphi _{j}({\bf r})f^{\rm xc}({\bf r},{\bf r}^{\prime})\varphi_{j}^{*}({\bf r}^{ \prime})\varphi_{i}({\bf r}^{\prime})d{\bf r}d{\bf r}^{\prime};\\ &&\mbox{where}\\ f^{\rm xc}({\bf r},{\bf r}^{\prime})&=&\frac{\delta v^{\rm xc}({\bf r})}{\delta \rho({\bf r}^{\prime})}\end{array} \tag{26}\]
Namely these quantities must be inserted in the expressions for the exchange correlation energies to get these later consistent with the total spin/permutation symmetry of the underlying many-electronic ground state. This is also in agreement with the way of constructing the coupling operators by Filatov and Shaik in their version of ROKS Ref. and a similar procedure can be easily designed for the \(\Upsilon\upsilon\) labelled states.
### Further examples
As we mentioned many times the spin in general does not suffice to distinguish many electronic states with the same one-electron density which produces problems in describing corresponding states in the DFT. The only example of the usage of the unitary group formalism given so far was however the simplest case when the total spin labeling was sufficient. Below we briefly exemplify the features one should expect in general case when there exist Young tableaux differing by the positions of the orbital symbols in these tableaux. In this case one can say that for given \(M,N,S\) uniquely defining the irreducible representation \(\Upsilon\) of the group \(U(M)\) and for the row \(\upsilon\) of the latter defined by a specific location of the orbital symbols in the tableau a "Hartree-Fock-like" energy functional can be written whose electron-electron interaction part is given by eq.. It can be optimized with respect to the expansion coefficients of the involved orbitals over the AO's basis yielding an effective Fockian matrix whose eigenvectors are precisely the orbitals involved in the construction of the Young tableau state in the same manner as it is in the ROHF/ROKS.
It is easy to check that the positions of the orbital indices in the tableaux really matter. For example, for two Young tableaux states:
\[\left|\Upsilon\upsilon\right\rangle=\left\|\begin{array}{c|c}\mbox{a}&\mbox {b}\\ \hline\mbox{c}&\mbox{d}\end{array}\right\rangle;\;\left|\Upsilon\upsilon^{ \prime}\right\rangle=\left\|\begin{array}{c|c}\mbox{a}&\mbox{c}\\ \hline\mbox{b}&\mbox{d}\end{array}\right\rangle \tag{27}\]
both representing singlet states of four electrons in four orbitals with equal one-electron density matrices, the contributions to the energy functionals of the form eq, proportional to the exchange integrals, respectively, are:
\[\begin{array}{ll}\left|\Upsilon\upsilon\right\rangle:&\;\;(ab|ba)+(cd|dc)- \frac{1}{2}\left[(ac|ca)+(ad|da)+(bc|cb)+(bd|db)\right]\\ \left|\Upsilon\upsilon^{\prime}\right\rangle:&-(ab|ba)-(cd|dc)+\frac{1}{2} \left[(ac|ca)+(ad|da)+(bc|cb)+(bd|db)\right]\end{array} \tag{28}\]
Remarkably enough neither of the expressions eq. (combined with other necessary temrs) yields a lower energy _a priori_: which one is lower depends on the relations between the molecular integrals involved. At this point one can return to the qualitative interpretation of the Young tableaux with different positions of the orbital symbols as of reflecting different "pairing schemes". Indeed, the states in eq. can be respectively treated (and this is in accord with the energy expressions eq.) as pairwisely coupling electrons in the states \(a\) and \(b\) and \(c\) and \(d\) to the singlets and triplets then coupling these intermediate states to the final singlet states.
On the other hand one can easily conclude that for the above pair of Young tableaux \(\Upsilon\upsilon\) and \(\Upsilon\upsilon^{\prime}\) for which \(n_{i}=n_{j}=1\) and the difference between them is only the positions of the orbital symbols \(b\) and \(c\) in the tableaux eq. the operators \({\bf E}_{ij}^{\Upsilon}{\bf E}_{ji}^{\Upsilon}\) entering as multipliers of the \((ij|ji)\) exchange integrals in the exact Hamiltonian yield also an off-diagonal matrix element of the Hamiltonian:
\[\left\langle\Upsilon\upsilon\left|{\bf H}^{\Upsilon}\right|\Upsilon\upsilon^{ \prime}\right\rangle=-\frac{\sqrt{3}}{2}\left((ac|ca)-(ad|da)-(bc|cb)+(bd|db )\right)\neq 0, \tag{29}\]
Thus the energy functional becomes a square root irrational function of the two-electron integrals rather than a linear one. Additional symmetry relations may produce the energy expression linear in the Coulomb and the exchange integrals (in the above example it suffice that exchange integrals \((ac|ca),(ad|da)\) and \((bc|cb),(bd|db)\) are pair-wisely equal (which makes eq. be zero) or alternatively that the exchange integrals satisfy the equality:
\[(ab|ba)+(cd|dc)=\frac{1}{2}\left[(ac|ca)+(ad|da)+(bc|cb)+(bd|db)\right] \tag{30}\]
Both symmetries yield specific forms of the 2\(\times\)2 configuration interaction matrix and by this allow the diagonalization to be feasible on the purely symmetry grounds).
State-specific exchange-correlation functionals for atomic \(d\)-shells
The above notion of irrationality shows that even the unitary group formalism does not solve the problem of constructing density functionals for the specific correlated states. Although the unitary group formalism allows to significantly contract the expansions of the states of the definite total spin (in fact the \(\Upsilon v\) labeled states become single-configuration, albeit each of them is a combination of many Slater determinants) the nonlinearity of the energy expression with respect to the two-electron integrals hinders constructing the symmetry adapted functionals along the lines suggested above. This problem manifests itself in the description of the many-electronic states in the \(d\)-shells of transition metal ions. Using the unitary goup formalism also in this case does not allow to go further than the Roothaan _old_ MC SCF theory as described in Ref.. There the spin/angular momentum dependent coupling coefficients \(a_{ij}\) and \(b_{ij}\) had been introduced ultimately to express the cumulant of the two-electron density matrix using symmetry considerations. They are valid only if the multiplet states can be uniquely obtained by applying operators projecting the Young tableau states to the specific rows of the irreducible representations of the \(SO\) or \(SO\) groups (atoms and linear molecules, respectively). In these two cases moderately simple expressions for the classifying operators (respectively, \(L^{2}\) and \(L_{z}\)) in terms of the generators \({\bf E}_{ij}^{\Upsilon}\) can be written and used for constructing the required symmetry adapted combinations of the Young tableau states. In the case of the atomic \(p\)-shells (and molecular \(\pi\)- and \(\delta\)-shells) the number of the \(SO\) (\(SO\)) symmetry labels (different values of the orbital momentum \(L\)) produced by the projection of the Young tableau states to the definite \(L^{2}\) states suffice to distinguish all different energies in these shells. In the case of atomic \(p^{n}\)-states the symmetry \(SO\) reduces also the number of independent two-electron integrals (including both the Coulomb and exchange ones) to only two independent Slater-Condon parameters \(F_{k}(pp);k=0,2\). This allowed the authors of Ref. to develop state-specific functionals for the atomic \(p^{n}\)-states. It turns out, however, that for the \(d\)-shells it does not suffice for a major part of the atomic electronic terms of the transition metal ions. Even in free ions where the multiple terms having the same spin and orbital momentum do exist and their energies cannot be expressed linearly through the two-electron integrals. Despite the high-symmetry sit uation of a free atom (ion) which reduces all the two-electron integrals to a limited number (three) Slater-Condon parameters \(F_{k};k=0,2,4\) in the free ions the energies of the multiplets require 2\(\times\)2-diagonalization and thus their analytical expressions contain square roots (for a handy reference see). This moment is crucial - it is not possible to get rid out of the irrationality (square root) in the expression for the energy by linearly combining the parameters of the Hamiltonian.
The situation clearly becomes less favorable in lower symmetries or in larger subshells (_e.g._ partially filled \(f\)-shells) where the terms of the same spin and symmetry span the subspaces of dimensionalities higher than two. For example, in the octahedral environment the LS states of \(d^{4}\)- (\(d^{6}\)-) configuration span up to seven-dimensional subspaces of many-electronic states. Clearly, at an arbitrarily low symmetry the problem of linearly expressing the exact energy of many-electronic terms through the Coulomb and exchange integrals cannot be solved and obviously the energy of any of such multiple terms cannot be expressed as a linear combination of Coulomb and exchange integrals. In what follows below we restrict ourselves to the case of atomic \(d\)-shells and the square root irrationalities in the state-specific expressions for the energy trying to squeeze the simplest thinkable irrationality reflecting nontrivial correlations in a kind of generalized density functional.
### The example of Fe\({}^{2+}\) ion
We concentrate on the free Fe\({}^{2+}\) (\(d^{6}\)) ion which is an important object in the studies of biologically active transition metal complexes and following provides a rich system of nontrivially correlated multiple states in its \(d\)-shell. Namely this kind of behavior is known to systematically evade from any DFT-based treatment. The energy expressions of the states in a free Fe\({}^{2+}\) (\(d^{6}\)) ion are given in Table 1. They nontrivially depend on two Slater-Condon parameters: \(F_{2}\) and \(F_{4}\). The ground state follows the Hund's rule and for the Fe\({}^{2+}\) (d\({}^{6}\)) ion it is the \({}^{5}D\) state. According to the data published Ref. the states \({1\over\pm}S\) and \({1\over\pm}D\) are not resolved from the spectra. Also the \({}^{2}F\) state cannot be present in the spectrum of an even-electron system. Thus we exclude three uppermost rows of the Table 1 and finally arrive to the set of data Table 2 which can be used for analysis.
In order to get an impression of what can be (and should be) possibly achieved in terms of describing the multiple states of the \(d\)-shells we determine parameters \(F_{2}\) and \(F_{4}\) from experimental data. This can be done in a number of ways. The semi-empirical approach is to assume that \(F_{2}\) and \(F_{4}\) are independent parameters. At the first stage we neglect the correlation and take into consideration only the average energies of the \({}^{(2S+1)}_{\pm}L\) states. The corresponding set of energies is given in Table 3. These energies are linear in the parameters \(F_{2}\) and \(F_{4}\). Applying the standard linear least squares procedure yields the experimental "non-correlated" estimate of the parameters (in cm\({}^{-1}\)):
\[\begin{array}{ccl}F_{2}^{\rm exp}&=&1411.0,\\ F_{4}^{\rm exp}&=&120.25,\\ F_{2}^{\rm exp}/F_{4}^{\rm exp}&=&11.734.\end{array} \tag{31}\]
The quality of this result can be assessed by the value of mean square deviation which is 686.80 cm\({}^{-1}\) which must be compared with the range of the energies described by the model being _ca._ 45000 cm\({}^{-1}\).
Next step consists in estimating the manifestations of correlations in the available data set. The most direct way to do that is to consider the square root contributions to the energies of the multiple terms of the same spin and symmetry. These come from the diagonalization of the symmetry adapted CI matrices. Technically the correlations of that sort are responsible for the splitting within the pairs of states of the same spin and symmetry which do not have any counterpart in the DFT and describe the nontrivial (non-dynamical) part of the correlation. We can see from the Table 4 that the correlation splitting between the double states is by approximately 10% underestimated when calculated with use of the non-correlated experimental estimates of the \(F_{2}^{\rm exp}\) and \(F_{4}^{\rm exp}\) parameters eq.. The overall picture as coming from the non-correlated estimate can be characterized by its mean square deviation 1228.0 cm\({}^{-1}\). This fit seems to be improvable by performing another (nonlinear) one for the entire set of available excitation energy expressions and the corresponding experimental values as given in Table 2. The result of this new fit is, however, twofold. The resulting values (cm\({}^{-1}\)) of the parameters \(F_{2}^{\rm exp}\) and \(F_{4}^{\rm exp}\) eq.:
\[\begin{array}{ccl}F_{2}^{\rm exp}&=&1468.92,\\ F_{4}^{\rm exp}&=&113.30,\\ F_{2}^{\rm exp}/F_{4}^{\rm exp}&=&12.960,\end{array} \tag{32}\]
Meanwhile, although the overall picture is improved the description of the average multiplet energies is deteriorated as compared to the "non-correlated" parameters eq. so that the corresponding mean square deviation somewhat increases to the value of 859.37 cm\({}^{-1}\).
The above results deserve thorough attention. First of all we notice following Refs. that the parameters \(F_{2}\) and \(F_{4}\) are by definition some functionals of radial density:
\[\begin{array}{rcl}F_{k}&=&\frac{e^{2}}{D_{k}}\int\limits_{0}^{ \infty}\int\limits_{0}^{\infty}\frac{\left[\min(r_{1},r_{2})\right]^{k}}{\left[ \max(r_{1},r_{2})\right]^{k+1}}R^{2}(r_{1})R^{2}(r_{2})r_{1}^{2}r_{2}^{2}dr_{1 }dr_{2}\\ &&\hskip 142.26378ptD_{0}=1;D_{2}=49;D_{4}=441\end{array} \tag{33}\]
However, according to the Theorem 2 of Ref. for whatever spatial multiplet the one-electron density is spherically symmetric. Thus the quantities \(F_{k}\) are the functionals of one-electron density which in the said case have only the radial dependence \(r=|{\bf r}|\). For that reason the energies in Table 1 can be also treated as functionals of the one-electron density representing the averages of the electron-electron interaction energy for each specific many-electron state in the \(d\)-shell.
\[T[R^{2}(r)]+V_{ne}[R^{2}(r)]+\frac{n_{d}\left(n_{d}-1\right)}{2}A[R^{2}(r)]+XC_ {nLS}[R^{2}(r)] \tag{34}\]
They can be treated according to the variational principle (in some analogy with Ref.) this is going to yield some integrodifferential equations for the functions \(R(r)\). This option will be considered in details elsewhere. Here we notice that assuming the model Slater orbital form for the functions \(R(r)\) in the \(d\)-shell:
\[R(r)=\frac{(2\zeta)^{n+\frac{1}{2}}}{\sqrt{(2n)!}}r^{n-1}\exp(-\zeta r)\]allows one to evaluate the integrals in eq. thus leading to the linear dependence of the latter on the orbital exponent \(\zeta\):
\[\begin{array}{ccc}F_{0}^{\rm th}&=&\frac{793}{3072}\zeta\\ F_{2}^{\rm th}&=&\frac{2093\cdot 5}{49\cdot 76800}\zeta\\ F_{4}^{\rm th}&=&\frac{91\cdot 9}{441\cdot 9216}\zeta\end{array} \tag{35}\]
(here \(n=3\)).
\[F_{2}^{\rm th}/F_{4}^{\rm th}=\frac{2093\cdot 5}{49\cdot 76800}/\frac{91\cdot 9 }{441\cdot 9216}=13.8 \tag{36}\]
Since as we mentioned the density is spherically symmetric for whatever of the states listed in the above Tables the only parameter characterizing the density is the orbital exponent \(\zeta\), provided the said multiplets are constructed on the Slater radial orbitals. In view of the linear dependency of \(F_{k}^{\rm th}\) on \(\zeta\) the state specific expressions for the energies and energy differences in Tables 1 and 2 become linear functions of \(\zeta\) as well.
The excitation energy expressions can be converted to the full scale density functionals for the \(d\)-shell if one complements the above electron interaction energies by the one-electron terms for the \(d\)-shell with six electrons in it. The one-electron terms are (i) the kinetic energy per electron:
\[\frac{\zeta^{2}}{2}; \tag{37}\]
(ii) the potential energy of attraction to the nucleus per electron where the 3 in the denominator stands for the principal quantum number of the \(d\)-shell under consideration:
\[-\frac{Z}{3}\zeta \tag{38}\]
The average electron-electron interaction value common for all electronic terms is proportional to the Racah \(A\) parameter whose expression in terms of \(F_{0}^{\rm th}\) and \(F_{4}^{\rm th}\) is given in the footnote to Table 1. For the iron(II) ion we can set \(Z=8\) and the number of \(d\)-electrons \(n_{d}=6\) to take care about the core screening, then the expression for the energy becomes:
\[3\zeta^{2}-16\zeta+15\frac{143}{576}\zeta. \tag{39}\]
The value of \(\zeta\) comes then as one providing the minimum to the above functional, so that:
\[\zeta=\frac{8}{3}-\frac{5}{2}\frac{143}{576}\approx 2.0460... \tag{40}\]
The screening increment coming form the formula for \(A\) amounts 0.387.
Including further contributions for the electron-electron interaction energy which are now state specific yields for the ground state:
\[\zeta\approx 2.0621 \tag{41}\]
On the other hand taking one of the higher excited states \({}^{1}I\) whose energy is about 30000 cm\({}^{-1}\) above the ground state gives:
\[\zeta\approx 2.0533 \tag{42}\]
From these estimates one can derive the following conclusion: The orbital exponent and thus the radial density is very weakly sensitive to whatever correlations. This finding is in agreement both with the accepted concept of correlation which attributes it exclusively to the cumulant of the two-electron density matrix so that there is no need to reload its manifestations on the density as well as with numerous demonstrations of _no relation_ between the one-electron density and the correlations known in the literature (see _e.g._ Ref.).
Further analysis can be based on the observation that inserting the theoretical definitions for the \(F_{2}^{\rm th}\) and \(F_{4}^{\rm th}\) parameters eq. into expressions for the excitation energies result in linear models for these energies with the single fitting parameter \(\zeta\). Two such models can be constructed: the non-correlated which uses only the average energies of multiple states with equal \(L\) and \(S\) and the correlated one which covers all ten available excitation energies. Fitting the excitation energies to the non-correlated model yields the value of 2.4823 for \(\zeta\). The quality of fitting with only one parameter is certainly somewhat worse than that using two independent parameters \(F_{2}^{\rm exp}\) and \(F_{4}^{\rm exp}\) and the mean square deviation becomes 1058.0 cm\({}^{-1}\) for the set of average energies of the multiplets (non-correlated fit). The value of \(\zeta\) which comes from the linear fitting procedure with the correlated energy expressions is 2.4604 and the mean square deviation is 1013.3 cm\({}^{-1}\). We see that also in this case the correlations only marginally affect the one-electron density distribution and that despite some deterioration of the precision as compared with the two-parameter models the overall quality of the fit is surprisingly good.
### Summary
Let us summarize the findings of this Section. We managed to obtain simple expressions for the energies of the nontrivially correlated ionic states (these expressions include non-dynamic correlation through the square root terms) with definite values of \(L\) and \(S\) as functions of a single parameter \(\zeta\) - the Slater orbital exponent for the \(d\)-shell. In the context of the accepted model it is the only quantity characterizing the density in the \(d\)-shell. In a sense there is one-to-one correspondence between the electron density of the \(d\)-shell and \(\zeta\) thus the expressions for the energy can be considered as state specific energy functionals of the form:
\[n_{d}\frac{\zeta^{2}}{2}-n_{d}\frac{Z}{n}\zeta+\frac{n_{d}\left(n_{d}-1\right) }{2}\frac{143}{576}\zeta+XC_{nLS}(n_{d},\zeta) \tag{43}\]
in the expressions given in Table 1 or analogous expressions for other \(d\)-shell fillings Ref..
As one can see our estimates of the characteristic quantity \(\zeta\) yield the values which fall into two classes depending on the type of the estimate: those coming from the variational estimate for the total energy of each respective state give the values close to \(\zeta=2.08\) coming from the Slater rules. The estimates based on fitting the excitation energies to \(\zeta\) yield much larger value (much less diffuse \(d\)-shell) about 2.5 with with extremely weak influence of electron correlation on the estimates of either of these types. These numerical results must be compared with other (empirical) values of the orbital exponents. These, however, demonstrate a wide range of values. _E.g._Ref. report the value \(\zeta=3.7266\); Ref. suggests \(\zeta=3.152\); Ref. gives \(\zeta=2.722\); and Ref. provide \(\zeta=3.15\) basically repeating the value of Ref.. This indicates that either the correlated or non-correlated estimates, coming from the excitation energies only, fall in the range defined by the Slater rules and other semi-empirical estimates. Although, the deviations between the density parameter estimates coming from different types of procedures also are expectable (we remind the existence of distinct thermochemical and spectral semi-empirical parameterizations) the true source of observed deviations is of certain interest.
## 6 Range-separated treatment of electronic Coulomb interaction in atomic \(d\)-shells
Based on the idea that the short-range behavior of the e-e interactions can be efficiently transferred from the homogeneous e-gas to arbitrary many-electron systems, while the long-range e-e interactions being much more system specific (less transferable), Savin and Stoll suggested a generalization of the Kohn-Sham theory by splitting explicitly the short- and long-range e-e interactions. The non-transferable long-range interactions can be assimilated to a wave function treatment, just like the kinetic energy in the conventional Kohn-Sham model, which results in a replacement of the non-interacting Kohn-Sham reference system by a "long-range-interacting" one. While in conventional KS theory the effective KS Hamiltonian has an exact single-determinant solution, the generalized, range-separated variant includes a certain amount of explicit e-e interaction and the corresponding effective Schrodinger equation has to be solved in a multi-determinant form. However, due to the nonsingular nature of the lr Coulomb operator, the solution can be converged considerably faster in both the one-electron and many-electron basis.
Recent works on the range-separated hybrid methods were mostly based on the the separation of the Coulomb potential into short- and long-rangeparts was performed according to:
\[\frac{1}{r_{12}} = \left(\frac{1}{r_{12}}\right)_{s}+\left(\frac{1}{r_{12}}\right)_{l}\] \[\left(\frac{1}{r_{12}}\right)_{s} = \frac{\mathrm{erf}(\mu r_{12})}{r_{12}} \tag{44}\] \[\left(\frac{1}{r_{12}}\right)_{l} = \frac{\mathrm{erfc}(\mu r_{12})}{r_{12}}\] \[1 = \mathrm{erf}(x)+\mathrm{erfc}(x)\]
The treatment of the long-range exchange has been done in the Hartree-Fock framework, while the correlation could be treated by MP2 or CCSD(T) level, leading to a successful description of London dispersion forces in vdW complexes or by MCSCF level to treat typical non-dynamic correlation problems, like the case of the H\({}_{2}\) dissociation. A simpler model, where long-and short-range correlations are both handled by density functional approximations (RSHX - exchange-only range separated hybrid, like LC-\(\omega\)PBE of Scuseria) has been recently shown to be quite successful in predicting magnetic coupling constants in transition metal systems.
In the following, we examine the behavior of the range-separated approach on the simple Fe(II) ion model system.
### Range separated hybrid approach
In order to make easier the evaluation of analytical integrals and obtain the \(F_{k}^{\mathrm{th}}\) parameters, we decided to employ the "Yukawa"-like separation as proposed in Ref.:
\[\left(\frac{1}{r_{12}}\right)_{s} = \frac{\exp(-\beta r_{12})}{r_{12}}, \tag{45}\] \[\left(\frac{1}{r_{12}}\right)_{l} = \frac{1-\exp(-\beta r_{12})}{r_{12}}.\]
The value of \(\beta\to 0\) corresponds to the absence of the long-range part. By contrast \(\beta\rightarrow\infty\) corresponds to the evanescence of the short-range part. The reach of the short-range interactions is roughly inversely proportional to the value of \(\beta\) measured in inverse bohr units.
The initial assumption is that only the long-range part of the Coulomb interaction contributes to the non-dynamical correlations in the \(d\)-shells so that only the matrix elements of \(\left(\frac{1}{r_{12}}\right)_{l}\) must be taken into account when the CI matrices describe the nontrivial correlation in the \(d\)-shells. In order to check this assumption we have performed the following. With use of analytical results of Refs. to get for the Yukawa potential the following expansion:
\[\left|{\bf r}_{1}-{\bf r}_{2}\right|^{-1}\exp(-\beta\left|{\bf r} _{1}-{\bf r}_{2}\right|) = 4\pi\sum_{l=0}^{\infty}\sum_{m=-l}^{l}\left(r_{<}r_{>}\right)^{- \frac{1}{2}}\] \[I_{l+\frac{1}{2}}\left(\beta r_{<}\right)K_{l+\frac{1}{2}}\left( \beta r_{>}\right)Y_{l}^{-m}\left(\frac{{\bf r}_{<}}{r_{<}}\right)Y_{l}^{m} \left(\frac{{\bf r}_{>}}{r_{>}}\right)\]
The above expression must be inserted in the definition of the matrix elements of the electron-electron interaction (see _e.g._ Ref.) which due to the spherical symmetry of the Yukawa potential allows us to express these latter in terms of the short range analogs of the Slater-Condon parameters. For the \(3d\) Slater orbitals with the orbital exponent \(\zeta\) the estimates for the short range \(F_{2}^{\rm(s)}\) and \(F_{4}^{\rm(s)}\) and long-range \(F_{2}^{\rm(l)}\) and \(F_{4}^{\rm(l)}\) contributions to the \(F_{k}^{\rm th}\) parameters eq.
\[t=\frac{\beta}{\beta+2\zeta} \tag{48}\]
we get somewhat simpler expressions for the long-range scaling coefficients
\[g^{} = -\frac{256t^{12}}{91}+\frac{2304t^{11}}{91}-\frac{8832t^{10}}{91}+ \frac{18432t^{9}}{91}-240t^{8}\] \[+144t^{7}-16t^{6}-\frac{144t^{5}}{7}-\frac{6t^{4}}{7}+\frac{30t^{ 3}}{7}+\frac{15t^{2}}{7}\] \[f^{} = -\frac{6400t^{12}}{6279}+\frac{75520t^{11}}{6279}-\frac{134528t^{ 10}}{2093}+\frac{1285120t^{9}}{6279}\] \[-\frac{384880t^{8}}{897}+\frac{14160t^{7}}{23}-\frac{23600t^{6}}{ 39}+\frac{2451808t^{5}}{6279}\] \[-\frac{23150t^{4}}{161}+\frac{6910t^{3}}{483}+\frac{3455t^{2}}{483}\]
Numerical optimization of the sum of square deviations, where the theoretical values are obtained under the condition that the parameters \(F_{2}\) and \(F_{4}\) under the square roots are respectively replaced by the long range contributions \(F_{2}^{}\) and \(F_{4}^{}\), with respect to \(\zeta\) and \(t\) results in the values:
\[\zeta = 2.46425\] \[t = 0.725436 \tag{50}\] \[\beta = 13.0218\]
The range separation parameter \(\beta\) is obtained by inverting the definition of \(t\). These values correspond to the following scaling parameters:
\[f^{} = 0.965879 \tag{51}\] \[g^{} = 0.912733\]
The precision of this estimate can be characterized as previously by the mean square deviation which amounts to 996 cm\({}^{-1}\). Taking into account that the short-range e-e potential corresponding to \(\beta=13.02\) falls down to a negligibly small value, say 0.001, for \(r=0.5\) bohr, it can be concluded that electron repulsion at shorter than 0.5 bohr direct space distance has an insignificant effect on the multiplet structure.
By contrast, if the theoretical values are obtained under the condition that the parameters \(F_{2}\) and \(F_{4}\) under the square roots are respectively replaced by the short range contributions \(F_{2}^{\rm(s)}\) and \(F_{4}^{\rm(s)}\), the optimization of the sum of square deviations with respect to \(\zeta\) and \(t\) results in the values:
\[\zeta = 2.4653\] \[t = 0.0482271 \tag{52}\] \[\beta = 0.249838\]
These values correspond to the magnitudes of the scaling parameters:
\[f^{\rm(s)} = 0.982441 \tag{53}\] \[g^{\rm(s)} = 0.994545\]
Incidentally, the precision of the procedure singling the short-range part characterized as previously by the mean square deviation yields the value 985 cm\({}^{-1}\), quite similar to the long-range estimate. The reach of the "short-range" interactions, measured by analogous criteria as before (falling off the short-range Coulomb potential below 0.001) is about \(r=16\) bohr, which englobes practically the full range for the significant densities of the \(d\)-electrons. It means that the long-range "tail" of the electron-electron interactions is essential to recover the correct muliplet structure. Furthermore, one can see that the renormalization of the \(F\) functions is in the order of 1%, confirming that the use of the optimal \(\beta\) implies the involvement of practically the full range of interactions (cf. previous Section).
A further lesson drawn from this simple model study is that the short range/long range separation of the Coulomb potential is not sensitive to the correlations as well: the characteristic parameter of the density distribution \(\zeta\) in all cases equals to 2.46 with variations in the third digit after the decimal point. Thus the short range/long range separation does not lift thus the strong contradiction between the estimates of the orbital exponent by the Slater rules or variationally from the state specific functionals eq. and those from linear fit for the excitation energies. Thunkable way out looks out twofold: First, the Slater rules can be thought to overestimate the screening (for the \(d\)-shell the screening by the inner shells is treated to be complete, which yields the value 8 for the effective charge) thus leading to the values of \(\zeta\) too small as compared to those extracted from the fitting of the experimental data on excitation energies. Second, one can think that the value of parameter \(F_{0}\) is for some reason much stronger renormalized as compared to its theoretical value eq. than those of the parameters \(F_{2}\) and \(F_{4}\). If we apply long-short range separation and calculate \(F_{0}\) at the values of \(\beta\) and \(\zeta\) eq. extracted from the fitting of excitation energies with the short-range parts \(F_{2}^{\rm(s)}\) and \(F_{4}^{\rm(s)}\) under the square roots the fraction of the short range part in \(F_{0}\) amounts \(h^{\rm(s)}=0.688608\) of the latter. Now if we assume that the short range part for some reason renormalizes to zero and thus only the long range part of the e-e potetial contributes to the real value of \(F_{0}\) then the variational estimate of the orbital exponent reads:
\[\zeta=\frac{8}{3}-\frac{5}{2}\cdot\frac{143}{576}\cdot(1-h^{\rm(s)})\approx 2.4743 \tag{54}\]
Of course, this may well be a pure coincidence, but possible consequences of the above hypothesis on the way of renormalization of the Slater-Condon parameters will be considered elsewhere.
## 7 Conclusion
In the present paper we discussed a few possible ways of avoiding the dead-locks of the pragmatic methods of molecular electronic structure theory based on the DFT, which appear due to the non-sensitivity of the basic quantity of the DFT - the one-electron density - to the differences in the spin (permutational) or/and spatial symmetry of the underlying many-electronic states. This non-sensitivity is reflected by two theorems (recent Theorems 1 and 2 of Ref.) which formalize two basically known facts that (i) the one-electron density does not depend on the total spin of the many-electron state, and, that (ii) the one-electron density in a many-electronic state, which transforms according to any irreducible representation of the group acting on the spatial coordinates of electrons (\(SO\), \(SO\), or their point subgroup), transforms according the fully symmetric irreducible representation of the corresponding group.
These theorems imply that necessarily the information concerning the symmetry of the respective many-electronic states at hand is to be introduced into any DFT-based treatment extraneously. When it goes about the total spin (or equivalently about the permutational symmetry) of a many-electron state, we suggest to use state-specific functionals labeled by the Young tableaux \(\Upsilon\upsilon\) (the rows of the irreducible representations of the unitary group \(U(N)\)) and to develop a procedure analogous to ROKS for each of them. In the particular case of multiple states sharing the same \(L\) and \(S\) in the \(d\)-shells of transition metal ions we suggest state-specific correlated functionals of the density and their model based on the assumption of a Slater orbital form of the radial density distribution. This procedure reduces the functional-type density dependence to function-type dependence on the orbital exponent. With the use of these expressions the excitation energies of the many-electron states of the Fe\({}^{2+}\) ion are reproduced with remarkable accuracy. The variational treatment of the proposed functionals reproduces with similar precision the values of the orbital exponent of the Fe\({}^{2+}\) ion prescribed by the Slater rules. Nevertheless, the estimates of the orbital exponent coming from the variational principle and from the fit of the excitation energies differ significantly although they fall in the range provided by different semi-empirical estimates. Some ideas related to conciliation of these two groups of estimates have been derived from analysis of the short range/long range separation of the electron-electron interaction potential.
|
10.48550/arXiv.0902.3559
|
Classes of admissible exchange-correlation density functionals for pure spin and angular momentum states
|
A. L. Tchougréeff, J. G. Ángyán
| 842
|
10.48550_arXiv.1904.08040
|
## Chapter 1 Electrochemical Systems: Electrodes and Double Layers
Asghar Aryanfar\({}^{1}\)\({}^{*,}\)\({}^{\dagger}\), Agustin J. Colussi\({}^{*}\), Laleh M. Kasmaee\({}^{*}\), Michael R. Hoffmann\({}^{*}\)
\(*\) _California Institute of Technology, 1200 E California Blvd, Pasadena, CA 91125_
\(\dagger\) _Balcsehir University, 4 Ciragan Cad, Besiktas, Istanbul, Turkey 34349_
### 1.1 Abstract
Electropolymerization plays a critical role in the electrochemical systems. In this chapter, we address such role within the context of interplay between kinetics and energetics. The trains of chin radical reactions leads to the formation of thin films in electrochemical devices. The structure of so-called solid electrolyte interphase (SEI) during the initial charge/discharge cycles of the device of any kind (i.e. rechargeable battery) on the surface of electrode directly controls the the ultimate stability and longevity. In this chapter, we study the morphological evolution of SEI, both in termsof transport and thermodynamics within quantitative and qualitative contexts.
### 1.2 Introduction
Electropolymerization plays an important role in the operation of rechargeable batteries in portable electronics and electric vehicles. As an alkaline metal can react with the most organic solvents, a surface film is formed during the initial charging/discharging processes. This electrically insulating and ionically conductive interface is named as the solid electrolyte interphase (SEI).
Electropolymerization typically occurs in the double layer region (DL herein after). Also called as electrical double layer (EDL), such structure appears on the surface of an object when it is exposed to a fluid. The object might be a solid particle, a gas bubble, a liquid droplet, or a porous body. The DL refers to two parallel layers of charge surrounding the object. The first layer, the surface charge (either positive or negative), consists of ions adsorbed onto the object due to chemical interactions. The second layer is composed of ions attracted to the surface charge via the Coulomb force, electrically screening the first layer. This second layer is loosely associated with the object. It is made of free ions that move in the fluid under the influence of electric attraction and thermal motion rather than being firmly anchored. It is thus called the "diffuse layer".
Interfacial DLs are most apparent in systems with a large surface area to volume ratio, such as a colloid or porous bodies with particles or pores (respectively) on the scale of micrometers to nanometers. However, DLs are important to other phenomena, such as the electrochemical behavior of electrodes.
This layer and includes various organic and inorganic components. On one hand, the formation of the SEI intrinsically consumes the anode and electrolyte, leading to a low efficiency. Consequently, the SEI effectively prevents the further physical contact between Li and the solvent, therefore making Li dynamically stable in certain organic electrolytes. In particular, the SEI can adjust the distribution of Li ions from the bulk electrolyte to the anode. This layer is merely the result of competitive desolvation of ionic compounds on the organic electrolytes.
The SEI ultimately covers the Li electrodes in multilayer surface films composed of organic or inorganic Li salts. Thereby, applying an electrical field to Li electrodes enables electrochemical Li dissolution and deposition to occur through these surface films.
The Figure 1.1 shows that upon formation of SEI layer, it interferes the morphology of deposition and therefore the upcoming lithium ions cannot afford to perform uniform deposition.
Looking closer to the morphology, SEI is thin and fragile film and stabilizes the redox reaction on the electrode surface. While this film doesn't let bigger organic compounds to reduce further, it is conductive to smaller charge carrier candidate ions such as lithium (\(Li^{+}\)), Nickel (\(Ni^{+}\)), Magnesium (\(Mg^{2+}\)) or Zinc (\(Zn^{+}\)).
The morphology of SEI is highly effective on the rechargeable lithium metal batteries as an optimal energy storage devices. \(Li^{0}\) has an excep
Figure 1.1: Various scenarios of SEI formation on lithium electrodes from organic carbonates.
Inhibition of such microstructures hinge on SEI generated from the decomposition of most organic solvents at the negative potentials required to reduce \(Li^{+}\). The importance generally ascribed to SEI is most objectively attested by recent reports which emphasized that '..._SEI formation is the most crucial and least understood phenomena impacting battery technology..'_, and '.._constructing stable and efficient SEI is among the most effective strategies to inhibit the dendrite growth and achieve superior cycling performance...'_. Previous attempts at improving SEI properties have variously resorted to '..._electrolyte additives and surface modification of the cathode...(which) have been shown to improve the formation of an effective SEI layer...'_ and led to the conclusion that '.._the formation of the SEI depends largely on electrode materials, electrolyte salts, and solvents involved...'_.
The composition and structure of SEI have also been intensively investigated by diverse techniques, such as XPS, solid state NMR, ellipsometry, sum-frequency generation spectroscopy, electron microscopies, neutron scattering, AFM, electron paramagnetic spectroscopy and matrix assisted laser desorption ionization (MALDI) time of flight mass spectrometry. Recent reviews, however, have acknowledged that '..._many strategies have been proposed to modify SEI structure. However, the modifying process is still out of control in a bulk cell because the thickness, density and ion conductivity cannot yet be rationally designed_'.
One interpretation of this impasse is that SEI properties depend not only on initial conditions, such as electrode materials, electrolyte salts, solvents and additives, but on the procedure by which SEI are generated.
\begin{table}
\begin{tabular}{|c|c|} \hline
## I
& \(1MLiCLO_{4}\) \\ \hline
## II
& \(0.9M\)\(LiClO_{4}+0.1M\)\(LiF\) \\ \hline
## III
& \(0.99M\)\(LiClO_{4}+0.01M\)\(LiF\) \\ \hline \end{tabular}
\end{table}
Table 1.1: Electrolyte compositions.
## Chapter 1 Electrodes and Double Layers
### 1.1 Introduction
The _Cv_ is a _Cv_-type (Cv) system of _Cv_-type (Cv) systems of _Cv_-Thus, if the mechanisms of generation that would allow us to rationally design SEI are still elusive it is simply because mechanisms cannot be deduced from information on initial and final states alone. Here, we address this issue in study of the kinetics of electropolymerization of propylene carbonate (PC) into SEI on metal electrodes, in conjunction with a fundamental analysis of the results obtained. Our goal is to gain insight into the mechanism of SEI generation.
The comparative electropolymerization has been performed through 3 electrolytes given in Table 1.1. We investigate via cyclic voltammetry, impedance spectroscopy and chronoamperometry the role of kinetics in controlling the properties of the SEI generated from the reduction of propylene carbonate.
Figure 1.3: Nyquist diagrams of \(Li|electrolyte|Li\) cells at OCV after getting charged galvanostatically at \(0.05mA/cm^{2}\) for various charge amounts. **A**: with \(1M\)\(LiClO_{4}\). **B**: with addition of \(+0.1M\)\(LiF\). A insets: the equivalent circuit, lithium dendrites that short-circuited the cell.
### 1.3 Cyclic Voltammetry
The CV diagrams for the symmetric cells of electrolytes given in Table 1.1 with two scanning rates are shown in Figure 1.2. We chose \(ClO_{4}^{-}\) because it is a stable, weakly coordinating anion and \(F^{-}\) because it is shown to improve the electropolymerization morphology and PC is widely used as a base solvent. In figure 1.2A and 1.2B, peaks between \(0.8V\) and \(1V\) correspond to PC reduction (PCR hereafter). Peaks at \(0.3V\) are assigned to the underpotential deposition (UPD) of \(Li^{0}\) on the basis of reported similar peaks within \(0.4\)-\(0.6V\), and the fact that the peak at 1.3 V associated with the anodic stripping of UPD \(Li^{0}\) deposits does not appear following cathodic scans that were reversed at 0.9 V to avoid \(Li^{0}\) deposition.
### 1.4 Electrochemical Impedance Spectroscopy
The electrical characteristics of the SEI produced can be analyzed by electrochemical impedance spectroscopy (EIS). The EIS measurements can be performed at open circuit voltage (OCV). Therefore they only provide information about the static electrical properties of preformed SEI. A typical diagrams consist of a single depressed semicircle at medium and high frequencies, which merges at low frequencies into a straight line associated with \(Li^{+}\) diffusion through SEI layers. The presence of a single semicircle excludes significant contributions from multiple SEI layers. Thus SEI properties can be accounted by a single layer despite their complex, heterogeneous morphology and chemical composition. In the equivalent circuit shown in the diagram \(R_{Bulk}\) is the sum of ohmic drops across the electrolyte and other cell components, \(R_{SEI}\) and \(C_{PE-SEI}\) are the resistance and capacitance of preformed SEI layers, and \(W\) is the impedance arising from \(Li^{+}\) diffusion through SEI layers.
The implication is that the SEI production is followed by homogeneous chemical reactions that incorporate substantial amounts of additional PC into the layers. Constant phase element \(Z_{r}\) vs \(\omega\) plots:
\[Z_{r}=Z_{0}+\sigma\omega^{-n^{\prime}} \tag{1.1}\]
The morphology of SEI layers is evidently sensitive to scan and \(PCR\) rates. The kinetics of SEI formation. \(Li^{+}\) diffusion coefficient, \(D^{+}\) in the SEI formed is obtained from
\[D_{Li}^{+}=\frac{R^{2}T^{2}}{2\theta^{2}n^{4}F^{4}[Li^{+}]^{2}\sigma^{2}} \tag{1.2}\]
The resistance of SEI layers is directly proportional to thickness l, and electrical resistivity \(\rho\):
\[R_{SEI}=\frac{\rho l}{\theta} \tag{1.3}\]
In contrast, the Nyquist diagrams of SEI grown at \(v=0.5^{mV}/_{s}\) (Figure 1.2D) are qualitatively and quantitatively different from those in Figure 1.2C.
\begin{table}
\begin{tabular}{|c|c c c c c c c c|} \hline CV cycle & Electrolyte & R-Bulk (\(\Omega\)) & R-SEI (\(\Omega\)) & C-SEI (\(10^{5}\Omega\)s\({}^{n}\)) & n & \(\sigma\)(\(10^{4}\Omega^{-1}\)s\({}^{n^{\prime}}\)) & \(n^{\prime}\) & \(D_{Li^{+}}\)(\(10^{13}cm^{2}s^{-1}\)) \\ \hline \multirow{2}{*}{\(1^{st}\)} & I & 4 & 543 & 1.3 & 0.74 & 2.1 & 0.62 & 3.0 \\ & II & 3.1 & 459 & 1.5 & 0.73 & 3.6 & 0.60 & 1.1 \\ \hline \multirow{2}{*}{\(5^{th}\)} & I & 4.7 & 846 & 1 & 0.76 & 2.1 & 0.67 & 2.9 \\ & II & 3 & 489 & 1.6 & 0.72 & 3.9 & 0.62 & 0.9 \\ \hline \end{tabular}
\end{table}
Table 1.2: Equivalent circuit parameters from spectra of Figure 1.2C.
The presence of fluoride has a significant effect on the long-term stability of electropolymerization upon galvanostatic charging at \(0.05^{mA}/_{cm2}\). Note that \(PC\) and \(Li^{+}\) are simultaneously reduced during galvanostatic charging. Figures 1.2A and 1.2B show the evolution of Nyquist diagrams as functions of circulated charge. Noteworthy is the fact that the resistance of cells filled with **II** (containing \(F^{-}\)) decreases by only 25% after the circulation of \(Q>17^{C}/_{cm^{2}}\), whereas the resistance of cells filled with I (without \(F^{-}\)) already drops eightfold at \(Q>5^{C}/_{cm^{2}}\), as an indication that \(Li^{0}\) dendrites had pierced SEI layers, reached the cathode and short-circuited the cell. Fluoride additions also enhance the persistence of electropolymerization.
### 1.5 Chronoamperometry
PCR at slow scan rates generates PC-impermeable SEI layers (Figures 1.2C and 1.2D) led us to test the dependence of PCR rates on applied potential by growing SEI under potentiostatic conditions. CA experiment at 1.0, 1.1 and 1.7 V (vs \(Li^{+}/Li^{0}\))) applied potentials in cells filled with electrolyte I are shown in Fig. 1.4A. Faradaic currents associated with PCR (i.e., those circulating after the decay of initial capacitive currents) markedly increase at more negative overpotentials: \(\eta=E-E_{p}\) (PCR rates peak at \(E_{p}\sim 1.3V\), Fig. 1.4B), as expected.
\begin{table}
\begin{tabular}{|c|c c c c c c c c|} \hline CV cycle & Electrolyte & R-Bulk (\(\Omega\)) & R-SEI (\(\Omega\)) & C-SEI (\(10^{5}\Omega^{5n}\)) & n & \(\sigma(10^{4}\Omega^{-1}s^{n^{\prime}})\) & \(n^{\prime}\) & \(D_{Li^{+}}(10^{13}cm^{2}s^{-1})\) \\ \hline \multirow{2}{*}{\(1^{st}\)} & I & 5.3 & 5439 & 2.7 & 0.79 & 3.2 & 0.50 & 1.3 \\ & II & 4.3 & 503 & 2.9 & 0.79 & 3.2 & 0.49 & 1.2 \\ \hline \multirow{2}{*}{\(5^{th}\)} & I & 4.7 & 543 & 3.1 & 0.82 & 3.0 & 0.46 & 1.5 \\ & II & 5.6 & 534 & 2.8 & 0.82 & 3.2 & 0.46 & 1.3 \\ \hline \end{tabular}
\end{table}
Table 1.3: Equivalent circuit parameters from spectra of Figure 1.2D.
SEI layers, currents circulating in the CA at a \(\eta=1.7V-1.3V=0.4V\) underpotential vanish after \(\sim 5000s\), in contrast with experiments carried at \(1.0V\) and \(1.1V\). Past the initial stages where currents are partially due to the capacitive charging of double layers (and also at \(i>50\mu A\), partially controlled by PC desolvation, cf. Fig. 1.2A and 1.2B), the slopes of faradaic currents vs \((time)^{-1/2}\):
\[i=\frac{nF\theta[PC]\sqrt{D_{PC}}}{\sqrt{\pi t}} \tag{1.4}\]
those grown at \(1.7\) V: \(D_{PC}(1.7V)=7.7\times 10^{-17}cm^{2}s^{-1}\), which are compatible with the \(D_{PC}\sim 10^{-12}\) to \(10^{-16}cm^{2}s^{-1}\) values reported in porous and compact SEI layers, respectively. Note that Eq. 1.4 for PC diffusion through a growing solid SEI layer is the analogue of Cottrell's equation for ion diffusion through a widening, solvent-filled double layer. In both cases layer thicknesses increase with \(t^{1/2}\), and the corresponding current densities decrease with \(t^{-1/2}\). Most remarkably, \(D_{PC}(1.7V)\) is \(\sim 1100\) times smaller than \(D_{PC}(1.0V)\) through SEI layers that were seeded by a small fraction of the charge: \(Q_{1.7V}/Q_{1.0V}=0.04\) (\(Q=\int Idt\))
The relevant electrochemical characteristics of the SEI layers grown po
\begin{table}
\begin{tabular}{|c c c c c c c c|} \hline V (vs \(Li^{+}/Li^{0}\)) & Electrolyte & \(R_{Baik}(\Omega)\) & \(R_{SEI}(\Omega)\) & \(C_{SEI}(10^{6}\Omega s^{n})\) & n & \(\sigma(10^{4}\Omega^{-1}s^{n^{\prime}})\) & \(n^{\prime}\) & \(D_{Li^{+}}(10^{13}cm^{2}s^{-1})\) \\ \hline \multirow{2}{*}{\(1.0\)} & 1st & 30 & 800 & 0.89 & 0.77 & 0.35 & 0.85 & 110 \\ & 2nd & 35 & 1641 & 0.77 & 0.72 & 0.34 & 0.90 & 120 \\ \hline \multirow{2}{*}{\(1.1\)} & 1st & 14 & 920 & 0.95 & 0.73 & 0.79 & 0.78 & 21 \\ & 2nd & 15 & 1336 & 0.72 & 0.73 & 0.71 & 0.77 & 26 \\ \hline \multirow{2}{*}{\(1.7\)} & 1st & 16 & 616 & 1.8 & 0.79 & 1.4 & 0.73 & 7.3 \\ & 2nd & 16 & 565 & 1.6 & 0.80 & 2.0 & 0.71 & 3.4 \\ \hline \end{tabular}
\end{table}
Table 1.4: Equivalent circuit parameters from spectra of Figure 1.2A. 1.2C.
Nyquist diagrams of the SEI layers produced in successive chronoamperometry experiments under 1.0, 1.1 and 1.7 applied potentials are shown in Fig. 1.4A. The parameters derived from their analysis are compiled in Table 1.4. It is apparent that the SEI produced in the first 1.7 V potentiostatic experiment does not grow upon further charging, in contrast with those produced at 1.0 and 1.1 V. This conclusion is corroborated by CV scans (Fig. 1.4B).
Inspection of Table 1.4 reveals that:
1. \(R_{SEI}\) of the layers grown in initial cycles is comparable values despite the fact that the amount of PC reduced from electrolyte. \(R_{SEI}\) remains nearly constant in 1.7 V experiments but increases by a factor of 2 in the second CA at 1.0 V.
2. \(C_{SEI}\) at 1.7 V is about \(\times 2\) times larger than those at 1.0 and 1.1 V suggesting (since \(C\propto\frac{1}{thickness}\)) that they are about half as thick.
3. \(Li^{+}\) diffusion, with \(n_{0}>0.7>0.5\), is anomalous in all cases. Noteworthy is that \(D^{+}\) for SEI layers produced at 1.7 V is comparable to the \(D^{+}\) values in the SEI obtained in potentiodynamic CV experi
Figure 1.4: **A**: Chronoamperograms in Cu \(\mid\) electrolyte \(\mid\) Li cells filled with electrolyte **1** under 1.0\(V\), 1.1\(V\) and 1.7\(V\) applied voltages (vs 1M \(Li^{+}/Li^{0}\) in PC). **B**: Cottrell current vs. \(t^{-0.5}\) plots (Eq. 1.4).
Summing up, the above findings are consistent with:
1. SEI layers that incorporate PC molecules in larger numbers than those undergoing reduction at the electrode surface, i.e., SEI are essentially polymer materials.
2. SEI properties strongly depend on the kinetics of the generation process.
Since SEI behave as polymeric materials, our findings suggest that the potential impact of experimental conditions on their properties should be evaluated on the basis of polymer science concepts.
Figure 1.5: **A**: Nyquist diagrams at open circuit voltage of Cu | electrolyte | Li cells filled with electrolyte I after the first and second chronoamperometries at 1.0 V, 1.1 V and 1.7 V applied voltages vs. 1 M \(Li^{+}/Li^{0}\) in PC. Dotted line: Warburg’s \(n^{\prime}=0.5\) slope as a reference. **B**: CV at \(v=5mV/s\) in cells with electrolyte I after being charged potentiostatically at 1.0 V, 1.1 V and 1.7 V for 2 h.
We show that slow initiation rates via one-electron PC reduction at underpotentials consistently yields compact, electronically insulating, Li+-conducting, PC- impermeable SEI films.What to expect for SEI generated in a polymerization process initiated by PC reduction, reaction eq:PC:
\[PC+e^{-}\to PC^{-} \tag{1.5}\]
Following previous reports, PC is deemed to open its ring into an alkoxycarbonyl radical followed by decomposition into \(CO\) or \(CO_{2}\), plus simpler radical anions, \(X^{-}\), which initiates radical chain-growth polymerizations propagated by reactions 1.6:
\[X^{-}+PC\to X-(PC)^{-}\rightarrow\to X-(PC)^{-}_{n} \tag{1.6}\]
and terminated via bimolecular radical recombination, reaction 1.7:
\[X-(PC)^{-}_{n}+X-(PC)^{-}_{m}\to X_{2}-(PC)^{2-}_{n+m} \tag{1.7}\]
SEI permeability, ionic and electronic conductivity, solubility and mechanical properties are essentially determined by the degree of solvent polymerization \(\lambda\), i.e., by the number of monomers incorporated into polymer units. \(\lambda\) is controlled by the competition between radical propagation (Reaction 1.6) vs. radical termination (Reaction 1.7). Thus we arrive at Reaction 1.8:
\[\lambda=\frac{k_{2}[PC^{-}][PC]}{2k_{3}[PC^{-}]^{2}}=\frac{k_{2}[PC]}{2k_{3}[ PC^{-}]} \tag{1.8}\]
Because initiation rates \(r_{i}\) and termination rates balance at steady state, we have:\[r_{i}=2k_{3}[PC^{-}]^{2}\Rightarrow[PC^{-}]=\left(\frac{r_{i}}{2k_{3}}\right)^{0.5} \tag{1.9}\]
Therefore we arrive at Reaction 1.10:
\[\lambda=\frac{k_{2}[PC]}{\sqrt{2k_{3}r_{i}}} \tag{1.10}\]
The fundamental \(\lambda\propto r^{\frac{1}{2}}\)relationship for a radical chain polymerization should apply whether \(Li^{+}\) are present in the SEI, as in the present case, or not.
Thus, our CA experiments at 1.7 V are deemed to produce functional SEI layers because the low current densities exclusively associated with PC reduction provide the slow initiation rates required to generate long polymerization chains. Furthermore, as a result, the overall slow polymerization process they bring about may not be limited by the availability of the free PC monomers released from the slow desolvation of \(Li(PC)^{+}_{n}\). The very low value of the PC diffusion coefficient \(D_{PC}(1.7V)=7.7\times 10^{-17}cm^{2}s^{-1}\) determined in the SEI generated at underpotential is clearly consistent with transport through a compact material comprising few, long and possibly linked or intertwined polymer chains. From this perspective, the PC reduction rates at the \(\sim 1V\) overpotentials prevailing under conventional LMB charging conditions, where the full voltage required to plate the anode is applied from the onset, may not be ideal because they are likely to generate short, disjoint polymer domains rather than compact, interconnected polymer films extending over the electrode surface. We believe that our results and analysis provide new insights into the outstanding questions formulated in a recent review on the subject: 'how does SEI form?' and 'what parameters control SEI properties?'.
### 1.6 Experimental
Experiments were performed in two types of electrochemical cells filled with electrolyte solutions I, II and III of three different compositions (Table 1). Studies on SEI layers were carried out in \(Cu|electrolyte|Li\) coin cells whereas the deposition of \(Li^{0}\) films was investigated in \(Li|electrolyte|Li\) coin cells. Round disk electrodes (\(A=1.6cm^{2}\)) were punched from \(Li^{0}\) foil (Aldrich, 99.9%, \(0.38mm\) thick) that had been polished by scraping with a blade and rinsed with dimethyl carbonate. Electrodes were mounted on a transparent poly-methyl methacrylate separator that kept them \(L=3.175mm\) apart. All operations were carried out in a glove box sparged with argon. Chronoamperometry (CA), electrochemical impedance spectroscopy (EIS) and cyclic voltammetry (CV) measurements were made with a Bio-Logic VSP potentiostat. Galvanostatic experiments were performed with an ARBIN BT2000 battery tester. EIS experiments (\(5mV\) modulation signal amplitude) covered the \(100mHz\) to \(1MHz\) frequency range. Impedance data were analyzed using _Zview_ software. All reported potentials are relative to \(Li^{+}/Li^{0}\) under working conditions. \(Li^{0}\) and \(Cu^{0}\) foils (Sigma-Aldrich) were used as-received. Lithium per- chlorate (\(LiClO_{4}\), Aldrich, battery grade, 99.9%) and lithium fluoride (\(LiF\), Aldrich, 99.99% trace metal basis) were dried at 90C under vacuum for \(24h\) and dissolved in propylene carbonate (PC) (Aldrich, 99.7% anhydrous). Further details can be found in previous publications from our laboratory.
|
10.48550/arXiv.1904.08040
|
Electropolymerization: Studies and Applications / Electrodes and Double Layers
|
Asghar Aryanfar, Agustin J. Colussi, Laleh M. Kasmaee, Michael R. Hoffmann
| 3,592
|
10.48550_arXiv.1101.2253
|
## 1 Introduction
R Coronae Borealis stars (RCBs, hereafter) stars are a rare class of supergiants whose atmospheres are extremely hydrogen-deficient - the H-deficiency ranges from about a factor of 10\(-\)100 to at least 10\({}^{8}\) - with helium the most abundant element and often the stars are carbon-rich (e.g., Lambert & Rao 1994). RCBs at unpredictable intervals form thick carbon dust clouds which if they form above the Earth-facing stellar surface can obscure the star causing a decrease up to eight magnitudes in the visual band and a decline can last from a few weeks to many months.
The RCB's hydrogen deficiency together with the helium- and carbon-rich character of the gas may facilitate the formation of molecular and dust species not seen in circumstellar envelopes of normal stars. In particular, these envelopes have been considered to be possible environments for the formation of the buckminsterfullerene molecule C\({}_{60}\) (e.g., Goeres & Sedlmayr 1992). The remarkable stability of C\({}_{60}\) against intense radiation, ionization, etc. (e.g., Kroto 1987) reinforces the idea that fullerenes such as C\({}_{60}\) could be present in the interstellar medium were they formed in and ejected from circumstellar envelopes of one or more kinds of mass-losing star.
Are RCB stars a source of C\({}_{60}\) molecules? Early observations of three RCB stars (R CrB, RY Sgr, and V854 Cen) at a resolution R=1000 around 8.6\(\mu\)m searched for the 8.4\(\mu\)m C\({}_{60}\) feature and reported a negative result (Clayton et al. 1995). With the advent of the Spitzer Space Telescope, we undertook a new search for C\({}_{60}\) around RCBs. Identification of the C\({}_{60}\) molecule can be done in the infrared domain, where there are infrared transitions centered at \(\sim\)7.0, 8.4, 17.4 and 18.8 \(\mu\)m according to gas-phase laboratory spectroscopy (Frum et al. 1991; Nemes et al. 1994). In this paper, we report two results. The first is that fullerenes are not seen around most RCBs. Thesecond is that the C\({}_{60}\) IR transitions are present along with transitions due to PAHs in the spectrum of DY Cen and possibly also in the spectrum of V854 Cen, which are the two least H-deficient RCBs known. Spitzer observations have very recently provided evidence for C\({}_{60}\) and C\({}_{70}\) from planetary nebulae (PNe; Cami et al. 2010; Garcia-Hernandez et al. 2010) and reflection nebulae (Sellgren et al. 2010). None of these environments is H-deficient. Although Cami et al. declare their observations refer to a H-poor region of the PN Tc 1, the literature does not support their claim (see below). These detections of fullerenes with our detection of C\({}_{60}\) from RCBs with a modicum of H suggests that formation of fullerenes may require some H in addition to C.
## 2 Spitzer observations and infrared spectra
We have recently conducted an infrared spectral survey with the Infrared Spectrograph (IRS) on the Spitzer Space Telescope for a complete sample (\(\sim\)30) of RCB stars spanning a full range of hydrogen content, temperature and composition. The spectral energy distributions (\(\sim\)0.4\(-\)40 \(\mu\)m) were constructed for all RCB stars in this sample (see Garcia-Hernandez et al. 2011 for more details). Since here we are interested in the emission and absorption features present in the 6-25 \(\mu\)m window, we interpolate between several points in the dust continuum and subtract this baseline to provide the residual spectrum, where the dust and gas features may be easily identified.
Residual spectra of those RCBs with a H-deficiency in excess of a factor of about 10\({}^{3}\) (Garcia-Hernandez et al. 2011) show only a broad \(\sim\)6-10 \(\mu\)m emission feature which is attributable to C-C stretching modes of amorphous carbon grains (Colangeli et al. 1995). The profile of this feature may vary slightly from RCB-to-RCB but this variation may not entirely be intrinsic to the circumstellar envelopes. Extension immediately longward of 10\(\mu\)m is sensitive to the adopted interstellar reddening which determines the correction for the 9.7\(\mu\)m silicate absorption feature. shows the average residual spectrum for the sample of the least reddened (E\({}_{B-V}\)\(\sim\)0.05\(-\)0.40) RCB stars (UW Cen, RT Nor, RS Tel, V CrA, V1157 Sgr, V1783 Sgr, S Aps, U Aqr, and Z Umi; Garcia-Hernandez et al. 2011).
The RCB stars DY Cen and V854 Cen display apparently similar but very different residual spectra from the majority of the RCBs. DY Cen and V854 Cen are the two stars in our sample that are the least H-deficient with H-deficiencies of factors of \(\sim\)10 for DY Cen (T\({}_{eff}\)=19500K; Jeffery & Heber 1993) and \(\sim\) 100-1000 for V854 Cen (T\({}_{eff}\)=6750K; Asplund et al. 1998). DY Cen and V854 Cen display strong features at \(\sim\)6.3, 7.7, 8.6, and 11.3 \(\mu\)m together with weaker features at \(\sim\)11.9 and 12.7 \(\mu\)m, all of which may be identified with polycyclic aromatic hydrocarbons (PAHs) (e.g., Allamandola et al. 1989; Bauschlicher et al. 2008, see Table 1). The PAH features (with the exception of the strong 7.7 \(\mu\)m band) are quite narrow (\(\sim\)0.2\(-\)0.4 \(\mu\)m) showing that they probably arise from free gas-phase PAHs, not PAHs in particles or clusters (Peeters et al. 2004). In addition, the wavelength positions of the strong 7.7 \(\mu\)m band are 8.0 and 7.8 \(\mu\)m for V854 Cen and DY Cen, respectively. The mean position of the 7.7 \(\mu\)m band is correlated apparently with effective temperature of the exciting star (see in Tielens 2008): the 7.8 to 8.0 \(\mu\)m shift between the two stars is consistent with the difference in their stellar temperatures.
The intriguing result from is the presence in the DY Cen residual spectrum of three features at \(\sim\)7.0, 17.4 and 18.8 \(\mu\)m not readily attributable to PAHs according to wavelength and/or intensity considerations. These features, which are real as they are detected in available low- and high-resolution Spitzer spectra at all slit positions, are attributable to C\({}_{60}\), as we show below. To establish their carrier as C\({}_{60}\), it is necessary to discuss the anticipated spectra of the fullerene and their possible blending PAHs. Table 1 lists features seen in DY Cen and V854 Cen as well as their identification.
## 3 Identifications: C\({}_{60}\) and/or PAH?
Laboratory gas-phase spectroscopy of neutral C\({}_{60}\) molecules was reported by Frum et al. and Nemes et al.. Infrared C\({}_{60}\) features are expected at \(\sim\)7.0, 8.4, 17.4 and 18.8 \(\mu\)m at a temperature of 0 K while these wavelengths are predicted to shift longward a maximum of 0.2 \(\mu\)m at a temperature of 1083 K (Nemes et al. 1994). Thereare apparently no reliable estimates of the relative strengths of the four bands in the gas phase. Cami et al. estimated Einstein A-values from published band absorption strengths for C\({}_{60}\) in rare gas matrices: these values according to our calculations are A(s\({}^{-1}\)) \(\simeq\) (1.9, 1.1, 4.2, 5.2) for the bands (18.8, 17.4, 8.4, 7.0). Intensities of circumstellar features will depend also on the level populations and radiative transfer effects.
There is a plethora of astronomical, laboratory, and theoretical data on the PAHs. Our primary argument applicable to DY Cen is that three of the four C\({}_{60}\) features are not blended beyond recognition with PAH features and that a reasonable case may be made that the fourth feature at 8.4 \(\mu\)m is a blend of the C\({}_{60}\) transition and the 8.6 \(\mu\)m PAH feature.
In terms of upper state excitation energy, the C\({}_{60}\) bands are ordered by decreasing wavelength and it is in this order that we discuss their presence in DY Cen. The two longest wavelength C\({}_{60}\) features match emission features in DY Cen at 17.40 \(\mu\)m and 18.98 \(\mu\)m in good agreement with laboratory measurements at about 1000K and with their extrapolation to 0K; one expects the temperature of the C\({}_{60}\) molecules to be within these limits. The line widths (dominated by the instrumental width) are consistent with the laboratory measurements, 0.31 and 0.36 \(\mu\)m (observed) versus 0.40 \(\mu\)m (laboratory).
In DY Cen's spectrum there are no other features between 15 and 30 \(\mu\)m but V854 Cen shows the 17.4 \(\mu\)m feature not only stronger than its 18.8 \(\mu\)m counterpart but shifted to shorter wavelengths (17.33 \(\mu\)m) and accompanied by emission extending to about 15 \(\mu\)m. These additional features are presumed to be PAH features known to fall in the 13 to 19 \(\mu\)m interval (Boersma et al. 2010). Except for a 16.4 \(\mu\)m feature, the intensities of these PAH contributions are uncorrelated with the stronger PAH features in the 5 to 13 \(\mu\)m interval. Additionally, the relative intensities of features in the 13 to 19 \(\mu\)m region may vary from object to object. Often, the 16.4 \(\mu\)m feature is the strongest in this window with an intensity correlated with that of the 11.2 \(\mu\)m PAH. With a slight extrapolation of Boersma et al's showing a correlation between the intensity ratio of 6.2/11.2\(\mu\)m and 16.4/11.2\(\mu\)m PAH features, the 16.4\(\mu\)m feature is predicted to have an intensity 2% that of DY Cen's 11.2\(\mu\)m feature, an expectation consistent with the feature's absence and, since other PAH features in this interval are expected to be weaker, it is not surprising that the C\({}_{60}\) longest wavelength transitions appear uncontaminated by PAH blends. For V854 Cen, the 6.2/11.2\(\mu\)m intensity ratio is higher than for DY Cen and predicts an intensity of about 5% for the 16.4\(\mu\)m PAH, a value approximately consistent with the presence of the 16.4\(\mu\)m and other weak PAH features in With a correction for weak 15.8, 16.4, and 17.0 \(\mu\)m PAH contaminants, a C\({}_{60}\) feature at about 17.4 \(\mu\)m is obtained that is consistent with that of the 18.8 \(\mu\)m feature.
Interestingly, DY Cen shows a unique spectrum across the 15 to 20 \(\mu\)m interval among spectra exhibiting PAH features. Other features at 15.8, and 16.4 \(\mu\)m are not seen, as they are, for example, in reflection nebulae such as NGC 7023 (Sellgren et al. 2007, 2010; Boersma et al. 2010). Several PAHs from the Ames spectral database show features near 17.4 and 19 \(\mu\)m but they are always accompanied by other stronger features (e.g., at 16.4 \(\mu\)m: Bauschlicher et al. 2008) but these features are absent from DY Cen's spectrum. Some may be present in V854 Cen's spectrum. Sellgren et al. show that the 17.4 \(\mu\)m feature has two components: one that correlates with the 18.9 \(\mu\)m C\({}_{60}\) feature and another one correlating with the 16.4 \(\mu\)m PAH feature. The absence of the 16.4 \(\mu\)m PAH feature from the DY Cen spectrum indicates that the 17.4 \(\mu\)m feature is dominated by C\({}_{60}\) emission. However, in the case of V854 Cen since both 16.4 and 18.9 \(\mu\)m features are seen, the 17.4 \(\mu\)m feature is due to a combination of C\({}_{60}\) and PAH emission.
The 7.0 and 8.4 \(\mu\)m C\({}_{60}\) transitions are in the interval spanned by common PAH features at 6.2, 7.7, and 8.5 \(\mu\)m. Of particular concern is the blending of the 8.4\(\mu\)m C\({}_{60}\) and 8.5\(\mu\)m PAH features where the PAH's contribution can be assessed only by comparison of relative strengths of this and adjacent PAH features, an uncertain exercise owing to considerable source-to-source variation in relative strengths.
The 7.0\(\mu\)m C\({}_{60}\) line is partially resolved in the DY Cen spectrum with a measured wavelength of 7.0 \(\mu\)m in good agreement with the laboratory spectroscopic value of 7.11 \(\mu\)m for a temperature of about 1000K. An estimate of the feature's width requires a correction for the overlapping wing of the strong 7.7\(\mu\)m PAH feature: a rough upper limit for the FWHM of the C\({}_{60}\) line is \(<\)0.16 \(\mu\)m, a value consistent with the laboratory measurement of 0.06 \(\mu\)m. In addition, we estimate an intensity ratio 7.0/18.9 of \(\sim\)0.7, which - according to Sellgren et al. - corresponds to excitation of C\({}_{60}\) molecules by photons with energies slightly less than 10 eV, in agreement with the effective temperature of DY Cen. The 8.4\(\mu\)m C\({}_{60}\) line is blended with the 8.5\(\mu\)m PAH. There seems to be no way in which to make a firm estimate of the PAH's contribution to the feature seen in DY Cen's spectrum. Cerrigone et al. provide intensities for the principal PAH features for both C-rich and O-rich post-AGB stars. Relative to the 11.2\(\mu\)m PAH, the mean intensity of the 8.5\(\mu\)m PAH is 0.40 for the four C-rich stars (range 0.19 to 0.91), and 0.38 for the entire sample of 13 stars (range 0.04 to 0.91). At the relative intensity of 0.4, the 8.5\(\mu\)m PAH is expected to have an intensity close to the observed value. However, with respect to the 6.2\(\mu\)m PAH, the mean intensity of the 8.5\(\mu\)m feature is 0.33 for the four C-rich stars (range 0.23 to 0.5) and 0.19 for the entire sample (range 0.05 to 0.5). With these mean intensities, the predicted intensity of the 8.5\(\mu\)m PAH is less than the observed value in the DY Cen spectrum. Different average relative intensities among PAH features very likely reflect the different origins of the features (Table 1). Sellgren et al.'s calculations indicate that the 7.0 and 8.4 \(\mu\)m C\({}_{60}\) features should be roughly similar in intensity. The predicted intensity of the 8.5\(\mu\)m PAH from the 11.2\(\mu\)m PAH would satisfy this expectation but the prediction from the 6.2\(\mu\)m PAH would not. However, the considerable star-to-star variation in relative intensities of these PAHs rules out making a reliable separation of C\({}_{60}\) and PAH contributions.
Inspection of V854 Cen's spectrum in Fig.2 reveals some differences with DY Cen's spectrum in the 6 to 10 \(\mu\)m interval. In particular, the strong 7.7\(\mu\)m PAH shows obvious blending to longer wavelengths - this feature is broader in V854 Cen than in DY Cen. This may result from intrusion by an additional PAH but another possibility is that V854 Cen includes a contribution from the broad feature seen in the more H-deficient RCBs. In Fig.3, we show the V854 Cen spectrum and the RCB feature scaled to provide an approximate possible fit. The profile of the RCB feature seems to include some contributions at 7 and 8.5\(\mu\)m. In contrast to V854 Cen, the DY Cen spectrum does not appear to be contaminated by the RCB feature and additional contributions at 7.0 and 8.5 \(\mu\)m are required which are most probably attributable to C\({}_{60}\).
## 4 Discussion
The original laboratory studies on the formation of fullerenes showed that fullerenes are clearly favored in environments which are H-deficient (Kroto et al. 1985; Kratschmer et al. 1990) and that H-poor conditions are a prerequisite for efficient fullerene formation (de Vries et al 1993). Thus, in the circumstellar envelopes of cool evolved stars having a normal H abundance (e.g., C-rich Asymptotic Giant Branch stars) and in dense interstellar clouds, acetylene (C\({}_{2}\)H\({}_{2}\)) and its radical derivatives are believed to be the precursors of complex C-based molecules such as PAHs, and fullerenes are probably not formed (e.g., Cherchneff & Cau 1999). However, in hydrogen-poor but C-rich environments, fullerene molecules may be formed from the coalescence of large monocyclic rings in the gas phase and PAHs are likely not formed as possible intermediates (e.g., Cherchneff et al. 2000). Furthermore, more recent laboratory studies show that at low temperatures (\(<\) 1700 K) soot formation proceeds through or involves the formation of PAH intermediaries while fullerenes are involved at temperatures in excess of 3500 K (Jager et al. 2009). This shows that the high temperatures rather than the H-poor conditions may also be a defining factor for efficient fullerene formation. Given the ease with which fullerenes are formed in a carbon and helium rich atmosphere in laboratory experiments, it is puzzling that fullerenes do not seem to form in great abundance in the carbon and helium rich environments of the very H-poor RCB stars, but, in fact, fullerenes are detected only around stars containing some hydrogen (this study and Garcia-Hernandez et al. 2010). Evidently, fullerenes formation is inefficient in the highly H-deficient RCB stars; fullerene destruction is expected to occur at a slow rate.
Our detection of C\({}_{60}\) around DY Cen and possibly also around V854 Cen occurs in conjunction with the presence of PAHs. As we have mentioned above, high temperature condensation (even in H-rich environments) will lead to efficient formation of fullerenes (Jager et al. 2009) and this may be relevant for these RCB stars since the gas in which molecules and dust form will be much hotter than in red giant star environments. However, laboratory of high temperature condensates have shown that no PAHs are formed as intermediates (Jager et al. 2009). An alternative explanation for the simultaneous presence of PAH and C\({}_{60}\) molecules is that they may be formed by the decomposition of hydrogenated amorphous carbon (HAC) (Scott et al. 1997a). Laser vaporization of HAC films produces a wide range of large aromatic carbon molecules including PAHs and fullerenes such as C\({}_{60}\) (Scott et al. 1997a). Indeed, the C\({}_{60}\) molecules do not dominate the mass distribution of molecules seen in laboratory experiments (Scott et al. 1997a), an observation qualitatively consistent with the infrared spectra of DY Cen and V854 Cen. The UV radiation field around these two RCB stars is unlikely intense enough to cause HAC destruction. However, high velocity strong winds are typical in RCB stars and the collisional environment (i.e., grain-grain collisions) of these stars may lead to HAC vaporization.
Interestingly, the Infrared Space Observatory's 1996 spectrum of V854 Cen showed a correspondence with the laboratory emission spectrum of HAC at 773 K (see Lambert et al. 2001). These HAC features are weaker, even absent, from our Spitzer spectrum and the C\({}_{60}\) and PAH features present in V854 Cen's Spitzer spectrum are weaker or absent in the ISO spectrum (see Fig. 4). Although the absolute flux level is different for both spectra, the dust continuum emission seems to be unchanged and it can be well fitted by a blackbody at a temperature of \(\sim\)1000 K. displays the ISO and Spitzer spectra of V854 Cen after the subtraction of the dust continuum emission at \(\sim\)1000 K. This contrast between ISO and Spitzer spectra necessarily prompts the speculation that the principal ingredient in the circumstellar envelope evolved from HAC grains to molecules such as the PAHs and C\({}_{60}\)1. If so, formation of molecules such as PAHs and fullerenes in the circumstellar envelopes of the more H-rich RCB stars is a time-dependent phenomenon. A certain concentration of hydrogen is presumably needed to form HAC grains, which may be then destroyed by shocks in the circumstellar envelope. One product of destruction of HAC grains may be C\({}_{60}\) molecules which being hardy may survive for longer periods of time than the HAC grains and PAHs. In RCBs with a recurring series of dust-forming events with replenishment of HACs and additional formation of C\({}_{60}\) is a possibility. If these speculations have merit, one may expect to find C\({}_{60}\) molecules un-accompanied by HACs and PAHs in environments where grain formation is a non-recurring event.
Footnote 1: Note that V854 Cen underwent minima between the time of the ISO and Spitzer spectra and thus the HACs seen with ISO were not - in all probability - the HACs that led to the fullerenes and PAHs in the Spitzer spectrum.
Contrary to a conclusion drawn by Cami et al., our speculations may account well for Cami et al.'s pioneering detection of C\({}_{60}\) and C\({}_{70}\) molecules from the inner region of the PN Tc 1. Noting that the Spitzer spectrum of Tc 1 shows no PAHs, Cami et al. drew the conclusion that the fullerenes were formed in very H-poor gas ejected a few thousand years by an AGB star following an even earlier ejection of the star's H-rich envelope. This conclusion overlooks two key observations. First, the nebula is not H-poor (Koppen et al. 1991; Milanova & Kholtygin 2009) and, in particular, optical and ultraviolet spectroscopy of the inner regions confirm that the gas has a normal mix of H and He (Williams et al. 2008; R. Williams - private communication); the gas is not H-poor. Second, the central star CoD -46\({}^{\circ}\) 11816 is not H-poor and He-rich (Mendez 1991). In short, the fullerenes were in all probability formed in H-rich and presumably C-rich gas. Dust formation occurred in the circumstellar wind but now presumably the wind has lessened or ceased. The hot central star may have already destroyed the less hardy grains and molecules (e.g., PAHs) from times when a cool AGB star fed the then stronger wind. In this picture, hydrogen is essential to form fullerenes but the absence of PAHs is not proof that fullerene production occurs in a H-deficient region.
## 5 Concluding remarks
In summary, contrary to general expectation, the formation of large fullerenes, specifically C molecules around RCB stars takes place efficiently only in the presence of some hydrogen and HACs may be the precursors of fullerenes. However, the absence of fullerene features in highly H-deficient RCB stars is puzzling. Carbon chemistry in H-deficient environments should include the formation of fullerenes, as it has been shown by the early laboratory experiments (e.g., Kroto et al 1985; Kratschmer et al. 1990; de Vries et al. 1993). More laboratory experiments at different temperatures and hydrogen compositions are encouraged in order to learn about the formation of fullerenes. In particular, laboratory experiments in H-poor atmospheres could explore higher temperature formation routes of fullerenes. In addition, one would expect that grain-grain collisions of pure carbon grains would lead to fullerenes, and in this sense, laser vaporization experiments of amorphous carbon films could help to solve this puzzle.
We thank the anonymous referee for useful comments that help to improve this manuscript. We would like to thank Jack Baldwin and Rob Williams for their quick clarification about Tc 1. This work is based on observations made with the Spitzer Space Telescope, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under NASA contract 1407. D.A.G.H. acknowledges support by the Spanish Ministry of Science and Innovation (MICINN) under a JdC grant and under grant AYA-2007-64748. D.L.L. acknowledges support for this work provided by NASA through an award for program GO 50212 issued by JPL/Caltech. D.L.L. also wishes to thank the Robert A. Welch Foundation of Houston, Texas for support through grant F-634. _Facilities:_Spitzer:IRS.
# Table 1 Mid-infrared features in DY Cen's spectrum.
\begin{tabular}{l l l l l l} \hline \hline \multicolumn{1}{c}{ Feature} & \multicolumn{1}{c}{DY Cen} & \multicolumn{1}{c}{V854 Cen} & \multicolumn{1}{c}{Identification} & \multicolumn{1}{c}{Mode} & \multicolumn{1}{c}{Ref.a } \\ \hline
6.3 \(\mu\)m & yes & yes & PAHs & C-C stretching & 1 \\
7.0 \(\mu\)m & yes & \(\cdots\) & C\({}_{60}\) & F\({}_{1u}\) & 2 \\
7.7 \(\mu\)m & yes & yes & PAHs & C-C stretching & 1 \\
8.6 \(\mu\)m & yes & \(\cdots\) & C\({}_{60}\), PAHs & F\({}_{1u}\), C-H bending in-plane & 2,1 \\
11.3 \(\mu\)m & yes & yes & PAHs & C-H bending out-of-plane & 1 \\
11.9 \(\mu\)m & yes & yes & PAHs & C-H bending duo & 1 \\
12.7 \(\mu\)m & yes & yes & PAHs & C-H bending trio & 1 \\
15.8 \(\mu\)m & no & yes & large PAHs? & C-C-C & 3 \\
16.4 \(\mu\)m & no & yes & large PAHs? & C-C-C & 3 \\
17.0 \(\mu\)m & no & yes & large PAHs? & C-C-C & 3 \\
17.4 \(\mu\)m & yes & yesb & C\({}_{60}\) & F\({}_{1u}\) & 2 \\
18.8 \(\mu\)m & yes & yes & C\({}_{60}\) & F\({}_{1u}\) & 2 \\ \hline \end{tabular}
-
References. -- Allamandola et al.; Frum et al.; Boersma et al. 2010.
Observed Spitzer/IRS spectrum of DY Cen (in black) together with a polynomial fit (in green) to continuum points free from any gas and dust feature. The corresponding residual spectrum (in blue) is shown in the bottom panel.
Spitzer/IRS residual spectra in the wavelength range \(\sim\)5\(-\)20 \(\mu\)m of the RCB stars DY Cen (in red) and V854 Cen (in blue). The average residual spectrum of nine extremely H-deficient RCBs with little reddening (in black) is also shown. The expected temperature-dependent positions of the neutral C\({}_{60}\) features are marked with black dashed vertical lines.
V854 Cen residual spectrum compared with the average residual RCB spectrum.
Residual ISO 1996 September 9 (in blue) and Spitzer/IRS 2007 September 7 (in red) spectra in the wavelength range \(\sim\)2-25 \(\mu\)m for the RCB star V854 Cen. A blackbody of \(\sim\)1000 K was subtracted from both spectra. The ISO spectrum (R\(\sim\)1000) has been smoothed with a 13 box car in order to be compared with the Spitzer spectrum. The laboratory emission spectrum of HAC at 773 K (in green; Scott et al. 1997b) is shown for comparison. The main laboratory HAC emission features (Scott et al. 1997b) are marked with black dashed vertical lines.
|
10.48550/arXiv.1101.2253
|
Are C60 molecules detectable in circumstellar shells of R Coronae Borealis stars?
|
D. A. Garcia-Hernandez, N. Kameswara Rao, D. L. Lambert
| 408
|
10.48550_arXiv.2205.05102
|
###### Abstract
Solar-to-fuel direct conversion devices are a key component to realize a full transition to a renewable-energy based chemistry and energy, but their limits and possibilities are still under large debate. In this perspective article, we focus on the current density as a fundamental figure of merit to analyse these aspects and to compare different device configurations and types of solar fuels produced from small molecules such as H\({}_{2}\)O, CO, and N. Devices with physical separation of the anodic and cathodic zones, photoelectrochemical-type (PIC) or with a photovoltaic element integrated in an electrochemical cell (PV/EC), are analyzed. The physico-chemical mechanisms involved in device operation that affect the current density and relations with device architecture are first discussed. Aspects relevant to device design in relation to practical use are also commented. Then discussion is moved towards the relevance of these aspects to compare the behaviour in the state-of-the-art of the conversion of these small molecules, with focus on solar fuels from CO and N. conversion, analysing the gaps and perspectives. The still significant lack of crucial data, notwithstanding the extensive literature on the topic, has to be remarked, particularly in terms of the need to operate these cells in conjunction with our concentration (in the 50-100 SUN range) which emerges as the necessary direction from this analysis, with consequent aspects in terms of cell and materials design to operate in these conditions. The work provides a guide for the optimisation of the investigated technology and the fixing of their practical limits for large-scale applications.
## Introduction
The transition towards renewable energy sources offers the possibility of solving the environmental burdens due to increasing greenhouse gas emissions, but it also represents a clear economic and innovation opportunity.1 In this context, the use of solar fuels (SFs), _i.e._, chemicals produced through reactions triggered directly or indirectly by sunlight, is pivotal for the widespread use of renewable energy (above ~50-60%)2 because they play an essential role in (i) mitigating renewable energy fluctuations, (ii) storing energy on seasonal/yearly bases and (iii) transporting energy to long distances, allowing the implementation of a world-scale renewable-energy economy. They will serve for a smoother and progressive substitution of fossil fuels as primary energy source,3,4 since SFs use in current energy infrastructure can take place with minimal adaptation.
The main advantage is that the combustion product is N2 which can be released to the atmosphere. The associated gravimetric and volumetric energies for liquid NH3 are 18.8 M+kg-1 and 11.5 M+L+, respectively, which are comparable to those of CH4OH. With respect to formic acid, HCOOH another H2 vector obtained from CO2 reduction,22 NH3 presents several advantages: the hydrogen content by weight is higher in NH3 (-18%) with respect to HCOOH (4%); NH3 decomposes producing N2 (NH3 - 1/2 N2 + 3/2 H2), while HCOOH decomposition releases CO2 (HCOOH - CO2 + H2); (iii) the production and transport of liquid NH3 are already implemented large-scale processes, so such infrastructures can be readily used for distributing NH3. Thus, NH3 is an attractive carbon-free energy or hydrogen vector, which can be implemented as SFs by using sunlight to trigger N2 reduction reaction (NRR).23
Solar-to-fuel conversion is thus a complementary crucial technology to green H2 produced by electrolysis to move toward a carbon-neutral future. In this review only integrated systems which combine in a single device the functions of harvesting sunlight and its use to produce SFs will be discussed.24-31 From the first report by Fujishima and Honda in 1972,24 several architectures have been proposed for the efficient synthesis of SFs.24,23 We can turn the proposed architectures in three main classes: particulate photocatalysis (PP), (ii) photoelectrochemical (PEC) devices and iii) photovoltaic-driven-electrocatalysis (PV/EC). In PP, nanoparticles of semiconductors are suspended in the electrolyte and conduct both redox reactions on their surfaces. Note that the term PP is not commonly used, but we believe it is useful to differentiate the photocatalysis approach from those such as PEC or PV/EC where there is physical separation between the anodic and cathodic processes. In PFC systems reduction and oxidation take place at two electrodes (a photoelectrode and a counter electrode or two photoelectrodes). In PV/EC devices a solar cell is connected to an electrolyser which produces the SF. Focus in this review is on these devices having formation of anodic and cathodic products in separate streams, because those actually more promising to produce solar fuels, as commented later.
These devices are complex with many aspects determining their behaviour. It is thus necessary to use figures of merit (FOM) for their direct and proper comparison.24,25 Systems efficiencies are usually employed as FoMs, which are defined as the ratio of the total output power (in the form of electricity and/or energy stored in the SFs) to the total input power (_i.e._, electricity and/or solar energy).26 For the case of H2 production, the solar-to-hydrogen conversion efficiency (_e1sm_) is often employed as FoM, which is defined as:
\[\eta_{STM} = \frac{\left\| {J_{EC} \in 1 + 2\,\text{V}\,\eta_{sp}} \right\|}{P_{in}}_{AM,1.5G}\]
These FoMs are defined for the whole device, so they do not provide information about the occurring physico-chemical processes and how to improve the resulting performances. Several other FoMs have been introduced to characterise half-cell performances, such as the ideal regenerative cell efficiency, the ratiometric power-saved and the applied bias photon-to-current component metric.26
Usually, calculating these FoMs is not trivial, since corrections are needed to account for many phenomena (such as mass-transport, resistance due to the device geometry, etc.) and the choice of the counter electrode/device assembly may lead to significant variations in the computed results.26 This can make comparisons of results from different research groups challenging, especially if all the experimental conditions and calculations are not explicitly reported. A fair comparison with already existing SF generation systems can become cumbersome, as different working principles can be involved.
In this perspective, we propose current density (_I_) as a FoM to analyse performances and compare SFs approaches and devices. This choice is due to several reasons: (i) a detailed analysis of \(J\) can provide many information about the physico-chemical processes occurring during the device operation; (ii) it is readily obtained from simple electrical or electrochemical characterizations (thus, it is easily accessible) and (iii) it represents a useful FoM for comparisons with already existing technologies (for example, commercial electrolyser work with \(J\) ~ 0.5-2 A cm-2).24,25 Furthermore, from eq. 1 it is clear that by fixing the operating voltage and maximising _e__P_r (ideally 100%), _e__P_r is a function of \(J\) only. As will emerge later from the discussion, focusing attention on \(J\) as FoM allows to bring out the need to operate devices in conjunction with sun concentration, an aspect crucial for practical implementation, but essentially not investigated and that has also significant implications in terms of materials and cells needed to operate under this mode. Moreover, \(J\) is also an effective FoM from an industrial perspective, because it provides a direct indication on the productivity and other cell design characteristics.
To properly discuss the use of \(J\) and relevance in comparing devices for the different indicated SFs, this paper is organised as follows: the first part deals with the basic principles behind SF generation and the impact of device architecture. For convenience we divide the processes involved during SF generation in four steps, similarly to other literature studies.24,25 All these aspects influence the value of \(J\), so in the following parts we discuss the physics and chemistry behind the optimization of such processes and emphasise the information that can be obtained by a careful analysis of \(J\). The last part is dedicated to analysing the different design configurations of PEC cells, their use in relation to the different SFs which can be produced and their potential as high-_J_ generating systems.
Note that the aim is not to make a systematic review and discussion of the state-of-the-art, but to offer concepts and clues to rethink the topic from a different viewpoint. Only selected references have been considered.
### Solar to fuel conversion
#### Working principles
The synthesis of SFs involves a complex machinery of events which, in a simplified approach, can be divided in:
a) Photon absorption and generation of free charge carriers
b) Separation and transport of charge carriers
c) Catalytic conversion
d) Mass transfer
In PFC systems, all these steps are performed within the PEC cell, whereas in PV/EC design charge carriers are photo-produced and separated in the PV component, while catalytic conversion and mass transfer take place in the EC unit.
Herein, we consider a PFC system comprising a photoanode, an electrolyte and a metallic counter electrode. When these three components are connected together, a charge flow is triggered that brings the whole system to equilibrium (_i.e._, to the condition where the chemical potential is constant throughout the device).4,4,4 For the case of a _n_-type semiconducting photoanode, the equilibrium at dark conditions is reached through an electron flow from the semiconductor to the solution.4,4,4 This causes the formation of a depletion layer inside the photoanode and an Helmoltz layer in the electrolyte.
The region depleted from electrons is characterised by the presence of fixed positive ions, leading to the formation of a built-in electric field which induces the "up" bending of the electronic bands. The same phenomenon occurs for _p_-type semiconducting photocathodes, with the difference that a "down" band bending occurs. In both _n_- and _p_-type cases, the band bending induces an electric field which opposes to the motion of electrons (for _n_-type) or holes (for _n_-type) from the photoelectrodes to the electrolyte once equilibrium of the chemical potential is reached (_i.e._, when the Fermi level of the semiconductor equals the chemical potential of the electrolyte redox couple).
Upon illumination, the light-harvester of the photoelectrode absorbs photons whose energy equals or exceeds that of its band gap (_E_b). When this happens, electron-hole pairs are generated which (depending on the properties of the semiconductor, as described in the following sections) can form a stable bound state (exciton) or dissociate into free charge carriers. Such non-equilibrium density of electrons and holes is described through the concept of quasi-Fermi level (one per type of charge carrier, Figure 1c).4,4,45 The difference between the values of these quasi-Fermi levels represents the photo-produced voltage, _i.e._, the photovoltage or open circuit voltage \(V_{\text{OC}}\).4,45 It is worth emphasising that photo-generated carriers are responsible to trigger the reactions for SFs production. However, such processes can be activated only if several constraints are met. In particular, splitting reactions are endergenic (_i.e._, they require energy to take place). For the case of H2 from H2O, CH2OH and CH4 from CO2 and NH3 from N2 these equations formally describe the splitting processes:
\[\text{H}_{2} 0 \rightarrow \text{H}_{2} + \frac{1}{2}\text{O}_{2}\]
\[\text{CO}_{2} + 2\text{H}_{2} 0 \rightarrow \text{C}_{\text{H}}\text{OH} + \frac{3}{2}\text{O}_{2}\]
\[\text{C}_{\text{O}_{2}} + 2\text{H}_{2} 0 \rightarrow \text{C}_{\text{H}}\text{4} + 2\text{O}_{2}\]
\[\text{N}_{2} + 3\text{H}_{2} 0 \rightarrow 2\text{NH}_{3} + \frac{3}{2}\text{O}_{2}\]
The overall water splitting (OWS) reported in (eq. 2) has a Gibbs free energy \(\Delta G^{+}\) = 237 kJ mol+1, the NH3 synthesis (eq. 4) needs \(\Delta G^{+}\) = 680 kJ mol+1, while the CO2 reduction into CH2OH (eq. 3a) or CH4 (eq. 3b) require \(\Delta G^{+}\) = 689 kJ mol+1 and \(\Delta G^{+}\) = 800 kJ mol+1 respectively (reported values correspond to 25 "C and 1 bar conditions).2,4,46,47 The \(\Delta G^{+}\) values can be readily converted into potential difference through the equation
\[\Delta\text{E}^{0} = \frac{\Delta G^{0}}{\text{nF}}\]
Consequently, the minimum photon energy needed to trigger these solar-to-fuel reactions are 1.23, 1.17, 1.19 and 1.03 eV.2,46However, these values consider only the thermodynamic limits from the reaction itself. Indeed, the minimum energy required for the practical production of SFs is actually higher, mainly because of ohmic and charge transport losses, kinetic overpotentials and reactor geometry determining cell resistances(r) (for example in the case of OWS values span between 1.8 and 2.4 eV).10,10 in general, increasing the potential with respect to RHE (reversible hydrogen electrode), favours higher formation rates and thus the electrode productivity. However, in NRR and Co,RR, higher potentials also favour the side reaction of H+/e' recombination to form H. Thus, to maintain high the selectivity to the target product, the overall overpotential is kept low (from 0.1-0.2 V vs RHE,31 to 0.4-0.5 V) which implies low reaction rates and electrode productivities.
Thus, semiconductors with such band gaps must efficiently produce enough photo-voltage to conduct the desired reactions. The band edge absorption lies at wavelengths higher than 600 nm, which reduces drastically the utilization of sunlight, consisting of IR (2500 - 700 nm), visible (700 - 400 nm) and UV (400 - 300 nm) regions which represent 52%, 43% and 5% of solar radiation.4)
Once electrons and holes are photo-generated, they must be kept separated and driven towards the interface formed with the electrolyte (in the PV/EC design charges are driven to the PV current collector and then to the EC component). When charges reach the active sites of the surfaces, electrons and holes are used for the redox reactions of cathodic conversion of protons and electrons to H2 (OWS) or reduction of N2 or CO2 (NRR and CO2RR, respectively) while generation of O2 (OWS) or other oxidation reaction at the anode side. For these processes to take place favourably, charge carriers inside the photoelectrode must lie into proper energy levels, _i.e._, the conduction band minimum (CBM, for electrons) and valence band maximum (VBM, for holes).2. Thus, electrons can reduce a species if the CBM lies above the reduction potential, while holes can drive oxidation of a reactant if the VBM lies below the oxidation Finally, the circuit is electrically closed by the motion of the ions within the electrolyte. This step is crucial to obtain high reaction rates because it is responsible to bring "new" active species from the electrolyte to the particle surface where they can be used for the redox processes. Thus, a fast mass transfer is fundamental to enhance the kinetics of the reactions and keep high \(J\) values.
This sequence of processes occurs in both PP and PFC cases, but in the latter the anodic and cathodic processes occurs in physically separated zones, allowing thus a separation of the anodic and cathodic products of reaction with the many inherent advantages (reduction of the costs of separation and safety, etc.). A measurable photocurrent is present between the anodic and cathodic compartments, and thus current density (_J_) is a direct measure of the process. In PP, instead, no measurable \(J\) is present and for this reason no longer discussed here. The PV/EC architecture shares with PFC the compartmentalisation in anodic and cathodic zones, but the process of photogeneration of the current occurs in a separate PV element and is not integrated in one of the cell zones, typically the anodic one. Also for PV/EC the photocurrent can be measured and is an indication of the process effectiveness.
This brief discussion shows that SFs production can be highly efficient only if these four steps are wisely optimised (a thorough discussion of the implications of each step is presented in the following sections). There are, however, other aspects that must be considered. A careful engineering of devices for SFs generation is needed. Many architectures are based on heterojunction systems, _i.e._, different materials are used synergistically with the aim to optimise each process. For example, charge transporting layers are used to selectively drive electrons or holes to one specific direction, interlayers are added to improve charge transport and passivate defects formed at the interfaces between two materials, and electrocatalysts are deposited onto the interface in contact with electrolyte to increase the catalytic activity of the reactions. Thus, the architecture of devices for SFs production shows several additional possible resistance and interfaces limiting the overall \(J\), as discussed later. The understanding of these effects is crucial to obtain high performances, but often not specifically investigated.
In addition, it must be remembered that high \(J\) implies fast charge carrier transport for each step of device machinery. However, there is a significant difference in the timescales (r) involved during light harvesting, charge separation and transport, redox reactions, and electrolyte transport. In fact, charge generation and separation are fast processes (r ~ fs) while charge transport, redox reactions and mass transfer take place at slower timescales (r ~ ms) and ms/s, respectively) as shown in Figure 2.4 Thus, when carriers reach the active sites of the reactions, they may "wait" for a long time before being used to trigger reduction or oxidation processes because of slow catalytic activity or stagglision motion of the ions inside the electrolyte. The use of an electrocatalyst can accelerate the redox reaction.
Indicative timescale (t) of the photophysical and (photo)electrochemical processes in water splitting. GK: conduction band; WS: valence band; trans: transfer. Elaborated from ref.4 Copyright Springer ©2015.
facets and defects, surface functionalization), rather than effectively catalysing the process of charge transfer.32-35
### Technologies and approaches
In the previous sections, the three main approaches used to produce SFs were introduced: PP, PEC and PV/EC. Although the PP case, as commented, is not discussed in detail here because a \(J\) cannot be measured and used as FoM, we recall it here briefly as a comparative element to understand better the differences between PEC or PV/EC and PP case, the latter being one of the most investigated, because simpler (even with many limitations). Herein, we thus discuss the pros and cons of these different device configurations for solar-to-fuel conversion.
In PP systems, nano- or micro-particles (made up by semiconductor materials) are suspended inside a water solution and use light to directly drive both the oxidation and reduction reactions on their surfaces.32,36,57 Typically, the two redox reactions occur on different crystal facets of the particle modified by adding different co-catalyst nanoparticles. Note that these added nanoparticles (used to promote the redox electron transfer reactions) are typically indicated in literature with the term co-catalyst (rather than only catalyst), because acts as co-elements to enhance the electron transfer capability of the semiconductor substrate itself. For OWS, PP is the simpler approach to produce H2 requiring only semiconductor particles (suspended in an aqueous electrolyte) and light to operate, but many issues limit its large-scale use. First, it is difficult to find a material that can meet the energetics requirements for both hydrogen and oxygen evolution reactions (HER and OER, respectvely) while being an optimal light harvester and stable against corrosion.
Using co-catalysts can mitigate some of these constraints and several approaches have been proposed (such as defect engineering, Z-scheme, elevated temperature photocatalysis, etc.)39, with the aim to increase the performances of such devices but so far, the conversion efficiencies are low (for the case of OWS, _e__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p____p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p____p__p__p__p__p__p__p__p__p__p__p____p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p____p__p__p__p__p__p__p__p__p__p__p__p__p____p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p____p__p__p__p__p__p____p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p____p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__p__pphotoelectrode/electrolyte interface to promote the redox reactions. Protective layers can also be needed to prevent the direct contact between the photoelectrode and the electrolyte (this configuration is usually referred to buried junction) since the former can suffer from corrosion.
2. _A cell system_: it comprises the hosting structure with separate anodic and cathodic zones, a membrane separating these two compartments and wires to transport the electrons from (typically) the anodic to the cathodic zone. Membranes and wires are necessary to close the ionic and electronic circuit, while maintaining the physical separation between the anodic and cathodic zones.
3. _A counter electrode_: in OWs, this could be a simple metallic electrode where HER occurs, but in Co,RR and NRR cases, the electrode has a more sophisticated multicomponent structure, being necessary to maximize the aimed reaction, while minimizing the side HER reaction.
It is worth noting that since reduction and oxidation reactions take place in different compartments, issues related to separation and recombination of the products, side reactions and safety are avoided, overcoming the intrinsic limitations of the PP case. However, the higher complexity of the resulting devices results in additional steps and sources of resistances, briefly outlines as follow:
1. _At the photoelectrode_ a) Photon absorption and generation of free charge carriers b) Separation and transport of charge carriers c) Catalytic conversion (into the reduced or the oxidised species, depending on the reaction occurring at the photoelectrode) d) Evolution of the products in the form of gaseous bubbles (avoiding the stable deposition on the surface of photoelectrode) and inhibition of side and recombination reactions e) Collection and transport, through an external wire, of the other charge carrier at the cathode
2. _At the membrane_ a) Mass transfer through the membrane b) Inhibition of cross-over from the counter electrode zone.
3. _At the counter electrode_ a) Mass transfer of active species from the membrane to the catalyst b) Transport of the charge carriers (generated by the photoelectrode and incoming through the wire) to the catalyst sites c) Evolution of the products in the form of gaseous bubbles (avoiding the stable deposition on the surface of photoelectrode) and inhibition of side and recombination reactions.
This intrinsically complex machinery, where all the steps are highly correlated and not independent (as often assumed), require special care and optimal engineering of the whole device to obtain high efficiencies. As a result, the structure of PFC photodetectors is complex, resulting in higher fabrication costs. For the case of OWS, currently PFC systems show 2 % < n_{\text{INT}} < ~ 19 %,9 depending on the photoelectrode components.
The third device architecture is that of PV/EC, where a solar cell is connected in series to an electrolyser: the redox reactions are carried out by the latter, using the charge carriers photo-generated in the PV component. It's worth mentioning that, with this kind of architecture, the solar cell is not in direct contact with the electrolyte, preventing detrimental effects due to corrosion. Furthermore, since both photovoltaic and electrolyser technologies have reached high efficiencies, the PV/EC approach is the most promising solution for sustainable 5S production,9 with the current record for OWS of 19.8% = 30%,00 although performances are not stable.
Another promising design, which exploits the possibility for a two-step light absorption somewhat similar to Z-scheme in PP,61,62 is the tandem PEC configuration,63,64 in this case, two photoelectrodes are used synergistically to increase both photocurrent and photovoltage, i.e., by maximising the absorped portion of the solar spectrum and optimising the electronic band gaps/band edges positions.
The main configuration of PFC tandem cells is the photocatable/photoanode (PC/PA), consisting of two photoelectrodes connected with an external wire. To optimise the light harvesting, one of the photoelectrodes and the separation membrane must be transparent to the light absorbed by the second photoelectrode, a requirement that limits the choice of materials used in such architecture. Thus, although tandem configurations have the potential to deliver high efficiencies, the expensive costs of fabrication often hinder their practical use.
In addition to configurations where the two electrodes are immersed in the electrode far from the membrane separating the two hemicels (Figures 3b-d), a compact configuration where the two electrodes are positioned on the two sides of the membrane can be realised. This "monolithic" configuration minimizes resistance mass transfer and ohmic losses, as described in the following sections.
As a final consideration, note that the efficiency of the systems can be increased by the addition of low-cost Fresnel lenses acting as a sun-concentrator. Depending on the design, sun concentration factors up to three orders of magnitude can be reached,68 but even a quite simple design can allow to increase light intensity by a factor up to 30-50 at a minimal additional cost. However, this assembly can easily induce temperature rise even above 100degC. Since most of the current cell designs and materials cannot run in a temperature range of around 100-200degC, further research is needed to address the possible applications of (low cost) sun-concentration systems to enhance the PFC performances.
## Reactions
Herein, we focus our attention to the conversion of H2O, N2 and CO2 into SFs. Such reactions present different features, but all are multi-electron reactions. If only the cathodic reaction is considered, HER is a 2-electrons reaction (eq. 6), but a 4-electron reaction by considering full OWS (eq. 9), the formation of OJS. NRR is a 6-electron reaction (eq. 7), while CO2+RR depends on the target product. For example, CO and HCOOH formation (eq. 3a and 8b) are 2-electrons reactions, methanol production is a 6-electrons reaction (eq. 9), while CH\({}_{4}\) formation is a 8-electrons reaction (eq. 10). More complex reactions have been proven to be feasible involving the formation of C-C bonds, such as the 8\(e\) reduction to acetic acid (CH\({}_{4}\)COOH, eq. 11) of the 18\(e\) reduction to isopropanol (CH\({}_{4}\)CH(OH)CH\({}_{4}\), eq. 12).
\[2\text{H}^{+}+2\text{e}^{-}\rightarrow\text{H}_{2}\]
\[\text{N}_{2}+6\text{H}^{+}+6\text{e}^{-}\rightarrow2\text{NH}_{3}\]
\[\text{CO}_{2}+2\text{H}^{+}+2\text{e}^{-}\rightarrow\text{CO}+\text{H}_{2}\text{O}\]
\[\text{CO}_{2}+2\text{H}^{+}+2\text{e}^{-}\rightarrow\text{HCOOH}\]
\[\text{CO}_{2}+6\text{H}^{+}+6\text{e}^{-}\rightarrow\text{CH}_{3}\text{OH}+ \text{H}_{2}\text{O}\]
\[\text{CO}_{2}+6\text{H}^{+}+8\text{e}^{-}\rightarrow\text{CH}_{4}+2\text{H}_{2}\]
\[\text{2CO}_{2}+8\text{H}^{+}+8\text{e}^{-}\rightarrow\text{CH}_{3}\text{COOH}+2\text{H}_{2}\]
\[\text{3CO}_{2}+18\text{H}^{+}+18\text{e}^{-}\rightarrow\text{CH}_{3}\text{CH}(\text{OH})\text{CH}_{3}+\frac{5}{2}\text{H}_{2}\text{O}\]
The formation of various other CO2RR products have been also documented, such as ethanol, ethylene and other hydrocarbons, oxalic acid, acetone and methyl formate. There is currently a large scientific interest on these multi-carbon CO2 reduction reactions as a valuable route to produce high-value chemicals in substitution of current routes from fossil fuels. The rich electrocatalytic chemistry of CO2 reduction opens new possibilities to build a new chemistry on it, although understanding how to control the reaction selectivity is still an issue. For conciseness, we will limit our discussion only to few type of CO-RR, which are relevant for S/S used as renewable energy storage system or as hydrogen vector.
The NRR is a reaction of recent large interest, with over 50 reviews on this topic published in the last few years, mainly focused on the electrocatalytic synthesis and related materials, a selection of which is given in ref.[78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 199, 199, 198, 199, 199, 199, 190, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 1999, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 1999, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 1999, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 1999, 199, 199, 199, 199, 1999, 199, 199, 199, 199, 199, 1999, 199, 199, 199, 199, 199, 1999, 199, 199, 199, 199, 1999, 199selectivity does not represent an issue. This is especially important in electrocatalysts because the reaction rate could be improved by increasing the applied overpotential, up to the maximum determined by limitations due to ohmic and concentration polarization, mass transport losses, activation losses. When a selectivity issue is present, the overpotential to be applied is instead determined from the compromise between reaction rate and selectivity. Productivity (and optimal _J_) is thus typically lower than in the OWS case.
### Factors limiting the current density
In the previous sections it was shown that solar-to-fuel conversion is a multi-step process, comprising many mechanisms which take place at different timescales. The following sections are dedicated to an analysis of the physical and chemical implications of each process and how they affect the output \(J\) of devices for SFs production.
#### Photon absorption and generation of free charge carriers
The absorption onset of semiconductors depends on their energy bandgap, _i.e._, the energy difference between VBM and CBM. The absorption coefficient (_a_) is a measure of how efficiently the incident radiation is absorbed. Consequently, good light harvesters must show panchromatic (to increase the part of absorbed sunlight) and enough thickness to capture all the absorbable light. A high \(a\) value has a twofold advantage: reduces the amount of required photo-material (thus, the costs of fabrication) and limits the distance that charge carriers must travel to reach the interfaces of the light harvester. A comparison of \(a\) for different photo-materials is shown in Figure 4a, revealing that the absorption onsets usually are not sharp and display sub-band gap absorption because of energetic disorder arising from the presence of defects, trap states and other aspects.10
However, high \(a\) values do not straightforwardly imply that a photo-material can produce high \(J\) values. The ability of a photo-material to generate electron-hole pairs is affected by other processes accounted by the incident-photon-to-current-collected-efficiency (_IPCE_), or external quantum efficiency, defined as:13,45
\[\text{IPCE} = \text{\Pi}_{\text{Hall}}\text{\overline{He}}_{\text{app}}\text{\overline{Li}}_{\text{cat}}\]
Here \(e_{t}\) represents the light harvesting efficiency (which is typically identified with the absorbance), _In__in_ is the efficiency of electron-hole pairs separation and \(n_{t}\) is the probability that electrons and holes reach the interfaces of the light harvester. The product \(e_{t}\) identifies the flux of collected charge carriers per absorbed photon which is termed the absorbed photon-to-current conversion efficiency (_APC_) or internal quantum efficiency.46,45 The photo-produced \(J\) arises from the direct use of solar photons, one-by-one, to generate electron-hole pairs.
The maximum attainable values of photo-produced \(J\) (which equals the short circuit current density \(J\),17) can be obtained by multiplying the elementary charge \(e\) by the number of photons which successfully produce collected charge carriers. In mathematical terms, \(J_{c}\) results from the integral, over all the emitted photon energies, of the solar spectral photon flux _ph_(_E_) times the _IPCE_:19
\[\text{I}_{\text{sc}} = \text{e}_{0}^{T_{\text{0}}}\text{\overline{I}}\text{PCE} \cdot \text{\varphi}(\text{E})\text{\overline{d}}\text{E}\]
In the ideal case where _IPCE_ = 1, eq. 14 can be used to calculate the maximum theoretical \(J_{c}\) by using the AML5 spectral features in _ph_(_E_). The resulting \(J_{c}\) is shown in Figure 4b, as a function of the photo-material \(E_{p}\) fit is worth noting that \(J_{c}\) increases with decreasing _E_g (maximum 73 mA-cm-2 for \(E_{p}\) = 0 eV), confirming that small band gaps are needed to achieve high \(J\). However, charge carriers with energy exceeding that of the band edges (_i.e._, hot carriers) lose their energy firstly by scattering events with other carriers (thermalization) and lastly by the emission of phonons (cooling).105 The relaxation of charge carriers to the band edges occurs with time scales ranging between _fs_ to _ns_ (depending on the occurring mechanisms).106 Actually, there is a large interest in hot carriers properties for many applications such as photodetection,107 lasing10 and photocatalysis.109 However, a thorough discussion of these aspects goes beyond the scope of this review. The thermodynamic restrictions due to SFs production pose a sever limit to the maximum reachable \(J_{c}\) (for example, in OWS, a minimum photo-voltage of about 1.6-1.8 eV is required to evolve H2 and O2 with a sufficient rate which correspond to ~25 mA-cm-2).
The efficient generation of charge carriers does not depend only on \(a\), but on \(e_{t}\), too. In fact, when a photon is absorbed, an electron and a hole are generated at the same location and can attract each other through their electrostatic interaction.10, it is then possible for them to form a neutral bound state (_i.e._, an exciton), thus reducing the number of free charge carriers contributing to \(J\). Excitons can have binding energies (_E_s) ranging between tens (Wannier-Mott excitons) to thousands (Frenkel excitons) meV.10 Since thermal energy at room temperature is ~25 meV, Wannier-Mott excitons can be readily dissociated, leading to the generation of free charges. Wannier-Mott excitons can be treated as hydrogenic systems with quantized energy levels given by:10
\[\text{E}_{\text{b}}(\text{n}) = - \frac{\text{m}_{\text{r}}\text{e}^{\text{n}}}{\text{n}^{2}\text{e}^{2}\text{n}^{2}}\]
The exciton energy levels strongly depend on how the electron-hole pair coulombic interaction is screened by the medium (_i.e._, \(E_{s}\)(_m_) is a function of \(e_{r}\), as shown in eq. 15). The \(e_{r}\), describes the dynamics of the medium and can span over a wide range of frequencies (from 104 to 105 Hz), depending on the resonances such as internal molecular vibrations, polar oscillations, internal diffusion of ions and other aspects. This is shown in where the real part of \(E_{s}\), is reported for the case of the perovskite CH3NH4Pb3.102
However, efficient dielectric screening may arise only if the dynamic responses have energy close to that of \(E_{b}\) which, for the case of Wannier-Mott excitons falls in the frequency interval ~ 2-20 THz. This frequency range pertains lattice vibrations arising from phonon modes, so the proper engineering of the vibrational dynamics (_i.e._, the careful choice of the lattice chemical composition) is a useful tool for tuning _E_b and consequently for boosting \(J\).
### Transport of charge carriers
After the absorption of radiation and the formation of free electron-hole pairs, charge carriers must travel through the bulk semiconductor to be collected at the interfaces. The quantity of charges that reaches the contacts gives \(J\), so high values can be obtained only with optimal transport properties of the semiconductor. The motion of electrons and holes is hindered by many aspects which can be conveniently grouped in:
1.
2.
### Diffusion/drift factors
The motion of charge carriers in the bulk of semiconductors can be due to a gradient of electron/hole concentrations (called diffusion and described by the diffusion coefficients _D_n or _D_n) and/or by an electric field (called drift and described by the mobilities _m_n and _m_n). Drift and diffusion are related to each other through the Einstein relation:111
\[\text{H} = \frac{\text{e}\text{D}}{\text{k}_{\text{B}}\text{T}}\]
16 holds for both electrons and holes by substituting the corresponding values of \(m\) and _D_). The factors which limit the transport of charge carriers can be easily identified by considering the definition of \(m\) on the basis of the Drude model.112 In this framework, the transport of electrons and holes is hindered by scattering events which randomize the charge carriers flow within the material. The average time between two scattering phenomena is called relaxation time r and the mobility can be expressed as:112
\[\text{H} = \frac{\text{er}}{\text{m}^{\ast}}\]
Each material will show different \(m\) since both _m_n' and \(r\) strongly depend on the nature of the investigated material. Indeed, _m_n' is inversely proportional to the curvature of the CB or VB (depending on the considered charge carrier), while \(r\) relates to the underlying scattering mechanisms. Thus, _m_n can be potentially tuned by optimising _m_n' (_i.e._, by varying the electronic energy bands with proper engineering of the chemical composition) and r. For the latter, scattering may occur because of other charge carriers, imperfections in the crystal (_i.e._, grain boundaries, impurities, energetic disorder, etc.) and interactions with phonons.112,123 Scattering from imperfections can be potentially avoided (or at least largely reduced) through proper material fabrication routes. On the contrary, phonon scattering is a fundamental process which can only be partially mitigated. In general, there are two main phonon scattering mechanisms which limit _m_: deformation potential and local polarization.112
The former affects all materials, since it results from deformations (_i.e._, variations) of the crystal potential because of lattice vibrations. This phenomenon can arise from both longitudinal and acoustic phonons. The local polarization mechanism is, instead, experienced by polar and partly ionic materials.112 Such scattering event is due to anti-phase oscillations of oppositely charged ions, which generate a dipole moment and consequently an electric field which affects charge carriers mobilities. In this case, acoustic and longitudinal optic (LO) phonons can contribute to this phenomenon (through the so called piezoelectric acoustic scattering and Frohlich interaction, respectively).114 The transverse optic (TO) phonon does not contribute to Frohlich interaction because it does not produce a macroscopic polarization of the lattice. Since phonon dispersion curves depend on the chemical nature of the crystal, tuning phonon properties is a fundamental strategy to increase \(m\) (and therefore _r_).
Note that these effects depend also on phonon and carrier confinement due to the dimensionality of the semiconductor.135 There is thus a dependence on the nanostructure of the materials, which determines the transport of carriers and photocatalytic behaviour. Note also that defects associated to strains created at the interface between grains, for example anatase and rutile phases of TiO2, have a significant effect on phonon lifetime.118
### Recombination factors
Photo-excited electron-hole pairs (with concentration _m_) move through the semiconductor and may recombine (through several mechanisms) reducing the current flow. inorganic semiconductors such as Si and GaAs show three main recombination pathways: monomolecular, bimolecular and Auger (schematically represented in Figure 4d).10,117
The total rate of recombination can be written as:111,117
\[\text{R}_{\text{T}} = \text{R}_{\text{m}} + \text{R}_{\text{b}} + \text{R}_{\text{A}} = \text{m} + \text{B}^{\text{n}} + \text{C}\text{n}^{3}\]
The time evolution of \(n\) can be written as:
\[\text{\frac{dn}{dt}} = \text{G} - \text{R}_{\text{T}} = \text{G} - \text{A} - \text{B}^{\text{n}} - \text{C}\text{n}^{3}\]
Extracting the values of \(A\), \(B\) and \(C\) (for example from time resolved spectroscopies)111 gives fundamental information about the recombination mechanisms occurring in the photo-material. This kind of information is fundamental to boost \(J\), as recombination processes result in the reduction of free charge carriers within devices for SrS synthesis.
From eq. 19 another parameter can be introduced, the average recombination lifetime _t_rec, _i.e._, the average time in which photo-generated electrons or holes can move within the semiconductor, before recombining:10,119
\[\text{t}_{\text{rec}} = \frac{\text{p}_{\text{T}}}{\text{n}} = A + \text{Bn} + \text{C}\text{n}^{2}\]
Finally, it can be shown that the average length travelled by charge carriers before recombining, which is the diffusion length _L_o is given by:106
\[\text{L}_{\text{D}} = \sqrt{\text{D}_{\text{T}\text{rec}}}\]
Thus, _L_o gives several information about the semiconductor features, as it depends on both the mobility (through _D_) and the recombination mechanisms. It's worth noting that knowledge of _Lo_ is fundamental for photo-materials since there is a trade-off in the choice of their thickness. High thicknesses increase the probability of photon absorption but thicknesses exceeding _Lo_ will cause materials with potentially high recombination rates (and consequently low _l_).
Table 1 presents the values of several parameters herein introduced, for many benchmark semiconductors used as light harvester in both photovoltaics (PV) and SFs production devices. It is evident that PV-grade materials (_i.e._, Si, III-V semiconductors, and chalcogenides) exhibit the highest values for both \(m\) and _Lo_. However, these materials usually suffer from degradation in the electrolyte environment, so they are poorly stable. Quite differently, oxide materials are far more resistant to electrolyte corrosion, but they show poorer charge carrier transport properties, leading to lower \(J\) values.10
The discussion presented so far describes the processes involved for the generation and efficient separation of charge-carriers inside the light harvester. It must be remembered that when electrons and holes reach the absorber interfaces, they must migrate through them and eventually through other layers (used as protective barriers or to facilitate charge transport) before reaching the active sites (for the reaction of interest) or the back contact. To ensure the production of high \(J\), it is essential to minimize both resistance to transport and possible charge recombination, both highly depending on the interface present. All components of a device must be carefully optimized from this perspective, although often this aspect is not accounted in simple device configurations. Thus, while the light harvester is essential to both generate and transport efficiently electrons and holes, all the other device components (such as transporting, passivating or buffer layers and contacts) must guarantee a low recombination probability of charge carriers, to prevent photo-current loss.
Nanostructuring the semiconductor is relevant from this perspective.139 For example, a 1D-type ordered nanostructure (as in TiO2 ordered arrays of nanotubes)140 allows to realize a vectorial transport with holes-electrons moving on the internal or external sides of the nanotubes, reducing thus the possibility of recombination.141 However, this positive effect is often masked from the presence of shallow trap states that instead favour recombination and reduces charge-carrier mobility.142 A proper mastering of all these factors is thus crucial to enhance the device performances.143
### Redox (catalytic) reactions
Once electron-hole pairs are formed, separated, and transported to the interfaces of the absorber, they must reach the back contact or the active sites (at the interface with the electrolyte) to trigger oxidation or reduction reactions. However, these reactions can take place only if the electronic band edges straddle the redox levels, _i.e._, the CBM must lay above the reduction potential while the VBM must lie below the oxidation potential.
This condition would pose a further limit to the choice of light harvesting materials, which now must realise these conditions: (i) maximise the part of absorbed light, (ii) produce enough photovoltage and (iii) show proper band edges energetics with respect to the desired reactions. Since finding a material that can meet all these requirements is particularly challenging, usually the processes of charge carrier generation and electrocatalysts are decoupled by using an optimised light harvester (which photo-produces carriers) and electrocatalysts (to made the considered chemical reactions).
The optimization of the electrocatalysts properties is pivotal to achieve high \(J\), as many processes may hinder charge carrier motion. First, electrons or holes must be driven efficiently from the light absorber/electrocatalyst interface to the electrocatalyst/electrolyte interface, so the transport must be optimal to avoid loss due to recombination effects at both absorber/electrocatalyst interface and within the electrocatalyst itself.
Moreover, when charge carriers reach the electrocatalyst/electrolyte interface they can finally trigger the chemical reactions, but if the carrier transfer (from the electrocatalyst to the reactant/electrolyte) is too slow, the chances of recombination events increase, leading again to a reduction of \(J\). Such charge transfer can be quantitatively assessed through the exchange current density, which is another fundamental parameter that must be considered when systems for SFs production are designed. The exchange current density can be also related to the chemisorption energy of the products on the octahedral surface, as observed by Trasati for the case of metallic cocatalysts.144 The resulting trend resembles a triangle, thus the name given to such shape is "volcano plot". Note that this trend is based on several assumptions (for example that the exchange current depends on the strength of the M-H bond), and thus that various experimental results deviate from this trend.145 In oxide electrodes, where large deviations from the outer sphere electron transfer mechanism are present (larger than in metallic electrodes), deviations from the predicted volcano-type trend are large.145 Moreover, in complex reactions, such as CO,RR and NMR, volcano-type approaches fail in giving correct predictions.
\begin{table}
\begin{tabular}{c c c c c c} Material & E\({}_{\text{a}}\) & _N_a & _N_a & _Lo_ & _Ref_ \\ eV & cm\({}^{3}\)V\({}^{4}\) s\({}^{-1}\) & cm\({}^{3}\)V\({}^{4}\) s\({}^{-1}\) & cm & _N_a \\ \hline _c_-Si & 1.12 & 1450 & 500 & 100 & 31,032 \\ GaAs & 1.42 & 8500 & 400 & 100-900 & 31,033 \\ CuGe/SiGe, & 1.3-1 & 100 & 2.5 & 0.3-0.52 & 31,034 \\ SiO,Si,Si & 1.1-1.2 & -10 & -10 & 0.28 & 10,035 \\ TiO2 & 3.02 & 0.01 & 100 & 0.01-0.1 & 100,037 \\ WO1 & 2.8 & 10 & 20-104 & 0.15-0.5 & 100,038 \\ Fe2O & 1.9-2.2 & 0.5 & 10–104 & 2–140 & 100,038 \\ Cu2O & 2.14 & 6 & 256 & 0.025 & 100,031,110 \\ BVO3 & 2.4 & 0.044/12 & 2 & 0.07 & 100,031,110 \\ TaAs & 2.01-2.12 & 1.3-1.4 & 0.08 & 18 & 100,031,110 \\ TaON & 2.08/2.4 & 17-21 & 0.01 & 8 & 100,031,110 \\ ChNH(Pb) & 1.6 & 1.4 & 0.9 & 0.3-1 & 0.3-1 \\ \end{tabular}
\end{table}
Table 1: Comparison of the energy band gap (_Lo_), electron (_Lo_) and hole (_no_) modules and diffusion length (_Lo_) of semiconductors used as light harvesters in PV and SFs devices.
would be necessary,146-148
In practical terms, a methodology largely used for the investigation of these aspects is the Tafel plot,149 which is an experimental measure of the current density (usually in a logarithmic form) and the applied overpotential. The onset of the curve defines the exchange current density. The Tafel equation approximates the Butler-Volmer equation, when the concentrations at the electrode are equal to the concentrations in the bulk electrolyte, allowing the current to be expressed as a function of potential only.
Finally, electrocatalysis should exhibit long-time operational stability, _i.e._, they should maintain their performances unchanged for long-time use at high \(J\). In fact, instability arises from many processes such as oxidative decomposition, metal leaching and peeling of the electrocatalysis from the photoelectrode. Currently, electrocatalysis exhibit a stability of tens of hours (while very few survive for more than hundreds of hours) at \(J\) -10 mA*cm2, while for practical applications current densities of at least one order of magnitude higher are requested (commercial electrostypers work at \(J\) -0.52 A*cm2),39,40 A very encouraging and interesting result about long-living electrocatalysis recently came from layered double hydroxides, which show an excellent and stable catalytic activity for over 6000 hours at \(J\) = 1 A cm2,136 which represents (to the best of our knowledge) the current record for highly stable and efficient electrocatalysis.
### Mass transfer
Other aspects that need to be considered in electrochemical processes are the limitations related to mass transfer, related to the ion mobility in the electrolyte (which consists of ionic active species dissolved in a solvent). For OWS, the active species coincide with the solvent, _i.e._, water. However, the resistivity of pure water is high (distilled water ~105 O, purified water ~107 O),37 so supporting ions are usually introduced to avoid Ohmic losses throughout the electrolyte environment. The potential drop due to the electrolyte resistivity is given by:
\[V_{loss} = i \cdot R_{E}\]
where \(i\) is the current flowing between the working and counter electrodes, while \(R_{E}\) is the electrolyte resistance, defined as:
\[R_{E} = \frac{K_{ext}}{\sigma}\]
Such resistivity influences the transport of the ions through the solvent. These transfer processes are intrinsically limited by the electrolyte properties and can be described by considering three mechanisms: (i) diffusion, due to chemical potential gradients, (ii) migration, induced by electric fields and (iii) convection, resulting from the electrolyte motion. Mass transport is usually described theoretically through the Nernst-Planck equation, which is given, for the \(k\) species along the x-direction (one-dimensional case) by:48,831
\[J_{k}(x) = - D_{k}\frac{a_{0}(x)}{\Delta x} - \frac{x^{2}}{\delta T}D_{k}a_{k}\frac{a_{0}(x)}{\Delta x} - a_{k}v(x)\]
The right-hand side comprises three terms which refer, respectively, to diffusion, migration, and convection phenomena. It is worth mentioning that the diffusion coefficient can be expressed, through the Stokes-Einstein relation as:151
\[D = \frac{k_{B}v}{\mathit{s}\mathit{n}\mathit{d}v}\]
It is then clear that several physico-chemical properties must be considered to optimise mass transport in the electrolyte (resistivity, activity and dimension of the ions, viscosity of the solution and other aspects) to optimize the performances in terms of \(J\), scalability, and industrialization of the devices.153-154 Fluidodynamic modelling of the electrolyte in the PFC cell is complex due to the presence of an electrical field, a (typically) multicomponent porous electrode, the formation of gas bubbles and their adhesion to the electrode, the possible crossover over the membrane and other effects, but is a crucial element to optimize PFC and PV/EC cell engineering and scale-up.154
Moreover, CO2RR and NRR suffer from limitations related to the low solubility of CO2 (~3-10-2 M135 and N2 (~7-10-4 M) in H2O, that intrinsically limits the maximum achievable \(J\) in the range of tens mA*cm2. For this reason, other configurations have been investigated that favour high-/ working conditions, such as pressurized electrolyte systems and gas-diffusion-layer (GDL) architecture.156 In particular, in GDL-based devices the electrodes are in contact directly with the membrane acting as separation element between the two cells, but also as element to provide the ionic closure of the circuit.28,157,158 The photo- and electrocatalysis are present in the GDL electrodes put in contact with the two sides of the membrane.28,157 Among the advantages of this configuration (called electrolyte-less or gas-phase design) are the (i) not-energy-intensive recover of the reaction products, (ii) easier scale-up and lower costs, (iii) more compact design, and (iv) elimination of the problems of reactant solubility in the electrolyte and diffusional limitations through the double electrical layer. Due to the latter aspect, the electrodes operate in the presence of different surface coverages of the reactants, giving rise to different products in CO2 electrocatalytic conversion, for example.158
This concept is presented schematically in Figure 5, where the comparison between the architectures with and without electrolyte is presented (in this latter case, the closure of the electric circuit is provided from the membrane itself, which thus acts also as a solid electrolyte).28
The absence of a liquid electrolyte makes scale-up easier, and cell manufacture cheaper. Corrosion issues and material stability are largely overcome. However, mass transfer limitations in the two PFC cell design are significantly different, due either, in one case, to the absence of a liquid electrolyte, and in general to the different transport mechanisms for the two configurations.
In the compact design, oxygen evolution occurs by water oxidation (eq. 26a), while hydroxyl oxidation (eq. 26b) occurs at the anode for the conventional cell design of Notethat in the latter case, protons migrate through the membrane separating the two half-cells, while in the compact design protons migrate firstly through the porous oxide semiconductor, then through the membrane between the anodic and cathodic gas-diffusion-layer type electrodes.
\[2\text{H}_{2} + 4\text{h}^{+}\rightarrow\text{O}_{2} + 4\text{H}^{+}\]
\[2\text{OH}^{-} + 4\text{h}^{+}\rightarrow\text{O}_{2} + 4\text{H}^{+}\]
The compact design PEC cell requires to have an oxide semiconductor nanostructure allowing surface diffusion of protons (like based on ordered arrays of vertically aligned oxide nanotubes).199,190
### The impact of the device architecture
Herein, the discussion will focus on PEC and PV/EC design, which represent currently the most promising technologies for SFs generation. As remarked before, the case of PP architecture is not addressed here, being not possible to measure a _J._ in addition, this architecture still shows low efficiencies. For example, Domen and collaborators reported recently a 2-scheme system (La- and Rh-codoped SrTiO3 and Mo-doped BiVO4 powders embedded into a gold layer) as a superior system for OWS, although still having efficiency of 1.1%.2011 We want to emphasise that our aim is to give a viewpoint from which new considerations and studies can be carried out. We will present only some relevant results about these topics since making a detailed state-of-the-art review is out of the scope.
The production of SFs requires the use of energy which can be harvested by sunlight (unassisted SFs production) or be partially provided by an external bias (assisted SFs production). From a practical and cost perspective, only the first case is relevant and could be used as first generation artificial-leaf devices. Note that a device having one light harvester (without sun concentrators) produces at the best _J._CE < 70 mA-cm2 (_i.e._, the maximum photo-generated _J._; which depend on the material's properties). This holds true for both photoelectrodes and solar cells, so PEC and PV/EC designs. The advantage of PEC devices is that they are well-compact integrated systems (_i.e._, they can both harvest light and make chemical conversion), thus potentially having reduced cost of SFs production.
In addition, as commented later, they are better suited for use in conjunction with sun concentrators than PV/EC devices. On the other hand, there are various other aspects to consider which are in favour of PV/EC design. For example, the substantial projected cost of running electrolyte over large areas and providing non-fouling transparent flow channels for these devices which operate at low _J._ Also joined multi-junction solar cells are currently the most efficient PV devices. As discussed later, PV/EC cell design is that currently giving the best performances, but a detailed techno-economic comparative analysis of the different cell configurations, and their pro/cons is still missing in literature.
In terms of industrial use, many issues must be assessed, such as: the development of efficient absorbers at low cost and from large-scale fabrication methods,1992 (ii) increasing stability especially towards photocorrosion,1993 (iii) improving the performances. These criteria translate into the great challenge of finding optimized materials and device architectures.1994 With the aim to increase _J._, the \(g\) of the photo-absorber must be small thus decreasing the generated photovoltage, but which may result too small to trigger the redox reactions. In this case, an external bias is needed (assisted SFs production).
Unassisted SFs production can be realised through PC/PA architecture, when both are photoactive. By connecting the two photoelectrodes in series, the resulting voltage will be roughly the sum of the photovoltages produced by the PC and PA components while \(J\) will be the same for both (_i.e._, it will equal the smallest produced _J._). There are several possible reactor designs for PC/PA systems, depending on both the illumination setup and the photoelectrodes assembly.1994 The simplest model is the one introduced in the previous sections, with PC and PA facing each other and connected by wires, so the front photoelectrode needs to be transparent.
bi) Scheme of a compact (electrolyte-less) design for PEC operations.
Such wireless configuration shows lower performances than the wired counterparts, because of poor charge transfer and ohmic contacts between the different layers.5,144 A tilted configuration can optimise the spatial arrangement, moreover the PC and PA can also be put side by side, avoiding the need for transparency but doubling the space requirements. A tilted configuration is also possible (with PC facing PA at 45deg),168 but it makes difficult product separation.166 Thus, whatever PC/PA series architecture is considered, the highest \(J\) is always limited to ~70 mA*cm2. Only the parallel connection of two photoelectrodes results in a \(J\) which equals the sum of the \(J\) produced by the single photoelectrodes.168 However, the output voltage would be almost equal to the mean value of the voltage produced by the two photoelectrodes, so an external bias would be needed to reach the minimum potential to drive the reaction process.
Note finally that if materials are robust enough, PEC cells could be integrated with light concentration elements (Figure 6e-f). Here, higher \(J\) can be obtained (since in eq. 14 the intensity of _qE_) over but he that of one SUN illumination ~100 W-cm2 - but higher), according to the level of sum concentration.168 for PV cell, the effectiveness in increasing the \(J\) by solar concentrations is limited from the negative effect of the increase of the temperature on the PV cell performances. However, using photoelectrodes in PEC cell makes this limitation likely less relevant (the dependence from the temperature of the performances of oxide semiconductor photoanodes is different from that of PV cells), although thermal robust components for the PEC cell are necessary. This is a largely unexplored area, but crucial to enhance PEC \(J\) and performances.
Materials and cells able to operate in a medium temperature range (50-200degC), which is not compatible to most of existing technologies, should be developed. However, the drastic possible reduction in costs (which largely depend on \(J\), being a main parameter determining cell productivity) is a main driver from an industrial perspective.
A related challenge is that the devices should operate efficiently also at variable operational temperatures, because of fluctuations of the SUN intensity due to seasonal variations as well as geographic location. It is thus an area which can open new application perspective, but which is still largely unrecognized.
As concerns the large scale implementation of PEC systems, many practical issues must be properly assessed such as performance of scaled-up electrodes (which may reach ~80% loss of current with respect to lab-scale devices and increased ohmic resistances,9 although the effective loss of performances could be lower than 10%), the need of avoiding critical raw materials for the realization of the PEC system and especially cost-effective solutions (including stability in the analysis) and safety concerns for operations. While commercialization of the systems for OWS is expected within 5-10 years, those for CO2RR and NRR will take longer time for the higher challenges in their development.
The other main architecture for SFs production is the PV/EC assembly, which presents several practical advantages with respect to other approaches. First, the carrier generation/transport and catalytic conversion processes are decoupled, so they can be optimised by the proper choice of the PV and EC components. Many PV technologies are already commercially available, each one being characterised by different performances (in terms of PV parameters such as _Jsc_, open circuit voltage, power conversion efficiency and other aspect), stability and cost.168,167 Furthermore, in such architecture the PV component lies outside the electrolyte environment, avoiding
Architectures for PC systems. Monolith (a) planar and (b) tilted PC/PA configuration. Side by side configuration of (a) planar and (b) tilted PC and PA electrodes. e-f) Monolith PC/PA devices coupled with sunlight concentration systems. Re-elaborates from ref.168
Current density-potential behaviour for a photovindicator and a water electrolyte: elaborated from ref.168 Copyright Elsevier ©2017
Finally, it has the potential to deliver high \(J\) by connecting in parallel more solar cells. The resulting current will be the sum of all the currents produced by the single devices. However, the working point of the PV/EC assembly equals the intersection of the _J-V_ curves of the PV and EC components, as shown schematically in (thus the power input to the EC equals the power output of the PV).189
For an ideal system, the intersection would occur at the maximum power point of the PV unit (_Jap_ ideal, reported in Figure 7), but the overpotential associated to the considered reaction can shift such point to lower \(J\) (_Jap_ actual, shown in Figure 7). In the PV/EC design _J_SC indicates the photocurrent produced by the PV unit, while _Jap_ refers to the practical current exploited during the electrochemical process. Thus, a correct definition, not always present, is fundamental.
8 with proper considerations on the \(J\), to derive reliable conclusions on the more promising materials and approaches.
The indication emerging from this combined analysis is that for OWS case, the PV/EC architecture allows to obtain high-_J_ and _rSNR_, especially when coupled to a concentrator system and a proper heat management system, but which increase the costs. Furthermore, performance losses after long-lasting use must be accurately assessed (such as the detachment of catalyst particles and membrane degradation).40 The introduction of low-cost solar concentration elements coupled to an efficient thermal management of the PV/cell, together with the use of thermal robust materials for the cell itself, will be likely the crucial factor for commercialization, although extremely few papers (of the very abundant literature on the topic) address these three key aspects.
Through the years, the PFC architecture was the most investigated, since it would provide (I) 2% < _rSNR_ < 19%, (ii) easier and cheaper separation of the reaction products and (iii) potential lower costs for the fabrication of the photoelectrodes with respect to the PV/EC assembly. However, to the best of our knowledge, no large-scale demonstration of PFC powered systems has been reported yet.
The best performances have been obtained for tandem PFC systems, mainly based on III-V semiconductors light harvesters. For example, Cheng et al.178 achieved _rSNR_ = 19%, at \(J_{K}\) = 15.7 mA-cm-2, using a monolithic device made up by a GainP/GainAs two-junction absorbing layer using Rh and RuO, as HER and DER catalysts, respectively. Interestingly, the authors performed some calculations to assess the origin of losses in their device. According to their results, kinetic, diffusion and ohmic contributions to \(J\) and _rSNR_ do not represent the dominant losses mechanisms. In fact, by measuring the IPCE of their device, the performances were found to be limited by \(J_{K}\), since it should potentially reach ~17.5 mA-cm-2. Thus, this work emphasises that a careful analysis of the origin of losses in \(J\) in the device can guide the choice and optimization to maximize the production of SFs. Often these aspects are more critical than the typical aspects related to catalyst design discussed in most of the reviews on the topic. However, there are only few literature results that present such kind of analysis, making comparisons with other studies rather difficult.
For OWS using PEC architecture, several materials have been intensively studied, for example binary oxides, oxynitrides and photovoltaic-grade materials among others. They were discussed in various reviews,3,12,10,13 while an analysis of their performances is beyond the scope of this review. As mentioned earlier, for all these systems the maximum \(J_{K}\) is limited by the \(E_{S}\) of the material used as light harvester, typically not exceeding -20 mA-cm-2 (Table 2). Since for industrial applications a \(J_{K}\) > 300-400 mA-cm-2 should be indicatively targeted, the PV/EC approach currently represent the only choice to meet this goal, in the absence of using concentration elements. Even for PV/EC approach, this target could be reached only by using sun-concentrations elements as shown in Table 2. There are intrinsic limits both in PFC and PV/EC approaches to reach these targets without a sun-concentration component and the associated system of thermal management, although an optimal trade-off between costs and performances is necessary, as well as considerations on the stability of the systems. On the other hand, it is evident that without an analysis of \(J\), these critical aspects in the discussion about solar devices are not emerging.
### Preliminary considerations on CO2-RR and NRR
The realization of efficient CO2-RR and NRR suffer from additional relevant limitations with respect to OWS. In particular, differently from OWS where a single product is formed (H-I), a range of chemicals can be synthesised in both O2-RR and NRR. Furthermore, for both CO2-RR and NRR the H2 side formation influences the Faradatic selectivity.
For CO2-RR, it should be distinguished between carbon Faradatic selectivity (which refers to the selectivity with respect to only the carbon-based products formed) and total Faradatic selectivity, which includes H2. Both are relevant, but carbon Faradatic selectivity is often more important for industrial exploitability.
\begin{table}
\begin{tabular}{c c c c} _Design_ & _Components_ & \(J_{K}\) / \(J_{K}\) & _rSNR_ & \(R_{F}\) \\ & & (mÅ+cm−2) & (_K_) & \\ \hline \multirow{2}{*}{PV/EC} & GainPto NH4, such as NH4NH2, but their contributions are usually negligible. Thus, total Faradai selectivity is the typically only considered.
For CO2RR, the resulting products have a different value (in terms of performances as SFs) and a mixture of chemicals can be obtained. This is an aspect often ignored, although from the application perspective of these devices the downstream separation and purification processes could be the determining cost element for implementing the SFs production. Recovery of the products during continuous operations and avoiding electrolyte degradation (also related to by-products accumulation) are other crucial elements scarcely considered in literature.
Finally, the low solubility of CO2 in H2O-based electrolytes intrinsically limits the maximum \(J\) to -35 mA-cm-2, about one order of magnitude lower of an indicative value of > 200-300 mA-cm-2 to have productivities in the range of industrial relevance.156
A starting point to compare with OWS case is to have a graph analogous to that presented in Figure 8, because this kind of comparison is absent in literature. summarises these results for CO2RR case, with an indication of also some of the possible products of reaction. Note that the limits of range of efficiency versus system complexity are larger for CO2RR with respect to H2 production, because the products have in average a higher added value and usability. In addition, they contribute directly to the reuse of CO2 from emissions (in perspective from the air). However, more precise estimations should be made case by case, by considering the specific products formed.
Also note that despite the abundant literature on CO2RR, only few results provide data in terms of reliable indication of _TSF_ in the absence of external bias and added sacrificial agents. Thus, can stimulate the readers to provide in future publications data which could be directly put in relation to this graph.
However, a similar figure cannot be made for NRR for the absence of enough reliable data. All results of photocatalytic N2 reduction do not provide indications on _TSF_. Those for which an estimation can be made (from available data of external quantum yield) indicate _TSF_ significantly lower than 0.1%. Some papers report a PEC of PV/EC approach in NRR, but with an applied external bias. Without this additional voltage (ranging typically between 0.3-0.6V), the formation of NH3 is quite negligible, indicating extremely low _TSF_ values. Quantitative indications of \(J\) for NH4 production are extremely limited, even for the case of a pure electrochemical approach (thus with external bias) for which a maximum around 0.5 mA-cm-2 was demonstrated (with Faradai efficiency to ammonia below 30%).88,205
For unassisted PEC or PV/EC approaches without sacrificial agents, \(J\) values are lower, nearly negligible in the few results reported up to now (see the following section for more details). Thus, NRR stage of development is still largely below that for CO2RR and OWS, preventing an analysis as that attempted in Figures 8 - 9.
In addition, in CO2RR and NRR the operational \(J\) is determined from the need to limit the side reaction of HER and thus productivity is limited. For CO2RR and NRR, the reaction mechanism is more complex than that for H2 production because more electrons/protons are involved, and reaction intermediates and possible pathways are competitive to the SF production. All these steps should be optimized and will depend on the applied potential. PV/EC approach, which is the best option for the OWS case, does not seem preferable also for CO2RR or NRR. In fact, does not evidence a clear distinction between the performances of PEC and PV/EC technologies. Also in this case, testing essentially only at room temperature, and without the use of sun-concentrations has been reported.
In principle, maximizing the overpotential difference between that necessary for CO2RR or NRR reactions, and that for the side HER reaction (which should be the highest possible) would allow to operate at high \(J\), while maintaining high the Faradai selectivity (typically indicated also as Faradai efficiency), at least with respect to the side HER reaction. However, this is not a trivial task because it is necessary to activate hydrogen to allow the hydrogenation of the adsorbed CO2 (or the products of further reduction) or N2 while avoiding the recombination between activated hydrogen species that would form H2. Thus, an optimal compromise is difficult to obtain.
Moreover, HER is a "simple" reaction since it involves only two electrons (eq. 6), while CO2RR and NRR usually require a higher number of electrons, except for CO or HCOOH production (eq.s 8a and 8b). Faradai efficiencies for these reactions above 90% have been achieved with high \(J\) in electrochemical reactions. For example, Wu et al.206 reported for a PbO2 electrode in the electroreduction of CO2 to HCOOH in an ionic liquid electrolyte a Faradai efficiency of 95.5% and a current density of 40.8 mA-cm-2 simultaneously. However, selectivity decreases when products of CO2 reduction, involving >2 electrons, are targeted such as CH2OH (a 6e reduction, eq. 9) or products involving also C-C bond formation (eq.s 10-12 as an example). In addition, the surface coverage by adspecies also plays a crucial role, and in fact by increasing the pressure of operation in CO2RR, the Faradai selectivity can be improved.207 The lower surface coverage in N2 with respect to CO2, due to strong N=N bond, explains why the Faradai selectivity is typically lower in NRR with respect to CO2RR.87,208
Note also that reliable data on carbon or nitrogen balance in CO2RR and NRR are usually not given. Significant crossover from the cathodic zone (where CO2RR and NRR occurs) to the anodic zone of the cell may be present (or viceversa), but often this is not an aspect considered. Analyzing the experimental \(J\) with respect to the \(J\) calculated from the current density needed to form all the detected products is a straight manner to have indication on these aspects as well as on other effects (for example, other side reactions including the reduction of the electrocostably). It should be thus part of the protocol of experimentation, although typically not made.
As final comment, it may be argued that an alternative strategy, giving the better performances in OWS than CO2RRand NMR, is to produce H2 by PV/EC approach, and then use H2 in catalytic (thermal) processes of CO2 or N2 conversion. This solution, however, is not efficient in terms of integration in small-scale devices, and thus does not adapt to the development of artificial-leaf type (distributed) devices.
Direct CO2RR and NMR in PFC (or analogous) devices remain a preferable strategy for several reasons: (i) the overpotentials necessary in H2 formation (HER) and in use H2 (need to activate and compress H2) are eliminated, due to the direct use of protons/electrons in N2 or CO2 reduction, (ii) mild operative conditions are used, and (iii) costs are potentially lower using an integrated device rather than two separate units (electrolyte) and catalytic unit). If improved electrocatalysts could be developed, a potential reduction of both fixed and operative costs can be achieved.
### The case of CO2
Galan-Mascaros,11 Creissen and Fontecave155 recently reviewed the best performing results about PV/EC and PFC systems for CO2RR commenting the low number of examples of unbiased solar CO2 reduction processes. Table 3 reports the comparisons of _e.g._737 and \(J\) of representative literature examples for CO2RR.
Cheng et al.2009 reported top performances for PV/EC approach of a _e.g._737 up to 19% (at 1 SUN) with current density of about 14 mA-cm-2 but only towards CO formation and using a costly triple-junction solar cell. The proof of the formation of only CO (and carbon balance) is not fully convincing. The authors claim long-term stability. Outdoor operations allowed a peak solar-to-CO efficiency of 18.7%, stable for >150 h and _J_-15 mA-cm-2. Schreier et al.189 reported a _e.g._737 of about 133%, also to only CO (but slightly lower Faradatic selectivity) and with a slight lower current density (about 14 mA-cm-2). They also used a Galp/Gantas/Ge costly triple-junction solar cell, but earth-abundant materials for the electrocatalysts, differently from Cheng et al.2009
Zho et al.199 reported a _e.g._737 of about 100% to formate, with current density of about 9 mA-cm-2 also using a costly InGaP/GaAs PV component. A _e.g._737 of about 8.5% to formate with current density about 10 mA-cm-2 was reported by Pia et al.2134 also using a triple-junction PV cell, but based on less-costly Si. However, a costly component such as IrO2 was used as electrocatalyst. Kato et al.213 reported a _e.g._737 - 73% to formate using a S-based PV cell and costly noble-metal based electrodes. The cell consists of five stacked electrodes (electrically parallel connected) and six series-connected single-crystalline Si PV cells (area ~1,000 cm2). The electrodes contain quite expensive elements (anodes based on IrO2 and cathodes on Ru complex polymer) whose stability is unclear. The cell is without a membrane and operates at 1.85 V and 6.30 A.
Lower _e.g._737 values have been reported in producing SFs involving a higher number of electrons for the reduction than the 2e for CO and HCOOH. For example, ethylene was obtained at _e.g._737 = 2.9% combining Si solar panels with an electrolyzer.207 Membrane-less electrospapers powered by Si PV cells reached a _e.g._737 >336 to C091 using earth-abundant catalysts with _J_-1 mA-cm-2.
The integrated PFC approach for CO2RR results in lower efficiencies, with _e.g._737 typically below 1% (Table 3).11 Zhou et al.212 reported _e.g._737 >100%, but for a PFC device incorporating a photonode made of a TiO2/InGaP/GaAs heterostructure decorated with a Ni OER catalyst which drives a dark Pd/C/Ti cathode producing formic acid at >94% Faradatic efficiency.
Current density is about 8.7 mA-cm-2. Incorporation of Si-heterojunctions in the photomode and photocathode yielded _TNN_ > 4% with CO as major product using ionic liquids as electrolyte and a WSe2-based cathode.196 A TiO2/p-Si photocathode modified by cobalt bis(terpyridine) has been reported as active and stable system for CO2 to CO production,200 but \(J\) were quite low (about 0.1 mA-cm-3). Data for \(J\) were not always reported directly, so they could be deduced from the other results, but indications in Table 3 evidence that typically the values are <1 mA-cm-2.
Few results have been reported for integrated PEC devices in producing C2+ products,194 although a recent review discussed the photocatalytic CO2 reduction to C2+ products.22 Combined with an external PV cell, an electrolyser based on a Cu/Ag/TiO2 photocathode results in 60% Faraadic efficiency to form a mixture of C2/C3 products. A tandem device was developed by coupling a Si photocathode with two series-connected semi-transparent CH3NH2+phs perovskite solar cells, achieving an efficiency for the conversion of sunlight to hydrocarbons and oxygenates of 1.5% (3.5% for all products, which include H2, CO - the dominant products -, ethylene, ethane, formate, propionaldehyde, alkyl alcohol, ethanol, 1-propanol, and other C2+ products).
Values of \(J\) range from 2 to 3 mA-cm-2 depending on the electrolyte concentration. Table 3 evidences two other aspects: (i) focus is on two electrons reductions (CO, HCOOH), while more complex reactions will be likely preferable; one case,194 however, reports good performances to C2/C3 products, although producing a range of different chemicals, and (ii) \(J\) are generally low, and only slightly better when a PV component is integrated in the system, although typically with problems of stability, mainly related to photoconversion.
