Machine learning–driven prediction and analysis of lifetime and electrochemical parameters in graphite/LFP batteries

Abstract This study proposed a novel transformer-based regression model for predicting the lifetime coefficient, using specific energy, specific power, and the remaining capacity of three cylindrical graphite/LFP batteries. Its predictive capabilities were methodically evaluated against six widely used machine learning approaches—M5, random forest, gradient boosting, stacked regressor, XGBoost, and CatBoost to benchmark in the small-data regime. A comprehensive dataset was used with 239 different cyclic conditions for 18,650 and 26,650 form factors, with form factor, capacity, cycling temperature, cycling depth, test duration, and full cycles as the input features. The seven models were pre-processed, hyperparameter-tuned, trained, and optimized to predict the target variables accurately. The study revealed vital insights into the correlation among the input features and the key trends among the target variables via violin plots, Pearson’s correlation heatmap, SHAP analysis, and feature importance analysis. The effectiveness of the proposed transformer-based regression over the commonly employed decision tree- and ensemble-based approaches was measured in terms of R 2 , RMSE, SMAPE, MASE, and MAE. The proposed model exhibited superior performance with an R 2 of 0.8653, 0.8657, 0.4516, and 0.3285 for lifetime coefficient, used specific energy, used specific power, and remaining capacity, respectively. The results from the study can pave the way for extending the robustness of the proposed model by integrating time-series cycling behavior, impedance spectra, feature engineering, or aging profiles to complement and advance the existing experimental findings in the energy storage landscape.