Geometry-informed neural operator for predicting surface flow field of 3D variable-geometry aerospace vehicles across a wide-speed range

The rapid iteration of modern aerospace vehicle design necessitates fast and reliable aerodynamic evaluation during the conceptual design phase. Computational fluid dynamics (CFD) methods, while reliable, are computationally expensive for evaluating hundreds or thousands of configurations. This work presents a data-driven geometry-informed neural operator (GINO) for the end-to-end prediction of surface flow fields for a specific class of 3D aerospace vehicles across a wide-speed range. Specifically, in GINO, graph neural operators (GNOs) are employed to encode and address arbitrary vehicle topologies and surface flow fields, and Fourier neural operator (FNO) is utilized to approximate the solution operator of the Reynolds-Averaged Navier-Stokes (RANS) equations while capturing the multi-scale flow field features. The model is trained on a limited dataset consisting of 120 high-fidelity CFD samples. The dataset covers 5 different 3D vehicle geometries, each across a wide range of Mach numbers and angles of attack. The results demonstrate that the GINO model is able to handle variable geometries across subsonic, transonic, and supersonic speed regimes, and predicts flow fields that agree well with CFD simulations, successfully capturing key flow features and achieving acceptable accuracy in global aerodynamic drag coefficients C_D on both training and test sets. Notably, through operator fusion and other optimizations, GINO attains second-scale inference times for aerospace vehicles with hundreds of thousands of surface grid elements. The GINO model shows promising potential in terms of generalization capability and computational efficiency within the design space, and provides as an attractive alternative for rapid aerodynamic evaluation in the early-stage design of aerospace vehicles.