Scholarlink

<p>In the field of nuclear fusion fuelling, the precise characterization of supersonic molecular beam parameters is essential for the accurate control of plasma conditions and the examination of experimental observations. However, the direct measurement of such supersonic beam distribution properties in high-vacuum environments is challenging and complex. The establishment of a comprehensive diagnostic platform for these measurements not only involves substantial costs but also requires technical expertise. Although density profiles can be assessed using schlieren imaging techniques, the measurement of velocity profiles continues to be a complex task. This paper presents an innovative approach using deep learning strategies to create a neural network model based on the U-Net architecture. The model was developed to infer beam velocity profiles from measured beam density data with high precision. Using the simulated beam profile data, a dataset of beam density and velocity profiles under varying gas source pressures (5, 10, and 50 bar) was created for training and verification of the neural network. To improve the predictive fidelity of the model, the influence of incorporating spatial attention, channel attention, and their hybrid form, the convolutional block attention module (CBAM), into the predictive model was examined in this study. Furthermore, this research demonstrates the generalization capabilities of the model by predicting the fluid parameter profiles under gas source pressures of 20, 30, and 40 bar, which are outside the range of the training dataset. The performance of the neural network was evaluated using established metrics such as mean squared error and structural similarity index measure. In addition, a Grad-CAM visualization of the influence of attentional mechanisms on the prediction of the model helps to understand the internal mechanisms of the model and thus improve its interpretability. CBAM attention has been shown to outperform other attention mechanisms in predicting fuel beam profiles.</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

Integrated profiling of fusion fuel beams using a deep learning approach

  • 期刊论文

  • 发表日期：

    2025年7月17日

查看全文
  • Ke Xu,
  • Guoliang Xiao,
  • Chengyuan Chen,
  • Jiao Yin,
  • Jun Zhao,
  • Xingzhong Xiong,
  • Xiaolan Zou,
  • Zongyu Yang,
  • Yiren Zhu,
  • Chiyu Wang,
  • Beibin Feng,
  • Wulyu Zhong

摘要

In the field of nuclear fusion fuelling, the precise characterization of supersonic molecular beam parameters is essential for the accurate control of plasma conditions and the examination of experimental observations. However, the direct measurement of such supersonic beam distribution properties in high-vacuum environments is challenging and complex. The establishment of a comprehensive diagnostic platform for these measurements not only involves substantial costs but also requires technical expertise. Although density profiles can be assessed using schlieren imaging techniques, the measurement of velocity profiles continues to be a complex task. This paper presents an innovative approach using deep learning strategies to create a neural network model based on the U-Net architecture. The model was developed to infer beam velocity profiles from measured beam density data with high precision. Using the simulated beam profile data, a dataset of beam density and velocity profiles under varying gas source pressures (5, 10, and 50 bar) was created for training and verification of the neural network. To improve the predictive fidelity of the model, the influence of incorporating spatial attention, channel attention, and their hybrid form, the convolutional block attention module (CBAM), into the predictive model was examined in this study. Furthermore, this research demonstrates the generalization capabilities of the model by predicting the fluid parameter profiles under gas source pressures of 20, 30, and 40 bar, which are outside the range of the training dataset. The performance of the neural network was evaluated using established metrics such as mean squared error and structural similarity index measure. In addition, a Grad-CAM visualization of the influence of attentional mechanisms on the prediction of the model helps to understand the internal mechanisms of the model and thus improve its interpretability. CBAM attention has been shown to outperform other attention mechanisms in predicting fuel beam profiles.