Charge-Dependent Machine-Learned Interatomic Potentials, Atomic Cluster Expansions, and Applications

Machine-learned interatomic potentials (MLIPs) predominantly treat interatomic interactions as local or semi-local. These approaches are often suitable for molecular dynamics simulations of material and chemical systems where there is minimal charge transfer and long-range interactions are effectively screened. Developments in machine-learned electronic property prediction and dynamic charge equilibration enable molecular dynamics simulations that incorporate long-range electrostatics while leveraging improvements from ML approaches. It is shown that such approaches may be used in regimes where long-range interactions and charge transfer cannot be neglected. Furthermore, these approaches reach the efficiency and accuracy needed to simulate systems with charged species at appropriate length scales. It is shown that this is beneficial for modeling systems like water. In this work, we explore the benefits of charge-dependent local and semi-local MLIPs. The methodology for using these approaches in practice is outlined.