DrivAerNet++ is a comprehensive multimodal dataset for aerodynamic car design, featuring 8,000 diverse car designs with high-fidelity CFD simulations. It targets researchers and engineers in automotive design and AI, enabling data-driven optimization, generative design, and surrogate modeling for aerodynamic performance prediction.

How It Works

The dataset leverages parametric modeling with 26 design parameters and morphing techniques to generate diverse car geometries, covering various configurations like fastbacks and estates, along with underbody and wheel variations. Each design is accompanied by detailed 3D meshes, volumetric CFD data, surface fields, and key aerodynamic coefficients. The dataset also includes 2D sketches and photorealistic renderings, bridging conceptual creativity with engineering data.

Quick Start & Requirements

• Dataset access is via Harvard Dataverse, requiring Globus for download.
• The full dataset requires 39 TB of storage.
• Computational cost for generating the dataset was substantial (3 × 10⁶ CPU-hours).
• Links to dataset subsets, leaderboard, and tutorials are provided.

Highlighted Details

• Largest and most comprehensive multimodal dataset for aerodynamic car design.
• Includes 8,000 diverse car designs with high-fidelity CFD simulations.
• Features 2D sketches and photorealistic renderings alongside engineering data.
• Integrated into NVIDIA Modulus and PaddleScience for SciML applications.

Maintenance & Community

Maintained by the DeCoDE Lab at MIT. Issues can be reported via GitHub issues.

Licensing & Compatibility

The DrivAerNet/DrivAerNet++ dataset is licensed under CC BY-NC 4.0, strictly for non-commercial research and educational purposes. Commercial use requires explicit licensing inquiries. The accompanying code is distributed under the MIT License.

Limitations & Caveats

The dataset is exclusively for non-commercial use, with strict enforcement against unauthorized commercial applications. Commercial licensing options are available via direct inquiry.