Predicting the Heat of Hydride Formation by Graph Neural Network - Exploring the Structure-Property Relation for Metal Hydrides

Theoretical tools or structure-property relations that enable the prediction of metal hydrides are of enormous interest in developing new hydrogen storage materials. Density functional theory (DFT) is one such approach that provides accurate hydride formation energies, which, if complemented with vibrational zero-point energy and other contributions, provides accurate hydride formation enthalpies. However, this approach is time consuming and, therefore, often avoided, hindering the modeling of experimental behavior. The recent implementation of graph neural networks (GNN) in materials science enables fast prediction of crystal formation energy with a DFT accuracy. Starting from the MatErials Graph Network (MEGNet), transfer learning is applied to develop a model for predicting hydride formation enthalpy based on the crystal structure of the starting intermetallic. Excellent accuracy is achieved for Mg-containing alloys, allowing the screening of the Mg-Ni-M ternary intermetallics. In addition, data containing matching experimental properties and crystal structure of metal hydrides are provided, enabling future development.

Automated summary

Theoretical tools or structure-property relations play an important role in developing new hydrogen storage materials. Density functional theory (DFT) provides accurate hydride formation energies, but it is time consuming, hindering experimental behavior modeling. The recent use of graph neural networks (GNN) enables fast prediction of crystal formation energy. Transfer learning is applied to predict hydride formation enthalpy based on the crystal structure of starting intermetallics, achieving excellent accuracy for Mg-containing alloys.