A unified deep neural network potential capable of predicting thermal conductivity of silicon in different phases

Abstract

Molecular dynamics (MD) simulations have been extensively used to predict thermal properties, but simulating different phases with similar precision using a unified force field is often difficult because of the lack of accurate and transferrable interatomistic potential fields. As a result, this issue has become a major barrier to predicting the phase change of materials and their transport properties with atomistic-level modeling techniques. Recently, machine learning–based algorithms have emerged as promising tools to develop accurate potentials for MD simulations. In this work, we approach the problem of predicting the thermal conductivity of silicon in different phases by performing MD simulations with a deep neural network potential (NNP). This NNP is trained with ab initio data of silicon in the crystalline, liquid, and amorphous phases. The accuracy of our potential is first validated through reproducing the atomistic structures during the phase transition, where other empirical potentials usually fail. The thermal conductivity of different phases is then calculated, showing a good agreement with the experimental results and ab initio calculation results. Our work shows that a unified neural network–based potential can be a promising tool for studying phase change and thermal transport of materials with high accuracy.