Dataset for the paper "Graph-Convolutional Autoencoder Frameworks for Aerodynamic Shape Predictions of the Agard Wing"

This repository provides the CFD dataset of wing shape deformations used in: D. Massegur, A. Da Ronch, G. Immordino, A. Vaiuso, M. Righi."Graph-Convolutional Autoencoder Frameworks for Aerodynamic Shape Predictions of the Agard Wing."AIAA SciTech 2025 Forum.https://doi.org/10.2514/6.2025-XXXX The dataset consists of steady Reynolds-averaged Navier–Stokes (RANS) simulations of the AGARD 445.6 wing subjected to structural deflections derived from finite-element (FEM) mode shapes. These data were used to train and validate the graph-convolutional autoencoder frameworks for predicting distributed aerodynamic surface fields on parametrically deformed wing shapes. Test Case Geometry: AGARD 445.6 wing (sweep, NACA 65A004 airfoil). CFD mesh: Unstructured type, 45,943 surface nodes, y+≈1.. Flow conditions: Mach number=0.96, Reynolds number=4.51×105. Solver: SU2 v7.5, Spalart–Allmaras turbulence model, JST scheme with viscous damping, BiCGStab with ILU preconditioner, multigrid V-cycle. Surface outputs: pressure coefficient (CP), shear-stress components (CFx,CFy,CFz). Integrated outputs: lift (CL), drag (CD), pitching moment (CMy). Wing Shape Dataset Six structural deflection modes included (first/second bending, first/second/third torsion, in-plane bending). Mode amplitudes scaled to max ±75 mm (~10% semi-span). Latin Hypercube Sampling (LHS) used to generate a design of experiments (DOE). Total of 251 wing shapes simulated (including undeflected baseline). CFD solutions provided for all configurations. Geometry Generation WorkflowFor each sample in the dataset, the following pipeline was used to generate a deformed geometry prior to CFD analysis: Mode amplitude definition: assign a set of amplitudes m=[m1,m2,...,m6] to the six FEM deflection modes. Structural deformation: compute the displacement of the FEM structural nodes based on the chosen modal amplitudes. Interpolation to CFD mesh: use the Moving Weighted Least Squares Interpolation (MWLSI) method to map FEM node displacements onto the CFD surface mesh nodes. Deformation file generation: write a displacement file containing the deformed coordinates of the CFD surface mesh. Mesh morphing: apply the deformation to the full CFD volume mesh using SU2_DEF. CFD simulation: run SU2 RANS simulations on the deformed geometry to obtain aerodynamic fields. PurposeThis dataset supports research in geometric deep learning and reduced-order modelling for aerodynamic shape prediction. It enables benchmarking of machine learning-based architectures against traditional CFD, demonstrating efficient surrogate models for optimisation workflows. Contents CFD snapshots of pressure and shear-stress fields on the Agard wing for 251 deflected geometries. Associated mode amplitude parameters for each sample. Mesh coordinates and connectivity for graph-based learning. KeywordsAGARD 445.6; transonic aerodynamics; shape deformation; geometric deep learning; graph neural networks; autoencoder; reduced-order modelling; CFD; SU2.本仓库提供用于以下研究的翼型变形计算流体动力学（Computational Fluid Dynamics, CFD）数据集： D. Massegur、A. Da Ronch、G. Immordino、A. Vaiuso、M. Righi. 《面向AGARD翼气动外形预测的图卷积自编码器框架》，AIAA SciTech 2025论坛. https://doi.org/10.2514/6.2025-XXXX 该数据集包含AGARD 445.6翼型在基于有限元（Finite Element Method, FEM）振型得到的结构变形下的定常雷诺平均纳维-斯托克斯（Reynolds-averaged Navier–Stokes, RANS）仿真结果。上述数据被用于训练并验证图卷积自编码器框架，以实现参数化变形翼型表面分布式气动场的预测。 ### 测试用例 #### 几何外形：AGARD 445.6翼型（后掠翼，NACA 65A004翼型） #### CFD网格：非结构化网格，包含45943个表面节点，y+≈1。 #### 流动条件：马赫数=0.96，雷诺数=4.51×10^5。 #### 求解器：SU2 v7.5，采用Spalart–Allmaras湍流模型、带粘性阻尼的JST格式、ILU预条件的BiCGStab迭代器以及多重网格V循环。 #### 表面输出量：压力系数（Pressure Coefficient, CP）、切应力分量（CFx、CFy、CFz）。 #### 积分输出量：升力系数（CL）、阻力系数（CD）、俯仰力矩系数（CMy）。 ### 翼型外形数据集 本数据集涵盖6种结构变形振型（一阶弯曲、二阶弯曲、一阶扭转、二阶扭转、三阶扭转以及面内弯曲）。 振型幅值被缩放至±75mm（约为半展长的10%）。 采用拉丁超立方抽样（Latin Hypercube Sampling, LHS）生成试验设计（Design of Experiments, DOE）样本集。 共完成251种翼型外形的仿真（包含未变形的基准构型）。 所有构型均提供CFD求解结果。 ### 几何外形生成流程 针对数据集中的每个样本，在开展CFD分析前需通过以下流程生成变形几何外形： 1. 振型幅值定义：为6种FEM变形振型分配幅值集合$oldsymbol{m}=[m_1,m_2,...,m_6]$。 2. 结构变形：基于选定的振型幅值计算FEM结构节点的位移。 3. 映射至CFD网格：采用移动加权最小二乘插值（Moving Weighted Least Squares Interpolation, MWLSI）方法将FEM节点位移映射至CFD表面网格节点。 4. 变形文件生成：写入包含CFD表面网格变形后坐标的位移文件。 5. 网格变形：通过SU2_DEF工具将变形应用至全CFD体网格。 6. CFD仿真：在变形后的几何外形上运行SU2 RANS仿真以获取气动场数据。 ### 研究用途 本数据集可支撑面向气动外形预测的几何深度学习与降阶建模研究，能够实现基于机器学习的架构与传统CFD方法的基准对比，为优化流程提供高效替代模型。 ### 数据集内容 - 251种变形构型下AGARD翼表面压力与切应力场的CFD快照 - 每个样本对应的振型幅值参数 - 适用于图学习的网格坐标与连接关系 ### 关键词 AGARD 445.6；跨声速气动；外形变形；几何深度学习；图神经网络；自编码器；降阶建模；CFD；SU2