Scholarlink

<p>Finite element method-based hemming simulations are prone to high setup effort and computational time, leading to delays in industrial product development. In this study, we propose a layer-based surrogate model using graph convolutional network (GCN) to predict stress distribution through the sheet metal thickness. The inputs comprise material properties, geometric and process parameters. The model is then trained on stress tensors derived from the end of hemming simulations generated with varied geometries and thicknesses. Using the trained GCN model, stress distributions are predicted with nearly 90% accuracy on and near the surface, where tension and compression usually dominate, indicating a clear geometric influence there. With a reduction in computation time of over 99%, these results indicate a practical surrogate for hemming simulations and motivate future work on enhanced nodal features and model generalization.</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

Layer-Based Stress Prediction Using Graph Neural Networks for Hemming Simulations

  • 期刊论文

  • 发表日期：

    2026年5月24日

查看全文
  • Naveen Paramasivam,
  • Saleh Yazdani,
  • Marcus Wagner

摘要

Finite element method-based hemming simulations are prone to high setup effort and computational time, leading to delays in industrial product development. In this study, we propose a layer-based surrogate model using graph convolutional network (GCN) to predict stress distribution through the sheet metal thickness. The inputs comprise material properties, geometric and process parameters. The model is then trained on stress tensors derived from the end of hemming simulations generated with varied geometries and thicknesses. Using the trained GCN model, stress distributions are predicted with nearly 90% accuracy on and near the surface, where tension and compression usually dominate, indicating a clear geometric influence there. With a reduction in computation time of over 99%, these results indicate a practical surrogate for hemming simulations and motivate future work on enhanced nodal features and model generalization.