Automatic Generation of Inorganic Solid-State Electrolytes via Unified Multi-Modal Network.

Accurate prediction of ion transport properties remains a key bottleneck in computational discovery of solid-state electrolytes, particularly because experimentally measured ionic conductivities are sparse and strongly dependent on crystal structure. Here we introduce USMNet, a unified structure-aware multimodal neural network that combines crystal-structure representations, compositional features, and physically motivated descriptors to predict lithium-ion conductivity. We benchmark the model against composition-only and structure-based baselines, showing that explicit structural information improves prediction accuracy and enables data-efficient and reliable exploration of candidate materials. We then apply USMNet to 168 675 candidate materials spanning 19 lithium solid-state electrolyte frameworks, identifying 21 high-potential inorganic electrolytes. Selected candidates are synthesized and experimentally validated, with Li16ZnSiP2S16 enabling all-solid-state batteries that retain 80% capacity after 600 cycles under 0.5 C at room temperature. These results establish a generalizable structure-aware learning strategy for electrochemical property prediction and demonstrate its utility in accelerating closed-loop materials discovery under limited data regimes.