Atomistic insights into room-temperature ion conduction mechanisms in Li10GeP2S12 via machine learning interatomic potentials

Abstract

Solid-state electrolytes (SSEs), as a key component of all-solid-state lithium-ion batteries, face significant hurdles for commercialization due to their inherently lower ionic conductivity compared to conventional liquid electrolytes. Among various SSE candidates, Li10GeP2S12 (LGPS) is distinguished by its exceptionally high ionic conductivity, which arises from its abundant mobile lithium ions and unique three-dimensional conduction channels. However, ab initio molecular dynamics (AIMD) simulations of LGPS require elevated temperatures (>800 K) due to timescale constraints, limiting accurate prediction of room-temperature behavior. To address these limitations, we develop a graph-based machine learning interatomic potential (MLIP) trained on diverse, small-scale structures from AIMD. This fine-tuned MLIP enables large-scale molecular dynamics simulations of up to 20 ns at 300 K timescales previously unreachable by conventional approaches. Through comprehensive structural analysis and Van Hove correlation function calculations, we elucidate the room-temperature Li+ transport mechanisms, revealing the intricate interplay between individual and collective ionic motion. This work establishes our MLIP as a physically accurate and computationally scalable framework, paving the way for multi-scale modeling of energy materials with near-density functional theory accuracy.