Crystallization of amorphous GeTe simulated by neural network potential addressing medium-range order

Abstract For the last decade, crystallization kinetics of phase change materials has been intensively investigated by molecular dynamics simulations. In particular, recent machine-learning potentials have advanced microscopic understanding of crystallization behavior of phase change materials by overcoming the computational limit of the density functional theory (DFT). Here we develop neural network potentials (NNP) for GeTe, an archetypal phase change material, and study the crystallization of amorphous GeTe. Consistently with the previous literature, we find that NNP results in a very short incubation time and the crystallization completes within a few nanoseconds, which is at variance with experimental measurement as well as DFT simulations. We show that such deficiencies of NNP originate from overly flat fourfold rings in the amorphous structure. By including explicitly relaxation paths from flat to puckered fourfold rings, we generate a modified NNP, which produces medium-range orders that are more consistent with DFT. Using the modified NNP, crystallization simulations are performed at two densities that represent partially or fully amorphized devices, and temperatures ranging from 500 to 650 K. At both densities, finite incubation times are clearly observed. In particular, the incubation time under the partially amorphized device condition is found to be 7 or 17 ns, which is consistent with experiments. By proposing a method to develop NNPs with correct medium-range order, this work will contribute to simulating phase change materials more accurately and realistically.