Highly reliable and large-scale simulations of promising argyrodite solid-state electrolytes using a machine-learned moment tensor potential

The high ionic conductivity of argyrodite makes it an attractive candidate for solid-state electrolytes (SSEs) in all-solid-state Li-ion batteries (ASSBs). Although great effort has been devoted to using ab initio molecular dynamics (AIMD) to evaluate ionic conductivity and elucidate the Li-ion diffusion mechanism of argyrodite-based SSEs, limitations in system size, simulation temperatures, and time associated with AIMD make accurate predictions and analysis of Li-ion diffusion challenging. Here, we present a reliable, large-scale computational approach to realistic simulation of SSEs in the bulk and at the grain boundary (GB) based on moment tensor potentials (MTPs) trained at the van der Waals optB88 level of theory. MTPs enable sufficiently large-scale and long-time simulations that reflect all possible configurational disorder of experimental crystal structures and provide accurate ionic conductivities that are close to values measured experimentally in halogenated Li-argyrodite (Li6PS5X [X = Cl, Br, I]). Our simulations show that the vibrational motion of a PS4 polyhedron has a positive effect on ionic conductivity. We also developed an accurate MTP using an active-learning approach to exploring Li-ion diffusion at the GB in polycrystalline SSEs. Simulations of the molecular dynamics of large & sum;5[100](021) (>10,000-atom) GB models reveal that Li-ion accumulation around the GB region retards ionic conductivity and extends into an interior region approximately 20 & Aring; from the GB interface. This work provides a practical approach to realistic large-scale and interfacial GB simulations that are otherwise inaccessible through ab initio calculations by developing accurate machine-learned MTPs.