Structure prediction of iron hydrides at high pressures with machine-learned interatomic potentials

The structure and properties of iron hydrides under pressure have been of interest to geoscientists. At ambient conditions, there are no stable solid iron hydrides. Previous theoretical and experimental studies suggest that the double hcp phase of FeH is stable at low pressures with phase transitions to the hcp and fcc phases up to 80 and 140 GPa, respectively. Here, we present a theoretical investigation of the potential energy surfaces of FeH at high pressure. We construct a highly transferable machine-learned interatomic potential with a hierarchical approach using the PyFLAME code. Then, using this fast and accurate neural network potential, we systematically explore the potential energy surfaces of bulk structures of FeH by global sampling using the minima hopping method, to predict stable and metastable iron hydrides up to 200 GPa. We have carried out density functional theory calculations to refine the predicted structures and to evaluate the dynamical stability of selected structures as well. In an automated and systematic approach, we are going to show how a transferable machine-learned interatomic potential can be trained and validated using global optimization and analyze the phase diagram of the stoichiometric Fe-H system under pressure.