EmmiAI/Emmi-Wing

Emmi-Wing Dataset

This repository contains the dataset proposed in Going with the Speed of Sound: Pushing Neural Surrogates into Highly-turbulent Transonic Regimes, presented at the Workshop on ML for the Physical Sciences at NeurIPS 2025.

Our dataset follows widely-used industrial standards:

• OpenFOAM-v2506 used for simulations and mesh generation
• Steady-state compressible solver (rhoSimpleFoam)
• Body-fitted mesh using snappyHexMesh, y^+ values in the range [50 − 200]
• Spalart-Allmaras turbulence model with wall functions
• Spatial discretization using second-order schemes for momentum, energy and pressure terms. First order used for turbulence quantities.

This repository contains a subset of the parameter scans that we used for evaluating our best surrogate model AB-UPT. In total there are 70 cases compressed in the scans_reduced.zip file which contains a directory for each case according to the following structure:

run_X
├── design_parameters.pt
├── surface_position.pt
├── surface_pressure.pt 
├── surface_wall_shear_stress.pt 
├── volume_position.pt
├── volume_pressure.pt
├── volume_velocity.pt
├── volume_vorticity.pt
├── cell_areas.npy
├── cell_centers.npy
├── cell_normals.npy
└── wing.stl

The pytorch tensors contain the positions and field values for the respective fields and volume and surface quantities. The numpy files contain necessary information to compute drag and lift coefficients. Finally, the STL file contains the surface mesh of each wing.

The full dataset can be downloaded at this link: https://data.emmi.ai/s/qTgKFQCRNnFTgXN

Data Quality

As mentioned in the paper, we used our best neural surrogate for quality control and were able to identify a bunch of cases that are affected by numerical artifacts. We provide a list of the potentially most severe cases in the erroneous_cases.npy file.

Please note that we attached this file to raise awareness, but they do not significantly impair the performance of neural surrogates trained on those cases. In fact, all surrogate models trained in the paper used this data during training and we found that the trained surrogates usually smooth out thos artifacts hence making them suitable as anomaly detectors. For more infos, please check out the paper.

Example usage

We provide a Github repository including basic scripts to illustrate dataloading and visualization to further facilitate training of neural surrogates on our data.

Citation

If you find our work useful or use our dataset, please consider citing it

@inproceedings{
    paischer2025going,
    title={Going with the Speed of Sound: Pushing Neural Surrogates into Transonic and Highly Turbulent Regimes},
    author={Anonymous},
    booktitle={Machine Learning and the Physical Sciences Workshop @ NeurIPS 2025},
    year={2025},
    url={https://openreview.net/forum?id=36Tpmdy1Cu}
    }