Scholarlink

<p>Tungsten plasma facing materials (W-PFMs) are essential in tokamak divertor systems, where efficient heat removal and control of retained hydrogen isotopes (HIs) directly affect component lifetime and reactor safety. The finite element method (FEM) incurs long computation times and high memory usage when simulating temperature and deuterium (D) diffusion-trapping in W-PFMs used in nuclear fusion tokamak devices. To address these challenges, this study proposes an efficient physics-informed neural network (PINN) approach. We first optimized the network architecture and hyperparameters to develop an optimal PINN model. The optimized PINN was then validated against FEM results under both interpolative and extrapolative heat flux and D irradiation conditions. The results show that the relative error between the PINN and FEM remains below 5%, satisfying accuracy requirements while significantly reducing computation time and memory usage. Furthermore, we discuss potential industrial applications and future development prospects. Based on existing literature, we discuss the advantages of the PINN in the nuclear fusion industry, the limitations of this work, and outline directions for improvement. Finally, drawing on recent advancements in machine learning and signal processing for the industrial field, we propose potential future research directions to integrate this study with emerging industrial technologies.</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

Application of physics-informed neural network to calculate heat transfer and deuterium diffusion-trapping in tungsten plasma facing materials

  • 期刊论文

  • 发表日期：

    2025年8月2日

查看全文
  • Pengcheng Guo,
  • Hanfeng Song,
  • Yunhui Tang,
  • Bing Wang,
  • Bo Wang

摘要

Tungsten plasma facing materials (W-PFMs) are essential in tokamak divertor systems, where efficient heat removal and control of retained hydrogen isotopes (HIs) directly affect component lifetime and reactor safety. The finite element method (FEM) incurs long computation times and high memory usage when simulating temperature and deuterium (D) diffusion-trapping in W-PFMs used in nuclear fusion tokamak devices. To address these challenges, this study proposes an efficient physics-informed neural network (PINN) approach. We first optimized the network architecture and hyperparameters to develop an optimal PINN model. The optimized PINN was then validated against FEM results under both interpolative and extrapolative heat flux and D irradiation conditions. The results show that the relative error between the PINN and FEM remains below 5%, satisfying accuracy requirements while significantly reducing computation time and memory usage. Furthermore, we discuss potential industrial applications and future development prospects. Based on existing literature, we discuss the advantages of the PINN in the nuclear fusion industry, the limitations of this work, and outline directions for improvement. Finally, drawing on recent advancements in machine learning and signal processing for the industrial field, we propose potential future research directions to integrate this study with emerging industrial technologies.