Ensemble Tree-Based Machine Learning for Predicting Volumetric Change in Lithium-Based Battery Electrode Materials

Abstract

Volumetric instability in lithium-based electrode materials remains a persistent challenge in electric vehicle battery development, as identifying stable material combinations through conventional laboratory methods is both time-consuming and resource-intensive. This study develops and compares three ensemble tree-based machine learning models, namely Random Forest, XGBoost, and CatBoost, to predict the maximum volume change percentage of lithium-based electrode materials. A dataset of 52,503 samples was constructed by integrating electrode pair data with structural and electronic features from the Materials Project API, enriched with compositional descriptors extracted using the Matminer Magpie preset. Each model underwent baseline evaluation followed by hyperparameter tuning using Optuna with Bayesian optimization over 100 trials, assessed using RMSE, MAE, and R². All three models achieved R² test above 0.98, with Random Forest yielding the best performance at RMSE of 17.1713, MAE of 4.0954, and R² test of 0.9900. SHAP analysis identified density discharge as the most determinant predictor across all models, reflecting its physicochemical role in representing the final crystal structure state following lithium intercalation. These findings confirm that ensemble tree-based models offer a reliable and efficient alternative to wet laboratory experimentation for lithium-based electrode material discovery.