The CO2 electrocatalytic reduction (thus without direct coupling with a photoanode or a PV cell) is an area of extensive current studies and several works were reported about the electroreduction of CO2 at elevated \(J\) (in the range of hundreds mA-cm-2).21,223 Such results have been obtained by exploiting device architecture that overcome the solubility problem of CO2 into water-based electrolyte. For example, gas-phase flow cells have been used for the production of alcohols (methanol, ethanol and n-propanol) and CO at, respectively, \(J\) of 180 and 60 mA-cm-2 and FE-40 (cumulative) and >90%.21,223 Furthermore, Li et al.226 demonstrated that gas-fed flow cells deliver up to twice more \(J\) (reaching also 200 mA-cm-2 with a FE ~50% towards CO) with respect to those based on liquid electrolyte (as shown in Figure 10).
Weekes et al.222 and Endr/ddi et al.223 discussed the role of cell architecture in the electroreduction of CO2, while more detailed aspects are out of the scope as well as the analysis of the recent development in electrocatalytic CO2 reduction, which for formate production from CO2 reached recently extremely high current densities (>930 mA-cm-2) with high Faradac efficiency (FFsCO0 = 93%) on defective Bi2O2CO2 nanosheets.227
While thus high \(J\) and FE are possible in the electrocatalytic approach (with an external source of potential and electrical current), the direct integration of a photostoxeture element in the device (photoanode, PV cell) allows largely lower \(J\) values, although with other advantages as earlier discussed.
Based on the presented results, the PV/EC design can be the best solution for reaching high-_J_ also for CO2+,RR, but the EC component must be properly engineered to account for the mass transport limitations that influence this cell performances. Cheng et al.200 (Table 3) reported an example of such synergistic approach using a triple junction solar cell (GainP/GainAs/Ge) to power an Ag-based GDE. The coupling between the photostoxeture element and the devices is optimal (_I_0_a_-_k_) and the electrocatalyst shows high selective FE (99th). This result was obtained by optimizing many parameters (electrolyte pH and compositions) in relation to \(J\). Another work, by Huan et al.2008 used a perovskite-based solar cell and a CuO-based EC to reduce CO2. Although the performances are modest (as shown in Table 3), the use of a perovskite solar cell is an interesting path towards low-cost PV components for PV/EC systems.
However, as pointed out by Burdyny and Smith,196 current studied on CO2RR mainly focus on low-\(\wedge\) conditions, thus the information obtained within this framework may not be useful when considering systems delivering high-_J_. In particular, it is not still clear if CO2RR proceeds as a three- or two-phases reaction, moreover some electrolytes (used for high selectivity) can exhibit large ohmic losses when operating at high-_J_ conditions, which hinders their use in practical industrial-scale devices.
### Nitrogen reduction reaction
Reducing N2 to NH3 is a challenging and intriguing chemical reaction since the N2 molecule is rather inert due to the high stability of the N=N triple bond. In contrast to many papers on electrochemical/catalytic approaches for NNR,87,88,28,289,290,291,292,293,294,295,296,297,298,299,299,290,291,294,295,296,297,298,299,299,299,299,299,299,291,299,299,299,299,299,299,291,299,299,2992,299,2993,294,2995,296,2997,298,2999,2998,2999,2999,2999,2999,2999,2999,2999,299,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,2999,29999,2999,2999,29999,29999,29999,2999,29999,2999,2999,29999,2999,2999,2999,2999,2999,29999,29999,2999,29999,2999,2999,29999,29999,29999,2999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,299999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,299999,29999,29999,299999,29999,29999,299999,29999,29999,299999,299999,29999,29999,29999,29999,29999,299999,29999,29999,299999,29999,299999,29999,29999,299999,29999,299999,29999,29999,29999,299999,299999,29999,299999,29999,299999,29999,29999,29999,29999,29999,29999,29999,299999,299999,299999,299999,29999,299999,29999,299999,299999,299999,299999,29999,29999,29999,29999,29999,29999,299999,299999,29999,29999,29999,29999,29999,29999,299999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,2999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,2999,29999,29999,29999,29999,29999,29999,299999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,29999,2999,29999,29999,29999,2999,29999,29999,29999,29999,29999,29999,29999,29999,29999,299999,29999,299999,29999,29999,299999,29999,29999,29999,299999,29999,299999,29999,29999,299999,2999,29999,299999,29999,29999,29999,29999,299999,29999,29999,299999,29999,29999,299999,299999,29999,299999,29999,29999,299999,299999,29999,29999,299999,299999,29999,299999,29999,299999,29999,29999,29999,299999,299999,29999,29999,29999,299999,29999,299999,299999,299999,29999,299999,299999,299999,299999,299999,299999,299999,29999,299999,299999,299999,299999,299999,299999,299999,2999999,299999,299999,299999,2999999,299999,2999999,2999999,299999,299999,299999,2999999,2999999,2999999,2999999,2999999,2999999,2999999,2999999,2999999,2999999,2999999,29999999,2999999,2999999,2999999,2999999,299999,2999999,2999999,2999999,2999999,29999999,29999999,29999999,2999999,2999999,299999999,299999999,29999999,29999999,29999999,29999999,29999999,299999999,299999999,29999999,29999999,299999999,299999999,299999999,299999999,2999999999,299999999,299999999,2999999999,299999999,2999999999,2999999999,2999999999,2999999999,29999999999,2999999999,29999999999,29999999999,2999999999,29999999999,2999999999,2999999999999,299999999999999respectively)232
* - Ag nanoparticles on black silicon photocathodes238 giving a current density (at the maximum NH3 production rate of 2.87 mmol-h-1-cm-3) at -0.2 V, thus bias assisted) of about 0.12 mA-cm-3 with FEsu ~41% (deactivation is observed in about 2h of experiments); \(e\)3_PV is not given, but estimated to be <0.1%
* - Au-poly (tetrafluoro ethylene)-Si photocathodes taking advantage of a hydrophobic environment to preclude HER234; ammonia yield at the best (-0.2 V, thus bias assisted) is about 19 mmol-h-1-cm-3 with FEsuae about 39%; \(e\)3_PV is not given, but estimated to be <0.1%; for J is nearly zero without applied potential and increases by applying a negative potential, but still remaining below ~1 mA-cm-3
* - Au-SrTiO3 plasmonic photocathodes, associated with a Zr cathode235; without applied bias the FEsuos is nearly zero and J about 2 mA-cm-3 with an IPCE at the best of 0.07% at 600 nm (thus much lower \(e\)3_PV).
Therefore, current state-on-the-art in NRR in PEC devices indicates extremely low / (< 0.1 mA-cm-3) and \(e\)3_PV (estimated significantly below 0.1%), and the typical need of applying an external bias to obtain measurable results.
In some case, the term photoelectrochemical was used to indicate an electrocatalytic cell, to which an external bias was mandatory for its function. For example, Mushta et al.236 claimed to have investigated the PEC N2 reduction to NH3 in a two-compartment H-type cell separated by the Nafion membrane. However, there is no photoanode or external PV cell, and it is thus an electrochemical (or electroneutralic, EC) cell, not a PFC device. Similarly, Zheng et al.217 in reviewing the photoelectrochemical conversion of CO2 and N2 into useful products on TiO2 (titanate) based nanomaterials, indicated a series of studies of PEC NRR.235,238-242 However, the results were instead referring to different situations. Xu et al.238 used a 3D-printed hierarchical porous TiO2 scaffold immersed in a solution and illuminated, thus like PP case, but no values of \(J\) or \(e\)3_PV were reported.
Shi et al.239 indicated photoelectrochemical measurements but ensuring the photocurrent of an Au/TiO2 electrode applying <0.3 V potential and the observed photocurrent was very low (< 0.01 mA-cm-3). Oshikri et al.230 used Au-doped SrTiO3 photoelectrode with a two-compartment cell, but separated by the NbSrTiO3 dense layer, illuminated by a Xe lamp. The closure of the electronic circuit is unclear, and a sacrificial agent (ethanol) is present in the anodic chamber. Li et al.230 used plasmon-enhanced Au-TiO2 (nanorod array) photoelectrode immersed in an aqueous solution, which is a PP-like approach, rather than a photoelectrochemical approach as reported and the resulting \(e\)3_PV is below 0.002%. Ye et al.241 also indicated a photoelectrochemical approach based on nanojunctions assembled from MoS2 nanosheets and TiO2. They used a two-chamber H-type cell divided by a Nafion 211 membrane, with the (deposited on carbon substrate) on the illuminated side. However, to observe NH3 formation it was necessary to apply an external bias and the results were unclear in separating the contribution of illumination from that of the electrochemical cell.
Li et al.242 recently reported PEC NRR using a Mo2C/C heterostructure, and BiVO4 as photoanode. Also here an external bias (0.2 V) was applied which doubles the production of NH3. No indication is given in terms of \(e\)3_PV and the estimated specific current density to ammonia was below ~0.03 mA-cm-3 in the absence of an external bias and thus low.
Wang et al.243 also reported the photoelectrochemical NH3 synthesis using a Co-phosphate/Ti-Fe2O3 photoanode and Co-SAC (single-atom catalyst) cathode. However, the PEC device has no separate compartments. In addition, an external bias should be applied to observe both photocurrent (which is enhanced with respect to case without illumination) and formation of NH3. Thus, it is a light enhanced electrochemical NRR, rather than a PEC NRR as indicated.
As emerges from this analysis, data on the PEC approach for NRR are limited, and results without an external bias, or the use of sacrificial agents, are not available. Moreover, both \(J\) and \(e\)3_PV are either not given or extremely low. A question is thus why these poor performances not only with respect to OWS, but also to CO2RR.
One of the differences between CO2 and N2 reduction cases is related to the coordination of the molecules on the surface of the electrocatalyst. Both the strength of C=O and N=N bonds is different, as well as the availability of \(d\) orbitals for coordination, making more challenging N2 molecule activation. In Fe or Ru catalysts for heterogeneous (thermal) catalysis, dissociative chemisorption of N2 is the first and rate limiting step. The general agreement is that in electrocatalysis, hydrogenation of undislocated coordinated N2 molecules is the rate determining step.27 End-on N2 molecule coordination at mono- or bi-metallic sites is the first step of N2 coordination, followed by sequential or simultaneous multi proton-assisted electron transfer.
Note that, differently from conventional electrochemical electron transfer, this process is typically considered to occur on chemisorbed CO2 or N2 molecules, rather than through an outer shell electron transfer mechanism as in electrochemistry theory.
The difference in CO2 versus N2 chemisorption on the electrocatalyst is a main reason explaining the different behaviour, both in terms of activity (and _l_) and Faradalc efficiency (lower surface coverage by N2 would imply also easier side HER).
In addition, a further limiting factor for \(J\) is N2 solubility, which is also related to NRR selectivity because a higher selectivity increases the surface coverage by N2.78 Note, however, that as commented for CO2 and outlined in Figure 5, it is possible to operate with GDL-type electrodes and cell configurations, which largely overcome the problem of N2 solubility (as assessed previously for CO2RR). This aspect, however, is rarely recognized.28 Thus, the design of devices should be focused at maximizing sites coordinating molecular N2 in a form susceptible for proton-assisted electron transfer. On the other hand, cell design, which can force the creation of an enhanced local pressure of N2 (or enhanced chemisorption) at the electrocatalyst surface would be equally relevant. Creating a suitable overlayer on the catalyst which enhances the N2 adsorption and create a pseudo-virtual higher N2 pressure at the electrocatalyst is a way to explore, but not attempted as far as we know.
An improved design of the electrocatalystysis is necessary for an efficient (photo) electrochemical NHA synthesis, since current results are still far from those required in industrial applications (for both FE and \(J\) values).38 While undoubtfully there is the need to a better mechanistic understanding of the reaction and the identification of the active sites, current approaches (including theoretical methods) are largely unable to properly identify these aspects. In front of an exceptionally large range of different indications on the active sites and reaction mechanisms, all the results fall within a limited range, pointing out that the aspects identified are likely not those relevant.4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,17,18,19,19,19,17,18,19,19,19,17,18,19,19,19,17,19,
## Conclusions and perspectives
The production of SFs (H2 from water and products of CO2 and N2 reduction) is a promising solution towards the foundation of a green and sustainable economy. Several technologies have been proposed with the aim to optimise the performances of the resulting devices such as particulate photocatalysts (PP), photoelectrochemical (PEC) catalysis and photovoltaic driven electrocatalysis (PV/EC). In this perspective, we analysed the role of current density (_V_) as a fundamental figure of merit because (i) it allows a more direct comparison between different approaches for SFs production and ii) it gives valuable information about the physico-chemical processes involved during the device operations. The latter aspect is crucial since a thorough knowledge of the working mechanisms (and associated limits) is fundamental to progress in the developments of technologies for SFs production. For convenience, the physico-chemical processes behind the conversion of sunlight into chemical fuels were divided in four steps: i) photon absorption and generation of free charge carriers, ii) separation and transport of charge carriers, iii) catalytic conversion and iv) mass transfer. Careful analysis of J gives information about each step, in particular:
* The short circuit current density (I_SC) is the maximum current that can be photo-generated by the light harvester used in the device. Many devices use one single light-harvester, thus the maximum theoretical I_SC is 73 mA-cm2 for a material with zero energy bandgap and one SUN irradiation. This has two main consequences low band-gap materials are needed to achieve high \(J\) (but this requirement contrasts that of minimum photovoltage needed to trigger the desired reaction) and ii) \(J\) comparable to those used in existing technologies for chemical fuels production (~0.5-2 A-cm2) cannot be practically achieved. When values of I_SC are lower with respect to the theoretical ones, the production of free electron/hole pairs is hindered by other processes such as exciton formation, that can be mitigated by proper engineering of the dielectric properties of the light-harvester (which influence the exciton binding energy).
* Several mechanisms can be responsible for \(J\) values lower than I_SC (at applied biases + OV), arising from the other three steps herein discussed. The separation and collection of the photo-generated charge carriers can suffer from recombination phenomena: monomolecular, bimolecular and Auger processes are the main ones. Thus, a careful assessment of the recombination constants is fundamental to set realistic goals for the maximum \(J\) obtained by a device assembly. The probability of occurrence of such phenomena is further increased by a low catalytic activity of the cocatalysts: if charge carriers are not quickly used to trigger the reaction of interest, they have a greater chance to recombine. Finally, sluggish motion of the ions in the electrolyte hinders the efficient "refuel" of the reactants adsorbed on the cocatalyst surface, which slows down charge carrier motion and so \(J\). As a result, all these steps must be properly accounted for to maximise the performances of the device.
Each technology has its pros and cons to produce a particular SF. For example, with H2, the most promising architecture is the PV/EC which can deliver remarkably high \(J\) thanks to the proper series/parallel connection of several PV units. Large scale demonstrations of such systems have been studied, with very encouraging results. However, the costs associated to both the PV and EC components have questioned the economic impact that such technology would pose on the market. PEC approach still is a hot topic among the scientific community (for converting small molecules, i.e. H2O, CO2 and N2)2(see in this case both charge carrier generation and catalytic processes are carried out by a single component. Here, the produced J would be limited by the absorption properties of the light harvester and in general it cannot exceed 73 mA-cm2 so, there is a trade-off between architecture complexity/cost and performances.
Differently, for the case of products of CO2 and N2 reduction, due to various additional issues (with respect to OWS) making these reactions (CQ,RR and NRR) more challenging and still far from the application. As commented before in analysing the state-of-the-art, PEC approach appears the more promising for these reactions, but \(J\) values are low. There are scattered results and not systematic regarding how to improve cell design and \(J\). There is the need to intensity R&D on these aspects, but it requires to first go back in understanding better the crucial parameters determining the performances and how to design an optimal cell to minimize the current limitations.
A general conclusion which can be derived from the discussion presented here is that from an implementation perspective of producing solar fuels from CO2 and N2 conversion in PEC or PV/EC devices, the need to operate these cells in conjunction with sun concentration (in the 50-100 SUN range) clearly emerges. However, this is not just an additional parameter to investigate, because it requires a fully redesign of both materials and cell to operate efficiently and stable under these conditions. Most of the systems and design approaches development up to now have severe limitations to operate under these conditions, and thus a general comment emerging from the discussion is the need to rethink current approaches to produce solar fuels. For NRR, there is also the additional requirement of developing efficient solution increasing the surface coverage of the electrocatalyst by N2. The novel possibility to use PEC devices to oxidize N2 to NO, on the anodic side and then reduce to NH2 the NO, on the cathodic side has been shortly introduced, although is a novel unexplored direction.
In conclusion, this perspective paper provides indications and clues to identify relevant directions to explore to implement solar fuels distributed production in PEC or PV/EC type of devices. The concept of current density is a useful figure of merit to address these aspects.
## Acronyms
CB conduction band CBM conduction band minimum CES chemical energy storage
|
10.48550/arXiv.2205.05102
|
Current density in solar fuel technologies
|
Valentino Romano, Giovanna D'Angelo, Siglinda Perathoner, Gabriele Centi
| 3,532
|
10.48550_arXiv.0909.3824
|
###### Abstract
The consequences of source charge and surface modulation are studied within the framework of the Poisson-Boltzmann theory of electrolyte solutions. Through a consideration of various examples, it is found that inherent modulation can lead to both like-charge attraction and overcharging effects.
pacs: 82.70.-y, 61.20.Qg
## I Introduction
As in any conductor, the potential due to a source distribution is screened within an electrolyte solution. The screening charge carriers, or counterions, in these classical systems are ionized atoms or molecules. The qualitative result of the screening is typically only to reduce the distance out to which the potential from a source distribution is significant. However, experiments have shown that multivalent counterions can allow for more exotic behavior. For example, over-efficient screening due to multivalent counterions can sometimes result in the charge reversal of a molecule or substrate. In a related effect, multivalent counterions have been observed to mediate attractions between similarly charged macromolecules. Study of these counterintuitive effects is pertinent to the understanding of many biological and colloid systems, in which molecular conformations and interactions are often sensitively controlled by counterion concentration.
The Poisson-Boltzmann (P-B) mean-field theory provides one of the simplest descriptions of electrolyte response and often allows for accurate modelling of these systems. However, it has been proven for mirror-symmetric geometries that like-charge attraction is impossible for a pair of identical molecules within this theory. This has bolstered the general opinion that the physical mechanisms behind like-charge attraction and overcharging cannot be understood within a mean-field description of electrolytes. Most recent works have thus attempted to understand these observations through a consideration of fluctuation corrections to the mean-field response. Certainly, for homogeneous systems, it seems clear that fluctuations provide the only possible mechanism for bringing about the charge correlations necessary for these effects. This conclusion does not immediately follow for heterogeneous systems, however, since static modulation allows for the introduction of some inherent charge correlation. Indeed, studies have shown that heterogeneity can play an important role in these systems. For example, numerical simulations have shown that modulation of molecular geometry can significantly affect interactions. In addition, it has been shown perturbatively that within the P-B theory, charge condensation onto a surface in an electrolyte solution may be increased when the surface charge contains certain limiting forms of modulation. This extra condensation can in turn lead to a decrease in the osmotic pressure between two surfaces, demonstrating that mean-field effects can at least help to facilitate both overcharging and like-charge attraction.
It is the purpose of the present paper to further explore the effects of inherent, quenched modulations within electrolyte solutions. Fluctuation effects are neglected and the electrolyte response is considered within the P-B formalism. Significantly, it is shown that inherent modulations alone, can, in fact, allow for both like-charge attraction and overcharging mechanisms. The key to like-charge attraction within the P-B theory is the breaking of mirror symmetry, which was assumed to hold in Ref.. A simple example is given which demonstrates this point. Heterogeneity-induced condensation is studied using a perturbation approach similar in spirit to that considered in Ref.. The present approach consists of expanding about solutions to the linearized P-B equation and allows for the consideration of additional charge modulation limits, boundary surface modulations, and the effects of having more than one species of screening ion. Various examples are considered in order to provide a brief survey of these different set-ups. Of particular interest is the counterion concentration dependence exhibited by systems containing multiple species of ion. In principle, this dependence should often allow one to determine whether modulation is a significant contributor to overcharging in any particular experimental observation.
An outline of this paper is as follows. In section II the P-B theory is reviewed and solutions to the linearized P-B, or Debye-Huckel, equation are presented. Charge and surface modulations are considered in sections III and IV, respectively, while section V contains concluding remarks.
## II The Poisson-Boltzmann equation
If one assumes that the mobile charge distribution in an electrolyte solution is given by a sum over ion speciesof Boltzmann factors multiplying bulk charge densities, then the Poisson equation reads
\[\nabla^{2}\phi = -4\pi\sum_{i}e^{-q_{i}\phi/kT}q_{i}n_{i}-4\pi\rho_{s} \tag{1}\] \[\approx \gamma^{2}\phi-4\pi\rho_{s}.\]
Here, \(\phi\) is the electrostatic potential, the \(q_{i}\) and \(n_{i}\) are the charge and bulk densities of screening species \(i\), \(kT\) is the thermal energy, \(\rho_{s}\) is the source distribution, and \(\gamma\equiv l_{D}^{-1}\) is the inverse Debye length. The first equation above is the full P-B equation. The linearized approximation is written out in the second line. Solutions to these equations provide mean-field approximations to the potentials of physical systems.
It is possible to derive the P-B equation in a rigorous manner which clarifies the physical approximations and assumptions required to arrive at this expression. In the weak potential limit, when the linearized form applies, it has been shown that the equation is exact, while errors result at each non-linear order. However, recent simulations have shown that a partial cancellation of two neglected effects, counterion correlations and finite ion sizes, can result in approximate agreement between the P-B solutions and experiments. Scattering experiments have also shown that the full non-linear solutions are in reasonable agreement with observed charge distributions when only a single screening ion species is present, though some extra divalent condensation was observed when both monovalent and divalent ion species were present. These results justify use of the non-linear equation, but the solutions should be considered qualitatively accurate only outside the innermost, Stern layer surrounding highly charged distributions.
Solutions to the inhomogeneous, linearized P-B equation will be needed throughout this paper. The Poisson sum rule provides a convenient method to determine the potential of a lattice of point charges, from which one can obtain the potentials of more general distributions through superposition.
\[\sum_{\mathbf{R}}f(\mathbf{r}-\mathbf{R})=\frac{1}{V}\int\sum_{\mathbf{G}}f( \mathbf{r}-\mathbf{R})\exp(i\mathbf{G}\cdot\mathbf{R})d^{n}R. \tag{2}\]
Here, \(n\) is the dimension of the lattice, \(V\) is the \(n\)-dimensional volume per unit cell of the direct lattice, the vectors \(\mathbf{R}\) are the direct lattice vectors of the distribution, and the vectors \(\mathbf{G}\) are the reciprocal lattice vectors of the distribution. Shifting the origin of integration in Eq. to the projected position of the observation point \(\mathbf{r}\) immediately gives formal integral representations for the Fourier coefficients of the potential.
To obtain the potential from a lattice of Coulomb point charges which are linearly screened, one may plug in the Yukawa potential function \(f=q\exp[-\gamma r]/r\), which is the solution to the linearized P-B equation with point charge source \(q\).
\[\Phi_{1,Y} = 2\lambda\sum_{\mathbf{G}}K_{0}((|\mathbf{G}|^{2}+\gamma^{2})^{1 /2}d)\exp[i\mathbf{G}\cdot\mathbf{r}], \tag{3}\] \[\Phi_{2,Y} = 2\pi\sigma\sum_{\mathbf{G}}\frac{\exp[i\mathbf{G}\cdot\mathbf{r} -(|\mathbf{G}|^{2}+\gamma^{2})^{1/2}d]}{(|\mathbf{G}|^{2}+\gamma^{2})^{1/2}},\] \[\Phi_{3,Y} = 4\pi\rho\sum_{\mathbf{G}}\frac{1}{|\mathbf{G}|^{2}+\gamma^{2}} \exp[i\mathbf{G}\cdot\mathbf{r}], \tag{5}\]
For large arguments, the modified Bessel function \(K_{0}\) above may be expanded as
\[K_{0}(z)\sim\sqrt{\frac{\pi}{2z}}e^{-z}\sum_{j=0}^{\infty}\frac{(-1)^{j}\prod_ {k=0}^{j}(2k+1)^{2}}{j!(8z)^{j}}. \tag{6}\]
It follows that each component will be exponentially damped with both the frequency and the distance from the distribution for one and two-dimensional systems. To relate these expressions to their unscreened analogs, one need only replace \((|\mathbf{G}|^{2}+\gamma^{2})^{1/2}\) by \(|\mathbf{G}|\) in the non-zero frequency components and then also replace the zero frequency component by the appropriate continuous charge distribution expression. This is given by \(-2\lambda\log r\) for a linear distribution, for example.
## III Source charge modulation
Consider now the charge modulated system depicted in Two model polyelectrolytes are shown, each of which consists of alternating point particles of charge \(+2\) and \(-1\) fixed on a rigid backbone. The period of the modulation is taken to be \(1\). Note that each polyelectrolyte has a net charge of \(+1\) per period. If the two polyelectrolytes are shifted with respect to one another by half of a period, as shown in the figure, then they will attract one another when their separation is small, demonstrating that mean-field like-charge attraction is possible.
One can approximate the interaction energy of this system in the Debye-Huckel limit by superposing the potential for an infinite line of \(-1\) charges with that for a line of \(+2\) charges, each obtained by plugging into Eq., and truncating the series after two terms.
\[E(d)\approx 2\text{K}_{0}[\gamma d]-36\text{K}_{0}[\sqrt{(2\pi)^{2}+\gamma^{2}}d], \tag{7}\]
A plot of this function is shown in for \(\gamma=3.0\). Attraction is indeed observed at small separations while repulsion is observed at large separations. This form of attraction is not salt mediated, but as \(\gamma\) is increased, the repulsive effect at large distances is reduced. This trivial example demonstrates that like-charged molecules can be attracted to one another even when the electrolyte is considered at the mean-field level. Note, however, that if the model polyelectrolytes are arranged in a mirror-symmetric orientation, as was explicitly assumed in Ref., then the coefficient of the second term in Eq. would also be positive. This is consistent with the conclusion that identical molecules cannot be attracted to one another, within the P-B formalism, when they are arranged in a mirror-symmetric orientation.
We now turn to a consideration of charge modulation-induced condensation. As mentioned above, this was previously studied in Refs. and. These previous works considered solutions to the full non-linear P-B equation for systems containing a planar source distribution and only a single species of screening ion. Solutions were obtained up to second order in the parameters \(\epsilon(\vec{q})/\bar{\sigma}\), for all \(\vec{q}\neq 0\). Here, \(\bar{\sigma}\) was the average charge density of the planar source distribution and the \(\epsilon(\vec{q})\) were the amplitudes of the oscillatory components of the source distribution's Fourier decomposition. Thus, these solutions were valid in the limit of relatively weak modulation relative to net surface charge. Interestingly, these solutions indicated that each non-zero Fourier component independently led to an increase of screening charge condensation at second order in perturbation theory. This universal result was also shown to hold for long wavelength modulations. Here it is shown that a similar conclusion also holds true in the general weak potential limit.
For simplicity, we consider first a system which contains only one species of counterion.
\[\nabla^{2}\phi = -4\pi qn_{0}e^{-q\phi/kT}-4\pi\rho_{s}\] \[\approx -4\pi qn_{0}(1-\frac{q}{kT}\phi+\frac{1}{2}(\frac{q}{kT})^{2}\phi ^{2})-4\pi\rho_{s}.\]
Here, \(\rho_{s}\) is a perturbing planar charge distribution. The first term in the above Boltzmann factor expansion corresponds to the uniform background charge density. This does not effect an electric field and so may be dropped for now. We assume that the potential can be expanded in a series as \(\phi=\phi_{1}+\phi_{2}+...\), where the term \(\phi_{k}\) is of order \(\rho_{s}^{k}\).
\[\nabla^{2}\phi_{1} = \frac{4\pi q^{2}n_{0}}{kT}\phi_{1}-4\pi\rho_{s}\] \[\nabla^{2}\phi_{2} = \frac{4\pi q^{2}n_{0}}{kT}\phi_{2}-\frac{2\pi q^{3}n_{0}}{(kT)^{2 }}\phi_{1}^{2} \tag{9}\] \[...\]
The solution to the first equation above may be obtained from our earlier expressions if the distribution \(\rho_{s}\) is expanded in a Fourier series as
\[\rho_{s}=\frac{1}{2\pi}\sum_{\bf G}A_{\bf G}\exp[i{\bf G}\cdot{\bf r}]. \tag{10}\]
Note that while extra, oscillatory condensation results at this first order, it averages to zero when integrated over the modulation directions. The source term for \(\phi_{2}\) is proportional to the square of the first order potential, however, and a non-zero condensation will result at this order. To obtain the net condensation, one need only consider the average charge density above the surface.
\[(\frac{\partial^{2}}{\partial z^{2}}-\gamma^{2})\overline{\phi_{2}} = -\frac{\gamma^{2}}{2}(\frac{q}{kT})\sum_{\bf G}|A_{\bf G}|^{2}\frac {e^{-2(G^{2}+\gamma^{2})^{1/2}z}}{G^{2}+\gamma^{2}},\]
The solution is
\[\overline{\phi_{2}}=\gamma(\frac{q}{kT})\sum_{\bf G}\frac{|A_{ \bf G}|^{2}}{\sqrt{G^{2}+\gamma^{2}}(4G^{2}+3\gamma^{2})}\] \[\times \{\exp[-\gamma z]-\frac{\gamma}{2\sqrt{G^{2}+\gamma^{2}}}\exp[-2 \sqrt{G^{2}+\gamma^{2}}z]\}.\]
Portions of two long, identical polyelectrolytes are shown. At large distances they repel but at short distances they attract.
A plot of the approximate interaction energy versus separation for the model polyelectrolytes.
A plot demonstrating the resulting \(z\) dependence of a typical term of the second order average charge density is shown in The total averaged second order counterion charge density at the surface is given by the following sum over modulation components:
\[\overline{\rho_{2}}(z=0) = -\frac{1}{4\pi}\frac{\partial^{2}}{\partial z^{2}}\overline{\phi_ {2}}\] \[= qn_{0}(\frac{q}{kT})^{2}\sum_{\mathbf{G}}\frac{|A_{\mathbf{G}}|^{2 }}{4G^{2}+3\gamma^{2}}(2-\frac{\gamma}{\sqrt{G^{2}+\gamma^{2}}}).\]
This demonstrates that indeed there is extra condensation proportional to the square of the amplitude for each mode of the modulation, consistent with the long wavelength and small modulation results of Refs. and. The present analysis extends the conclusion that universal modulation-induced condensation is expected at second order to all relative values and wavelengths of modulation. In particular, the result applies to systems which have little or no net charge but large modulations. In systems such as these, the modulation-induced condensation could easily result in overcharging. This demonstrates that mean-field overcharging is possible in certain limits.
A single component electrolyte is somewhat non-physical in that the solution contains charges of only one sign and thus cannot be net neutral. Accordingly, it is of interest to consider systems which contain multiple species of ions. This would be difficult when considering solutions to the full P-B equation but the present solution technique is readily applied to such systems.
\[\nabla^{2}\phi\approx\gamma^{2}\phi-\zeta\phi^{2}-4\pi\rho_{s}. \tag{14}\]
Here, the coefficients \(\gamma^{2}\) and \(\zeta\) are now given by sums over ions species as
\[\gamma^{2} = \frac{4\pi}{kT}\sum_{i}q_{i}^{2}n_{i} \tag{15}\] \[\zeta = \frac{2\pi}{(kT)^{2}}\sum_{i}q_{i}^{3}n_{i}. \tag{16}\]
Charge neutrality requires \(\sum q_{i}n_{i}=0\), if the charge of the polyelectrolyte itself is neglected, for simplicity.
\[\overline{\rho}(z=0)=\rho_{0}\] \[+\frac{\zeta}{2\pi}\sum_{\mathbf{G}\neq 0}\frac{|A_{\mathbf{G}}|^{ 2}}{4G^{2}+3\gamma^{2}}(2-\frac{\gamma}{\sqrt{\gamma^{2}+G^{2}}}), \tag{17}\]
It is interesting to consider some particular examples. Consider first a net neutral electrolyte system containing screening charges of both signs and equal magnitudes. For this system, the quadratic terms in the expansion of the Boltzmann factors will cancel and no second order potential or condensation will result. This implies that any extra condensation would have to occur at higher order for such systems and the resulting effect should be much weaker. The quadratic term in the expansion is retained, however, for a net neutral system which contains charges of different sign and different valence.
\[\zeta=\frac{2\pi q_{1}n_{0}}{(kT)^{2}}(q_{1}^{2}-q_{2}^{2}). \tag{18}\]
The result is that extra charge condensation still occurs at the lowest non-linear order for a net neutral system, provided the free charges have different valences. Note that the net sign of the resulting modulation-induced condensation does not depend on \(A_{0}\). Extra condensation occurs for both species of ion, and the sign of the net condensation depends only on which species dominates. This, in turn, is determined by the sign of the quadratic term in the expansion of the P-B equation; if \(\zeta\) is large and positive, the resulting condensation will be large and positive, and vice versa. This obviously holds for any distribution of counterion species. Further comments on this point appear in the discussion section of the paper.
One and three-dimensional charge modulated distributions may be analyzed in a similar manner. For a three-dimensional lattice, the concept of condensation is not well-defined. In the one-dimensional case, however, an analogous condensation to that discussed above for surface distributions also occurs. Each modulation component leads to additional condensation at second order, the sign of which depends only on the sign of the quadratic term in the expansion of the P-B equation. A consideration of similar surface modulation effects is contained in the following section.
## IV Surface modulation
In the above, we saw how charge modulation can allow for oscillatory potential components that lead to a net charge condensation at non-linear orders. A modulated surface can also allow for the introduction of potential modulation and similar condensation effects must naturally result. It is worthwhile to consider some examples, and two are presented here which are susceptible to perturbation expansions. These examples are of interest, not because they accurately model physical systems, but because they demonstrate the mechanisms involved for different boundary conditions and geometries.
A constant potential boundary condition problem will be considered first, where the surface takes the form
\[Z(x,y)=\beta\sum_{\mathbf{G}\neq 0}A_{\mathbf{G}}e^{i\mathbf{G}\cdot\mathbf{r}}. \tag{19}\]
Here \(\beta\) is a small constant and the \(A_{\mathbf{G}}\) are assumed to be \(O(\beta^{0})\).
\[\begin{cases}(\nabla^{2}-\gamma^{2})\phi_{1}=0&\text{if }z\geq Z(x,y),\\ \phi_{1}=1&\text{if }z=Z(x,y),\\ \phi_{1}\to 0&\text{as }z\rightarrow\infty,\end{cases} \tag{20}\]
and the term \(\phi_{1}\) is assumed expandable as
\[\phi_{1}=\sum_{k=0}^{\infty}\phi_{1,k}\beta^{k}. \tag{21}\]
Each of the \(\phi_{1,k}\) satisfy the Debye-Huckel equation individually and may be determined iteratively by expanding the boundary condition as
\[1 = \phi_{1}(z=Z) \tag{22}\] \[= \sum_{n=0}^{\infty}(\sum_{\mathbf{G}}A_{\mathbf{G}}e^{i\mathbf{G} \cdot\mathbf{r}})^{n}\frac{\beta^{n}}{n!}\frac{\partial^{n}\phi_{1}}{\partial z ^{n}}|_{z=0}.\]
Equating coefficients at order \(\beta^{k}\) on both sides of Eq. generates the boundary conditions for the \(\phi_{1,k}\) at the simple surface \(z=0\).
\[\phi_{1,0} = \exp[-\gamma z] \tag{23}\] \[\phi_{1,1} = \sum_{\mathbf{G}}\gamma A_{\mathbf{G}}e^{i\mathbf{G}\cdot\mathbf{ r}}\exp[-\sqrt{G^{2}+\gamma^{2}}z]\] \[\overline{\phi_{1,2}} = \sum_{\mathbf{G}}|A_{\mathbf{G}}|^{2}(\gamma\sqrt{G^{2}+\gamma^{2 }}-\frac{\gamma^{2}}{2!})\exp[-\gamma z]. \tag{25}\]
The last term has once again been averaged over the modulation directions to simplify the calculation. As in the charge modulation discussion, each modulation component is observed to independently result in extra second order condensation. Here, however, the surface modulations generate the extra charge condensation at linear order in the P-B theory. Since the effect takes place at linear order, the sign of the condensation does not depend on the sign and distribution of the counterions, but instead is always opposite to that of the surface potential.
The first order solution above can easily be plugged back into the non-linear P-B equation to obtain the next order term in the solution. Although this won't be done here, we note that the non-linear terms in the P-B expansion can result in extra condensation at the same order in the potential and in \(\beta\) as that obtained above. These terms, however, are sensitive to the sign of \(\zeta\). In particular, if \(\zeta=0\), the linearized P-B solution will be correct to lowest order. Alternatively, the higher order solutions can result in charging of the wrong sign, or negative screening, if \(\zeta\) is chosen appropriately.
We now consider a nearly cylindrical geometry with boundary conditions which are more applicable to biological molecules.
\[R=1+\beta\cos(Gz). \tag{26}\]
Outside the dielectric near-cylinder the potential is assumed to satisfy the linearized P-B equation. The following system of equations specifies the unique solution for the potential:
\[\begin{cases}(\nabla^{2}-\gamma^{2})\phi=0&\text{if }r\geq R(z),\\ \nabla^{2}\phi=-\frac{4\pi}{\epsilon_{i}}\rho&\text{if }r<R(z),\\ \phi\to 0&\text{as }r\rightarrow\infty,\\ \phi_{i}=\phi_{o}&\text{at }r=R(z),\\ \mathbf{D}_{i}\cdot\hat{\mathbf{n}}=\mathbf{D}_{o}\cdot\hat{\mathbf{n}}&\text {at }r=R(z).\end{cases} \tag{27}\]
Here, the subscripts \(i\) and \(o\) stand for inside and outside, respectively, \(\epsilon\) is the dielectric coefficient, \(\mathbf{D}\) is the electric displacement, \(\hat{\mathbf{n}}\) is the outward normal to the near cylinder's surface, and \(\rho=\delta(r)/2\pi r\).
The boundary conditions may be expressed at the surface \(r=1\) in a similar manner to that presented in Eq..
\[\epsilon_{i}\sum_{k=0}^{\infty}\frac{(\beta\cos(Gz))^{k}}{k!} \frac{\partial^{k}\nabla\phi_{i}\cdot\hat{\mathbf{n}}}{\partial r^{k}}\mid_{r=1} \tag{28}\] \[= \epsilon_{o}\sum_{k=0}^{\infty}\frac{(\beta\cos(Gz))^{k}}{k!} \frac{\partial^{k}\nabla\phi_{o}\cdot\hat{\mathbf{n}}}{\partial r^{k}}\mid_{r=1},\]
A plot of the one component fluid’s second order charge condensation versus \(z\), the distance from the plane. Here \(G=2\gamma=2/l_{D}\). The condensation is particularly localized and then screened at larger distances. The localization increases exponentially with the magnitude of \(G\). The total integrated charge from \(z=0\) to \(\infty\) is zero for each component, as expected.
where the normal is given explicitly by
\[\hat{\bf n}=\frac{\bf r}{\sqrt{1+\beta^{2}G^{2}\sin^{2}(Gz)}}+\frac{ \beta G\sin(Gz){\bf z}}{\sqrt{1+\beta^{2}G^{2}\sin^{2}(Gz)}}. \tag{29}\]
Up to second order, the boundary conditions generate terms of the form
\[\phi_{i} = A-\frac{2\lambda}{\epsilon_{i}}\log r+\beta BI_{0}(Gr)\cos(Gz) \tag{30}\] \[+\beta^{2}\bigg{\{}C+DI_{0}(2Gr)\cos(2Gz)\bigg{\}},\]
and
\[\phi_{o} = EK_{0}(\gamma r)+\beta FK_{0}((G^{2}+\gamma^{2})^{1/2}r)\cos(Gz)\] \[+\beta^{2}\bigg{\{}GK_{0}(\gamma r)+HK_{0}((4G^{2}+\gamma^{2})^{1 /2}r)\cos(2Gz)\bigg{\}},\]
The boundary condition equations further enable one to obtain a system of linear equations which specify the unique solution for the coefficients \(A\) through \(H\). The algebra is most easily carried out in a software package such as _Mathematica_. The resulting expressions are easy to obtain but are too lengthy to report here. Once the coefficients for the potential have been determined, one can read off the effective charge of the near-cylinder, defined here to be \((E+\beta^{2}G)/2\). This is proportional to the coefficient of \(K_{0}(\gamma r)\) above, a natural choice given that the potential due a screened line of charge is given by \(2\lambda K_{0}(\gamma r)\), from Eq..
We now examine how the resulting effective charge varies with the wavelength of the surface modulation. At long wavelength the surface curvature is low and the cylinder has a slowly varying local radius. The effective charge at long wavelength is thus expected to be that which one would obtain through an appropriate average over radii of the effective charge for an unmodulated cylinder.
\[\bar{\lambda}=\int_{0}^{1}\frac{du}{\gamma\epsilon_{o}(1+\beta\cos 2\pi u)K_ {1}[\gamma(1+\beta\cos 2\pi u)]}. \tag{32}\]
This integral is over a convex function, which, by Jenson's inequality, implies that the modulation will result in an increase rather than a decrease in the effective charge of the cylinder. A plot of the effective charge, valid to second order in \(\beta\), versus wavelength is shown in The effective charge does approach the radii-averaged unmodulated value at long wavelengths, as expected. Note also that the effective charge decreases as the modulation wavelength decreases, and below wavelength values of \(l_{0}\approx 2\), the resulting condensation works to reduce the net effective charge of the cylinder. However, the perturbation expansion converges most quickly when the modulation wavelength is large. Indeed, the expansion will break down when \(\beta G>1\) owing to the fact that the \(\beta\) power series of \(\hat{\bf n}\) in Eq. will diverge in this limit. Thus, convergence concerns indicate that the plot should not be considered quantitatively accurate at small values of \(l_{0}\). A decrease in the effective charge is also observed at short modulation wavelengths for very small \(\beta\), however, where the second order solution is likely to be very accurate. It is thus reasonable to expect the observed decrease in effective charge at small modulation wavelengths to be qualitatively accurate at larger \(\beta\) values as well.
These two examples demonstrate that the sign of surface modulation-induced condensation can be sensitive to both counterion valence and modulation wavelength. A modulation wavelength dependence was observed only in the second example but it also holds true for some other systems, linear geometries with constant potential boundary conditions providing one example. We have focused above on the net condensation for these systems. It is worth noting, however, that the oscillatory components of these potentials can introduce attractive terms to the interaction between two shifted, identical molecules, just as in the charge modulated example presented earlier.
## V Discussion
In this paper, a collection of simple examples have been presented which demonstrate how inherent modulations can alter the potentials within electrolyte solutions. It was first shown that charge modulation can result in like-charge attraction between identical molecules within the P-B formalism.
A plot of the effective charge density versus modulation wavelength for the near-cylinder. The thin line is the unmodulated charge density, the dashed line is the modulated effective charge valid to second order in \(\beta\), and the thick line is the radii-averaged value discussed in the text. The parameter values were set to \(\beta=0.1\), \(\epsilon_{i}=1\), \(\epsilon_{0}=5\), and \(\gamma=5\).
Next, it was demonstrated that charge modulation results in extra screening charge condensation when the P-B equation is expanded to second order in the potential. This result, when taken together with the results of Ref., implies that charge modulation-induced condensation universally occurs within the P-B formalism when the potential is expanded to second order in the modulation amplitudes. In particular, the result applies to systems which have large modulations but little net charge. In systems such as these, charge modulations can easily result in overcharging. It was further shown that screening charges of higher valence allow for a reduction in the net cancellation of modulation-induced condensation which occurs in neutral, univalent electrolyte systems. This is consistent with the fact that overcharging has only been observed in systems which contain multivalent counterions. Similar results were shown to apply for surface modulated systems and related effects can be expected whenever non-linearities appear in either the differential equation or the boundary conditions which determine the potential.
We conclude by elaborating upon how one can test whether charge modulation significantly contributes to any experimental observation of overcharging. In the above we noted that mobile charges of equal valence but opposite sign are expected to condense in equal numbers onto a substrate due to inherent charge modulations. This is not the case for fluctuation-induced condensation, however, where charges are only expected to condense onto substrates of the opposite sign. This suggests that an overcharging observation can be identified as modulation-induced if the overcharging is negated through the introduction of equal quantities of multivalent counterions of each sign into the system. We note that this test would not apply to systems where small surface modulations cause condensation in the weak potential limit, where the linearized P-B equation applies. It would, however, also apply to systems where surface modulations bring about condensation at higher orders in the P-B expansion.
|
10.48550/arXiv.0909.3824
|
Modulation effects within the mean-field theory of electrolyte solutions
|
Jonathan Landy
| 6,699
|
10.48550_arXiv.0903.4948
|
## Introduction
If many detailed kinetic models are available for the oxidation of mixtures representative of gasolines, they are much less numerous in the case of diesel fuels because of their more complex composition. The constituents of diesel fuel contain from 10 to 20 carbon atoms and include about 30% (by mass) of alkanes, the remaining part being mainly alkylcyclohexanes (24%), alkyldecalines (15%), alkylbenzenes (10%) and polycyclic naphtenoaromatic compounds. If the oxidation of alkanes has been extensively studied, the abundance of models diminishes considerably when other families of components are considered and very few models exist for substituted cycloalkanes and aromatics compounds, except for toluene. The oxidation of alkylbenzenes with alkyl side-chains from C\({}_{2}\) to C\({}_{4}\) has been studied in a flow reactor in Princeton and that of n-propylbenzene in a jet-stirred reactor at Orleans. The autoignition of n-butylbenzene has also been investigated in a rapid compression machine in Lille. While premixed flames of ethylbenzene and non-premixed methane flames doped with ethylbenzene and isomers of propylbenzene and butylbenzene have been studied, no measurement in a premixed flame containing butylbenzene has yet been reported.
The first purpose of the present paper is to experimentally investigate the structure of a premixed laminar methane flame doped with n-butylbenzene. The use of a methane flame will allow us to be more representative of the combustion mixtures containing large hydrocarbons, such as those present in a diesel fuel, than hydrogen or unsaturated C\({}_{2}\) flames. Large alkanes decompose to methyl radicals and the use of a methane flame will capture the involved chemistry. This study will be performed using a lean flame for the chemistry to be better representative of that occurring in engines controlled via Homogenous Charge Combustion Ignition (HCCI), which are under development. The second objective is to use these results in order to develop a new mechanism for the oxidation of n-butylbenzene based on our experiencein modeling the reactions of both alkanes and light aromatic compounds (benzene and toluene).
## Experimental Procedure
The experiments were performed using an apparatus developed in our laboratory to study temperature and stable species profiles in a laminar premixed flat flame at low pressure and recently used in the case of rich methane flames doped by light unsaturated soot precursors.
FIGURE 1
The body of the flat flame matrix burner, provided by McKenna Products, was made of stainless steel, with an outer diameter of 120 mm and a height of 60 mm (without gas/water connectors). This burner was built with a bronze disk (95% copper, 5% tin). The porous plate (60 mm diameter) used to assist the flame stabilization was water cooled (water temperature: 333 K) with a cooling coil sintered into the plate. The burner could be operated with an annular co-flow of argon to favor the stabilization of the flame.
This horizontal burner was housed in a water-cooled vacuum chamber evacuated by two primary pumps and maintained at 6.7 kPa by a regulation valve. This chamber was equipped of four quartz windows for optical access, a pressure transducer (MKS 0-100 Torr), a microprobe for samples taking and a thermocouple for temperature measurements. The burner could be vertically translated, while the housing and its equipments were kept fixed. A sighting telescope measured the position of the burner relative to the probe or the thermocouple with an accuracy of 0.01 mm. The flame was lit using an electrical discharge.
Gas flow rates were regulated by RDM 280 Alphagaz and Bronkhorst (El-Flow) mass flow regulators. Methane (99.95 % pure) was supplied by Alphagaz - Air Liquide. Oxygen (99.5%pure) and argon (99.995% pure) were supplied by Messer. Liquid butylbenzene, supplied by Alfa Aesar (purity 99%), was contained in a glass vessel pressurized with argon. After each load of the vessel, argon bubbling and vacuum pumping were performed in order to remove oxygen traces dissolved in the liquid hydrocarbon fuel. The liquid reactant flow rate was controlled by using a liquid mass flow controller, mixed to the carrier gas and then evaporated by passing through a single pass heat exchanger, the temperature of which was set above the boiling point of the mixture. Carrier gas flow rate was controlled by a gas mass flow controller located before the mixing chamber.
Temperature profiles were obtained using a PtRh (6%)-PtRh (30%) type B thermocouple (diameter 200 \(\upmu\)m). The thermocouple wire was supported by an arm and crossed the flame horizontally to avoid conduction heat losses. The junction was located at the centre of the burner. The thermocouple was coated with an inert layer of BeO-Y\({}_{2}\)O\({}_{3}\) to prevent catalytic effects. The ceramic layer was obtained by dipping the thermocouple in a hot solution of Y\({}_{2}\)(CO\({}_{3}\))\({}_{3}\) (93% mass.) and BeO (7% mass.) followed by drying in a Meker burner flame. This process was repeated about ten times until the whole metal was covered. Radiative heat losses are corrected using the electric compensation method.
The sampling probe was constructed of quartz with a hole of about 50 \(\upmu\)m diameter (d\({}_{i}\)). The probe was finished by a small cone with an angle to the vertical of about 20\({}^{\circ}\). For temperature measurements in the flames perturbed by the probe, the distance between the junction of the thermocouple and the end of the probe was taken equal to two times d\({}_{i}\), i.e. to about 100 \(\upmu\)m. The sampling quartz probe was directly connected to a heated transfer line made with a passivated stainless steel tube and heated at 423 K. This line was itself connected to a heated pressure transducer (MKS 0-100 Torr) and through heated valves to a turbomolecular pump, a pyrex line and a heated 10 ml stainless steel loop located inside a gas chromatograph. The pressure in the transfer line was always below 1.3 kPa (10 Torr) so that the pressure drop between the flame and the inlet of the probe ensured rapid quenching of reactions. The on-line connection to a chromatograph was made via a heated transfer line in order to analyse compounds above C\({}_{6}\).
The analyses could be then performed according to two methods:
Gas samples of compounds with a sufficient vapour pressure were directed through the pyrex line towards a volume which was previously evacuated by the turbo molecular pump down to 10\({}^{-7}\) kPa and which was then filled up to a pressure of 1.3 kPa and collected in a pyrex loop after compression by a factor 5 by using a column with rising mercury. Pressures in the pyrex line were measured by a MKS 0-10 Torr pressure transducer before compression and by a MKS 0-1000 Torr pressure transducer after compression. A chromatograph with a Carbosphere packed column and helium or argon as carrier gas was used to analyse O\({}_{2}\), H\({}_{2}\), CO and CO\({}_{2}\) by thermal conductivity detection and CH\({}_{4}\), C\({}_{2}\)H\({}_{2}\), C\({}_{2}\)H\({}_{4}\), C\({}_{2}\)H\({}_{6}\) by flame ionisation detection (FID). Water was detected by TCD but not quantitatively analyzed. Calibrations were performed by analysing a range of samples containing known pressures of each pure compound to quantify. Mole fractions were derived from the known total pressure of the gas included in the pyrex loop. This method of sampling could allow us to analyze heavier hydrocarbons from allene to toluene using a chromatograph with a Haysep packed column with FID and nitrogen as gas carrier gas as in our previous study, but it was found more accurate and faster to use the second method described hereafter. For the compounds for which both methods could be used, a very good agreement was obtained between both measurements.
Gas samples were directed towards the loop located inside the on-line gas chromatograph and previously evacuated by the turbo molecular pump down to 10\({}^{-7}\) kPa. The loop was then filled up to a pressure of 1.3 kPa and contents were immediately analysed using FID and helium as gas carrier gas by using a capillary column (a HP-Plot Q or a HP-1 column). This method of sampling was used to analyze methane, all the hydrocarbons from C3 and the oxygenated products other than carbon oxides.
The identification of these compounds was performed using a gas chromatograph with mass spectrometry detection (GC-MS) fitted with the same column or by comparison of retention times when injecting the product alone in gas phase. Four C3 species (propyne, allene, propene and propane), nine C4 species (diacetylene, vinylacetylene, 1-butyne, 2-butyne, 1,3-butadiene, 1,3-butadiene, 1-butene, iso-butene and butane) and four C5 species (cyclopentene, cyclopentadiene, 1,3-pentadiene and isoprene (2-methyl-1,3-butadiene) were observed. The butenes, mainly 1-butene and iso-butene, were not well separated. In this lean flame, it was also possible to quantify seven oxygenated species (methanol, ketene, acetaldehyde, ethanol, acrolein, propanal and acetone). For methane, C3 species, 1,3-butadiene and butenes, calibrations were performed by analyzing each pure compound for a range of known pressures in the transfer line and mole fractions were derived from the known total pressure. For other species, we have used the fact that the response of an FID detector is linked to the Effective Carbon Number (ECN) in the molecule. The detector response coefficient (DRCGP-i) of each compound \(i\) can be calculated from the response of a molecule of close structure (DRCGP-mod), the ECN of this molecule (ECNmod) and that of the compound \(i\) (ECNi):
\[\mathrm{DRCGP\_i}=\mathrm{DRCGP\_mod}\times(\mathrm{ECN}_{\mathrm{i}}\,/\, \mathrm{ECN}_{\mathrm{mod}})\]
ECN is obtained as the sum of the contributions of all the atoms of carbon and oxygen included in the molecules. These contributions have been taken to 1 for aliphatic and aromatic atoms of carbon, 0.95 for olefinic ones, 1.3 for acetylenic ones and 0 for carbonyl ones (C=O) and to -1, -0.6 and -0.25 for atoms of oxygen included in ethers, in primary alcohols and in tertiary alcohols, respectively.
FIGURE 2
The identification of these compounds was performed using a gas chromatograph with GC-MS detection fitted with the same column. Two C\({}_{6}\) cyclic species (methylcyclopentene and methylcyclopentadiene), twelve monocyclic aromatic compounds (benzene, toluene, xylenes, phenylacetylene, ethylbenzene, styrene, allylbenzene, n-propylbenzene, cumene (iso-propyl-benzene), methylstyrenes, butenylbenzenes (compounds with C\({}_{10}\)H\({}_{12}\) as chemical formula) and butadienylbenzene) and 5 bicyclic aromatic compounds (indene, indane, methylindene, dihydronaphthalene and naphthalene) were identified as the most probable products according to their mass spectrum. Several peaks have been observed for xylenes, methylstyrenes and compounds with C\({}_{10}\)H\({}_{10}\) as chemical formula, which have been identified as butadienylbenzenes, methylindene and dihydronaphthalene. In this last case, the separate quantification of the different isomers was not possible. Five oxygenated aromatic compounds (phenol, benzaldehyde, benzylalcohol, anisole (methylphenylether) and benzofuran) were also quantified. For those compounds having a low vapour pressure, calibrations were performed in two steps. First, a calibration was made for gas-phase cyclohexane, a cyclic compound with a much larger vapour pressure, by analysing this species for a range of known pressures introduced in the transfer line.
\(\text{P}_{\text{GP-CHX}}\): the cyclohexane partial pressure in the transfer line.
The detector response coefficient (DRC\({}_{\text{GP-P}}\)) of each gas-phase aromatic compound of interest, P, was derived from its relative detector response coefficient (DRC\({}_{\text{L-P}}\)) compared to that of cyclohexane (DRC\({}_{\text{L-CHX}}\)) obtained by injecting an acetone solution containing known mass of both cyclohexane (m\({}_{\text{CHX}}\)) and the P (m\({}_{\text{P}}\)) in liquid phase:
\[\text{DRC}_{\text{GP-P}}=(\text{DRC}_{\text{L-P}}/\text{DRC}_{\text{L-CHX}}) \times\text{DRC}_{\text{GP-CHX}}\times(\text{M}_{\text{P}}/\text{M}_{\text{CHX}})\]
with:
\(\text{DRC}_{\text{L-CHX}}\) = A\({}_{\text{L-CHX}}\) / m\({}_{\text{CHX}}\), A\({}_{\text{L-CHX}}\) is the area of the cyclohexane peak in the case of the liquid injection,
\(\text{DRC}_{\text{L-P}}\) = A\({}_{\text{L-P}}/\) m\({}_{\text{P}}\), A\({}_{\text{L-P}}\) is the area of the P peak in the case of the liquid injection,
\(\text{M}_{\text{CHX}}\) the molar mass of cyclohexane,
Mp the molar mass of P.
This method of calibration has been used for all major aromatic compounds, except from methylstyrenes, butenylbenzenes, compounds with C\({}_{10}\)H\({}_{10}\) as chemical formula, phenol, benzylalcohol and benzofuran for which the ECN method has been used.
FIGURE 3
Calculated uncertainties on the species quantifications were about \(\pm\) 5% for major hydrocarbon compounds and \(\pm\) 10% for oxygen, hydrogen, carbon oxides and the minor hydrocarbon products. The detection limit of the FID is about 2 ppm.
## Experimental Results
A laminar premixed flat flame has been stabilized on the burner at 6.7 kPa (50 Torr) with a gas flow rate of 5.44 l/min corresponding to a gas velocity at the burner of 49.2 cm/s at 333 K and with mixtures containing 7.1% (molar) of methane, 36.8% of oxygen and 0.96% of n-butylbenzene corresponding to an equivalence ratio of 0.74.
Without the probe, the lowest temperatures measured the closest to the burner (0.47 mm above) were around 1095 K. Due to the thinness of this lean flame and the size of the thermocouple, it was not possible to measure lower temperatures. The highest temperatures were reached from 5 mm above the burner and were around 1960 K.
FIGURE 4
FIGURE 5
Figures 6 and 7 present the profiles of oxygen, hydrogen, water and of the main C0-C2 species involved in the combustion of methane versus the height above the burner. The mole fraction of water has been obtained from a material balance on the other major species. In this lean flame, the profiles of carbon monoxide and hydrogen display a marked maximum at 2 mm height and the major final products are for a large extent carbon dioxide and water. Ethylene is the most abundant C2 species and is produced first. It reaches its maximum concentration close to the burner around 1 mm. The profiles of ethane and of acetylene peak around 1.1 mm and 1.2 mm, respectively.
FIGURE 6 AND 7
The peak mole fractions observed for propene and propane are more than ten times larger than that of allene and propyne.
1-butyne, 1-2 butadiene, butenes and butane are produced first and reach their maximum concentration close to the burner, around 1.1 mm. The profiles of diacetylene, vinylacetylene, 2-butyne and 1,3-butadiene peak around 1.5 mm. The most abundant C\({}_{4}\) compounds are vinylacetylene, 1,3-butadiene and butanes with peak concentrations between 60 and 80 ppm. Butynes and 1,2-butadiene are present in very low amounts with peak concentrations well below 10 ppm.
FIGURE 9
Cyclopentene, 1,3-pentadiene and isoprene reach their maxima first around 1 mm above the burner. The profiles of methylcyclopentadiene peak further around 1.1 and the maxima of that of cyclopentadiene, the most abundant species of this series with a peak concentration of 50 ppm, and methylclopentene are the last ones around 1.5 mm.
FIGURE 10
Methanol, ketene, acetaldehyde and ethanol, which are intermediate products of the combustion of methane, as well as propanal, are produced very early and reach their maximum concentration close to the burner, around 0.9 mm. The profiles of acrolein and acetone peak around 1.2 mm. The most abundant of these species are methanol and acetaldehyde with peak concentrations above 150 ppm.
FIGURE 11
Figures 12 and 13 present the profiles of monocyclic aromatic products.
The most abundant of these species are benzene with a peak concentration of 800 ppm, toluene with a peak concentration of 800 ppm and styrene with a peak concentration of 1000 ppm)and to a lesser extent allylbenzene and butenylbenzenes with peaks concentration about 100 ppm. Xylenes, phenylacetylene and n-propylbenzene are present in low amounts with peak concentrations below 15 ppm.
FIGURES 12 AND 13
The maxima of the profiles of indene and the isomers of C\({}_{10}\)H\({}_{10}\) are observed first around 1 mm above the burner, followed by that of indane around 1.3 mm and finally that of naphthalene around 1.8mm. The most abundant of these species is indene with a peak concentration of 50 ppm. Indane and naphthalene are present in very low amounts with peak concentrations well below 10 ppm.
FIGURE 14
Finally presents the profiles of oxygenated aromatic species. Benzaldehyde, the most abundant of these species with a peak concentration of 150 ppm, and benzylalcohol reach their maxima first around 1 mm above the burner. The profiles of phenol and anisole peak further around 1.4 mm and the maximum of that of benzofuran is the last one around 1.5 mm. This last compound, which can be a precursor of dioxins, corresponds to a very low peak concentration of 2 ppm.
## Description of the Proposed Mechanism
The mechanism proposed here to model the oxidation of n-butylbenzene includes the previous mechanisms that were built by our team to model the oxidation of C\({}_{3}\)-C\({}_{5}\) unsaturated hydrocarbons benzene and toluene. Thermochemical data were estimated by the software THERGAS developed in our laboratory, which is based on the group additivity methods proposed by Benson, apart from the heat of formation of biaromatic species which were taken from Burcat and Ruscic. The complete mechanism is available as supplementary material.
## Reaction base for the oxidation of C\({}_{3}\)-C\({}_{5}\) unsaturated hydrocarbons
This C\({}_{3}\)-C\({}_{5}\) reactions database was built from a review of the literature and is an extension of our previous C\({}_{0}\)-C\({}_{2}\) reactions database. This C\({}_{0}\)-C\({}_{2}\) reactions database includes all the unimolecular or bimolecular reactions involving radicals or molecules including carbon, hydrogen and oxygen atoms and containing less than three carbon atoms. The kinetic data used in this base were taken from the literature and are mainly those proposed by Baulch _et al._ and Tsang _et al._.
The C\({}_{3}\)-C\({}_{5}\) reactions database includes reactions involving C\({}_{3}\)H\({}_{2}\), C\({}_{3}\)H\({}_{3}\), C\({}_{3}\)H\({}_{4}\) (allene and propyne), C\({}_{3}\)H\({}_{5}\), C\({}_{3}\)H\({}_{6}\), C\({}_{4}\)H\({}_{2}\), C\({}_{4}\)H\({}_{3}\), C\({}_{4}\)H\({}_{4}\), C\({}_{4}\)H\({}_{5}\), C\({}_{4}\)H\({}_{6}\) (1,3-butadiene, 1,2-butadiene, 1-butyne and 2-butyne), C\({}_{4}\)H\({}_{7}\) (6 isomers), some linear and branched C\({}_{5}\) compounds and well as cyclopentene and derived species, and the formation of benzene and toluene.
In this reactions base, pressure-dependent rate constants follow the formalism proposed by Troe and efficiency coefficients have been included. It has been validated by modeling experimental results obtained in a jet-stirred reactor for methane and ethane, profiles in laminar flames of methane, acetylene and 1,3-butadiene and shock tube autoignition delay times foracetylene, propyne, allene, 1,3-butadiene, 1-butyne and 2-butyne. The agreement in modeling concentration profiles in laminar flames of pure methane is satisfactory at an equivalence ratio of 1.55, but deteriorates for the formation of ethylene and acetylene at an equivalence ratio of 0.4. An improved version has recently been used to model the structure of a laminar premixed flame of methane doped with allene, propyne, 1,3-butadiene and cyclopentene. Compared to this last version, reactions of formation and consumption of acetone have been added.
## Mechanisms for the oxidation of benzene and toluene
Our mechanism for the oxidation of benzene contains 135 reactions and includes the reactions of benzene and of cyclohexadienyl, phenyl, phenylperoxy, phenoxy, hydroxyphenoxy, cyclopentadienyl, cyclopentadienoxy and hydroxycyclopentadienyl free radicals, as well as the reactions of ortho-benzoquinone, phenol, cyclopentadiene, cyclopentadienone and vinylketene, which are the primary products yielded. Validations have been made using for comparison experimental results obtained in a jet-stirred and a plug flow reactors, profiles in a laminar flame of benzene and shock tube autoignition delay times.
The mechanism for the oxidation of toluene contains 193 reactions and includes the reactions of toluene and of benzyl, tolyl, peroxybenzyl (methylphenyl), alcoxybenzyl and cresoxy free radicals, as well as the reactions of benzaldehyde, benzyl hydroperoxyde, cresol, benzylalcohol, ethylbenzene, styrene and biphenyl. Validations have been made using for comparison experimental results obtained in a jet-stirred and a plug flow reactors, and shock tube autoignition delay times.
Compared to these published mechanisms, reactions of formation and consumption of some products have been added. Cumene is obtained by combination of methyl and phenylethyl radicals or of phenyl and iso-propyl radicals. Methylstyrene is produced by addition of methyl radicals to styrene or by combination of phenylvinyl and methyl radicals. Anisole and benzofuran both derive from phenoxy radicals.
## Mechanism proposed for the oxidation of n-butylbenzene
Table 1 presents the reactions of n-butylbenzene and derived species. The complete mechanism is then composed of this submechanism and of the submechanisms described above. It involves 210 species in 1478 reactions and can be used to run simulations using CHEMKIN softwares. Table 2 presents the names, the formulae and the heats of formation of the aromatic species included in this mechanism and containing at least 9 atoms of carbon.
TABLE S1 AND II
The primary mechanism contains the reactions of n-butylbenzene and of the radicals directly deriving from it. We have considered the unimolecular reactions of n-butylbenzene, the ipso-additions and the H-abstractions by oxygen molecules and radicals present with important concentrations. The rate constants of the unimolecular decompositions involving the breaking of a C-C bond (reactions 1-4) have been calculated from the modified collision theory and thermokinetic relationships. The rate constants of the unimolecular decompositions to give H atoms and phenylbutyl radicals (reactions 5-8) have been deduced from that of the reverse reaction, k = 1.0 \(\times\) 10\({}^{14}\) s-1 according to Allara _et al._. The kinetic parameters of the bimolecular initiations with oxygen molecules (reactions 9-12) have been obtained using the correlation proposed by Ingham _et al._. The rate constants of the ipso-additions of hydrogen and oxygen atoms and of hydroxyl and methyl radicals (reactions 13-17) have been deduced from value proposed for the similar reactions in the cases of benzene and toluene. The rate constants for the abstractions of alkylic H-atoms (reactions 19-21, 23-25, 27-29, 31-33, 35-37) were deduced from the correlations proposed by Buda _et al._ for alkanes and those for the abstractions of allylic H-atoms were deduced from the correlationsproposed by Heyberger _et al._ and by analogy with similar reactions of ethylbenzene.
The reactions of phenylbutyl and phenylpropyl radicals were mainly derived from the reactions of alkyl and alkenyl radicals generated by EXGAS software. In the case of phenylbutyl radicals, reactions involved isomerizations (reactions 42-46), decompositions by breaking of a C-C bond to form styrene, ethylene, propene and 1-butene as stable molecules, the formation of phenylbutenes by breaking a C-H bond (reactions 48, 51-52, 54-55 and 57) or by oxidation with oxygen molecules (reactions 58 to 63). Termination steps (reactions 64 to 67) were only written for the resonance stabilized 4-phenylbut-1-yl radicals: combinations with HO\({}_{2}\) radicals led to phenylbutoxy radicals, combinations with CH\({}_{3}\) radicals to 2-phenyl-n-pentane and disproportionnations with benzyl and allyl radicals to 1-phenyl-1-butene, toluene and propene, respectively. Phenylpropyl radicals were considered to react by isomerisation followed by decomposition (reaction 68) to lead ultimately to styrene, by cyclisation (reaction 69) to give indane, by \(\beta\)-scission decompositions (reactions 70-71) yielding ethylene and allylbenzene as stable molecules, by oxidation (reaction 72) to produce allylbenzene and by combination with H-atoms to form propylbenzene (reaction 73). The reactions of 1-butyl radicals (reactions 74 to 78) are a reduced version of those generated by EXGAS software for their high temperature oxidation.
The reactions of oxygenated aromatic radicals are still very uncertain. The globalized reactions considered for butylphenoxy radicals (reactions 79-80) are derived from those of phenoxy and started as an elimination of carbon monoxide. Propylbenzylalcoxy radicals have been assumed to react by isomerizations with a rapid decomposition of the obtained radicals (reactions 81-82) or by direct \(\beta\)-scission decompositions (reactions 83-84).
The secondary mechanism includes the reactions of the primary products which are not considered in the mechanisms of the oxidation of benzene and toluene, namely the three isomersof phenylbutene (reactions 85 to 164), allylbenzene (reactions 165 to 182), propylbenzene (reactions 183 to 200), 2-phenyl-n-pentane (reactions 201 to 204) and indane (reactions 205 to 216), as well as those of the derived products, namely butadienylbenzenes (reactions 217 to 226), the bicyclic isomers of C\({}_{10}\)H\({}_{10}\) (reactions 227 to 231), indene (reactions 232 to 243) and naphthalene (reactions 244 to 255).
In the case of the isomers of phenylbutene and allylbenzene, we have considered the bimolecular initiations with oxygen molecules, the ipso-additions of H-atoms and methyl radicals to give benzene and toluene, the additions to the double bond followed by the decompositions of the obtained adducts and the H-abstractions to give phenylalkenyl radicals, which can react by decompositions by \(\beta\)-scission, oxidations, cyclizations and combinations. The rate constants of the additions to the double bond and of the formation and the consumption of alkenyl radicals were derived from our previous work on the oxidation of alkenes. The cyclizations of phenylbutenyl and phenylpropenyl radicals lead to bicyclic isomers of C\({}_{10}\)H\({}_{10}\) (dihydronaphthalene and methylindene) and indene, respectively. Their rate constant has been assumed equal to that proposed by Gierzak for the cyclization of pentenyl radicals to give cyclopentyl radicals. For propylbenzene, unimolecular and bimolecular initiations, ipso-additions of H-atoms and methyl radicals and H-abstractions have been written. In the case of 2-phenyl-pentane, which has not been experimentally observed, only four globalized H-abstractions were taken into account. Indane has been considered as yielding indene by bimolecular initiations with oxygen molecules and H-abstractions by small radicals followed by \(\beta\)-scission decompositions of the obtained indanyl radicals.
As only very few studies concerning their oxidation have been published, the reactions of the other products are still very uncertain. For butadienylbenzene, globalized reactions have been written starting by bimolecular initiations with oxygen molecules or H-abstractions by O- and H-atoms and OH radicals, both followed by the cyclization of the obtained radicals yielding naphthalene. The reactions of the bicyclic isomers of C\({}_{10}\)H\({}_{10}\) have also been assumed as leading to naphthalene by molecular dehydrogenation or by H-abstractions. In the case of indene, the addition of OH radicals to the double bond included in the five members cycle has been written, as well as bimolecular initiations with oxygen molecules and H-abstractions by small radicals to give resonance stabilized indenyl radicals, the combinations of which has been assumed to be similar to those written by Da Costa _et al._ for cyclopentadienyl radicals. Finally, as previously assumed in the work of Bounaceur _et al._, the reactions of naphthalene and naphthyl radicals can be derived from those of benzene and phenyl radicals.
## Comparison between experimental and simulated results
Simulations were performed using PREMIX from CHEMKIN II using the experimental temperature profile as an input. To compensate the perturbations induced by the quartz probe and the thermocouple, the temperature profile used in calculations is an average between the experimental profiles measured with and without the quartz probe, shifted 0.4 mm away from the burner surface, as shown in
Figures 5 to 7 and 11 show that the model reproduces satisfactorily the consumption of reactants and the formation of the main C\({}_{0}\)-C\({}_{2}\) products, including the oxygenated ones, related to the consumption of methane in the flame doped with n-butylbenzene. Only the formation of hydrogen in the burned gas is largely overestimated, that of ethylene is underpredicted by a factor almost 2 and that of ketene and ethanol are overpredicted by factors 10 and 20, respectively.
To decouple the effect due to the increase of equivalence ratio (\(\phi\)) and that induced by the presence of n-butylbenzene, figures 5 to 7 and 11 display also the results of a simulation performed for a flame containing 7.1% methane and 15.64 % oxygen (with no additive) for \(\phi\)=0.74, i.e. equal to that of the doped flame. As the temperature rise is mainly influenced by \(\phi\), we have used the same temperature profile as that used to model the doped flame. The profile of methane is very similar for the doped and undoped flames at the same \(\phi\). However the content in atoms of carbon is 2.35 lower in the pure methane flame compared to the doped one (the C/O ratio is equal to 0.096 in the undoped flame and to 0.225 in the seeded one) which is well reflected by the profiles of carbon oxides and ethane. While the profile of ethane is not much affected, the formation of ethylene and acetylene is increased by the addition of n-butylbenzene by a factor much larger than the increase in the number of atoms of carbon showing clearly to what extent these products derive from the decomposition of this additive. The presence of the additive does not influence the formation of methanol, but increase also considerably that of ketene, acetaldehyde and ethanol.
Figures 8 to 11 present the comparison between experimental and simulated data for the other non-aromatic products, which are all modeled with a factor better than 2, except from n-butane and methylcyclopentene, for which a deviation of a factor 3 is observed, and isoprene, methylcyclopentadiene, and acetone, the formation of which is more considerably underestimated. Comparison with the simulation of a pure methane flame at \(\Phi\)=0.74 shows that C\({}_{3}\) compounds, which are formed in significant amounts in the doped flame, are almost non-existent in the lean mixture without additive, except for propane. Specific reactions leading to unsaturated C\({}_{3}\) products are then also induced by the presence of the cyclic additive.
Figures 12 to 15 display the comparison for the profiles of aromatic compounds. Concerning the most abundant species, the formation of benzene, toluene and allylbenzene is very well reproduced, that of styrene is underestimated by a factor 1.5, that of benzaldehyde is overestimated by the same factor and that of butenylbenzenes is overestimated by a factor 2.5.
The poor separation of the C\({}_{10}\)H\({}_{12}\) isomers on the GC column that we used is a possible explanation for this discrepancy. While the simulated profiles of bicyclic compounds notably differ from the experimental ones, the maximum mole fractions obtained for minor products are all modeled with a factor better than 4, except from benzofuran, the formation of which is underestimated by a factor 10 showing that that a production pathway is missing in the model. The production of phenol is overestimated by a still larger factor.
It is worth noting that the largest disagreements are obtained for several oxygenated compounds. For ethanol, ketene and phenol which are strongly overestimated by the model, it is probable that, despite the heated transfer line, some of these oxygenated compounds may be absorbed on the walls and result in lower measured values than those predicted by our model. While the reactions of ketene are still very uncertain, those of ethanol and phenol are rather well determined with pressure effects taken into account.
In order to extend the temperature ranges of these simulations, the results of Litzinger _et al._, which were obtained in a flow reactor at 1060 K, at atmospheric pressure, with nitrogen as bath gas, for an initial concentration of n-butylbenzene of 620 ppm and for an equivalence ratio of 0.98, have also been modeled. displays comparisons between the experimental and computed mole fraction of reactants and main products. This figure shows that a globally correct agreement can be observed. The consumptions of hydrocarbon and oxygen are correctly reproduced, as well as the formation of the major products, methane, ethylene, benzene, toluene, styrene, allylbenzene. The production of carbon monoxide is overestimated by a factor 1.5 at the longest residence times.
FIGURE 16
## Discussion
A large enough conversion has been chosen, so that the major ways of consumption of the primary products can be observed. Under these conditions, butylbenzene is mainly consumed by H-abstractions (40 % if its consumption) to give the resonance stabilized 4-phenylbut-4-yl radical which is the main source of styrene through a b-scission decomposition. At this temperature, this radical is consumed faster than it is produced which is only possible because of its accumulation at lower temperature.
FIGURE 17
The other important channels of consumption of n-butylbenzene are the H-abstractions leading to the other 4-phenylbutyl radicals (globally 41 % of its consumption). A part of 4-phenylbut-1-yl and 4-phenylbut-2-yl radicals isomerizes to give resonance stabilized 4-phenylbut-4-yl radicals. However 4-phenylbut-1-yl radicals mainly decompose to give ethylene and 2-phenyleth-1-yl radicals yielding phenyl radicals and ethylene or, for a smaller part, styrene and H-atoms. The decomposition of 4-phenylbut-1-yl radicals explains most part of the formation of ethylene. 4-phenylbut-2-yl radicals are consumed to produced allylbenzene and methyl radicals, butenylbenzene and H-atoms and 1-butene and phenyl radicals. This last channel is the main way to form 1-butene. Phenyl radicals are an important source of toluene and phenoxy radicals yielding phenol, and to a much lesser extent of benzene. Anisol and benzofuran are produced from minor reactions of phenoxy with methyl radicals and acetylene, respectively. As benzofuran is strongly underestimated by the model, it is probable that important way of formation of this compound are missing. Finally 4-phenylbut-3-yl radicals are almost completely consumed to form propene and benzyl radicals, which are the major source of toluene, benzaldehyde, ethylbenzene, benzylalcohol and n-propylbenzene. The formation of xylenesderive from toluene through the ipso-addition of methyl radicals. The largest part of the formation of propene is due to reactions of 4-phenylbut-3-yl radicals.
There are two additional minor channels of consumption of n-butylbenzene by ipso-additions. The ipso-addition of H-atoms (6 % of its consumption) is the main source of benzene. Butyl radicals are also produced through this way and mostly decompose to form ethylene and ethyl radicals and to a lesser extent 1-butene and H-atoms. A very small fraction of them also combines with H-atoms to give butane. Ethyl radicals are a source of acetaldehyde by reaction with O-atoms and of ethanol by combinations with OH radicals. The ipso-addition of O-atoms (3.7 % of its consumption) leads ultimately to carbon monoxide, as well as to cyclopentadienyl radicals and 1-butene or to H-atom, propene and benzene.
Styrene reacts mainly by additions of OH radicals or of H-atoms and by H-abstractions. The additions of OH radicals are a source of formaldehyde and benzyl radicals and of benzaldehyde and methyl radicals. The additions of H-atoms produce resonance stabilized 2-phenylethyl-2-yl radicals, the combinations of which yield cumene and ethylbenzene. Phenylvinyl radicals which are obtained by H-abstractions from styrene react mainly with oxygen molecules yielding resonance stabilized benzoyl radicals which decompose to produce phenyl radicals and carbon monoxide. Minor channels consuming phenylvinyl radicals involve the formation of methylstyrene and of phenylacetylene.
Butenylbenzenes and allylbenzene react mainly by H-abstractions to give resonance stabilized radicals the cyclization of which leads to methylindene and indene, respectively. The cyclization of the alkylic alkenyl radicals, which are formed by another minor channel consuming butenylbenzenes, produces dihydronaphthalene. The cyclization of the phenylpropyl radicals, which are obtained by addition of H-atoms to allylbenzene, explains the formation of indane (not shown in fig. 17, as it corresponds to a very small flow rate). The flow rate of the addition to allybenzene is much smaller than that of the H-abstractions explaining the lower production of indane compared to indene. The reactions of methylindene and dihydronaphthalene lead to naphthalene or to resonance stabilized cyclopentadienyl radicals. These cyclic C\({}_{5}\) radicals mostly yield cyclopentadiene by combination with H-atoms, but also, to a lower extent, methylcyclopentadiene by combinations with methyl radicals, and acetylene or 1,3-butadiene by opening of the ring. A small part of cyclopentadienyl radicals derives also from the CO elimination from phenoxy radicals.
Let us now describe the major ways of formation of the minor species which do not derives directly from butylbenzene. Ethane, propane and methanol are formed by combination of methyl radicals with themselves, ethyl and OH radicals, which explains why their formation is not much affected by the presence of butylbenzene. Allene derives from propene via the formation of resonance stabilized allyl radicals. Propyne is formed by addition of O-atoms to vinylacetylene, which is obtained from cyclopentadienone which derives from benzoquinone, a minor product of the reaction of phenyl radicals with oxygen molecules. The additions of methyl radicals to propyne yield iso-butyl radicals, the recombination of which is the main source of iso-butene. Diacetylene and butynes derive from vinylacetylene. 1,2-butadiene is obtained from resonance stabilized 1-buten-3-yl radicals, which are produced by H-abstractions from 1-butene. 1,3-pentadiene derives from cyclopentene after H-abstractions and opening of the ring and isoprene from the addition of vinyl radicals to allene. Ketene and acetone both derive from CH\({}_{3}\)CO radicals obtained by H-abstraction from acetaldehyde. Propanal is produced by addition of OH radicals to 1-butene or butenylbenzenes. Acrolein is mainly obtained by reaction of allyl radicals with O-atoms.
## Conclusion
This paper presents new experimental results for a lean premixed laminar flame of methane seeded with n-butylbenzene, as well as a new mechanism developed to reproduce the combustion of this substituted aromatic compound, which can be considered as a model molecule of an important class of components of diesel fuels. Profiles of temperature have been measured and mole fraction profiles have been obtained for 55 identified stable species from C\({}_{0}\) to C\({}_{10}\), including 20 aromatic products and 12 oxygenated compounds other than the reactants. Several of these species are considered as toxic pollutants, this is the case of oxygenated compounds (acrolein). Several of these products are also known as soot precursors, this is the case of all the aromatic compounds.
Satisfactory agreement has been obtained between experimental results and simulations, apart from 4 oxygenated species (ethanol, ketene, phenol, benzofuran). The prediction of the profiles of bicyclic aromatic compounds (indene, indane, naphthalene, benzofuran) could also be improved when a better knowledge of their chemistry will be available.
## AKNOWLEDGEMENT
This work has been supported by PSA Peugeot Citroen and TOTAL
# Figure Captions
Scheme of the apparatus of low-pressure laminar premixed flame. The thick lines correspond to the heated lines.
Typical chromatogram of light compounds obtained at a distance of 1.40 mm from the burner at 1380 K (oven temperature program: 333 K during 10 min, then a rise of 5 K/min until 523 K).
Typical chromatogram of heavy compounds obtained at a distance of 1.31 mm from the burner at 1310 K (oven temperature program: 313 K during 30 min, then a rise of 5 K/min until 453 K).
Temperature profiles: experimental measurements performed without and with the sampling probe and profile used for simulation.
Profiles of the mole fractions of both hydrocarbon reactants including a computed comparison between doped and undoped flames. Points are experiments and lines simulations. In the case of methane, the full lines correspond to the flame seeded with n-butylbenzene and the broken one to a simulated flame of pure methane at \(\Phi\)=0.74 (see text).
Profiles of the mole fractions of oxygen and the main oxygenated products including a computed comparison between doped and undoped flames. Points are experiments and lines simulations. Full lines correspond to the flame seeded with n-butylbenzene and broken lines to a simulated flame of pure methane at \(\Phi\)=0.74 (see text).
Profiles of the mole fractions of hydrogen and C\({}_{2}\) species including a computed comparison between doped and undoped flames. Points are experiments and lines simulations. Full lines correspond to the flame seeded with n-butylbenzene and the broken lines to a simulated flame of pure methane at \(\Phi\)=0.74 (see text).
Profiles of the mole fractions of C\({}_{3}\) species including a computed comparison between doped and undoped flames. Points are experiments and lines simulations. Full lines correspond to the flame seeded with n-butylbenzene and the broken lines to a simulated flame of pure methane at \(\Phi\)=0.74 (see text).
Profiles of the mole fractions of C\({}_{4}\) species. Points are experiments and lines simulations.
Profiles of the mole fractions of C\({}_{5}\)-C\({}_{6}\) non-aromatic species. Points are experiments and lines simulations.
Profiles of the mole fractions of oxygenated C\({}_{1}\)-C\({}_{3}\) species. Points are experiments and lines simulations. Full lines correspond to the flame seeded with butylbenzene and the broken lines to a simulated flame of pure methane at \(\Phi\)=0.74 (see text).
Profiles of the mole fractions of C\({}_{6}\)-C\({}_{8}\) aromatic species. Points are experiments and lines simulations.
Profiles of the mole fractions of C\({}_{9}\)-C\({}_{10}\) monoaromatic species. Points are experiments and lines simulations.
Profiles of the mole fractions of biaromatic species. Points are experiments and lines simulations.
Profiles of the mole fractions of oxygenated aromatic species. Points are experiments and lines simulations.
Computed species mole fractions versus residence time during the oxidation of n-butylbenzene in a flow reactor at 1060 K, P= 1atm and \(\Phi\) = 0.98, with an initial concentration of hydrocarbon of 620 ppm. Points are experiments and lines simulations.
Flow rate analysis for the consumption of the n-butylbenzene for a distance of 1.15 mm from the burner corresponding to a temperature of 1144 K and a 87 % conversion of butylbenzene. The size of the arrows is proportional to the relative flow rates.
Distance to the burner (mm)
Distance to the burner (mm)
Distance to the burner (mm)
Distance to the burner (mm)
Distance to the burner (mm)
Distance to the burner (mm)
Figure 17:
|
10.48550/arXiv.0903.4948
|
A Lean Methane Prelixed Laminar Flame Doped witg Components of Diesel Fuel. Part I: n)Butylbenzene
|
Emir Pousse, Pierre-Alexandre Glaude, René Fournet, Frédérique Battin-Leclerc
| 5,570
|
10.48550_arXiv.2002.05085
|
###### Abstract
Chemically-active droplets exhibit complex avoiding trajectories. While heterogeneity is inevitable in active matter experiments, it is mostly overlooked in their modelling. Exploiting its geometric simplicity, we fully-resolve the head-on collision of two swimming droplets of different radii and demonstrate that even a small contrast in size critically conditions their collision and subsequent dynamics. We identify three fundamentally-different regimes. The resulting high sensitivity of pairwise collisions is expected to profoundly affect their collective dynamics.
Exploiting the simplicity of a two-sphere geometry, we were recently able to fully-resolve the symmetric collision of two droplets, and provided a critical insight on how the nonlinear convective transport conditions the collision. To note, controlling experimentally the exact droplet radius, which slowly varies over time, is however impossible, giving rise to heterogeneous systems and non-symmetric in practice.
In this Letter, using a fully-coupled model of the detailed solute and flow dynamics, we demonstrate how even a small contrast in the droplets' size critically conditions their head-on collision and subsequent dynamics, thus likely affecting substantially the complex collective behavior of active droplets suspensions. To this end, we consider the axisymmetric dynamics of two droplets of fixed radii \(R_{1}\) and \(R_{2}\), and of identical chemical properties, with a minimum surface-to-surface distance \(d\). The chemical activity of the droplet is generically modelled here as a fixed-rate surface release of solute - coined \(\mathcal{A}\) - into the outer fluid. The interaction of this solute with the droplets' interface modifies its effective local surface tension, introducing a chemically-induced Marangoni stress.
\[D\mathbf{n}\cdot\nabla c=-\mathcal{A},\qquad\qquad\mathbf{n}^{\perp}\cdot[ \boldsymbol{\sigma}-\boldsymbol{\tilde{\sigma}}]\cdot\mathbf{n}=-\gamma_{c} \nabla_{\parallel}c\;, \tag{1}\]
\(c\) denotes the solute concentration and \(D\) its molecular diffusivity. In addition for small concentration variations, the solute-induced change in surface tension, \(\gamma_{c}=(\partial\gamma/\partial c)\), is constant, and \(\boldsymbol{\sigma}\) and \(\boldsymbol{\tilde{\sigma}}\) are respectively the Cauchy stress tensors of the outer and inner Newtonian fluids of viscosities \(\eta\) and \(\tilde{\eta}\). Droplets must be negatively auto-chemotactic in order to self-propel: they must be repelled by the solute they release, which is advected into their wake by the surface Marangoni flows. This, in turn, imposes \(\mathcal{A}\gamma_{c}>0\); both are taken positive in the following.
Chemical solutes involved in the spontaneous emulsification of active droplets are large molecular compounds and diffuse slowly so that their concentration \(c\) in the outer phase follows an advection-diffusion equation
\[\frac{\partial c}{\partial t}+\mathbf{u}\cdot\nabla c=D\nabla^{2}c\;. \tag{2}\]
Due to the microscopic size of the droplets, inertial effects are negligible, and the velocity of the fluid is found by solving Stokes' equations in each phase, coupled through the Marangoni condition, Eq.. Each droplet is hydrodynamically force-free, which determines uniquely their velocities, \(V_{1}\) and \(V_{2}\), measured positively along the axis of symmetry \(z\). The full nonlinear coupling of the Stokes and chemical transport problems is solved numerically for arbitrary distance using an efficient scheme based on a time-dependent bispherical grid presented in detail in Ref..
The problem is now re-scaled using \(R_{2}\), \(\mathcal{V}_{2}\) and \(\mathcal{A}R_{2}/D\) as reference length, velocity and concentration respectively, where \(\mathcal{V}_{i}=\mathcal{A}\gamma_{c}R_{i}/[D(2\eta+3\tilde{\eta})]\) is the passive drift velocity of a droplet of radius \(R_{i}\) in an externally-imposed gradient \(\mathcal{A}/D\). In addition to the viscosity ratio \(\eta/\tilde{\eta}\) (which is set to unity for simplicity here), the problem is fully-determined by two independent non-dimensional parameters, namely the specific Peclet number (or advection-diffusion ratio) \(\mathrm{Pe}_{i}=R_{i}\mathcal{V}_{i}/D\) of each droplet. Together they are related to the size ratio \(\xi=R_{1}/R_{2}=\sqrt{\mathrm{Pe}_{1}/\mathrm{Pe}_{2}}\). The Peclet number plays a key role for an individual droplet's dynamics, and for \(\mathrm{Pe}_{i}\geq\mathrm{Pe}_{c}=4\), a droplet of radius \(R_{i}\) starts swimming spontaneously as a result of a transcritical bifurcation.
To identify the effect of the size ratio, \(\xi\), the results are reported in the following in terms of \(\xi\geq 1\) and of the Peclet number of the smaller droplet, \(\mathrm{Pe}_{2}\geq 4\) for the time-dependent dynamics of two droplets initially located far apart (\(d\gg 1\)) and swimming toward each other (\(V_{2}>0\) and \(V_{1}<0\)). Depending on the value of \((\xi,\mathrm{Pe}_{2})\), three dynamic collision regimes are observed, and are analyzed in detail in the following.
_Asymmetric rebound -_ Regardless of the advection-to-diffusion ratio, \(\mathrm{Pe}_{2}\), and for sufficiently small contrast in size \((\xi-1\ll 1)\), the collision always leads to a reversal of the swimming direction of both droplets (Fig. 2, blue). The general dynamics associated with this regime is similar to that identified for symmetric collisions: as the droplets get closer, solute accumulates between them, reducing the polarity of their surface concentration, and effectively acting as a chemical repulsion. This chemical asymmetry vanishes and reverses sign for a relative distance \(d\) that decreases with \(\mathrm{Pe}_{2}\); as a result, the reversed mechanical stresses propel the droplets away from each other. It should be noted however that the variations of each droplet's axial velocity are not symmetric with respect to the collision. Specifically, the droplets temporarily swim away more slowly than in their final approach phase, and this effect is even more pronounced for the larger droplet. For symmetric collisions, this delayed dynamic reversal of the concentration polarity was identified as the result of slower diffusion (larger \(\mathrm{Pe}\)); this observation likely explains the difference of behaviour between the two droplets (\(\mathrm{Pe}_{1}\geq\mathrm{Pe}_{2}\)).
_Chasing -_ For moderate \(\mathrm{Pe}_{2}\) and larger size contrast \(\xi\), a chasing regime is observed where instead both droplets eventually swim in the same direction. While the smaller droplet still experiences a similar repulsion and rebound dynamics, the effective repulsion it exerts on the larger droplet is not sufficient to reverse its swimming direction (Fig. 2, green). As the smaller droplet swims away, the polarity of the larger droplet still allows it to accelerate and recover its original velocity, effectively chasing the smaller droplet. The absolute self-propulsion velocity of an isolated droplet increases with its radius (when rescaled by its radius it increases linearly with \(\mathrm{Pe}_{i}\) before reaching a plateau). As a result, when far enough apart, the larger droplet will swim faster than and catch up with the smaller one. A bound state may then arise, where the two droplets maintain a fixed distance \(d_{\mathrm{eq}}\) and swim together. Their common velocity results from the balance of their self-propulsion properties and mutual chemical repulsion.
For fixed \(\mathrm{Pe}_{2}\), the rebound-to-chasing transition with increasing \(\xi\), illustrated in Fig. 3, depends on the detailed dynamics of the collision and thus occurs at a precise value \(\xi_{c}(\mathrm{Pe}_{2})\). This transition as \(\xi\) increases results from two effects that prevent the larger droplet's rebound. The smaller droplet being exposed to an increased chemical gradient is quickly repelled away, reducing the interaction time with the larger one. Additionally, the increased size (or \(\mathrm{Pe}\)) of the larger droplet enhances the persistence of its chemical polarity and self-propulsion, as already observed for symmetric collisions.
The emergence of the stationary chasing regime is analysed here further. In this bound state, both droplets have the same swimming velocity \(V_{2}^{\infty}<V_{b}<V_{1}^{\infty}\) with \(V_{i}^{\infty}\) the velocity of each droplet when isolated.
Transition between the rebound (blue) and chasing (green) regimes: influence of the size ratio \(\xi\) on the evolution of the larger droplet velocity, \(V_{1}\), parameterized by the droplet position.
(a) Regime selection for the asymmetric dynamic collision of two droplets for varying \((\xi,\mathrm{Pe}_{2})\): rebound (blue square), chasing (green disk) and pausing regimes (red, triangle). (b) Snapshots of the corresponding dynamics for each regime, showing the droplets’ instantaneous velocity (arrow) and solute concentration distribution (color) (movies available in Supplemental Material). (c) Evolution of the small (dashed) and large (solid) droplets’ velocity with their position during three representative collision regimes at \(\mathrm{Pe}=12\): \(\xi=1.3\) (blue, rebound), \(\xi=1.5\) (green, chasing) and \(\xi=1.8\) (red, pausing).
Increasing the distance between the droplets reduces both effects leading to the larger droplet swimming faster than, and catching up with, the smaller droplet. Conversely, a reduction in \(d\) induces larger chemical repulsions and the resulting acceleration (resp. deceleration) of the smaller (resp. larger) droplet.
Following Ref., these arguments can be formulated quantitatively in the asymptotic limit of small supercriticality, i.e. when \(\delta_{i}=\mathrm{Pe}_{i}-\mathrm{Pe}_{c}\) are both small (which imposes \(\xi-1\ll 1\)). The velocity of droplet \(i\) is then obtained at leading order as a function of the leading chemical gradient generated by droplet \(j\),
\[|V_{i}|=\frac{V_{i}^{\infty}}{2}\left(1+\sqrt{1+\mathrm{sgn}(V_{i})\frac{256G_ {j\to i}}{\delta_{i}^{2}d^{2}}}\right), \tag{3}\]
In Eq., \(G_{j\to i}\) is proportional to the chemical repulsion exerted by droplet \(j\) on droplet \(i\) and may either slow down or speed up the motion of the latter depending on the relative signs of \(V_{i}\) and \(G_{j\to i}\). Its exact expression depends on the swimming direction of \(j\) to account for the exponential (resp. algebraic) decay of concentration in front (resp. behind) that swimming droplet:
\[G_{1\to 2}= \left\{\begin{array}{ll}-1&\mbox{if $V_{1}>0$},\\ -\mathrm{exp}\left[-4d|V_{1}|\right](1+4d|V_{1}|)&\mbox{if $V_{1}<0$},\end{array} \right.\qquad G_{2\to 1}=\quad\left\{\begin{array}{ll}1&\mbox{if $V_{2}<0$},\\ \mathrm{exp}\left[-4d|V_{2}|\right](1+4d|V_{2}|)&\mbox{if $V_{2}>0$}.\end{array}\right. \tag{4}\]
Solving Eqs.- for each droplet, and considering the situation of a leading smaller droplet, the leading order value when \(\xi-1\ll 1\) of the equilibrium distance \(d_{\rm eq}\) for which \(V_{1}=V_{2}<0\) is obtained as
\[d_{\rm eq}=\frac{2\sqrt{2}}{\sqrt{(\xi-1)(\mathrm{Pe}_{2}-4)}}. \tag{5}\]
This result quantifies the physical divergence of \(d_{\rm eq}\) when both droplets have the same size or swim near the critical threshold: in both cases, the swimming velocities are small so that their difference can only be compensated by weak chemical repulsion (or large inter-droplet distance).
The results of the full model, Fig. 4, are consistent with the \((\xi-1)^{-1/2}\) scaling despite the large value of \(\delta_{2}\), which is the most likely origin of any difference in prefactor. It should be further noted from that \(d_{\rm eq}\) presents a minimum for intermediate \(\xi\), which is consistent with its divergence for small \(\xi\). For large \(\xi\), the influence of the chemical repulsion on the velocity of the larger droplet is negligeable and the whole assembly swims at a velocity \(V=V_{1}^{\infty}\gg V_{2}^{\infty}\). This velocity scales in dimensionless units as \(\xi\), and is much larger than the spontaneous self-propulsion of the smaller droplet \(V_{2}^{\infty}\). Its motion is therefore dominated by its passive drift in the chemical gradient of the larger droplet which scales as \(\xi^{2}/d^{2}\): as a result, \(d_{\rm eq}\sim\xi^{1/2}\) becomes an increasing function of \(\xi\).
While the bound states emerging from this
Evolution with the size ratio \(\xi\) of the equilibrium distance \(d_{\rm eq}\) between the droplets in the bound state for \(\mathrm{Pe}_{2}=6\). The predictions of Eq. are also reported (dashed line in the log-log inset). Results are obtained from the time-dependent dynamics of two droplets initially swimming in the same direction (black crosses) or toward each other (green dots).
The exact history of the droplets' motion therefore appears critical in setting their long-term dynamics, thus underlining the nonlinearity and complexity of the transition to such bound states.
## 4 Pausing
For larger \(\mathrm{Pe}_{2}\), and even for relatively small contrast in size, a pausing regime arises. The larger droplet is slowed down by the excess concentration generated by the approach with the smaller droplet. Very much like for the other regimes, the smaller droplet quickly reverses direction and propels away from the collision site. However, the chemical repulsion experienced by the larger droplet during this brief encounter is neither sufficiently large nor long to reverse its chemical polarity and provoke its rebound; but it is sufficient to accumulate solute at its front, resulting in a symmetric but non-uniform surface concentration, driving pusher-like Marangoni flows from the equatorial plane toward its poles. Such quadrupolar flow balances diffusion and maintains a stationary regime where the droplet acts as a symmetric pump (see Fig. 5). The specific Peclet number of the larger droplet in this regime, \(\mathrm{Pe}_{1}=\xi^{2}\mathrm{Pe}_{2}\), is greater than the instability threshold of the quadrupolar mode for a non-deformable active droplet. The emergence of the pausing regime can therefore be interpreted as the non-linear transition between two branches of the single droplet dynamics, in response to the finite perturbation induced by the collision.
The stability analysis of this pumping state lies beyond the scope of the present Letter and is left for future research. Yet, regardless of the detailed stability properties, the dominance of the dipolar propelling mode in the dynamics of a single non-deformable droplet, suggests that sufficiently large perturbations of the concentration or flow fields (e.g. by another droplet) will likely provoke a new mode switching and self-propulsion of the larger droplet. This may however occur long after the first collision so that memory of the initial propulsion direction will be lost, reminiscent of the run-and-tumble motion of swimming bacteria. This is in stark contrast with the bouncing and chasing regimes, which do preserve the collision's directionality. Such memory loss is expected to significantly affect the long-term collective dynamics of droplets, by introducing an effective rotational diffusion.
In summary, we showed in this Letter how variability in the size of active droplets profoundly affects their collective dynamics as a result of the sensitivity of the collision outcome to the precise droplet characteristics. Although not considered here explicitly, variability in chemical properties likely has a similar effect, since regime selection results mainly from the droplet's chemical signature intensity and specific Peclet number. Once again, the strength of the advective coupling is a key factor: for moderate advection, the dynamics is only weakly modified from the symmetric collision, and large size contrast is required for more complex regimes. However, for large advective effects, non-symmetric bouncing, chasing or scattering may develop, even for droplets of comparable sizes, stressing the extreme sensitivity of the interaction.
A major limitation of the present study lies in the axisymmetric assumption, which locks the droplets' dynamics along a single axis. As a result, a \(180^{\circ}\) flip of at least one droplet's motion is the only possible outcome of the collision, which may be avoided through small scattering in more generic situations. The stability of these axisymmetric regimes (i.e. their sensitivity to small fluctuations in the initial droplets' alignment) also remains elusive. Yet, by allowing for a complete resolution of the fully-coupled problem, this framework provides significant in-depth physical insight on the nonlinear chemo-hydrodynamic interactions of active droplets, which may prove essential in future theoretical characterization and experimental interpretation of more generic collisions, as well as for understanding the resulting scattering and collective dynamics.
Pausing regime: Streamlines (top) and solute concentration (bottom) around the arrested larger droplet after collision in the pausing regime obtained for \(\mathrm{Pe}=12\) and \(\xi=1.6\).
|
10.48550/arXiv.2002.05085
|
Bouncing, chasing or pausing: asymmetric collisions of active droplets
|
Kevin Lippera, Michael Benzaquen, Sebastien Michelin
| 5,507
|
10.48550_arXiv.2006.12597
|
## 1 Introduction
The computational description of structural patterns and acidity constants in condensed molecular environment is a challenging problem. It relies upon both accurate _ab initio_ quantum chemistry (_e.g._, hybrid density functional theory (DFT)) and converged statistical sampling involving Born-Oppenheimer molecular dynamics. Yet, the numerical evaluation of energies and forces at the hybrid DFT level is too computationally demanding to achieve long timescale simulations of explicitly solvated species. Structural and chemical properties of acids and bases are routinely measured experimentally but their estimate for strong or weak acids as well as unstable species and molecules with multiple tautomeric equilibria is not straightforward.
Protonated hydrogen peroxide is a prototypical example that serves as a strong oxidizing agent in acidic media. It is, for instance, used for the catalytic hydroxylation of phenol, as an industrial route towards the production of catechol and hydroquinone, key chemicals in the manufacturing of cosmetic, pharmaceutical, and agrochemical products. The reaction (see Fig. 1) involves protonated hydrogen peroxide - formed _in situ_ in the presence of an acid - that reacts with phenol via electrophilic aromatic substitution. One of the possible acids for this reaction is methanesulfonic acid (CH\({}_{3}\)SO\({}_{3}\)H, pK\({}_{a}\)(H\({}_{2}\)O) = -1.9), which is a non-volatile liquid at ambient temperature, soluble in organic solvents and amphiphilic media.
Identifying the protonation state of methanesulfonic acid and its solvation shells in mixtures of phenol and hydrogen peroxide is complicated by the amphiphilic character of phenol. Phenol acts as a weak acid, which causes a decrease of the common acid strength in comparison to the aqueous medium. CH\({}_{3}\)SO\({}_{3}\)H interacts with phenol through both hydrogen bonding and apolar interactions owing to presence of the hydroxyl group and an aromatic ring. From the computational perspective, achieving an accurate description of the non-trivial interplay between hydrogen bonds and/or \(\pi\)-interactions between the acid and hydrogen peroxide should help rationalizing the stability of the possible reaction intermediates and even the regioselectivity of the reaction.
Here, we demonstrate how the combination of data-driven and enhanced sampling techniques helps characterizing molecular patterns in amphiphilic media. First, we develop a set of neural-network-based reactive force fields, which retain (in the interpolative regime) the accuracy of the hybrid DFT energy and force computations they are trained on, at a fraction of their computational cost.
Reaction scheme of the acid-catalyzed hydroxylation of phenol by hydrogen peroxide to form catechol or hydroquinone and water.
The trained neural networks (_i.e.,_ baselined and direct) are integrated into a molecular dynamic driver, making use of a multiple time-stepping (MTS) approach, to perform replica exchange molecular dynamics (REMD), metadynamics and path integral MD (PIMD). Finally, we use data-driven analysis techniques to characterize and estimate the occurrence of recurrent structural patterns in the solvation environment of the different species.
The article starts with a general description of the proposed workflow, followed by technical details associated with the theory and implementation of the reference forces and energies, the training of the neural network, and the enhanced sampling schemes. The framework is then validated on a test set, and applied to the analysis of the solvation of hydrogen peroxide and methanesulfonic acid in phenol, taken as an illustrative example.
## 2 Operational Workflow
We begin by providing an overview of the workflow we introduce to achieve an accurate yet computationally affordable exploration of the free energy landscape of a complex, fully-solvated chemical system. The workflow is illustrated in Fig. 2, and is based on the following steps:
Graphical summary of the essential ingredients of the workflow we used to train a robust MLP, and to use it to sample the free-energy landscape of a complex solution
* **Preliminary phase space exploration:** A relevant portion of the phase space is explored at a low-cost electronic structure level (_e.g._, semi-empirical method).
* **Database selection:** A set of distinct structures is selected by a farthest point sampling (FPS) algorithm to avoid structural redundancy.
* **Reference energies and forces:** The forces and energies of the configurations selected at the previous step are computed at the reference electronic structure level (_e.g._, beyond DFT or hybrid DFT including London dispersion corrections).
* **Training of the Neural Networks:** A NN model is trained to reproduce DFT energies and forces, both directly, and using the semiempirical method as a baseline.
* **Free energy surface exploration:** Extensive sampling is performed by combining direct and baselined NNs in MD simulation runs integrated through a MTS scheme to achieve the accuracy of the latter and exploit the efficiency of the former.
* **On-the-fly uncertainty estimate:** On-the-fly uncertainty estimations based on committee models are used to monitor the extrapolation error and, if necessary, feed novel structures to the NN, to improve its accuracy and reliability along the simulations.
## 3 Methods
### Machine Learning Potentials
Machine learning potentials (MLP) trained on DFT data are increasingly used to achieve fast-and-accurate prediction of molecular energetics involved in complex atomistic systems. They have often been employed to investigate chemical reactions in the gas phase, and the properties of materials, but recent focus has also been placed on reactions occurring in aqueous media. Specifically, mechanisms and energetics associated with proton transfer in aqueous system have been investigated for the case of zinc-oxide water interfaces, Na+ or glycine solvated in water, as well as for water clusters. Many frameworks have been proposed to construct machine learning potentials. These differ by the choice of the atomic structure representation and by the regression scheme.
Atomic symmetry functions and Neural Network potentialsHere we employ Behler-Parrinello atomic symmetry functions (ASF) as an input to a feed-forward neural network. The global energy \(E(\mathcal{R})\) is approximated as the sum of local atomic energy contributions \(\epsilon(\mathbf{q}_{ASF}^{i})\)
\[E(\mathcal{R})=\sum_{i}\epsilon(\mathbf{q}_{ASF}^{i})\, \tag{1}\]
The local energy \(\epsilon(\mathbf{q})\) is expressed as a two-layers feed-forward neural network (NN), whose parameters are optimized to minimize the error on an appropriately constructed training set. The feature vector \(\mathbf{q}_{ASF}^{i}\) is built to provide a symmetry-invariant representation of the environment of the \(i\)-th atom, and contains atom-centered symmetry functions (ASF) \(G_{2}\) and \(G_{3}\), defined as in Ref.. The choice of a reasonably complete, yet non redundant set of ASFs is one of the most delicate and time-consuming aspects in the construction of a Behler-Parrinello style MLP. Here we automate this selection using CUR decomposition as described in Ref.. We start from a large set of 192 2-body and 800 3-body symmetry functions per element, with parameters that span evenly distances up to \(\sim 7\) A and angles from 0 to \(360^{\circ}\). We then use CUR decomposition to select the most descriptive of these functions, choosing 64 for each element.
Neural Network simulationsAn MLP trained only on configurations sampled across an _ab initio_ trajectory at standard temperature and pressure would often fail when a thermal fluctuation generates highly distorted structures (_e.g._, atoms in close contact), since it would then enter an extrapolative regime for which no regularizing effect enforces short-distance repulsion. The use of a physics-informed surrogate model as a baseline prevents this possible shortcoming and facilitates the learning of forces and energy predictions.
The baselined NN is however significantly more computationally demanding because of the cost of performing a baseline level computation at each simulation step. An additional \(n\)-fold gain in speed is achieved by integrating the dynamics via a multiple-time-stepping scheme, in which direct neural network predictions are corrected every \(n^{th}\) step by the baselined NN estimate.
Uncertainty QuantificationMachine learning potentials yield accurate out-of-sample predictions only when used in an interpolative regime. To probe whether the neural network predictions take place in an extrapolative regime, it is useful to have a scheme that provides uncertainty estimates. Here we use a scheme based on a committee model, _i.e._, an ensemble of \(M\) neural networks that are trained on different subsets of the training set. The additional cost connected with the use of multiple models is offset by the increased reliability afforded by a model with uncertainty quantification. Furthermore, the use of an ensemble based on the same symmetry functions and architecture could allow a substantial reduction of the overhead, which however we do not exploit due to limitations of the current implementation. The average of the force/energy predictions for a configuration \(\mathcal{X}\), \(\bar{y}(\mathcal{X})=\sum_{i}y^{(i)}(\mathcal{X})/M\) is taken as the best estimate, and is used to drive the dynamics; the standard deviation across the committee, \(\sigma(\mathcal{X})\), is taken as a qualitative measure of the uncertainty. In order to improve the quantitative accuracy of such estimate, the standard deviation is scaled by a factor \(\alpha\), \(\sigma(\mathcal{X})\leftarrow\alpha\sigma(\mathcal{X})\), that is determined by maximizing the log-likelihood of the predictive distribution over a validation set of size N\({}_{val}\):
\[\alpha^{2}=\frac{1}{N_{val}}\sum_{n}\frac{(y_{n}-\bar{y}(\mathcal{X}_{n}))^{2} }{\sigma(\mathcal{X}_{n})^{2}}. \tag{2}\]
### Phase space exploration and characterization
In addition to the computational gain provided by the machine learning potentials, enhanced sampling MD schemes allow for the efficient exploration of complex free energy landscapes, thus providing insights on species relative stability and kinetics in _e.g._, proton transfer reactions. The combination of MLPs and enhanced sampling techniques such as meta-dynamics (MetaD) makes the computational exploration of reactions in explicit solvents possible even over extended time and length scales.
The long time-scale trajectories that can be achieved pose an additional challenge when it comes to identify the most relevant species, molecular motifs and reactive events. Fortunately, data-driven techniques provide a framework that can also be used for unbiased structural characterization. Here we use the sketch-map dimensionality reduction algorithm. Similar to multi-dimensional scaling, sketch-map tries to find a low-dimensional representation of a set of configurations, matching the distances between high-dimensional sets of features that describe each structure, and those between their projections. A non-linear transformation of the distances helps disregarding uninteresting features (_e.g._, thermal fluctuations) and obtaining a map in which recurring structural motifs are clearly identified as separate clusters.
## 4 Computational details
After having summarized the overall methodological framework, we give specific details of the reference electronic structure computations, the training of the MLP and the statistical sampling strategy.
Quantum chemistryDFTB3 with 3OB parameters and the D3H5 correction is used as a baseline, which is robust enough to avoid completely unphysical configurations, but is known to occasionally yield qualitatively incorrect predictions, _e.g._ a planar equilibriumstructure for gas-phase H\({}_{2}\)O\({}_{2}\). DFTB computations are performed with the DFTB+ 18.2 software interfaced with the dynamic driver i-PI. The reference energies and forces used for the training are obtained at the PBE0-D3BJ level as implemented in CP2K 6.1. All elements are described with the TZV2P-MOLOPT basis set with cores represented by the dual-space Goedecker-Teter-Hutter pseudopotentials (GTH PBE). The plane-waves cut-off is set to 700 Ry with a relative cut-off of 70 Ry. All computations employ a Coulomb operator truncated at R = 6 A and the auxiliary density matrix method with a cpFIT3 fitting basis set. We use converged PBE-D3BJ wave functions as the initial guess for the PBE0-D3BJ computations. We evaluate the energies and forces of 3304 carefully selected configurations (see next section).
Training setThe various mixtures used to train the NN comprise 20 phenol molecules, one methanesulfonic acid molecule, up to four water and hydrogen peroxide molecules. A set of 3048 configurations is selected from REMD trajectories performed at the DFTB level (see subsection 4 for further details) using the FPS scheme relying upon Hausdorff distances in the metric described by the symmetry functions used for the neural network. 256 additional structures are extracted from high temperature (600 K), high pressure (above 1000 atm), and PIMD simulations such as to expand the NN training set with a small albeit informative number of highly distorted configurations.
Neural network trainingThe NNs are trained with 2 layers and 22 nodes per layer to predict PBE0-D3BJ level forces and energies. The weights in the NN are optimized via Kalman Filtering routines. We use 100% of the energies and 0.5% of the force components per configuration in the training fraction, with force weights 8-fold larger relative to energy weights. The ASF calculation and NN training is performed with the n2p2 package. Overfitting is avoided by using an early stopping criterion. A total of 400 training iterations is allowed. An ensemble of 5 neural networks is used to compute the baselined force and energy predictions. The uncertainty calibration for the energy predictions carried out by the -baselined NN yields \(\alpha=5.8\).
Molecular DynamicsPreliminary replica exchange molecular dynamics (REMD) simulations are performed with 8 constant-temperature replicas (333 K, 339 K, 347 K, 358 K, 370 K, 384 K, 403 K, 423 K), initiated from the same structure, with randomly sampled momenta from the appropriate distribution at each temperature. Swaps among replicas are attempted every 10 steps. Simulations were run for 15 - 40 ps per replica (in total 2 368 ps).
Two specific mixtures are considered for further study, using the trained NNs to achieve more thorough sampling. One of the mixtures contains 20 phenol molecules and one methanesulfonic acid molecule, the second also contains one molecule of hydrogen peroxide.
MD trajectories are integrated using i-PI with the DFTB+ and LAMMPS drivers for forces and energy computations. Equations of motion are integrated using a multiple time step (MTS) scheme, with an outer time step of 3 fs (involving a DFTB calculation and a NN correction) and an inner one of 0.5 fs (involving a direct MLP fitted to DFT calculations). We improve the stability of the trajectories using a BAOAB splitting, and apply a velocity rescaling thermostat with a frequency of 10 fs, together with a thermostat based on Generalized Langevin Equation (GLE). In order to stabilize MTS trajectories without slowing down diffusion, the GLE is designed so that modes above 100 THz are affected by a friction of 125 ps\({}^{-1}\), while the effective frictions diminishes to 0.01 ps\({}^{-1}\) for frequencies approaching zero. MD sampling is performed for 720 ps using REMD, with the same temperature distribution used for the preliminary DFTB simulations. Nuclear quantum effects (NQEs) are investigated for systems at 363 K using data gathered from independent path integral molecular dynamics trajectories for a total of 720 ps. The PIGLET technique is used to reduce the number of replicas needed for convergence. Consistently with previous simulations of molecular liquids at room temperature, 6 path integral beads were found to be sufficient. Generalized Langevin equation parameters were obtained from the GLE4MD website.
Collective variables and metadynamicsWe focus on the solvation structure around the acid, on the protonation states of the different species, and on the possibility for proton transfer events. For this reason, we use two collective variables (CVs) for the accelerated sampling simulations. The coordination number (CN) of the sulfonyl group. The CN, which has been frequently applied to promote the sampling of proton transfer reactions, is defined here in terms of a smooth switching function of the distances between the oxygen atoms of the acid and hydrogen atoms of the hydroxyl groups (either of phenol or hydrogen peroxide); the CN is normalized so as to avoid attributing the same H atom to multiple species. We employ also a collective variable corresponding to the minimum distance, D, between the hydroxyl hydrogen and the oxygen atoms in the acid, to characterize the H-bonds between the anion and nearby protons. Detailed definitions of the CVs are discussed in section 2.3 in the SI.
MetaD trajectories (with and without incorporation of NQEs) are carried on by depositing gaussians of height 0.8 kJ/mol and width 0.1 and 0.2 in the CV1 and CV2 space respectively every 36 fs. Well-tempering is enforced by using a Gaussian height damping factor \(\Delta T\) such that the ratio \(\frac{\Delta T+T}{T}\) is equal to 3. The parameters of the normalized coordination numbers of the oxygen of the methanesulfonyl hydroxyl group, are set to \(p_{0}=0.9\), \(q_{0}=0.4\), \(n=6\), and \(m=12\), while the minimum distance between the acid oxygen and the hydrogens bonded to any oxygen is computed with \(\delta=24\) (see the SI for a definition of the order parameters). The metadynamics simulations (100 ps for each mixture simulated classically and including NQEs) are performed with PLUMED version 2.5.1 interfaced with i-PI. A representative plumed input is stored in the plumed nest repository at [https://www.plumed-nest.org/eggs/20/008/](https://www.plumed-nest.org/eggs/20/008/).
## 5 Results
### Framework Validation
Before discussing the chemical outcome of the simulations, we present some diagnostics that assess the training performance of the MLP. reports learning curves for the model, _i.e._, the root mean square error (RMSE) for the energy (top) and force (bottom) as a function of the number of training structures fed to the neural networks. The RMSE is computed for a test set of structures randomly extracted from the database described in Section 3.1. The left-most panels show curves for the direct MLP, that reach an accuracy of approximately 4 meV/atom for the energy, and 300 meV/A for the forces, with the largest training set size considered. The power-law decay of the learning curves indicates that the accuracy is data-limited, and that a more accurate model could be obtained, at the price, however, of performing a much larger number of reference computations.
Energy (direct NN in panel (A), baselined NN in panel (B)) and forces (direct NN in panel (C), baselined NN in panel (D)) learning curves (RMSE vs. relative number of training structures, with the full database comprising 3304 structures). All data are taken after 400 iterations of the neural network loss function optimization, except when the training set comprises 99% of the structures, for which results after 600 iterations are reported.
Parity plots for the direct and baselined-learning with 80% training set, are shown in The direct learning has a RMSE of \(\sim\) 4 meV/atom and \(\sim\) 300 meV/A notwithstanding the complexity of the investigated mixture. The baselined model achieves an energy RMSE around 2.5 meV/atom, and a force RMSE below 150 meV/A. The lower RMSE is also associated with faster-decaying tails in the error distribution (see Fig. S1 in the SI), underscoring the better stability in comparison with a direct MLP.
DFTB is more than 1000 times faster than hybrid DFT, for a system of this size, but
Parity plots (True vs. Predicted values) for DFTB+baselined NN and direct NN for energies (direct NN in panel (A), baselined NN in panel (B)) and forces (direct NN in panel (C), baselined NN in panel (D)) predictions. Results are reported by taking the average energy and force predictions of the ensembles of baselined and direct models trained with 80% of the database structures.
The respective timings of PBE0-D3BJ, PBE-D3BJ, DFTB-D3H5, NN baselined on DFTB-D3H5, and NN computations are gathered in Table 1. In order to reduce the cost of a baselined model, we use MTS integration, computing the direct MLP with an inner time step of 0.5 fs, and the DFTB correction at the outer loop. Thanks also to the use of a targeted GLE thermostat, that dampens high frequency modes associated with the OH vibrations, the integration is stable up to an outer step of 4 fs (see conserved quantity and temperature time evolution in Fig. S3 in SI). We choose a more conservative value of 3 fs for our production runs. The stability of the dynamics is verified in different mixtures comprising 20 phenol molecules, 20 phenol molecules and 1 methanesulfonic acid, and 20 phenol molecules and 1 hydrogen peroxide. The RDFs obtained from using the MTS scheme were also benchmarked against trajectories obtained by integrating the equations of motion only with the baselined-NN force-field. As reported in Fig. S4 in the SI, the RDFs using both approaches are nearly identical. The length of the outer step could be further increased by improving the relative accuracy of the direct NN with respect to the baseline corrected potential. This comes at the computational cost of refining the direct NN training. In addition, such improvement may not be possible _ad libitum_ if the direct NN learning saturates with the number of iterations of the NN optimization or the number of training structures. Overall, the use of a NN MTS force field leads to a \(10^{4}\) gain in speed with respect to PBE0-D3BJ.
\begin{table}
\begin{tabular}{c c c c c c} PBE0 & PBE & DFTB-D3H5 & DFTB & MTS NN & NN \\ -D3BJ & -D3BJ & + NN comm. & -D3H5 & comm. & single \\ \hline \hline
67\(\times 10^{3}\) & 21\(\times 10^{3}\) & 16 & 12 & 6.7 & 0.8 \\ \end{tabular}
\end{table}
Table 1: CPU time (core seconds) required to advance MD simulations by 0.5 fs. DFT timings (PBE0-D3BJ, PBE-D3BJ) are determined based on computations performed on a single computing node with two 14 cores Intel Broadwell processors running at 2.6 GHz. DFTB-D3H5 and neural-network force-fields timings are computed based on single-core execution on the same hardware. The “NN comm.” label indicates a prediction obtained with a committee of five members, as opposed to “NN single” that indicates results from a single member. The MTS cost is computed based on a \(\sim 40\) CPU s timing for a 3 fs outer time step, consisting of 1 DFTB + 35 NN computations (5 DFTB corrections + 30 “direct” NNs).
### Uncertainty quantification
The use of an ensemble of neural networks for force and energies allows to estimate an uncertainty on the prediction. shows the evolution of the 5 energy predictions for the members of the MLP ensemble (rescaled according to \(\alpha\) as discussed in Ref. 14), during one of the classical MetaD trajectories discussed in Section 3.1. The spread in the predictions is consistent with a mean estimated uncertainty around 2.5 meV/atom, which is comparable to the accuracy estimated on the validation set. The largest uncertainty observed along the trajectory is of the order of 7 meV/atom, corresponding to a momentary fluctuation rather than to a systematically larger error along an extended section of the trajectory. The inset in the figure shows a conditional average of the uncertainty as a function of the coordination number of the sulfonyl oxygen atoms, that is used to sample the deprotonation reaction. On average, the uncertainty is larger for configurations along the proton dissociation pathway, which correspond to higher free energy values and are poorly represented in the training set. While the predicted error is still very low, \(\approx 3\) meV/atom, its magnitude serves to identify under-sampled regions, and could be used in an active learning setting. Thanks to the rather extensive preliminary REMD sampling, however, we did not encounter large-error regions of configuration space that required retraining.
### Ch\({}_{3}\)So\({}_{3}\)H and H\({}_{2}\)O\({}_{2}\) in phenol
The data from the REMD trajectories (720 ps for each replica) are analyzed addressing three specific questions:
* What is the probability for the acid to be deprotonated? Are proton transfer reactions likely to occur?
* What are the characteristic features in the H-bond patterns involving the acid, phenol, and hydrogen peroxide?
* What are the structural signatures associated with apolar interactions, _i.e._, CH..\(\pi\)?Oxygen environments in a complex mixtureThe different species in the mixture are identified and analyzed with a sketch-map representation that relies on a set of features describing the atomic environment of the oxygen atoms, making the distinction between the different moieties and their possible protonation states (for further details see Section 2.3 in the SI). represents the chemical environment of oxygen atoms in a phenol solution containing one methanesulfonic acid and one hydrogen peroxide molecule at the DFTB-D3H5 and NN-corrected levels. With the DFT-quality MLP, the methanesulfonic acid does not promote the protonation of any other species and remains protonated along the REMD simulation. This is reflected in the sketch-map representation by the presence of four clusters shown in One cluster regroups the phenol oxygen atoms, one is formed by the H\({}_{2}\)O\({}_{2}\) oxygens, whereas the last two corresponds to the OH and sulfonyl group of the acid.
Time evolution of the potential energy during a metadynamics simulation sampling the protonation states of CH\({}_{3}\)SO\({}_{3}\)H. For each member in the ensemble of neural networks we report the corresponding energy prediction in colored lines. Their average, which is used to propagate the dynamics, is shown in black. The inset shows the mean uncertainty, conditionally averaged by binning over different values of the collective variable CN\({}^{O}\).
Such a discrepancy highlights the necessity of using higher level (_e.g._, PBE0-D3BJ) energies and forces for capturing the correct qualitative protonation state of the acid.
Deprotonation free energyA quantitative estimate of the pKa of methanesulfonic acid with respect to the protonated phenol and hydrogen peroxide can be extracted from the MetaD sampling at 363 K. A and B show the reconstructed free energy profile for the deprotonation of CH3SO3H in MetaD simulations. The free energy minimum in the presence of H2O2 corresponds to the protonated acid, which is about 8 kJ/mol lower than the other species. In pure phenol, the free energy difference between the neutral and deprotonated acid is almost twice this value. This comparison highlights that H2O2 facilitates the deprotonation of CH3SO3H solvated in phenol, _i.e._, that the relative acidity of the protonated species is PhOH2+\(>\)H3O2+\(>\)CH3SO3H. Accounting for NQEs by performing path integral MetaD simulations (panels C and D) stabilizes the deprotonated species by up to 2 kJ/mol in both mixtures, which corresponds roughly to a change of 0.5 pH units.
Sketch-map displaying species in the mixture containing 1 H2O2 for classical REMD sampling carried out using NN-corrected (A) and DFTB-D3H5 (B) energy and force predictions. 358-K replica sampled for 52 ps (DFTB-D3H5) and 240 ps (NN) are used. Points represent different oxygen atoms. The color code corresponds to the sum of the reciprocal distances of hydrogen atoms within 1.5 Å (DFTB) and 2 Å (NN) radius. Details on the description of the atomic environments are provided in the SI (Section 2.3).
NQEs would be important to determine quantitatively the pKa values, but they are not as crucial as the use of a high-quality _ab initio_ MLP.
Characterization of the H-bond networkAll the species present in the mixture are potential hydrogen bond donors or acceptors but some bonding patterns are more frequent than others. The focus is especially placed on the H-bond network involving the acid. In its protonated form, the acid carries two types of oxygen atoms - the sulfonyl oxygens, which are hydrogen bond acceptors, and the hydroxyl group which can both donate and accept H-bonds.
Methanesulfonic acid acts both as a HB acceptor and as a donor, with the sulfonyl oxygens
2D Free energy landscape projections without (upper panels) and with NQEs (lower panels) for 1 CH\({}_{3}\)SO\({}_{3}\)H in phenol (A, C) and 1 CH\({}_{3}\)SO\({}_{3}\)H and 1 H\({}_{2}\)O\({}_{2}\) in phenol (B, D). The minimum at corresponds to CH\({}_{3}\)SO\({}_{3}\)H, the deprotonated acid is mapped around CN\({}^{O}\)\(\sim 0\).
Panel (a) in shows that NQEs strengthen the HB donated by the acid OH, increasing slightly the height of the peak and shifting it towards smaller distances, and weaken the HB accepted by the sulfonyl oxygens. It has been observed consistently that NQEs tend to strengthen strong HBs and weaken weak HBs, suggesting that the HBs donated by the acid are stronger than those it accepts. Panel (b) shows that in classical REMD simulations H\({}_{2}\)O\({}_{2}\) binds strongly to the acid hydroxyl group, remaining within \(\approx\) 6A for the whole trajectory. H\({}_{2}\)O\({}_{2}\) competes with phenol for binding, reducing the height of the corresponding peak by \(\approx\) 15%. The CH\({}_{3}\)SO\({}_{2}\)OH-H\({}_{2}\)O\({}_{2}\) RDFs show a split first-neighbor peak, demonstrating that the acid binds preferentially to one of the two O atoms in the hydrogen peroxide molecule, with the second O staying further apart. H\({}_{2}\)O\({}_{2}\) can also donate HBs to the sulfonyl groups, but remains at a somewhat larger distance than phenol, and the \(g(r)\) between PhOH and the sulfonyl oxygens remain largely unchanged. Panel (c), that shows the same mixture as in panel (b), modeled by PIMD simulations, demonstrates that NQEs have a substantial impact. The binding of H\({}_{2}\)O\({}_{2}\) with the acid hydroxyl group is greatly enhanced, with the phenol acceptor peak almost completely suppressed. At the same time, one observes a substantial increase of the peak that corresponds to phenol donating a HB to the sulfonyl oxygens, indicating that the combined effect of the presence of H\({}_{2}\)O\({}_{2}\) and of nuclear quantum effects changes the structure of the solvation shell around CH\({}_{3}\)SO\({}_{3}\)H.
The hydrogen bond patterns discussed in the previous paragraphs are visualized more explicitly by the 3D density distributions reported in The acid hydroxyl group is a strong donor towards both phenol and H\({}_{2}\)O\({}_{2}\) with a stronger affinity towards the latter. Specifically, the density distribution of H\({}_{2}\)O\({}_{2}\) oxygen atoms around the hydroxyl group is high and directional (right Fig. 9). A similar localized region is observed for the phenol distribution around the hydroxyl group both in pure phenol and in the phenol/H\({}_{2}\)O\({}_{2}\) mixture (left and middle Fig. 9). Hydrogen peroxide binds almost exclusively by accepting a HB from the acid OH. Although binding occurs preferentially to one of the two oxygens, the distribution is broad, and one also observes bifurcated HBs, in which the acid hydroxyl group is shared between the two oxygen atoms of hydrogen peroxide (see illustrative snapshots in Fig. S7). The trends observed in the O-O \(g(r)\) sampled at 358K are qualitatively consistent with those observed over the 333K-423K temperature range (see Figure S8-S13), which indicates that temperature does not affect substantially the solvation environment of CH\({}_{3}\)SO\({}_{3}\)H. Nuclear quantum effects have a relatively minor effect on the distribution in pure phenol, but lead to larger changes in the presence of a peroxide molecule. When including NQEs, the latter binds very strongly to the methanesulfonic acid OH group, displacing completely phenol and triggering a rearrangement of the solvation shell, in which phenol binds more strongly to the sulfonyl oxygen atoms. It should be stressed that quantitative convergence of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic of the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the solvation shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to the thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic shell to thermodynamic of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic equilibrium of thermodynamic shell to thermodynamic thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of thermodynamic shell to thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of thermodynamic shell to thermodynamic thermodynamic thermodynamic equilibrium of the thermodynamic shell to thermodynamic thermodynamic equilibrium of thermodynamic shell to thermodynamic shell the 3D distribution function is not trivial even with several hundred ps of simulations. The distributions in are obtained from several independent runs, all showing qualitatively consistent trends.
Density distribution of oxygen (red contour) and hydrogen (grey contour) atoms around the acid: Ph**OH** in pure phenol (left) and in a mixture of phenol and one H\({}_{2}\)O\({}_{2}\) (middle); **H\({}_{2}\)O\({}_{2}\)** in the same mixture of phenol and hydrogen peroxide (right). Panel A corresponds to the classical REMD simulation, while panel B corresponds to path integral MD, and incorporate nuclear quantum effects. The density isocontours correspond to density values of 0.004 (transparent) and 0.012 (opaque). The density histogram is accumulated in a reference frame defined by the SOH atoms: given the flexibility of methanesulfonic acid, the sulfonyl oxygens are not fixed in the ideal positions (see also Fig. S4), contributing to the larger spread of the density in the direction of the SO\({}_{2}\) group compared to that in the direction of the acid hydroxyl, which defines the orientation of the axes. Distributions computed in a reference frame defined by the sulfonyl oxygens are reported in the SI, Fig. S5 and S6. The interested reader is referred to the Supplementary Information Section 5 for a detailed discussion on the degree of localization of the atoms in the acid and the density distribution calculation.
CH-\(\pi\) interactionsApart from the rich hydrogen bond network, the methanesulfonic acid may form apolar interactions with the aromatic phenol ring through its methyl group. The relevance of CH-\(\pi\) interactions between the acid methyl group and phenol ring can be further analyzed from the density distribution given in Fig. S14. The fact that the distribution of the phenol ring around the acid is very delocalized with no evident preferential arrangement close to the methyl group suggests weak interactions without directionality. This is observed both in classical simulations as well as in PIMD. Unlike the H-bond network, these apolar interactions are not associated with a clear structural CH-\(\pi\) signature influencing the mixture.
## 6 Conclusions
This work demonstrates a framework that combines several traditional and data-driven atomistic modeling techniques to enable the simulation of complex, multi-component mixtures, taking as example the study of the solvation and acidity of CH\({}_{3}\)SO\({}_{3}\)H and H\({}_{2}\)O\({}_{2}\) in phenol.
We use a neural-network machine-learning interatomic potential to reproduce accurate reference energetics based on dispersion-corrected hybrid DFT calculations. In order to obtain a robust model, we use semiempirical DFTB-D3H5 energies and forces as a baseline, achieving close-to-DFT accuracy (RMSE = 2 meV/Atom and 146 meV/ A for the energies and forces respectively) while being 4000 times less expensive for the simulations we perform here. Using a multiple time step integrator with a direct MLP yields a further 3-fold speedup with no loss in accuracy. As a final component in our framework, we build a committee model to obtain accurate uncertainty estimation, that we use to monitor the error during simulations and for active learning. The robustness of our framework, combining MLP, accelerated sampling techniques, and on-the-fly uncertainty estimates, in turn, demonstrates a great potential to tackle reactions in condensed phase environments such as phenol hydroxylation.
We use this framework to study the solvation and the deprotonation of methanesulfonicacid in phenol, with and without the presence of a hydrogen peroxide molecule, using accelerated sampling techniques such as replica exchange molecular dynamics and metadynamics, and including quantum nuclear fluctuations using the PIGLET technique. While the modeling of phenol hydroxylation catalyzed by CH\({}_{3}\)SO3H is in fact beyond the scope of this manuscript, the characterization of the acidity and solvation of a potential acid catalyst is a necessary prior step in the study of this reaction.
CH\({}_{3}\)SO\({}_{3}\)H is found to be less acidic than protonated phenol and hydrogen peroxide, and to remain in its protonated state in unbiased simulations. The behavior is qualitatively different with respect to DFTB simulations, in which methanesulfonic acid readily loses its proton, which underscores the need for a MLP to reach hybrid DFT accuracy.
An analysis of the solvation environment of CH\({}_{3}\)SO\({}_{3}\)H reveals that the methanesulfonic hydroxyl group acts as a strong hydrogen bond donor toward phenol, while the sulfonyl oxygen atoms act as acceptors. Hydrogen peroxide binds strongly to the acid OH, competing with phenol and substituting it almost completely in simulations that include nuclear quantum fluctuations. Contrary to phenol, H\({}_{2}\)O\({}_{2}\) interacts very weakly with the sulfonyl oxygen atoms. The highly structured environment, with directional hydrogen bonds and competition between phenol and hydrogen peroxide, underscore the need for an explicit, condensed-phase treatment of the system. These insights into the acid and H\({}_{2}\)O\({}_{2}\) solvation shells also suggest that the reaction will never involve an isolated H\({}_{3}\)O\({}_{2}\)\({}^{+}\) but rather an acid-H\({}_{2}\)O\({}_{2}\) complex, which could potentially control the regioselectivity. In addition to the hydrogen bond network, we also evaluated the importance of CH-\(\pi\) interactions between the acid methyl group and the phenol ring but did not identify any relevant pattern influencing the mixture.
These simulations demonstrate that in order to obtain quantitative insights into the behavior of complex mixtures it is necessary to describe explicit solvation, achieve hybrid DFT accuracy using state-of-the-art MLPs, sample thoroughly thermal and quantum fluctuations of the nuclei, and analyze trajectories with automatic data-driven techniques. Wehave brought together all these ingredients into a robust, flexible framework that is readily applicable to the study of chemical reactions in complex condensed-phase environments.
The authors thank Venkat Kapil, Giulio Imbalzano, Federico Giberti, Piero Gasparotto and Raimon Fabregat for discussions. KR and VJ were supported by an industrial grant with Solvay. This project was initiated within the framework of the NCCR MARVEL, funded by the Swiss National Science Foundation (SNSF).
The following files are available free of charge. Supplementary Information (SI) available: Neural network training and simulation set ups - Technical details - Error distribution of baselined training - MTS benchmarking - Density distribution computations - Temperature dependence of oxygen-oxygen pair correlation functions from REMD simulation - Hydrogen bonding network of the acid with hydrogen peroxide - NQEs in apolar interaction. See DOI: 10.1039/cXCP00000x/
|
10.48550/arXiv.2006.12597
|
Simulating solvation and acidity in complex mixtures with first-principles accuracy: the case of CH$_3$SO$_3$H and H$_2$O$_2$ in phenol
|
Kevin Rossi, Veronika Juraskova, Raphael Wischert, Laurent Garel, Clemence Corminboeuf, Michele Ceriotti
| 706
|
10.48550_arXiv.1305.7451
|
###### Abstract
A new method that accurately describes strongly correlated states and captures dynamical correlation is presented. It is derived as a modification of coupled-cluster theory with single and double excitations (CCSD) through consideration of particle distinguishability between dissociated fragments, whilst retaining the key desirable properties of particle-hole symmetry, size extensivity, invariance to rotations within the occupied and virtual spaces, and exactness for two-electron subsystems. The resulting method called the distinguishable cluster approximation, smoothly dissociates difficult cases such as the nitrogen molecule, with the modest \(N^{6}\) computational cost of CCSD. Even for molecules near their equilibrium geometries, the new model outperforms CCSD. It also accurately describes the massively correlated states encountered when dissociating hydrogen lattices, a proxy for the metal-insulator transition, and the fully dissociated system is treated exactly.
## I Introduction
Coupled-cluster theory is the most successful method used for the treatment of many-body correlations in quantum chemistry.
\[|\Psi\rangle=e^{\hat{T}}|0\rangle \tag{1}\]
Even when the expansion of \(\hat{T}\) is truncated, coupled-cluster theory retains important properties of the exact solutions to the Schrodinger equation, primarily size extensivity, and includes all contributions from all orders of perturbation theory up to the given excitation rank. Coupled-cluster theory provides a systematically improvable hierarchy of polynomially scaling methods that approach full configuration interaction (or exact diagonalization).
In quantum chemistry, coupled-cluster theory is typically truncated at the doubles level (to give the so-called CCSD method), and the effect of triple excitations is handled perturbatively in the CCSD(T) method. These truncated coupled-cluster theories achieve extremely high accuracy for a very wide range of problems, but fail to describe the strongly correlated states encountered in molecular dissociation or other highly degenerate situations. For example, CCSD famously fails to repair the incorrect physics of a restricted Hartree-Fock treatment of the dissociating N\({}_{2}\) molecule, and the perturbative triples correction only makes matters worse.
An enormous range of approaches have been devised to address strong correlation in quantum chemistry. Those that build on coupled-cluster theory do so by using a more flexible reference function; by symmetry breaking; by including leading terms from variational coupled-cluster theory; through renormalization; or through combining valence-bond theory and coupled-cluster.
One could assume that CCSD is the most accurate theory possible that includes only single and double excitations from a single reference determinant, because it includes all possible diagrammatic contributions to the correlation energy within that constraint. But consider some quantity in the full theory \(x=y+z\) that is approximated in CCSD as \(x_{\mathrm{SD}}=y_{\mathrm{SD}}+z_{\mathrm{SD}}\). Now if \(z\approx 0\) (for example through cancellation of \(z_{\mathrm{SD}}\) with higher-order terms), it could be better to exclude the \(z_{\mathrm{SD}}\) terms altogether and use the approximation \(x_{\mathrm{SD}}=y_{\mathrm{SD}}\) instead.
In this spirit, there is a growing body of evidence that improved doubles-based schemes can be produced by leaving certain terms out of the amplitude equations, like various versions of the coupled-electron pair approximation, the \(n\)CC hierarchy and pCCSD.
## II Antisymmetry of the wavefunction
The traditional description of antisymmetry in quantum mechanics states that
\[\Psi(\mathbf{r}_{1},\mathbf{r}_{2})=-\Psi(\mathbf{r}_{2},\mathbf{r}_{1})\]
The soundness of this statement has been called into question, because the notion that the classical configuration \((\mathbf{r}_{1},\mathbf{r}_{2})\) is different from \((\mathbf{r}_{2},\mathbf{r}_{1})\) distinguishes the particles from the outset. This subtlety appears to have motivated the analysis by Leinaas and Myrheim of the classical configuration space of indistinguishable particles, famously leading to the prediction of anyon statistics.
A concrete meaning can be attached to the concept of particle interchange, as being the result of closed pathsin the classical configuration space for indistinguishable particles, in which the points \((\mathbf{r}_{1},\mathbf{r}_{2})\) and \((\mathbf{r}_{2},\mathbf{r}_{1})\) are identified. For the present work, the important result from Leinaas and Myrheim is that although the topology of the configuration space for multiple identical particles is globally different from the Euclidean product space, it is locally isometric in regions where particles are not close. To express it another way, there are no physical consequences of the indistinguishability of particles during processes in which the particles do not become close.
The problems with truncated coupled-cluster theories are encountered when molecules are fragmented into separate pieces. All possible exchange processes between the fragments are considered, because antisymmetry is cemented into the structure of the theory right from the beginning, through the use of second quantization. In exact theory it makes no difference whether these exchange processes are considered or not (but of course the exchange processes _within_ the fragments must be treated exactly). Here we speculate that an incomplete treatment of exchange processes, whose effect must ultimately cancel out in an exact treatment, could be the source of the pathological failure to describe dissociation in coupled-cluster theory.
## III Theory
We have motivated removal of exchange terms between isolated fragments. There is no systematic way to do so in an orbital-invariant theory, so as a substitute we instead investigate removal of exchange terms between the 2-particle clusters formed through the application of the \(\hat{T}_{2}\) cluster operator, whilst retaining all terms arising from particle indistinguishability _within_ clusters.
We therefore turn our attention to the terms in the amplitude equation that couple together two \(T_{2}\) amplitudes, shown in in the form of nonantisymmetrized Goldstone diagrams. Contrary to the usual convention, our interaction lines correspond to bare electron repulsion integrals, rather than antisymmetrized integrals.
Consider diagrams A and A\({}^{\prime}\). In A, particle-hole pairs formed as a result of two separate double excitation processes interact through the Hamiltonian. In A\({}^{\prime}\) there is an additional process in which the particles (or equivalently holes) from each double excitation are exchanged. In an infinitely separated system the only physically relevant double excitation processes are those where the two particles and two holes are associated with a single fragment. But the very cases where CCSD fails to describe dissociation are those where the restricted Hartree-Fock orbitals cannot be localized on to the fragments.
We hypothesize that the approximate description of physically irrelevant exchange processes could be the source of the poor behaviour of CCSD for dissociation.
Nevertheless _within_ clusters antisymmetry is handled correctly (the amplitudes have the correct antisymmetry); it is only between clusters that parity violation is permitted.
We now proceed by restoring the other desirable properties of CCD in the new theory. The quadratic portion of the CCD \(T_{2}\) amplitude equation reads A+A\({}^{\prime}\)+B+C+D. The amplitude equation can be modified without compromising exactness for two-electron systems by adding \(\alpha^{\prime}(2\textsf{A}^{\prime}+\textsf{C})+\beta(2\textsf{B}+\textsf{D})\) for any parameters \(\alpha^{\prime}\) and \(\beta\). Diagram A\({}^{\prime}\) is removed by setting \(\alpha^{\prime}=-1/2\), and particle-hole symmetry then requires \(\beta=-1/2\), leading to the modified quadratic contribution A + C/2 + D/2. This defines the distinguishable cluster approximation with double excitations (DCD).
The diagrams in can be interpreted as a Coulombic interaction between fluctuations in separate clusters (A), exchange interactions between clusters (A\({}^{\prime}\) and B), and scattering of a particle (or hole) of one cluster with a Fock-like potential arising from fluctuations in the other (C and D). DCD can be viewed as a method in which each two-particle cluster is treated exactly in an embedding produced by other fluctuations in the system, but with neglect of exchange processes between the two-particle cluster and the environment.
To avoid the complications introduced by a full consideration of single excitations we use a Brueckner formulation that rotates the occupied and virtual spaces such that the \(T_{1}\) amplitudes vanish. The BDCD program is implemented as a variation of the BCCD program in Molpro.
Nonantisymmetrized Goldstone diagrams contributing to the CCD amplitude equation that are quadratic in the amplitudes. The solid interaction lines denote a \(T_{2}\) amplitude; the dashed lines a Coulombic interaction from the Hamiltonian; and the thin lines particles (downward arrows) or holes (upward arrows) in the spinorbital representation.
## IV Results
We first consider the chemically relevant situation where a low-cost correlation method that can handle strongly correlated states is of immediate importance: molecular dissociation. Calculations were performed using the cc-pVDZ gaussian basis set and were based on a restricted Hartree-Fock reference. In the BDCD results for N\({}_{2}\) and H\({}_{2}\)O dissociation are compared with other coupled cluster calculations and benchmark data from the Davidson-corrected internally contracted multireference configuration interaction method (denoted MRCI+Q). Whereas CCSD, CCSD(T) and CCSDT show increasingly poor behaviour in the strongly-correlated region, BDCD smoothly and systematically dissociates the molecules, resulting in the physically correct description of zero net force between the dissociated fragments. This is despite the fundamentally flawed description in the closed-shell restricted reference state.
The modification to improve description of a fragmented system could well have made things worse in the bonded region, but in fact this turns out not to be the case. On the contrary, the shape of the curve around the minimum is considerably better in BDCD than in BCCD (or CCSD). For N\({}_{2}\), the difference in harmonic frequency relative to MRCI+Q is 99 cm\({}^{-1}\) for BCCD but only 13 cm\({}^{-1}\) for BDCD, and the discrepancy in the equilibrium bond length falls from 0.90 pm to 0.16 pm. Moreover reaction energies are well described by BDCD. For a test set of small-molecule reactions calculated using the aug-cc-pVTZ basis set, the root mean squared deviation compared to CCSD(T) is reduced from 7.6 kJ/mol for BCCD to 5.0 kJ/mol for BDCD.
We next consider hydrogen systems, which are important model problems that connect the challenges of strong correlation in the quantum chemistry of dissociating molecules with those in the description of the metal-insulator transition in condensed-matter physics.
First we consider H\({}_{4}\) using the cc-pVQZ basis set. The four atoms sit on the circumference of a circle of radius 1.738 A, with opposite pairs connected by diameters subtending an angle \(\theta\) at the centre. For \(\theta<90^{\circ}\) the electronic structure is dominated by a reference state which pairs the hydrogen atoms into molecules one way, and for \(\theta>90^{\circ}\) the other. The MRCI+Q data can be regarded as essentially exact, and the energy goes through a maximum at \(\theta=90^{\circ}\) with zero derivative. Neither CCSD nor BDCD reproduce this feature exactly, both producing two different energies depending on the choice of reference state, but in BDCD the maximum is much more accurately captured, with the energy at \(\theta=90^{\circ}\) being reproduced almost exactly.
Next we consider dissociation of the H\({}_{6}\) ring and linear H\({}_{50}\) using the STO-6G basis set. In the former case the exact result (from full configuration interaction) is available. In the latter case exact diagonalization is unfeasible -- the Hilbert space has over \(10^{28}\) dimensions -- but highly accurate results can be found using the density matrix renormalization group (DMRG) approach, owing to the linear structure. For H\({}_{6}\) it can be seen that BCCD (approximately equivalent to CCSD) rapidly diverges from the exact result, whereas BDCD closely follows the exact curve at all distances, reproducing it exactly in the long range. For H\({}_{50}\) agreement with DMRG is also spectacular, and to emphasize the extent of the error that BDCD has to correct we also present the reference energy from restricted Hartree-Fock theory.
A significant challenge for many-body theory is the description of the onset of strong correlation during the metal-insulator transition; this has often been investigated using the model system of cubic lattices of hydrogen atoms. In the equilibrium geometry the electronic structure is relatively benign, but as the nearest-neighbour distance is increased enormous degeneracy is introduced until, at dissociation, the wavefunction is composed of around \(10^{18}\) equally weighted determinants, from a Hilbert space of \(10^{36}\) dimensions.
For this system, very few approximations work even qualitatively; the restricted Hartree-Fock energy rises rapidly to spuriously high values; perturbative methods produce infinities; and coupled-cluster calculations appear impossible to converge.
Potential energy curves for N\({}_{2}\) dissociation (top); symmetric double dissociation of water (bottom). Shown are BDCD and, for comparison, a variety of alternative methods. In each case the hierarchy CCSD, CCSD(T), CCSDT breaks down in the strongly correlated region, but BDCD is systematically smooth and well behaved.
## V Conclusion
Coupled-cluster theory is an exceptionally powerful framework for describing many-body correlation effects with polynomial cost. But its structure -- forming a wavefunction through excitations from a single reference Slater determinant -- seems to preclude treatment of strongly correlated states, in which potentially vast numbers of determinants contribute with comparable weights.
Motivated by speculations about the role of antisymmetry in the description of dissociated fragments, we have argued for a simple modification of the CCD amplitude equation to provide a theory that neglects certain exchange terms that must cancel out in an exact description of the dissociated system. By doing so, we also threw out terms important for the description of correlations within the fragments themselves, or in a non-dissociated molecule; but it transpired that all of the desirable properties of CCD could be restored by other modifications. Thus BDCD has \(N^{6}\) cost, is extensive and invariant to orbital transformations; is exact for two-electron systems, and treats particles and holes symmetrically.
The resulting method does indeed describe dissociation of molecules. But perhaps more surprisingly it also improves the energetics in the bonding region, and improve energy differences such as reaction energies. Most astonishing of all, though, is the fact that BDCD smoothly dissociates hydrogen lattices producing exactly the correct energy at infinite separation, a feat that should apparently only be possible using a wavefunction of immense and computationally intractable complexity.
