neashton/ahmedml

AhmedML: High-Fidelity Computational Fluid Dynamics dataset for incompressible, low-speed bluff body aerodynamics

Contact:

Neil Ashton (NVIDIA) - contact@caemldatasets.org

website:

https://caemldatasets.org

Summary:

This dataset contains 500 different geometric variations of the Ahmed Car Body - a simplified car-like shape that exhibits many of the flow topologies that are present on bluff bodies such as road vehicles. The dataset contains a wide range of geometries that exhibit fundamental flow physics such as geometry and pressure-induced flow separation of flows as well as 3D vortical structures. Each variation of the Ahmed car body were run using a time-accurate hybrid Reynolds-Averaged Navier-Stokes (RANS) - Large-Eddy Simulation (LES) turbulence modelling approach using the open-source CFD code OpenFOAM. The dataset contains both surface boundary, 3D volume, geometry STL and forces/moments in open-source formats (.vtu,.vtp).

CFD Solver:

All cases were run using the open-source finite-volume code OpenFOAM v2212. Each case was run transiently for approximately 80 convective time units (CTU) on meshes of approximately 20M 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/2407.20801

@article{ashton2024ahmed, title = {{AhmedML: High-Fidelity Computational Fluid Dynamics dataset for incompressible, low-speed bluff body aerodynamics}}, year = {2024}, journal = {arxiv.org}, author = {Ashton, Neil and Maddix, Danielle and Gundry, Samuel and Shabestari, Parisa} }

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:

• ahmed_i.stl : geometry stl (~5mb):
• geo_parameters_1.csv (missing run 500): parameters that define the geometry
• boundary_i.vtp : Boundary VTP (~500mb)
• boundary_cell_area_i.npy: Native polygon areas (m^2) for boundary_i.vtp, in exact CellData tuple order.
• volume_i.vtu : Volume field VTU (~5GB)
• force_mom_i.csv : forces (Cd,Cl) time-averaged with constant reference area
• force_mom_varref_i.csv : forces (Cd,Cl) time-averaged with varying reference area
• slices : folder containing .vtp slices in x,y,z that contain flow-field variables
• images : (folder) that contains images of the following variables (CpT, UxMean) for slices of the domain in the X,Y & Z locations. In addition we provide:
• force_mom_all.csv : run, cd,cl for all runs in a single file
• force_mom_varref_all.csv : run, cd,cl for all runs in a single file with varying reference area
• geo_parameters_all.csv : all the geometry parameters for each run inside a single file
• ahmedml.slvs : SolveSpace input file to create the parametric geometries
• stl : folder containing stl files that were used as inputs to the OpenFOAM process
• openfoam-casesetup.tgz : complete OpenFOAM setup that can be used to extend or reproduce the dataset
• validation : folder containing full outputs from all four mesh levels that were used to validate the methodology
• splits/: deterministic benchmark manifests, methods documentation, derived metrics, diagnostic figures, and generation code

Recommended dataset splits

AhmedML provides eight deterministic train/validation/test split families in splits/manifest.json. Identifiers correspond to the top-level run_N directories.

Split	Type	Train	Validation	Test	Intended evaluation
full	In-distribution	400	50	50	Seed-42 public baseline
medium	In-distribution	133	50	50	Intermediate data efficiency
scarce	In-distribution	67	50	50	Low-data evaluation
super_scarce	In-distribution	11	50	50	Extreme low-data evaluation
geometry	OOD	350	50	100	STL-surface geometry extrapolation
high_drag	OOD	350	50	100	High-drag extrapolation
low_drag	OOD	350	50	100	Low-drag extrapolation
image_wake	OOD	350	50	100	Image-derived wake extrapolation

The full assignment follows the established seed-42 AhmedML baseline used by Noether. 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, while the held-out extreme is reserved for final testing.

Download only the split package with:

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

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

Native boundary polygon areas

Each run_N/boundary_cell_area_N.npy is a one-dimensional little-endian float32 array containing one area in m^2 per native polygon—and therefore per CellData tuple—of run_N/boundary_N.vtp, in its original cell order. The array is valid only for that exact raw VTP. The definition, generator, hashes, and complete 500-run manifest are in surface_cell_areas/.

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 (~2TB)

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/ahmedml

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 paths
HF_OWNER="neashton"
HF_PREFIX="ahmedml"

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

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

# Loop through the run folders from 1 to 500
for i in $(seq 1 500); do
    RUN_DIR="run_$i"
    RUN_LOCAL_DIR="$LOCAL_DIR/$RUN_DIR"

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

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

• OpenFOAM solver and workflow development by Neil Ashton (Amazon Web Services, now NVIDIA)
• Geometry parameterization by Samuel Gundry (Amazon Web Services) and Parisa Shabestari (Amazon Web Services)
• Guidance on dataset preparation for ML by Danielle Madix (Amazon Web Services)
• 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.

version history:

• 17/08/2026 - Added native boundary polygon-area arrays for all 500 runs, with reproducibility metadata.
• 17/08/2026 - Added deterministic benchmark train/validation/test splits, including nested data-efficiency and out-of-distribution evaluation protocols.
• 15/02/2025 - files uploaded to HuggingFace
• 12/11/2024 - added validation folder that contains the full output from all four mesh levels that were used to validate the methodology used.
• 04/08/2024 - updates to the file description and arxiv paper
• 05/06/2024 - global forces/geo added for all runs
• 01/05/2024 - force/moments corrected (prior version had incorrect Cs data)
• 18/04/2024 - draft version produced