Accuracy and Uncertainty Quantification of Phase Diagrams Calculated from Universal Machine Learning Interatomic Potentials

Authors:
Preprint
Publisher website
Google Scholar
Abstract
State-of-the-art universal machine learning interatomic potentials (UMLIPs) provide accurate surrogate potential energy surfaces based on Density functional theory (DFT) data and enable rapid and efficient calculations of phase diagrams and prediction of thermodynamic stability. In this work, we evaluate the performance of the most recent UMLIPs for phase diagram prediction in different representative binary systems and compare them with phase diagrams from both DFT calculations and experimental assessments. While DFT results are more robust, we show that the best-performing UMLIPs achieve statistically similar accuracy to DFT for phase diagrams in the tested binary systems, as their errors are similar in magnitude. Errors in finite-temperature free energies from MLIP calculations are assessed, and uncertainty quantification is used to evaluate their impact on phase boundaries. We demonstrate that uncertainty quantification can reveal potentially stable phases within MLIP uncertainties, leading to more robust predictions and systematic integration with experimental data.