Our attention is now focused on the derivation of this theory in a more rigorous theoretical framework, with the aim of treating singles, doubles and higher excitations in one systematically improvable hierarchy of distinguishable cluster theories.
|
10.48550/arXiv.1305.7451
|
The distinguishable cluster approximation
|
Daniel Kats, Frederick R. Manby
| 6,496
|
10.48550_arXiv.2404.09883
|
## I Introduction
Forster resonance energy transfer (FRET) experiments are commonly used in biochemistry and bio-physics to measure distances at the molecular level to resolve conformational states. This is done by exploiting the physical mechanism of FRET wherein energy is transferred non-radiatively between a _donor_ and an _acceptor_ fluorophore due to electric dipole coupling. For simplicity, we refer to these fluorophores as dyes. In FRET, the rate of energy transfer depends on the distance between the dipoles to the sixth power, as well as an orientational factor and several other time-independent factors. Due to this strong dependence on distance, FRET is often used as a _molecular ruler_. However, due to the many distance-related interpretations in FRET and potential pitfalls, assignments of FRET derived distances to specific structural properties is not trivial. For example, measuring distances within single molecules (smFRET) is a well-established concept in structural biology. Hence, accurately understanding FRET measurements is crucial for interpreting experimental data and forming concrete hypotheses about the conformational dynamics of the molecule of interest. This understanding is essential for developing accurate biosensors, elucidating signal transduction pathways, studying biomolecular interactions, advancing drug development, and creating fluorescence spectroscopic toolkits.
FRET can be measured via photon counting or time-resolved methods. Previously, we and others demonstrated that the joint distribution between FRET efficiency--the proportion of excited donor molecules transferring energy to acceptor molecules--and fluorescence lifetime--the time a molecule remains in the excited state--provides extensive information on the structural dynamics of the underlying molecule. Thus, we introduced the term "dynamic shift" to account for deviations from the ideal Forster relationship, with multiple example on how to the dynamic shift concept helped understand the structural dynamics of biomolecules. Due to thermal noise, the dye position fluctuates randomly, altering the joint FRET-lifetime distribution. Therefore, it is crucial to understand how these stochastic fluctuations influence the resulting FRET measurements. A comprehensive understanding of the dynamic shift induced by dye dynamics is needed to achieve accurate estimates of structural dynamics.
Current dye models in the literature lead to significant uncertainties in FRET measurements. Moreover, the available models are either simplistic or require high computational costs. For example, all-atom molecular dynamics (MD) simulations are often conducted on a sample-by-sample basis to calculate the accessible volume of the dyes, thereby estimating the range of motion of the dye and the labeled molecule while providing a method for uncertainty quantification in the FRET-lifetime distribution. However, a key critique of this method is that MD simulations do not last for the entire sampling process of typical smFRET experiments. Additionally, MD simulations are designed to approximate the equilibrium distribution of the dye motion, ignoring any time heterogeneity in the energy transfer process itself. Consequently, the full range of dye motion during a sample period is not captured.
Here, we present a semi-analytic model for fluorescent dye motion that addresses the following questions: First, will an isotropic Gaussian process provide a correct model? Second, what is the impact of linker length? Third, what is the role of dipole motion in FRET measurements? We answer these questions in the context of simulated smFRET experiments. Our results challenge the previous assumption that the dynamic shift induced by the dyes is solely due to the accessible volume of the dye (i.e., translational motion). This work demonstrates that the dynamic shift actually depends on the full state dynamics of the dyes, including translational and orientational movements. Therefore, to reduce uncertainty in FRET-derived distances or determine the structural dynamics of biomolecules, we can decouple the individual dye dynamics from the dynamics of the molecule under study by their dynamic shift signature.
### Confocal smFRET
A time-resolved confocal smFRET experiment uses the FRET phenomena by attaching fluorescent dyes to a molecule of interest by continuously exciting the donor molecule and measuring the resulting emitted light; one may estimate the FRET efficiency and, from that, estimate the distance between the molecules.
The estimation of FRET efficiency is done in two ways deemed _intensity_ based FRET and _lifetime_ based FRET. Both methods may be used in time-resolved confocal smFRET, and indeed, the subject of this paper concerns the relationship between the two.
The experiment is carried out by diluting the sample of interest so that, on average, less than one molecule of interest lies within the confocal volume of the microscope. By freely diffusing through the confocal volume, the attached donor dye becomes repeatedly excited, and the resulting fluorescence is measured. The random amount of time spent within the confocal volume is termed a burst time due to the burst of photons seen in the measurement process. Moreover, by recording the time between the laser pulse and measurement, one obtains the _lifetime_ measurement. The lifetime measures how long the molecule remains excited, a quantity that will change in the presence of other molecules.
A _burst time_ is defined as the amount of time the molecule diffuses through the confocal volume. During this time, the inter-photon arrival time decreases sharply from the background level. Through this stochastic time frame, measurements are made and one obtains a sample of lifetimes and photons by which the FRET efficiency and lifetime distribution may be estimated. By collecting multiple burst samples, one obtains the joint FRET-lifetime distribution. During each burst, the donor molecule repeatedly enters an excited state wherein it may fluoresce, transfer energy via the FRET mechanism, or relax due to some other relaxation pathway. We refer to the event of excitation and the following process dictating the observation as an _excitation event_. In simple terms, a burst time will give a data point in the FRET-lifetime distribution and represents _multiple_ photons. Whereas an excitation event gives the color and lifetime of _one_ photon.
### FRET Model
Consider two completely static dyes with normalized dipole moments \(\hat{\mu}_{A}\in S^{2}\) and \(\hat{\mu}_{D}\in S^{2}\) for the acceptor and donor, respectively. Further, let the inter-dye displacement vector be \(r\in\mathbb{R}^{3}\). The energy transfer rate is defined in Eq.,
\[k_{ET}(r)=k_{D}\bigg{(}\frac{R_{0}}{\|r\|}\bigg{)}^{6}, \tag{1}\]
Note that the energy transfer rate increases steeply as the distance decreases and inversely as the distance increases. However, no matter the distance, no energy transfer can happen if \(\hat{\mu}_{D},\hat{\mu}_{A}\), and \(\hat{r}\) are mutually orthogonal. The Forster radius can be written as \(R_{0}^{6}(t)=C\kappa^{2}(t)\) where \(C\) is a constant depending on the environment surrounding the dye.
\[\kappa(t)=(\hat{\mu}_{D}(t)\cdot\hat{\mu}_{A}(t))-3(\hat{r}\cdot\hat{\mu}_{D}( t))(\hat{r}\cdot\hat{\mu}_{A}(t)) \tag{2}\]
Since the dipole moments are known to reorient on timescales faster than the energy exchange rate, kappa square is treated as time-dependent. This is in contrast to previous models wherein the dipole moment is chosen from the equilibrium distribution of the rotational diffusion. In this model, the initial distribution of the dipoles is chosen according to the equilibrium distribution, but rotational processes evolve during the energy transfer.
\[\mathcal{E}(T)=\frac{\int_{0}^{T}k_{ET}(s)ds}{\int_{0}^{T}k_{ET}(s)ds+k_{D}T}. \tag{3}\]In this way, the FRET efficiency process, \(\mathcal{E}(t)\), is non-Markovian. It is important to note that each excitation event's fluorescence process will still be Markovian. As noted in, the inter-arrival time for the photon count process need not be exponentially distributed. Therefore, the photon arrival process cannot be seen as a time-homogeneous Poisson process in contrast to previous common assumptions. Depending on the rate at which the dyes reorient, each vector may be treated as uniformly distributed on the unit sphere or a cone. In this case the average value of \(\kappa^{2}\) is given by \(\frac{2}{3}\). This is referred to as the dynamic averaging regime.
Using the fact that exponential random variables can well model fluorescence times and accounting for the time dependence of the Forster radius on \(\kappa^{2}\) the energy transfer process in FRET is modeled as a time-inhomogeneous continuous-time Markov chain (CTMC), illustrated in Figure 1, with rate matrix defined in;
\[Q(t)=\begin{pmatrix}-(k_{D}+k_{ET}(t))&k_{ET}(t)&k_{D}&0\\ 0&-k_{A}&0&k_{A}\\ 0&0&0&0\\ 0&0&0&0\end{pmatrix}, \tag{4}\]
The state space is defined as \(S=\{D,A,F_{D},F_{A}\}\), where \(D\) is the donor position, \(A\) is the acceptor position, \(F_{D}\) is the donor fluorescence, and \(F_{A}\) is the acceptor fluorescence.
Note that if \(\tau=0\), the CTMC is reduced to a two-state system transitioning between states \(A\) and \(F_{A}\) with rate \(k_{A}\).
Assuming the dynamic averaging regime, one may easily derive the common time-homogeneous FRET efficiency. Observation of an acceptor photon only occurs when energy transfer occurs, i.e., if we have a transition from \(D\to A\). Let \(\tau_{D}\) and \(\tau_{ET}\) be the transfer times of \(D\to F_{D}\) and \(D\to A\) respectively.
\[\mathcal{E}(t) =\mathbb{P}(\min(\tau_{D},\tau_{ET})=\tau_{ET})\] \[=\frac{\int_{0}^{t}k_{ET}(t)}{\int_{0}^{t}k_{ET}(t)+k_{D}t}\] \[=\frac{k_{D}(\frac{R_{0}}{r})t}{k_{D}(\frac{R_{0}}{r})t+k_{D}t}\] \[=\frac{1}{\left(\frac{r}{R_{0}}\right)^{6}+1}.\]
Therefore, the theoretical time-homogeneous FRET efficiency is given by
\[\mathcal{E}=\frac{1}{\left(\frac{r}{R_{0}}\right)^{6}+1}. \tag{5}\]
One may approximate this value using two methods: intensity-based FRET and lifetime-based FRET. For _intensity-based FRET_, the measurements are counts of observed photons from each dye. Effectively, the experiment measures the probability of success of a binomial random variable with a probability of success \(p\) given by the FRET efficiency, \(\mathcal{E}\). The best estimator in the absence of experimental corrections is given by the number of successes observed divided by the total number of trials, denoted in Equation;
\[\mathcal{E}_{I}=\frac{I_{A}}{I_{A}+I_{D}}. \tag{6}\]
For _lifetime-based FRET_, consider
\[\mathcal{E}+P(\min(\tau_{D},\tau_{ET})=\tau_{D})=1.\]
Noting that since \(P(\tau_{D}>t|\min(\tau_{D},\tau_{ET})=\tau_{D})\sim\exp(k_{D}+k_{ET})\), the FRET efficiency can be calculated in terms of the lifetimes,
\[\mathcal{E}=1-\frac{\tau_{D}^{\prime}}{\tau_{D}}, \tag{7}\]
Hence, the measurements are observed lifetimes and an estimate for the mean lifetime of the donor, \(\tau_{D}\). The FRET efficiency is estimated by approximating the mean, and hence the rate, of this exponential random variable.
### The Dynamic Shift
Consider a sample drawn from a population with a distribution of fluorescence rates \(K(x)\) such that the probability of an individual having a specific rate is given by the distribution \(\pi(x)\).
\[\overline{\tau}=\mathbb{E}[\tau]=\int_{\mathbb{R}^{d}}\mathbb{E}[\tau|K(x)]d \pi(x)=\int_{\mathbb{R}^{d}}\frac{1}{K(x)}d\pi(x). \tag{8}\]
Representation of FRET by CTMC.
Comparison of the CTMC states with the Jablonski diagram for the FRET process.
However, the lifetime resulting from the average rate is given by
\[\underline{\tau}=\frac{1}{\mathbb{E}[K(x)]}=\frac{1}{\int_{\mathbb{R}^{d}}K(x)d\pi( x)}. \tag{9}\]
Therefore, by Jensen's inequality, using the fact that \(\phi(x)=\frac{1}{x}\) is convex for \(x\in[0,\infty)\), it must be that
\[\underline{\tau}=\frac{1}{\mathbb{E}[K(x)]}\leq\mathbb{E}\bigg{[}\frac{1}{K(x) }\bigg{]}=\overline{\tau}. \tag{10}\]
Consequently, the average lifetime for a mixture of states will be greater than that of the associated average state. This phenomenon is known as the dynamic shift.
We introduce a new quantitative definition of the dynamic shift \(\Delta\) for a point \((\mathcal{E}^{\prime},\tau^{\prime})\) in the plane, given by the signed distance from the point to the static line, \(S=\{(\mathcal{E},\tau):\mathcal{E}=1-\tau\}\) as shown in
One can find an expression for the dynamic shift using standard analytic geometry.
\[\Delta(\mathcal{E},\tau)=\frac{\mathcal{E}+\tau-1}{\sqrt{\mathcal{E}^{2}+ \tau^{2}}}. \tag{11}\]
The definition of the dynamic shift becomes the signed length of the orthogonal projection of the point onto the static line - how much it deviates from the static line. Under the constraint that the FRET - lifetime pair resides within the unit square, this implies that the dynamic shift has extreme values at \(\pm\frac{1}{\sqrt{2}}\) at \(\) and \(\). This definition provides a means by which each data point from a smFRET experiment may be assigned a dynamic shift value, and the resulting distribution may be examined. The average dynamic shift can be seen as an average deviation from the static line. With two state transitions, this definition agrees with the definition present in. Furthermore, when the average dynamic shift is \(0\), one may use the dynamic shift distribution to quantify shot noise inherent in the measurements.
Another way to view the dynamic shift introduced in is the moment difference approach. In this method, one investigates the behavior of the difference between the first and second moments of the FRET distribution, \(\mathbb{E}[\mathcal{E}(1-\mathcal{E})]=\mathbb{E}[\tau(1-\tau)]\). In this way, the effects of multiple states are linearized, while the static line is non-linear. In this case, the dynamic shift can be seen as a consequence of Jensen's inequality but for concave functions. When dynamic mixing is present in the sample, the moment difference should fall below the static line of \(\mathcal{E}_{\tau}(1-\mathcal{E})\). Note that when this difference is negative, it implies that the covariance between \(\mathcal{E}_{\tau}\) and \(\mathcal{E}\) is larger than the average of \(\mathcal{E}\). This can occur from shot noise or when the lifetime distribution has a large variance but maintains the same mean. Conditions for this to occur are discussed in Section III.2 To define the dynamic shift from the static moment difference line, one again takes the distance from the point to the static line. The vector between the point and the static line with a length equal to the moment difference dynamic shift will be orthogonal to the tangent line of the static line at the point closest to the point.
The dynamic shift introduced in considers an underlying distribution dependent on two separate states. Consider two FRET efficiency states denoted by \(\mathcal{E}_{i},\ i=1,2\) with equal transition rate between the states \(\lambda\) for simplicity. Such a two-state system provides valuable insight into the nature of the dynamic shift. When two states are separated on long time scales, \(\lambda<<1\), the dynamic shift is slight due to the small amount of mixing during a burst or sample. As the two states mix, corresponding to an increase in \(\lambda\), an arc forms between the static FRET-lifetime coordinates, following \((1-\mathcal{E}_{1}-\mathcal{E}_{2})\mathcal{E}-\mathcal{E}_{1}\mathcal{E}_{2}\). As \(\lambda\rightarrow\infty\), this process culminates in a point mass FRET-lifetime distribution with a dynamic shift at the maximum of this arc. Therefore, the dynamic shift can be seen as a metric of the amount of mixing between states.
Visualization of the definition of the dynamic shift using normalized values: \(S=\{(\mathcal{E},\tau)\in^{2}:\mathcal{E}+\tau-1=0\}\) in the bottom. The top figure is a visualization of the moment difference dynamic shift. For the moment difference we used \(\mathcal{E}_{\tau}=1-\frac{\tau_{D(A)}}{\tau_{D}}\).
For the current purpose, it provides a convenient method for interpreting the dynamic shift induced by the dyes. The dynamic shift will most readily be present when there are mixing states, and understanding the influence of mixing between an uncountable number of states is of current interest. It will be shown in Section III that the dynamic shift induced by dye dynamics can be viewed as a consequence of the fluctuations in the energy transfer rate during the FRET process. Under common circumstances, the energy transfer rate can be approximated by a two-state system corresponding to the modes of the distribution, essentially leading to a quickly transitioning two-state system.
## II Stochastic models of fluorescence dynamics
This section presents several models of stochastic fluorescence dynamics related to the smFRET dynamic shift and associated molecular probes. In this way, estimation of the mixture of states, \(\pi(x)\) as seen in Section I.3, is accomplished. shows the basic coordinate expression for the dye motion.
Throughout, the dynamics are assumed to evolve on different timescales. Letting \(T_{P}\), \(T_{D}\), \(T_{O}\) represent the timescales of biomolecules dynamics, dye translational dynamics, and dipole orientational dynamics. The order of timescale separation assumed in this work is given by \(T_{P}>>T_{D}>>T_{O}\). Further, as in, it is assumed that the orientation process and the translational process are independent processes. Note that this is an extremely common assumption since the independence of \(\kappa\) and \(r\) dynamics is implicitly assumed whenever the average \(\kappa\) value is used and whenever static \(\kappa\) distributions are employed. Moreover, to provide a clear and succinct picture of the influence of dye dynamics on FRET measurements, the timescale \(T_{P}\) is not considered in the current discussion. However, an extension of this analysis to include this timescale is in development.
### Spring Models
The simplest possible model to describe a stationary mean-reverting process is an Ornstein - Uhlenbeck (OU) process. This physically represents an overdamped harmonic oscillator subject to noise. The OU process is a Gauss-Markov process and, therefore, provides a simple model for thermal fluctuations of the fluorescent dyes.
\[d\mathbf{X}_{t}=K(\mathbf{X}_{t}-\mathbf{X}_{eq})dt+\sigma\mathbb{I}\circ dB_{ t}, \tag{12}\]
The notation \(\circ dB_{t}\) denotes the use of Stratonovich integration, where \(B_{t}\) is Brownian motion. \(\sigma>0\) is the volatility of the random fluctuations that are modeled as Brownian motions. We refer to systems such that the spring matrices can be written in the form \(K=k\mathbb{I}_{3\times 3}\), as isotropic springs. Otherwise, the system is called anisotropic.
Both isotropic and anisotropic spring systems with a diagonal spring matrix are considered. The spring coefficients are calculated using the linker chemistry. Utilizing the vibrational frequency of a \(C-C\) bond, we find that the spring constant for a single \(C-C\) bond is \(k=1010N/nm\). Therefore, a system of \(N\geq 1\), \(C-\bar{C}\) links is treated as a system of springs in series.
\[\frac{1}{k_{eff}}=\sum_{i=1}^{N}\frac{1}{k}\to k_{eff}=k\frac{k}{N}.\]
Finally, to find the length of the linker, we investigate the equilibrium bond length, \(L\), in a \(C-C-C\) link. Using the law of cosines, we find that \(2L=\sqrt{2l^{2}-2l^{2}\cos\theta}\) with \(l\) being the length of a \(C-C\) bond. Therefore, the effective length in the linker for each link can be calculated using \(l=1.54\AA\) and \(\theta=109.5^{\circ}\).
In the isotropic case, illustrated in Figure 4A, the spring matrix is given by \(k_{eff}\mathbb{I}_{3\times 3}\). This provides a symmetric three-dimensional Gaussian as the stationary distribution for the isotropic spring. It can be seen in that the variance of this distribution will be given by \(\Sigma=\frac{\sigma}{k_{eff}}\mathbb{I}_{3\times 3}\).
In the anisotropic case, illustrated in Figure 4C, we use a diagonal spring matrix with two entries being \(pk_{eff}\) and the third being \(k_{eff}\) with \(p\in\). Therefore, the stationary distribution is an ellipse with major axes determined by the entries of the spring matrix. In this section rotational dynamics have not yet been considered; it will be covered in Section II.3.
Cartoon showing the coordinate references for the processes.
The two-dimensional dynamics in the anisotropic case can be used to investigate the influence of the orientation of the stationary distribution on the resulting dynamic shift. Such a scenario is exemplified in the case when the planes formed by the major axes of each stationary ellipse are mutually orthogonal. Since the stationary distribution for the isotropic case is a sphere and is perfectly symmetric, this can only arise in the anisotropic case.
Furthermore, these models have the added benefit of having an analytical expression for the inter-dye displacement, especially in the isotropic case. Since the coordinates will be Gaussian distributed the distance between them is simply Rayleigh distributed. This distribution is unimodal, and therefore the only mixing present is due to the variance of the stationary distributions. This mixing is therefore strongly dependent on the flexibility of the dyes.
### Elastic Pendulum Model
The next model for the dye linker dynamics takes a stochastic geometric mechanics approach. Consider the motion of a rigid body attached to a spring that is free to move in space. This system forms an elastic pendulum. The following system of Langevin equations describes the motion of a point mass elastic pendulum system subject to white noise;
\[\begin{cases}dr_{t}&=-k_{r}(r_{t}-r_{eq})+\frac{1}{r_{t}}dt+\sigma_{r}\circ dB _{t}^{r}\\ d\theta_{t}&=-k_{\theta}\sin{(\theta_{t})}+\frac{\sigma_{r}^{2}}{r_{t}^{2}\tan {\theta_{t}}}\circ dB_{t}^{\theta}\\ d\phi_{t}&=\frac{\sigma_{\phi}}{r_{t}\sin{(\theta_{t})}}\circ dB_{t}^{\phi}. \end{cases} \tag{13}\]
Similar to Equation, \(B_{t}\) is Brownian motion, \(\sigma>0\) is the volatility of the random fluctuations that are modeled as Brownian motions, and \(k\) is the spring constant. Note that each superscript/subscript is indexed by each of the three components \((r,\theta,\phi)\) explained in the next sentences. Importantly, the system is considered in spherical coordinates. The radial dynamics, \(r_{t}\) evolve according to the spring dynamics explained in Section II.1, with slight alterations due to the change of coordinates. The angular parts of the motion are given by the standard nonlinear pendulum force in the polar direction \(\theta_{t}\) and free diffusion in the azimuthal direction \(\phi_{t}\). shows a sample dye trajectory.
Dye model trajectories and resulting FRET efficiency vs normalized mean fluorescence lifetime and moments difference for A-B) Isotropic Spring, C-D) Anisotropic Spring, E-F) Elastic Pendulum, as described in Section II. G-H) Sample trajectory of a spherical Brownian motion on the unit sphere \(S^{2}\). Such processes are used to model the diffusion of the electric dipole moment. Each of the plots B,D,F,H shows the FRET efficiency vs normalized mean fluorescence lifetime, and the moments difference when the dipole orientation for donor and acceptor is included in the FRET process. Each color in A,C,E,G is an individual burst trajectory. Each case simulates 25000 trajectories over 7 hours. The values in plots B,D,F,H correspond to each simulated state.
The flexibility in the angular components is reminiscent of the wobble in a cone model used in previous investigations, and angular flexibility can be explained via the angular flexibility of \(C-C\) bonds themselves. However, unlike the classical wobble in a cone model, thermal noise and dye linker chemistry drive the dynamics and present a purely stochastic system. Moreover, by varying the parameters used, the system shows various behaviors.
Moreover, this model presents a possible explanation for the dynamic shift induced by dye motion due to the non-Gaussian inter-dye distributions, as shown in
This bimodality presents a mixing of two distinct states that are frequently needed to present a dynamic shift. Further, this provides a much larger change in inter-dye displacement than the spring models, which, as mentioned in Section II.1, do not possess any strong range of translational motion.
### Orientational Dynamics
The final consideration involves the orientational dynamics of the electric dipole moments of the dyes. As discussed in Section I, the Forster radius is dependent on the \(\kappa\) parameter, which is dependent on the mutual orientations of the electric dipole moments \(\hat{\mu}_{A},\hat{\mu}_{D}\) and the inter-dye displacement unit vector \(\mathbf{R}\). Typically, \(\kappa^{2}\) is taken as the mean value of \(2/3\) when the unit vectors are considered uniformly distributed on the sphere \(S^{2}\). This assumption ignores the temporal aspect of the fluorescent process. Since dyes reorient on timescales faster than fluorescent lifetimes, the energy exchange rate changes during the FRET process. This changes the original CTMC model to a time-inhomogeneous CTMC, and thus, the transfer rates are dependent on the time integral of the infinitesimal transfer rates.
To incorporate the influence of orientational dynamics on the lifetime distribution and FRET efficiency, consider the dipoles to be fixed to a reference frame of some rigid body with tensor of inertia \(\mathbf{I}\).
\[\begin{cases}\mathbf{I}d\omega_{t}+\omega_{t}\times\mathbf{I}\omega_{t}=-\nu \omega_{t}+d\mathbf{W_{t}}\\ \omega_{t}=d\mathbf{\Phi_{t}}\end{cases} \tag{14}\]
Assuming the dye is overdamped and hence \(d\omega_{t}=0\), one obtains the simplified equations
\[\begin{cases}\omega_{t}\times\mathbf{I}\omega_{t}=-\nu\omega_{t}+d\mathbf{W_{t }}\\ \omega_{t}=d\mathbf{\Phi_{t}}.\end{cases} \tag{15}\]
Making the assumption that the dye is spherical and therefore the inertia tensor may be replaced with a scalar value and using the fact that \(v\times v=0\) for any vector \(v\) we obtain the simple formula
\[\nu d\mathbf{\Phi_{t}}=d\mathbf{W_{t}} \tag{16}\]
Spherical Brownian motions can be expressed in terms of the Langevin equations below
\[\begin{cases}d\theta_{t}&=\frac{\sigma_{\theta}^{2}}{\tan\left(\theta_{t} \right)}dt+\sigma_{\theta}\circ dB_{t}\\ d\mathbf{\Phi_{t}}&=\frac{\sigma_{\theta}}{\sin\theta_{t}}\circ dB_{t}\end{cases} \tag{17}\]
The rotational diffusion coefficients depend on the hydrodynamic radius of the dye \(R_{h}\) by the classical relation \(D=kT/(8\pi\nu R_{h}^{3})\), where \(kT\) denotes the product of the Boltzmann constant and the temperature.
Note that the stationary distribution for such a system is the uniform distribution, providing an ideal starting stochastic process to test the time-dependent behavior of orientational dynamics. The key idea is that the excitation of the fluorophores provides a single initial \(\kappa\) value. The relaxation effects are the object of interest, especially with regard to lifetime duration. The notion that \(\kappa\) may be close to \(0\) during the entire FRET process for one excitation but higher for another in the same sampling time provides an additional source of variance in the lifetime distribution. Every FRET process can lead to a different equilibrium, which should depend on the rotational diffusion of the dipole moment, with faster reorientation causing an averaging out effect as mentioned in.
Histogram of inter-dye distances for the elastic pendulum model during an excitation event.
## III Sources of observed dynamic shift
### Dye Configuration
This section compares the dye models mentioned in Section II. First, inspecting the joint FRET-lifetime distributions using the contour plots and marginal histograms seen in Figure 4, one can see the influence of the differing models. These FRET-lifetime distributions were generated by simulating the above-mentioned models in a confocal smFRET environment. As can be seen, the spring models show similar characteristics to the anisotropic model, which shows slightly more dynamic shifting. In addition, the elastic pendulum model produces a noticeable shift where the bulk of the distribution lies off of the static line. Despite the dynamic mixing involved in the purely translational elastic pendulum model, the dynamic shift produced is still not representative of experimental observations. Only through the addition of the orientational motion and time-inhomogenous energy transfer rates does the distribution show the hallmark dynamic shift in both the moment difference as well as direct FRET-lifetime distribution.
The dynamic shift distributions of each dye configuration are calculated using the definition of the dynamic shift shown in Equation. It has been known from experimental data that the average dynamic shift of dye motion is \(\mu(\Delta)\approx 0.2\). Using this quantity, the average dynamic shift for the associated models is examined to determine the model that captures the appropriate mean dynamic shift.
In addition, these simulations have no burst noise from background radiation, as this could potentially cloud the impact of the dye motion. The noise is solely from the experimental photon loss considerations and the dye motion as dictated by the models and simulation methods. Therefore, the only source of dynamic shifting must be from the dye models.
Curiously the dynamic shift densities shown in exhibit similar variances but differing mean dynamic shift values, as seen in Table 1.
While the elastic pendulum model does not completely capture the mean dynamic shift, with the addition of a dynamic \(\kappa^{2}\) value, the proper dynamic shift appears. This is in stark contrast to previous suggestions that the dynamic shift is a result of the accessible volume of the dye. Despite the elastic pendulum having an accessible volume comparable to those found in all-atom molecular dynamic simulations and exhibiting dynamic mixing between two states, the resulting dynamic shift is less than expected. This suggests that the inclusion of time-inhomogeneities in the Forster radius due to orientational factors is essential to the description of the dynamic shift.
It is important to note that the spring models can be made to incorporate a dynamic \(\kappa^{2}\) parameter. However, the model parameters based on linker composition prevent these models from being used in earnest, as the accessible volumes are smaller than physically reasonable. This can be seen by investigating the variance of their associated stationary distributions. As mentioned in Section II.1, the variance along each axis in the isotropic case will be \(\sigma/k_{eff}\), or \(K_{B}T/\gamma k_{eff}\), where \(\gamma\) is the local friction. This produces a stationary distribution with a 3 standard deviation radius of less than an Angstrom.
### \(\kappa^{2}\) Dynamics
An important consideration brought to attention in Section III.1 is the influence of \(\kappa^{2}\) dynamics on FRET lifetime pairs. The key issue in incorporating \(\kappa^{2}\) dynamics into FRET uncertainty quantification has been to assess \(\kappa^{2}\) as a stationary object during energy transfer. This is done by considering \(\kappa^{2}\) as being chosen from its equilibrium distribution, assuming that it follows a discrete state Markov chain, or simply using the 2/3 approximation. From there, one can use the mean and standard deviation in uncertainty quantification. However, these approximations ignore the time inhomogeneous nature of the FRET process, as mentioned in Section I.2.
\begin{table}
\begin{tabular}{|c|c|c|} \hline Model & \(\mu(\Delta)\) & \(\sigma(\Delta)\) \\ \hline Isotropic Spring & 0.10 & 0.01 \\ \hline Anisotropic Spring & 0.01 & 0.01 \\ \hline Elastic Pendulum \(\kappa\approx 2/3\) & 0.12 & 0.01 \\ \hline Elastic Pendulum Dynamic \(\kappa\) & 0.20 & 0.10 \\ \hline \end{tabular}
\end{table}
Table 1: **Dynamic Shift:** Mean and standard deviation of the dynamic shift for each model in
Linear Dynamic shift, \(d\), histogram comparison of dye modelsof a donor fluorescence event, as shown in Section I.2, is dependent on the integral of the energy transfer rates as well as the lifetime distribution. While the use of an ergodic approximation might mitigate these concerns, this relies on the convergence of the long run time average to the long run spatial average. The short time span involved in the FRET process, especially in slow rotational diffusion regimes, should not be considered long enough to invoke such an approximation. Furthermore, this mitigation ignores the dependence of the energy transfer rate and FRET efficiency on the history of the \(\kappa^{2}\) process. It is important to note that experimental bursts are one-millisecond in length, equivalent to one molecule. The equivalent measurement used in the simulations is one burst per second, leading to approximately 25,000 bursts. As seen in Section I.2, the energy transfer rate is dependent on the integral of the \(\kappa^{2}\) path. In we keep track of the \(\kappa^{2}\) value and display it's stochasticity. For example, in Figures 7A and B, the trajectories have similar mean values after the \(1ns\) run. However, if one were to investigate the energy transfer rate at time 0.5 one would find that the probability of energy transfer is much lower for path B rather than being similar to path A. One can see how this integral fluctuation may be reduced in the presence of fast dyes. As shown in Figure 7C, the rapid oscillations induced by the swift rotational diffusion of the dyes produce highly clustered \(\kappa^{2}\) trajectories. However, as can be seen in the associated average \(\kappa^{2}\) distribution in Figure 7F, the resulting average distribution becomes bimodal. This bimodality can be seen in the trajectories in which the oscillations can vary between smaller values or larger values. Notably, this is heavily dependent on the radial-dipole dot product term in the definition of \(\kappa^{2}\). Furthermore, with the decrease in rotational diffusion, the \(\kappa^{2}\) process predictably oscillates much slower. This yields trajectories such that larger values are maintained for longer. These larger stretches of high \(\kappa^{2}\) values can lead to much shorter donor lifetimes, but similarly, this also leads to longer stretches of small \(\kappa^{2}\) values. Overall, slower rotational diffusion will show a larger temporal correlation. Hence, the stationary distribution will play a much more important role, since the \(\kappa^{2}\) trajectory is not as likely to change as drastically from its initial value.
The skew of the mean \(\kappa^{2}\) value shows the flaw in the averaging assumption. While the mean \(\kappa^{2}\) average is indeed 2/3, the most common values for the mean \(\kappa^{2}\) are below this value.
Comparison of \(\kappa^{2}\) trajectories for three common cases. Figure \(A-C\) shows the sample paths of \(\kappa^{2}\) during an excitation event, these can be seen as four realizations of \(\kappa^{2}\) during the same burst. Each path is representative of a single energy transfer event. Figures \(D-F\) show the associated distribution of average \(\kappa^{2}\), the red vertical lines in figures \(D-F\) are the isotropic 2/3 average, where the colored lines are the mean of the path average. Figures \(A\) and \(D\) show a pair of dyes for which the rotational diffusion is only an order of magnitude greater than the translational diffusion of the dye. Figures \(B\) and \(E\) show the sample behavior when one dye has a rotational diffusion three orders of magnitude greater than translational and one dye one order of magnitude greater. Finally, Figures \(C\) and \(F\) show the case in which both dyes have rotational diffusions three orders of magnitude greater than translational.
The dynamic shift produced by dye motion is then a product of the variability of the \(\kappa^{2}\) trajectories. This temporal heterogeneity _during_ the energy exchange process provides the crucial mixing element to significantly slow the the average sampled donor lifetime. While it can be seen in that inter-dye radial displacement dynamics can produce a slight dynamic shift, the mixing is not strong enough to produce the lifting seen by considering the direct change in energy transfer rate caused by rotational dynamics. Figure 7D- 7E can be seen as representative of the average \(\kappa^{2}\) values during a single burst. If one simplifies the path-based description above to these distributions, it is simple to see that in the case of two rapidly rotating dyes, the bimodality of the distribution provides two populations of energy transfer rate. The switching between the two during the burst provides a direct comparison to the two-state systems investigated in. However, it must be stressed that the path-by-path heterogeneity of the FRET process is crucial to the analysis. While the average values shown can be representative of normalized time integrals, one must consider that the plots shown in each last for one nanosecond, and therefore do not capture the stopping time of fluorescence or energy transfer. As noted, the time-dependent FRET efficiency \(\mathcal{E}(t)\) will vary path-by-path, producing the observed lifetime distribution.
## IV Discussion
In this work, we showed how dynamic shift results from dye motion. By presenting the first physics-based model for fluorescence dynamics, incorporating dye linker chemistry and fluorescent dye composition, it has become clear that the time-inhomogeneous nature of the Forster radius is an essential part of the dynamic shift. This is shown by noting that models with proper accessible volumes do not demonstrate experimentally observed dynamic shifts and noting that the time-inhomogeneous nature of the FRET process depends on the path space dynamics of \(\kappa^{2}\) trajectories. Furthermore, it was shown that the dynamics of \(\kappa^{2}\) trajectories change with the selection of fluorescent dyes used, with common situations such as an organic dye paired with a fluorescent biomolecules exhibiting noticeable differences from a pair of organic dyes. While the average \(\kappa^{2}\) stays consistent with the 2/3 isotropic assumption, the fluctuations during the FRET process cause fluctuations in the FRET-lifetime distribution during the burst. These fluctuations provide a source of dynamic mixing and, hence, a dynamic shift. In short, FRET efficiency does not change, but the lifetime does.
In this work the results are for only spherical dyes, however, the rotational Langevin equations used in section II.3 can be utilized for arbitrary intertia tensors. While this is simple to state, the resulting equations become arbitrarily difficult to work with, as well as simulate. Non-spherical dyes require different intertia tensors and rotational processes, which produce unique \(\kappa^{2}\) trajectories.
An additional consideration for future work is the coupling between translational motion and orientational motion. While this is a common assumption, as it is invoked each time dynamic averaging is used for \(\kappa^{2}\) and is used in other theoretical analyses such as in, it is an important direction of consideration. As noted in there is a marked correlation between translational motion and \(\kappa^{2}\). This can be further emphasized by the considerations of Sections II.2 and II.3 therein one may couple the orientational rigid body dynamics and the dynamics of the elastic pendulum. Such a system is known to have non-trivial coupling in the classical case that is likely to remain in the overdamped case. These considerations, however, increase the difficulty of analysis and simulation considerably as one must consider the dynamics of a full state space model on the non-compact, non-abelian Euclidean group \(SE\) rather than independent processes as used here and in.
Furthermore, the work shown here provides an important framework for analyzing other sources of the dynamic shift. The algorithms used to simulate the experiments can easily be scaled to incorporate another timescale. This opens the door to studying the influence of biomolecules dynamics on the dynamic shift. By using reaction coordinate Langevin models, one may investigate the influence of differing energy landscapes on the dynamic shift, as well as the influence of a more dynamic inter-dye distance vector. This can be done swiftly due to the low computational cost incurred by the stochastic simulation of the FRET process. Additionally, by introducing a new form for the dynamic shift, distributional quantities may be accessed more readily, providing a new tool to analyze the motion of underlying molecules of interest as well as dye dynamics. Additionally, while our simulated experiments were conducted in the confocal environment the results carry over to TIRF measurements as well.
In conclusion, this work has demonstrated the importance of time-inhomogeneities in the FRET process and their influence on resulting measurements. The path dependence of the FRET efficiency, as well as the influence of a time-varying Forster radius during the FRET process, is shown to be non-negligible. In order to conduct optimal uncertainty quantification for smFRET measurements the path dynamics of \(\kappa^{2}\), not just the average values, must first be understood. Therefore, the anisotropic behaviors of fluorescent dyes are of vital importance. By gaining an understanding of the orientational dynamics of fluorescent dyes, better uncertainty quantification for FRET measurements can be done. Additionally, this relationship between the full state space dynamics of a fluorescent dye and resulting smFRET measurements emphasizes the need for more stochastic geometric mechanical considerations in fluorescent measurements and biology.
|
10.48550/arXiv.2404.09883
|
Time-Heterogeneity of the Förster Radius from Dipole Orientational Dynamics Impacts Single-Molecule FRET Experiments
|
David Frost, Keisha Cook, Hugo Sanabria
| 1,740
|
10.48550_arXiv.1209.5882
|
###### Abstract
We analyze a specific class of random systems that are driven by a symmetric Levy stable noise. In view of the Levy noise sensitivity to the confining "potential landscape" where jumps take place (in other words, to environmental inhomogeneities), the pertinent random motion asymptotically sets down at the Boltzmann-type equilibrium, represented by a probability density function (pdf) \(\rho_{*}(x)\sim\exp[-\Phi(x)]\). Since there is no Langevin representation of the dynamics in question, our main goal here is to establish the appropriate path-wise description of the underlying jump-type process and next infer the \(\rho(x,t)\) dynamics directly from the random paths statistics. A priori given data are jump transition rates entering the master equation for \(\rho(x,t)\) and its target pdf \(\rho_{*}(x)\). We use numerical methods and construct a suitable modification of the Gillespie algorithm, originally invented in the chemical kinetics context. The generated sample trajectories show up a qualitative typicality, e.g. they display structural features of jumping paths (predominance of small vs large jumps) specific to particular stability indices \(\mu\in\).
pacs: 05.40.Jc, 02.50.Ey, 05.20.-y, 05.10.Gg
## I Introduction
Many random processes in real physical systems admit a simplified description based on stochastic differential equations. In such case there is a routine passage procedure from microscopic random variables to macroscopic (statistical ensemble) data. The latter are encoded in the time evolution of an associated probability density function (pdf) which is a solution of a deterministic transport equation. A paradigm example is the so-called Langevin modeling of diffusion-type and jump-type processes. The presumed microscopic model of the dynamics in external force fields is provided by the Langevin (stochastic) equation whose direct consequence is the Fokker-Planck equation, Risken and Risken. We note that in case of jump-type processes the familiar Laplacian (Wiener noise generator) needs to be replaced by a suitable pseudo-differential operator (fractional Laplacian, in case of a symmetric Levy-stable noise).
We pay a particular attention to jump-type processes which are omnipresent in Nature (see Risken and references therein). Their characterization is primarily provided by jump transition rates between different states of the system under consideration. However our major focus is on a specific class of random systems which are plainly incompatible with a straightforward Langevin modeling of jump-type processes and, as such, are seldom addressed in the literature.
To this end we depart from the concept, coined in an isolated publication Risken, of Levy flights-driven models of disorder that, while at equilibrium, do obey detailed balance. The corresponding research line has been effectively initiated in Refs. Risken-Risken. It has next been expanded in various directions, with a special emphasis on so-called Levy-Schrodinger semigroup reformulation of the original probability density function (pdf) dynamics, Risken; Risken; Risken and Risken-Risken, c.f. also Risken; Risken; Risken. We note in passing that the familiar Fokker-Planck equation can be equally well formulated in terms of the Schrodinger semigroup and this property is universally valid in the standard theory of Brownian motion, Risken; Risken. Its generalization to Levy flights is neither immediate nor obvious. It is often considered in the prohibitive vein following Risken; Risken.
In fact, in relation to Levy flights, a novel fractional generalization of the Fokker-Planck equation has been introduced in Refs. Risken-Risken to handle systems that are randomized by symmetric Levy-stable drivers. In this case, contrary to the popular lore about properties of (Langevin-based) Levy processes c.f. Refs. Risken-Risken and Risken, the pertinent random systems are allowed to relax to (thermal) equilibrium states of a standard Boltzmann-Gibbs form.
The underlying jump-type processes, in the stationary (equilibrium) regime, respect the principle of detailed balance by construction Risken. Their distinctive feature, if compared with the standard Langevin modeling of Levy flights, is that they have a built-in response _not_ to external forces _but rather to_ external force potentials. These potentials are interpreted to form confining "potential landscapes" that are specific to the environment. Levy jump-type processes appear to be particularly sensitive to environmental inhomogeneities, Risken; Risken.
Levy flights are pure jump (jump-type) processes. Therefore, it seems useful to indicate that various model realizations of standard jump processes (jump size is bounded from below and above) can be thermalized by means of a specific scenario of an energy exchange with the thermostat. It is based on the _principle of detailed balance_. We have discussed this issue in some detail before Risken along with an extension of this conceptual framework to Levy-stable processes. Not to reproduce easily available arguments of past publications, we shall be very rudimentary in our motivations.
We quantify a probability density evolution, compatible with a jump-type process on \(R\) (this limitation may in principle be lifted in favor of \(R^{n}\)), in terms of the masterequation:
\[\partial_{t}\rho(x)=\int\limits_{\varepsilon_{1}\leq|x-y|\leq\varepsilon_{2}}[w_{ \phi}(x|y)\rho(y)-w_{\phi}(y|x)\rho(x)]dy, \tag{1}\]
where \(\varepsilon_{1}\) and \(\varepsilon_{2}\) are, respectively, the lower and upper bounds of jump size and
\[w_{\phi}(x|y) = C_{\mu}\frac{\exp[(\Phi(y)-\Phi(x))/2]}{|x-y|^{1+\mu}},\] \[C_{\mu} = \frac{\Gamma(1+\mu)\sin(\pi\mu/2)}{\pi} \tag{2}\]
We stress that \(w_{\phi}(x|y)\) is a non-symmetric function of \(x\) and \(y\).
An implicit Boltzmann-type weighting involves a square root of a target pdf \(\rho_{*}(x)\sim\exp[-\Phi(x)]\) and accounts for the a priori prescribed "potential landscape" \(\Phi(x)\) whose confining features affect the jump-type process. What matters is a relative impact of a confinement strength of \(\Phi(x)\) (level of attraction, see Ref.) upon jumps of the size \(|x-y|\), both at the point of origin \(y\) and that of destination \(x\). In principle, \(\Phi(x)\) may be an arbitrary function that secures a \(L^{1}(R)\) normalization of \(\exp(-\Phi(x))\). In this case, the resultant pdf \(\rho_{*}\) is a stationary solution of the transport equation with unbounded jump length, e.g. \(\varepsilon_{1}\to 0\) and \(\varepsilon_{2}\to\infty\).
We note that the presence of lower and upper bounds of the jump size \(\varepsilon_{1,2}\), that are necessary for an implementation of numerical algorithms, enforces a truncation of the jump-type process (without any cutoffs) to a standard jump process. The transition rates of the latter, however, are ruled by Levy measures of symmetric Levy stable noises with \(\mu\in\). A lower bound for the jump size is usually removed while evaluating the corresponding integrals in the sense of their Cauchy principal values. An upper bound is less innocent and its effects need to be controlled by long tailed pdfs which stands for a distinctive feature of Levy flights, see a discussion of Levy stable limits of step processes in Ref.. There is also pertinent discussion of a long time behavior of (unconfined, e.g. free) truncated Levy flights in Ref..
In contrast to procedures based on the Langevin modeling of Levy flights in external force fields,, there is no known path-wise approach underlying the transport equation. With no direct access to sample trajectories of the stochastic process in question, a method must be devised to generate random paths directly from jump transition rates. The additional requirement here is that we set a priori a "potential landscape" \(\Phi(x)\) for a chosen jump-type (symmetric Levy stable) noise driver.
The outline of the paper is as follows. First we describe our modification of the Gillespie algorithm which entails a numerical generation of random paths for the dynamics determined by Eqs and. Next the statistics of random paths is addressed and various accumulated data are analyzed with a focus on inherent compatibility issues.
We analyze generic (Cauchy, quadratic Cauchy) and non-generic (Gaussian and locally periodic) examples of target pdfs for the jumping dynamics. Random paths are generated in conjunction with representative Levy stable drivers, like e.g. those indexed by \(\mu=1/2,1,3/2\). Their qualitative typicality is emphasized.
Statistical data, acquired from our modification of Gillespie algorithm, have been employed to generate the dynamical patterns of behavior \(\rho(x,t)\to\rho_{*}(x)\), to demonstrate the compatibility of the transport (master) equation, and its underlying path-wise representation. Both coming from the predefined knowledge of the target pdf and non-symmetric (biased) jump transition rates.
## II Random paths: modified Gillespie algorithm.
Here we adopt (and properly adjust to handle Levy flights) basic tenets of so-called Gillespie's algorithm. Originally, this algorithm had been devised to simulate random properties of coupled chemical reactions. The advantage of the algorithm is that it permits to generate random trajectories of the corresponding stochastic process directly from its (jump) transition rates, with no need for any stochastic differential equation and/or its explicit solution. We emphasize that this feature of Gillespie's algorithm is vitally important, since Langevin modeling is not operational in our framework.
We rewrite Eq. in the form \((x-y=z)\)
\[\partial_{t}\rho(x)=\int\limits_{\varepsilon_{1}\leq|z|\leq \varepsilon_{2}}\bigg{[}w_{\phi}(x|z+x)\rho(z+x)-\] \[-w_{\phi}(z+x|x)\rho(x)\bigg{]}dz. \tag{3}\]
To construct a reliable path generating algorithm consistent with Eq. we first note that chemical reaction channels in the original Gillespie's algorithm may be re-interpreted as jumps from one spatial point to another, like transition channels in the spatial jump process. An obvious provision is that the set of possible chemical reaction channels is finite (and generically low), while we are interested in all admissible jumps from a chosen point of origin \(x_{0}\) to any of \([x_{0}-\varepsilon_{2},x_{0}-\varepsilon_{1}]\cup[x_{0}+\varepsilon_{1},x_{0}+ \varepsilon_{2}]\). It is clear that such jumps form an infinite continuous set. With a genuine computer simulation in mind, we must respect standard numerical assistance limitations. Surely we cannot admit all conceivable jump sizes. As well, the number of destination points, even if potentially enormous, must remain finite for any fixed point of origin.
Our modified version of the Gillespie's algorithm, appropriate for handling of spatial jumps is as follows:1. Set time \(t=0\) and the point of origin \(x=x_{0}\).
2. Create the set of all admissible jumps from \(x_{0}\) to \(x_{0}+z\) that is compatible with the transition rate \(w_{\phi}(z+x_{0}|x_{0})\).
3. Evaluate \[W_{1}(x_{0}) =\int_{-\varepsilon_{2}}^{-\varepsilon_{1}}w_{\phi}(z+x_{0}|x_{0} )dz,\] \[W_{2}(x_{0}) =\int_{\varepsilon_{1}}^{\varepsilon_{2}}w_{\phi}(z+x_{0}|x_{0})dz\] and \(W(x_{0})=W_{1}(x_{0})+W_{2}(x_{0})\).
4. Using a random number generator draw \(p\in\) from a uniform distribution.
5. Using above \(p\) and identities \[\left\{\begin{array}{ll}\int\limits_{-\varepsilon_{2}}^{b}w_{ \phi}(z+x_{0}|x_{0})dz=pW(x_{0}),&p<W_{1}(x_{0})/W(x_{0});\\ W_{1}(x_{0})+\int\limits_{\varepsilon_{1}}^{b}w_{\phi}(z+x_{0}|x_{0})dz=pW(x_{0 }),&p\geqslant W_{1}(x_{0})/W(x_{0}),\end{array}\right.\] find \(b\) corresponding to the "transition channel" \(x_{0}\to b\).
6. Draw a new number \(q\in\) from a uniform distribution.
7. Reset time label \(t=t+\Delta t\) where \(\Delta t=-\ln q/W(x_{0})\).
8. Reset \(x_{0}\) to a new value \(x_{0}+b\).
9. Return to step (ii) and repeat the procedure anew.
## Comment 1
The original Gillespie algorithm employs a discrete label \(\nu\) (with a finite range) enumerating possible chemical reactions channels. To identify a channel, one must look for estimates of a double inequality (see Eq. (21b) of Ref.)
\[\sum_{\nu=1}^{\mu-1}a_{\nu}<r_{2}a_{0}\leqslant\sum_{\nu=1}^{\mu}a_{\nu},\qquad a _{0}=\sum_{\nu=1}^{M}a_{\nu}, \tag{6}\]
To adjust this recipe to our settings, we need to enumerate the infinite number of (infinitely close) channels. This corresponds to passing from summation to integration in Eq.. As it has been pointed out above, such situation corresponds to possible transitions from \(x_{0}\) into an interval \([x_{0}-\varepsilon_{2},x_{0}-\varepsilon_{1}]\cup[x_{0}+\varepsilon_{1},x_{0} +\varepsilon_{2}]\). As the Lebesgue measure of a point equals zero, we can replace inequalities by identities in, see step (v) of the above algorithm. Formally, we can say that although the number of jumps destinations is finite, their number is so large that it is consistent to approximate it by a dense subset of intervals \([x_{0}-\varepsilon_{2},x_{0}-\varepsilon_{1}]\cup[x_{0}+\varepsilon_{1},x_{0} +\varepsilon_{2}]\). This justifies a replacement of finite sum by an integral over corresponding interval. The following jump size bounds (integration boundaries) were adopted in the numerical procedure: \(\varepsilon_{1}=0.001\) and \(\varepsilon_{2}=1\).
## III Statistics of random paths: pdf time evolution and compatibility issues.
Our main task in the present section is to select, to some extend generic, transition rates for jump-type processes that will prove to be amenable to the outlined random path generation procedure. Once suitable path ensemble data are collected, we shall verify whether statistical (ensemble) features of generated random trajectories are compatible with the master equation. That includes a control of an asymptotic behavior \(\rho(x,t)\rightarrow\rho_{*}(x)\) when \(t\rightarrow\infty\).
### Harmonic confinement (Gaussian target)
Let us consider an asymptotic invariant (target) pdf in the Gaussian form:
\[\rho_{*}(x)=\frac{1}{\sqrt{\pi}}e^{-x^{2}}. \tag{7}\]
The corresponding \(\mu\)-family of transition rates reads
\[w_{\phi}(z+x|x)=C_{\mu}\frac{e^{-z^{2}/2+xz}}{|z|^{1+\mu}}. \tag{8}\]According to the step (iii) of the simulation algorithm, we must evaluate integrals with transition rates in the intervals \([-\varepsilon_{2},-\varepsilon_{1}]\) and \([\varepsilon_{1},\varepsilon_{2}]\). To execute the step (iv) of the algorithm, we employ the Mersenne-Twister random number generator,. We find \(b\) by a numerical solution of the transcendental equation in conformity with step (v) of the algorithm. The C-codes for trajectory generating algorithm,, were finally employed to get the trajectory statistics data for three specific choices of Levy drivers, namely \(\mu=0.5,1,1.5\).
## Comment 2
For small \(z\), transition rates vary rapidly so that adaptive numerical integration algorithms become very time-consuming. To speed up the calculations, we propose a more efficient procedure (than adaptive numerical integration algorithms with huge number of subdivisions for small \(z\)), that amounts to separate integrations for "large" and "small" \(z\) subintervals. At small \(z\) we can expand the corresponding integrand in Taylor series and truncate it at, say, quadratic term. Generally speaking, the number of terms to be left depends on the accuracy which should be retained during the integration. At "large" \(z\) the transition rates vary gradually so that standard adaptive numerical integration works fine.
As an example, we consider the \([\varepsilon_{1},\varepsilon_{2}]\) integration with \(\mu=1\). We set \(\varepsilon_{12}=0.05\), so that for \(z\in[\varepsilon_{1},\varepsilon_{12}]\) we get
\[\int\limits_{\varepsilon_{1}}^{\varepsilon_{12}}\frac{e^{-z^{2}/ 2-xz}}{z^{2}}dz\approx\int\limits_{\varepsilon_{1}}^{\varepsilon_{12}}\frac{1- xz+\frac{x^{2}-1}{2}z^{2}}{z^{2}}dz=\] \[=\frac{\varepsilon_{12}-\varepsilon_{1}}{2\varepsilon_{1} \varepsilon_{12}}\left[2+(x^{2}-1)\varepsilon_{1}\varepsilon_{12}\right]-x\ln \left|\frac{\varepsilon_{12}}{\varepsilon_{1}}\right|. \tag{9}\]
In the interval \(z\in[\varepsilon_{12},\varepsilon_{2}]\) we evaluate the integral numerically. The proposed hybrid procedure (integrating analytically in the "most dangerous" small \(z\) interval and numerically otherwise) permits to speed up the calculation drastically and has actually been used in our simulations.
The results of our numerical simulations are reported in Figs. 1 through 3. We note, that on the second moment oscillates near its equilibrium value \(1/2\). The oscillations are smoothed out with the growth of the number of random trajectories that contribute to the statistics. A numerical convergence to \(<X^{2}>=\)1/2 is consistent with an analytic equilibrium value of the second moment of the chosen \(\rho_{*}(x)\). The rate of this convergence is higher for larger \(\mu\in\). Clearly, for small \(\mu\) the big jumps are frequent which enlarges the inferred time intervals \(\Delta t\) in the Gillespie's algorithm, see the trajectories on left panel of Thus, the relaxation to equilibrium is slow. It gets faster for larger \(\mu\), when big jumps are rare and time intervals \(\Delta t\) are generically very small.
All of them have started form the same point \(x=0\). Although the data fidelity grows with the number of contributing paths, we have not found significant qualitative differences to justify a presentation of data for 100 000, 200 000, 250 000 and more trajectories. The relaxation time rate dependence on \(\mu\) is clearly visible as well. It suffices to analyze differences between three curves for \(t=0.2\) and/or \(t=1\). We observe a conspicuous lowering of their maxima with the growth of \(\mu\) (take care of different scales on the vertical axes on Fig.2 panels). The simulated pdfs at \(t=10\) are practically indistinguishable from an exact analytical asymptotic pdf. The convergence of \(\rho(x,t)\) towards \(\rho_{*}(x)\) appears to be relatively fast irrespective of the chosen \(\mu\)-driver.
Although our reasoning is definitely path-wise and all data have been extracted from trajectory ensembles, it is instructive to visualize generic sample paths. That is accomplished in Fig. 3, basically to indicate their (paths) qualitative typicalities. The structural impact of larger against smaller jumps can be visually compared and has been found to conform with standard simulations of Levy stable sample paths (with no forces or potentials involved), c.f..
### Logarithmic confinement
#### iii.2.1 Quadratic Cauchy target
Let us consider a long-tailed asymptotic pdf which is a special \(\alpha=2\) case of the one-parameter \(\alpha\)-family of equilibrium (Boltzmann-type) states, associated with a logarithmic potential \(\Phi(x)\equiv\alpha\ln(1+x^{2})\), \(\alpha>1/2\), see :
\[\rho_{*}(x)=\frac{2}{\pi}\frac{1}{(1+x^{2})^{2}}. \tag{10}\]
The transition rate \(w_{\phi}(z+x|x)\) for any \(\mu\in\) takes the form
\[w_{\phi}(z+x|x)=\frac{C_{\mu}}{|z|^{1+\mu}}\frac{1+x^{2}}{1+(z+x)^{2}}. \tag{11}\]
Similar to the previous Gaussian case, the simulations can be speed up by analytical evaluation of some integrals. Such acceleration of numerical routines permits to handle (in the same timescale) trajectories for much longer running times (\(t\sim 400\)) than in the previous harmonic case (\(t\sim 10\)). Each \(\mu\) -driver case (\(\mu=0.5,1,1.5\)) will be addressed separately.
## Case of \(\mu=1/2\)
For \(z>0\) we need to evaluate
\[f_{1/2}(x,z)=\int\frac{1}{z^{3/2}}\frac{1+x^{2}}{1+(z+x)^{2}}dz. \tag{12}\]
As
\[\frac{1}{z^{3/2}}\frac{1+x^{2}}{1+(z+x)^{2}}=\frac{1}{z^{3/2}}-\frac{2x+z}{ \sqrt{z}\left[1+(x+z)^{2}\right]}, \tag{13}\]Gaussian target. Qualitative typicalities of sample paths for \(\mu=0.5\) (left panel), \(\mu=1\) (middle panel) and \(\mu=1.5\) (right panel). All trajectories originate form \(x=0\). Right panel comprises two trajectories instead of three (for better visibility).
Gaussian target: Time evolution of \(\rho(x,t)\) inferred from 75 000 trajectories: \(\mu=0.5\) (left panel), \(\mu=1\) (middle panel) and \(\mu=1.5\) (right panel). All trajectories originate from \(x=0\), i.e. refer to the \(\delta(x)\)-type initial probability distribution.
Gaussian target: Time evolution of the pdf \(\rho(x,t)\) second moment for 25 000 (left panel), 50 000 (middle panel) and 75 000 (right panel) trajectories. Insets visualize the oscillations smoothing in the asymptotic regime for \(10\leq t\leq 15\); figures near curves correspond to \(\mu\) values.
\(f_{1/2}(x,z)\) can be written as
\[f_{1/2}(x,z)=-2z^{-1/2}-\int\frac{2x+z}{\sqrt{z}(1+(x+z)^{2})}dz. \tag{14}\]
For small \(z\) the dominant contribution comes from the first term so that the second term can efficiently be evaluated numerically. We note here that although the integration in the second term of can be performed analytically, the result appears to be quite cumbersome. Therefore, for our purposes it is more profitable to integrate this term numerically. For \(z<0\) the corresponding integral can be expressed as \(-C_{\mu}f_{1/2}(-x,-z)\).
## Case of \(\mu=1\)
Here, we need to evaluate the integral
\[f_{1}(x,z)=\int\frac{1+x^{2}}{1+(z+x)^{2}}\cdot\frac{dz}{z^{2}}=\frac{x}{1+x^{ 2}}\ln\left(1+(z+x)^{2}\right)+\frac{x^{2}-1}{x^{2}+1}\arctan(z+x)-\frac{2x}{ 1+x^{2}}\ln|z|-\frac{1}{z}. \tag{15}\]
We pay attention to the fact that the integrand is rational, hence the integral can be evaluated analytically in a simple form. This permits to use the analytical answer in our simulations without the need to divide the integration range into "small" and "large" \(z\) domains.
## Case of \(\mu=3/2\)
In this case, we have for \(z>0\)
\[f_{3/2}(x,z)=\int\frac{1}{z^{5/2}}\frac{1+x^{2}}{1+(z+x)^{2}}dz. \tag{16}\]
Since
\[\frac{1}{z^{5/2}}\frac{1+x^{2}}{1+(z+x)^{2}}=\frac{1}{z^{3/2}}\left(\frac{1}{ z}-\frac{2x+z}{1+(x+z)^{2}}\right), \tag{17}\]
there holds
\[f_{3/2}(x,z) = -\frac{2}{3z^{3/2}}+\frac{4x}{(1+x^{2})z^{1/2}}+ \tag{18}\] \[\int\frac{2xz+3x^{2}-1}{(1+x^{2})(1+(x+z)^{2})z^{1/2}}dz.\]
Again, the third term in can efficiently be evaluated numerically. For \(z<0\) we encounter \(-f_{3/2}(-x,-z)\).
Simulation results are displayed in Figs. 4 and 5. If we compare with we see the existence of small oscillations in the asymptotic regime about the value \(1/2\). Those from Fig.1 were relatively small and were quickly smoothed out with the growth of the number of trajectories used to extract statistical data. In the oscillations are more noticeable and persist even for \(200000\) trajectories and more. This is related to much slower decay of transition rates (determined by slow-decaying asymptotic pdf) as compared to those for Gaussian case.
The second moment of the present \(\rho_{*}(x)\),, equals \(1\) and the convergence towards this value is clearly seen in This convergence is much slower than in the Gaussian (harmonic confinement) case which is not a surprise: and indicate that the present rate of convergence should be logarithmically slower. Fig. 5, quite alike Fig. 2, convincingly demonstrates a convergence of \(\rho(x,t)\) to the asymptotic \(\rho_{*}(x)\). For definitely large times around \(t=400\), \(\rho(x,t)\) and \(\rho_{*}(x)\) become practically indistinguishable. Similarly to the Gaussian case, the rate of convergence becomes larger with the growth of \(\mu\in\).
#### iii.2.2 Cauchy target
Now we consider an asymptotic pdf of the form :
\[\rho_{*}(x)=\frac{1}{\pi}\frac{1}{1+x^{2}}. \tag{19}\]
In this case, the transition rate from \(x\) to \(x+z\) reads
\[w_{\phi}(z+x|x)=\frac{C_{\mu}}{|z|^{1+\mu}}\sqrt{\frac{1+x^{2}}{1+(z+x)^{2}}}. \tag{20}\]
We consider Cauchy driver corresponding to \(\mu=1\). The transition rate integral can be evaluated analytically.
\[f(x,z)=\int\sqrt{\frac{1+x^{2}}{1+(z+x)^{2}}}\cdot\frac{dz}{z^{2}}=\frac{- \sqrt{1+x^{2}}\sqrt{1+(x+z)^{2}}+xz\ln\left((1+x^{2}+xz+\sqrt{1+x^{2}}\sqrt{1+ (x+z)^{2}})/z\right)}{(1+x^{2})z}. \tag{21}\]
For \(z<0\) the outcome is \(-f(-x,-z)\).
In Fig. 6, we report the time evolution of the statistically inferred \(\rho(x,t)\), its half-width (as second moment does not exist for asymptotic pdf) and simulated cumulative probability distributions (CPD) for different time instants. An approach to the asymptotic pdf is clearly seen, together with a convergence of a half-width to its asymptotic value 1. The same convergence pattern is observed for CPD which approaches the asymptotic function \(F(x)=\frac{1}{2}+\frac{\arctan x}{\pi}\).
## Comment 3
Displayed empirical (numerically retrieved) curves in are hampered by certain errors. The figures have been read from a histogram of randomly sampled data. Its partitioning into subintervals is a source of inaccuracies. In case of a small number of intervals, the read-out error would be large, with a size of about half-interval length. A finer partitioning (large number of small subintervals) would still produce an error which is close to the half-maximum of the curve. The error bound would be smaller or equal to the half-length of subintervals corresponding to roughly the same histogram values. One more inaccuracy source in the finer partition case comes from the maximum read-out imprecision. Namely, we can have a conspicuous peak, whose close vicinity displays much (half or less) smaller values. Therefore the partitioning finesse must be slightly optimized.
### Locally periodic confinement
To set firm grounds for future research it is instructive to study our model for more complicated forms of confining potentials. In view of their physical relevance, it is appealing to address an issue of confining (trapping) environments with a periodic spatial structure. Here, we encounter a major difficulty with a \(L^{1}(R)\) integrability of the Boltzmann-type weighting function \(\exp(-\Phi)\). Periodicity and integrability can here be reconciled either on compact sets or by means of locally periodic potentials that take a definite confining form (harmonic or polynomial) for larger values of \(x\in R\).
\[\rho_{*}(x)=\left\{\begin{array}{ll}\frac{1}{C}e^{-\sin^{2}(2\pi x)},&|x| \leqslant 2;\\ \frac{1}{C}e^{-(x^{2}-4)},&|x|>2,\end{array}\right. \tag{22}\]
Quadratic Cauchy target: Time evolution of the pdf \(\rho(x,t)\) second moment for 50 000 (upper left panel), 100 000 (upper right panel), 150 000 (lower left panel) and 200 000 (lower right panel) trajectories.
The transition rate from \(x\) to \(x+z\) reads
\[w_{\phi}(z+x|x)=\frac{C_{\mu}}{|z|^{1+\mu}}\exp{[(\phi(x)-\phi(x+z))/2]}, \tag{23}\]
where the potential \(\phi\) has the form
\[\phi(x)=\left\{\begin{array}{ll}\sin^{2}(2\pi x),&|x|\leqslant 2;\\ (x^{2}-4),&|x|>2.\end{array}\right. \tag{24}\]
We consider \(\mu=1\). To optimize the simulation, here we use the same trick of isolating of "most dangerous" small \(z\) terms in the integrals involved in the Gillespie algorithm. For small \(z\) we expand the term \(\exp[(\sin^{2}(2\pi x)-\sin^{2}(2\pi(x+z)))/2]\) in Taylor series. We choose \(\varepsilon_{12}=0.05\). In the vicinity of \(|x|=2\) due attention must be paid to the proper power series truncation, to correctly choose the intervals where integration should be performed numerically. For example at \(x\in(1.95,2)\) we have
Quadratic Cauchy target: Time evolution of \(\rho(x,t)\) inferred from 200000 trajectories for \(\mu=0.5\) (left panel), \(\mu=1\) (middle panel) and \(\mu=1.5\) (right panel). All trajectories are started from \(x=0\). Note scale differences on vertical axes.
Cauchy target: Time evolution of pdf \(\rho(x,t)\) (left panel), half-width (HW) of \(\rho(x,t)\) (middle panel). Right panel reports the cumulative probability distributions (CPD) for different time instants. Here \(\mu=1\) and all data are inferred from 200000 trajectories, starting from \(x=0\).
\[\frac{1}{C_{1}}\int\limits_{\varepsilon_{1}}^{\varepsilon_{12}}w_{ \phi}(z+x|x)dz=\int\limits_{\varepsilon_{1}}^{2-x}\exp\bigg{[}\frac{\sin^{2}(2\pi x )-\sin^{2}(2\pi(x+z))}{2}\bigg{]}\frac{dz}{z^{2}}+\int\limits_{2-x}^{ \varepsilon_{12}}\exp\bigg{[}\frac{\sin^{2}(2\pi x)-(x+z)^{2}+4}{2}\bigg{]} \frac{dz}{z^{2}}\] \[=\int\limits_{\varepsilon_{1}}^{2-x}\exp\bigg{[}\frac{\sin^{2}(2 \pi x)-\sin^{2}(2\pi(x+z))}{2}\bigg{]}\frac{dz}{z^{2}}+\exp\bigg{[}\frac{\sin^{ 2}(2\pi x)-x^{2}+4}{2}\bigg{]}\int\limits_{2-x}^{\varepsilon_{12}}\exp\bigg{[} \frac{x^{2}-(x+z)^{2}}{2}\bigg{]}\frac{dz}{z^{2}}. \tag{25}\]
The numerators of integrand fractions have been expanded into Taylor series and (safely) truncated at the quadratic terms.
Time evolution of the inferred pdf \(\rho(x,t)\) is reported in All sample trajectories were started form \(x=0\) which corresponds to the \(\delta(x)\)-type initial distribution. The probability density spreads out with time in conformity with the trapping (confining) properties of the locally periodic enclosure (environment or "potential landscape"). For large running times t=400 the trajectory statistics produces data that are indistinguishable from those for the asymptotic pdf. We have checked that beginning from about 100 000 trajectories, further accumulation of the trajectories number like e.g. 200 000 (displayed) and 300 000 (not displayed) for the data statistics is inessential. In such cases the curves are almost the same, we merely improve a fidelity of the statistics.
## IV Conclusions
If a random process does not admit the description in terms of a stochastic differential equation (e.g. Langevin modeling), its direct numerical simulation becomes impossible by means of existing popular algorithms. In the present paper, for the first time in the literature, we propose a working method to generate stochastic trajectories (sample paths) of a random jump-type process without resorting to any explicit (or numerical) solution of a stochastic differential equation. To this end we have modified the Gillespie algorithm, normally devised for sample paths generation if the transition rates refer to a finite number of states of a system.
The essence of our modification is that we take into account the continuum of possible transition rates, thereby changing the finite sums in the original Gillespie algorithm into integrals. The corresponding procedures for stochastic trajectories generation has been changed accordingly. In other words, here we "extract" the background sample paths of a jump process, whose pdf obeys the transport equation (generalized Fokker-Planck dynamics),. We emphasize once more here, that we have focused on those background jump-type processes that cannot be modeled by any stochastic differential equation of the Langevin type.
Although heavy-tailed Levy stable drivers were involved in the present considerations, we have clearly confirmed that an enormous variety of stationary target distributions is dynamically accessible in each particular \(\mu\in\) case. That comprises not only a standard Gaussian pdf, casually discussed in relation to the Brownian motion (e.g. the Wiener process). Among heavy-tailed distributions, we have paid attention to the Cauchy pdf which can stand for an asymptotic target for any \(\mu\neq 1\) driver, provided a steering environment is properly devised. In turn, the Cauchy driver in a proper environment may lead to an asymptotic pdf with a finite (in fact arbitrarily large) number of moments, the Gaussian case being included ().
An example of the locally periodic environment has been considered as a toy model for more realistic physical systems. Our major hunch are strongly inhomogeneous "potential landscapes",, being sufficiently smooth to avoid a direct reference to random potentials,. Even if various mean field data are available in such (experimentally realizable) systems, it is of interest to have some knowledge about the microscopic dynamics (random paths) for the system under consideration. The detailed analysis of sample path data (ergodicity, mixing or lack of those properties) deserve a separate analysis.
We mention possible generalizations of our method to the Brownian motor concept (see, e.g., Ref.
Time evolution of \(\rho(x,t)\) inferred from 200000 trajectories at \(\mu=1\). The data for 100000 and 300000 trajectories (not displayed) do not show qualitative differences.
In those systems it is the properly tailored periodic "potential landscape" which enforces a conversion of a homogenous stochastic process (Brownian motion for reference) into the directed motion of particles at nanometer scales. That is closely related to the problem of so-called sorting in periodic potentials. Other problem to be addressed concerns ultracold atoms in optical lattices subject to random potentials, which might promising not only from a purely scientific point of view, but also with prospects for many technological applications. We note that the theoretical description of the above mentioned topics relies essentially on the Langevin-like equation input.
Our approach offers an immediate generalization for generalizations, where systems with non-Langevin response to external potentials may come into consideration, along with more traditional ones. What we actually need to implement our version of Gillespie's algorithm is the knowledge of jump transition rates of those random systems only.
A preliminary work (in progress) shows that an extension of our algorithm to higher dimensions is operational. In particular, the planar case is worth exploration, possibly with more complex "potential landscapes". While departing from final comments of Ref. we expect that the presented methodology can be effectively adopted to construct optimal random search routines, see in this connection.
|
10.48550/arXiv.1209.5882
|
Levy flights in confining environments: Random paths and their statistics
|
M. Zaba, P. Garbaczewski, V. Stephanovich
| 2,020
|
10.48550_arXiv.2103.06285
|
###### Abstract
We derive a formulation of mixed quantum-classical dynamics for describing electronic carriers interacting with phonons in reciprocal space. For dispersionless phonons, we start by expressing the real-space classical coordinates in terms of complex variables. A Fourier series over these coordinates then yields the reciprocal-space coordinates. Evaluating the electron-phonon interaction term through Ehrenfest's theorem, we arrive at a reciprocal-space formalism that is equivalent to mean-field mixed quantum-classical dynamics in real space. This equivalence is numerically verified for the Holstein and Peierls models, for which we find the reciprocal-space Hellmann-Feynman forces to involve momentum derivative contributions in addition to the position derivative terms commonly seen in real space. We close by presenting a proof of concept for the inexpensive modeling of low-momentum carriers interacting with phonons by means of a truncated basis in reciprocal space, which is not possible within a real space formulation.
_Introduction._ Real space and reciprocal space provide two alternative representations for describing quantum-mechanical phenomena in the condensed phase. Typically, finite-sized disordered solids such as molecules and molecular aggregates are characterized by localized quantum excitations and carrier dynamics dominated by an incoherent hopping between sites. Both aspects are described most effectively in real space. For (quasi)infinite and periodic solids such as highly-ordered crystals, on the other hand, quantum states take the form of Bloch waves, and carrier dynamics proceeds through "bandlike" transport, for which a reciprocal space representation is the most effective.
Materials for which reciprocal space representations have traditionally been adopted are typically characterized by a weak coupling between electronic carriers and nuclear vibrations (phonons). This coupling, which underpins the nonequilibrium dynamics of such materials, could therefore be adequately represented by theories truncating higher-order electron-phonon correlations. This is markedly different for molecular materials, where electron-phonon coupling is intermediate to strong. The need to accurately model interacting electrons and phonons in such materials has resulted in a plethora of quantum-dynamical methods, among which those involving an expansion and/or projection of the electron-phonon interaction terms, polaron transforms, and tensor-network decompositions, each conventionally represented in real space. These methods in principle retain a quantum treatment for all explicit degrees of freedom, which comes with high computational cost, especially when interactions are treated nonperturbatively. An inexpensive alternative is provided by mixed quantum-classical methods where phonons are treated classically, which enables one to describe nonperturbative and Markovian dynamics at the expense of taking the classical approximation. Mixed quantum-classical dynamics has proven successful in describing a broad variety of molecular phenomena.
The recent years have seen an increased interest in materials that are best represented in reciprocal space, but which feature intermediate to strong electron-phonon coupling. Examples of such materials include monolayer and few-layer variants of transition-metal dichalcogenides as well as layered and bulk hybrid metal-halide perovskites, both of which have risen to prominence due to their potential application as optoelectronic and quantum information devices. The rational engineering of such materials for technological purposes relies on a thorough understanding of their nonequilibrium properties, which requires the development of reciprocal-space models accounting for electron-phonon interactions beyond a truncated/perturbative level.
Here, we derive a mixed quantum-classical formalism tailored to interacting electronic carriers and phonons in reciprocal space. Assuming the local sites to form a periodic lattice, and the phonon modes to be dispersionless and harmonic, we first express the real-space classical phonon coordinates in terms of complex variables, which when taken as a Fourier series yields classical expressions in reciprocal space. These expressions are shown to be identical to those obtained by a Fourier transform of the quantum Hamiltonian followed by the classical approximation applied in reciprocal space. We evaluate the transformed electron-phonon interaction term by invoking Ehrenfest's theorem, effectively describing the Hellmann-Feynman forces acting on the classical modes of means of a mean-field average of the quantum state. Whereas in real space these forces commonly involve a position derivative term, an additional momentum derivative contribution is shown to appear in reciprocal space, which contributes to the classical equations of motion. We apply the resulting approach to the Holstein and Peierls models, both of which offer a straightforward comparison between our approach and the conventional real-space formulation of mean-field mixed quantum-classical dynamics, verifying the equivalence between the two formalisms. For both models, we present a proof of concept for the inexpensive modeling of low-momentum carriers interacting with phonons by means of a truncation of the Brillouin zone, which can only be realized in a reciprocal-space formulation.
_Classical phonons in reciprocal space._ For a one-dimensional lattice of harmonic, noninteracting, and dispersionless (Einstein) phonons the real-space quantum Hamiltonian is given by
\[\hat{H}_{\text{ph}}=\omega\sum_{n}\left(\hat{b}_{n}^{\dagger}\hat{b}_{n}+\frac{1}{ 2}\right)=\sum_{n}\left(\frac{1}{2}\hat{p}_{n}^{2}+\frac{1}{2}\omega^{2}\hat{q} _{n}^{2}\right)\!, \tag{1}\]
Applying the classical approximation to this Hamiltonian amounts to replacing the position and momentum operators by their classical coordinates,
\[\hat{q}_{n}\to q_{n},\quad\hat{p}_{n}\to p_{n}. \tag{2}\]
Introducing the classical equivalent of the ladder operators,
\[z_{n}\equiv\sqrt{\frac{\omega}{2}}\Big{(}q_{n}+i\frac{p_{n}}{\omega}\Big{)}, \tag{3}\]
the classical Hamiltonian can be expressed as
\[H_{\text{ph}}=\sum_{n}\left(\frac{1}{2}p_{n}^{2}+\frac{1}{2}\omega^{2}q_{n}^{ 2}\right)=\omega\sum_{n}z_{n}^{*}z_{n}. \tag{4}\]
Using Hamilton's equations for \(q_{n}\) and \(p_{n}\) it can be shown that the time derivative of \(z_{n}\) is given by \(\dot{z}_{n}=-i\omega z_{n}\).
Similarly to the ladder operators, a Fourier series over \(z_{n}\) yields its reciprocal-space equivalent as
\[z_{k}=\frac{1}{\sqrt{N}}\sum_{n}e^{i\hbar\omega}z_{n}. \tag{5}\]
Here, the lattice is assumed to be periodic and to consist of \(N\) sites, and \(k\) is the wavevector in units of \(2\pi/a\), with \(a\) as the lattice constant. (Note that in the following we will consistently use \(n\) to denote sites, and \(k\) and \(\kappa\) to denote wavevectors.) The classical Hamiltonian can then be expressed in reciprocal space as
\[H_{\text{ph}}=\omega\sum_{k}z_{k}^{*}z_{k}=\sum_{k}\left(\frac{1}{2}p_{k}^{2} +\frac{1}{2}\omega^{2}q_{k}^{2}\right)\!, \tag{6}\]
with the reciprocal "position" and "momentum" coordinates given by
\[\begin{split} q_{k}&=\frac{1}{\sqrt{2}\omega}(z_{k }+z_{k}^{*})=\frac{1}{\sqrt{N}}\sum_{n}\Big{(}q_{n}\cos(kn)-\frac{p_{n}}{ \omega}\sin(kn)\Big{)},\\ p_{k}&=-i\sqrt{\frac{\omega}{2}}(z_{k}-z_{k}^{*}) =\frac{\omega}{\sqrt{N}}\sum_{n}\Big{(}\frac{p_{n}}{\omega}\cos(kn)+q_{n}\sin (kn)\Big{)}.\end{split} \tag{7}\]
It is straightforward to show that \(\dot{z}_{k}=-i\omega z_{k}\), and that the time-evolution of \(q_{k}\) and \(p_{k}\) is governed by Hamilton's equations using \(H_{\text{ph}}\) given by Eq. 6. The result is formally equivalent to the real-space equations of motion, but is expressed entirely within reciprocal space.
It is worth noting that Eq. 6 is identical to the result obtained when first Fourier-transforming the quantum Hamiltonian to reciprocal space,
\[\hat{H}_{\text{ph}}=\omega\sum_{k}\left(\hat{b}_{k}^{\dagger}\hat{b}_{k}+\frac {1}{2}\right)=\sum_{k}\left(\frac{1}{2}\hat{p}_{k}^{2}+\frac{1}{2}\omega^{2} \hat{q}_{k}^{2}\right)\!, \tag{8}\]
This observation is a manifestation of the equivalence of canonical representations in classical-limit quantum mechanics.
_Mixed quantum-classical system._ In the following, we consider a system of interacting electronic carriers and phonons on a periodic lattice.
\[\hat{H}=\hat{H}_{\text{el}}+\hat{H}_{\text{ph}}+\hat{H}_{\text{el-ph}}. \tag{9}\]
1 or 8.
\[\hat{H}_{\text{el}}=-J\sum_{n}\left(\hat{c}_{n+1}^{\dagger}\hat{c}_{n}+\hat{c }_{n}^{\dagger}\hat{c}_{n+1}\right), \tag{10}\]
(Note that due to periodic boundaries, site \(N+1\) couples to site \(1\).) The reciprocal analog is given by
\[\hat{H}_{\text{el}}=-2J\sum_{k}\hat{c}_{k}^{\dagger}\hat{c}_{k}\cos(k). \tag{11}\]
Within the classical approximation for the phonons \(\hat{H}_{\text{ph}}=H_{\text{ph}}\), and the electron-phonon quantum Hamiltonian \(\hat{H}_{\text{el-ph}}\) depends parametrically on the phonon coordinates. Not only do the phonon coordinates impact the electronic quantum states through this parametric dependence, they also experience a "quantum force" due to the electronic states, which affects their classical equations of motion. The phonon momentum term is offdiagonal in real space, as a result of which only the phonon position contributes to \(\hat{H}_{\text{el-ph}}\).
\[F_{n}=-\left\langle\Psi|\nabla_{q_{n}}\hat{H}_{\text{el-ph}}|\Psi\right\rangle. \tag{12}\]
Here, \(\Psi\) is the electronic state that provides "feedback" to the phonon coordinates. Mixed quantum-classical methods vary in their choice of feedback state.
\[\dot{\Psi}=-i(\hat{H}_{\text{el}}+\hat{H}_{\text{el-ph}})\Psi. \tag{13}\]
This choice of feedback state is motivated by Ehrenfest's theorem stating that the classical analog of a quantum state behaves as the quantum expectation value with respect to that state.
From Eq. 7 it can be seen that the classical position and momentum coordinates become scrambled under the transformation to their reciprocal analogs.
\[\dot{p}_{k} =-\frac{\partial(H_{\text{ph}}+H_{\text{el-ph}})}{\partial q_{k}}=- \omega^{2}q_{k}-\langle\Psi|\nabla_{q_{k}}\hat{H}_{\text{el-ph}}|\Psi\rangle\,,\] \[\dot{q}_{k} =\frac{\partial(H_{\text{ph}}+H_{\text{el-ph}})}{\partial p_{k}}=p _{k}+\langle\Psi|\nabla_{p_{k}}\hat{H}_{\text{el-ph}}|\Psi\rangle\,. \tag{14}\]
Holstein and Peierls models.Using the above approach, we proceed by considering the Holstein and Peierls models, which differ by the form of the electron-phonon interaction Hamiltonian.
\[\hat{H}_{\text{el-ph}} =g\omega\sum_{n}\hat{c}_{n}^{\dagger}\hat{c}_{n}\left(\hat{b}_{n} ^{\dagger}+\hat{b}_{n}\right)=g\sqrt{2\omega^{3}}\sum_{n}\hat{c}_{n}^{\dagger }\hat{c}_{n}\hat{q}_{n}. \tag{15}\]
Here, \(g\) is the dimensionless coupling parameter, which relates to the vibrational reorganization energy as \(g^{2}\omega^{43}\).
\[\hat{H}_{\text{el-ph}} =\frac{g\omega}{\sqrt{N}}\sum_{k,\kappa}\hat{c}_{k+\kappa}^{ \dagger}\hat{c}_{k}\left(\hat{b}_{-\kappa}^{\dagger}+\hat{b}_{\kappa}\right)\] \[=\frac{g\sqrt{\omega}}{\sqrt{2N}}\sum_{k,\kappa}\hat{c}_{k+ \kappa}^{\dagger}\hat{c}_{k}\left(\omega(\hat{q}_{-\kappa}+\hat{q}_{\kappa})-i (\hat{p}_{-\kappa}-\hat{p}_{\kappa})\right). \tag{16}\]
Taking the classical approximation (either in real space or reciprocal space) yields the position and momentum derivative terms appearing in the reciprocal-space Hamilton's equations
\[\langle\Psi|\nabla_{q_{k}}\hat{H}_{\text{el-ph}}|\Psi\rangle =g\sqrt{\frac{2\omega^{3}}{N}}\Re\{C_{k}\},\] \[\langle\Psi|\nabla_{p_{k}}\hat{H}_{\text{el-ph}}|\Psi\rangle =-g\sqrt{\frac{2\omega}{N}}\Im\{C_{k}\}, \tag{17}\]
with the autocorrelation function
\[C_{k}\equiv\sum_{k^{\prime}}\langle\Psi|\hat{c}_{k+k}^{\dagger}\hat{c}_{k^{ \prime}}|\Psi\rangle\,. \tag{18}\]
Note that while \(\hat{H}_{\text{el-ph}}\) is complex-valued, the position and momentum derivative terms are purely real.
For the Peierls model (also known as Su-Schrieffer-Heeger model), the real-space phonon position coordinates couple linearly to the electronic nearest-neighbor interaction terms as
\[\hat{H}_{\text{el-ph}} =g\omega\sum_{n}\left(\hat{c}_{n}^{\dagger}\hat{c}_{n+1}+\hat{c} _{n+1}^{\dagger}\hat{c}_{n}\right)\left(\hat{b}_{n}^{\dagger}+\hat{b}_{n}- \hat{b}_{n+1}^{\dagger}-\hat{b}_{n+1}\right)\] \[=g\sqrt{2\omega^{3}}\sum_{n}\left(\hat{c}_{n}^{\dagger}\hat{c}_{n +1}+\hat{c}_{n+1}^{\dagger}\hat{c}_{n}\right)\left(\hat{q}_{n}-\hat{q}_{n+1} \right), \tag{19}\]
which in reciprocal space yields
\[\hat{H}_{\text{el-ph}} =2i\frac{g\omega}{\sqrt{N}}\sum_{k,\kappa}\hat{c}_{k+\kappa}^{ \dagger}\hat{c}_{k}\left(\hat{b}_{-\kappa}^{\dagger}+\hat{b}_{\kappa}\right) \left(\sin(k+\kappa)-\sin(k)\right)\] \[=\frac{g\sqrt{2\omega}}{\sqrt{N}}\sum_{k,\kappa}\hat{c}_{k+ \kappa}^{\dagger}\hat{c}_{k}\left(i\omega(\hat{q}_{-\kappa}+\hat{q}_{\kappa})+ \left(\hat{p}_{-\kappa}-\hat{p}_{\kappa}\right)\right)\] \[\qquad\times\left(\sin(k+\kappa)-\sin(k)\right), \tag{20}\]
This yields the position and momentum derivative terms
\[\langle\Psi|\nabla_{q_{k}}\hat{H}_{\text{el-ph}}|\Psi\rangle =-g\sqrt{\frac{2\omega^{3}}{N}}\Im\left\{C_{k}^{\prime}\right\},\] \[\langle\Psi|\nabla_{p_{k}}\hat{H}_{\text{el-ph}}|\Psi\rangle =-g\sqrt{\frac{2\omega}{N}}\Re\left\{C_{k}^{\prime}\right\}, \tag{21}\]
with the modulated autocorrelation function
\[C_{k}^{\prime} \equiv\sum_{k^{\prime}}\langle\Psi|\hat{c}_{k^{\prime}+k}^{ \dagger}\hat{c}_{k^{\prime}}|\Psi\rangle\left(\sin(k^{\prime}+k)-\sin(k^{ \prime})\right). \tag{22}\]
As seen above, within the Holstein and Peierls models we find the mixed quantum-classical equations of motion to assume simple forms in both real space and reciprocal space. It is therefore straightforward to numerically verify the equivalence between the real-space and reciprocal-space formulations.
\[|\Psi\rangle=|k=0\rangle=\frac{1}{\sqrt{N}}\sum_{n}|n\rangle\,. \tag{23}\]
Such an initial condition could be representative of a tightly-bound electron-hole pair (Frenkel exciton) created upon impulsive optical excitation, in which case \(n\) represents the exciton location.
\[P(\{q_{n},p_{n}\})\propto\prod_{n}\exp\left(-\beta\frac{1}{2}\left(p_{n}^{2}+ \omega^{2}q_{n}^{2}\right)\right), \tag{24}\]
Through a Fourier transform of this expression one finds an identical distribution for the reciprocal-space coordinates \(q_{k}\) and \(p_{k}\).
Parameters are expressed without units to keep the discussion general, but we note that when taking the thermal energy at room temperature (\(T=293\) K) as a reference, a unit of energy amounts to 25 meV and a unit of time to 164 fs. Results were obtained by propagating the classical and quantum coordinates using a fourth-order Runge-Kutta algorithm with a time step of \(\Delta t=0.02\). All results have been averaged over 12 000 thermal initial conditions for the classical coordinates. Shown in (a) are time-dependent reciprocal-space populations of the electronic carrier \(P_{k}(t)\equiv|\langle k|\Psi(t)\rangle|^{2}\), calculated by solving the equations of motion for all coordinates in reciprocal space. The electronic state, upon initiating at \(k=0\), can be seen to rapidly broaden in reciprocal space due to scattering with the phonon modes, while equilibrating within a time span of \(t\sim 2\). Throughout, the populations are seen to remain symmetric with respect to inversion of \(k\), as a result of the underlying Hamiltonians and (thermally-averaged) initial conditionsbeing conserved under this symmetry operation. Also notable is that the majority of the populations remain concentrated at \(k=0\). The reason for this behavior is that the electronic band is minimized at this Brillouin zone location, and that the relatively large number of phonon modes forces the electronic carrier to relax to a quasi-thermal equilibrium, which favors low energies. It should be pointed out that mean-field mixed quantum-classical dynamics is known violate detailed balance, but nevertheless induces a low-energy bias.
(b) compares the time-dependent populations at \(k=0\), \(P_{0}(t)\), resulting from the reciprocal-space and real-space mixed quantum-classical formalisms. (The local populations coming out of the real-space formulation have been Fourier-transformed in order to yield \(P_{0}(t)\).) The lack of conceivable differences between the shown data confirms the equivalence between the equations of motion in both formalisms. We found this result to be independent on the choice of parameters.
Shown in are results analogous to Fig. 1, but for the Peierls model. Again, a rapid broadening in reciprocal space is observed. Similarly to the Holstein model, the formal equivalence between the real-space and reciprocal-space mixed quantum-classical equations of motion yields a perfect agreement between results obtained in both representations.
_Brillouin zone truncation._ The appeal of a real-space representation is that local phenomena can be accurately described at manageable computational cost by truncating the degrees of freedom beyond those within the spatial domain of interest. A similar truncation can be applied to reciprocal space for phenomena that occur within a limited domain of the Brillouin zone. Electronic carriers may remain confined to a limited domain when the time scales under consideration do not allow them to cover the full Brillouin zone, or when they thermally relax to band minima. This principle has been been utilized in static calculations of exciton and trion states in monolayer transition-metal dichalcogenides, where a truncation radius around the \(K\) points in the Brillouin zone was imposed. To the best of our knowledge, applications of similar reciprocal-space truncation schemes in dynamical calculations have remained limited.
From the results in Figs. 1 (a) and 2 (a) it can be seen that the electronic populations continuously remain concentrated around \(k=0\) for both Holstein and Peierls models, owing to the \(k=0\) initial condition and as well as the band minimum being located here. This suggests that accurate results are retained by restricting the Brillouin zone to within a truncation radius, \(k_{0}\), such that the electronic basis states and phonon modes are limited to those having \(|k|<k_{0}\). To demonstrate that this is indeed the case, we show in Figs. 1 (b) and 2 (b) results for various values of \(k_{0}\). As can be seen, for the Holstein model accurate dynamics is obtained even upon halving the Brillouin zone. For a similar truncation within the Peierls model, discrepancies can be seen to emerge around \(t\approx 2\), but the short and long-time dynamics are still reasonably accurate.
Same as but for the Peierls model.
Transient electronic populations \(P_{k}(t)\) calculated within the mean-field mixed quantum-classical method for the Holstein model with \(J=1.0\), \(\omega=0.1\), \(g^{2}=5.0\), and \(T=1.0\). Shown in (a) are \(k\)-dependent populations obtained through the reciprocal-space formulation of the method. Shown in (b) is \(P_{0}(t)\) obtained through the real-space (markers) and reciprocal-space (solid) formalisms. Also shown are reciprocal-space results obtained upon truncating the Brillouin beyond \(|k|<k_{0}\).
23.
_Discussion and conclusions._ While mixed quantum-classical methods are conventionally formulated in real space, we have shown here that for a periodic lattice an equivalent reciprocal-space formulation can be obtained by taking a Fourier series over the real-space classical coordinates expressed in terms of complex variables. As such, we have arrived at an approach tailored to describing bandlike phenomena such as electronic carriers interacting with phonons in crystalline solids. Some of the benefits offered by a reciprocal-space representation are illustrated in the present study by its proof of concept for the accurate modeling of low-momentum carriers using a truncated Brillouin zone.
We have specifically considered a mean-field approach to self-consistently describe the electron-phonon interactions, in which case we find the real-space and reciprocal-space mixed quantum-classical formalisms to be equivalent. In verifying this equivalence through numerical calculations for the Holstein and Peierls models, we have chosen the parameters to be such that the intermediate coupling regime is reached (where perturbative approaches loose validity), and that the classical approximation should hold reasonably well (\(T\gg\omega\)). It is important to reiterate, however, that the real-space and reciprocal-space results come out identically regardless of the choice of parameters.
While it is tempting to assess the accuracy of the dynamics obtained in this work against exact results, it is important to note that the reciprocal-space formulation is merely a convenient representation of a method that has been well-characterized in real space, and as such its accuracy should be no different than that of the real-space variant. In particular, mean-field mixed quantum-classical dynamics has previously been shown to yield inaccurate equilibrium populations as a result of its violation of detailed balance. Indeed, for the parameters employed in the present work we found the quantum populations to equilibrate towards an effective temperature significantly exceeding \(T\). It would therefore be worthwhile to consider alternative formulations for the electron-photon interaction term known to yield an improved description of detailed balance. Interestingly, this would considerably enhance the long-time accuracy for the truncated data shown in (b), as a decrease in the effective temperature will localize the quantum populations to a narrower domain around \(k=0\). It is also noteworthy that the numerical efficiency for obtaining accurate results upon Brillouin zone truncations is expected to increase with higher dimensionality, for which a generalization of our formalism is trivial.
Upon obtaining the reciprocal-space mixed quantum-classical equations of motion through a Fourier transform of the real-space classical phonon coordinates, we observed that identical equations would be obtained upon taking the classical approximation within reciprocal space (i.e., upon first Fourier-transforming the quantum phonon coordinates). This opens the opportunity to perform mixed quantum-classical modeling of reciprocal space Hamiltonians that have no simple analog in real space, such as those having dispersed phonons. More broadly speaking, it would be interesting to explore the ideas presented in the present work in the context of anharmonic modes and nonlinear electron-phonon coupling. It is conceivable that some of the many benefits of classical molecular dynamics can be harnessed while taking advantage of a reciprocal-space representation.
|
10.48550/arXiv.2103.06285
|
A Reciprocal-Space Formulation of Mixed Quantum-Classical Dynamics
|
Alex Krotz, Justin Provazza, Roel Tempelaar
| 3,699
|
10.48550_arXiv.1609.01201
|
###### Abstract
Using infrared spectroscopy combined with _ab initio_ methods we study reactions of H\({}_{2}\)O and CO inside the confined spaces of Zn-MOF-74 channels. Our results show that, once the water dissociation reaction H\({}_{2}\)O \(\rightarrow\) OH+H takes place at the metal centers, the addition of 40 Torr of CO at 200 \({}^{\circ}\)C starts the production of formic acid via OH+H+CO \(\rightarrow\) HCO\({}_{2}\)H. Our detailed analysis shows that the overall reaction H\({}_{2}\)O+CO \(\rightarrow\) HCO\({}_{2}\)H takes place in the confinement of MOF-74 without an external catalyst, unlike the same reaction on flat surfaces. This discovery has several important consequences: It opens the door to a new set of catalytic reactions inside the channels of the MOF-74 system, it suggests that a recovery of the MOF's adsorption capacity is possible after it has been exposed to water (which in turn stabilizes its crystal structure), and it produces the important industrial feedstock formic acid.
## I Introduction
Metal organic framework (MOF) materials are porous crystals widely studied for important applications and industrial processes such as gas storage and sequestration, molecular sensing, polymerization, luminescence, non-linear optics, magnetic networks, targeted drug delivery, multiferroics, and catalysis, In particular, MOF-74 [\(\mathcal{M}_{2}\)(dobdc), \(\mathcal{M}=\) Mg\({}^{2+}\), Zn\({}^{2+}\), Ni\({}^{2+}\), Co\({}^{2+}\), and dobdc=2,5-dihydroxybenzenedicarboxylic acid] has shown great potential for the adsorption of small molecules such as H\({}_{2}\), CO\({}_{2}\), N\({}_{2}\), and CH\({}_{4}\), among others.
The favorable reactivity of MOF-74 has been widely studied. For example, Co-MOF-74 exhibits a catalytic activity towards CO oxidation, originating from the high density of lewis acidic coordinatively unsaturated sites and the MOF's porosity. The inclusion of Co atoms into Ni-MOF-74 results in a mixed system (Co/Ni-MOF-74) that shows activity towards the oxidation of cyclohexene, where the catalytic performance of the mixed system is higher than the one of pure Co-MOF-74. On the other hand, our previous results have shown that several members of the MOF-74 family are able to catalyze the dissociation of water into H and OH groups (H\({}_{2}\)O \(\rightarrow\) OH+H, see Fig. 1) at low temperatures and pressures, i.e. above 150 \({}^{\circ}\)C and at 8 Torr of H\({}_{2}\)O. This particular catalytic reaction is responsible for the loss of crystal structure and adsorption capacity after exposure of MOF-74 to water, and constitutes one of the main hurdles for wide-spread applications of MOFs in general and MOF-74 in particular. This challenge has motivated our efforts to look for new catalytic reactions inside the confined channels of MOF-74, further reacting the undesirable products of the H\({}_{2}\)O \(\rightarrow\) OH+H reaction in order to overcome these hurdles.
In this work, we show that introducing CO molecules into the pores of MOF-74--after the H\({}_{2}\)O \(\rightarrow\) OH+H reaction has taken place--enables the reaction OH+H+CO \(\rightarrow\) HCO\({}_{2}\)H. Our results show that the overall reaction H\({}_{2}\)O+CO \(\rightarrow\) HCO\({}_{2}\)H takes place in the confinement of the MOF without an external catalyst, with a number of important consequences: First, it showcases the reactivity inside the well-controlled and isolated environment of the MOF-74 channels. This aspect is very important, as the confinement of the MOF-74 environment catalyzes reactions that would otherwise require very high pressure, bringing significant simplifications for experiments and possible MOF applications.
Zn-MOF-74 with its hexagonal channels clearly visible. The open-metal sites at the corners form the primary adsorption sites. The arrow indicates how the H of the water is transferred to the O of the linker during the H\({}_{2}\)O \(\rightarrow\) OH+H reaction. Black, red, white, and blue spheres represent C, O, H, and Zn atoms. The box shows the portion of MOF-74 visible in other figures, albeit from a slightly different angle.
In turn, it increases the crystal structure stability of MOF-74 by removing the OH and H groups that cause the instability (note that, due to their strong binding, those groups cannot be removed by thermal activation). And finally, it binds the toxic CO and produces formic acid, a non-toxic liquid with 4.4 wt% hydrogen and thus a promising hydrogen carrier and an important feedstock medical/industrial chemical. The use of Pd as catalytic material in direct formic acid fuel cells has brought interesting developments in this area, highlighting formic acid as a valuable asset for a hydrogen economy.
## II Experimental and theoretical methods
### Zn-Mof-74
Out of the isostructural \(\mathcal{M}\)-MOF-74 family, Zn-MOF-74 exhibits the highest catalytic activity towards the H\({}_{2}\)O \(\rightarrow\) OH+H reaction. We thus use this system to study the H\({}_{2}\)O+CO \(\rightarrow\) HCO\({}_{2}\)H reaction through a combination of _ab initio_ simulations and experiments.
### Hydrogen vs.
Only recently, our work showed direct evidence of the water dissociation reaction H\({}_{2}\)O \(\rightarrow\) OH+H at the metal centers of MOF-74 above 150 \({}^{\circ}\)C. In this reaction, the water first binds to an open-metal site and then donates one H to the nearby O at the linker; the remaining OH group stays at the open-metal site, see Interestingly, this reaction can only be observed when heavy water D\({}_{2}\)O is used. Its fingerprint is a sharp peak at 970 cm\({}^{-1}\) in the IR spectrum, corresponding to the O-D vibration at the linker. When H\({}_{2}\)O is used instead, the peak appears at a higher frequency, where it couples with and is masked by the vibrational modes of the MOF and becomes impossible to detect. Therefore, the main focus of our experiments is on the water reaction with D\({}_{2}\)O. We refer to the resulting deuterated formic acid as FA(D). Nonetheless, we do show that the reaction also occurs with H\({}_{2}\)O, referring to the resulting formic acid as FA(H). For simplicity, throughout the text we may generally say _water_, even when experiments are done with _heavy water_.
### Experimental Details and Procedure
Our experiments are divided into 3 steps:
## (i) Preparation and activation of the sample:
Zn-MOF-74 powder (\(\sim\)2 mg) was pressed onto a KBr pellet (\(\sim\)1 cm diameter, 1-2 mm thick). The sample was placed into a high-pressure high-temperature cell (product number P/N 5850c, Specac Ltd, UK) at the focal point of an infrared spectrometer (Nicolet 6700, Thermo Scientific, US). The sample was activated under vacuum at 180 \({}^{\circ}\)C for 4 hours and then cooled down to room temperature to measure CO\({}_{2}\) absorption by introducing 6 Torr of CO\({}_{2}\) into the cell until saturation (30 minutes). Then, the area under the peak at 2338 cm\({}^{-1}\) was determined, which is a characteristic peak of CO\({}_{2}\) adsorbed on the Zn site and thus a quantitative measure of the CO\({}_{2}\) uptake. Thereafter, the cell was evacuated under vacuum (\(<\) 20 mTorr) at a temperature of 150 \({}^{\circ}\)C for a period of 4 hours.
## (ii) Dissociation reaction:
The sample was heated to 200 \({}^{\circ}\)C. 8 Torr of D\({}_{2}\)O were then introduced into the cell until saturation occurred (8 hours) to start the dissociation reaction. Spectra were recorded as a function of time during the adsorption process to evaluate the 970 cm\({}^{-1}\) peak, i.e. the fingerprint of the D\({}_{2}\)O \(\rightarrow\) OD+D reaction. Thereafter, evacuation under vacuum (\(<\) 20 mTorr) for a period of 4 hours at 150 \({}^{\circ}\)C was required to evacuate the water gas phase completely and avoid further reaction. Note that this temperature is not high enough to also remove the OD and D products of the dissociation reaction. Then, at room temperature, CO\({}_{2}\) adsorption was measured again and the cell was evacuated as in step (i).
## (iii) Formic acid production and removal:
The temperature in the cell was raised back to 200 \({}^{\circ}\)C and 40 Torr of CO were introduced for 1 hour to start the formic acid production. Spectra were recorded. Thereafter, the cell was evacuated for 3 hours under vacuum (\(<\) 20 mTorr) at 200 \({}^{\circ}\)C, removing the formic acid and unreacted CO, while spectra were recorded. Then, CO\({}_{2}\) adsorption at room temperature was measured and the cell was evacuated as in step (i). This production and removal step was repeated two times and we refer to each occurrence as _removal 1_ and _removal 2_.
### Computational Details
_Ab initio_ modeling was performed at the density functional theory level, using quantum espresso with the vdW-DF functional. Ultrasoft pseudo potentials were used with cutoffs of 544 eV and 5440 eV for the wave functions and charge density. Due to the large dimensions of the unit cell, only the \(\Gamma\)-point was used. During relaxations all atom positions were optimized until forces were less than 2.6\(\times\)10\({}^{-4}\) eV/A. Reaction barriers were found with a transition-state search algorithm, i.e. the climbing-image nudged-elastic band method. The primitive cell of our pristine Zn-MOF-74 system contained 54 atoms and has space group R3. Additional atoms/molecules were added as appropriate for the reactants. The rhombohedral axes are \(a=b=c=15.105\) A and \(\alpha=\beta=\gamma=117.78^{\circ}\).
## III Results and Discussion
### Confirming Formic Acid Production and Removal
We begin by showing experimental evidence that the reactive environment inside the MOF-74 channels catalyses the formic acid production (through water dissociation) OH+H+CO \(\rightarrow\) HCO\({}_{2}\)H after the water dissociation H\({}_{2}\)O \(\rightarrow\) OH+H has taken place. To this end, we follow the three-step procedure outlined in Sec. II.3. After the introduction of CO in step (iii), our IR spectra in clearly show the presence of FA(D) and FA(H) molecules. As expected, due to the deuterium presence, the FA(D) peaks (CD\({}_{\rm beno}\) and OD\({}_{\rm beno}\)) are red shifted with respect to FA(H) peaks (CH\({}_{\rm beno}\) and OH\({}_{\rm beno}\)) by a factor of \(\sim\)1.4. C-O and C=O modes are less disturbed (shifted), as they are not directly affected by the presence of deuterium or hydrogen. The OH\({}_{\rm beno}\) vibrational mode signal at \(\sim\)1550 cm\({}^{-1}\) for FA(H) appears very close to a strong MOF mode at \(\sim\)1530 cm\({}^{-1}\), and this vibrational mode may be contributing to the OH\({}_{\rm beno}\) signal. On the other hand, the signal at \(\sim\)1530 cm\({}^{-1}\) in the FA(D) spectrum may be due to a hydrogen contamination of the deuterated water, increased by the vibrations of the MOF modes.
In we show how the characteristic peaks of FA(D) disappear as a function of time during the removal in step (iii), showing that the produced formic acid can readily be removed. Note that these experiments rely on the detection of the linker O-D mode at 970 cm\({}^{-1}\) and are thus only performed for the deuterated case (see Sec. II.2). We will henceforth only discuss the deuterated case. It is interesting to note that--while the starting point for _removal 1_ and \(2\) are comparable to within 6%--the desorption becomes faster. For example, in the former case 35% of FA(D) was removed after 20 min, while in the latter 62% was removed during the same time. This fact, together with the fact that several removals are necessary to react all OD and D groups suggests a bottleneck in diffusion of the reactants and products, discussed further below.
After the water dissociation reaction happens, its products (OD or OH) are strongly bound to primary adsorption site in the MOF and take up valuable adsorption sites. This undesirable decrease of the MOF's adsorption capacity is well known and unfortunately limits the applicability of MOF materials to non-humid environments. Note that the water dissociation products bind so strongly to the MOF that a simple removal through activation is not possible before the MOF dissoci integrates. Other means to recover the uptake capacity of MOFs after exposure to water are thus highly desirable. Our production and removal of formic acid reacts those unwanted groups that are otherwise bound to the MOF after the water dissociation reaction. We now show that this process also partially restores the MOF's small-molecule uptake capacity. In we track the 970 cm\({}^{-1}\) peak (a measure for the amount of dissociated heavy water present in the MOF cavity) as well as the 2338 cm\({}^{-1}\) peak (a measurement of the CO\({}_{2}\) adsorption capacity) at different stages of our experiment. We see that the former decreases as we introduce CO into the system, i.e by 1.6% after _removal 1_ and 7.9% after _removal 2_. This confirms that we have successfully removed the D groups from the linkers of the MOF. On the other hand, the latter--after an expected big reduction in the CO\({}_{2}\) uptake capacity after the D\({}_{2}\)O dissociation (22%)--increases by 1.5% and 5.1% after _removal 1_ and \(2\).
IR absorption spectra of Zn-MOF-74 reacted with D\({}_{2}\)O (top panel) and H\({}_{2}\)O (bottom panel) followed by the addition of 40 Torr of CO for 1 hour at 200 \({}^{\circ}\)C. The panels show the characteristic peaks of FA(D) and FA(H). Both samples are referenced to pure Zn-MOF-74.
IR spectra of the desorption of FA(D) as a function of time for _removal 1_ and _removal 2_. Both figures are referenced to pure Zn-MOF-74.
As expected, the decrease in the amount of dissociated water (area under the peak at 970 cm\({}^{-1}\)) goes hand-in-hand with the increase of the CO\({}_{2}\) uptake capacity (area under the peak at 2338 cm\({}^{-1}\)). However, it is interesting to see that more than one removal cycle is necessary to restore a significant amount of uptake capacity. In principle, the partial pressure of 40 Torr CO introduced into the system should be more than enough (we estimate that it results in at least 6 CO molecules per unit cell) to react all OD and D groups. However, this is not the case, see We conclude that the produced formic acid inhibits diffusion of CO deeper into the bulk. After each removal of formic acid and the renewed introduction of CO, the process picks up where it had left off earlier, working from the MOF surface into the bulk until, eventually, all OD and D groups have been reacted. Work to reduce the number of cycles thus needs to focus on diffusion in MOF-74 as well as using similar reactions with different products.
The CO region in the IR spectrum also provides information on the mechanism of formic acid formation. shows the CO region during _removal 1_ at several stages. When CO gas is still inside the MOF, the predominant signal is at 2173 cm\({}^{-1}\). However, after 1 min of desorption, a shift to lower frequencies (2150 cm\({}^{-1}\)) is observed. This indicates that the majority of the CO gas phase has been evacuated and now the IR spectrum is dominated by a signal that suggests a stronger interaction between the CO and the MOF, such as in the CO\({}_{2}\)H+H state (see Fig. 6). After longer periods of desorption, the intensity of the signal is reduced as the chamber is evacuated.
Before we continue to study the nature of the reaction, we give an estimate of how much formic acid is produced. From our calculations we know that the crystal density of Zn-MOF-74 is 1.231 g/mL, with a volume of 3944.65 A\({}^{3}\) for the hexagonal cell (note that the hexagonal cell contains 18 metal centers and is three times bigger than the rhombohedral representation). Based on that, we calculate that in our sample (2 mg of Zn-MOF-74) we have 4.12\(\times\)10\({}^{17}\) hexagonal unit cells. According to Fig. 4, we observe a reduction of 22% in the CO\({}_{2}\) adsorption capacity, suggesting that we produced \(\sim\)4 OD+D groups every 18 metal centers. Therefore, when CO is introduced into the cell and 5.1% of the CO\({}_{2}\) adsorption capacity is recovered, we estimate a production of 3.95\(\times\)10\({}^{17}\) formic acid molecules. This corresponds to 2.311\(\times\)10\({}^{-5}\) mL of formic acid in the 2 mg of Zn-MOF-74, or 11.55 \(\mu\)L/gMOF. Clearly, this is a small quantity, but as mention before, our goal is to investigate the chemistry in the confined spaces of Zn-MOF-74 and not the mass production of formic acid.
### Pathway of the Formic Acid Reaction
We now investigate the nature of the formic acid reaction H\({}_{2}\)O+CO \(\rightarrow\) HCO\({}_{2}\)H and give and explanation of how it takes place. We know that the first step is the dissociation of water at the metal centers H\({}_{2}\)O \(\rightarrow\) OH+H, which we have studied in detail before. We find that the water dissociation takes place above 150 \({}^{\circ}\)C with an energy barrier up to 1.09 eV, depending on the number of water molecules involved in the reaction. The second step of the reaction starts by the introduction of CO at 200 \({}^{\circ}\)C, which catalyses the OH+H+CO \(\rightarrow\) HCO\({}_{2}\)H reaction. Based on this information, and taking into account that the metal centers are poisoned by the OH groups after the H\({}_{2}\)O \(\rightarrow\) OH+H reaction, we propose the following mechanism for the overall reaction: Once the H\({}_{2}\)O \(\rightarrow\) OH+H reaction takes place, the added CO molecules interact with the OH groups at the metal centers to form CO\({}_{2}\)H adsorbed at the metal center.
Integrated areas of the 970 cm\({}^{-1}\) peak (a measure of the amount of dissociated water) and 2338 cm\({}^{-1}\) peak (a measure of the CO\({}_{2}\) uptake capacity).
CO region of the IR spectrum during _removal 1_. The black line is taken just before CO evacuation and is measured on a scale of 0.005, since the CO gas-phase signal is very strong. Thereafter, IR spectra are taken at 1, 10, 20, and 60 minutes during evacuation, measured on the smaller scale of 0.0002.
Overall, the reaction pathway follows H\({}_{2}\)O+CO \(\rightarrow\) OH+H+CO \(\rightarrow\) CO\({}_{2}\)H+H \(\rightarrow\) HCO\({}_{2}\)H, as depicted in
We now use our _ab initio_ transition-state search to find the energetically most favorable pathway (i.e. lowest energy barriers) for our proposed reaction pathway. Results for the structures of reactants, stable states, and products are depicted in and the energy profile along the entire reaction is plotted in The first step of the reaction is the endothermic water dissociation H\({}_{2}\)O+CO \(\rightarrow\) OH+H+CO. We have previously calculated its reaction barrier (1.09 eV) and confirmed the separate OH (bound to the open-metal site) and H (bound to the O of the linker) experimentally inside MOF-74 above 150 \({}^{\circ}\)C. Thereafter, the reaction proceeds exothermic via OH+H+CO \(\rightarrow\) CO\({}_{2}\)H+H \(\rightarrow\) HCO\({}_{2}\)H, resulting in the formation of formic acid adsorbed on the metal centers of Zn-MOF-74. Our calculations show that the energy barrier between the states OH+H+CO and CO\({}_{2}\)H+H is 0.8 eV, while the barrier between CO\({}_{2}\)H+H and HCO\({}_{2}\)H is 1.04 eV. The final state, i.e. HCO\({}_{2}\)H, has an energy 0.23 eV lower than the energy of the initial state H\({}_{2}\)O+CO, and the formic acid binds to the metal centers with an energy of 0.68 eV (65.61 kJ/mol), comparable to the binding of other molecules to Zn-MOF-74.
The overall barrier for the reaction in is significant and explains why only a small amount of formic acid is produced. But, that barrier corresponds to the presence of only one water molecule. In related work, we show that the barrier to the first step of the reaction is lowered by 37% when the water molecules create clusters above the linkers. It is conceivable that the presence of several CO and/or H\({}_{2}\)O molecules can also lower the energy barrier of the OH+H+CO \(\rightarrow\) HCO\({}_{2}\)H reaction. However, due to the large number of stable geometries and possible paths for the H\({}_{2}\)O+CO \(\rightarrow\) HCO\({}_{2}\)H reaction when more than one molecule is involved, a comprehensive _ab initio_ transition-state search becomes computational prohibitively expensive.
## IV Conclusions
Our experimental and theoretical work confirms that we can use the OH and H groups--produced by the H\({}_{2}\)O \(\rightarrow\) OH+H reaction--to start a new reaction mechanism catalyzed inside the confined environment of the Zn-MOF-74 channels through water dissociation and produce formic acid via H\({}_{2}\)O+CO \(\rightarrow\) HCO\({}_{2}\)H. This discovery has several important consequences: It opens the door to a new set of catalytic reactions inside a well controlled system (MOF-74), it provides a proof-of-principle that a recovery of the adsorption capacity and structural stability of Zn-MOF-74 is possible after exposure to water, and finally it produces the important medical/industrial feedstock formic acid.
Structures for the reactants (H\({}_{2}\)O+CO), stable states (OH+H+CO, CO\({}_{2}\)H+H), and products (HCO\({}_{2}\)H). First, the water is adsorbed at the open-metal site, while CO is adsorbed at a secondary site. Then, the water dissociates into OH+H; OH remains at the open-metal site and H is transferred to the O at the linker; CO is still adsorbed at the secondary site. Then, CO reacts with OH to form CO\({}_{2}\)H at the metal center. Finally, the H from the linker reacts with the CO\({}_{2}\)H to form HCO\({}_{2}\)H at the open-metal site.
Energy of the stable and transition states (TS) along the H\({}_{2}\)O+CO \(\rightarrow\) HCO\({}_{2}\)H reaction.
Acknowledgements
This work was supported by DOE Grant No. DE-FG02-08ER46491. Furthermore, this research used computational resources of the OLCF at ORNL, which is supported by DOE grant DE-AC05-00OR22725.
|
10.48550/arXiv.1609.01201
|
Chemistry in confined spaces: Reactivity of the Zn-MOF-74 channels
|
S. Zuluaga, E. M. A. Fuentes-Fernandez, K. Tan, C. A. Arter, J. Li, I. J. Chabal, T. Thonhauser
| 2,114
|
10.48550_arXiv.1706.08509
|
## Introduction:
Room-temperature ionic liquids (RTILs) are a vast class of ionic systems, usually consisting of an organic cation and either an organic or inorganic anion, whose melting temperature falls below the conventional limit of 100 \({}^{\circ}\)C (Welton 1999). They have been intensively investigated for their potential applications as solvents, non-aqueous electrolytes, high-performance lubricants, and advanced engineering materials (Plechkova et al. 2009; Ranke et al. 2007b; Ghandi 2014). The widespread appeal of RTILs to some extent relies on their perceived low environmental impact, making these compounds one of the bases of the so-called "green-chemistry". Their introduction in industrial processes together with their organic character, in turn, motivated the first studies of their interaction with biomolecules, and bio-organisms (Petkovic et al. 2011). As a result, several studies have highlighted their toxicity on living organisms (Pretti 2006 ; Bernot et al. 2005; Ranke et al. 2006, 2007a; Stolte et al. 2007; Kulacki and Lamberti 2008). Toxicity is also a measure of the high affinity between RTILs and bio-systems. This affinity, in turn, together with the extremely tunable chemistry of RTILs, is at the basis of the great future of RTILs in applications from pharmacology, to bio-medicine, and, broadly speaking, in bio-nano-technology (Stoimenovski et al. 2012; Hough et al. 2007). It has been already shown, for example, that RTILs are able to:
* kill bacteria;
* extract, purify and even store DNA at ambient temperature;
* stabilize proteins and enzymes;
* either help or prevent protein amyloidogenesis, in same cases even turning amyloid fibres back to functional proteins;
* penetrate, create pores, and destroy biomembranes;
* dissolve cellulose and other complex polysaccharides.
A general overview of the interaction between RTILs and several classes of biomolecules (e.g. proteins and peptides, mono- and poli-saccharides, nucleic acids, and biomembranes) has been presented in two mini-review published in 2016 (Benedetto et al. 2016a, b). The main aim of these studies has been to link the biological effects of RTILs - usually detected by bio-chemical approaches - to the microscopic mechanisms of interaction between RTILs and biomolecules, with the idea that only a complete understanding of their microscopic mechanisms can provide the basis to (a) synthetize _greener_ RTILs for industrial applications, and (b) develop breakthrough applications in bio-nano-technology.
Having this _big-picture_ in mind, the interaction of RTILs with biomembranes is one of the most relevant topics. Since the first encounter of any foreign chemical species with a living cell is likely to occur at its protective biomembrane, this subject merges the two original aims of these investigations, i.e. assessing and reducing RTILs cytotoxicity, and developing bio-nano-technologies. For example, it has been shown and suggested that:
* controlling the portion of biomembranes by RTILs could result in the development of new drug-delivering methods, with an apparent impact in pharmacology;
* even more, since their chemistry can be finely tuned in such a way RTILs can be lethal to bacteria at the same doses that are harmful to eukaryotic cells, they can open new antibacterial strategies, something that is quite needed nowadays;finally, tuning their chemistry already gave RTILs able to kill cancer cells and leave healthy cells almost unaffected.
Moreover, the water-like environment in which biomembranes reside are rich in a variety of (inorganic) ions, that play a major role in promoting and regulating the biomembrane functions. The effect of simple ions such as Na\({}^{*}\), K\({}^{+}\), Cl-, etc. has been studied extensively with experimental and computational methods (Berkowitz and Vacha 2012; Pabst et al. 2007; Bockmann et al. 2003) as well as by empirical modeling (Aroti et al. 2007). It is natural, therefore, to turn the attention to RTILs, i.e. organic salts, whose complex structure and larger size provide many more ways to tune their interaction.
This mini-review is dedicated to recent results obtained in the study of the interaction between RTILs and biomembranes, with a special focus on their chemical-physical properties.
## From complex biomembranes to phospholipids and their similarities with room-temperature ionic liquids:
Biomembranes are complex and diverse biological supramolecular structures, which divide the inner part of cells where proteins, bio-complexes and biomachineries, and nuclei reside, from the external environment. Moreover, they role in the biochemistry of cells is far more relevant than being just a physical barrier. Biomembranes, for example, regulate the diffusion of any chemical species into cells, either through specific (protein) channels or simply by absorption into the phospholipid region, and play a major role in cell replication processes as well. Biomembranes are also the target of several antibiotics, since even small modification of their structure, kinetics, and elastic properties can drastically affect the cells' stability up to their death. It is then not surprising that biomembranes are one of the major and quite broad subjects of studies.
Chemical sketches of some selected RTILs cations: (a,b,c) the most common imidazolium and pyridinium RTILs cations; (d,e) the double-tail lipid-mimic imidazolium-based RTILs of (Wang et al. 2015a); (f,g) the ethylammonium and guanidinium RTILs cations that help and contrast protein amyloidosis, respectively (Byrne 2007, 2008, 2009); (h) a phosphonium-based RTIL cation; and (i,l) the choline and phosphocholine cations also used in RTILs made of amino acids (Benedetto et al. 2014a).
Phospholipid bilayers are a well-accepted first order model of biomembranes: they can be seen as the skeleton of any biomembranes into which proteins, extra lipids, saccharides, and, in general, bio-complexes can be absorbed to create more detailed and specific models of real biomembranes. As a result, assessing the effect of RTILs on phospholipid bilayers is the required first step along the path of the molecular-level comprehension of the biological effects of RTILs on biomembranes, and, in turn, on cells. This _modus operandi_ is also corroborated by the fact that the cytotoxicity of RTILs measured by a variety of bioassay shows a clear positive correlation with their lipo-philicity (Ranke et al. 2006, 2007a; Stolte et al. 2007). Moreover, phospholipid bilayers can be easily prepared in a controlled and, thus, reproducible way in labs. Moreover, their cost is reasonable and affordable even for small research groups; this makes experimental investigations possible and accurate.
The basic units of any phospholipid bilayer are phospholipid molecules. They can be either zwitterionic or ionic, and are made by two hydrocarbon tails having a high hydrophobic character, and a hydrophilic head. As a result, when in a water environment, they arrange themselves in superstructures in such a way to minimize the contact between their hydrophobic tails and the water molecules, and maximize, on the other hand, the contact between their hydrophilic heads and the water molecules. Due to different factors, e.g. concentration of phospholipids in water, physico-chemical conditions, and geometry constrains, they can form uni- and multi-lamellar vesicles and micelles, and supported single-, bi- and multi-layers. Almost all of these geometries have been employed in the investigation of the effect of RTILs on biomembranes. Among all those systems, (i) supported bilayers, and (ii) vesicles are those that reproduce the double layer of biomembranes and, in turn, can be considered the best models to mimic them.
Two of the most common phospholipids: (a) POPC and (b) DMPC. They differentiate for the length of their (hydrophobic) hydrocarbon tails, whereas they share exactly the same (hydrophilic) head. When dispersed in aqueous environments, the hydrophobic-hydrophilic competition generates supramolecular structures like uni-lamellar (c) liposomes, (d) micelles, and (e) bilayer sheets (Wikipedia). Also multi-lamellar structures can be formed.
Comparing the phospholipids of with the RTIL cations of highlights their common similarities. Both types of molecules have an ionic / polar character, and both display hydrophilic and hydrophobic regions. The closest similarities is between phospholipids and the new RTILs synthetized by H.J. Galla and co-workers sketched in (Wang et al. 2015a), which have two hydrocarbon chains as phospholipids. A second intriguing overlap is with the ionic liquids of amino-acids (AAIL) in which anions are deprotonated amino acids, and cations either a protonated choline or phosphocholine group that are also present in phospholipid head groups (Benedetto et al. 2014a).
In these two cases, but also in general, the similarities between phospholipids and RTILs are certainly responsible for their mutual affinity; moreover, they also suggest that a combination of (i) electrostatic, (ii) dispersion interactions, (iii) hydrophobic and hydrophilic effects, and (iv) hydrogen bonds structure and dynamics at the interface have to be taken into account to describe the mechanisms of interaction that perhaps should result to be a fine balance between all those forces in competition one to the other. It should then also be clear that even small changes in the chemistry of the molecules could affect the total balance of the forces, and, in turn, dramatically change the system properties. This circumstance highlights how a full comprehension of the microscopic mechanisms of RTILs -biommbranes interaction, together with the tunable character of the RTILs chemistry, is a key step for any major progress in this field. This observation is somehow in contrast with one of the main motivations at the basis of these investigations aiming at finding general rules to assess the effect of RTILs on biomembranes. However, we believe that this "unfortunate" circumstance can be balanced by the almost immense playground we have in front of us to develop new breakthrough applications in bio-nano-technology.
## A _snapshot_ of RTILs - biomembranes interaction: a joint neutron scattering and computational study.
Different techniques, both experimental and computational, have been used in the last decade to study the interaction between RTILs and (model) biomembranes, with the aim to determine the microscopic mechanisms behind the observed biological effects. Initial experimental studies carried out by K.O. Evans (Evans 2008) pointed to the marginal stability of phospholipid bilayers in contact with water solutions of imidazolium and pyrrolidinium RTILs. Both floating lipid vesicles and supported phospholipid bilayers geometries have been studied by means of photoluminescence, atomic force microscopy (AFM) and quartz microbalance, respectively. All measurements revealed damage to the bilayer from moderate to substantial, increasing with increasing length of the cation hydrocarbon tail. To the best of our knowledge, these are the first chemical-physics investigations of the interaction between RTILs and model biomembranes. However, the first experimental study able to characterize at molecular level the penetration of RTILs into model biomembranes has been done by means neutron reflectometry (Benedetto et al. 2014b). More specifically, by giving access to the density distribution of the chemical species, neutron reflectometry has been used to investigate the changes in the microscopic structure and stability of two model phospholipid bilayers made by POPC and DMPC, respectively, in contact with water solutions of two RTILs([bmim][Cl] and [Cho][Cl]). As a result it was observed that: (i) phospholipid bilayers maintain their characteristic 2D structure at the RTIL concentrations of the experiments (up to 0.5M); (ii) RTIL cations penetrate into the phospholipid region staying in the first leaflet at the junction between the phosphonium polar head and neutral hydrocarbon tail of phospholipid molecules (red curve in Fig. 3); (iii) the phospholipid bilayer thickness shrinks by about 1 A, and the area per lipid increases; (iv) the amount of cations absorbed in the lipid region is more for DMPC than POPC; (v) the position of cations into the lipid region is independent of either the RTILs or the phospholipids choice; however (vi) for DMPC the [bmim]\({}^{*}\) cations have diffused in the inner layer as well. (vii) The penetration of the RTIL cations is not fully reversible, since after rinsing with pure water a non-negligible amount of cations remains in the lipid region (about 8% and 2.5% for DMPC and POPC, respectively).
Density distribution profiles as a function of height \(z\) from the surface of the substrate obtained by fitting the neutron reflectivity data taken from (Benedetto et al. 2014b). Neutron reflectometry has allowed to model each single supported phospholipid bilayers with 4 different density distributions accounting for (i) the inner lipid heads layer (cyan); (ii) the inner lipid tail layer (blue); (iii) the outer lipid tail layer (blue); and (iv) the outer lipid tail layer (cyan). (v) In red the density distribution of the cations, whereas the anion (Cl) is almost invisible to neutrons. Three cases are here reported where two different phospholipid bilayers intercar with water solutions of two different RTILs at 0.5M: (a) POPC and [Chol][Cl], (b) POPC and [bmim][Cl], and (c) DMPC and [bmim][Cl]; the RTILs cations absorption accounts for 8%, 6.5%, and 11% of the lipid bilayer volume, respectively. In (c) the diffusion of the cations into the inner leaflet is apparent. In all the cases phospholipid bilayers are in the liquid phase.
Starting from the above-mentioned experimental results, A. Benedetto and co-workers carried out classical atomistic molecular dynamics (MD) simulations on the same systems (Benedetto et al. 2015). The MD simulations have reproduced the density distributions obtained by neutron reflectivity, something that certifies the ability and good quality of the model. This is something that is not obvious and has to be checked, since the empirical force fields used in these MD studies, have never been tested properly for these ternary systems, i.e. phospholipids, water, and RTILs. Whereas the empirical potentials of lipids and water have been tested to work together, and the same is also true for those of RTILs and water, they never have been tested all three together. In a simplified way, we can conclude that the agreement between the MD simulations and the neutron reflectivity profiles (compare with Fig. 4) is a proof of validity of the empirical potential used.
The MD simulations were able to determine or at least suggest the microscopic mechanism of the RTIL-biomembrane interaction, consisting of the following steps: (i) cations enter the phospholipid bilayer after 1-2ns of simulations; (ii) the penetration of cations into the bilayer is driven by the Coulombic attraction between the positive charge of the cation itself with the negative charged groups in the lipid head region (e.g. the negative oxygens in the carbonyl group at the matching point of the hydrocarbon tails), and (iii) it is stabilized by substantial dispersion forces between the cation and phospholipid tails; (iv) the absorption of the cations drives the penetration of a small amount of water into the polar portion of the phospholipid bilayer, and (v) stabilizes the hydrogen bonds at the lipid-water interface. presents some of the MD results.
In the MD study presented above, the systems were made of 1.360 POPC molecules, and 26.000 water molecules, and the computational "production" time was about 100-150 ns. The simulation box had a length of about 200 x 100 x 120 Angstrom along x, y, and z, respectively. For classical full-atom MD simulations, these are "big numbers", and to the best of our knowledge this is the biggest computer simulation study of phospholipids and RTILs. Both longer time scales and bigger systems are important to catch several features that otherwise cannot be detected. For example, only undulations of the bilayer surface whose wavelength is lower that twice the length of the simulation box can be detected, and only dynamical relaxations whose characteristic time is few times shorter that the total computational time can be properly identified. Needless to say, the simulation box sizes and computational "production" time affect also the error bars of all the observables extracted from the MD trajectories. As a result, our MD simulations allow to identify several trends due to the absorption of the RTILs cations in the lipid phase, from the shrink of the bilayer thickness, to the variation of diffusion coefficients and of elastic properties.
Density distribution profiles as a function of height z obtained from the full-atom classical MD trajectories of Ref. (Benedetto et al. 2015) for (a) neat POPC bilayers, and (b) doped with two RTILs. The computed profiles agreed with the one measured by neutron reflectivity reported in RTILs cations are absorbed in the lipid region whereas the anions remain in the water solution in contact with the bilayers.
(v) decreasing in [bmim][BF\({}_{6}\)] the diffusion coefficient of the lipids. Neutron scattering can be used to probe the most of these observables: elastic and quasi-elastic neutron scattering can access the pico-to-nanosecond dynamics (Benedetto and Kearley 2016; Magazu et al. 2010; Magazu et al. 2008; Bee 1988; Volino 1978); neutron spin-echo can be used to investigate slower relaxation processes (Mezei 1972), and also to probe structural fluctuations (Nagao 2009; Woodka 2012); small angle neutron scattering and diffraction can finally used for further structural characterization.
The joint experimental - simulation study presented above has its strength in combining two powerful approaches into one study, allowing a very detailed description of the microscopic mechanisms of RTIL-bilayer interaction. In the following we will present a series of results recently appeared in literature that either confirm or extend the scenario outlined above. Apart from few cases all the results are in agreement each other, and what is emerging is the extreme importance of the chemistry of lipids and RTILs, suggesting that case-by-case studies are needed. However, maintaining almost fixed the chemistry, some general trends emerge, like the correlation between RTILs chain lengths, concentration, and cytotoxicity.
## What matters in the microscopic world of RTILs-biomembranes interaction? From chain length and RTIL concentration to chemistry: a huge playground of challenges and opportunities.
The shrinkage of the phospholipid bilayer thickness upon the interaction with RTILs water solutions has been also confirmed by X-ray experiments (Bhattacharya et al. 2017, Kontro et al. 2016). In one case (Bhattacharya et al. 2017), single supported bilayers of DPPC (a zwitterionic lipid) interacting with water solutions of [bmim][BF\({}_{4}\)] have been measured by means of X-ray reflectivity at several RTIL concentrations.
Schematic view of one of the sample configurations used in the MD simulations of (Benedetto et al. 2015). POPC domains in grey, water layers in red, [C*mim]\({}^{*}\) in blue, and [PF\({}_{6}\)] in green. Inset (a): Representative configuration of POPC and [C*mim]\({}^{*}\). Inset (b): water density profiles: the difference (area in red) points to a water excess in the POPC doped with [C*mim]\({}^{*}\).
2017). The effect of concentration has also been studied by monitoring the survival percentage of _E. coli_ versus the concentration of [bmim][BF\({}_{4}\)], which shows an inverse correlation relationship. In the same paper S.K. Ghosh and co-workers, report also that the in-plane elasticity of supported monolayers decreases upon addition of the RTIL, suggesting that the rigid structure of the well-packed lipid monolayer relaxes in the presence of the RTIL. The in-plane elasticity has been determined by the pressure-area isotherm profiles on monolayers, with the water solution of RTIL added to the monolayer-forming lipid solution. Another way to probe how the mechano-elastic properties are modified by the addition of RTILs, is offered by AFM, where both the Young's moduli and the rupture forces can be determined for single supported phospholipid bilayers and vesicles (Monocles et al. 2010; Garcia-Manyes and Sanz 2010; Roa et al. 2011; Ferenc et al. 2012; Stetter et al. 2016; Ding et al. 2017). To the best of our knowledge, there are not AFM studies of this kind so far. The elastic properties of phospholipid bilayers with and without RTILs have been studied by MD simulations as well; as discussed below. AFM can also image the surface of the bilayers and tell us something about the degree of homogeneity of the surface and/or the presence of pores and defects.
The second X-ray study cited above (Kontro et al. 2016) used small-angle X-ray scattering (SAXS) to study multi-lamellar liposomes of eggPC and eggPG (80:20 mol%) - and also with cholesterol (60:20:20 mol%) - in interaction with water solutions of phosphonium-based RTILs. The results have been compared with those of the more common imidazolium-based RTILs. The interest of this study is increased by the fact that some phosphonium-based RTILs have shown antibacterial activity (0"Toole et al. 2012). The lamellar spacing of liposomes decreases with increasing RTIL concentration. It is then clear that in this case RTILs pass through the phospholipids multilayer superstructures. In the same study S.K. Wiedmer and co-workers probe the effect of RTILs also by means of dynamic light scattering (DLS) and zeta potential measurements on large uni-lamellar vesicles. They conclude that the ability of RTILs to affect liposomes is related to the length of the hydrocarbon chains of their cations.
Inhibition percentage of _E. coli_ versus the concentration of the RTIL [BMIM][BF\({}_{4}\)] taken from (Bhattacharya et al. 2017).
This result was confirmed by a more recent work again by S.K. Wiedmer and co-workers, where by means of nanoplasmonic sensing (NPS) measurement technique, they characterized the interaction between supported phospholipid uni-lamellar vesicles with amidinium- and phosphonium-based RTILs (Witos et al. 2017). NPS is a label-free optical technique that allows the study of surfaces and interfaces of metals, relying on surface plasmons, with a penetration depth of around 10 nm (quartz crystal microbalance has a penetration depth of about 250 nm).
S.K. Wiedmer and co-workers have also studied how the addition of cholesterol in the lipid phase changes the RTIL-biomembrane interaction (Kontro et al. 2016). This is a quite important, since it is well known that cholesterol is present in several biomembranes and has the ability to stabilize them. As a result, vesicles without cholesterol ruptured at lower concentrations than vesicles with cholesterol. At concentration the effects on the cholesterol-containing liposomes were more severe: the ruptured liposomes reassembled into organized lamellae, so under certain conditions, phosphonium-based ionic liquids have the ability to create new self-assembled structures from phospholipids.
Phase diagram of [Ca mim][Cl] ionic liquid induced morphological changes to a supported α-PC bilayer taken from ref. (Yoo et al. 2016a). The EC50 toxicity line (in magenta) for IPC-cell shows a negative correlation between the toxic concentration and the RTIL cation chain length (Ranke et al. 2007a). The blue and lines are the predicted EC50 lines for wild-type (with cell wall)
SAXS pattern of multilamellar POPC vesicles in interaction with RTILs from (Kontro et al. 2016). Reference MLV (light gray), MLV treated with [P4441][OAc] (black), and MLV treated with [emim][OAc] (dark gray). Insert: magnification of the first diffraction peak.
The green line corresponds to the RTIL critical micelle concentration (CMC) of (Blesic et al. 2007). The symbols correspond to the specific morphologies as in the right image. Black square: neat bilayer; red circle: multilayer; blue triangle: multilayer and fiber/tube; pink diamond: multilayer, fiber/tube and vesicle; green hexagon: vesicle; navy star: disrupted bilayer. The solid black and red lines correspond to the onset of supported lipid bilayer disruption and the total disruption of the supported lipid bilayer, respectively.
Further investigations on the connection between cytotoxicity, cation chain length, and RTIL concentration has been done by E.J. Maginn and co-workers by combining several experimental techniques with computer simulations (Yoo et al. 2014, 2016a, b). One of their main results is about the concentration-dependent effect of RTILs on biomembranes: they found that RTIL cations nucleate morphological defects on biomembranes at concentrations near the half maximal effective concentration (EC50) of several microorganisms, and that RTILs destroy biomembranes at the RTIL critical micelle concentration. The results of E.J. Maginn and co-workers suggest that the molecular mechanism of RTIL cytotoxicity may be linked to the RTIL-induced morphological reorganization of cell membranes initially caused by the insertion of RTILs into the membrane. In their studies, they have also shown that cytotoxicity increases with increasing alkyl chain length of the cation and relate this observation to the higher ability of alkyl chain to penetrate, and ultimately disrupt, cell membranes.
E.J. Maginn and co-workers performed also full-atom and coarse-grained MD simulations. Coarse-grained simulations, in particular, by sacrificing some atomic-level details, are able to explore longer time-scales and larger length-scales than atomistic MD simulations; this is quite important to study both the absorption of RTILs, and the changes in the bilayers structures. Full-atom simulations usually can access a time scale of 100ns and a spatial scale of tens of nanometers, whereas with coarse-grained simulations it is possible to access time scales of microseconds and spatial scales of micrometers. As a result, their coarse-grained MD simulations show that the short-tail RTIL [C4mm] cation spontaneously inserted into the lipid bilayer with the same orientation as that in the atomistic simulations, whereas the long-tail RTIL [C10mm] cation self-assembled into micelle eventually absorbed onto the upper bilayer leaflet forming a RTIL monolayer. In the case of the short-tail cation, the number of inserted cations into the upper bilayer leaflet saturates at about 0.6 cations per lipid, after which the bilayer bends in response to the asymmetric distribution of inserted cations; moreover, while RTIL cations are absorbed, the bending modules of the bilayer drops from 22.6 \(\pm\) 1.7 \(\times\) 10-20 J to 9.3 \(\pm\) 0.9 \(\times\) 10-20 J. Since this asymmetric situation is not occurring when cations are absorbed in both the two leaflets, E.J. Maginn and co-workers comment on the inability of cations to diffuse from one leaflet to the other, that is at origin of the bending fluctuations of the bilayer.
The inability of the RTILs to diffuse into the whole bilayer after inserted into the closest leaflet seems in contrast, however, with the SAXS data of (Kontro et al. 2016), and in our opinion is something that requires more investigations. Having a closer look to our neutron reflectivity data of Fig. 3, for example, we can conclude that it seems to depend on the phospholipids and RTILs used. It seems, moreover, that it depends on the difference between the lipid and the cation chain lengths. Since [bmim][Cl] diffuses in the inner leaflet of DMPC, but does not with POPC, we could conclude that the shorter the lipid chain, the higher the ability of RTILs to diffuse from one leaflet to the other. Since E.J. Maginn and co-workers did not observe _inter-leaflet_ diffusion of [bmim][Cl] in POPC, and S.K. Wiedmer and co-workers did observe _inter-leaflet_ diffusion of different RTILs in POPC, we can conclude that the disagreement is due to the important role played by the RTIL chemistry, in line with our neutron reflectivity results (Benedetto et al. 2014b).
In our full-atom MD trajectories (Benedetto et al. 2015) done on POPC in interaction with [bmim][Cl], moreover, we did not observe any diffusion of cations from one layer to the other (result not yet published). However, longer simulation times may be needed to properly assess this _inter-leaflet_ diffusion of RTILs, and its implication in RTILs cytitoxicity. AFM could also be a good technique to double-check indirectly this _in-bilayer_ diffusion of RTILs by measuring the distance between the bilayer surface and the support: since we know that RTILs shrink the bilayer thickness, any increment of such distance could be related to the diffusion of RTILs into the inner bilayer leaflet and the water interlayer on the top of the support. The ability of RTILs to diffuse in multi-layer geometries can be useful not only in setting-up experiments (like diffraction experiments in which multi-layer lipid structures can be used) but also in applications. In our opinion this subject deserves further investigations.
Eplifuorescence images of mixed monolayers of DPPC with double-tail imidazolium-based RTILs of different chain-length at different molar fractions at the air-water interface at room temperature taken from (Wang et al. 2016).
Coarse-grained MD simulations of Ref. (Yoo et al. 2016b) suggest that is the inability of some RTILs to diffuse from the outer leaflet to the inner leaflet of the phospholipid bilayer at the origin of the bilayer disruption.
On the contrary, the biological activity of these RTILs goes the other way around, since the shortest one is the more toxic and the longest one stabilize the bilayer phase.
An important contribution to the saga has been made by H.J. Galla and co-workers (Wang et al. 2015a, b, 2016; Drucker et al. 2017, Ruhling et al. 2017). They have designed and synthesized a series of backbone-alkylated imidazolium-based RTILs composed of a hydrophilic N,N'-dimethylated imidazolium headgroup and two hydrophobic alkyl chains located at the 4- and 5-positions of the imidazole ring (Fig. 1d-e). The structure of these two-tail imidazolium-based RTILs (C\({}_{\text{n}}\)1Me\(\cdot\)HI) is similar to that of phospholipids (Fig. 2a-b), explaining the similarity of their physicochemical properties. The interest in this new class of RTILs is increased by their significant antitumor activity and cellular toxicity: in comparison with the more common one-tail imidazolium RTILs, they show approximately 3 order of magnitude higher antitumor activity. Having a global picture in mind, the most intriguing aspect is that their toxicity is negatively correlated with their chain lengths: the C?1Me\(\cdot\)HI has the highest toxicity and antitumor activity, whereas the C\({}_{15}\)1Me\(\cdot\)HI has the lowest toxicity. This circumstance is in contrast with the "general picture" presented above, and highlights how the chemistry of the molecules plays an extremely important rule. Moreover, the biological activity of these RTILs is inversely correlated with their lipo-philicity, something that also plays against the general picture following which lipo-philicity and toxicity are directly correlated. From it emerges that C\({}_{15}\)1Me\(\cdot\)HI has the higher membrane activity, and by increasing its concentration the RTIL-lipid system shows a reorganization at 0.3 molar fraction (more details in the referred paper). shows that C\({}_{11}\)1Me\(\cdot\)HI has the ability to inhibit the supramolecular reorganization of the phospholipids, whereas shows how the membrane activity of C\({}_{7}\)1Me\(\cdot\)HI is negligible.
Model for membrane interaction and structure formation of double-tail imidazolium-based RTILs taken from (Drucker et al. 2017). Liposomes (blue) are tethered via biotin linkers (green) and streptavidin (purple) on a self-assembled monolayer (brown), which itself is chemisorbed on a gold-coated sensor surface (orange). (a) C\({}_{15}\)-1Me\(\cdot\)HI is able to form vesicles in
(b) C11-IMe-HI is able to form both vesicles and micelles while in water, which can then intercalate and lyse bilayer membranes. Bilayer disintegration is accompanied by the formation of micelles and mixed micelles. (c) C7-IMe-HI dissolves to micelles and single molecules and can pass though the membrane without disintegration.
H.J. Galla and co-workers combined several experimental techniques as film balance, quartz crystal microbalance, confocal laser scanning microscopy, calorimetry, epifluorescence microscopic measurements, with MD simulations. They have also measured the interaction of RTILs with biomembranes enriched with cholesterol.
Several other studies have been done on this new and promising subject, which more or less agreed with the results presented above: (i) positive correlation between RTILs chain lengths, RTILs concentration, and cytotoxicity (Witos et al. 2017; Losada-Perez et al. 2016; Dusa et al. 2015; Galletti et al. 2015; Galluzzi et
Bilayer domain fluidization of small bulged domains to flat large domains with enhanced dye specificity in the presence of 10% the double-tail imidazolium-based RTIL C15Me-HI from (Drucker et al. 2017). Giant uni-lamellar vesicles of (a) DOPC/SSM/Chol (33:33:33), and (b) DOPC/SSM/Chol/ C151Me-HI (33:23:33:10) at 38\({}^{\circ}\)C, scale 20 \(\mu\)m.
Effect of RTILs on gramicidin A ion channel from (Ryu et al. 2015). (a) Neat system; (b) system doped with RTIL. The RTIL cation C10min (i) stabilizes the membrane-channel interaction by reducing the bilayer thickness and, in turn, its curvature closer to the channel location; and (ii) reduce the channel activity by electronic repulsion as sketched in (a). The function of the channel seems also affected by the amount of the inorganic salt NaCl in the solution: the higher the amount the higher the ion permeability.
2013; Kulacki and Lamberti 2007; Jeong et al. 2012; Mikkola et al. 2015; Jing et al. 2016); (ii) shrinkage of the bilayer thickness (Lim et al. 2014; Lim et al. 2015), and (iii) variation of its elasticity (Dusa et al. 2015) upon the absorption of RTILs; and (iv) importance of the chemistry of the molecules (Weaver et al. 2013; Gal et al. 2012; Lee et al. 2015). The effect of RTILs on the thermotropic behavior of biomembranes have also probed in several works (Weaver et al. 2013; Wang et al. 2016; Jeong et al. 2012): at low concentrations of RTILs the variation of the main phase transition of phospholipids bilayers is negligible, and at high concentration the variation is of the order of 5 to 10 degrees, sometimes reaching 20 degrees, but this just before the collapse of the bilayers. Several other MD simulations studies have also done on this subject (e.g. Bingham and Ballone 2012; Cromie et al. 2009; Lim et al. 2015), which give access to the detailed mechanisms of interactions. In Lim et al. 2015, for example, the mutual interaction between RTILs cations once absorbed into the membrane have been studied, and the measured increment of permeability due to the absorption of RTILs has been related to their antibacterial activity. Investigations focuses on specific cases and more complex systems, moreover, have enriched even more the broad and diverse panorama presented above (Patel et al. 2016; Modi et al. 2011; Jeong et al. 2012; Ryu et al. 2015; Lee et al. 2015). An interesting example is the study of T-J Jeon, H. Lee and co-workers on the effect of RTILs on the ion channel function of Gramicidin A embedded in a phospholipid bilayer (Jeong et al. 2012; Ryu et al. 2015; Lee et al. 2015). Their results show for the first time how the changes of the physical properties of the biomembrane (e.g. thickness) induced by the absorption of RTILs cations can influence the activities of membrane proteins. These effects are more significant with RTILs with longer alkyl chains, and at higher RTILs concentration. Interestingly, they figured out how the concentration of inorganic salts (i.e. NaCl) can play a major role also in this case. The picture that emerges is the following: RTILs disorder phospholipids and shrink the bilayer, which yields less membrane curvature around the gramicidin A and thus stabilizes it, leading to the increased ion permeability. However, this effect occurs at 1 M of NaCl, where RTILs only slightly increase the phospholipids dynamics because of the strong electrostatic interactions between NaCl and lipids. On the other hand, at 0.15 M of NaCl, RTILs also significantly increase the lateral mobility of both phospholipids and gramicidin A, which leads to the decreased ion permeability.
