Multi-scale geometry-driven neural operator method for predicting aerodynamic performance of aircraft

To address the limitations of conventional aerodynamic simulations, which require numerous repeated iterations and thus struggle to support rapid aircraft design evaluation. This paper proposes a novel neural operator-based method with multi-scale geometric priors. A grid-aware geometric remapping (GAGR) module is first employed to map the original point cloud into a locally refined representation with concentrated information distribution, thereby enabling more effective encoding of critical geometric features under a limited point. In addition, an adaptive sampling under spatial constraints (ASSC) mechanism is introduced, which regulates the point density across different regions of the input geometry via voxel partitioning and local point set reconstruction, yielding an importance-aware, adaptively controlled geometrically reasonable spatial layout of the point cloud. Finally, a geometry-informed neural operator is constructed, which operates in a latent regular grid space and exploits local geometric adjacency together with a Fourier-based operator structure to achieve an end-to-end mapping from exterior geometry to aerodynamic performance. Experiments on a blended wing body configuration demonstrate that the proposed method accurately recovers the distribution characteristics of surface pressure coefficients, achieving correlation coefficients above 0.95. On a custom dataset of dorsal inlet configurations, the correlation coefficients exceed 0.91, validating the robustness of the proposed method for complex configurations and providing an efficient and practical solution for rapid aerodynamic performance evaluation of aircraft.