Machine Learning for Hydrogen Storage Properties of TiFe Alloys Based on Literature-Curated Data Set: Prediction and Limitation

This study establishes an interpretable machine learning framework to elucidate the composition–property relationships of TiFe-based hydrogen storage alloys. Beyond elemental compositions, two physical descriptors (Size 1/Size 2) were first introduced. XGBoost performed superiorly in predicting hydrogen absorption and desorption plateau pressures. SHAP analysis revealed that Fe content and temperature are the principal variables affecting platform pressure, a result that is consistent with empirical observations. However, hydrogen storage capacity prediction showed limited accuracy In addition to the features utilized in the present study, there is a lack of supplementary descriptive features that can robustly characterize hydrogen storage capacity. Potentially relevant attributes may encompass phase composition, synthesis methodology, and chemical information on the material. The lack of incorporation of such data constrains the predictive performance and utility of machine learning data sets of TiFe alloys. This work provides actionable insights for alloy design while exposing inherent limitations of the Literature-curated Data set by coupling machine learning with experimental data. The convergence of interpretable AI and materials science demonstrates transformative potential to overcome longstanding challenges in hydrogen storage material development, particularly through hybrid data-physics modeling strategies.