Scholarlink

<p>Geometric filtering with deep-learning-based validity constraints has proven useful in improving optimization convergence efficiency via shrinking geometric space. Nevertheless, in challenging aerodynamic shape design scenarios, for example, reducing the drag of bluff shapes, it is ineffective to directly embed geometric validity constraints into optimizers, and the optimization tends to fail even after a long iteration struggle. To address the issue, we propose a hybrid gradient-based aerodynamic shape optimization method that effectively leverages deep-learning-based geometric filtering to enhance optimization convergence. In this method, the optimization problem is first approximately solved using aerodynamic sensitivities and geometric validity constraints. To enhance filtering effectiveness, we improve the validity model generalizability by augmenting airfoil shape datasets. After the optimization reaches an optimal geometric sub-domain, an improved gradient-based optimization is switched on to accurately solve the problem. The effectiveness of this approach is validated in aerodynamic shape optimization of two-dimensional and three-dimensional cylindrical shapes. Both optimization cases converge smoothly, highlighting much higher effectiveness compared with the state-of-the-art aerodynamic shape optimization framework. This work facilitates effective aerodynamic shape optimization in extensively large geometric design spaces, establishing a foundation to achieve more efficient aerodynamic designs.</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

Aerodynamic shape optimization with effective deep-learning-based geometric filtering: from a cylinder to a wing

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Yan Chen,
  • Jichao Li,
  • Chao Song,
  • Hongyang Liu,
  • Zhu Zhou,
  • Jiangtao Huang,
  • Jinsheng Cai,
  • Boo Cheong Khoo

摘要

Geometric filtering with deep-learning-based validity constraints has proven useful in improving optimization convergence efficiency via shrinking geometric space. Nevertheless, in challenging aerodynamic shape design scenarios, for example, reducing the drag of bluff shapes, it is ineffective to directly embed geometric validity constraints into optimizers, and the optimization tends to fail even after a long iteration struggle. To address the issue, we propose a hybrid gradient-based aerodynamic shape optimization method that effectively leverages deep-learning-based geometric filtering to enhance optimization convergence. In this method, the optimization problem is first approximately solved using aerodynamic sensitivities and geometric validity constraints. To enhance filtering effectiveness, we improve the validity model generalizability by augmenting airfoil shape datasets. After the optimization reaches an optimal geometric sub-domain, an improved gradient-based optimization is switched on to accurately solve the problem. The effectiveness of this approach is validated in aerodynamic shape optimization of two-dimensional and three-dimensional cylindrical shapes. Both optimization cases converge smoothly, highlighting much higher effectiveness compared with the state-of-the-art aerodynamic shape optimization framework. This work facilitates effective aerodynamic shape optimization in extensively large geometric design spaces, establishing a foundation to achieve more efficient aerodynamic designs.