## Summary and outlook for the future
The analysis of the current literature on the interaction between RTILs and biomembranes shows that the first target of these studies has been to determine the microscopic mechanism behind the cytotoxicity of RTILs with the aim to design _greener_ RTILs for industrial applications. Whereas a positive correlation exists between RTILs chain lengths, RTILs concentration and cytotoxicity, the are several exceptions that highlight how the RTILs effect on biomembranes is a complex balance of different interaction where the chemistry of each molecule is playing a non-negligible role. On the one hand, this observation implies that the effect of RTILs on biomembranes has to be assessed on a case-by-case basis. On the other hand, the same observation opens a huge playground of new opportunities for applications in bio-nano-technology. In our opinion these opportunities are more important than the difficulties preventing an _a priori_assessment of the potential danger. Just few nano-bio-technology studies have been done so far, and we are just at the beginning of this promising field of research.
|
10.48550/arXiv.1706.08509
|
Room-Temperature Ionic Liquids Meet Bio-Membranes: the State-of-the- Art
|
Antonio Benedetto
| 5,261
|
10.48550_arXiv.2105.07379
|
###### Abstract
## Abstract
We discuss the extension of the empirical equation: \(\left\langle s_{N}^{2}\right\rangle_{0}\propto g\,l^{2}\), where the subscript \(0\) denotes the ideal value with no excluded volume and \(g\) the generation number from the root to the youngest (outermost) generation. By analogy with the linear chain problem, we introduce the assumption that the scaling relation, \(\left\langle s_{N}^{2}\right\rangle_{0}\propto g^{2\lambda}\,l^{2}\), exists for arbitrary polymeric architectures, where \(\lambda\) is an exponent for the backbone structure. Then, making use of the relationship between \(g\) and \(N\) (monomer number), we can deduce the exponent, \(\nu\), for polymers with various architectures. The theory of the excluded volume effects impose the severe restriction on the quantities: \(\nu_{0}\), \(\nu\), and \(\lambda\); for instance, the inequality, \(\nu_{0}\geq\frac{1}{d+1}\), must be satisfied for isolated polymers in good solvents. An intriguing question is whether or not there exists an actual molecule that violates this inequality. We take up two examples having \(\nu_{0}=1/4\) for \(d=2\) and \(\nu_{0}=1/6\) for \(d=3\), and discuss this question.
## Key Words
Comb Polymers/ Nested Architectures/ Internal Density/ Exponent \(\nu\)/
+
Footnote †: footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † footnotetext: † †
## 2 Theoretical
To consider the generalization of Eq., let us take up the regular comb polymer having side chains of the length, \(n\). This comb polymer has the total monomer number, \(N=g+(g-1)n\), as mentioned above. Let each side chain be a part of the corresponding monomer on the main backbone, and index accordingly so that, for instance, \(3_{2}\) denotes the second monomers on the side chain emanating from the branching unit on the third generation, and so forth. According to Isihara, the end-to-end vector from the center of gravity to the \(p\)th monomer is written in the form:
\[\vec{r}_{Gp}=\vec{r}_{1p}-\frac{1}{N}\sum_{p=1}^{N}\vec{r}_{1p} \tag{2}\]
Following the formula, it is easy to show that the component vectors have the expressions:
\[\vec{r}_{Gh}=\frac{1}{N}\left\{\sum_{k=1}^{h-1}[N-(N-k-(k-1)n)]\, \vec{l}_{(k+1)}-\sum_{k=1}^{g-1}[N-k-(k-1)n]\,\,\vec{l}_{(k+1)}\right.\\ \left.-\sum_{k=1}^{g-1}\sum_{i=1}^{n}(n-i+1)\,\vec{l}_{(k+1)_{i}} +\sum_{k=1}^{h-1}[N-k-(k-1)n]\,\vec{l}_{(k+1)}\right\} \tag{3}\]
\[\vec{r}_{Gh_{j}}=\frac{1}{N}\left\{\sum_{k=1}^{h-1}[N-(N-k-(k-1)n) ]\,\vec{l}_{(k+1)}+\sum_{i=1}^{j}[N-n+i-1]\,\vec{l}_{h_{i}}-\sum_{k=1}^{g-1} \,[N-k-(k-1)n]\,\,\vec{l}_{(k+1)}\right.\\ \left.-\sum_{k=1}^{g-1}\sum_{i=1}^{n}(n-i+1)\,\vec{l}_{(k+1)_{i}} +\sum_{k=1}^{h-1}[N-k-(k-1)n]\,\vec{l}_{(k+1)}+\sum_{i=1}^{j}[n-i+1]\,\vec{l}_ {h_{i}}\right\} \tag{4}\]
It is seen from Eqs.
\[\vec{r}_{Gp}=\frac{1}{N}\sum_{q=2}^{N}c_{q}(p)\,\vec{l}_{q} \tag{5}\]
In Eq., starting from \(q=2\) is necessary since this polymer has \(N-1\) bonds. Eq. is the typical Pearson random walk with unequal step lengths, \(\{c_{q}(p)/N\}\). It is noteworthy that Eq. is of the same form as the end-to-end distance of a linear chain which corresponds to the special case of \(c_{q}(p)/N=1\)
A comb polymer having side chains of the length, \(n\).
Let all bonds have an equal length, \(|\vec{l}_{q}|=l\), and \(\vec{e}\) be the unit vector. The mean square of the radius of gyration is written in the form:
\[\left\langle s_{N}^{2}\right\rangle=\frac{1}{N}\sum_{p=1}^{N}\left\langle\vec{r }_{Gp}\cdot\vec{r}_{Gp}\right\rangle=\frac{1}{N}\sum_{p=1}^{N}\left[\sum_{q=2} ^{N}\frac{c_{q}(p)^{2}}{N^{2}}+\sum_{i\neq j}^{N}\frac{c_{i}(p)c_{j}(p)}{N^{2} }\left\langle\vec{e}_{i}\cdot\vec{e}_{j}\right\rangle\right]l^{2} \tag{6}\]
If there is no correlation between bonds, we have \(\left\langle\vec{e}_{i}\cdot\vec{e}_{j}\right\rangle=0\) for \(i\neq j\). Then, Eq. reduces to the freely-jointed-chain model to yield the quantity for the ideal polymer without the excluded volume. For the comb polymer under discussion, this has the form: in terms of \(g\),
\[\left\langle s_{N}^{2}\right\rangle_{0} = \frac{1}{N}\sum_{p=1}^{N}\sum_{q=2}^{N}\left[\frac{c_{q}(p)}{N} \right]^{2}l^{2} \tag{7}\] \[= \frac{1}{6}\cdot\frac{(n+1)^{2}g^{3}+3n^{2}(n+1)g^{2}-(n+1)(8n^{ 2}+2n+1)g+n(n+1)(5n+1)}{(n+1)^{2}g^{2}-2n(n+1)g+n^{2}}\,l^{2}\]
which, as \(g\rightarrow\infty\), approaches the asymptotic form:
\[\left\langle s_{N}^{2}\right\rangle_{0}\doteq\frac{1}{6}\,g\,l^{2} \tag{8}\]
If we take into account the fact that Eq. has the same form as the linear chain model and that \(g\) acts as the size of the molecule, the result of Eq. seems to be a natural consequence. Remarkable is the fact that while, by Eqs. and, \(\left\langle s_{N}^{2}\right\rangle\) represents the statistical average of \(N-1\) bonds with unequal lengths, it behaves, phenomenologically, as though it is a linear chain linked by \(g\) bonds with an equal length, \(l\).
For the excluded volume polymer, we have \(\left\langle\vec{e}_{i}\cdot\vec{e}_{j}\right\rangle\neq 0\). In that case, by analogy with the linear chain model, it seems reasonable to expect that Eq. has the asymptotic form of \(g\rightarrow\infty\):
\[\left\langle s_{N}^{2}\right\rangle\propto g^{2\lambda} \tag{9}\]
To deduce the real exponent, \(\nu\), let us consider the nested structures put forth in the preceding paper. In Fig. 2, the starting polymer, \(N_{1}\), may be an arbitrary polymer. On each nesting, a linear polymer with \(g\) monomers is newly introduced as a base polymer, on which the preceding structure is linked with each monomer on the base polymer. Such nesting is repeated successively to create deeper structures. Let \(z\) represent the depth of the nest. Then, we can write generally \(N_{z}=g+(g-1)N_{z-1}\). The solution to this recurrence relation is:
\[N_{z}=(g-1)^{z-1}N_{1}+g\frac{(g-1)^{z-1}-1}{g-2}\ \ \ \ (z\geq 1) \tag{10}\]
In this paper, we consider the simplest case in which \(N_{1}\) is a linear polymer with \(g\) monomers, so that \(N_{1}=g\). Substituting into Eq., we have \(N_{z}\doteq g^{z}\) for \(g\rightarrow\infty\). Substituting further this asymptotic relation into Eq., we have
\[\left\langle s_{N}^{2}\right\rangle\propto N^{2\frac{\lambda}{z}}\ \ \ \ \ \ \ (z=1,2,3,\dots) \tag{11}\]
which yields the required relation between the exponents:
\[\nu=\frac{\lambda}{z}\ \ \ \ \ \ \ \ (z=1,2,3,\dots) \tag{12}\]
According to Eqs.
\[\nu_{0}=\frac{1}{2z}\ \ \ \ \ \ \ \ (z=1,2,3,\dots) \tag{13}\]From Eqs.
\[\nu=2\lambda\nu_{0} \tag{14}\]
Since, according to Fig. 2, the full contour length of the backbone is, at most, \(\dot{=}\)\(zg\), clearly \(\lambda\) must satisfy \(\lambda\leq 1\), so that
\[\nu\leq 2\nu_{0} \tag{15}\]
Let us apply the inequality to the classic formula, the excluded volume equation for the dilution limit in good solvents:
\[\nu=\frac{2(1+\nu_{0})}{d+2} \tag{16}\]
Substituting Eq. into Eq., we have
\[\nu_{0}\geq\frac{1}{d+1} \tag{17}\]
While Eq. indicates that there exist a large number of architectures that have different ideal exponents of the interval, \(0\leq\nu_{0}\leq 1/2\), Eq. implies that, of those architectures, only those satisfying the inequality are compatible with Eq.. For instance, \(\nu_{0}\geq 1/4\) is required for \(d=3\), whereas, as is seen from Eq., the nesting architectures with the depth greater than \(z=3\) can not satisfy Eq. since those architectures have the exponents of the interval \(0\leq\nu_{0}=\frac{1}{2z}\leq 1/6\).
It is obvious, on the other hand, that the \(z=3\) polymer with \(\nu_{0}=\frac{1}{6}\) can be accommodated in the three-dimensional space, for instance, by arranging the backbone on the \(XY\) plane and the combs in parallel to the \(YZ\) plane so as to extend perpendicularly to each other. As this example shows, in the viewpoint of synthetic chemistry, the \(z=3\) polymer can be a real compound! Then, what happens to this polymer when placed in the dilution limit? One possible conjecture is that, by virtue of Eq. and \(\lambda=1\), \(\nu\) simply takes the exponent, \(1/3\), which marginally satisfies the accommodation condition, \(1/d\). Under this critical circumstance, according to the thermodynamic theory of the excluded volume effects, there still remains the strong Gibbs potential to enlarge the molecular dimensions due to the diffusive-flow of segments along the density gradient, which is, however, exactly canceled out by the elastic counterforce of the stretched backbone having a stronger spring constant, \(\mathpzc{k}\).
The first (\(z=1\)), the second (\(z=2\)), and the third (\(z=3\)) nested structures. In \(z=3\), \(g-1\) comb polymers branch off from the monomers on the new backbone (red solid-line). Such nesting may be repeated successively, which ultimately approaches the dendrimer structure with \(f=3\).
In the following section, we will examine the validity of this conjectured exponent, \(\nu_{z=3}=1/3\), through the direct calculation with the help of the lattice model.
Recall that the above argument is a story for the limiting case of \(N\to\infty\). A real molecule having a finite \(N\) has greater freedom. Such an example of the \(z\)=3 polymer with \(g=4\) (\(N=52\)) is displayed in There is a possibility that one can experimentally determine the exponent by extrapolating to \(N\to\infty\), making use of samples with finite \(N\)'s.
### Two-dimensional case
The same situation arises in the two-dimensional case of the \(z=2\) (extended comb) polymer. Since a real molecule is build up based on the tetrahedral structure, a genuine two-dimensional polymer is generally unavailable.
\begin{table}
\begin{tabular}{c c c c c} Polymers & \multicolumn{4}{c}{Exponents} \\ \cline{2-5} & \(z\) (depth of the nest) & \(\lambda\) & \(\nu_{0}\) & \(\nu\) \\ Nested structures & \(z\geq 1\) & \(\lambda\leq 1\) & \(\frac{1}{2z}\) & \(\frac{\lambda}{z}\) \\ \(z=1\) (linear polymer) & 1 & \(\frac{1}{2}\) & \(\frac{1}{2}\) & \\ & 1 & \(\frac{3}{5}\) & & \(\frac{3}{5}\) \\ \(z=2\) (extended comb)\({}^{\,\rm a}\) & 2 & \(\frac{1}{2}\) & \(\frac{1}{4}\) & \\ & 2 & 1 & & \(\frac{1}{2}\) \\ \(z=3\)\({}^{\,\rm b}\) & 3 & \(\frac{1}{2}\) & \(\frac{1}{6}\) & \\ & 3 & 1 & & \(\frac{1}{3}\)\({}^{\,\rm c}\) \\ \hline \end{tabular}
* a. equivalent to the comb polymer with \(n=g\) in * b. incompatible with Eq., but the polymer can be embedded in the real space.
* c. a conjectured value according to the equations-.
\end{table}
Table 1: Exponents, \(\nu\), for the nesting structures in the dilution limit in good solvents (\(d=3\)).
The \(z=2\) cluster has \(\nu_{0}=1/4\), the same value as the triangular polymer without the excluded volume, so it does not satisfy Eq. in \(d=2\). In contrast, according to Fig. 2, this cluster can clearly be arranged on the two-dimensional square lattice. Thus, by Eqs.-, we expect \(\lambda=1\) and \(\nu=1/2\), which is again equal to the critical packing density. However, we note that the exponent, \(\nu=1/2\), thus estimated is not consistent with other researchers' results that scatter from 0.5 to 0.687 and conflict with each other. There is a possibility that one can settle this puzzling problem through simulation experiments on lattices by using pure polymers with fixed architectures, such as the comb polymer with \(n=g\) and the triangular polymer.
In the following, we consider this problem through the direct calculation of lattice clusters, assuming that molecules are constructed from the stretched backbone and stretched side chains.
## 3 Mean Squares of Radii of Gyration of Self-avoiding Polymers
The mean square of the radius of gyration of an excluded volume polymer can generally be evaluated by means of the thermodynamic theory. Here, we evaluate the corresponding quantity in an algebraic manner, albeit for special polymers. For this purpose, we make use of the formulae for the regular comb polymer, which has the end-to-end vectors given by Eqs. and.
We are interested in the excluded volume phenomena of polymers that violate the inequality. The \(z\)=2 (extended comb) polymer in \(d=2\) and the \(z\)=3 polymer in \(d=3\) belong to this category, as discussed above. We are going to approach this problem according to the lattice model. Suppose these polymers take the configurations with the stretched backbone and stretched side-chains extending perpendicularly to each other; in that case, it seems evident that, as \(g\rightarrow\infty\), these, when arranged on the square and the cubic lattices, respectively, will cover all the lattice sites. Our aim is to calculate the radii of the gyration of such fully expanded polymers and estimate the exponent, \(\nu\). The main point of the calculation is that only the scalar products between the bond vectors parallel to each other survive, and all the perpendicular components vanish.
### The \(z=1\) (Rod) Polymer in the One-Dimensional Space
The one-dimensional freely-jointed polymer has \(\nu_{0}=1/2\), and hence, it marginally satisfies the inequality. The end-to-end distance vector, \(\vec{r}_{Gh}\), of the one-dimensional polymer corresponds to the case of \(n=0\) in Eq., so that only the scalar product, \(\vec{l}_{k}\cdot\vec{l}_{h}\), between bonds on the backbone is mathematically useful. Let this polymer be comprised of \(g\) monomers, and all the bonds have the equal length, \(|\vec{l}_{k}|=l\). According to the definition of the mean square of the radius of gyration and Eq., we have
\[\left\langle s_{g}^{2}\right\rangle_{z=1}=\frac{1}{N}\left\langle\sum_{h=1}^{g }\vec{r}_{Gh}\cdot\vec{r}_{Gh}\right\rangle=\frac{1}{12}\left(g^{2}-1\right)l ^{2} \tag{18}\]
For \(g\rightarrow\infty\), this gives \(\left\langle s_{g}^{2}\right\rangle_{z=1}\doteq\frac{1}{12}\,g^{2}\,l^{2}\). Since \(N=g\), we have \(\nu=1\), in accord with what is predicted by the equations-, and also agrees with the prediction by Eq..
### The \(z=2\) (Extended Comb) Polymer in the Two-Dimensional Space
In Eqs. and, we put \(n=g\).
\[\left\langle s_{g}^{2}\right\rangle_{z=2}=\frac{1}{N}\left\{\left\langle\sum_ {h=1}^{g}\vec{r}_{Gh}\cdot\vec{r}_{Gh}\right\rangle+\left\langle\sum_{h=2}^{g }\sum_{j=1}^{g}\vec{r}_{Gh_{j}}\cdot\vec{r}_{Gh_{j}}\right\rangle\right\}= \frac{1}{6}\frac{\left(g^{2}-1\right)\left(g^{2}+3\right)}{g^{2}}l^{2} \tag{19}\]
Since \(N=g^{2}\) for the \(z\)=2 polymer, we have \(\nu=1/2\), in agreement with the prediction of the equations-, and also with the critical packing density, \(1/d\).
### Mathematical Check
For \(g=1\) (\(N=1\)), Eq. gives the obvious result: \(\left\langle s_{g}^{2}\right\rangle_{z=2}=0\). For \(g=2\) (\(N=4\)), the molecule has the configuration (\(\clubsuit\)\(\clubsuit\)) like the capital letter, L. It is easy to calculate, by the elementary geometry, the mean square of the radius of gyration of this molecule to yield \(\left\langle s_{g}^{2}\right\rangle_{z=2}=\frac{7}{8}\,l^{2}\), in exact agreement with the prediction of Eq..
### The \(z=3\) Polymer in the Three-Dimensional Space
To solve this problem, it is necessary for us to make an extension of Eqs. and; the result is shown in Appendix A. Using the extended equations (A.2)-(A.4), we can calculate, in quite the same manner, the corresponding quantity for the \(z\)=3 polymer to obtain
\[\left\langle s_{g}^{2}\right\rangle_{z=3} =\frac{1}{N}\left\{\left\langle\sum_{h=1}^{g}\vec{r}_{Gh}\cdot \vec{r}_{Gh}\right\rangle+\left\langle\sum_{h=2}^{g}\sum_{j=1}^{g}\vec{r}_{Gh _{j}}\cdot\vec{r}_{Gh_{j}}\right\rangle+\left\langle\sum_{h=2}^{g}\sum_{j=2}^ {g}\sum_{p=1}^{g}\vec{r}_{Gh_{jp}}\cdot\vec{r}_{Gh_{jp}}\right\rangle\right\}\] \[=\frac{1}{12}\frac{\left(g-1\right)\left(3g^{5}-5g^{4}+16g^{3}+6g ^{2}+g-5\right)}{\left(g^{2}-g+1\right)^{2}}\,l^{2} \tag{20}\]
Since, by Eq., \(N\doteq g^{3}\) for the \(z\)=3 polymer, we have \(\left\langle s_{N}^{2}\right\rangle_{z=3}\doteq\frac{1}{4}\,N^{\frac{2}{3}}\,l^ {2}\) to yield \(\nu=1/3\), in agreement with the conjectured value in Table 1.
An essential point of the above calculation results- for the fully expanded polymers is that the mean squares of the radii of gyration are proportional to \(g^{2}\), namely, \(\langle s_{g}^{2}\rangle\propto g^{2}\), in contrast to the case of the ideal branched polymers that obey \(\langle s_{g}^{2}\rangle_{0}\propto g\).
\[\langle s_{g}^{2}\rangle\propto\left\{\begin{array}{ll}g&\mbox{ (for freely jointed molecules)}\\ g^{2}&\mbox{ (for fully expanded molecules)}\end{array}\right. \tag{21}\]
The results are in harmony with the initial scaling assumption, \(\left\langle s_{N}^{2}\right\rangle\propto g^{2\lambda}\), introduced in Eq.. From these examples, and taking dimensional symmetry into consideration, the relations in Eq. are probably a general rule and holds irrespective of the difference in architectures.
The exponent, \(\nu=1/2\), deduced above for the lattice analog of the isolated \(z\)=2 (extended comb) polymer conflicts with the thermodynamic theory, 5/8, the simulation experiments, \(\simeq 0.61\), on the square lattice, and the other theoretical methods, for the randomly branched polymer with the same ideal exponent \(\nu_{0}=1/4\). We notice that, because of the constraint condition, all those approaches are not applicable to the polymers such as the \(z\)=2 polymer (\(\nu_{0}=1/4\)) in the two-dimensional space and the \(z\)=3 polymer (\(\nu_{0}=1/6\)) in the three-dimensional space. The real exponent, \(\nu\), in the isolated state of these special polymers that violate the inequality can be evaluated by the use of Eq.; namely, for polymers that obey \(\nu_{0}<\frac{1}{d+1}\), we have \(\lambda=1\), so it must be that
\[\nu=2\nu_{0} \tag{22}\]
for the \(z\)=2 polymer, and in Eq. for the \(z\)=3 polymer.
In spite of the advent of such unprecedented polymers that violate the constraint condition, the fundamental equation of force for the excluded volume phenomena is invariably valid for almost all actual polymers synthesized in laboratories.
\[\frac{dG}{d\alpha}=\left(\mu_{c_{2I\!I}}-\mu_{c_{2I}}\right)\frac{dc_{2I\!I}} {d\alpha}+\frac{dG_{\rm elasticity}}{d\alpha}=0 \tag{23}\]
(In Eq., the subscript 2 denotes polymer segments; the subscripts \(I\) and \(I\!I\) a more dilute region and a more concentrated region, respectively.)Discussion
Through the previous to the present papers, we have shown that there are an infinite number of architectures having different exponents, \(\nu_{0}\). We showed that, of those architectures, only those that satisfy the inequality are compatible with the classic formula. We have investigated two examples that deviate from the inequality: one is the \(z=2\) nested structure (\(\nu_{0}=1/4\)) in \(d=2\), and the other is the \(z=3\) nested structure (\(\nu_{0}=1/6\)) in \(d=3\). Despite the deviation from the inequality, for instance, the \(z=2\) polymer in \(d=2\) is expected to have a reality, because it can be embedded in the two-dimensional space. We showed that the same is true for the \(z=3\) nested structure in \(d=3\). An intriguing aspect is that the mean squares of the radii of gyration obey the relationship: \(\left\langle s_{g}^{2}\right\rangle\propto g^{2}\) for the fully expanded polymers. From the viewpoint of the dimensional symmetry, there is a reason to conjecture that this relationship is independent of architectures.
An important fact clarified through this work is that the thermodynamic theory of the excluded volume effects and the scaling formula complement each other and are, by no means, exclusive. Eq. is useful for the architectures that satisfy \(\nu_{0}\geq\frac{1}{d+1}\) (most of the real polymers in \(d=3\) belong to this category!), while Eq. is useful in predicting the exponent, \(\nu\), of the architectures that fulfill \(\nu_{0}\leq\frac{1}{d+1}\), which necessarily leads to \(\lambda=1\) and \(\nu=2\nu_{0}\), under the condition of the dilution limit in good solvents. Whereas the scaling formula is a law restricted to \(N\to\infty\), the thermodynamic theory covers a broader range of \(N\) and a broader range of polymer concentration from the dilution limit to the melt; it is useful in extracting the physicochemical facets of the excluded volume phenomena.
## Appendix A Spatial Configuration of the \(z=3\) Polymer
We calculate the spatial configuration and the mean square of the radius of gyration of this polymer, which can be accomplished with the help of the Isihara formula:
## Figure A1
The \(z\)=3 polymer having the stretched backbone and the stretched side chains of the length, \(g\).
\[\vec{r}_{Gp}=\vec{r}_{1p}-\frac{1}{N}\sum_{p=1}^{N}\vec{r}_{1p}\] (A.1)
The key point is to rearrange the end-to-end vector, \(\vec{r}_{Gp}\), from the center of mass to the \(p\)th monomer into the grand sum of all the bond vectors that constitute the polymer: the result being
\[\vec{r}_{Gh}=\frac{1}{N}\left\{\sum_{k=1}^{h-1}[N-(N-k-(k-1)g^{2} )]\,\vec{l}_{(k+1)}-\sum_{k=1}^{g-1}\left[N-k-(k-1)g^{2}\right]\,\vec{l}_{(k+1)}\right. \\ \left.-\sum_{k=1}^{g-1}\left(g^{2}\,\vec{l}_{(k+1)_{1}}+\sum_{i=2} ^{g}\left\{g^{2}-[i-1+(i-2)g]\right\}\,\vec{l}_{(k+1)_{i}}\right)-\sum_{k=1}^{ g-1}\sum_{i=2}^{g}\sum_{p=1}^{g}(g-p+1)\vec{l}_{(k+1)_{ip}}^{\prime}\right.\\ \left.+\sum_{k=1}^{h-1}\left[N-k-(k-1)g^{2}\right]\,\vec{l}_{(k+1 )}\right\}\] (A.2)
\[\vec{r}_{Gh_{j}}=\frac{1}{N}\Bigg{\{}\sum_{k=1}^{h-1}[N-(N-k-(k-1)g ^{2})]\,\vec{l}_{(k+1)}+\left(N-g^{2}\right)\vec{l}_{h_{1}}+\sum_{i=2}^{j} \left[N-\left(g^{2}-[i-1+(i-2)g]\right)\right]\,\vec{l}_{h_{i}}\\ -\sum_{k=1}^{g-1}\left[N-k-(k-1)g^{2}\right]\,\vec{l}_{(k+1)}-\sum _{k=1}^{g-1}\left(g^{2}\,\vec{l}_{(k+1)_{1}}+\sum_{i=2}^{g}\left\{g^{2}-[i-1+(i -2)g]\right\}\,\vec{l}_{(k+1)_{i}}\right)\\ -\sum_{k=1}^{g-1}\sum_{i=2}^{g}\sum_{p=1}^{g}(g-p+1)\vec{l}_{(k+1) _{ip}}+\sum_{k=1}^{h-1}\left[N-k-(k-1)g^{2}\right]\,\vec{l}_{(k+1)}\\ +g^{2}\,\vec{l}_{h_{1}}+\sum_{i=2}^{j}\left(g^{2}-[i-1+(i-2)g] \right)\,\vec{l}_{h_{i}}\Bigg{\}}\] (A.3)
\[\vec{r}_{Gh_{jp}}=\frac{1}{N}\Bigg{\{}\sum_{k=1}^{h-1}[N-(N-k-(k- 1)g^{2})]\,\vec{l}_{(k+1)}+\left(N-g^{2}\right)\vec{l}_{h_{1}}+\sum_{i=2}^{j} \left[N-\left(g^{2}-[i-1+(i-2)g]\right)\right]\,\vec{l}_{h_{i}}\\ +\sum_{q=1}^{p}[N-(g-q+1)]\,\vec{l}_{h_{jq}}-\sum_{k=1}^{g-1}\left[ N-k-(k-1)g^{2}\right]\,\vec{l}_{(k+1)}\\ -\sum_{k=1}^{g-1}\left(g^{2}\,\vec{l}_{(k+1)_{1}}+\sum_{i=2}^{g} \left\{g^{2}-[i-1+(i-2)g]\right\}\,\vec{l}_{(k+1)_{i}}\right)-\sum_{k=1}^{g-1} \sum_{i=2}^{g}\sum_{q=1}^{g}(g-q+1)\vec{l}_{(k+1)_{iq}}\\ +\sum_{k=1}^{h-1}\left[N-k-(k-1)g^{2}\right]\,\vec{l}_{(k+1)}+g^{ 2}\,\vec{l}_{h_{1}}+\sum_{i=2}^{j}\left(g^{2}-[i-1+(i-2)g]\right)\,\vec{l}_{h_{ i}}+\sum_{q=1}^{p}(g-q+1)\,\vec{l}_{h_{jq}}\Bigg{\}}\] (A.4)
A1). In Eqs. (A.2)-(A.4)
\[\begin{array}{ccccc}\bullet&&1\leq h\leq g&&\mbox{for}&\vec{r}_{Gh}\\ \bullet&&2\leq h\leq g&\mbox{and}&1\leq j\leq g&&\mbox{for}&\vec{r}_{Gh_{j}}\\ \bullet&&2\leq h\leq g,&2\leq j\leq g,&\mbox{and}&1\leq p\leq g&&\mbox{for}& \vec{r}_{Gh_{jp}}\end{array}\] (A.5)
First, let us consider the fully expanded molecule. Assume that, in the limit of a large \(g\), this polymer takes the configuration with the stretched backbone and the stretched side chains (Fig. A1). Then, the polymer can marginally be arranged on the simple cubic lattice with the side length \(l\).
\[\left\langle s_{g}^{2}\right\rangle_{z=3} =\frac{1}{N}\left\{\left\langle\sum_{h=1}^{g}\vec{r}_{Gh}\cdot \vec{r}_{Gh}\right\rangle+\left\langle\sum_{h=2}^{g}\sum_{j=1}^{g}\vec{r}_{Gh_ {j}}\cdot\vec{r}_{Gh_{j}}\right\rangle+\left\langle\sum_{h=2}^{g}\sum_{j=2}^{g }\sum_{p=1}^{g}\vec{r}_{Gh_{jp}}\cdot\vec{r}_{Gh_{jp}}\right\rangle\right\}\] \[=\frac{1}{12}\frac{\left(g-1\right)\left(3g^{5}-5g^{4}+16g^{3}+6g ^{2}+g-5\right)}{\left(g^{2}-g+1\right)^{2}}\,l^{2}\] (A.6)
As \(g\rightarrow\infty\), Eq. (A.6) reduces to
\[\left\langle s_{g}^{2}\right\rangle_{z=3}\doteq\frac{1}{4}\,g^{2}\,l^{2}\] (A.7)
Applying the relation, \(N=g(g^{2}-g+1)\), in the text, we have \(\left\langle s_{N}^{2}\right\rangle_{z=3}\doteq\frac{1}{4}\,N^{\frac{2}{3}}\,l ^{2}\).
\[\nu_{z=3}=\frac{1}{3}\ \ \ \ (d=3)\] (A.8)
The result is in accord with the conjectured value in the text (Table 1).
## Mathematical Check
Apply \(g=2\) (\(N=6\)) to Eq. (A.6), and we have: \(\left\langle s_{N}^{2}\right\rangle_{z=3}=\frac{55}{36}\,l^{2}\). According to Fig. A1, each monomer on the \(z\)=3 polymer with \(g=2\) can be allocated to the following coordinates on the simple cubic lattice (\(l=1\)) shown in Fig. A2: \(\vec{r}_{O1}=(-1,0,0),\vec{r}_{O2}=,\vec{r}_{O3}=,\vec{r}_{O4}=( 0,2,0),\vec{r}_{O5}=,\vec{r}_{O6}=\). From this information, we find quickly \(\vec{r}_{OG}=(\frac{-1}{6},\frac{7}{6},\frac{3}{6})\) and hence, \(\left\langle s_{6}^{2}\right\rangle=\frac{55}{36}\) which is exactly the result that Eq. (A.6) predicts./
For the freely jointed model, Eqs. (A.2)-(A.5) give
\[\left\langle s_{g}^{2}\right\rangle_{0,z=3}=\frac{1}{6}\frac{\left(g-1\right) \left(7g^{5}-4g^{4}-13g^{3}+13g^{2}-2g+1\right)}{g\left(g^{2}-g+1\right)^{2}} \,l^{2}\] (A.9)
As \(g\rightarrow\infty\), this reduces to
\[\left\langle s_{g}^{2}\right\rangle_{0,z=3}\doteq\frac{7}{6}\,g\,l^{2}\] (A.10)
|
10.48550/arXiv.2105.07379
|
Branched Polymers with Excluded Volume Effects/ Relationship between Polymer Dimensions and Generation Number
|
Kazumi Suematsu, Haruo Ogura, Seiichi Inayama, Toshihiko Okamoto
| 4,561
|
10.48550_arXiv.2007.08026
|
###### Abstract
We introduce a machine learning method in which energy solutions from the Schrodinger equation are predicted using symmetry adapted atomic orbitals features and a graph neural-network architecture. OrbNet is shown to outperform existing methods in terms of learning efficiency and transferability for the prediction of density functional theory results while employing low-cost features that are obtained from semi-empirical electronic structure calculations. For applications to datasets of drug-like molecules, including QM7b-T, QM9, GDB-13-T, DrugBank, and the conformer benchmark dataset of Folmsbee and Hutchison, OrbNet predicts energies within chemical accuracy of DFT at a computational cost that is thousand-fold or more reduced.
+
Footnote †: preprint: AIP/123-QED
## I Introduction
The potential energy surface is the central quantity of interest in the modelling of molecules and materials. Calculation of these energies with sufficient accuracy in chemical, biological, and materials systems is in many - but not all - cases adequately described at the level of density functional theory (DFT). However, due to its relatively high cost, the applicability of DFT is limited to either relatively small molecules or modest conformational sampling, at least in comparison to force-field and semi-empirical quantum mechanical theories. A major focus of machine learning (ML) for quantum chemistry has therefore been to improve the efficiency with which potential energies of molecular and materials systems can be predicted while preserving accuracy.
In the context of quantum chemistry, many applications have focused on the use atom- or geometry-specific feature representations and kernel-based or neural-network (NN) ML architectures. Recent studies focus on the featurization of molecules in abstracted representations - such as quantum mechanical properties obtained from low-cost electronic structure calculations - and the utilization of novel graph-based neural network techniques to improve transferability and learning efficiency.
In this vein, we present a new approach (OrbNet) based on the featurization of molecules in terms of symmetry-adapted atomic orbitals (SAAOs) and the use of graph neural network methods for deep-learning quantum-mechanical properties. We demonstrate the performance of the new method for the prediction of molecular properties, including the total and relative conformer energies for molecules in a range of datasets of organic and drug-like molecules. The method enables the prediction of molecular potential energy surfaces with full quantum mechanical accuracy while enabling vast reductions in computational cost; moreover, the method outperforms existing methods in terms of its training efficiency and transferable accuracy across diverse molecular systems.
## II Method
The target of this work is to machine-learn a transferable mapping from input features values \(\{\mathbf{f}\}\) to the regression labels that are quantum mechanical energies,
\[E\approx E^{\text{ML}}\left[\{\mathbf{f}\}\right]. \tag{1}\]
The key elements of OrbNet include the efficient evaluation of the features in the SAAO basis, the utilization of a graph neural-network architecture with edge and node attention and message passing layers, and a prediction phase that ensures extensivity of the resulting energies. We summarize these elements in the current section and discuss the relationship between OrbNet and other ML approaches. Although results in the current paper are presented for the mapping of features from semi-empirical-quality features to DFT-quality labels, the method is general with respect to the mean-field method used for features (i.e., also allowing for Hartree-Fock, DFT, etc.) and the level of theory used for generating labels (i.e., also allowing for coupled-cluster and other correlated-wavefunction-method reference data).
### SAAO Features
Let \(\{\phi_{n,l,m}^{A}\}\) be the set of atomic orbital (AO) basis functions with atom index \(A\) and the standard principal and angular momentum quantum numbers, \(n\), \(l\), and \(m\). Let \(\mathbf{C}\) be the corresponding molecular orbital coefficient matrix obtained from a mean-field electronic structure calculation, such as HF theory, DFT, or a semi-empirical method.
\[P_{\mu\nu}=2\sum_{i\in\text{occ}}C_{\mu i}C_{\text{\tiny{vi}}} \tag{2}\]
(for a closed-shell system).
\[\mathbf{P}_{nl}^{A}\mathbf{V}_{nl}^{A}=\mathbf{Y}_{nl}^{A}\text{diag}(\lambda_ {nlm}^{A}) \tag{3}\]
For s orbitals (\(l=0\)), this symmetrization procedure is obviously trivial, and can be skipped. By construction, SAAOs are localized and consistent with respect to geometric perturbations of the molecule, and in contrast with localized molecular orbitals (LMOs) obtained from minimizing a localization objective function (Pipek-Mezey, Boys, etc.), SAAOs are obtained by a series of very small diagonalizations, without the need for an iterative procedure. The SAAO eigenvectors \(\mathbf{Y}_{nl}^{A}\) are aggregated to form a block-diagonal transformation matrix \(\mathbf{Y}\) that specifies the full transformation from AOs to SAAOs:
\[|\hat{\phi}_{p}\rangle=\sum_{\mu}Y_{\mu p}|\phi_{\mu}\rangle, \tag{4}\]
We employ ML features \(\{\mathbf{f}\}\) comprised of tensors obtained by evaluating quantum-chemical operators in the SAAO basis. Hereafter, all quantum mechanical matrices will be assumed to represented in the SAAO basis, including the density matrix \(\mathbf{P}\) and the overlap matrix \(\mathbf{S}\). Following our previous work, the features include expectation values of the Fock (\(\mathbf{F}\)), Coulomb (\(\mathbf{J}\)), and exchange (\(\mathbf{K}\)) operators in the SAAO basis. In this work, we additionally include the SAAO density matrix, \(\mathbf{P}\), the orbital centroid distance matrix, \(\mathbf{D}\), the core Hamiltonian matrix, \(\mathbf{H}\), and the overlap matrix, \(\mathbf{S}\); other quantum-mechanical matrix elements are also possible for featurization.
Summary of the OrbNet workflow. (a) A low-cost mean-field electronic structure calculation is performed for the molecular system, and (b) the resulting SAAOs and the associated quantum operators are constructed. (c) An attributed graph representation is built with node and edge attributes corresponding to the diagonal and off-diagonal elements of the SAAO tensors. (d) The attributed graph is processed by the embedding layer and message passing layers to produce transformed node and edge attributes. (e) The transformed node attributes for the encoding layer and each message passing layer are extracted and (f) passed to MPL-specific decoding networks. (g) The node-resolved energy contributions \(\epsilon_{\text{w}}\) are obtained by summing the decoding networks outputs node-wise, and (h) the final extensive energy prediction is obtained from a one-body summation over the nodes.
### Approximated Coulomb and exchange SAAO features
When a semi-empirical quantum chemical theory is employed, the computational bottleneck of SAAO feature generation becomes the \(\mathbf{J}\) and \(\mathbf{K}\) terms, due to the need to compute four-index electron-repulsion integrals.
\[(pq|rs)^{\text{MNOK}}=\sum_{A}\sum_{B}Q_{pq}^{A}Q_{rs}^{B}\gamma_{AB}. \tag{5}\]
Here, \(A\) and \(B\) are atom indices, \(p,q,r,s\) are SAAO indices, and
\[\gamma_{AB}^{\{\mathbf{J},\mathbf{K}\}}=\left(\frac{1}{R_{AB}^{\gamma(\mathbf{ J},\mathbf{K})}+\eta^{-\gamma(\mathbf{J},\mathbf{K})}}\right)^{1/\gamma_{\{ \mathbf{J},\mathbf{K}\}}}, \tag{6}\]
In this work, we used \(y_{\mathbf{J}}=4\) and \(y_{\mathbf{K}}=10\) similar to which employed in the sTDA-RSH method.
\[Q_{pq}^{A}=\sum_{\mu\in A}Y_{\mu p}^{\prime}Y_{\mu q}^{\prime}, \tag{7}\]
This yields approximated \(\mathbf{J}\) and \(\mathbf{K}\) matrices for featurization,
\[J_{pq}^{\text{MNOK}}=(pp|qq)^{\text{MNOK}}=\sum_{A,B}Q_{pp}^{A}Q_{qq}^{B} \gamma_{AB}^{\mathbf{J}} \tag{8}\]
\[K_{pq}^{\text{MNOK}}=(pq|pq)^{\text{MNOK}}=\sum_{A,B}Q_{pq}^{A}Q_{pq}^{B} \gamma_{AB}^{\mathbf{K}} \tag{9}\]
A naive implementation of Eqs. 8 and 9 is \(\mathcal{O}(N^{4})\), the leading asymptotic cost. However, this scaling may be reduced to \(\mathcal{O}(N^{2})\) with negligible loss of accuracy through a tight-binding approximation; for molecules in this study, computation of \(\mathbf{J}^{\text{MNOK}}\) and \(\mathbf{K}^{\text{MNOK}}\) is not the leading order cost for feature generation and such tight-binding approximation is thus not employed.
### OrbNet
OrbNet encodes the molecular system as graph-structured data and utilizes a graph neural network (GNN) machine-learning architecture. The GNN represents data as an attributed graph \(G(\mathbf{V},\mathbf{E},\mathbf{X}\mathbf{^{e}})\), with nodes \(\mathbf{V}\), edges \(\mathbf{E}\), node attributes \(\mathbf{X}:\mathbf{V}\rightarrow\mathbb{R}^{n\times d}\), and edge attributes \(\mathbf{X}^{\mathbf{e}}:\mathbf{E}\rightarrow\mathbb{R}^{m\times e}\), where \(n=|V|\), \(m=|E|\), and \(d\) and \(e\) are the number of attributes per node and edge, respectively.
Specifically, OrbNet employs a graph representation for a molecular system in which node attributes correspond to diagonal SAAO features \(X_{u}=[F_{uu},J_{uu},K_{uu},P_{uu},H_{uu}]\) and edge attributes correspond to off-diagonal SAAO features \(X_{uv}^{\mathbf{e}}=[F_{uv},J_{uv},K_{uv},D_{uv},P_{uv},S_{uv},H_{ur}]\). By introducing an edge attribute cutoff value for edges to be included, non-interacting molecular systems separated at infinite distance are encoded as disconnected graphs, thereby satisfying size-consistency.
The model capacity is enhanced by introducing nonlinear input-feature transformations to the graph representation via radial basis functions,
\[\mathbf{h}_{u}^{\text{RBF}}=[\phi_{1}^{\text{h}}(\tilde{X}_{u}),\phi_{2}^{ \text{h}}(\tilde{X}_{u}),...,\phi_{n_{t}}^{\text{h}}(\tilde{X}_{u})] \tag{10}\]
\[\mathbf{e}_{uv}^{\text{RBF}}=[\phi_{1}^{\text{e}}(\tilde{X}_{uv}^{\text{e}}), \phi_{2}^{\text{e}}(\tilde{X}_{uv}^{\text{e}}),...,\phi_{m}^{\text{e}}(\tilde {X}_{uu}^{\text{e}})], \tag{11}\]
Sine basis functions \(\phi_{n}^{\text{h}}(r)=\sin(\pi nr)\) are used for node embedding.
\[\phi_{m}^{\text{e}}(r)=j_{0}^{m}(r/c_{\mathbf{X}})\cdot I_{\mathbf{X}}(r)= \sqrt{\frac{2}{c_{\mathbf{X}}}}\frac{\sin(\pi nr/c_{\mathbf{X}})}{r/c_{ \mathbf{X}}}\cdot I_{\mathbf{X}}(r), \tag{12}\]
To ensure that the feature varies smoothly when a node enters the cutoff, we further introduce the mollifier \(I_{\mathbf{X}}(r)\):
\[I_{\mathbf{X}}(r)=\begin{cases}\exp\left(-\frac{c_{\mathbf{X}}^{2}}{(|r|-c_{ \mathbf{X}})^{2}}+1\right)&\quad\text{if }0\leq|r|<c_{\mathbf{X}}\\ 0&\quad\text{if }|r|\geq c_{\mathbf{X}}\end{cases} \tag{13}\]
Note that \(\phi_{m}^{\text{e}}(r)\) decays to zero as an edge approaches the cutoff to ensure size-consistency, and the mollifier is infinite order differentiable at the boundaries, which eliminates representation noise that can arise from geometric perturbation of the molecule.
\[\mathbf{e}_{uv}^{\text{aux}}=\mathbf{W}^{\text{aux}}\cdot\mathbf{e}_{uv}^{ \text{RBF}}, \tag{14}\]
The radial basis function embeddings are transformed by neural network modules to yield 0-th order node and edge attributes,
\[\mathbf{h}_{u}^{0}=\text{Enc}_{\text{h}}(\mathbf{h}_{u}^{\text{RBF}}),\ \mathbf{e}_{uv}^{0}=\text{Enc}_{\text{e}}(\mathbf{e}_{uv}^{\text{RBF}}). \tag{15}\]
In contrast to atom-based message passing neural networks, this additional embedding transformation captures the interactions among the physical operators.
The node and edge attributes are updated via the Transformer-motivated message passing mechanism in For a given message passing layer (MPL) \(l+1\), the information carried by each edge is encoded into a message function \(\mathbf{m}_{uv}^{l}\) and associated attention weight \(\mathbf{m}_{uv}^{l}\), and is accumulated into node features through a graph convolution operation.
\[\mathbf{m}_{uv}^{l}=\sigma(\mathbf{W}_{\text{m}}^{l,2}\cdot\big{[}(\mathbf{W}_ {\text{m}}^{l,1}\cdot\mathbf{h}_{u}^{l})\odot(\mathbf{W}_{\text{m}}^{l,1} \cdot\mathbf{h}_{v}^{l})\odot\mathbf{e}_{uv}^{l}\big{]}+\mathbf{b}_{\text{m}}^ {l}) \tag{17}\]
and the convolution kernel weights, \(w_{uv}^{l,i}\), are evaluated as (multi-head) attention scores to characterize the relative importance of an orbital pair,
\[w_{uv}^{l,j}=\sigma_{\text{a}}(\sum[(\mathbf{W}_{\text{a}}^{l,j}\cdot\mathbf{ h}_{u}^{l})\odot(\mathbf{W}_{\text{a}}^{l,j}\cdot\mathbf{h}_{v}^{l})\odot\mathbf{e}_{ uv}^{l}\odot\mathbf{e}_{uv}^{\text{aux}}]/n_{\text{e}}), \tag{18}\]
Here, the index \(i\) specifies a single attention head, and \(n_{\text{e}}\) is the dimension of hidden edge features \(\mathbf{e}_{uv}^{l}\), \(\bigoplus\) denotes a vector concatenation operation, \(\odot\) denotes the Hadamard product, and \(\cdot\) denotes the matrix-vector product.
The edge attributes are updated according to
\[\mathbf{e}_{uv}^{l+1}=\sigma(\mathbf{W}_{\text{e}}^{l}\cdot\mathbf{m}_{uv}^{l} +\mathbf{b}_{\text{e}}^{l}), \tag{19}\]
\(\mathbf{W}_{\text{m}}^{l,1},\mathbf{W}_{\text{m}}^{l,2},\mathbf{W}_{\text{h}}^ {l},\mathbf{W}_{\text{e}}^{l},\mathbf{b}_{\text{m}}^{l},\mathbf{b}_{\text{h}}^ {l},\mathbf{b}_{\text{e}}^{l}\) and \(\mathbf{a}^{l}\) are MPL-specific trainable parameter matrices, \(\mathbf{W}_{\text{a}}^{l,i}\) are MPL- and attention-head-specific trainable parameter matrices, \(\sigma(\cdot)\) is an activation function with a normalization layer, and \(\sigma_{\text{a}}(\cdot)\) is the activation function used for generating attention scores.
The decoding phase of OrbNet (Fig. 1f-h) is designed to ensure the size-extensivity of energy predictions. The employed mechanism outputs node-resolved energy contributions for the embedding layer (\(l=0\)) and all MPLs (\(l=1,2,...,L\)) to predict the energy components associated with all nodes and MPLs. The final energy prediction \(E^{\text{ML}}\) is obtained by first summing over \(l\) for each node \(u\) and then performing a one-body sum over nodes (i.e., orbitals), such that
\[E^{\text{ML}}=\sum_{u\in\mathbb{V}}\varepsilon_{u}=\sum_{u\in\mathbb{V}}\sum_{ l=0}^{L}\text{Dec}^{l}(\mathbf{h}_{u}^{l}), \tag{20}\]
### Comparison with other methods that use quantum mechanical features
Several ML methods have been developed for the prediction of high-level (i.e., coupled-cluster) correlation energies based on quantum mechanical features from a mean-field-level (i.e., HF theory or DFT) electronic structure calculation. An example from our own work includes the molecular-orbital-based machine-learning (MOB-ML) approach to predict molecular properties using localized molecular orbitals for input feature generation. Localized molecular orbitals are obtained via an orbital localization procedure (Boys, IBO, etc), with the orbitals obtained from a mean-field electronic structure calculation. Feature vectors are then calculated for diagonal and off-diagonal molecular orbital pairs from matrix elements of the molecular orbitals with respect to various operators (i.e., Fock, Coulomb, and exchange operators) within the basis and using a feature sorting scheme. Gaussian-process or clustering-based regressors are trained for the pair correlation energy labels associated to the MOB feature vectors.
Closer in spirit to OrbNet are NeuralXC and DeePHF, which employ AO-based features obtained from electronic structure calculations to perform the regression and prediction of molecular energies. Both NeuralXC and DeePHF utilize the electronic density and orbitals obtained from either a Hartree-Fock (HF) (in DeePHF) or low-level density functional theory (DFT) (in NeuralXC) calculation using cc-pVDZ or larger atomic-orbital basis sets. However, these methods typically require a mean-field calculation in the same-sized atomic orbital basis set as that of the high-level correlation method (i.e., they do not directly make predictions on the basis of features that are obtained in a minimal basis), and they have not been applied for the prediction of DFT-quality results on the basis of lower-level semi-empirical methods, such as GFN-xTB, as is done here.
Summary of the OrbNet MPL update. For the \(l+1\) MPL, the attributes of a given node (blue) are updated due to interactions with nearest-neighbor nodes (red and gold), which depend on both the nearest-neighbor node attributes and the nearest-neighbor edge attributes. The node and edge features (i.e., \(\mathbf{h}_{u}^{l}\), \(\mathbf{h}_{v}^{l}\), and \(\mathbf{e}_{uv}^{l}\)) combine to produce a message \(\mathbf{m}_{uv}^{l}\) (Eq. 17) and multi-head attention score \(\mathbf{w}_{uv}^{l}\) (Eq. 18) which undergo attention mixing. The attention-weighted messages from each nearest-neighbor node and edge are combined and passed into a dense layer, the result of which is added to the original node attributes to perform the update (Eq. 16).
In terms of featurization methods, OrbNet differs from NeuralXC and DeePHF by providing a more information-rich quantum mechanical representation. Unlike NeuralXC, OrbNet avoids shell-averaging of the AOs, and unlike both NeuralXC and DeePHF, OrbNet includes all off-diagonal operator matrix elements (including both intra- and inter-atom elements, as well as intra- and inter-shell elements) within the features, thereby preserving information content while also enabling description of long-range contributions. Unlike DeePHF, OrbNet includes interactions between different shells on the same atom and avoids the need for a pre-determined weighting function based on inter-atomic distances. OrbNet additionally includes quantum-chemical matrices including \(\mathbf{F},\mathbf{J},\mathbf{K}\) which are valuable components for energy prediction tasks. Other differences arise in the way in which rotational invariance is enforced within the features. In NeuralXC, the rotational invariance of the features is guaranteed by summing all sub-shell components of the AO-projected density \(d^{ml}=\sum_{m=-l}^{l}c_{nlm}^{2}\) (i.e. the trace of the local density matrix), such that the information content is not preserved. In DeePHF, the rotational invariance of the features is enforced by using the eigenvalues of the local density matrix instead of the trace to build the feature vector for each shell. By contrast, OrbNet achieves the rotational invariance of features through the use of SAAOs, which involve no loss of information content.
In terms of ML regression methods, OrbNet also differs from NeuralXC and DeePHF. For NeuralXC, the ML regression is performed using a Behler-Parrinello type dense neural network. Similarly, for DeePHF, the ML regression is performed using a dense neural network, with the labels associated with a one-body summation over the atoms to yield the total correlation energy. In contrast, OrbNet uses a GNN for the ML regression. Specifically, we report results using a multi-head graph attention mechanism and residual blocks to improve the representation capacity of the model, to learn complex chemical environments. Unlike the pre-tuned aggregation coefficients in DeePHF, OrbNet also offers a flexible framework for learning orbital interactions and could be naturally transferred to downstream tasks.
## III Computational details
Results are presented for the QM7b-T dataset (which has seven conformations for each of 7211 molecules with up to seven heavy atoms of type C, O, N, S, and Cl), the QM9 dataset (which has locally optimized geometries for 133885 molecules with up to nine heavy atoms of type C, O, N, and F), the GDB-13-T dataset (which has six conformations for each of 1000 molecules from the GDB-13 dataset with up to thirteen heavy atoms of type C, O, N, S, and Cl), DrugBank-T (which has six conformations for each of 168 molecules from the DrugBank database with between fourteen and 30 heavy atoms of type C, O, N, S, and Cl), and the Hutchison conformer dataset from Ref. (which has up to 10 conformations for each of 622 molecules with between nine and 50 heavy atoms of type C, O, N, F, P, S, Cl, Br, and I). Except for DrugBank-T, all of these datasets have been described previously; thermalized geometries from the DrugBank dataset are sampled at 50 fs intervals from _ab initio_ molecular dynamics trajectories performed using the B3LYP/6-3g level of theory and a Langevin thermostat at 350 K. The structures for the DrugBank-T dataset are provided in the Supporting Information, and all other employed datasets are already available online. For results reported in Section IV.1, the pre-computed DFT labels from Ref. were employed. For results reported in Section IV.2, all DFT labels were computed using the \(\omega\)B97X-D functional with a Def2-TZVP AO basis set and using density fitting for both the Coulomb and exchange integrals using the Def2-Universal-JKFIT basis set; these calculations are performed using Psi4. Semi-empirical calculations are performed using the GFN1-xTB method using the Entos Qcore package, which is also employed for the SAAOs feature generation.
For the results presented in this work, we train OrbNet models using the following training-test splits of the datasets. For results on the QM9 dataset, we removed 3054 molecules due to a failed a geometric consistency check, as recommended in Ref.; we then randomly sampled 110000 molecules for training and used 10831 molecules for testing. The training sets of 25000 and 50000 molecules in section IV.1 are subsampled from the 110000-molecule dataset. For the QM7b-T dataset, two sets of training-test splits are generated; for the model trained on the QM7b-T dataset only (Model 1 in Section IV.2), we randomly selected 6500 different molecules (with 7 geometries for each) from the total 7211 molecules for training, holding out 500 molecules (with 7 geometries for each) for testing; for Models 2-4 in Section IV.2, we used a 361-molecule subset of this 500-molecules set for testing, and we used the remaining 6850 molecules of QM7b-T for training. For the GDB13-T dataset, we randomly sampled 948 different molecules (with 6 geometries for each) for training, holding out 48 molecules (with 6 geometries for each) for testing. For the DrugBank-T dataset, we randomly sampled 158 different molecules (with 6 geometries for each) for training, holding out 10 molecules (with 6 geometries for each) for testing. No training on the Hutchison conformer dataset was performed, as it was only used for transferability testing. Since none of the training datasets for OrbNet included molecules with elements of type P, Br, and I, we excluded the molecules in the Hutchison dataset that included elements of these types for the reported tests (as was also done in Ref. and in for the ANI methods). Moreover, following Ref., we excluded sixteen molecules due to missing DLPNO-LCCSD(T) reference data; an additional eight molecules were excluded on the basis of DFT convergence issues for at least one conformer using Ps14. The specific molecules that appear in all training-test splits are listed in the Supporting Information.
Table 1 summarizes the hyperparameters used for training OrbNet for the reported results.
\(-\ln(|X_{ur}|)\) for \(\mathbf{X}\in\{\,\mathbf{F},\mathbf{J},\mathbf{K},\mathbf{P},\mathbf{S},\mathbf{H }\,\}\), and \(\tilde{D}_{ur}=D_{ur}\). The model hyperparameters are selected within a limited search space; the cutoff hyperparameters \(c_{\mathbf{X}}\) are obtained by examining the overlap between feature element distributions between the QM7b-T and GDB13-T datasets. The same set of hyperparameters is used throughout this work, thereby providing a universal model.
To provide additional regularization for predicting energy variations from the configurational degree of freedom, we performed training on loss function of the form
\[\mathcal{L}(\hat{\mathbf{E}},\mathbf{E}) =(1-\alpha)\sum_{t}\mathcal{L}_{2}(\hat{E}_{i},E_{i})\] \[+\alpha\sum_{i}\mathcal{L}_{2}(\hat{E}_{i}-\hat{E}_{t(i)},E_{i}- E_{t(i)}). \tag{21}\]
For a conformer \(i\) in a minibatch, we randomly sample another conformer \(t(i)\) of the same molecule to be paired with \(i\) to evaluate the relative conformer loss \(\mathcal{L}_{2}(\hat{E}_{i}-\hat{E}_{t(i)},E_{i}-E_{t(i)})\), putting additional penalty on the prediction errors for configurational energy variations. \(\hat{\mathbf{E}}\) denotes the ground truth energy values of the minibatch, \(\hat{\mathbf{E}}\) denotes the model prediction values of the minibatch, and \(\mathcal{L}_{2}\) denotes the L2 loss function \(\mathcal{L}_{2}(y,y)=||\hat{y}-y||_{2}^{2}\). For all models in Section IV.1, we choose \(\alpha=0\) as only the optimized geometries are available; for models in Section IV.2, we choose \(\alpha=0.9\) for all training setups.
All models are trained on a single Nvidia Tesla V100-SXM2-32GB GPU using the Adam optimizer. For all training runs, we set the minibatch size to 64 and use a cyclical learning rate schedule. that performs a linear learning rate increase from \(3\times 10^{-5}\) to \(3\times 10^{-3}\) for the initial 100 epochs, a linear decay from \(3\times 10^{-3}\) to \(3\times 10^{-5}\) for the next 100 epochs, and an exponential decay with a factor of 0.9 every epoch for the final 100 epochs. Batch normalization is employed before every activation function \(\sigma\) except for that used in the attention heads, \(\sigma_{\text{a}}\).
## IV Results
We present results that focus on the prediction of accurate DFT energies using input features obtained from the GFN1-xTB method. The GFN family of methods have proven to be extremely useful for the simulation of large molecular system (1000s of atoms or more) with time-to-solution for energies and forces on the order of seconds. However, this applicability can be limited by the accuracy of the semi-empirical method, thus creating a natural opportunity for "delta-learning" the difference between the GFN1 and DFT energies on the basis of the GFN1 features.
\[E^{\text{ML}}\approx E^{\text{DFT}}-E^{\text{GFN1}}-\Delta E^{\text{fit}}_{ \text{atoms}}, \tag{22}\]
This approach yields the direct ML prediction of total DFT energies, given the results of a GFN1-xTB calculation.
### The QM9 dataset
We begin with a broad comparison of recently introduced ML methods for the total energy task, \(U_{0}\), from the widely studied QM9 dataset. QM9 is composed of organic molecules with up to nine heavy atoms at locally optimized geometries, so this test (Table 2) examines the expressive power of the ML models for systems in similar chemical environments. Results for OrbNet are presented both without ensemble averaging of independently trained models (i.e., predicting only on the basis of the first of trained model) and with ensemble averaging the results of five independently trained models (OrbNet-ens5). As observed previously, ensembling helps in this and other learning tasks, reducing the OrbNet prediction error by approximately 10-20%.
Also included in the table are previously published methods utilizing graph representations of atom-based features, including SchNet, PhysNet, DimeNet, and DeepMoleNet. We note that DimeNet employs a directional message passing mechanism and PhysNet and DeepMoleNet employ supervision based on prior physical information to improve the model transferability, which could also be employed within OrbNet; it is clear that without these additional strategies and even without model ensembling, OrbNet provides greater accuracy and learning efficiency than all previous deep-learning methods.
### Transferability and Conformer Energy Predictions
A more realistic and demanding test of ML methods is to train them on datasets of relatively small molecules (for which high-accuracy data is more readily available) and then to test on datasets of larger and more diverse molecules. This provides useful insight into the transferability of the ML methods and the general applicability of the trained models.
To this end, we investigate the performance of OrbNet on a series of dataset containing organic and drug-like molecules. presents results in which OrbNet models are trained with increasing amounts of data. Using the training-test splits described in Section III, Model 1 is trained using data from only the QM7b-T dataset; Model 2 is trained using data from the QM7b-T, GDB13-T, and DrugBank-T datasets; Model 3 is trained using data from the QM7b-T, QM9, GDB13-T, and DrugBank-T datasets; and Model 4 is obtained by ensembling five independent training runs with the same data as used for Model 3. Predictions are made for total energies and relative conformer energies for held-out molecules from each of these datasets, as well as for the Hutchison conformer dataset.
As expected, it is seen from that the OrbNet predictions improve with additional data and with ensemble modeling. The median and mean of the absolute errors consistently decrease from Model 1 to Model 4 except for a non monotonicity in the DrugBank-T MAE, likely due to the relatively small size of that dataset. It is nonetheless striking that Model 1, which includes only data from QM7b-T yields relative conformer energy predictions on the DrugBank-T and Hutchison datasets (which include molecules with up to 50 heavy atoms) with an accuracy that is comparable to the more heavily trained models. Note that all of the OrbNet models predict relative conformer energies with MAE and median prediction errors that are well within the 1 kcal/mol threshold of chemical accuracy, across all four test datasets. Predictions for QM9 using Models 1 and 2 are not included, since QM9 includes F atoms whereas the training data in those models do not; relative conformer energies are not predicted for QM9 since they are not available in this dataset. Although total energy prediction error for the OrbNet is slightly larger per heavy atom on the Hutchison dataset than for the other datasets, the relative conformer energy prediction error for the Hutchison dataset is slightly smaller than for GDB13-T and DrugBank-T; this is due to the fact that the Hutchison dataset involves locally minimized conformers that are less spread in energy per heavy atom than the conformers of the thermalized datasets. This relatively small energy spread among conformers in the Hutchison dataset is a realistic and challenging aspect of drug-molecule conformer-ranking prediction, which we next consider.
For the Hutchison conformer dataset of drug-like molecules which range in size from nine to 50 heavy atoms, the accuracy of the various methods was evaluated using the median R\({}^{2}\) of the predicted conformer energies in comparison to DLPNO-CCSD(T) reference data and with computation time evaluated on a single CPU core.
The OrbNet conformer energy predictions (Fig. 4, black) are reported using Model 4 (i.e., with training data from QM7b-T, GDB13-T, DrugBank-T, and QM9 and with ensemble averaging over five independent training runs). The solid black circle indicates the median R\({}^{2}\) value (0.81) of the OrbNet predictions relative to the DLPNO-CCSD(T) reference data, as for the other methods; this point provides the most direct comparison to the accuracy of the other methods. The open black circle indicates the median R\({}^{2}\) value (0.90) of the OrbNet predictions relative to the \(\omega\)B97X-D/Def2-TZVP reference data against which the model was trained;
\begin{table}
\begin{tabular}{c c c c c c c} Training size & SchNet & PhysNet & PhysNet-ens & DimeNet & DeepMoleNet & OrbNet & OrbNet-ens \\ \hline
25,000 & - & - & - & - & - & **11.6** & **10.4** \\ \hline
50,000 & 15 & 13 & 10 & - & - & **8.22** & **6.80** \\ \hline
110,000 & 14 & 8.2 & 6.1 & 8.02 & 6.1 & **5.01** & **3.92** \\ \end{tabular}
\end{table}
Table 2: MAEs (reported in meV) for predicting the QM9 dataset of total energies at the B3LYP/6-31G(2df,p) level of theory. Results from the current work are reported for a single model (OrbNet) and with ensembling over 5 models (OrbNet-ens5).
\begin{table}
\begin{tabular}{c c c} Hyperparameter & Meaning & Value or name \\ \hline \(n_{\mathrm{r}}\) & Number of basis functions for node embedding & 8 \\ \hline \(m_{\mathrm{r}}\) & Number of basis functions for edge embedding & 8 \\ \hline \(n_{\mathrm{h}}\) & Dimension of hidden node attributes & 256 \\ \hline \(n_{\mathrm{e}}\) & Dimension of hidden edge attributes & 64 \\ \hline \(n_{\mathrm{a}}\) & Number of attention heads & 4 \\ \hline \(L\) & Number of message passing layers & 3 \\ \hline \(L_{\mathrm{enc}}\) & Number of dense layers in \(\mathrm{Enc}_{\mathrm{h}}\) and \(\mathrm{Enc}_{\mathrm{e}}\) & 3 \\ \hline \(L_{\mathrm{dec}}\) & Number of dense layers in a decoding network & 4 \\ \hline & Hidden dimensions of a decoding network & 128, 64, 32, 16 \\ \hline \(\sigma\) & Activation function & Swish \\ \hline \(\sigma_{\mathrm{a}}\) & Activation function for attention generation & TanhShrink \\ \hline \(\gamma\) & Batch normalization momentum & 0.4 \\ \hline \(c_{\mathrm{F}}\) & Cutoff value for \(\tilde{F}_{\mathrm{av}}\) & 8.0 \\ \hline \(c_{\mathrm{J}}\) & Cutoff value for \(\tilde{J}_{\mathrm{av}}\) & 1.6 \\ \hline \(c_{\mathrm{K}}\) & Cutoff value for \(\tilde{R}_{\mathrm{av}}\) & 20.0 \\ \hline \(c_{\mathrm{D}}\) & Cutoff value for \(\tilde{D}_{\mathrm{av}}\) & 9.45 \\ \hline \(c_{\mathrm{P}}\) & Cutoff value for \(\tilde{P}_{\mathrm{av}}\) & 14.0 \\ \hline \(c_{\mathrm{S}}\) & Cutoff value for \(\tilde{S}_{\mathrm{av}}\) & 8.0 \\ \hline \(c_{\mathrm{H}}\) & Cutoff value for \(\tilde{H}_{\mathrm{av}}\) & 8.0 \\ \hline \hline \end{tabular}
\end{table}
Table 1: Model hyperparameters employed in OrbNet. All cutoff values are in atomic units.
We performed timings for OrbNet on a single core of an Intel Core i5-1038NG7 CPU 2.00GHz, finding that the OrbNet computational cost is dominated by the GFN1-xTB calculation for the feature generation. In contrast to Ref. which used the xtb code of Grimme and coworkers, we used Entos Qcore for the GFN1-xTB calculation calculations. We find the reported timings for GFN1-xTB to be surprisingly slow in Ref., particularly in comparison to the GFN0-xTB timings. For GFN0-xTB, our timings with Entos Qcore are very similar to those reported in Ref., which is sensible given that the method involves no self-consistent field (SCF) iteration. However, whereas Ref. indicates GFN1-xTB timings that are 43-fold slower than GFN0-xTB, we find this ratio to be only 4.5 with Entos Qcore, perhaps due to differences of SCF convergence. To account for the issue of code efficiency in the GFN1-xTB implementation and to control for the details of the single CPU core used in the timings for this work versus in Ref., we normalize the OrbNet timing reported in with respect to the GFN0-xTB timing from Ref.. The CPU neural-network inference costs for OrbNet are negligible contribution to this timing.
The results in make clear that OrbNet enables the prediction of relative conformer energies for drug-like molecules with an accuracy that is comparable to DFT but with a computational cost that is 1000-fold reduced from DFT to realm of semiempirical methods. Alternatively viewed, the results indicate that OrbNet provides dramatic improvements in prediction accuracy over currently available ML and semiempirical methods for realistic applications, without significant increases in computational cost.
## V Conclusions
Electronic structure methods typically face a punishing trade-off between the prediction accuracy of the method and its computational cost, across all areas of the chemical, biological, and materials sciences. We present a new machine-learning method with the potential to substantially shift that trade-off in favor of _ab initio_-quality accuracy at low computational cost. OrbNet utilizes a graph neural network architecture to predict high-quality electronic-structure energies on the basis of features obtained from low-cost/minimal-basis mean-field electronic structure methods. The method is demonstrated for the case of predicting \(\omega\)B97X-D/Def2-TZVP energies using GFN1-xTB input features, although it is completely general with respect to both the choice of high-level (including correlated wavefunction) method used for generating reference data and the choice of mean-field method used for feature generation. In comparison to state-of-the-art GNN methods for the prediction of total molecule energies for the QM9 dataset, it is shown that OrbNet provides a 33% improvement in prediction accuracy with the same amount of data relative to the next-most accurate method (DeepMoleNet). And in comparison to the wide array of methods used for predicting relative conformer energies in a realistic and diverse dataset of drug-like molecules, as compiled by Folmsbee and Hutchison, it is shown that OrbNet provides a prediction accuracy that is similar to DFT and much improved over existing ML methods, but at a computational cost that is reduced by at least three orders of magnitude relative to DFT. Natural future directions for development will include the expansion of OrbNet to a broader set of chemical elements, incorporation of directional message-passing and model supervision using prior physical information, and end-to-end refitting of the semi-empirical method used for feature generation.
## VI Supplemental Material
The supplemental material includes the structures for the DrugBank-T dataset, as well as specification of molecules that appear in all training-test splits for the trained models.
Prediction errors for (a) molecule total energies and (b) relative conformer energies performed using OrbNet models trained using various datasets. The mean absolute error (MAE) is indicated by the bar height, the median of the absolute error is indicated by a black dot, and the the first and third quantities for the absolute error are indicated as the lower and upper bars. Model 1 uses training data from QM7b-T; Model 2 additionally includes training data from GDB13-T and DrugBank-T; Model 3 additionally includes training data from QM9; and Model 4 additionally includes ensemble averaging over five independent training runs. Testing is performed on data that is held-out from training in all cases. Training and prediction employs energies at the \(\omega\)B97X-D/Def2-TZVP level of theory. All energies in kcal/mol.
|
10.48550/arXiv.2007.08026
|
OrbNet: Deep Learning for Quantum Chemistry Using Symmetry-Adapted Atomic-Orbital Features
|
Zhuoran Qiao, Matthew Welborn, Animashree Anandkumar, Frederick R. Manby, Thomas F. Miller III
| 5,610
|
10.48550_arXiv.1511.07126
|
## 1 Introduction
Increasing the sensitivity of nuclear magnetic resonance (NMR) experiments is always an ongoing research direction in the NMR area which can significantly broaden NMR applications, especially for molecular and cellular imaging. Hyperpolarization methods including chemically induced dynamic nuclear polarization (CIDNP), para-hydrogen-induced polarization (PHIP), spin-exchange optical pumping (SEOP), dynamic nuclear polarization (DNP) have opened a new avenue in the field of NMR. They can enhance the sensitivity by several orders of magnitude where the nuclear polarization is greatly increased by themanipulation of spin states. In those methods, CIDNP and PHIP are generated by spin-sensitive chemical reactions and it has so far remained limited to specialized chemical systems and nuclei. SEOP are achieved by optical pumping with circularly polarized laser light which is a way of hyper-polarizing noble gases such as helium, neon, krypton, and xenon only. These three methods all have some drawbacks either complex experimental conditions (low temperature, laser polarization), or only for specific circumstances (a specific chemical reaction or nucleus). DNP is an electron-nuclear double resonance technique, where the spin polarization of high-\(\gamma\) electrons is transferred to the surrounding nuclei to enhance the NMR sensitivity using microwave (MW) irradiation under general ambient conditions. In general, DNP can be utilized to enhance the polarizations of all kinds of nuclei, such as \({}^{1}\)H, \({}^{19}\)F, \({}^{13}\)C, \({}^{31}\)P, \({}^{23}\)Na.
Magnetic resonance imaging (MRI), as a non-radioactive and non-invasive means, has become an important tool in clinical diagnostics and biomedical research. MRI, in most case, is based on the proton signal, because the other nuclei signals of biological small molecules, for example, \({}^{13}\)C and \({}^{15}\)N are too weak to be imaged. DNP enhancement is an important method to enhance the biological small molecules MRI sensitivity _in vitro_ and _in vivo_. DNP-MRI molecular imaging technologies can provide the unique information to study the generation and development mechanism of the disease. Especially, it can diagnose and treat of disease by imaging special molecules. For example, we can obtain the spatial distribution and the changes over time of several related metabolic disease molecules by direct imaging. Such an important information is difficult to yield by other microscopic molecular imaging methods.
Parallel acquisition with multiple receivers was originally introduced into MRI to reduce the experiment time. The multiple receiver coils which cover the object surface in MRI systems usually detect the same nuclear species at different spatial regions. Parallel acquisition of NMR signals for two, three, or multiple different nuclear species has proven its enormous potential for unambiguous structure determination of organic molecules and fast multidimensional NMR spectra. A new type of experiment, named PANACEA (parallel acquisition NMR and all-in-one combination of experimental applications) allows structure determination from a single measurement and includes an internal field/frequency correction routine.
DNP mechanisms in the dielectric solid state and liquid state are different. For dielectric solid where the paramagnetic centers (free radicals or unpaired electrons) are fixed, there are three main enhancement processes: solid effect (SE), cross effect (CE) and thermal mixing (TM). On the other hand, DNP in liquid and metal where the paramagnetic centers (radicals) are movable is achieved through Overhauser effect (OE). To understand different DNP enhancement mechanisms in solid and liquid, we assume here a 2 spin coupled system of electron and nuclear spins resulting in the standard four-energy level model. The maximum DNP enhancement of the NMR signal for the nuclei with frequency \(\omega_{\rm n}\) in solid is obtained by irradiating double or zero quantum transition with microwave frequencies at \(\omega_{\rm e}\)+\(\omega_{\rm n}\) or \(\omega_{\rm e}\)-\(\omega_{\rm n}\). Clearly, solid DNP enhancement is highly selective of nuclear resonance frequency, and the microwave irradiating with frequencies \(\omega_{\rm e}\)+\(\omega_{\rm n}\) can only enhance the polarization of one kind of nuclei with frequency \(\omega_{\rm n}\) when the EPR line width is narrow. However, the enhancement of Overhauser DNP effect in liquid is yielded by irradiating electron single-quantum transition with electron Larmor frequency \(\omega_{\rm e}\). The maximum Overhauser DNP enhancement as shown in appears at the same MW frequency \(\omega_{\rm e}\) for all different kinds of nuclei and is thus independent of nuclear frequency. That is to say, the polarizations of all different kinds of nuclei can be enhanced simultaneously by saturating electron transition.
This fascinating advantage of simultaneous multiple nuclear Overhauser enhancements, however, has not been exploited experimentally up to now. The simultaneous multiple nuclear DNP enhancements once combined with parallel acquisition method will extremely promotes NMR developments in both MR spectroscopy (MRS) and MR image (MRI). In this paper, by parallel acquisition we have observed simultaneously DNP-enhanced spectra of different nuclei, such as \({}^{1}\)H and \({}^{23}\)Na, \({}^{1}\)H and \({}^{31}\)P, \({}^{1}\)H and \({}^{13}\)C, \({}^{19}\)F and \({}^{31}\)P, especially for the first time to report sodium ion enhancement in liquid. The first phosphorous image at low field by solution state DNP, together with parallel \({}^{1}\)H image, is also obtained, which may advance the development of MRI of X nuclei (non-proton) at low field.
## 2 Materials and methods
To demonstrate the principle and superiority of simultaneous acquisition of Overhauser-enhanced multi-nuclei spectra and images in solution, here we present two simply illustrative examples of two-nuclei co-acquisition on our homebuilt multi-channel DNP spectrometer at 0.35T, although the method is clearly applicable to more sophisticated experiments involving three or more different nuclear species. After the saturation pulse of MW at electron Larmor frequency, all of the conventional pulse sequences can be used in each channel. Signals from different channels can be separately acquired either simultaneously or at different stages of their respective pulse sequences. A series of in-house-built double resonance probes are used for various nuclei with simultaneous acquisition in NMR and MRI
Four level scheme of a coupled electron (S = -1/2)-nucleus (I = 1/2) spin pair (a). Nuclear polarization enhancement as a function of the microwave frequency (b), the maximum DNP in solid is at \(\omega_{\rm e}\)+\(\omega_{\rm n}\) and \(\omega_{\rm e}\)+\(\omega_{\rm n}\) (black), in liquid is at \(\omega_{\rm e}\) (red).
shows the simplest pulse sequences for NMR spectra (a) and images (b) used in this work.
Here, we present the experimental data obtained simultaneously from the samples in Table 1.
The detail processes for samples configuration are described in following:
Sample 1 for \({}^{1}\)H and \({}^{31}\)P resonances: the free radicals \(\alpha\), \(\gamma\)-bisdiphenylene-\(\beta\)-phenylallyl (BDPA complex with benzene (1:1)) was dissolved directly into 20 \(\mu\)L benzene. And then 30 \(\mu\)L PCl\({}_{3}\) is dissolved into the mixed solution. The concentration of free radicals is about 60mM.
Sample 2 for \({}^{1}\)H and \({}^{31}\)P resonances: The same free radicals BDPA was dissolved directly into 60 \(\mu\)L benzene. And then 50 mg triphenyl phosphorous ((C\({}_{6}\)H\({}_{5}\))\({}_{3}\)P, PPh\({}_{3}\)) is dissolved into the mixed solution. The concentration of free radicals is about 90mM. Volumes of approximately 30\(\mu\)L of sample 1 and sample 2 were loaded into each 1.7 mm inner diameter glass capillaries and sealed and used for both DNP spectra and image measurements.
\begin{table}
\begin{tabular}{|p{56.9pt}|p{56.9pt}|p{56.9pt}|p{56.9pt}|p{56.9pt}|} \hline & Sample 1 & Sample 2 & Sample 3 & Sample 4 \\ & (\({}^{1}\)H and \({}^{31}\)P) & (\({}^{1}\)H and \({}^{31}\)P) & (\({}^{1}\)H and \({}^{23}\)Na) & (\({}^{19}\)F and \({}^{31}\)P) \\ \hline radicals & BDPA (60mM) & BDPA (90mM) & TEMPOL (120mM) & BDPA (62mM) \\ \hline solute & 30 \(\mu\)L PCl\({}_{3}\) & 50 mg (C\({}_{6}\)H\({}_{5}\))\({}_{3}\)P & NaCl & 40\(\mu\)L C\({}_{6}\)F\({}_{6}\), 80 mg (C\({}_{6}\)H\({}_{5}\))\({}_{3}\)P \\ \hline solvent & 20 \(\mu\)L C\({}_{6}\)H\({}_{6}\) & 60 \(\mu\)L C\({}_{6}\)H\({}_{6}\) & H\({}_{2}\)O & 50 \(\mu\)L C\({}_{6}\)D\({}_{6}\) \\ \hline sample volumes & 30\(\mu\)L & 30\(\mu\)L & 25\(\mu\)L & 30\(\mu\)L \\ \hline \end{tabular}
\end{table}
Table 1: samples used in this work.
Two-channel acquired pulse sequences used in this work. (a) \(\pi\)/2 pulse enhancement sequence; (b) 2D spin echo image enhancement sequence; p1, p2: \(\pi\)/2 pulse, p1’, p2’: \(\pi\) pulse.
Sample 3 for \({}^{1}\)H and \({}^{23}\)Na resonances: the concentrations of NaCl is 4.8M and the free radical is TEMPOL with 120mM, volumes of approximately 25\(\upmu\)L were loaded into 1 mm inner diameter glass capillaries and sealed.
Sample 4 for \({}^{19}\)F and \({}^{31}\)P resonances: The free radicals \(\alpha\),\(\gamma\)-bisdiphenylene-\(\beta\)-phenylallyl (BDPA complex with benzene(1:1)) was dissolved directly into 50 \(\upmu\)L deuterated benzene and 40\(\upmu\)L hexafluorobenzene. And then 80 mg triphenyl phosphorous ((C\({}_{6}\)H\({}_{5}\))\({}_{3}\)P, PPh\({}_{3}\)) is dissolved into the mixed solution. The concentration of free radicals is about 62 mM. Volumes of approximately 30\(\upmu\)L were loaded into 1.7 mm inner diameter glass capillaries and sealed.
Hereafter in the text, the samples will be referred to as sample 1, sample 2 and sample 3, sample 4 without mentioning the concentrations of free radicals and the kinds of the solute. The structures of several molecules used in our studies for \({}^{1}\)H and \({}^{31}\)P are shown in Fig.3. This structures is determined the J-coupling effect between \({}^{1}\)H and \({}^{31}\)P in PPh\({}_{3}\) (Sample 2) but not in PCl\({}_{3}\) (Sample 1).
## 3 Result
### \({}^{1}\)H and \({}^{31}\)P spectra
Phosphorus is especially interesting because of its chemical diversity of different oxidation states in organic solvent molecules and its significance in biological materials. The DNP \({}^{31}\)P enhancement of a number of phosphorus compounds has been reported early by Potenza et al. Fig.4 is the simultaneously acquired \({}^{31}\)P and \({}^{1}\)H spectra of sample 1 with (Fig.4a and b) and without (Fig.4c and d) DNP enhancement at one scan using the pulse sequence in FIG2a. The time of microwave saturation is 2s and microwave power is 8W (MW resonator input terminal). We observe a large \({}^{1}\)H signal enhancement of up to 60 fold in benzene which is determined from the ratio of the areas of the spectra. We also observe an intensive Overhauser-enhanced \({}^{31}\)P signal at MW irradiation but such a \({}^{31}\)P signal is not observable without MW irradiation even it was accumulated for 1024 scans. The normalized \({}^{31}\)P DNP enhancement factor \(\varepsilon\) of sample 1 as a function of the microwave power is shown in Fig.5.
Structure of various molecules used in our studies for \({}^{1}\)H and \({}^{31}\)P
enhancement factor of \({}^{23}\)Na since \({}^{23}\)Na signal is undetectable without MW irradiation. Our next goal is to obtain the spectra or image of \({}^{23}\)Na and \({}^{39}\)K simultaneously, which is significant for clinical diagnostics and biomedical research.
## Fig.6. Simultaneously acquired \({}^{1}\)H and \({}^{23}\)Na spectra after one scan. (a) and (c) are \({}^{1}\)H spectra with (red) and without (blue) MW irradiation. (b) and (d) are \({}^{23}\)Na spectra with (red) and without (blue) MW irradiation. The thermally polarized spectrum is multiplied 8 times.
### \({}^{19}\)F and \({}^{31}\)P spectra
## Fig.7. Simultaneously acquired \({}^{19}\)F and \({}^{31}\)P spectra with and without Overhauser enhancement for benzene/hexafluorobenzene/triphenyl phosphorous solutions. (a) and (c) are \({}^{19}\)F spectra with (red) and without(blue) MW irradiation. (b) and (d) are \({}^{31}\)P spectra with (red) and without (blue) MW irradiation. The thermally polarized spectrum is magnified 8 times.
Due to the relatively high sensitivity, no interfering background signals and much broader chemical shift range, \({}^{19}\)F NMR spectroscopy has a potential application in molecular biology and biochemistry investigations including the structure and function of proteins and nucleic acids, enzymatic mechanisms, metabolic pathways and biomolecular interactions \({}^{26}\). DNP enhanced \({}^{19}\)F NMR could compensate the loss of sensitivity at lower fields as compared to high field measurements. Fig.7 shows the simultaneously acquired \({}^{19}\)F and \({}^{31}\)P spectra with enhancement (Fig.7a and b) and without enhancement (Fig.7c and d) at 0.35T. The microwave pumping time is 2s and microwave power is about 30W (resonator input terminal). Very strong \({}^{19}\)F and \({}^{31}\)P NMR signals are seen in the DNP-enhanced co-acquisition spectra, but in the co-acquisition spectra without DNP enhancement \({}^{19}\)F signal is very weak and \({}^{31}\)P signal is totally unobservable. \({}^{19}\)F DNP enhancement is about 20.
### \({}^{1}\)H and \({}^{31}\)P image
**Fig.8.**\({}^{31}\)P (a-e) and \({}^{1}\)H (f-j) images obtained at 0.35T. Sample 1: PCI\({}_{3}\) and benzene, sample 2: triphenyl phosphorous and benzene. a), b), c), and f), g), h) are acquired separately from \({}^{31}\)P and \({}^{1}\)H; d) and i), e) and j) are from parallel acquisitions. Independent control of TE to obtain the image of T\({}_{2}\)-weight, a), b), d), and f), g), i): TE=11ms; c), e), h), j): TE=35ms. Other experiment parameters: TR=3s, phase encode = 128, MW power = 8W.
An in-house-built user interface software, which supports experimental operations, data processing and graphic editing pulse sequences, was used for MRI. Because the same gradient and sampling bandwidth were utilized for spatial encoding in the \({}^{31}\)P and \({}^{1}\)H co-image (Fig.2b), the different nucleus spatial frequency sampling leads to different FOV and resolution. This difference in FOV and resolution can be corrected by scaling the k-space trajectory or interpolating data in image space during image reconstruction. Fig.8 shows a series of \({}^{31}\)P and \({}^{1}\)H images with (Fig.8b-e, g-j) and without (Fig.8a, f) microwave irradiation, in the cases of parallel acquisition (Fig.8d-e, i-j) and single-channel acquisition (Fig.8b-c, g-h). The numbers 1 and 2 in Fig.8 denote sample 1 and sample 2 respectively. One sees clearly the shape of the sample from DNP-enhanced images (Fig.8b and g), but can not discern the profile of the sample without DNP enhancement from either \({}^{31}\)P or \({}^{1}\)H (Fig.8a and f). Since the \({}^{31}\)P \(T_{2}\) of PCl\({}_{3}\) is much shorter than that of the triphenyl phosphorous ((C\({}_{6}\)H\({}_{5}\))\({}_{3}\)P, PPh\({}_{3}\)), we can only observe the image of \({}^{31}\)P in PPh\({}_{3}\) but not in PCl\({}_{3}\) (Fig.8c) when the echo time TE is long. Owing to the J-coupling effect between \({}^{1}\)H and \({}^{31}\)P in PPh\({}_{3}\), we can see the parallel images of \({}^{31}\)P and \({}^{1}\)H only in the case of short TE during which \({}^{31}\)P evolution caused by \({}^{31}\)P-\({}^{1}\)H J-coupling in PPh\({}_{3}\) can be neglected. Whereas for long TE period during which a pair of \(\pi\) pulses of \({}^{1}\)H and \({}^{31}\)P was applied simultaneously and the J-coupling transforms the \({}^{31}\)P in-phase coherence into the anti-phase coherence, the \({}^{31}\)P image from parallel acquisition becomes unobservable in PPh\({}_{3}\) (Fig.8e). But the \({}^{1}\)H image is still clear in the latter case because most of the protons are not J-coupled with the \({}^{31}\)P. In addition, we can also obtain the chemical shift of \({}^{31}\)P nuclei with different position shifting from the center of the image (Fig.8b).
### \({}^{19}\)F and \({}^{31}\)P image
We have also simultaneously obtained the co-images of \({}^{19}\)F and \({}^{31}\)P of sample 4. Fig.9 shows the images of \({}^{19}\)F and \({}^{31}\)P with (Fig.9a, b) and without (Fig.9c, d) microwave irradiation. We see clearly the \({}^{31}\)P and \({}^{19}\)F images of the sample in the case of DNP enhancement (Fig.9a, b) but can not discern the profile of the sample without DNP enhancement from either \({}^{31}\)P or \({}^{19}\)F (Fig.9c, d).
## 4 Discussion and Conclusion
In summary, we have realized the parallel spectra and images of two nuclei based on simultaneous multiple-nuclei Overhauser DNP enhancement. These DNP enhanced multi-nuclei NMR spectra/images may provide more significant information of metabolism _in vivo_. For example, we may simultaneously monitor the sodium ion transport and phosphorus metabolism by analyzing the spectral data. Without doubt, simultaneous acquisition of multi-nuclear enhanced magnetic resonance through DNP offers a number of advantages over classical simultaneous acquisition and DNP
\({}^{19}\)F (a,c) and \({}^{31}\)P (b,d) co-images with BDPA radical from parallel acquisitions. Other experiment parameters: TR=3s, TE=40ms, phase encode = 128, MW power = 30W.
It could significantly broaden NMR applications in biomedical systems especially for studying low-\(\gamma\) nuclei which are important for metabolism. Structural information, ion transport pathway and metabolic processes can be simultaneously observed with various nuclear spectroscopy and image. The spin-spin coupled relationship between the different nuclei can be extracted either from two-dimensional and multidimensional spectra or image and used for three-dimensional structure refinement. At the same time, it significantly reduces the acquisition time for multidimensional and quantitative analysis experiments. This method can also be used for many fast acquisition techniques, for instance, Hadamard NMR spectroscopy, computer optimized aliasing and projection reconstruction method of protein with enhanced signal at low field. What's more, we can image with different nuclei at a single molecule of different position or at different molecules to obtain considerably complementary information. We can obtain two different images if the distribution of the sample is inhomogeneous or the chemical shift of nuclei can be resolved distinctly. This method can also be used in high field dynamic nuclear polarization to obtain the high-resolution and high-sensitivity of NMR spectra and image for biological samples.
|
10.48550/arXiv.1511.07126
|
Simultaneous Acquisition of Multi-nuclei Enhanced NMR/MRI by Solution State Dynamic Nuclear Polarization
|
Yugui He, Zhen Zhang, Jiwen Feng, Chongyang Huang, Fang Chen, Maili Liu, Chaoyang Liu
| 4,221
|
10.48550_arXiv.1712.01789
|
# Abstract
We analyze the Coulomb hole of Ne from highly-accurate CISD wave functions obtained from optimized even-tempered basis sets. Using a two-fold extrapolation procedure we obtain highly accurate results that recover 97% of the correlation energy. We confirm the existence of a shoulder in the short-range region of the Coulomb hole of the Ne atom, which is due to an internal reorganization of the _K_-shell caused by electron correlation of the core electrons. The feature is very sensitive to the quality of the basis set in the core region and it is not exclusive to Ne, being also present in most of second-row atoms, thus confirming that it is due to _K_-shell correlation effects.
## 1 Introduction
Electron correlation remains as a central issue for the physico-chemical description of the electronic structure. Its study often provides physical insights to develop new computational methods to tackle the electronic structure of molecules. The primitive description provided by the Hartree-Fock (HF) wave function has been improved by consideration of different types of electron correlation, such as dynamic and nondynamic correlation, in the so-called post-HF methods as well as in methods that do not employ wave functions, such as the density and reduced-density matrix functional theories (DFT\({}^{\text{GH}}\) and RDMFT\({}^{\text{}}\)). The improvement of computational methods, the correct choice of a computational protocol to address a given problem, and our understanding of the electron correlation, hinge on the development of appropriate descriptors of electron correlation.\({}^{\text{}}\) Lately, our efforts have concentrated in this direction, resulting in the development of simple electron correlation descriptors capable of separating dynamic and nondynamic correlation.\({}^{\text{}}\)
The Coulomb hole stands among the classical descriptors that are used to study electron correlation due to its conceptual simplicity and its connection with the electron-electron interaction energy.\({}^{\text{}}\) The Coulomb hole provides a practical picture of how the electron correlation affects the interelectronic separation. Namely, it reflects the change of the electron-electron distance distribution upon the inclusion of electron correlation. From this quantity the correlation effects on the average interelectronic distance, its variance, and the electron-electron repulsion are easily assessed. The topological features of the Coulomb hole have also been studied, leading to some relevant conclusions about the nature of electron correlation.\({}^{\text{}}\) Some of us have also recently used the long-range part of the Coulomb hole to characterize van der Waals interactions.\({}^{\text{}}\)
In this work, we analyze a key feature of the Coulomb hole of the Ne atom that, thus far, has been largely ignored by many quantum mechanics practitioners. In 1969, Bunge and co-reorganization of the _K_-shell caused by electron correlation of the core electrons. The feature is very sensitive to the quality of the basis set in the core region and it is not exclusive to Ne, being also present in most of second-row atoms, thus confirming that it is due to _K_-shell correlation effects.
We have found that the shoulder is very sensible to the quality of the basis sets employed in the calculation, turning into a minimum or vanishing depending on the basis set. In order to confirm the presence of the shoulder we have performed CSD and FCI calculations employing large optimized even-tempered basis sets, which provide energy estimates that compare well with the most accurate values obtained by Bunge.\({}^{\text{}}\) Our results provide a thorough study on the origin of the shoulder, identifying the causes that are responsible for its presence. Finally, we prove that this feature is not exclusive to Ne atom.
## 2 Methodology
There are mainly three different ways to define correlation holes: McWeeny's,\({}^{\text{}}\) Ros\({}^{\text{}}\) and Coulson's.\({}^{\text{}}\) The former is statistically motivated and it does not employ reference wave functions, whereas the other two use HF as the uncorrelated reference. In this work, we are concerned with Coulson's definition, which is connected with an experimental observable, the X-ray scattering intensity. Coulson's Coulomb hole is obtained from the difference between the exact and the HF intracule densities.
\[l(u)=\iint d\mathbf{r}_{12}d\mathbf{r}_{12}\,n_{2}(\mathbf{r}_{1},\mathbf{r}_{2})\delta(u-r_{12}), \tag{1}\]
The X-ray scattering intensity is essentially determined by the Fourier-Bessel transform of the intracule pair density\({}^{\text{}}\) and it is employed in the study of elastic and inelastic scattering of electrons.\({}^{\text{}}\) The total X-ray scattering intensity for short wavelengths is actually governed by the value of the intracule at the coalescence points, \(l\).\({}^{\text{}}\)The difference between the exact pair density and an uncorrelated reference, represents the change in the electron pair distribution upon the introduction of electron correlation.
\[h_{\text{c}}(u)=I(u)-I_{\text{HF}}(u) \tag{2}\]
The integration of \(h_{\text{c}}(u)\) over \(u\) gives zero.
Since the quality of the basis set is crucial for the description of the holes at short electron-electron distances, we have generated an optimized set of basis functions. The optimization of the basis sets employs an analoguous procedure to the one developed elsewhere.23 This procedure has been successfully used to generate highly-accurate basis functions to test model systems and calibrate a number of methods.16,22,48,56 First of all, a family of uncontracted basis sets consisting of spherical Gaussian primitives is constructed by selecting the optimized exponents that minimize the CISD energies (the coefficients that multiply the primitives are equal to 1 and do not enter the optimization procedure). From these values, the complete-basis set (CRS) estimate of the CISD energies are obtained by a two-fold extrapolation procedure.
The family of basis sets employs functions with exponents \(\xi_{k,N}^{4}\) that are even-tempered51 according to the expression
\[\xi_{k,N}^{4}=a_{k,N}\left[\beta_{k,N}\right]^{k-1},\quad 1\leq k\leq N. \tag{3}\]
Each basis set is characterized by the maximum angular momentum, \(L\), and the number of basis functions for each function type, \(N\). For instance, 65P (\(L\)=1, \(N\)=6) basis set consists of six groups of functions containing one 5 and three P functions (\(p_{r}\), \(p_{r}\) and \(p_{r}\)) sharing the same exponent. The exponent assigned to each group is given by \(k\) in Eq. 3, which runs from 1 to \(N\). \(a_{k,N}\) and \(\beta_{k,N}\) are, therefore, unique for each basis set and determined by minimization of the CISD energy of We with a simple method (minimal accuracy \(10^{-7}\) a.u.). The family includes basis sets with angular momentum between 0 and \(L\) (\(1\leq L\leq 4\)) and involve equal numbers \(N\) (\(6\leq N\leq 16\)) of spherical Gaussian primitives with exponents \(\xi_{k,N}^{4}\) giving rise to 44 different basis sets.
The computed energies \(E_{k,N}\) have been extrapolated to the respective \(N\rightarrow\infty\) limits \(E_{i}\) by fitting the actual energy values for \(N\)=12, 13, 14, 15 and 16 with the double-exponential expression
\[E_{k,N}=E_{k}+a_{k}e^{-a_{k}u,N}+b_{k}e^{-a_{k}u,N}, \tag{4}\]
which generalizes the Dunning extrapolation.56 The resulting system of five non-linear equations has been solved analytically with Mathematica88 employ the Ramanujan algorithm.10
Footnote 10: The authors, published by Wiley-VCH Verlag GmbH & Co., KGa.
In turn, the estimates \(E_{k}\) have been extrapolated to the respective CBS limits \(E\) by fitting the values of \(E_{k}\) for \(L\)=2, 3, and 4 to the expression112,48,56
\[E_{k}=F+\frac{B}{[L+1]^{3}}. \tag{5}\]
These extrapolations, \(E_{k}\) and \(E_{r}\) provide lower-energy estimates of the total energy that are not variational.
In the case of HF, the energy results are almost converged using only S and P basis functions. Therefore, we take the SP-energy limit as a good estimate of the CBS-extrapoled result. The numerical estimate is obtained from \(N\)=16, 17, 18, 19 and 20 calculations applying the fitting of Eq. 5.
The full-configuration interaction (FCI) calculations have been carried out with a modified version of the FCI program of Knowles and Handy41 and the CISD calculations have been performed with Gaussian.114 The calculations of the second-order reduced density matrices (2-RDM) have been calculated from the FCI/CISD expansions coefficients using the in-house DMN code.16,14 The radial intractive density was computed with the in-house RHO2_OPS code,10 which uses the algorithm proposed by Gioslowski and Liu.148
Footnote 10: The authors, published by Wiley-VCH Verlag GmbH & Co., KGa.
## 3 Results
### Benchmark Data
Following the procedure described in the previous section we have obtained a CISD extrapolated energy of \(-128.9254609\) a.u., which represents an energy lowering of \(-0.0143843\) a.u. with respect to the best variational estimate, \(E_{k,16}\) (see Table 1). These results compare well with the best non-relativistic FCI estimate available in the literature, \(-128.937588\) a.u.21
Footnote 21: The authors, published by Wiley-VCH Verlag GmbH & Co., KGa.
Our CISD SP-energy limit, \(-128.8984284\) a.u. is in good agreement with the FCI value \(-128.897\pm 0.002\) a.u. calculated by Bunge.23 This and the other partial waves reported in Table 1 are also in accord with the second-order correlation energies of Lindgren and Salomonson.22
Footnote 22: The authors, published by Wiley-VCH Verlag GmbH & Co., KGa.
Our extrapolated HF energy, \(-128.547100\) a.u., which also corresponds to the SP-energy limit, is in excellent agreement with the numerical HF results, \(-128.547098\) a.u., reported elsewhere.23 Our best CISD estimate of the correlation energy
\begin{table}
\begin{tabular}{c c c c c} \(N\) & \(E_{k,N}\) & \(E_{k,N}\) & \(E_{k,N}\) & \(E_{k,N}\) \\ \hline
5 & \(-127.8146757\) & \(-127.9311638\) & \(-127.9638254\) & \(-127.9755176\) \\
6 & \(-128.3326841\) & \(-128.4543077\) & \(-128.4887366\) & \(-128.5010355\) \\
7 & \(-128.5692718\) & \(-128.692709\) & \(-128.7276231\) & \(-128.7395973\) \\
8 & \(-128.6612022\) & \(-128.68290592\) & \(-128.8299039\) & \(-128.8335145\) \\
9 & \(-128.6987672\) & \(-128.8240698\) & \(-128.8599894\) & \(-128.8725885\) \\
10 & \(-128.7167209\) & \(-128.843421\) & \(-128.8974215\) & \(-128.892215\) \\
11 & \(-128.7265762\) & \(-128.4582151\) & \(-128.8888828\) & \(-128.1907158\) \\
12 & \(-128.7304973\) & \(-128.5867898\) & \(-128.8932237\) & \(-128.90562091\) \\
13 & \(-128.233966\) & \(-128.8588008\) & \(-128.9854104\) & \(-128.905610\) \\
14 & \(-128.734578\) & \(-128.806036\) & \(-128.897452\) & \(-128.9096489\) \\
15 & \(-128.7340477\) & \(-128.8607425\) & \(-128.8974582\) & \(-128.9106198\) \\
16 & \(-128.73443430\) & \(-128.8611063\) & \(-128.8978764\) & \(-128.9110766\) \\
17 & \(-128.7346499\) & \(-128.861534\) & \(-128.8984284\) & \(-128.9117007\) \\ \end{tabular}
\end{table}
Table 1: CISD energies (a.u.) for the basis set family developed in this work and the corresponding partial waves.
Our best variational estimate of the correlation energy, based on the CISD/165PDFG calculation (including 400 basis functions), recovers 93% of the correlation energy. Our calculations on the angular and the radial correlation indicators of Kutzelnigg show no qualitative improvement in the description of correlation beyond the 115PDF basis set (see Figures S1 and S2) and Bunge and coworkers report very small effects upon introduction of the triple and quadruple excitations (less than 0.01% change on the density). Therefore, we conclude that our CISD calculations provide a satisfactory description of electron correlation in Ne.
We have also explored the convergence of certain properties related to the Coulomb hole with the size of the basis set. Our results indicate that the average interelectronic distance and its variance are much more affected by the number of basis functions than by the inclusion of functions of large angular momentum. In this respect, the use of 9SP basis functions provides a reasonable description of these indicators (see Figures S3 and S4). For this reason, we have chosen the CISD/95P wave function to provide a qualitative explanation of the Coulomb hole in Ne atom. In a number of selected cases, analysis with larger basis sets have been performed to confirm our conclusions.
### The Coulomb Hole of the Ne Atom
In his seminal paper, Bunge reported a small shoulder of the Coulomb hole of Ne in the short interelectronic distances domain that he attributed to the electron correlation within the _K_-shell. This calculation was based on a FCI wave function that yield an electronic energy of -128.8602 a.u. and, thus, only retrieved 85% of the correlation energy. Thirty years later, Closlowski and Liu confirmed this result using 2-RDMs obtained from energy derivatives of MP2 calculations with a non-optimized even-tempered basis set of 50 functions (20s10_p_). We have tried to reproduce the results of Bunge and Closlowski and have encountered a major difficulty choosing the appropriate basis set. We have performed over hundred CISD calculations (and some FCI calculations as well using different basis with and without the frozen core approximation, finding that the shoulder is only reproduced in about half of the cases (see Tables S1 and S2). No frozen-core calculation could reproduce the shoulder structure regardless the size of the basis set, supporting the idea that this feature, if not an artifact due to inaccurateness of the wave function, is a result of the correlation of the core electrons. The basis set families show similar results among its members. Pople's 6-311G and larger basis of this family as well as the first family of basis sets developed by Dunning (n2) and the core correlated-consistent basis sets cc-pCVn2 display the shoulder structure. Conversely, the family of correlated-consistent basis sets of Dunning (cc-pVn2) and the series of basis sets of Petersson (n2aP) cannot reproduce the shoulder structure (see for some examples).
In order to solve this controversy, we have built a series of even-tempered basis sets following the procedure described above. For all these basis sets, regardless the size, the shoulder structure shows at ca. 0.1 A (see Figure 2). The whole profile of the Coulomb hole is very sensitive to the basis set. Increasing the number of basis functions improves the description of interelectronic cusps, shifting the hole to shorter electron-electron distances. For small basis sets, including only S and P
The CISD Coulomb hole of Ne for some selected basis sets.
The CISD Coulomb hole of Ne for some even-tempered basis sets.
Augmenting with F functions does not produce a large change, and the addition of G functions barely changes the Coulomb hole, thus suggesting that the presence of the shoulder is not due to a basis set completeness problem (see Figure 2). Actually, the presence of the shoulder structure was also reported using Monte Carlo calculations.58 The shoulder always appears when we use basis sets with enough flexibility to afford a correct description of core electrons, indicating that those electrons are causing the shoulder. The role of core orbitals is confirmed by the corresponding frozen-core CISD (fc-CISD) calculations which do not show any shoulder structure (see in the Supporting Information). Correlation effects on the _K_-shell of the have been previously studied by Buijse and Baerends. They used a modified version of the Coulomb hole proposed by Ros58 based on conditional probability densities. They showed that core-electron correlations have a major impact on the Coulomb hole plots for Ne when the reference electron is located on the _K_-shell.22
### The Origin of the Shoulder
In this section we analyze the reasons for the existence of the shoulder in the Coulomb hole of Ne. We have already established that the correlation of core electrons is responsible for it. Let us now consider the importance of different configurations by removing some of them from the CISD expansion calculated with the 9SP basis set. In we have plotted the Coulomb hole generated with this wave function and other wave functions in which we have truncated the CISD expansions including only some excitations from the 1s orbital (see caption of for more details). The truncation of the determinant expansion has been performed after a standard CISD calculation, normalizing the resulting wave function after the truncation. The CISD expansion in which we have removed all the excitations from the core orbital except the single excitations (CISD(nc) + A in Figure 3) produces a Coulomb hole that is virtually identical to the fc-CISD one. The double excitations involving only one electron in 1s2 produce likewise a Coulomb hole qualitatively similar to the fc-CISD wave function (CISD(nc) + B). Among the double excitations the most important ones are those exciting simultaneously both 1s2 electrons as evidenced from the shoulder structure of the Coulomb hole of the CISD wave function where only these excitations from the core orbital are retained (CISD(nc) + C). A detailed analysis of the double excitations from the 1s orbital shows that the preferred virtual orbitals are 4s, 5s, 5\(p\), and 6\(p\) (see CISD(nc) + D Coulomb hole in Figure 3). These results have been qualitatively confirmed with the CISD/165PDFG wave function (see in the Supporting Information).
Unlike the previous CISD expansions, these ones only include correlation effects due to the core electrons in Ne and, therefore, should reflect the importance of certain configurations in retrieving the shoulder. The inclusion of double excitations from the core orbitals gives rise to a hole structure (see HF + E in Figure 4) that is responsible for the shoulder structure of the complete CISD expansion. From this plot is also evident that double excitations and particularly those involving 4s, 5s, 5\(p\), and 6\(p\) are mostly responsible for the shoulder structure.
Thus far, we have firmly established the presence of the shoulder in the Coulomb hole of Ne, which is due to the
The CISD/9SP Coulomb hole in terms of several expansions. The groups of configurations included involve excitations from the core orbitals to some particular virtual orbitals (see the caption of for C and D). The E group includes configurations involving double excitations from 1s to all virtual orbitals.
The CISD/9SP Coulomb hole in terms of several expansions. fc-CISD calculations were obtained from a CISD calculation in which no excitations from core orbitals were allowed, whereas CISD(nc) is a regular CISD calculation in which the configurations involving excitations from the core orbital have been removed _a posteriori_. _A_–_C_ are groups of configurations including various excitations from the core orbital (A) single excitations, (B) double excitations involving only one electron in the core orbital, (C) double excitations involving the two electrons in the core orbital excited to one single orbital, and (D) double excitations involving the two electrons in the core orbital excited to orbitals 4s, 5s, 5\(p\), and 6\(p\). After removal and addition of these configurations, the expansion coefficients have been rescaled to attain the normalization of the wave function.
In the following, we will analyze how the correlation affects the electronic structure of Ne and the particular role that the core electrons play in this context using the CISD/165P wave function. First of all, we will consider the shell-structure of Ne. There has been some controversy in the literature concerning the descriptor that should be employed to identify the shell structure and shell numbers in atoms,25,26 in our opinion, the one-electron potential (OEP) of Kohout being the most robust suggestion made thus far.27 According to the OEP, we find that the radius of the \(K\) shell does not change upon inclusion of electron correlation effects (_r_x=0.138 A) and the _K_-shell number only increases 3 * 10-3 electrons due to correlation (\(n_{e}^{\text{ss}}=2.0019\) e.). Therefore, according to the shell structure determined by the OEP, we conclude that electron correlation does not cause an expansion or contraction of the \(K\) shell, but a small reorganization within the \(K\) shell.
We have also checked the convergence of the electron-electron repulsion and the electron-nucleus attraction to see how these quantities are affected by the frozen-core approximation. These energy components show a convergence pattern that alternates fc-CISD results with CISD results, suggesting that wave functions that do not show the shoulder structure do not converge these properties differently (see Figures 56 and 57). Conversely, as one could expect, we have found that the intracule of the pair density at the coalescence point divided by the charge-concentration index, \(\int\rho^{2}(\mathbf{r})d\mathbf{r}_{1}^{2}\),58,60 is affected by the inclusion of core correlation (see Figure S8).
Finally, let us assess the type of correlation affecting the shoulder structure. We will use our recently introduced separation of dynamic and nondynamic correlation scheme24,46 that we have lately extended to separate the correlation in Coulomb holes.18 In we can see that the short-range part of the Coulomb hole corresponds mostly to dynamic correlation and that the shoulder structure is also present in this part of the Coulomb hole.
### Second-row Atoms and Molecules
In this section we investigate whether the shoulder is a feature of the Coulomb hole of Ne or other second-row atoms also show a signature of core-electron correlation at short interelectronic distances of their holes. From previous studies,24,36 it is known that the Coulomb holes of He and Li do not present such a shoulder. For the rest of second-row atoms in their ground states, a shoulder or a minimum is always obtained as we show in (the complete holes can be found in Figure S10). In we plot the Coulomb hole divided by the square of the atomic charge (\(Z^{2}\)) in order to make all the holes fit in the same scale. The Coulomb holes reported for the open-shell systems were obtained using an unrestricted formalism (i.e. they correspond to the difference between the UCISD and the UHF radial intracule densities). Our study reveals that for Be, B, and F atoms, a minimum of the Coulomb hole is observed, while for C, N, and O atoms, a shoulder is produced. The shoulder or minimum vanish when they are calculated employing a fc-CISD wave function (see Figure S11), proving that the features observed correspond to correlation effects of the core electrons. The analysis of the OEP reveals that in all cases the radius of the \(K\) shell does not change upon inclusion of electron correlation effects, and only an internal small reorganization within the \(K\) shell is produced (see Table 54 for more details).
## 4 Conclusions
We have analyzed the Coulomb hole of Ne from highly-accurate CISD wave functions. Our energy estimates have been obtained from a two-fold extrapolation of optimized even-tempered basis sets and compare well with the best estimates available in the literature (we recover 97 % of the correlation energy of Ne).
The dynamic part (\(h_{0}\)) and the total (\(h_{s}\)) Coulomb hole of Ne at the CISO/165PDFG level of theory.
Zoom of the short-range CISO/6-311G\({}^{\circ}\) Coulomb holes of the second-row atoms.
Double excitations from the core orbital give rise to the most important configurations in the CISD expansion that contribute to the shoulder. The shoulder is due to an internal reorganization of the \(K\) shell, where electrons are pushed towards the \(K\)-shell boundary. The correlation nature of the shoulder is dynamic, as one would expect. This feature is very sensitive to the basis set in the core region. Finally, we have proven that for the rest of second-row atoms, except Li, a shoulder or a maximum in the short-range region of the Coulomb hole is obtained, which is due to the correlation of the core electrons in the \(K\) shell. In all cases, the shoulder or the minimum corresponds to an internal reorganization of the \(K\) shell.
|
10.48550/arXiv.1712.01789
|
The Coulomb Hole of the Ne atom
|
Mauricio Rodríguez-Mayorga, Eloy Ramos-Cordoba, Xabier Lopez, Miquel Solà, Jesus M. Ugalde, Eduard Matito
| 3,649
|
10.48550_arXiv.2410.19445
|
## 1 Introduction
As an active reagent, water plays a crucial role in regulating biomolecular functions, for instance, to stabilize their structures and facilitate dynamic interactions in proteins and nucleic acids. Water also exhibits anomalous characteristics for nanomaterials, such as unique structural arrangements and fast diffusion rates within confined nanotubes compared to the bulk phase. The polarization effect, which is the electron rearrangement in response to different local environments, has been shown to profoundly influence both static and dynamic properties of water. For example, the induced polarization by electric field is considered to be the major source of the nonadditive cooperativity in water hydrogen bonding interactions. Moreover, the considerable electron transfer between water molecules enhances the proton donor-acceptor orbital interactions and hence the covalency, directionality as well as relevant spectroscopic properties of water hydrogen bonds. In recent years, microdroplets have been shown to accelerate chemical reactions by up to 10 million times compared to bulk solutions, widely attributed to the strong electric fields formed on their surfaces. One promising origin of the interfacial electric fields is the large number of H\({}_{2}\)O\({}^{\delta+}\)/H\({}_{2}\)O\({}^{\delta-}\) radical pairs on the microdroplet surface created by the intermolecular charge transfer. A recent second-order Moller-Plesset perturbation theory (MP2) based molecular dynamics (MD) study indicates that substantial and inhomogeneous interfacial water charge transfer dramatically promotes the reactions of Criegee intermediates with water.
Atomic partial charge is a qualitative representation of the environment-dependent electron distribution for understanding polarization effects and rationalizing chemical phenomena. Nevertheless, partial charge is neither experimentally nor theoretically observable. In quantum chemistry, atomic charges vary significantly with different electron density partition methods, such as Mulliken, Lowdin, natural population analysis (NPA), quantum theory of atoms in molecules (QTAIM), Hirshfeld, etc. Notably, the Hirshfeld population analysis precisely describes the density redistribution during bonding. Furthermore, the charge model 5 (CM5) corrects the underestimated Hirshfeld charges to achieve accurate dipole moments, which has been shown to provide accurate electronic distribution and to be insensitive to the choice of basis sets and theories. Calculating accurate electron densities to produce partial charges is computationally demanding using density functional theory (DFT) or higher level correlated wavefunction methods (e.g. MP2 and coupled cluster singles and doubles (CCSD)), making it cost-prohibitive for large molecular systems. In the past decades, efforts have been made to develop machine learning (ML) charges by mapping the atomistic local environment descriptors onto QM reference atomic charges. However, these charge models were mainly tested on single molecules and small clusters, casting doubt on their accuracy when applied to larger molecular systems. A recent study reports a kernel ridge regression (KRR) model accurately predicts iterative Hirshfeld atomic charges of water molecules in bulk liquid represented by QM/molecular mechanical (QM/MM) water clusters. Remarkably, the hydrogen-bond stretch infrared (IR) peak is successfully retrieved using dynamic ML charges, whereas this weak peak disappears for fixed charges, marking the importance of polarization and intermolecular charge transfer in simulating liquid water. However, this model has several limitations. Firstly, the training and testing were conducted with an identical system (23 QM water molecules with 1977 MM water molecules), leaving the model's generalizability to larger QM regions uncertain. Secondly, the charge model was specifically tailored for water in bulk solution, which makes it inadequate for predicting charges for interfacial water molecules that experience more complex local environments. Finally, while ML charges were used to compute dipole moments and IR spectra using trajectories prepared with TIP4P/2005, they have not been integrated into a water model for simulating polarization in molecular dynamics.
Simulating water has been challenging for decades. Since the introduction of the rigid non-polarizable water models with fixed charges, they have been substantially improved by re fining parameters, introducing intramolecular flexibility, and integrating implicit polarizability. Moreover, tremendous efforts have been made to recover the explicit polarization effects. Some polarizable models incorporate induced multipoles in response to the local electric fields using atomic polarizability tensors or Drude oscillators. In contrast, the charge-flow models enable geometry-dependent charge redistribution. Although charge-flow polarizable models appear enticing due to their explicit incorporation of charge exchange between atoms, they suffer from several deficiencies, including high computational costs associated with iteratively updating charges, non-linear computational scaling of the polarizability with system sizes, inadequate description of intermolecular charge transfer, and a lack of out-of-plane polarization. Furthermore, despite extensive fine-tuning of the charge functional parameters to fit the DFT electrostatic potentials, the absence of spontaneous quantum mechanical information during the solution of dynamical charges may result in questionable electron density arrangements. Beyond the empirical force fields, energy potentials from first-principles for large water systems can be built by employing many-body expansion (MBE) for potentials of small water fragments obtained through fitting with functionals or machine learning. Nevertheless, constructing an accurate potential requires a tremendous amount of training data to cover the vast chemical space. The applicability of potentials is severely limited due to their system-specific nature that does not generalize well to unseen molecules. Furthermore, important electronic properties such as electron density and multipole moments are absent in the potentials. Notably, the QM-polarizable force fields that incorporate instantaneous QM-level polarizability enabled by machine-learning, such as FFLUX using machine learned QTAIM atomic multipoles, have been successfully applied to various chemical systems.
In this study, we present a novel QM-polarizable water model that incorporates dynamical atomic charges and charge transfer exclusively trained on QM data acquired at the _ab initio_ MP2 level of theory. The charges are produced using a deep neural network charge model called ChargeNN. A set of Interaction Classified Functions (ICFs) was employed as the ChargeNN features, explicitly capturing the types of atomic interactions to accurately depict local environments.
Trained on diverse (H\({}_{2}\)O)\({}_{25}\) structures sampled from MD trajectories, the ChargeNN is capable of accurately predicting charges for much larger water clusters, for example, (H\({}_{2}\)O)\({}_{190}\), as compared to computed MP2 charges. The ChargeNN has also been integrated into a water model for dynamical polarization simulations. The intermolecular charge transfer has been explicitly accounted for using machine-learned pair-wise water charge transfer, which effectively captures the out-of-plane polarization. We have implemented analytical energy gradients in the periodic boundary condition (PBC) within the shared-memory message passing interface (MPI) framework for efficient MD simulations. Our results validate the capability of ChargeNN-derived water model to reproduce a variety of water properties in excellent agreement with experimental data, for instance, the model generalizes well to a few thousand water molecules in unit cell for predicting the liquid water properties. Finally, the statistical analysis of liquid water and large water microdroplets using ChargeNN reveals that the partial breakage of hydrogen-bond networks, along with water charge transfer, concurrently promotes the layer electric fields at the air/water interface. The ChargeNN water model is thus of broad applicability in studying condensed ice, liquid water, nano-phase, large microdroplets, and solvent interactions with materials and biomolecules.
## 2 Methods
### Charge Model 5
Charge Model 5 (CM5) based on Hirshfeld population analysis (HPA) was implemented for third-order many-body-expansion orbital-specific-virtual MP2 (MBE-OSV-MP2) to investigate the water polarization and charge transfer in response to the variation in the local environment.
\[q_{i}^{\rm HPA}=Z_{i}-\int\mathrm{d}\mathbf{r}\frac{\rho_{i}^{0}(\mathbf{r})}{ \sum_{j}\rho_{j}^{0}(\mathbf{r})}\rho(\mathbf{r}), \tag{1}\]where \(Z_{i}\) is the nuclear charge. \(\rho^{0}(\mathbf{r})\) and \(\rho(\mathbf{r})\) denote the electron densities of the promolecule and real molecule at position \(\mathbf{r}\), respectively. In contrast to the hard-boundary partitioning approaches like QTAIM, where electron density at each grid point is attributed to a single atom, the boundary-less Hirshfeld population analysis allows electron density to contribute to multiple atoms. It provides a clear partitioning of the electron density and accurately describes the density redistribution from promolecule to real molecule during bonding. Nevertheless, Hirshfeld charges are significantly underestimated due to the Hirshfeld weighting factor, which causes the atomic population to closely resemble that of the isolated atom.
By introducing a unified set of parameters optimized by fitting to the experimental or DFT dipole moments of 614 molecules, the CM5 model maps the underestimated Hirshfeld charges onto a new set of charges, providing a more accurate description of the electrostatic potential:
\[\begin{split} q_{i}^{\text{CM5}}&=q_{i}^{\text{HPA} }+\sum_{j\neq i}A_{ij}B_{ij},\\ B_{ij}&=\text{exp}\left[-\alpha(r_{ij}-R_{Z_{i}}-R_ {Z_{j}})\right],\end{split} \tag{2}\]
Constant \(R_{Z}\) denotes the atomic covalent radius. CM5 has been shown to deliver precise electronic charge distributions and be insensitive to different theoretical levels and basis sets. As demonstrated in Figure 1**a**, CM5 charges outperform iterative Hirshfeld (Hirshfeld-I), NPA, and QTAIM charges in reproducing the dipole moments of water clusters containing 1-20 molecules. Furthermore, CM5 exhibits numerically consistent dipole moments between cc-pVTZ and aug-cc-pVTZ basis sets, as shown in Figure 1**b**. The CM5 charges in this work were obtained from the MBE-OSV-MP2/cc-pVTZ relaxed density matrix.
### Deep Neural Networks for Atomic Charges
Tremendous efforts have been made to design descriptors of local atomic environments for building machine learning models of energy potentials and atomic properties.
\[\begin{split} G_{i}^{\text{rad}}&=\,\sum_{j}e^{- \eta(r_{ij}-r_{s})^{2}}s_{c}(r_{ij}),\\ G_{i}^{\text{ang}}&=2^{1-\zeta}\sum_{j,k\neq i}\,(1 +\lambda\cos\theta_{ijk})^{\zeta}\cdot e^{-\eta\left(r_{ij}^{2}+r_{ik}^{2}+r_{ jk}^{2}\right)}\\ &\quad\cdot s_{c}\left(r_{ij}\right)s_{c}\left(r_{ik}\right)s_{c} \left(r_{jk}\right),\end{split} \tag{3}\]
\(s_{c}\) denotes the cutoff function that enforces a smooth transition of atoms \(j\) entering and exiting the selection region for the atom
**Water dipole moments derived from charges (a)** Water dipole moments obtained from CM5, Hirshfeld-I, NPA and QTAIM charges, using electron density of MP2/aug-cc-pVTZ computed with Gaussian 16. CM5 and NPA charges were obtained with Gaussian 16, while Multiwfn was used to compute Hirshfeld-I and QTAIM charges. (b) Water dipole moments obtained from CM5 charges, computed with MBE-OSV-MP2 using aug-cc-pVTZ and cc-pVTZ basis sets. Experimental structure of water monomer and structures from Cambridge water clusters were used for this test. (H\({}_{2}\)O)\({}_{6\text{c}}\) and (H\({}_{2}\)O)\({}_{6\text{p}}\) refer to cage and prism conformers of water hexamers, respectively. Experimental dipole moment of the water monomer is from ref.126.
\[s_{c}(r_{ij})=\begin{cases}0.5\left[\text{cos}\left(\frac{\pi r_{ij}}{r_{c}}\right) +1\right],&r_{ij}\leq r_{c}\\ 0,&r_{ij}>r_{c}\end{cases}. \tag{4}\]
However, ACSF descriptors omit the dependence of the atomic interactions on element type, which can significantly impact training performance and data efficiency.
To incorporate the information of atomic interaction types, we propose an interaction classified function (ICF) to describe the interaction-specific (\(\alpha\beta\)) contributions to the charge of atom \(i\) of the atom type \(\alpha\), which satisfies translational, rotational and permutational invariance but requires no angular ACSF parameterization for angularly independent atomic charges:
\[f_{i}^{\alpha\beta}=\sum_{j\in\beta}e^{-\frac{r_{ij}^{2}}{k_{f}}}s_{c}(r_{ij}), \text{ where }i\in\alpha. \tag{5}\]
In the above formula, a single parameter \(k_{f}\) controls the interaction decay with respect to the atomic distance, effectively alleviating the heavy process of the parameter tuning. Despite their simplicity, the radial ICFs have been found to sufficiently describe local interactions needed for reproducing accurate atomic charges with excellent data efficiency (see section 3.1). To determine the atom selection cutoff (\(r_{c}\)) in eq. 4, we tested the charge convergence for both bulk water and air/water interface, and found that 24 surrounding water molecules sufficiently form the local environment to produce charges comparable to those surrounded by 127 molecules, as demonstrated in Figure S1. The selected cutoff \(r_{c}\) was accordingly set to be 4.4 A, which is the minimum radius of multiple (H\({}_{2}\)O)\({}_{25}\) droplets extracted from well-equilibrated liquid water boxes.
The ICF feature set for the charge of atom \(i\) in the molecule \(I\) reads
\[\mathbf{F}_{i_{I}}=\{f_{i_{I}}^{\alpha_{I}\beta_{I}},f_{i_{I}}^{\alpha_{I} \beta_{J}},f_{j_{I}}^{\alpha_{I}\beta_{J}}\}. \tag{6}\]
In addition, the intermolecular ICFs of other atoms in the same molecule (\(f_{j_{I}}^{\alpha_{I}\beta_{J}}\)) are also appended to the feature set to account for the perturbation arising from the interactions for the neighboring atoms, which notably improves the prediction accuracy. Three decay parameters \(k_{f}\) were employed for each ICF in eq. 6 to simulate varying interactions with the atomic distance, as illustrated in Figure S2. To better handle the distinct charge ranges and distributions, two separate neural networks were specifically trained for the charges of oxygen and hydrogen.
Although the electrostatic interaction, modeled by Coulomb potential with flexible atomic charges, implicitly accounts for environment-dependent charge flows, it is substantially underestimated in the absence of the intermolecular charge transfer (CT) contribution. For recovering the missing interactions, an extra neural network is used to predict intermolecular CT between water pairs (\(IJ\)) by supplying a feature set composed of intermolecular ICFs:
\[\begin{split}\mathbf{F}_{IJ}&=\{f^{\alpha_{I}\beta_ {J}}\},\\ f^{\alpha_{I}\beta_{J}}&=\sum_{\begin{subarray}{c}i \in\alpha_{I}\\ j\in\beta_{J}\end{subarray}}e^{-\frac{r_{ij}^{2}}{k_{f}}}s_{c}^{\text{CT}}(r_{ ij}).\end{split} \tag{7}\]
We examined the decay of the CT with the atomic distance, as the decay of the CT energy in eq. 9 specifically depends on CT. Figure S3 shows that CT deteriorates to only \(\sim\)\(10^{-5}\) e at the hydrogen bond distance of 5.5 A for a water dimer. Considering that CT would also be hindered by the water molecules filled between distant water pairs, 5.5 A is sufficiently large for the CT cutoff.
The CM5 atomic charges of (H\({}_{2}\)O)\({}_{25}\) were utilized to train the neural networks. Reference charges were computed using the low-scaling MBE-OSV-MP2 method with the cc-pVTZ basis set, as MP2 has been shown to accurately predict the energetic and electronic properties of water. To benchmark the charge model, charges were prepared for 20000 (H\({}_{2}\)O)\({}_{25}\), 200 (H\({}_{2}\)O)\({}_{32}\), 50 (H\({}_{2}\)O)\({}_{64}\), 10 (H\({}_{2}\)O)\({}_{64}\) and 1 (H\({}_{2}\)O)\({}_{190}\) cluster conformations that were randomly sampled from the equilibrated MD trajectories using SWM4-NDP with harmonic bonds and angles on OpenMM. The \(NVT\) trajectories were simulated in the gas phase at 298.15 K, with a time step of 1 femtosecond.
Four hidden layers are employed to effectively capture intricate relationships between featuresand atomic charges, which also mitigates overfitting issues. The Swish function is adopted as the activation function:
\[A(x)=\frac{x}{1+e^{-\beta x}}. \tag{8}\]
The Swish function is differentiable and ensures smooth charge functions for new molecular configuration, which avoids discontinuity on the potential energy surface. Following multiple trials, we discovered that utilizing \(\beta=3\) allows \(A(x)\) to retain good linearity around \(x=0\), enhancing the learning performance while circumventing erratic charge functions caused by overfitting. The mean squared error (MSE) was chosen as the loss function. Among the 20,000 configurations of (H\({}_{2}\)O)\({}_{25}\), 8,000 were randomly selected for training, with 50 used as the validation data set. The remaining data points were used for evaluating the prediction performance. After extensive testing, we applied the "early stopping" technique with 200 epochs to ensure satisfactory loss convergence and to prevent overfitting.
### Water model with ChargeNN charges
The potential energy functional of the 3-site flexible water model incorporating ChargeNN charges and charge transfer contributions is expressed as a sum of the intramolecular energy (\(U^{\rm intra}\)) and the intermolecular energy (\(U^{\rm inter}\)),
\[U^{\rm intra} =\sum_{I}\left[\sum_{i}\frac{k_{b}}{2}(r_{{\rm O}_{I}{\rm H}_{I} }-r_{\text{OH}}^{0})^{2}+\frac{k_{a}}{2}(\theta_{I}-\theta^{0})^{2}\right],\] \[U^{\rm inter} =\sum_{i<j}\varepsilon_{ij}\left[7\left(\frac{\sigma_{ij}}{r_{ij }}\right)^{9}-5\left(\frac{\sigma_{ij}}{r_{ij}}\right)^{6}\right] \tag{9}\] \[+\sum_{i<j}k_{e}\frac{q_{i}q_{j}}{r_{ij}}-\sum_{I<J}k_{c}\frac{ \delta q_{IJ}^{2}}{\delta q_{IJ}^{2}+3(\delta q^{0})^{2}}.\]
In the formula for intramolecular energy, \(k_{b}\) and \(k_{a}\) represent the coefficients of bond stretching and bending, respectively. \(r_{{\rm O}_{I}{\rm H}_{I}}\) denotes the O-H bond length, while \(\theta_{I}\) refers to the H-O-H angle of the molecule \(I\), with \(r_{\text{OH}}^{0}\) and \(\theta^{0}\) being the equilibrium values. For the intermolecular energy, \(\epsilon_{ij}\)and \(\sigma_{ij}\) are standard 12-6 Lennard-Jones (LJ) parameters. \(k_{e}\) and \(k_{c}\) pertain to coulomb constant and charge-transfer constant, respectively. \(q_{i}\) denotes the charge of atom \(i\). \(\delta q_{IJ}\) represents the charge transferred between molecules \(I\) and \(J\), with the parameter \(\delta q^{0}\) as the reference CT that yields the strongest CT force.
We observed that atomic charges exhibit significant sensitivity to stretching bonds and bending angles, necessitating the addition of flexibility in the intramolecular forces. Moreover, this flexibility has been reported to improve predicted water properties, and also permits recovering intramolecular vibrational spectra. Consequently, we adopted harmonic bonds and angles for better description of the dynamical behavior of the water molecules. The equilibrium bond length \(r_{\text{OH}}^{0}\) and angle \(\theta^{0}\) were tuned to ensure that the average values in the liquid phase align with experimental measurements at 298.15 K and 1 atm. The coefficients of bond length (\(k_{b}\)) and angle (\(k_{a}\)) were adjusted to produce accurate positions of the bond stretching and bending peaks in the IR spectrum.
The electrostatic energy is calculated using ChargeNN charges. To facilitate simulations under periodic boundary conditions (PBC), Ewald summation is employed, utilizing the minimum image convention to construct the local environment for charge generation. It is important to note that the total charge of a cluster or unit cell could be non-zero due to the accumulative error of atomic charges. To ensure a zero net charge of the whole system, all atomic charges must be subtracted by a small constant (\(\Delta q=\sum_{i}q_{i}/N\), where \(N\) is the number of atoms) of approximately \(\pm\)0.0001 e. Additionally, non-homogeneous correction can be accomplished by considering the weights of atomic charges or integrating an explicit penalizing term to enforce electroneutrality into the loss function. There have been attempts to include charge transfer interactions with empirical charge transfer. For our model that uses CT obtained quantum mechanically, the charge transfer term formulates a smooth ratio function of squared CT. The formula offers several merits, for instance, it is symmetric and the CT energy converges to zero in the absence of CT, preventing abrupt changes in potential energy around the CT cutoff. Additionally, the formula effectively captures the rapid intensification of CT interactions when initial charges are small, as well as the gradual attenuation due to the electron migration from equilibrium. Moreover, since the inter-water CT occurs away from the nuclear sites, the CT term naturally accounts for out-of-plane polarization, resulting in more accurate water conformations.
As the inclusion of the CT term strengthens the short-range interactions, the traditional 12-6 LJ potential for Van der Waals (VDW) forces yielding a steep repulsive wall constrains inter-water distances to a compact range, which leads to an excessively high first peak in the oxygen-oxygen radial distribution function (RDF). Instead, we employ a 9-6 functional form with minor revisions to the original LJ formula, retaining the 12-6 LJ parameters \(\varepsilon\) and \(\sigma\) for compatibility with the existing force fields. In the process of the parameter tuning, the VDW equation was written as \(U_{ij}^{\mathrm{vdw}}=A_{ij}/r_{ij}^{9}-B_{ij}/r_{ij}^{6}\) for convenience. The intermolecular parameters \(k_{c}\), \(k_{d}\), \(A_{ij}\) and \(B_{ij}\) were optimized for the accurate heat of vaporization, density and radial distribution functions (RDF) for liquid water at the ambient conditions. All parameters can be found in Table S1.
### Implementation
**Implementation flowchart.** Schematic illustration for the implementation of the polarizable water model using the charges produced by the ChargeNN model.
In each MD step, atomic coordinates adjusted by the MD solver are utilized to compute the ML features interaction-classified functions \(\mathbf{F}\) and their analytical derivatives with respect to nuclear positions \(\mathrm{d}\mathbf{F}/\mathrm{d}\mathbf{R}\). The ML features are then fed into the pre-trained neural networks for the predictions of atomic charges and charge transfer. By employing auto-differentiation through back-propagation, the gradients of the charges concerning the features \(\mathrm{d}q/\mathrm{d}\mathbf{F}\) can be conveniently evaluated. Subsequently, the predicted atomic charges and charge transfer are used to determine the potential energy as described in eq. 9, where total gradients are obtained using the chain rule:
\[\frac{\mathrm{d}U}{\mathrm{d}\mathbf{R}}=\frac{\partial\bar{U}}{\partial \mathbf{R}}+\frac{\partial U}{\partial q}\cdot\frac{\mathrm{d}q}{\mathrm{d} \mathbf{F}}\cdot\frac{\mathrm{d}\mathbf{F}}{\mathrm{d}\mathbf{R}}. \tag{10}\]
Finally, the energy and gradients are passed to the MD solver to update the positions.
The ChargeNN program was written in Python with double precision, integrated with Tensorflow for handling neural networks. The algorithm has been parallelized based on Message Passing Interface (MPI) standard of version 3. The shared memory windows built in MPI-3's remote memory access module enable significantly lower-latency data communications compared to traditional point-to-point communications.
Results and discussion
### Performance of the charge model
The ChargeNN model demonstrates exceptional data efficiency, requiring only 8000 training (H\({}_{2}\)O)\({}_{25}\) clusters to achieve converged mean absolute errors (MAEs) for the remaining testing clusters, as shown in Figure S4. According to Figure 3, when trained on 8000 (H\({}_{2}\)O)\({}_{25}\), the charge model yields testing MAEs of 0.0025 e for hydrogen and 0.0039 e for oxygen within the same system. These values are markedly lower than the respective MAEs of 0.0036 e and 0.0045 e for (H\({}_{2}\)O)\({}_{23}\) predicted by a recent KRR water charge model trained with the same number of data points. In addition, our model shows exceptionally high prediction reliability with the coefficients of determination (\(R^{2}\)) of 0.96 for hydrogen and 0.93 for oxygen, outperforming the 0.88 reported for the KRR model. Notably, shows that the ChargeNN model considerably improves the transferability, only with a slow deterioration in accuracy as the water cluster size increases.
**Benchmark results for the ChargeNN model. Comparisons of ChargeNN charges and MP2 charges for a. hydrogen and b. oxygen atoms of testing water clusters containing 25, 32, 64, 128 and 190 molecules. The charge models were trained with (H\({}_{2}\)O)\({}_{25}\). The MAEs in elementary charge (e) and coefficients of determination (\(R^{2}\)) are given in the right brackets.**
(H\({}_{2}\)O)\({}_{190}\), yielding MAEs of 0.0028 e for hydrogen and 0.0042 e for oxygen, both with \(R^{2}\) values above 0.9. Figure S5 shows that predicting molecular charges by summing the ChargeNN atomic charges within each molecule is less satisfactory, with \(R^{2}>0.92\) and marginally larger MAEs of \(\sim\)0.007 e due to the accumulative errors. This error can be improved in future by imposing constraints on the loss function to penalize the predicted molecular charges. For assessing ChargeNN's performance in predicting charge transfer, we prepared charge transfer of 40000 water pairs extracted from the MD trajectory of (H\({}_{2}\)O)\({}_{25}\). The ChargeNN trained on 20000 water pairs achieves an MAE of only 0.001 e and \(R^{2}\) of 1.00 for the remaining testing pairs.
### Performance of the water model
To assess the performance of the ChargeNN water model, we calculated various properties of
\begin{table}
\begin{tabular}{l l l l l l} \hline \hline & Units & ChargeNN & SPC/FW\({}^{c}\) & SWM4-NDP\({}^{d}\) & Expt.\({}^{e}\) \\ \hline
## Monomer
& & & & & \\ \(\mu_{\rm total}\) & Debye & 1.79 & 2.19 & 1.85 & 1.85 \\
## Dimer
& & & & & \\ \(\mu_{\rm total}\) & Debye & 2.397 & 3.594 & 2.062 & 2.643 \\ \(r_{\rm OO}\) & Angstrom & 2.81 & 2.73 & 2.83 & 2.98 \\ \(\theta_{A}\) & degree & 65 & 22 & 71 & 58 \\ \(E_{\rm int}\) & kcal/mol & -4.92 & -7.14 & -5.15 & -5.4 \\
## Liquid
& & & & & \\ \(\langle r_{\rm OO}\rangle\) & Angstrom & 0.97 & 1.03 & 0.96 & 0.97 \\ \(\langle\angle_{\rm HOH}\rangle\) & degree & 106.1 & 107.7 & 104.52 & 106.5 \\ \(\rho\) & g/cm\({}^{3}\) & 0.996 & 1.012 & 0.998 & 0.997 \\ \(E_{\rm int}\) & kcal/mol & -9.917 & -11.926 & -9.927 & -9.92 \\ \(H_{\rm vap}\) & kcal/mol & 10.50 & 10.72 & 10.52 & 10.52 \\ \(D\) & 10\({}^{-5}\) cm\({}^{2}\)s\({}^{-1}\) & 2.08 & 2.32 & 2.33 & 2.3 \\ \(\langle\mu_{\rm mol}\rangle\) & Debye & 1.76\({}^{a}\) & 2.39 & 2.46 & 2.9 \\ \(\epsilon\) & & 50.9\({}^{b}\) & 79.6 & 79.0 & 78.4 \\
## Ice
& & & & & \\ \(T_{\rm melt}\) & K & 288 & 190 & 185 & 273 \\ \hline \hline \end{tabular}
* The origin was set to be the center of the nuclear charge of each molecule. \({}^{b}\)Molecules were wrapped to the unit cell. \({}^{c}\)The gas-phase properties, interaction energy and melting temperature for SPC/FW were computed in this work, while other data were from ref.61. \({}^{d}\)SWM4-NDP results from ref.79, 80. \({}^{e}\)Experimental results from ref.140–145.
\end{table}
Table 1: **Benchmark results for the ChargeNN water model**, compared to SPC/FW, SWM4-NDP, and experiments.
Computational details are provided in the SI. We compare our results with those of SPC/FW, which is the fixed charge counterpart of the 3-site ChargeNN model, and those with a widely used polarizable model SWM4-NDP. As shown in Table 1, the ChargeNN's predictions of water monomer and dimer properties agree well with experimental results, demonstrating its capability to accurately simulate gas-phase water clusters. Figure 4**a** illustrates that the angle \(\theta_{A}\) of the water dimer obtained with ChargeNN (65\({}^{\circ}\)) closely matches the experimental value of 58\({}^{\circ}\), in contrast to the unreasonable 22\({}^{\circ}\) by SPC/FW, highlighting the significant impact of polarization and charge transfer on the equilibrium geometries. Despite of the inclusion of intermolecular charge-transfer interactions in our model, enhanced electrostatics using high-order polarization may further improve the structure predictions. For instance, the O-O distance of the water dimer optimized by the multipolar AMOEBA (2.91 A) aligns more closely with experimental measurement (2.98 A) compared to ChargeNN (2.81 A).
Table 1 also demonstrates that the ChargeNN water model accurately reproduces bulk water properties at room temperature and standard pressure, such as average bond length and angle, density, interaction energy, vaporization heat, and diffusion constants. As shown in Figure 4**b**, ChargeNN produces an RDF for liquid oxygen-oxygen distances in good agreement with the MP2 and early experimental data. Nevertheless, the latest x-ray diffraction experiment has measured a less pronounced first peak of approximately 2.57 than the predicted one of 2.98, indicating the need for further parameter refinement to mitigate the "over-structured" hydrogen-bond network. The predicted infrared spectra in Figure 4**c** were obtained through a Fourier transformation of the autocorrelation function of the total dipole moments, in which the one produced by ChargeNN matches the experimental spectrum closely in both band positions and intensities. Notably, ChargeNN successfully captures the hydrogen-bond stretching peak at \(\sim\)200 cm\({}^{-1}\), which is absent in the spectra by SPC/FW and SWM4-NDP. Our results and a previous study indicate that accurately describing hydrogen-bond stretching requires precise polarization and intermolecular CT.
Nevertheless, we observe that the O-H stretching band at around 3400 cm\({}^{-1}\) is considerably narrower than the reference, which may be improved by replacing the harmonic intramolecular forces with anharmonic forces or reactive force fields.
We computed the molecular dipole moments by setting the origin as the nuclear charge center of each partially charged water molecule, following the reported protocol. While ChargeNN accurately reproduces dipole moments for water monomer and dimer, as well as the liquid IR spectrum, it yields an average molecular dipole moment of approximately 1.76 Debye, which is notably lower than the experimental estimation of 2.9 Debye.
**Assessment of the ChargeNN water model.****(a)** Geometries of water dimers optimized by SPC/FW and ChargeNN, compared to the experimental structure. **(b)** The liquid radial distribution functions of the oxygen-oxygen distance obtained with ChargeNN, MP2 and experiments under ambient conditions. **(c)** Liquid infrared spectra generated by SPC/FW, SWM4-NDP, ChargeNN and experiment under ambient conditions. **(d)** Temperature dependent densities from SPC/FW, SWM4-NDP, ChargeNN and experiment at 1 atm.
A recent comprehensive study discovered that none of the selected charge models (Mulliken, Hirshfeld, QTAIM, RESP, ChelpG, Hirshfeld-I, and NPA) is capable of reproducing both total dipole moments and average molecular dipole moments of water clusters. Similarly, CM5 charges agree well with the reference values for total dipole moments, but fall in short to provide a reasonable average molecular dipole moment. Additionally, our water model underestimates the dielectric constant obtained from the total liquid dipole moments, which is largely due to the non-uniqueness of the dipole moment of an extended system according to the modern theory of polarization.
The accurate characterization across a range of temperatures is important to the performance of a water model. Figure 4**d** demonstrates that ChargeNN successfully predicts the temperature-dependent density of liquid water, with the temperature of maximum density around 280 K, very close to the experimental value of 277 K. In contrast, this prediction remains a challenge for the fixed charge model SPC/FW and other well-tuned 3-site models, emphasizing the substantial improvement resulting from ChargeNN polarization. The polarizable SWM4-NDP also falls short in predicting the density variation with temperature. Additionally, Figure S8 presents that the computed temperature-dependence of vaporization enthalpy by ChargeNN agrees well with the reference. For the simulation of ice, we calculated the orientational tetrahedral order parameters of the equilibrium configurations across various temperatures, starting from a perfect hexagonal (\(I_{h}\)) ice structure. The ChargeNN yields a melting temperature of 288 K, close to the experimental value of 273 K. While this result is significantly better than SPC/FW at 190 K and SWM4-NDP at 185 K, there is still room to fine-tune the parameters to achieve ice properties comparable to TIP4P/Ice.
As demonstrated in Figure 5**a**, ChargeNN exhibits exceptional computational efficiency, making it highly promising for large-scale water simulations. For instance, using a single CPU core for a PBC box containing 10044 water molecules, ChargeNN's force calculation time is only one-eighth of that for the economical polarizable SWM4-NDP and merely five times that of the 3-site fixed charge model SPC-FW, both implemented in OpenMM for CPU platforms. Moreover, while the incorporation of machine-learned charges involves additional steps such as feature preparation, charge prediction, and charge derivatives, these ChargeNN-related steps only account for approximately 40% of the total time. Furthermore, the non-ChargeNN components (equivalent to fixed charge computations) are three times slower than OpenMM's SPC-FW, indicating room for algorithmic improvement. In addition, Figure 5**b** reveals suboptimal parallel efficiency of the ChargeNN program, necessitating further optimization. We intend to transfer the algorithm to C++ for enhanced performance, with plans to adapt it to CUDA or OpenCL for extensive parallelization on graphics processing units (GPUs) in future implementations.
**Timing performance of ChargeNN.****(a)** Elapsed time for force calculations on varying water box sizes using a single CPU core (AMD EPYC 7H12, 2.6GHz), comparing in-house ChargeNN and fixed point charges models with OpenMM’s SPC-FW and SWM4-NDP. (b) Speedup of ChargeNN calculations of periodic (H\({}_{2}\)O)\({}_{10044}\) with respect to the number of the CPU cores.
### Demonstrative application
As an illustrative application, we probe the origin of the emerging on-water reactivity on the air/water interface using ChargeNN. The electronic process for creating the strong electric fields on the water microdroplet surface remains elusive, and has been attributed to the abundance of dangling OH bonds at the interface, the presence of OH\({}^{-}\)/H\({}^{+}\) due to proton transfer and electron transfer. To understand the source of the interfacial electric fields, we carried out statistical analysis to identify differences in geometries and charge distributions between the droplet and bulk liquid.
**Statistical results for the illustrative application with a large droplet.** (**a**) Geometry of the studied droplet colored according to molecular charges. (**b**) Correlation between VCD and PHA as a function of the layer distance to the droplet center. (**c**) Charge distributions of water molecules (**d**) Volume charge density using net molecular charges. The center of mass of each water molecule was used to compute the layer distance to the center.
For the liquid simulation, we analyzed the last 500 snapshots from a 1-ns MD trajectory of a large unit cell containing 2123 water molecules.
Our results reveal that the primary factor influencing the layer charge density is the hydrogen-to-oxygen ratio. The volume charge density (VCD) has been extensively used as an indicator for the strength of the layer electric fields, and is highly correlated with the proportion of the hydrogen atoms (PHA) (see Figure 6**b**). For deep layers inside the droplet at distances less than 37 A from the center, the PHA fluctuates around an equilibrium percentage of 67% indicating an homogeneous distribution, with the VCD oscillating near the liquid VCD of 0 e/nm\({}^{3}\). The PHA on the intermediate layer at the distance of 41 A from the center drops significantly to 66.2%, and coincides with the most negatively charged layer with a VCD of -0.35 e/nm\({}^{3}\). Beyond this layer, the PHA grows continuously to break the theoretical limit of 66.7% (that is, 2 H atoms and 1 O atom per water) and even exceed 68% at the air/water interface, revealing an inhomogeneous distribution of H-bond network near the surface, where the VCD increases consistently and reaches a maximum of 0.18 e/nm\({}^{3}\) at 45 A. Our results align with the "dangling OH theory" that more OH bonds dangle on the droplet surface due to the partial collapse of the hydrogen-bond network near the surface, leading to a higher proportion of hydrogen atoms and a positively charged surface layer. Consequently, the inward layer near the interface is abundant in oxygen atoms, and overall negatively charged.
In addition, the considerable charge transfer also significantly contributes to the layer charge density. As shown in Figure 6**c**, in the inner shells at distances shorter than 37 A from the center, the molecular charges are closely clustered around zero, similar to the distribution in the liquid phase. In contrast, the vast interfacial charge separation results in a substantially increasing population of the charged water molecules. For understanding the explicit CT effect on the charge density, we obtained the water VCDs using net molecular charges of the water molecules whose centers of mass reside within the layers, as demonstrated in Figure 6**d**. The interior of the droplet is overall negatively charged with a water VCD of \(\sim\)-0.015 e/nm\({}^{3}\), indicating electron migration from the droplet surface to the inter layers. This is drastically different from the uniform charge density distribution in liquid phase. Furthermore, the water VCD of the positive droplet surface is approximately 0.045 e/nm\({}^{3}\), constituting 25% of the total VCD at the air/water interface.
## 4 Application Outlook
Beyond the simulation of water, we can envisage broader applications with ChargeNN. Firstly, the ChargeNN water model can be seamlessly combined with the existing force fields to simulate interactions between proteins and water solvents. Such interactions markedly affect the structures and dynamic behavior of proteins. Additionally, integrating chargeNN water model with QM electron structures via a QM/MM interface will allow accurate descriptions of outer solvent shells for reactive centers. Moreover, incorporating empirical valence bond (EVB) and reactive force fields (ReaxFF) into ChargeNN will enable simulations of bond breaking and formation, facilitating efficient modeling of proton transfers through hydrogen-bond networks and chemical reactions. Furthermore, accurate simulation of protein-ligand electrostatic interactions requires incorporating electronic polarizability, which induces screening effects that weaken electrostatics in the buried environment. A recent study also shows that QM electronic polarization is essential for accurately producing spectral densities for proteins. Therefore, we aim to extend the ChargeNN to other atoms and molecular systems, enabling generalizable predictions of atomic charges across diverse local environments and achieving QM-level charge assignments for complex macromolecules. Combining accurate charges and force fields will allow for the depiction of instantaneous polarization and transient charge transfer in molecular dynamics. Coupled with well-designed energy functionals, fast prediction of _ab initio_ polarization from neural networks would significantly enhance the accuracy of atomistic simulations for large chemical systems other than water.
Conclusions
Modeling water has been pivotal task in theoretical chemistry, aiming at understanding its unique thermodynamic and electronic characteristics and propensities. Nevertheless, simulating polarization remains a challenge due to the volatile nature of electron distribution and its high sensitivity to local electric fields, which necessitates quantum mechanical treatments for accurate descriptions. Despite significant advancements in low-scaling QM electronic structures and advanced computing technology, QM calculations for large water systems remain prohibitively expensive. Machine learning offers promise for overcoming the cost obstacles and predicting atomic charges with QM accuracy. However, the existing ML charge models are limited to static calculations of specific systems, incapable of simulating polarization in both liquid and gas phase water during the structural and dynamical evolution.
