A machine-learned interatomic potential for the study of surface-mediated defect formation in GaN

The growth of thin film semiconductors by deposition is mediated by competitive surface mechanisms. Understanding processing-mechanism relationships is critical to the prediction of defects and morphology in novel thin film materials. GaN, one of the most technologically important and studied semiconductors, took years to mature to its current relevancy through guess-and-check experimental optimization. Interatomic potentials can describe the forces controlling diffusion, reconstruction, and defect formation on surfaces.  We demonstrate a machine-learned interatomic potential (MLIP) to simulate the homoepitaxial molecular beam epitaxy synthesis of GaN. The potential, based on the atomic cluster expansion formalism, is fit to a database of experimentally informed configurations calculated by DFT and ab initio molecular dynamics. Simulated binding, diffusion, and nucleation behavior is validated against experiment and ab initio calculations. The MLIP provides an enhanced understanding of the potential energy landscape and produced improved binding energetics and morphology development when benchmarked against a classical potential. This MLIP enables the large scale study of surface dependent properties during GaN deposition, approaching the accuracy of first principles calculations. We demonstrate that MLIPs can provide deep insight into surface dynamics and establish the capability of machine-learned potentials for the inverse design of thin film semiconductors.