A robust artificial neural network potential for Si(001)

Despite being is a powerful method for material research, density functional theory (DFT) finds limited applications at large length scale or long time scale, at which realistic properties emerge, thanks to its computational expensiveness. Thus, development of simple potentials whose accuracies are at the level of DFT method and that are fast enough for large length scale and long time scale simulations is of interest. We will report our development of a robust artificial neural network (ANN) potential for Si. The ANN potential was trained by using about 15000 data points obtained from ab initio molecular dynamic simulations of 2×2×6 and 2×2×8 Si(001) slabs at 500 K and 1000 K. The trained ANN potential is capable of producing energy in agreement with DFT one (within few meV per atom). More importantly, it can reproduce different reconstruction phases of Si(001) surface, i.e. (2×1), p(2×1), p(2×2), and c(4×2). Molecular dynamics simulations of Si(001) using the ANN potential show the existence of p(2×1) phase at low temperature (10-20K) and p(2×2) or c(4×2) phase at high temperature (100-150K).