In this work, we address the aforementioned problems by introducing a dynamic and polarizable water model that leverages machined-learned MP2-level partial charges. The charge model, termed ChargeNN, employs deep neural networks to map the interaction classified functions that accurately characterize the local environment onto the CM5 charges computed with MP2 electron density, demonstrating high accuracy and generalizability. Equipped with quantum atomic charges and charge transfer predicted by the neural networks, the ChargeNN water model successfully reproduces a variety of water properties across different temperatures and phases in excellent agreement with the experimental measurements, validating its capacity to generate electronic and thermodynamic quantities via fast and long molecular dynamics.
Additionally, we probed the origin of the strong local electric fields on the droplet surface by conducting molecular dynamics of both liquid water and a large droplet using the ChargeNN water model. The findings clearly indicate that the layer electric fields on the droplet surface are primarily induced by breakage of the hydrogen-bond network, which leads to distinct proportion of hydrogen atoms from the liquid phase. Furthermore, surface-to-interior charge transfer also considerably contributes to the layer electron density.
## 3 MBE-OSV-MP2 electronic structure
The atomic charges were calculated using the MBE-OSV-MP2 method. The OSV approximation has demonstrated a significant reduction in the computational effort required for correlated methods without considerable loss in accuracy. The OSVs \((\mathbf{Q}_{k})\) can be obtained through a simple singular value decomposition of the diagonal pair MP2 amplitudes \(\mathbf{T}_{kk}\):
\[[\mathbf{Q}_{k}^{\dagger}\mathbf{T}_{kk}\mathbf{Q}_{k}]_{\overline{\mu}_{k} \overline{\nu}_{k}}=\omega_{\overline{\mu}_{k}}\delta_{\overline{\mu}\overline {\nu}}. \tag{11}\]
The inherent sparsity allows the OSV space to be reduced by setting a cutoff for the singular values \(\omega_{\overline{\mu}_{k}}\), which indicates the significance of OSVs. To further address the computational scaling challenge, a many-body expansion of OSV-MP2 amplitudes and density matrices with orbital-specific partitioning was introduced:
\[\mathbf{T}_{(ii,ii)}=\mathbf{T}_{(ii,ii)}^{i}+\sum_{k}\Delta\mathbf{T}_{(ii, ii)}^{i,k}+\sum_{k>l}\Delta\mathbf{T}_{(ii,ii)}^{i,k,l}. \tag{12}\]
The population analysis can be performed using the total relaxed one particle density matrix, which can be computed as the sum of the density matrices of Hartree Fock, relaxed MP2 and molecular orbital (MO) relaxation:
\[\mathbf{P}=\mathbf{P}^{\mathrm{HF}}+\mathbf{P}^{\mathrm{MP2}}+\mathbf{P}^{ \mathrm{MO}}. \tag{13}\]
The relaxed OSV-MP2 density matrix in MO basis can be obtained with
\[\mathbf{P}^{\mathrm{MP2}}=\sum_{ij}\mathbf{T}_{ii}\left\langle\mathbf{X}_{ij}^ {\mathrm{T}}+\mathbf{X}_{ji}^{\perp}\right\rangle, \tag{14}\]
Additionally, we developed a large-scale parallelism strategy utilizing efficient memory management and data communication, which enables computations of MP2-level charges for large molecular systems.
The authors acknowledge financial supports from the Hong Kong Research Grant Council, Hong Kong Quantum AI Lab through program of Hong Kong Government, and the Hung Hing Ying Physical Sciences Research Fund of the University of Hong Kong. Q.L. acknowledges Dr. Xinyan Wang and Dr. Ruiyi Zhou for valuable discussions.
The Supporting Information is available free of charge. The SI document comprises the results of the convergence of atomic charges and charge transfer with respect to atomic distances, testing MAEs relative to the training set size, parity graphs of molecular charges and charge transfer, optimized parameters for the ChargeNN water model, computational details for benchmarking ChargeNN, and some supplementary benchmark results.
|
10.48550/arXiv.2410.19445
|
Polarizable Water Model with Ab Initio Neural Network Dynamic Charges and Spontaneous Charge Transfer
|
Qiujiang Liang, Jun Yang
| 239
|
10.48550_arXiv.1712.05991
|
###### Abstract
We measure the electron spin resonance spectrum of the endohedral fullerene molecule at pressures ranging from atmospheric pressure to 0.25 GPa, and find that the hyperfine coupling increases linearly with pressure. We present a model based on van der Waals interactions, which accounts for this increase via compression of the fullerene cage and consequent admixture of orbitals with a larger hyperfine coupling. Combining this model with theoretical estimates of the bulk modulus, we predict the pressure shift and compare it to our experimental results, finding fair agreement given the spread in estimates of the bulk modulus. The spin resonance linewidth is also found to depend on pressure. This is explained by considering the pressure-dependent viscosity of the solvent, which modifies the effect of dipolar coupling between spins within fullerene clusters.
## I Introduction
In the central-field approximation, atomic nitrogen has no hyperfine coupling because the combined spin density of the three unpaired \(p\) orbitals vanishes at the nucleus and is spherically symmetric outside it. In reality, interactions beyond this approximation lead to a non-zero hyperfine coupling for the free atom. In the presence of other atoms or molecules, distortion of the electron orbitals further modifies the hyperfine coupling, which therefore offers an insight into the nature of the interatomic interactions.
The endohedral fullerene offers a nearly unique system in which the fullerene cage stabilises the encapsulated nitrogen such that it behaves like a free atom. However, the hyperfine coupling of is enhanced by approximately 50% relative to the free atomic value, which reflects the effect of confinement of the nitrogen orbitals by the cage. This increase has previously been attributed to nitrogen-cage interactions that mix in excited states with larger hyperfine couplings, although without proposing a microscopic model.
Understanding the hyperfine coupling of this molecule is important due to its potential use as a molecular spin qubit or frequency reference in an atomic clock. In particular, is suitable as a frequency reference due to its sharp resonances and the existence of a clock transition in its low-field spectrum. Since the frequency of this clock transition depends solely on the isotropic hyperfine coupling constant \(A\), it is important to characterise the mechanisms that affect it. For example, the hyperfine coupling is known to depend on temperature, which could reduce the long term stability of a fullerene clock against environmental temperature fluctuations.
Here, we measure the hyperfine coupling strength and the electron spin resonance (ESR) linewidth over a pressure range up to 0.25 GPa. The hyperfine coupling increases linearly with pressure, which we explain using a microscopic model of electron wavefunction distortion mediated by van der Waals interactions between the nitrogen atom and the cage. Using this model and theoretical estimates of the bulk modulus of isolated fullerene molecules, we then estimate the pressure shift and compare it with our experimental results. The predicted shift is smaller than the observed shift, which may indicate contributions beyond the van der Waals interaction as well as uncertainty in the the theoretical bulk modulus. To the best of our knowledge, the value deduced from our model is the first experimental estimate of bulk modulus for an individual C\({}_{60}\) molecule under conditions of hydrostatic pressure.
We find that the measured linewidth increases non-linearly with pressure. This is explained by the pressure-dependent viscosity of the solvent and its effect on rotational diffusion of fullerene clusters. At low pressure, the solvent is sufficiently non-viscous that dipole-dipole coupling between spins is averaged out by rotation of the cluster. At high pressure, the solvent viscosity increases, leading to an increase in the linewidth towards the rigid lattice limit.
## II Hyperfine coupling
### Experiment
A sample was prepared by ion implantation, dissolved in toluene, and purified by high-performance liquid chromatography to a purity of \(\sim\)0.6 %. The sample used for high-pressure measurements was then concentrated by bubbling nitrogen gas through it to evaporate some solvent. This is because the small sample volume of the pressure cell leads to a low resonator filling factor, which reduces the signal-to-noise ratio (SNR) of the measurement. This sample was then injected into an yttria-stabilized zirconia cell attached to a barocyler and loaded into a dielectric resonator (Bruker ER4123D). Measurements were performed at the \(X\) band (frequency \(f\sim 9.5\) GHz) using a Bruker EMXmicro spectrometer. All measurements were performed atroom temperature, with sufficient time between increasing the pressure and performing the measurement to allow thermal equilibration.
The energy levels of the spin system are described by the Hamiltonian
\[\mathcal{H}=g_{\mathrm{e}}\mu_{\mathrm{B}}S_{z}B_{0}-g_{\mathrm{N}}\mu_{\mathrm{N }}I_{z}B_{0}+A\hat{\mathbf{S}}\cdot\hat{\mathbf{I}}, \tag{1}\]
The electronic and nuclear \(g\) factors are denoted by \(g_{e}\) and \(g_{N}\), respectively. The final term describes the hyperfine interaction, with coupling strength \(A\). At high fields, where the Zeeman interaction dominates, the energy levels are labelled by the quantum numbers \(m_{S}\) and \(m_{I}\), which give the projection of the relevant spin onto the axis defined by the direction of \(B_{0}\). The ESR signals arise from magnetic dipole transitions between different \(m_{S}\) levels, with selection rules \(\Delta m_{S}=\pm 1\).
The ESR spectrum, measured as a function of magnetic field, exhibits a pair of spin resonances corresponding to the two states of the \({}^{15}\)N nuclear spin, as shown in the upper inset of The resonances are separated in the field domain by the hyperfine splitting \(\Delta B\). We extract \(\Delta B\) by fitting the spectrum to identify the extrema of each resonance observed in the first-derivative spectra. This splitting is converted to a hyperfine coupling using the relationship \(A=g\mu_{B}\Delta B\), where the \(g\) factor \(g=hf/\mu_{\mathrm{B}}B_{\mathrm{av}}\). Here, \(B_{\mathrm{av}}\) is the centre field of the spectrum and \(f\) is the frequency of the applied microwave radiation. This frequency depends weakly on pressure via the changing dielectric constant of toluene, which alters the cavity frequency.
Data for the hyperfine coupling constant as a function of pressure are plotted in Fitting the data by a linear relationship, such that \(A(P)=A_{0}+A_{1}P\), gives \(A_{0}=22.263\) MHz and \(A_{1}=1.1\times 10^{-4}\) Hz Pa\({}^{-1}\). The bracketed number gives the statistical error in the last digit of the quoted value.
### Theory
We will now explain the pressure-dependent hyperfine coupling displayed in A similar effect has been measured previously for atomic nitrogen generated using a radio-frequency discharge plasma, where the hyperfine coupling is proportional to the buffer gas pressure. In contrast to the pressure-dependent coupling observed for atoms with unpaired \(s\) electrons, which can either be positive or negative, the magnitude of the nitrogen hyperfine coupling always increases with pressure.
This observation has been explained by treating van der Waals interactions between the nitrogen atom and the buffer gas atoms as the dominant contribution to the increased hyperfine coupling. These interactions lead to spin-dependent excitation of the \(s\) electrons, which causes spin polarization of the \(s\) orbital at the nucleus and an increased hyperfine coupling. Following an analogous model, we hypothesise that the increase in hyperfine coupling for the nitrogen endohedral fullerene is due to van der Waals interactions between the incarcerated nitrogen and the fullerene cage. There should also be a pressure-dependent exchange coupling between the cage and the \(p\) orbitals. However, because of spherical symmetry this does not mix \(p\) and \(s\) orbitals and is therefore expected to make a weaker contribution to the hyperfine coupling. We therefore model the shift as arising entirely from van der Waals interactions.
To apply this model to the endohedral fullerene system, we consider the increase in hyperfine coupling \(\Delta A=A-A_{\mathrm{free}}\), where \(A_{\mathrm{free}}\) is the value for free atomic nitrogen. A perturbative calculation of the van der Waals interaction between atomic nitrogen and a nearby particle predicts the relationship (Eq. of Ref.):
\[\Delta A=\frac{k}{R^{6}}, \tag{2}\]
Since the nitrogen atom is located at the centre of the cage, we set this distance equal to the fullerene radius. The constant \(k\) parameterises the strength of the interaction between the two charge clouds. Performing an ab initio calculation of the value of \(k\) is beyond the scope of this work; instead, we treat it as an experimentally determined parameter.
Hyperfine coupling constant \(A\) as function of pressure. The data (dark green circles with error bars) are fitted by a linear function (solid black line). The data points are the mean value of \(A\) at each pressure, and the error bars are the standard deviation of this measurement. Upper inset: ESR signal as function of magnetic field at frequency \(f\approx 9.334\) GHz, demonstrating the hyperfine splitting \(\Delta B\), peak-to-peak linewidth \(\delta B\), and the centre field \(B_{\mathrm{av}}\). Lower inset: schematic of high-field energy levels showing the two sets of transition frequencies corresponding to the two values of \(m_{I}\).
We test this model by using Eq. to predict the pressure shift given the bulk modulus \(B\).
\[\left(\frac{\partial\Delta A}{\partial P}\right)=\frac{2\Delta A}{B}. \tag{3}\]
For we take\(A_{\mathrm{free}}=-14.65\) MHz and \(A=-22.35\) MHz and hence \(\Delta A=7.70\) MHz. The bulk modulus of crystalline C\({}_{60}\) samples has been measured experimentally, but the value for an isolated C\({}_{60}\) molecule under conditions of hydrostatic pressure is only known via simulations. The simulated values range from 300 to 1200 GPa, with a grouping in the range 700 to 900 GPa. Taking \(B\sim 800\) GPa, Eq. 3 predicts \(\partial\Delta A/\partial P\sim 2\times 10^{-5}\). This value is approximately six times smaller than the value \(\partial\Delta A/\partial P=1.1\times 10^{-4}\) Hz Pa\({}^{-1}\) determined from the fit in
The discrepancy could have at least two causes. It may indicate that exchange interaction between the nitrogen and the cage, not captured by our van der Waals model, in fact contributes significantly to the nitrogen hyperfine coupling. Alternatively, the discrepancy could indicate that previous predictions have overestimated the fullerene's bulk modulus. To explain the measured pressure shift using Eq. 3 would require a value \(B=140\pm 13\), which is outside the range of theoretical predictions. However, these predictions themselves cover a wide range, which may indicate a need for further modelling. One source of uncertainty in the modelling is the effect of the solvent, which is expected to modify the molecule's bulk modulus. In particular, the effect of toluene solvent has not yet been modelled.
### Solvent effects
The most direct interpretation of the pressure dependence is that the nitrogen wavefunction is modified by compression of the cage. However, for some dissolved radicals, the hyperfine coupling constant depends on the solvent due to interactions between the solvent molecules and the unpaired electrons of the radical. For such an effect should be small, since the fullerene cage effectively isolates the atomic nitrogen from the environment. However, changes in the electronic properties of the solvent could plausibly alter the electronic properties of the fullerene cage; considering Eq. 2, this would correspond to altering \(k\). Therefore, we must exclude the possibility that the pressure-dependent hyperfine coupling reflects changes in solvent properties, such as dielectric constant \(\epsilon_{r}\) or molecular electric dipole moment \(\mu\), that are known to vary with pressure.
At room temperature and atmospheric pressure, toluene has dielectric constant \(\epsilon_{r}=2.39\) and dipole moment \(\mu=0.36\) D. At 0.25 GPa, \(\epsilon_{r}\) increases to \(\sim\)2.6, while \(\mu\) is likely to vary by less than 2%. To measure the effect of changes in \(\epsilon_{r}\) and \(\mu\), we therefore compare the spectrum of dissolved in toluene with those of dissolved in chlorobenzene (\(\epsilon_{r}=5.69\) and \(\mu=1.69\) D) and carbon disulfide (\(\epsilon_{r}=2.63\) and \(\mu=0\) D). The ESR spectra of dissolved in these solvents were measured at room temperature and atmospheric pressure at the \(X\) band using a Bruker ER4122-SHQE-W1 resonator and EMXmicro spectrometer. Data from these experiments are presented in Table 1.
The data show that the effect of altering the dielectric properties of the solvent is negligible to within experimental error. Moreover, the effect of using different solvents is less than the observed change in hyperfine coupling as a function of pressure, despite varying the dielectric constant and dipole moment by considerably more than the variation achieved by pressurising toluene. The data were obtained using a different spectrometer to the data presented in Fig. 1, and the small difference in \(A\) between comparable values presented in and Table 1 presumably reflects differences in magnet calibration. Such a systematic error does not affect the validity of the comparison between different solvents shown here. Therefore, we conclude that the increased hyperfine coupling at high pressures is not caused by pressure-dependent solvent properties, and that the cage compression model is correct.
## III Linewidth
In addition to the pressure-dependent hyperfine coupling, we also observe a pressure-dependent linewidth. The field domain peak-to-peak linewidth \(\delta B\), as shown in the upper inset of Fig. 1, was determined by measuring the splitting between the extrema of each resonance. As shown in Fig. 2, the linewidth increases with pressure.
Pressure-dependent linewidths are typically governed by spin-exchange processes. However, for exchange interactions between the incarcerated nitrogen atoms are suppressed by the cage, which prevents overlap of the electronic wavefunctions. We therefore use a model that instead considers the dipole-dipole interactions between spins in fullerene clusters.
\begin{table}
\begin{tabular}{c c c c} \hline \hline Solvent & \(\epsilon_{r}\) & \(\mu\) (D) & \(|\mathcal{A}|\) (MHz) \\ \hline Toluene & 2.39 & 0.375 & 22.247 \\ Carbon disulfide & 2.63 & 0 & 22.240 \\ Chlorobenzene & 5.69 & 1.69 & 22.253 \\ \hline \hline \end{tabular}
\end{table}
Table 1: Hyperfine coupling at room temperature and pressure in solvents with different dielectric constants \(\epsilon_{r}\) and dipole moments \(\mu\). Values of the dielectric constant are given for atmospheric pressure and \(T=293.2\) K. The bracketed number gives the one standard deviation error calculated by taking the sample standard deviation of six measurements.
At low pressures, the viscosity of the toluene solvent is low. The fullerene cluster therefore rotates sufficiently rapidly to average out the dipolar coupling between the spins, leading to a narrow linewidth. At high pressures, however, the viscosity increases, which reduces the rate of rotation. The linewidth then tends toward the rigid lattice limit imposed by dipole-dipole coupling between the spins in the cluster.
In this model, the dephasing time \(T_{2}^{*}\) obeys the implicit equation
\[\left(T_{2}^{*}\right)^{-2}=(2/\pi)C^{2}\tan^{-1}(2\tau_{c}/T_{2}^{*}), \tag{4}\]
the rigid lattice limit. The rotational correlation time \(\tau_{c}=4\pi\eta a^{3}/3k_{B}T\), where \(\eta\) is the viscosity of the solvent, \(a\) is the effective hydrodynamic radius of the rotating cluster, \(k_{B}\) is the Boltzmann constant, and \(T\) is the temperature. We fit the field domain linewidths \(\delta B\) presented in as a function of solvent viscosity, which is known from the previously measured equation of state. From the fit, we extract \(C=590\pm 100\) kHz and \(a=15.5\pm 0.8\) nm, which is comparable to the size of previously observed clusters.
By considering the interaction energy of two spins, we determine the average spin-spin separation \(r\approx\left(\mu_{0}S(S+1)g_{e}^{2}\mu_{B}^{2}/hC\right)^{1/3}=16.1\pm 0.9\) nm.
The value for \(a\) is much greater than the radius of an individual C\({}_{60}\) molecule, while the value of \(r\) is much less than the expected separation for unclustered fullerenes given the spin density. Both these facts are evidence for the clustering of fullerenes in our sample. However, the observed spin-spin separation is larger than the separation expected if the cluster were formed of close packed fullerenes. Such a cluster would have a spin-spin separation of approximately 5 nm given the purity of the sample and the diameter of an individual fullerene. However, the values are consistent with a porous structure for the fullerene clusters, which would reduce the effective spin density in the cluster. The structure of fullerene clusters depends on the formation process. Slow aggregation leads to densely packed clusters, whereas mechanical agitation and exposure to light lead to the formation of fractal clusters. It is therefore possible that the concentration procedure, during which nitrogen was bubbled through the solution under ambient light conditions, promoted the formation of porous clusters. Furthermore, the sample used in this work was stored under ambient conditions for approximately one year after the initial purification procedure. The fullerene molecules may have oxidised during this storage period, which further promotes the formation of large clusters by forming epoxide bonds between individual fullerenes.
This rotating cluster model explains why the observed linewidth is much greater than the minimum achievable linewidth set by relaxation mechanisms inherent to the molecule itself, such as the Orbach process. The narrowest linewidths measured previously required careful sample preparation to inhibit clustering, which occurs at concentrations above 0.06 mg/mL, and minimise paramagnetic impurities such as dissolved oxygen. However, using such a low concentration in our experiments was not feasible given the small sample volume of the pressure cell and the available sample purity.
The field-independent clock transition in the low-field spectrum of should suppress the dipolar broadening. However, the clock transition does not protect against all dipolar decoherence mechanisms. The maximum achievable stability of a fullerene clock could therefore be constrained by the need to compromise between increasing spin density to increase signal intensity and the need to minimise linewidth broadening caused by dipole-dipole interactions.
## IV Conclusions
The proposed model, based on van der Waals interactions between the nitrogen atom and the cage, explains the pressure dependence of the hyperfine coupling. The model predicts a smaller pressure shift than is observed experimentally, which may indicate that considering only van der Waals interactions between the nitrogen atom and the cage is insufficient. However, the model provides reasonable agreement with the data given the spread in predictions of the bulk modulus. While the small magnitude of the pressure shift likely precludes using it to offset the temperature shift of the clock frequency, it ensures that a based frequency reference is insensitive to atmospheric pressure fluctuations.
The pressure-dependent linewidth is well fitted by a model based on dipole-dipole interactions between spins embedded in fullerene clusters. Dipolar broadening of the spin resonance depends on pressure via the viscosity of the solvent, which modifies the rotational correlation time of the cluster. This model provides insight into spin relaxation processes in concentrated solutions that may limit the stability of a fullerene clock.
Peak-to-peak linewidth and toluene viscosity as a function of pressure. The linewidth data (red circles with error bars) are fitted (black dashed line) using a model based on the rotation of fullerene clusters. The viscosity of toluene (solid orange line) over this pressure range is plotted from the measured equation of state given in reference.
|
10.48550/arXiv.1712.05991
|
High pressure electron spin resonance of the endohedral fullerene $^{15}\mathrm{N@C}_{60}$
|
R. T. Harding, A. Folli, J. Zhou, G. A. D. Briggs, K. Porfyrakis, E. A. Laird
| 4,304
|
10.48550_arXiv.1906.00922
|
### Orbital angular momentum constraints
Consider the Hamiltonian for an atomic many-electron system. At the non-relativistic limit, the operators corresponding to the square of the orbital angular momentum (\(\hat{L}^{2}\)) and its projection onto the \(z\)-axis (\(\hat{L}_{z}\)) commute with this Hamiltonian.
\[\langle\Psi|\hat{L}^{2}|\Psi\rangle=L(L+1), \tag{10}\]and
\[\langle\Psi|\hat{L}_{z}|\Psi\rangle=M_{L}, \tag{11}\]
These constraints can be expressed in terms of the elements of the 1- and 2-RDM as
\[\sum_{\xi=x,y,z} \bigg{(} \sum_{\sigma\tau}\sum_{pqrs}{}^{2}D_{q_{x}s_{r}}^{p_{x}r_{r}}[L_ {\xi}]_{q}^{p}[L_{\xi}]_{s}^{r} \tag{12}\] \[+ \sum_{\sigma}\sum_{pq}{}^{1}D_{q_{x}}^{p_{x}}[L_{\xi}^{2}]_{q}^{p }\bigg{)}=L(L+1),\]
and
\[\sum_{\sigma}\sum_{pq}{}^{1}D_{q_{\sigma}}^{p_{\sigma}}[L_{z}]_{q}^{p}=M_{L}, \tag{13}\]
A 1-RDM that satisfies Eq. 13 is not guaranteed to represent a wavefunction that is an eigenfunction of \(\hat{L}_{z}\).
\[(\Delta L_{z})^{2} = \sum_{\sigma\tau}\sum_{pqrs}{}^{2}D_{q_{x}s_{r}}^{p_{x}r_{r}}[L_{ z}]_{q}^{p}[L_{z}]_{s}^{r} \tag{14}\] \[+ \sum_{\sigma}\sum_{pq}{}^{1}D_{q_{\sigma}}^{p_{\sigma}}[L_{z}^{2} ]_{q}^{p}-M_{L}^{2}.\]
Here, we have assumed that the 1-RDM satisfies Eq. 13, and, thus, \(\langle\hat{L}_{z}\rangle^{2}=M_{L}^{2}\). Similar arguments could be made for RDMs that satisfy Eq. 12, so a constraint on the variance of \(\hat{L}^{2}\), \((\Delta L^{2})^{2}=\langle\hat{L}^{4}\rangle-\langle\hat{L}^{2}\rangle^{2}\), might also be desirable. However, the evaluation of this quantity requires knowledge of the four-particle RDM, so this constraint will not be considered in this work.
Since the angular momentum operator is pure imaginary, the RDMs that enter our computations can only represent states with non-zero \(M_{L}\) if they are allowed to take on complex values. Although the boundary-point SDP algorithm was initially defined using real matrices, its extension to the optimization of complex and even quaternion matrices is a purely technical challenge.
\[\mathrm{Re}(\mathbf{M})+i\,\mathrm{Im}(\mathbf{M})\simeq\left[\begin{array}[] {cc}\mathrm{Re}(\mathbf{M})&-\mathrm{Im}(\mathbf{M})\\ \mathrm{Im}(\mathbf{M})&\mathrm{Re}(\mathbf{M})\end{array}\right], \tag{15}\]
As discussed in Refs., the boundary-point SDP solver for the v2RDM problem is a two-step procedure.
\[\mathbf{A}\mathbf{A}^{T}\mathbf{y}=\mathbf{A}(\mathbf{c}-\mathbf{z})+t( \mathbf{b}-\mathbf{A}\mathbf{x}) \tag{16}\]
Here, \(\mathbf{x}\) represents the primal solution vector (which maps onto the RDMs), \(\mathbf{y}\) and \(\mathbf{z}\) represent dual solution vectors, \(\mathbf{c}\) represents a vector containing the one- and two-electron integrals that define the quantum system, and \(\mathbf{A}\) and \(\mathbf{b}\) represent the constraint matrix and vector, respectively, which encode the \(N\)-representability conditions. The symbol \(t\) represents a penalty parameter. In the second step, the primal solution \(\mathbf{x}\) and the secondary dual solution \(\mathbf{z}\) are updated via the solution of an eigenvalue problem. The rate-limiting step in this algorithm is the latter one, and its computational cost increases with the third-power of the dimension of the RDMs. As such, expanding the complex RDMs as is done in Eq. 15 will increase the number of floating-point operations required by the boundary-point SDP algorithm by a factor of eight.
We have performed numerical tests to determine the relative efficiency of real symmetric (DSYEV) and complex Hermitian (ZHEEV) eigensolvers. The wall time required to diagonalize a complex matrix of dimension 4000 is roughly 30% of that required for the diagonalization of a real symmetric matrix of twice the dimension, when using Intel's MKL library and one core of an Intel Core i7-6850K CPU. Hence, we elect to retain the use of complex RDMs and modify the boundary-point solver accordingly. The only substantive change is that the number of coupled linear equations represented by Eq. 16 increases by a factor of two; one set of equations is used to update \(\mathrm{Re}(\mathbf{y})\), while the other determines \(\mathrm{Im}(\mathbf{y})\). Because the constraints we consider do not directly couple the real and imaginary components of the RDMs, these equations can be solved independently.
## III Computational details
The boundary-point SDP solver for the complex v2RDM problem was implemented as a plugin to the Psi4 electronic structure package. Optimized RDMs obtained from this plugin satisfied the PQG \(N\)-representability conditions and the spin angular momentum constraints outlined in Sec. II. Energies from v2RDM computations were compared to those from full CI and multireference CI (MRCISD+Q) computations performed with the Psi4 and ORCA packages, respectively. All orbitals were considered active within all v2RDM and full CI computations, while the reference computations for MRCISD+Q considered only full valence active spaces. All computations on atomic systems employed the cc-pVDZ basis set, while linear molecular systems were described by the STO-3G, Dunning-Haydouble zeta (D95V), 6-31G*, and cc-pVDZ basis sets; the reader is referred to Sec. IV.2 for additional details.
For atomic systems, the v2RDM procedure was considered converged when \(\epsilon_{\rm error}<1.0\times 10^{-5}\) and \(\epsilon_{\rm gap}<1.0\times 10^{-4}\)\(E_{\rm h}\), with the exception of two cases identified in Table 2 for which the convergence were achieved at least at \(\epsilon_{\rm error}<4.4\times 10^{-6}\) and \(\epsilon_{\rm gap}<5.6\times 10^{-4}\)\(E_{\rm h}\). Here, \(\epsilon_{\rm error}\) refers to the maximum of the primal error (\(||{\bf A}{\bf x}-{\bf b}||\)) and ground error (\(||{\bf A}^{T}{\bf y}-{\bf c}+{\bf z}||\)), and the primal/dual energy gap, \(\epsilon_{\rm gap}\), is defined as \(|{\bf x}^{T}{\bf c}-{\bf b}^{T}{\bf y}|\). For linear molecular systems, the v2RDM procedure was considered converged when \(\epsilon_{\rm error}<1.0\times 10^{-4}\) and \(\epsilon_{\rm gap}<1.0\times 10^{-4}\)\(E_{\rm h}\), with the exception of several calculations used to produce The most challenging calculation could only be converged to \(\epsilon_{\rm error}<1.4\times 10^{-5}\) and \(\epsilon_{\rm gap}<2.0\times 10^{-3}\)\(E_{\rm h}\), and six other calculations were converged to at least \(\epsilon_{\rm error}<1.2\times 10^{-5}\) and \(\epsilon_{\rm gap}<8.3\times 10^{-4}\)\(E_{\rm h}\). The reader is referred to the Supporting Information for additional details.
All v2RDM computations exploited the block structure of the RDMs resulting from spin and abelian point-group symmetry considerations, but it should be noted that the point group was chosen in each case such all operators belonged to the totally symmetric irreducible representation. Hence, computations in which we constrained the expectation values of \(\hat{L}_{z}\) were performed within the \(C_{2h}\) point group, and computations in which we constrained the expectation value of \(\hat{L}^{2}\) were performed within the \(C_{i}\) point group.
The orbital angular momentum constraints outlined in Sec. II.2 involve molecular integrals that do not usually arise in quantum chemical energy calculations. The molecular integrals over the orbital angular momentum operator, \(\hat{L}_{z}\), were obtained from the standard molecular integral library in Psi4. On the other hand, the integrals over the square of the angular momentum operator are not implemented in this package. We evaluated integrals of the form \([L_{\xi}^{2}]_{q}^{p}=\langle\chi_{p}|\hat{L}_{\xi}^{2}|\chi_{q}\rangle\) numerically, where \(\xi\in\{x,y,z\}\), and \(\chi_{p}\) represents an atomic basis function. Numerical integrals were evaluated on the same quadrature grids employed with density functional theory (DFT) computations in Psi4. We use the Lebedev-Trueutler grid, which is the default grid for all DFT computations in Psi4.
## IV Results and discussion
In this Section, we numerically evaluate the effects of orbital angular momentum constraints in v2RDM computations on systems with well-defined orbital angular momentum symmetry. Table 1 provides the designations used to describe the constraints applied in calculations on atomic systems, as well as the complexity of the RDMs. Note that the consideration of \(\hat{L}^{2}\) symmetry does not require the use of complex RDMs, but L\({}^{2}\) computations were performed using our complex-valued v2RDM algorithm nonetheless.
### Atomic systems
First, as a technical note, the error incurred when using complex- and real-valued RDMs is nearly indistinguishable on this scale, which suggests that our complex-valued boundary-point SDP algorithm is implemented correctly. Second, we note that the error increases, in general, with the number of electrons. This observation is consistent with the fact that v2RDM methods with approximate \(N\)-representability constraints are not strictly size extensive. However, in the absence of orbital angular momentum constraints, the error does not increase monotonically with system size; it is exaggerated for states with non-zero orbital angular momentum. For these states, the application of \(\hat{L}^{2}\) constraints results in a minor improvement. On the other hand, constraints on the expectation value of \(\hat{L}_{z}\) lead to a significant improvement in accuracy. Here, these non-zero angular momentum states are taken to have the maximal orbital angular momentum, which results in complex-valued RDMs. The subsequent application of variance constraints [\((\Delta L_{z})^{2}=0\)] leads to essentially no improvement in the description of these maximal orbital angular momentum projection states.
Clearly, orbital angular momentum constraints play an important role in the v2RDM-based description of ground states with non-zero total angular momentum. The data in indicate that, in some cases (boron, carbon, and oxygen), the application of such constraints reduces the error in the v2RDM energy by more than a factor of two. Moreover, angular momentum constraints also allow us to directly optimize 2-RDMs for excited states that are not otherwise accessible by v2RDM methods. Table 2 illustrates energy differences between excited spin and orbital angular momentum states and the ground electronic states for all second-row atoms, except lithium and neon. Note that all results tabulated under the heading "L\({}_{z}\)" correspond to the maximum orbital angular momentum projection.
\begin{table}
\begin{tabular}{l c c} designation & RDM complexity & constraints enforced \\ \hline real & real & \\ complex & complex & \\ L\({}^{2}\) & complex & \(\langle\hat{L}^{2}\rangle\) & \\ L\({}_{z}\) & complex & \(\langle\hat{L}^{2}\rangle\), \(\langle\hat{L}_{z}\rangle\) & \\ \((\Delta\)L\({}_{z})^{2}\) & complex & \(\langle\hat{L}^{2}\rangle\), \(\langle\hat{L}_{z}\rangle\), \((\Delta L_{z})^{2}\) \\ \end{tabular}
\end{table}
Table 1: Designation of the v2RDM computations on atomic systems according to the complexity of the RDMs and the orbital angular momentum constraints enforced.
For the beryllium atom, the \({}^{1}\)S \(\rightarrow\)\({}^{3}\)P transition is equally well-described by all combinations of angular momentum constraints considered. On the other hand, the description of every other transition energy is improved by the consideration of angular momentum constraints, sometimes dramatically so. In particular, the consideration of \(\hat{L}^{2}\) symmetry improves the almost 1 eV error in the description of the \({}^{4}\)S \(\rightarrow\)\({}^{2}\)D transition in nitrogen by 0.32 eV. The subsequent application of the constraint on \(\langle\hat{L}_{z}\rangle\) reduces the error to only 0.15 eV.
Now, consider those cases in Table 2 where no numerical values are given under the heading "real;" the excited states in question are inaccessible to the v2RDM approach unless angular momentum constraints are imposed. In one case, the \({}^{4}\)S \(\rightarrow\)\({}^{4}\)P transition in nitrogen, a constraint on the expectation value of \(\hat{L}^{2}\) yields a terrible estimate of the excitation energy; it is too low by 5.78 eV. However, subsequent application of the constraint on \(\langle\hat{L}_{z}\rangle\) yields an excitation energy that agrees with that from the full CI to within less than 0.01 eV. We also observe that the application of the \(\hat{L}_{z}\) constraint improves over the consideration of the \(\hat{L}^{2}\) constraint alone for the \({}^{4}\)S \(\rightarrow\)\({}^{2}\)P transition in nitrogen, although the improvement is less dramatic in this case. On the other hand, it appears that the application of the \(\hat{L}^{2}\) constraint alone gives superior results to the application of both \(\hat{L}^{2}\) and \(\hat{L}_{z}\) constraints in the cases of the \({}^{3}\)P \(\rightarrow\)\({}^{1}\)S transitions in carbon and oxygen. We believe this behavior stems from an inconsistency in the description of different \(S\) and \(L\) states in v2RDM methods in general. For example, for linear chains of hydrogen atoms, we have found that large-\(S\) states are more well-constrained than low-\(S\) states. That effect, combined with an apparent complementary effect regarding the relative description of large-\(L\) and small-\(L\) states, results in estimates of the absolute energies of the \({}^{1}\)S states that are relatively poor, as compared to estimates of the absolute energies of higher angular momentum states in the same atoms (the absolute energies for all states considered here are tabulated in the Supporting Information). The application of \(\hat{L}^{2}\) constraints alone (i.e., without constraints on \(\langle\hat{L}_{z}\rangle\)) overstabilizes the \({}^{3}\)P states, resulting is a fortuitous cancellation of error in the description of the \({}^{3}\)P \(\rightarrow\)\({}^{1}\)S transitions in carbon and oxygen.
To this point, all computations enforcing constraints on \(\langle\hat{L}_{z}\rangle\) considered only the maximal orbital projection
\begin{table}
\begin{tabular}{c c c c c c} atom & transition & real & L\({}^{2}\) & L\({}_{z}\) & full CI \\ \hline Be & \({}^{1}\)S \(\rightarrow\)\({}^{3}\)P & 2.75 & 2.75 & 2.75 & 2.75 \\ B & \({}^{2}\)P \(\rightarrow\)\({}^{4}\)P & 3.56 & 3.56 & 3.52 & 3.51 \\ C & \({}^{3}\)P \(\rightarrow\)\({}^{1}\)D & 0.86 & 1.18 & 1.44 & 1.49 \\ C & \({}^{3}\)P \(\rightarrow\)\({}^{1}\)S & – & 2.80 & 2.68 & 2.93 \\ C & \({}^{3}\)P \(\rightarrow\)\({}^{5}\)S & 4.11 & 4.10 & 3.98 & 3.93 \\ N & \({}^{4}\)S \(\rightarrow\)\({}^{2}\)D & 1.75 & 2.07 & 2.57 & 2.72 \\ N & \({}^{4}\)S \(\rightarrow\)\({}^{2}\)P & – & 2.92 & 3.40\({}^{b}\) & 3.31 \\ N & \({}^{4}\)S \(\rightarrow\)\({}^{4}\)P & – & 5.46 & 11.24 & 11.24 \\ O & \({}^{3}\)P \(\rightarrow\)\({}^{1}\)D & 1.57 & 1.71 & 2.03 & 2.14 \\ O & \({}^{3}\)P \(\rightarrow\)\({}^{1}\)S & – & 4.28 & 3.82 & 4.30 \\ F & \({}^{2}\)P \(\rightarrow\)\({}^{4}\)P & 34.96 & 34.97 & 35.00\({}^{b}\) & 35.00 \\ \end{tabular} \({}^{1}\)S transitions in carbon and oxygen.
\end{table}
Table 2: Energy differences (eV) between ground and excited spin and orbital angular momentum states calculated by at the v2RDM\({}^{a}\) and full CI levels of theory. The lack of numerical data under the “real” heading indicates that the excited state in question is not accessible by v2RDM methods without considering angular momentum symmetry.
The v2RDM energy (E\({}_{\rm h}\)) for different \(L_{z}\) projection states corresponding to the \({}^{3}\)P and \({}^{1}\)D terms of the carbon and oxygen atoms.
Errors in ground-state energies (mE\({}_{\rm h}\)) of second row atoms computed at the v2RDM/cc-pVDZ level of theory, as compared to results from full CI.
Here, we demonstrate that, for a given \(L\)-state, different orbital angular momentum projections are not treated on equal footing by the v2RDM approach. illustrates the energy for each \(M_{L}\) state within the manifold of states associated with the \({}^{3}\)P and \({}^{1}\)D terms of the carbon and oxygen atoms. For comparison, the horizontal lines represent the corresponding full CI energies for each state. Clearly, the v2RDM approach fails to recover the proper degeneracy of different angular momentum projection states. Rather, the v2RDM energy is a convex function of the expectation value of \(\hat{L}_{z}\), with the maximal projection states giving the best lower-bound to the full CI energy. Similar observations were made by van Aggelen _et al._, regarding the treatment of spin projection states within v2RDM theory. The consideration of \(\langle\hat{L}_{z}\rangle=0\) constraint does not improve the quality of the v2RDM results over the case in which a real-valued algorithm is applied; this result is not too surprising, since any purely real-valued 1-RDM satisfies this constraint. What is more interesting is that forcing the variance (\(\Delta L_{z}\))\({}^{2}\) to vanish substantially improves the quality of the non-maximal orbital angular momentum projections, most dramatically so for the \(\langle\hat{L}_{z}\rangle=0\) state; such a constraint could be applied within a real-valued v2RDM optimization. On the other hand, variance constraints do not appear to improve the quality of the maximal orbital angular momentum projection states. Again, this behavior is similar to that observed in Ref. for spin projection states. In that work, the application of pure-state and ensemble spin conditions yielded comparable results for maximal spin projection states.
### Linear molecular systems
Unlike the Hamiltonian for atomic systems, the Hamiltonian for linear molecular systems does not commute with \(\hat{L}^{2}\), so, in this case, the only good orbital angular momentum quantum number is \(\Lambda=\langle\hat{L}_{z}\rangle\), the projection of the orbital angular momentum on the internuclear axis (which we have chosen to be aligned in the \(z\)-direction). The results presented above for atomic systems suggest that orbital angular momentum projection constraints may play a similarly important role in the v2RDM-based description of states with non-zero \(\Lambda\) (e.g., \(\Pi\), \(\Delta\), \(\Phi\), etc. states). Hence, in this Section, we explore the utility of constraints on \(\hat{L}_{z}\) and \((\Delta L_{z})^{2}\) in linear molecular systems, beginning with a simple question: at the v2RDM level of theory, is the ground state of molecular oxygen a singlet or a triplet?
Table 3 illustrates the energy gap between the \({}^{3}\Sigma\) and \({}^{1}\Lambda\) states of molecular oxygen, as computed at the v2RDM, full CI, and MRCISD+Q levels of theory, in various basis sets. Here, a positive value for the gap indicates that the triplet is lower in energy. Note that values labeled as "real" were generated without the consideration of orbital angular momentum constraints, so the orbital angular momentum is technically unspecified in these cases. In a minimal (STO-3G) basis, such a real-valued v2RDM computation predicts a triplet/singlet gap of 0.914 eV, which is in reasonable agreement with that from full CI (1.042 eV). However, the v2RDM result is surprisingly sensitive to the size of the basis set; in a 3-21G basis, the triplet/singlet gap reduces to 0.424 eV, and, in a cc-pVDZ basis, the singlet is actually predicted to be _lower_ in energy than the triplet by almost 0.2 eV. Table 3 also provides results from complex-valued v2RDM computations in which we have placed constraints on the expectation value and variance of \(\hat{L}_{z}\), where \(\Lambda=0\) for the triplet state (\({}^{3}\Sigma\)) and \(\Lambda=2\) for the singlet state (\({}^{1}\Delta\)). The application of orbital angular momentum constraints significantly improves the v2RDM results, in all basis sets. In particular, \(\hat{L}_{z}\) and \((\Delta\hat{L}_{z})^{2}\) constraints remedy the qualitative failure of the v2RDM approach within the cc-pVDZ basis. In this case, the predicted triplet/singlet gaps are 0.924 eV and 0.940 eV, respectively, which are both in reasonable agreement with the value of 1.049 eV predicted by MRCISD+Q.
In the cc-pVDZ basis set, the imposition of orbital angular momentum constraints is clearly important for obtaining the correct ordering of the spin angular momentum states of molecular oxygen. However, these constraints cannot guarantee the correct ordering of orbital angular momentum states within a given spin manifold; this trend is evident in energy diagrams depicted in In these diagrams, the energy levels in all cases are shifted such that the energy of the \({}^{3}\Sigma\) state is zero. In a minimal basis set [Fig. 3(a)], the full CI, v2RDM [L\({}_{z}\)], and v2RDM [(\(\Delta\)L\({}_{z}\))\({}^{2}\)] approaches all predict that the \({}^{3}\Sigma\) is the ground state. When constraining only the expectation value of \(\hat{L}_{z}\), the v2RDM approach incorrectly predicts that the three singlet states considered are nearly degenerate, and the energy of the \({}^{1}\Pi\) state in particular is severely underestimated. Further, the energies of the \({}^{5}\Sigma\) and \({}^{3}\Pi\) states are far too low. With variance constraints, the v2RDM approach recovers the correct ordering for all spin and orbital angular momentum states, but the spacing between the ground and \({}^{1}\Pi\) state is still underestimated by more than 1 eV. In the D95V and cc-pVDZ basis sets [Figs. 3(b) and 3(c), respectively], we observe similar dramatic failures of the v2RDM approach (with constraints on the expectation value of \(\hat{L}_{z}\)) to yield the
\begin{table}
\begin{tabular}{l c c c} \hline \hline & STO-3G & 3-21G & cc-pVDZ \\ \hline MRCISD+Q & 1.042\({}^{a}\) & 1.113 & 1.049 \\ real & 0.914 & 0.424 & -0.196 \\ L\({}_{z}\) & 1.031 & 1.132 & 0.924 \\ \((\Delta\)L\({}_{z})^{2}\) & 1.037 & 1.162 & 0.940 \\ \hline \hline \end{tabular} \({}^{a}\) For values labeled as “real,” the specification of the spin angular momentum state is meaningful, while the specification of the orbital angular momentum state is not.
\({}^{b}\) This value was obtained from the full CI.
\end{table}
Table 3: The relative energies (eV) of the \({}^{3}\Sigma\) and \({}^{1}\Delta\) states of molecular oxygen,\({}^{a}\) with an inter-atomic distance of 1.208 Å.
In the cc-pVDZ basis in particular, constraints on the expectation value of \(\hat{L}_{z}\) alone are insufficient to yield the correct ground state; the \({}^{1}\Sigma\) and \({}^{1}\Pi\) states are _both_ predicted to lie below the \({}^{3}\Sigma\) state. Fortunately, the application of variance constraints leads to the correct prediction that the ground state of molecular oxygen is a triplet. Nonetheless, in both the D95V and cc-pVDZ basis sets, the singlet and triplet states are not ordered correctly amongst themselves; energies of the \({}^{1}\Pi\), \({}^{1}\Sigma\), and \({}^{3}\Pi\) states are all severely underestimated. The relative energies of all of the states considered in are tabulated in the Supporting Information.
Here, the v2RDM curves were generated under orbital angular momentum constraints (\((\hat{L}_{z})=\Lambda\) and \((\Delta L_{z})^{2}=0\)), as well as the spin angular momentum constraints outlined in Sec. II for the maximal spin projection states. As observed in Table 3, the \({}^{3}\Sigma\) / \({}^{1}\Delta\) energy gap is well-predicted by the v2RDM approach at the equilibrium geometry, but the overall shapes of the v2RDM-derived curves are not particularly accurate. It is clear that the v2RDM approach suffers from some serious deficiencies, particularly in the limit of dissociation. The \({}^{3}\Sigma\), \({}^{1}\Delta\), and \({}^{5}\Pi\) curves should all share the same energy at dissociation, but they do not, regardless of the imposition of angular momentum constraints.
The lack of degeneracy of the \({}^{3}\Sigma\), \({}^{1}\Delta\), and \({}^{5}\Pi\) states in the limit of dissociation is similar to the behavior observed in Ref..
The relative energies (eV) of the spin and orbital angular momentum states of molecular oxygen described by the (a) STO-3G, (b) D95V, and (c) cc-pVDZ basis sets. All energies are given relative to that of the \({}^{3}\Sigma\) state.
Dissociation curves for molecular oxygen, calculated within the D95V basis set. The v2RDM computations enforced constraints on the expectation value and variance of \(\hat{L}_{z}\).
Here, we can draw similar conclusions regarding the orbital angular momentum projections. In the limit of dissociation, the ground state should have an energy equal to twice that of a single oxygen atom in its ground state (\({}^{3}\)P). Two such atoms could couple to form nine states with \(S\) = 0, 1, or 2 and \(\Lambda\) = 0, 1, or 2, all of which should be degenerate at large O-O bond distances. illustrates the energy of these nine states at an O-O bond length of 5.0 A; in all cases, the spin-projection state is chosen to be the maximal one. The dashed line represents twice the energy of an isolated oxygen atom in the \({}^{3}\)P state, as described by the v2RDM method (constraining the maximal spin and orbital angular momentum projection states, but not the expectation value of \(\hat{L}^{2}\)). We can draw two conclusions from these data. First, for a given spin state, higher orbital angular momentum projection states are more well constrained. Second, for a given orbital angular momentum projection state, the highest-multiplicity state is the most well constrained. Indeed, the highest energy is obtained for the \({}^{5}\Delta\) state; the size consistency error (\(E_{\rm O_{2}}\) - 2 \(E_{\rm O}\)) is only 2.9 m\(E_{\rm h}\) in this case.
Lastly, we consider dissociation curves for the \({}^{1}\Delta\) and \({}^{1}\Sigma\) states of another linear molecular system, C\({}_{2}\). It is well known that a proper description of these states requires a sophisticated treatment of electron correlation effects, and, in the absence of orbital angular momentum constraints, v2RDM methods can only describe whichever state lies lower in energy. What is more problematic is that, because the potential energy curves for the \({}^{1}\Sigma\) and \({}^{1}\Delta\) states should cross, a real-valued v2RDM computation may yield RDMs for different electronic states at different C-C bond lengths. illustrates v2RDM and full CI potential energy curves for C\({}_{2}\) computed within the 6-31G* basis set. Full CI results were taken from Ref.. The application of orbital angular momentum constraints facilitates the description of both states via the v2RDM approach, and, near the equilibrium geometry for the ground state, we observe reasonable splittings between the ground and excited states. At a C-C bond length of 1.25 A, full CI predicts that the \({}^{1}\Delta\) state lies 2.43 eV above the \({}^{1}\Sigma\) state, while the v2RDM approach predicts that these states are separated by 2.90 eV. The relative overstabilization of the \({}^{1}\Sigma\) state is consistent with our observation that, for a given spin state, higher orbital angular momentum projection states are more well-constrained. Unfortunately, the v2RDM method exhibits two qualitative failures for this system. First, it predicts that the \({}^{1}\Sigma\) state is the ground state for all C-C bond lengths; that is, the potential energy cures for the two states are predicted to never cross. Second, as was observed above for molecular oxygen, the two electronic states considered here do not share the same dissociation limit.
## V Conclusions
In systems with well-defined orbital angular momentum symmetry, the application of orbital angular momentum constraints facilitates the direct variational determination of 2-RDMs for multiple electronic states. Moreover, without such considerations, the v2RDM approach cannot qualitatively describe states with non-zero \(z\)-projection of the orbital angular momentum, even if the state in question is the lowest-energy state of a given spin symmetry. Indeed, we demonstrated that, in the absence of orbital angular momentum constraints, the v2RDM approach incorrectly predicts that the ground state of molecular oxygen (described by the cc-pVDZ basis set)
The energy of molecular oxygen (\(E_{\rm h}\)), as described by the D95V basis set, at an O–O distance of 5 Å. The v2RDM computations enforced constraints on the expectation value of \(\hat{L}_{z}\) or both the expectation value and variance of \(\hat{L}_{z}\).
Dissociation curves for the \({}^{1}\Sigma\) and \({}^{1}\Delta\) states of molecular carbon, calculated using the 6-31G* basis set. The v2RDM computations enforced constraints on the expectation value of \(\hat{L}_{z}\), and the full CI results were taken from Ref..
The application of appropriate constraints, which necessitates the consideration of complex-valued RDMs, recovers the correct spin-state ordering.
The v2RDM energy appears to be a convex function of the expectation value of \(\hat{L}_{z}\), and, for a given magnitude of the orbital angular momentum, maximal orbital angular momentum projection states are the most well-constrained. This result reveals a qualitative failure of v2RDM methods: they do not to recover the correct degeneracy for different \(L/M_{L}\) states, at least when the RDMs satisfy the ensemble \(N\)-representability conditions considered in this work. This behavior suggests that the conclusions of Ref. regarding the description of different spin projection states apply to angular momentum projection states in general. Presumably, should one consider the direct optimization of 2-RDMs corresponding to different _total_ angular momentum states, similarly incorrect behavior would emerge.
|
10.48550/arXiv.1906.00922
|
The role of orbital angular momentum constraints in the variational optimization of the two-electron reduced-density matrix
|
Run R. Li, A. Eugene DePrince III
| 3,298
|
10.48550_arXiv.2202.07354
|
###### Abstract
We investigate a class of diffusion-controlled reactions that are initiated at the time instance when a prescribed number \(K\) among \(N\) particles independently diffusing in a solvent are simultaneously bound to a target region. In the irreversible target-binding setting, the particles that bind to the target stay there forever, and the reaction time is the \(K\)-th fastest first-passage time to the target, whose distribution is well-known. In turn, reversible binding, which is common for most applications, renders theoretical analysis much more challenging and drastically changes the distribution of reaction times. We develop a renewal-based approach to derive an approximate solution for the probability density of the reaction time. This approximation turns out to be remarkably accurate for a broad range of parameters. We also analyze the dependence of the mean reaction time or, equivalently, the inverse reaction rate, on the main parameters such as \(K\), \(N\), and binding/unbinding constants. Some biophysical applications and further perspectives are briefly discussed.
pacs: 02.50.-r, 05.40.-a, 02.70.Rr, 05.10.Gg +
Footnote †: : _J. Phys. A: Math. Gen._
## 1 Introduction
Diffusion-controlled processes and reactions play the central role in microbiology, physiology and many industrial procedures. In a common setting of bimolecular reactions, two particles (e.g., a ligand and a receptor) need to meet each other to initiate a reaction event. As the encounter results from the stochastic motion of one or both particles, the reaction time is random. Since the seminal work by von Smoluchowski, such first-encounter or first-passage problems have been thoroughly investigated. Among various studied aspects, one can mention the impact of structural organization and dynamical heterogeneities of the medium, the asymptotic behavior of the reaction rate and the mean first-passage time in the small-target limit, distinct features of the whole distribution, and the effect of target mobility.
However, there exist more sophisticated processes (that we will still call "reactions") involving multiple particles. In microbiology, there are many activation mechanisms controlled by a threshold crossing such as signalling in neurons, synaptic plasticity, cell apoptosis caused by double strand DNA breaks, cell differentiation and division. For instance, binding of five calcium ions to a calcium-ion-sensing protein initiates a release of neurotransmitters in the signalling process between two neurons. Similarly, the ryanodine receptor is activated when two calcium ions bind to the receptor binding sites. In these examples, the biochemical event such as signal transmission starts when a fixed number \(K\) among \(N\) diffusing particles are simultaneously bound to the target region for the first time. If \(\mathcal{N}(t)\) denotes the number of bound particles at time \(t\), the reaction time \(\mathcal{T}_{K,N}=\inf\{t>0\;:\;\mathcal{N}(t)=K\}\) is the first-crossing time of a fixed threshold \(K\) by the stochastic non-Markovian process \(\mathcal{N}(t)\). In the idealized case of irreversible binding when any particle after its binding to the target stays bound forever, this is the problem of finding the \(K\)-th fastest first-passage time \(\mathcal{T}^{0}_{K,N}\) to the target. If the particles diffuse independently, the distribution of \(\mathcal{T}^{0}_{K,N}\) can be easily expressed in terms of the survival probability for a single particle (see A). In most cases, however, binding is reversible so that some particles can unbind and resume their diffusion before the binding of the \(K\)-th fastest particle that renders the problem of such "impatient" particles much more challenging. Recently, Lawley and Madrid proposed an elegant approximation, in which the first-binding time and the rebinding time \(\tau\) after each unbinding event were assumed to obey an exponential law. The process \(\mathcal{N}(t)\) could thus be approximated by a Markovian birth-death process, for which the distribution of the first-crossing time is known explicitly (see also). In the special case \(K=N\), we derived the exact solution of the problem of impatient particles and showed both advantages and limitations of the Lawley-Madrid approximation (LMA). Despite its crucial role in providing us with analytical insight into the problem of impatient particles and the validity of its approximate treatments, the case when all particles have to bind the target is not so common in applications.
In this paper, we investigate the general problem of impatient particles in a common setting when all particles start from independent uniformly distributed positions. First,we revisit the Lawley-Madrid approximation and discuss its validity range. In particular, we argue that the key assumption of the LMA requires that the target is small _and_ weakly reactive. The condition of weak reactivity, which was not emphasized on in, limits the applicability of this approximation. To overcome this limitation, we develop an alternative approach to the general problem. Our approximate solution is confronted to Monte Carlo simulations and shown to be remarkably accurate for a broad range of parameters. It allowed us to investigate the short-time and long-time behaviors of the probability density of the reaction time \(\mathcal{T}_{K,N}\), the dependence of the mean reaction time on the unbinding rate, and the role of the numbers \(K\) and \(N\).
The paper is organized as follows. In Sec. 2, we formulate the problem of impatient particles and discuss the LMA. Section 3 presents the main steps of our approach and summarizes the approximate formulas for the probability density of the reaction time \(\mathcal{T}_{K,N}\), its short-time and long-time behaviors, and the mean reaction time. In Sec. 4, we illustrate these results for an emblematic model of restricted diffusion between concentric spheres. We discuss the accuracy of our approximation and its limitations. Section 5 concludes the paper and suggests further perspectives. As our derivations are technically elaborate, most mathematical details are re-delegated to Appendices in order to facilitate the main text for a wider audience.
## 2 Problem of impatient particles
We consider \(N\) particles that independently diffuse with diffusion coefficient \(D\) inside a bounded domain \(\Omega\subset\mathbb{R}^{d}\) with a smooth boundary \(\partial\Omega\) that is reflecting everywhere except for a target region \(\Gamma\) with a finite reactivity \(\kappa\). For instance, \(\Omega\) may represent the cytoplasm of a living cell, surrounded by a plasma membrane \(\partial\Omega\) that is impermeable for diffusing particles, and \(\Gamma\) be the boundary of an organelle or a sensor protein on that membrane. The reactivity \(\kappa\) (in units m/s) is related to the binding probability and characterizes how easily the particle can bind the target upon their encounter, ranging from \(\kappa=0\) for an inert target (no binding) to \(\kappa=\infty\) for a perfectly reactive target (binding upon the first encounter). The finite reactivity may represent the effect of an energetic or entropic barrier for binding, stochastic switching between open and closed states of the target (e.g., an ion channel), microscopic heterogeneity of the target, etc.. In (bio)chemistry, the reactivity is usually expressed in terms of the forward (bimolecular) reaction rate \(k_{\rm on}\) via \(\kappa=k_{\rm on}/(|\Gamma|N_{A})\), where \(|\Gamma|\) is the surface area of the target and \(N_{A}\) is the Avogadro number. After binding, each particle stays on the target region for a random exponentially distributed waiting time, characterized by the unbinding rate \(k_{\rm off}\), and then resumes its diffusion from a uniformly distributed point on \(\Gamma\). The particle diffuses in \(\Omega\) until the next binding, and so on. In other words, each particle alternates between free and bound states. We aim at describing the random reaction time \(\mathcal{T}_{K,N}\), i.e., the first instance when \(K\) particles among \(N\) are simultaneously in the bound state on the target region that is considered as a trigger of the underlying biochemical process (a reaction event). As binding and unbinding eventsof all particles are independent from each other and thus asynchronized, finding the probability density \(\mathcal{H}_{K,N}(t)\) of the \(\mathcal{T}_{K,N}\) is a challenging open problem. Note that the above problem of impatient particles resembles some stochastic models of multi-channel particulate transport with blockage.
The first-binding time \(\tau_{0}\) and the consequent rebinding times \(\tau_{1},\tau_{2},\ldots\) of any particle are random variables, which are characterized by the survival probabilities \(S(t|\mathbf{x}_{0})=\mathbb{P}_{\mathbf{x}_{0}}\{\tau_{0}>t\}\) and \(S(t)=\mathbb{P}\{\tau_{i}>t\}\), where \(\mathbf{x}_{0}\) is the starting point of the particle, and \(\mathbb{P}\{\ldots\}\) denotes the probability of a random event between braces.
\[\epsilon=\frac{\kappa\left|\Gamma\right|\left|\Omega\right|}{D|\partial\Omega| ^{2}}\ll 1, \tag{2}\]
For clarity, we focus here on a three-dimensional setting, \(d=3\), but the arguments are valid in higher dimensions as well.
**(a)** A planar illustration of a bounded domain \(\Omega\) between two concentric spheres of radii \(\rho=1\) and \(R=2\), whose disjoint boundary \(\partial\Omega=\partial\Omega_{0}\cup\Gamma\) is composed of the reflecting outer sphere \(\partial\Omega_{0}\) and the partially reactive inner sphere \(\Gamma\). **(b)** A numerical simulation for three diffusing particles. Upper plot shows the radial coordinate, \(\left|\mathbf{X}_{t}\right|\), of simulated trajectories of three particles that start from a fixed initial position with \(\left|\mathbf{x}_{0}\right|=1.5\) and diffuse independently, with eventual bindings to the target. Arrows indicate the first-crossing times \(\mathcal{T}_{1,3}\), \(\mathcal{T}_{2,3}\), and \(\mathcal{T}_{3,3}\). Bottom plot illustrates the number of bound particles at time \(t\), \(\mathcal{N}(t)\). At the beginning, all three particles are free, and \(\mathcal{N}=0\). At \(\mathcal{T}_{1,3}\), the “red” particle binds, switching the counter \(\mathcal{N}(t)\) to \(1\). At \(\mathcal{T}_{2,3}\), the “green” particle binds, switching the counter \(\mathcal{N}(t)\) to \(2\). Few moments later, the “red” particle unbinds, diffuses and rebinds to the target. Finally, the last “blue” particle binds at time \(\mathcal{T}_{3,3}\), switching the counter \(\mathcal{N}(t)\) to \(3\).
LMA and discuss the validity of the condition, which actually combines two distinct properties of the target: its relative size and reactivity. We argue that the LMA is applicable when the target is small _and_ weakly reactive. For instance, when the target is a sphere of radius \(\rho\), the following two conditions should be fulfilled:
\[\rho\ll R=\frac{|\partial\Omega|^{2}}{4\pi|\Omega|}\,,\qquad\frac{\kappa\rho}{D }\ll 1. \tag{3}\]
The first condition is purely geometrical (smallness of the target as compared to the confining domain), while the second condition involves both the reactivity and the size of the target but does not depend on the confining domain. These two conditions evidently imply Eq., but the opposite claim is not true. In particular, if the target is small but highly reactive, the second condition may not be valid, even if Eq. is fulfilled. This situation will be illustrated in Sec. 4.
## 3 Approximate solution
To overcome the constraint on weak reactivity, we develop an alternative approach, which does not rely on the approximation. For this purpose, we extend the derivation in Ref.
\[\mathcal{P}_{t}(N|0)=\int\limits_{0}^{t}dt^{\prime}\,\mathcal{H}_{N,N}(t^{ \prime})\,\mathcal{P}_{t-t^{\prime}}(N|N), \tag{4}\]
Expressing both \(\mathcal{P}_{t}(N|0)\) and \(\mathcal{P}_{t-t^{\prime}}(N|N)\) in terms of known occupation probabilities for a single particle and applying the Laplace transform led to the probability density \(\mathcal{H}_{N,N}(t)\) in the Laplace domain.
A direct extension of this equation to the general case \(K<N\) fails. In fact, the probability \(\mathcal{P}_{t}(K|0)\) can still be expressed as an integral of \(\mathcal{H}_{K,N}(t^{\prime})\) with the probability \(\mathcal{P}_{t-t^{\prime}}(K|K)\) of transition from a state with \(K\) bound particles to another state with \(K\) bound particles. However, this probability also depends on _random_ positions of the remaining \(N-K\) free particles at time \(t^{\prime}\) that should be averaged out. Even for independently diffusing particles, an exact computation of this average remains an open problem (see C for further discussion). Moreover, the resulting probability would be a function of both \(t-t^{\prime}\) and \(t^{\prime}\) so that an extension of Eq. would be no longer a convolution, and thus would not be simplified in the Laplace domain.
This fundamental difficulty can be partly resolved in the case when the starting positions of \(N\) particles are uniformly distributed in the confining domain. The key point is that the distribution of any _free_ particle that started uniformly remains to be almost uniform at all times, except for a boundary layer near the target region. When the target is small and not too highly reactive, this boundary layer is narrow and can be neglected so that all free particles can be approximately treated as uniformly distributed at any time \(t^{\prime}\). As a consequence, the average of \(\mathcal{P}_{t-t^{\prime}}(K|K)\) turns out to be only a function of \(t-t^{\prime}\), thus keeping the convolution form of the renewal equation:
\[\mathcal{P}_{t}(K|0)=\int\limits_{0}^{t}dt^{\prime}\,H_{K,N}(t^{\prime})\, \overline{\mathcal{P}_{t-t^{\prime}}(K|K)}, \tag{5}\]
In other words, this integral equation determines an approximation \(H_{K,N}(t)\) of the probability density \(\mathcal{H}_{K,N}(t)\) of the reaction time \(\mathcal{T}_{K,N}\). Both transition probabilities in Eq.
\[\mathcal{P}_{t}(K|0)=\binom{N}{K}[P(t|\circ)]^{K}[1-P(t|\circ)]^{N-K} \tag{6}\]
and
\[\overline{\mathcal{P}_{t}(K|K)} = \sum_{j=0}^{K}\binom{K}{j}[Q(t)]^{K-j}[1-Q(t)]^{j}\binom{N-K}{j} \tag{7}\] \[\times [P(t|\circ)]^{j}[1-P(t|\circ)]^{N-K-j},\]
Here \(P(t|\circ)\) (resp., \(Q(t)\)) is the probability of finding a particle that was free with uniform initial distribution (resp., bound) at time \(0\), in the bound state at time \(t\). For instance, the term with \(j=0\) in Eq. describes the configuration when all \(K\) initially bound particles are found to be bound at time \(t\) (note that they can unbind and rebind in the meantime), while \(N-K\) initially free particles are found to be free at time \(t\) (they can also bind and unbind in the meantime). Similarly, the term with \(j=1\) describes the configuration when \(K-1\) initially bound particles are found to be bound at time \(t\), one initially bound particle is found to be free at time \(t\), \(N-K-1\) initially free particles are found to be free at time \(t\), while one initially free particle is found to be bound at time \(t\) (and all these particles can undertake an arbitrary number of binding/unbinding events in the meantime).
\[P(t|\circ)=\frac{1-Q(t)}{k_{\mathrm{off}}\langle\tau\rangle}\,, \tag{8}\]
In, we derived a very simple and general expression for this quantity:
\[\langle\tau\rangle=\frac{|\Omega|}{\kappa|\Gamma|}=\frac{N_{A}|\Omega|}{k_{ \mathrm{on}}} \tag{9}\]
(we reproduce its derivation in C). Here, it is expressed in terms of the volume \(|\Omega|\) of the confining domain, the surface area \(|\Gamma|\) of the target region, and its reactivity \(\kappa\) or, equivalently, in terms of the forward reaction constant \(k_{\mathrm{on}}\). Counter-intuitively, the mean rebinding time does not depend on the diffusion coefficient \(D\). This is a particular example of the invariance property of general random walks in bounded domains that the mean traveled distance (and thus the mean exit time) does not depend on the dynamics of the diffusing particles that enter and exit the domain through the same subset of the boundary (here, the target). Solving the convolution equation in the Laplace domain, we obtain the approximate probability density \(H_{K,N}(t)\) of the reaction time \(\mathcal{T}_{K,N}\):
\[H_{K,N}(t)=\mathcal{L}^{-1}\left\{\frac{\mathcal{L}\{\mathcal{P}_{t}(K|0)\}}{ \mathcal{L}\{\mathcal{P}_{t}(K|K)\}}\right\}, \tag{10}\]
This approximate solution of the general problem of impatient particles constitutes the main result of the paper. For \(K=N\), one has \(\mathcal{P}_{t}(K|0)=[P(t|\circ)]^{N}\) and \(\overline{\mathcal{P}_{t}(K|K)}=[Q(t)]^{N}\) and thus retrieves an extension of the exact solution from Ref. to the case of the uniform initial distribution of the particles.
In addition to a direct numerical way of computing the approximate probability density \(H_{K,N}(t)\) (see E for details), Eq.
\[T_{K,N}\approx\frac{1}{\overline{\mathcal{P}_{\infty}(K|K)}}\int\limits_{0}^{ \infty}dt\Big{(}\overline{\mathcal{P}_{t}(K|K)}-\overline{\mathcal{P}_{\infty} (K|K)}\Big{)}, \tag{13}\]
with \(P(\infty|\circ)=Q(\infty)=1/(1+k_{\rm off}\langle\tau\rangle)\). In addition, our approximate solution allows us to evaluate the moments of the reaction time \(\mathcal{T}_{K,N}\). For instance, we derived the following approximation for the mean reaction time (see G)
\[\langle\mathcal{T}_{K,N}\rangle\approx\frac{1}{\mathcal{P}_{\infty}(K|0)}\int \limits_{0}^{\infty}dt\Big{(}\overline{\mathcal{P}_{t}(K|K)}-\mathcal{P}_{t}( K|0)\Big{)}. \tag{14}\]
Note that this expression is similar to Eq. for the decay time, and they usually yield very close results.
The dimensionless parameter \(\eta=k_{\rm off}\langle\tau\rangle\propto k_{\rm off}/k_{\rm on}\) determines whether the reversible binding kinetics is relevant (\(\eta\gtrsim 1\)) or not (\(\eta\ll 1\)). As discussed in G, Eq. fails as \(\eta\to 0\) but gets more and more accurate as \(\eta\) increases. For \(\eta\gg 1\), the integral in Eq.
\[\langle\mathcal{T}_{K,N}\rangle\approx\langle\tau\rangle\,\frac{(k_{\rm off} \langle\tau\rangle)^{K-1}}{K\binom{N}{K}}\qquad(\eta\gg 1). \tag{15}\]
For \(K=1\), the approximate mean reaction time \(\langle\mathcal{T}_{1,N}\rangle\approx\langle\tau\rangle/N\) does not depend on \(k_{\rm off}\), as the first-binding event is independent of the unbinding kinetics. This mean value decreases inversely proportional to \(N\), as discussed earlier in Ref. in the context of the fastest first-passage time problem.
\[\langle\mathcal{T}_{K,N}\rangle\approx k_{\rm off}(K-1)!\left(\frac{\kappa|\Gamma| N}{k_{\rm off}|\Omega|}\right)^{K}, \tag{16}\]
## 4 Discussion
To illustrate our general results, we consider restricted diffusion inside a confining reflecting sphere of radius \(R\) towards a small concentric partially reactive spherical target of radius \(\rho\) (Fig. 1(a)). This domain can be considered as an idealized model for the intracellular transport towards the nucleus or a model of the presynaptic bouton.
Probability density of the reaction time \(\mathcal{T}_{K,N}\) for restricted diffusion between concentric spheres of radii \(\rho\) and \(R=10\rho\), with \(N=4\), \(\kappa\rho/D=1\), a timescale \(\delta=\rho^{2}/D\), three values of \(k_{\rm off}\) (see legend), and four values of \(K\): \(K=1\)**(a)**, \(K=2\)**(b)**, \(K=3\)**(c)**, and \(K=4\)**(d)**. Symbols show empirical histograms from Monte Carlo simulations with \(10^{6}\) particles. Thick solid line presents the exact solution for irreversible binding; thick dashed lines indicate our approximation evaluated numerically as described in E. Thin lines show the Lawley-Madrid approximation, with \(\nu\) given by Eq.; note that the thin line for the case \(k_{\rm off}\delta=0.03\) in panel **(d)** is not visible as it appears below the figure (i.e., \(\tilde{H}_{4,4}(t|\diamond)\delta<10^{-6}\)). Thin gray solid line presents the short-time asymptotic behavior.
As the unbinding kinetics can only be initiated after the first binding, the reaction time \(\mathcal{T}_{1,N}\) is equal to the first-binding time of the fastest particle and thus does not depend on the unbinding rate \(k_{\rm off}\) (see also A). Expectedly, three curves with different \(k_{\rm off}\) coincide on the panel Fig. 2(a). Moreover, the short-time behavior does not depend on \(k_{\rm off}\) for any \(K\). In turn, the long-time decay is strongly affected by \(k_{\rm off}\) when \(K>1\): the decay time \(T_{K,N}\) increases with \(k_{\rm off}\) and thus the distribution is getting broader for faster unbinding kinetics. In all cases, the approximate solution is in a remarkable agreement with Monte Carlo simulations over a broad range of times. We also stress that our solution is exact for \(K=N\). The Lawley-Madrid approximation (see B) captures correctly the overall behavior but overestimates the decay time. The agreement is better for smaller \(k_{\rm off}\) and smaller \(K\). In turn, the disagreement for larger \(k_{\rm off}\) or \(K\) is caused by moderate reactivity of the target, for which the second condition in Eq. is not satisfied. Note that the parameter \(\epsilon\) from Eq. is equal to \(0.03\), wrongly suggesting the validity of the LMA. This example clearly illustrates that the single condition is not sufficient and should be replaced by two separate conditions in. from H illustrates that the disagreement is getting even bigger for a small target with higher reactive \(\kappa\rho/D=10\). In contrast, the LMA is very accurate for weakly reactive targets (see, e.g., in Ref., which was plotted for the case \(\kappa\rho/D=0.01\) and \(k_{\rm off}\rho^{2}/D=0.001\)). Finally, we emphasize that the short-time asymptotic relation is not accurate in the considered range of times, requiring many correction terms for amendment (see F for details). Similar behavior was observed for \(N=2\) and \(N=3\) (see Figs. 12 and 13 from H).
The impact of unbinding kinetics and the consequent rebinding events can be characterized by the dimensionless parameter \(\eta=k_{\rm off}\langle\tau\rangle\), which is proportional to the ratio \(k_{\rm off}/k_{\rm on}\) (or \(k_{\rm off}/\kappa\)), see Eq.. In particular, this parameter fully determines the steady-state probability \(P(\infty|\circ)=1/(1+\eta)\) for a particle to be in the bound state. Intuitively, one might expect that \(\eta\) mainly controls the statistics of the reaction times \(\mathcal{T}_{K,N}\). To emphasize on the respective roles of binding and unbinding effects, we fix \(\eta=1\) and compare the probability densities for three combinations of \(\kappa\) and \(k_{\rm off}\). shows that two curves with larger unbinding rates \(k_{\rm off}(\rho^{2}/D)=0.03\) and \(k_{\rm off}(\rho^{2}/D)=0.3\) (and, accordingly, larger reactivities) almost coincide. This effect can be attributed to a sort of statistical averaging due to multiple rebinding events. In contrast, the curve with the lowest \(k_{\rm off}\) and \(\kappa\) differs from the others, due to a limited number of rebinding events. We conclude that the parameter \(\eta\) plays an important role but does not fully determine the statistics of the reaction time. Expectedly, the Lawley-Madrid approximation gets less and less accurate as the reactivity increases.
We complete this section by looking at the mean reaction time \(\langle\mathcal{T}_{K,N}\rangle\). shows the dependence of \(\langle\mathcal{T}_{K,N}\rangle\) on the unbinding rate \(k_{\rm off}\) (rescaled by \(\langle\tau\rangle\)) for a fixed reactivity \(\kappa\rho/D=1\). When \(\eta=k_{\rm off}\langle\tau\rangle\) is small, the mean reaction time is almost constant and close to \(\langle\mathcal{T}_{K,N}^{0}\rangle\) for irreversible binding (\(k_{\rm off}=0\)), as expected. In turn, for \(\eta\gtrsim 1\), the mean reaction time starts to rapidly increase with \(\eta\).
Mean reaction time \(\langle\mathcal{T}_{K,3}\rangle\) for restricted diffusion between concentric spheres of radii \(\rho\) and \(R=10\rho\), with \(\kappa\rho/D=1\), a timescale \(\delta=\rho^{2}/D\), \(N=3\), and three values of \(K\) (see legend). Thick lines show our approximation, thin lines present the Lawley-Madrid approximation with \(\nu\) given by Eq., while symbols illustrate the results of Monte Carlo simulations with \(10^{6}\) realizations. Thin straight solid lines present the large-\(\eta\) asymptotic behavior.
Probability density of the reaction time \(\mathcal{T}_{K,N}\) for restricted diffusion between concentric spheres of radii \(\rho\) and \(R=10\rho\), with \(N=4\), a timescale \(\delta=\rho^{2}/D\), three combinations of \(k_{\rm off}\) and \(\kappa\) (\(k_{\rm off}\delta=0.003,0.03,0.3\) corresponding to \(\kappa\rho/D=1,10,100\), respectively, such that \(\eta=1\) in all cases), and four values of \(K\): \(K=1\)**(a)**, \(K=2\)**(b)**, \(K=3\)**(c)**, and \(K=4\)**(d)**. Symbols show empirical histograms from Monte Carlo simulations with \(10^{6}\) particles. Thick lines indicate our approximation evaluated numerically as described in E, whereas thin lines show the Lawley-Madrid approximation, with \(\nu\) given by Eq.; note that the thin line for the case \(k_{\rm off}\delta=0.03\) in panel **(d)** is not visible as it appears below the figure (i.e., \(\bar{H}_{4,4}(t|\circ)\delta<10^{-6}\)).
## 5 Conclusion
In this paper, we investigated diffusion-controlled reactions or events that are triggered on a target region after binding a prescribed number \(K\) among \(N\) independently diffusing particles. The reversible target-binding kinetics, which is so common for most applications, presented the major mathematical difficulty. We developed a powerful theoretical approach to derive a new approximation \(H_{K,N}(t)\) for the probability density of the reaction time \(\mathcal{T}_{K,N}\) in the case when the particles were initially released uniformly. Under the assumption that the random positions of free particles at time \(\mathcal{T}_{K,N}\) remain to be uniform, we derived a renewal equation that determines \(H_{K,N}(t)\). This convolution-type equation was then solved in the Laplace domain to relate the probability density via Eq. to two occupancy probabilities, which were in turn expressed in terms of the survival probability for a single particle. In this way, we managed to describe the collective effect of multiple diffusing particles in terms of the diffusive dynamics of a single particle and thus to extend the well-known extreme statistics for the \(K\)-th fastest first-passage time to a more general and much more challenging setting with reversible binding. In other words, the knowledge of the survival probability \(S(t|\circ)\) (or, equivalently, \(S(t)\)) of a single particle was sufficient for approximating the probability density of the reaction time \(\mathcal{T}_{K,N}\).
The assumption of uniform positions was the crucial step and the only source of eventual deviations between the exact probability density and our approximation. Strictly speaking, this assumption is fulfilled exactly only for an inert non-reactive target (\(\kappa=0\)). When the target is reactive, binding events lead to a formation of a depletion boundary layer near the target, in which the probability density of finding a diffusing particle is lower, and thus not uniform. In contrast, unbinding events tend to homogenize the probability density and thus render our assumption more accurate. As a consequence, our approximation is applicable whenever the binding/unbinding kinetics ensure a nearly uniform distribution of free particles. A systematic study of quantitative conditions for the validity of our approximation presents an important perspective of this work in the future. Meanwhile, Monte Carlo simulations that we realized in this paper indicate that the approximation is remarkably accurate when \(\eta=k_{\mathrm{off}}\langle\tau\rangle\) is not too small. As the limit \(\eta=0\) corresponds to irreversible binding (with either \(k_{\mathrm{off}}=0\), or \(\kappa=\infty\)), our approximation complements this well-studied setting and thus provides the overall insight onto diffusion-controlled reactions with multiple particles.
We also emphasize on the conceptual difference between our approach and the Lawley-Madrid approximation. The latter relied on the exponential approximation for the survival probability of a single particle, which is valid only for small and weakly reactive targets. This restriction concerns only binding events and does not involve unbinding kinetics. In turn, our approximation deals with the exact form of the survival probability, while the underlying assumption depends on binding/unbinding kinetics. As a consequence, it yields accurate results even for highly reactive targets, if the unbinding rate is not too small. In summary, the validity range of our approximationis different from that of the Lawley-Madrid approximation (see details in I), and it allows one to deal with highly reactive targets. At the same time, we outline that the LMA is much more explicit and easier to implement and to analyze, even in sophisticated geometric settings. Moreover, the LMA provides bounds to the first-crossing times for impatient particles. These two approximations present therefore valuable and complementary theoretical tools for studying diffusion-controlled reactions with reversible target-binding kinetics.
The present work can be extended in several directions. First, one can further analyze and possibly relax the assumption of uniform positions, beyond the discussion presented in C. This analysis can potentially lead to an exact solution of the general problem of impatient particles, which remains open for \(1<K<N\). Second, one can consider more sophisticated diffusive dynamics such as diffusing-diffusivity and switching models that allow one to incorporate dynamic heterogeneities of the medium or reversible binding to buffer molecules. Similarly, more elaborate target-binding mechanisms beyond that described by a constant reactivity \(\kappa\) can be investigated. For instance, one can consider encounter-dependent reactivity that may describe saturation effects after a number of reaction attempts that are relevant to some chemical or biological reactions. Moreover, one can incorporate surface diffusion in the bound state that was shown to enhance the overall reaction rate for a single particle. Finally, while the present paper focused on theoretical aspects of the problem of impatient particles, its application to relevant examples of diffusion-controlled events with multiple particles is a promising perspective. For this purpose, one needs further progress on the numerical implementation of our approximation to deal with a large number \(N\) of diffusing particles (e.g., several hundred of calcium ions). A large-\(N\) asymptotic analysis of the approximate solution would also be beneficial.
|
10.48550/arXiv.2202.07354
|
First-passage times of multiple diffusing particles with reversible target-binding kinetics
|
Denis S. Grebenkov, Aanjaneya Kumar
| 6,712
|
10.48550_arXiv.1009.4031
|
## Abstract
The analytical relations in position, momentum and four-dimensional spaces are established for the expansion and one-range addition theorems of relativistic complete orthonormal sets of exponential type spinor wave functions and Slater spinor orbitals of arbitrary half-integral spin. These theorems are expressed through the corresponding nonrelativistic expansion and one-range addition theorems of the spin-0 particles introduced by the author. The expansion and one-range addition theorems derived are especially useful for the computation of multicenter integrals over exponential type spinor orbitals arising in the generalized relativistic Dirac-Hartree-Fock-Roothaan theory when the position, momentum and four-dimensional spaces are employed.
## Key words:
Exponential type spinor orbitals, Slater type spinor orbitals, Addition theorems, Relativistic Dirac-Hartree-Fock-Roothaan theory
## 1.
The solutions of the Dirac equation for hydrogen-like systems play a significant role in theory and application to relativistic quantum mechanics of atoms, molecules and nuclei. However, the relativistic hydrogen-like position orbitals and their extensions to momentum and four-dimensional spaces cannot be used as basis sets because they are not complete unless the continuum is included. In Ref. we have constructed in position, momentum and four-dimensional spaces the complete orthonormal sets of two- and four-component relativistic spinor wave functions based on the use of complete othonormal sets of nonrelativistic orbitals. By the use of this method, in a previous work, we introduced the new complete orthonormal sets of relativistic \(\Psi^{as}\) -exponential type spinor orbitals ( \(\Psi^{as}\) - ETSO) and \(\mathrm{X}^{s}\) -Slater type spinor orbitals ( \(X^{s}\) -STSO) for particles with arbitrary half-integral spin in position, momentum and four-dimensional spaces through the correspondingnonrelativistic \(\psi^{\alpha}\) -exponential type orbitals (\(\psi^{\alpha}\) -ETO) and \(\chi\) -Slater type orbitals (\(\chi\) - STO). The elaboration of algorithm for the solution of generalized Dirac equations in linear combination of atomic spinor orbitals (LCASO) approach necessitates progress in the development of theory for one-range addition theorems of spinor orbitals of multiple order.
Addition theorems play a more and more important role in nonrelativistic and relativistic atomic and molecular electronic structure calculations. Two fundamentally different types of addition theorems occur in the literature. The first type of the addition theorems has the two-range form of Laplace expansion for the Coulomb potential. There is second class of addition theorems which can be constructed by expanding a function located at a center \(a\) in terms of a complete orthonormal set located at a center \(b\). The use of one-range addition theorems in electronic structure calculations would be highly desirable since they are capable of producing much better approximations than the two-range addition theorems. In Refs. we have developed the method for constructing in position, momentum and four-dimensional spaces the one-range addition theorems of complete orthonormal sets of nonrelativistic \(\psi^{\alpha}\) -ETO and \(\chi\) -STO. The aim of this work is to derive the relevant expansion and one-range addition theorems of complete orthonormal sets of relativistic \(\Psi^{as}\) -ETSO and X\({}^{s}\) -STSO in position, momentum and four-dimensional spaces through the corresponding theorems for nonrelativistic orbitals \(\psi^{\alpha}\) -ETO and \(\chi\) -STO. These theorems might be useful for the calculation of multicenter integrals which appear in relativistic MO LCASO theory of arbitrary half-integral spin particles when the spinor orbitals basis sets in position, momentum and four-dimensional spaces are employed.
## 2 Definitions and basic formulas
In order to derive the expansion and one-range addition theorems for 2(2s+1)-component spinor orbitals in position, momentum and four-dimensional spaces we use the following definitions:
Complete orthonormal sets of nonrelativistic orbitals
\[k^{\alpha}_{nlm}(\zeta,\vec{x}) = \psi^{\alpha}_{nlm}(\zeta,\vec{r}),\phi^{\alpha}_{nlm}(\zeta,\vec{ k}),z^{\alpha}_{nlm}(\zeta,\vec{\alpha}_{k}) \tag{1}\] \[\vec{k}^{\alpha}_{nlm}(\zeta,\vec{x}) = \vec{\psi}^{\alpha}_{nlm}(\zeta,\vec{r}),\vec{\phi}^{\alpha}_{nlm }(\zeta,\vec{k}),\vec{z}^{\alpha}_{nlm}(\zeta,\vec{\alpha}_{k})\,, \tag{2}\]Slater type nonrelativistic spinor orbitals
\[k_{nlm}(\zeta,\vec{x})=\chi_{nlm}(\zeta,\vec{r}),u_{nlm}(\zeta,\vec{k}),v_{nlm}( \zeta,\vec{\omega}_{k})\,, \tag{3}\]
Complete orthonormal sets of 2(2s+1)-component relativistic spinor orbitals
\[{}^{t}K^{\alpha\alpha}_{nlm_{j}}(\zeta,\vec{x})={}^{t}{}^{\nu}\Psi^{\alpha\alpha }_{nlm_{j}}(\zeta,\vec{r}),{}^{t}{}^{\rho}\Phi^{\alpha\alpha}_{nlm_{j}}(\zeta, \vec{k}),{}^{t}Z^{\alpha\alpha}_{nlm_{j}}(\zeta,\vec{\omega}_{k}) \tag{4a}\] \[{}^{t}K^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{x})={}^{t}\Psi^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{r}),{}^{t}\Phi^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{k}),{}^{t}Z^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{\omega}_{k})\] (4b) \[{}^{t}\vec{K}^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{x})={}^{t}\vec{\Psi}^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{r}),{}^{t}\vec{\Phi}^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{k}),{}^{t}Z^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{\omega}_{k})\] (5a) \[{}^{t}\vec{K}^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{x})={}^{t}\vec{\Psi}^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{r}),{}^{t}\vec{\Phi}^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{k}),{}^{t}\vec{Z}^{\alpha\alpha}_{nl_{j}}(\zeta,\vec{\omega}_{k})\,, \tag{5b}\]
Slater type 2(2s+1)-component relativistic spinor orbitals
\[{}^{t}K^{s}_{nlm_{j}}(\zeta,\vec{x})={}^{t}{}^{\chi}\mathbf{X}^{s}_{nlm_{j}}(\zeta,\vec{r}),{}^{t}U^{s}_{n(lm_{j}}(\zeta,\vec{k}),{}^{t}V^{s}_{n(lm_{j}}(\zeta,\vec{\omega}_{k}) \tag{6a}\] \[{}^{t}K^{s}_{nl_{j}}(\zeta,\vec{x})={}^{t}\mathbf{X}^{s}_{nl_{j}m_{j}}(\zeta,\vec{r}),{}^{t}U^{s}_{n(lm_{j}}(\zeta,\vec{k}),{}^{t}V^{s}_{n(lm_{j}}(\zeta,\vec{\omega}_{k})\,, \tag{6b}\]
See Refs. and for the exact definition of quantities occurring in Eqs-.
We shall also use the following formulas for 2(2s+1)-component spinor orbitals through the independent sets of two-component spinors defined as a product of complete orthonormal sets of radial parts of nonrelativistic scalar \(\psi^{\alpha}\) -ETO and modified Clebsch-Gordan coefficients appearing in two-component tensor spherical harmonics (see Refs. and):
\[\mbox{for }\,K^{\alpha\alpha}-ETSO\] \[{}^{t}K^{\alpha\alpha}_{nlm_{j}}(\zeta,\vec{x})=N_{nl_{j}}\left[{}^{t}K^{\alpha\alpha 0}_{nl_{j}}(\zeta,\vec{x})\right]\] \[{}^{t}K^{\alpha\alpha}_{nlm_{j}}(\zeta,\vec{x})=N_{nl_{j}}\left[{}^{t}K^{\alpha\alpha 2}_{nl_{j}}(\zeta,\vec{x})\right]\] \[{}^{t}K^{\alpha\alpha}_{nl_{j}m_{j}}(\zeta,\vec{x})=N_{nl_{j}}\left[{}^{t}K^{\alpha\alpha 2}_{nl_{j}m_{j}}(\zeta,\vec{x})\right]\] (7a) \[{}^{t\[{}^{t}K^{as\lambda}_{nlm_{s}}\left(\zeta,\vec{x}\right)=\begin{bmatrix}\eta_{t} \,^{t}a^{ls}_{jm_{s}}(\lambda)k^{\alpha}_{nlm(\lambda)}\left(\zeta,\vec{x}\right) \\ -\eta_{t}\,^{t}a^{ls}_{jm_{s}}(\lambda+1)k^{\alpha}_{nlm(\lambda+1)}\left(\zeta, \vec{x}\right)\end{bmatrix} \tag{7b}\]
\[{}^{t}K^{as\lambda}_{nlm_{s},m_{s}}\left(\zeta,\vec{x}\right)=\begin{bmatrix}- i\,^{t}a^{ls}_{jm_{s}}(2s-\lambda)k^{\alpha}_{nl,m(\lambda)}\left(\zeta,\vec{x} \right)\\ -i\,^{t}a^{ls}_{jm_{s}}(2s-(\lambda+1))k^{\alpha}_{nl,m(\lambda+1)}\left(\zeta,\vec{x}\right)\end{bmatrix} \tag{7c}\]
for \(\vec{K}^{as}-ETSO\)
\[{}^{t}\vec{K}^{as\lambda}_{nlm_{s}}\left(\zeta,\vec{x}\right)=N_{nl_{s}} \begin{bmatrix}\,^{t}\vec{K}^{as\lambda}_{nlm_{s}}\left(\zeta,\vec{x}\right) \\ \,^{t}\vec{K}^{as\lambda}_{nlm_{s}}\left(\zeta,\vec{x}\right)\\ \,^{t}\vec{K}^{as\lambda}_{nlm_{s}}\left(\zeta,\vec{x}\right)\\ \,^{t}\vec{K}^{as\lambda}_{nl,m_{s}}\left(\zeta,\vec{x}\right)\end{bmatrix} \tag{8a}\] \[{}^{t}\vec{K}^{as\lambda}_{nlm_{s}}\left(\zeta,\vec{x}\right)= \begin{bmatrix}\eta_{t}\,^{t}a^{ls}_{jm_{s}}(\lambda)\vec{k}^{\alpha}_{nlm (\lambda)}\left(\zeta,\vec{x}\right)\\ -\eta_{t}\,^{t}a^{ls}_{jm_{s}}(\lambda+1)\vec{k}^{\alpha}_{nlm(\lambda+1)} \left(\zeta,\vec{x}\right)\end{bmatrix} \tag{8b}\]
\[{}^{t}\vec{K}^{as\lambda}_{nlm_{s},m_{s}}\left(\zeta,\vec{x}\right)= \begin{bmatrix}-i\,^{t}a^{ls}_{jm_{s}}(2s-\lambda)\vec{k}^{\alpha}_{nl,m( \lambda)}\left(\zeta,\vec{x}\right)\\ -i\,^{t}a^{ls}_{jm_{s}}(2s-(\lambda+1))\vec{k}^{\alpha}_{nlm(\lambda+1)}\left( \zeta,\vec{x}\right)\end{bmatrix} \tag{8c}\]
for \(K^{s}-STSO\)
\[{}^{t}K^{s}_{nlm_{s}}\left(\zeta,\vec{x}\right)=N_{nl_{s}}\begin{bmatrix}\,^ {t}K^{s0}_{nlm_{s}}\left(\zeta,\vec{x}\right)\\ \,^{t}K^{s2}_{nlm_{s}}\left(\zeta,\vec{x}\right)\\ \,^{t}K^{s2,s-1}_{nlm_{s}}\left(\zeta,\vec{x}\right)\\ \,^{t}K^{s2}_{nl,m_{s}}\left(\zeta,\vec{x}\right)\\ \,^{t}K^{s2}_{nl,m_{s}}\left(\zeta,\vec{x}\right)\\ \,^{t}K^{s2}_{nl,m_{s}}\left(\zeta,\vec{x}\right)\\ \,^{t}K^{s0}_{nl,m_{s}}\left(\zeta,\vec{x}\right)\end{bmatrix} \tag{9a}\]\[{}^{t}K_{nlm_{j}}^{s\lambda}\left(\zeta,\vec{x}\right)=\begin{bmatrix}\eta_{t}\,^{ \iota}\,a_{jm_{j}}^{ls}\left(\lambda\right)k_{nlm\left(\lambda\right)}\left( \zeta,\vec{x}\right)\\ -\eta_{t}\,^{\iota}\,a_{jm_{j}}^{ls}\left(\lambda+1\right)k_{nlm\left(\lambda+1 \right)}\left(\zeta,\vec{x}\right)\end{bmatrix} \tag{9b}\]
\[{}^{t}K_{nl_{j}m_{j}}^{s\lambda}\left(\zeta,\vec{x}\right)=\begin{bmatrix}-i\,^ {\iota}\,a_{jm_{j}}^{ls}\left(2s-\lambda\right)k_{nl_{j}m_{j}}\left(\zeta,\vec {x}\right)\\ -i\,^{\iota}\,a_{jm_{j}}^{ls}\left(2s-\left(\lambda+1\right)\right)k_{nl_{j}m \left(\lambda+1\right)}\left(\zeta,\vec{x}\right)\end{bmatrix}, \tag{9c}\]
## 2. Expansion and one-range addition theorems for ETSO and STSO
With the derivation of expansion and one-range addition theorems for 2(2s+1)-component spinor orbitals in position, momentum and four-dimensional spaces, we use the method set out in previous papers described for the nonrelativistic cases. Then, using Eqs.- and carrying through calculations analogous to those for the nonrelativistic basis sets we obtain the following relations in terms of nonrelativistic cases:
## EXPANSION THEOREMS:
for ETSO
\[{}^{t}K_{nlm_{j}m_{j}}^{as^{\lambda}}\left(\zeta,\vec{x}\right)^{t}K_{n^{ \prime}l^{\prime}m_{j}}^{as}\left(\zeta^{\prime},\vec{x}\right)=\sum_{\lambda=0 }^{2s-1}\cdot\left[\,^{\iota\prime}F_{nlm_{j}m_{j}}^{as\lambda}\left(\zeta, \zeta^{\prime*};\vec{x}\right)+\,^{\iota\prime}\,\boldsymbol{F}_{nl_{j}m_{j}, n^{\prime}l^{\prime}m_{j}}^{as\lambda}\left(\zeta,\zeta^{\prime*};\vec{x} \right)\,\right] \tag{10a}\]
\[{}^{a^{\prime}}F_{nlm_{j}m_{j},n^{\prime}l^{\prime}m_{j}}^{as\lambda}\left( \zeta,\zeta^{\prime*};\vec{x}\right)=\eta_{t}\eta_{t^{\prime}}\left[\,^{\iota} a_{jm_{j}}^{ls}\left(\lambda\right)^{t^{\prime}}a_{jm_{j}}^{ls}\left(\lambda \right)k_{nlm\left(\lambda\right)}^{a^{*}}\left(\zeta,\vec{x}\right)k_{n^{ \prime}l^{\prime}m^{\prime}\left(\lambda\right)}^{\alpha}\left(\zeta^{\prime },\vec{x}\right)\,\right. \tag{10b}\]
\[{}^{a^{\prime}}\boldsymbol{F}_{nl_{j}m_{j},n^{\prime}l^{\prime}m_{j}}^{as\lambda }\left(\zeta,\zeta^{\prime*};\vec{x}\right)=\,^{\iota}a_{jm_{j}}^{ls}\left(2s -\lambda\right)\,^{\iota^{\prime}}a_{jm_{j}}^{l^{\prime}s}\left(2s-\lambda \right)k_{nl_{j}m\left(\lambda\right)}^{\alpha^{*}}\left(\zeta^{\prime},\vec{x} \right)k_{n^{\prime}l^{\prime}m^{\prime}\left(\lambda\right)}^{\alpha}\left( \zeta^{\prime},\vec{x}\right) \tag{10c}\] \[+\,^{\iota}\,a_{jm_{j}}^{l^{\prime}s}\left(2s-\left(\lambda+1\right)\right)\,^{ \iota^{\prime}}a_{jm_{j}}^{l^{\prime}s}\left(2s-\left(\lambda+1\right)\right)k_ {nl_{j}m\left(\lambda+1\right)}^{\alpha^{*}}\left(\zeta^{\prime},\vec{x} \right)k_{n^{\prime}l^{\prime}m^{\prime}\left(\lambda+1\right)}^{\alpha}\left( \zeta^{\prime},\vec{x}\right),\]
for STSO
\[{}^{t}K_{nl_{j}m_{j}}^{s^{\lambda}}\left(\zeta,\vec{x}\right)^{t}K_{n^{\prime}l ^{\prime}m_{j}}^{s}\left(\zeta^{\prime},\vec{x}\right)=\sum_{\lambda=0}^{2s-1} \cdot\left[\,^{\iota\prime}F_{nlm_{j},n^{\prime}l^{\prime}m_{j}}^{s\lambda} \left(\zeta,\zeta^{\prime*};\vec{x}\right)+\,^{\iota\prime}\,\boldsymbol{F}_{nl_{j} m_{j},n^{\prime}l^{\prime}m_{j}}^{s\lambda}\left(\zeta,\zeta^{\prime*};\vec{x} \right)\,\right] \tag{11a}\]
\[{}^{a^{\prime}}F_{nlm_{j}m_{j},n^{\prime}l^{\prime}m_{j}}^{s\lambda}\left(\zeta,\zeta^{\prime*};\vec{x}\right)=\eta_{t}\eta_{t^{\prime}}\left[\,^{\iota}a_{jm_{j}}^{ ks}\left(\lambda\right)^{t^{\prime}}a_{jm^{\prime}}^{l^{\prime}s}\left(\lambda \right)k_{nlm\left(\lambda\right)}^{*}\left(\zeta,\vec{x}\right)k_{n^{\prime} l^{\prime}m^{\prime}\left(\lambda\right)}\left(\zeta^{\prime},\vec{x}\right)\,\right. \tag{11b}\]\[\begin{split}&{}^{\alpha^{s}}\mathbf{F}_{n_{l},j_{m},\,\,\,\,\,\,\,\, \,\
|
10.48550/arXiv.1009.4031
|
Expansion and one-range addition theorems for complete orthonormal sets of spinor wave functions and Slater spinor orbitals of arbitrary half-integral spin in position, momentum and four-dimensional spaces
|
I. I. Guseinov
| 3,014
|
10.48550_arXiv.1010.2700
|
###### Abstract
It is shown that natural boundary conditions for non-relativistic wave functions are of periodic or of homogeneous Robin type. Using asymptotic central symmetry of Hamiltonian and theory of singular differential equations the many-electron wave function is expanded in series both in the vicinity of Coulomb singularities and at infinity. Hydrogenic angular dependence of three leading terms of expansion about Coulomb singularities is found. Exact first- and second-order cusp conditions are obtained demonstrating redundancy of spherical average in Kato's cusp condition. Our first-order cusp condition exhibits \(CP\) symmetry. Homogeneous Robin boundary conditions are obtained for aperiodic many-electron systems from the expansions. Use of our explicit boundary conditions improves both speed and accuracy of numerical calculations. A confluent hypergeometric series defining arbitrarily high order cusp conditions for the spherically averaged Hamiltonian is presented.
## 1 Introduction
Boundary conditions play important role in eigenvalue problems of mathematical physics even if they are imposed implicitly. Explicit use of boundary conditions is crucial to numerical calculations. The role of regularity and boundary conditions in _existence_ of quantum mechanical eigenvalue problems was first recognized by Schrodinger and von Neumann. Boundary and regularity conditions for the Schrodinger equation were first studied in detail by Jaffe who recognized that both \(\psi\) and its first derivatives should be bounded even at singular points of the potential. McCrea and Newing discussed boundary conditions for the \(H_{2}^{+}\) molecular ion.
In numerical calculations of many-electron systems, explicit boundary conditions should be imposed on the wave function at each singular point of the Hamiltonian, namely, at Coulomb singularities and at infinity The exact many-electron wave function satisfies boundary conditions not only at the nuclei but also at electron-electron coalescence points \(r_{ij}=0\) (\(i\neq j\)).
Asymptotic behaviour of many-electron wave functions was discussed in the framework of Hartree-Fock (HF) approximation by Handy, Marron and Silverstone and in the general case by Katriel and Davidson and by Patil. Vitanov and Panev extended the investigation to excited atomic states. The long-range behaviour of density was studied by Tal and by Levy, Perdewand Sahni. Ernzerhof, Burke and Perdew discussed the long-range behaviour of ground-state wave functions, one-electron density matrices, and pair densities including their angular dependence.
The study of many-electron wave functions about Coulomb singularities dates back to Lowdin who first recognized their cusp-like behaviour. He obtained, by analysing tabulated atomic HF data, hydrogenic \(r\)-dependence \(R_{n\ell}(r)=r^{\ell}f_{n\ell}(r)\propto r^{\ell}[1-Zr/(\ell+1)+\ldots]\) of radial wave functions in the vicinity of nucleus resulting in _cusp condition_
\[\left.\frac{f^{\prime}_{n\ell}}{f_{n\ell}}\right|_{r=0}=-\frac{Z}{\ell+1}. \tag{1}\]
Similar conditions were obtained both for electron-nucleus and electron-electron coalescences in the ground state of the \(He\) atom by Roothaan and Weiss by analysing singularities of Hamiltonian which were used as constraints on trial functions in their variational calculation in order to avoid divergence of local energy \(E_{loc}\equiv\Psi^{-1}\hat{H}\Psi\) at coalescence points \(r_{1}=0\), \(r_{2}=0\) and \(r_{12}=0\). They have introduced the _reduced mass_ of coalescent particles into the cusp condition (see quotation from p. 196 of Ref.):
Each cusp refers to a pair of Coulombic particles; the sign of the cusp is positive for a repulsion, negative for an attraction, and the magnitude is equal to the product of the two charges and the reduced mass of the two particles.
Roothaan, Sachs and Weiss used above cusp condition in variational calculation of light atoms and ions (up to \(Z=10\)). The first variational molecular calculation using the cusp constraint (for \(H_{2}\)) was performed by Kolos and Roothaan.
The near-nucleus behaviour of an arbitrary many-electron eigenfunction \(\Psi=\Psi({\bf r}|{\bf r}_{2},\ldots,{\bf r}_{N})\) was given by Kato's theorem \(IIb\)
\[\left.\frac{\partial\overline{\Psi}}{\partial r}\right|_{r=0}=-Z\Psi(r=0), \tag{2}\]
Steiner added a corollary to the theorem which relates the charge density and its derivative at the nucleus of an atom in a similar manner. M. Hoffmann-Ostenhof and Seiler generalized Kato's result to multiple coalescences; their fixed-nucleus approximation was removed by Johnson using Jacobi coordinates. Validity of Eq. is limited to the \(s\) states since wave function \(\psi\propto r^{\ell}\) has a root of multiplicity \(\ell\) at the origin. Unfortunately, Eq. and its generalizations cannot be used as exact boundary conditions due to the spherical average. It will be shown in section 4 that the _spherical average is redundant in Kato's cusp condition_ which was already indicated for above mentioned few-electron systems by Roothaan and his co-workers.
Results of our thorough and rigorous analysis differ from that of three highly-cited papers but are in accordance with that of Lowdin and of Roothaan and his co-workers. Pack and Brown added the phrase "all \(l,m\)" to Lowdin's cusp condition by solving many-electron Schrodingerequation in the vicinity of coalescence points which is not justified by our more rigorous solution. Bingel introduced an additional linear term proportional to the cosine of an unknown angle in the series expansion of wave function of which spherical average is always zero. In our rigorously derived expansion, all three leading terms exhibit the same hydrogenic angular dependence \(Y_{\ell m}(\vartheta,\varphi)\) as a result of isotropy of three leading terms of many-electron Hamiltonian in the vicinity of Coulomb singularities. Perhaps some recent research (e.g. Refs.) relying on these old results should be revised.
As a result of spherical average introduced by Kato the higher-order cusp conditions were derived only for the spherically averaged Hamiltonian by Rassolov and Chipman which were generalized to the excited states by Nagy and Sen. The near-nucleus anisotropy of Hamiltonian was first taken into consideration by Qian and Sahni, nevertheless omitting term \(\mathrm{O}\left(r^{0}\right)\). We will show in sections 3.3 and 4.2 that first and second order cusp conditions are exact without spherical averaging and only that of third and of higher orders need spherical average. In addition, we present a confluent hypergeometric series in section 3.4 defining arbitrarily high order cusp conditions for the spherically averaged Hamiltonian.
Fournais, M. and T. Hoffmann-Ostenhof and Sorensen recently decomposed the many-electron wave function as \(\psi(\mathbf{x})=\psi^{}(\mathbf{x})+\left|x_{1}\right|\psi^{}(\mathbf{ x})\) in the neighbourhood of Coulomb singularities using inconvenient Cartesian coordinates \(\mathbf{x}=(x_{1},\ldots,x_{N})\in\mathbb{R}^{3N}\) inhibiting the recognition of asymptotic central symmetry of Hamiltonian. In addition, their component functions \(\psi^{}\) and \(\psi^{}\) are real-valued. Our decomposition is slightly different due to the isotropy of singularity taken into consideration: the first term is an eigenfunction of the spherically averaged Hamiltonian, the second term reflects the anisotropy of molecular or crystalline potential.
Since Kato's rigorously proved spherically averaged cusp condition contradicts Eq. which was successfully used in few-electron calculations by Roothaan and his co-workers, a different but similarly rigorous treatment is needed to decide between the two forms. The primary aim of this paper is to derive boundary conditions by solving many-electron Schrodinger equation both at large distances and in the neighbourhood of Coulomb singularities using a well-tried formalism of mathematical physics, the theory of singular differential equations. The key idea of this paper is that due to _asymptotic isotropy_1 of many-electron Hamiltonian both at Coulomb singularities and as \(r\rightarrow\infty\) the angular momentum asymptotically commutes with the Hamiltonian leading to asymptotically hydrogen-like wave functions with definite values of quantum numbers \(\ell\) and \(m\).
Footnote 1: Term _asymptotic equality_ is used in this paper in the sense of \(\lim_{x\to x_{0}}\left[f(x)/g(x)\right]=1\). Isotropic and anisotropic terms of expansions are distinguished throughout this paper so a boldface argument of _ordo_ symbol \(\mathrm{O}\left(\mathbf{r}^{n}\right)\) indicates the anisotropy of omitted terms of expansions.
The outline of the paper is as follows. In section 2, the allowed boundary conditions of non-relativistic one-particle quantum mechanics are studied. In section 3.1, the isotropy of singularities of many-electron Hamiltonian is discussed. In section 3.2, the many-electron wave function is expanded at large distances in terms of irregular solid spherical harmonics. At the beginning of section 3.3, the many-electron Hamiltonian is studied in the vicinity of Coulomb singularities, where explicit form of potential contribution of non coalescent particles is obtained. Then it is shown that due to isotropy of leading terms of Hamiltonian, all three leading terms of many-electron wave function exhibit hydrogenic angular dependence, therefore first and second order cusp conditions (21a, 21b) are exact without spherical averaging. The section ends with the decomposition of many-electron wave function as the sum of two functions, where the first term is an eigenfunction of the spherically averaged Hamiltonian, the second term reflects the anisotropy of molecular or crystalline potential. In section 3.4, eigenfunctions of the spherically averaged Hamiltonian are discussed leading to recurrence relations (26a, 26b) defining cusp conditions of arbitrarily high orders for this special case. In section 4.2, exact first and second order cusp conditions (30, 31a, 31b, 32) are presented demonstrating the redundancy of spherical average in Kato's cusp condition. As a result of asymptotic hydrogenic behaviour of many-electron wave function both at Coulomb singularities and as \(r\to\infty\) we obtain boundary conditions of homogeneous Robin type similarly to Eqs. and derived in section 2 for the \(H\) atom. In section 5, physical and numerical consequences of our results are discussed.
## 2 Boundary conditions in one-particle quantum mechanics
The problem of hydrogen atom was solved by Schrodinger in his historically famous paper without imposing explicit boundary conditions on the wave function. In the first version of manuscript, the requirement of stationarity of current flux was used as a constraint on the variational problem which was changed to the weaker normalization condition by an addendum (cf. equations 6 and 24 of Ref.). Latter form is more conventional mathematically since it is compatible with Sturm-Liouville theory of eigenvalue equations, where \(\int\left|\psi\right|^{2}dv\) represents the denominator of Rayleigh quotient. Von Neumann has concluded that normalization condition for the wave function and requirement of self-adjointness of the Hamiltonian is equivalent to imposing both boundary and regularity conditions on the wave function.
In fact, these conditions are too weak to enforce unique regular solutions of Schrodinger equation. The normalization condition does not exclude irregular particular solution \(\psi\propto r^{-\ell-1}\) for \(s\) states of the Coulomb problem hence it is excluded by hand both in Schrodinger's paper and in the textbooks. The requirement of self-adjointness does not lead to a unique eigenvalue problem since a little-known theorem of Sturm-Liouville theory of differential equations states that _any_ of following two types of boundary conditions are consistent with self-adjointness of the Liouville operator:
1. _periodic_ boundary conditions \[\psi(a)-\psi(b) = 0,\] (3a) \[\psi^{\prime}(a)-\psi^{\prime}(b) = 0,\] (3b)
2. _homogeneous Robin_ boundary conditions \[\alpha_{1}\psi^{\prime}(a)+\beta_{1}\psi(a) = 0,\] (4a) \[\alpha_{2}\psi^{\prime}(b)+\beta_{2}\psi(b) = 0,\] (4b)where \(\alpha\)'s and \(\beta\)'s are real constants, \(a\) and \(b\) denote endpoints of the interval 2. The theorem can be generalized to partial Sturm-Liouville equations by taking function values and normal derivatives over hypersurfaces of the domain (the proof is based on the definition of self-adjointness and the use of Green's theorem).
Footnote 2: The natural boundary condition for the momentum operator \(-\mathrm{i}\frac{\mathrm{d}}{\mathrm{d}x}\), discussed by von Neumman as an example, is simply (3a). Widely used substitution \(\psi=u(r)/r\) eliminates not only the first derivative from the Laplacian but also the periodic boundary conditions since \(r\) equals the square root of coefficient \(p=r^{2}\) of Liouville operator \(\frac{\mathrm{d}}{\mathrm{d}r}\left[p(r)\frac{\mathrm{d}}{\mathrm{d}r}\right]+ q(r)\), i.e. only one branch of \(\pm\sqrt{p}\) is used which illustrates the importance of single-valuedness of coefficients in the theory of differential equations. In \(d\)-dimensional hyperspherical coordinates, the substitution eliminating the first derivative is of the form \(\psi=u(r)/r^{(d-1)/2}\), where the square root appears explicitly.
Equations (3a, 3b) are known as Born - von Karman or Bloch boundary conditions of solid state physics. In view of above and Bloch's theorem we can state that _eigenfunctions of aperiodic systems satisfy homogeneous Robin boundary conditions_. Homogeneous Dirichlet and Neumann boundary conditions for model problems of the textbooks are special cases of (4a, 4b). Boundary conditions (4a, 4b) can be divided by arbitrary constants so coefficients \(\alpha_{i}/\sqrt{\alpha_{i}^{2}+\beta_{i}^{2}}\equiv\sin\gamma_{i}\) and \(\beta_{i}/\sqrt{\alpha_{i}^{2}+\beta_{i}^{2}}\equiv\cos\gamma_{i}\) define angles \(\gamma_{1}\) and \(\gamma_{2}\) representing the boundaries. Random coefficients for amorphous materials, in the Wannier representation, may be interpreted as random walk of these "phase points" around the unit circle which leads to a band structure similarly to the periodic boundary conditions.
As an example of Eqs. (4a, 4b) let us recover hidden boundary conditions for a non-relativistic \(H\)-like ion with nuclear charge \(Z\) using known properties of hydrogenic bound-state wave functions. The normalization condition guarantees a part of boundary conditions, namely, vanishing at infinity.
\[\lim_{r\to\infty}\frac{1}{\psi}\frac{\partial\psi}{\partial r}=-\sqrt{-2E}, \tag{5}\]
The normalization condition results in a mild singularity (hyperconical cusp) of the wave function: whereas it is continuous everywhere, its directional derivatives are bounded but discontinuous at the Coulomb singularity.
\[\frac{\mathbf{e}\cdot\nabla\psi}{\psi}=\frac{\mathbf{e}\cdot\mathbf{e}_{r}}{R _{n\ell}}\frac{\mathrm{d}R_{n\ell}}{\mathrm{d}\,r}+\frac{\mathbf{e}\cdot \mathbf{e}_{\vartheta}}{Y_{\ell m}\,r}\frac{\partial Y_{\ell m}}{\partial \vartheta}+\frac{\mathbf{e}\cdot\mathbf{e}_{\varphi}}{Y_{\ell m}\,r\sin \vartheta}\,\frac{\partial Y_{\ell m}}{\partial\varphi},\]
In radial directions defined by \(\mathbf{e}\cdot\mathbf{e}_{r}=\pm 1\) and \(\mathbf{e}\cdot\mathbf{e}_{\vartheta}=\mathbf{e}\cdot\mathbf{e}_{\varphi}=0\), above expression reduces to
\[\frac{\mathbf{e}\cdot\nabla\psi}{\psi}=\frac{\pm 1}{R_{n\ell}}\frac{\mathrm{d}R_{ n\ell}}{\mathrm{d}\,r}=\frac{\pm 1}{\psi}\frac{\partial\psi}{\partial r}.\]
Since radial wave function \(R_{n\ell}=r^{\ell}u_{n\ell}(r)\) of the central field problem has a root of multiplicity \(\ell\) at the origin the l'Hospital rule should be applied \(\ell\) times in order to obtain a definite limit
\[\lim_{r\to 0}\frac{1}{R_{n\ell}}\frac{\mathrm{d}R_{n\ell}}{\mathrm{d}\,r}=\lim_{r \to 0}\frac{R_{n\ell}^{(\ell+1)}}{R_{n\ell}^{(\ell)}}=(\ell+1)\frac{u_{n\ell} ^{\prime}}{u_{n\ell}}=-Z,\]
The discontinuity of the radial logarithmic derivative at the nucleus is then characterized by
Footnote 3: The power series expansion of the hydrogenic radial wave function about the nucleus is of the form \(R_{n\ell}=C_{n\ell}\,r^{\ell}\left[1-\frac{Zr}{\ell+1}+\frac{2n^{2}+\ell+1}{2( \ell+1)(2\ell+3)}\left(\frac{Zr}{n}\right)^{2}+\ldots\right]\).
\[\lim_{\mathbf{r}\rightarrow\pm 0}\frac{1}{\psi}\frac{\partial\psi}{ \partial r}=\lim_{\mathbf{r}\rightarrow\pm 0}\frac{\partial_{r}^{\ell+1}\psi}{ \partial_{r}^{\ell}\psi}=\mp Z, \tag{6}\]
where
\[\mathbf{r}\rightarrow+0 \equiv (r\to 0,\,\vartheta,\,\varphi),\] \[\mathbf{r}\rightarrow-0 \equiv (r\to 0,\,\pi-\vartheta,\,\pi+\varphi).\]
We emphasize that all steps of above derivation, except the last one, rely on isotropy of singularity of Hamiltonian at \(r=0\). It is interesting to observe the \(CP\) invariance of cusp relation. Pair of and obviously represent homogeneous Robin boundary conditions of the form (4a, 4b).
Similar boundary conditions will be obtained for many-electron wave functions in section 4 as a result of asymptotic central symmetry of many-electron Hamiltonian both in the vicinity of nuclei and at large distances.
## 3 Behaviour of many-electron wave function at singular points of Hamiltonian
Let us consider non-relativistic Hamiltonian describing \(N\) particles interacting with each other by Coulomb potentials
\[\hat{H}=-\frac{1}{2}\sum_{i=1}^{N}\frac{\Delta_{i}}{m_{i}}+\sum_{i=1}^{N-1}\sum _{j=i+1}^{N}\frac{q_{i}q_{j}}{r_{ij}}, \tag{7}\]
Electrons and nuclei will be distinguished only in the final results. Spin coordinates are omitted for simplicity.
\[\hat{H}\Psi=E\Psi \tag{8}\]
Singularities of the Coulomb potential are removable and singularity of the kinetic energy at infinity is essential. The singularities and other leading terms of the Hamiltonian are isotropic in both limiting cases. In the vicinity of coalescence points, the singular potential contribution of coalescent particles \(\mathrm{O}\left(r^{-1}\right)\) and bounded leading term of potential due to remaining non-coalescent particles \(\mathrm{O}\left(r^{0}\right)\) are isotropic. At infinity, only Coulombic monopole term \(\mathrm{O}\left(r^{-1}\right)\) is isotropic, whereas multipole terms \(\mathrm{O}\left(\mathbf{r}^{-2}\right)\) are anisotropic.
### Isotropy of singularities
Particular solutions of a linear ordinary differential equation in the neighbourhood of an isolated singular point \(x_{0}=0\) are of product form \(y=f(x)g(x)\), where \(f(x)\) ensures the correct behaviour of \(y\) at the singularity and \(g(x)\) is a single-valued analytic function (or has at most a logarithmic singularity) which is non-zero at the singular point. Function \(f(x)\) is typically power, exponential or Gaussian function. For _isotropic singularities_ of a linear partial differential equation, the solution has the form \(y=f(r)g(\mathbf{r})\) with \(\mathbf{r}\in\mathbb{R}^{n}\) and \(r=|\mathbf{r}|\), where \(f(r)\) is responsible for correct behaviour of \(y\) about the singular point and \(g(\mathbf{r})\) (reflecting anisotropy of the coefficient functions) is non-zero at the singular point. (One may expand it in 3 dimensions in terms of regular \(r^{\lambda}Y_{\lambda\mu}(\vartheta,\varphi)\) or irregular \(r^{-\lambda-1}Y_{\lambda\mu}(\vartheta,\varphi)\) solid spherical harmonics.) It will be shown in this section that due to isotropy of singularities of Hamiltonian, the wave function about singular points has the following limiting forms
\[\Psi\underset{\mathbf{r}\rightarrow\mathbf{r}_{i}=0}{\longrightarrow}r^{\ell_ {i}}u_{i}(\mathbf{r}),\quad u_{i}\neq 0,\;i=1,2,\ldots,N,\]
\[\Psi\underset{r\rightarrow\infty}{\longrightarrow}e^{\alpha r}r^{\beta}v( \mathbf{r}),\quad\lim_{r\rightarrow\infty}v(\mathbf{r})\neq 0,\;\alpha<0.\]
In many-body systems, in contrast to the central-field problem, only energy \(E\) is conserved throughout the configuration space, square \(\mathbf{L}^{2}\) and projection \(L_{z}\) of angular momentum are conserved only at the singularities
\[[\hat{H},\,\hat{\mathbf{L}}^{2}]\underset{\mathbf{r}\rightarrow \mathbf{r}_{i}}{\longrightarrow}0, [\hat{H},\,\hat{L}_{z}]\underset{\mathbf{r}\rightarrow\mathbf{r}_{i}}{ \longrightarrow}0, \tag{9a}\] \[[\hat{H},\,\hat{\mathbf{L}}^{2}]\underset{r\rightarrow\infty}{ \longrightarrow}0, [\hat{H},\,\hat{L}_{z}]\underset{r\rightarrow\infty}{ \longrightarrow}0, \tag{9b}\]
In the neighbourhood of singular points angular momentum quantum number \(\ell\) and magnetic quantum number \(m\) have definite values. Eigenfunctions of many-electron Hamiltonian approach eigenfunctions of angular momentum when approaching singularities of the Hamiltonian leading to _asymptotic hydrogenic angular dependence_
\[\Psi\underset{\mathbf{r}\rightarrow\mathbf{r}_{i}=0}{\longrightarrow}R_{i}(r )Y_{\ell_{i}m_{i}}(\vartheta,\varphi), \tag{10a}\] \[\Psi\underset{r\rightarrow\infty}{\longrightarrow}R_{\infty}(r)Y_ {00} \tag{10b}\]
In other words, molecular symmetries manifest only at molecular distances, where many-body Hamiltonian does not commute with angular momentum.
Due to asymptotic \(H\)-like behaviour of many-electron wave functions the equations of this section and their solution methods are similar to the mathematically correct original treatment of \(H\) atom by Schrodinger and Weyl (see acknowledgement in footnote on p. 363 of Ref.) based on theory of singular differential equations which differs from the simplistic textbook-derivations. The Coulomb singularities are treated using Fuchs' theorem exactly the same way as in Ref.. The singularity at \(\infty\) is treated using Hamburger's theorem which is shorter than Schrodinger's solution based on the Laplace transform. Of course, our equations lead to wave functions of the \(H\) atom in the special case of two Coulombic particles.
### Asymptotic behaviour at infinity
Let us consider an electron, say particle 1, separated from the rest of the system (\(m_{1}=1\), \(q_{1}=-1\) and \(r_{1}>r_{2},\ldots,r_{N}\)). Let us introduce reduced mass \(M^{\prime}\) and center of mass \(\mathbf{R}^{\prime}\) of the whole system, center of mass \(\mathbf{R}^{\prime\prime}\) of particles except the electron located at \(\mathbf{r}_{1}\) and their separation \(\mathbf{r}\):
\[\frac{1}{M^{\prime}}\equiv\sum_{i=1}^{N}\frac{1}{m_{i}},\quad \mathbf{R}^{\prime}\equiv\frac{\sum_{i=1}^{N}m_{i}\mathbf{r}_{i}}{\sum_{i=1}^{ N}m_{i}},\quad\mathbf{R}^{\prime\prime}\equiv\frac{\sum_{i=2}^{N}m_{i}\mathbf{r }_{i}}{\sum_{i=2}^{N}m_{i}},\] \[\mathbf{r}\equiv\mathbf{r}_{1}-\mathbf{R}^{\prime\prime}\equiv(r,\vartheta,\varphi)\equiv(r,\omega).\]
Using Laplace expansion for \(r>r_{i}\)
\[\frac{1}{|\mathbf{r}-\mathbf{r}_{i}|}=\sum_{\lambda=0}^{\infty}\sum_{\mu=- \lambda}^{\lambda}\frac{4\pi}{2\lambda+1}\frac{r_{i}^{\lambda}}{r^{\lambda+1} }Y_{\lambda\mu}^{*}(\omega)Y_{\lambda\mu}(\omega_{i})\]
the potential energy of the system can be expressed at large distances as
\[U(\mathbf{r}) = \sum_{\lambda,\mu}\frac{4\pi q_{1}}{2\lambda+1}\frac{Y_{\lambda \mu}^{*}(\omega)}{r^{\lambda+1}}\sum_{i=2}^{N}q_{i}r_{i}^{\lambda}Y_{\lambda\mu }(\omega_{i})\] \[= \frac{q_{1}}{r}\sum_{i=2}^{N}q_{i}+\mathrm{O}\left(\mathbf{r}^{-2 }\right)=-\frac{Q+1}{r}+\mathrm{O}\left(\mathbf{r}^{-2}\right),\]
Due to asymptotic isotropy (9b) of Hamiltonian the many-particle Schrodinger equation is asymptotically separable (10b) in terms of spherical polar coordinates as \(r\rightarrow\infty\). Asymptotic radial wave function \(R=R\left(r|\mathbf{r}_{2},\ldots,\mathbf{r}_{N}\right)\) satisfies differential equation 4
Footnote 4: Centrifugal kinetic energy term is omitted since it has the same order as omitted anisotropic dipole term of the potential energy.
\[\left[-\frac{\Delta_{r}}{2M^{\prime}}-\frac{Q+1}{r}-E+\mathrm{O}\left(\mathbf{ r}^{-2}\right)\right]R\mathop{\longrightarrow}_{r\rightarrow\infty}0, \tag{11}\]
The equation has an isolated _essential_ singularity at infinity since transformation of variable \(z\equiv 1/r\) leads to second order differential equation
\[z^{4}R^{\prime\prime}(z)+2M^{\prime}\left[E+(Q+1)z+\mathrm{O}\left(\mathbf{z}^ {2}\right)\right]R(z)\mathop{\longrightarrow}_{z\to 0}0\]
The pole being independent of potential is a consequence of the Laplacian.
\[\frac{\mathrm{d}^{2}R}{\mathrm{d}\,r^{2}}+\left(a_{0}+\frac{a_{1}}{r}+\frac{a _{2}}{r^{2}}+\ldots\right)\frac{\mathrm{d}R}{\mathrm{d}\,r}+\left(b_{0}+\frac {b_{1}}{r}+\frac{b_{2}}{r^{2}}+\ldots\right)R=0\]
with
\[a_{1}=2,\quad a_{0}=a_{2}=a_{3}=\ldots=0,\quad b_{0}=2M^{\prime}E,\quad b_{1}= 2M^{\prime}(Q+1),\]\(b_{2}\) and higher order coefficients are anisotropic. Due to essential singularity of the equation the solution vanishes transcendentally as \(r\to\infty\).
\[R=\mathrm{e}^{\alpha r}u(r),\quad\lim_{r\to\infty}u(r)\neq 0\]
we obtain
\[\frac{\mathrm{d}^{2}u}{\mathrm{d}r^{2}}+2\left(\alpha+\frac{1}{r}\right)\frac{ \mathrm{d}u}{\mathrm{d}r}+\left[\alpha^{2}+b_{0}+\frac{2\alpha+b_{1}}{r}+ \mathrm{O}\left(\mathbf{r}^{-2}\right)\right]u\underset{r\to\infty}{ \longrightarrow}0.\]
By equating leading term of coefficient of \(u(r)\) to zero we obtain _indicial equation_ of which roots are
\[\alpha=\pm\sqrt{-b_{0}},\]
Non-essential singularity of the latter differential equation can be removed by substitution
\[u=r^{\beta}v(r),\quad\lim_{r\to\infty}v(r)\neq 0\]
yielding
\[\frac{\mathrm{d}^{2}v}{\mathrm{d}r^{2}}+2\left(\alpha+\frac{\beta+1}{r}\right) \frac{\mathrm{d}v}{\mathrm{d}r}+\left[\frac{2\alpha(\beta+1)+b_{1}}{r}+ \mathrm{O}\left(\mathbf{r}^{-2}\right)\right]v\underset{r\to\infty}{ \longrightarrow}0.\]
By equating leading term of coefficient of \(v(r)\) to zero we obtain indicial equation with only root
\[\beta=-\frac{b_{1}}{2\alpha}-1=\frac{b_{1}}{2\sqrt{-b_{0}}}-1.\]
In case of a central-symmetric problem the solution of latter differential equation would be of the form
\[v=v_{0}+v_{1}r^{-1}+v_{2}r^{-2}+\ldots,\quad(v_{0}\neq 0)\]
The radial wave function at large distances is then
\[R\underset{r\to\infty}{\longrightarrow}\mathrm{e}^{-\sqrt{-2M^{\prime}E}}r^{ \frac{M^{\prime}(Q+1)}{\sqrt{-2M^{\prime}E}}-1}\left[v_{0}+\mathrm{O}\left( \mathbf{r}^{-1}\right)\right]. \tag{12}\]
This behaviour is in accordance with results of Katriel and Davidson who derived it less rigorously in two different ways.
\[\Psi=\mathrm{e}^{-\sqrt{-2M^{\prime}E}}r^{\frac{M^{\prime}(Q+1)}{\sqrt{-2M^{ \prime}E}}}\sum_{\lambda,\mu}\frac{v_{\lambda\mu}\left(r^{-1}|\mathbf{r}_{2}, \ldots,\mathbf{r}_{N}\right)}{r^{\lambda+1}}Y_{\lambda\mu}(\vartheta,\varphi), \tag{13}\]
Molecular or crystalline symmetries are reflected by relations between coefficient functions \(v_{\lambda\mu}\left(r^{-1}|\mathbf{r}_{2},\ldots\mathbf{r}_{N}\right)\) for \(\lambda>0\).
### Local behaviour at Coulomb singularities
Let us focus now our attention on coalescence of any two particles, say 1 and 2, while keeping remaining particles separated from them: \(r_{1},\,r_{2}<r_{3},\ldots,r_{N}\)5. In order to explore symmetry properties of the Hamiltonian it is convenient to use Jacobi coordinates by introducing reduced mass \(M\), center of mass \({\bf R}\) and separation \({\bf r}\) of these two particles:
Footnote 5: Other authors e.g. of Refs. assume well separated particles \(r_{1},\,r_{2}\ll r_{3},\ldots,r_{N}\) which is unnecessary.
\[M\equiv\frac{m_{1}m_{2}}{m_{1}+m_{2}},\quad{\bf R}\equiv\frac{m_ {1}{\bf r}_{1}+m_{2}{\bf r}_{2}}{m_{1}+m_{2}},\] \[{\bf r}\equiv{\bf r}_{1}-{\bf r}_{2}\equiv(r,\vartheta,\varphi) \equiv(r,\omega).\]
Many-body Hamiltonian can be partitioned as
\[\hat{H}=-\frac{\Delta}{2M}+\frac{q_{1}q_{2}}{r}+\hat{W}+\hat{G}, \tag{14}\]
where
\[\hat{W} \equiv \sum_{i=3}^{N}\left(\frac{q_{1}}{r_{1i}}+\frac{q_{2}}{r_{2i}} \right)q_{i},\] \[\hat{G} \equiv -\frac{\Delta_{\bf R}}{2(m_{1}+m_{2})}-\sum_{i=3}^{N}\frac{ \Delta_{i}}{2m_{i}}+\sum_{i=3}^{N-1}\sum_{j=i+1}^{N}\frac{q_{i}q_{j}}{r_{ij}}.\]
Use of Laplace expansion for \(r<r^{\prime}\)
\[\frac{1}{|{\bf r}-{\bf r}^{\prime}|}=\sum_{\lambda=0}^{\infty}\sum_{\mu=- \lambda}^{\lambda}\frac{4\pi}{2\lambda+1}\frac{r^{\lambda}}{r^{\prime\lambda+1 }}Y_{\lambda\mu}^{*}(\omega)Y_{\lambda\mu}(\omega^{\prime})\]
yields
\[W= \sum_{\lambda,\mu}\frac{4\pi w_{\lambda\mu}}{2\lambda+1}\left[q_{1 }r_{1}^{\lambda}Y_{\lambda\mu}^{*}(\omega_{1})+q_{2}r_{2}^{\lambda}Y_{\lambda \mu}^{*}(\omega_{2})\right],\] \[w_{\lambda\mu}\equiv w_{\lambda\mu}\left({\bf r}_{3},\ldots,{\bf r }_{N}\right)\equiv\sum_{i=3}^{N}\frac{q_{i}}{r_{i}^{\lambda+1}}Y_{\lambda\mu} (\omega_{i}).\]
Expressing \({\bf r}_{1}\) and \({\bf r}_{2}\) with Jacobi coordinates \({\bf r}\) and \({\bf R}\), putting origin of the coordinate system to center of mass \({\bf R}\), using inversion property
\[Y_{\lambda\mu}(\pi-\vartheta,\,\pi+\varphi)=(-1)^{\lambda}\,Y_{\lambda\mu}( \vartheta,\varphi)\]
and addition theorem
\[\frac{4\pi}{2\lambda+1}\sum_{\mu=-\lambda}^{\lambda}Y_{\lambda\mu }(\vartheta_{i},\varphi_{i})Y_{\lambda\mu}^{*}(\vartheta,\varphi)=P_{\lambda}( \cos\gamma_{i}),\] \[\cos\gamma_{i}\equiv\cos\vartheta\cos\vartheta_{i}+\sin\vartheta \sin\vartheta_{i}\cos(\varphi-\varphi_{i})\]we obtain
\[W = \sum_{\lambda=0}^{\infty}\left[\frac{q_{1}}{m_{1}^{\lambda}}+(-1)^{ \lambda}\frac{q_{2}}{m_{2}^{\lambda}}\right]\left(Mr\right)^{\lambda}\sum_{i=3}^{ N}\frac{q_{i}}{r_{i}^{\lambda+1}}P_{\lambda}(\cos\gamma_{i}) \tag{15}\] \[= W_{0}(r_{3},\ldots,r_{N})+W_{1}(\mathbf{r}/r,\mathbf{r}_{3}, \ldots,\mathbf{r}_{N})\,r+\ldots\]
_For identical coalescent particles, all odd powers of separation vanish_ in the vicinity of the coalescence point due to the inversion symmetry. This property has a profound consequence in the behaviour of electron-electron potentials which will be discussed in a separate paper.
\[W_{0}\equiv W_{0}\left(r_{3},\ldots,r_{N}\right)\equiv(q_{1}+q_{2})\sum_{i=3} ^{N}\frac{q_{i}}{r_{i}} \tag{16}\]
In practical calculations, the terms of above sum should be evaluated as expectation values \(\left\langle\Psi\left|q_{i}/r_{i}\right|\Psi\right\rangle\), however the fixed-nucleus approximation leaves them in their original form for the nuclei.
\[\overline{W}=W_{0} \tag{17}\]
Consequences of this property will be discussed in detail in section 3.4. illustrates the behaviour of potential \(W\) in the vicinity of a Coulomb singularity.
Since term \(\hat{W}\) of Hamiltonian acts on separation \(\mathbf{r}\) of particles 1 and 2 and term \(\hat{G}\) acts on their center of mass \(\mathbf{R}\) the many-particle Schrodinger equation is separable resulting in an effective one-body problem.
\[\left[-\frac{\Delta}{2M}+\frac{q_{1}q_{2}}{r}+W_{0}-E+\mathrm{O}\left(\mathbf{ r}\right)\right]\Psi\mathop{\longrightarrow}\limits_{r\to 0}0\]
Due to local isotropy (9a) of Hamiltonian the equation is locally separable in terms of spherical polar coordinates and the wave function exhibits local hydrogenic angular dependence of the form (10a) about the origin.
\[\left\{\frac{-1}{2Mr^{2}}\left[\frac{\mathrm{d}}{\mathrm{d}r}\left(r^{2}\frac {\mathrm{d}}{\mathrm{d}r}\right)-\ell(\ell+1)\right]+\frac{q_{1}q_{2}}{r}+W_{ 0}-E+\mathrm{O}\left(\mathbf{r}\right)\right\}R\mathop{\longrightarrow} \limits_{r\to 0}0\]
Frobenius normal form of this equation is
\[r^{2}R^{\prime\prime}+rP(r)R^{\prime}+Q(r)R=0,\]
where
\[P\equiv 2,\quad Q\equiv-\ell(\ell+1)-2Mq_{1}q_{2}r-2M(W_{0}-E)r^{2}+\mathrm{O} \left(\mathbf{r}^{3}\right).\]The singular point is _removable_ since \(P(r)\) and \(Q(r)\) are single-valued analytic functions. _Fuchs' theorem_ states that in the neighbourhood of removable singularities, the fundamental system of solutions is
\[R_{1} = r^{\lambda_{1}}u(r),\] \[R_{2} = r^{\lambda_{2}}\left[v(r)+\alpha\,u(r)\ln r\right],\]
The theorem distinguishes three cases for existence of logarithmic term of the second solution depending on difference \(\lambda_{1}-\lambda_{2}\). In order to determine exponents \(\lambda_{1}\) and \(\lambda_{2}\) we seek the solution in the form \(r^{\lambda}u(r)\). The substitution gives _indicial equation_
Footnote 6: Transformation of \(r\) to a dimensionless variable required by the logarithmic term is omitted for brevity.
\[\left[\lambda(\lambda-1)+2\lambda-\ell(\ell+1)\right]u(r)+\mathrm{O}\left(r \right)\mathop{\longrightarrow}\limits_{r\to 0}0\]
Since difference \(\lambda_{1}-\lambda_{2}=2\ell+1\) is a non-zero integer there is no general rule for existence of logarithmic term hence value of \(\alpha\) should be determined individually by substituting \(R_{2}\) into the differential equation leading to
\[\alpha\mathop{\longrightarrow}\limits_{r\to 0}-r\frac{2Mq_{1}q_{2}v(r)+2\ell v ^{\prime}(r)+\mathrm{O}\left(r\right)}{(2\ell+1)u(r)+\mathrm{O}\left(r\right)} \mathop{\longrightarrow}\limits_{r\to 0}0,\]
the logarithmic term of the second solution vanishes at the origin hence the fundamental solutions are simply \(R_{1}=r^{\ell}u(r)\) and \(R_{2}=r^{-\ell-1}v(r)\).
Illustrative example of a linear chain of homo-nuclear atoms showing the behaviour of potential term \(W\) of Hamiltonian in the vicinity of an electron-nucleus coalescence point \(x=0\).
unbounded at \(r=0\) the physical solution is
\[R\mathop{\longrightarrow}\limits_{r\to 0}r^{\ell}u(r). \tag{18}\]
Substituting this expression into the radial equation we obtain differential equation
\[u^{\prime\prime}+\frac{2\ell+2}{r}u^{\prime}-2M\left[\frac{q_{1}q_{2}}{r}+W_{0} -E+\mathrm{O}\left(\mathbf{r}\right)\right]u\mathop{\longrightarrow}\limits_{r \to 0}0 \tag{19}\]
of which solution should be analytic according to Fuchs' theorem hence it can be expanded in power series
\[u(r)=u+u^{\prime}r+\frac{u^{\prime\prime}}{2!}r^{2}+\ldots,\quad u \neq 0. \tag{20}\]
Inserting it into the differential equation we obtain algebraic equation
\[-2\left[Mq_{1}q_{2}u-(\ell+1)u^{\prime}\right]r^{-1}\] \[+\left[2M(E-W_{0})u-2Mq_{1}q_{2}u^{\prime}+(2\ell+3)u^{ \prime\prime}\right]r^{0}+\mathrm{O}\left(\mathbf{r}\right)\mathop{ \longrightarrow}\limits_{r\to 0}0\]
which can be satisfied only if
\[a \equiv \frac{u^{\prime}}{u}=\frac{Mq_{1}q_{2}}{\ell+1}, \tag{21a}\] \[b \equiv \frac{1}{2}\frac{u^{\prime\prime}}{u}=\frac{(\ell+1)a^{2}+ M\left(W_{0}-E\right)}{2\ell+3}, \tag{21b}\]
We have to note that \(b=b\left(r_{3},\ldots,r_{N}\right)\) depends only on charges and distances of the non-coalescent particles similarly to \(W_{0}=W_{0}\left(r_{3},\ldots,r_{N}\right)\).
Footnote 7: Leading expansion coefficients of radial wave functions of the \(H\) atom are the same as (21a, 21b) with \(W_{0}=0\).
Since singularity of many-particle Hamiltonian is isotropic at \(r=0\) its eigenfunctions have the form
\[\Psi=r^{\ell}u\left(\mathbf{r}|\mathbf{r}_{3},\ldots,\mathbf{r}_{N}\right)=r^{ \ell}\sum_{\lambda,\mu}r^{\lambda}u_{\lambda\mu}(r)Y_{\lambda\mu}(\vartheta,\varphi) \tag{22}\]
In view of local commutativity (9a) of angular momentum with Hamiltonian both \(\ell\) and \(m\) have definite values at the origin. Since many-particle wave function is antisymmetric under interchange of any two electrons, \(\ell\) takes only even values for relative singlet spin states \(s_{1}+s_{2}=0\) and odd values for relative triplet spin states \(s_{1}+s_{2}=1\) of the coalescent electrons.
Footnote 8: Due to ignoring the asymptotic central symmetry of Hamiltonian at \(r=0\), spatially generalized power series \(\Psi=\sum_{\lambda,\mu}r^{\lambda}u_{\lambda\mu}(r;\mathbf{r}_{3},\ldots, \mathbf{r}_{N})Y_{\lambda\mu}(\vartheta,\varphi)\) was used by former authors.
\[\Psi\mathop{\longrightarrow}\limits_{r\to 0}r^{\ell}u_{\ell}\left(1+ar+br^{2} \right)Y_{\ell m}(\vartheta,\varphi)+\ldots.\]Therefore the many-electron wave function can be decomposed in the vicinity of Coulomb singularities as
\[\Psi=r^{\ell}\left[u_{\ell}(r|\mathbf{r}_{3},\ldots,\mathbf{r}_{N})Y_{\ell m}( \vartheta,\varphi)+r^{3}v(\mathbf{r}|\mathbf{r}_{3},\ldots,\mathbf{r}_{N}) \right], \tag{23}\]
where
\[u_{\ell} = u_{\ell}\left(0\right)\left(1+ar+br^{2}+cr^{3}+\ldots\right),\] \[v = \sum_{\lambda=\ell+1}^{\infty}\sum_{\mu=-\lambda}^{\lambda}r^{ \lambda-\ell-1}v_{\lambda\mu}(r|\mathbf{r}_{3},\ldots,\mathbf{r}_{N})Y_{ \lambda\mu}\left(\vartheta,\varphi\right).\]
Term \(\mathrm{O}\left(\mathbf{r}^{\ell}\right)\) of decomposition is an eigenfunction of the spherically averaged Hamiltonian, term \(\mathrm{O}\left(\mathbf{r}^{\ell+3}\right)\) reflects anisotropy of molecular or crystalline potential, so group theoretical considerations apply only to \(v=v(\mathbf{r}|\mathbf{r}_{3},\ldots,\mathbf{r}_{N})\).
### Local behaviour at Coulomb singularities of spherically averaged Hamiltonian
Let us investigate the local behaviour of first term of decomposition in more detail. In view of Eq.
\[\overline{\widetilde{H}}\mathop{\longrightarrow}\limits_{r\to 0}-\frac{ \Delta}{2M}+\frac{q_{1}q_{2}}{r}+W_{0} \tag{24}\]
can be rewritten as
\[ru^{\prime\prime}+\left(2\ell+2\right)u^{\prime}-\left(2\alpha+\beta^{2}r \right)u=0,\]
Seeking the solution in the form \(u=\mathrm{e}^{-\beta r}w(r)\) and then by making change of variable \(x\equiv 2\beta r\) we obtain following confluent hypergeometric equation
\[xw^{\prime\prime}+\left(2\ell+2-x\right)w^{\prime}-\left(\ell+1+\alpha/\beta \right)w=0\]
which is of Kummer type
\[xw^{\prime\prime}+\left(\mathfrak{b}-x\right)w^{\prime}-\mathfrak{a}w=0\]
Regular solution of this equation is the following Kummer function 9
Footnote 9: The solution reduces to associated Laguerre polynomial in case of \(W_{0}=0\). We may formally consider \(-\alpha/\beta\) as a non-integer principal quantum number.
\[w={}_{1}F_{1}\left(\mathfrak{a};\,\mathfrak{b};\,x\right)\equiv\sum_{k=0}^{ \infty}\frac{\left(\mathfrak{a}\right)_{k}}{\left(\mathfrak{b}\right)_{k}} \frac{x^{k}}{k!},\]
Therefore the physical solution of Eq.
\[u_{\ell}(r)=\mathrm{e}^{-\beta r}{}_{1}F_{1}\left(\ell+1+\alpha/\beta;\,2\ell+ 2;\,2\beta r\right). \tag{25}\]Since Eq. is non-singular, its analytic solution can be expanded in terms of a power series.
\[\left[(2\ell+2)\,a_{1}-2\alpha a_{0}\right]r^{0}\] \[+ \sum_{k=1}^{\infty}\left[(k+1)(2\ell+2+k)a_{k+1}-2\alpha a_{k}- \beta^{2}a_{k-1}\right]r^{k}=0\]
leading to recurrence relations
\[a_{1} = \frac{\alpha}{\ell+1}a_{0}, \tag{26a}\] \[a_{k+1} = \frac{2\alpha\,a_{k}+\beta^{2}a_{k-1}}{(2\ell+2+k)(k+1)},\quad k= 1,2,\ldots\,. \tag{26b}\]
The first one is equivalent to Lowdin's first order cusp condition and the second one provides a simple way of generating higher order cusp conditions for eigenfunctions of the spherically averaged Hamiltonian. Use of these recurrence relations is more efficient numerically than evaluating the confluent hypergeometric function in Eq..
In view of Eqs. (10a), and the eigenfunction of the spherically averaged Hamiltonian is of the form
\[\psi=r^{\ell}u_{\ell}\mbox{e}^{-\beta r}{}_{1}F_{1}\left(\ell+1+\alpha/ \beta;\,2\ell+2;\,2\beta r\right)Y_{\ell m}(\vartheta,\varphi) \tag{27}\]
This analytic function fully characterizes cusps of spherically symmetric or spherically averaged systems since arbitrarily high order cusp relations can be derived from it. Higher order cusp conditions obtained from Kummer type wave function agree with that of Refs. and (after typographic correction \(Z_{\alpha}\mapsto Z_{\alpha}^{2}\)). For anisotropic systems, only first and second order cusp conditions (21a, 21b) are exact which can also be obtained from recurrence relations (26a, 26b). This bound-state local solution becomes non-physical at larger distances satisfying \(q_{1}q_{2}/r+W_{0}\geq 0\) since non-coalescent particles within the sphere of radius \(r\) are not included. Function is not square-integrable but knowledge of its higher order derivatives at \(r=0\) is useful in numerical calculation of first term of Eq..
## 4 Many-electron cusp and boundary conditions
Using expansions and of many-electron wave function about singular points of Hamiltonian we are able to recover explicit forms of natural boundary conditions for the many-electron Schrodinger equation. Due to local isotropy (9a, 9b) of Hamiltonian at singular points we can follow the same procedure as we did in section 2 for the \(H\) atom. One of our resulted boundary conditions may be considered as an _exact_ cusp condition without spherical average which contradicts Kato's result. Using our Kummer function solutionRadial density obtained from compared with that of computed by Koga et al. for ground state of the \(He\) atom. Our near-nucleus approximation is accurate up to surprisingly large distances. Effective Bohr radius \(r_{0}\) of the orbital is marked on the \(r\) axis.
Wave function of Kummer type compared with Hartree-Fock-Roothaan (HFR) wave function computed by Koga, Kanayama, Watanabe and Thakkar for ground state of the \(He\) atom. Radius \(r=Z/W_{0}\) of bound-state region of spherically averaged Hamiltonian is marked on the \(r\) axis.
### Boundary conditions
In view of expansion of many-electron wave function at large distances, its logarithmic derivative with respect to \(r\) is
\[\frac{1}{\Psi}\frac{\partial\Psi}{\partial r}\underset{r\rightarrow\infty}{ \longrightarrow}-\sqrt{-2M^{\prime}E}+\left[\frac{M^{\prime}(Q+1)}{\sqrt{-2M^ {\prime}E}}-1\right]\frac{1}{r}+\mathrm{O}\left(\mathbf{r}^{-2}\right)\]
which defines our first boundary condition:
\[\lim_{r\rightarrow\infty}\frac{1}{\Psi}\frac{\partial\Psi}{\partial r}=-\sqrt{ -2M^{\prime}E}. \tag{28}\]
The directional logarithmic derivative of the many-electron wave function in arbitrary direction \(\mathbf{e}\) can be expressed in terms of spherical polar coordinates as
\[\frac{\mathbf{e}\cdot\nabla\Psi}{\Psi}=\frac{\mathbf{e}\cdot\mathbf{e}_{r}}{ \Psi}\frac{\partial\Psi}{\partial r}+\frac{\mathbf{e}\cdot\mathbf{e}_{\vartheta }}{\Psi\,r}\frac{\partial\Psi}{\partial\vartheta}+\frac{\mathbf{e}\cdot \mathbf{e}_{\varphi}}{\Psi\,r\sin\vartheta}\,\frac{\partial\Psi}{\partial \varphi},\]
In radial directions, defined by \(\mathbf{e}\cdot\mathbf{e}_{r}=\pm 1\) and \(\mathbf{e}\cdot\mathbf{e}_{\vartheta}=\mathbf{e}\cdot\mathbf{e}_{\varphi}=0\), above expression reduces to
\[\frac{\mathbf{e}\cdot\nabla\Psi}{\Psi}=\frac{\mathbf{e}\cdot\nabla_{r}\Psi}{ \Psi}=\frac{\pm 1}{\Psi}\frac{\partial\Psi}{\partial r}.\]
Since wave function \(\Psi=r^{\ell}u(\mathbf{r}|\mathbf{r}_{3},\ldots,\mathbf{r}_{N})\equiv r^{\ell }u(\mathbf{r})\) given by Eq.
\[\frac{1}{\Psi}\frac{\partial\Psi}{\partial r}\underset{r\to 0}{ \longrightarrow}\frac{\partial_{r}^{\ell+1}\Psi}{\partial r^{\ell}_{\Psi}} \underset{r\to 0}{\longrightarrow}\frac{\ell+1}{u(\mathbf{r})} \frac{\partial u(\mathbf{r})}{\partial r}+\mathrm{O}\left(\mathbf{r}\right),\]
Since expansion exhibits hydrogenic angular dependence \(r^{\ell}u_{\ell}(r)Y_{\ell m}(\vartheta,\varphi)\) in the vicinity of Coulomb singularities the radial logarithmic derivative of \(\Psi\) has the following one-sided limits:
Footnote 10: In general, radial partial derivative of an anisotropic function is anisotropic which explains appearance of spherical average in Kato’s cusp condition. See discussion after the statement of theorem \(IIa\) in Ref..
\[\lim_{\mathbf{r}\rightarrow\pm 0}\frac{1}{\Psi}\frac{\partial\Psi}{\partial r}= \lim_{\mathbf{r}\rightarrow\pm 0}\,\frac{\partial_{r}^{\ell+1}\Psi}{ \partial_{r}^{\ell}\Psi}=\pm\left(\ell+1\right)a, \tag{29}\]
(21a) and
\[\mathbf{r}\rightarrow+0 \equiv \left(r\to 0,\,\vartheta,\,\varphi\right)\!,\] \[\mathbf{r}\rightarrow-0 \equiv \left(r\to 0,\,\pi-\vartheta,\,\pi+\varphi\right)\!.\]
Equation is our second boundary condition.
