Graph Neural Network-Based Discovery Of Novel Materials For Solid-State Batteries

Abstract

Rapid identification of solid electrolytes is a central challenge in the development of all-solid-state batteries because high ionic conductivity must coexist with thermodynamic stability, electrochemical compatibility, electronic insulation, mechanical integrity, and scalable processing. Although graph neural networks (GNNs) have transformed prediction on large computed materials databases, their reliability is less established for small and heterogeneous experimental conductivity datasets. This study presents a leakage-controlled and externally validated public-data framework for solid-electrolyte screening. A stoichiometry-aware composition GNN was trained on the public OBELiX dataset containing 599 synthesized lithium solid electrolytes and its official grouped split (478 training; 121 testing). Element types were represented as graph nodes, atomic and stoichiometric properties as node features, and chemically weighted relationships as message-passing edges. A five-seed GNN ensemble was benchmarked against ridge regression, random forest, extremely randomized trees, and histogram gradient boosting (HGB). On the official hold-out set, HGB was the strongest standalone model (MAE = 1.158 log10(S cm-1), 95% bootstrap CI [0.955, 1.385]; R2 = 0.565; Spearman rho = 0.692), whereas the GNN ensemble achieved MAE = 1.700, R2 = 0.174, and rho = 0.299. The paired bootstrap difference favored HGB by 0.542 log units (95% CI 0.287-0.804), demonstrating that architectural complexity does not guarantee improved generalization in sparse experimental regimes. Nevertheless, the GNN captured complementary transferable signal. On 204 non-overlapping room-temperature records from Liverpool Ionics, an unoptimized equal-weight GNN-HGB consensus achieved MAE = 1.081 (95% CI [0.921, 1.253]), R2 = 0.440, and rho = 0.709. Rule-based isovalent substitutions generated 392 compositions absent from both source datasets, and uncertainty-aware ranking identified Li10.35Ge1.08P1.65S12Sn0.27 as the leading hypothesis, with predicted conductivity of 1.39 x 10-2 S cm-1 and combined uncertainty of 0.258 log units. These formulas are prioritized hypotheses rather than experimentally discovered materials. The study establishes a rigorous pathway in which leakage-resistant benchmarking, simple-model comparison, external validation, uncertainty analysis, and staged DFT and experimental verification jointly support credible data-driven materials discovery.