Geometry-Based Surrogate Modeling for Vehicle Drag Prediction Using Multi-View Images and Dynamic-Neighborhood Point Clouds

Rapid aerodynamic drag prediction is an important surrogate-modeling task in simulation-based vehicle design because high-fidelity computational fluid dynamics is often too expensive for screening large numbers of candidate geometries. We investigate two complementary geometric representations: multi-view rendered images and sampled three-dimensional point clouds. The image branch converts vehicle meshes into six-view Depth and Normal-related renderings and evaluates single-modality and fusion architectures under a leakage-resistant base-vehicle split. Across three training seeds, cross-attention fusion achieves a mean test R2 of 0.7793±0.0157 and a mean MAE of 0.026350±0.000712. The principal methodological contribution is ImprovedRegDGCNN, an architecture that combines layer-wise Dynamic-K neighborhoods with attention-based global aggregation and a lightweight regression head. On the Independent Robustness Split with three matched training seeds, ImprovedRegDGCNN achieves the highest mean test R2 and the lowest mean test MSE and MAE among RegDGCNN, PointNet++, an xyz-only Point Transformer, and the proposed model, although its mean Max AE is not the lowest; its mean test R2 is 0.3921±0.0327, compared with 0.2785±0.0178 for RegDGCNN. In a complementary experiment across three data partitions with the training seed fixed at 42, ImprovedRegDGCNN increases mean R2 from 0.2877±0.0259 to 0.4522±0.0328 and reduces mean MAE from 0.011912±0.000277 to 0.010525±0.000247. Post hoc analyses show larger errors at both ends of the drag-coefficient range and the clearest improvements in the four minority vehicle configurations. Because the image and point-cloud branches use different geometry sources and data partitions, their metrics are interpreted within each branch rather than as a direct numerical ranking.