Constructing Machine-learned Interatomic Potentials for SiC: MD Analyses of Dislocation

Abstract

In Molecular Dynamics (MD), the motion of atoms is calculated based on the interatomic potential, which is a mathematical model of the interatomic interaction. Recently, some types of machine-learning interatomic potentials (ML-IAPs) are being proposed, which will reproduce adequately results of first-principles calculation without any empirical assumption or data. First-principles calculations should be avoided for large simulation models, such as edge dislocations, due to its high calculation costs. However, it is expected for ML-IAPs to analyze crystalline defects in large systems with the same accuracy as first-principles calculations. In this research, SNAP, which is one of ML-IAPs, was created for silicon carbide with cubic crystalline structure (3C-SiC), and it was examined on the reproducibility when it is applied to MD computation. It is confirmed that the SNAP we built from scratch has higher reproducibility than any empirical one in terms of lattice constant, elastic modulus, and bulk modulus. The edge dislocation cores properly moved along the predicted slip plane of 3C-SiC. In addition, the Peierls stress estimated by SNAP agrees well with that of covalent materials which is usually high.