Evaluating mechanical property prediction across material classes using molecular dynamics simulations with universal machine-learned interatomic potentials.

Simulating the mechanical and thermal properties of materials requires accurate treatment of interatomic interactions, yet quantum-mechanical methods can be computationally prohibitive for the time scales needed. Universal machine-learned interatomic potentials (MLIPs) offer a promising alternative, but their reliability for dynamics across diverse material classes remains largely untested. Here, we assess the accuracy of six universal MLIPs for predicting the temperature and pressure response of 13 diverse materials (nine metal-organic frameworks and four inorganic compounds), computing bulk modulus, thermal expansion, and thermal decomposition. These MLIPs employ three architectures (graph neural networks, graph network simulators, and graph transformers) with varying training datasets. We observe qualitative agreement with experiment, outperforming UFF4MOF, but also systematic underestimation of bulk modulus and overestimation of thermal expansion across all models, consistent with potential energy surface softening. From all tested models, three top performers arise; 'MACE-MP-0a', 'fairchem_OMAT', and 'Orb-v3', with average error across metrics and materials of 41%, 43%, and 43%, respectively. Beyond overall performance, dataset homogeneity and structural representation dominate model accuracy, while certain architectures can compensate for biases, a step closer to truly universal MLIPs.