Machine Learning Insights Into Functional ABH ₃ Perovskite Hydrides for Hydrogen Storage Applications

ABSTRACT

Here, we develop and validate a machine‐learning (ML) framework for rapid screening and prediction of hydrogen‐storage performance in perovskite hydrides (ABH ₃ ), using 17 physicochemical descriptors derived from 1000 DFT‐calculated compounds. Efficient solid‐state hydrogen storage is critical for sustainable energy, but exhaustive DFT screening of the thousands of possible perovskite‐hydride (ABH ₃ ) compositions is computationally prohibitive. Here a ML framework was developed and validated on 1000 DFT‐derived ABH ₃ compounds described by 17 physicochemical descriptors, benchmarking Random Forest, XGBoost, SVR, Decision Tree, Linear Regression, and a stacking ensemble across five target properties. On a genuinely held‐out 200‐compound test set, the optimized stacking ensemble reached R ²  = 0.994 for hydrogen‐storage capacity and F1 = ROC‐AUC = 1.000 for stability and magnetic‐ordering classification; formation energy ( R ²  = 0.755) and bulk modulus ( R ²  = 0.749) proved harder to predict, reflecting genuine descriptor‐property limits rather than model failure. A SISSO‐style symbolic‐regression search distilled the ensemble into a transparent two‐term formula (summed A‐/B‐site atomic mass and its product; R ²  = 0.838, test), recovering the model's learned physics with no black‐box component. SHAP, LIME, permutation importance, and UMAP consistently identify A‐/B‐site atomic mass as the dominant driver of storage capacity, yielding a design threshold (M _a  + Mᵇ ≤ 202.5 amu) satisfied by all 15 top‐ranked candidates (KScH ₃ , NaCoH ₃ , CaNiH ₃ , among others). Benchmarked against AB5 (LaNi ₅ , 1.2–1.5 wt%) and AB2 (TiMn ₂ , 1.8–2.2 wt%) intermetallic hydrides, the explored ABH ₃ space reaches up to 8.19 wt%. Inference across all 1000 compounds completes in under 100 ms after an 18‐s training step, versus hours‐to‐weeks per DFT calculation. A statistically significant but weak, mass‐confounded relationship between magnetic ordering and storage capacity is reported, correcting an earlier overstated trade‐off.