Scholarlink

<p>To address the poor accuracy of empirical potentials of silicon nitride material under high temperature, and ab initio molecular dynamics (AIMD) efficiency is low. The deep learning potential of silicon nitride material with a high accuracy-to-cost ratio under high temperature is established, based on the coupling method of first-principles and deep learning. Combined with the first-principles and density functional theory (DFT), the initial configuration of silicon nitride material is constructed. Molecular dynamics simulations are performed of silicon nitride material under high temperature, as well as atomic coordinates, interatomic forces and energies are collected of each femtosecond to construct a dataset. The deep learning network is constructed to fit the initial deep learning potential of silicon nitride material. Based on DP-GEN, it is optimized to obtain the final deep learning potential. The deep learning potential is used to simulate molecular dynamics and compare with AIMD. The predicted root mean square error (RMSE) of atomic energy is 0.38&#xa0;meV/atom and the RMSE of interatomic force is 46&#xa0;meV/Å, while the speed is improved by 5–6 orders of magnitude.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Study of the Deep Learning Potential of Silicon Nitride Materials Under High Temperatures

  • 期刊论文

  • 发表日期：

    2025年6月30日

查看全文
  • Dongling Yu,
  • Liangyu Zhu,
  • Xiaobin Cui,
  • Xiang Ning,
  • Jiao Li,
  • Dahai Liao

摘要

To address the poor accuracy of empirical potentials of silicon nitride material under high temperature, and ab initio molecular dynamics (AIMD) efficiency is low. The deep learning potential of silicon nitride material with a high accuracy-to-cost ratio under high temperature is established, based on the coupling method of first-principles and deep learning. Combined with the first-principles and density functional theory (DFT), the initial configuration of silicon nitride material is constructed. Molecular dynamics simulations are performed of silicon nitride material under high temperature, as well as atomic coordinates, interatomic forces and energies are collected of each femtosecond to construct a dataset. The deep learning network is constructed to fit the initial deep learning potential of silicon nitride material. Based on DP-GEN, it is optimized to obtain the final deep learning potential. The deep learning potential is used to simulate molecular dynamics and compare with AIMD. The predicted root mean square error (RMSE) of atomic energy is 0.38 meV/atom and the RMSE of interatomic force is 46 meV/Å, while the speed is improved by 5–6 orders of magnitude.