Universal Machine Learning Potentials as Effective Tools for Crystal Structure Prediction

Author: Olesya Gorbunova

Field of Study: Materials Science and Engineering

Faculty Mentors: Alexey Kolmogorov

Easel: 74

Timeslot: Afternoon

Abstract: Universal machine learning potentials (uMLPs) are a recently developed class of atomistic modeling methods that offer substantial gains in speed and computational efficiency relative to density functional theory (DFT), the conventional reference for materials simulations. Although their utility in applications such as molecular dynamics has been widely demonstrated, their performance in crystal structure prediction (CSP) has not yet been systematically benchmarked. This work evaluates nine uMLPs for CSP of inorganic crystals using global optimization across 12 compounds, including materials outside the models’ training sets. To account for the stochastic nature of global optimization, all promising candidate structures generated during the searches, including the ground state structure, are pooled and reevaluated with each uMLP. Model performance is then assessed using root-mean-square error, structural similarity, and energy proximity metrics. The results range from essentially non predictive behavior to near-DFT accuracy, indicating that some uMLPs are already promising tools for inorganic materials discovery workflows.