Scholarlink

<p>In recent years, physics-informed neural network (PINN) has emerged as a major research focus for solving physical systems. However, existing research primarily addresses solving partial differential equations under ideal conditions, with limited exploration into its application for engineering analysis in complex scenarios. To address this gap, we present a novel deep learning framework (called FEM-PINN) that integrates the finite element method (FEM) with PINN to build surrogate models for predicting the performances of engineering structures. First, the finite element models of the structures, obtained after discretization and comprising various elements, are efficiently represented as graph data. Graph neural network (GNN) is subsequently introduced to tackle the crucial scalability challenge inherent in traditional deep learning models like multilayer perceptron (MLP), empowering FEM-PINN to predict the performances of engineering structures with different topologies simultaneously. Furthermore, a well-defined GNN design space and a controlled random search approach were employed to derive the optimal GNN model for this performance prediction task. Finally, rich physical information is constructed directly using the FEM equations to provide additional supervision for the training of our GNN model. In numerical experiments involving car frames, car roofs and wheel hubs, models with FEM-PINN achieved higher precisions (99.63, 99.56, and 99.37%) than models without FEM-PINN, demonstrating that incorporating the supervisory information provided by FEM equations can enhance the performance of the model and accelerate the convergence of the training process. Through comparisons with MLP and FEM, we highlight the significant advantage of FEM-PINN in terms of inference efficiency.</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

FEM-PINN: integrating finite element method and physics-informed neural network for performance prediction of engineering structures via graph neural network

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Wenbin Hou,
  • Shuyu Lv,
  • Yongcheng Li,
  • Changsheng Wang

摘要

In recent years, physics-informed neural network (PINN) has emerged as a major research focus for solving physical systems. However, existing research primarily addresses solving partial differential equations under ideal conditions, with limited exploration into its application for engineering analysis in complex scenarios. To address this gap, we present a novel deep learning framework (called FEM-PINN) that integrates the finite element method (FEM) with PINN to build surrogate models for predicting the performances of engineering structures. First, the finite element models of the structures, obtained after discretization and comprising various elements, are efficiently represented as graph data. Graph neural network (GNN) is subsequently introduced to tackle the crucial scalability challenge inherent in traditional deep learning models like multilayer perceptron (MLP), empowering FEM-PINN to predict the performances of engineering structures with different topologies simultaneously. Furthermore, a well-defined GNN design space and a controlled random search approach were employed to derive the optimal GNN model for this performance prediction task. Finally, rich physical information is constructed directly using the FEM equations to provide additional supervision for the training of our GNN model. In numerical experiments involving car frames, car roofs and wheel hubs, models with FEM-PINN achieved higher precisions (99.63, 99.56, and 99.37%) than models without FEM-PINN, demonstrating that incorporating the supervisory information provided by FEM equations can enhance the performance of the model and accelerate the convergence of the training process. Through comparisons with MLP and FEM, we highlight the significant advantage of FEM-PINN in terms of inference efficiency.