High-Accuracy Machine-Learned Interatomic Potentials for the Phase Change Material Ge3Sb6Te5

Ge3Sb6Te5 with an emergingoff-stoichiometriccomposition has been proven to have characteristic properties of phasechange materials (PCMs) by experiments. However, the detailed mechanismof the phase transition and the highly temperature-dependent kineticsof its crystallization process have yet to be resolved at the atomicscale. In this work, we develop an artificial neural network-basedpotential (NNP) to accelerate the molecular dynamics (MD) simulationof Ge3Sb6Te5 without sacrificingthe quantum mechanical accuracy. Overall, the comprehensive structuralinformation predicted by NNP shows an excellent agreement with thatof ab initio MD (AIMD), indicating the reliability of the proposedmethod. Based on the well-trained NNP, an MD simulation can be adoptedto simulate Ge3Sb6Te5 with over 10,000atoms with high efficiency and accuracy, which is beyond the reachof AIMD. Subsequently, a further NNP-based MD simulation with longtimescales is carried out, which successfully captures the rapid transitionof the crystallization and perfectly reproduces the crystallizationprocess consistent with experiments. This work provides a novel atomic-levelsimulation and analysis approach to the complex Ge3Sb6Te5 and makes it possible to simulate the realnonvolatile memory device.