A theoretical study on crystallization of chalcogenides via neural network potential

Phase-change materials (PCM) have attracted wide interests in fields such as data storage and neuromorphic computing. Ge-Sb-Te alloys are a representative PCM, which show large contrast of optical and electrical properties between crystalline and amorphous phases. Atomic scale modeling of PCM has relied on ab-initio molecular dynamics (AIMD), but its large computational costs have limited simulation size and time. Neural network potential (NNP) can deal with more than thousands of atoms upto microseconds, while accurate potential energy surface can be obtained by learning the data of density functional theory (DFT).
In this work, we suggest the accurate and effective training scheme to develop reliable NNP for chalcogenides (e.g. GeTe and Ge₂Sb₂Te₅). It seems that NNP trained in the conventional way moderately predicts the basic properties but rapidly crystallizes an amorphous phase without any incubation time. We then propose a simple descriptor to diagnose it and an active learning scheme to manage it. Dependence of crystallization kinetics upon temperature and density is investigated using about 4000-atom cells. It shows that incubation and crystallization time is dependent on temperature and pressure, qualitatively consistent with experiments.