Hydrogen is a clean energy carrier, but its largescale use hinges on safe, highdensity storage. Solidstate hydrogen storage (metal hydrides, complex hydrides, metal–organic frameworks) is promising, yet traditional experimental and computational methods are slow and costly.
The Digital HydrogenS platform, developed by the team, consolidates data from over 1000 peerreviewed publications. It contains more than 3000 unique material entries and exceeds 254,000 structured records across ten material classes (Mgbased, AB₅, highentropy alloys, borohydrides, MOFs, etc.). Unlike earlier singlemodal databases, Digital HydrogenS integrates multimodal data: numerical thermodynamic/kinetic parameters, PCT (pressurecompositiontemperature) curves, kinetic curves, and synthesis metadata.
Data analytics reveal persistent gaps. Clustering analysis of V–Ti–Cr, Mg–RE–Ni, and Ti–Fe/Ti–Mn alloys shows that almost no material simultaneously meets the US Department of Energy targets for onboard storage: gravimetric capacity >5.5 wt% (2025) and 6.5 wt% (ultimate), delivery temperature –40 to 85 °C, and delivery pressure 5–12 bar. V–Ti–Cr alloys suffer from low capacity (<4 wt%); Mg–RE–Ni alloys have high capacity (>6 wt%) but dehydrogenation temperatures above 500 K; Ti–Fe/Ti–Mn alloys have too low capacity and too high pressure. The data are also unevenly distributed, causing outofdistribution problems for ML models.
ML is now widely applied. Models (random forest, gradient boosting, Gaussian process regression, graph neural networks) predict hydrogen storage capacity, formation enthalpy, equilibrium pressure, and even full PCT curves with R² often >0.95. For example, a GPR model for AB₂ alloys achieved R² = 0.969 for hydride formation enthalpy. Highthroughput computation combined with ML has screened millions of candidate alloys and MOFs, identifying promising materials such as a Vbased MOF with 9.0 wt% H₂ at 77 K and 150 bar.
For mechanistic understanding, neural network potentials (NNPs) offer nearDFT accuracy at greatly reduced cost. NNPs have simulated H₂ dissociation on Cu surfaces, hydrogen diffusion in Pd(111), and dehydrogenation of MgH₂ slabs, capturing subsurface H₂ formation and longtime dynamics. Current limitations include poor handling of longrange electrostatics, lack of transferability across materials, and the absence of a general NNP for complete absorptiondesorption cycles.
The review outlines a forward roadmap: (1) build openaccess multimodal databases combining numbers, text, spectra, and images; (2) develop multimodal foundation models integrating experiment, computation, and data; (3) implement applicationdriven inverse design (generative models, genetic algorithms); (4) construct generalized NNPs covering full hydrogen storage cycles. These advances will accelerate the transition from empirical to intelligent materials discovery, supporting the clean energy transition.
DOI
10.1007/s11705-026-2649-3