Pressure-transferable universal machine-learning potentials: the WBM-MatPES benchmark and Tuga-FF

Abstract

Universal machine-learning interatomic potentials hold the promise to replace costly ab initio calculations in materials science workflows. However, the accuracy and transferability of current models are limited by the availability of training data. We release WBM-MatPES, a dataset of 182 thousand structures recomputed with consistent MatPES settings, together with two reference convex hulls of different depths and test sets at six pressures up to 150 GPa. These datasets reproduce a realistic discovery workflow and allow potentials to be ranked systematically at this level of theory, similar to the Matbench Discovery leaderboard at the Materials Project settings. In this benchmark, five published MatPES potentials agree within 12 meV/atom at 25 GPa, yet their errors increase dramatically to 0.4-5 eV/atom at 150 GPa. To address this limitation, we develop Tuga-FF, a graph transformer trained on datasets spanning ambient and high-pressure configurations alike. Tuga-FF is the most accurate model in the benchmark at ambient pressure, with an energy error of 13.6 meV/atom, and maintains an error of just 21 meV/atom at 150 GPa. We believe this benchmark will encourage the community to train higher-accuracy models and generate richer datasets, advancing machine-learned interatomic potentials as a whole.