CarBench: A comprehensive benchmark for neural surrogates on high-fidelity 3D car aerodynamics

Benchmarking has been the cornerstone of progress in computer vision, natural language processing, and the broader deep learning domain, driving algorithmic innovation through standardized datasets and reproducible evaluation protocols. The growing availability of large-scale Computational Fluid Dynamics (CFD) datasets has opened new opportunities for applying machine learning to aerodynamic and engineering design. Yet, despite this progress, large-scale standardized benchmarks remain limited for high-resolution three-dimensional automotive aerodynamics, making it difficult to evaluate and compare models systematically. In this work, we introduce CarBench, a large-scale standardized benchmark dedicated to high-resolution 3D automotive aerodynamics, and perform a large-scale evaluation of state-of-the-art models on DrivAerNet++, the largest public dataset for automotive aerodynamics, containing over 8,000 high-fidelity car simulations. We assess eleven representative model configurations spanning multiple modeling paradigms: classical neural operator methods (e.g., Fourier Neural Operator), geometric deep learning approaches (PointNet, RegDGCNN, PointMAE, PointTransformer), transformer-based neural solvers (Transolver, Transolver++, AB-UPT), and implicit field networks leveraging triplane representations (TripNet). Beyond standard interpolation tasks, we perform cross-category experiments in which representative transformer-based solvers trained on a single car archetype are evaluated on unseen categories to probe generalization. Our analysis covers predictive accuracy, pressure-field fidelity relative to the CFD reference, computational efficiency, and statistical uncertainty in evaluation metrics, providing a unified comparison across modeling families and quantifying the robustness of performance estimates under shifts in test geometry distributions. To accelerate progress in data-driven engineering, we open-source the complete benchmark framework, including training pipelines, metric-level uncertainty estimation routines based on bootstrap resampling, and pretrained model weights, establishing a reproducible foundation for large-scale learning from high-fidelity external automotive aerodynamics simulations, available at https://github.com/Mohamedelrefaie/CarBench .