Rare-earth-based hydrogen storage alloys are key materials for clean energy storage because of their high hydrogen storage capacities and good stability. However, further performance improvements remain challenging, as conventional trial-and-error methods and linear models often fail to capture complex nonlinear relationships between composition and performance, leading to low predictive accuracy and slow progress. In this study, an interpretable machine learning (ML) method was developed to predict the maximum hydrogen absorption capacity (AbMax) of rare-earth-based alloys with various crystal structures. We selected a total of 267 data entries from the ML-HydPARK v0. 0. 5 database and literature research, covering systems such as AB, AB 2, AB 5, and La–Mg–Ni. Four ML models were employed for data analysis, with the CatBoost model significantly enhancing the prediction accuracy compared with the linear baseline model (R 2 ≈ 0. 84). Through SHAP and LIME analyses, the key compositional features governing the hydrogen storage performance were identified. The optimal ranges for these features were identified while considering the high hydrogen storage capacity, thereby providing clear guidance for alloy design. The final model is simple and computationally efficient, training can be conducted under the regular configuration, Auxiliary tools that can be used for candidate pre-screening. Overall, this study demonstrates that combining advanced predictive modeling with interpretability analysis can accelerate the discovery of high-performance hydrogen-storage alloys and provide valuable mechanistic insights.
Qi et al. (Fri,) studied this question.