Multi-Level Structural Descriptors for Deep Learning–Based Prediction of Hydrogen Storage in Metal–Organic Frameworks

Safe and efficient hydrogen storage is a critical barrier to realizing a carbon-neutral energy system. Metal–organic frameworks (MOFs) are promising candidates owing to their adjustable porosity and high surface areas, yet the vast compositional design space makes exhaustive molecular simulation impractical. We developed a hybrid deep-learning framework using multi-level structural descriptors—geometric, atomic, and crystallographic—to predict hydrogen uptake in MOFs. Training on 10,736 experimentally synthesised MOFs with hydrogen storage capacities computed using Grand Canonical Monte Carlo simulations, we benchmarked three architectures (a multilayer perceptron, a graph transformer, and an edge-free materials graph network). All three models performed comparably, with high prediction accuracies across gravimetric and volumetric capacities under both pressure-swing and temperature–pressure-swing conditions. Rapid screening of 137,953 hypothetical MOFs (hMOFs) resulted in the identification of 630 candidates exceeding 50 g H2/L under working conditions (77 K/100 bar to 160 K/5 bar). Top performers are characterised by multinuclear Zn clusters, aromatic carboxylate linkers, nitrogen incorporation, and selective halogen functionalisation. These findings demonstrate that the use of multi-level structural descriptors accelerates computational screening and prioritisation of candidate MOFs for further simulation and experimental evaluation.