Compositionally-informed machine learning for solid-state electrolyte design: a structure-free approach

Next-generation energy storage demands high-performance solid-state electrolytes (SSEs), where machine learning (ML) promises to be a powerful design tool. Typically, ML requires atomic positions to construct essential material descriptors, which limits its utility at the earliest design stages when structural data are absent. Composition-based ML offers an alternative by relying solely on chemical formulas and elemental properties, yet accurately capturing ionic size effects and handling compositional complexity remain open challenges. Here, we introduce a compositionally-informed ML (CI-ML) framework that combines ionic radius mismatch (IonicRad_Mis) as a central composition-only descriptor with a compositionally-stratified modeling strategy. Applied to halide SSEs, the XGBoost model achieves robust accuracy on the global dataset (R²: 0.847 training, 0.707 test) and exceptional performance on compositionally-stratified subsets (R² of 0.991/0.863 for ternary halides, 0.892/0.836 for quaternary halides). SHAP analysis identifies IonicRad_Mis as the paramount descriptor and reveals its negative correlation with ionic conductivity, a global trend mapped across 1194 halides from the Materials Project database. Guided by this insight, Br-substituted Li₃InCl₆ is designed to reduce IonicRad_Mis (e.g. from 0.20318 to 0.17494 Å), which boosts the conductivity from 0.88 to 1.30 mS cm^-1 that closely matches the model's prediction (1.39 mS cm^-1). This structure-free CI-ML approach provides a generalizable pathway to overcome the structural-dependency bottleneck, paving the way for accelerated early-stage discovery of SSEs as well as other functional materials.