Journal of Chemical Theory and Computation最新文献

We present a directed electrostatics strategy integrated as a graph neural network (DESIGNN) approach for predicting stable nanocluster structures on their potential energy surfaces (PESs). The DESIGNN approach is a graph neural network (GNN)-based model for building structures of large atomic clusters with specific sizes and point-group symmetry. This model assists in the structure building of atomic metal clusters by predicting molecular electrostatic potential (MESP) topography minima on their structural evolution paths. The DESIGNN approach is benchmarked on the prototype Mg_n clusters with n < 150. The predicted MESP topography minima of Mg_n clusters, n < 70, fairly agrees with the whole-molecule MESP topography calculations. Moreover, the ground-state structures of Mg_n (n = 4-32) clusters generated through the DESIGNN approach corroborate well with the global minimum structures reported in the literature. Furthermore, this approach could generate novel symmetric isomers of medium to large Mg_n clusters in the size regime, n < 150, by constraining the point-group symmetry of the parent clusters. The parent growth potential (GP) of a cluster gives a measure of its parent cluster to accommodate more atoms and characterize the structures on the DESIGNN-guided path. The GP of a cluster can also be interpreted as a measure of the cooperative interaction relative to its parent cluster. Along the highest GP paths, the DESIGNN approach is further employed to generate stable Mg_n nanoclusters with n = 228, 236, 257, 260. Therefore, the DESIGNN approach holds great promise in accelerating the structure search and prediction of large metal clusters guided through MESP topography.

The light-harvesting complex II (LHCII) in green plants exhibits highly efficient excitation energy transfer (EET). A comprehensive understanding of the EET mechanism in LHCII requires quantum chemical, molecular dynamics (MD), and statistical mechanics calculations that can adequately describe pigment molecules in heterogeneous environments. Herein, we develop MD simulation parameters that accurately reproduce the quantum mechanical/molecular mechanical energies of both the ground and excited states of all chlorophyll (Chl) molecules in membrane embedded LHCII. The present simulations reveal that Chl a molecules reside in more inhomogeneous environments than Chl b molecules. We also find a narrow gap between the exciton energy levels of Chl a and Chl b. In addition, we investigate the nature of the exciton states of Chl molecules, such as delocalization, and analyze the optical spectra of LHCII, which align with experimental results. Thus, the MD simulation parameters developed in this study successfully reproduce the excitonic and optical properties of the Chl molecules in LHCII, validating their effectiveness.

Classical reaction barriers in density-functional theory are considered from the perspective of the density-fixed adiabatic connection. A 'reaction adiabatic-connection integrand', $R_{\lambda }$, is introduced, where λ is the electron-electron interaction strength, for which ${\int }_0^1{R}_{\lambda }d\lambda$ equals the barrier, meaning the barrier can be easily visualized as the area under a plot of $R_{\lambda }$ vs λ. For five chemical reactions, plots of reference $R_{\lambda }$, calculated from Lieb maximizations at the coupled-cluster level of theory, are compared with approximate $R_{\lambda }$, calculated from common exchange-correlation functionals using coordinate scaling, for coupled-cluster densities. The comparison provides a simple way to visualize and understand functional-driven errors and trends in barriers from approximate functionals, while allowing a clean separation of the role of exchange and correlation contributions to the barrier. Specifically, the accuracy of $R_0$ is determined entirely by the accuracy of the exchange functional, while the shape of $R_{\lambda }$ is determined entirely by the correlation functional. The results clearly illustrate why the optimal amount of exact (orbital) exchange in hybrid functionals differs between reactions, including forward and reverse directions in the same reaction, and hence why simply introducing larger amounts of exact exchange may not be a reliable approach for improving barriers. Instead, the shape of $R_{\lambda }$ must be captured more accurately through more accurate correlation functionals, and the numerical data presented may be useful for this purpose. Density-driven errors are then considered, and possible cancellation with functional-driven errors in barriers─noted in prior studies when Hartree-Fock densities are used─is illustrated from the perspective of $R_{\lambda }$.

