Crystal structures can be predicted from first-principles using ab initio random structure searching (AIRSS) to sample the potential energy landscape, and density functional theory (DFT) to robustly and accurately describe it. While classical interatomic potentials lower computational costs at the expense of robustness and accuracy, modern machine-learning potentials offer a compromise, pro- viding both at a fraction of the DFT cost. In this work, we use Ephemeral Data-Derived Potentials (EDDPs) to accelerate AIRSS calculations for binary hydrides at 100 GPa. Since the training data is generated iteratively using AIRSS, the searches require no prior knowledge of hydrides. These potentials allow more diverse searches, sampling diverse compositions, larger unit cells, and many more structures. Besides recovering known structures, the searches reveal hydrogen-rich phases of H22(BrH), H23Pb, and H32Mg, supermolecular phases of H25Cs and H26Rn, and many ‘substoi- chiometric’ variants of known hydrides. Our results indicate that using the current generation of universal MLIPs to search for novel high-pressure hydrides is less effective due to model instabilities or markedly slower inference speeds and highlight the necessity of generating new, targeted data.
Conway et al. (Fri,) studied this question.