Equations and represent homogeneous Robin boundary conditions for the many-electron wave function and have the same form as Eqs. and obtained for the \(H\) atom which is a consequence of asymptotic isotropy (9a, 9b) of Hamiltonian at singular points.
### Cusp conditions
Boundary condition is the exact form of first order cusp condition which does not contain spherical average. By substituting specific values of \(M\), \(q_{1}\) and \(q_{2}\) into Eq.
\[\lim_{\mathbf{r}\rightarrow\pm 0}\frac{\partial_{r}^{\ell+1}\Psi}{\partial_{r}^{ \ell}\Psi}=\mp\frac{M_{\nu}}{M_{\nu}+1}Z_{\nu},\quad\ell=0,1,2,\ldots, \tag{30}\]
u.).
\[\lim_{\mathbf{r}\rightarrow\pm 0}\frac{\partial_{r}^{\ell+1}\Psi_{ \uparrow\downarrow}}{\partial_{r}^{\ell}\Psi_{\uparrow\downarrow}}=\pm\frac{1 }{2},\quad\ell=0,2,4,\ldots, \tag{31a}\] \[\lim_{\mathbf{r}\rightarrow\pm 0}\frac{\partial_{r}^{\ell+1} \Psi_{\uparrow\uparrow}}{\partial_{r}^{\ell}\Psi_{\uparrow\uparrow}}=\pm\frac{ 1}{2},\quad\ell=1,3,5,\ldots, \tag{31b}\]
First order many-body cusp conditions (30, 31a, 31b) are obviously \(CP\)-invariant similarly to Eq. obtained for the \(H\) atom.
Since second directional derivative \(\left(\mathbf{e}\cdot\nabla\right)^{2}\) is of definite sign the second order cusp condition has the form
\[\lim_{r\to 0}\frac{1}{\Psi}\frac{\partial^{2}\Psi}{\partial r^{2}}=\lim_{r \to 0}\frac{\partial_{r}^{\ell+2}\Psi}{\partial_{r}^{\ell}\Psi}=\left(\ell+1 \right)\left(\ell+2\right)b, \tag{32}\]
(21b).
Coefficient \(a\) in Eqs. depends only on charges and masses of the coalescent particles. In view of and (21b), coefficient \(b=b\left(r_{3},\ldots,r_{N}\right)\) in Eq. depends only on charges and distances of non-coalescent particles hence it has the same value for the spherically averaged Hamiltonian. Arbitrarily high order cusp conditions can be obtained from Eq. by means of recurrence relation (26a, 26b) in the framework of central field approximation. Third and higher order cusp conditions for exact wave functions require spherical average similarly to Kato's cusp condition since higher order coefficients of expansion depend both on distances and directions of the non-coalescent particles.
In view of Eq. the electron pair density exhibits short-range hydrogenic angular dependence in the vicinity of coalescence points hence satisfies exact first and second order cusp conditions similar to (31a, 31b) and.
## 5 Consequences
### Physical consequences
Since jump of logarithmic derivative of wave function is \(2Z\) at the nuclei both in the one-electron and in the many-electron cases, its discontinuity is caused by common terms of the two Hamiltonians, namely, by singular Coulomb potential and by singular kinetic energy of opposite sign as if no other particles were present except the coalescent ones. An electron-electron potential taking part in cancellation of the nuclear Coulomb singularity would require a more singularwave function than the square-integrable functions hence cusp condition shows an evidence for the normalization condition.
Potential energy term \(W_{0}\) defined by Eq. is related to the NMR chemical shift since it is proportional to ratio \(H_{i}/H\) of induced diamagnetic shielding field to the applied external magnetic field.
Expansion characterizes the short-range behaviour of many-electron wave function both in the vicinity of electron-nucleus and electron-electron coalescences hence it can be used to describe short-range electron correlation which depends on the chemical environment in second order through parameter \(b=b\left(r_{3},\ldots,r_{N}\right)\).
First order cusp condition (21a, 29) is responsible for boundedness of many-electron wave function at Coulomb singularities by exactly cancelling singular kinetic and potential energy terms. Second order cusp conditions (21b, 32) enforce correct value \(E\) of local energy \(E_{loc}\equiv\Psi^{-1}\hat{H}\Psi\) at Coulomb singularities by imposing constraints on curvature of many-electron wave function. These conditions are exact for arbitrary values of \(\ell\) and do not need spherical average in contrast to Kato's cusp condition.
One can see from Eq. that all radial partial derivatives of eigenfunctions of spherically averaged Hamiltonian at coalescence points depend on the same constant quantities: \(\ell\), \(q_{1}\), \(q_{2}\), \(M\), \(E\) and \(W_{0}\).
Third and higher derivatives of exact eigenfunctions depend on anisotropic multipole potential terms of expansion as well. In view of Eq., symmetry considerations apply only to terms \(\mathrm{O}\left(\mathbf{r}^{\ell+3}\right)\).
The real component of Bloch functions exhibits cusps at the nuclei similarly to molecular wave functions since the cosine function does not vanish at \(r=0\).
### Numerical consequences
If both boundary conditions are known the numerical solution of Schrodinger equation over a grid yields an algebraic eigenvalue problem which can be solved by means of Jacobi - Goldstine - Murray - von Neumann diagonalization algorithm. Before Kato's cusp condition only long-range behaviour of wave function was known hence a numerical trick called _shooting method_ was used to substitute the missing boundary condition at Coulomb singularities. This method is best suited to equidistant grids, where inward and outward numerical integrations are equally accurate. Since most of energy is concentrated at near-nucleus regions the practical grids used in quantum chemistry are substantially finer in the vicinity of nuclei than in the interstitial and exterior regions. The grid is coarsest at large distances representing \(\infty\) in the numerical calculation hence an outward integrated solution starting from a guessed initial condition is fitted at a midpoint to an inward integrated solution based on an inaccurate initial condition. Kato's cusp condition is suitable only to the central-field approximation and cannot be used as an exact boundary condition. Our boundary conditions and are exact regardless of molecular or crystalline symmetry since the wave function exhibits hydrogenic angular dependence in both limiting cases. A shooting method based on our cusp condition would be more accurate than the traditional one due to the finer grid at nuclei. Since both boundary conditions are known one can transform Schrodinger equation to an algebraic eigenvalue problem instead of performingthe time-consuming shooting loop. Approximate wave function can be used to find optimal near-nucleus step size for the grid.
Basis sets satisfying cusp conditions improve convergence of Hartree-Fock-Roothaan variational calculations.
\[R=c_{0}r^{\ell}\mathrm{e}^{-\frac{2r}{\ell+1}}+\sum_{\lambda=1}^{L}c_{\lambda}r ^{\ell+\lambda}e^{-\zeta_{\lambda}r}\]
Similar basis set was used by Roothaan and Kelly however their summation inexplicably starts from \(\lambda=3\) resulting in a slow convergence for \(\ell>0\) hence use of cusp condition was limited to the \(s\) orbitals in their atomic calculations 11.
Footnote 11: We suppose that their summation starting from \(\lambda=3\) arises from some unpublished expansion similar to our Eq., where our \(u_{\ell}\) was simply replaced by \(e^{-Zr/(\ell+1)}\).
\[R=c_{0}r^{\ell}\left[1+\left(b-\frac{a^{2}}{2}\right)r^{2}\right]\mathrm{e}^{ ar}+\sum_{\lambda=3}^{L}c_{\lambda}r^{\ell+\lambda}\mathrm{e}^{-\zeta_{\lambda}r}, \tag{33}\]
(21a, 21b). In order to preserve asymptotic behaviour of wave function as \(r\to\infty\) a Slater function of the form should be added to the basis set.
It is widely believed that Gaussian basis sets are not suitable to describe nuclear cusps since they have zero gradients at the nuclei which is true only for individual Gauss functions but not for their linear combinations By equating three leading expansion coefficients of linear combination of three Gauss functions with equal exponential factors to that of expansion we obtain Gaussian basis set
\[\psi(x,y,z) = c_{0}r^{\ell}\left[1+ar+\left(b+g_{0}\right)r^{2}\right]\mathrm{ e}^{-g_{0}r^{2}}Y_{\ell m}(\vartheta,\varphi) \tag{34}\] \[+\sum_{i+j+k\geq\ell+3}^{L}c_{ijk}x^{i}y^{j}z^{k}\mathrm{e}^{-g_{ ijk}r^{2}}\]
There is no finite linear combination of Gauss functions which exhibits asymptotic behaviour of wave function as \(r\to\infty\). Use of basis functions satisfying both first and second order electron-electron cusp conditions provides the simplest way to include short-range correlation effects. Above basis set is more efficient numerically and requires less modification of existing Gaussian computer codes than implementing a Jastrow-type correlation.
Asymptotic hydrogenic angular dependence of three leading terms of expansion of many-electron wave function about Coulomb singularities explains the success of central-field approximation used in atomic calculations and muffin-tin approximation of solid state physics. The Kummer-type confluent hypergeometric function intended to characterize the short-range behaviour of eigenfunctions of the spherically averaged Hamiltonian is surprisingly accurate even at relatively large distances, e.g. one can see from that relative error of this approximation to radial density \(4\pi r^{2}\left|\psi\right|^{2}\) is \(5.8\%\) at the effective Bohr radius and is \(0.4\%\) at its half. Therefore the spherically averaged part of Hamiltonian is responsible for most of the effects and anisotropic terms can be considered as perturbations.
## Appendix: Some differentiation rules
Leibniz's theorem for differentiation of products states that
\[\left[f(x)g(x)\right]^{(n)}=\sum_{k=0}^{n}\binom{n}{k}f^{(n-k)}(x)g^{(k)}(x).\]
For \(f(x)=x^{\ell}\) one obtains
\[\left[x^{\ell}g(x)\right]^{(n)}=\sum_{k=0}^{n}\binom{n}{k}\frac{\ell!\,x^{\ell -n+k}g^{(k)}(x)}{(\ell-n+k)!}.\]
Specific higher order derivatives of the above type used in this paper are
\[\left[x^{\ell}g(x)\right]^{(\ell)} \underset{x\to 0}{\longrightarrow}\ell!\,g+\mathrm{O}\left(x \right), \tag{35a}\] \[\left[x^{\ell}g(x)\right]^{(\ell+1)} \underset{x\to 0}{\longrightarrow}(\ell+1)!\,g^{\prime}+ \mathrm{O}\left(x\right),\] (35b) \[\left[x^{\ell}g(x)\right]^{(\ell+2)} \underset{x\to 0}{\longrightarrow}(\ell+2)!\,g^{\prime\prime}+ \mathrm{O}\left(x\right). \tag{35c}\]
For higher order radial partial derivatives of \(r^{\ell}u(\mathbf{r})\) we obtain similarly
\[\frac{\partial^{\ell}r^{\ell}u(\mathbf{r})}{\partial r^{\ell}} \underset{r\to 0}{\longrightarrow}\ell!\,u+\mathrm{O}\left( \mathbf{r}\right), \tag{36a}\] \[\frac{\partial^{\ell+1}r^{\ell}u(\mathbf{r})}{\partial r^{\ell+1}} \underset{r\to 0}{\longrightarrow}(\ell+1)!\,\left.\frac{\partial u}{ \partial r}\right|_{r=0}+\mathrm{O}\left(\mathbf{r}\right),\] (36b) \[\frac{\partial^{\ell+2}r^{\ell}u(\mathbf{r})}{\partial r^{\ell+2}} \underset{r\to 0}{\longrightarrow}(\ell+2)!\,\left.\frac{\partial^{2}u}{ \partial r^{2}}\right|_{r=0}+\mathrm{O}\left(\mathbf{r}\right). \tag{36c}\]
Quotients of above derivatives become isotropic if \(u(\mathbf{r})\) can be written as a product of radial and angular parts.
|
10.48550/arXiv.1010.2700
|
Boundary conditions for many-electron systems
|
Péter V. Tóth
| 1,511
|
10.48550_arXiv.1602.05615
|
## I Introduction
Highly efficient and fast vibrational energy transport on a molecular scale has been a subject of theoretical and experimental investigations in recent decades. The possible applications in biochemistry, organic chemistry and nanotechnology include development of efficient cooling in microscopic and nanoscopic molecular systems, such as nanowires and optical limiters, designing efficient energy transport schematics for energy signaling, as well as optimizing and even promoting chemical reactions by concentrating the excess energy at the reaction center. It is suggested that quantum vibrational excitations can be manipulated similarly to electrons and photons, thus enabling controlled heat transport. Moreover, delocalized excitations (phonons) can be used to carry and process quantum information. The highest transport speed was found in alkanes (1.44 km/s).
Possible candidates capable to maintain fast and efficient energy transport are oligomers because of their periodic structure. In such systems vibrational states can be substantially delocalized because of the strong interaction of equivalent site states, so that ballistic energy transport takes place as a free-propagating wavepacket. The ballistic constant-speed transport has been observed in bridged azulene-anthracene compounds, polyethylene glycol oligomers, alkanes, and perfluoroalkanes and the theory describing this transport and its possible breakdown due to decoherence has been suggested.
Phonon wavepackets in carbon based polymer chains can propagate with the group velocity as high as 10 km/s because of a high strength of covalent bonds. Yet the maximum energy of singly excited vibrations does not exceed ca. 3000 cm\({}^{-1}\), as the motion is associated with displacements of rather heavy nuclei. Thus, the ballistically transferred energy is much smaller than a typical bond energy exceeding 1 eV (\(\sim 10^{5}\) cm\({}^{-1}\)). Involvement of multi-phonon transport to increase the amount of transferred energy is expected to enhance the energy relaxation/dissipation. Much larger energy can be carried by excitons, delocalized electronic states. However molecular excitons are usually strongly coupled to the environment resulting in incoherent energy transport (see e. g. exciton transport in DNA).
Here we propose to exploit the special vibrational modes of entirely electronic nature capable of efficient delivering energies in the eV range. Such modes can exist in molecules having all atoms aligned along the single axis (see Figure 1) and they are formed by propagating torsional oscillations of electronic nature. Nuclei do not participate in these oscillations because their rotation about the axis they located on is degenerate.
Considering a linear molecule as an elastic rod, four gapless phonon branches are expected based on symmetry, including longitudinal, two transverse, and one torsional modes. The longitudinal and torsional oscillations of frequency \(\omega\) and wavevector \(q\) are characterized by an acoustic spectrum \(\omega=cq\) with a relatively high speed of sound, \(c\). As oppose to an elastic rod, for a molecular chain with all atoms located on the same axis, there is no nuclear contribution to the torsional vibrations, since the chain is completely linear. Nevertheless, the system can possess a remarkable torsional stiffness due to anisotropic arrangement of its electronic clouds. Such situation is found in cumulenes, featuring a chain of carbon atoms coupled to each other by double bonds, where the anisotropy results from the \(\pi\)-bond anisotropy between carbon atoms (see Figures 1(a), 2). Similar conditions can be realized in transition metals where atomscan form chain bridges between junctions.
The torsional sound should exist in such system and we expect it to be of a purely electronic nature because nuclei are positioned along the primary axis and cannot participate in the torsional motion. Since electrons are much lighter than atoms it is natural to expect the speed and a single quantum energy to be much higher than those for nuclei vibrations.
In the present study we performed a first principle investigation of the electronic torsional waves in cumulene chains. We found that the speed of sound in cumulenes to be as high as 1000 km/s and a maximum energy of the quantum as high as 10 eV. Because this type of motion is not related to the atomic vibrations and differs from Langmuir waves in plasma, the corresponding quantum quasi-particle is neither a phonon nor a plasmon. To avoid confusion we will call a quantum of torsional electronic oscillations a _torsion_. The effect of zero point vibrations leading to the _torsion_ spectral gap of the order of 0.03 eV is estimated. The possible ways to observe _torsitons_ experimentally are discussed.
We consider the torsional oscillations of a cumulene in a dielectric environment, so the electronic excitations can be neglected. Although the metallic behavior of cumulene was predicted theoretically it is not confirmed experimentally so the nature of electronic excitations remains unclear. Here we ignore electronic excitations assuming that there is the significant spectral gap (cf. Ref.).
## II The system
A linear cumulene chain is a compound containing a sequence of \(n\) carbon atoms with \((n-1)\) double bonds between them R=C=(C=)\({}_{n-2}\)C=R. Quantum chemistry calculations were performed for the simplest termination of cumulene chain by two hydrogen atoms on each side; an example of cumulene molecule H\({}_{2}\)C=(C=)\({}_{3}\)CH\({}_{2}\) is shown schematically in Figure 1(a). One can see that orthogonal \(\pi\)-bonds between carbon atoms can provide rigidity with respect to twisting with remarkable torsional stiffness, while much smaller stiffness is expected in another carbyne modification, polyyne, which is a chain of carbon atoms with alternating single and triple bonds between them (see Figure 1(b)). For cumulene molecules the shortened notation H\({}_{2}\)C\({}_{n}\)H\({}_{2}\) (without bond type specification) will be used.
## III Electronic torsional mode
To estimate the speed of sound for electronic torsional wave we consider a model of elastic rod (torsion spring) which can be described by the Lagrangian
\[\mathcal{L}_{e}=\frac{1}{2}\int\limits_{z_{l}}^{z_{r}}dz\Bigg{\{}j_{e}\left( \frac{d\theta}{dt}\right)^{2}-\kappa\left(\frac{\partial\theta}{\partial z} \right)^{2}\Bigg{\}} \tag{1}\]
Here \(z_{l,r}=\mp L/2\), \(L\) -molecule length, \(\kappa\) stands for torsional stiffness and \(j_{e}\) is an average linear density of electronic moment of inertia with respect to z-axis.
We estimate parameters of interest as \(j_{e}=1.73\ m_{e}\cdot\)A (\(0.95\cdot 10^{-3}\) u\(\cdot\)A) and torsional stiffness \(\kappa=10.6\) eV\(\cdot\)A as described in the next two sections. Our estimate for the torsional stiffness is consistent with the previous estimate of 10.3 eV\(\cdot\)A reported in Ref..
With the angle \(\theta(z,t)\) and the related angular velocity \(d\theta/dt\) considered as dynamical variables Eq.
\[j_{e}\frac{\partial^{2}\theta}{\partial t^{2}}=\kappa\frac{\partial^{2}\theta} {\partial z^{2}} \tag{2}\]
which is a wave equation with the dispersion relation \(\omega(q)=q\sqrt{\kappa/j_{e}}\) and the speed of _torsion_ wave
\[c=\frac{d\omega}{dq}=\sqrt{\frac{\kappa}{j_{e}}}\simeq 1.0\cdot 10^{6}\,\mathrm{ m/s} \tag{3}\]
Carbyne modifications: (a) cumulene molecule with orthogonal double bonds; (b) polyyne molecule with alternating single and triple bonds
Cumulene molecule torsionally strained along primary axis
This velocity exceeds the typical phonon propagation velocity in polymers by two or three orders of magnitude. Next we also estimate the maximum energy transfered by the electronic torsional mode.
The dispersion relation for longitudinal vibration in a uniform chain with nearest neighbor coupling and the lattice period \(a\) has the standard form \(\omega(q)=\omega_{*}\sin(aq/2)\).
It should be a good approximation for the torsional mode under consideration because the interaction responsible for the torsional stiffness is due to short range covalent bonding.
\[\hbar\omega_{*}=2\frac{\hbar c}{a}\simeq 10.5\,\mathrm{eV} \tag{4}\]
(the lattice period in cumulenes is given by \(a=1.28\) A). This result corresponds to the nearest neighbor effective coupling \(\hbar\omega_{*}/4\sim 2.6\) eV, which is about the \(\pi\)-bond strength of 2.74 eV in C=C bonds.
We can also roughly estimate the _torsion_ mean free path order of magnitude. In Ref. the decoherence rate has been estimated analyzing the switch between ballistic and diffusive transport in perfluoroalkanes. It is found to be \(W=2\) ps\({}^{-1}\), so the phonon decoherence time has been estimated as \(T\sim 1/W\sim 0.5\) ps. Assuming similar decoherence time for the _torsion_, one can estimate its mean free path as \(l_{0}=cT\sim 0.5\)\(\mu\)m.
Thus we found the electronic torsional sound wave velocity and energy unprecedentedly high compared to typical phonon parameters which makes this system very attractive for energy transport applications. The energy transferred by a single quantum is sufficiently large for chemical applications: bond making-bond breaking, energy release, and energy transfer to reaction center.
Below we derive our estimate for electronic moment of inertia and for torsional stiffness, discuss the limitations of our result due to zero point atomic vibrations and propose the way to observe the ultrafast energy transport due to electronic torsional sound.
## IV Electronic moment of inertia
The linear density of electronic moment of inertia is defined as
\[j_{e}(z)=\iint\rho(x,y,z)\left(x^{2}+y^{2}\right)dxdy \tag{5}\]
We calculated the linear density of electronic moment of inertia for cumulene molecule using density functional theory with B3LYP hybrid functional and standard 6-31(d, p) basis set, as implemented in a Gaussian 09 software package. The electron density as a function of coordinates is extracted with the uniform grid of 0.1 Bohr radius (0.0529 A), the symmetric limits in X-Y plane (the plane perpendicular to the prime axis) were chosen \(\pm 6.5\) Bohr radius (\(\pm 3.44\) A). Either doubling of the limits or decrease of the grid by the same factor change the result by less than 1%.
To estimate the accuracy of the numerical result we tested the same approach on the hydrogen atom. The theoretical value of the moment of inertia of hydrogen electron cloud in the ground state can be calculated using the electron wave-function as \(J_{0}=2m_{e}a_{0}^{2}\), where \(m_{e}\) is electron mass and \(a_{0}\) is the Bohr radius. The result of numerical calculations obtained using the same method as for cumulene is \(J_{num}=1.90m_{e}a_{0}^{2}\), which is within 5% accuracy.
In we show dependence \(j_{e}(z)\) obtained from DFT-calculations for H\({}_{2}\)C\({}_{11}\)H\({}_{2}\). One can see that \(j_{e}(z)\) is a smooth function weakly deviating from its average \(J_{e}/L\) (\(\sim 5\%\)), so for simplicity coordinate dependent moment of inertia density \(j_{e}(z)\) can be replaced with the constant \(j_{e}\simeq J_{e}/L\simeq 1.73\)\(m_{e}\cdot\mathrm{\AA}\) (\(0.95\cdot 10^{-3}\) u\(\cdot\)A). As shown in Figure 3, we define molecular length \(L\) as a distance between the second left and second right carbon atoms, where \(j_{e}(z)\) is still not affected by boundaries.
## V Torsional stiffness calculation
Combining the proposed model with the first principles calculations of the hydrogen atom torsional vibrational mode associated with the relative torsional oscillations of pair of hydrogen atoms ("whiskers", see Figure 2), we
Linear density of the electronic inertia moment \(j_{e}(z)\) as a function of coordinate along primary axis \(z\) (blue solid line). The corresponding molecule H\({}_{2}\)C\({}_{11}\)H\({}_{2}\) is represented below. One can see that \(j_{e}(z)\) is a smooth function with low deviations from the average (magenta dash-line), except of the boundaries. The effective length \(L\) is defined as a distance between second left and second right carbon atoms.
introduce Lagrangian
\[\mathcal{L}=\frac{J_{l}}{2}\left(\frac{d\Phi_{l}}{dt}\right)^{2}+\frac{J_{r}}{2} \left(\frac{d\Phi_{r}}{dt}\right)^{2}-\int\limits_{-L/2}^{+L/2}dz\frac{\kappa} {2}\left(\frac{\partial\theta}{\partial z}\right)^{2} \tag{6}\]
\(\theta(z,t)\) is the same as in Eq. with boundary conditions \(\theta(\mp L/2)=\Phi_{l,r}\).
In this model the potential energy is originated from the torsional strain of the electronic spring and kinetic energy is entirely defined by the motion of hydrogen atoms, so long as the kinetic energy of electrons is neglected. The latter assumption is justified as long as \(J_{e}\ll J\) (0.028 vs 3.43 u\(\cdot\)A for \(n=25\)).
The torsional energy has a minimum at constant torsional angle gradient \((\partial\theta/\partial z)=(\Phi_{r}-\Phi_{l})/L\) suggesting that electrons adiabatically follow atomic motion. For the only hydrogen torsional oscillator mode one can assume antisymmetric condition \(-\Phi_{l}=\Phi=\Phi_{r}\).
\[\frac{d^{2}\Phi}{dt^{2}}=-\frac{4\kappa}{JL}\Phi \tag{7}\]
This equation describes the harmonic oscillator with the frequency defined as
\[\omega_{\tau}^{2}=\frac{4\kappa}{J}\frac{1}{L} \tag{8}\]
Using the same DFT calculation, from harmonic vibrational analysis one can find \(\omega_{\tau}\) of H\({}_{2}\)C\({}_{n}\)H\({}_{2}\) for different \(n\). In we represented the related frequency \(\omega_{\tau}\) for \(n=5,6,8,10,12,16,24,25\). Since the choice of length \(L\) includes some arbitrariness (our choice is illustrated in Figure 3), the correct fit should include some length parameter \(B\sim a\), so that \(\omega_{\tau}^{2}=A/(B+L)\). Using optimum fitting analysis we found \(B=4.22\) A and the torsional stiffness is given by \(\kappa=AJ/4\simeq 2.89\cdot 10^{6}\) cm\({}^{-2}\)\(\cdot\)u\(\cdot\)A\({}^{3}\simeq 10.6\) eV\(\cdot\)A (u stands for the atomic mass unit), while atomic moment of inertia \(J\) is defined by end groups only and does not depend on \(n\). These estimates were used to evaluate the speed of _torsitons_. Below we analyze the effect of zero-point atomic vibrations on their spectrum.
## VI Effect of zero-point atomic vibrations
In our description of the electronic torsional mode we implicitly used Born-Oppenheimer approximation, considering electronic motion in an axially symmetric field of motionless nuclei, positioned along the z-axis. This axial symmetry is reflected by the symmetry of the Lagrangian in Eq. with respect to the change of the function \(\theta(z)\) by arbitrarily constant.
In reality, the nuclei participate in zero-point vibrations in the ground state, which does not possess an axial symmetry because this ground state is adjusted to the electronic ground state where this symmetry is broken (see Figure 2). Indeed, to find this ground state, one needs to consider interacting nuclei in the field of electronic cloud with already calculated anisotropic electronic density.
Thus, the potential energy depends on the angle \(\theta\) even in the absence of torsion and the energy minimum is realized at some angle \(\theta_{0}\) which we can set to zero. The potential energy can be expanded over the small displacement from this minimum as \(\alpha\theta^{2}/2\). This term incorporated to the Lagrangian in Eq. as \(-\alpha\theta^{2}/(2L)\) leads to the gap in the spectrum of torsional waves. Correcting Eq.
\[\omega^{2}(q)=\frac{\kappa}{j_{e}}q^{2}+\frac{\alpha}{j_{e}L} \tag{9}\]
Since \(\omega\neq 0\) the mode is not exactly acoustic due to the gap \(\Delta\omega=\sqrt{\alpha/j_{e}L}\).
To estimate the parameter \(\alpha\) consider the change of classical energy \(\delta E(\theta)=\left\langle\mathbf{\hat{\mathcal{H}}}(\theta)-\mathbf{\hat{\mathcal{ H}}}\right\rangle_{g}\), where \(\mathbf{\hat{\mathcal{H}}}\) is the atomic chain quantum Hamiltonian and \(\left\langle\dots\right\rangle_{g}\) is an average over the ground state of carbon atoms considering their zero-point vibrations.
All the normal modes of carbon atoms in the molecule, which don't include motion of hydrogen atoms with respect to the adjacent carbon atoms, can be either longitudinal or transverse. For \(D_{2d}\) symmetry point group with coordinate system defined above there are only two possible second order invariants: \(z^{2}\) and \(x^{2}+y^{2}\), so the transverse modes of the harmonic Hamiltonian of the atomic chain are expected to be double-degenerate and the corresponding eigenfunctions possess axial symmetry. Practically, this degeneracy is observed in DFT-calculated IR spectra of H\({}_{2}\)C\({}_{n}\)H\({}_{2}\) for odd \(n\), while for even \(n\) all energy levels are split, because such molecules belong to \(D_{2h}\) symmetry group. Indeed, the splitting is entirely an effect of sides, because in even \(n\) molecules the side CH\({}_{2}\) groups lie in the same plane (while in odd \(n\) they are orthogonal), so X-Z and Y-Z plane become distinguishable, while for an infinite chain this effect would disappear.
The break of axial symmetry takes place in the third order anharmonic interaction. To express potential energy in normal modes representation introduce notations \(u_{x_{i}}\) and \(u_{y_{i}}\) for transverse modes with energy \(\hbar\omega_{x_{i}}=\hbar\omega_{y_{i}}=\hbar\omega_{i}\) and \(u_{z_{k}}\) for \(k-th\) longitudinal mode.
\[\boldsymbol{\hat{\mathcal{V}}}_{3}=\sum_{ijk}\Big{\{}V_{x_{i}y_{j }z_{k}}u_{x_{i}}u_{y_{j}}+\\ \frac{V_{x_{i}x_{j}z_{k}}}{2}\Big{(}u_{x_{i}}u_{x_{j}}+u_{y_{i}}u _{y_{j}}\Big{)}\Big{\}}u_{z_{k}} \tag{10}\]
With \(\boldsymbol{\hat{\mathcal{V}}}_{3}\) as a perturbation, a meaningful correction to the ground state in the first order of perturbation theory is given by
\[|\psi_{0}\rangle=|0\rangle-\frac{1}{\hbar\sqrt{8}}\sum_{ijk}\frac{V_{x_{i}y_{j }z_{k}}}{\omega_{i}+\omega_{j}+\omega_{z_{k}}}\left|1_{i}^{x}1_{j}^{y}1_{k}^{z}\right\rangle \tag{11}\]
The rotation of electronic cloud about the z-axis by an angle \(\theta\) changes the energy in diabatic approximation by
\[\delta\boldsymbol{\hat{\mathcal{H}}}(\theta)=-2\sum_{ijk}V_{x_{i}y_{j}z_{k}}u_ {x_{i}}u_{y_{j}}u_{z_{k}}\sin^{2}\theta \tag{12}\]
Assuming for the small displacement from the minimum \(\sin\theta\simeq\theta\), one can find \(\delta E(\theta)=\langle\psi_{0}|\,\delta\boldsymbol{\hat{\mathcal{H}}}(\theta )\,|\psi_{0}\rangle=\alpha\theta^{2}/2\), with \(\alpha\) given by
\[\alpha=\frac{1}{2\hbar}\sum_{k}\Big{\{}\sum_{i}\frac{V_{x_{i}y_{i}z_{k}}^{2}}{ 2\omega_{x_{i}}+\omega_{z_{k}}}+\sum_{i<j}\frac{2V_{x_{i}y_{j}z_{k}}^{2}}{ \omega_{x_{i}}+\omega_{y_{j}}+\omega_{z_{k}}}\Big{\}} \tag{13}\]
Using anharmonic frequency analysis of H\({}_{2}\)C\({}_{5}\)H\({}_{2}\) we calculated third-order anharmonicity constants. To exclude effect of hydrogen atoms we considered the only transverse and longitudinal normal modes with the nearest integer of reduced mass greater or equal 2 atomic units. Applying Eq. we found \(\alpha=3.2\) cm\({}^{-1}\), so that the energy gap can be estimated using Eq. as \(\Delta\varepsilon\simeq 130\) cm\({}^{-1}\).
To answer a question how crucial is the described effect for the acoustic mode, one can find the length \(L\) of a cumulene chain where this energy becomes comparable to the minimum _torsion_ energy, which can be estimated as
\[\hbar\omega_{min}\simeq\hbar cq_{min}=\hbar c\frac{\pi}{L} \tag{14}\]
Thus the length required to make the gap value of the same order as \(\hbar\omega_{min}\) is \(L_{*}\simeq\hbar\pi c/\Delta\varepsilon\simeq 128\) nm, that is much larger than the real molecule length.
Axial symmetry can be violated also by the forth-order anharmonic interaction, however its contribution into the energy gap does not change qualitatively the presented estimate.
## VII Experiment suggestions
As shown above, the electronic torsional mode features an unprecedented speed of 1000 km/s = 1 nm/fs and can transfer energy up to 10 eV, which is comparable to the energies of the strongest chemical bonds (C=C, N=N, etc.). Such high transferred energy brings an opportunity of performing chemistry at distances, including chemical bond breaking reactions. shows a schematic of the compound suitable for the proof of principle experiment on remote chemistry initiation. The compound features two surface-anchored end-groups connected by a cumulene chain. Laser initiated bond breaking at the initiation (left) end-group can result in generation of a strong torque at the chain which will propagate as a wave-packet along the chain and can result in bond breaking at another end group, the target. The energy released by the initiation end-group can be tuned by selecting convenient functional groups. Spectroscopic observation of the transported energy can aim at detecting the formation of the products at the target or detection of the excess energy at the target. In the latter case a longer cumulene chain is required as for the chain length of 50 carbon atoms the transport time is only ca. 5 fs. Compounds with such long chains have been synthesized for polyynes and we hope that this should be possible for cumulenes as well.
## VIII Conclusion
In this article we considered electronic torsional waves in cumulene chains (_torsitons_) which are torsional sound
A schematic experiment set up on remote chemistry initiation. The compound features two surface-anchored end-groups connected by a cumulene chain. Laser initiated bond breaking at the initiation (left) end-group can result in generation of a strong torque at the chain which will propagate as a wave-packet along the chain and can result in bond breaking at another end group, the target.
We evaluated the speed of _torsion_ propagation as high as 1000 km/s. Single _torsion_ can carry energy up to 10 eV. Similar waves should exist in other atomic chain with anisotropic bonds including recently discovered transition metal linear chains. While the largest band energy computed for cumulenes at 10 eV, the computations neglected electronic excitation, which will likely be contributing at such high energies. It will be interesting to see how the ground electronic state _torsitons_ are perturbed by electronic excitations at higher _torsion_ band energies and how the quasi-particles of two types, _torsitons_ and excitons, interact. Nevertheless, the presented band calculations are expected to be free of electronic excitation effects at smaller energies. Importantly, the transport speed supported by the lower half of the _torsion_ band is similar to that of the full band (with small corrections due to the _torsion_-vibron coupling).
|
10.48550/arXiv.1602.05615
|
Electronic torsional sound in linear atomic chains: chemical energy transport at 1000 km/s
|
Arkady A. Kurnosov, Igor V. Rubtsov, Andrii O. Maksymov, Alexander L. Burin
| 1,699
|
10.48550_arXiv.1106.4899
|
###### Abstract
The shielding of the nuclear magnetic moment by the bound electron in hydrogen-like ions is calculated _ab initio_ with inclusion of relativistic, nuclear, and quantum electrodynamics (QED) effects. The QED correction is evaluated to all orders in the nuclear binding strength parameter and, independently, to the first order in the expansion in this parameter. The results obtained lay the basis for the high-precision determination of nuclear magnetic dipole moments from measurements of the \(g\)-factor of hydrogen-like ions.
Other methods such as atomic beam magnetic resonance, collinear laser spectroscopy, and optical pumping (OP) have also been used. The measured quantities are usually the ratio of the frequencies (or the \(g\)-factors) for the nucleus of interest and the reference nucleus. Such ratios can be experimentally determined with a part-per-billion (ppb) accuracy. However, magnetic moments of _bare_ nuclei extracted from these experiments are much less accurate. This is because the experimental data should be corrected for several physical effects, which are difficult to calculate. The main effect is the diamagnetic shielding of the external magnetic field by the electrons in the atom. The NMR results should be also corrected for the paramagnetic chemical shift caused by the chemical environment and the OP data are sensitive to the hyperfine mixing of the energy levels. Significant (and generally unknown) uncertainties of calculations of these effects often lead to ambiguities in the published values of nuclear magnetic moments.
Since the accuracy of calculations of the chemical shifts cannot be reliably assessed, the means of comparison of nuclear moments shielded by different environments in NMR measurements are rather limited. Independent determinations of nuclear magnetic moments would define uncertainties of theoretical calculations of the chemical shifts and help to assess the accuracy of NMR standards.
Reliable determination of the nuclear magnetic moments is also prompted by a new generation of QED calculations of the hyperfine splitting in highly charged ions. It was demonstrated that the magnetic sector of bound-state QED can be tested in these systems to all orders in the binding field, if the nuclear magnetic moments are accurately known. Alternatively, comparing theoretical predictions with experimental results, one can determine nuclear properties and set benchmark tests for nuclear-structure theory. A recent example is the spectroscopic determination of the nuclear charge radii of the neutron-halo nuclei \({}^{8}\)He, \({}^{11}\)Li, and \({}^{11}\)Be, which yielded unique information about the properties of these extraordinary systems.
A way to a high-precision determination of nuclear magnetic moments is to study the simplest atomic systems, the hydrogen-like ions. Measurements of the bound-electron \(g\)-factor in these systems progressed dramatically during the recent years and reached the ppb level. They led not only to a stringent test of sophisticated QED calculations but also to an improved determination of the electron mass. Extensions of these experiments to ions with a nonzero nuclear spin will provide a determination of the nuclear magnetic moments from a simple system that can be described theoretically up to a very high accuracy.
It is well known that the nuclear-spin-dependent part of the atomic \(g\)-factor \(g_{F}\) is suppressed by about 3 orders of magnitude as compared to the leading effect due to the bound-electron \(g\)-factor (see Eq. below). This imposes limitations on possible determinations of the nuclear magnetic moment from \(g_{F}\). We show here, however, that the leading effect cancels exactly in the sum of the \(g\)-factors for two hyperfine-structure levels (see Eq. below). This sum is proportional to the nuclear \(g\)-factor and, therefore, is much better suited for extracting the nuclear magnetic moment. Its calculation can be conveniently parameterized in terms of the nuclear shielding constant \(\sigma\), as given by Eq..
In this work we perform an _ab initio_ calculation of the nuclear magnetic shielding for the ground state of hydrogen-like ions. The relativistic, QED, and nuclear effects are accounted for. The main challenge is the calculation of the QED correction. To the best of our knowledge, the only attempt to address it was the estimate reported in Ref.. In this Letter, we calculate the QED correction rigorously to all orders in the binding nuclear strength parameter \(Z\alpha\) (where \(Z\) is the nuclear charge and \(\alpha\) is the fine-structure constant) and, independently, we derive the leading term of its \(Z\alpha\) expansion.
We now turn to the theory of the \(g\)-factor of a hydrogen-like ion with a nonzero spin.
\[g_{F}^{}=g_{j}\,\frac{\langle\mathbf{j}\cdot\mathbf{F}\rangle}{F(F+1)}-\frac{m}{m_{p }}\,g_{I}\,\frac{\langle\mathbf{I}\cdot\mathbf{F}\rangle}{F(F+1)}\,, \tag{1}\]where \(F\) is the total angular momentum, \(I\) is the nuclear spin, \(j\) is the electron angular momentum, \(g_{j}\) is the Dirac bound-electron \(g\)-factor, \(g_{I}=\mu/(\mu_{N}I)\) is the nuclear \(g\)-factor, \(\mu\) is the nuclear magnetic moment, \(\mu_{N}=|e|/(2m_{p})\) is the nuclear magneton, \(m\) and \(m_{p}\) are the electron and proton masses, respectively, \(\langle\mathbf{j}\cdot\mathbf{F}\rangle=[F(F+1)-I(I+1)+j(j+1)]/2\), and \(\langle\mathbf{I}\cdot\mathbf{F}\rangle=[F(F+1)+I(I+1)-j(j+1)]/2\). The higher-order corrections enter into Eq. in two ways: (i) the Dirac electron \(g\)-factor \(g_{j}\) is modified by QED and recoil effects that do not depend on nuclear spin, (ii) the free-nucleus \(g\)-factor \(g_{I}\) is shielded by the bound electron. Additional corrections, e.g., those due to the electric quadrupole interaction, are small and can be absorbed into the definition of the nuclear shielding.
For the ground state of an ion with a nuclear spin \(I>\nicefrac{{1}}{{2}}\), we introduce the combination of \(g\)-factors \(\overline{g}\),
\[\overline{g}\equiv g_{F=I+\nicefrac{{1}}{{2}}}+g_{F=I-\nicefrac{{1}}{{2}}}=-2 \frac{m}{m_{p}}\frac{\mu}{\mu_{N}I}\left(1-\sigma\right), \tag{2}\]
If both \(g\)-factors \(g_{F=I+\nicefrac{{1}}{{2}}}\) and \(g_{F=I-\nicefrac{{1}}{{2}}}\) are measured and \(\sigma\) is known from theory, the above formula determines the nuclear magnetic moment \(\mu\). For the ions with a nuclear spin \(I=\nicefrac{{1}}{{2}}\), Eq. is not applicable and the nuclear magnetic moment has to be determined from Eq..
The nuclear shielding constant \(\sigma\) defined by Eq.
\[\sigma=\sigma^{}+\delta\sigma_{\rm QED}+\delta\sigma_{\rm rec}+\delta\sigma_ {\rm BW}+\delta\sigma_{Q}\,, \tag{3}\]
The exact relativistic result for the leading-order magnetic shielding \(\sigma^{}\) was obtained analytically (for a point nucleus) and numerically.
\[\delta\sigma_{Q}=-\frac{\alpha\,(Z\alpha)^{3}\,Q\,m}{I(2I-1)\,g_{I}\,m_{p}} \,\frac{6\left[35+20\gamma-32(Z\alpha)^{2}\right]}{45\,\gamma(1+\gamma)^{2} \left[15-16(Z\alpha)^{2}\right]}\,, \tag{5}\]
We now turn to the QED correction to the nuclear magnetic shielding. It consists of the self-energy (SE) and vacuum-polarization parts, the SE being the most difficult one. The Feynman diagrams representing the SE correction contain two magnetic interactions, one with the external magnetic field (in what follows, the Zeeman interaction), \(V_{\rm zee}(r)=\frac{|e|}{2}\,\mathbf{B}\cdot(\vec{r}\times\vec{\alpha})\), and the other with the magnetic dipole nuclear field (in what follows, the hfs interaction), \(V_{\rm hfs}(r)=\frac{|e|}{4\pi}\,\mathbf{\mu}\cdot(\vec{r}\times\vec{\alpha})/r^{3}\). Formal expressions for the corresponding energy shifts can be obtained by the two-time Green's function method. Irreducible parts of the diagrams in Fig. 1(a)-(c) give rise to the _perturbed orbital_ contribution,
\[\delta E_{\rm po}=2\,\langle a|\Sigma(\varepsilon_{a})|\delta^{}a\rangle+2 \,\langle\delta^{}_{\rm hfs}a|\Sigma(\varepsilon_{a})|\delta^{}_{\rm zee }a\rangle\,, \tag{6}\]
The SE operator is defined by
\[\langle i|\Sigma(\varepsilon)|k\rangle = \frac{i}{2\pi}\int_{-\infty}^{\infty}d\omega\sum_{n}\frac{\langle in |I(\omega)|nk\rangle}{\varepsilon-\omega-u\varepsilon_{n}}\,, \tag{7}\]
The diagram in Fig. 1(d) gives rise to the _hfs-vertex_ contribution,
\[\delta E_{\rm vr,hfs}=2\,\langle a|\Gamma_{\rm hfs}(\varepsilon_{a})|\delta^{( 1)}_{\rm zee}a\rangle+2\,\langle a|\Sigma^{\prime}|\delta^{}_{\rm zee}a \rangle\,\langle V_{\rm hfs}\rangle\,, \tag{8}\]
where the prime denotes the derivative of the operator with respect to the energy argument and
\[\langle i|\Gamma_{\rm hfs}(\varepsilon)|k\rangle =\frac{i}{2\pi}\int_{-\infty}^{\infty}d\omega \tag{9}\] \[\times\sum_{n_{1}n_{2}}\frac{\langle in_{2}|I(\omega)|n_{1}k \rangle\langle n_{1}|V_{\rm hfs}|n_{2}\rangle}{(\varepsilon-\omega-u \varepsilon_{n_{1}})(\varepsilon-\omega-u\varepsilon_{n_{2}})}\,.\]
The diagram in Fig. 1(e) induces the _Zeeman-vertex_ contribution, in analogy with its hfs-vertex counterpart,
\[\delta E_{\rm vr,zee}=2\,\langle a|\Gamma_{\rm zee}|\delta^{}_{\rm hfs}a \rangle+2\,\langle a|\Sigma^{\prime}|\delta^{}_{\rm hfs}a\rangle\,\langle V _{\rm zee}\rangle. \tag{10}\]
Finally, Fig. 1(f) together with the remaining derivative terms yields the _double-vertex_ contribution,
\[\delta E_{\rm d.vr} = 2\,\langle\Lambda\rangle+\langle\Sigma^{\prime\prime}\rangle \langle V_{\rm zee}\rangle\langle V_{\rm hfs}\rangle+\langle\Gamma^{\prime}_{ \rm hfs}\rangle\langle V_{\rm zee} \tag{11}\] \[+\langle\Gamma^{\prime}_{\rm zee}\rangle\langle V_{\rm hfs} \rangle+2\,\langle\Sigma^{\prime}\rangle\,\langle a|V_{\rm zee}|\delta^{}_{ \rm hfs}a\rangle,\]
Self-energy correction to the nuclear magnetic shielding. Double line represents the electron in the binding nuclear field. Wave line terminated by a triangle represents the dipole hyperfine interaction with the nucleus and wave line terminated by a cross represents the interaction with the external magnetic field.
where \(\Lambda\equiv\Lambda(\varepsilon_{a})\) is the 4-point vertex operator,
\[\langle i|\Lambda(\varepsilon)|k\rangle =\frac{i}{2\pi}\int_{-\infty}^{\infty}d\omega\sum_{n_{1}n_{2}n_{3}}\] \[\times\frac{\langle in_{3}|I(\omega)|n_{1}k\rangle\langle n_{1}|V_{ 2\rm ee}|n_{2}\rangle\langle n_{2}|V_{\rm hfs}|n_{3}\rangle}{(\varepsilon- \omega-u\varepsilon_{n_{1}})(\varepsilon-\omega-u\varepsilon_{n_{2}})( \varepsilon-\omega-u\varepsilon_{n_{3}})}\,. \tag{12}\]
The formulas reported so far refer to the energy shifts. The corrections to the magnetic shielding are related to the energy shifts by \(\delta\sigma_{i}=\delta E_{i}\,IF(F+1)/(\mu\,B\,M_{F}\,\langle\mathbf{I}\cdot\mathbf{F} \rangle)\), where \(M_{F}\) is the projection of the total momentum \(F\). It can be shown that for the \(j=\nicefrac{{1}}{{2}}\) reference states, \(\delta\sigma_{\rm QED}\) does not depend on nuclear quantum numbers. The numerical calculation of \(\delta\sigma_{\rm SE}\) was performed along the lines developed in Ref.; its details will be reported elsewhere.
The remaining part of the QED effect is the vacuum polarization (VP). In our calculation, we include two dominant VP corrections induced by (i) modification of the electron line by the Uehling potential and (ii) modification of the hfs interaction by the free-loop VP.
Our calculational results for the SE and VP corrections are listed in Table 1, expressed in terms of the function \(D(Z\alpha)\),
\[\delta\sigma_{\rm QED}=\alpha^{2}\,(Z\alpha)^{3}\,D(Z\alpha)\,. \tag{13}\]
The SE correction is calculated for the point nucleus, whereas the VP part accounts for the finite nuclear size as well as higher-order iterations of the Uehling potential. Because of large numerical cancellations, we were able to perform our numerical SE calculations for \(Z\geq 10\) only. In order to extend our calculations to the lower-\(Z\) ions and to cross-check the numerical procedure, we also performed an analytical calculation of the leading term of the \(Z\alpha\) expansion.
\[D_{n}(Z\alpha)=\frac{8}{9\pi n^{3}}\,\left[\ln(Z\alpha)^{-2}+2\,\ln k_{0}-3\, \ln k_{3}-\frac{1817}{480}\right]\,, \tag{14}\]
Details of the analytical calculation will be reported elsewhere. The results of the numerical and the analytical calculations are in good agreement.
We now turn to the effect induced by the spatial distribution of the nuclear magnetic moment, also known as the Bohr-Weisskopf (BW) correction. Following Ref., our treatment of the BW effect is based on the effective single-particle model of the nuclear magnetic moment. Within this model, the magnetic moment is assumed to be induced by the odd nucleon with an effective \(g\)-factor, which is fitted to yield the experimental value of the nuclear magnetic moment. Under these assumptions, the BW effect can be described by the magnetization-distribution function \(F(r)\) that multiplies the standard point-dipole hfs interaction \(V_{\rm hfs}(r)\). The function \(F(r)\) is induced by the wave function of the odd nucleon, which is obtained by solving the Schrodinger equation with the Woods-Saxon potential (see Ref. for details). The BW correction \(\delta\sigma_{\rm BW}\) is obtained by reevaluating the leading-order magnetic shielding \(\sigma^{}\) with the hfs interaction \(V_{\rm hfs}\) multiplied by the magnetization-distribution function \(F(r)\). The relative uncertainty of 30% is ascribed to this correction, which is consistent with previous error estimates for this effect.
Numerical results of our calculations are presented in Table 2 and The error of the QED correction comes from the numerical uncertainty of the SE part and the estimate of uncalculated VP terms (30% of the total VP part). The error of the quadrupole contribution comes from the nuclear quadrupole moments. The largest error is due to the BW correction. Since this effect cannot be presently accurately calculated, this uncertainty sets the practical limit to which the nuclear magnetic moment can be determined from an atomic system.
\begin{table}
\begin{tabular}{c c c} \(Z\) & SE & VP \\ \hline
10 & \(-0.51\,\) & \(0.229\) \\
14 & \(-0.710\,\) & \(0.256\) \\
16 & \(-0.789\,\) & \(0.271\) \\
20 & \(-0.927\,\) & \(0.302\) \\
26 & \(-1.110\,\) & \(0.355\) \\
32 & \(-1.283\,\) & \(0.417\) \\
40 & \(-1.519\,\) & \(0.520\) \\
54 & \(-2.029\,\) & \(0.775\) \\
82 & \(-4.457\,\) & \(1.996\) \\
92 & \(-7.107\,\) & \(2.954\) \\ \end{tabular}
\end{table}
Table 1: QED corrections to the nuclear magnetic shielding.
(Color online) Individual contributions to the nuclear shielding. “NR” is the nonrelativistic contribution, “REL” is the relativistic point-nucleus contribution, “FNS” is the finite nuclear size correction, “QED” is the QED correction, “BW” is the Bohr-Weisskopf correction, “REC” is the recoil correction, and “QUAD” is the electric quadrupole correction. Note that the QED correction changes its sign between \(Z=4\) and \(5\).
For very light ions, the theoretical accuracy is limited by the recoil effect (see Fig. 2), which is known in the nonrelativistic limit only. Note that some of corrections to \(\sigma\) depend on the nuclear \(g\)-factor. This dependence, however, is so weak that it can be safely ignored in the determination of the magnetic moments.
Summarising, we have presented _ab initio_ calculations of the nuclear shielding in hydrogen-like ions, which account for relativistic, nuclear, and QED effects. The present theory permits determination of nuclear magnetic moments with fractional accuracy ranging from \(10^{-9}\) in the case of \({}^{17}\)O\({}^{7+}\) to \(10^{-5}\) for \({}^{209}\)Bi\({}^{82+}\). This Letter is primarily focused on nuclei with spin \(I>\nicefrac{{1}}{{2}}\), but the case of \(I=\nicefrac{{1}}{{2}}\) is only slightly more complicated. Then, the nuclear-spin-independent part of \(g_{F}\) in Eq. can be cancelled approximately, by taking a difference of the \(g\)-factors \(g_{F}\) for two different isotopes of the same element.
Modern experiments on \(g\)-factors of hydrogen-like ions have achieved the accuracy of a few parts in \(10^{11}\) but so far have been restricted to ions with spinless nuclei. Their extention to the nuclei with spin requires driving the hfs transition and measuring the \(g\)-factor of an atom in a hyperfine excited state. These are significant complications but they do not make an experiment prohibitively difficult.
Stimulating discussions with K. Blaum and G. Werth are gratefully acknowledged. Z.H. was supported by the Alliance Program of the Helmholtz Association (HA216/EMMI). V.A.Y. was supported by the Helmholtz Association (Nachvuchsgruppe VH-NG-421). K.P. acknowledges support by NIST Precision Measurement Grant PMG 60NANB7D6153.
|
10.48550/arXiv.1106.4899
|
QED theory of the nuclear magnetic shielding in hydrogen-like ions
|
V. A. Yerokhin, K. Pachucki, Z. Harman, C. H. Keitel
| 2,521
|
10.48550_arXiv.1706.07589
|
###### Abstract
We study marginally compact macromolecular trees that are created by means of two different fractal generators. In doing so, we assume Gaussian statistics for the vectors connecting nodes of the trees. Moreover, we introduce bond-bond correlations that make the trees locally semiflexible. The symmetry of the structures allows an iterative construction of full sets of eigenmodes (notwithstanding the additional interactions that are present due to semiflexibility constraints), enabling us to get physical insights about the trees' behavior and to consider larger structures. Due to the local stiffness the self-contact density gets drastically reduced.
## I Introduction
In nature, many objects can be successfully represented through fractal models. Examples are provided by lungs, plants, proteins, and chromatin, to name only a few. Also man-made materials, such as super-repellent surfaces, porous cements, super-lenses, and supercapacitors, can be build in a fractal way in order to make a better performance. The purpose of many of these examples requests an effective usage of the space provided for them. This challenge is usually connected to a very dense packing of the objects and at the same time to a huge surface needed for their function (e.g., surface available for charge in case of supercapacitors). Thus, in the best case almost all their constituents build a surface, e.g., for compact objects in three dimensional space consisting of \(N\) units and having size \(R\sim N^{1/3}\), the surface \(A\) scales as \(A\sim R^{3}\sim N\).
With respect to the biological and technological examples listed above, it is worth mentioning another actively studied system - the melt of nonconcatenated and unknotted ring polymers - that have been surmised to be marginally compact. However, the marginal compactness of ring melts is controversially argued, partly due to the clever theoretical argument that the marginal compactness leads to a logarithmic divergence of the self-contact density. In a recent work by some of us, it was suggested a practical way out of this difficulty. There we have studied the fractal trees of Ref. (see tree \(\mathcal{T}_{1}\) of Fig. 1) that are by construction marginally compact. These toy-structures, not aiming to describe the full complexity of examples such as given by Refs., allowed us to show that a simple ingredient that can suppress the divergent behavior of the self-contact density \(\hat{\rho}_{c}\) is the linear spacers between branching points of the trees.
The present study focuses on another aspect of marginally compact trees, namely on the role of local semiflexibility. The recent studies have considered Gaussian, marginally compact trees with interactions between topologically nearest neighboring beads, i.e. in the framework of generalized Rouse model. In particular, this assumption implies that the orientations of bonds are uncorrelated. However, the price one has to pay for the bond-correlations is a more complex structure of the dynamical matrix, that then in the easiest case (under freely-rotating bonds assumption for the non-adjacent bonds) contains also the elements related to the next-nearest neighboring beads. Notwithstanding this difficulty, the framework of semiflexible treelike polymers (STP) of Ref., where the semiflexibility is introduced at all beads (also at branching nodes), turned out to be very helpful in studying the relaxation dynamics of semiflexible dendrimers and fractals. Moreover, inclusion of bond-bond correlations has been shown to have a fundamental importance for NMR relaxation of dendrimers. Therefore, the semiflexiblity should also be an important ingredient for marginally compact trees.
In this work we consider marginally compact trees which are locally semiflexible. The topology of the trees is sketched in Fractal tree \(\mathcal{T}_{1}\) consists of beads of functionality 1, 2, and 3; the generalized Rouse behavior (i.e. in the absence of bond-bond correlations) of these trees has been studied in Refs.. In order to make our results more rigorous and to exemplify the role of functionality of branching nodes we introduce another fractal generator that builds marginally compact trees \(\mathcal{T}_{2}\) (see Fig. 1), which do not have any linear spacers but contain beads of functionality 4. Both trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) show all relevant scalings of marginally compact, flexible trees, when one introduces local bending rigidity. At the same time the semiflexibility leads to a swelling of the structures and hence to an increase of the higher relaxation times and to a significant suppression of self-contacts. Yet, the underlying STP framework allows us to perform a detailed analysis of eigenmodes and to reduce the computational work.
The paper is structured as follows. In the next section we provide theoretical formulas and details for the dynamical matrix in the STP framework, whose spectrafor trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) are analyzed then in Sec. III (the technical details are relegated to the Appendix). The static and dynamical properties of the trees are presented in Sec. IV. Section V closes the paper with a summary and conclusions.
## II Theoretical model
We start this section with a brief recall of the theory of semiflexible treelike polymers (STP). The STP framework allows to introduce local bending rigidity for Gaussian trees with arbitrary topology. The resulting dynamical matrix of the trees is sparse and has an analytically closed form.
In the STP theory the edges of the treelike structures represent Gaussian bonds \(\{\mathbf{d}_{i}\}\), whose orientations are constrained. For any two adjacent bonds \(\mathbf{d}_{i}\) and \(\mathbf{d}_{j}\) one has \(\langle\mathbf{d}_{i}\cdot\mathbf{d}_{j}\rangle=\pm b^{2}q_{m}\), where \(b^{2}=\langle\mathbf{d}_{i}\cdot\mathbf{d}_{i}\rangle=\langle\mathbf{d}_{j} \cdot\mathbf{d}_{j}\rangle\) is the mean-square length of each bond and \(q_{m}\) is the so-called stiffness parameter related to bead \(m\) connecting bonds \(\mathbf{d}_{i}\) and \(\mathbf{d}_{j}\). The sign determines connection of the bonds, plus sign corresponds to a head-to-tail connection and minus to two other configurations. The connection between non-adjacent bonds is taken in a freely-rotating manner, i.e., for bonds connected through the path \(\mathbf{d}_{k_{1}},...,\mathbf{d}_{k_{s}}\) the relation \(\langle\mathbf{d}_{i}\cdot\mathbf{d}_{j}\rangle=\langle\mathbf{d}_{i}\cdot \mathbf{d}_{k_{1}}\rangle\langle\mathbf{d}_{k_{1}}\cdot\mathbf{d}_{k_{2}} \rangle\cdots\langle\mathbf{d}_{k_{s}}\cdot\mathbf{d}_{j}\rangle b^{-2s}\) holds.
Given that each bond \(\mathbf{d}_{i}\) has a zero mean, the average scalar products \(\langle\{\mathbf{d}_{i}\cdot\mathbf{d}_{j}\}\rangle\) represent the covariance matrix \(\mathbf{\Sigma}=\langle(\mathbf{d}_{i}\cdot\mathbf{d}_{j})\rangle\) that fully determines the Gaussian distribution of the bonds. Furthermore, each bond vector \(\mathbf{d}_{i}\) can be represented through a difference of position vectors of beads connected through \(\mathbf{d}_{i}\), \(\mathbf{d}_{i}=\mathbf{r}_{n}-\mathbf{r}_{m}\).
\[V=\frac{3}{2}k_{B}T\sum_{i,j}(\mathbf{\Sigma}^{-1})_{ij}\mathbf{d}_{i}\cdot \mathbf{d}_{j}=\frac{3k_{B}T}{2b^{2}}\sum_{m,n}A_{nm}\mathbf{r}_{n}\cdot \mathbf{r}_{m}, \tag{1}\]
Based on the potential energy \(V\), the dynamics of a polymer can be described by a set of Langevin equations, e.g., for the position of the \(k\)th bead one has
\[\zeta\frac{\partial}{\partial t}\mathbf{r}_{k}(t)+\frac{3k_{B}T}{b^{2}}\sum_{n }A_{kn}\mathbf{r}_{n}=\mathbf{g}_{k}(t), \tag{2}\]
The conditions on the averaged scalar products used in the STP framework lead to an analytic form of \(\mathbf{A}\). Moreover, under these conditions the matrix \(\mathbf{A}\) turns out to be very sparse. Its non-vanishing elements are either diagonal or related to nearest-neighboring and next-nearest neighboring beads. For a bead of functionality \(f\) (i.e.
\[\mu_{f_{1}\ldots f_{f}}^{(f)}=\frac{f}{1-(f-1)q_{f}}+\sum_{s=1}^{f}\frac{(f_{s }-1)q_{f_{s}}^{2}}{1-(f_{s}-2)q_{f_{s}}-(f_{s}-1)q_{f_{s}}^{2}}. \tag{3}\]
For two directly connected beads of functionalities \(f_{1}\) and \(f_{2}\) one has
\[\nu_{f_{1}f_{2}}=-\frac{1-(f_{1}-1)(f_{2}-1)q_{f_{1}}q_{f_{2}}}{(1-(f_{1}-1)q_{ f_{1}})(1-(f_{2}-1)q_{f_{2}})} \tag{4}\]
and for two next-nearest neighboring beads connected through a bead of functionality \(f\) the corresponding element of \(\mathbf{A}\) is
\[\rho_{f}=\frac{q_{f}}{1-(f-2)q_{f}-(f-1)q_{f}^{2}}. \tag{5}\]
In Eqs.- the stiffness parameters \(q_{f_{i}}\) are related to the beads (junctions) of functionality \(f_{i}\). Each stiffness parameter \(q_{f_{i}}\) is bounded from above by \(1/(f_{i}-1)\); if all stiffness parameters are zero one recovers fully-flexible structures so that the dynamical matrix \(\mathbf{A}\) transforms into the connectivity (Laplacian) matrix.
We note that the STP theory allows to choose the stiffness parameters at every junction separately.
Fractal trees \(\mathcal{T}_{1}\) (a) and \(\mathcal{T}_{2}\) (b) studied in this work. Both structures are at iteration \(I=2\). Beads shown as open circles represent the trees of inital iteration \(I=1\). The sketch is aimed only to present the topology of the fractal trees; their spatial conformations may appear in vastly different forms.
Moreover, here we assume a linear dependence of the stiffness parameters from each other by taking \(q_{f}=q/(f-1)\) (with \(f\geq 2\)), so that the limits \(0\) and \(1/(f-1)\) are reached simultaneously for all junctions by varying \(q\) from \(0\) to \(1\). For beads of functionality \(1\) no stiffness parameter can be assigned. This fact is automatically taken by Eqs.- into account, where the corresponding terms due to prefactors like \((f_{i}-1)\) disappear.
Needless to say, the information about the behavior of STP in completely encoded in the eigenvalues and eigenvectors of the dynamical matrix \(\mathbf{A}\). Moreover, the symmetry of the structures allows to reduce computational efforts and to get physical insights of the relaxation behavior, as we proceed to show in Sec. III.
## III Spectrum of the dynamical matrix and the corresponding eigenmodes
The symmetry of trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) allows an iterative construction of a full set of eigenvectors. The construction procedure is rooted in the work of Cai and Chen for flexible dendrimers, which has been extended to STP treatment of semiflexible dendrimers and regular fractals.
We start with tree \(\mathcal{T}_{1}\) at iteration \(I=1\). displays the eigenmodes of the structure. Those of Fig. 2(a)-(d) leave some beads immobile, whereas in the eigenmodes of Fig. 2(e)-(i) all beads are involved. The modes (a) and (b) represent two vectors, which contain only two non-zero entries \(1/\sqrt{2}\) and \(-1/\sqrt{2}\). The ensuing (double degenerate) eigenvalue is equal to \(\mu_{3}^{}-\rho_{3}\), i.e.
\[\mathbf{A}^{}(\mathcal{T}_{1})=\left(\mu_{3}^{}-\rho_{3}\right). \tag{6}\]
Next, we consider the modes displayed in Fig. 2(c)-(d) that have the shape \((x,x,\mp y,0,0,0,\pm y,-x,-x)^{\intercal}\).
\[\tilde{\mathbf{A}}^{}(\mathcal{T}_{1})=\begin{pmatrix}\mu_{3}^{}+\rho_{3 }&\nu_{13}\\ 2\nu_{13}&\mu_{113}^{}-\rho_{3}\end{pmatrix}. \tag{7}\]
Thus, diagonalization of \(\tilde{\mathbf{A}}^{}(\mathcal{T}_{1})\) given by Eq. leads to two eigenvalues of \(\mathbf{A}\); the smallest one is related to Fig. 2(d) and the other one to Fig. 2(c).
\[\mathbf{B}^{}(\mathcal{T}_{1})=\begin{pmatrix}\mu_{3}^{}+\rho_{3}&\nu_{ 13}&0&0&\rho_{3}\\ 2\nu_{13}&\mu_{113}^{}+\rho_{3}&0&\rho_{3}&\nu_{33}\\ 0&0&\mu_{2}^{}&\nu_{12}&\rho_{2}\\ 0&2\rho_{3}&\nu_{12}&\mu_{13}^{}&\nu_{23}\\ 4\rho_{3}&2\nu_{33}&\rho_{2}&\nu_{23}&\nu_{23}^{}\end{pmatrix}. \tag{8}\]
This matrix is related to the eigenmodes of Fig. 2(e)-(i). For each of these modes the beads that are symmetric with respect to the core (blue bead) move in the same direction and with the same amplitude. Figure 2(i) depicts the translational mode \((1,\dots,1)^{\intercal}/\sqrt{N}=(1/3,\dots,1/3)^{\intercal}\) related to the eigenvalue \(\lambda_{0}=0\).
The construction of eigenmodes for tree \(\mathcal{T}_{2}\) goes in a similar manner, see The modes of Fig. 3(a)-(b) are related to the reduced matrix
\[\mathbf{A}^{}(\mathcal{T}_{2})=\left(\mu_{3}^{}-\rho_{3}\right) \tag{9}\]
Schematic sketch of eigenmodes of \(\mathcal{T}_{2}\) for \(I=1\). Beads having the same amplitude are color-coded. Black beads are immobile.
Schematic sketch of eigenmodes of \(\mathcal{T}_{1}\) for \(I=1\). Beads having the same amplitude are color-coded. Black beads are immobile.
The matrix \(\tilde{\mathbf{A}}^{}(\mathcal{T}_{2})\) corresponding to Fig. 3(c)-(d) differs slightly from \(\tilde{\mathbf{A}}^{}(\mathcal{T}_{1})\) of Eq. due to the core bead of functionality 4,
\[\tilde{\mathbf{A}}^{}(\mathcal{T}_{2})=\begin{pmatrix}\mu_{3}^{}+\rho_{3}& \nu_{13}\\ 2\nu_{13}&\mu_{114}^{}-\rho_{4}\end{pmatrix}. \tag{10}\]
Differently from \(\mathcal{T}_{1}\), tree \(\mathcal{T}_{2}\) has for \(I=1\) five eigenmodes that leave some beads (incl. the core) immobile. So the mode of Fig. 3(e) leads to the eigenvalue \(\mu_{4}^{}-\rho_{4}\), that can be formulated as \(1\times 1\) matrix
\[\tilde{\mathbf{A}}^{}(\mathcal{T}_{2})=\left(\mu_{4}^{}-\rho_{4}\right). \tag{11}\]
The remaining four eigenvalues related to Fig. 3(f)-(i) come from the diagonalization of
\[\mathbf{B}^{}(\mathcal{T}_{2})=\begin{pmatrix}\mu_{1133}^{}&2\nu_{14}&4 \rho_{3}&2\nu_{34}\\ \nu_{14}&\mu_{4}^{}+\rho_{4}&0&2\rho_{4}\\ \rho_{3}&0&\mu_{3}^{}+\rho_{3}&\nu_{13}\\ \nu_{34}&2\rho_{4}&2\nu_{13}&\mu_{114}^{}+\rho_{4}\end{pmatrix}. \tag{12}\]
As for \(\mathcal{T}_{1}\), this matrix has one vanishing eigenvalue related to the translational mode of Fig. 3(i).
The above procedure of construction of the sets of eigenmodes can be extended for higher \(I>1\). The respective reduced matrices can be build iteratively, see Appendix. Here we discuss the sizes of the reduced matrices and the degeneracy of the corresponding eigenvalues.
As it is observed for \(I=1\), the modes (a) and (b) of Figs. 2 and 3 lead to a double degenerate eigenvalue (\(\mu_{3}^{}-\rho_{3}\)). Going to the next iteration each bond gets replaced through a tree of iteration \(I=1\) (see Fig. 1), hence each bead of functionality 1 at iteration \(I=1\) leads to a pattern as displayed in Fig. 2(a) at iteration \(I=2\). At iteration \(I-1\) trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) have \((3\cdot 8^{I-1}+11)/7\) and \((4\cdot 8^{I-1}+10)/7\) beads with functionality 1, respectively. Thus the degeneracy of eigenvalue (\(\mu_{3}^{}-\rho_{3}\)) at iteration \(I\) is \((3\cdot 8^{I-1}+11)/7\) for \(\mathcal{T}_{1}\) and \((4\cdot 8^{I-1}+10)/7\) for \(\mathcal{T}_{2}\). For tree \(\mathcal{T}_{2}\) each bond of the previous iteration will lead to the pattern of Fig. 3(e). Hence the degeneracy of eigenvalue (\(\mu_{4}^{}-\rho_{4}\)) at iteration \(I\) is equal to the number of bonds in \(\mathcal{T}_{2}\) at iteration \(I-1\), i.e. to \(8^{I-1}\).
Now, going from one iteration to the next (\(I-1\to I\)), two next-nearest neighboring beads both of functionality 1 [such as in involved in the eigenmode of Fig. 2(a)] lead to two directly connected trees \(\mathcal{T}_{1}\) or \(\mathcal{T}_{2}\) of \(I=1\) (called leaves in the following, see Appendix). These leaves are involved in the eigenmodes, where each bead of one leaf has an opposite amplitude to that of the symmetrically equivalent bead of the other leaf. Moreover, in these modes all symmetrically equivalent beads belonging to the same leaf have the same amplitude and phase. In general, these modes lead to reduced matrices \(\mathbf{A}^{(n)}(\mathcal{T}_{1})\) and \(\mathbf{A}^{(n)}(\mathcal{T}_{2})\) whose iterative construction for \(n=2,\ldots,I\) is discussed in the Appendix.
\[S(n)=\begin{cases}\frac{\sqrt{13}-1}{6\sqrt{13}}(4+\sqrt{13})^{n}+\frac{\sqrt{ 13}+1}{6\sqrt{13}}(4-\sqrt{13})^{n}\text{ for }\mathcal{T}_{1},\\ \frac{\sqrt{37}-1}{6\sqrt{37}}\left(\frac{7+\sqrt{37}}{2}\right)^{n}+\frac{ \sqrt{37}+1}{6\sqrt{37}}\left(\frac{7-\sqrt{37}}{2}\right)^{n}\text{ for } \mathcal{T}_{2}.\end{cases} \tag{13}\]
Following the above discussion, the degeneracy of each eigenvalue stemming from \(\mathbf{A}^{(n)}\) appearing for the trees at iteration \(I\geq n\) is
\[D(n)=\begin{cases}(3\cdot 8^{I-n}+11)/7\text{ for }\mathcal{T}_{1},\\ (4\cdot 8^{I-n}+10)/7\text{ for }\mathcal{T}_{2}.\end{cases} \tag{14}\]
The size \(\hat{S}(n)\) of \(\tilde{\mathbf{A}}^{(n)}(\mathcal{T}_{2})\) is equal to \(S(n)\) of \(\mathcal{T}_{2}\),
\[\hat{S}(n)=\frac{\sqrt{37}-1}{6\sqrt{37}}\left(\frac{7+\sqrt{37}}{2}\right)^{ n}+\frac{\sqrt{37}+1}{6\sqrt{37}}\left(\frac{7-\sqrt{37}}{2}\right)^{n}, \tag{15}\]
and the degeneracy of each ensuing eigenvalue at iteration \(I\geq n\) is (_vide supra_)
\[\hat{D}(n)=8^{I-n}. \tag{16}\]
Apart from matrices \(\mathbf{A}^{}(\mathcal{T}_{1}),\ldots,\mathbf{A}^{(I)}(\mathcal{T}_{1})\) for \(\mathcal{T}_{1}\) or \(\mathbf{A}^{}(\mathcal{T}_{2}),\ldots\mathbf{A}^{(I)}(\mathcal{T}_{2})\) and \(\tilde{\mathbf{A}}^{}(\mathcal{T}_{2}),\ldots,\tilde{\mathbf{A}}^{(I)}( \mathcal{T}_{2})\) for \(\mathcal{T}_{2}\), there appear for each tree (at iteration \(I\)) one matrix \(\tilde{\mathbf{A}}^{(I)}\) and one matrix \(\mathbf{B}^{(I)}\).
\[\widetilde{S}(I)=\begin{cases}\frac{\sqrt{13}+2}{6\sqrt{13}}(4+\sqrt{13})^{I}+ \frac{\sqrt{13}-2}{6\sqrt{13}}(4-\sqrt{13})^{I}\text{ for }\mathcal{T}_{1},\\ \frac{\sqrt{37}+5}{6\sqrt{37}}\left(\frac{7+\sqrt{37}}{2}\right)^{I}+\frac{ \sqrt{37}-5}{6\sqrt{37}}\left(\frac{7-\sqrt{37}}{2}\right)^{I}\text{ for } \mathcal{T}_{2}\end{cases} \tag{17}\]
and of \(\mathbf{B}^{(I)}\) is
\[S_{B}(I)=1+\begin{cases}2\widetilde{S}(I)\text{ for }\mathcal{T}_{1},\\ \widetilde{S}(I)+\hat{S}(I)\text{ for }\mathcal{T}_{2}.\end{cases} \tag{18}\]
Finally, it is a simple matter to check that for \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) the total number of eigenvalues, \(S_{B}(I)+\widetilde{S}(I)+\sum_{n=1}^{I}D(n)S(n)\) and \(S_{B}(I)+\widetilde{S}(I)+\sum_{n=1}^{I}[D(n)S(n)+\widehat{D}(n)\hat{S}(n)]\), respectively, is exactly equal to the number of beads at iteration \(I\), \(N(I)=8^{I}+1\). This shows that the constructed sets of eigenmodes are complete.
In we exemplify the spectra for \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) having stiffness parameter \(q=0\) (fully-flexible case) and \(q=0.9\) (semiflexible case). As it is typical for semiflexible trees, switching on the stiffness leads to an increase of higher eigenvalues (due to the restricted local vibrations) and a decrease of the lower ones (due to the growth of the trees' size). Here, the lower eigenvalues scale with the mode number \(p\) as \(\lambda_{p}\sim p^{5/3}\), notwithstanding their non-smooth behavior reflecting the degeneracy of eigenvalues. The exponent \(5/3\) is directly related to the spectral dimension \(d_{s}=6/5\), \(2/d_{s}=5/3\), that determines the scaling of density of states, \(h(\lambda)\sim\lambda^{d_{s}/2-1}\). Thus, we observe that the local bending rigidity does not affect the spectral dimension.
For many quantities related to global physics the lowest eigenvalues play a major role. Looking at one can observe that the lowest non-vanishing eigenvalue \(\lambda_{1}\) is (almost) equal for \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) in case of \(q=0\) and it is slightly higher for \(\mathcal{T}_{2}\) for \(q=0.9\), see also Table 1. This eigenvalue comes from the matrix \(\tilde{\mathbf{A}}^{(I)}\) and related to the eigenmode in which the largest branches move as whole, such as depicted in case (d) of Figs. 2-3 for \(I=1\). Going to the second smallest eigenvalue \(\lambda_{2}\), one observes large deviations between the structures, see Table 1 and Especially in the semiflexible case (\(q=0.9\)) the difference is almost given by factor two. Eigenvalue \(\lambda_{2}\) follows from matrix \(\mathbf{B}^{(I)}\) and related to the mode such as displayed in Figs. 2-3(h). This mode involves motion of side-chains as whole that are longer in case of tree \(\mathcal{T}_{1}\) leading hence for this tree to a smaller \(\lambda_{2}\).
The gyration radius does not provide information about deviations from the spherical shape. For this one has to consider the eigenvalues \((\sigma_{1},\sigma_{2},\sigma_{3})\) of the gyration tensor, such that \(\sigma_{1}>\sigma_{2}>\sigma_{3}\) and \(\sigma_{1}+\sigma_{2}+\sigma_{3}=R_{g}^{2}\) hold.
\[\langle A_{d}\rangle=\left\langle\frac{\sum_{i<j}^{3}(\sigma_{i}-\sigma_{j})^{ 2}}{2R_{g}^{4}}\right\rangle \tag{20}\]
and
\[\langle P_{d}\rangle=\left\langle\frac{\prod_{i=1}^{3}(3\sigma_{i}-R_{g}^{2}) }{2R_{g}^{6}}\right\rangle. \tag{21}\]
The limiting values for asphericity \(\langle A_{d}\rangle\) are \(0\) for spherical shape and \(1\) for rodlike shape. The prolateness \(\langle P_{d}\rangle\) takes negative values from \((-1/8,0)\) for oblate shapes and positive values from \(\) for prolate shapes. As for asphericity, if prolateness is zero, the shape of the structure is spherical.
\[\langle A_{d}\rangle=\frac{15}{2}\int_{0}^{\infty}\mathrm{d}y\sum_{k=1}^{N-1} \frac{y^{3}}{(\lambda_{k}+y^{2})^{2}}\left[\prod_{j=1}^{N-1}\frac{\lambda_{j}} {\lambda_{j}+y^{2}}\right]^{\frac{3}{2}} \tag{22}\]
and
\[\langle P_{d}\rangle=\frac{105}{8}\int_{0}^{\infty}\mathrm{d}y\sum_{k=1}^{N-1} \frac{y^{5}}{(\lambda_{k}+y^{2})^{3}}\left[\prod_{j=1}^{N-1}\frac{\lambda_{j}} {\lambda_{j}+y^{2}}\right]^{\frac{3}{2}}, \tag{23}\]
In we plot average asphericity \(\langle A_{d}\rangle\) and prolateness \(\langle P_{d}\rangle\) for trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) of different size \(N(I)\) and stiffness \(q\). First, one can see that both trees have an aspheric shape that for high iterations \(I\) saturates to an universal value \(\langle A_{d}\rangle\simeq 0.22\) for all considered values of \(q\). Thus, the trees are less aspherical than ideal linear chains or combs and more aspherical than ideal stars with \(f>4\) arms. Furthermore, both trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) are prolate, given that \(\langle P_{d}\rangle>0\). For larger iterations the data collapse for all considered values of \(q\) on \(\langle P_{d}\rangle\simeq 0.088\) for \(\mathcal{T}_{1}\) and \(\langle P_{d}\rangle\simeq 0.092\) for \(\mathcal{T}_{2}\). The latter prolateness value of \(0.092\) for \(\mathcal{T}_{2}\) is close to that of the 4-arm-star. Tree \(\mathcal{T}_{1}\) is less prolate (that is also evident from the topology of the tree, Fig. 1), the corresponding value \(0.088\) lies between that of the 4-arm and 5-arm-stars.
While the mean-square gyration radius and the shape parameters can be calculated based on the eigenvalues
Double-logarithmic representation of the mean-square gyration radii \(\langle R_{g}^{2}\rangle\) of trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) as function of total number of beads \(N(I)=8^{I}+1\) for different values of the stiffness parameter \(q\). Inset shows rescaled radii with the minimal nonvanishing eigenvalue \(\lambda_{1}\).
Half-logarithmic representation of the asphericity \(\langle A_{d}\rangle\) and prolateness \(\langle P_{d}\rangle\) of trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) as function of total number of beads \(N(I)=8^{I}+1\) for different values of the stiffness parameter \(q\). As expected for large self-similar objects (here for \(N>500\)), both characteristics saturate to a plateau.
Here we consider the equilibrium density of contacts and the form factors of the trees. Both characteristics can be calculated based on the matrix of equilibrium mean-square distances \(\mathbf{L}=(L_{ij})\), where \(L_{ij}\) gives the mean-square distance between monomers \(i\) and \(j\) (in the units of \(b^{2}\)).
\[L_{ij}=A_{ii}^{\dagger}+A_{jj}^{\dagger}-2A_{ij}^{\dagger}, \tag{24}\]
Given that the singularity of the matrix \(\mathbf{A}\) comes from the translational mode \(\mathbf{v}_{0}=(1,1,\ldots,1)/\sqrt{N}\) [such as depicted in Figs. 2-3(i)] that leads to the eigenvalue \(\lambda_{0}=0\), the pseudo-inverse of \(\mathbf{A}\) can be readily computed,
\[\mathbf{A}^{\dagger}=(\mathbf{A}-\mathbf{v}_{0}\otimes\mathbf{v}_{0})^{-1}+ \mathbf{v}_{0}\otimes\mathbf{v}_{0}. \tag{25}\]
Now, the probability \(p_{ij}\) that two monomers (say, \(i\) and \(j\)) are in contact is given by \(p_{ij}=(2\pi L_{ij}/3)^{-3/2}\). With this, the contact density (i.e., number of contacts per monomer) reads
\[\hat{\rho}_{c}=\frac{1}{N}\sum_{i<j}\left(\frac{3}{2\pi L_{ij}}\right)^{3/2}. \tag{26}\]
In we show the contact density for different values of stiffness parameter \(q\). Introducing stiffness leads to a tremendous reduction of the number of contacts. Moreover, this effect is more striking for larger trees. For fully-flexible (\(q=0\)) tree \(\mathcal{T}_{1}\) at iteration \(I=4\) the number of contacts per bead is higher than two, whereas introducing semiflexibility to this tree leads, e.g. for \(q=0.9\), to less than one contact per bead. Generally, tree \(\mathcal{T}_{1}\) has lower contact density than \(\mathcal{T}_{2}\) of the same size \(N\) and stiffness \(q\).
The internal organization of macromolecules is studied in scattering experiments by looking at the coherent intramolecular form factor \(F(k)=\frac{1}{N}\sum_{i,j}^{N}(\exp[\mathrm{i}\mathbf{k}\cdot(\mathbf{r}_{i}- \mathbf{r}_{j})])\). For Gaussian distributed \(\{\mathbf{r}_{i}\}\) the form factor \(F(k)\) can be formulated in terms of the distance matrix \(\mathbf{L}\),
\[F(k)=\frac{1}{N}\sum_{i,j}^{N}\exp\left[-\frac{k^{2}b^{2}}{6}L_{ij}\right]. \tag{27}\]
In we plot the form factor of trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) at iteration \(I=4\) for different values of the stiffness parameter \(q\) using Kratky representation. Moreover, we rescale the wave vector by taking \(Q=k\sqrt{\langle R_{g}^{2}\rangle}\). In this representation all data for \(Q\lesssim 4\) collapse. For higher \(Q\) the data for stiffer structures lie above those of the flexible ones, reflecting more swollen local organization of the trees. In the intermediate region of \(1<Q<10\) the data approach scaling, \(F(k)\sim k^{-3}\). The differences at rather large \(Q\simeq 10\) reflect their local character, hence for higher iterations \(I\) they are expected to be less relevant.
We close the discussion of static properties of the trees and proceed to the dynamics of the structures. First, we consider the mean-square displacement (MSD) of monomers averaged over the whole structure, that follows from Eq.
\[\overline{\langle(\mathbf{r}(t)-\mathbf{r})^{2}\rangle}=\frac{2b^{2}}{N} \left(\frac{t}{\tau_{\mathrm{mon}}}+\sum_{p=1}^{N-1}\frac{1-e^{-t\lambda_{p} /\tau_{\mathrm{mon}}}}{\lambda_{p}}\right), \tag{28}\]
Half-logarithmic representation of the number of self-contacts per monomer for trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) as function of total number of beads \(N(I)=8^{I}+1\) for different values of the stiffness parameter \(q\).
The results for MSD of the trees at iteration \(I=5\) are presented in Apart from evident scaling \(t^{1}\) for \(t\ll\tau_{\rm mon}\) and \(t\gg\tau_{\rm mon}N^{5/3}\), there is subdiffusion \(t^{2/5}\) at intermediate times. The exponent \(2/5\) is closely related to the spectral dimension \(d_{s}=6/5\): The relation \(2/5=1-d_{s}/2\) follows straightforwardly from Eq. if one replaces there the sum through an integral, \(\sum\cdots\rightarrow\int{\rm d}\lambda h(\lambda)\dots\), where \(h(\lambda)\sim\lambda^{d_{s}/2-1}\) is the density of states. The subdiffusive exponent is robust under introduction of stiffness, the MSD of beads belonging to stiffer structures is slightly higher at intermediate times.
In the mechanical relaxation experiments one measures responses to external strain fields.
\[G(t)=\frac{ck_{B}T}{(N-1)}\sum_{p=1}^{N-1}\exp\left[-\frac{2\lambda_{p}t}{\tau _{\rm mon}}\right], \tag{29}\]
The development of \(G(t)\) with time is exemplified for trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) on Also there we plot experimentally relevant frequency representatives of \(G(t)\), the storage \(G^{\prime}(\omega)\) and loss \(G^{\prime\prime}(\omega)\) moduli,
\[G^{\prime}(\omega)=\frac{ck_{B}T}{(N-1)}\sum_{p=1}^{N-1}\frac{(\omega\tau_{\rm mon }/2\lambda_{p})^{2}}{1+(\omega\tau_{\rm mon}/2\lambda_{p})^{2}} \tag{30}\]
and
\[G^{\prime\prime}(\omega)=\frac{ck_{B}T}{(N-1)}\sum_{p=1}^{N-1}\frac{\omega\tau _{\rm mon}/2\lambda_{p}}{1+(\omega\tau_{\rm mon}/2\lambda_{p})^{2}}. \tag{31}\]
The initial value of the shear-stress relaxation modulus, \(G(t=0^{+})=ck_{B}T\), is given by the affine shear elasticity of a system of ideal springs. At the intermediate times the \(G(t)\) decays algebraically (here with the exponent \(-3/5=-d_{s}/2\)) that readily follow from the behavior of the density of states \(h(\lambda)\). At long times due to the finite size of structures one gets an exponential cut-off related to the minimal eigenvalue \(\lambda_{1}\), see Table 1. Exceptionally at initial times, the \(G(t)\) for semiflexible (\(q=0.9\)) trees decays faster than that of the flexible trees \(q=0\). (One finds corresponding deviations for \(G^{\prime}(\omega)\) or \(G^{\prime\prime}(\omega)\) at high frequencies.) This behavior shows fast local vibrations in semiflexible trees due to the locally restricted bonds, that are also manifested in the eigenvalues for large mode number \(p\) in Correspond
(top) Double-logarithmic representation of the shear-stress relaxation modulus \(G(t)\) for trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) at iteration \(I=5\) having different values of stiffness parameter \(q\). Inset shows \(G(t)\) with rescaled by factor \(\lambda_{1}\) time. (bottom) Double-logarithmic representation of the storage \(G^{\prime}(\omega)\) and loss \(G^{\prime\prime}(\omega)\) moduli corresponding to \(G(t)\) of the main top plot. For all curves \(ck_{B}T=1\) is taken.
Double-logarithmic representation of the monomeric MSD for trees \(\mathcal{T}_{1}\) and \(\mathcal{T}_{2}\) at iteration \(I=5\) having different values of stiffness parameter \(q\).
Moreover, as one expects for self-similar fractal objects of spectral dimension \(d_{s}\), we find in the intermediate frequency regime that
\[G^{\prime}(\omega)\approx G^{\prime\prime}(\omega)\sim\omega^{d_{s}/2}. \tag{32}\]
## V Summary and conclusions
In summary, in this work we have studied marginally compact trees that are created by means of two fractal generators. We focused on the role of local stiffness for the typical static and dynamical characteristics of the trees. We have shown that introduction of stiffness leads to an increase of size \(R\) of the structures. Nevertheless the structures remain compact, by showing a \(R\sim N^{1/3}\) scaling. Moreover, the static form factor approaches for large structures an intermediate \(F(k)\sim k^{-3}\) behavior. The ensuing exponent can be assigned, from one side, to the fractal dimension \(d_{f}=3\) and, from another side, to a fractal surface with dimension \(d_{A}=3\). (We remind that the objects with a smooth surface, e.g., a ball, have \(d_{A}=2\).) Furthermore, the shape of the trees is not spherical and the corresponding asphericity and prolateness parameters for large enough structures are independent of the stiffness and the tree structure. At the same time the semiflexibility influences tremendously the density of self-contacts that gets drastically reduced with growing stiffness. In the dynamics, the scaling of the relaxation times, \(\tau_{p}\sim(N/p)^{5/3}\), is reflected in the monomeric mean-square displacement or in the shear-stress relaxation modulus by showing at intermediate times the behavior \(t^{2/5}\) or \(t^{-3/5}\), respectively.
Coming back to recent paper by some of us, where we have shown that the linear spacers reduce the number of contacts, here we have suggested another recipe for suppression of the self-contact density by introducing local stiffness. We note that so far these findings were demonstrated for ideal trees. In this respect it will be interesting to look on the excluded volume and finite extensibility effects in the future.
|
10.48550/arXiv.1706.07589
|
Marginally compact fractal trees with semiflexibility
|
Maxim Dolgushev, Adrian L. Hauber, Philipp Pelagejcev, Joachim P. Wittmer
| 6,741
|
10.48550_arXiv.2502.17349
|
## 1 Introduction
Fragment-based drug discovery strategically assembles molecular fragments to create stable and effective drug candidates. A key challenge is linker generation, which involves designing molecular linkers that connect fragments while ensuring their validity. In this work, we categorize existing linker generation models into Point Cloud Free (PC-Free) and Point Cloud Aware (PC-Aware) models, distinguished by whether they account for the 3D conformation of fragments when determining the bonding topology of the complete molecule.
(a) Qualitative comparison of our method and existing pipelines on diversity of valid molecule samples measured on five metrics. (b) Trade-off between diversity and validity in baseline models and our hybrid approach overcoming these limitations.
## Point Cloud Free Models (\(\uparrow\) Diversity \(\downarrow\) Validity).
These models, such as FFLOM (Jin et al., 2023) and DeLinker (Imrie et al., 2020), generate diverse bonding topologies based on fragment connectivity while excluding fragment 3D conformations, followed by 3D conformation prediction for the sampled topology using conformation generators (Riniker and Landrum, 2015; Xu et al., 2022, 2021). By disregarding 3D fragment conformations during topology generation, they maximize entropy in the sampling distribution, enhancing diversity. However, this often results in poor alignment between the linker's conformation and the predefined fragment geometry, leading to invalid, high-energy molecules.
## Point Cloud Aware Models (\(\downarrow\) Diversity \(\uparrow\) Validity).
DiffLinker (Igashov et al., 2024) and 3DLinker (Huang et al., 2022) achieve high validity by incorporating fragment conformations when determining bonding topology, ensuring spatial alignment between fragments and generated linkers. However, their strict spatial constraints lower sampling entropy, limiting the exploration of diverse topologies. This constrained search space increases the risk of overfitting, making it challenging to generate topologically diverse drug candidates.
To overcome the diversity-validity trade-off, we propose **HybridLinker**, a framework that integrates the strengths of PC-Free and PC-Aware models. By guiding a PC-Aware model with bonding topologies sampled from a PC-Free model, HybridLinker enhances diversity while preserving validity. At its core, we introduce LinkerDPS, the first diffusion posterior sampling (DPS) method that bridges molecular topology and point cloud space via an energy-inspired function. This approach transfers the highly diverse samples of PC-Free models into the validity-focused distribution of PC-Aware models, achieving a balanced trade-off between diversity and validity. illustrates how HybridLinker surpasses existing methods.
We evaluate HybridLinker on the ZINC (Irwin et al., 2020) test dataset, a standard benchmark in drug discovery, demonstrating its ability to generate diverse and valid molecules from fragment inputs. By leveraging zero-shot cooperation between PC-Free and PC-Aware models, HybridLinker surpasses existing methods in diversity of valid molecules and its enhanced diversity also drives superior performance in property optimization tasks, highlighting its potential as a foundational model. Moreover, HybridLinker's success validates LinkerDPS, showcasing its versatility across various domains and applications.
Our contributions can be summarized as follows:
* We are the first to highlight the trade-off between diversity and validity in a fair comparison of Point Cloud Free and Point Cloud Aware linker generation models.
* We present HybridLinker, a simple yet effective framework that integrates pretrained Point Cloud Free and Point Cloud Aware linker generation models within a two-step generation pipeline, enabling zero-shot inference inheriting their strengths.
* We introduce LinkerDPS, the first DPS method beyond the image domain, operating across molecular topology and point cloud spaces. It bridges these domains through an energy-inspired cross-domain function, enabling effective topology-guided molecular point cloud generation. Its cross-domain guidance overcomes challenges in point cloud space by leveraging topology space as an intermediary, offering a wide range of potential applications.
* We evaluate HybridLinker across both the fundamental task of diverse sampling and the application-driven task of property optimization, demonstrating its potential as a foundation model. Furthermore, validated performance of LinkerDPS in linker generation show its potential for broader applications, particularly in challenging point cloud tasks benefiting from topological guidance.
\begin{table}
\begin{tabular}{l|c c c c} \hline \hline
## Task** & **Linker Gen. (Our Setup)** & **2D Linker Gen.** & **Scaffold Hop.** & **PROTAC Design
\\ \hline
## Input
& \(G_{1}\), \(G_{2}\) & \(\mathcal{T}_{1}\), \(\mathcal{T}_{2}\) & \(F_{1}\), \(F_{2}\) & \(G_{1}\), \(G_{2}\) \\
## Output
& \(G^{\prime}\mid G_{1},G_{2}\subset G^{\prime}\) & \(\mathcal{T}^{\prime}\mid\mathcal{T}_{1},\mathcal{T}_{2}\subset\mathcal{T}^{\prime}\) & \(G^{\prime}\mid F_{1},F_{2}\subset G^{\prime}\) & \(G^{\prime}\mid\mathbf{T}_{1}[G_{1}],\mathbf{T}_{2}[G_{2}]\subset G^{\prime}\) \\ \hline
## Task** & **Target-Aware Drug Design** & **Fragment Growing** & **De Novo Gen.** & **Conformation Gen.
\\ \hline
## Input
& \(G_{1}\), \(G_{2}\), \(M^{\prime}\) & \(G^{\star}\) & None & \(\mathcal{T}^{\star}\) \\
## Output
& \(G^{\prime}\mid G_{1},G_{2}\subset G,M^{\prime}\) & \(G^{\prime}\mid G^{\star}\subset G^{\prime}\) & \(G^{\prime}\) & \(G^{\prime}=(\mathcal{T}^{\star},R^{\prime})\mid\mathcal{T}^{\star}\) \\ \hline \hline \end{tabular}
\end{table}
Table 1: Comparison of molecular generation tasks based on inputs and outputs. It reveals that our setup, linker generation, is a general formulation of fragment-based molecular generation. \(G\) represents a complete molecule represented as a 3D graph, while \(F\), another 3D graph, is a partial molecule with \(|F|\ll|G|\) in most cases. \(M\) denotes the binding pocket of target proteins, providing structural constraints. \(\mathcal{T}\) refers to the bonding topology of a molecule, defined by atom types and chemical bonds. \(\mathbf{T}\) represents an arbitrary transformation in SE (e.g., rotations and translations).
## 2 Background
### Molecule Generation
Molecular RepresentationsA molecule is commonly represented as a **molecular graph**, where nodes correspond to atoms, and edges represent covalent bonds (Huang et al., 2024). We formalize this below.
## Definition 2.1** (3D Molecular Graph).: **A 3D molecular graph
is defined as a triplet \(G=(V,E,R)\), where:
* \(V\) is the list of **atoms** in the molecule.
* \(E\) is the set of **chemical bonds**, represented as an adjacency tensor.
* \(R\in\mathbb{R}^{|V|\times 3}\) denotes the **3D conformation**, specifying the spatial coordinates of each atom.
Given a molecular graph with \(N\) atoms, we adopt the following encoding scheme: \(V\) is one-hot encoded as \(V\in\{0,1\}^{N\times A}\), where \(A\) is the number of atom types. \(E\) is stored as an adjacency tensor \(E\in\{0,1\}^{N\times N\times B}\), where \(B\) is the number of bond types.
We also introduce two key sub-representations:
* The **bonding topology**, denoted as \(\mathcal{T}=(V,E)\), which captures the connectivity of atoms without considering spatial information.
* The **point cloud representation**, given by \((V,R)\), which encodes atom types and 3D positions but omits bond information.
Thus, a 3D molecular graph can be equivalently expressed as \(G=(\mathcal{T},R)\), combining both connectivity and geometry. We write \(\mathbb{P}_{G}=\mathbb{P}_{\mathcal{T},R}\) to denote the distribution (dataset) of valid 3D molecular graphs which can be seen as the joint distribution over the bonding topology and 3D conformation.
Recent molecular diffusion models (Hoogeboom et al., 2022; Igashov et al., 2024; Corso et al., 2023; Schneuting et al., 2024; Corso et al., 2023a) represent molecules as point clouds \((V,R)\), disregarding explicit chemical bonds (\(E\)) during the denoising steps. In these methods, the discrete point cloud representation is embedded in a continuous space \(\mathbb{R}^{d}\), where \(d=NA+3|V|\) accounts for both atom types (one-hot encoded) and spatial coordinates. Diffusion denoising steps operate in \(\mathbb{R}^{d}\), gradually refining the continuous representation. After sampling, the discrete atom representation is recovered using an argmax operation on \(V\). Chemical bonds are inferred via a post-hoc bond predictor as \(E=\mathcal{E}(\mathcal{F},V,R)\) or \(\mathcal{E}(\emptyset,V,R)\) depending on fragment conditions \(\mathcal{F}\), producing \(G=(V,E,R)\).
Tasks in Molecule GenerationThis paper addresses the linker generation task, a subfield within the broader domain of large molecule generation. As shown in Table 1, we contextualize this task by comparing it with other related tasks in the domain, including topology linker generation, scaffold hopping, PROTAC design, fragment growing, de novo molecule generation, and conformation generation. Related works are summarized in Appendix H.
ValidityIn prior works, molecular validity has been primarily defined in terms of satisfying the valence rule in molecular topology. In this paper, we extend its definition to also consider molecular conformation and the resulting energetic stability. A detailed discussion on validity is provided in Appendix A.
### Problem Setup: Linker Generation
Suppose we are given two molecule fragments \(G_{1}=(\mathcal{T}_{1},R_{1})\) and \(G_{2}=(\mathcal{T}_{2},R_{2})\), where \(G_{1}\) and \(G_{2}\) are subgraphs of a reference molecule \(G_{\text{ref}}=(\mathcal{T}_{\text{ref}},R_{\text{ref}})\). The reference molecule \(G_{\text{ref}}\), which contains \(N_{\text{ref}}=|V_{\text{ref}}|\) atoms, is chemically valid, and belongs to a molecular graph dataset.
Comparison of generation pipelines for PC-Free and PC-Aware models. HybridLinker is designed to leverage the strengths of both approaches, inheriting high diversity from the PC-Free model and high validity from the PC-Aware model.
Further, we set \(\mathcal{T}_{\text{cond}}=\mathcal{T}_{1}\cup\mathcal{T}_{2}\), and \(R_{\text{cond}}=\mathcal{R}_{1}\cup\mathcal{R}_{2}\).
## Linker Generation
is the task of generating new complete 3D molecular graphs \(G^{\prime}=(\mathcal{T}^{\prime},R^{\prime})\), from the following conditional distribution1 (see Figure 3):
Footnote 1: We write \(G_{1}\subset G^{\prime}\) to denote that \(G^{\prime}\) contains \(G_{1}\) as its subgraph, meaning that \(V_{1}\subset V^{\prime},E_{1}\subset E^{\prime},R_{1}\subset R^{\prime}\). We analogously write \(\mathcal{T}_{1}\subset\mathcal{T}\) to mean \(V_{1}\subset V^{\prime},E_{1}\subset E^{\prime}\).
\[G^{\prime}\sim\mathbb{P}_{G}(G\mid\underbrace{G_{1},G_{2}\subset G,|V|=N_{ \text{ref}}}_{\mathcal{F}=\mathcal{T}_{\text{cond}}\cap R_{\text{cond}}:\, \text{fragment conditions}}) \tag{1}\]
For simplicity, we use the shorthand \(\mathbb{P}_{G}(G\mid\mathcal{F})\) to denote this joint distribution which captures the topological (\(\mathcal{T}_{\text{cond}}\)) and geometric (\(R_{\text{cond}}\)) constraints imposed by a fragment condition \(\mathcal{F}\). We analogously write the marginal distributions of \(\mathcal{T},R\) as \(\mathbb{P}_{\mathcal{T}}\) and \(\mathbb{P}_{R}\), respectively.
In this setup, \(G_{1}\) and \(G_{2}\) serve as endpoints, and the task is to generate the missing linker atoms and bonds that connect \(G_{1}\) and \(G_{2}\) while maintaining its validity.
### PC-Free and PC-Aware Approaches
Existing works in linker generation are divided into two approaches, Point Cloud Free (PC-Free) and Point Cloud Aware (PC-Aware) approaches, based on awareness of fragments' conformation \(R_{\text{cond}}\) on determining the bonding topology of output molecule \(\mathcal{T}\). To clarity, the ways they sample \(\mathcal{T}\) are distinguished as following:
\[\text{PC-Free}:\mathbb{P}_{\mathcal{T}|\mathcal{T}_{\text{cond}} }(\mathcal{T}\mid\mathcal{T}_{\text{cond}}) \tag{2}\] \[\text{PC-Aware}:\mathbb{P}_{\mathcal{T}|\mathcal{T}_{\text{cond}},R_{\text{cond}}}(\mathcal{T}\mid\mathcal{T}_{\text{cond}},\boxed{R_{\text{ cond}}})\]
In the following, we explain each approach in detail with related models.
## Point Cloud Free (PC-Free) Approach.** PC-Free approaches (Jin et al., 2023; Imrie et al., 2020) model the conditional distribution as in, sequentially sampling the bonding topology \(\mathcal{T}\) conditioned on **only the fragments' bonding topology \(\mathcal{T}_{\text{cond}}\)
\[\begin{split}\mathbb{P}_{G}(G\mid\mathcal{F})&\stackrel{{ \text{Bayes}}}{{=}}\mathbb{P}_{\mathcal{T}}(\mathcal{T}\mid \boxed{\mathcal{F}})\cdot\mathbb{P}_{R|\mathcal{T}}(R\mid\mathcal{T},\mathcal{ F})\\ &\stackrel{{\text{PC-Free}}}{{\approx}}\mathbb{P}_{ \mathcal{T}}(\mathcal{T}\mid\boxed{\mathcal{T}_{\text{cond}}})\cdot\mathbb{P}_ {R|\mathcal{T}}(R\mid\mathcal{T},\mathcal{F})\\ &=:\mathbb{P}_{\text{2D}}(G\mid\mathcal{F})\end{split} \tag{3}\]
However, the approximation in, which assumes that the linker's topology is independent of \(R_{\text{cond}}\), is often inaccurate. In reality, the topology of the linker generally depends on \(R_{\text{cond}}\), as its topology-driven conformation must align with \(R_{\text{cond}}\) to ensure energetic stability. To emphasize the distinction between the true distribution \(\mathbb{P}_{G}\) and its approximation, we denote the latter as \(\mathbb{P}_{\text{2D}}\), referring to it as the **surrogate distribution**.
PC-Free models sample from by using a neural network to approximate \(\mathbb{P}_{\mathcal{T}}(\mathcal{T}\mid\mathcal{T}_{\text{cond}})\) and an off-the-shelf conformation generator \(\mathcal{R}\) which produces the 3D conformation of the molecule conditioned on \(R_{1}\) and \(R_{2}\).
\[R^{\prime}=\mathcal{R}(\mathcal{T}^{\prime},\underbrace{R_{1},R_{2}}_{R_{ \text{cond}}})\in\mathbb{R}^{|V|\times 3} \tag{4}\]
## Point Cloud Aware (PC-Aware) Approach.
In contrast, PC-Aware models like DiffLinker (Igashov et al., 2024) and 3DLinker (Huang et al., 2022) learn the distribution of bonding topology conditioned on both 3D conformation and bonding topology of fragments, as in. In particular, DiffLinker and 3DLinker incorporate atom-wise distances between linker atoms and fragments to account for \(R_{\text{cond}}\) in determining both topology and conformation, thereby resulting in a non-approximated sampling framework.
\[\mathbb{P}_{G}(G\mid\mathcal{F})=\underbrace{\mathbb{P}_{V,R}(V,R\mid\mathcal{ F})}_{\text{Generative Model}}\cdot\underbrace{\mathbb{P}_{E|V,R}(E\mid\mathcal{F},V,R)}_{E= \mathcal{E}(\mathcal{F},V,R)} \tag{5}\]
This formulation allows generative models to focus on sampling atomic positions while leveraging external bond inference models.
### Trade-Off in Diversity and Validity
Due to different sampling strategy, PC-Free and PC-Aware models exhibit an inverse relationship between diversity and validity. Specifically, PC-Free models demonstrate **high diversity but low validity**, whereas PC-Aware models achieve **high validity but low diversity**.
## Diversity
PC-Free models generate \(\mathcal{T}\) without conditioning on \(R_{\text{cond}}\), leading to greater diversity compared to PC-Aware models, which incorporate \(R_{\text{cond}}\) in sampling \(\mathcal{T}\).
\[H(\mathcal{T}\mid\mathcal{T}_{\text{cond}},R_{\text{cond}})\leq H(\mathcal{T} \mid\mathcal{T}_{\text{cond}}). \tag{6}\]
## Validity
PC-Free models face validity issues due to the inaccurate assumptions in. Since the 3D conformations of fragments, \(R_{1}\) and \(R_{2}\), are ignored when determining the bonding topology \(\mathcal{T}^{\prime}\), even a model that perfectly learns the target distribution \(\mathbb{P}_{\text{2D}}\) will still deviate from the true molecular distribution \(\mathbb{P}_{G}\). In contrast, PC-Aware models inherently overcome this limitation by explicitly training on the linker generation task in.
Problem setup for linker generation, illustrating the fragments \(G_{1}\) and \(G_{2}\) embedded in a candidate molecule \(G^{\prime}\). \(N_{\text{ref}}\) is the size of reference molecule.
### Diffusion-based Molecule Generation
The molecular diffusion model (Hoogeboom et al., 2022; Igashov et al., 2024; Corso et al., 2023b) has emerged as a leading architecture for various molecule generation tasks, including PC-Aware linker generation (Igashov et al., 2024). It leverages the reverse diffusion process to iteratively denoise a randomly sampled \(x_{T}\sim\mathcal{N}(0,I)\) into a valid point cloud \(x_{0}=(V,R)\) in \(T\) timesteps.
\[\begin{cases}x_{T}\sim\mathcal{N}(0,I)\\ dx_{t}=\left[f(x_{t},t)+g^{2}(t)\nabla_{x_{t}}\log p_{t}(x_{t})\right]dt+g(t)d \bar{W}_{t}\end{cases} \tag{7}\]
The drift coefficient is given by \(f(x_{t},t)=\frac{1}{2}\beta_{t}x_{t}\), and the diffusion coefficient is \(g(t)=\sqrt{\beta_{t}}\), where \(\beta_{t}\) is a predefined noise schedule. The term \(\nabla_{x}\log p_{t}(x)\), referred to as the score, is estimated using a trained network \(s_{\theta^{*}}\). The output denoised graph \(x_{0}\) is converted into a one-hot encoded molecular point cloud using argmax followed by a post-hoc bond predictor, as described in Definition 2.1.
## 3 Method
Rather than directly sampling \(G^{\prime}\sim\mathbb{P}_{G}\) in a single step, we propose a hybrid two-step approach that leverages both PC-Free and PC-Aware linker generation models.
### Motivation for a Hybrid Approach
Experimental ComparisonTo validate this trade-off, we evaluate PC-Free and PC-Aware models on the ZINC dataset using standard metrics: **Validity** for validity, **Uniqueness** for diversity and **V + U** for diversity of valid molecules. The experimental setup follows Section 4.1. As shown in Table 2, PC-Free models excel in diversity, while PC-Aware models perform best in validity, confirming our hypothesis. As observed in V + U score, the trade-off between diversity and validity significantly limits the practical utility of existing linker generation models. To overcome this limitation, we propose a **hybrid approach** that integrates the strengths of both PC-Free and PC-Aware models without requiring any training. Our hybrid pipeline leverages pretrained PC-Free and PC-Aware models cooperatively to achieve both high diversity and validity. illustrates the generation processes of PC-Free and PC-Aware models, along with goal of our proposed hybrid strategy of inheriting strength of two models.
### Hybrid Approach
We develop a straightforward and intuitive pipeline that integrates two pretrained models, leveraging a novel theoretical methodology to execute the process.
\[\mathbb{P}_{G,\tilde{G}}(G,\tilde{G}\mid\mathcal{F})=\mathbb{P}_{\tilde{G}}( \tilde{G}\mid\mathcal{F})\cdot\mathbb{P}_{G\mid\tilde{G}}(G\mid\tilde{G}, \mathcal{F}) \tag{8}\]
Now, we plug-in the PC-Free and PC-Aware models to implement the two steps in. In the first stage, we set \(\mathbb{P}_{\tilde{G}}\) as the PC-Free distribution \(\mathbb{P}_{\text{2D}}\), which captures the topological diversity of \(\mathbb{P}_{G}\) but also includes invalid samples.
\begin{table}
\begin{tabular}{l c c c c} \hline \hline
## Method** & **F / A** & **Uniquess (\(\uparrow\))** & **Validity (\(\uparrow\))** & **V + U (\(\uparrow\))** \\ \hline FFLOM & F & **0.840** (High) & 0.370 (Low) & 0.313 \\ DeLinker & F & 0.638 (High) & 0.575 (Low) & **0.386** \\ DiffLinker & A & 0.349 (Low) & **0.711
(High) & 0.269 \\
3DLinker & A & 0.443 (Low) & 0.654 (High) & 0.326 \\ \hline \hline \end{tabular}
\end{table}
Table 2: Trade-off between diversity and validity in PC-Free and PC-Aware models. The **F/A** column indicates whether the model belongs to PC-Free (F) or PC-Aware (A).
Conceptual comparison of sampling distributions in PC-Free, PC-Aware, and HybridLinker models. (a) PC-Free models explore a broad molecular space but often generate invalid molecules. (b) PC-Aware models focus on validity but suffer from low diversity due to spatial constraints. (c) HybridLinker leverages LinkerDPS to balance diversity and validity, enhancing exploration while maintaining correctness.
Correspondingly, \(\mathbb{P}_{G|\tilde{G}}\) in the second stage becomes sampling of valid molecule equal or similar to \(\tilde{G}\). To clarify, if \(\tilde{G}\) is invalid, we run a molecular refinement process to obtain a similar but valid molecule; otherwise, we set \(G\) to be \(\tilde{G}\). To implement the molecular refinement process, we perform posterior sampling using the validity-focused prior distribution of \(G\), learned by the PC-Aware model, and our likelihood of \(\tilde{G}\), which favors affinity with \(G\). To summarize, we implement as.
\[\mathbb{P}_{G,\text{2D}}(G,\tilde{G}\mid\mathcal{F})=\underbrace{\mathbb{P}_{ \text{2D}}(\tilde{G}\mid\mathcal{F})}_{\text{Step 1: PC-Free}}\cdot\underbrace{\mathbb{P}_{G|\text{2D}}(G\mid\tilde{G}, \mathcal{F})}_{\text{Step 2: PC-Aware}} \tag{9}\]
At a high level, our implementation can be viewed as transferring the high-entropy of the PC-Free distribution into the validity-focused PC-Aware distribution, leading to a coordinated distribution that ensures both high diversity and validity of \(G\). Specifically, our process produces a tuple comprising a valid 3D molecule sample \(G\sim\mathbb{P}_{G}\) and a surrogate \(G\sim\mathbb{P}_{\text{2D}}\). We describe the algorithmic difference of HybridLinker to PC-Aware model and PC-Free model in Appendix C.
### Posterior Sampling From \(\mathbb{P}_{G|\text{2D}}\)
The challenge in sampling \(\mathbb{P}_{G|\text{2D}}\) lies in refining an invalid surrogate molecule into a valid one, which corresponds to the inverse problem (Gutha et al., 2024; Yang et al., 2024; Chung et al., 2024a) in the molecular domain. Since this area remains unexplored, we introduce LinkerDPS, the first DPS (Chung et al., 2024a) method designed for the molecular domain. By adopting the reverse process of the diffusion-based PC-Aware model, LinkerDPS samples a refined point cloud \((V,R)\) favored by the pretrained prior distribution of valid point clouds.
\[p(V,R\mid\tilde{G},\mathcal{F})=\overbrace{\frac{p(V,R\mid\mathcal{F})}{p( \tilde{G}\mid\mathcal{F})}\cdot\overbrace{p(\tilde{G}\mid\mathcal{F})}^{\text{ validity}}}^{p(\text{V,R,\mathcal{F})}} \tag{10}\]
As shown in, leveraging Bayes' rule, we decompose \(p(V,R\mid\tilde{G},\mathcal{F})\) into the prior distribution \(p(V,R\mid\mathcal{F})\) and the likelihood \(p(\tilde{G}\mid V,R,\mathcal{F})\), which respectively ensure the validity of \((V,R)\) and its affinity with \(\tilde{G}\). Beyond the prior learned by the pretrained PC-Aware model, we design the likelihood of \(\tilde{G}=(\tilde{V},\tilde{E},\tilde{R})\) as follows. Note that we disregard the condition \(\mathcal{F}\) by considering \(V\) and \(R\) as stricter constraints. Additionally, we assume mutual independence of \(\tilde{V},\tilde{E},\tilde{R}\) given \(V\) and \(R\).
Likelihood of AtomTo ensure consistency between the surrogate and generated molecules, we enforce that the atoms of both molecules remain the same.
\[\mathbb{P}_{\tilde{V}\mid V}(\tilde{V}\mid V):=\mathbb{I}(\tilde{V}=V) \tag{11}\]
Likelihood of Bond and ConformationTo account for the cross-domain nature of bond likelihood given a point cloud, we introduce a molecular energy-inspired function \(U\), defined as
\[U(E^{*},R^{*}):=\sum_{1\leq i,j\leq|V|}\mathbb{1}_{E^{*}_{i},j\neq 0}\|R^{*}_{i }-R^{*}_{j}\|, \tag{12}\]
Two-step generation pipeline of HybridLinker. (Step 1.) diverse molecular samples are generated using pretrained PC-Free models. (Step 2.) posterior sampling is performed for each molecule from the previous step. If a molecule is invalid, LinkerDPS is applied to a diffusion-based PC-Aware model to refine it into a valid structure while maintaining similarity to the guiding molecule. Valid molecules from the first step are directly used as the final output.
We define the affinity of \(\tilde{E}\) with a given \(R\) and its likelihood as the probability that the system encompassing \(R\) adopts the bond \(\tilde{E}\). Utilizing the Boltzmann distribution, this quantity is expressed as being proportional to \(\exp\Bigl{(}-U(\tilde{E},R)\Bigr{)}\). Similarly, to model the likelihood of \(\tilde{R}\), we employ a Gaussian kernel \(\kappa\) to quantify the similarity between \(\tilde{R}\) and \(R\) as in, and we define the likelihood of \(\tilde{R}\) to be proportional to \(\kappa(\tilde{R},R)\):
\[\kappa(\tilde{R},R)=e^{-\|\tilde{R}-R\|^{2}}. \tag{13}\]
Combining these components, we formulate the joint likelihood of \((\tilde{E},\tilde{R})\) as
\[p(\tilde{E},\tilde{R}\mid R):=\frac{1}{Z}\kappa(\tilde{R},R)e^{-U(\tilde{E},R)}, \tag{14}\]
LinkerDPS Formulation.Utilizing and to express the likelihood in, sampling \(p(V,R\mid\tilde{G})\) reduces to sampling \(R\) while keeping \(V=\tilde{V}\) fixed, formulated as
\[p(R\mid\tilde{G},V,\mathcal{F})\propto p(R\mid V,\mathcal{F})\cdot p(\tilde{E },\tilde{R}\mid R). \tag{15}\]
Accordingly, LinkerDPS implements conditional sampling of the conformation \(R\) via the reverse diffusion process:
\[dr_{t}=\left[f(r_{t},t)+g^{2}(t)\nabla_{r_{t}}\log p_{t}(r_{t}\mid\tilde{E}, \tilde{R},V)\right]dt+g(t)d\tilde{W}_{t} \tag{16}\]
Applying Bayes' rule, we express:
\[\nabla_{r_{t}}\log p(r_{t}|\tilde{E},\tilde{R},V)=\nabla_{r_{t}}\log p_{t}(r_ {t}|V)+\nabla_{r_{t}}\log p_{t}(\tilde{E},\tilde{R}|r_{t},V) \tag{17}\]
To compute \(\nabla_{r_{t}}\log p_{t}(r_{t}|V)\), we employ an inpainting strategy (Lugmayr et al., 2022) and introduce a conditional score estimator \(s^{r}_{\theta^{*}}\), derived from the pretrained score estimator \(s_{\theta^{*}}\) of the PC-Aware model:
\[s^{r}_{\theta^{*}}(r_{t},t,V,\mathcal{F})\approx\nabla_{r_{t}}\log p_{t}(r_{t }|V). \tag{18}\]
The detailed operation of this estimator is described in Appendix D.5. To evaluate \(\nabla_{r_{t}}\log p_{t}(\tilde{E},\tilde{R}\mid r_{t},V)\), we extend the DPS approximation (Chung et al., 2024) for our likelihood in, obtaining:
\[p_{t}(\tilde{E},\tilde{R}\mid r_{t},V)\approx p(\tilde{E},\tilde{R}\mid\hat{r }), \tag{19}\]
where
\[\hat{r}=\frac{1}{\sqrt{\alpha_{t}}}\left(r_{t}+(1-\alpha_{t})\cdot s^{r}_{ \theta^{*}}(r_{t},t,V,\mathcal{F})\right) \tag{20}\]
The adaptation of the DPS approximation is detailed in Appendix D.3.
\[\nabla_{r_{t}}\log p_{t}(r_{t}|\tilde{E},\tilde{R},V)=s^{r}_{\theta^{*}}(r_{t },t,V,\mathcal{F})+\nabla_{r_{t}}\log p(\tilde{E},\tilde{R}\mid\hat{r}). \tag{21}\]
The detailed computation of \(\nabla_{r_{t}}\log p(\tilde{E},\tilde{R}\mid\hat{r})\) is presented in Appendix D.4. Finally, we apply **ancestral sampling**(Ho et al., 2020) to sample from the reverse diffusion process. The sampled conformation is concatenated with \(V\) to construct the point cloud, which is subsequently converted into a complete molecular structure using the post-hoc bond predictor \(\mathcal{E}\), following the PC-Aware model. The full algorithm for LinkerDPS is provided in Appendix D.6.
## 4 Experiments
### Experiment Setup
We evaluate linker generation algorithms on for 400 fragment-linker pairs from ZINC-250K (Gomez-Bombarelli et al., 2018), which are the test data in prior linker generation work (Imrie et al., 2020). Further, we run 50 times sampling for each fragment and use 20,000 samples in total for our experiments. To evaluate generated samples' validity, we use the metric **Validity** described in Section 2.1. To score the diversity of samples, we use the followings to capture diversity in multiple perspective: **Uniqueness, Novelty, V+U, V+N, V+HD, V+FG**, and V+BM**. Uniqueness and Novelty accounts for all samples satisfying topological valence rule, but the others measures the diversity of valid molecules by counting only the valid samples.
\begin{table}
\begin{tabular}{l c c c c c c c c c} \hline \hline & \multirow{2}{*}{**F/A**} & \multirow{2}{*}{**Validity (\%)**} & \multicolumn{3}{c}{**Diversity w/ Validity**} & \multicolumn{3}{c}{**Diversity w/ Validity**} \\ \cline{4-10} & & & **Uniqueness (\%)** & **Novelty (\%)** & **V+U (\%)** & **V+N (\%)** & **V+HD** & **V+FG** & **V+BM** \\ \hline FFLOM & F & 37.07 & **84.04** & **58.29** & 31.25 & 31.32 & 6.38 & 16.37 & 14.48 \\ DeLinker & F & 57.44 & 63.78 & 31.63 & 38.57 & 38.53 & 7.32 & 16.20 & 17.33 \\ DiffLinker & A & 71.08 & 34.94 & 16.32 & 26.90 & 26.90 & 5.08 & 11.37 & 10.91 \\
3DLinker & A & 65.31 & 44.31 & 21.83 & 32.61 & 32.64 & 6.14 & 15.40 & 14.23 \\ \hline HybridLinker (FFLOM) & F+A & 69.03 & 68.39 & 44.67 & **55.10** & 55.09 & **10.81** & **23.92** & **24.59** \\ HybridLinker (DeLinker) & F+A & **77.27** & 68.09 & 35.70 & **55.02** & **55.28** & 10.21 & 21.82 & 24.23 \\ \hline \hline \end{tabular}
\end{table}
Table 3: Comparison of linker generation models across diversity and validity metrics. The F/A column indicates whether the model belongs to a PC-Free (F), PC-Aware (A), or hybrid (F+A) model. The ”Diversity w/o Validity” columns represent diversity without considering validity (only satisfying topological valence rules), while the ”Diversity w/ Validity” columns represent diversity of valid molecules.
Furthermore, we present the results of the ablation study on the design of LinkerDPS in Appendix E.
For baseline comparisons, we use all pretrained PC-Free and PC-Aware models trained on the ZINC dataset: FFLOM (Jin et al., 2023) and DeLinker (Imrie et al., 2020) for PC-Free models, and 3DLinker (Huang et al., 2022) and DiffLinker (Igashov et al., 2024) for PC-Aware models. We evaluate two implementations of HybridLinker, HybridLinker(FFLOM) and HybridLinker(DeLinker), both utilizing DiffLinker as the diffusion-based PC-Aware model while incorporating FFLOM and DeLinker as their respective PC-Free models. For all approaches, we use ETKDGv3 (Riniker and Landrum, 2015) algorithm provided by RDkit (Landrum, 2022) as conformation predictor \(\mathcal{R}\), and Obabel (O'Boyle et al., 2011) as post-hoc bond predictor \(\mathcal{E}\). For further implementation details, see Appendix B.2.
### Result: Advanced Diversity of Valid Molecules
We run quantitative comparison of each linker generation algorithm in Table 3, demonstrating that HybridLinker successfully balance both high validity and diversity while all the baselines fail. The high scores in all diversity counting only valid molecules (V+U, V+N, V+HD, V+FG, V+BM), indicate that its diversity is driven by meaningful substructural variations. Notably, the high V+N score highlights HybridLinker's ability to discover non-trivial linkers for given fragments. When using DeLinker as the surrogate generator, HybridLinker achieves the highest validity, while utilizing surrogates from FFLOM boosts diversity while maintaining competitive validity. Importantly, HybridLinker also maintains high Uniqueness and Novelty scores, showing that its superior diversity of valid molecules stems from a balanced enhancement of both validity and diversity.
### Application: Property Optimization
Property optimization, which aims to identify molecules with high scores in specific measurements, is a direct application benefitted by sample diversity. Here, we demonstrate that the high diversity of HybridLinker enhances property optimization results through the following two experiments: **Druglikens Optimization** and **Molecule Descriptor Optimization**.
This work aims to advance drug discovery by leveraging the remarkable computational power of modern artificial intelligence. While we acknowledge both the significant benefits and potential risks associated with our approach, such as accelerating drug development for emerging viral threats like COVID-19 or inadvertently facilitating the discovery of unintended compounds, we emphasize the importance of responsible application. We hope this technology will be utilized ethically to drive positive advancements in healthcare and pharmaceutical research.
# Appendix H Related Works: Molecule Generation and Guidance on Diffusion Models
Fragment-Based Drug DesignFragment-Based Drug Design (FBDD) (Jin et al., 2023; Igashov et al., 2024; Torge et al., 2023; Imrie et al., 2020; Huang et al., 2022) is a drug discovery approach that utilizes small molecular fragments and optimizes them into larger, more potent drug candidates. Deep learning-based FBDD encompasses a variety of tasks, each distinguished by its learning objective. **Linker Generation**(Igashov et al., 2024; Huang et al., 2022) is a fundamental task in FBDD, where two molecular fragments are connected to form a complete molecule. Similarly, **Topology Linker Generation**(Jin et al., 2023; Imrie et al., 2020; Zhang et al., 2024) focuses on linking the topological graphs of fragments to generate a complete molecular topology graph. **Scaffold Hopping**(Torge et al., 2023; Zhou et al., 2024) involves replacing the core structure of a given molecule while preserving its biological activity. **PROTAC Design**(Guan et al., 2024; Kao et al., 2023) focuses on generating molecules that incorporate fragment linkers with flexible rotation and translation in 3D space. **Fragment Growing**(Maziarz et al., 2024; Ghorbani et al., 2023) expands single small molecular fragment into larger drug-like structure. Our work specifically addresses Linker Generation, tackling the critical trade-off between diversity and validity observed in existing models. Table 1 Summarizes the tasks in FBDD along with two standard molecular generation tasks--De Novo Generation and Conformation Generation.
Molecule GenerationMolecule generation (Hoogeboom et al., 2022; Vignac et al., 2023; Wang et al., 2024; Liu, 2021; Madhawa et al., 2019; Schneuing et al., 2024) based on deep learning plays a crucial role in drug discovery and is broadly categorized into **De Novo Molecule Generation**(Peng et al., 2023; Geng et al., 2023; Vignac et al., 2023; Jo et al., 2024), **Fragment-Based Drug Design**(Jin et al., 2023; Igashov et al., 2024; Torge et al., 2023; Imrie et al., 2020; Huang et al., 2022), **Target-Aware Drug Design**(Guan et al., 2023; Corso et al., 2023; Schneuing et al., 2024; Peng et al., 2022), and **Conformer Generation**(Xu et al., 2022; Kim et al., 2024; Jing et al., 2023; Xu et al., 2021), based on their input and output formulations. Molecule generation can also be classified into three sub-tasks: **Topology Generation**(Shi et al., 2020; Geng et al., 2023; Jin et al., 2023; Jo et al., 2024; Ji et al., 2020; Wang et al., 2024b;a), **Point Cloud Generation**(Hoogeboom et al., 2022; Xu et al., 2023; Igashov et al., 2024; Guan et al., 2023; Schneuing et al., 2024), and **3D Graph Generation**, where deep learning models learn the distributions of molecular bonding topology, spatial coordinates, and 3D molecular graph representations, respectively. Our work falls under Fragment-Based Drug Design, introducing a hybrid approach that leverages pretrained models for topology and point cloud generation in a zero-shot manner.
Guidance on Diffusion ModelsDiffusion models (Ho et al., 2020; Song et al., 2022; Rombach et al., 2022; Ho et al., 2022; Jo et al., 2024; Hoogeboom et al., 2022; Vignac et al., 2023) have demonstrated exceptional performance across various generative tasks, including image, video, graph, and molecular generation. A recent advancement in diffusion models is **conditional generation**(Ho & Salimans, 2022; Chung et al., 2024a;b; Dhariwal & Nichol, 2021), which enables sampling from a conditional distribution based on desired properties. This is achieved by incorporating a guidance term into the backward diffusion process. To compute the guidance term, **Classifier Guidance**(Dhariwal & Nichol, 2021) employs a classifier trained to estimate the likelihood of a given property, while **Classifer-Free Guidance** (CFG) (Dhariwal & Nichol, 2021) replaces the classifier with a conditional diffusion model. **CFG++**(Chung et al., 2024b) is designed to mitigate off-manifold sampling issues, and **Diffusion Posterior Sampling** (DPS) (Chung et al., 2024a) was developed to solve nonlinear noisy inverse problems. In this paper, we introduce the first DPS-based method for guiding diffusion models in molecular point cloud generation using molecular topology. Our approach introduces a novel energy-based function that effectively bridges topological and spatial molecular representations.
|
10.48550/arXiv.2502.17349
|
HybridLinker: Topology-Guided Posterior Sampling for Enhanced Diversity and Validity in 3D Molecular Linker Generation
|
Minyeong Hwang, Ziseok Lee, Kwang-Soo Kim, Kyungsu Kim, Eunho Yang
| 4,538
|
10.48550_arXiv.1612.07930
|
## 3 Data Analysis
This study used the BYTe variational line list and experimental energies determined using the MARVEL procedure.
### Calculating Experimental Absorption Spectra
Experimental transmission spectra \(\tau_{\rm exp}(\nu,T)\) at a temperature \(T\) [K] and a line position \(\nu\) [cm\({}^{-1}\)] are calculated from four SB spectra. This four measurements scheme is needed because the IR light source used is not modulated but modulation of the IR light appears in the FTIR spectrometer. Therefore, the MCT detector in the FTIR spectrometer will "see" both modulated emissions from the light source and the cell. Moreover at 1027 \({}^{\circ}\)C many of NH\({}_{3}\) bands appear in emission as well and the additional measurements with the cold beam stopper allow one to separate emission from absorption. The four measurements are two reference (N\({}_{2}\) in the central part of the cell) measurements \(\mathbf{I}_{\rm ref+BB}\) and \(\mathbf{I}_{\rm ref}\) and two sample (N\({}_{2}\) + NH\({}_{3}\) mixture) measurements \(\mathbf{I}_{\rm gas+BB}\) and \(\mathbf{I}_{\rm gas}\), one with and one without signal from the BB (at 1800 K):
\[\tau_{\rm exp}(\nu,T)=\frac{\mathbf{I}_{\rm gas+BB}-\mathbf{I}_{\rm gas}}{ \mathbf{I}_{\rm ref+BB}-\mathbf{I}_{\rm ref}} \tag{1}\]
Spectra without signal from the BB are measured from a cold (room temperature) beam stopper placed at 90 degrees from the optical axis of the setup using a movable mirror in the BB adapter.
The method for calculating theoretical absorption spectra follows Ref.. First the 'true' transmission spectrum was computed as:
\[\tau_{\rm calc}^{\rm true}(\nu,T)=\exp{(-\sigma(\nu,T)lc)} \tag{3}\]
Lorentz half-widths were estimated from the experimental spectra and with reference to measured widths compiled in the HITRAN database. The measured (effective) transmittance spectrum is derived by convolving \(\tau_{\rm calc}^{\rm true}(\nu,T)\) with the instrument line shape (ILS) function \(\Gamma(\nu-\nu_{0})\):
\[\tau_{\rm calc}^{\rm eff}(\nu,T)=\int_{0}^{\infty}\tau_{\rm calc}^{\rm true} (\nu_{0},T)\Gamma(\nu-\nu_{0})d\nu_{0} \tag{4}\]
For boxcar apodization, the ILS is a sinc function:
\[\Gamma(\nu)=\Lambda\ {\rm sinc}(\Lambda\pi\nu)=\Lambda\frac{\sin(\Lambda\pi\nu) }{(\Lambda\pi\nu)} \tag{5}\]
For triangular apodization, the ILS is a sinc\({}^{2}\) function:
\[\Gamma(\nu)=\Lambda\ {\rm sinc}^{2}(\Lambda\pi\nu)=2\Lambda\frac{\sin^{2}( \Lambda\pi\nu)}{(\Lambda\pi\nu)^{2}} \tag{6}\]where \(\Lambda\) is commonly termed the FTIR retardation and is generally defined as the inverse of the nominal resolution of the spectrometer.
The theoretical absorption spectrum is then computed as:
\[A_{\text{calc}}(\nu,T)=\log_{10}\left[\frac{1}{\tau_{\text{calc}}^{\text{eff}}( \nu,T)}\right] \tag{7}\]
### The Assignment Procedure
First a list of observable BYTe lines for the experimental conditions was compiled. For this purpose the absorbance of each line, \(j\), was approximated as:
\[A_{\text{calc}}^{\text{approx}}=\frac{S_{j}^{a}lc}{\Delta L\ln} \tag{8}\]
If both the upper and lower energies involved in a observable transition were known experimentally, the BYTe line position was replaced by the MARVEL line position generated by subtracting upper and lower state energies. This hybrid line list, which retains all the BYTe transitions, will be presented elsewhere; it shall henceforth be referred to as BARVEL.
Taking the resolution of the measurements and the accuracy of BYTe intensities into account (see Section 4.1), experimental peaks and BARVEL line positions were coupled using python scripts to produce a 'trivial' assignment list. In cases where multiple BARVEL lines corresponded to a single peak, the peak was assigned to the strongest line.
Trivial assignments for the same vibrational band provide an expected observed minus calculated (obs. - calc.) difference for all lines in that band. Lines present in the list of observable BYTe lines, but not in BARVEL, were shifted by this residual to make future assignments by the method of branches.
A list of all trivial and branch assignments, the final assignment list, was then compared to previous studies, namely those catalogued in the HITRAN database.
## 4 Results and Discussion
The absorption measurements were performed at a temperature of 1027 \({}^{\circ}\)C for the NH\({}_{3}\) volume concentration of 10%.
The measurements were used to test the accuracy of BYTe then analysed using BYTe to generate an assignment list for the data. Central wavenumbers for assigned peaks are compared to line positions measured by Hargreaves et al. where possible.
The absorption spectra, a peak list (partially assigned) including line positions from Hargreaves et al. for assigned lines where available, and new energy level information derived from the assignments are presented in the supplementary data.
### Direct Comparison with BYTe
A comparison between the experimental and theoretical absorption spectra at 1027 \({}^{\circ}\)C for the whole region (2100 - 5500 cm\({}^{-1}\)) is shown in Overall, taking into account the experimental noise, there is good agreement.
However there are shifts in line position of the order \(\pm\) 0.2 cm\({}^{-1}\) across the entire spectral range and shifts up to \(\pm\) 1 - 2 cm\({}^{-1}\) in a few regions, particularly at higher wavenumbers. Hence it was decided that assignments should only be made using MARVEL line positions or BYTe line positions corrected for the expected obs. - calc. difference derived from trivial assignments, and not by simple line list comparison. BARVEL line positions should have an obs. - calc. difference smaller than the nominal resolution of the measurements, 0.09 cm\({}^{-1}\), whilst the wavenumber threshold for the BYTe line positions was taken to within 0.1 cm\({}^{-1}\) of the expected obs. - calc. difference. On the whole experimental line intensities are reproduced within 30 %. This is illustrated in for the region 4860 - 4900 cm\({}^{-1}\). As such experimental lines were coupled to BARVEL or BYTe lines using an intensity threshold of 30 %.
### Assignments
Out of 3701 measured experimental peaks 2308 lines have been assigned. The remaining peaks either did not correspond to a BARVEL or BYTe line within the set wavenumber and intensity thresholds or corresponded to multiple lines with roughly equal contribution to the total intensity such that it could not be confidently assigned. 553 lines were previously assigned by studies included in the HITRAN database (see Table 1). The full 1027 \({}^{\circ}\)C peak list with assignments is available as supplementary material to this article.
Hargreaves et al. presented high temperature line lists for the region 2100 - 4000 cm\({}^{-1}\) constructed from emission spectra recorded at a resolution of 0.01 cm\({}^{-1}\). These line lists are currently being updated and extended (private communication) and hence were not the focus of the current work.
Comparison between experimental (upper) and calculated (BYTe, lower) absorption spectra at 1027 \({}^{\circ}\)C for the range 2100 - 5500 cm\({}^{-1}\).
Comparison between experimental (upper) and calculated (BYTe, lower) absorption spectra at 1027 \({}^{\circ}\)C for the range 4860 - 4900 cm\({}^{-1}\).
Of the 1755 newly assigned lines in this work, 990 are also present in the line lists of Ref.. In these cases line positions from Ref. are included with the current central peak wavenumbers in the supplementary data and employed in the computation of upper state energies described below, as these were measured at a higher resolution.
For branch assignments with an experimentally known lower energy state, energies for the upper state were computed using MARVEL energies and the line position of the strongest assigned transition to that state. The calculated energies are available as supplementary material to this article.
As in our previous study, lines were assigned to a large number of different bands. Table 2 gives a summary of the observed bands including the number of lines assigned to each and whether the band was observed for the first time in this work. Bands are listed in order of theoretical vibrational band centre (VBC). VBC = VBO\({}^{\prime}\)- VBO\({}^{\prime\prime}\) where VBO is the vibrational band origin from BYTe, in wavenumbers.
\begin{table}
\begin{tabular}{l c} \hline \hline & Lines \\ \hline Experimental & 3701 \\ HITRAN & 553 \\ New trivial & 272 \\ New branch & 1483 \\ Total Assigned & 2308 \\ \hline \hline \end{tabular}
\end{table}
Table 1: Summary of NH\({}_{3}\) lines assigned in the region 2100 - 5500 cm\({}^{-1}\).
If the observed \(J_{\rm max}\) in this work is bigger that quoted in the literature, the previous \(J_{\rm max}\) is also given. The full 26 quantum labels for each transition, 13 per vibration-rotation state as recommended by Down _et al._, will be given in the partially assigned peak list and energies files.
\begin{table}
\begin{tabular}{c c c c c c c} \hline Band & VBC & \(N\) & \(J^{\prime}_{\rm max}\) & \(J^{\prime\prime}_{\rm max}\) & o\(-\)c & Note \\ \hline \(\nu_{3}^{1,-}-\nu_{2}^{-}\) & 2475.50 & 59 & 17 & 16 & \(-0.1\) & \\ \(\nu_{3}^{1,+}-\nu_{2}^{+}\) & 2511.55 & 52 & 19 & 18 & 0.1 & \\ \((\nu_{2}+\nu_{3}^{1})^{-}-2\nu_{2}^{-}\) & 2553.27 & 14 & 16 & 15 & 0.1 & New \\ \((\nu_{2}+\nu_{3}^{1})^{+}-2\nu_{2}^{+}\) & 2553.27 & 6 & 9 & 9 & 0.1 & New \\ \(3\nu_{2}^{-}-0^{+}\) & 2895.53 & 6 & 9 & 9 & 0.1 & \\ \hline \multicolumn{8}{c}{Continued on next page} \\ \hline \end{tabular}
\end{table}
Table 2: Summary of observed bands in the region 2100 - 5500 cm\({}^{-1}\) in order of theoretical (BYTe) vibrational band centre (VBC = VBO\({}^{\prime}\) - VBO\({}^{\prime\prime}\) where VBO = vibrational band origin, in cm\({}^{-1}\)) with maximum upper and lower \(J\) rotational quantum number (\(J^{\prime}_{\rm max}\) and \(J^{\prime\prime}_{\rm max}\) respectively). \(N\) is the number of lines assigned to the band. If \(J_{\rm max}\) in this work is higher than given in the literature, the previously known \(J_{\rm max}\) is given in parentheses. o\(-\)c gives the band shift used, in cm\({}^{-1}\), in making branch assignments. VBO of \(0^{+}\) is set to 0.0 cm\({}^{-1}\) in line with the MARVEL study.
## Table 2 - continued from previous page
\begin{tabular}{c c c c c c} \hline Band & VBC & \(N\) & \(J^{\prime}_{\max}\) & \(J^{\prime\prime}_{\max}\) & o\(-\)c & Note \\ \hline \((\nu_{1}+2\nu_{2})^{+}-2\nu_{2}^{-}\) & 3120.69 & 8 & 16 & 16 & \(-2.0\) & \\ \((\nu_{2}+2\nu_{4}^{0})^{+}-\nu_{2}^{+}\) & 3147.49 & 1 & 8 & 8 & \(-0.6\) & New \\ \((\nu_{2}+2\nu_{4}^{2})^{+}-\nu_{2}^{+}\) & 3167.81 & 19 & 18 & 19 & 0.0 & New \\ \((2\nu_{2}+\nu_{4}^{1})^{+}-0^{+}\) & 3189.04 & 17 & 15 & 14 & 1.0 & \\ \(2\nu_{4}^{0,+}-0^{-}\) & 3215.21 & 95 & 21 & 20 & 0.3 & \\ \(2\nu_{4}^{0,+}-0^{+}\) & 3216.00 & 2 & 6 & 7 & 0.3 & \\ \(2\nu_{4}^{0,-}-0^{-}\) & 3216.75 & 13 & 17 & 18 & \(-0.2\) & \\ \(2\nu_{4}^{0,-}-0^{+}\) & 3217.55 & 72 & 18 & 17 & 0.4 & \\ \(2\nu_{4}^{2,+}-0^{-}\) & 3239.39 & 5 & 9 & 8 & \(-0.1\) & \\ \(2\nu_{4}^{2,+}-0^{+}\) & 3240.18 & 82 & 24 & 25 & \(-0.2\) & \\ \(2\nu_{4}^{2,-}-0^{-}\) & 3240.78 & 53 & 21 & 22 & \(-0.2\) & \\ \((\nu_{2}+2\nu_{4}^{0})^{-}-\nu_{2}^{-}\) & 3240.81 & 1 & 7 & 8 & 0.0 & New \\ \(2\nu_{4}^{2,-}-0^{+}\) & 3241.58 & 23 & 17 & 18 & 0.1 & \\ \((\nu_{2}+2\nu_{4}^{2})^{-}-\nu_{2}^{-}\) & 3260.70 & 8 & 16 & 17 & 0.0 & New \\ \(2\nu_{1}^{+}-2\nu_{4}^{0,-}\) & 3296.58 & 1 & 4 & 4 & 1.0 & New \\ \((\nu_{1}+\nu_{2})^{+}-\nu_{2}^{-}\) & 3326.39 & 62 & 22 & 22 & 0.05 & \\ \((\nu_{1}+\nu_{4}^{1})^{+}-\nu_{4}^{1,-}\) & 3328.35 & 19 & 14 & 15 & 0.0 & New \\ \((\nu_{1}+\nu_{4}^{1})^{-}-\nu_{4}^{1,-}\) & 3329.51 & 11 & 15 & 16 & 0.0 & New \\ \((\nu_{1}+\nu_{4}^{1})^{-}-\nu_{4}^{1,+}\) & 3330.61 & 10 & 13 & 14 & 0.0 & New \\ \(\nu_{1}^{+}-0^{-}\) & 3335.28 & 158 & 21 & 21 & \(-0.4\) & \\ \(\nu_{1}^{-}-0^{+}\) & 3337.07 & 155 & 21 & 21 & \(-0.4\) & \\ \((\nu_{1}+\nu_{2})^{-}-\nu_{2}^{+}\) & 3387.59 & 85 & 19 & 19 & \(-0.05\) & \\ \hline \end{tabular}
Continued on next page15 bands have been assigned for the first time in this work, although some of the energy levels involved are known from observations of other bands.
All trivial assignments are secure, as the MARVEL energies (and hence BARVEL line positions) are known to very high accuracy (of the order \(10^{-4}\) cm\({}^{-1}\) for the energies). The accuracy of branch assignments depends on the determination of the obs. - calc. difference for a given vibrational band.
For bands with many (\(>10\)) assignments the obs. - calc. difference can be tracked through the band. As this remains relatively stable we have confidence in our assignments.
Bands for which only a few lines could be assigned are more tentative, although every observed band in this work has at least one associated trivial assignment.
It is worth noting that the single lines assigned to \((\nu_{2}+2\nu_{4}^{0})^{+}-\nu_{2}^{+}\), \((\nu_{2}+2\nu_{4}^{0})^{-}-\nu_{2}^{-}\), \(2\nu_{1}^{+}-2\nu_{4}^{0,-}\), \((\nu_{2}+\nu_{3}^{1})^{-}-\nu_{2}^{+}\) and \((\nu_{3}^{1}+\nu_{4}^{1})^{+}-0^{+}\) are all trivial.
## 5 Summary
High-resolution absorption measurements of NH\({}_{3}\) in the region 2100 - 5500 cm\({}^{-1}\) at atmospheric pressure and a temperature of 1027 \({}^{\circ}\)C have been reported and analysed.
A comparison between the measurements and BYTe shows in general good agreement through there are some shifts in line position (up to 2 cm\({}^{-1}\)) and overall BYTe reproduces experimental intensities only within 30 %. Work towards a new NH\({}_{3}\) line list is currently being carried out as part of the ExoMol project.
The use of BYTe and MARVEL has allowed the assignment of 2308 lines. 553 lines were previously assigned by studies included in the HITRAN database. 1755 lines have been assigned for the first time in this work. The 272 lines assigned using MARVEL line positions, also known as trivial assignments, are secure as the accuracy of MARVEL energies is of the order \(10^{-4}\) cm\({}^{-1}\). Of the 1483 branch assignments, those associated with bands which have numerous assignments in this work should be reliable because the observed-calculated differences remain relatively stable within a given band. The remaining assignments should also be valid, as all observed bands have at least one verified assignment in this work which provides an expected observed-calculated difference for the band, however these are more tentative.
|
10.48550/arXiv.1612.07930
|
High-resolution absorption measurements of NH3 at high temperatures: 2100 - 5500
|
Emma J. Barton, Sergei. N. Yurchenko, Jonathan Tennyson, Sonnik Clausen, Alexander Fateev
| 5,506
|
10.48550_arXiv.2201.02599
|
## 1 Introduction
The uniform electron gas (UEG) plays an iconic role in condensed matter physics. It is the simplest example of an interacting electronic system in the thermodynamic limit. It has also been used as a simple model for metals, especially simple metals like Na and Al. Moreover, its exchange-correlation (XC) energy is a vital input to the local density approximation (LDA) for inhomogeneous systems, an approximation that dominated DFT calculations for a generation, and remains in wide use today. The ground-state energy was first calculated accurately by Ceperley and Alder, and modern parameterizations largely agree at the 1% level. On the other hand, the finite temperature case is still actively being calculated today.
Recently, it was pointed out that, for any electronic system, one could extract the unknown XC energy by a sequence of Kohn-Sham DFT (KS-DFT) calculations (one for each point on a grid in the system) if a well-defined but unknown potential, the conditional probability potential, were known. A simple LDA to this potential yields surprisingly accurate results for systems as disparate as the binding energy curve of H\({}_{2}\) and the uniform gas. This approximation automatically has no self-interaction error for one-electron systems, correctly dissociates the H\({}_{2}\) singlet into two separate H atoms, and its accuracy does not deteriorate as the temperature is raised in a uniform gas.
However, many questions were left unexplored in the original publication. The formal underpinnings, for any quantum system, are being presented elsewhere. In the present work, we focus exclusively on the uniform gas at zero temperature. We give details on how CP calculations are performed, which differ substantially from a traditional DFT calculation, both in the nature of the external potential and the boundary conditions. We first perform the crudest possible CP calculation, by adding a simple Coulomb repulsion at the reference point. This is called a blue electron approximation, thinking of the reference electron as distinguishable from all others (painted blue). The CP is approximated as the ground-state density of an \(N-1\)-electron system with the blue electron treated as an external potential generating a simple Coulomb repulsion. This scheme works remarkably well for densities with Wigner-Seitz radii, \(r_{s}>2\), with errors of less than 12% for the XC energy. However, it fails badly at high densities where exchange dominates, and so is poorly modeled by a classical approximation. Essentially, the on-top exchange hole and its environs is about minus half the density for an unpolarized gas, while the blue electron potential digs almost no ontop hole at all.
We next show how this difficulty was overcome in Ref. First, the pure blue electron approximation was shown to violate the electron-electron cusp condition by a factor of 2. When this is restored, accurate results for highly quantum systems were achieved, but it has little or no impact on the high-density uniform gas. Then, a large repulsive Gaussian was added to the CP potential, with parameters chosen to reproduce the exchange hole, i.e., essentially an accurate CP potential in this limit. As this does not require any many-body calculation, this is deemed acceptable. But it was also necessary to smoothly turn this Gaussian off as a function of increasing \(r_{s}\), a somewhat empirical procedure, but one which yields the high accuracy presented in Ref.
