Data-driven composition-only machine learning for high-performance solid-state electrolytes

Abstract

As a pivotal advancement in energy storage technology, all-solid-state batteries represent a transformative direction for next-generation lithium-ion batteries. To address the critical challenge of low ionic conductivity in solid-state electrolytes (SSEs), we propose a machine learning-driven screening workflow to search for SSEs with high ionic conductivity. By leveraging an experimental database of lithium-ion SSEs, we trained five ensemble boosting models using exclusive elemental composition and temperature parameters. The CatBoost algorithm emerges as the optimal predictor, achieving superior accuracy in ionic conductivity estimation. By implementing this model, we systematically screened 3311 lithium-containing materials from the Materials Project database, identifying 22 promising candidates with the predicted ionic conductivity exceeding 1 mS cm⁻¹. Especially, the predicted conductivity of Li₈SeN₂ (2.72 mS cm⁻¹) is well consistent with the AIMD measurement (2.85 mS cm⁻¹). This data-driven approach accelerates SSE discovery while providing fundamental insights into structure–property relationships, establishing a robust framework for next-generation electrolyte development.