Scholarlink

<p>Accurate prediction of transonic aerodynamics is critical for aircraft design, especially for shockwave-dominated flows. While Computational Fluid Dynamics (CFD) delivers high-fidelity solutions, its prohibitive computational cost motivates the development of efficient surrogates. This study proposes MeshGIN, a Graph Isomorphism Network (GIN)-based surrogate with global flow conditioning and learned edge feature updates, tailored for unstructured CFD meshes. The model was trained on CFD-generated datasets (121), validated against wind-tunnel data and evaluated across a range of Mach numbers and angles of attack in transonic flight regime. The model’s generalization was assessed using K-Fold Cross-Validation, benchmarked against alternative GNN architectures and a PODI reduced-order model, achieving an average <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> of 0.945 on 13 test cases and less than 8% error in lift (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(C_L\)</EquationSource> </InlineEquation>) and drag coefficients (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(C_D\)</EquationSource> </InlineEquation>). Despite these strong results, the model encountered challenges in accurately resolving shocks and predicting pitching moments (<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(C_M\)</EquationSource> </InlineEquation>), with an average error of 20%. These results demonstrate the potential and limitations of GIN-based surrogates for transonic aerodynamic analysis, offering substantial computational savings while maintaining reasonable fidelity compared to conventional CFD.</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

MeshGIN: a graph neural network framework for transonic aerodynamic predictions

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Saad Hussain,
  • Yang Pei,
  • Bagh Hussain,
  • Fahad Nawaz

摘要

Accurate prediction of transonic aerodynamics is critical for aircraft design, especially for shockwave-dominated flows. While Computational Fluid Dynamics (CFD) delivers high-fidelity solutions, its prohibitive computational cost motivates the development of efficient surrogates. This study proposes MeshGIN, a Graph Isomorphism Network (GIN)-based surrogate with global flow conditioning and learned edge feature updates, tailored for unstructured CFD meshes. The model was trained on CFD-generated datasets (121), validated against wind-tunnel data and evaluated across a range of Mach numbers and angles of attack in transonic flight regime. The model’s generalization was assessed using K-Fold Cross-Validation, benchmarked against alternative GNN architectures and a PODI reduced-order model, achieving an average \(R^2\) of 0.945 on 13 test cases and less than 8% error in lift ( \(C_L\) ) and drag coefficients ( \(C_D\) ). Despite these strong results, the model encountered challenges in accurately resolving shocks and predicting pitching moments ( \(C_M\) ), with an average error of 20%. These results demonstrate the potential and limitations of GIN-based surrogates for transonic aerodynamic analysis, offering substantial computational savings while maintaining reasonable fidelity compared to conventional CFD.