Atomic-scale phase-field modeling with universal machine learning potentials
Abstract
Atomic-scale phase-field modeling formulates the probability densities of atomic vibrations as Gaussian distributions and derives a free energy functional using variational Gaussian theory and interatomic potentials. This framework permits per-Gaussian decomposition of the free energy, providing a description of local thermodynamic states with atomic resolution. However, existing formulations are limited to classical pairwise interatomic potentials, restricting their applicability to specific materials and compromising quantitative accuracy. In this work, we extend the atomic-scale phase-field methodology by incorporating universal machine learning interatomic potentials, thereby generalizing the free energy functional to many-body systems. This extension enhances both the accuracy and transferability of the approach. We demonstrate the method by applying it to bulk copper under NVT and NPT ensembles, where the predicted pressures and equilibrium lattice constants show excellent agreement with molecular dynamics simulations, validating the theoretical framework. Furthermore, we apply the method to {\Sigma}5(310)[001] grain boundaries in copper, enabling the visualization of local free energy distributions with atomic-scale resolution. The results reveal a pronounced free energy concentration at the grain boundary core, capturing the thermodynamic signature of the interface. This study establishes a versatile and accurate framework for atomic-scale thermodynamic modeling, significantly broadening the scope of phase-field approaches to include complex materials and defect structures.