How well do various QM-derived net atomic charges reproduce the electrostatic potential surrounding a material across multiple geometric conformations?

Abstract

Atom-centered point-charge models are computationally efficient and commonly used electrostatic models to build forcefields used in classical simulations of materials. To assess their performance, we evaluated various atomic charge assignment methods across the following material types: (a) organic molecules, (b) inorganic molecules, (c) heterodiatomic molecules, (d) transition metal complexes, and (e) nanoporous crystals. We compared 12 atomic charge assignment methods for molecular systems and 6 for nanoporous crystals. In this article, we introduce a computationally efficient quadrupole-dipole-resorption (QDR) method that improves the accuracy of stockholder-partitioning models (e.g., DDEC6) for approximately reproducing the electrostatic potential and molecular dipole and quadrupole moments. For each electrostatic model, we computed the electrostatic potential's root-mean-squared error (RMSE) and relative RMSE (RRMSE) using the material's QM-computed electrostatic potential as a reference. The electrostatic RMSE and RRMSE were computed for a training dataset containing 21 geometric conformations per material and a validation dataset containing 20 new geometric conformations per material. Raincloud plots were prepared to visualize the resulting data distributions. For each charge assignment method in the nonperiodic materials, we also computed and compared the root-mean-squared charge transfer magnitude, correlations to other charge assignment methods, conformational sensitivity, etc. Among point-charge models, multiframe ESP methods provided the best overall accuracy for reproducing the electrostatic potential across different system conformations, but they require a training set containing many geometric conformations. The QDR-DDEC6 and CM5 methods provided good conformational transferability and electrostatic potential accuracy across various conformations even when trained only on a single optimized ground-state geometry. A Pareto plot was prepared illustrating the tradeoff between conformational sensitivity and accuracy for reproducing the electrostatic potential of individual conformations. Including atomic dipoles (e.g., QDR-DDEC6_ad, DDEC6_ad, and MBIS_ad) greatly improved the electrostatic model for individual conformations, and QDR-DDEC6_ad outperformed all atom-centered point-charge models. We recommend that more computationally efficient methods be developed to use electrostatic models containing atom-centered point charges plus atomic dipoles in flexible forcefields. Finally, some electron-density partitioning approaches have the key advantage of providing accurate results even when applied to dense solids under high pressures, and we demonstrated this using 11 sodium chloride crystals having various stoichiometries.