GeoFormer: Mesh-free geometry-to-flow alignment framework for real-time aerodynamics on non-watertight vehicle geometries

Accurate aerodynamic evaluation of complex vehicle shape remains expensive, mainly due to geometry-specific meshing and the high computational cost of large-scale simulations. We present GeoFormer, a mesh-free surrogate model that directly maps arbitrary, even non-watertight, computer aided design (CAD) geometries to both surface and volumetric flow fields. The framework decouples the geometry representation from the Eulerian flow domain and reconnects them through a geometry-to-flow cross-attention mechanism. Vehicle surfaces are encoded from STereoLithography samples, flow fields are represented on a fixed Cartesian grid, and their interaction is handled by a kernelized linear attention operator that ensures scalable O(N) complexity on million-point inputs. We train and test GeoFormer on industrial-scale DrivAer variants generated by high-fidelity delayed detached eddy simulation. The model jointly predicts surface quantities and full volumetric fields in a single forward pass, and it shows consistent improvements over strong baselines such as Transolver, three-dimensional-GeoCA, and TripNet. Attention maps naturally highlight physically relevant regions (e.g., stagnation zones and separation lines), providing interpretability aligned with aerodynamic principles. These results demonstrate a practical CAD-to-Flow pipeline that eliminates meshing, scales efficiently, and bridges geometry and physics through cross-attention, offering a promising direction for industrial aerodynamic design.