DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization

Vehicle aerodynamics optimization is fundamental to automotive engineering, drag reduction, noise minimization, and vehicle body stability through complex fluid dynamics simulations.

Traditional approaches rely on computationally expensive CFD simulations that limit design exploration or simplified models that compromise accuracy. Machine learning methods offer promising alternatives but require high-fidelity training data that has been largely unavailable in the public domain.

The gap between academic machine learning research and industrial CFD applications remains unbridged due to the absence of datasets meeting rigorous engineering standards.

Here we present DrivAerStar, a comprehensive and reproducible dataset of 12,000 high-precision automotive CFD simulations, created by 3 basic rear designs and 20 fine-tuned CAD parameters by the FFD algorithm, with all configurations simulated using the industry-standard STAR-CCM+ software.

Unlike existing datasets, DrivAerStar provides complete engineering data that has been thoroughly validated against wind tunnel experiments with discrepancies below 5\%, including aerodynamic coefficients, surface pressures, and velocity fields. Our benchmarks demonstrate that machine learning models trained on this dataset achieve industrial-grade prediction accuracy while reducing computational costs by orders of magnitude.

This dataset establishes a foundation for data-driven aerodynamic design methodologies that can transform automotive development processes.

Beyond automotive applications, DrivAerStar represents a paradigm for integrating high-fidelity industrial-grade physics-based simulations with artificial intelligence, potentially extending to diverse engineering disciplines where computational constraints currently limit design optimization.