SiC-TGAP: A machine learning interatomic potential for radiation damage simulations in 3C-SiC

Abstract: Silicon carbide (SiC) has long been a subject of study for its application in harsh environments. Existing empirical interatomic potentials for 3C-SiC show significant discrepancies in predicting the properties that are crucial in describing the evolution of defects generated in collision cascades. We present a Gaussian approximation potential model for 3C-SiC (TGAP) trained by two-body and the turboSOAP many-body descriptors. The dataset covers crystalline, liquid and amorphous phases. To accurately capture defect dynamics, twenty-one defect types have been included in the dataset. TGAP captures the experimentally observed decomposition of carbon atoms in the liquid phase at atmospheric pressure, while also accurately reproducing the radial distribution function of the high-temperature homogeneous liquid phase across a range of densities. Moreover, it predicts the melting point in very good agreement with density functional theory and experiments. The potential is equipped with the Nordlund-Lehtola-Hobler repulsive potential to capture the high repulsion of recoils in the collision cascades. TGAP provides an accurate tool for atomistic simulation of radiation damage in cubic SiC.