Metal Hydride Composition-Derived Parameters as Machine Learning Features for Alloy Design and H2 Storage

S. Nations, T. Nandi, A. Ramazani, S. Wang, Y. Duan
National Energy Technology Laboratory,
United States

Keywords: H2 storage, Machine learning, metal hydride


The green energy technology revolution has at its core the problems of finding efficient and sustainable energy generation, transportation, and storage which can be integrated to take the place of legacy energy solutions. Hydrogen holds great promise as a transportation and storage solution due fundamentally to its high energy density and lack of greenhouse gas emission. Though hydrogen is a promising energy carrier for a green future, many challenges persist in its implementation. One such challenge is the difficulty in engineering viable storage solutions, with metal hydrides being a leading contender among solid-state strategies. To facilitate efficient searching of candidate alloys, in this study, the ridge regression, simple decision trees, random forest ensembles and gradient boosting ensembles were employed to predict sparse energy of formation data, with the random forest ensemble resulting in the lowest test set error. First, two public databases, Materials Project and HydPark, were combined for metal hydride entries. Feature engineering was performed before the models were developed, resulting in electronegativity, density, atomic density, d-character, f-character, band gap, hydrogen weight fraction, magnetization, temperature, and pressure being retained. Principal component analysis and polynomial feature creation were tested but found to be detrimental to predictions. The model hyperparameters were optimized via a Bayesian cross-validation search and the models were benchmarked against one another. The lowest test error model, a random forest ensemble, was then used to populate sparse entries missing energy of formation before all were scored by hydrogen storage capacity and energy of formation suitability, with the latter being a proxy for viable operating temperatures. A robust model was trained for predicting energy of formation from readily available features including several able to be computed from just the chemical formula. These features derived from chemical formula were found to be highly predictive and so are promising for screening many candidates just by composition. Candidates were scored for hydrogen storage potential and promising ones were identified. Finally, a tool for screening of arbitrary metal hydrides by composition, band gap, density, atomic density, magnetization, temperature, and pressure is presented. The features presented in this database can be expanded and applied for H2 storage and alloy design.