neashton/windsorml

WindsorML: High-Fidelity Computational Fluid Dynamics dataset for automotive aerodynamics

Contact:

Neil Ashton (NVIDIA) - contact@caemldatasets.org

website:

https://caemldatasets.org

Summary:

This work presents a new open-source high-fidelity dataset for Machine Learning (ML) containing 355 geometric variants of the Windsor body, to help the development and testing of ML surrogate models for external automotive aerodynamics. Each Computational Fluid Dynamics (CFD) simulation was run with a GPU-native high-fidelity Wall-Modeled Large-Eddy Simulations (WMLES) using a Cartesian immersed-boundary method using more than 280M cells to ensure the greatest possible accuracy. The dataset contains geometry variants that exhibits a wide range of flow characteristics that are representative of those observed on road-cars. The dataset itself contains the 3D time-averaged volume & boundary data as well as the geometry and force & moment coefficients. This paper discusses the validation of the underlying CFD methods as well as contents and structure of the dataset. To the authors knowledge, this represents the first, large-scale high-fidelity CFD dataset for the Windsor body with a permissive open-source license (CC-BY-SA).

CFD Solver:

All cases were run using the Volcano Platforms commerical CFD solver, which is based upon a GPU-native cartesian immersed-boundary method Wall-Modelled Large-Eddy Simulation (WMLES) approach. Each case was run transiently for approximately 80 convective time units (CTU) on meshes of approximately 300M cells. Please see the paper for full details on the code and validation:

How to cite this dataset:

In order to cite the use of this dataset please cite the paper below which contains full details on the dataset. It can be found here: https://arxiv.org/abs/2407.19320

@article{ashton2024windsor, title = {WindsorML: High-Fidelity Computational Fluid Dynamics dataset for automotive aerodynamics}, year = {2024}, journal = {arxiv.org}, url={https://arxiv.org/abs/2407.19320}, author = {Ashton, Neil and Angel, Jordan and Ghate, Aditya and Kenway, Gaetan and Long Wong, Man and Kiris, Cetin and Walle, Astrid and Maddix, Danielle and Page, Gary} }

Files:

Each folder (e.g run_1,run_2…run_“i” etc) corresponds to a different geometry that contains the following files where “i” is the run number:

• windsor_i.stl : geometry stl (~5mb)
• windsor_i.stp : geometry step (~1mb)
• geo_parameters_1.csv : parameters that define the geometry (explained in the associated paper)
• boundary_i.vtu : Boundary VTU (~500mb)
• boundary_dual_area_i.npy: Native-point surface quadrature weights (m^2) for boundary_i.vtu, in exact point order.
• volume_i.vtu : Volume field VTU (~20GB)
• force_mom_i.csv : forces/moments time-averaged (Cd,Cs,Cl,Cmy)
• force_mom_varref_i.csv: forces/moments time-averaged (Cd,Cs,Cl,Cmy) using unique reference area per geometry
• images (folder) that contains images of the following variables (pressure, velocityX,ReynoldsStressXX,YY,ZZ) for slices of the domain in the X,Y & Z locations as well as an image of the geometry itself (windsor_i.png)
• force_mom_all.csv: contains force/moments for all runs in a single file
• force_mom_varref_all.csv: contains force/moments for all runs in a single file using a reference frontal area that is unique to each geometry
• geo_parameters_all.csv: contains all the geometry parameters for all the runs in a single file
• splits/: deterministic benchmark manifests, methods documentation, derived metrics, diagnostic figures, and generation code

Recommended dataset splits

WindsorML provides eight deterministic train/validation/test split families in splits/manifest.json. Identifiers follow the run_N convention used by the dataset.

