2025 AIChE Annual Meeting
(58g) Evaluating the Accuracy of Machine Learning Interatomic Potentials in Calculating Elastic Properties
Authors
Luis Barroso-Luque, Fundamental AI Research (FAIR) Meta
Muhammed Shuaibi, Fundamental AI Research (FAIR) Meta
John Kitchin, Carnegie Mellon University
Computing elastic properties is essential for designing materials with desirable mechanical properties. Ab initio methods, such as density functional theory (DFT), are commonly used to compute elastic properties, guiding experimental alloy design. Machine learning interatomic potentials (MLIPs) have emerged as efficient and generalizable surrogates for DFT, enabling rapid screening across a vast compositional and elemental space. MLIPs are typically evaluated based on the accuracy of predicting energies and forces in atomistic simulations, but these metrics do not always translate to more accurate physical property prediction. Instead, we evaluate their accuracy in predicting relevant elastic properties, such as shear and bulk moduli. In this work, we assess the accuracy of MLIPs using a dataset from the Materials Project, comprising 12,000 structures, including unary, binary, ternary, and higher-order alloys. Additionally, we evaluate the parallelization efficiency of multiple software packages that integrate MLIPs into elastic property calculation workflows. Our analysis shows that the accuracy in predicting the elastic properties is strongly influenced by the smoothness of the potential energy surface (PES) constructed by the MLIP. We find that the eSEN MLIP, which was designed to have a smoothly-varying PES, has the lowest errors with 5.6 and 7.6 GPa on the bulk and shear moduli, respectively. This work introduces a new experimentally relevant metric for evaluating the accuracy of MLIPs and highlights their potential for elastic property calculation, advancing machine learning in computational materials science.