Stacking Ensemble Learning with Genetic Algorithm Optimization for Multi-Property Prediction and Inverse Design of BCC-Type V-Based Hydrogen Storage Alloys

Accurate prediction of hydrogen storage properties is needed for accelerating the design of body-centered cubic (BCC)-type V-based alloys, where the composition–property space is too large for trial-and-error experimentation alone. Here we report a stacking ensemble framework that combines XGBoost, Random Forest, Extra Trees, and Gradient Boosting as base learners with linear or ridge meta-learners, with hyperparameters tuned by a genetic algorithm (GA). Three descriptor strategies are compared across seven target properties: composition-only (C), property-only (P), and composition–property fusion (CP). On target-specific subsets containing 95–901 experimental records and 17 alloying elements (the smallest subset, 95 records, corresponds to maximum hydrogen capacity Cmax, which may limit model stability for that property), the best model for each target gives R2 values from 0.865 to 0.981 (all metrics are from five-fold cross-validation; no independent external test set was employed); the P-type model for desorption plateau pressure reaches R2=0.981. SHapley Additive exPlanations (SHAP) analysis shows that physically derived descriptors, including valence electron concentration, atomic size mismatch, and electronegativity difference, dominate in P and CP models, whereas Ti and Cr contents are the leading compositional features. A Non-dominated Sorting Genetic Algorithm II (NSGA-II) optimizer then ranks 3000 final candidate alloy compositions in six alloy families using four objectives: high predicted retention, high predicted cycle count, high predicted BCC phase ratio, and a low cost proxy; these candidates represent model-based predictions and await experimental synthesis and cycling validation. The V–Ti–Cr–Fe–Mn system contains the largest number of candidates with predicted retention above 99%, while Ti-free V–Cr–Fe–Mn–Al alloys provide low-cost alternatives in the model search space. First-principles calculations on four representative alloys only show that BCC structures are lower in energy than FCC structures by about 0.08–0.13 eV/atom and that hydrogenated structures exhibit clear charge accumulation around H sites, supporting the physical plausibility of the data-driven screening results, though density functional theory (DFT) validation does not replace experimental measurement of PCI curves and cyclic stability.