DrivAerNet++: High-Fidelity Computational Fluid Dynamics & Deep Learning Benchmarks

<p align="center"> <a href="https://neurips.cc/virtual/2024/poster/97609"><img src="https://img.shields.io/badge/NeurIPS-2024-blue.svg" alt="NeurIPS 2024"></a> <a href="https://arxiv.org/abs/2406.09624"><img src="https://img.shields.io/badge/arXiv-2406.09624-b31b1b.svg" alt="arXiv"></a> <a href="https://dataverse.harvard.edu/dataverse/DrivAerNet"><img src="https://img.shields.io/badge/Dataset-Harvard%20Dataverse-orange.svg" alt="Dataset"></a> <a href="https://creativecommons.org/licenses/by-nc/4.0/"><img src="https://img.shields.io/badge/License-CC%20BY--NC%204.0-lightgrey.svg" alt="License"></a> </p> <p align="center"> <b>The largest and most comprehensive multimodal dataset for aerodynamic car design</b> </p>

We present DrivAerNet++, comprising 8,150 diverse car designs modeled with high-fidelity computational fluid dynamics (CFD) simulations, covering configurations such as fastback, notchback, and estateback.

---

📢 Latest News

| Date | News | |------|------| | 🆕 2024 | CarBench Released — A unified benchmark for high-fidelity 3D car aerodynamics and generalization testing |

• 🏆 Leaderboard: CarBench Leaderboard
• 📄 Paper: CarBench Paper

---

🔗 Quick Links

| Resource | Description | Link | |----------|-------------|------| | DrivAerNet++ Paper | NeurIPS 2024 Full Paper | arXiv | | Dataset Download | Hosted on Harvard Dataverse | Access Data | | Leaderboard | Submit models & compare results | DrivAerNet++ Leaderboard | | Video Summary | Overview of the project | YouTube | | Podcasts | Deep dive discussions | DrivAerNet++ | | Podcasts | Deep dive discussions | AI Design Agents |

---

🏎️ Design & Shape Variation

<p align="center"> <img src="https://github.com/user-attachments/assets/1c305975-f825-4a11-85f4-357f97fe134f" alt="Design Variation" width="80%"> </p>

Design Parameters

Several geometric parameters with significant impact on aerodynamics were selected and varied within a specific range. These parameter ranges were chosen to avoid values that are either difficult to manufacture or not aesthetically pleasing.

Shape Variation

DrivAerNet++ covers all conventional car designs. The dataset encompasses various underbody and wheel designs to represent both:

• Internal Combustion Engine (ICE) vehicles
• Electric Vehicles (EV)

<table> <tr> <td><img src="https://github.com/Mohamedelrefaie/DrivAerNet/assets/86707575/98064523-1a12-4ab3-9be4-8b745d1d1072" width="100%"></td> <td><img src="https://github.com/Mohamedelrefaie/DrivAerNet/assets/86707575/0fc97e2a-f06c-4036-a9de-8d9d1c5e6a91" width="100%"></td> </tr> </table>

💡 Each 3D car geometry is parametrized with 26 parameters that completely describe the design.

Importance of Diversity

By providing a wide range of car shapes and configurations with high-fidelity CFD, DrivAerNet++ enables:

• ✅ Models to generalize better
• ✅ Exploration of unconventional designs
• ✅ Enhanced understanding of how geometric features impact aerodynamic performance

---

📦 Dataset Contents & Modalities

✅ Available Modalities

| Modality | Description | |----------|-------------| | Parametric Models | Structured tabular design parameters | | Volumetric Fields | Full 3D CFD (pressure, velocity, turbulence) | | Surface Fields | Coefficient of pressure (Cp) and Wall Shear Stress (WSS) | | Streamlines | Flow visualization data illustrating streamlines | | Point Clouds | Dense and sparse point cloud representations | | Meshes | High-resolution 3D surface triangulations | | Aerodynamic Coefficients | Drag (Cd), Lift (Cl), and moment coefficients | | Annotations | Per-part semantic labels | | Renderings | High-quality photorealistic 2D renderings | | Sketches | Hand-drawn style sketches (Canny edge & CLIPasso) |

🚧 Coming Soon

• 📐 2D Slices: Planar field extractions
• 📊 Signed Distance Fields (SDF): For occupancy modeling
• 💥 Deformations: Simulation outputs under crash/pressure conditions

Dataset Annotations

The dataset includes detailed annotations for various car components (29 labels), such as wheels, side mirrors, and doors. These are instrumental for:

• Classification
• Semantic segmentation
• Automated meshing

---

✏️ Sketch-to-Design Extension

We bridge the gap between conceptual creativity and computational design with 2D hand-drawn sketches and photorealistic renderings.

<table> <tr> <td><img src="https://github.com/user-attachments/assets/f0ca86ae-f903-46d0-8ee5-9e63e83d88cf" width="100%"></td> <td><img src="https://github.com/user-attachments/assets/e1e4ec63-c08c-496e-ba5b-2888ba637df0" width="100%"></td> </tr> </table>

🔍 For details, check out our recent Design Agents paper: AI Agents in Engineering Design

---

💾 Dataset Access & Download

The dataset is hosted on Harvard Dataverse (CC BY-NC 4.0).

| Specification | Value | |--------------|-------| | Total Size | 39 TB | | Subsets | 3D Meshes, Pressure, Wall Shear Stress, Full CFD Domain |

We provide instructions on how to use Globus to download the dataset efficiently.

Performance Data

| Data | Download | |------|----------| | Drag Values | Download CSV | | Frontal Projected Areas | Download CSV |

---

Datasets Comparison

DrivAerNet++ stands out as the largest and most comprehensive dataset in the field.

---

🏆 Leaderboard & Comparisons

DrivAerNet++ serves as a valuable benchmark dataset due to its size and diversity. It provides extensive coverage of various car designs and configurations, making it ideal for testing and validating machine learning models in aerodynamic design. We provide the train, test, and validation splits in the following folder: train_val_test_splits.

Drag values for the 8k car designs can be found Here, and the frontal projected areas Here.

Researchers and industry practitioners can submit their models to the leaderboard to compare performance against state-of-the-art baselines. The benchmark promotes transparency, reproducibility, and innovation in AI-driven aerodynamic modeling.

For submission guidelines and current rankings, visit CarBench.

📄 Read CarBench Paper

---

📚 Related Research & Extensions

TripOptimizer

A fully differentiable deep-learning framework for rapid aerodynamic analysis and shape optimization on industry-standard car designs.

📄 Read Paper

AI Agents in Engineering Design

A multi-agent framework leveraging VLMs and LLMs to accelerate the car design process—from concept sketching to CAD modeling, meshing, and simulation.

📄 Read Paper

RegDGCNN

We have open-sourced the RegDGCNN pipeline for surface field prediction on 3D car meshes.

🔗 View Code

🛠️ Framework Integrations

DrivAerNet++ is integrated into leading Scientific Machine Learning (SciML) frameworks:

NVIDIA Modulus

• FIGConvUNet Example
• AeroGraphNet Example

PaddleScience (Baidu)

🔗 IJCAI 2024 Competition 🔗 PaddleScience DrivAerNet Example 🔗 [PaddleScience