ratiolabs/downforce-airfoils

Downforce Airfoils

Steady, incompressible, two-dimensional subsonic RANS CFD simulations of downforce-generating (inverted-camber) NACA 4-digit airfoils, released by Ratio Labs.

Paper: Airfoil2Vec: Spectral Geometry-Conditioned Neural Surrogate Models for Airfoil Aerodynamics and a Downforce-Generating CFD Dataset (arXiv:2609.38213)

The collection targets a regime that is comparatively underrepresented in existing public airfoil CFD datasets for machine learning: geometries and operating conditions relevant to road-vehicle / motorsport aerodynamics (rear wings, front wings, and other inverted-lift-generating surfaces), rather than aircraft wings.

Dataset summary

• 907 valid airfoil geometries, generated from the NACA 4-digit family with negative camber (m < 0, m∈[1,8], p∈[1,8], t∈[4,18]) and a sharp trailing edge, to produce downforce-oriented sections.
• 11 angles of attack per geometry, α ∈ [-10°, 0°].
• 9,977 converged CFD cases in total (907 × 11, minus a few non-converging combinations).
• Solver: steady incompressible RANS with the k-ω SST turbulence model, OpenFOAM (simpleFoam).
• Flow conditions: Re = 4.66e6, M∞ = 0.2 (clean, straight-line operating condition).
• Mesh: automatically generated structured C-type mesh, ~276k cells per case (converged per a grid-independence study on NACA6412 at ~177k cells; a finer mesh was used for robustness across the full geometry sweep).

Splits

Split	# files	# geometries	Description
train	8,600	860	Training split
test	1,377	47	Held-out NACA geometries (no overlap w/ train)
non_naca	55	5	Non-NACA airfoils, pure geometry extrapolation

non_naca contains 5 airfoils typical of motorsport aerodynamics — NASA LS(1)-0413, Selig S1223, Eppler E423, Clark Y, and Wortmann FX 63-137 — each at the same 11 angles of attack, none seen during any training. Use it to test generalization beyond the NACA 4-digit family, not as a training or validation split.

File format

Each case is one HDF5 file, named <geometry>_AoA_<angle>deg.h5, e.g. NACA6412_inverted_AoA_-5deg.h5. Every file contains:

Dataset	Shape	Description
flow_field	(N, 9)	Per-cell volume field. Columns (see column_headers attr): x, y, z, pressure, u_x, u_y, u_z, temperature, stanton_number
extra_flow_field	(N, 10)	Additional per-cell derived/turbulence quantities
surface_field	(M, 10)	Per-panel surface quantities (Cp, Cf, wall shear, ...)
spline_coordinates	(1999, 2)	Ordered (x, y) points describing the airfoil contour (chord-normalized)
summary	attrs	Integrated aerodynamic coefficients (Cl, Cd, Cm, ...)

N ≈ 276,000 cells per case (varies slightly by mesh).

Browsing the dataset (viewer)

Each .h5 file stores several arrays of different lengths per case (flow_field has ~276,000 rows, spline_coordinates has 1,999, surface_field has ~410, summary is scalar attributes) — a heterogeneous scientific HDF5 layout, not a single flat table, so the Hub cannot render an .h5 file itself as a table.

Instead, this repo also ships a metadata config: one row per CFD case (geometry, angle_of_attack_deg, reynolds_number, mach_number, lift_coefficient_cl, drag_coefficient_cd, pitch_moment_coefficient_cm, n_cells, and file_name pointing at the corresponding full HDF5 file). This is what renders in the Hub's dataset viewer above, and is also the fastest way to filter/sort cases (e.g. by Cl/Cd) before pulling the full per-cell HDF5 file for the ones you actually need:

from datasets import load_dataset

meta = load_dataset("ratiolabs/downforce-airfoils", "metadata", split="train")
meta = meta.filter(lambda r: r["drag_coefficient_cd"] < 0.05)
print(meta[0]["file_name"])  # -> download this specific case, see below

Loading a file

import h5py

with h5py.File("train/NACA6412_inverted_AoA_-5deg.h5", "r") as f:
    flow_field = f["flow_field"][:]                       # (N, 9)
    columns = [c.decode() if isinstance(c, bytes) else c
               for c in f["flow_field"].attrs["column_headers"]]
    spline = f["spline_coordinates"][:]                    # (1999, 2)
    cl = f["summary"].attrs.get("Cl")

x = flow_field[:, columns.index("x")]
y = flow_field[:, columns.index("y")]
pressure = flow_field[:, columns.index("pressure")]

Downloading

The dataset is not (yet) wrapped in a Hugging Face datasets loading script — it's a plain collection of HDF5 files, best pulled with the Hub client directly:

from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="ratiolabs/downforce-airfoils",
    repo_type="dataset",
    allow_patterns=["test/*", "non_naca/*"],  # drop this to also get train/ (~338GB)
)

Citation

If you use this dataset, please cite the accompanying paper (arXiv:2609.38213):

@misc{borde2026airfoil2vecspectralgeometryconditionedneural,
      title={Airfoil2Vec: Spectral Geometry-Conditioned Neural Surrogate Models for Airfoil Aerodynamics and a Downforce-Generating CFD Dataset}, 
      author={Haitz Sáez de Ocáriz Borde and Flavio Savarino and Andrei Cristian Popescu and Pietro Innocenzi and Pantelis Papageorgiou and Xerxes Xian Chong},
      year={2026},
      eprint={2609.38213},
      archivePrefix={arXiv},
      primaryClass={physics.flu-dyn},
      url={https://arxiv.org/abs/2609.38213}, 
}

License

CC BY 4.0.