Abstract

Materials design enables technologies critical to humanity. While many properties of a material are determined by its atomic crystal structure, prediction of the atomic crystal structure for a given material’s chemical formula is a long-standing grand challenge that remains a barrier in materials design. We build a novel dataset of more than 100,000 random structure relaxations of battery anode materials using high-throughput density functional theory (DFT) calculations, which includes calculations with varying quantum mechanical settings for out-of-domain generalization. We modify graph neural network force fields to also predict stress information, which allows them to effectively simulate relaxations. We show that models trained on data conventionally used to train interatomic potentials fail to simulate relaxations from random structures, and random structure relaxations data is crucial for crystal structure search. We find that models trained with data augmentation via random perturbations improves both the accuracy and out of domain generalization, and is able to find an experimentally verified structure of a new stoichiometry.