DrivAerNet++ Dataset Overview

Dataset Description

• Scale and Content: DrivAerNet++ is the largest multimodal automotive aerodynamic design dataset, containing 8,000 diverse car designs modeled via high‑fidelity computational fluid dynamics (CFD) simulations.
• Design Diversity: The dataset covers various body types such as fastback, hatchback, and wagon, as well as different chassis and wheel designs.

Design Parameters

• Parameter Selection: The generation of the dataset considered multiple geometric parameters that significantly affect aerodynamics, and their ranges were set to avoid values that are difficult to manufacture or aesthetically unpleasing.
• Parametric Model: Each 3D car geometry is fully described by 26 parameters, using two primary deformation methods: deformation boxes and direct deformation.

Dataset Content

• CFD simulation data: High‑fidelity CFD simulation data for each car design, including 3D flow fields.
• 3D car mesh: Detailed 3D meshes suitable for various machine learning applications.
• Parametric model: Parameter tables enabling extensive exploration of automotive design variables.
• Aerodynamic coefficients: Key metrics such as drag coefficient (Cd), lift coefficient (Cl), etc.
• Segmentation parts: Segmented components of the car model for classification tasks.
• Point cloud data: Point clouds for each car design.

Application Domains

• Data‑driven design optimization
• Generative AI
• Surrogate model training
• CFD simulation acceleration
• Geometry classification

Computational Resources

• Computational Cost: CFD simulations were performed on MIT Supercloud using 60 nodes (total 2,880 CPU cores), with each CFD case using 256 cores and 1,000 GB memory.
• Storage Requirements: The full dataset requires 39 TB of storage.
• Computation Time: Simulations consumed approximately 3 × 10⁶ CPU‑hours.