Multi-task diffusion graph model for aero-electromagnetic analysis of blended-wing-body configurations

Blended-wing-body configurations require coordinated assessment of aerodynamic loading and electromagnetic scattering characteristics; however, high-fidelity computational fluid dynamics and computational electromagnetics solvers remain too expensive for large-scale computations. We propose a multi-task diffusion graph neural network (GNN) defined on unstructured surface meshes that jointly predicts the surface pressure-coefficient field Cp, the global aerodynamic coefficients (CL,CD), and mean monostatic radar cross section (RCS) levels under horizontal and vertical polarizations (R¯H,R¯V). The node field and global scalars are concatenated into a joint diffusion state and denoised by a hierarchical, mesh-aware graph neural network; physical consistency between Cp and (CL,CD) is encouraged by a surface-integral regularizer, and task-appropriate conditioning is enforced through hard-gated masking for the mean-RCS branch. On a dataset of 3800 geometries evaluated at Mach numbers 0.60, 0.85, and 1.50 and angles of attack 2° and 6°, the model attains R2=0.969 for Cp, mean absolute errors (MAEs) of 0.0065/0.0013 for CL/CD, and MAEs of 0.381/0.358 dB for (R¯H,R¯V) on a geometry-disjoint test set. The reported errors should therefore be interpreted as solver-level surrogate fidelity with respect to the label-generation pipelines, rather than as independent experimental validation. In the long-step setting, inference provides an approximately 35× wall-clock speedup for the coupled aerodynamic–electromagnetic outputs relative to the solver pipelines used for label generation. In a short-step setting focused on Cp, the full pressure field is produced in 2.705 s per sample.