Attention-based point cloud learning for aerodynamic performance prediction of wing–fuselage configurations

Advances in artificial intelligence and computational fluid dynamics (CFD) have created new opportunities for rapid aerodynamic design, yet traditional CFD-driven workflows remain costly and slow. We propose an end-to-end framework that predicts the aerodynamic performance of complex aircraft wings directly from raw point-cloud geometry, thereby bypassing mesh generation and expensive flow-field solvers. We provide a large-scale dataset assembled by coupling parametric Free-Form Deformation (FFD) with high-fidelity CFD analyses: 255 baseline wing–fuselage configurations are morphed into > 10,000 variants, each annotated with lift and drag coefficients computed in ANSYS Fluent. Second, a lightweight point-cloud network is introduced. It combines hierarchical set-abstraction layers for multi-scale geometric encoding, an attention-guided neighborhood aggregator that focuses learning capacity on flow-sensitive regions such as the leading edge, and a dual-head regression branch optimised with a weighted MSE–Huber compound loss. On a held-out test set, the model achieves RMSE = 0.063 for lift and 0.011 for drag — an average 15 % error reduction over PointNet++ and a 98 % runtime saving relative to CFD evaluation of the same shapes.