Deep learning for fast aerodynamic estimation of road vehicles

Abstract

The high computational cost and long runtimes of traditional evaluation methods often slow automotive aerodynamic design. Computational Fluid Dynamics (CFD) simulations and wind tunnel tests, while accurate, are resource-intensive and impractical for real-time feedback during iterative design. This thesis addresses the need for faster aerodynamic estimation by developing deep learning-based surrogate models for predicting aerodynamic quantities directly from 3D geometry. Specifically, the performance of PointNet and Geometry-Informed Neural Operators (GINO) is evaluated for predicting global drag coefficients (Cd) and local pressure distributions over complex automotive geometries. Using the DrivAerNet dataset, systematic experiments investigate the influence of total sample size, point cloud resolution, batch size, and hyperparameters on predictive accuracy. Results demonstrate that PointNet achieves strong drag prediction performance, reaching an R2 of 0.957, with a mean error percentage of approximately 1.6% and a maximum error percentage of under 8% for unseen data of around 3500 samples, when the model is trained with 400 samples which is 80% of 500 total samples, 100,000 vertices per sample, and a batch size of 16. However, PointNet shows limited sensitivity to training variations in pressure prediction, with Rel L2 errors consistently within the range of 0.35–0.37. In contrast, GINO significantly outperforms PointNet in pressure prediction tasks, achieving a test R2 of 0.873, Rel L2 errors below 0.28, and demonstrating robust data efficiency and sensitivity to latent space configurations. This study establishes a rigorous baseline for deep learning-driven aerodynamic prediction, highlighting the suitability of PointNet for global scalar quantities and the potential of GINO for accurate field-level predictions. The findings support the future development of hybrid models for fast, data-driven aerodynamic design optimization in the automotive industry.