We present an application of our new theoretical formulation of quantum dynamics, moment propagation theory (MPT) (Boyer et al., J. Chem. Phys. 160, 064113 (2024)), for employing machine-learning techniques to simulate the quantum dynamics of electrons. In particular, we use real-time time-dependent density functional theory (RT-TDDFT) simulation in the gauge of the maximally localized Wannier functions (MLWFs) for training the MPT equation of motion. Spatially localized time-dependent MLWFs provide a concise representation that is particularly convenient for the MPT expressed in terms of increasing orders of moments. The equation of motion for these moments can be integrated in time, while the analytical expressions are quite involved. In this work, machine-learning techniques were used to train the second-order time derivatives of the moments using first-principles data from the RT-TDDFT simulation, and this MPT enabled us to perform electron dynamics efficiently. The application to computing optical absorption spectrum for various systems was demonstrated as a proof-of-principles example of this approach. In addition to isolated molecules (water, benzene, and ethene), condensed matter systems (liquid water and crystalline silicon) were studied, and we also explored how the principle of the nearsightedness of electrons can be employed in this context.

We present flow matching for reaction coordinates (FMRC), a novel deep learning algorithm designed to identify optimal reaction coordinates (RC) in biomolecular reversible dynamics. FMRC is based on the mathematical principles of lumpability and decomposability, which we reformulate into a conditional probability framework for efficient data-driven optimization using deep generative models. While FMRC does not explicitly learn the well-established transfer operator or its eigenfunctions, it can effectively encode the dynamics of leading eigenfunctions of the system transfer operator into its low-dimensional RC space. We further quantitatively compare its performance with several state-of-the-art algorithms by evaluating the quality of Markov state models (MSM) constructed in their respective RC spaces, demonstrating the superiority of FMRC in three increasingly complex biomolecular systems. In addition, we successfully demonstrated the efficacy of FMRC for bias deposition in the enhanced sampling of a simple model system. Finally, we discuss its potential applications in downstream applications such as enhanced sampling methods and MSM construction.

The electronic structures of lanthanide monoxides (LnO/LnO⁺) and monosulfides (LnS/LnS⁺) for all lanthanide series elements (Ln = La-Lu) have been systematically analyzed with sophisticated quantum chemical calculations. The ground electronic configuration has been determined to be Ln 4fⁿ6s¹ or 4fⁿ⁺¹ for the neutral molecules and Ln 4fⁿ for the cations. The low-lying energy states resulting from spin-orbit coupling and ligand field effects have been resolved using spin-orbit multiconfiguration quasi-degenerate second-order perturbation theory calculations. The ionization energies of LnO (5.20-7.06 eV) are about 0.3-2.2 eV lower than those of LnS (5.54-9.22 eV) due to the difference in the Ln 6s and 4f orbital energies from which an electron is removed during the ionization process. The bond dissociation energies (BDEs) have been computed by the state-averaged general multiconfigurational perturbation theory and the completely renormalized coupled-cluster [CR-CC(2,3)] methods. The BDEs are highly dependent on the lanthanide elements as several factors of the lanthanides affect the bond dissociation. The calculated bond lengths and energies agree well with available experimental values and are systematically predicted for the series of lanthanide monoxides and monosulfides where experimental values are not available. Furthermore, the LS terms of low-lying energy states and their corresponding bond properties have been clarified in detail to systematize the similarities and differences of the lanthanide compounds.

We introduce an enhanced sampling algorithm to obtain converged free energy landscapes of molecular rare events, even when the collective variable (CV) used for biasing is not optimal. Our approach samples a time-dependent target distribution by combining the on-the-fly probability enhanced sampling and its exploratory variant, OPES Explore. This promotes more transitions between the relevant metastable states and accelerates the convergence speed of the free energy estimate. We demonstrate the successful application of this combined algorithm on the two-dimensional Wolfe-Quapp potential, millisecond time-scale ligand-receptor binding in the trypsin-benzamidine complex, and folding-unfolding transitions in chignolin mini-protein. Our proposed algorithm can compute accurate free energies at an affordable computational cost and is robust in terms of the choice of CVs, making it particularly promising for the simulation of complex biomolecular systems.

