Machine Learning for Predicting Hydrogen Storage Capacity: Uncovering Key Material Descriptors Governing Metal Hydride Performance

Abstract

Amongst the various means of hydrogen storage, metal hydride storage with solid-state is considered the most promising way for safe and efficient hydrogen utilization. But, identifying the storage capacity governing descriptors remains a major challenge simply because of complex thermodynamic and compositional effects. This study aimed to predict hydrogen storage capacity in metal hydride materials and identify the key descriptors influencing material performance. A cleaned dataset of 741 observations representing 713 unique metal hydride compositions was analyzed. Hydrogen storage capacity was modeled using Linear Regression, Random Forest, and Gradient Boosting. Predictor variables included material class, heat of formation, entropy of formation, logarithmic equilibrium pressure, and H/M ratio. Model performance was evaluated using MAE, RMSE, and R², while Random Forest feature importance was used to interpret descriptor contributions. Gradient Boosting achieved the best predictive performance, with MAE = 0.1457, RMSE = 0.2400, and R² = 0.8634. Random Forest also performed strongly, with R² = 0.8590, while Linear Regression achieved R² = 0.7978. Feature importance analysis identified the H/M ratio as the dominant predictor, accounting for 47.0% of total importance, followed by the Mg-based material class at 36.7%. Together, these variables contributed approximately 83.7% of total predictive importance. Interpretable machine learning can effectively predict hydrogen storage capacity and reveal descriptor-level insights for metal hydride screening. The findings highlight the importance of hydrogen accommodation capacity and material class in guiding future hydrogen storage material design.