Umbrella sampling with machine learning potentials applied for solid phase transition of GeSbTe

Abstract

The free energy difference is the key role to predict reaction rates of solid phase transition. Umbrella sampling is an effective method for enhanced sampling, especially in the region with high potential energy, and has been widely applied in studies of protein folding and aggregation, etc. Here, we demonstrated a strategy that integrates umbrella sampling with machine learning potentials (US–MLPs) for the study of solid phase transition. The method succeeded in theoretically investigating the phase transition process of GeSbTe from vacancy ordered cubic phase to quasi–amorphous phase. The initial and final geometries of the phase transition process were obtained by using ab initio molecular dynamics (AIMD). Subsequently, the free energy barrier and location of transition state of the above phase transition process were investigated by the US–MLPs method. During the sampling process, the constrained molecular dynamics (MD) combining with MLPs lasted 51 ns and generated 510,000 structure. For comparing with experiment, our method shows advantages over previous simulation method such as the climbing image nudged elastic band (CI–NEB) method.