Molecular oxygen (O₂) is essential for life, and continuous effort has been made to understand its pathways in cellular respiration with all-atom (AA) molecular dynamics (MD) simulations of, e.g., membrane permeation or binding to proteins. To reach larger length scales with models, such as curved membranes in mitochondria or caveolae, coarse-grained (CG) simulations could be used at much lower computational cost than AA simulations. Yet a CG model for O₂ is lacking. In this work, a CG model for O₂ is therefore carefully selected from the Martini 3 force field based on criteria including size, zero charge, nonpolarity, solubility in nonpolar organic solvents, and partitioning in a phospholipid membrane. This chosen CG model for O₂ (TC3 bead) is then further evaluated through the calculation of its diffusion constant in water and hexadecane, its permeability rate across pure phospholipid- and cholesterol-containing membranes, and its binding to the T4 lysozyme L99A protein. Our CG model shows semiquantitative agreement between CG diffusivity and permeation rates with the corresponding AA values and available experimental data. Additionally, it captures the binding to hydrophobic cavities of the protein, aligning well with the AA simulation of the same system. Thus, the results show that our O₂ model approximates the behavior observed in the AA simulations. The CG O₂ model is compatible with the widely used multifunctional Martini 3 force field for biological simulations, which will allow for the simulation of large biomolecular systems involved in O₂'s transport in the body.

The photocatalytic efficiency of materials such as graphene and noble metal nanoclusters depends on their plasmon lifetimes. Plasmon dephasing and decay in these materials is thought to occur on ultrafast time scales, ranging from a few femtoseconds to hundreds of femtoseconds and longer. Here we focus on understanding the dephasing and decay pathways of excited states in small lithium and silver clusters and in plasmonic states of the π-conjugated molecule anthracene, providing insights that are crucial for interpreting optical properties and photophysics. To do this, we study the time dependence of the electronic density matrix of these molecules using a new approach that expresses the density matrix in terms of TDDFT eigenstates (ESs) of the TDDFT Hamiltonian. This approach, which involves combining linear response time-dependent density functional theory (LR-TDDFT) and real-time time-dependent density functional theory (RT-TDDFT), leads to an analysis of the electron dynamics in terms of ESs, rather than individual molecular orbital (MO) transitions as has typically been done. This circumvents the complexities and subjective biases that traditional MO-based analysis provides. We find in an analysis of the induced dipole moment in these molecules that what had previously been considered to be energy relaxation is actually dephasing associated with the eigenstates that are stationary after the excitation pulse is turned off. We conclude that the ES-basis analysis has significant potential to advance understanding of the electron dynamics of plasmonic nanomaterials, aiding their optimization for photocatalytic and technological applications.

Training accurate machine learning potentials requires electronic structure data comprehensively covering the configurational space of the system of interest. As the construction of this data is computationally demanding, many schemes for identifying the most important structures have been proposed. Here, we compare the performance of high-dimensional neural network potentials (HDNNPs) for quantum liquid water at ambient conditions trained to data sets constructed using random sampling as well as various flavors of active learning based on query by committee. Contrary to the common understanding of active learning, we find that for a given data set size, random sampling leads to smaller test errors for structures not included in the training process. In our analysis, we show that this can be related to small energy offsets caused by a bias in structures added in active learning, which can be overcome by using instead energy correlations as an error measure that is invariant to such shifts. Still, all HDNNPs yield very similar and accurate structural properties of quantum liquid water, which demonstrates the robustness of the training procedure with respect to the training set construction algorithm even when trained to as few as 200 structures. However, we find that for active learning based on preliminary potentials, a reasonable initial data set is important to avoid an unnecessary extension of the covered configuration space to less relevant regions.