Dataset Card for DrivAerCH3

A dataset to develop and test shape-to-drag models for automotive applications. The geometries are based on the DrivAer model. Drag coefficients are provided, as well as surface pressures, available in two formats: (i) on the STL triangles, for point cloud or graph-based approaches, and (ii) planar projections of the automobile from six directions, for image-based approaches. Five samples have been withheld for a blind test; see blind_cases.txt.

For details about the dataset generation procedure, please refer to this paper.

Dataset Details

Dataset Description

• Curated by: Mark Benjamin and Gianluca Iaccarino
• License: MIT

Dataset Sources

• Repository: https://github.com/benjamark/shapech
• Paper: https://doi.org/10.1063/5.0233367

Dataset Structure

There are 1275 samples. For each sample XXXX, the geometry is provided under

stls/XXXX.stl

Images of the geometry colored by normalized distance from the camera are provided under

images/XXXX/x_*.png

The pressure value on each triangle of XXXX.stl is provided under

pressures/XXXX.txt

Images of the geometry colored by normalized pressure are provided under

images/XXXX/p_*.png

The drag coefficient value is in the XXXX+1th row of

cd_values.txt

Data Collection and Processing

The large-eddy simulations used to obtain the pressure data were run using the CharLES solver. All image data have been normalized on a per-sample basis.

Citation

BibTeX:

@article{benjamin2025systematic,
  title={A systematic dataset generation technique applied to data-driven automotive aerodynamics},
  author={Benjamin, Mark and Iaccarino, Gianluca},
  journal={APL Machine Learning},
  volume={3},
  number={1},
  year={2025},
  publisher={AIP Publishing}
}

APA:

Benjamin, M., & Iaccarino, G. (2025). A systematic dataset generation technique applied to data-driven automotive aerodynamics. APL Machine Learning, 3(1).

Dataset Card Contact

Please reach out to markben@stanford.edu if you need any assistance.