Analyzing lithium-ion conduction in thiophosphate glassy electrolytes via machine learning

Glasses such as lithium thiophosphates (Li2S-P2S5) show promise as solid electrolytes in lithium-ion batteries, but a poor understanding of the impact of the disordered structure on lithium transport properties limits the further development of glassy electrolytes. Here, we simulate glassy Li2S-P2S5 electrolytes with varying fractions of polyatomic anion clusters using classical molecular dynamics. Based on the determined variation in ionic conductivity, we use a classification-based machine learning metric termed “softness” – a structural fingerprint that is correlated to the atomic rearrangement probability – to unveil the structural origin of lithium-ion mobility. To derive a real-space origin of the machine-learned softness metric, we analyze the energy barrier of softness-coded lithium ions migrating between two sites, showing that soft lithium-ion migration requires a smaller energy barrier to be overcome relative to that observed for hard lithium-ion migration.