Evaluating Mechanical Property Prediction across Material Classes using Molecular Dynamics Simulations with Universal Machine-Learned Interatomic Potentials

TL;DR

This study evaluates the accuracy of six universal machine-learned interatomic potentials in predicting material responses across diverse classes using molecular dynamics, highlighting their strengths, limitations, and transferability issues.

Contribution

The paper compares different MLIP architectures across multiple materials, identifying top performers and analyzing factors affecting model transferability and accuracy.

Findings

Top models achieved average errors of 41-47%.

All models underestimated bulk modulus and overestimated thermal expansion.

Model transferability depends heavily on dataset homogeneity and structural representation.

Abstract

We assess the accuracy of six universal machine-learned interatomic potentials (MLIPs) for predicting the temperature and pressure response of materials by molecular dynamics simulations. Accuracy is evaluated across 13 diverse materials (nine metal-organic frameworks and four inorganic compounds), computing bulk modulus, thermal expansion, and thermal decomposition. These MLIPs employ three different architectures (graph neural networks, graph network simulators, and graph transformers) with varying training datasets. We observe qualitative accuracy across these predictions but 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%, 44%, and 47%, respectively. Despite strong overall performance, questions arise about the limits of model transferability: dataset homogeneity and structural representation dominate model accuracy. Our results show that certain architectures can compensate for biases, a step closer to truly universal MLIPs.