Modular Composition-to-Structure Machine-Learning Workflow for Hypothesis Generation in Lithium Solid Electrolytes

Abstract

Machine-learning screening of lithium solid electrolytes is often performed either at the composition level, where polymorph-dependent transport cannot be represented, or within fixed crystal-structure databases, where unexplored compositions and structures are inaccessible. Here, we present a modular two-stage workflow that connects constrained composition generation and composition-based property prediction with stochastic crystal-structure generation, machine-learning-potential relaxation, and structure-aware ranking. The individual algorithms are established; the methodological contribution is their integration into a composition-to-structure hypothesis-generation pipeline that operates beyond a fixed list of known structures. The workflow was executed through M3GNet prerelaxation and internal ranking of 4600 PyXtal-generated trial structures. Because the structure-aware model has an MAE of 2.72 log10 units and no candidate-specific uncertainty analysis, applicability-domain analysis, DFT reoptimization, convex-hull analysis, transport simulation, or experimental validation was performed, the generated structure-level scores are not interpreted as quantitative conductivities or evidence of physical viability. The demonstrated contribution is therefore a validation-aware early triage architecture that produces hypotheses for subsequent high-fidelity assessment rather than a validated solid–electrolyte discovery.