AeroGTO (AAAI 2025)

🚩News (2024.12) Our model achieved second place in the CIKM 2024 AnalytiCup competition.

AeroGTO: An Efffcient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries [Paper]

In the automotive industry, achieving high-precision aerodynamics requires large-scale computational fluid dynamics (CFD) simulations, which are both time-consuming and computationally expensive. To capturing intricate physical correlations across complex geometries while balancing large-scale discretization with computational costs,we propose AeroGTO, an efficient graph-transformer operator designed specifically for large-scale aerodynamics with the following features:

• Local and Global Feature Extraction: By combining message passing and projection-inspired attention, AeroGTO isolates and captures physical correlations at both local and global levels, enhancing interpretability.
• Efficient Graph Neural Network: The frequency-enhanced GNN with kNN handling 3D geometries allows for efficient local feature extraction.
• Transformer Architecture with Linear Complexity: The model’s ability to handle multi-level dependencies with linear complexity relative to mesh points allows for fast, scalable inference.
• Performance Gains: AeroGTO reduces error by 7.36% on average, achieves a 10.71% improvement in drag coefficient estimation, and provides fast, low-parameter predictions with an overall R² of 0.9250 on unseen data.

Figure 1. Overview of AeroGTO.

AeroGTO v.s. Previous Operators

Compared to previous state-of-the-art models, our model demonstrates superior performance and reduces computational resource usage significantly by leveraging attention method.

As shown below, AeroGTO can accurately capture the vehicle's geometric information, enabling precise and efficient surface pressure prediction.

Figure 2. Visualization of ground-truth pressure and corresponding prediction.

Get Started

Add dataset folder if it does not exist, add data to corresponding dataset. To make it easier for everyone to run the program, we have included a sample in ShapeNet dataset for each data collection. You can directly run the following code to test it in code folder:

# For DP training
bash run_dp.sh
# For DDP training
bash run_ddp.sh

You should change the environment settings in the file according to your own hardware configuration.

Data Format:

The format of our dataset (AeroGTO_Dataset) should be as follows:

Node_pos = [
    [X1, Y1, Z1],
    [X2, Y2, Z2],
   ...
]
Cells = [
    [p1, p2, p3],
    [p4, p5, p6],
   ...
]

• X,Y,Z: (X_dim x Y_dim x Z_dim) numpy array, representing input mesh points

• X_dim, Y_dim, Z_dim: input dimension of geometry

• Cells: (N_cells x 3) numpy array, representing Geometric Cells, which contains three points. The minimum index is 0, and the maximum value is N_points-1

• Note:
      I. For a single sample, The number of points must match, i.e, X.shape[0]=Y.shape[0], but it can vary with different samples.
      II. If additional preprocessing is required for the data, please modify it in the corresponding dataset file in the directory code/src
  

For test, You can directly run the following code to test it in code folder to get infer result, saving as *.npy in result folder:

# For infer on a single GPU
bash infer.sh

Requirements

• torch==2.1.0
• torch_scatter==2.1.0
• numpy==1.24.3
• pandas==2.0.1
• plyfile==0.7.4
• h5py==3.9.0
• vtk==9.2.6
• tensorboardX==2.6

Citation

If you find this repo useful, please cite our paper.

@inproceedings{liu2025aerogto,
  title={AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries},
  author={Liu, Pengwei and Wang, Pengkai and Ren, Xingyu and Yuan, Hangjie and Hao, Zhongkai and Xu, Chao and Cai, Shengze and Ni, Dong},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={18},
  pages={18924--18932},
  year={2025}
}

Contact

If you have any questions or want to use the code, please contact liupw@zju.edu.cn.

Contributing

We welcome contributions to improve the dataset or project. Please submit pull requests for review.

doing list:

• This paper uses [PaddlePaddle] as the deep learning framework. PR[#1]
• ...

Acknowledgement

We appreciate the following contents a lot for their valuable code base or datasets:

https://github.com/echowve/meshGraphNets_pytorch

https://github.com/HaoZhongkai/GNOT

https://github.com/thuml/Transolver

https://github.com/7tl7qns7ch/IPOT

https://github.com/Mohamedelrefaie/DrivAerNet

https://catalog.ngc.nvidia.com/orgs/nvidia/teams/modulus/resources/modulus_datasets-ahmed_body_test