Mohamedelrefaie/DrivAerNet
A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks
What it solves
DrivAerNet++ provides a large-scale, high-fidelity dataset to bridge the gap between conceptual car design and aerodynamic performance analysis. It eliminates the need for expensive, time-consuming computational fluid dynamics (CFD) simulations for every design iteration by providing a comprehensive benchmark for training neural surrogates.
How it works
The project consists of 8,150 diverse car designs (including fastback, notchback, and estateback configurations) modeled with high-fidelity CFD simulations. Each design is described by 26 geometric parameters. The dataset is multimodal, providing data in several formats:
- 3D Representations: Point clouds, meshes, and volumetric fields (pressure, velocity, turbulence).
- Performance Metrics: Aerodynamic coefficients like drag (Cd) and lift (Cl).
- ** uma 2D Visuals**: Photorealistic renderings and hand-drawn sketches.
- Annotations: Semantic labels for 29 different car components.
Who it’s for
- AI Researchers: Those developing generative AI for vehicle design or surrogate models for aerodynamic prediction.
- Engineering Practitioners: Designers looking to optimize car shapes for better performance without relying solely on traditional CFD.
- SciML Framework Users: Users of NVIDIA Modulus or Baidu PaddleScience.
Highlights
- Massive Scale: Includes 8,150 designs with a total dataset size of 39 TB.
- Multimodal Approach: Combines parametric models, 3D meshes, CFD fields, and 2D sketches.
- Comprehensive Benchmarking: Features a dedicated leaderboard (CarBench) for comparing neural surrogate performance.
- Broad Coverage: Represents both Internal Combustion Engine (ICE) and Electric Vehicles (EV).