Machine Learning-Assisted Crystal Structure Prediction of Solid-State Electrolytes Reveals Superior Ionic Conductivity in Metastable Edge-Sharing Phases.

Significant attention has been devoted to developing novel solid-state electrolytes (SSEs) with high ionic conductivity for all-solid-state batteries (ASSBs). However, most studies have primarily focused on compositional substitutions, often overlooking the fundamental role of inherent crystal structures on ion transport. To address this, we introduce a theoretical crystal structure prediction (CSP) approach based on the machine-learning moment tensor potential (MTP). The proposed approach successfully identifies novel SSE structures and reproduces 12 experimental crystal structures. Using a phase-diagram-guided strategy, CSP is applied to four promising SSE candidates, Li2SiS3, Li2GeS3, Li4SiGeS6, and Li4SiSnS6, to assess their polyhedral connectivity, relative stability, and Li-ion transport properties. The results reveal that metastable edge-sharing phases exhibit superior Li-ion mobility compared with their stable corner-sharing counterparts. This superior conductivity is attributed to the Li-ion accessible volume, quantified by the packing ratio (fraction of the unit cell volume occupied by nonconductive volume) and by the dynamic distortion of the Li-S4 sublattice, which represents the local environment encountered by migrating Li-ions. The metastable phases feature higher packing efficiency, larger Li-S4 sublattice volume, and greater distortion, all of which contribute to improved Li-ion transport. This study highlights the potential of CSP to design novel SSEs and high-performance ASSBs.