Physics-Informed Composition-to-Tank Modeling of Metal Hydride Hydrogen Storage

Composition-based machine-learning models can estimate metal hydride thermodynamics, but their errors are reported at the material level, whereas storage systems are judged at the tank level, and this study connects the two. Gradient-boosting models trained on 420 hydrides with a van ’t Hoff-consistent parameterization predict the room-temperature plateau, ln Peq, with cross-validated mean absolute errors of 0.90 for AB5 and 1.75 for AB2 alloys. A reduced-order heat- and mass-transfer model of an open 7.8 kg AB2 tank was calibrated on the pressure, temperature and flow time series of one test and reproduced two further tests within 3.5–5.5 °C and 0.4–1.0 bar. The tank data place the alloy’s plateau at ln Peq = 2.59–2.68, 2.0 units above the prediction, while the predicted enthalpy is within about 2 kJ mol−1. With predicted thermodynamics, the simulated tank absorbs 60 g in 1.1 h instead of 8.8 h, peaks at 106 °C instead of 58 °C and delivers 31% less hydrogen; delivered capacity tolerates plateau errors of about −1.5/+0.25, charging behaviour only ±0.15. Composition models nevertheless rank substitution series correctly, as shown for La1−xCexNi5 and LaNi5−xAlx, and combined with one measured isotherm they offer a fast route from alloy composition to tank design.