A physics-informed surrogate model for 3D vehicle aerodynamics via boundary gradient constraints

In automotive design, aerodynamic optimization still relies on costly computational fluid dynamics (CFD) simulations, which are heavily bottlenecked by complex volumetric mesh generation. To address this limitation, a physics aware point cloud based surrogate framework is proposed for direct drag prediction from 3D unstructured point clouds. A Curvature-Aerodynamic Weighted Sampling (CAWS) strategy is introduced to preserve critical geometric boundary features—such as the front stagnation zone and rear wake region—during the preprocessing of the DrivAerNet++ dataset. The model adopts a dual-branch architecture, where an explicit Point Transformer captures local geometric variations and an implicit field represents global flow structures. More importantly, to enforce physical consistency without requiring full flow-field supervision, a Streamline Gradient Alignment Loss (SGAL) is designed based on automatic differentiation. SGAL acts as a macroscopic soft boundary constraint, penalizing gradient directions that violate basic stagnation flow behaviors near the vehicle's windward boundary surface. Experimental results demonstrate that the proposed point cloud surrogate achieves a mean R 2 score of 0.8962 ± 0.0080 and millisecond level inference on an independent test set. By integrating boundary physical constraints into a data driven point cloud representation, the framework offers an efficient and physically consistent alternative to traditional mesh-based CFD for early-stage aerodynamic design exploration.