Decoupling correlation from causality in material design: A causal machine learning framework for Mg-based hydrogen storage and hydrolysis

Magnesium-based hydrolysis is promising for portable hydrogen supply, yet its optimization is hindered by costly trial-and-error experiments and the weaknesses of conventional machine learning, which may overfit spurious correlations and fail to control confounders, offering limited intervention-oriented guidance. This study develops a causal machine-learning framework integrating Tree-structured Parzen Estimator optimization, SHAP interpretability, and causal inference for Mg-based hydrogen storage and hydrolysis. Using 568 hydrogen-storage-density and 279 hydrogen-production records, optimized NGBoost and XGBoost models achieved R2 of 0.868 and 0.989, respectively. Causal inference, adjusting for measured covariates, quantified elemental effects: for storage density, In showed a strong positive effect (0.682), Ce a negative effect (−0.175), and Ca a negligible effect (0.018); for hydrogen production, positive effects followed In (105.98) > Ce (42.89) > Ca (17.71), identifying In as the most influential variable within the experimental domain. SHAP analysis further revealed practical thresholds: Fe@C loading above 1.5 g improves storage density, while In content above 2.5 g enhances hydrogen production. Finally, a Streamlit-based GUI supporting parameter input, real-time prediction, and candidate comparison provides a transparent, quantitative, intervention-oriented tool for accelerating Mg-based hydrogen-generation material design.