Research on aerodynamic performance prediction of point cloud wings based on transfer learning

Acquiring extensive aerodynamic performance data in the current wing design phase incurs substantial costs, necessitating the development of efficient and accurate predictive methods. This study focused on a low-speed double-trapezoid wing. A dataset under subsonic flight conditions was constructed using OpenVSP and CFD simulations. Based on this dataset, a PointNet-based aerodynamic performance prediction network (PointNet-AP) was designed, which incorporated transfer learning utilizing multi-fidelity data. The PointNet-AP network took wing point cloud data, conventional planar geometric features, and flight conditions as inputs to predict the corresponding lift coefficient ($C_L$), drag coefficient ($C_D$), and pitching moment coefficient ($C_m$). Experimental results demonstrated that the PointNet-AP model with transfer learning achieved satisfactory predictive performance, with mean relative prediction errors of 2.88%, 4.75%, and 6.74% for $C_L$, $C_D$, and $C_m$, respectively.The proposed network provides a technical foundation for developing large-scale aerodynamic performance prediction models by offering solutions for data feature extraction and model transfer.