Machine Learning-Driven Discovery of Sustainable Battery Materials for Next-Generation Energy Storage

Abstract

Solid-state electrolytes are central to safer, high-energy lithium batteries, yet their discovery remains slow because ionic conductivity depends on coupled compositional, structural, processing, and measurement factors. At the same time, screening solely for conductivity can reinforce dependence on scarce, costly, or environmentally problematic elements. This study develops an interpretable, sustainability-aware machine-learning framework for prioritizing lithium solid electrolytes using the open OBELiX database of experimentally measured room-temperature ionic conductivities. After removing records without a valid positive target, 562 entries were retained, comprising 449 training observations and 113 observations in the official held-out test set. Composition-derived descriptors, lattice parameters, symmetry information, and expert-assigned structural families were used to train Elastic Net, support vector regression, random forest, Extra Trees, XGBoost, and LightGBM models. Grouped five-fold cross-validation was performed by reduced composition to limit leakage from closely related formulations. Extra Trees achieved the strongest independent-test performance (R2 = 0.373, MAE = 1.184, RMSE = 1.784 log10 S cm-1, and Spearman rho = 0.581). Feature analysis indicated that space-group number, number of constituent elements, structural family, lattice parameter c, cell volume, and selected elemental fractions were influential. A separate sustainability screening score was constructed from crustal abundance, relative supply risk, price burden, toxicity flags, non-lithium critical-material fraction, and compositional complexity. Multi-objective analysis produced six Pareto-efficient held-out records. Lithium-lanthanum-titanate perovskites offered a favorable performance-sustainability balance, whereas a high-conductivity thio-LISICON composition, Li10Ge0.95Si0.05P2S12, illustrated the trade-off between conductivity and germanium-related resource burden. The moderate external accuracy demonstrates that the framework is most appropriate for ranking and experimental triage rather than precise conductivity prediction. The study provides an auditable route for integrating performance, uncertainty, and sustainability into battery-material selection while clearly separating data-driven prioritization from experimental proof.