Here, we bypass this messy procedure, but still without using any many-body results for the uniform gas. We present a more sophisticated and satisfactory solution, by performing a CP calculation for antiparallel spin only, fixing the parallel CP density to that of exchange (which has long been known analytically). By construction, this recovers the correct exchange in the high-density limit, and yields reasonable accuracy for all other \(r_{s}\), but is less accurate in the low density limit. Combining this procedure at high densities with the original calculation at low densities yields an accurate curve for all \(r_{s}\) values.
Lastly, we consider the Thomas-Fermi (TF) solution to the blue electron problem, and evaluate its accuracy relative to a full KS calculation. We solve the TF equation analytically using a high-density approximation, which turns out to be remarkably accurate for all densities. For atoms, inaccurate TF densities are the chief source of error in TF energies. But here we are calculating a simple impurity in an otherwise uniform background, thereby avoiding evanescent regions, etc. Moreover, since the XC energy involves a double spatial integral in general (and a single integral), it may be more forgiving of errors than an explicit density functional might be. We find that TF CP densities yield reasonably accurate XC energies in CP theory, and their accuracy improves with increasing \(r_{s}\). We expect TF-CP calculations to be particularly useful at temperatures beyond which the KS equations fail to converge, a regime which inspired many of these ideas (see Ref. and subsequent work).
## 2 Theory
### Background
Standard KS-DFT calculations solve the KS equations within some approximation for the XC energy as a functional of the density, \(E_{\mathrm{xc}}[n]\):
\[\left[-\frac{1}{2}\nabla^{2}+v_{\mathrm{s}}[n](\mathbf{r})\right]\phi_{i}( \mathbf{r})=\varepsilon_{i}\,\phi_{i}(\mathbf{r}), \tag{1}\]
Hartree atomic units are used throughout.
\[v_{\mathrm{s}}[n](\mathbf{r})=v(\mathbf{r})+v_{\mathrm{u}_{\mathrm{XC}}}[n]( \mathbf{r}), \tag{2}\]
In practice, spin-DFT is used. Most calculations of total energies yield densities sufficiently close to the exact density that the error in the total energy is dominated by the error in \(E_{\mathrm{XC}}\) itself, rather than on its evaluation on the approximate density. While there have been many improvements and refinements in such approximations over the last half century, there are many known systematic limitations of present-day functionals, such as their inability to correctly dissociate bonds (strong correlation effects) and electron self-interaction errors.
Recently, CP-DFT, an alternative approach to electronic structure calculation, was suggested. This takes advantage of the well-known expression for the XC energy in terms of the (coupling-constant averaged) XC hole, which in turn is simply related to the conditional probability density for finding an electron at \(\mathbf{r}^{\prime}\), given an electron at \(\mathbf{r}\). Knowledge of this function (or in fact just some particular integrals over it) determines \(E_{\mathrm{XC}}\) exactly. The aim in CP-DFT is to _calculate_ this CP probability density using a standard KS calculation at every point \(\mathbf{r}\) in the system. In principle, if it exists, there is a unique correction to the one-body potential such that the ground-state of \(N-1\) electrons yields the desired CP density. In practice, we find a simple local density approximation works very well most of the time, as our results show.
Thus, writing
\[E_{\mathrm{XC}}=\frac{1}{2}\int_{0}^{1}d\lambda\int\,d^{3}r\int\,d^{3}r^{\prime }\,\,\frac{n(\mathbf{r})n_{\mathrm{XC}}^{\lambda}(\mathbf{r},\mathbf{r}^{ \prime})}{|\mathbf{r}-\mathbf{r}^{\prime}|}, \tag{3}\]
The XC hole density is determined via
\[n_{\mathrm{XC}}^{\lambda}(\mathbf{r},\mathbf{r}^{\prime})=\tilde{n}_{\mathbf{ r}}^{\lambda}(\mathbf{r}^{\prime})-n(\mathbf{r}^{\prime}), \tag{4}\]
We define the CP potential as the potential that has ground state density \(\tilde{n}_{\mathbf{r}}^{\lambda}(\mathbf{r}^{\prime})\) for \(N-1\) electrons. It can be written as \(v[\tilde{n}_{\mathbf{r}}^{\lambda}](\mathbf{r}^{\prime})\), where \(v[n](\mathbf{r})\) is the functional dependence of the potential on the ground state density, first described by Hohenberg and Kohn. We then define:
\[\Delta\tilde{v}_{\mathbf{r}}^{\lambda}(\mathbf{r}^{\prime})=v[\tilde{n}_{ \mathbf{r}}^{\lambda}](\mathbf{r}^{\prime})-v^{\lambda}(\mathbf{r}^{\prime}) \tag{5}\]
This is a functional of the original (\(N\)-electron) gas density \(n(\mathbf{r})\), by the Hohenberg-Kohn theorem.
This theory is formally exact if the potential exists, but the exact CP potential is not known in general and must be approximated. A simple approximation, \(\Delta\tilde{v}_{\mathbf{r}}^{\lambda}=\lambda/|\mathbf{r}-\mathbf{r}^{ \prime}|\), accounts for the majority of the CP correlation, and becomes exact as the distance between \(\mathbf{r}\) and \(\mathbf{r}^{\prime}\) increases. This can be considered the correct potential in the classical limit, in which particles can be distinguished from one another, and where the missing electron provides a simple Coulomb repulsive impurity potential.
But this cannot be exact for quantum systems in general. For example, recovery of the electron-electron cusp conditions at small \(|\mathbf{r}-\mathbf{r}^{\prime}|\) distances requires instead \(\lambda/(2|\mathbf{r}-\mathbf{r}^{\prime}|)\).
\[\Delta\tilde{v}_{\mathbf{r}}^{\lambda}[n](\mathbf{r}^{\prime})\approx\frac{ \lambda}{2|\mathbf{r}-\mathbf{r}^{\prime}|}\left(1+\mathrm{Erf}\left(\frac{| \mathbf{r}-\mathbf{r}^{\prime}|}{r_{s}(n(\mathbf{r}))}\right)\right), \tag{6}\]
While this mirrors the longstanding challenge of approximating \(E_{\mathrm{XC}}\) with a density functional, DFT methods often produce highly accurate densities even when their energies are incorrect, allowing CP-DFT to produce accurate solutions for several systems that are challenging for standard DFT methods.
For those familiar with classical DFT of distinguishable particles, this procedure is very like the Percus-Yevick closure of the Ornstein-Zernike equation. In the truly classical case, the blue electron approximation would be exact, if the exact inhomogeneous XC functional were used when solving the KS equations for the impurity. In practice, we expect the use of approximate XC in such calculations to produce only very slight errors, much of which will be forgiven by the multiple integrals in Eq.. Moreover, we anticipate that errors will reduce as the density becomes lower, the electrons become more spread out, and the behavior is dominated by pure Coulomb repulsion. Likewise, in the high density limit, we expect difficulties, where the entirely non-classical exchange effect dominates.
This paper reports only results for the uniform gas at zero temperature. It gives the full details of the potentials and computational methods that were used in, plus an extension of the method used there to extract spin-decomposed holes, and better justify the smooth turn-off of the exchange barrier needed to recover the high-density limit. We also discuss the quasi-analytic solution of the Thomas-Fermi equation (instead of the KS equations) for this case. This is of interest, as clearly a TF solution is far less computationally expensive than the KS approach. In cases where the additional cost of solving KS equations at every point is prohibitive, use of TF theory may provide a practical alternative, but only if errors are not substantially increased.
## 3 Methods
### Cp-Dft
We now consider applying the CP concept to the spin unpolarized uniform electron gas. By symmetry, the CP potential is spherical, and we perform calculations within a finite sphere of radius \(R\). Because the UEG is translationally invariant, we choose \(r=0\) for the reference point. Further, we drop the prime from \(r^{\prime}\), so that \(r\) now denotes the distance from the reference point, as is conventional for UEG calculations. In practice, we perform calculations in a finite sphere with a large but finite number of electrons \(N\). The boundary of the sphere produces non-uniformities in the density, which are minimized by a judicious choice of boundary condition, Eq.. Moreover, by calculating our XC hole density, Eq., by subtracting densities with \(N\) and \(N-1\) electrons, the effect of non-uniformity near the surface largely cancels. Thus all our calculations are converged with respect to the size of the sphere. The systems have a maximal radial coordinate \(R\), volume of \(V=4\pi R^{3}/3\), and average electron density of \(\bar{n}=N/V\) and \((N-1)/V\). The Kohn-Sham (KS) equations are solved in the local density approximation (LDA) for both systems.
\[v_{\rm s}[n]({\bf r})=v_{\rm n}[n_{0}]({\bf r})+\Delta\tilde{v}[n]({\bf r})+v_ {\rm n}[n]({\bf r})+v_{\rm XC}[n]({\bf r}), \tag{7}\]
where \(v_{\rm n}[n]({\bf r})\) is a Hartree repulsion
\[v_{H}(r)=\int d^{3}r^{\prime}\,\frac{n({\bf r}^{\prime})}{|{\bf r}-{\bf r}^{ \prime}|}, \tag{8}\]
The exchange-correlation potential, \(v_{\rm XC}[n]({\bf r})\) is from LDA using the PW92 parametrization.
\[\Delta\tilde{v}(r^{\prime})=\left\{\begin{array}{l}\frac{1}{r^{ \prime}}\\ \frac{1}{2r^{\prime}}\left(1+\,{\rm Erf}\left(\frac{r^{\prime}}{\bar{r}_{s}} \right)\right).\end{array}\right. \tag{9b}\]
Here, \(\bar{r}_{s}\) refers to the Wigner-Seitz radius of the system as a whole. As was discussed in the introduction, the simple approximation of \(1/r^{\prime}\) for the CP potential is not sufficient for quantum systems. However, it is still instructive to see where it fails. Further details of the CP-DFT and the numerical solution method are contained in Appendix A.
### Cp-Dft results
We first show results for calculations at \(r_{s}=2.5\) and \(r_{s}=10.0\) using the simple \(1/r\) potential, Eq. (9a), for the blue electron. plots the density of the \(N\) electron system and the \(N-1\) electron system as a function of the radial coordinate for these two values of \(r_{s}\). The blue system density in both cases is depressed near \(r=0\) by the repulsion as we would expect. The density difference is the XC hole (at \(\lambda=1\)) from which one deduces the pair distribution function (PDF), with Eq. and Eq.. This is shown in where we have also included calculations for the modified CP potential Eq. (9b).The advantage of calculating both the \(N\) and \(N-1\) electron systems is apparent in the cancellation of boundary effects. The modified CP potential is significantly better than the simple \(1/r\) potential. We can use Eq. to evaluate the XC energy. This energy is _not_ averaged over the coupling constant, i.e., it is the potential contribution only. shows \(r_{s}\) times the potential XC energy \(\varepsilon_{xc}^{(\lambda=1)}\) plotted vs \(r_{s}\) for the same blue potentials as before, along with the exact values. The latter is calculated using the PW92 expressions for \(g(r)\) along with Eq. and Eq.. For values of \(r_{s}\gtrapprox 2.5\), the modified blue potential does much better than the simple Coulomb potential, with relative errors less than 3% and 11%, respectively. However, both do badly as the system density increases (\(r_{s}\to 0\)), as shown for \(r_{s}=0.02\) in The model fails to adequately depress the density near \(r=0\). This is due to the dominance of exchange at high density, which is not included in our approximate CP potential. In practical terms, the blue electron is not "repulsive" enough. A more satisfactory solution can be obtained by performing a CP calculation in one spin channel only where the other spin CP density is fixed to that given by exchange alone. We turn to the details of this approach in the next section.
### Spin CP-DFT
In this section, we perform a spin CP-DFT calculation instead of a CP-DFT calculation, showing that the high-density limit can be treated accurately by this method.
The spin conditional probability densities for a system are defined as the probability density for finding an electron at \(\mathbf{r}^{\prime}\), given an electron _of spin_\(\sigma\) at \(\mathbf{r}\). Spin conditional probability densities are total densities (not spin densities), but are conditional on an electron having a given spin as well as given position. We can further decompose a spin CP density into a sum of two spin densities.
\[\bar{n}_{\mathbf{r}\uparrow}(\mathbf{r}^{\prime})=\bar{n}_{\mathbf{r}\uparrow }(\mathbf{r}^{\prime}\uparrow)+\bar{n}_{\mathbf{r}\uparrow}(\mathbf{r}^{ \prime}\downarrow). \tag{10}\]
The first is the parallel spin CP spin density, the latter being the anti-parallel. We perform a spin KS-DFT calculation in which the former is fixed at its exchange form, which is well-known for a uniform gas, given explicitly in (B2). In principle, the CP potential should now be two different potentials, one for each spin channel, and the functional dependence could differ for parallel and antiparallel spins. Here we simply use the same CP potential we have used for non-spin CP-DFT, (Eq.). For finite systems, using spin densities in CP-DFT is complicated, as in general they do not integrate up to integer particle numbers. Here, this is not a concern, as the particle number is infinite. Moreover, in our calculation, we use the exchange pair correlation function, which does yield an integer particle number.
We consider a UEG with spin densities \(\bar{n}_{\uparrow}\) and \(\bar{n}_{\downarrow}\). Placing a _blue_\(\uparrow\) electron at the origin, \(r=0\) and dropping the prime, it will be surrounded by its hole of \(\uparrow\) electrons given by:
\[\delta n_{\uparrow}(r)=\bar{n}_{\uparrow}\left[g_{\uparrow\uparrow}(\bar{n}_{ \uparrow},\bar{n}_{\downarrow};r)-1\right]. \tag{11}\]
Here, \(g_{\uparrow\uparrow}(\bar{n}_{\uparrow},\bar{n}_{\downarrow};r)\) is the PDF for like-spin electrons separated by a distance \(r\) in a uniform gas with given spin densities, \(\bar{n}_{\uparrow}\) and \(\bar{n}_{\downarrow}\). The up-spin density is then given by:
\[\bar{n}_{\uparrow}(\tau\uparrow)=\bar{n}_{\uparrow}+\delta n_{\uparrow}(r)= \bar{n}_{\uparrow}g_{\uparrow\uparrow}(\bar{n}_{\uparrow},\bar{n}_{\downarrow };r) \tag{12}\]
With a known \(\bar{n}_{\uparrow}(\tau\uparrow)\), we can write down a KS equation for \(\bar{n}_{\uparrow}(r\downarrow)\) and solve for it within the local spin density approximation. Thus, _given_\(g_{\uparrow\uparrow}\), we can calculate \(g_{\uparrow\downarrow}\) via CP-DFT. As \(g_{\uparrow\downarrow}\) contains no exchange, we can hope to be more accurate in the high density limit.
Normalized density vs \(r/r_{s}\) for (a) \(r_{s}=2.5\) and (b) \(r_{s}=10.0\). Solid line is the \(N\)-electron system and the dashed line is the blue electron, \(N-1\) system.
Pair distribution function, \(g(r)\) plotted vs \(r/r_{s}\) for (a) \(r_{s}=2.5\) and (b) \(r_{s}=10.0\). The dashed blue line is for the CP potential equal to \(1/r\), the dotted red line is for the modified CP Potential and the solid black line is the exact solution.
The numerical solution for the fixed spin case is carried out in a manner analogous to that for CP-DFT. Again, we consider two systems. First, the \(N\) electron system with numbers of up and down spins, \(N_{\sigma}=N/2\) and average spin densities, \(\bar{n}_{\sigma}=N_{\sigma}/V\) and secondly, the \(N-1\) electron system with \(N_{\uparrow}=N/2-1\) up spins and \(N_{\downarrow}=N/2\) down spins. The \(N\) electron system is initialized as \(n(r\uparrow)=\bar{n}_{\uparrow}\), \(n(r\downarrow)=\bar{n}_{\downarrow}\) and the \(N-1\) electron system is initialized as \(\tilde{n}_{\uparrow}(r\uparrow)=\bar{n}_{\uparrow}g_{\rm X}(\zeta,k_{F}r)\) and \(\tilde{n}_{\uparrow}(r\downarrow)=\bar{n}_{\downarrow}\). Here, \(\bar{n}_{\uparrow(\downarrow)}\) is understood as referencing the \(N\) or \(N-1\) electron system accordingly. Further details are contained in Appendix B. By using only \(g_{x}\), which can be derived analytically, we avoid using any many-body results. For high densities, this will be extremely accurate.
### Computational details
Calculations can be carried out on personal computers, typically lasting minutes or to hours. Our model was written in Python 3 utilizing NumPy, SciPy, Numba, and an exchange-correlation module "exec.py" written by Kristjan Haule (Physics Dept. Rutgers University).
### Spin CP-DFT results
The calculations for the spin CP-DFT use the modified version of the CP potential Eq. (9b). Additionally we are setting the correlation potential to zero. We first consider an electron density of \(r_{s}=0.02\). We plot the result for the PDF in The agreement is excellent as at such a high density the interactions are dominated by exchange. We also plot \(g_{\uparrow\downarrow}(r)\) in and these results are also quite good.
We next consider a lower density of \(r_{s}=2.5\). We show results for the total PDF in Fig. 6(a) and for the antiparallel PDF in Fig. 6(b). Although there is some small difference with the exact result for the antiparallel PDF, the overall PDF is good and the error in potential XC energy is about 0.2%.
We see a more serious departure from the exact result at \(r_{s}=10.0\). Figs. 7(a) and 7(b) plot the PDF and antiparallel
Pair distribution function, \(g(r)\) plotted vs \(r/r_{s}\) for \(r_{s}=0.02\). The dashed blue line is for the CP potential equal to \(1/r\), the dotted red line is for the modified CP Potential, and the solid black line is the exact soultion.
Pair distribution function (a) and antiparallel PDF-1 (b) vs \(r/r_{s}\) for the spin CP-DFT with \(r_{s}=0.02\). Solid black line is the exact result and the orange dotted line is the spin CP-DFT.
\(r_{s}e_{xc}^{(\lambda=1)}\) vs \(r_{s}\) for the CP potential equal to \(1/r\) (dashed blue line), modified CP potential (dotted red line) and the exact solution (solid black line). The inset shows an expanded view of the region for \(r_{s}\lessapprox 2.0\)
PDF for this low density. Here we can see more clearly the departure from the exact result and the error in the potential XC energy is about 6%.
In Fig.8 we plot the potential XC energy times \(r_{s}\) as a function of \(r_{s}\) for both models. The agreement with the exact result is quite good for the spin CP-DFT, being worst at low density.
Since, as noted above, for values of \(r_{s}\gtrapprox 2.5\), the fractional error for the CP-DFT is on the order of 3%, the question arises as to whether an additional potential can be added to the CP-DFT to correct for its deficiencies at high density.
\[v_{G}(r)=A(r_{s})e^{-r^{2}/2\sigma^{2}(r_{s})}, \tag{13}\]
The \(r_{s}\)-dependence was chosen to replicate exchange in the high density limit, and to slowly turn off as \(r_{s}\) grows, vanishing beyond \(r_{s}=2.5\). This is described in Appendix C. The results of using the Gaussian potential, Eq. with the strength and range parameters given by Eq. (C3) are shown in where the maximum difference with the exact result for the 2-channel model is now less than 3%. This was used in Ref.
We can obtain the coupling-constant averaged values of the XC energy, \(\varepsilon_{xc}\), by an integral over \(r_{s}\), i.e.,
\[\varepsilon_{\mathrm{xc}}(r_{s})=\frac{1}{r_{s}^{2}}\int_{0}^{r_{s}}r_{s}^{ \prime}\varepsilon_{xc}(r_{s}^{\prime})dr_{s}^{\prime}. \tag{14}\]
The comparison with PW92 for the XC energy is shown in where the relative difference is below 6% for the single channel model and for the 2-channel model with Gaussian repulsion is below 3%.
\(r_{s}\varepsilon_{\mathrm{xc}}^{(\lambda=1)}\) vs \(r_{s}\). The solid black line is the exact result, the dotted orange line is the spin CP-DFT, and the dotted blue line is for the CP-DFT.
Same as for but for \(r_{s}=2.5\)
## 4 Thomas-Fermi solution for the blue electron problem
The Thomas-Fermi functional for the blue electron system is
\[F^{TF}[n]=\int d^{3}r\left[A_{s}n^{5/3}(\mathbf{r})+v_{n}(\mathbf{r })n(\mathbf{r})\right.\\ \left.+\frac{1}{2}v_{n}(\mathbf{r})n(\mathbf{r})+\Delta\tilde{v} (\mathbf{r})n(\mathbf{r})\right] \tag{15}\]
The potentials, \(v_{n}\), Eq. and \(v_{n}\), Eq. are as before and we use \(\Delta\tilde{v}(r)=1/r\). The charge densities are normalized such that \(\int d^{3}r\)\(n_{0}=N\) and \(\int d^{3}r\)\(n(\mathbf{r})=N-1\).
Defining the total electrostatic potential as \(V(\mathbf{r})=v_{n}(\mathbf{r})+v_{n}(\mathbf{r})+\Delta\tilde{v}(\mathbf{r})\) minimizing the functional, Eq. with respect to the density, \(n(\mathbf{r})\) subject to the normalization condition, and formally solving for the density, we have (just as for bare atoms or ions):
\[n(\mathbf{r})=\left(\frac{5}{3}A_{s}\right)^{-3/2}\left[\mu-V(\mathbf{r}) \right]_{+}^{3/2}, \tag{16}\]
As \(r\rightarrow\infty\), \(n(r)\to n_{0}\), and thus \(V(r)\to 0\). The latter can be determined from a Taylor expansion of the denominator in \(V(\mathbf{r})\). From Eq., we find
\[\mu=\frac{5}{3}A_{s}n_{0}^{2/3}. \tag{17}\]
But as \(r\to 0\), \(V\sim 1/r\), so \(\mu-V\rightarrow-\infty\). Thus, \(\mu=V\) at some \(r_{0}\), so \(n=0\) for \(r\leq r_{0}\) in Eq..
We relate the Laplacian of the electrostatic potential to the charge densities using Poisson's equation
\[\nabla^{2}V((\mathbf{r}))=-4\pi\left[n(\mathbf{r})-n_{0}+\delta(\mathbf{r}) \right]. \tag{18}\]
Defining
\[\frac{\phi(r)}{r}=\mu-V(r) \tag{19}\]
we have
\[\nabla^{2}\left(\frac{\delta\phi}{r}\right)=4\pi\left[-n_{0}+\left(\frac{5}{3} A_{s}\right)^{-3/2}\left(\frac{\delta\phi}{r}\right)^{3/2}\Theta(r-r_{0})\right], \tag{20}\]
where \(\Theta(x)\) is a step function in \(x\), and the \(\delta-\)function at \(r=0\) implies:
\[\delta\phi=-1,\hskip 42.679134pt\delta\phi(r_{0})=-\mu r_{0}. \tag{21}\]
Introducing dimensionless variables \(z=r/r_{s}\) and \(y=\delta\phi/\mu r_{s}\) we can write Eq. as
\[\frac{d^{2}y}{dz^{2}}=x_{s}z\left(-1+\left(1+\frac{y}{z}\right)_{+}^{3/2} \right), \tag{22}\]
The boundary conditions are:
\[y=-\frac{x_{s}}{3},\quad y(z_{0})=-z_{0},\quad y\to 0\ \mathbf{as}\ z\rightarrow\infty. \tag{23}\]
For \(z<z_{0}\),
\[y(z)=-\left(\frac{x_{s}}{2}\right)z^{3}+Az+B, \tag{24}\]
where
\[A=\left(\frac{x_{s}}{3z_{0}}\right)\left[1+\frac{z_{0}^{3}}{2}\right]-1, \hskip 28.452756ptB=-\,\frac{x_{s}}{3}. \tag{25}\]
For \(z>z_{0}\), we have, for large \(z\), to first order
\[\frac{d^{2}y}{dz^{2}}\approx 3x_{s}\frac{y}{2}, \tag{26}\]
Coupling-constant averaged exchange-correlation energies plotted as \(r_{s}\varepsilon_{xc}\) vs \(r_{s}\). The solid black line is the exact result, the dashed orange line is for the spin CP-DFT, and the dotted blue line is for the CP-DFT with the added Gaussian potential.
Same as for but where the CP-DFT has the added Gaussian potential.
By matching the log derivatives at \(z=z_{0}\) between Eq.
\[z_{0}^{3}+\frac{3k}{x_{s}}z_{0}^{2}+\frac{3}{x_{s}}z_{0}-1=0. \tag{27}\]
The solution to Eq. depends on the value of \(r_{s}\). For \(r_{s}<a/\sqrt{2}\approx 0.4872\), there are 3 real roots but only 1 is positive and for larger \(r_{s}\) there is only 1 real root.
\[z_{0}=\left\{\begin{array}{ll}Q^{1/2}\left[\;2\cos\left(\frac{\theta+\pi}{3} \right)+\sqrt{3}\;\right]&r_{s}<a/\sqrt{2}\\ A+Q/A-\sqrt{3}Q&r_{s}>a/\sqrt{2},\end{array}\right. \tag{28}\]
The numerical solution to the TF blue electron model is detailed in Appendix D.
We performed calculations for \(r_{s}\) values ranging from \(r_{s}=0.02\) to \(r_{s}=100\). Plots of the normalized hole densities for these \(r_{s}\) values are shown in As can be seen in this figure, the normalized hole density tends toward a step function at \(r=r_{s}\) in the low density limit and the location of \(z_{0}\) moves farther out. In we plot the value of \(z_{0}\) as a function of \(r_{s}\) as determined from the numerical calculations, as well as the analytic approximation of Eq.. The analytic solution for \(z_{0}\) is surprisingly accurate with a maximum error of 2.3% around \(r_{s}=5.0\). A table of values for the numerical and analytical results for \(z_{0}\) can be found in Table 1.
We can calculate the potential XC energy from the CP hole density, \(n_{\mathrm{xc}}(r)=n(r)-n_{0}\) using Eq.. In we plot potential energy \(r_{s}\varepsilon_{\mathrm{xc}}^{(\lambda=1)}\) and compare with the the exact result as well as the KS blue electron result using CP-DFT with \(\Delta\tilde{v}(r)=1/r\). We also plot the fractional error in For mid- to low-densities (\(r_{s}\geq 1.0\)) the results are surprisingly good, with errors less than 10%. It is instructive at this point to plot the PDF vs \(r/r_{s}\) for different values of \(r_{s}\) and compare with the exact result. This comparison is shown in for \(r_{s}=0.02,1.0,5.0,10.0\). The results are poor at \(r_{s}=0.02\) but improve with decreasing density. Interestingly enough, although the shape of the hole is inaccurate at \(r_{s}=1.0\), the integrated value of the hole energy is quite good when compared to the exact result.
## V Conclusions
We have shown how to implement the blue-electron concept for a uniform gas at low temperature. The simple classical idea of an impurity capture the basic physics for moderate to low densities, and its accuracy is improved by a local density approximation for the CP potential that respects the electron-electron cusp condition at short distances.
The blue-electron idea naturally fails at high densities, where exchange effects dominate.
Normalized hole density vs \(r/r_{s}\) for a range of densities. Solid lines are the numerical solutions and the dashed lines are for the analytic solutions.
\begin{table}
\begin{tabular}{|c|c|c|c|} \hline \(r_{s}\) & \(z_{0}\) & \(z_{0}^{A}\) & \((z_{0}^{A}-z_{0})/|z_{0}|\) \\ \hline
0.02 & 0.0108 & 0.0108 & -0.0000 \\ \hline
0.1 & 0.0529 & 0.0529 & -0.0004 \\ \hline
0.5 & 0.2168 & 0.2186 & -0.0081 \\ \hline
1.0 & 0.3404 & 0.3460 & -0.0161 \\ \hline
5.0 & 0.6322 & 0.6472 & -0.0233 \\ \hline
10.0 & 0.7272 & 0.7425 & -0.0205 \\ \hline
50.0 & 0.8704 & 0.8809 & -0.0119 \\ \hline
100. & 0.9071 & 0.9153 & -0.0089 \\ \hline \end{tabular}
\end{table}
Table 1: Values of \(z_{0}\) as a function of \(r_{s}\) exactly and analytic calculations along with the relative error between them.
Our procedure still avoids any many-body or QMC input, and provides useful accuracy for all densities, including the high-density limit. Moreover, it provides formal justification for the large Gaussian repulsion used when the entire hole is simulated for high-densities, and its smooth turning off. Errors no greater than 10% in the XC energy density are then available.
As the method is based on classical DFT ideas, we expect errors to lessen with increasing temperature, and overall, they do. But we have also given results for TF blue electron calculations, as higher-temperature simulations often must use TF in place of KS because of convergence issues with the KS scheme. Here we present our analysis of the TF equation just at zero temperature.A simple analytic approximation is remarkably accurate relative to a fully converged numerical solution.We use the numerical solution to compare with KS blue electron calculations at zero temperature. Again, we expect the error introduced by the TF approximation to reduce with increasing temperature.
Even at low temperature, where convergence of KS is not an issue, the TF blue electron might prove a pragmatic alternative to full CP-DFT calculations, as these require a KS calculation at every point in the system.
Regardless of the accuracy of these specific calculations, the principle that one could calculate XC holes and their energies exactly via the densities of many KS calculations remains valid. Moreover, the CP approach provides a useful bypass of the need to find more accurate XC functionals, albeit at higher computational cost. It remains to be seen whether CP-DFT can evolve into a practical alternative to existing DFT calculations, especially in situations where standard DFT fails.
|
10.48550/arXiv.2201.02599
|
Correlation energy of the uniform gas determined by ground state conditional probability density functional theory
|
Dennis Perchak, Ryan J. McCarty, Kieron Burke
| 132
|
10.48550_arXiv.1804.04660
|
###### Abstract
We consider nonadiabatic systems in which the classical Born-Oppenheimer approximation breaks down. We present a general theory that accurately captures the full transmitted wavepacket after multiple transitions through either a single or distinct avoided crossings, including phase information and associated interference effects. Under suitable approximations, we recover both the celebrated Landau-Zener formula and standard surface-hopping algorithms. Our algorithm shows excellent agreement with the full quantum dynamics for a range of avoided crossing systems, and can also be applied to single full crossings with similar accuracy.
The Born-Oppenheimer approximation (BOA) is one of the most widely used methods used to study the quantum dynamics of molecules. Intuitively, it is motivated by the fact that the electrons are much lighter, and therefore much faster, than the nuclei, and hence rapidly adjust their positions with respect to those of the nuclei. This scale separation allows, in many cases, for the electronic and nuclear dynamics to be decoupled. In particular, if the electrons start in a certain bound state, for a fixed set of nuclei positions, then they should remain in this bound state even though the nuclei are slowly moving. Hence the nuclear dynamics can be determined by considering their motion on only one (electronic) potential energy surface.
However, there are interesting situations in which the BOA breaks down. For example, in many photochemical processes the nuclear motion cannot be restricted to a single potential energy surface because, for some nuclear configurations, two such surfaces become close, or even cross. In the former case, known as an avoided crossing, the BOA is still valid to leading order (in the small parameter \(\epsilon\), which is the square root of the ratio of the electronic and nuclear masses), but the remaining corrections are of fundamental interest and, in fact, determine the associated chemistry. In the latter case, which generally takes the form of conical intersections, the BOA breaks down completely.
Here we are primarily interested in cases where the transmitted wavepacket is (exponentially) small, for example when there is an avoided crossing, or when the wavepacket does not pass directly over the conical intersection. Such regimes are, in some sense, generic, as avoided crossings are generic in 1D, and in higher dimensions the probability of an arbitrary wavepacket exactly hitting a conical intersection is vanishingly small. In particular, we consider cases where the wavepacket passes through multiple avoided crossings, or repeatedly through the same crossing. In such cases the transmitted wavepackets can interfere, and thus it is necessary to understand their phases. This suggests that a full quantum mechanical treatment of the problem is required. However, in even moderate dimensions, such treatments are numerically intractable, especially for multiple, coupled electronic potential surfaces.
In order to overcome this, a range of coupled quantum-classical and semiclassical methods have been developed. These include the multiple-spawning wavepacket method, the frozen Gaussian wavepacket method, Ehrenfest dynamics, and the semiclassical initial value representation. The main advantage of such schemes is the significantly reduced computational cost. The main disadvantage, at least with respect to the problem at hand, is the lack of phase information from almost all such schemes. Along with those mentioned above, one of the most widely-used quantum-classical approaches is surface hopping, in which particles are evolved under classical dynamics on a single surface and can 'hop' to other surfaces with a specified probability. Perhaps the most common approach is to only allow hops at points in the trajectory where the gap between energy surfaces has a local minimum (i.e. at an avoided crossing), and the probability of the hop is given by a Landau-Zener (LZ) formula. Such methods give good results for a single transition, especially when the transmitted wavepacket is reasonably large, but fail completely when multiple transitions are involved, due to the complete lack of phase information. We note here that there is at least one such scheme that does aim to retain the phase information, but this is limited to small gaps between the potential energy surfaces, which in turn leads to large transmitted wavepackets. The same restriction is true for other mathematical approaches that lead to explicit formulae for the transmitted wavepacket; see e.g. Ref.. It has been shown that, if the gap scales with \(\epsilon\), then the transitions are of order one and dominated by the Landau-Zener factor.
An alternative approach, inspired by the work of Berry on superadiabatic representations, considers the full quantum mechanical wavepacket. These results, which are restricted to the semiclassical regime where the nuclei move classically, were later made rigorous. It was later shown that, through the use of such superadiabatic representations (which are generalisations of the well-known adiabatic representation), it is possible to derive a formula for the transmitted wavepacket, including phase, at an avoided crossing. The associated algorithm requires only the quantum evolution of wavepackets on single energy surfaces. Whilst this is still computationally demanding if one wants to solve the full Schrodinger equation, there are approximate methods, such as Hagedorn wavepackets or standard quantum chemistry techniques such as MCTDH, which make small relative errors and are computationally much more tractable. The associated algorithm has so far been applied to single transitions through avoided crossings, and to multiple transitions of a single crossing in the case of the photodissociation of NaI. The main goals here are to extend the methodology to multiple transitions through different avoided crossings and to systematically study the effects of making various approximations that lead to a LZ-like transition probability. We will also demonstrate that, although not designed to tackle such problems, the methodology can be successfully applied to single transitions of full crossings.
We present an algorithm that has a number of advantages. We have already mentioned: (i) Preservation of phase information, which allows the accurate study of interference effects; (ii) Only evolution on a single surface is required, which significantly reduces the computational cost when compared to a fully-coupled system, whilst also allowing the use of state-of-the art numerical schemes. The main other benefits are: (iii) Only the adiabatic surfaces (which are the most commonly obtained surfaces from quantum chemistry calculations) are required, in particular there is no need for a diabatization scheme, or the determination of the adiabatic coupling elements; (iv) Such surfaces are only required locally, and thus can be computed on-the-fly; (v) The transmitted wavepacket is created instantaneously, and hence there is no reliance on complicated numerical cancellations of highly-oscillatory wavepackets, which are generally present in the adiabatic representation; (vi) The methodology is easily extended to multiple adiabatic surfaces; (vii) The derived formula is accurate for a wide range of potential energy gaps and small parameters \(\epsilon\), and for any semiclassical wavepacket, i.e. one of typical width or order \(\sqrt{\epsilon}\).
There are, of course, also some disadvantages when compared to the more widely-used schemes: (i) In order to capture the phase information, the one-level dynamics must retain at least some of their quantum nature, and this is inherently more computationally demanding than the analogous classical dynamics; (ii) In the full formalism, it is necessary to be able to extend the potential surface into the complex plane, at least in the region of an avoided crossing. This is essential to be able to accurately compute the transition probabilities. However, in some regimes, for example when the LZ formula is accurate, we can bypass this requirement; (iii) The scheme is, in principle, restricted to wavepackets that are semiclassical near the avoided crossing. However, due to the linearity of the Schrodinger equation, and as demonstrated in, it is possible to'slice' the wavepacket at the crossing. However, this may be more problematic in higher dimensions; (iv) As it stands, the method is restricted to 1D. However, we have successfully extended it to higher dimensions through a slicing procedure [[REF 2D PAPER]].
To outline our approach, we will first review the standard model for nonadiabatic transitions (Section II) and avoided crossings (Section III). We will then, in Section IV, give a brief overview of existing surface hopping models and LZ formulas. We then outline the superadiabatic approach and give the resulting formula in Section V, before describing the associated algorithm in Section VI. In Section VII we systematically investigating its accuracy, and the effects of replacing the true transition probability by two LZ-like approximations. Finally, in Section VIII, we summarize our results and discuss some open problems.
## II The model
The Schrodinger equation governing the quantum dynamics of a molecular system can be written as
\[\mathrm{i}\hbar\partial_{t}\psi(\mathbf{x}_{\mathrm{n}},\mathbf{x}_{\mathrm{e}},t)=H_{ \mathrm{mol}}\psi(\mathbf{x}_{\mathrm{n}},\mathbf{x}_{\mathrm{e}},t),\]
where \(\mathbf{x}_{\mathrm{n}}\) and \(\mathbf{x}_{\mathrm{e}}\) are the nuclear and electronic positions, respectively, and the Hamiltonian is given by
\[H_{\mathrm{mol}}=-\frac{\hbar^{2}}{2m_{\mathrm{n}}}\Delta_{\mathbf{x}_{\mathrm{n}}} -\frac{\hbar^{2}}{2m_{\mathrm{e}}}\Delta_{\mathbf{x}_{\mathrm{e}}}+V_{\mathrm{n}}( \mathbf{x}_{\mathrm{n}})+V_{\mathrm{e}}(\mathbf{x}_{\mathrm{e}})+V_{\mathrm{n,e}}(\mathbf{ x}_{\mathrm{n}},\mathbf{x}_{\mathrm{e}}).\]
Here the first two terms are the kinetic energies of the nuclei and electrons with masses \(m_{\mathrm{n}}\) and \(m_{\mathrm{e}}\), respectively. Note that the masses of the nuclei may all be chosen to be the same by a rescaling of the nuclear coordinates. The potentials \(V_{\mathrm{n}}\) and \(V_{\mathrm{e}}\) denote the nuclear and electronic Coulomb repulsions, respectively, whilst \(V_{\mathrm{n,e}}\) is the attraction between the nuclei and electrons.
We now change to atomic units (\(\hbar=m_{\mathrm{e}}=e=1\)) and define \(\epsilon=1/\sqrt{m_{\mathrm{n}}}\) and the electronic Hamiltonian for fixed nuclear positions \(\mathbf{x}_{\mathrm{n}}=\mathbf{x}\),
\[H_{\mathrm{e}}(\mathbf{x})=-\frac{1}{2}\Delta_{\mathbf{x}_{\mathrm{e}}}+V_{\mathrm{n}}( \mathbf{x})+V_{\mathrm{e}}+V_{\mathrm{n,e}}(\mathbf{x},\cdot).\]
Suppose that \(U^{\pm}(\mathbf{x})\) are two eigenvalues of the electronic Hamiltonian (i.e. two adiabatic potential energy surfaces) of multiplicity one and well-separated from the rest of the electronic spectrum.
\[\mathrm{i}\epsilon\partial_{t}\psi(\mathbf{x},t)=\Big{(}-\frac{\epsilon^{2}}{2} \Delta_{\mathbf{x}}+V(\mathbf{x})\Big{)}\psi(\mathbf{x},t), \tag{1}\]
a diabatic matrix.
\[V(\mathbf{x})=\begin{pmatrix}V_{1}(\mathbf{x})&V_{12}(\mathbf{x})\\ V_{12}(\mathbf{x})&V_{2}(\mathbf{x})\end{pmatrix}.\]
For notational convenience, and to connect back to previous work, we find it useful to define
\[Z=(V_{1}-V_{2})/2,\quad X=V_{12},\quad d=(V_{1}+V_{2})/2,\quad\rho=\sqrt{X^{2} +Z^{2}}\]
and so
\[V(\mathbf{x})=d(\mathbf{x})+\begin{pmatrix}Z(\mathbf{x})&X(\mathbf{x})\\ X(\mathbf{x})&-Z(\mathbf{x})\end{pmatrix}.\]
It is easy to see that the adiabatic surfaces are then given by \(U^{\pm}(\mathbf{x})=d(\mathbf{x})\pm\rho(\mathbf{x})\) and so \(\rho\) is half the energy gap between the two surfaces.
## III Avoided crossings
In the adiabatic representation, an explicit unitary transformation \(U_{0}\) is applied to the system such that the electronic Hamiltonian is diagonal at each choice of \(\mathbf{x}\).
\[H_{0}=-\frac{\epsilon^{2}}{2}\partial_{x}^{2}+\begin{pmatrix}U^{+}(x)&- \epsilon\kappa(x)(\epsilon\partial_{x})\\ \epsilon\kappa(x)(\epsilon\partial_{x})&U^{-}(x)\end{pmatrix}+\mathcal{O}( \epsilon^{2}). \tag{2}\]
Here \(\kappa=(X^{\prime}Z-Z^{\prime}X)/\rho^{2}\) is an explicit 'kinetic coupling' function and we have grouped the terms such that it is more obvious that the off-diagonal elements of the above matrix are of order \(\epsilon\). This can be seen from the fact that wavepackets typically oscillate with frequency \(1/\epsilon\) (see Section III), so \(\epsilon\partial_{x}\psi(x)\) is of order one. Hence we see that, naively, the transitions are of order \(\epsilon\) globally in time. However, as discussed previously, the transitions are exponentially small in \(1/\epsilon\) away from the avoided crossings.
Typically, when the adiabatic potentials are well-separated, the coupling elements are small and then two levels may be treated separately via the Born-Oppenheimer approximation. However, if the adiabatic surfaces become close, but do not cross, the coupling terms typically become large (but do not diverge). Such nuclear configurations are known as avoided crossings. As a result of the large coupling elements, a small, but not negligible, part of the nuclear wavepacket is transferred between the adiabatic surfaces.
Suppose, for clarity of exposition, that the wavepacket initially occupies the upper adiabatic level. The aim of this work is to determine the transmitted wavepacket (on the lower adiabatic level) well away from the crossing (in the scattering regime). Whilst one can, in principle, compute this by a standard numerical solution of the Schrodinger equatiion, there are a number of challenges that prevent this from being a realistic option for most systems of interest:
1. In order to compute the dynamics, one needs an accurate representation of the potential energy surfaces. Typically the adiabatic surfaces are calculated using quantum chemistry methods, such as Density Functional Theory, but it is computationally expensive to determine such surfaces, especially when the number of degrees of freedom (dimension of \(\mathbf{x}\)) is large. In such cases, it is desirable to design methods that can utilise on-the-fly surfaces, determined only locally. Additionally, practical methods for determining surfaces for excited states are still in their infancy, and one also needs to determine the off-diagonal coupling elements. Finally, we note that diabatic representations are not unique, and those obtained in two- and multiple-level cases may differ significantly.
2. The wavepackets we wish to compute are highly oscillatory, typically oscillating with frequency of order \(\epsilon^{-1}\) in space. This can be seen by comparing the kinetic and potential terms in. When using a standard numerical scheme, such as Strang splitting, correctly resolving such oscillations requires very fine grids in both position and momentum space. The curse of dimensionality (for \(N\) points in \(d\) dimensions, one requires \(N^{d}\) points) results in such approaches being impractical for all but very small dimensional systems.
3. Away from avoided crossings, the transmitted wavepacket is typically exponentially small in both the gap size \(\delta\) and \(1/\epsilon\). This can be seen from the formula in Section V.2 or the standard LZ transition probabilities and, where \(\rho_{x_{c}}=\delta\). In contrast, globally in time, the transitions in the adiabatic representation are of order \(\epsilon\), which we have already seen from the Hamiltonian 2. The necessary cancellations in the transmitted wavepacket occur through Stuckelberg oscillations. See Figure 1for an example. There are two challenges here. The first is to correctly resolve these cancellations, which can require very small time steps. The second is the more general challenge of computing an exponentially small quantity; any absolute errors in the numerical scheme must also be exponentially small or they will overwhelm the desired results.
Inset: The mass of the transmitted wavepacket against time as the wavepacket on the original adiabatic surface moves through an avoided crossing. Main figure: Zoom for clarity of the final transmitted mass and Stückelberg oscillations. The time at which the centre of mass of the original wavepacket reaches the avoided crossing is marked with a dashed vertical line and coincides with the maximum transmitted mass.
Existing Approaches and Landau-Zener
In this section we discuss some existing approaches to calculate the transition probability or the transmitted wavepacket.
### Surface Hopping Algorithms
Here we present a brief overview of surface hopping methods, which are one of the most successful approaches for simulating nonadiabatic dynamics. Surface hopping is a mixed quantum-classical approach, where particles are transported classically on the adiabatic surfaces and hop between them under certain conditions, which simulates the quantum effects. A general surface hopping algorithm consists of four steps:
1. Sampling of the initial condition.
2. Classical evolution via \(\dot{x}=p\), \(\dot{p}=-\nabla U^{\pm}(x)\).
3. Surface hopping.
4. Computation of observables.
There are many such schemes, both deterministic and probabilistic and we refer to for further details.
Of particular interest here is the surface hopping step. Typically this is performed when the gap between the two adiabatic surfaces is minimal along a classical trajectory. Whenever such a trajectory reaches a local minimum, a transition to the other surface is performed with a certain probability, usually derived from a simplified quantum mechanical model. The standard approach is to use a LZ formula, which we describe in the next section. The choice of this hopping probability is the main distinguishing feature of different surface hopping models.
The principal advantage of surface hopping algorithms is their simplicity. Due to their use of classical dynamics, which only require local properties of the potential energy surfaces, the methods can be applied in relatively high dimensions, using on-the-fly surfaces. As mentioned previously, such high dimensional systems are beyond the reach of full quantum mechanical methods.
The principal disadvantage is that they lose all phase information, and so cannot treat systems in which interference effects are important, or determine observables in which the relative phase of the wavepackets on the adiabatic surfaces is required. Additionally, they are accurate only when the specified hopping probability is accurate; we will investigate this in Section VII.
### The Landau-Zener Formula
As mentioned previously, in order to compute the transition probability, it is common to use a LZ formula. Whilst the LZ model provides a simple formula for the transition probability, it is generally obtained from a one-dimensional, two-level model in the diabatic representation. However, practical applications occur in multiple dimensions and the potential energy surfaces are usually calculated in the adiabatic representation. There are a number of formulations of the LZ probability, including the extension to multiple dimensions in the diabatic formalism, and versions which only require knowledge of the adiabatic potentials. Here we restrict ourselves to two such formalisms, the first is a diabatic representation. which requires knowledge of the diabatic matrix elements, whilst the second is an adiabatic representation, which only requires the gap between the adiabatic potentials. From now on, we consider only 1D systems; see Section VIII for some discussion of progress in higher dimensions.
Consider a classical particle with trajectory \(\big{(}x(t),p(t)\big{)}\) in phase space. Denote the position where \(\rho\) attains a minimum by \(x_{c}\), and the momentum of the particle at the corresponding time \(t_{c}\) by \(p_{c}\).
\[P_{\rm d}=\exp\Big{(}-\frac{\pi}{\epsilon}\frac{\rho(x_{c})^{2}}{|p_{c}|\sqrt{ X^{\prime}(x_{c})^{2}+Z^{\prime}(x_{c})^{2}}}\Big{)}. \tag{3}\]
The corresponding adiabatic transition probability is given by
\[P_{\rm a}=\exp\Big{(}-\frac{\pi}{\epsilon}\sqrt{\frac{\rho(x_{c})^{3}}{\frac{ \rm d^{2}}{\rm d\tau^{2}}\rho(x(t))|_{t=t_{c}}}}\Big{)}.\]
If one has knowledge of the diabatic matrix elements, and hence \(X\) and \(Z\), this can be rewritten as
\[P_{\rm a}=\exp\Big{(}-\frac{\pi}{\epsilon}\frac{\rho(x_{c})^{2}}{|p_{c}|\sqrt{ X^{\prime}(x_{c})^{2}+Z^{\prime}(x_{c})^{2}+X(x_{c})X^{\prime\prime}(x_{c})+Z(x_ {c})Z^{\prime\prime}(x_{c})}}\Big{)}. \tag{4}\]Note that in the corresponding multidimensional formula there is an additional term, which in 1D would be \([X(x_{c})X^{\prime}(x_{c})+Z(x_{c})Z^{\prime}(x_{c})](U^{\pm})^{\prime}(x_{c})\). However, since an avoided crossing is defined as a minimum of \(\rho\), and \(\rho^{\prime}=(XX^{\prime}+ZZ^{\prime})/\rho\), this term is zero in 1D.
## V Superadiabatic representations and the formula
In this section we will briefly review the ideas behind the use of superadiabatic representations to compute the transmitted wavepacket and refer the reader to the cited works for more details. We will then present a generalisation of a previously-derived formula, which is applicable to 1D avoided crossings not centred at the origin.
### Superadiabatic Representations
Superadiabatic representations were first introduced by Berry, under the additional approximation that the nuclei move classically. More recently this has been extended to the full BOA. As suggested by the name, superadiabatic representations are refinements of the adiabatic representation, which we described in Section refS:avoidedCrossings. In the adiabatic representation, transitions can be very complicated, as demonstrated by the population on the lower level during a typical transition, see This reliance on large cancellations to leave an exponentially small wavepacket suggests that the adiabatic representation may not be the ideal frame of reference in which to study transitions at avoided crossings.
Superadiabatic representations improve on the adiabatic one by simplifying the dynamics near an avoided crossing, at the expense of introducing computational complexities. The superadiabatic representations can be enumerated, and, initially, moving to successively higher superadiabatic representations reduces the spurious oscillations in the dynamics until the transmitted population builds up monotonically as the wavepacket travels through the avoided crossing. This is known as the optimal superadiabatic representation. However, moving to even higher representations results in the spurious oscillations returning. Previous results give a reliable method to determine the optimal superadiabatic representation. However, computing the unitary operators for this representation is highly challenging, and performing the numerical computations in such a representation is similarly difficult.
The main benefit of superadiabatic representations for our purposes is that they allow the derivation of an explicit formula for the transmitted wavepacket in the optimal superadiabatic representation, without requiring the associated unitary matrix. By general theory, all of the superadiabatic representations agree with the adiabatic one away from any avoided crossing. This leads to a simple algorithm to compute the transition through an avoided crossing in the adiabatic representation, as described in Section VI.
### The Formula
Following Ref., it is useful to introduce a nonlinear rescaling in which the adiabatic coupling elements obtain a universal form, known as the natural scale,
\[\tau(x)=2\int_{x_{c}}^{x}\rho(s)\mathrm{d}s,\]
We now extend \(\rho\) and \(\tau\) into the complex plane and, by the theory of Stokes lines, the analytic continuation of \(\rho\) has a pair of complex conjugate zeros, close to \(x_{c}\), at \(x_{cz}\) and \(x_{cz}^{*}\).
\[\tau_{x_{c}}=\tau(x_{cz})=\tau_{r}+\mathrm{i}\tau_{c}.\]
Let \(\phi^{\pm}(x,t_{c})\) be the incoming wavepacket on the corresponding adiabatic surface \(U^{\pm}\) at time \(t_{c}\) when the centre of mass coincides with an avoided crossing at \(x_{c}\).
\[\psi(x,t)=\,\mathrm{e}^{-(\mathrm{i}/\epsilon)(t-t_{c})H^{\mp}}\,\psi^{\mp}(x)\]
where \(H^{\mp}\) are the BOA Hamiltonians for the two levels and \(\psi^{\mp}(x)\) is a wavepacket instantaneously created at time \(t_{c}\), which is more easily expressed in Fourier space via
\[\hat{\psi}^{\mp}(p)=\Theta(p^{2}\mp 4\delta)\frac{p+\eta^{\mp}}{2| \eta^{\mp}|}\exp\Big{(}-\frac{\tau_{c}}{2\delta\epsilon}|p-\eta^{\mp}|\Big{)} \exp\Big{(}-\mathrm{i}\frac{\tau_{r}}{2\delta\epsilon}(p-\eta^{\mp})\Big{)}\] \[\times\exp\Big{(}-\mathrm{i}\frac{x_{c}}{\epsilon}(p-\eta^{\mp}) \Big{)}\hat{\phi}^{\pm}(\eta^{\mp}). \tag{5}\]
Here
\[\delta=\rho(x_{c}),\quad\eta^{\mp}=\mathrm{sign}(p)\sqrt{p^{2}\mp 4\delta},\]
\(\Theta\) is the Heaviside function and the Fourier transform needs to be performed under the correct scaling:
\[\hat{\psi}(p)=\frac{1}{2\pi\epsilon}\int\,\mathrm{e}^{-(\mathrm{i}/\epsilon) px}\,\psi(x)\mathrm{d}x.\]We note that the principle difference from previous presentations of the formula is the final exponential factor involving \(x_{c}\), the position of the avoided crossing. In previous work, this position has been taken to be zero, in which case the factor is simply 1. The new term arises from the approximation \(x(\tau)=\tau/(2\delta)+x_{c}+\mathcal{O}(\tau^{3})\) (which is a simple generalisation of the calculation for \(x_{c}=0^{44}\) (p. 2258)).
### Analysis of the Formula
We now present a brief analysis of the formula in 5, which allows us to connect to surface hopping approaches, as well as to LZ formulas.
Firstly we note that the formula involves the same momentum adjustment that is phenomenologically introduced in surface hopping algorithms. We note that \(\eta\mp\) is precisely the classical incoming momentum required to give outgoing momentum \(p\) when moving down/up, respectively, a potential energy gap of \(2\delta\) and requiring (classical) energy conservation. Relatedly, when passing from the upper to the lower level, the Heaviside function ensures that the transmitted wavepacket has (absolute) momentum at least \(2\delta\), whereas when passing from the lower to upper level it is trivially 1, indicating no restriction on the transmitted momentum. The analogous restriction that a classical particle can only be transmitted to the upper level if it has sufficient kinetic energy is accounted for by the \(\hat{\psi}^{\pm}(\eta^{\mp})\) term.
We now discuss how, in appropriate limits, the formula essentially reduces to a LZ transition for each point in momentum space. We make a number of independent approximations:
1. \(x_{c}=0\). For a single avoided crossing we may do this without loss of generality by shifting the space variable.
2. \(\tau_{r}=0\). This is the case, for example, when the potential is symmetric around the avoided crossing.
3. \(\delta\) is small. This produces two simplifications to the formula using that \(\eta^{\mp}\approx p\mp 2\delta/p\): * The prefactor simplifies to \(\frac{\eta^{\mp}+p}{2|\eta^{\mp}|}\approx p/|p|=\text{sign}(p)\); ** The factor in the exponential simplifies to \(|p-\eta^{\mp}|\approx 2\delta/|p|\). Note that the small parameter in these expansions is actually \(\delta/p_{0}\), and so we expect these approximations to be more accurate for either small \(\delta\) or large incoming momentum.
4. Second order expansion of \(\rho\). It is well known\({}^{39}\) that a naive second order expansion of \(\rho\) is incorrect as the analytic continuation of \(\rho\) must vanish like a square root at its complex zeros. We therefore approximate \(\rho\) via \[\rho(x)\approx\sqrt{\delta^{2}+g(x-x_{c})}\] with \(g\) a smooth function such that \(g=g^{\prime}=0\). Performing a second order expansion of \(g\) then gives \[\rho(x)\approx\sqrt{\delta^{2}+\mbox{$\frac{1}{2}$}g^{\prime\prime}(x-x_{c })^{2}}.\] In this case both \(x_{cz}\) and \(\tau_{c}\) can be computed analytically to give \[\tau_{c}\approx\mbox{$\frac{\pi\delta^{2}}{2\alpha}$}\] where \(\alpha^{2}=\mbox{$\frac{1}{2}$}g^{\prime\prime}\). To connect purely to \(\rho\), we note that \(\mbox{$\frac{1}{2}$}g^{\prime\prime}=\delta\rho^{\prime\prime}(x_{c})\) and hence \[\tau_{c}\approx\mbox{$\mbox{i}$}\frac{\pi\delta^{3/2}}{2\sqrt{\rho^{\prime \prime}(x_{c})}}.\] Finally, in order to connect to the LZ formulas, an explicit computation using that \(\rho^{\prime}(x_{c})=(X(x_{c})X^{\prime}(x_{c})+Z(x_{c})Z^{\prime}(x_{c}))/ \rho(x_{c})=0\) and \(\rho(x_{c})=\delta\) gives \[\rho^{\prime\prime}(x_{c})=\frac{X^{\prime}(x_{c})^{2}+Z^{\prime}(x_{c})^{2}+X (x_{c})X^{\prime\prime}(x_{c})+Z(x_{c})Z^{\prime\prime}(x_{c})}{\delta}\] and so \[\tau_{c}\approx\mbox{$\mbox{i}$}\frac{\pi\delta^{2}}{2\sqrt{X^{\prime}(x_{c})^ {2}+Z^{\prime}(x_{c})^{2}+X(x_{c})X^{\prime\prime}(x_{c})+Z(x_{c})Z^{\prime \prime}(x_{c})}}.\]
Suppose now that we make all four approximations.
\[\hat{\psi}^{\mp}(p)=\mbox{sign}(p)\Theta(p^{2}\mp 4\delta)\exp\left(-\frac{\pi \delta^{2}}{2\epsilon|p|\sqrt{X^{\prime}(x_{c})^{2}+Z^{\prime}(x_{c})^{2}+X(x_ {c})X^{\prime\prime}(x_{c})+Z(x_{c})Z^{\prime\prime}(x_{c})}}\right)\hat{\psi}^ {\pm}(\eta^{\mp}). \tag{6}\]It is now clear that the exponential factor corresponds precisely to the adiabatic LZ transition probability in, with the additional factor of \(1/2\) accounting for the fact that we are determining the size of the transmitted wavepacket rather than the transition probability, which is proportional to the square of the wavepacket. The Heaviside function is also included indirectly in surface hopping models, which explicitly exclude classically-forbidden transitions, see e.g..
Note that if we are interested solely in the transition probability then the first two approximations are irrelevant as they only affect the phase. However, when dealing with multiple transitions these terms are crucial in understanding interference effects. In Section VII we will investigate the effects of these approximations in some example systems.
After approximations- have been made, the resulting formula can be thought of as a surface hopping algorithm that retains phase information. This can be seen by noting that the formula decouples in momentum space. Thus, if we replace the classical transport of individual particles, the ensemble of which represents the initial wavepacket, with quantum evolution of the initial wavepacket, and then replace particle hopping with hopping of momentum components of the wavepacket, then we have a clear analogue of the surface hopping methods. One promising avenue of further work is to investigate the use of formula 5 for the transmission probability (instead of the LZ one) in traditional surface hopping algorithms. Alternatively, we can recover the surface hopping methodology (but retaining phase information) by dividing the wavepacket into small pieces (the surface hopping particles), evolving them classically on the initial level (e.g. using Hagedorn's wavepacket approach) until they reach an avoided crossing, and then applying the formula either with the full transition probability, or the LZ approximation, and reconstructing the wavepacket on the other level.
## VI The algorithm
The general algorithm described below is similar to that presented in previous work, but here it is extended to multiple transitions and to different levels of approximation, which ultimately lead to an analogue of the LZ formula, but applied to wavepackets, rather than simply as a transition probability. The transmitted wavepacket is computed via the following algorithm. For clarity, we present the algorithm for two BOA surfaces, but its extension to multiple surfaces is straightforward due to the linearity of the Schrodinger equation.
1. **Initial Condition:** The initial wavepacket should be specified on either the upper or lower adiabatic level, well away from any of the avoided crossings. Note that, in such regions, the adiabatic, superadiabatic, and diabatic levels are very close, so one may instead specify the wavepacket on a single diabatic level. If the initial wavepacket is given close to an avoided crossing, for example as the result of a laser excitation, then it must be evolved away (into the scattering regime) on the corresponding adiabatic level under the BO approximation to obtain an appropriate initial wavepacket.
2. **One-Level Dynamics:** The initial wavepacket is now evolved under the BOA until the final, specified time, or until another termination condition is satisfied (such as the wavepacket reaching a minimum distance from an avoided crossing). This can be done using any one-level scheme that provides sufficient accuracy, such as Strang splitting, Hagedorn wavepackets, or MCTDH. The wavepacket on the other BOA level is evolved simultaneously; the level is initially unoccupied.
3. **Detection of Avoided Crossings:** Here an avoided crossing is defined as a (local) minimum of the gap \(\rho\). Whenever the centre of mass of the wavepacket reaches such a minimum, apply the formula as described in the next step. Such local minima may be determined _a priori_, for example when the potentials are given analytically, or on-the-fly by monitoring \(\rho\).
4. **Application of the Formula:** Apply the formula to the wavepacket at the avoided crossing and add the resulting wavepacket to the lower level. Note that the formula implicitly requires the potentials to be extended into the complex plane in order to compute \(\tau\). However, as described in the following Section, this requirement may be bypassed by using an analogue of the Landau-Zener formula, at a cost to accuracy which is investigated for some examples in Section VII.
5. **Computation of Observables:** At any time step the wavepackets on the two levels may be used to compute observables, such as mean position, momentum and the level populations, including those which require phase information such as inter-level observables. Note, however (as discussed in Section V.1), that these will only agree with the corresponding quantities computed for the adiabatic populations well away from any avoided crossings.
In the following section we will investigate the accuracy of this algorithm. One restriction for its application to multiple crossings is that the transmitted wavepacket must be small, or, more precisely, the wavepacket remaining on the original surface must not change significantly when compared to its evolution under the BOA. This is due to the perturbative nature of the derivation, which assumes that the original wavepacket is unchanged during a transition.
## VII Numerics
Note that the MATLAB code used to produce the results in this section is available from [https://bitbucket.org/bdgoddard/qmd1dpublic/](https://bitbucket.org/bdgoddard/qmd1dpublic/).
### Jahn-Teller
We consider first a simple example in order to demonstrate the effects of the approximations in Section V.3.
\[V(x)=\begin{pmatrix}x&\delta\\ \delta&-x\end{pmatrix}\]
There is a single avoided crossing at \(x_{c}=0\), with gap \(2\delta\). It is clear that \(x_{cz}=\mathrm{i}\delta\) and a straightforward calculation shows that \(\tau_{x_{c}}=\mathrm{i}\delta^{2}\pi/2\). Note, therefore, that assumptions, and of Section V.3 hold exactly. Furthermore, since \(X(x_{c})X^{\prime\prime}(x_{c})+Z(x_{c})Z^{\prime\prime}(x_{c})=0\), the diabatic and adiabatic LZ transition probabilities given in and are identical in this case. This simple model allows us to investigate the effects of approximation, i.e. the difference between the full formula and the LZ approximation for a range of values of \(\delta\). From the arguments in Section V.3, we expect the two results to agree to high accuracy when \(\delta/p_{0}\) is small, and hence the transition is large, but we expect the full formula result to be more accurate in the regime of interest (relatively large \(\delta\) and small transitions).
We choose to specify the wavepacket at the avoided crossing, and determine the initial condition by evolving it backwards in time away from the crossing on a single adiabatic surface. This ensures that the wavepacket is semiclassical (i.e. of width order \(\sqrt{\epsilon}\)) when it reaches the avoided crossing. As noted above, due to the linearity of the Schrodinger equation, if this were not the case then we could apply a slicing procedure to obtain similarly accurate results.
\[\hat{\psi}(p)=\frac{1}{(\pi\epsilon)^{1/4}}\exp\Big{(}-\frac{\mathrm{i}}{ \epsilon}p_{0}x_{0}-\frac{1}{2\epsilon}(p-p_{0})^{2}-\frac{\mathrm{i}}{ \epsilon}x_{0}(p-p_{0})\Big{)} \tag{7}\]
For this example we fix \(\epsilon=1/50\), which is similar to the value chosen in surface hopping works e.g. (and approximately correct for real-world systems e.g.) and \(p_{0}=8\). We could, in principle, vary these parameters, and we will do so in later examples. We note that due to the nature of the potential, in order to start sufficiently far away from the avoided crossing (such that the adiabatic and superadiabatic representations agree) the initial potential energy must be reasonably large, leading to a correspondingly large minimum value of \(p_{0}\) at the avoiding crossing. See for the adiabatic potentials.
Here we evolve backwards to a start time of \(-40/p_{0}\) with timestep \(1/(1000p_{0})\) and then forwards through the avoided crossing for time \(80/p_{0}\) with the same timestep. We perform the numerics with a spatial grid with \(2^{15}\) points and endpoints \(\pm 60\). We observe that halving the time step and doubling the number of grid points does not significantly affect the results.
In we show the relative error in the transmitted wavepacket and the transmitted mass. This clearly demonstrates that, for small \(\delta\) (and large transmitted mass), both our formula and the LZ-like version give very good results. However, as \(\delta\) increases, the simplified version becomes increasingly inaccurate.
### Simple Avoided Crossing
We now consider a simple example, which will both allow us to systematically investigate the accuracy of our method for different parameter regimes, and provide a benchmark for the accuracy of a single transition; this is, at least heuristically, a lower bound for the accuracy for multiple transitions.
\[V(x)=\begin{pmatrix}\frac{1}{2}\tanh(x)&\delta\\ \delta&-\frac{1}{2}\tanh(x)\end{pmatrix}\]
See for the adiabatic potentials with \(\delta=1/2\).
As in the previous example, in order to control the (mean) momentum of the wavepacket when it reaches the crossing, we specify the wavepacket in momentum space at the avoided crossing and then evolve it backwards in time on a single adiabatic surface to obtain an initial wavepacket for the computations. In particular, we take a Gaussian wavepacket as given in for a range of values of \(\epsilon\) and \(p_{0}\). We compute the results for a single transition of the avoided crossing, both using the full formula and the LZ-like one. As can be seen from Figure 4, the relative error when using 5 is typically of the order of a few percent, with increasing accuracy as \(\delta\) and/or \(p_{0}\) increase. The deviation of the green curve, which corresponds to \(\epsilon=1/10\) is a result of the asymptotic nature of the formula. The odd behaviour of the blue curve for \(p_{0}=3\), \(\epsilon=1/50\) and \(\delta\approx 1\) seems to be a result of parts of the wavepacket becoming 'trapped' near the avoided crossing, which violates the assumption of a single transition.
Figures 5 and 6 demonstrate the effects of using the algorithm with the approximate
Left: The relative error between the ‘exact’ numerical solution and the application of the algorithm using formula [red, solid] and [blue, dashed]. Right: The ‘exact’ transmitted mass, which is in excellent agreement with that computed using for all values of \(\delta\).
As can be seen from Figure 5, for moderate values of \(\delta\), the results become very poor. However, as expected, shows that, for small \(\delta\), the results are very similar to those using the full formula.
For completeness, we give the numerical details: The spatial grid uses \(2^{14}\) points with limits \(\mp 40\). We use a time step of \(1/(100p_{0})\) and obtain the initial wavepacket by evolving the wavepacket backwards from the crossing for time \(20/p_{0}\). The system is then evolved forwards for time \(40/p_{0}\). Again, we note that halving the time step and doubling the number of grid points does not significantly affect the results.
As a further test of the accuracy of the algorithm we perform the same calculation as for the Gaussian wavepackets in the previous example, but with a wavepacket on the upper
The adiabatic potentials for [clockwise from top left] the Jahn-Teller, Simple, Dual and Quadratic Potentials
level at the crossing given by
\[\hat{\psi}(p)=\sum_{j=1}^{3}w_{i}\hat{\psi}(x_{0,i},p_{0,i},p), \tag{8}\]
We choose \(\epsilon=1/50\), \(w=[0.7,1,0.9]\), \(p_{0}=[4.6,5,5.3]\) and \(x_{0}=[0.1,0,-0.05]\). However, we note that the results are robust under these choices for a wide range of values. We show the resulting transmitted wavepacket in which for convenience of displaying the phase, we have evolved backwards to the avoided crossing on the lower level. Note that the relative error in this case is \(0.0057\).
The relative error between the ‘exact’ numerical solution and the application of the algorithm using formula. Each subplot shows the result for a different value of \(p_{0}\) for a range of \(\delta\) values. Different colour curves {green, purple, yellow, red, blue} correspond to \(\epsilon=\{1/10,1/20,1/30,1/40,1/50\}\), respectively. Note that, apart from the largest value \(\epsilon=1/10\), the errors are very similar.
We note here that the results for wavepackets starting on the lower level are very similar, and we will investigate such a situation in the following Section.
#### iv.3.1 Full Crossings
Here we consider if the algorithm is applicable to full crossings (with \(\delta=0\)). In such a case, the approximations made in Section V.3 lead to the conclusion that the transmitted wavepacket is approximately equal to the incoming wavepacket. Applying this in the case \(p_{0}=5\), \(\epsilon=1/50\) and \(\delta=0\) gives a relative of \(0.0856\) for both the formula and LZ approximation. This is comparable to the relative error for small, but non-zero \(\delta\) (see Figure 4). This indicates that the methodology can also be used for full crossings.
As but using formula.
is important in higher dimensions, where part of the wavepacket may travel across a full crossing (conical intersection), whilst other parts experience an effective avoided crossing, in which case we need only one algorithm to accurately treat the whole wavepacket.
#### iv.1.2 Diabatic vs. Adiabatic LZ
In the previous examples, we have \(X(x_{c})X^{\prime\prime}(x_{c})+Z(x_{c})Z^{\prime\prime}(x_{c})=0\) and hence the diabatic and adiabatic LZ transition probabilities, and, respectively, are identical. However, here we briefly consider an example in which this is not the case. We note that such a situation was also investigated in Ref. for two-dimensional crossings and the results when using the two formalisms were found to be very similar.
Zoom of
always the case.
\[V(x)=\begin{pmatrix}\frac{1}{2}\tanh(x)&\delta+\frac{1}{10}\tanh^{2}(x)\\ \delta+\frac{1}{10}\tanh^{2}(x)&-\frac{1}{2}\tanh(x)\end{pmatrix},\]
We choose \(\delta=0.2\), \(\epsilon=1/50\) and \(p_{0}=5\) where these parameters, are chosen such that we are in a regime where we expect both the formula and the (adiabatic) LZ approximation to be reasonably accurate, whilst simultaneously the results are not dominated by \(\delta\) being very small.
The wavepacket in momentum space. ‘Upper’ denotes the wavepacket on the upper level at the avoided crossing, as given by. ‘Lower’ denotes the transmitted wavepacket, computed using the algorithm, evolved backwards on the lower level to the avoided crossing. ‘Phase’ shows the phase of the upper (red), lower (blue) and error, for the lower, transmitted phase (black, dashed). ‘Relative Error’ displays the relative error between the transmitted wavepackets given by the ‘exact’ solution and the result of the algorithm.
0.0219 and 0.0217, respectively. In contrast, when using the diabatic approximation, the results are much worse, with a relative error of 0.1081. This, along with previous results, demonstrates a clear motivation to use the transition formula in surface hopping algorithms.
### Multiple Transitions of a Single Crossing
We now demonstrate the algorithm when the wavepacket makes multiple transitions of a single avoided crossing. Here we add a quadratic confining potential, which causes the wavepacket to oscillate backwards and forwards through the avoided crossing:
\[V(x)=\alpha x^{2}+\begin{pmatrix}\frac{1}{2}\tanh(x)&\delta\\ \delta&-\frac{1}{2}\tanh(x)\end{pmatrix},\]
We choose \(\alpha=0.05\), which gives a relatively weak confining potential. We use the same grid and time step as for the simple case in Section VII.2 but here evolve back to \(t=-5\) and forwards to \(t=30\), which gives 3 complete transitions of the avoided crossing. Here we start with a wavepacket of the form on the lower level with \(x_{0}=0\) and \(p_{0}=5\). Again we choose \(\epsilon=1/50\). See for the adiabatic potentials.
As can be seen in Figure 8, the 'exact' dynamics require extreme numerical cancellations at each transition in order to produce the true wavepacket. Although not shown in the Figure, the maximum transmitted mass is 0.0028, which is around 200 times larger than the final mass. Note that the results of both formulas are of a similar accuracy to the results for a single crossing, with relative errors 0.0123 and 3.637 for and, respectively. In particular, the agreement between the 'exact' and formula results is excellent whilst, in this case, significantly underestimates the size of the transmitted wavepackets.
Whist, in principle, we would expect the results of using to improve when \(\delta\) decreases (i.e. when the transmitted wavepacket is larger) this adds a complication to the algorithm: When the transmitted wavepacket is large, the transition significantly affects the wavepacket on the original level, which is used explicitly in the formula for the transmitted wavepacket at the next avoided crossing. Due to the perturbative nature of the derivation of the formula (see e.g.), the wavepacket on the original level is not treated explicitly, and so we do not have access to this unless it can be assumed that it is largely unaffected by the transition. A necessary requirement for this, due to mass conservation, is that the transmitted wavepacket is small.
### Dual Avoided Crossings
As we have seen, for multiple transitions at a single avoided crossing, the algorithm described in Section VI works as expected, determining the correct phase of the wavepackets, and therefore also the correct interference effects. However, for transitions at separate avoided crossings there is an extra difficulty that arises from the definition of the diabatic and adiabatic potentials.
Mass of transmitted wavepacket on the upper adiabatic level over time for the ‘exact’ dynamics [black, dotted] and using the algorithm with formulas [red, solid] and [blue, dashed]. The centre of mass of the wavepacket on the lower level reaches the avoided crossing three times, at approximately \(t=5,15,25\), as indicated by the jumps in the formula masses.
Hence, treating the two crossings independently, the dynamics through the second crossing could be computed by flipping the surfaces in space (which gives the diabatic surfaces associated with the first crossing) and reversing the momentum of the wavepacket. From, we see that reversing the momentum introduces a sign change in the transmitted wavepacket, which must be taken into account when computing the total transmitted wavepacket. (This issue is related to the diabatic eigenfunctions only being defined up to their sign.) Note that this argument generalises to the case where there are multiple non-identical crossings; the case here was chosen for clarity.
Here we choose potentials
\[V(x)=\begin{pmatrix}\frac{1}{2}\big{(}\tanh(x-5)+\tanh(x+5)+1\big{)}&\delta\\ \delta&-\frac{1}{2}\big{(}\tanh(x-5)+\tanh(x+5)+1\big{)}\end{pmatrix}\]
We choose \(\epsilon=1/50\), \(\delta=1/2\) and an initial Gaussian condition using with \(x_{0}=0\) and \(p_{0}=5\). We use the same numerical scheme as in Section VII.2], first evolving backwards for \(t=5\) and then forwards for \(t=10\). See for the adiabatic potentials.
As can be seen from Figure 10, the results using are once again very good (with a relative error of \(0.0295\)), whilst those using the approximate formula are much poorer (relative error \(2.193\)). Note that if we do not include the additional phase correction described above then the results using are also very poor.
This is in contrast to standard surface-hopping algorithms that lose all phase information, and cannot hope to treat systems with interference effects.
Open problems, which will be the subject of future works, are (i) Approximation of the wavepacket that remains on the original level when the transmitted wavepacket is not small, which would allow the study of multiple transitions of general crossings; (ii) Extension to higher dimensions. This can be done via a slicing algorithm; preliminary results for model systems are presented in; (iii) Implementation of our more accurate transition rate in surface hopping models, which should extend their range of validity to systems when the transmitted wavepacket is significantly smaller than those which can be accurately captured by existing LZ schemes.
|
10.48550/arXiv.1804.04660
|
Multiple Superadiabatic Transitions and Landau-Zener Formulas
|
Benjamin D. Goddard, Tim Hurst
| 994
|
10.48550_arXiv.1310.1006
|
###### Abstract
The phase diagram of ice is studied by a quasi-harmonic approximation. The free energy of all experimentally known ice phases has been calculated with the flexible q-TIP4P/F model of water. The only exception is the high pressure ice X, in which the presence of symmetric O\(-\)H\(-\)O bonds prevents its modeling with this empirical interatomic potential. The simplicity of our approach allows us to study ice phases at state points of the \(T-P\) plane that have been omitted in previous simulations using free energy methods based on thermodynamic integration. The effect in the phase diagram of averaging the proton disorder that appears in several ice phases has been studied. It is found particularly relevant for ice III, at least for cell sizes typically used in phase coexistence simulations. New insight into the capability of the employed water model to describe the coexistence of ice phases is presented. We find that the H-ordered ices IX and XIV, as well as the H-disordered ice XII, are particularly stable for this water model. This fact disagrees with experimental data. The unexpected large stability of ice IX is a property related to the TIP4P-character of the water model. Only after omission of these three stable ice phases, the calculated phase diagram becomes in reasonable qualitative agreement to the experimental one in the \(T-P\) region corresponding to ice Ih, II, III, V, and VI. The calculation of the phase diagram in the quantum and classical limits shows that the most important quantum effect is the stabilization of ice II due to its lower zero-point energy when compared to that one of ices Ih, III, and V.
pacs: 64.60.-i,64.60.De, 63.20.-e, 63.20.Ry
## I Introduction
Sixteen different crystalline ice phases have been identified so far in the phase diagram of water. In all phases, except ice X, the water molecules are part of a network connected by H-bonds. In most ice lattices there appears a unique H-bond network that fills the whole volume. However four phases (ices VII, VIII, VI, and XV) are made of two identical and independent networks that interpenetrate one into another. Within a H-bond network, each oxygen atom is coordinated to four oxygen neighbors in a more or less distorted tetrahedral arrangement. The protons are distributed according to the Bernal-Fowler ice rules. They state that in a network there must be one and only one proton between two adjacent oxygen atoms and that each oxygen is part of two OH covalent bonds characteristic of the water molecule. These rules are compatible with either ordered or disordered spatial distributions of H atoms. In fact order-disorder transitions have been observed for most pairs of ice phases (Ih-XI, III-IX, V-XIII, VI-XV, VII-VIII, XII-XIV). Only for H-ordered ice II and H-disordered ices Ic and IV, the other member of the corresponding pair has not been yet experimentally found.
A comprehensive review of the calculation of free energies of water phases with the thermodynamic integration (TI) method can be found in Ref.. The classical phase diagram of water, simulated with the rigid TIP4P/2005 model, shows a reasonable qualitative agreement to the experimental one, in particular in the complex region of stability of ices Ih, II, III, V, and VI. The coexistence of these ice phases has been also studied by quantum path integral simulations with the rigid TIP4PQ/2005 model. The phase diagram of ice Ih, II, and III was additionally investigated using a flexible water model (q-TIP4P/F) in the classical limit. Singer and Knight have analyzed the order-disorder transition in ices Ih-XI, III-IX, V-XIII, VI-XV, VII-VIII, and XII-XIV by the calculation of the small energy differences between the innumerable H-bond configurations possible in a large simulation cell. Since the lattice energy is a scalar, it can be related to topological properties of the H-bond configurations that are invariant to the symmetry operations of the lattice. This link between H-bond topology and energetics is used to extrapolate from electronic calculations on small unit cells to larger cells that approximate the thermodynamic limit. Thus accurate results for the order-disorder transitions in ice are obtained just by focusing on the dependence of the lattice energy with the H-bond configurations. The vibrational energy was assumed to play a secondary role in these transitions. H-bond order-disorder transitions are understood as discontinuous changes in the H-bond topologies sampled by the system, while the oxygen lattice changes minimally. Note that for phase transitions other than order-disorder ones, the change in the oxygen lattice is drastic. Therefore for such transitions the vibrational free energy is expected to play a significant role.
The quasi-harmonic (QH) approximation (QHA) allows us to compute the free energy of a solid as an analytic function of the volume and the temperature for a given interatomic potential. The prediction of the volume, enthalpy, kinetic energy, and heat capacity, of ice Ih, II, and III by the QHA has been compared to results of quantum path integral molecular dynamics (PIMD) simulations using the q-TIP4P/F model. A remarkable overall agreement was found in temperature (\(T\)) and pressure (\(P\)) ranges up to 300 K and 1 GPa, respectively. Moreover, the QHA offers a simple alternative to TI methods to study the phase diagram of solid phases.
The phase boundary between ice VII and VIII has been studied by a QHA in a 16-molecule supercell with _ab initio_ density-functional theory (DFT) calculations of total energies and phonon frequencies. The calculation shows that the coexistence line in the \(P-T\) diagram has negative Clapeyron slope and a noticeable isotope effect, both facts in good agreement to experimental data. The phase diagram of ices Ih, II, and III has been recently calculated by a QHA. The studied models were based on both flexible (q-TIP4P/F) and rigid (TIP4P/2005 and TIP4PQ/2005) descriptions of the water molecule. The QHA was able to reproduce, for each of the studied models, the available coexistence lines Ih-II, II-III, and Ih-III of the phase diagrams derived by TI methods. Moreover, the simplicity of the QHA allowed to uncover new information by considering conditions that had not been treated in previous TI simulations. In particular, for the typical cell sizes used in phase coexistence simulations, the averaging over the proton disorder of ice III was an essential step to obtain a converged phase diagram. Thus, the common procedure of using only one randomly selected ice III structure makes the calculated phase diagram affected by an uncontrolled factor, that can be highly significant for the final result.
The purpose of the present paper is to derive the phase diagram of all experimentally known ice phases of ice (except ice X) using the QHA in combination with the flexible q-TIP4P/F water model. The layout of the manuscript is as follows. A summary of the employed computational conditions is presented in Sec. II. The generation of the ice structures is introduced in Sec. III. The effect of H-disorder averaging in the lattice energy is discussed in Sec. IV. The calculated phase diagram of ice is compared to the experimental one in Sec. V. The pressure dependence of the free energy of several ice phases is presented in Sec. VI. A comparison of classical and quantum phase diagrams is given in Sec. VII. The paper closes with the conclusions.
## II Computational conditions
The employed QHA has been introduced in Refs.. We present here a brief summary.
\[F(V,T)=U_{S}(V)+F_{v}(V,T)-TS_{H}+\triangle U_{ave}\;, \tag{1}\]
\(F_{v}(V,T)\) is the vibrational contribution to \(F\).
\[F_{v}(V,T)=\sum_{k}\left(\frac{\hbar\omega_{k}}{2}+\frac{1}{\beta}\ln\left[1- \exp\left(-\beta\hbar\omega_{k}\right)\right]\right)\;. \tag{2}\]
Here \(\beta\) is the inverse temperature: \(1/k_{B}T\). \(\omega_{k}\) are the wavenumbers of the harmonic lattice vibrations for the volume \(V\), with \(k\) combining the phonon branch index and the wave vector within the Brillouin zone. The anharmonicity of the interatomic potential enters in the QHA only through the volume dependence of \(\omega_{k}\).
\[F_{v,cla}(V,T)=\sum_{k}\frac{1}{\beta}\ln\left(\beta\hbar\omega_{k}\right)\;. \tag{3}\]
The entropy \(S_{H}\) and the energy \(\triangle U_{ave}\) are related to the disorder of hydrogen and they vanish for the ordered ice phases (i.e., ices XI, II, IX, XIII, XV, VIII, and XIV). \(S_{H}\) was estimated by Pauling for fully disordered phases as
\[S_{H}=Nk_{B}\ln\frac{3}{2}\;. \tag{4}\]
A comparison of the Pauling estimate to accurate numerical determinations has been recently presented for several ice phases.\(\triangle U_{ave}\) is a constant energy that depends on the average of the lattice energy over the proton disorder of the ice phase (see below).
\[G(T,P)=F(V_{min},T)+PV_{min}\;. \tag{5}\]
The implementation of the QHA for an ice phase follows these steps:
\(i\)) Find the reference cell that minimizes the static energy \(U_{S}\). This minimization implies optimization of both cell shape and atomic positions. The resulting volume is \(V_{ref}\) and the corresponding static energy \(U_{S,ref}\).
\(ii\)) Only for H-disordered phases: generate a random set of structures with different H-configurations and calculate the constant energy \(\triangle U_{ave}\) as
\[\triangle U_{ave}=\overline{U}_{S,ref}-U_{S,ref}\;, \tag{6}\]
The number of random H-isomers is set so large that the estimated error of the mean value \(\overline{U}_{S,ref}\) is lower than 0.02 kJ/mol. It is also sensible to take as reference cell in step \(i)\) the structure whose lattice energy \(U_{S,ref}\) is closest to the average \(\overline{U}_{S,ref}\).
\(iii)\) Select a grid of 50 volumes in a range of interest \([V_{min},V_{max}]\). The ice cell with volume \(V_{i}\) is set by isotropic scaling of the reference cell. Subsequently, each ice cell is held fixed while minimizing the static energy \(U_{S}(V_{i})\) with respect to the atomic positions. The crystal phonons, \(\omega_{k}(V_{i})\), are obtained after the minimization.
\(iv)\) Calculate the function \(F(V_{i},T)\) by Eq.. The minimum of \(F(V_{i},T)\) as a function of \(V\) is determined by a fit to a 5th degree polynomial in \(V\).
The phase diagram of ice is derived by a brute force method, i.e., given a state point \((T,P)\) one calculates the Gibbs free energy of all ice phases. The stable phase is selected as the one with the lowest value of \(G\).
The crystal phonon calculation has been performed by the small-displacement method. For the flexible water model the atomic displacement employed in this work is \(\delta x=10^{-6}\) A along each Cartesian direction. We have used a \(\Gamma\) sampling (\(\mathbf{k}=\mathbf{0}\)) of the crystal phonons, which is a reasonable approximation for the sizes of the employed supercells. The Lennard-Jones interaction between oxygen centers was truncated at a distance of \(r_{c}=8.5\) A, and standard long-range corrections for both potential energy and pressure were computed assuming that the pair-correlation function is unity for \(r>r_{c}\). Long-range electrostatic potential and forces were calculated with the Ewald method.
The assumption of isotropic scaling of the reference cell made in step \(iii\) was checked for ice II in Ref.. By relaxing this constraint the QHA free energy of ice II changes slightly, by about 0.01 kJ/mol, having a small effect in the phase diagram.
## III Ice Structures
Supercells of similar size to those employed in recent simulations have been used in the QH derivation of the phase diagram. In Table 1 we summarize the crystallographic references used in the generation of the ice structures. Supercells are defined by translation vectors applied along the crystallographic axes of the lattice. The total number \(N\) of water molecules generated in the supercell is also given. The potential energy (\(U_{S,ref}\)) and volume (\(V_{ref}\)) obtained in the minimization of the supercell structures with the q-TIP4P/F model are presented, together with the volume interval \([V_{min},V_{max}]\) used in the QHA for each phase. The optimized reference cells and the corresponding fractional coordinates of the water molecules for each studied ice phase are made available as Supplementary Material.
The algorithm proposed by Buch _et al._ in Ref. was applied for the random generation of full proton disordered structures with vanishing cell dipole moment of ices Ih, Ic, III, IV, V, VI, VII, and XII. The reason for choosing a vanishing cell dipole moment is that the disordered ice phases are not ferroelectric. In the case of ice III and ice V the neutron diffraction experiments show the existence of partial H-disorder, i.e., fractional H-occupancies different from 0.5. The Buch's algorithm has been then slightly modified for the generation of random structures with partial H-disorder. For these phases the proton disorder entropy is somewhat lower than the Pauling result, \(S_{H}\). We have employed the estimations of \(0.9S_{H}\) and \(0.94S_{H}\) for ice III and V, respectively.
In the generation of the crystal structures the following particularities were considered. For ice Ih the reference supercell was orthorhombic with parameters \((\mathbf{4a},3\sqrt{3}\mathbf{b},3\mathbf{c})\), with \((\mathbf{a},\mathbf{b},\mathbf{c})\) being the standard hexagonal lattice vectors of ice Ih. For ice XI, the H-ordered form of ice Ih, we have generated two different structures with crystal symmetry \(Cmc2_{1}\) and \(Pna2_{1}\). The former corresponds to the experimental phase ice XI. The latter is associated to the global energy minimum predicted by TIP4P-like models, that is not in accord with experiment. For ice V the orthorhombic cell setting \((\mathbf{a},\mathbf{b},\mathbf{c})\) used in Ref. corresponds to the space group symbol \(A2/a\), but was changed here to a more standard setting \((\mathbf{c},-\mathbf{b},\mathbf{a})\) with space group symbol \(C2/c\). In the case of the high pressure phases ice VII and VIII, the energy minimization of a flexible supercell did not lead to a stable crystal lattice at zero pressure. To overcome this difficulty the form of the supercell was constrained to that one obtained by classical \(NPT\) simulations of ice VII and VIII using a flexible cell at \(P=\)2 GPa and \(T=\)50 K, i.e., at conditions where instability problems are fully absent. Then subsequent energy minimizations of ices VII and VIII maintaining the rigid form of the ice cell do always lead to stable crystal structures even at volumes corresponding to small negative pressures.
The validity of the QHA in ice is restricted by the presence of anharmonic effects not included in the approximation. Such effects are expected to increase at high temperature. A direct check of the QHA is the comparison to numerical simulations that fully consider the anharmonicity of the interatomic interactions. PIMD results of the density of ice phases for a number of state points are compared to the corresponding QHA as well as to available experimental data in Table 2. The studied state points appear in a temperature range between 77 and 300 K. We find a reasonable agreement between PIMD and QHA densities even at high temperature. Similar temperature behavior was reported for the volume, enthalpy, and heat capacity of ices Ih, II, and III in Ref. up to 300 K. Note that the thermal energy at 300 K corresponds to a wavenumber (\(k_{B}T/\hbar\)) of about 200 cm\({}^{-1}\), so that at 300 K most ice phonons remain in their vibrational ground state. In particular, those related to the molecular stretching and bending modes, as well as the H-bond librations. The comparison between calculated and experimental ice densities in Table 2 displays a satisfactory overall agreement. The largest deviation is found for the high-pressure ice VII, where the calculated q-TIP4P/F density is about 5% lower than the experimental one. Such error has been previously reported in ice VII by classical and quantum Monte Carlo (MC) simulations using TIP4P-like models.
## IV Disorder averaging
An interesting practical question is the importance of proton disorder in the stability of H-disordered cells. Two opposite strategies have been used to address this point. The first one is to use a cell so small that an explicit calculation of the internal energy of all existing H-isomers is possible. The lattice symmetry may be exploited by considering only symmetry inequivalent H-isomers with the corresponding multiplicity. This approach leads to an exact average over proton disorder, but at the cost of introducing an unspecified finite size effect as a consequence of the small unit cell.
The second strategy is to define a large supercell and calculate the internal energy of a _single H-isomer_ of the disordered phase. The average over proton disorder is then introduced _ad hoc_ by adding the proton disorder entropy \(S_{H}\), as in Eq. 1. An implicit assumption here is that the supercell is so large that both the lattice energy and the vibrational free energy of the single H-isomer have already converged with respect to any change in the proton configuration. Such a change is understood to be compatible with the Bernal-Fowler rules and with the corresponding (full or partial) proton disorder. This procedure has been adopted in most TI studies of the phase diagram of ice.
\begin{table}
\begin{tabular}{c c c c c c c} \hline Phase & \(T\) (K) & \(P\)(GPa) & \(\rho\)(PIMD) & \(\rho\) (QHA) & \(\rho\) (exp.) & Ref. \\ \hline Ih & 250 & 0.0 & 0.925 & 0.917 & 0.920 & \\ II & 123 & 0.0 & 1.190 & 1.191 & 1.190 & \\ III (full) & 250 & 0.28 & 1.168 & 1.177 & 1.165 & \\ IV & 110 & 0.0 & 1.290 & 1.296 & 1.272 & \\ V (full) & 237.7 & 0.53 & 1.269 & 1.272 & 1.271 & \\ VI & 225 & 1.1 & 1.397 & 1.382 & 1.373 & \\ VII & 300 & 10 & 1.783 & 1.785 & 1.880 & \\ VIII & 77 & 2.4 & 1.590 & 1.592 & 1.628 & \\ IX & 165 & 0.28 & 1.187 & 1.191 & 1.194 & \\ XI & 77 & 0.0 & 0.931 & 0.930 & 0.934 & \\ XII & 260 & 0.5 & 1.301 & 1.299 & 1.292 & \\ XIII & 80 & 0.0 & 1.241 & 1.242 & 1.244 & \\ XIV & 250 & 0.28 & 1.308 & 1.311 & 1.332 & \\ XV & 80 & 0.0 & 1.329 & 1.336 & 1.326 & \\ \hline \end{tabular}
\end{table}
Table 2: Ice densities derived by PIMD simulations of ice phases using the q-TIP4P/F model are compared to the corresponding QHA, as well as to experimental data. Simulation results for ices III and V correspond to cells with full H-disorder. The employed H-isomer of ice III had a static lattice energy of -60.86 kJ/mol. Density unit is g cm\({}^{-3}\).
\begin{table}
\begin{tabular}{l c c c c c c c} \hline Space symmetry & Supercell & H-disorder & \(N\) & \(U_{S,ref}\) & \(V_{ref}\) & \(V_{min}\) & \(V_{max}\) \\ \hline Ic (\(Fd\overline{3}m\)) & & yes & 216 & -62.00 & 31.00 & 20.44 & 35.02 \\ Ih (\(P6_{3}/mmc\)) & (4, 3\(\sqrt{3}\), 3) & yes & 288 & -61.98 & 30.96 & 29.47 & 35.05 \\ XI (\(Cmc_{2}\)) & & no & 288 & -61.95 & 31.03 & 29.48 & 35.05 \\ XI (\(Pna2_{1}\)) & & no & 288 & -62.02 & 30.90 & 29.30 & 34.96 \\ II (\(R\overline{3}\)) & & no & 324 & -60.84 & 24.14 & 21.75 & 27.31 \\ III (\(P4_{4}\),\(2_{1}2\)) & & yes (full) & 324 & -60.96 & 24.90 & 23.58 & 28.14 \\ III (\(P4_{4}\),\(2_{1}2\)) & & yes (partial) & 324 & -60.72 & 25.05 & 24.07 & 28.31 \\ IX (\(P4_{1}\),\(2_{1}2\)) & & no & 324 & -61.52 & 24.63 & 23.55 & 27.83 \\ IV (\(R\overline{3}c\)) & & yes & 128 & -59.77 & 22.10 & 18.49 & 24.31 \\ V (\(C2/c\)) & & yes (full) & 504 & -60.28 & 22.84 & 20.55 & 25.81 \\ V (\(C2/c\)) & & yes (partial) & 504 & -60.04 & 22.99 & 20.80 & 25.98 \\ XIII (\(P2_{1}/a\)) & & no & 504 & -60.15 & 23.15 & 20.83 & 26.16 \\ VI (\(P4_{2}/nmc\)) & & yes & 360 & -59.58 & 21.44 & 18.14 & 22.51 \\ XV (\(P\overline{1}\)) & & no & 360 & -59.43 & 21.53 & 18.43 & 22.61 \\ VII (\(Pn\overline{3}m\)) & & yes & 432 & -53.08 & 19.57 & 14.02 & 20.98 \\ VIII (\(I4_{1}/amd\)) & & no & 600 & -53.19 & 19.47 & 13.92 & 20.83 \\ XII (\(I\overline{2}d\)) & & yes & 288 & -60.06 & 21.99 & 18.81 & 24.19 \\ XIV (\(P2_{1}2_{1}2_{1}\)) & & no & 192 & -60.62 & 21.90 & 18.32 & 24.09 \\ \hline \end{tabular}
\end{table}
Table 1: The space group of the studied phases is shown with the reference used to generate the ice supercell. H-disordered phases (except ice IV and ice Ic) are found in a row immediately above the H-ordered counterpart. The number of water molecules in each supercell is \(N\). The static lattice energy (\(U_{S,ref}\)) and volume (\(V_{ref}\) ) of the optimized supercells were derived with the q-TIP4P/F model. \([V_{min},V_{max}]\) is the volume interval studied by the QHA. The data for ices III and V correspond to both full and partial H-disorder. Energy unit is kJ/mol, volumes in Å\({}^{3}\)/molec.
Note that the assumption of the convergence of the lattice energy with the proton disorder in a large supercell is only correct in the thermodynamic limit, as the relative fluctuation of thermodynamic quantities is expected to decrease as \(1/\sqrt{N}\). For finite \(N\), the Buch's algorithm will produce a set of H-isomers with different lattice energies and therefore also different statistical weights. Only in the thermodynamic limit will the Buch's algorithm produce proton configurations with the same statistical weight, as it was assumed by Pauling in his estimation of the residual entropy of ice in Eq.. In this respect, Eq. can be understood as a finite size correction for the estimation of the lattice energy associated to the thermodynamic limit.
The simplicity of the QHA renders possible to check the convergence of the lattice energy, for a given cell size, with respect to the proton disorder. In our previous QH study of ice Ih and III, the convergence of the static lattice energy, \(U_{S,ref}\), was studied for a small set of six random H-isomers in a cell with 324 molecules. In the following we present a more accurate account of the convergence of \(U_{S,ref}\) for ice III.
### Ice III
In we have represented the results of \(U_{S,ref}\) for a set of 50 random H-isomers of ice III. The supercell contains 324 water molecules with full H-disorder. The static lattice energy \(U_{S,ref}\) is plotted as a function of the corresponding cell volume, \(V_{ref}\). For comparison, we show also the data obtained when ice III has partial H-disorder, and the value for ice IX, the H-ordered counterpart of ice III. The minimized lattice energy, \(U_{S,ref}\), and volume, \(V_{ref}\), are related in a nearly linear way. We note that all isomers having partial H-disorder display _larger_ static energy than those with full H-disorder.
An important result of is that for the employed supercell with 324 molecules the dispersion of \(U_{S,ref}\) is rather large (\(\sim 0.4\) kJ/mol). We have chosen a threshold of 0.05 kJ/mol as criterion to qualify if a given energetic difference can be considered as significant in the calculation of the phase diagram. Then, following this criterion, the spreading of the lattice energy of ice III may appreciably affect the phase diagram whenever it is calculated with a _single random_ H-isomer of ice III. As a remedy to this uncertainty, the average term \(\triangle U_{ave}\) was introduced in Eq. 1 to improve the convergence of the internal energy of ice with respect to the proton disorder.
The effectiveness of this averaging procedure is illustrated in It shows the phase diagram of ice Ih, II, and III calculated for six random H-isomers of ice III with the q-TIP4P/F model. The lattice energy \(U_{S,ref}\) of these H-isomers scatters in an interval of about 0.3 kJ/mol. In the free energy of ice III was calculated _without_ the H-disorder averaging term (\(\triangle U_{ave}\)) of Eq. 1. These results are identical to those presented in Fig.4 of Ref.. Coexistence lines of different ice III structures are clearly separated.
Lattice energy and volume of a set of H-isomers of ice III generated randomly according to the Bernal-Fowler ice rules. The H-isomers display either full (open circles) or partial (closed squares) H-disorder. The results were derived with the q-TIP4P/F model for a supercell with 324 molecules. The close triangle shows the result for ice IX, the H-ordered counterpart of ice III. The line is a guide to the eye.
_a_) QHA phase diagram of ice Ih, II, and III. The coexistence lines were calculated for six random H-isomers of ice III with full H-disorder. The free energy of ice III does not include the disorder averaging term \(\triangle U_{ave}\) of Eq.. _b_) Same diagram after including the disorder averaging term \(\triangle U_{ave}\) in the free energy of ice III. Note the decrease in the dispersion of the coexistence lines. All results derived with the q-TIP4P/F model in the quantum limit.
The phase diagram obtained after considering the term \(\triangle U_{ave}\) in Eq. 1 is shown in The spreading of the coexistence lines is now appreciably reduced. The remaining dispersion is related to the vibrational free energy, that is also affected by the disorder of protons in the employed supercell. However, this effect of H-disorder in the vibrational energy is comparatively less important than in the lattice energy. A similar conclusion has been presented in the analysis of order-disorder transitions in Ref..
### Other disordered phases
The mean static lattice energy, \(\overline{U}_{S,ref}\), and its standard deviation, \(\sigma(U_{S,ref})\), was calculated by sampling a set of random H-isomers for all H-disordered phases (Ih, Ic, III, IV, V, VI, VII, and XII). The results are summarized in Table 3. A large value of the standard deviation, \(\sigma\), implies that H-disorder strongly affects the value of the static energy \(U_{S,ref}\), of the supercell and therefore also the stability of the ice phase. The largest value of \(\sigma\) corresponds to ice III with full H-disorder (\(\sigma\) =0.1 kJ/mol ). Accordingly the static energy, \(U_{S,ref}\), of a single random H-isomer of ice III can be found in an energy window of about \(4\sigma\sim 0.4\) kJ/mol, as shown in
The standard deviations \(\sigma(U_{S,ref})\) given in Table 3 decrease along the series of ices: III (full disorder) > XII > V (full disorder) > III (partial disorder) > IV > VI. For these phases, the _no consideration_ of H-disorder averaging may introduce arbitrary shifts in the lattice energy larger than 0.05 kJ/mol, at least for the supercell sizes employed here. For ices Ih, Ic, and VII the energetic effect of H-disorder is smaller than this threshold so that it can be safely neglected for the studied supercells.
For ices III and ice V, the mean lattice energy, \(\overline{U}_{S,ref}\), is significantly larger (\(\sim\) 0.24 kJ/mol) in the case of partial than in the case of full H-disorder. This behavior is in contradiction to the experimental finding that ice III and V display both partial H-disorder. This unphysical result is in line with the reported limitations of the effective potentials to reproduce the energetics of the H-bond rearrangement in the ice phases.
Our analysis on the disorder averaging of the lattice energy of ice has omitted several factors that might be relevant. The consideration of reference cells with non-zero dipole moment should increase the standard deviation, \(\sigma(U_{S,ref})\), and affect also the mean static lattice energy of the cell. This behavior has been demonstrated in the classical MC simulation of the dielectric constant of ice using several water models. Another factor is the fractional occupation of H-sites in the ice structures with partial H-disorder (ices III and V), that may depend on the employed water model. This was shown in Ref. where the fractional occupancies (H\({}_{\alpha}\), H\({}_{\beta}\)) of ice III, experimentally found as (0.35, 0.5), change to (0.5, 0.25) by using the TIP4P/2005 model. In our treatment of partial disorder in ice III and ice V we have considered only fractional occupancies derived from the experimental data.
## V Quasi-harmonic phase diagram
The experimental phase diagram of ice at temperatures in the range K and pressures below 10 GPa is presented in The broken line shows the boundary between ice an liquid water.
Phase diagram of ice. _a_) experimental result from Ref.. The broken line is the boundary between ice and liquid (L) water. Full lines denote coexistence conditions between ice phases. _b_) QHA result derived in the quantum limit with the q-TIP4P/F potential model.
\begin{table}
\begin{tabular}{c c c c c} \hline Ice & \(N\) & \(N_{ave}\) & \(\overline{U}_{S,ref}\) & \(\sigma(U_{S,ref})\) \\ \hline Ih & 288 & 6 & -61.98 & 0.00 \\ Ic & 216 & 9 & -62.00 & 0.00 \\ III (full) & 324 & 50 & -60.98 & 0.10 \\ III (partial) & 324 & 6 & -60.73 & 0.03 \\ IV & 128 & 9 & -59.77 & 0.02 \\ V (full) & 504 & 10 & -60.27 & 0.03 \\ V (partial) & 504 & 6 & -60.03 & 0.01 \\ VI & 360 & 10 & -59.57 & 0.02 \\ VII & 432 & 6 & -53.08 & 0.01 \\ XII & 288 & 9 & -60.07 & 0.04 \\ \hline \end{tabular}
\end{table}
Table 3: Result of the averaging of the lattice energy, \(U_{S,ref}\), of H-disordered ices using the q-TIP4P/F model. For each phase we show the number of molecules in the supercell (\(N\)), the number of random H-isomers in the average (\(N_{ave}\)), the mean value of the lattice energy, \(\overline{U}_{S,ref}\), and its standard deviation, \(\sigma(U_{S,ref})\). Results for ices III and V are shown for both full and partial H-disorder. The last two columns in units of kJ/mol.
In the displayed region there appear seven different stable ice phases: Ih, II, III, V, VI, VII, and VIII. Note that ice XV, the H-ordered counterpart of ice VI, appears at temperatures lower than 130 K and is not shown in the figure.
The QH result derived with the q-TIP4P/F model in the quantum limit is presented in The free energy was calculated for the ice phases listed in Table 1 by using Eqs. and. There are striking differences between the calculated phase diagram and the experimental data. The main deviations of the model are:
* the most stable phase at low pressures is ice Ic instead of ices Ih or XI,
* the H-ordered ice IX is stable in the region where experimentally appears the H-ordered ice II,
* the H-ordered ice XIV occupies part of the stability region of the H-disordered ice VI and H-ordered ice XV,
* the H-disordered ice XII occupies the stability region of ice V.
We stress that ices II and V are not stable phases in the calculated phase diagram. Instead the ice polymorphs IX, XIV, and XII occupy large regions of stability. Such behavior has not been reported in any TI simulations of the phase diagram of ice using models based on the TIP4P potential.
One may wonder if these unexpected findings are a pathology of the QHA. Against this point of view it can be argued that the QH phase diagram of ices Ih, II, and III studied in Ref. is in reasonable agreement with TI simulations. Deviations found between QH and TI methods for several models (rigid TIP4P/2005 and TIP4PQ/2005, as well as flexible q-TIP4P/F) were more likely due to structural differences in the supercell employed for ice III than to limitations of the QHA. For this reason, we consider plausible that the QHA is providing valid information about the potential model for \((T,P)\) regions and ice phases that have not been previously studied by TI methods. Thus, the understanding of these unexpected findings is worth the effort.
In addition to the information displayed in the phase diagram of Fig. 3b, it is interesting to know the free energy differences between stable and metastable phases in several regions of the phase diagram. Free energy differences lower than the threshold of 0.05 kJ/mol are considered within the numerical error of the method and therefore will not be given a large physical significance. In the following Section we present a closer look into the calculated free energies at state points where the most stable phase is either ice Ic, IX, XIV, or XII.
## VI Quasi-harmonic free energies
### \(T\)=0 K and \(P\)=0
The QH Gibbs free energy, \(G_{0}\), of an ice phase at \(T=0\) K and \(P=0\) is the sum of two energy contributions
\[G_{0}=U_{S,0}+U_{Z,0}\;. \tag{7}\]
\(U_{S,0}\) is the static lattice energy for the equilibrium volume, \(V_{0}\), that includes the averaging term for the proton disorder,
\[U_{S,0}=U_{S}(V_{0})+\triangle U_{ave}\;. \tag{8}\]
\(U_{Z,0}\) is the zero-point energy calculated as
\[U_{Z,0}=\sum_{k}\frac{\hbar\omega_{k}(V_{0})}{2}\;. \tag{9}\]
In Table 4 we collect the values of \(U_{S,0}\), \(U_{Z,0}\), and \(V_{0}\) of the ice phases studied with the q-TIP4P/F model.
The most stable (i.e. lowest \(G_{0}\)) phases are ice Ic, XI, and Ih. Although the predicted order of increasing stability is: Ic\(>\)XI(\(Pna2_{1}\))\(>\)XI(\(Cmc2_{1}\))\(>\)Ih, the free energy differences between them are lower than the threshold of 0.05 kJ/mol. Such small differences are also conserved at higher pressures and temperatures. Therefore, the stability of ice Ih, Ic, and XI is nearly identical for the employed model.
\begin{table}
\begin{tabular}{c c c c c} \hline q-TIP4P/F & \(G_{0}\) & \(U_{S,0}\) & \(U_{Z,0}\) & \(V_{0}\) \\ \hline Ic & 6.97 & -61.77 & 68.74 & 32.35 \\ XI(\(Pna2_{1}\)) & 6.99 & -61.77 & 68.76 & 32.19 \\ XI(\(Cmc2_{1}\)) & 7.01 & -61.73 & 68.73 & 32.26 \\ Ih & 7.01 & -61.74 & 68.75 & 32.23 \\ IX & 7.22 & -61.40 & 68.62 & 25.60 \\ II & 7.47 & -60.60 & 68.08 & 25.11 \\ III (full) & 7.82 & -60.89 & 68.71 & 25.90 \\ XIV & 7.90 & -60.39 & 68.29 & 22.78 \\ XIII & 8.18 & -59.97 & 68.15 & 24.01 \\ V (full) & 8.28 & -60.07 & 68.35 & 23.72 \\ XII & 8.35 & -59.84 & 68.20 & 22.85 \\ IV & 8.60 & -59.56 & 68.17 & 22.93 \\ VI & 8.71 & -59.35 & 68.06 & 22.25 \\ XV & 8.74 & -59.21 & 67.95 & 22.34 \\ VIII & 14.05 & -52.81 & 66.86 & 20.55 \\ VII & 14.15 & -52.80 & 66.95 & 20.43 \\ \hline \end{tabular}
\end{table}
Table 4: QH Gibbs free energy (\(G_{0}\)) calculated at \(T=0\) K and \(P=0\) with the q-TIP4P/F model in the quantum limit. The partition of \(G_{0}\) into lattice (\(U_{S,0}\)) and zero-point (\(U_{Z,0}\)) energy is given. \(V_{0}\) is the equilibrium volume in Å\({}^{3}\)/molec. The data for ice III and V correspond to full H-disorder. Energy units in kJ/mol.
Then, when compared with other ice polymorphs, ice Ih displays several distinct properties at \(T=0\) K and \(P=0\). It has the largest volume (\(V_{0}\)) per water molecule, the lowest lattice energy (\(U_{S,0}\)) and the highest zero-point energy (\(U_{Z,0}\)) of all ice phases. The leading factor for the stability of ice Ih at \(T=0\) K and \(P=\)0 is its lowest lattice energy.
It is interesting to note that the nuclear quantum effect causes an expansion of the equilibrium volume of ice at low temperatures. This anharmonic effect in the structure of ice is predicted by the QHA due to zero-point contribution to the free energy in Eq.. This energy term is absent in the classical limit, where the free energy is equal to the static lattice energy at \(T=0\) K. Therefore, the equilibrium volume, associated to the minimum of the Gibbs free energy, is different in the quantum and classical limits. We find that the quantum volume, \(V_{0}\), of the ice phases at \(T=0\) K (see Table 4) are typically 4% larger than those ones derived in the classical limit (see \(V_{ref}\) values in Table 1).
### Increasing pressure at _T_=0 K
The values of \(G_{0}\) and \(V_{0}\) in Table 4 allow us to rationalize the changes in the stability of the ice phases upon an increase of the pressure at \(T=0\) K. A positive pressure will add a \(PV\) term to the free energy \(G_{0}\). Obviously the larger the ice volume the larger the increase in the free energy. Then ice Ih (with the largest volume) will be destabilized with respect to all other ice phases upon an increase of pressure. Ice IX is the best candidate to become stable. It has the lowest value of \(G_{0}\) after that one of ice Ih, and its equilibrium volume is significantly lower (20%) than that of ice Ih.
In we have represented the pressure dependence up to 0.14 GPa of the Gibbs free energy of the ice phases with lowest \(G\) at \(T=0\) K. We have plotted the free energies of ices Ic, Ih, and XI to show that their small free energy differences at \(P=0\) are conserved as pressures increases. We observe that at low pressures the phase having the minimum free energy is ice Ih. However as the pressure increases above 0.06 GPa, ice IX becomes more stable than ice Ih.
A further increase of the pressure will stabilize another ice phase with even lower volume than ice IX. In we show the Gibbs free energy, \(G\), of several ice phases in the range \(0.42<P<0.48\) GPa. The crossing of the free energy lines of ice IX and XIV at 0.45 GPa is the fingerprint for a phase transition from ice IX to ice XIV.
Given a pressure \(P\), for thermodynamic consistency in the low-temperature limit (\(T\to 0\) K) one expects the stable ice polymorph to be an H-ordered phase. This does not appear to be always the case in our calculations. Thus, at \(P<60\) MPa we find cubic ice Ic to be the low \(T\) stable phase. However, as indicated above, free-energy differences between ice Ic and the H-ordered ice XI are smaller than our sensitivity limit (\(\sim 0.05\) kJ/mol). Also, ice VI appears to be the stable phase at low temperatures in the region between 3 and 5 GPa. In this case one would expect the corresponding H-ordered phase (ice XV) to be the stable polymorph, but its free energy for \(T\to 0\) K is higher than that of ice VI.
Gibbs free energy, \(G\), of the ice phases with lowest \(G\) at \(T=0\) K and pressures below 0.14 GPa. The results correspond to the QHA using the q-TIP4P/F model. The curves for ices Ih, XI, and Ic are nearly identical at the displayed energy scale. A phase transition from ice Ih to ice IX is predicted at \(P\sim 0.06\) GPa.
Gibbs free energy, \(G\), of the ice phase with lowest \(G\) at \(T=0\) K and pressures in the interval [0.4,0.52] GPa. The results are derived by the QHA using the q-TIP4P/F model. A phase transition from ice IX to ice XIV is predicted at \(P\sim 0.45\) GPa.
### Stability of ice IX versus ice II
A consequence of the stability of ice IX is that ice II does not appear (i.e., it is metastable) in the QH phase diagram of This metastability of ice II disagrees obviously with the experimental phase diagram. In Figs. 4 and 5 the free energy \(G\) of ice II is larger than that of ice IX. The difference is nearly independent of the pressure, as the \(G(P)\) curves of ices IX and II are approximately parallel. If we measure the free energy difference between these phases at \(T=0\) by the ordinates in the origin of (i.e., the \(G_{0}\) values of ice IX and II in Table 4), one gets
\[\Delta G_{0}(\mbox{II-IX})=0.25\;\mbox{kJ/mol}\;. \tag{10}\]
The positive value implies that ice IX is more stable than ice II. The energy partitioning of \(G_{0}\) in Table 4 shows that the leading term for the larger stability of ice IX is its lower lattice energy.
Note that if the \(G(P)\) curve of ice IX were omitted from Fig. 4, then the first transition as a function of pressure would correspond to the crossing of the Ih and II free energy curves at 0.11 GPa.
Given the large stability of ice IX predicted by the q-TIP4P/F model, we want to address the following question: Is this stability a consequence of the flexibility of the model or it has its origin in the TIP4P-character of the potential?
To this aim we have calculated the QH free energies of several ice phases by using the rigid TIP4P/2005 model. This model was parameterized for water simulations in the classical limit.
\[G_{0,cla}\equiv\overline{U}_{S,ref}\;, \tag{11}\]
\[V_{0,cla}\equiv V_{ref}\;. \tag{12}\]
For reference, the values of \(G_{0,cla}\) and \(V_{0,cla}\) calculated with the TIP4P/2005 model for ices Ih, XI, II, and XIV are shown in Table 5.
\[\Delta G_{0,cla}(\mbox{II-IX})=0.58\;\mbox{kJ/mol}\;. \tag{13}\]
This free energy difference between ice IX and II is even larger than that found for the flexible q-TIP4P/F model.
The classical QH phase diagram of ice Ih, II, and IX was calculated with the rigid TIP4P/2005 model at temperatures up to 300 K and pressures below 0.6 GPa. We find that ice IX is more stable than ice II in the whole studied \((T,P)\) region. Therefore ice II is metastable in the classical phase diagram of the TIP4P/2005 model. This result is identical to that found for the flexible q-TIP4P/F model. Our conclusion is that the large stability of ice IX is a property of the TIP4P-character of the model, and not a consequence of the added molecular flexibility.
We believe that limitations inherent to the exploration of the phase diagram by TI methods is the reason why the stability of ice IX and the metastability of ice II has not been detected in previous studies using TIP4P-like models. An advantage of the QHA is that the brute force calculation of free energies allows a thorough exploration of state points for all ice phases.
### Increasing pressure at _T_=250 K
An unexpected result of the calculated phase diagram at temperatures around 250 K is that ice XII, the H-disordered counterpart of ice XIV, occupies the stability
\begin{table}
\begin{tabular}{c c c} \hline TIP4P/2005 & \(G_{0,cla}\) & \(V_{0,cla}\) \\ \hline Ih & -62.99 & 31.34 \\ IX & -62.71 & 24.87 \\ II & -62.13 & 24.30 \\ XIV & -61.72 & 22.10 \\ \hline \end{tabular}
\end{table}
Table 5: Gibbs free energy (\(G_{0,cla}\) in kJ/mol) calculated with the rigid TIP4P/2005 model for several ice phases at \(T=0\) K and \(P=0\) in the classical limit. \(V_{0,cla}\) is the equilibrium volume in Å\({}^{3}\)/molec.
Gibbs free energy, \(G\), of the ice phase with lowest \(G\) at \(T=250\) K and pressures in the interval [0.44,0.46] GPa. The results correspond to the q-TIP4P/F model and the QHA. A phase transition from ice III to ice XII appears at \(P\sim 0.45\) GPa.
3). The pressure dependence of the free energy of the ice phases with lowest \(G\) is presented in at 250 K. The crossing of the \(G(P)\) curves of ice III and ice XII at \(P=0.45\) GPa indicates that ice XII becomes the stable phase at 250 K for pressures higher than 0.45 GPa.
It is interesting to note that at the pressures shown in the free energy of ice V (with full H-disorder) is only slightly higher (\(\sim 0.06\) kJ/mol) than that of ice XII. Besides, ice XII displays lower volume than ice V. Therefore an increase in the pressure will always stabilize ice XII with respect to ice V. At \(T=250\) K the QHA predicts that the coexistence pressure for ices V-XII is 0.33 GPa. The equilibrium volumes at this state point (\(T=250\) K, \(P=0.33\) GPa) are 23.1 A\({}^{3}\)/molec. and 23.8 A\({}^{3}\)/molec., for ice XII and ice V, respectively.
Note that at 250 K the pressure interval where ice V appears as stable phase in the experimental phase diagram is about [0.35,0.6] GPa (see Fig. 3a). In this pressure range the q-TIP4P/F model predicts that the free energy difference between ices V and XII increases from a vanishingly small value (at 0.35 GPa) to a maximum value of 0.1 kJ/mol (at 0.6 GPa). Therefore free energy differences between ice XII and V are relatively small. Similar values for the free energy of ice V and XII have been already reported for the TIP4P/2005 model at \(P=0.5\) GPa by TI simulations in the classical limit.
## VII Quantum vs.
The QH phase diagram of ice shown in for the q-TIP4P/F model is characterized by the stability of ice IX, XII, and XIV over large regions of the \((T,P)\) plane. However, these phases have not been reported as stable ones in previous studies. For this reason our calculated phase diagram has little resemblance to previous ones derived by using rigid TIP4P-like potentials in combination with TI methods.
Then, it is interesting to recalculate our phase diagram of ice under omission of the phases IX, XII, and XIV. In addition we will consider ice Ih as unique representative of the phases Ih, Ic, and XI.
The quantum limit of the new phase diagram of the q-TIP4P/F model is presented in Now we find the following stable phases: Ih, II, III, V, VI. VII and VIII. Both ices III and V correspond to full H-disordered lattices. This phase diagram is in reasonable qualitative agreement to the experimental one. Moreover, it is also in reasonable agreement to those phase diagrams derived for TIP4P-like potentials.
Concerning the thermodynamic consistency of these results for \(T\to 0\) K, we emphasize that the H-disordered ices Ih and VI cannot strictly be the low-temperature stable phases for any pressure \(P\). For this question, the arguments are the same as those given above in Sec. VI.2 and are not repeated here.
Our results in include the averaging of the lattice energy over H-disorder for ice III, V, and VI. None of the previously published phase diagrams calculated with TIP4P-like potentials include any kind of disorder averaging. In fact, different single H-isomers of ice III seem to have been employed for the calculations with several TIP4P-like potentials (i.e., TIP4P, TIP4P/2005, TIP4PQ/2005). If the ice III structure employed for each potential model has a different H-configuration, then the lattice energy of ice III may be affected by an uncontrolled factor. This situation makes it difficult to draw definitive conclusions about differences found in calculated phase diagrams with different H-isomers of ice III. This uncertainty should affect not only ice III, but also the stability region of other phases (Ih, II, V) having a boundary with ice III.
It is interesting to compare the quantum phase diagram in with that one calculated within the classical limit. The main difference between both limits is related to the stability region of ice II. It is much larger in the quantum case. The origin of this quantum effect is related to the lower zero-point energy of ice II, in comparison to ices Ih, III, and V (see the \(U_{Z,0}\) values given in Table 4). The quantum stabilization of ice II has been already discussed in Ref..
An additional difference between quantum and classical phase diagrams is that the coexistence lines that are nearly horizontal (i.e., Ih-III, III-V, V-VI) appear shifted to higher pressures in the classical limit.
QHA of the phase diagram of ice using the q-TIP4P/F model in the quantum limit. Ices III and V are full H-disordered. The following ice phases were omitted from the QH calculation: ice IX, XIV, and XII.
The rising of the pressure occurs at all studied temperatures. This quantum effect is most easily explained at \(T=0\) K. Let us consider ices VI and V as an example. In the quantum limit, the free energy difference between ice VI and V at \(T=0\) K and \(P=0\) is (see the \(G_{0}\) data in Table 4),
\[\triangle G_{0}(\mathrm{VI}-\mathrm{V})=0.4\;\mathrm{kJ/mol}\;. \tag{14}\]
However, in the classical limit one gets using Eq. and the values of Table 1,
\[\triangle G_{0,ela}(\mathrm{VI}-\mathrm{V})=0.7\;\mathrm{kJ/mol}\;. \tag{15}\]
In both cases \(\Delta G_{0}>0\), i.e., ice VI is less stable than ice V. However, in the quantum case \(\Delta G_{0}\) is lower. The reason is that the zero-point energy (see the \(U_{Z,0}\) data in Table 4) tends to stabilize ice VI with respect to ice V by \(\sim-0.3\) kJ/mol.
On the other side, although the volume of ice VI is lower than that of ice V (see \(V_{0}\) data in Table 4), the volume differences are found to be nearly the same in the quantum (\(\Delta V_{0}=-1.5\) A\({}^{8}\)/molec.) and classical cases (\(\Delta V_{0,ela}=-1.4\) A\({}^{8}\)/molec., see the \(V_{ref}\) data in Table 1).
By increasing the pressure, the phase with lower volume (ice VI) will become more stable.
\[P\thicksim-\frac{\Delta G_{0}}{\Delta V_{0}}\;. \tag{16}\]
Note that while the denominator is similar in both quantum and classical limits, the numerator is lower in the quantum case. Therefore the coexistence pressure for c V-VI at \(T=0\) K is reduced in the quantum limit with respect to the classical one.
A similar argument explains why the transitions Ih-III and III-V are displaced to higher pressures in the classical case.
## VIII Conclusions
The phase diagram of ice has been studied by a quasi-harmonic approximation using the flexible q-TIP4P/F model of water. The simplicity of this approach allows us to include all known ice polymorphs (except ice X) and all state points for \(T<400\) K and \(P<10\) GPa.
Surprisingly the simple QHA seems to be accurate enough to reproduce free energy differences between ice phases, in spite of the large complexity and variety in their crystal structures. This conclusion about the success of the QHA is derived using a simple model potential, but its validity is expected to be largely independent of the model. Therefore it opens a route for the study of the whole phase diagram of ice by ab initio electronic structure methods.
The H-disorder of many ice phases is an additional complication in the calculation of their phase diagram. The QHA has allowed us to quantify the influence of this effect. The disorder averaging of the lattice energy of ice III has been proven to be important to obtain a converged phase diagram, at least using TIP4P-like models and ice III supercells with full H-disorder. Disorder averaging of vibrational free energies has been shown to be comparatively less important. In addition to ice III, the disorder averaging of the lattice energy of ice XII and V has been shown to be also significant for the ice stability.
We stress that phase diagrams calculated using a single random H-isomer of ice III may be affected by an uncontrolled energetic factor that can be highly significant in the final result. An immediate consequence of this is that comparison of phase diagrams calculated using a single, but different, H-isomer of ice III might not be physically sound. The reason is that the stability of the disordered phase may depend strongly on the employed H-isomer.
The QH phase diagram of ice with the flexible q-TIP4P/F model has been calculated by performing a disorder averaging of the lattice energy of the H-disordered ice phases. We have found an unexpected large stability of several ice phases, specially the H-ordered ices IX and XIV, and also the H-disordered ice XII. The presence of these phases in the calculated phase diagram implies that both ice II and V are metastable phases with the q-TIP4P/F model. This finding disagrees with the experimental phase diagram. We have checked that ice IX remains more stable than ice II if the phase diagram is calculated using the rigid TIP4P/2005 model in the classical limit.
The phase diagram of is calculated with the q-TIP4P/F model in the classical limit.
The QH free energy and volume of several ice phases have been analyzed at \(T=0\) K and \(P=0\). The free energy has been partitioned into lattice and zero-point energies. These contributions are important magnitudes in the analysis of the stability of the ice phases as a function of pressure.
By excluding ices IX, XIV, and XII from the calculation, we find that the phase diagram of the q-TIP4P/F model shows qualitative agreement to both experimental and previously simulated ones by using TIP4P-like models. The comparison of the quantum and classical limits shows several differences. The most important are the increase in the stability of ice II and the shift of the coexistence lines III-V and V-VI to lower pressures in the quantum case. Similar conclusions were reached previously in Ref.. Differences in the zero-point energies of the ice phases provide an explanation for these effects.
|
10.48550/arXiv.1310.1006
|
The phase diagram of ice: a quasi-harmonic study based on a flexible water model
|
R. Ramirez, N. Neuerburg, C. P. Herrero
| 5,170
|
10.48550_arXiv.2203.06682
|
## I Introduction
The van der Waals (vdW) density functional (vdW-DF) method for strictly nonempirical density functional theory (DFT) has been successfully applied in materials and chemistry for more than two decades. vdW-DF opened the door for early DFT predictions of adhesion among graphene sheets and in lubricants, weak molecular binding, and the weak adhesion of nuclae bases and other organics on graphene and oxides. The functionals of the vdW-DF method have no empirical parameters and avoid double counting of correlation. Predating the set of also-popular dispersion-corrected DFTs, the accuracy and robustness have systematically improved over time.
The success of vdW-DF motivates continued investments to design even better non-empirical versions of vdW-DF. The vdW-DF method is built from many-body perturbation theory (MBPT) analysis of the nature of the fully interacting electron ground state. This strategy led, for example, to a straightforward extension to include spin. We can also directly interpret the quality and performance differences in terms of the spatial variation in and hence nature of the different contributions to the exchange correlation (XC) energy.
The vdW-DF method provides a formally exact framework for a systematic inclusion of nonlocal-correlation effects. Part of the MBPT foundation for the vdW-DF method was first described in the same paper that established logic for correlations in the constraint-based generalized gradient approximation (GGA). As such, it represents a third generation of XC-energy functionals in an electron-gas tradition that started with the local spin density approximation (LDA) and led to the highly successful PBE and PBEsol GGAs. The overall logic of this tradition is to gradually introduce a controlled increase in flexibility so that we can reliably benefit from more of the pool of physics insight and trusted MBPT inputs.
Finding an accurate, general-purpose functional is important since theory often concerns complex materials, i.e., systems where the atomic structure is not fully established. There, DFT calculations must be used to first assert which are the most plausible of several candidate motifs, for example, as in Ref.. The consistent-exchange vdW-DF-cx version (abbreviated CX) is crafted to seek high accuracy simultaneously for molecules, bulk, and surfaces but (as discussed elsewhere) it uses a type of XC guidelines that favors dense-matter and noncovalent (NOC) interaction problems. The vdW-DF2-b86R (abbreviated B86R) uses a different nonlocal correlation but retains and, in fact, enhances the general-purpose char acter. Other vdW-DFs are often found better at some rather than other types of problems; See supplementary information (SI) material for an illustration of molecular-performance variations and see SI Table S.1 for a list of functional abbreviations that we shall systematically use below. All of the vdW-DFs, including the unscreened-hybrid forms, fail in some cases to correctly balance vdW attraction with the repulsion provided by the gradient correction to exchange, for example, in complex metal-organic-framework systems. Nevertheless, in DFT we seek to characterize and predict molecular reaction energies and transition barriers at the 1 kcal/mol (or 43 meV) limit that defines so-called chemical accuracy. Higher accuracy still is needed for understanding chemical fuels, CO\({}_{2}\) capture, batteries, and biochemistry. These are cases where we must understand the role of NOC interactions as they act in concert and competition. To get at the complex-materials challenge, we must correctly balance the XC terms in just one general-purpose, yet highly accurate, vdW-DF design that also avoids density-driven DFT errors.
This paper reports the design and testing of a range-separated hybrid (RSH) vdW-DF. It is termed vdW-DF2-ahbr and abbreviated AHBR because it builds on the vdW-DF2 nonlocal-correlation description and on an analytical-hole (AH) analysis of the nature of exchange in the B86R variant. We show that it stands out by having an exceptional general-purpose capability and clearly outperforms even the recent AHCX design of a RSH vdW-DF, for reasons we explain.
Figure 1, below, summarizes our assessment of performance over broad molecular properties, illustrating that AHBR is both highly accurate and has an excellent transferability. We find that AHBR can navigate generic density-driven functional errors that, for example, often affect molecular transition states.
We propose that the new AHBR be used to test the status of the vdW-DF method, as we also partly illustrate, because it is free of a performance bias. The performance of RSH HSE+D3 is independent of the benchmark type in the very broad GMTKN55 suite on broad molecular properties. The unscreened hybrid B3LYP+D3 is an improvement over HSE+D3 on molecular transition states and NOC interactions but not across the board, SI Tables S.2-3. HSE+D3 and B3LYP+D3 are widely used in materials science and chemistry, respectively, and their transferability sets the bar for the generic vdW-DF method testing. The new RSH AHBR (black curve in Fig. 1) outperforms HSE+D3 (gray) across all types of molecular properties and matches (clearly improves) the B3LYP+D3 performance for the important group 3 of molecular transition-state benchmarks (for the rest of the GMTKN55 benchmarks). As summarized in Fig. S 1 and documented in Tables S 2-3 of the SI material, these observations holds for either of the weighted-mean-absolute-deviation measures that are suggested and used in Ref.. Unlike AHCX (dark red), AHBR is more successful than HSE+D3 and B3LYP+D3 on, for example, the BH76 benchmark set on molecular barrier heights, problems that are sensitive to density-driven DFT errors. The AHBR provides systematic accuracy gains over present-standard hybrid choices.
The specific contributions of the paper can be summarized as follows. We first complete a robust planewave-based assessment across the full-GMTKN55 benchmark suite, documenting that the unscreened hybrid extension of B86R, abbreviated DF2-BR0, provides the best performance on molecular properties. This is true for the GMTKN55 suite and among all of the vdW-DFs, including the RSH vdW-DFs, Fig. S 2 of the SI material. Our broad documentation is consistent with a very recent independent observation for proton transition barriers. We proceed to define the AHBR, the RSH generalization of DF2-BR0, using an AH characterization for exchange effects in B86R. We validate that AHBR is an exceptional performer across GMTKN55, Fig. 1, and retains a strong performance for bulk and some layered materials. Finally, we illustrate the usefulness of AHBR for DNA assembly and molecular adsorption problems, finding good agreement on quantum-chemistry reference calculations, the correct site preference for the CO/Pt problem, and a good performance for characterization of CO\({}_{2}\) uptake in two metal organics framwors (MOFs).
The rest of the paper is organized as follows. The theory section II presents an overview of the vdW-DF method, analysis of the B86R exchange hole, and contains the formulation of the new AHBR. Section III contains results and discussions, including illustrations of AHBR accuracy. Section IV contains our conclusion and outlook. The paper has two appendices giving computational details, including the electrostatic-environment approach used to complete planewave benchmarking across the GMTKN55 suite.
## II Theory
Central in MBPT and in the electron-gas foundation of DFT is the screened density response \(\delta n(\omega)\) to some external-potential change \(\delta\Phi^{\omega}_{\rm ext}\) oscillating at frequency \(\omega\). In MBPT we can, at least in principle, compute the nonlocal response function \(\chi_{\lambda}({\bf r},{\bf r}^{\prime};\omega)\equiv\delta n({\bf r})/\delta \Phi^{\omega}_{\rm ext}({\bf r}^{\prime})\), often expressed as a function of a complex frequency \(\omega=iu\). We can also do that at a range of an assumed reduced strength \(0<\lambda<1\) of the electron-electron interaction \(\lambda\hat{V}\).
\[E_{\rm xc}=-\int_{0}^{1}\,d\lambda\,\int_{0}^{\infty}\,\frac{du}{2\pi}\,{\rm Tr }\{\chi_{\lambda}(iu)V\}-E_{\rm self}\,, \tag{1}\]of the XC energy functional \(E_{\rm xc}\). Here \(V=|{\bf r}-{\bf r}|^{-1}\) denotes the matrix element of the electron-electron interaction \(\hat{V}\), \(u\) is an imaginary frequency argument in the response description, while the last term is the electron self energy \(E_{\rm self}={\rm Tr}\{\hat{n}V\}/2\). The expressions for \(E_{\rm xc}\) and \(E_{\rm self}\) involve Coulomb-weighted traces, that is, integrations in spatial coordinates of \(|{\bf r}-{\bf r}^{\prime}|^{-1}\) times \(\chi_{\lambda}({\bf r}^{\prime},{\bf r};\omega)\) and times the electron density \(n({\bf r}^{\prime})\), respectively. Also we have (at every coupling-constant value \(\lambda\)) added an auxiliary potential that keeps the electron density \(n({\bf r})\) unchanged across the implied adiabatic turn on of the electron-electron interaction \(\hat{V}_{\lambda}=\lambda\hat{V}\). The actual XC potential used in the Kohn-Sham (KS) scheme for efficient DFT calculations is simply the \(\lambda\to 0\) limit of this auxiliary potential. It is given by a functional derivative of the XC energy, as discussed many places elsewhere.
In MBPT, we compute the response functions \(\chi_{\lambda}(\omega)\) as a ground-state expectation value of correlations between density fluctuations. As such, \(\chi_{\lambda}(\omega)\) is directly reflecting the Lindhard-type screening that exists in the electron gas at assumed coupling constant \(\lambda\), as discussed, for example, in Ref.. The screening is given by the dielectric function \(\kappa_{\lambda}(\omega)=(1+\lambda V\chi_{\lambda})^{-1}\). For practical DFT, we seek XC functional approximations that contain the most pertinent physics contents of the widely complex, many-body interacting processes that define \(\chi_{\lambda}(iu)\). The massive DFT usage allows us to get successively more adapt at this as long as we stay systematic and can interpret performance differences, for example, within MBPT.
In the electron-gas tradition for XC functional designs [17, 12, 14, 43, 46, 48, 99, 100, 102, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 199, 190, 191, 198, 199, 190, 192, 194, 195, 196, 197, 198, 199, 199, 198, 199, 199, 199, 190, 191, 190, 192, 193, 194, 195, 196, 197, 198, 199, 199, 199, 198, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 1999, 199, 199, 199, 1999, 199, 1999, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 199, 1999, 1999, 199, 1999, 199, 1999, 199, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 19999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 19999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 19999, 1999, 1999, 19999, 19999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 19999, 1999, 19999, 1999, 19999, 1999, 1999, 19999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 19999, 19999, 1999, 19999, 1999, 19999, 1999, 1999, 1999, 19999, 1999, 1999, 19999, 1999, 1999, 1999, 19999, 1999, 19999, 19999, 1999, 1999, 19999, 19999, 19999, 1999, 1999, 19999, 1999, 19999, 1999, 1999, 1999, 1999, 19999, 1999, 19999, 1999, 1999, 19999, 1999, 19999, 19999, 1999, 1999, 19999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 19999, 1999, 1999, 19999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 19999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 1999, 19999, 1999, 19999, 1999, 1999, 19999, 1999, 1999, 1999, 19999, 19999, 1999, 1999, 19999, 1999, 1999, 1999, 1999, 1999, 19999, 19999,exchange approximation \(n_{\rm x}^{\rm Fo}({\bf r};{\bf r}^{\prime})\) to \(n_{\rm x}\) results by considering one-particle density matrices formed from the KS-wavefunction solutions, for example, as summarized in the discussion provided in Ref..
### The vdW-DF framework
To begin a summary of vdW-DF, we note that since LDA and GGAs are completely set from a modeling of an underlying XC hole, we are also ready to capture vdW forces as defined from an electrodynamics coupling of electron-density fluctuations. Any XC functional can be seen as the net binding energy of the electrons and associated XC holes, Eq.. However, it is also clear that the electron and the associated (GGA-type) XC hole form an antenna of charged parts that have a mutual zero-point energy dynamics. The electron-XC-hole pairs will interact even across regions that have but a sparse or no electron density. In fact, this electron-gas electrodynamics coupling is a systematic generalization of the original London picture of dispersion forces among noble-gas atoms.
The vdW-DF method achieves a systematic extension of MBPT-based GGAs by recasting the exact XC functional as an electrodynamics problem, while counting (via a frequency contour integration) the coupling-induced shifts in energies for collective excitations. Thus, for the vdW-DF XC energy description, we rely on a formally exact recast of the ACF result,,
\[E_{\rm xc}=\int_{0}^{\infty}\,\frac{du}{2\pi}\,{\rm Tr}\{{\rm ln}(\kappa_{\rm ACF }(iu))\}-E_{\rm self}\,. \tag{3}\]
In Eq., we have introduced an effective, spatially nonlocal, dielectric functional function \(\kappa_{\rm ACF}(iu)\). The formal XC evaluation is given as a weighted \(\lambda\) average of \(\kappa_{\lambda}(\omega)\) and hence of \(\chi_{\lambda}\) and there is full equivalence of Eqs. and, given consistent approximations; The recast is also equivalent to the XC-hole formulation, Eq.. The electrodynamical recast, Eq., simplifies the porting of the ideas of the cumulant expansion to the vdW-DF development work. Specifically, we can link the Ashcroft-Langreth-Lundqvist picture of vdW-binding contributions, and the MBPT input to the constraint-based GGAs. In turn this allows the vdW-DF method to provide an effective (MBPT-guided) approximation to the \(\lambda\)-averaged response description, Eq., as discussed in Refs..
In the vdW-DF method, we furthermore use a so-called internal functional \(E_{\rm xc}^{\rm in}\) to first set a lowest-order (GGA-level) approximation for the screening. The screening is described by a truly nonlocal dielectric function \(\epsilon(\omega)\) (but we suppress spatial coordinates in this discussion). This dielectric function has collective excitations, plasmons, that sets the start of the response modeling, given by an effective electron-gas susceptibility \(\alpha(\omega)=(\epsilon(\omega)-1)/4\pi\).. First, we approximate the plasmon propagator as \(S_{\rm xc}(\omega)={\rm ln}(\epsilon(\omega))\) and rely on an explicit two-pole approximation that reflects plausible assumptions and all plasmon-related conservation laws.
\[E_{\rm xc}^{\rm in}=\int_{0}^{\infty}\,\frac{du}{2\pi}\,{\rm Tr}\{{\rm ln}( \epsilon(\omega))\}-E_{\rm self}\,, \tag{4}\]
In summary, we have a formal link between the GGA-level internal functional and structure of the starting approximation \(S_{\rm xc}(\omega)\) for an emerging description of the actual plasmon dynamics.
We also explicitly enforce a longitudinal projection of the response in the dielectrics approximation function
\[\kappa_{\rm ACF}(\omega)=-\nabla\cdot\epsilon(\omega)\nabla V/4\pi\,. \tag{5}\]
We note that having this projection inside the recast ACF, Eq., produces terms that capture the vdW attraction as described in the presence of electron-gas screening. Moreover, the use of \(\epsilon(\omega)=\exp(S_{\rm xc}(\omega))\) implies a cumulant-or-cluster-expansion logic in the response description; This allows the vdW-DF method to also pick up high-order susceptibility and screening effects, i.e., to balance the vdW attraction by other nonlocal-correlation effects.
Comparison of exchange enhancement factor \(F_{x}(s)\) (top panel) and of the Gaussian-exponent prefactor \(\mathcal{H}(s)\) that determines the main damping in the AH modeling for the density-density correlation defining the underlying exchange holes.
Finally, given the choice of GGA-level plasmon modeling, Eq., we repeat the contour integration evaluation with Eqs. and to secure efficient kernel-based evaluation of nonlocal-correlation energies. This approach means that the vdW-DFs versions have no discernible cost increase over GGA in planewave codes. Similarly, the new class of range-separated hybrid (RSH) vdW-DFs has the same costs as the HSE, i.e., a RSH that is based on the PBE GGA. However, use of the vdW-DFs sometimes requires a better convergence of the electron-density variation.
The resulting non-empirical vdW-DF description stands out, for example, in the class of vdW-inclusive functionals by treating all interaction contributions on the same electron-gas footing. We avoid auxiliary inputs beyond ground-state DFT and we avoid all need for semi-empirical adjustments, for example, to ameliorate double counting of nonlocal correlations. The vdW-DF method is set up to describe vdW interactions as they emerge in concert and competition with covalent and ionic binding and with orbital-interaction shifts produced by wider nonlocal-correlation effects. The latter effects are, for example, documented to counteract contributions to vdW interaction contributions from the high-density regions near the nuclei as well as from the saddle point of the electron-density variation that exists between fragments. Noting that formal MBPT sets the nature of the exact XC energy, we seek to use the MBPT as a guide. This is done, for example, by trying to recycle the accuracy gains that were made on the exchange description in the MBPT-based GGAs.
### vdW-DF versions
Prior vdW-DF versions involve a controlled introduction of systematic design changes. For example, vdW-DF1 and vdW-DF2 have the same overall structure but differ in whether we prioritize MBPT or scaling insight on exchange to model the collective-excitation response that forms the starting point of the \(E_{\rm sc}^{\rm nl}\) evaluation. They also differ in how we enforce a method criterion that the actual exchange component of a vdW-DF should not itself lead to spurious weak binding, since the vdW attraction is a correlation effect.
For a computationally efficient evaluation, the standard general-geometry formulation expands the recast ACF, Eq., to second order in the (nonlocal) plasmon propagator \(S_{\rm xc}(\omega)=\ln(\epsilon(\omega))\).
\[E_{\rm xc}^{\rm DFs} \approx E_{\rm xc}^{\rm in}+E_{\rm c}^{\rm nl} \tag{6}\] \[E_{\rm c}^{\rm nl} = \int_{0}^{\infty}\,\frac{du}{4\pi}\operatorname{Tr}\{S_{\rm xc}^ {2}(iu)-(\nabla S_{\rm xc}\cdot\nabla V/4\pi)^{2}\}\,, \tag{7}\]
set the details of \(S_{\rm xc}(iu)\).
In all present vdW-DFs, the internal \(E_{\rm xc}^{\rm in}\) functional is chosen semi-local (GGA-like), comprising LDA correlation and a simple choice of physics-motivated gradient-corrected exchange.
Radial variation in the scaled exchange hole \(J(s,y=k_{\rm F}({\bf r})|{\bf r}-{\bf r}^{\prime}|)=n_{\rm x}({\bf r},{\bf r} ^{\prime})/n({\bf r})\) for PBE, PBEsol, and B86R, all as described in an analytical-hole (AH) model parameterization. The shapes of these holes define the exchange components of PBE, of PBEsol, and of B86R, respectively. Using the latter exchange-hole model, we define the here-released RSH AHBR, following the design steps that we recently documented in crafting AHCX.
The gradient-exchange choices used in \(E_{\rm xc}^{\rm in}\) are defined by formal-MBPT input. That input is used through the Langreth-Perdew and Langreth-Vosko (LV) analysis of a screened-exchange nature (that is natural for bulk and metals) in the first general-geometry release, vdW-DF1. It is set as the Schwinger exact-exchange scaling analysis in the second, vdW-DF2. The Schwinger MBPT result is instead relevant for capturing exchange effects in molecules; This fact is, for example, revealed by a demonstration that it leads to a non-empirical derivation of Becke-88 exchange, when interpret in the GGA framework.
The ACF foundation, Eqs. through, does motivate the use of Eq. to pick the exchange. That is, we should ideally use the internal-exchange formulation, and hence the \(S_{\rm xc}(iu)\) form, to also define the actual exchange. However, the inner functional is deliberately kept simple, while the overall exchange design must also reflect other considerations. Taken together, these observations mean that it is presently not possible to directly implement this idea.
\[E_{\rm xc}^{\rm DFs}\equiv E_{\rm xc}^{0}+E_{\rm c}^{\rm nl}\,, \tag{8}\]
The top panel of compares the so-called exchange-enhancement factor \(F_{\rm x}\) that defines the nature of the gradient-corrected exchange in all GGA descriptions as well as in present vdW-DFs. We show the variation with the the scaled density gradient \(s=|\nabla n|/(2nk_{\rm F})\) (where \(k_{\rm F}=(3\pi^{2}n)^{1/3}\) is the local Fermi-wavevector) for the four XC functional cases that are of interest here.
\[E_{\rm x}^{\rm GGA}=\int_{\bf r}n({\bf r})\epsilon_{\rm x}^{\rm LDA}(n({\bf r} ))F_{\rm x}(s({\bf r}))\,. \tag{9}\]
Here \(\epsilon_{\rm x}^{\rm LDA}=-3k_{\rm F}/4\pi\) denotes the LDA exchange result, i.e., the exchange energy-per-particle value that characterizes a homogeneous system (at density \(n=n({\bf r})\)). The variations in this gradient corrected exchange is thus set alone by the enhancement form of factor \(F_{\rm x}(s)\). For \(E_{\rm xc}^{\rm in}\) the enhancement is set as a quadratic expansion, \(F_{\rm x}^{\rm in}(s)=1+\mu s^{2}\), Refs..
Exchange descriptions of the popular constraint-based GGAs arise when one imposes an exchange-hole-conservation criterion in the modeling of density gradient effects. This was first done in the (revised) PW86. It was repeated in the popular PBE and PBEsol designs while then also paying attention to preventing the exchange-hole depth from dramatically exceeding the local-electron density. The PBE and PBEsol also adhere to a local implementation of the so-called Lieb-Oxford bound on the high-\(s\)\(F_{\rm x}(s)\) variation, but the actual (globally-implemented) bound does not, in practice impact this discussion of picking a robust GGA-type (or hybrid-type) exchange for the vdW-DFs.
For the vdW-DFs we must craft an asymptotic \(F_{\rm x}(s)\) behavior that produces an adequate but not excessive repulsion by gradient-corrections to exchange for weakly interacting molecules. This is to ideally eliminate (without overcompensating) spurious weak-system binding by the errors in the LDA exchange description.
The actual exchange descriptions in vdW-DF1 and vdW-DF2 are set as in the revPBE variant of PBE and as the refitted PW86 form, respectively. In both cases the selections were made following analysis of the weak binding of noble-gas and small-molecule dimers, Refs.. Both of these exchange choices are more repulsive than the PBE exchange, i.e, they have a gradient-correction to exchange that gives a stronger push to separating fragments. That extra repulsion is needed since, in the vdW-DFs, we are also upgrading to a truly nonlocal correlation description \(E_{\rm c}^{\rm nl}\). That new term includes vdW forces and gives a stronger attraction mechanism than what exists in PBE.
The consistent-exchange CX - and hence with the CX0P and AHCX hybrid extensions - aligns the two ways that exchange insight underpins the vdW-DF details, in the inner-functional \(E_{\rm xc}^{0}\) and in Eq., as far as possible. The idea is to look at the impact of XC-balance on the binding-energy descriptions instead of on the total-energy descriptions. The move to reconcile the inner and actual exchange has the benefit that we use the Lindhard-screening logic, implied in the expansion Eqs. and, as well as current conservation, to effectively balance XC terms.
Our general design strategy is to maximize the role of MBPT inputs, like Lindhard screening, because it is a promising path to securing high accuracy broadly in one general-purpose XC design. In formal MBPT, we summarize the net impact on the electron-electron interactions in terms of a so-called, self-energy term \(\Sigma_{\rm xc}({\bf r},{\bf r}^{\prime},\omega)\). It determines how a single-electron excitations propagate in the fully interacting system. It plays a similar role as the DFT exchange-correlation potential \(v_{\rm xc}({\bf r})=\delta E_{\rm xc}/\delta n({\bf r})\) that defines the independent-particle dynamics of in the DFT representation of the same system (except that it is both frequency dependent and truly nonlocal). A key argument for the vdW-DF design strategy is that the formal MBPT description of the total energy is tolerant, i.e., one can get good results even when interaction effects are merely approximated by perturbation theory for the self-energies \(\Sigma_{\rm xc}({\bf r},{\bf r}^{\prime},\omega)\). We get a sound electron-response description as long as we keep those (so-called skeleton) diagrams that capture the essential physics and dominant features of the electron-gas response. Also, in principle, that robustness extents to the choice of \(E_{\rm xc}\)thanks to the Sham-Schluter relation between \(v_{\rm xc}({\bf r})\) and \(\Sigma_{\rm xc}({\bf r},{\bf r}^{\prime},\omega)\).
