Molecular dynamics simulations of Li6PS5Cl accelerated by machine learning interatomic potentials

Li₆PS₅Cl is regarded as a promising candidate for the solid electrolyte in all-solid-state lithium-ion batteries (ASSLIBs). Despite considerable attention, atomistic simulations of polycrystalline Li₆PS₅Cl remain rare due to the high computational cost. In this study, machine-learning interatomic potentials, specifically moment tensor potentials (MTPs),
are employed to accelerate the simulations while preserving the ab initio accuracy. In the initial stage, energies, forces and stresses of a small number of configurations are generated under the ab initio framework. Based on it, MTPs are fitted and validated with
experimental data. The usage of MTPs enables molecular dynamics (MD) simulations in larger system sizes (up to 10 000 atoms) and longer time scales (several ns). To this end, statistically more precise diffusion coefficients under wide ranges of temperature and pressure are obtained. The derivation of diffusion coefficients of the Li atoms crossing grain boundaries becomes possible. These data can be directly integrated for supporting large-scale simulations based on, e.g., the finite element method. Additionally, thermal properties of Li₆PS₅Cl can be derived from MD simulations, which benefits the research and application of ASSLIBs.