Active learning and explicit electrostatics enable accurate modeling of electrolytes

Abstract

Machine learning interatomic potentials (MLIPs) offer near-\textit{ab initio} accuracy with the efficiency of classical force fields, making them attractive for modeling electrolytes. Collecting a diverse training set is essential for their accuracy and reliability, and explicit treatment of strong electrostatic interactions may be necessary. In this work, we demonstrated that D-optimality-based active learning can automatically generate diverse training sets for moment tensor potentials (MTPs), enabling reliable molecular dynamics simulations of pure ethylene carbonate (EC), ethyl methyl carbonate (EMC), their mixtures, and LiPF$_6$ solutions. The resulting MTPs exhibit excellent transferability across various EC/EMC compositions, producing ionic conductivities within 11\% mean deviations from experiments. In addition, we assessed the impact of explicitly incorporating electrostatics by augmenting MTP with charge redistribution schemes using either fixed or environment-dependent charges. Our results show that the augmented MTP achieves the same or higher accuracy than standard model with fewer parameters, while environment-dependent charges further improve accuracy and the stability of simulations.