Constructing Machine-learned Interatomic Potentials for Covalent Bonding Materials and MD Analyses of Dislocation and Surface

Abstract

As machine learning potentials for molecular dynamics (MD) simulations, Spectral Neighbor Analysis Potential (SNAP) and quadratic SNAP (qSNAP) were constructed for silicon (Si) and silicon carbide (SiC). The reproducibility of the basic material properties about perfect crystal, free surface and dislocation cores in Si and 3C-SiC was investigated. The coefficients of SNAP and qSNAP were optimized using liner regression to present energy and force obtained by DFT. In addition, hyperparameters (cutoff length and weights for optimization, here) were determined using genetic algorithm to reproduce elastic moduli obtained by DFT. Lattice constant and elastic moduli of Si crystal by MD using our SNAP or qSNAP agree well with the values of DFT, and they have higher accuracy than those by any empirical potential. Additionally, melting point and specific heat at constant pressure were calculated by MD correctly. Especially in qSNAP of Si, the surface energy of {100} and {111} planes and the reconstructed {100} surface structure were almost reproduced. For 3C-SiC, SNAP reproduces lattice constant and elastic moduli of DFT. Furthermore, edge dislocation cores were generated successfully. However, the potentials we constructed have insufficient reproducibility in the plastic region, so it is necessary to continue development.

Full Paper