Go to ICLR 2025 Workshop AI4MAT homepageEvaluating Universal Interatomic Potentials for Molecular Dynamics of Real-World Minerals
Universal interatomic potentials (UIPs) have emerged as promising models for capturing complex atomic interactions across diverse material families through graph-based representations. Recent UIP architectures, trained on density functional theory (DFT) trajectories spanning the periodic table, have demonstrated accuracy in energy and force predictions for 0 K structures. However, their efficacy for finite temperature molecular dynamics (MD) simulations of experimentally verified materials under physical conditions remains unexplored. We present a comprehensive evaluation of six state-of-the-art UIPs (CHGNET, M3GNET, MACE, MATTERSIM, SEVENNET, ORB) on a curated dataset, namely AMCSD-MD-2.4K, comprising ~2,400 minerals with experimentally validated crystal structures and densities from the American Mineralogist Crystal Structure Database. Our analysis comprises two components: (1) a systematic comparison of model performance across the mineral dataset, and (2) a quantitative assessment of temporal evolution during MD simulations, analyzing structural properties including density and lattice parameters. Our evaluation reveals significant performance variations among UIPs, with ORB and SEVENNET achieving completion rates of 99.96% and 98.75% respectively, while CHGNET completed only 7% of simulations. Furthermore, none of the models achieved the empirically accepted structural variation threshold of ±2.5%, with MACE, MATTERSIM, SEVENNET, and ORB showing comparatively better accuracy (R$^2$ > 0.8) in density predictions. This evaluation framework establishes rigorous benchmarks for assessing UIP performance in realistic atomistic simulations of mineral systems.