Split	Type	Train	Validation	Test	Intended evaluation
full	In-distribution	284	35	36	Seed-42 random baseline, approximately 80/10/10
medium	In-distribution	95	35	36	Intermediate data efficiency
scarce	In-distribution	47	35	36	Low-data evaluation
super_scarce	In-distribution	8	35	36	Extreme low-data evaluation
geometry	OOD	248	36	71	STL-surface geometry extrapolation
high_drag	OOD	248	36	71	High-drag extrapolation
low_drag	OOD	248	36	71	Low-drag extrapolation
image_wake	OOD	248	36	71	Image-derived wake extrapolation

The full family is a reproducible seed-42 benchmark. It is not a reconstruction of the paper's preliminary 60/20/20 evaluation partition, whose case membership was not published. The reduced-data training sets are strictly nested and share the same validation and test cases. For the OOD families, validation is selected from the training-side population.

The aggregate tables and manifest cover run_0 through run_354. At the source revision used for this package, per-run STL and image files for run_350 through run_354 were unavailable. Their geometry and image-wake scores are estimated from nearby observed cases and are explicitly identified in the distributed metric CSVs.

Download only the split package with:

hf download neashton/windsorml \
  --type dataset \
  --include "splits/**" \
  --local-dir ./windsorml

Complete definitions, construction methods, missing-data treatment, diagnostic figures, and reproducibility instructions are provided in splits/README.md.

Downloads:

The dataset is now available on HuggingFace. Below are some examples of how to download all or selected parts of the dataset. Please refer to the HuggingFace documentation for other ways to accessing the dataset and building workflows.

Example 1: Download all files (~8TB)

Please note you’ll need to have git lfs installed first, then you can run the following command:

git clone git@hf.co:datasets/neashton/windsorml

Example 2: only download select files (STL,images & force and moments):

Create the following bash script that could be adapted to loop through only select runs or to change to download different files e.g boundary/volume.

#!/bin/bash

# Set the path and prefix
HF_OWNER="neashton"
HF_PREFIX="windsorml"

# Set the local directory to download the files
LOCAL_DIR="./windsor_data"

# Create the local directory if it doesn't exist
mkdir -p "$LOCAL_DIR"

# The currently available per-run folders span 0 to 349.
for i in $(seq 0 349); do
    RUN_DIR="run_$i"
    RUN_LOCAL_DIR="$LOCAL_DIR/$RUN_DIR"

    # Create the run directory if it doesn't exist
    mkdir -p "$RUN_LOCAL_DIR"

    # Download the windsor_i.stl file
    wget "https://huggingface.co/datasets/${HF_OWNER}/${HF_PREFIX}/resolve/main/$RUN_DIR/windsor_$i.stl" -O "$RUN_LOCAL_DIR/windsor_$i.stl"

    # Download the force_mom_i.csv file
    wget "https://huggingface.co/datasets/${HF_OWNER}/${HF_PREFIX}/resolve/main/$RUN_DIR/force_mom_$i.csv" -O "$RUN_LOCAL_DIR/force_mom_$i.csv"

done

Acknowledgements

• CFD solver and workflow development by Jordan Angel, Aditya Ghate, Gaetan Kenway, Man Long Wong, Cetin Kiris (Volcano Platforms) and Neil Ashton (Amazon Web Services - now NVIDIA)
• Geometry parameterization by Astrid Walle (Siemens Energy)
• Windsor advise and consultation by Gary Page (Loughborough University)
• Guidance on dataset preparation for ML by Danielle Maddix (Amazon Web Services - now NVIDIA)
• Simulation runs, HPC setup and dataset preparation by Neil Ashton (Amazon Web Services - now NVIDIA )

License

This dataset is provided under the CC BY SA 4.0 license, please see LICENSE.txt for full license text.

Native-point surface areas

Each run_N/boundary_dual_area_N.npy is a one-dimensional little-endian float32 array containing one barycentric dual-area weight in m^2 per native point of run_N/boundary_N.vtu, in its original point order. The array is valid only for that exact raw VTU. The definition, generator, hashes, and complete 350-run manifest are in surface_dual_areas/.

version history:

• 17/08/2026 - Added native-point surface-area arrays for all 350 available runs, with reproducibility metadata.
• 17/08/2026 - Added deterministic benchmark train/validation/test splits, including nested data-efficiency and out-of-distribution evaluation protocols; documented current per-run asset coverage.