Physically consistent and explainable crystal graph neural networks for hydrogen desorption enthalpy prediction in crystalline hydrides

Accurate prediction of hydrogen desorption enthalpy (ΔH) is essential for discovering crystalline hydrogen storage materials. Although crystal graph neural networks (CGNNs) show promise in materials informatics, differences in datasets and protocols hinder objective comparison. This study presents a unified benchmarking framework evaluating ten CGNN architectures using identical preprocessing, descriptors, optimization, and leakage-free five-fold cross-validation on 817 hydrides from OQMD. Performance was assessed using R 2 , MAE, RMSE, MedianAE, MaxAE, Pearson correlation, and Entropy Weight TOPSIS. Explainability (GNNExplainer, Integrated Gradients, GraphSVX) and physical consistency (crystal-chemical attribution, electronegativity trends) were also examined. SphereNet achieved the highest predictive accuracy (R 2 = 0.8135, MAE = 17.18 kJ mol −1 H 2 , RMSE = 34.56 kJ mol −1 H 2 ), while Equiformer ranked first in multi-criteria evaluation. Explainability highlighted local coordination environments and chemically meaningful atomic interactions as dominant factors. The proposed framework provides a standardized, explainable, and physically consistent strategy, supporting AI-assisted discovery of advanced hydrogen storage materials.