Scholarlink

<p>The ability to predict physical fields quickly and accurately is crucial for engineering research. Traditional methods for computing physical fields typically rely on simulations and experiments, which can face challenges like complex modeling, dependence on dense grids, and high computational costs. Engineering projects often involve complex data, such as unstructured grids with millions of nodes and high-dimensional point clouds. To overcome these challenges, this paper introduces a global information-guided framework for physical field reconstruction (GIG-GNN) that leverages graph neural networks. The proposed method builds on the rapidly developing field of graph neural networks. GIG-GNN utilize physical field data at mesh vertices, integrates global field features, and employs adjacency matrices to capture neighborhood relationships between vertices. This study demonstrates the generalization capability of GIG-GNN by reconstructing the flow field of the NAC0012 airfoil and the stress field of the head sheave. The results, based on assessments of prediction accuracy and computational efficiency, demonstrate that the model can quickly and accurately capture flow field characteristics under different boundary conditions. Furthermore, the model performs effectively in reconstructing physical fields.</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 global information-guided framework based on graph neural networks for physical field reconstruction

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Muchen Wang,
  • Pengwei Liang,
  • Yong Pang,
  • Shuai Zhang,
  • Zhuangzhuang Gong,
  • Xueguan Song

摘要

The ability to predict physical fields quickly and accurately is crucial for engineering research. Traditional methods for computing physical fields typically rely on simulations and experiments, which can face challenges like complex modeling, dependence on dense grids, and high computational costs. Engineering projects often involve complex data, such as unstructured grids with millions of nodes and high-dimensional point clouds. To overcome these challenges, this paper introduces a global information-guided framework for physical field reconstruction (GIG-GNN) that leverages graph neural networks. The proposed method builds on the rapidly developing field of graph neural networks. GIG-GNN utilize physical field data at mesh vertices, integrates global field features, and employs adjacency matrices to capture neighborhood relationships between vertices. This study demonstrates the generalization capability of GIG-GNN by reconstructing the flow field of the NAC0012 airfoil and the stress field of the head sheave. The results, based on assessments of prediction accuracy and computational efficiency, demonstrate that the model can quickly and accurately capture flow field characteristics under different boundary conditions. Furthermore, the model performs effectively in reconstructing physical fields.