A robust and interpretable machine learning framework for vanadium oxide supercapacitors

Abstract

As global energy demands intensify, the development of efficient, scalable and reliable energy storage systems becomes increasingly urgent. While lithium-ion batteries dominate the current market, their low power density makes them unsuitable for current fluctuations degrading their life expectancy. Supercapacitors (SCs) with pseudocapacitance materials such as vanadium oxide offer an attractive option, with high power density, long life cycle and fast charge-discharge rate. However, their low energy density remains a major bottleneck limiting broader adoption. Current supercapacitor research is focused on improving the specific capacitance and thus expanding their energy density, nevertheless this is mostly done on traditional trial and error experiments, making it time-consuming, slow and expensive. Materials Informatics offers a paradigm shift by implementing machine learning (ML) techniques to uncover patterns in existing data and accelerate the design of novel materials. Despite promising results, many current materials ML studies suffer from limitations such as small data range, improper data preprocessing, target leakage, and lack of reproducibility due to unshared code and datasets. In this work a robust machine learning framework was developed for vanadium oxide SCs, designed to extract interpretable insights from manually gathered literature data. A rigorous cross-validation (CV) pipeline was implemented to ensure reliable model evaluation, avoiding common pitfalls such as overfitting and data leakage. Among the evaluated models, a Voting Regressor combining Ridge Regression, Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) achieved the best performance with a mean absolute error (MAE), root mean squared error (RMSE), and 𝑅2 of 81 𝐹 𝑔 ⁄ , 104 𝐹 𝑔 ⁄ and 0.61, respectively. To extract insights from the models, interpretability algorithms, including permutation importance (PI) and SHapley Additive exPlanations (SHAP) values were employed. Binder-free electrodes, wider potential windows, and a low current density are consistently associated with higher specific capacitance predictions. These findings highlight the potential of interpretable methods to uncover the ML models behavior and lead guided design of SCs.