Scholarlink

<p>Recently, machine learning techniques that employ neural networks have been extensively utilized in the study of interatomic potentials, exhibiting notable accuracy and efficiency. Among various materials, the rare earth element yttrium (Y) has gained increasing significance due to its extensive applications in both high-carbon technologies, such as battery manufacturing, and low-carbon technologies, such as new energy solutions. However, the substantial computational expense associated with ab initio molecular dynamics (AIMD) limits extensive investigations into these materials, thereby hindering their practical applications. To address this challenge, we employed a machine learning molecular dynamics approach, which achieves at least a 100-fold improvement in computational efficiency compared to conventional AIMD calculations while maintaining high accuracy. We systematically compared the deep potential (DP) model's predictions of lattice constants, melting point, defect formation energies, surface energies, elastic constants, and thermal transport properties with density functional theory (DFT) results to validate the exceptional accuracy of the developed DP model in predicting physical properties, which are approximately consistent with DFT results. Specifically, the predicted equilibrium lattice constant by the DP model deviates by less than 0.15% from the DFT result, the vacancy formation energy is 0.3 eV lower, and the interstitial formation energy is 0.9 eV higher. In terms of elastic properties, the absolute error in the predictions compared to DFT results is less than 4.7 GPa for over 90% of the cases. Furthermore, molecular dynamics simulations based on the DP framework demonstrated that increasing temperature significantly enhances the diffusivity of Y atoms, reaching a self-diffusion coefficient of 4.9 × 10<sup>-12</sup>m<sup>2</sup>/s at the melting point, which exhibits diffusion behavior similar to that of typical hexagonal close-packed metals. Further analysis revealed that the phonon dispersion curves were accurately predicted, particularly in the low-frequency and high-frequency branches, with a relative error in lattice thermal conductivity as low as 3.1%. By developing a reliable and efficient structural prediction framework, our research paves the way for large-scale and open-ended exploration of complex material systems.</p> Graphical abstract <p></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

Deep learning molecular dynamics for large-scale insights into yttrium's structural, thermal and diffusion properties

  • 期刊论文

  • 发表日期：

    2025年5月16日

查看全文
  • Li Ma,
  • Haowen Wang,
  • Yongzhi Wu,
  • Keyuan Chen,
  • Haoxiang Zhang,
  • Xingkao Zhang,
  • Ju Rong,
  • Yudong Sui,
  • Xiaohua Yu,
  • Jing Feng

摘要

Recently, machine learning techniques that employ neural networks have been extensively utilized in the study of interatomic potentials, exhibiting notable accuracy and efficiency. Among various materials, the rare earth element yttrium (Y) has gained increasing significance due to its extensive applications in both high-carbon technologies, such as battery manufacturing, and low-carbon technologies, such as new energy solutions. However, the substantial computational expense associated with ab initio molecular dynamics (AIMD) limits extensive investigations into these materials, thereby hindering their practical applications. To address this challenge, we employed a machine learning molecular dynamics approach, which achieves at least a 100-fold improvement in computational efficiency compared to conventional AIMD calculations while maintaining high accuracy. We systematically compared the deep potential (DP) model's predictions of lattice constants, melting point, defect formation energies, surface energies, elastic constants, and thermal transport properties with density functional theory (DFT) results to validate the exceptional accuracy of the developed DP model in predicting physical properties, which are approximately consistent with DFT results. Specifically, the predicted equilibrium lattice constant by the DP model deviates by less than 0.15% from the DFT result, the vacancy formation energy is 0.3 eV lower, and the interstitial formation energy is 0.9 eV higher. In terms of elastic properties, the absolute error in the predictions compared to DFT results is less than 4.7 GPa for over 90% of the cases. Furthermore, molecular dynamics simulations based on the DP framework demonstrated that increasing temperature significantly enhances the diffusivity of Y atoms, reaching a self-diffusion coefficient of 4.9 × 10-12m2/s at the melting point, which exhibits diffusion behavior similar to that of typical hexagonal close-packed metals. Further analysis revealed that the phonon dispersion curves were accurately predicted, particularly in the low-frequency and high-frequency branches, with a relative error in lattice thermal conductivity as low as 3.1%. By developing a reliable and efficient structural prediction framework, our research paves the way for large-scale and open-ended exploration of complex material systems.

Graphical abstract