Computational fluid dynamics (CFD) is crucial for automotive design, requiring analysis of 3D point clouds to investigate how vehicle geometry affects pressure fields and drag. Running CFD on high-resolution 3D geometry quickly becomes computationally heavy, and many solvers slow down noticeably as the geometric detail increases. We therefore introduce a dual-task deep learning framework, named AeroFormer, that predicts aerodynamic quantities directly from the vehicle’s surface geometry and avoids the need for full CFD simulations. The model is organized into two parts. One branch, AeroFormer-Cd, predicts the overall drag coefficient (Cd), while the other, AeroFormer-Press, reconstructs the pressure distribution over the vehicle’s surface. Both branches rely on a shared curvature-guided adaptive sampling process and a physics-aware attention encoding module, which enable the network to emphasize fine geometric details in aerodynamically sensitive regions such as the front bumper, A-pillars, and wake area. By integrating geometric encoding with a Transformer module, AeroFormer can learn the complex spatial dependencies that exist in irregular surface meshes. Experiments conducted on the DrivAerNet++ datasets show that AeroFormer attains high accuracy in both Cd prediction and pressure field reconstruction. Compared with traditional CFD solvers and recent surrogate models, it offers a faster and more scalable solution for aerodynamic analysis.