Kenji Nishimura1,Ken-ichi Saitoh2
National Institute of Advanced Industrial Science and Technology1,Kansai University2
Kenji Nishimura1,Ken-ichi Saitoh2
National Institute of Advanced Industrial Science and Technology1,Kansai University2
Silicon carbide (SiC) is a promising candidate for next-generation power electronics materials because of its superior electrical properties. However, SiC is a brittle material at low temperatures and is known to be a hard-to-work material. Thus, ductile mode machining, which yields a highly efficient and smooth machined surface, has been developed over several decades. To grind brittle materials in ductile mode, a basic understanding of the material’s mechanical properties is necessary, that is, mechanisms of plastic deformation, phase transformation, and crystal defect formation. Specifically, since electronic materials require precise processing, mechanical phenomena occurring at the nanoscale should be focused on.<br/> Quantum electronic structure calculations are highly accurate, but due to its high calculation cost, the computational scale is often limited to hundreds of atoms. That is why it is difficult to elucidate the mechanical properties including nanoscale phenomena at an atomistic point of view by means of first-principles calculations. Recently, some types of machine-learning interatomic potentials (ML-IAPs) have been proposed, which can reproduce the results of the first-principles calculations adequately without any empirical information or data, and have been applied to molecular dynamics (MD) simulations. It is expected for ML-IAPs to analyze crystalline defects in large systems with the same accuracy as the first-principles calculations.<br/> In this study, we attempt to create a spectral neighbor analysis potential (SNAP) for SiC from reference data obtained by the first-principles calculations. The SNAP proposed by Thompson et al as one of ML-IAPs adopts bispectrum components as descriptors to express the energy of the atomic system. Then the SNAP potential for SiC we built is applied to MD simulations to examine its reproducibility. As a result, the SNAP potential developed in this study reproduces a lattice constant, elastic modulus, and bulk modulus with higher accuracy than any other empirical ones. We confirm that stable edge dislocation cores are generated as a dislocation dipole in crystalline 3C-SiC and they properly glide in a predicted slip plane. Additionally, the Peierls stress estimated by our MD simulations agrees well with that of the previous study.