MACHINE LEARNING ACCELERATED PROPERTY PREDICTION AND SCREENING OF STABLE LITHIUM-BASED BATTERY MATERIALS

DOI: 10.18173/2354-1059.2026-0005

Abstract

Rapid screening of battery materials is critical to meeting the growing demand for high-performance energy-storage solutions. In this study, we employ a data-driven approach to evaluate and characterize stable lithium-containing compounds with potential applications in next-generation batteries. We retrieved a dataset of 1,846 stable lithium-based materials from the Materials Project database using the mp-api. By leveraging the matminer library, we generated a comprehensive set of composition-based and structure-based features, including elemental- property and density descriptors. A Random Forest Regressor was trained to predict the band gap of these materials, a key electronic property that determines their suitability as electrolytes or electrodes. The model achieved a mean coefficient of determination (????2) of 0.77 ± 0.04 and a mean absolute error (MAE) of 0.48 ± 0.05 eV across 10-fold cross-validation, demonstrating robust predictive capability. Featureimportance analysis revealed that electronegativity and Mendeleev number are the most significant predictors of band-gap energy. These findings underscore the effectiveness of machine learning in accelerating property prediction for battery materials and provide a pathway to more efficient experimental targeting.