A machine learning force field for lithium thiophosphate solid electrolytes

We present a machine-learning force field (MLFF) for lithium thiophosphate (LPS) solid electrolytes, designed for large-scale atomistic simulations of ionic transport and structural properties. The model is trained on an extensive set of ab initio molecular dynamics data for twelve crystalline and glass–ceramic LPS compositions and refined to accurately reproduce structural and dynamical properties. The resulting MLFF shows low errors on energies, forces, and virials, remains stable over long molecular dynamics simulations, and demonstrates good generalizability across related compositions within the LPS chemical family, including compositions not explicitly represented during training. Large-scale simulations are used to investigate the rate and the geometric features of lithium diffusion, reproducing known conductivity trends across the LPS family. Analysis of lithium trajectories reveals a strong dependence of ionic mobility on anion speciation, with ortho-thiophosphate units promoting fast Li⁺ transport and pyro- and meta-thiodiphosphate motifs progressively hindering diffusion.