Using neural network potential to study point defect properties in multiple charge states of GaN with nitrogen vacancy

TL;DR

This paper introduces a neural network potential scheme that efficiently predicts properties of charged point defects in GaN, enabling large-scale and accurate simulations of defect behaviors in semiconductors.

Contribution

The study presents a minimally modified neural network potential scheme capable of accurately modeling multiple charge states of point defects in GaN.

Findings

Accurately predicted total energies and atomic forces for various charge states.

Reproduced phonon band structures and thermodynamic properties.

Enabled large-scale defect analysis with reduced computational cost.

Abstract

Investigation of charged defects is necessary to understand the properties of semiconductors. While density functional theory calculations can accurately describe the relevant physical quantities, these calculations increase the computational loads substantially, which often limits the application of this method to large-scale systems. In this study, we propose a new scheme of neural network potential (NNP) to analyze the point defect behavior in multiple charge states. The proposed scheme necessitates only minimal modifications to the conventional scheme. We demonstrated the prediction performance of the proposed NNP using wurzite-GaN with a nitrogen vacancy with charge states of 0, 1+, 2+, and 3+. The proposed scheme accurately trained the total energies and atomic forces for all the charge states. Furthermore, it fairly reproduced the phonon band structures and thermodynamics properties of the defective structures. Based on the results of this study, we expect that the proposed scheme can enable us to study more complicated defective systems and lead to breakthroughs in novel semiconductor applications.