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

Abstract

Ge3Sb6Te5 with an emerging off-stoichiometric composition has been proven to have characteristic properties of phase change materials (PCMs) by experiments. However, the detailed mechanism of the phase transition and the highly temperature-dependent kinetics of its crystallization process have yet to be resolved at the atomic scale. In this work, we develop an artificial neural network-based potential (NNP) to accelerate the molecular dynamics (MD) simulation of Ge3Sb6Te5 without sacrificing the quantum mechanical accuracy. Overall, the comprehensive structural information predicted by NNP shows an excellent agreement with that of ab initio MD (AIMD), indicating the reliability of the proposed method. Based on the well-trained NNP, an MD simulation can be adopted to simulate Ge3Sb6Te5 with over 10,000 atoms with high efficiency and accuracy, which is beyond the reach of AIMD. Subsequently, a further NNP-based MD simulation with long timescales is carried out, which successfully captures the rapid transition of the crystallization and perfectly reproduces the crystallization process consistent with experiments. This work provides a novel atomic-level simulation and analysis approach to the complex Ge3Sb6Te5 and makes it possible to simulate the real nonvolatile memory device.

Machine learning interatomic potential: Bridge the gap between small-scale models and realistic device-scale simulations

“… 78 These two methods now also are the main technologies of MLIPs. 79 , 80 , 81 , 82 , 83 , 84 , 85 , 86 , 87 , 88 , 89 , 90 , 91 , 92 , 93 , 94 Previously, there are some review articles on MLIPs, for example, Behler et al. 95 , 96 , 97 outlined the timeline for the evolution of neural network potentials by classifying schemes for MLIPs into four generations; Deringer et al.…”