MACHINE LEARNING-ASSISTED SCREENING OF THERMODYNAMICALLY STABLE LITHIUM-BASED BATTERY MATERIALS FOR SUSTAINABLE ENERGY STORAGE AND SOCIETAL BENEFITS

Abstract

Efficient, sustainable energy storage hinges on our ability to identify stable lithium-based materials. Standard first-principles calculations take an immense amount of time and computing power, though, when applied across broad materials databases. In this study, we built a machine-learning-assisted workflow to screen for viable candidates, using predicted per-atom formation energies across lithium-bearing and related inorganic compounds. Our initial dataset held 86,988 raw records across 11 descriptors, but filtering out duplicates brought that down to 21,747 unique materials. From chemical, structural, energetic, and electronic descriptors, we built an input matrix totaling 92 features. We evaluated five machine-learning models Random Forest, XGBoost, LightGBM, CatBoost, and an Artificial Neural Network (ANN) using mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R²). On the initial test set, Random Forest achieved the lowest MAE (0.1065), while the Artificial Neural Network achieved the lowest RMSE (0.2041) and the highest R2 (0.9577). According to the SHAP analysis, energy_per_atom influenced the predictions dominantly, while dropping it shifted the main weight onto element_O, band_gap, element_F, density, and volume. When evaluated under a stricter five-fold Group K-Fold split, R² held at 0.9547 ± 0.0065 with energy_per_atom included, but fell to 0.8366 ± 0.0074 without it. The out-of-fold predictions yielded 16 distinct candidate compounds. Cross-referencing these candidates against Materials Project energy-above-hull figures showed that 8 of them (50%) are thermodynamically stable, four lie above the convex hull, and four have no E_hull records at all. While DFT calculations and convex-hull analysis remain necessary to confirm thermodynamic stability across all 16 candidates, these findings point to the real utility of interpretable machine learning for screening energetics quickly.