Charge-dependent machine-learned interatomic potentials for modeling charged surfaces and defects

Machine learned interatomic potentials (MLIPs) are computational models that leverage machine learning techniques to predict properties of materials at the atomistic scale. Typical MLIPs rely on local descriptions of atomic interactions, but Coulombic interactions are long-range. In systems where Coulomb interactions are not effectively screened, the MLIPs may not predict material properties accurately. In this work, flexible-charge MLIPs for solid state systems such as Galium Nitride(GaN) are developed that contain explicit Coulomb interactions to enhance accuracy. It is shown that extended-Lagrangian charge equilibration schemes along with machine-learned charge properties can be used to predict emergent materials properties sensitive to charge, such as defect energetics and charged surface energetics of GaN. Important steps in the development of such models are outlined, including the curation of a training dataset for flexible-charge MLIPs. Benefits and limits of using these modeling approaches for defects and surfaces of semiconductors are explored. Ongoing model development and new approaches for molecular dynamics simulations of GaN are outlined.