It is important to observe, however, that we have not in CX (nor in AHCX) enforced a complete exchange alignment for all types of problems. The Lindhard screening and current-conservation are essential parts of the ACF result for the exact XC functional. The CX and associated hybrid designs allow us to use this idea, but only for descriptions of system processes where the important density changes are set by density regions with moderate values \(s\lesssim 2.5\). This criterion, i.e., that relevant binding or process-energy contributions to the XC energy differences should converge relatively fast with \(s\), holds for typical bulk and surface problems, where the use of the CX/CX0P/AHCX tool chain is suggested. It is a welcome bonus for CX and AHCX that the CX criterion (and CX accuracy) often seems to hold also for many molecular problems. However, there are also cases where we can document a large binding impact of density changes and where the interaction problem is not completely set by low-to-moderate \(s\) values. This happens for CO\({}_{2}\) uptake a diamine-functionalized MOF.
The fact that there is a condition on the CX/AHCX implementation of the Lindhard logic also suggests a potential susceptibility to density-driven errors. Such errors undercut overall arguments for a generic XC robustness of CX and AHCX, at least for the systems in systems where the sensitivity is identified. The translation of MBPT robustness into XC-design robustness is vulnerable because the Sham-Schluter equation also includes factors, i.e., independent-particle Green functions \(G_{0}\), that are set by KS energy levels and by the spatial variation in the KS orbitals that arise in the XC approximation. In order for an XC approximation to inherit the \(\Sigma_{\rm xc}({\bf r},{\bf r}^{\prime},\omega)\) robustness, it should delivers a near-exact density variation. Also, density-driven errors emerge, for example, by self-interaction error (SIE) effects in negatively-charged ions and they can, as such, reflect large density changes. The CX and AHCX rely on the ACF, but once the CX usage is pushed beyond the small-to-moderate-\(s\) criterion, the consistency benefit is gone. The inherent Lindhard-screening logic and current-conservation mechanism is then no longer able to enforce an automatic XC-hole conversation on the LV-exchange description. While the high-\(s\) form of CX exchange, i.e., the rPW86, also reflects a separately implemented (older) XC-hole conservation criterion, the strong MBPT connection is lost. In this type of large-\(s\) problems, we expect that the use of modern constraint-based PBE and PBEsol exchange would be the safer approach. In summary, we cannot expect that CX (and hence AHCX) will always remain a robust choice.
### The logic of the B86R variant
Improvement in accuracy generally followed from coupling the vdW-DF1 and vdW-DF2 correlation to other (less repulsive) exchange choices, for example, with the suggestions for OBK8, C09, OB86, B86R, and vdW-BEEF variants. The same is true for the CX release (and formal spin and hybrid extensions) that uses a Lindhard-screening logic to balance the XC components in typical binding cases. The balance question is also central for the DF3-opt1 and DF3-opt2 designs. Some of these vdW-DFs emphasize MBPT input on the gradient correction to exchange.
The introduction of variants has advantages for illustrating usefulness but complicates the search for systematic further progress. The variants (as well as CX) enhanced the range of applications that can easily be addressed with the vdW-DF method (beyond the reach of vdW-DF1 and vdW-DF2), as summarized in a number of reviews as well as perspectives. However, flexible variants might effectively be compensating for possible \(E_{\rm c}^{\rm nl}\) limitations since they fit the choice of \(\delta E_{\rm x}^{0}\) to a target or expected representative application. Having too much flexibility can diffuse the underlying drive for seeking increasingly more versatile XC functionals: We could inadvertently be hiding an actual method limitation.
Nevertheless, for our overall XC development goals we need to supplement CX and AHCX by a new RSH vdW-DF that is more robust towards density-driven errors. Unfortunately, simply creating a RSH vdW-DF right off of vdW-DF2 (from analysis included in Ref.) in a design termed DF2-AH, does not meet the need. This is made clear in the SI material with observations summarized in the discussion below.
Fortunately, the B86R variant of vdW-DF2 does offers a realistic path to craft a new general-purpose nonempirical RSH vdW-DF, the AHBR. Importantly, as explained below, the AHBR and B86R also offers a valuable contrast to AHCX and CX when it comes to prioritizing among possible MBPT inputs. That is, the combination of AHBR and AHCX gives us an option for a controlled 'functional-derivative' or 'functional-contrast' analysis: We can interpret and learn from observations of performance variations in terms of well-understood design differences. A similar idea of making a functional-difference analysis was also explored for adsorption studies in Refs.. Nicely enough, we discover that the resulting AHBR design also has a better-than-AHCX resilience towards the density-driven errors in molecular barrier-height problems as well as in some large-system isomerization problems, and Ref..
In practice, our AHBR development work starts from inspecting the B86R exchange and by providing an AH modeling of the B86R exchange hole, adapting Ref.. This work is an extension of the analysis presented for CX and for vdW-DF2 (and revisited for PBE and PBEsol)in Ref.. Like vdW-DF2 and CX, the B86R respects the lesson that the asymptotic form of the exchange-enhancement factor should rise as \(s^{2/5}\) asymptotically to appropriately counteract errors in the exchange contributions to (weak) binding. The B86R accomplishes that by relying on a revised Becke86b exchange.
\[F_{\rm x}^{\rm B86R}(s)=1+\frac{\mu_{\rm GEA}s^{2}}{(1+\mu_{\rm GEA}s^{2}/ \kappa)^{4/5}}\,, \tag{10}\]
This low-\(s\) form is aligned with the correct gradient-expansion result from a diagrammatic MBPT analysis, when interaction lines are interpreted as bare Coulomb interactions. We note that the use of \(\kappa=0.7114<1\) in Eq. implies a smaller prefactor on the asymptote, \(F_{\rm X}(s)\sim s^{2/5}\), than what applies for CX and OB86.
Since B86R relies on vdW-DF2 correlation, it has only a weaker consistency in that both exchange and correlation energies are set by MBPT inputs that are valid for molecular-type problems, above. The exchange enhancement of the internal functional is set as \(F_{\rm x}(s)=1+0.2097s^{2}\), while the expansion of Eq. is given by \(\mu_{\rm GEA}=10/81\). This means that B86R does not have full alignment of the inner-functional and the actual exchange-energy terms, something that CX maintains up to \(s=2-3\) by systematically relying on diagrammatic MBPT (assuming screened interaction lines). However, we do find that the use of \(\mu_{\rm GEA}=10/81\) and \(\kappa<1\) in Eq. brings the B86R exchange enhancement values, \(F_{\rm X}(2<s<10)\), closer to PBEsol exchange, without giving up the asymptotic \(F_{\rm X}(s)\sim s^{2/5}\) behavior that is also necessary. We observe that PBEsol reflects a more modern approach to set exchange by enforcing XC hole conservation than the rPW86 (that is, large-\(s\)) part of CX.
In summary, switching between CX-AHCX and B86R-AHBR means using different assumptions when setting the exchange impact on both the plasmon modeling and on the actual XC balance. However, the switching is still done while staying within the same framework of MBPT analysis. There are arguments for and against CX/AHCX and B86R/AHBR (just as there are for PBE and PBEsol in the GGA framework). We shall here employ broad testing to assert which priority brings the greater benefits within the present range of vdW-DF design ideas.
Since AHBR is intended as a key part of our performance-contrast strategy, we must also secure and validate a general-purpose capability in this new nonempirical RSH vdW-DF. Here we benefit from past investments: The logic of the unscreened-hybrid "vdW-DF+0" class (that generalizes PBE0) leads, for example, to the formulation of the B86R-based DF2-BR0. We included a code option for this in the Quantum Espresso (QE) code suite, while releasing the CX-based hybrid CX0 and CX0P, and we now report a full GMTKN55 assessment for DF2-BR0; See for example Figs. S.1-2 and Tables S.2-X of the SI material. We discover that there are very few outliers in the DF2-BR0 benchmark results and that the performance gain of DF2-BR0 is particularly strong for the important transition-state problems. The fact that DF2-BR0 delivers well-balanced progress across general types of molecular problems is an additional strong motivation to here complete and launch the screened hybrid AHBR.
### Analytical-hole design of AHBR
The bottom panel of compares the key exponent form \(\mathcal{H}(s)\) that defines the long-separation shape of the exchange hole that is assumed to be of a modified Gaussian type.
\[n_{\rm x}({\bf r};{\bf r}^{\prime})\propto\exp\left[-\mathcal{H}(s({\bf r}))( sy)^{2}\right]\,, \tag{11}\]
The ideas of the analytical exchange-hole modeling, as well as the logic and details of Eq., are presented and discussed in Refs..
The details of this Gaussian-suppression factor \(\mathcal{H}(s)\) must be asserted to complete this AH model of a given XC functional; Technical details for vdW-DFs are discussed in Ref.. The \(\mathcal{H}(s)\) variation is given by a rational function with parameters fitted subject to constraints so as to accurately describe, for example, the B86R exchange behavior without introducing any spurious variation (that cannot be ascribed any physical meaning). The procedure for setting the parameters of \(\mathcal{H}\), used previously to discuss and understand the PBE, PBEsol, CX and AHCX exchange, is here repeated for B86R. Table XI of the SI material reports \(\mathcal{H}(s)\) parametrizations that underpin the now extended range of RSHs, including AHBR (that is based on understanding B86R exchange).
The panels show spatial variations of the exchange holes that result in the AH exchange modeling at a set of increasing values for the scaled density gradient \(s\).
\[n_{\rm x}({\bf r},{\bf r}^{\prime})=n({\bf r})\times J(s({\bf r});{\bf r}^{ \prime}-{\bf r})\,. \tag{12}\]
In the GGA-exchange model framework that we work with, the density suppression (by exchange effect) induced at position \({\bf r}^{\prime}\) by an electron at \({\bf r}\), can be completely expressed in terms of a locally scaled distance \(y=k_{\rm F}({\bf r})|{\bf r}^{\prime}-{\bf r}|\) defined by the Fermi wavevector \(k_{\rm F}\),. As indicated in Eq., the shape of depends on the local value of the scaled density gradient \(s(\mathbf{r})\). However, the entire \(J(s,y)\) variation is set by finding parameterizations of the Gaussian-damping functions \(\mathcal{H}(s)\), functions, discussed above and plotted in the lower panels of Thus, by tracking the shape of \(J(s,y)\), panels of Fig. 3, we summarize the full detail of the exchange modeling. The actual exchange-hole modeling (for PBE, PBEsol, and B86R, respectively) at any given position \(\mathbf{r}\) is revealed by simply inserting the relevant local values for \(s(\mathbf{r})\) and \(k_{\mathrm{F}}(\mathbf{r})\).
The right panel of represents our new AH modeling for exchange effects in B86R; It is fully summarized in the \(J^{\mathrm{B86R}}\) variation that is, in turn, sufficient to both recoup the B86R-exchange term \(E_{\mathrm{x}}^{\mathrm{B86R}}(s)\) and set AHBR, below, adapting Refs. and. We find that, initially (at small \(s\) values,) the B86R exchange hole follows the PBEsol-exchange nature but does gradually roll over to a more PBE-exchange type behavior; It also eventually approaches a CX-like behavior at large \(s\) values where it respects the lessons of the analysis in Ref.. Interestingly, the B86R exchange does, for \(s\lesssim 3\), perform better than (CX and) PBE exchange in terms of avoiding the formation of deep exchange holes: The modelling of the B86R exchange hole means that the suppression remains smaller than the local value of the electron density (at \(s\lesssim 3\)).
We compare this AH analysis for the B86R exchange hole variation also with that for CX and AHCX, using the lower panel of and of Ref.. First, it is clear that setting the exchange enhancement by \(\mu_{\mathrm{GEA}}\) brings the B86R closer to PBEsol exchange than the CX exchange design. Since the B86R exchange is still constrained by the input from Ref., the large-\(s\) behavior rolls over towards that of CX. Moreover, being an intermediate of the PBEsol and of the CX exchange-hole modeling (of Ref.,) the B86R has an mid-\(s\)-range behavior (around \(s\approx 3-4\)) that is close to the PBE hole form. This is again a more trusted behavior than that for rPW86 (that enters in CX).
From the AH analysis of B86R, Figs. 2 and 3, we complete the RSH nonlocal-correlation functional AHBR, following the same steps as previously described for AHCX. The key observation is that our knowledge of \(J^{\mathrm{B86R}}\) variation allow us project out the short-range (SR) exchange-energy component \(E_{\mathrm{x}}^{\mathrm{B86R,SR}}[n;\gamma]\) from the exchange term \(E_{\mathrm{x}}^{\mathrm{B86R}}\) of B86R. The projection \(E_{\mathrm{x}}^{\mathrm{B86R,SR}}[n;\gamma]\) is again a density functional.
\[E_{\mathrm{FX}}^{\mathrm{SR}}(\gamma)=\frac{1}{2}\int_{\mathbf{r}}\int_{ \mathbf{r^{\prime}}}\frac{n(\mathbf{r})n_{\mathrm{x}}^{\mathrm{Fo}}(\mathbf{r^ {\prime}};\mathbf{r})}{|\mathbf{r}-\mathbf{r^{\prime}}|}\,\mathrm{erfc}(\gamma |\mathbf{r^{\prime}}-\mathbf{r}|)\,, \tag{13}\]
The overall RSH vdW-DF form is
\[E_{\mathrm{xc}}^{\mathrm{AHBR}}[n]=E_{\mathrm{xc}}^{\mathrm{B86R}}[n]+\alpha(E _{\mathrm{FX}}^{\mathrm{SR}}(\gamma)-E_{\mathrm{x}}^{\mathrm{B86R,SR}}[n;\gamma ])\,, \tag{14}\]
A robust and computationally efficient determination of \(E_{\mathrm{x}}^{\mathrm{SR,B86R}}[n;\gamma]\) is a key benefit of working with the AH modeling. To complete the RSH vdW-DF construction, Eq., we need \(E_{\mathrm{x}}^{\mathrm{SR,B86R}}[n;\gamma]\). It is given in analogy to Eq. but set by a modified exchange-enhancement factor \(F_{\mathrm{x}}^{\mathrm{SR}}(k_{\mathrm{F}}(\mathbf{r}),s(\mathbf{r}))\).
\[F_{\mathrm{x}}^{\mathrm{B86R,SR}}(k_{\mathrm{F}},s)=-\frac{8}{9}\int_{0}^{ \infty}yJ^{\mathrm{B86R}}(s,y)\mathrm{erfc}(\gamma y/k_{\mathrm{F}})dy\,. \tag{15}\]
In fact, we get all exchange details of the new AHBR from the corresponding AHCX details, Ref., by simply switching from the CX- to the B86R-specific parametrization of the AH modeling, SI Table S.XI.
Both RSH vdW-DFs, the new AHBR and the AHCX, can be used when screening of the Fock-exchange is essential, for example, for descriptions of adsorption at metal and high-dielectric-constant surfaces.
The AHBR is deliberately kept free of fitted parameters. The extent and nature of the screened-Fock-exchange inclusion could be adjusted but should then be given by physic inputs. Implementation of a formal thermodynamics criterion, in effect that the functional exhibits a piecewise-linearity with the addition of a fractional electron, can set the value of the inverse screening length \(\gamma\). Similarly, the extent of Fock-exchange mixing \(\alpha\) can be set by a coupling-constant analysis or by demanding that the resulting dielectric constant is consistent with that implied in the Coulomb range separation. In this AHBR launching work, we will only illustrate and contrast generic RSH vdW-DF usage (that is, at fixed, HSE-standard, values for \(\alpha\) and \(\gamma\)).
## III Results and discussion
The DFT-result data is obtained using our in-house version of the QE code suite. This code also has generic subroutines for calling RSH vdW-DFs (already released to QE-7.0) and it benefits from the adaptively compressed exchange evaluation of Fock exchange. Appendices A and B provide an overview of the computational details that we use in general demonstrations and for the GMTKN55 assessment work, partly summarized in already.
The CX and hence AHCX emphasis on screened LV exchange means that they are naturally set up for accuracy in metal systems and, we expect, broadly for bulk and many surface problems, including adsorption. The new RSH vdW-DF has an advantage in being a general-purpose choice for molecular properties, Here we assert and discuss whether the new AHBR will remain an option also for bulk and adsorption, and whether it also works in a biochemistry and green-technology context. Our full-GMTKN55 assessment, see SI material, is part the AHBR documentation and we extract a number of observations also from that mapping.
We note that the move to hybrids, including the RSH vdW-DFs, can help in counteracting excessive charge transfer and hence some density driven errors. This is true in the raw, fixed-parameters form presented above and because the AHCX/AHBR come with an option for \(\gamma\) tuning so as to also impose the thermodynamics (fractional-electron) constraint on the designs. We expect that such tuned-AHBR usage will help further on controlling (density-driven) errors. However, we have not used that potential for gaining additional XC-functional consistency in this first assessment.
### Bulk-structure performance
More broadly, contrasts the performance for bulk of a new, second tool chain (comprising AHBR-B86R) with that of the first (AHCX/CX) (and of DF2-AH). We do not report data for the unscreened hybrid components (DF2-BR0/CX0/CX0P) as we also consider metals.
The violin plots summarize deviations in percentage of computed results for lattice constants \(a\), cohesive energies \(E_{\rm coh}\), and bulk moduli \(B_{0}\) from back-corrected experimental values for five transition metals (Cu, Ag, Au, Pt, Rh), one simple metal (AI), four semiconductors (Si, C, SiC, GaAs) and three ionic insulators (LiF, MgO, NaCl) as in Ref.. Tables S XIII, S XIV, and S XV of the SI present a more complete quantitative presentation, contrasting values computed in the two tool chains with experimental values (that are back corrected for vibrational effects) and with those we obtain for the RSH DF2-AH. The subscript on one AHBR-data (and on one AHCX-data) label identifies the extent of Fock-exchange mixing; A corresponding specification is suppressed for 'AHBR=AHBR\({}_{0.25}\)' ('AHCX=AHCX\({}_{0.20}\)') since this mixing reflects a recommended default, as explained in the following subsection.
To contrast functional performance on bulk, we compare the position of the mean (median) deviation, shown by a central bar (diamond) and the so-called interquartile range, shown as a bar. This bar reflects the difference of positions for the first and third quartile of the performance distribution (for each functional). We also consider the presence or absence of outliers (open circles), by which we mean a performance that lies beyond markers (wiskers) that identify 1.5 times the interquartile range. Lattice-constant outliers are Au (Au and Ag) for AHBR\({}_{0.20}\) (AHBR) while cohesive-energy outliers are Au and Rh for B86R. There are more outliers for the bulk modulus: Rh for AHCX/AHCX\({}_{0.25}\), GaAs for B86R, and Au for AHBR and AHBR\({}_{0.20}\).
However, the B86R is more accurate than AHBR for predictions of the bulk cohesive energies.
This is consistent with findings that CX has a small bulk-performance edge over B86R. This observation holds, for example, for the lattice constant. However, AHBR is accurate on structure and hence useful also for substrate descriptions in heterogeneous system. For example, the lattice-constant accuracy on noble-metal and Pt metals remains within 0.5% deviations relative to (back
\begin{table}
\begin{tabular}{c|c c c|c c c|c} & CX & AHCX & AHCX\({}_{0.25}\) & B86R & AHBR\({}_{0.20}\) & AHBR & Exper.\({}^{*}\) \\ \hline Cu \(a\) & 3.576 & 3.587 & 3.592 & 3.602 & 3.613 & 3.617 & 3.599 \\ \(E_{\rm coh}\) & 3.781 & 3.348 & 3.264 & 3.582 & 3.160 & 3.064 & 3.513 \\ \(B_{0}\) & 163 & 148 & 146 & 151 & 141 & 136 & 144 \\ \hline Ag \(a\) & 4.065 & 4.078 & 4.082 & 4.104 & 4.115 & 4.118 & 4.070 \\ \(E_{\rm coh}\) & 2.955 & 2.774 & 2.737 & 2.779 & 2.592 & 2.549 & 2.964 \\ \(B_{0}\) & 115 & 105 & 104 & 102 & 95 & 95 & 106 \\ \hline Au \(a\) & 4.101 & 4.098 & 4.097 & 4.134 & 4.127 & 4.126 & 4.067 \\ \(E_{\rm coh}\) & 3.634 & 3.440 & 3.398 & 3.402 & 3.205 & 3.158 & 3.835 \\ \(B_{0}\) & 171 & 168 & 167 & 153 & 152 & 151 & 182 \\ \hline Pt \(a\) & 3.929 & 3.910 & 3.906 & 3.952 & 3.929 & 3.925 & 3.917 \\ \(E_{\rm coh}\) & 6.226 & 5.524 & 5.259 & 5.999 & 5.131 & 4.930 & 5.866 \\ \(B_{0}\) & 284 & 298 & 298 & 264 & 278 & 279 & 286 \\ \end{tabular}
\end{table}
Table 1: Comparison of vdW-DF performance on Pt- and noble-metal structure and elastic response: Lattice constants \(a\) (in Å), cohesive energies \(E_{\rm coh}\) (in eV), and bulk moduli \(B_{0}\) (in GPa). The SI material gives full listings for our computed \(a\), \(E_{\rm coh}\), and \(B_{0}\) results for these transition metals and 9 other materials. Experimental values, back-corrected for vibrational effects (as indicated by an asterisk), are taken from Ref..
As an interesting aside, we document that RSH DF2-AH (implicitly defined in Ref.) is not well suited for bulk-system use. This is not surprizing since there are real issues with using vdW-DF2 for bulk systems. Tables 5 33, 54, and 54 of the SI document that there are large deviations between DF2-AH (and vdW-DF2) predictions and back-corrected experimental values for all of the investigated bulk properties. In the violin-plot we do not even depict the full extent of the interquantile range which for DF2-AH is set by first/third-quartile relative deviations (0.78% and 3.5% for \(a\), 11% and 29% for \(E_{\text{coh}}\) and -3% and -34% for \(B_{0}\)). Unlike AHBR, the DF2-AH is simply not a reliable option for bulk and hence for substrates in adsorption studies.
We ascribe the absence of vdW-DF2 and DF2-AH robustness (for descriptions across general molecular properties and bulk) to the fact that their designs depart from some of the ideas that emerged in the electron-gas tradition. Use of diagrammatic-MBPT input, as summarized in Refs., are prioritized in the design of PBEsol (unscreened) exchange and in CX, with its emphasis on LV screened exchange. However, the design of (r)PW86, that defines vdW-DF2 and DF2-AH exchange, did not prioritize this MBPT input to the same extent.
Finally, shows that the AHBR and AHCX bulk-structure characterizations are relatively insensitive to the choice of the Fock-exchange mixing. The AHCX design is set with a default 0.20 Fock exchange mixing but the overall bulk performance does improve slightly by going from AHCX=AHCX\({}_{0.20}\) to AHCX\({}_{0.25}\).
\begin{table}
\begin{tabular}{l l|c|c c c|c c c c|c c} & Stacking & Benchmark & vdW-DF & CX & AHCX & vdW-DF2 & B86R & AHBR\({}_{0.20}\) & AHBR & RPA & DMC \\ \hline Graphite & AB-crystal & \(E_{\text{bind}}\) & 55 & 67 & 72 & 54 & 62 & 63 & 64 & \(48^{a}/62^{b}\) & \(60\pm 5^{c}\) \\ & & \(d_{\text{opt}}\) & 3.57 & 3.26 & 3.27 & 3.51 & 3.30 & 3.31 & 3.34\({}^{a}/3.34^{b}\) & \(3.43\pm 0.04^{d}\) \\ Graphite & AA-crystal & \(E_{\text{bind}}\) & 50 & 54 & 52 & 47 & 49 & 44 & 43 & - & \(36\pm 1^{e}\) \\ & & \(d_{\text{opt}}\) & 3.73 & 3.55 & 3.55 & 3.67 & 3.54 & 3.56 & 3.57 & - & \(3.63^{e}\) \\ \hline Graphene & AB-bilayer & \(E_{\text{bind}}\) & 25 & 30 & 33 & 25 & 28 & - & 29 & \(46^{f}\) & \(18\pm 1^{g}\) \\ & & \(d_{\text{opt}}\) & 3.61 & 3.29 & 3.29 & 3.53 & 3.33 & - & 3.33 & \(3.39^{f}\) & \(3.384^{g}\) \\ Graphene & AA-bilayer & \(E_{\text{bind}}\) & 22 & 24 & 26 & 21 & 22 & - & 23 & - & \(12\pm 1^{g}\) \\ & & \(d_{\text{opt}}\) & 3.76 & 3.58 & 3.57 & 3.69 & 3.57 & - & 3.58 & - & \(3.495^{g}\) \\ \hline \(\alpha\)-Graphyne & AB-bilayer & \(E_{\text{bind}}\) & 20 & 20 & 23 & 18 & 17 & - & 18 & - & \(23^{h}\) \\ & & \(d_{\text{opt}}\) & 3.47 & 3.30 & 3.26 & 3.36 & 3.26 & - & 3.25 & - & \(3.24^{h}\) \\ \(\alpha\)-Graphyne & Ab-bilayer & \(E_{\text{bind}}\) & 19 & 19 & 21 & 18 & 16 & - & 17 & - & \(22^{h}\) \\ & & \(d_{\text{opt}}\) & 3.65 & 3.49 & 3.43 & 3.52 & 3.42 & - & 3.41 & - & \(3.43^{h}\) \\ \hline hBN & AA’-bilayer & \(E_{\text{bind}}\) & 24 & 29 & 32 & 24 & 26 & - & 28 & \(19^{f}\) & \(20^{i}\) \\ & & \(d_{\text{opt}}\) & 3.58 & 3.26 & 3.24 & 3.51 & 3.31 & - & 3.28 & \(3.34^{f}\) & \(3.25/3.50^{i}\) \\ \hline Phosphorus & AB-crystal & \(E_{\text{bind}}\) & 79 & 127 & 124 & 83 & 117 & - & 112 & - & \(81\pm 6^{j}\) \\ & & \(d_{\text{opt}}\) & 5.69 & 5.19 & 5.26 & 5.67 & 5.27 & - & 5.33 & - & \(5.2^{j}\) \\ \hline \end{tabular} \({}^{a}\) Ref..
\({}^{b}\) Ref..
\({}^{c}\) Ref.; We report the raw DMC \(E_{\text{bind}}\) value (omitting an estimate for vibrational effects) for a relevant comparison.
\({}^{d}\) Ref.; The authors warn that (in-plane-size) convergence at large layer separations is not sufficient to accurately fit \(d_{\text{opt}}\).
\({}^{e}\) Ref..
\({}^{f}\) Ref..
\({}^{g}\) Ref..
\({}^{h}\) Ref..
\({}^{i}\) Ref.; \(E_{\text{bind}}\) value without correction for infinite-layer extension is presented in paranthesis. No fit for \(d_{\text{opt}}\) given.
\({}^{j}\) Ref.; The \(d_{\text{opt}}\) is extracted from of that reference.
\end{table}
Table 2: Comparison of vdW-DF performance on layered materials: Layer-binding \(E_{\text{bind}}\) (in meV/atom) and optimal layer separation \(d_{\text{opt}}\) (in Å) for graphite (sp\({}^{2}\)-hybridized carbon) as well as graphene and \(\alpha\)-Graphyne bilayers (with sp-sp\({}^{2}\)-hybridized carbon), hexagonal boron nitride (hBN) and phosphorus; See DMC references for summaries of in-plane atomic configurations. Stacking labels AA and AB (Ab) identify geometries with carbon layers in full alignment and displaced one-third (one-ninth) of the sum of in-plane lattice constants, respectively. Stacking label AA’ (for hBN) identifies in-plane alignment of boron atoms in one layer with nitrogen atoms in the other layer.
### Binding in layered matter
We assess the RSH vdW-DF performance on layered systems using diffusion Monte Carlo (DMC) results as reference data. We also discuss random-phase approximation (RPA) literature results and recent measurements of the cleavage energy in graphite. The graphite DMC study came with both a raw DMC result and with an estimate for the expected correction (4 meV/atom) for vibrational effects in graphite. Our computed results (as well as the RPA literature results) are obtained in the Born-Oppenheimer approximation and should be compared with the raw DMC values listed in Table 2.
For our RSH and nonhybrid RSH characterizations, we used a 2-step approach, seeking to compare all functionals at a fixed in-plane structure that is close to experiments. First, we determine the in-plane lattice constants using a set of CX calculations (as CX is strong on structure). This was generally done by variable-cell calculations. For the phosphorus crystal, the in-plane \(a_{1}\) lattice constant is soft (but the other in-plane lattice constant \(a_{2}\) is still set by covalent bonds); There we kept \(a_{1}\) fixed at the experimental (bulk) value (4.374 A) while we used CX calculations to determine \(a_{2}\). Further computational details are described in Appendix A.
Second, at fixed in-plane lattice vectors, we compute the total unit-cell energy \(E(d)\) as a function of layer separation \(d\). For the RSH and regular vdW-DFs, we thus determine the optimal separation value \(d_{\rm opt}\) and the asymptotic system value \(E_{\rm asymp}\) (taken as the system energy at \(d=20\) A).
\[E_{\rm bind}=\left[E(d_{\rm opt})-E_{\rm asymp}\right]/N\,, \tag{16}\]
We note that the typical DMC binding-energy definition, Eq., differs from the binding-energy definition (meV-per-layer-atom) that is used for the discussion of early bilayer vdW-DF studies. This has led to confusion in vdW-DF presentations, for example, in Ref.. We include new vdW-DF1 and vdW-DF2 characterization, now using the definition Eq. for comparisons.
Table 2 reports our comparison of CX/AHCX and B86R/AHBR performance relative to the DMC reference data. In graphite, the CX and AHCX functionals give \(E_{\rm bind}\) results that are larger than the DMC reference value, even allowing for error bars reported in Ref.. However, the B86R and AHBR results for \(E_{\rm bind}\), at 62-64 meV/atom, are in excellent agreement with the relevant (that is, the raw) DMC reference value, \(60\pm 5\) meV/atom for the regular (AB-stacked) graphite crystal.
We observe that the B86R and AHBR results for the optimal layer separation, at \(d_{\rm opt}=3.30-3.31\) A, are smaller that the reported DMC value, \(3.43\pm 0.04\) A and more in line with the experimental characterization, at \(d_{\rm opt}=3.35\) A. The graphite-AB DMC study has a lower convergence (with regards to in-plane extension) results for large layer separation and the authors warn that it impacts the \(d_{\rm opt}\) fit.
At the same time, there are good reasons to trust the graphite-AB DMC result of Ref. for \(E_{\rm bind}\). The trust comes from recent high-precision measurements of the graphite cleavage energy, giving binding-energy results of 54 and 55 meV/atom. Including the estimate of a 4 meV/atom vibrational correction, there is full alignment with the graphite-AB DMC \(E_{\rm bind}=60\pm 5\) meV/atom result. In turn, given the trust in the DMC result, we conclude that AHBR is highly accurate for graphite-AB binding, Table 2.
In fact, for the regular graphite crystal, CX, B86R, and AHBR provide an \(E_{\rm bind}\) description that is closer to DMC than one of the RPA results (at \(E_{\rm bind}^{\rm RPA}=48\) mev/atom).
Bulk-system performance as asserted in percentage deviations for the CX-AHCX and B86R-AHBR descriptions of lattice constants \(a\), bulk cohesive energies \(E_{\rm coh}\), and bulk moduli \(B_{0}\). We also show the impact of setting the Fock-exchange mixing; The default AHCX and AHBR mixing value is set at 0.20 and 0.25, respectively, while a subscript identifies an adjustment. For completeness, we furthermore include a performance assessment for DF2-AH, the RSH form of vdW-DF2 that is implicitly defined by analysis in Ref.. We compare our results (listed in the SI), computed in the Born-Oppenheimer approximation, with experimental values (back-corrected for vibrational effects). The violin plots summarize result-statistics data for 13 solids (5 transition metals, 1 simple metal, 4 semiconductors, and 3 ionic insulators), with the CX and AHCX results repeated from Ref.. The set of horizontal bars (of diamonds) reflects the mean (median) deviation in the distributions while the boxes identify the so-called interquartile range, see text. Outliers (identified in the text) are shown by open circles.
However, the authors of that second-RPA study warn that there can be an impact of RPA convergence.
Considering next the meta-stable graphite configuration with an AA-stacking, we find that CX/AHCX and B86R all make larger errors. However, we also find that AHBR, through its inclusion of a screened Fock-exchange component, is able to correct much of the B86R overestimations. The AHBR result lands at about a 7-8 meV/atom deviation from the DMC reference data. The AA configuration may be seen as the barrier of an in-plane slip process of graphite (generalizing a picture that applies for polymers). We therefore interpret the reasonable robustness of the AHBR characterization for graphite-AA as an another example of success at characterizing transition states.
We also assess the performance for bilayer graphene, in AB and AA stackings, and for the closely related hexagonal boron nitride (hBN), in AA' stacking, against DMC studies. These studies are again obtained for systems with a dense in-plane electron distribution, however, there is significantly less binding contribution in the unit-cells. Part of the reason for that is evident by inspecting of Ref., noting that the dominant contribution in layered materials arises from the moderately-low electron variation that exists between the layers: In a bilayer form there is only half as many such regions as in the corresponding bulk. The layer binding is further reduced because bilayer systems lack the coupling to layers that are further away, for example, as discussed in Refs..
For the graphene bilayer system, in both AB and AA stackings, we find that AHBR differs more (by about 11 meV/atom) from the DMC results. The B86R and AHBR do perform better than CX and AHCX and are still significantly closer than the one RPA description that we have found.
The AHBR description also differs from the DMC result for a hBN bilayer system; Here it is instead a RPA result that is close to the DMC value. Our AHBR result differs 8 meV/atom from the DMC reference, in line with the status for graphite in AA stacking.
We note in passing that vdW-DF and vdW-DF2 are often closer to the DMC values for binding energies than B86R and AHBR (or CX and AHCX). However, Table 2 makes it clear that vdW-DF and vdW-DF2 systematically overestimate the layer-binding separation.
We also consider the AHCX and AHBR performance at \(\alpha\)-graphyne bilayers, cases in which some of the carbon atoms in each plane are in the sp-hybridized form. Here, interestingly we find that AHCX slightly outperforms the AHBR description on both structure and binding energies. However, the performance of AHBR, as a new general-purpose RSH vdW-DF, is still good, landing within 5 meV/atom of the DMC reference description. We also note that both AHCX and AHBR correctly predict the 1 meV/atom preference that separates the two competing motifs (AB and Ab stacking) for these weakly bonded systems.
Finally, we assert the AHBR and AHCX performance for phosphorus bulk, in the stable AB stacking. The DMC reference data has \(E_{\rm bind}\) and we can extract an estimated value also for \(d_{\rm opt}\). We find, again, that vdW-DF2 and vdW-DF are closer than AHBR and AHCX for the layer binding energy but not for structure. The AHBR (AHCX) description is 2% or an estimated 0.13 A (1% or 0.06 A) too large on \(d_{\rm opt}\) while it overestimates \(E_{\rm bind}\) by 31 meV/atom (43 meV/atom) relative to the DMC result, \(81\pm 6\) meV/atom.
### Robust molecular benchmarking and setting the Fock-exchange mixing in AHBR
Our use of the planewave-code QE gives us the prerequisites for delivering a high-quality (in principle, complete-basis-set) assessment.
However, there are dramatic SIE effects or density-driven errors, and hence challenges with plane-wave benchmarking, in the study of negatively charged atoms and radicals. The last electron will not necessarily remain bonded, unless we work with small unit cells that artificially raise the vacuum floor in QE. Because we seek to approach the complete-basis set limit, we cannot provide a meaningful direct assessment of performance for the G21EA and WATER27 benchmark sets of the GMTKN55 suite. Nevertheless, in appendix B, we document that use of a dielectric-environment extension permits us to circumvent the SIE challenges and reliably complete general functional assessments.
Details of this GMTKN55-based assessment are given in the SI material Tables S.II through S.X; It covers almost all of vdW-DFs that are coded in the QE version 6.7, the related revised VV10, as well as the dispersion-corrected revPBE+D3 and HSE+D3. For comparison, we also list literature MAD results for revPBE+D3, HSE+D3, SCAN+D3, and B3LYP+D3, as obtained in orbital-based DFT.
The right panel focuses on the seven sets in the important GMTKN55 group 3 of molecular-barrier benchmarks, reporting mean absolute deviations (MADs) relative to coupled-cluster quantum chemistry calculations (in kcal/mol). The left panel presents the broader GMTKN55 performance overview (characterized by weighted MAD values, i.e., the WTMAD1 measure introduced in Ref.) and thus also covers the GMTKN55 group 1 and group 2 assessments for small- and large-molecules reactions and transformations, as well as groups 6, 4, and 5 of benchmarks covering total, inter- and intra-molecular NOC interactions.
This class of DFT problems affects barrier-height problems that in turn define the GMTKN55 benchmark group 3. Accuracy for transition states, and hence for predicting reaction rates, is considered a challenge even when the issue is considered in isolation. It is exciting that AHBR provides a balanced progress, i.e., it works just as well (maybe even better) for transition-state problems as for molecular-reaction energies.
For the benchmarking summarized in Fig. 1, and generally throughout the paper, we have deliberately kept the RSH parameters fixed at the default HSE-specification for the screening-length parameter, 0.106 inverse bohr, assuming also a fixed 0.25 (0.20) fraction for the mixing of short-range Fock exchange for AHBR (AHCX). These defaults are used throughout the paper, although we also sometimes illustrate the impact of switching between the two (standard, 0.20 and 0.25) choices for the extent of Fock-exchange mixing, as marked by subscripts (for example, AHBR\({}_{0.20}\)).
In fact, we have set this default recommendation for the AHBR Fock-exchange mixing, AHBR=AHBR\({}_{0.25}\), by directly relying on the broad GMTKN55 molecular benchmarking for this second generation RSH vdW-DF. For individual problems and benchmarks we could proceed to make a system-specific analysis to establish a plausible choice of the Fock-exchange mixing and screening, for example, as pursued in Refs.. Here we observe that use of AHBR\({}_{0.25}\) is systematically more accurate than AHBR\({}_{0.20}\) on molecular properties; We pick \(\alpha=0.25\) to give an impression of the accuracy that we can hope to get from AHBR usage.
We also note in passing that the suggested default AHCX 0.20 mixing came from a coupling-constant analysis of the contribution of CX correlation to the atomization energies. The logic of that specification needs not hold for CX0P and AHCX when it comes to large systems (or bulk), let alone for AHBR. Looking at the full survey, in the SI material, we find that moving the AHCX to 0.25 Fock-exchange mixing gives a small performance gain both overall and for all but the NOC-interaction benchmark groups 4 and 5. The impact is in any case limited.
### Navigating density-driven errors
It is natural to discuss the progress of AHBR as a molecule performer in terms of the resilience towards density-driven errors. The promise of AHBR success on these key challenges is implied in and here we provide details.
Any given XC approximation will cause an incorrect XC-energy, and thus DFT-total-energy evaluation, even if we had access to the exact density. However, there are additional challenges because the DFT calculations (based on the specific XC functional approximation) can sometimes lead to an electron density solution that is far from the exact density. This extra sensitivity causes performance outliers with a dramatically reduced accuracy of the system-specific DFT study. Important examples are the generic-DFT failure to correctly confine the last electrons in some negatively charged ions, charge trapping in oxide defects and color centers, molecular-reaction barrier heights, the CO-adsorption site-preference challenge on Pt, and, we expect, adsorption-induced dissociation in the presence of large charge transfer.
The BH76 molecular barrier-height benchmark has already been discussed as a key challenge for securing robustness, i.e., a driver for us to complete the design of the AHBR. It is also clear that the density-driven errors directly impact negatively charged ions and radicals, and hence the performance on, for example, the G21EA and the WATER27 sets in groups 2 and 4, see also Appendix B. Ref. additionally highlights the G2RC and RSE43 benchmark sets as being prone to density-driven errors.
For molecules it is possible to use calculations of the Hartree-Fock (HF) electron-density solution to spot when we can expect density-driven errors. For such cases, one would, in general, expect that moving to a hybrid is motivated, but there are also molecular cases where use of the unscreened hybrid PBE0 cannot by itself correct the issue.
The HF-based tests for density sensitivity are carried out using non-selfconsistent DFT-energy calculations. Our ppACF code contribution makes this option available in the QE code suite, given some manual adjustments of the xml file that the regular DFT solver provides. For a given problem, we first pursue selfconsistent HF calculations (for all reactants and products) to obtain the density variations, denoted \(n^{\rm HF}(\mathbf{r})\), in that approximation. We also compute density variations, denoted \(n^{\rm LDA}(\mathbf{r})\), using self-consistent LDA. Next, we obtain so-called post-PBE energies (for all reactants and products), denoted \(E^{\rm HF-PBE}\) and \(E^{\rm LDA-PBE}\), by evaluating the total DFT-PBE energies on the set of fixed \(n^{\rm HF}(\mathbf{r})\) and \(n^{\rm LDA}(\mathbf{r})\) density variations, respectively.
\[\tilde{S}_{\rm proc}=|\Delta E^{\rm HF-PBE}_{\rm proc}-\Delta E^{\rm LDA-PBE}_{ \rm proc}|\,. \tag{17}\]
This measure reveals whether there are fundamental differences between the HF and KS orbitals and hence whether we can expect density errors to significantly affect any given (GGA or vdW-DF) DFT characterization. For a benchmark set one can also define an average value \(\tilde{S}_{\rm avg}\) over case-specific \(\tilde{S}_{\rm proc}\) values, essentially adapting the ideas of benchmark MAD assessments. If \(\tilde{S}_{\rm avg}\) (\(\tilde{S}_{\rm proc}\)) is asserted as larger than 2 kcal/mol, then the benchmark set (process) should be considered as density sensitive.
Using this assessment procedure, we presently add documentation that the PX13 (barrier-height) and the C60ISO (large-system isomerization) sets are also prone to density-driven errors, i.e., given by characteristic sensitive measures \(\tilde{S}_{\rm avg}>2\) kcal/mol. The present \(\tilde{S}_{\rm avg}\)-based mapping supplements the literature identification of pronounced sensitivities for BH76, G2RC, and RSE43. In the literature cases the measures are comparable even to the PBE MAD values themselves; For C60ISO (and to some extent also for PX13), we find that the measure represents a significant fraction of the PBE MAD value.
Table 3 contrasts the PBE, CX, vdW-DF2, HSE+D3, AHCX, and AHBR performances, asserted as MAD values, on those density-sensitive benchmarks. The table shows that the AHBR here performs at the same level or better as HSE-D3, and systematically better than AHCX. This latter finding is expected because we have given arguments (in Section II) that the condition that exists for CX (to fully leverage its Lindhard screening foundation, discussed in Section II) might make the CX more susceptible to density-driven errors. The first observation is important for encouraging broad vdW-DF
\begin{table}
\begin{tabular}{l c c c c c c c c c} & Sensitive\({}^{a}\) & PBE0+D4\({}^{a}\) & revPBE+D3 & CX & vdW-DF2 & B86R & HSE+D3 & AHCX & AHBR \\ BH76 & yes & unclear & 7.38 & 9.15 & 6.90 & 9.22 & 4.21 & 5.15 & 4.14 \\ PX13 & yes & yes & 9.29 & 12.80 & 1.14 & 11.36 & 7.38 & 7.30 & 4.57 \\ BHPERI & no & yes & 5.74 & 7.20 & 3.08 & 6.08 & 2.83 & 4.05 & 1.84 \\ W4-11 & yes & yes & 5.88 & 8.55 & 18.69 & 6.97 & 6.77 & 4.99 & 9.50 \\ DC13 & yes & unclear & 9.38 & 7.88 & 24.21 & 7.26 & 8.24 & 8.35 & 6.18 \\ SIE4x4 & yes & yes & 21.67 & 23.80 & 21.73 & 23.52 & 13.58 & 17.00 & 15.09 \\ ISOL24 & yes & yes & 4.82 & 2.65 & 12.69 & 3.68 & 2.42 & 2.25 & 1.82 \\ \hline PNICO23 & yes & no & 0.84 & 0.66 & 0.39 & 0.56 & 0.86 & 0.44 & 0.25 \\ HAL59 & yes & no & 0.82 & 0.94 & 0.69 & 1.00 & 0.64 & 0.63 & 0.58 \\ WATER27 & yes & unclear & 2.63 & 2.88 & 1.75 & 5.10 & 5.73 & 2.71 & 2.52 \\ Amino20x4 & yes & yes & 0.35 & 0.25 & 0.39 & 0.22 & 0.29 & 0.22 & 0.19 \\ IDISP & - & - & 3.25 & 2.27 & 7.89 & 2.69 & 2.96 & 1.61 & 1.50 \\ \hline \end{tabular} \({}^{a}\) Ref..
\end{table}
Table 4: Comparison of revPBE+D3, CX, vdW-DF2, HSE+D3, AHCX and AHBR performance on benchmarks sets that are density-sensitive, here as asserted from comparing self-consistent PBE+D4 and HF-PBE+D4 performance measures. For key benchmark examples, the second column asks ‘Is the MAD value for HF-PBE+D4 (namely, the key element of DC(HF)-PBE+D4) better than for regular (selfconsistent) PBE+D4?’, using a GMTKN55 survey of density-corrected PBE. The answers are practical identifications of sensitivity. The third column uses that data to answer the question ‘Does PBE0+D4 outperform HF-PBE+D4 for this benchmark set?’, i.e., it assesses if an unscreened hybrid can itself be expected to correct for the density sensitivity. The first and bottom sections of the table contrast benchmark performances (listed as MAD values in kcal/mol) for GMTKN55 groups 1-3 and in the inter- & intra-molecular NOC interaction groups 4 & 5, respectively.
\begin{table}
\begin{tabular}{l c c c c c c c c} & \(\tilde{S}_{\rm avg}\) & PBE & CX & vdW-DF2 & DC(HF)-PBE/+D4 & HSE+D3 & AHCX & AHBR \\ BH76 & 8.0\({}^{a}\) & 8.46 & 9.15 & 6.90 & 4.4\({}^{a}\)/4.7\({}^{a}\) & 4.21 & 5.15 & 4.14 \\ PX13 & 4.3 & 12.12 & 12.80 & 1.14 & - & 7.38 & 7.30 & 4.57 \\ \hline G2RC & 11.3\({}^{a}\) & 5.85 & 6.77 & 9.43 & 4.3\({}^{a}\)/4.1\({}^{a}\) & 6.48 & 4.71 & 3.30 \\ G21EA & - & 3.07 & 2.80 & 9.66 & - & 3.40 & 2.17 & 2.32 \\ \hline RSE43 & 3.7\({}^{a}\) & 2.54 & 2.21 & 1.13 & 2.0\({}^{a}\)/1.9\({}^{a}\) & 1.25 & 1.01 & 0.74 \\ C60ISO & 5.1 & 10.06 & 12.01 & 10.43 & - & 2.51 & 3.99 & 2.72 \\ \hline \end{tabular} \({}^{a}\) Ref..
\end{table}
Table 3: Comparison of PBE, CX, vdW-DF2, HSE+D3, AHCX and AHBR performance on benchmarks sets that have a pronounced density sensitivity, as asserted by a Hartree-Fock (HF) sensitivity measure \(\tilde{S}_{\rm avg}>2\) kcal/mol. We also list, where available, literature values for the performance of density-corrected PBE, termed DC(HF)-PBE (using HF densities to improve the PBE description when the process or reaction is found sensitive). For this density-corrected PBE, we list literature results for the performance with a Grimme-D4 dispersion correction. The table section sorts the benchmark comparison acording to their inclusion in GMTKN55 Group 3, 1, and 2, respectively. The density sensitivity of G21EA is discussed in Appendix B. Benchmark results are represented in MAD values (in kcal/mol) asserted relative to the coupled-cluster results that define the GMTKN55 reference data.
In fact, Table 3 suggests that the AHBR may also compete favorably with PBE-based density-corrected DFT on density sensitive problems. That approach is denoted DC(HF)-PBE, or DC(HF)-PBE+D4 when supplemented with the Grimme-D4 dispersion correction; The D4 inclusion is not proven relevant in these GMTKN55 benchmark cases (BH76, G2RC, and RSE43). The DF(HF)-PBE approach takes off from the above-discussed HF-PBE description, however, in DC(HF)-PBE, the HF-PBE result only replaces a self-consistent PBE result when the process-specific measure \(\tilde{S}_{\text{proc}}\) is larger than 2 kcal/mol. We find that AHBR performs better than DC(HF)-PBE in all three cases and it is, for example, more robust for both small-molecule reaction energies (RC21) and for transition states (as probed in the barrier-height benchmark BH76).
Table 4 provides further documentation of the AHBR ability to navigate molecular DFT challenges. The table compares our assessments for the RSH vdW-DFs with regular functionals (revPBE+D3, CX, vdW-DF2, and B86R) and with HSE+D3. The selection to focus our additional discussion on these benchmark sets is based on two criteria: 1) the benchmark-specific MAD values are large enough to allow a reliably interpretation in terms of XC functional trends, and 2) the usage-oriented mapping of the density-sensitivity problems, Ref. 191, suggests that we here face interesting DFT challenges. That is, we primarily focus on the benchmark sets where a separate assessment found affirmative answers (listed in the second column) to the question of whether HF-PBE+D4 improves PBE+D4 (and we can again expect an impact of density-driven errors). However, we also direct attention to performance surprises. For all cases we simultaneously report the literature answer (third column) to whether the unscreened hybrid PBE0+D4 in selfconsistent calculations makes for a further improvement over HF-PBE+D4, i.e., whether use of a hybrid can be expected to resolve the underlying DFT issues.
We want the AHBR to succeed well in all such cases as we want it to be general purpose. Note that we supplement the selection in the table with the BHPERI barrier-height set and with the IDISP set. In the case of the BHPERI set, the HF-DFT+D4 was not found to help but the PBE0+D4 was, so something more than density-driven errors could be at play. We have not found any literature evidence for density sensitivity in the case of the IDISP set. However, we still include it in this discussion, for it is our experience that IDISP and WATER27 are the two NOC-interaction benchmark sets that primarily challenge the vdW-DFs, see SI material.
We find that the RSH vdW-DFs (like HSE+D3) almost always improve the description over comparable regular forms, compare AHCX and AHBR MAD values with those for CX and B86R, respectively. For HSE+D3, the observation is only infered: HSE+D3 should ideally be compared to PBE+D3, which we have not asserted, but Ref. does identify revPBE+D3 as being the best overall GGA+D3 performer. This trend of RSH strength is generally expected. However, atom descriptions are known to generally challenge hybrids and we do find that the W4-11 set on atomization energies is the exception to this trend.
For the W4-11 set, we also find that AHCX performs better than AHBR; This is not surprising since the two RSH vdW-DFs are born with different default values of the Fock-exchange mixing. We note that picking instead the same mixing \(\alpha=0.2\) (a choice that we have previously justified for atomization-energy descriptions using the CX0P and hence AHCX), significantly helps the
Performance comparisons of vdW-DFs (as well as dispersion-corrected revPBE and HSE) for NOC interactions. We report and correlate total MADs (TMADs,) formed as a simple average over benchmark MAD values, of intermolecular and intramolecular NOC interactions; Table S IV of the SI material shows a quantitative listing. The survey of RSH vdW-DF performance is tracked at two different choices of the Fock-exchange mixings (subscripts). The top panel relies on raw planewave-DFT results, as indicated by the asterisk; We must then omit the WATER27 benchmark set. The bottom panel presents the corresponding survey as it results when we use an electrostatic-handling procedure, Appendix B, to also assert the performance for the proton-transfer processes in the WATER27 set.
AHBR\({}_{0.20}\) performance on W4-11, see SI material.
Table 4 supports in showing that AHBR performs better than the HSE+D3 overall also for this set of externally-identified DFT challenges. The W4-11 is again an exception, but there are also many cases where the AHBR is the top performer for the sets that are flagged as density sensitive. The AHBR is also better on the BHPERI set even though the PBE-based gauge found that simply going to the HF-PBE+D4 correction was not a help in that class of problems. It is clear that the hybrid benefits for BHPERI is not confined to a selection of dispersion-corrected PBE0. We furthermore observe that Ref. found dispersion-corrected PBE0 insufficient by itself to recoup or improve on a dispersion-corrected HF-PBE description for BH76, DC13, PNICO23, HAL59, and WATER27 sets. It is encouraging that, in contrast, the AHBR does provide accuracy gains over both B86R and HSE+D3 in these special cases.
Taken together, tables 3 and 4 also show that vdW-DF2 is able to navigate density-driven errors extremely well in some type of problems, but is overall characterized by having highly uneven performance. The vdW-DF2 advantages seems to primarily manifest themselves in group 3 barrier-height problems, such as the BH76, PX13, and to a lesser extent the BHPERI sets. It is also good for the WATER27 set as well as, in fact, on most NOC-interaction problems. However, it is a weak performer on the IDISP and ISOL24 challenges and the SI material shows that the vdW-DF2 also has substantial limitations for many group 1 and group 2 sets on small- and large-molecule properties.
Some of us have recently documented that a good vdW-DF1 and vdW-DF2 performance can sometimes arise because it has an XC balance that is better set up to handle the case of more diffusive interactions. The CX is considerably more accurate than vdW-DF1 and DF2 for the simpler case of CO\({}_{2}\) adsorption in Mg-MOF-74, but the reverse is true for the more complex cases of CO\({}_{2}\) adsorption in diamine-appended Zn(dobpc). We find it plausible that the vdW-DF1 and vdW-DF2 successes (compared to other vdW-DFs) on the more complex m-2-m\(-\)Zn\({}_{2}\)(dobpdc) system in part result by minimizing density-driven errors, given the fact that they navigate such errors in, for example, the BH76 and PX13 sets, above. It is also possible that the vdW-DF1 and vdW-DF2 accuracy on m-2-m\(-\)Zn\({}_{2}\)(dobpdc) energies is simply fortuitous as these functionals worsen the description of structure, for example, in m-2-m\(-\)Zn\({}_{2}\)(dobpdc). However, the relevance of using the rPW86 exchange in vdW-DF2 (that, like AHBR, uses a Schwinger-scaling argument to set the nonlocal correlation) was asserted by documenting that rPW86 mimics a Fock-exchange description for intermolecular interactions. Since vdW-DF2 does excel at many types of transition-state and NOC-interaction problems, SI Tables 5.V-71, we find it wise to respect the lessons of vdW-DF2 progress. This is especially so now that AHBR offers us a chance to combine its nonlocal-correlation design with a screened Fock-exchange form. We therefore include, below, an additional AHBR test, asserting its performance for m-2-m\(-\)Zn\({}_{2}\)(dobpdc).
### Further lessons from small and large systems
The figure concentrates on the benchmark groups that give different indications for a convenient AHCX and AHBR choice of default Fock-exchange mixing. For the vdW-DFs and for rVV10, as well as dispersion-corrected revPBE+D3 and HSE+D3, the group-averaged performance is represented in a scatter plot that relies on taking a raw, so-called TMAD, performance indicator on NOC-interaction systems (from the GMTKN55 suite). This measure is defined by taking a simple average over the MAD values that we obtain for the individual benchmarks, as described in the SI material for Ref.. We use the intermolecular (intramolecular) NOC TMAD value for group 4 (group 5) to set the abscissa (ordinate); Table 5 of the SI materials contains a listing of the TMAD values that are reported in the panels of (as well as for those evaluated for the other GMTKN55 benchmark groups).
This benchmark is often excluded in functional comparisons on vdW problems, because it contains the negatively charged small radical OH\({}^{-}\) and it is therefore not accessible in a simple benchmarking. Like for the systems in the G21A benchmark set, this radical has pronounced self-interaction errors as well as convergence challenges that prompted us to pursue the more general planewave benchmarking procedure defined in Appendix B. Insight on water in general and accuracy in WATER27 benchmarking are essential on science grounds and when pursuing systematic XC development.
We show that the correct inclusion of the WATER27 benchmark set, Appendix B and Fig. 5, has dramatic impact on what we consider XC functional promise for NOC interactions. The assessment map is clearly affected as we switch from the top to the bottom panel of In fact, it is alone the consistent-exchange class (CX, CX0P/CX0 and AHCX), the unscreened DF2-BR0, and the new AHBR that remain good options for this challenge, at least as asserted by our planewave benchmarking of the GMTKN55 suite.
In a recent paper we suggested that there is value in using the tool chain of closely related CX/CX0P/AHCX designs for a systematic exploration of bulk materials. We here propose use of AHBR for explorations on molecules and for their adsorption. It is known that density-corrected DFT should not always be used. For traditional hybrids (like HSE) there is an expectation that such hybrid-PBE will worsen the thermophysical de scription of solids, especially for magnetic elements, and for transition-metal adsorption. The idea of exploring is that we want to know what limitations holds for the RSH vdW-DFs and for AHBR in particular. More broadly, we want to learn to assert, a priori, when one should use both nonlocal-correlation and nonlocal-exchange in combination and when the regular vdW-DF (CX or B86R) suffices. Interestingly, we are finding, here and in Refs., that AHCX and AHBR improve descriptions of bulk structure and, at least, of some adsorption problems (over the underlying regular versions, CX and B86R). For molecular properties, it is instead securing a balanced and robust progress that is the challenge for hybrid vdW-DFs.
In this light, it is interesting to note that the impact of including the WATER27 challenge on the first tool chain (comprising CX-CX0/CX0P-AHCX) is, overall, smaller than that on the B86R design. The B86R itself is, in fact, moved out off the figure range for intermolecular-NOC assessment values. However, the corresponding B86R-based hybrids, DF2-BR0 and the new AHBR, remain exceptionally well suited to meet the full set of NOC challenges, at least as presently asserted. Also, looking at the quantitative measures, Table S IV of the SI material, we find that AHBR\({}_{0.20}\) appears to perform slightly better than AHBR\({}_{0.25}\) on intermolecular NOC interactions in the approximate assessment (top panel), but the actual status is different (bottom panel).
Overall, we see the robustness of CX/CX0P/AHCX and DF2-BR0/AHBR as an indication of value and usefulness as we seek to map for and understand outstanding DFT challenges. For NOC-interaction, small-molecular and barrier cases, the AHBR is the more robust, error-resilient design, Figs. 1 and 5.
The C60ISO is a benchmark where all regular vdW-DFs fails and where hybrids is needed, see SI material. It is also a case where we document that there are pronounced density-driven errors at play, compare the C60ISO values that we compute for \(\tilde{S}_{\rm avg}\) and for PBE in Table 3.
The top panels of illustrate the nature of the C60ISO benchmark set in the large-system isomerization group 2 of the GMTKN55 suite. The benchmark set considers the energy differences among 10 meta-stable forms of C\({}_{60}\) of energy \(E_{n}\). The top-left panel shows the stable Fullerene form ('\(n=1\)') and the top right panel shows the oblate form ('\(n=10\)').
Performance of original vdW-DF releases (left panel), of recent regular vdW-DFs (middle panel) and of hybrid vdW-DFs (right panel) on a set of C\({}_{60}\) isomerization problems. The top left (right) panel shows the regular (highly distorted) configuration. The middle panel depicts one of the intermediate configurations in the transformation, as tracked in the C60ISO benchmarks set. The bottom left (middle) panel shows MAD values (in kcal/mol) characterizing the performance of the original vdW-DF releases (recent regular vdW-DFs) on describing energy differences between such configurations. The bottom right panel shows that the hybrid vdW-DFs are needed to substantially improve the description.
GMTKN55 suite provides reference data for 9 isomerization energy-difference problems, denoted '\(n\)' (for \(n=1\) through \(n=9\)'). These C60ISO problems are specified by reference values for the set of differences \(E_{n+1}-E_{1}\) (\(n=1\) through 9). The top middle panel shows the atomic geometry configuration for the frustrated structure '\(n=8\)', i.e., a form that determines isomerization problem 'I7' and that is often hard to accurately describe.
The set of bottom panels provides a radar-plot comparison that reveals both vdW-DF limitations and vdW-DF promise. Specifically, the bottom left (middle) panel shows that vdW-DF1 and vdW-DF2 both fail (that recent releases and variants including CX and B86R offer no improvements). As the benchmark is prone to density-driven errors, Table 3, it is no surprise that HSE+D3 provides an significantly improved description over regular vdW-DFs. As shown it the bottom-left panel (note the change in scale), the AHCX helps, but AHBR is needed to improve the description to the HSE-D3 level.
We interpret the nonhybrid-vdW-DF performance issues as arising because we must here describe stretched and frustrated binding. The fact that there are density-driven errors is not surprising, for this isomerization problem can also be seen as another transition-state case. In fact, the set of meta-stable C60 configurations can be seen as configurations that define some effective deformation paths taking us from configuration '1' to '10'.
We can expect that the DF2-BR0 and AHBR are accurate also for the C60ISO set (as it is documented in the lower-right panel). However, given the status for the (barrier-height) benchmark group 3 and given that this is a transition-state problem, it is perhaps surprising that the vdW-DF1 and vdW-DF2 are not acceptable performers here. It appears that there are more than one type of challenges in describing transition states. We can certainly still learn from vdW-DF2 and vdW-DF1, but is also clear that we need to cast a wider net to identify good development ideas. This can, as suggested here, be done by use of the more robust AHBR.
### Need for vdW-DF hybrids: Molecular examples
Transition-state problems stand out because they involve a comparison of energy terms that must simultaneously reflect several different types of binding, for example, relaxed and stretched or diffusive. Some of us have recently characterized functional performance in the description of CO\({}_{2}\)-MOF adsorption that happens in concert with a site-specific-reaction and resulting CO\({}_{2}\)-insertion in a diamine-appended MOF. There, the vdW-DF1 and vdW-DF2 also had an accuracy edge but it did not apply for CO\({}_{2}\) adsorption in the simpler Mg-MOF-74 system. The transition-state problems can be seen as key drivers for XC development.
S.1-2 of the SI material provide a broad illustration that the vdW-DF tool bag must include the set of new hybrid vdW-DFs. i.e., DF2-BR0,
Drivers for XC-development in the GMTKN55 suite of benchmarks: Performance of vdW-DF-based hybrids on traditional molecular challenges (identified in left panel) and on what we consider key transition-state challenges (right panel) in benchmark set PX13 (I), BHPERI (II), IDISP (III), WATER27 (IV), and WCPT18 (V). Self-interaction errors play an essential role not only in the SIE4x4 set but also in G21EA (WATER27) due to the presence of negative ions and radicals, requiring an electrostatic-environment approach for reliable assessments (appendix B). The traditional challenges include the so-called ‘difficult-for-DFT’ (DC13) and ‘mindless benchmarking’ (MB16-43) sets, as well as Fullerene isomerization problems (collected in benchmark set C60ISO).
AHCX and AHBR. The set of recent regular vdW-DFs (including CX and B86R) do not remain uniformly accurate when tracked across the 7 individual barrier-height benchmarks of GMTKN55, (right panel) and SI material. Inspecting SI Tables S V through S X makes it clear that problems of a transition-state nature (such as we deem those of the C60ISO and MB16-43 sets), in general, challenge the non-hybrid vdW-DFs. Nevertheless, we find that the DF2-BR0 performance stands out in terms of both accuracy and cross-benchmark resilience even at transition-state problems. As noted in the introduction, in turn, the DF2-BR0 resilience and transferability for molecular problems motivate the AHBR completion.
Meanwhile, the underlying benchmarking data, Tables S II through S X of the SI material, shows that the vdW-DF2-ah hybrid is not a reliable option for the study of broad molecular properties (just like it failed for bulk). Like the vdW-DF2, it has an acceptable performance for NOC-interaction and barrier-height problems. However, vdW-DF2 and vdW-DF2-ah have real shortcomings when the focus is instead moved to general properties of small- and large-molecule systems.
More broadly, the GMTKN55 survey suggests that we cannot hope for one gradient-corrected exchange form (as used in the present type of nonhybrid vdW-DF designs) to always succeed. We see that the assessments on GMTKN55 performance show a spread in performance - for all nonhybrid vdW-DFs - among different classes of problems, see SI material. The same is true, but to a significantly smaller extent, of the CX and B86R-based hybrids. There are many problems, and it is not trivial to secure a good XC balance in the present type of vdW-DF designs, across all of such problems. Specifically, in the vdW-DFs, we add an attractive nonlocal-correlation term to the spurious LDA-exchange overbinding. The nonlocal correlation contains more than pure dispersion effects while it also goes beyond GGA correlation at shorter distances. At the same time, in the nonhybrid vdW-DF we must rely on nothing but gradient-corrected exchange for stabilization in the presence of both LDA overbinding and the enhanced attraction of these nonlocal-correlation effects.
The fact that the strength of the vdW attraction may enhance in select situations suggests an additional mapping role for vdW-DF hybrids. The plan is simply to use and contrast AHCX and AHBR for explorations, for example, in cases with diffusive interactions, when interactions compete, and when phase-transformations compete. In the latter case, the presence of an actual or incipient ferroelectric transformation will itself affect the magnitude of the dielectric constant. For example, while the unscreened hybrid CX0P is highly accurate for the BaZrO\({}_{3}\), we need more to address the SrTiO\({}_{3}\). Access to RSH vdW-DFs means that we better map the nature of specific challenges, correlate progress with design choices, and eventually implement new ideas in XC developments.
In we identify a key set of molecular-type challenges that we think can be used to drive XC development, without going directly to more complex (but technologically relevant) cases. For example, an accurate description of some transition-state problems (benchmark sets 'I', 'II', 'IV', and 'V' in the left panel) requires a unique XC balance that vdW-DF2 occasionally provides, yet vdW-DF2 often fails spectacularly, as seen in the SI material. Ref. provides an adsorption case where vdW-DF2 also shines by delivering an exceptionally large repulsion by gradient-corrected exchange, but such repulsion is not always needed. It is therefore good to seek a simpler way to survey for XC issues, as in Our proposal is based on the experience that we have gained by planewave benchmarking across both the entire GMTKN55 and for CO\({}_{2}\) uptake in MOFs, for example, in Refs.. We mostly echo, but also simplify, the logic that led to the definition of the full GMTKN55 suite. In making this identification, we are assuming that we are working with a vdW-inclusive functional, like the vdW-DFs, so that the design is also capable of dealing with the group 4 and group 5 types of NOC-interaction problems.
Well known challenges for any DFT are, of course, still found in the SIE4x4 set (on self-interaction errors in of neutral and positively charged systems), the MB16-43 (mindless benchmarking) set and the DC13 (difficult-for-DFT) set. To these sets we add the G21EA and WATER27 sets because these two sets suffer from pronounced SIE effects, and in the latter case will also significantly impact the performance of any given candidate on the group 4 problems. Additionally, we include the IDISP set as it is almost always the challenge that dominates in setting the performance of an XC functional on the group 5 of intramolecular NOC interactions.
From the barrier-height class, we find that the performance varies prominently when inspecting the BHPERI, PX13, and WCPT18 sets. At least it is clear that these benchmarks allow the hybrid benefits to directly manifest themselves.
Schematics of the CO-on-Pt problem: Competition between TOP- and FCC-site adsorption. The image is generated using the VESTA program.
This is because the vdW-DF1 and vdW-DF2 barrier-height successes, at the WCPT18 and PX13 benchmark sets, do not port well to this C60ISO set.
Finally, illustrates the usefulness of simplifying functional comparisons while focusing on where we can learn more. The figure compares the performance of the CX/CX0P/AHCX and the B86R/DF2-BR0/AHBR functionals in two types of radar plot. The left panel makes the comparison based on G21EA as well as on the sets where there are often massive deviations between DFT results and reference data from quantum-chemistry calculation. The right panel makes the comparison on the selection of barrier problems and of NOC-interaction problems. An overall impression is that the robustness of the DF2-BR0 and AHBR is confirmed from the testing summary presented in Figs. 1 and S.I. However, we also see that the MB16-43 benchmark set identifies an example set of problems where we can still learn more from the Lindhard-screening logic (summarized in Section II) and the screened-exchange gradient expansion result that underpins CX and hence AHCX.
### CO adsorption on Pt
We note that the CX has a lattice constant that is in close agreement with experimental characterizations of Pt, Table 1. This observation motivates our assessment strategy: To keep the adsorption structure fixed at the CX description. We shall concentrate on directly comparing and discussing the new-XC-functional accuracy in the CO/Pt adsorption within a fixed-nuclei framework.
We compute the site variation in the adsorption or binding energy
\[E_{\rm bind}^{\rm TOP/FCC}=E_{\rm CO/Pt}^{\rm TOP/FCC}-E_{\rm Pt,surf}-E_{\rm CO,mol}\,, \tag{18}\]
We focus on discussing the site-preference energy
\[\Delta E_{\rm site}=E_{\rm bind}^{\rm TOP}-E_{\rm bind}^{\rm FCC}\,. \tag{19}\]
Experimental studies clearly indicate that TOP-site adsorption (at binding energy \(-1.30\) eV) applies for dilute coverage at room temperatures (corresponding to 0.026 eV). Results obtained using the random-phase approximation concurs with these observations.
Accordingly, DFT calculations with an accurate XC functional should find \(\Delta E_{\rm site}\) negative and with a magnitude that exceeds the 0.026 eV value. However, this problem is a long-standing challenge for DFT, in the sense that essentially all non-vdW-DFs (including HSE and PBE0) fail, when considered at or close to the actual Pt lattice constant. We observe that density-driven errors are expected to complicate the DFT setting of the correct CO/Pt adsorption-site preference.
Table 5 summarizes our comparison of RSH vdW-DF performance for the classic CO/Pt problems. We note that the AHBR lattice constant for Pt is slightly larger (and further from experiment) than the CX choice we have used. An AHBR adsorption study at the AHBR lattice constant will have a more narrow Pt \(d\) band and therefore yield smaller adsorption energies.
\begin{table}
\begin{tabular}{c|c c c|c c c} & CX & AHCX & AHCX\({}_{0.25}\) & B86R & AHHR\({}_{0.20}\) & AHBR \\ \hline \(E_{\rm bind}^{\rm TOP}\) & -1.830 & -1.949 & -1.974 & -1.683 & -1.801 & -1.824 \\ \(E_{\rm bind}^{\rm FCC}\) & -1.966 & -1.954 & -1.940 & -1.780 & -1.767 & -1.752 \\ \(\Delta E_{\rm site}\) & 0.133 & 0.005 & -0.034 & 0.097 & -0.034 & -0.072 \\ \hline \(E_{\rm gap}^{\rm CO}\) & 7.048 & 8.768 & 9.192 & 7.081 & 8.798 & 9.222 \\ \end{tabular}
\end{table}
Table 5: Comparison of TOP- and FCC-site CO/Pt binding energies \(E_{\rm bind}\), site-preference energies \(\Delta E_{\rm site}=E_{\rm bind}^{\rm TOP}-E_{\rm bind}^{\rm FCC}\), as well as of molecular-gap results \(E_{\rm gap}^{\rm CO}\) (all in eV). The first and second block are for CX-AHCX and for B86R-AHBR, respectively. For both tool chains we also illustrate the impact of the choice of the Fock-mixing (as identified in the functional-label subscript). All results are provided for the CX provided substrate lattice constant \(a_{0}=3.929\) Å, molecular structure, and adsorption-induced deformations. Experimental observations of CO adsorption find TOP site adsorption with binding energy \(\Delta E_{\rm site}=-1.32\) eV.
Correlation between results for the CO molecular gap \(E_{\rm gap}^{\rm CO}\) and for the CO/Pt site-preference energy \(\Delta E_{\rm site}\) as computed in the CX-AHCX functional chain and in the B86R-AHBR functional chain. We keep the adsorption geometry fixed in all studies.
We find that neither CX nor the default AHCX description (using a 0.20 fraction of Fock-exchange mixing) offers an improvement in the description of the site-preference challenge. In contrast, the new AHBR works (as do AHBR\({}_{0.20}\) and AHCX\({}_{0.25}\)), bringing the site-preference description in alignment with experimental observations. Still, the AHBR does not offer a complete resolution of the CO/Pt problem because it is over-estimating the actual adsorption energy \(E_{\rm bind}\).
The adsorption is often discussed in terms of the Bylholder model and, as such, controlled by the substrate electronic structure (which we must accurately characterize to correctly describe the molecule-to-substrate charge transfer) and the molecular gap (that we must accurately characterize to correctly describe the back donation). We keep the adsorption geometry fixed in all calculations, and thus contrast the direct effects that the functionals have on both the molecule gap and substrate electronic structure.
There are several lessons from First, there is a strong impact on the molecular gap by moving from a regular function to the associated RSH. Second, this gap variation effectively controls the prediction of site preference within a given tool chain. Third, there is also a systematic effect of switching between the tool chains and thus changing the substrate electronic structure. In effect, the figure documents that the well-defined AHCX-AHBR differences (in terms of nonlocal-correlation- and exchange-design details) have a direct impact on the Pt electronic description. This impact is making the AHBR better at reflecting the true site variation in the CO/Pt adsorption energy.
### CO\({}_{2}\) adsorption in Mg-MOF-74 and m-2-m\({}_{2}\)(dobpdc) MOFs
The panels show atomic coordinates as obtained in fully relaxed CX characterizations that form the starting points for our performance comparisons and discussion.
We here assert the AHBR (and AHCX) performance by systematically calculating the CO\({}_{2}\) binding enthalpy at room temperature \(H_{\rm ads}^{\rm room}\) for Mg-MOF-74 and the Born-Oppenheimer (frozen-atom) binding energy
\[E_{\rm ads,BO}=E_{\rm CO_{2}-MOF}-\left(E_{\rm MOF}+E_{\rm CO_{2}}\right), \tag{20}\]
In Eq., we evaluate the difference between the total energy of CO\({}_{2}\)-adsorbed MOF (\(E_{\rm CO_{2}-MOF}\)) and the total energy sum of Mg-MOF-74 or m-2-m-Zn\({}_{2}\)(dobpdc) (\(E_{\rm MOF}\)) and gas phase CO\({}_{2}\) (\(E_{\rm CO_{2}}\)). For the gas phase CO\({}_{2}\) energy, we optimize the geometry within a 20 A \(\times\) 20 A \(\times\) 20 A cubic cell. We have verified that our computational set up, Appendix A, converges the CO\({}_{2}\) binding-energy results for Mg-MOF-74 and m-2-m\(-\)Zn\({}_{2}\)(dobpdc) adsorptions to 1 kJ/mol.
To simplify comparison among functionals - and to limit computational costs - we provide all calculations with the MOF structures kept fixed at the CX results for the adsorption geometry; This strategy is similar to that we used for discussing the CO/Pt problem, above. However, in the case of carbon capture in Mg-MOF-74, we permit the CO\({}_{2}\) molecules to relax according to the forces that we compute in the specific functionals.
We seek to compare with room-temperature measurements of the CO\({}_{2}\) heat of adsorption, \(H_{\rm ads}^{\rm room}\). The DFT (and the true) internal-energy binding description are thus affected by vibrational zero-point energy (ZPE) and thermal-energy (TE) corrections:
\[H_{\rm ads}^{\rm room}=E_{\rm ads,BO}+\Delta\rm ZPE+\Delta\rm TE. \tag{21}\]
Optimized crystal structures of CO\({}_{2}\)-adsorbed Mg-MOF-74 (upper panel) and CO\({}_{2}\)-inserted m-2-m\(-\)Zn\({}_{2}\)(dobpdc) (lower panel). The gray, orange, red, skyblue, brown, and pink circles represent Zn, Mg, O, N, C, and H atoms, respectively. The images are generated using the VESTA program.
Accordingly, for the simpler Mg-MOF-74 system, we also compute and contrast the vibrational frequencies of adsorbed CO\({}_{2}\) and free CO\({}_{2}\) using a finite difference approach with the Phonopy package. Specifically, we displace each atom of CO\({}_{2}\) in twelve random directions with a constant displacement distance (0.03 bohr) to extract corrections \(\Delta\text{ZPE}+\Delta\text{TE}\) that hold at 298 K.
In the larger, more complex diamine-functionalized case, we stick with comparing the Born-Oppenheimer results, Eq., directly with the measured \(H_{\text{ads}}^{\text{room}}=-57\) kJ/mol value. Based on experience from asserting the performance of density-explicit vdW-DFs, in Ref., we expect that a plausible back-corrected experimental value would be approximately \(-63\) to \(-60\) kJ/mol as the computed \(\Delta\text{ZPE}+\Delta\text{TE}\) value is about 3-6 kJ/mol (see Ref.).
Table 6 compares our results, contrasting the vdW-DF tool chain descriptions of CO\({}_{2}\) uptake in the Mg-MOF-74 (in the diamine-functionalized MOF) with experiments in the upper (lower) section. In the case of Mg-MOF-74, we find that the AHCX result for the Mg\(-\)O distance is excellent but that AHBR is also accurate. For adsorption energies we find that AHCX systematically strengthens the CX binding results in both MOF cases. In the complex diamine-functional MOF case, we therefore find that moving to AHCX does not repair a clear overbinding tendency that we have very recently documented for present vdW-DFs.
In contrast, the results for CO\({}_{2}\) adsorption in new RSH vdW-DF shows a trend of vdW-DF repairing. The AHBR result \(H_{\text{ads}}^{\text{room}}=-46.6\) kJ/mol for the Mg-MOF-74 is in itself excellent, being in close agreement with the experimental value \(-43.5\) kJ/mol. The AHBR outperforms all of the vdW-DFs that we have previously tried (see SI material associated with Ref.), for example, lifting the underbinding B86R \(H_{\text{ads}}^{\text{room}}=-38.9\) kJ/mol value.
Meanwhile, in the case of the diamine-functionalized MOF, the AHBR \(E_{\text{ads,BO}}=-86.8\) kJ/mol result instead _lowers_ the clearly overbinding B86R \(H_{\text{ads}}^{\text{room}}=-92.1\) kJ/mol value towards the value of the measurement, at \(-57\) kJ/mol. According to previous vdW-DF experience in characterizing vdW-DF vibrational corrections to \(E_{\text{ads,BO}}\), this AHBR characterization leads to the estimate \(H_{\text{ads}}^{\text{room}}\approx-84\) to \(-81\) kJ/mol. The AHBR therefore outperforms SCAN, SCAN+rVV10 and the recent vdW-DFs (vdW-DF1 and vdW-DF2 are better on energies but have too long binding lengths). The AHBR has an accuracy that matches the semi-empirical rVV10 for this MOF challenge. Unlike in the case of Mg-MOF-74, the AHBR does not perform at the revPBE+D3 level for m-2-m\(-\)Zn\({}_{2}\)(dobpdc). However, unlike the AHCX, the AHBR is able to move the non-empirical vdW-DFs towards a binding softening. Robustness, a repairing behavior (documented here and for many density-driven challenges), is needed when the vdW-DF method faces significant charge relocations that, in turn, challenge the XC balancing.
### Base-pair stacking in a DNA model
DNA can be seen as a stacking of Watson-Crick (WC) base pairs that are essentially flat and therefore have a significant (eV-scale) vdW attraction from one base pair to the next. The WC pairs are steps in the resulting double-helix DNA structure. There are in total 10 possible combinations for 2 steps, i.e., base-pair combinations that are here denoted ApA, ApA, ApT, ApC, ApG, CpC, CpG, GpC, TpT, TpC, and TpG; Ref. identifies and illustrates a set of possible atomic positions for these base-pair combinations. We want to compute such base-pair stepping energies since the mutual vdW attraction might have driven the DNA self-assembly in the first place, as life emerged.
Some of the DNA cohesion comes, of course, from the presence of the sugar-phosphate backbones that incorporate and organize the WC bases into two strands, with the sequence of WC bases, adenine (A), thymine (T), cytosine (C), or guanine (G), setting the genetic code. The strands are complementary in the sense that the WC-base sequence must be exactly matched, with each of the individual bases having only one suitable counter part, i.e., forming steps that must have one of the A-T, T-A, C-G, or G-C forms.
\begin{table}
\begin{tabular}{l|c c c|c|c} & CX\({}^{a}\) & AHCX & AHCX\({}_{0.25}\) & AHBR & Exper. \\ \hline Mg\(-\)O & 2.29 & 2.27 & 2.26 & 2.25 & 2.27\({}^{b}\) \\ \(E_{\text{ads,BO}}\) & -53.7\({}^{a}\) & -58.5 & -59.7 & -50.3 & - \\ \(\Delta\)ZPE & 2.8 & 2.4 & 2.5 & 2.6 & \\ \(\Delta\)TE & 1.1 & 1.6 & 1.3 & 1.1 & \\ \(H_{\text{ads}}^{\text{room}}\) & -49.7 & -54.5 & -55.9 & -46.6 & -43.5\({}^{b}\) \\ \hline \(E_{\text{ads,BO}}\) & -94.6 & -98.7 & -99.7 & -86.8 & -57\({}^{c}\) (-63 - -60 ) \\ \hline \({}^{a}\) Ref. & & & & & \\ \({}^{b}\) Ref. & & & & & \\ \({}^{c}\) Ref. & & & & & \\ \end{tabular}
\end{table}
Table 6: Comparison of performance for AHBR and for the CX-based tool chain on the CO\({}_{2}\) adsorption energies in Mg-MOF-74, top section, and in m-2-m\(-\)Zn\({}_{2}\)(dobpdc), bottom section. All calculations are performed at the CX characterizations for the MOF geometries. For Mg-MOF-74, we permit the CO\({}_{2}\) to relax as described in the stated functional, and list the results for the characteristic bond length. We presents energy (bond length) results in kJ/mol (in Å) to facilitate an easy comparison with a recent MOF study covering regular vdW-DFs and revPBE+D3. The subscript ‘BO’ identifies Born-Oppenheimer results, while a superscript ‘room’ identifies adsorption free-energy values that include the effects of vibrations as described at room temperature, using Eq. to facilitate a direct comparison to experimental values for Mg-MOF-74. The values in parenthesis represent an estimate for how a back correction for vibrations would impact the measured m-2-m\({}_{2}\)(dobpdc) \(H_{\text{ads}}^{\text{room}}=-57\) kJ/mol value.
The DNA strands are mutually bonded, by a combination of hydrogen and vdW binding, but the energy of the WC pairings (A-T, T-A, C-G, G-C, among one base and its counter part directly across) is not our present focus. Instead, we seek to understand the extent that the mutual step-binding energies contribute to the DNA cohesion, using a DNA model that ignores the backbone but instead relies on hydrogen terminations of the bases.
To set us up for future, more general DNA explorations, we contrast the performance of the first and second vdW-DF tool chains relative to the reference descriptions provided in Ref.. That is, we compare with so-called domain-based pair natural orbital couple-cluster ('DLPNO-CCSD(T)') calculations at fixed reference base-pair combination structures. The DLPNO-CCST(T) method is also used for setting reference energies of the GMTKN55 suite.
\[\Delta E_{\mathrm{WC-step}}^{\mathrm{ApC}}=E_{\mathrm{ApC}}-E_{\mathrm{A-T}}- E_{\mathrm{C-G}}\,. \tag{22}\]
The reference work also provides data for the sum of pair interactions among the four bases, \(\Delta E_{\mathrm{B-pair}}^{\prime}\), excluding the two WC pairings (as indicated by the prime). This pair summation is illustrated in the abstract figures of Refs. and.
Tables S XVI and S XVII of the SI material compares the performance of CX/AHCX/AHCX\({}_{0.25}\) and of B86R/AHBR\({}_{0.20}\)/AHBR for each of the base-pair combinations, reporting (in kcal/mol) the \(\Delta E_{\mathrm{WC-step}}\) and \(\Delta E_{\mathrm{B-pair}}^{\prime}\), respectively. We also report mean deviation (MD) and MAD values relative to the DLPNO-CCSD(T) calculations. Table VII summarises the performance comparison (in terms of 10-base-pair-combination averages) and makes it clear that the AHBR is a strong performer for the description of the DNA base-pair assembly.
We find that B86R performs better that CX and that the AHBR functional design is in fact very accurate also for descriptions of the DNA stepping energies. This is especially true when it is used in the suggested default mode with a 0.25 Fock-exchange mixing. The description is significantly more accurate than a standard choice of dispersion-corrected hybrid DFT provides, as also listed in Table VII. This result is in itself encouraging.
We also observe that the AHBR is somewhat less accurate when it is instead used to study energies from the sum of pair contributions, \(\Delta E_{\mathrm{B-pair}}^{\prime}\). As in Ref., we find that the vdW-DF based descriptions of the stepping energies benefit from a cancellation of errors that affect the descriptions of pairing between individual bases.
Finally, Table VII reveals an important difference between the performance trend in the CX-AHCX and B86R-AHBR functional chains for this class of large molecular problems. We find that moving to a hybrid form significantly worsens the CX/AHCX descriptions, whereas we find that such a step slightly improves the B86R/AHBR-type description. In both tool chains, we find that including and increasing Fock exchange fraction strengthens the interactions but the changes are small in the AHBR case. Including Fock exchange makes AHBR very accurate but the most important lesson is perhaps that the B86R-AHBR functionals have an inherent stability here: They start and remain accurate on the DNA assembly energies.
## IV Conclusion and Outlook
We have developed a new, accurate, non-empirical RSH vdW-DF, termed vdW-DF2-ahbr, and we have documented general-purpose capabilities for molecular problems as well as promise for bulk and adsorption properties. Since AHBR is based on a range separation of the Coulomb interaction, it is setup (and coded) to allow use of a physics-based tuning of both the Fock-exchange mixing \(\alpha\) and of the RSH inverse screening length \(\gamma\). This means that it is also possible to ameliorate residual RSH vdW-DF errors in charge-transfer descriptions. In this first presentation, however, we have deliberate kept these values fixed to allow a simple demonstration of broad AHBR capabilities and usefulness.
Figures 1 and S.1-2 of the SI materials highlight two important take-away messages of this paper, namely
\begin{table}
\begin{tabular}{l c c} \hline \hline XC study & \(\Delta E_{\mathrm{WC-step}}\) & \(\Delta E_{\mathrm{B-pair}}^{\prime}\) \\ \hline B3LYP+D3a & 0.89 & - \\ CXb & 1.73 & 0.44 \\ CXGBRV & 1.17 & 1.49 \\ \hline CX & 1.48 & 1.75 \\ AHCX & 3.06 & 3.30 \\ AHCX\({}_{0.25}\) & 3.47 & 3.69 \\ \hline B86R & 0.38 & 0.82 \\ AHBR\({}_{0.20}\) & 0.12 & 0.60 \\ AHBR & 0.08 & 0.52 \\ \hline \hline \end{tabular}
\end{table}
Table VII: Comparison of the CX-based and B86R-based tool chain performance for DNA assembly. We report MAD values (in kcal/mol) as extracted by averaging the deviations of vdW-DF results relative to coupled-cluster calculations on the 10 different base-pair stacking configurations, using the specific configurations that are provided in Ref.. We report on ONCV-SG15/160 Ry results, but we also compare with literature values as well as an assessment obtained from a separate ultrasoft-PP study (‘GBRV’). The reference calculations use DLPNO-CCST(T) to compute the step energy \(\Delta E_{\mathrm{WC-step}}\) and \(\Delta E_{\mathrm{B-pair}}^{\prime}\) which is a sum over relevant molecular-pair contributions.
Figures 1 and 5 exemplify a key conclusion about the AHBR advantage: It has a robust ability to navigate so-called density-driven DFT errors.
This vdW-DF2-ahbr resilience, in combination with its emphasis on a MBPT foundation, suggests that it will be accurate also beyond the successes that we have here documented for a set of DFT challenges in molecule, layered, bulk, and surface systems. When using formal MBPT to compute the total energy, there is an inherent robustness towards making approximations. That robustness also extends to the exact XC energy functional. We rely on MBPT guidance in making XC energy approximations (such as CX, AHCX, and AHBR) so that these vdW-DFs can potentially benefit from that inherent robustness. However, as we also discuss, such benefits can only emerge in practice when the actual XC functional approximation delivers accurate orbitals, for example, as tested on the quality of its density description. That AHBR matches or exceeds B3LYP/HSE+D3 for transition-state problems means that it generally navigates density driven errors and therefore outperforms the B3LYP/HSE+D3 broadly - for example, as summarized in Figs. S 1-2 of the SI material. Like the HSE+D3 it has both a strong resilience to density errors and the MBPT foundation to benefit from the formal-MBPT robustness towards XC-functional approximations.
We argue that this general-purpose character suggests that AHBR should be used to map strengths and weaknesses of the vdW-DF method on molecules, just like the AHCX can serve us in that role for bulk systems.
An overall outcome of this vdW-DF2-ahbr work is also a roadmap with a DFT-usage feedback strategy for making further functional improvements in the vdW-DF framework. Since both the AHCX and AHBR are found to be fairly robust in all tested problems, we can contrast performance differences over a broad range of systems. Furthermore, since the pair of RSH vdW-DFs have systematic design differences, we can correlate the performance differences in terms of the nature of the underlying physics input to the XC designs. The AHBR and AHCX are particularly useful because they are complementary, representing one of two internally-consistent classes of MBPT input on how exchange effects impacts all XC components. Specifically, as explained in Section II.A and II.B, the AHBR and AHCX rely on valid but different interpretations of formal MBPT; They correspond to systematic reliance of molecular or a weakly-perturbed-bulk perspective on screening, respectively). Taken together, the observations allow us to interpret the DFT-usage feedback and draw development conclusions concerning which types of MBPT input to prioritize. This DFT-feedback strategy is in many ways just a continuation of the electron-gas tradition that, as we see it, has pushed MBPT-based DFT from LDA over constraint-based GGAs and to the vdW-DF method.
|
10.48550/arXiv.2203.06682
|
Accurate non-empirical range-separated hybrid van der Waals density functional for complex molecular problems, solids, and surfaces
|
Vivekanand Shukla, Yang Jiao, Jung-Hoon Lee, Elsebeth Schroder, Jeffrey B. Neaton, Per Hyldgaard
| 5,461
|
10.48550_arXiv.1301.6928
|
###### Abstract
**Status:**_To appear in JMO special issue "Ultrafast Dynamical Imaging of Matter", summer 2013._
**Version:**_RevTex 4.1 typeset version for arxiv._
Time-resolved coincidence imaging of photoelectrons and photoions represents the most complete experimental measurement of ultrafast excited state dynamics, a multi-dimensional measurement for a multi-dimensional problem. Here we present the experimental data from recent coincidence imaging experiments, undertaken with the aim of gaining insight into the complex ultrafast excited-state dynamics of 1,3-butadiene initiated by absorption of 200 nm light. We discuss photoion and photoelectron mappings of increasing dimensionality, and focus particularly on the time-resolved photoelectron angular distributions (TRPADs), expected to be a sensitive probe of the electronic evolution of the excited state and to provide significant information beyond the time-resolved photoelectron spectrum (TRPES). Complex temporal behaviour is observed in the TRPADs, revealing their sensitivity to the dynamics while also emphasising the difficulty of interpretation of these complex observables. From the experimental data some details of the wavepacket dynamics are discerned relatively directly, and we make some tentative comparisons with existing _ab initio_ calculations in order to gain deeper insight into the experimental measurements; finally, we sketch out some considerations for taking this comparison further in order to bridge the gap between experiment and theory.
## I Introduction
Coincidence imaging techniques, in which the full momentum vector of both photoelectron and photoion is measured, have been growing in popularity and sophistication over the last 20 years. The COLTRIMS (Cold Target Recoil Ion Momentum Spectroscopy) community, in particular, has developed significant expertise in relatively high-energy and multi-coincidence measurements, typically (although not exclusively) utilising synchrotron light sources in photoionization studies. The original aim of COLTRIMS was application to many-body collision dynamics, via kinematically complete measurements of collision systems, although the technique has since been applied to studies as diverse as photoelectron diffraction, probing entanglement and the investigation of tunnel ionization, as well as extensive studies of photoelectron angular distributions. Flat-field and VMI based coincidence imaging experiments have also been used by a handful of groups, usually with a focus on lower-energy processes as appropriate to experiments based around table-top laser sources. The first demonstration of femtosecond time-resolved coincidence imaging to study photochemical processes was over a decade ago, and this type of imaging measurement provides the fullest experimental dataset possible, which can be considered as a 7D measurement, or even 8D if one considers the fragment mass spectrum as a distinct observable to the fragment velocity distributions; time-resolved coincidence imaging therefore provides the best chance of elucidating complicated, multi-dimensional, excited state dynamics from experimental measurements.
Despite the potential of coincidence imaging, the technique has had only a small impact thus far to time-resolved measurements generally, and more specifically to measurements utilising UV sources. The difficulty of applying coincidence imaging in time-resolved, UV pump-probe experiments is partly due to the limitation of single-particle counting techniques - with consequent requirements for long experimental runs and long-term experimental stability - which makes time-resolved experiments particularly challenging; additionally there is the inherent difficulty of producing and controlling short-pulse UV light. A particular issue is the minimisation of background signal from scattered light, which becomes a problem on a per photon basis once photon energies are above the work function of the materials used in the spectrometer. The benefit of UV wavelengths is that, for many small molecules, the photon energies are sufficient for 1-photon pump, 1-photon probe experimental schemes, which are ideal for the study of the dynamics of electronically excited states of molecules. In these kind of schemes the laser intensities can be kept low and well within the perturbative regime (\(\ll 10^{12}\) Wcm\({}^{-2}\)), and the observables take their simplest form. The few successful studies to date have begun to explore the power of the technique, but much work remains to be done.
The observables provided in a full imaging study provide additional information beyond the energy-time mapping of a, now routine, 2D photoelectron or photoion measurement, which provide time-resolved photoelectron spectra (TRPES) or mass spectra (TRMS) respectively. In particular the time-resolved photoelectron angular distributions (TRPADs) are sensitive to the electronic structure of the ionizing state, so are expected to reveal subtle details of the non-adiabatic electronic dynamics. In the simplest case one might concoct, that of passage through a conical intersection (CI) leading to a change in the electronic symmetry of the excited state, the PAD is expected to change reflecting the non-adiabatic dynamics and map principally the electronic part of the dynamics. This expectation can be contrasted with the time-resolved photoelectron spectrum, which can be considered to be an observable dominated by vibrational motions, but may additionally map some aspects of the non-adiabatic electronic dynamics cleanly depending on the ionization correlations.
However, in the case of large amplitude motions on a single electronic state (adiabatic dynamics) changes in the PAD are also expected and, furthermore, the PAD is energy dependent, so changes to the vertical ionization potential (IP) as a function of nuclear coordinates will also couple into the form of the observed PAD. These factors make the mapping of the dynamics onto the TRPADs non-trivial to understand at both a qualitative and quantitative level, and obviate the (relatively) simple picture that the TRPADs map only the electronic dynamics for all but the simplest of cases; recent computational studies of excited state dynamics in NO\({}_{2}\) have illustrated the response of the TRPES and TRPADs to complicated excited-state dynamics, demonstrating the richness of the observable while also suggesting the difficulty of obtaining detailed insights into molecular dynamics via a purely experimental approach, with no _a priori_ knowledge of the underlying dynamics. Of particular note is the non-isomorphic nature of the nuclear configuration and observable mapping spaces: the wavepacket motion on the excited-state, which includes dispersion, bifurcation, interferences and other complex quantum-mechanical behaviours, does not allow for a direct mapping of a given dimension in nuclear coordinate space onto a given dimension of the observable (e.g. photoelectron energy, anisotropy parameter), although such mappings may be possible in low dimensionality problems such as vibrational wavepackets in diatomics. However unsurprising this conclusion is, the knowledge gap between the observable and the underlying wavepacket dynamics is often overlooked or ignored in treatments of time-resolved measurements.
Despite these difficulties, the TRPADs provide an additional observable - therefore more information - than the TRPES alone, and additional dimensionality to the dataset. One demonstration of the utility of this higher information content has been the interpretation of experimental measurements via qualitative/semi-quantitative modelling of TRPADs, which can provide insight into the mapping of the excited state wavepacket to the observables without the need for full _ab initio_ treatments, and yield deeper insight into the molecular dynamics than the TRPES alone. Many studies based on velocity map imaging (VMI) measurements have also illustrated the utility of TRPADs at a phenomenological level (in no small part because PADs come "for free" with the technique) and a recent review article has surveyed much of the work by such "users" of photoelectron angular distributions.
The ultrafast dynamics of 1,3-butadiene (\(C_{4}H_{6}\)) have been studied in considerable detail experimentally and theoretically, most recently by Boguslavskiy et. al. and Levine et. al. (for a more detailed overview of the butadiene literature to date see ref.). In the experimental photoelectron study, butadiene was excited to the bright \(1^{1}B_{u}\) state, with UV radiation around 216 nm (5.74 eV), and probed via ionization with a time-delayed UV pulse at 266 nm (4.66 eV). The computational studies, based around an _ab initio_ multiple spawning (AIMS) methodology, focussed on describing the excited state dynamics of population on the same bright state and with similar energy to the experimental case.1 The conclusions from these complementary experimental and theoretical studies are in accord, and point to rapid and complicated dynamics involving fast motion on the initially populated bright \(1^{1}B_{u}\) state (historically termed the \(S_{2}\) state, as it lies higher in energy in the Franck-Condon region), with twisting about the carbon backbone and out-of-plane bending motions leading to highly distorted geometries (relative to the planar ground state) on \(<\)40 fs timescales. At least two minimum energy conical intersections (CIs), coupling \(S_{2}\) to \(S_{1}\) (the optically dark \(2^{1}A_{g}\) state) and three CIs coupling \(S_{1}\) to \(S_{0}\), were found to play important roles in the relaxation of the excited state. In a wavepacket picture, the initial dynamics from the Franck-Condon region to the first CI would correspond to rapid passage down steep gradients, with little dispersion of the wavepacket along other coordinates. Once on \(S_{1}\), the topology would cause the wavepacket to split, with parts heading towards each CI, and there would be the possibility of more complex dynamical behaviour. Away from the CIs, the non-adiabatic coupling of the \(S_{2}\) and \(S_{1}\) states is strong over large regions of the nuclear configuration hyperspace, with significant interaction between the states even at the equilibrium geometry - for example, the "dark" \(2^{1}A_{g}\) state is actually found to carry non-negligible oscillator strength due to non-adiabatic coupling with the bright state. The strongly coupled nature of these states and short timescales involved suggests that butadiene can be considered as something of a prototypical, perhaps even limiting or pathological, case for rapid dynamics; as such butadiene is a good exemplar of a system where the information available from frequency resolved measurements is very limited,2 but the speed and complexity of the dynamics make time-resolved measurements technically demanding and difficult to interpret.
Footnote 2: For instance ref. discusses the experimental absorption spectra, and the lack of information obtainable from the very diffuse bands observed (with just three broad features visible over the 44000 - 51000 cm\({}^{-1}\) region studied); ref. provides details of _ab initio_ calculations of the optical spectra which require the inclusion of a phenomenological dephasing constant to match the experimental data, again indicating a gap in the understanding of the radiationless relaxation of the excited state.
Experimental measurements have probed the projection of these dynamics onto the time-resolved ion yields and time-resolved photoelectron spectrum. In the former case fast time-constants, \(<\)50 fs, were determined for the decay of the parent ion signal, and delayed onset of fragmentation was also observed; in the latter case, as well as a rapid decay of the photoelectron yield, a fast shift in the vertical ionization potential and a photoelectron band with little structure were observed. Recent work, in which the TRPES was measured in coincidence with the mass spectrum, has demonstrated the separation of the photoelectron spectra correlated with the \(S_{2}\) and \(S_{1}\) states due to the ionization correlations with \(D_{0}\) and \(D_{1}\) ion states respectively; the \(D_{0}\) state is stable, yielding parent ion signal only, while the \(D_{1}\) state fragments, thus producing a very different mass spectrum. In this way coincidence measurements have been able to help disentangle the TRPES data by providing complementary, correlated observables. The same photofragment coincidence technique has also been applied to strong-field ionization of butadiene in order to probe multi-electron effects. Measurement of the TRPES and time-resolved mass spectrum (TRMS) in coincidence provides a 3D dataset, with the additional dimension of ion time-of-flight and, depending on the details of the measurement, may also provide information on the kinetic energy release of the fragments. One might hope, therefore, that a full, 7D, coincidence imaging experiment provides a rich enough dataset to discern some specific, possibly mechanistic, details of excited-state dynamics in molecular systems, even in the case of very fast dynamics exemplified by butadiene.
In order to explore some of these issues we present here our recent experimental results, focussing on a presentation of a full dataset to illustrate the richness of the 7D coincidence measurements. Along with the data, we include a brief description of our coincidence imaging apparatus, and a qualitative analysis of the data, with a focus on the TRPADs. In future publications, aspects of our apparatus will be discussed in fuller detail, and a more detailed interpretation of the experimental data and comparison with recent ab initio dynamics calculations will be made.
## II Experimental
Time-resolved pump-probe measurements were carried out with sub-40 fs UV pulses, \(\lambda_{pump}\) =200 nm (6.20 eV) and \(\lambda_{probe}\) =266 nm (4.66 eV). Butadiene (1% in He) was introduced to the interaction region via a 1 kHz pulsed valve (Even-Lavie, 150 \(\mu\)m diameter conical nozzle) with a stagnation pressure of \(\sim\)5 bar. Full 7D coincidence measurements were performed on a coincidence imaging spectrometer (CIS). Details of the apparatus are provided in the following sections.
### Optical set-up
Short pulse infrared light (\(\lambda\)=800 nm, 35 fs, 1 kHz repetition rate) was generated by a standard titanium-sapphire based regenerative amplifier system, followed by a single pass amplifier stage (Coherent Legend Elite Duo). Approximately 700 \(\mu J\) of the output was used to pump a 3\({}^{rd}\) (3\(\omega\)) and 4\({}^{th}\) (4\(\omega\)) harmonic generation scheme, based on sum-frequency mixing in thin BBO crystals.
Calcium fluoride prism pairs were used to compress the generated UV pulses, and compensate for the dispersion of transmissive optics in the beam paths (harmonic separation and recombination optics, \(\lambda/2\) plate for 3
Spectra and autocorrelation traces for the (a) 266 nm probe pulses, (b) 200 nm pump pulses. Spectra show central wavelength and FWHM for the Gaussian fit. Autocorrelation traces show \(\tau_{XC}\) (the FWHM of the Gaussian fit), and the corresponding pulse \(\tau_{p}\) duration assuming a Gaussian envelope.
The output pulses were measured directly with an autocorrelator based on 2-photon absorption, and the pulse durations in the experimental chamber were additionally confirmed via the cross-correlation feature obtained from fitting the TRPES data. Output pulses on the order of 35 \(\pm\) 3 fs were measured.
Control over pump-probe delay was achieved via a high-precision linear motor stage (Newport XML210), for the data reported herein steps of 10 fs were used. The beams were recombined in a collinear geometry, sent through a spatial filter to clean the mode and increase the beam diameters by a factor of 2, and loosely focussed into the interaction region via an \(f\)= 1 m focussing mirror. Beam diameters of the UV at the focussing mirror were 4 mm at 3\(\omega\) and 2.6 mm at 4\(\omega\). Focal spot sizes in the interaction region were estimated to be \(\sim\)85 \(\mu\)m at 3\(\omega\) and \(\sim\)100 \(\mu\)m at 4\(\omega\). To avoid multi-photon effects, and minimise scattered light signal from the 4\(\omega\) light, relatively low pulse energies of \(\sim 200\,nJ\) at 3\(\omega\) and \(\sim 10\,nJ\) at 4\(\omega\) were used, corresponding to peak intensities of \(\sim\)1.9x10\({}^{11}\) Wcm\({}^{-2}\) and \(\sim\)7x10\({}^{9}\) Wcm\({}^{-2}\) respectively.
### Coincidence imaging spectrometer
The coincidence imaging spectrometer (CIS) is illustrated in An overview is provided in the following, and a more detailed description may be found in ref.. The imaging of both electrons and ions is based on a flat-field, Wiley-McLaren type geometry. A flat-field design was chosen to allow for a large turn-around time spread of photoelectrons along the time-of-flight (\(ToF\)) axis (labelled as the \(z\)-axis in figure 2), as compared with a VMI type configuration, and also to provide the most direct mapping of (\(x,y,ToF\)) data to initial velocity vector, thus allowing for a simple calibration and data back-transformation procedure. To create flat fields, the interaction region is enclosed by grids in the vertical direction; the open apertures in the horizontal plane, required to admit the supersonic molecular beam and laser beams, are removed as far from the interaction centre as possible and minimised in spatial extent in order to avoid aberrations and fringing fields. The bottom grid is biased (typically +10 to +20 V) relative to the top grid (0 V) in order to extract photoelectrons towards the lower flight tube. After the photoelectrons have cleared the interaction region a high voltage pulse (on the order of -0.5 to -0.8 kV, \(<\)15 ns rise time, supplied by a HVC-1000 pulser unit from GPTA) is applied to the top grid to eject the photoions to the upper detector. The photoion flight tube also contains an acceleration region, and an Einzel lens assembly (not used in this work), for further control over the ion imaging spectrometer conditions. The pulsed valve (not illustrated) is situated in a source chamber, separated from the CIS chamber by a conical molecular beam skimmer (Beam Dynamics, 1 mm aperture), approximately 45 cm from the interaction region. The molecular beam is further skimmed by an adjustable slit, discussed below. Partly shown in are the baffle arms, designed to limit the background photoelectron signal from scattered light, a particularly severe problem from the 200 nm pump pulse. Further details of the scattered light problem and the baffle system employed here will be given in a future publication.
Both detectors are comprised of a triple stack of 40 mm diameter MCPs and delay line anodes (Sensor Sciences LLC). Time-of-flight measurement is made via a pick-off from the MCP front face, and (\(x,y\)) position data is obtained from the delay line signals. The ion detector is offset slightly (15 mm) along the molecular beam axis to compensate for the initial translational velocity along this axis. The data read-out and storage is handled by an electronics rack comprising NIM modules on a CAMAC bus backbone, for full specifications see ref., with final output of the processed signals to a PC. The timing-resolution, averaged over the full detector area, is \(\sim\)200 ps, limited by the pulse propagation characteristics over the detector face; the spatial resolution is \(<\)80 \(\mu\)m for both \(x\) and \(y\) axes of the electron detector, and \(<\)60 \(\mu\)m for the ion detector. The PC also controlled other experimental variables such as the pump-probe delay, the shutters installed in the pump and probe beam paths, and dwell times at each delay.
Schematic of the coincidence imaging spectrometer, showing key aspects of the design. The inset shows details of the interaction region and electron flight region, including electron and ion trajectories; dashed lines represent the initial velocity vectors and the solid lines the paths in the guiding fields. Dimensions shown are in mm, and the axis definitions shown are used throughout this work.
Pump-only and probe-only signals were measured at each delay for dwell times of \(1/N\) less than the pump-probe signal, where \(N\) was the total number of delays set. Each set of \(N\) delays defines one experimental cycle, and the final dataset presented here comprises \(\sim 200\) cycles; cycles were kept short (\(\lesssim\)30 minutes of real-time) to minimise the effects of any drifts during the course of a cycle.
Instrument resolution is determined both by the overall mechanical design, optimised for low-energy photoelectrons and photoions (\(\lesssim\)2 eV), the acquisition electronics and the fields applied for a given measurement. Adjustment of the extraction fields provides experimental control over \((x,y)\) resolution via the choice of the energy cut-off (image size), in essence higher extraction fields increase the dynamic range of the image at the cost of energy resolution, while lower extraction fields maximise energy resolution over a reduced dynamic range. This consideration is common to VMI and other imaging techniques. In 3D imaging, in common with 1D time-of-flight measurements, the temporal resolution (\(z\)-axis) is also affected by the choice of extraction fields due to the influence of the fields on the turn-around time-spread of the particles, hence temporal spread at the detector. Under typical operating conditions, with photoelectrons up to 1 eV extracted with a +13.5 V field, the photoelectron energy resolution \(\Delta E/E\) was calculated from SIMION simulations to be 1% (10 meV) for the \(x\)-axis, 10% (100 meV) for the \(y\)-axis and 3% (30 meV) for the ToF axis. Similar figures of 1%, 15% and \(<\)1% were calculated for the ion \((x,y,ToF)\) resolution at 0.78 eV with a 200 V extraction field and 600 V acceleration field. In practice the \(y\)-resolution was improved from these design stage simulations by a reduction of the ionization volume along the laser axis, reducing \(\Delta y\) from the 2 mm assumed in the resolution figures provided above. This reduction was achieved by creating a pseudo-1D molecular beam source along the \(y\)-axis, via the use of a piezo-actuated razor blade slit, mounted after the skimmer which separates the source and spectrometer chambers. In the experiments detailed herein the slit width was set at 400 \(\mu\)m, providing a significant improvement on the \(y\)-axis resolution and allowing for energy slices of 100 meV to be used throughout the photoelectron data analysis over the full 2 eV energy range imaged.
### Data analysis & calibration
For each event - photoelectron and/or photoion detection - a data record is stored, consisting primarily of position and time-of-flight \((x,y,ToF)\), and anode charges \((Q_{x},Q_{y})\), for the ion and electron. Also stored for each data record are various indices allowing correlation of the event with pump-probe configuration, pump-probe delay, experimental cycle etc. In the current electronics configuration only a single electron and/or ion event is recorded per laser shot, although in principle multi-hit operation of the ion detector is possible and only limited by the dead-time of the detector, as defined by the time taken for pulses to clear the delay lines (\(\lesssim\)15 ns). The maximum count rate for electrons is therefore the limiting factor, and is the same as the repetition rate of the laser (1 kHz), although for operation in a true coincidence regime lower count rates may be required to limit false coincidence events to a reasonable level. In the experiments presented here the count rates were kept to \(\lesssim\)100 Hz for the dominant parent ion channel, which should limit false coincidences to \(<\)10%. It is interesting, although unsurprising, to note that higher total count rates are permissible in a channel-resolved experiment as compared to a single channel case. For example, the majority of background electrons from scattered light appear at early \(ToF\) relative to the main signal electrons, so are temporally resolved and do not contribute to false coincidences. A more thorough statistical analysis of the multi-channel case will be presented in another publication.
The data hypercube obtained experimentally may be filtered along any or all dimensions in order to examine correlations, throw out bad data, retrieve various mappings of the data and so on. For example, charge histograms allow for the determination of a window of good events, defined as single hit events, and the rejection of bad events where two or more hits were registered within the delay line pulse transit period, such bad events have higher \((Q_{x},Q_{y})\) than single hit events. Correlation of photoelectrons with photoions of high translational velocity eliminates signal from ionization of background gas (see section III.1); correlation with a given mass provides fragment-resolved data, and so forth.
Calibration and backtransformation of the data is performed via a three-step process: the image centre \((x_{0},y_{0},z_{0})\) is defined, either by manual inspection, or taken as the peak in the \((x,y,ToF)\) histograms; the \((x,y)\) position bins are converted to mm positions from the image centre, based on a (static) look-up table calibration, and converted to velocities \((V_{x},V_{y})\) by making use of the measured time-of-flight; a \(ToF\) look-up table, based on numerical trajectory calculations which account for the voltages applied, is computed and applied to convert the \(ToF\) data to \(V_{z}\). Because of the flat-field arrangement the trajectory calculations are straightforward and not computationally demanding.
Once converted to velocity space, energy and angular data can readily be extracted from the dataset. Obtaining these integrated observables is again just a case of creating histograms of the various quantities of interest, combined with filtering as described above, to provide maps of observables for given regions of the data hypercube. The histogram of events with energy (\(E=\frac{1}{2}m\mathbf{V}^{2}\)) versus time delay (\(t\)) provides the 2D map (\(E\), \(t\)) - the TRPES - with bins of width \(\Delta E\) and \(\Delta t\). Conversion of the data to spherical polar coordinates \((E,\theta,\phi)\) allows binning into volume elements \(\Delta E\sin\theta\Delta\theta\Delta\phi\) for each delay \(t\), to create 3D maps of the energy and angular dis tributions of events. Although the full, quasi-continuous, 3D distribution may be visualised as a set of isosurfaces or projection planes from these maps, the data is perhaps represented most tractably as a series of TRPADs at selected energy and time slices, with angular distribution \(I(\theta,\phi;\,E,\,t)\). For cylindrically symmetric distributions, further integration over \(\phi\) can also be performed leading to a reduced form \(I(\theta;\,E,\,t)\). Low event number datasets also benefit from this integration because without integration or smoothing the 3D maps may be too sparse or noisy for further analysis.
The extracted angular distributions can also be described phenomenologically by \(\beta_{LM}\) parameters,
\[I(\theta,\phi;\,E,\,t)=\sum_{L,M}\beta_{LM}Y_{LM}(\theta,\,\phi) \tag{1}\]
This distribution is general to any scattering system but the expansion is constrained, by the symmetry of the experiment, to only certain values of \(L,\,M\); in the case of the pump-probe experiment considered here, with total absorption of two photons, linearly polarised light and pump and probe polarisation vectors parallel, the final angular distributions contain terms \(L=0,\,2,\,4\) and are cylindrically symmetric (\(M\)=0) with symmetry axis defined by the laser polarisation. For the photoelectrons, the \(\beta_{LM}\) contain details of the ionization dynamics, and are complex functions of molecular geometry and photoelectron energy. For the photoions, the \(\beta_{LM}\) contain details of the fragmentation dynamics (although the expansion is not usually written in this exact form, see for example refs.).
## III Results & Discussion
In this section we present a complete dataset of photoelectron and photoion measurements. An overview of the data is first given, providing time-integrated 3D visualizations of the full datasets. Low dimensionality (TRPES, TRMS) and high dimensionality (TRPADs, correlated data) maps extracted from the data are then presented, followed by a brief discussion of the results in the context of gaining insight into the underlying molecular dynamics.
### Overview - visualizing multi-dimensional measurements
The most direct way to begin considering the data is via 3D maps of the electron and ion signals. Here we present only the time-integrated data in this form, although visualization of time-sliced data as 3D maps is also a useful technique for qualitative analysis. The 3D maps provide direct information on the performance of the instrument, a quick check for artefacts, some information on the shape of the data in terms of complexity, ion fragment yields, energy spectra and so on; they represent a framework within which to begin a more detailed analysis of the data.3
Footnote 3: Figures 3 & 4 are interactive in some versions of this manuscript. Interactive versions, and source data, are also available at [http://dx.doi.org/10.6084/m9.figshare.106343](http://dx.doi.org/10.6084/m9.figshare.106343).
The overall 3D photoelectron distribution obtained (integrated over \(t\)) is shown in A few of the major features in the dataset are labelled. Of particular note is the signal along the laser axis leading to blurring in this direction. Although the interaction region along the laser axis is minimized by use of a pseudo 1D molecular beam, as discussed in section II.2, ionization of background gas along the laser axis cannot be prevented. However, in a coincidence measurement this background signal can be removed at the analysis stage by filtering the electron data for coincidences with parent ions in the central ion spot (see for the corresponding ion distribution).
Electron imaging summary. One quadrant of the raw, time-integrated 3D electron data is shown as nested isosurfaces. The 2D image planes show the \((x,y)\), \((y,z)\) and \((x,z)\) projections of the raw data. For the position data, 1 bin \(\approx\) 5 \(\mu\)m. This is an interactive figure in some versions of this manuscript; the interactive version, and source data, is also available at [http://dx.doi.org/10.6084/m9.figshare.106343](http://dx.doi.org/10.6084/m9.figshare.106343).
The asymmetry is therefore ascribed to a combination of possible effects: slight inhomogeneities in the extraction fields, especially near the grids, and perturbation by weak external fields. Such effects could perturb the photoelectrons, and would become more significant for longer interaction time-scales. A central artefact is present around \((V_{x},V_{y})=0\) and to long \(ToF\)s, the source of this artefact is unknown, but such structure could arise from the creation of metastable autoioinzing states with ns lifetimes (e.g. high-lying Rydberg states). The strong probe-only signal, which appears in a narrow energy band \(E<0.3\) eV, is clearly visible near the centre of the distribution and peaked along the laser polarisation axis. For \(E>0.3\) eV all electrons are from the pump-probe signal and, due to the rapid dynamics resulting in a rapid shift of the IP (see figure 6), appear quite diffuse in the 3D representation with no obvious angular dependence in the \(t\) integrated data.
and (c) the fragment region around 39 a.m.u. The central panel (b) shows the calibrated mass spectrum obtained by integrating over \((x,y)\), excluding the background gas signal (\(x<4000\)). The 3D ion map provides much more direct information than the equivalent mapping of the electron data, and from the figure various key features are immediately visible: the main parent ion feature, localised in \((x,y,ToF)\) and intense; satellite features to the parent ion, assigned as isotopes to higher mass (parent+1 and parent+2) and hydrogen loss channels to lower mass (parent-1, parent-2 and parent-3); weak fragment channels appearing at shorter \(ToF\), assigned as fragmentation of the parent ion; a stripe of signal to the edge of the detector, arising from ionization of background gas, i.e. molecules not entrained in the molecular beam which have zero net translational velocity, so appear along the laser propagation axis. The benefits of \((x,y)\) sensitivity to the mass spectrum are immediately obvious: the background signal is easily gated out of the analysis. Mass resolution of \(\ll 1\) a.m.u. is readily obtained under these experimental conditions (ion extraction pulse of 550 V, acceleration voltage of 150 V, mass range 0 - 230 a.m.u.), and butadiene isotope features are cleanly resolved in the mass spectrum.
Figure 4(a), and the inset to 4(b), shows an expanded view of the parent ion and nearby features. Underlying the main feature, and centred on it, is a diffuse signal with the same \(ToF\). This feature is approximately 3 orders of magnitude weaker than the main feature. This diffuse feature is assigned to a combination of parent ions which fragment after extraction from the ionization region (i.e.
Ion imaging summary (time-integrated). (a) Raw 3D data \((x,y,ToF)\) for the parent ion region. Raw 3D data, sliced along the centre of the distribution, is shown by the nested isosurfaces, and 2D projections onto the \((x,y)\), \((y,z)\) and \((x,z)\) image planes are also shown. For the position data, 1 bin \(\approx\)5 \(\mu\)m. Colour mapping shows log\({}_{10}\)(counts), with a dynamic range of 5 orders of magnitude. (b) Time-integrated mass spectrum. Obtained by integrating the data over \((x,y)\) (background gas signal excluded). The inset shows a detailed view of the parent ion region, 50 - 58 a.m.u. (c) As (a) but for fragment region centred at 39 a.m.u. This is an interactive figure in some versions of this manuscript; the interactive version, and source data, is also available at [http://dx.doi.org/10.6084/m9.figshare.106343](http://dx.doi.org/10.6084/m9.figshare.106343).
In both cases there would be a kinetic energy release relative to the direct parent ion signal, which would lead to the broadened, isotropic feature observed. The dimer peak at 108 a.m.u. was also observed (not shown in figure 4(b), normalized intensity 2x10\({}^{-4}\)), and dimer-1 and dimer+1 features were also just visible (normalized intensities \(<\)5x10\({}^{-5}\)), confirming the presence of dimers in the molecular beam. In the work discussed herein no attempt was made to investigate the diffuse feature further by, for instance, varying the gas mixture or pulsed valve timing. The diffuse nature of this feature, relative to the main parent-ion feature, means that most of these ions can be gated out of the analysis, while those that remain under the main feature hardly contribute to the total signal.
Further weak, somewhat diffuse and isotropic features are observed at earlier \(ToF\) than the parent ion. These features are assigned to hydrogen loss channels, resulting in the species \(C_{4}H_{5}^{+}\), \(C_{4}H_{4}^{+}\) and \(C_{4}H_{3}^{+}\). These channels are 2 - 4 orders of magnitude weaker than the main parent ion signal. They show some kinetic energy release, resulting in their diffuse appearance in the imaging data and a broadening of the peaks in the \((x,y)\) integrated data, and this broadening in the \(ToF\) clearly increases with the number of \(H\)-atoms lost (see inset of figure 4(b)). The isotropic nature of these distributions indicates the slow release of the fragment, relative to the timescale of molecular rotations (\(\sim\)10s of picoseconds), as would be anticipated from the complex dissociation dynamics.
Figure 4(c) shows the imaging data for the fragment region around 39 a.m.u. The main feature is assigned to the methyl loss channel, resulting in a \(C_{3}H_{3}^{+}\) fragment. The higher mass fragment is assigned to the same channel for the +1 isotope, based on the intensity ratio, but could also contain contributions from \(C_{3}H_{4}^{+}\); the lower mass channel is assigned to \(C_{3}H_{2}^{+}\). As with the hydrogen loss features, these channels show isotropic angular distributions, but with a larger spread of kinetic energy release. At longer \(ToF\)s a diffuse tail is observed. The length of the tail - stretching to the parent ion feature - indicates a relatively long lifetime for the fragmenting complex, because fragmentation must occur in all regions of the spectrometer to result in the large smearing out of the fragment mass spectrum observed. Therefore the upper limit for the timescale of fragmentation is the time taken for ion extraction from the interaction and acceleration region of the spectrometer, hence ns to \(\mu\)s timescales. Conversely, ions which appear at the fragment \(ToF\) must fragment before extraction, so cannot take longer than a few ns to fragment. The large range in timescales here suggests that multiple fragmentation pathways may play a role.
Also shown in the mass spectrum, figure 4(b), is another fragment region centred at 28 a.m.u. This is assigned as ethylene cation, \(C_{2}H_{4}^{+}\). The satellite peaks, following the same pattern as the methyl loss region, are assigned to the +1 isotope at 29 a.m.u., and further hydrogen loss resulting in \(C_{2}H_{3}^{+}\) and \(C_{2}H_{2}^{+}\). This region has no diffuse tail, indicating more rapid fragmentation than for the methyl loss channels. No lower masses were observed in the mass spectrum (down to the mass cutoff at 3 a.m.u.), indicating that all lower mass fragments were produced as neutrals.
At this level of representation, a few key aspects of the dynamics can be discerned: the electron data appears relatively structureless, hence spectrally broad and possibly varying rapidly as a function of \(t\); the ion data shows several fragmentation products, and some limits to the timescales of fragmentation for different channels can be intuited; the fragment angular distributions appear near isotropic, although quantitative analysis is required to make this conclusion definitive.
### Low-dimensionality mappings: time-resolved mass spectrum and photoelectron spectrum
The next stage in analysis complexity is low-dimensionality mappings: here we consider 2D mappings of the dynamics.
Time-resolved ion yields for a selection of the mass channels observed. (a) Major ion channels. The parent ion signal is scaled down by a factor of 20 for plotting purposes. (b) Minor channels.
time for each mass channel; second, the time-resolved photoelectron spectrum, the mapping of electron yield vs. time for each kinetic energy channel.
To obtain this representation, the data was integrated over the \((x,y,ToF)\) coordinates for each of the features of interest at each pump-probe delay \(t\), and was converted from raw counts to count rates to allow for the correct weighting of the pump only and probe only signals, which could then be subtracted from the full pump-probe signal.
The parent ion signal is observed to rise with the cross-correlation, plateau and then fall with a Gaussian tail. The diffuse part of the parent ion signal shows a very similar response, as do the isotope peaks. The major hydrogen loss channel, \(C_{4}H_{5}^{+}\), rises at later \(t\), and peaks around 20 fs after the parent ion. The minor hydrogen loss channels, the parent-2 and parent-3 features shown in figure 5(b), both peak at around 30 fs after the parent ion. Therefore, the data indicates that evolution on the excited state of around 20 - 30 fs is required before these fragmentation channels are open. Similarly, the fragment channels assigned to \(C_{3}H_{3}^{+}\) and \(C_{2}H_{4}^{+}\) show peaks around 20 fs after the parent ion. In all cases the peak shapes are similar, with non-Gaussian tails. Assuming that fragmentation only occurs on the \(D_{1}\) surface, the timescales here indicate that direct ionization to \(D_{1}\) is possible very rapidly, and indicates a significant lowering of the vertical IP from the Franck-Condon region. At the ground state equilibrium geometry \(D_{0}\) and \(D_{1}\) lie at 9.07 eV and 11.39 eV respectively, so are separated by \(\sim\)2.3 eV, with \(D_{1}\) lying 0.53 eV above the available 1+1\({}^{\prime}\) photon energy of 10.86 eV. Hence the observed dynamics suggest that the vertical IP to \(D_{1}\) falls by \(\sim\)0.5 eV in \(\sim\)20 fs, and this inferred drop is very similar to the shift observed in the photoelectron signal, discussed below. Another possibility is that the cross-section for absorption of a second probe photon increases dramatically over the first 20 fs of the dynamics. Such 1+2\({}^{\prime}\) processes would provide 15.52 eV of energy, allowing population of several higher-lying cationic states, either via direct 2-photon absorption, or sequential absorption via \(D_{0}\).
The low laser fluences used experimentally suggest that all observed signals arise from 1+1\({}^{\prime}\) pump-probe processes, conversely the known appearance energies of the fragments (\(>\)11.3 eV) suggest 1+2\({}^{\prime}\) processes, may be energetically required for fragmentation to occur.4 In order to check more carefully for 1+2\({}^{\prime}\) processes the time-resolved ion yields were also extracted in coincidence with photoelectrons of energy 0 - 1.8 eV. This produced essentially the same traces as shown in figure 5, except slightly noisier. This does not completely rule out 1+2\({}^{\prime}\) processes, which could also produce photoelectrons in this energy region by population of cation states with internal energies of \(>\)13.72 eV, or via the sequential process of ionization and subsequent excitation of the cation. A careful probe power study would be required to firmly answer this question experimentally, but this remains for future work.
Footnote 4: Experimentally these studies could only probe the ground state minimum geometry, so do not rule out the appearance of fragments at lower energies as a function of nuclear coordinates, which would only be limited by the asymptotic energy of the fragments. However, the energetics of the dissociation pathways, including transition states and final products, were also considered computationally in ref., and the theoretical results also indicate that the observed fragments are not energetically accessible via 1+1\({}^{\prime}\) processes.
In summary, the fragment yields peak around 20 - 30 fs after the parent ion, indicating rapid dynamics lead to the opening of the observed fragmentation channels. Additional filtering of the data for coincidences with electrons in the 1-photon ionization region suggested, but cannot rigorously confirm, that there is no significant 2-photon ionization contribution. In a future publication, an extended analysis of the time-resolved ion data will be made in order to examine the kinetic energy release spectra of each fragment. Combined with the available fragment energetics data this information may be sufficient to accurately determine mechanistic details of the various dissociation pathways.
We next consider the (\(E,t\)) mapping of the electron data, the TRPES, shown in The TRPES shows an energetically broad photoelectron band up to the 1-photon cut-off at \(\sim\)1.8 eV. The low-energy cut-off at \(\sim\)0.3 eV corresponds to the region of probe-only signal. The main feature at \(\sim\)1.2 eV follows the cross-correlation
TRPES mapping. (a) Raw counts per (\(E\), \(t\)) bin. The dashed line follows the tilt of the band maximum, highlighting the chirp of the signal, with gradient \(\sim\)20 fs/eV. (b) As (a), but normalized to the maximum counts for each \(E\) bin in order to show the temporal behaviour at each energy independent of total counts. The second dashed line follows the low edge of the signal, showing the slight broadening as a function of energy, with gradient \(\sim\)30 fs/eV.
This stripe is slightly sloped or chirped (this is especially clear in the energy normalized representation in figure 6(b)), with the onset time varying as a function of energy. The chirp appears close to linear over the observed energy region, with a gradient of 20\(\pm\)10 fs/eV. The temporal width of the band increases slightly as a function of energy (shown by the second dashed line in figure 6(b)), and the trailing edge of the band follows a gradient of approximately 30\(\pm\)10 fs/eV, although there is increased, non-linear, broadening below \(\sim\)0.4 eV. For the linear region the broadening of the band is therefore \(\sim\)10 fs/eV; a higher temporal resolution measurement - smaller \(\Delta t\) - would provide a more accurate figure. Apart from the slope of the band there is little structure observed, and most of the population rapidly leaves the observation window of the measurement (on the order of \(\tau_{xc}\), \(\sim\)60 fs). This data is very similar to previous TRPES studies, but with shorter \(\tau_{xc}\) and obtained for only a small set of delays in order to ensure good statistics for the TRPADs. Although the clipping of the data along the temporal coordinate makes rigorous determination of the lineshapes impossible for the peak of the signal in this dataset, based on the previous TRPES data the temporal profile of the signal is assumed to be approximately Gaussian for most, if not all, energies (see also the lineouts in figure 8), with a small (few percent of the total counts) non-Gaussian tail outside of \(\tau_{xc}\).
In terms of the dynamics, the delayed onset as a function of energy is phenomenologically consistent with the picture of fast wavepacket propagation on a steep potential energy surface. Such motion would map primarily to the vertical IP (assuming that the excited state and ionic ground state potential energy surfaces are not topologically identical), hence the kinetic energy of the observed photoelectrons. The TRPES provides information on the speed of the IP change, and is consistent with the appearance of fragmentation channels with a delayed onset which require a significant drop in the vertical IP to \(D_{1}\) relative to the Franck-Condon region. Conversely, wavepacket dynamics which cause little change to the IP would be responsible for signal observed in a given energy region at long delays, arising from population which remains within the observation window of the measurement (although such population may still trace complex trajectories in energy space). Since the signal in all regions outside of the cross-correlation time-scale is small, it is clear that wavepacket motion along these coordinates is a very minor contribution to the dynamics. The data also shows that the observed signal stays near Gaussian for all energy slices, again indicating that there is little dispersion of the wavepacket.
In broad terms, the ion and electron 2D data give some insight into the wavepacket dynamics, with the general characteristics apparent and some clocking of this motion possible. The observed energetic shift in the photoelectron signal is consistent with the vertical IP drop required to access \(D_{1}\), hence observe fragmentation, as already inferred from the ion data. In order to assemble a more detailed picture we next consider the additional information available from the TRPADs.
### High-dimensionality mappings: time-resolved photoelectron angular distributions
In order to extract TRPADs the data is calibrated and rebinned in polar coordinates as detailed in section II.3 to give the intensity (counts) per 4D volume element \(\Delta\phi\sin\theta\Delta\Theta\Delta E\Delta t\), denoted \(I(\theta,\phi;\,E,\,t)\). The choice of binning is, naturally, limited by the experimental time-steps and instrument resolution; however coarser binning can be used in order to improve the statistics at the loss of resolution. For the data presented here \(\Delta t=10\) fs and \(\Delta E=0.1\) eV, and in all cases shown here the data was also integrated over \(\phi\) which, due to the cylindrical symmetry of this experiment, results in no loss of information but does improve the statistics per \(\theta\) bin. To obtain the cleanest possible TRPADs the upper hemisphere electrons were discarded at the cost of a factor of two in counts. The data was also filtered for coincidences with the main parent ion feature. Because the parent ion dominates the signal, this filtering was not required to obtain TRPADs correlated with a given ion channel (although, more generally, could be used in this way), but did serve to remove all signal from background gas and scattered light, resulting in cleaner TRPADs albeit at the cost of total counts. shows an example of the TRPADs obtained in this way for four energy slices. In this representation, the areas of the TRPADs are normalised to unity in order to allow comparison of the form of the PADs independent of total counts. The error bars and scatter of the data points should therefore be used as a guide to the statistical significance of the extracted PADs over the various energy and time slices shown. The solid lines show a fit to an expansion in spherical harmonics, as defined by eqn. 1.
From the data it is immediately clear that the TRPADs exhibit complex behaviour. The TRPADs change rapidly as a function of \(t\), with changes on the order of 10 - 20 fs apparent. Because the laser pulses used in this work were around 35 - 40 fs in duration, there is already significant temporal blurring in the measured data. Despite this, the changes are still clear and unambiguous in the data. For comparison, TRPADs extracted for the probe only background (figure 7(a)) show no significant changes temporally beyond the signal noise, and consequent statistical variation of the fit, signifying that no temporal artefacts are present in the raw data or introduced via the data processing. The scatter in the data is worst at the poles of the distribution due to the \(\sin\theta\) normalisation factor which serves to amplify noise near the poles; conversely the equatorial region shows very little variability. There is some asymmetry present in the data which is ascribed to a combination of noise (scatter) and detector artefacts/inhomogeneities. For this latter reason the form of the extracted PADs (as defined by the fitted \(\beta_{LM}\) parameters) may not be highly accurate, but the results do have high precision and reproducibility as shown by figure 7(a), so any relative temporal changes observed should be reliable and robust within the statistical uncertainty.
One challenge of high-dimensionality datasets is the presentation and/or reduction of the data to a more tractable form to allow for pattern recognition at a phenomenological or quantitative level. For the TRPADs one can reduce the full dataset to the fitted TRPADs, which can then be represented as \((E,I(\theta))\) or \((t,I(\theta))\) surfaces, or maps of \(\beta_{LM}(E,t)\).
TRPADs, \(I(\theta;t,E)\), extracted from the dataset as detailed in the main text. (a) Probe only data, 0.1 - 0.2 eV. (b) - (d) Pump-probe data for 0.7 -0.8 eV, 0.6 - 0.7 eV and 0.5 - 0.6 eV respectively. Data points are shown with statistical error bars, solid line shows fits to eqn. 1 with \(L=0\), 2, 4 and \(M=0\).
Such representation, as a function of energy or time, readily allows for comparison of the evolution of the shape of the PADs, although visual information about the quality of the raw data and fit fidelity are lost, so care must be exercised when drawing conclusions from such maps. For the TRPADs shown in figure 8, the comments pertaining to can be reiterated, namely that the TRPADs for the different energy regions are significantly different, and show complex temporal evolution. More specifically, for the 0.5 - 0.6 eV region, the observed TRPADs show a 4-lobed structure (significant \(\beta_{4,0}\)) at \(t=-10\) fs, which evolves to a 4-lobed structure with a different orientation at \(t=10\) fs, returns to near its initial shape at intermediate \(t\), and finally shows fast oscillations at \(t>60\) fs, although these later oscillations may not be reliable due to the low statistics in this region. The colour mapping on the surface plots emphasises the evolution of the signal intensity along the equator and at the poles of the TRPADs; the corresponding \(\beta_{LM}\) plots show that both \(\beta_{2,0}\) and \(\beta_{4,0}\) change significantly as a function of \(t\), with \(\beta_{2,0}\) displaying two peaks in the range \(20\leq t\leq 70\) fs. The \(\beta_{4,0}\) trace shows a minimum at \(t=40\) fs and a maximum at \(t=60\) fs, which coincide with a mimimum and maximum in the \(\beta_{2,0}\) trace, but at earlier times (\(t<40\) fs) and later times (\(t>60\) fs) the behaviour does not appear to be directly correlated to the \(\beta_{2,0}\) trace.
For the 0.6 - 0.7 eV energy slice the picture is quite different. For the first time step the PAD again shows a 4-lobed structure, but with intensity peaked at the poles and equator, as opposed to at \(45^{\circ}\) as per the 0.5 - 0.6 eV window. This corresponds to a large and positive \(\beta_{4,0}\), compared with a large and negative \(\beta_{4,0}\) for the lower energy slice. The \(\beta_{4,0}\) value goes negative, with small magnitude, for the following time slices, and appears to show a slight oscillation with minima at \(t=0,\ 40\) fs and a peak at \(t=20\) fs; the \(\beta_{2,0}\) parameter shows a correlated oscillation but centred around a mean value of \(\sim 0.5\). In the surface plots, this oscillation appears as a slight breathing of the TRPADs, with the largest changes at the poles. At later times, \(t>50\) fs, more complex behaviour is observed, with a significant beating around the equator of the distributions as well as at the poles.
For the 0.7 - 0.8 eV energy slice the behaviour is again different, with much less variability in the observed TRPADs over the peak of the signal. The \(\beta_{2,0}\) value decreases gradually from the local maxima at \(t=0\) fs until \(t=50\) fs, then increases gradually to \(t=70\) fs. The \(\beta_{4,0}\) trace shows correlated local maxima at \(t=0,\ 70\) fs, but much more variability between these peaks.
Fitted TRPADs, \(I(\theta;t,E)\), as per figure 7, represented as polar surface plots. White lines show the fits at 10 fs intervals, the surface and colour map interpolate between these discrete measurements. Lower panels show the fitted \(\beta_{20}(t)\) and \(\beta_{40}(t)\), as well as the counts, for each energy region.
However, the fact that the observed \(\beta_{2,0}\) dips at the same time (\(t=80\) fs) over all three energy slices, which were analysed independently, suggests this behaviour is genuine and not the result of random noise. In terms of the form of the TRPADs, the oscillations in the \(\beta_{L,M}\) describe significant oscillations which include large changes to the photoelectron flux around the equator of the distributions. Because the equatorial region is statistically more reliable than the noisier poles this again suggests that these observations are genuine.
This representation of the TRPADs is essentially one step further removed from the raw data, so again should be used in concert with plots showing the raw data before drawing firm conclusions, but also provides the most reduced and tractable form of the measurement. In this format, the 3 energy regions discussed in detail in the preceding can be readily compared. It is clear that the higher energy region correlates to larger \(\beta_{2,0}\) values, which change little over the main part of the signal, while the lower energy region shows marked oscillations; similar oscillations extend to the lower energy slices, with peaks around \(t=20\), 70 fs. At later times, \(t>60\) fs, the trend is for reduced magnitude \(\beta_{2,0}\), particular over the main signal in the 0.9 - 1.3 eV region. These times are outside of the cross-correlation of the laser pulses, so should be representative of only the parts of the excited state wavepacket which move orthogonal to the steep gradients on the potential energy surface responsible for the speed of the dynamics and the rapid IP change observed (i.e. population which remains in a given region of configuration space for a longer time than the main part of the wavepacket). The \(\beta_{4,0}\) map shows much more oscillatory behaviour, which appears to show no obvious correlation with energy or time, except for the trend towards larger positive values at \(t\geq 60\,fs\). For higher time resolution data it would be feasible to Fourier Transform this data to extract the frequency content, but for the dataset shown here the limited number of delays results in only a crude frequency spectrum of little utility.
In summary, the TRPADs presented here contain a plethora of information, and show very complex temporal evolution, in contrast to the TRPES which provides little information on the temporal evolution of a given energy slice over the probe pulse envelope, nor provides an observable as sensitive to the underlying dynamics. The challenge for the experimentalist is to determine whether the richness of the measured TRPADs can be interpreted in terms of the underlying dynamics without recourse to detailed theoretical treatments, that is to say without _a priori_ knowledge of the underlying molecular dynamics or a full _ab initio_ treatment of the dynamics and ionization. These points are discussed further below (section III.5).
### High-dimensionality mappings: correlated observables
The advantages of measuring in coincidence have already been discussed in terms of removal of background or other unwanted signals from the data hypercube. Naturally a further advantage is the ability to look for ion-electron correlations, and retrieve minor channels which would otherwise be inaccessible; for instance, the photoelectron spectra of the fragment channels, which would not be distinguishable in a non-coincidence measurement.
As expected, the parent+1 channel has an identical spectrum. However, the diffuse part of the parent ion signal has a very different spectrum, consistent with a different ionization process. Similarly, the fragmentation channels show different spectra; in all cases the spectrum is broad, and there is a drop or even disappearance of the 1 eV peak seen in the parent ion channel. These spectra therefore provide additional information towards understanding the excited state dynamics, with the spectra providing a fingerprint of different ionization channels. Here, the suppression of the cross-correlation peak in the fragment channels is consistent with the delayed opening of these channels, as observed in the time-resolved ion yields.
Furthermore, the similarity between the channels suggests ionization occurs from a similar region of configuration space in all cases. This is consistent with the similarity of the rise times and fall times observed in the time-resolved fragment yields, which intuitively suggests only a single dynamical pathway, shared by all the fragment channels, hinting at localization of the excited state wavepacket in configuration space. The similarity of the time-scales of the fragment channels, considered in light of the similarities in the correlated photoelectron spectra, therefore suggests that the fragmentation pathway of the ion is very sensitive to the form of the wavepacket at the time of ionization and, possibly, bifurcation on the ionic state(s) may lead to channels with apparently similar temporal response, but quite different fragmentation products. This is consistent with the picture of rapid wavepacket motion, with little dispersion, on the excited state and additionally indicates complex dissociation dynamics on the ionic surfaces. In terms of the experimentally accessed ionization pathways, the observation of similar photoelectron spectra for the fragment channels is consistent with the possibility of sequential 1+2\({}^{\prime}\) processes (ionization to \(D_{0}\) followed by absorption of a second probe photon) as discussed in section III.2, but suggests that direct 1+2\({}^{\prime}\) ionization to higher lying cationic states, and 1+1\({}^{\prime}\) ionization to \(D_{1}\), are unlikely channels. Direct ionization to excited ionic states would be expected to correlate generally with different spec tra, although it is always possible that the spectra would not be significantly different - especially given the diffuse, unstructured nature of the photoelectron bands of butadiene - so this is not a rigorous conclusion.
Various other correlated mappings are available from the data, and will be explored in a future publication. For example, the fragment correlated TRPES and TRPADs, although for weak channels the latter is very demanding statistically. For photodissociation studies energy correlation spectra - maps of electon vs. ion kinetic energy - are also useful.
### Discussion - mapping dynamics with multi-dimensional measurements
The data presented herein indicates that rapid and complex dynamics are present in butadiene, consistent with earlier experimental and theoretical works. The benefit of a multidimensional dataset is clear, with complementary information available from the coincident electron and ion data. The overall shape of the dynamics, that of a near-Gaussian wavepacket which moves along steep gradients on the potential energy surface which are strongly correlated with the vertical IP, and shows little dispersion along other coordinates, emerges rapidly from the data. This picture fits both the chirped TRPES data and the delayed-onset of fragmentation channels which are assumed to be correlated with ionization to \(D_{1}\). The observation of TRPADs which show very rich behaviour on rapid timescales (\(<\)20 fs) is striking, and indicates the sensitivity of the TRPADs to the dynamics under study.
However, more specific mechanistic details, such as the nuclear motions involved or the mapping of these motions onto the TRPADs, do not seem to be forthcoming. This is in contrast to simpler cases, such as previous work on the \(NO\) dimer or \(CS_{2}\), in which the limited dimensionality of the problem enabled a more direct empirical approach, coupled with symmetry-based modelling to understand the measurements in more detail. Understanding the information conveyed by the TRPADs from larger systems therefore remains a significant challenge.
In the case of butadiene, detailed _ab initio_ dynamics
Time-integrated photoelectron spectra correlated with the major mass channels (as shown in figure 5(a)). The parent ion correlated spectrum is scaled down by a factor of 20 for plotting purposes.
Energy-time maps of the total counts (TRPES) and extracted \(\beta_{LM}\) parameters.
It is therefore very tempting to compare the measurements to these calculations, despite the fact that these calculations do not, so far, explicitly include the ionization matrix elements. For instance, figures 5 and 6 of ref. show the behaviour and timescales associated with the bond alternation coordinate (a motion involving bonds along the carbon backbone streching and compressing) and the out-of-plane twisting motions (twisting about the central CC bond and twisting of the terminal methyl groups). The bond alternation coordinate shows oscillations with a period of \(\sim\)20 fs, and large amplitude motion with changes up to 0.4 A. The twisting coordinates show rapid changes over \(\sim\) 40 fs, followed by almost constant bond angles, with some fluctuations, for the remaining 160 fs of the calculations. These motions are very suggestive of the oscillations observed in \(\beta_{2,0}(t)\) for some of the low-energy slices, and the gradual rise over \(\sim\)60 fs observed in \(\beta_{4,0}(t)\) over all photoelectron energies. However, from the data presented in ref. it is not apparent how these motions map to the IP, which is a very significant factor in drawing these conclusions more firmly. As shown in ref. such mappings may be very complex, and difficult to determine from anything but a full calculation including the ionization matrix elements.
Another conclusion from the dynamics calculations, which is likely to be more robust, is the agreement between the timescale of the population dynamics and the gross changes in both \(\beta_{2,0}(t)\) and \(\beta_{4,0}(t)\) at all energies for \(t\geq 60\) fs. of ref. shows the populations of the bright and dark adiabatic states, defined there as \(S_{1}\) and \(S_{2}\) respectively. The \(S_{1}\) population dips almost immediately from its initial value at \(t=0\), with almost equal populations of \(S_{1}\) and \(S_{2}\) at \(t\approx 25\) fs, the \(S_{1}\) population then grows again and peaks at around 50 fs, while the \(S_{2}\) population drops rapidly. Unsurprisingly, given the similarity of the timescales to those mentioned above, the non-adiabatic coupling between these states is primarily mediated by the out-of-plane twisting motions. In terms of the electronic character of the states, the early time dynamics is best considered in a diabatic picture, with the bright state \({}^{1}B_{u}\) character maintained as population switches from the \(S_{1}\) to the \(S_{2}\) surface at early times, while over the longer timescale the electronic character becomes mixed by the out-of-plane twisting, and population is transferred non-adiabatically between \(S_{2}\) and \(S_{1}\).
Given that the adiabatic states are of different electronic character one would expect a different PAD from each state; additionally, interferences between photoelectrons from the two states (different parts of a split wavepacket) may also play a significant role in the observable if electrons of the same energy are created from the two states, and this could lead to more rapid modulations in the form of the TRPADs (this is essentially the concept of the two-state model previously applied to \(CS_{2}\)), particularly in the case of symmetry-breaking leading to mixing of the adiabatic electronic state characters. With this in mind, it appears that the (adiabatic) population dynamics may have a more direct link to the observed TRPADs, with the changes at \(t\geq 60\) fs corresponding to the region where most of the population is on a single adiabatic state, and the molecular geometry, hence electronic character, has stabilised after the initial, rapid, out-of-plane twisting. Although this picture seems reasonable, without further modelling or calculations these conclusions remain somewhat tentative.
The challenge for the experimentalist therefore remains: to determine how the TRPADs, and other correlated observables, can be interpreted without recourse to detailed theoretical treatments, hence without _a priori_ knowledge of the underlying molecular dynamics or a full _ab initio_ treatment of the dynamics and ionization. While such computational approaches are very powerful, they are also time consuming, difficult, and only computationally tractable for small molecules. Naturally the same applies to a purely experimental approach to the TRPES and HRMS data, but typically the lower dimensionality of this data lends itself to only broad interpretations, such as the presence of band switching or quantum beats, which may fully describe the wavepacket dynamics in small systems or arise from only the recurrent parts of the wavepacket in more complex cases. Moreover, TRPES data is routinely interpreted via fitting of temporal profiles in what is essentially a principle component or single-value decomposition analysis. While such analysis provides a quantitative breakdown of the data into spectral and temporal functions, it makes many assumptions regarding the underlying dynamics (vis. that the observables equate directly to state populations in the standard interpretations, or at the very least behave phenomenologically in a statistical manner) and cannot be used to treat more complex dynamics, such as the kind observed here, without recourse to many time constants. The TRPADs clearly convey more information on the wavepacket dynamics, but the mapping from wavepacket to observable is inherently complicated.
Aside from the TRPADs, the full imaging data provides further correlated observables which may also be considered in light of the computational results. In particular, the time-resolved ion yields and fragment-correlated photoelectron spectra may provide a way of experimentally probing the three \(S_{1}\to S_{0}\) CIs discussed in ref., which correspond to different molecular geometries. It is likely that ionization from these different regions of configuration space (if energetically possible in the limited observation window) would lead to different photoelectron spectra and different fragmentation products from \(D_{1}\). Even in the case where direct ionization from these regions is energetically forbidden, ionization of the parts of the wavepacket on \(S_{1}\) heading towards these distinct regions of configuration space would also result in different dynamics on the ion surfaces, hence possibly lead to different fragmentation products. Further analysis of ion-electron correlations should present deeper insight into the excited state dynamics than uncorrelated ion or electron measurements alone.
## IV Conclusions & Future Work
The data presented here illustrates the richness of coincidence imaging datasets, even in the case of very rapid and complex dynamics; it also highlights the difficulty of interpretation of such data. In this work, we have focussed on the TRPADs obtained in coincidence with the parent ion channel, and made some tentative comparisons with theory. In general the full 7D data provides the best hope of understanding the excited state dynamics of molecules but, unsurprisingly, significant effort is required to gain understanding directly from the data, or with the aid of theory. Despite the depth of material presented here there are still many avenues to explore in order to obtain more complete, unified experimental and theoretical approaches to problems in excited state dynamics.
In future work we hope to continue bridging this gap between experiment and theory via a number of routes: by the more direct and explicit comparison of the experimental data with dynamics calculations, and the inclusion of observables in the calculations; by exploring further the possibilities of modelling for a qualitative/semi-quantitative treatment of the dynamics, expanding upon previous work which employed a simple 2-state model, possibly to include a basic wavepacket treatment; through further analysis of the ion fragment distributions, which have not yet been fully explored; by making further experimental measurements for aligned systems and with different pump-probe polarisation geometries in order to obtain more detailed TRPADs (in both cases more \(L\), \(M\) terms are allowed in the PADs), with harder VUV probe photons to access a larger photoelectron energy range, and with a stability enhanced set-up to allow for longer experimental runs; by further use of fragment correlated TRPES/TRPADs, so far not statistically feasible due to the dominance of the parent ion channel and consequent low total counts in these fragment channels. This multi-dimensional approach to a multi-dimensional challenge will hopefully prove fruitful for a deeper understanding of excited state dynamics in polyatomic molecules and, in particular, help to make the experimental measurement of such dynamics a more routine and insightful tool.
|
10.48550/arXiv.1301.6928
|
Probing Ultrafast Dynamics with Time-resolved Multi-dimensional Coincidence Imaging: Butadiene
|
Paul Hockett, Enrico Ripani, Andrew Rytwinski, Albert Stolow
| 812
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.