Scholarlink

<p>Topology optimization plays a critical role in structural design. However, stress-related problems typically involve computationally intensive sensitivity and finite element analysis, making traditional iterative methods costly and inefficient. In this study, an efficient stress-minimizing topology optimization method is proposed using a conditional generative adversarial network (cGAN) based on the residual U-shaped convolutional neural network (ResUNet) model. The von Mises stress field computed from the first iteration of the Solid Isotropic Material with Penalization (SIMP) method is incorporated into the generator as a physical prior to improve the accuracy and mechanical consistency of the generated topologies. A dataset is constructed using the SIMP method under random boundary conditions, volume fractions, and external loads, with the optimization problem solved using the Method of Moving Asymptotes (MMA). Global stress is evaluated using the p-norm function. The generative performance of convolutional neural network (CNN)-cGAN, U-shaped (U-Net)-cGAN, ResUNet-generative adversarial network (GAN), and ResUNet-cGAN models is systematically compared. The proposed method is validated on cantilever and MBB beam cases. Results show that the topologies generated by ResUNet-cGAN closely resemble those produced by the SIMP method, while significantly reducing computation time. This study demonstrates the feasibility of deep learning for efficient stress-related topology optimization.</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

Generative adversarial network for stress-minimizing topology optimization

  • 期刊论文

  • 发表日期：

    2025年9月1日

查看全文
  • Xiaomei Huang,
  • Shijiang Li,
  • Congmin Miao,
  • Liang Hou,
  • Yun Chen

摘要

Topology optimization plays a critical role in structural design. However, stress-related problems typically involve computationally intensive sensitivity and finite element analysis, making traditional iterative methods costly and inefficient. In this study, an efficient stress-minimizing topology optimization method is proposed using a conditional generative adversarial network (cGAN) based on the residual U-shaped convolutional neural network (ResUNet) model. The von Mises stress field computed from the first iteration of the Solid Isotropic Material with Penalization (SIMP) method is incorporated into the generator as a physical prior to improve the accuracy and mechanical consistency of the generated topologies. A dataset is constructed using the SIMP method under random boundary conditions, volume fractions, and external loads, with the optimization problem solved using the Method of Moving Asymptotes (MMA). Global stress is evaluated using the p-norm function. The generative performance of convolutional neural network (CNN)-cGAN, U-shaped (U-Net)-cGAN, ResUNet-generative adversarial network (GAN), and ResUNet-cGAN models is systematically compared. The proposed method is validated on cantilever and MBB beam cases. Results show that the topologies generated by ResUNet-cGAN closely resemble those produced by the SIMP method, while significantly reducing computation time. This study demonstrates the feasibility of deep learning for efficient stress-related topology optimization.