Predicting drag coefficients of vehicle geometries: A PointNet++ point cloud surrogate model approach

In automotive aerodynamics, traditional computational fluid dynamics (CFD) simulations for predicting the drag coefficients (Cd) are computationally intensive and often require hours to days, thereby hindering rapid design iterations. In this study, we developed a deep-learning surrogate model using PointNet++ for fast and accurate Cd prediction from vehicle point-cloud geometries. The model features a hierarchical architecture with set abstraction layers using farthest point sampling, ball query grouping, and multilayer perceptrons with channel configurations [32, 32, 64], [64, 64, 128], and [128, 256, 512], followed by fully connected layers (256 × 128 × 64) with dropouts. Trained on the DrivAerNet++ dataset of 8000 designs with 2048-point clouds, it employs data augmentation via random scaling (0.9–1.1), Gaussian jitter (σ = 0.005), and 5% point dropout. The optimization used the mean squared error loss, an initial learning rate of 0.0001, and a cosine annealing scheduler over 200 epochs with a batch size of 64. The model achieves a mean squared error of 1.72 × 10−5 and an R2 of 0.7815 on the test set, with inference times of 10–15 ms on an NVIDIA RTX A4000 GPU, outperforming baseline PointNet (R2 = 0.643) and offering a ∼10.5-fold speedup over CFD. This mesh-free framework balances the accuracy and speed, with ablation studies confirming its robustness to geometric variations.