Enhancing universal machine learning potentials with polarizable long-range interactions

Long-range interactions are crucial in determining the behavior of chemical systems in various environments. Accurate predictions of physical and chemical phenomena at the atomic level hinge on accurate modeling of these interactions. Here, we present a framework that substantially enhances the predictive power of machine learning interatomic potentials by incorporating explicit polarizable long-range interactions with an equivariant graph neural network short-range potential. The pretrained universal model, applicable across the entire periodic table, can achieve first-principles accuracy. This versatile model has been further applied to diverse areas of research, including the study of mechanical properties, ionic diffusivity in solid-state electrolytes, ferroelectricity, and interfacial reactions, demonstrating its broad applicability and robustness.
Comment: 13 pages, 5 figures