HiLiftAeroML: High-Fidelity Computational Fluid Dynamics Dataset for High-Lift Aircraft Aerodynamics

Contact: Neil Ashton (contact@caemldatasets.org)

Summary

This dataset provides the first-ever open-source high-fidelity CFD dataset of a high-lift aircraft for the purpose of AI surrogate model development. The dataset is composed of 1,800 samples, arising from 180 geometry variants of the NASA Common Research Model (CRM-HL) across 10 angles of attack (from 4° to 22° in 2° increments).

One of the key novelties of this dataset is the use of a GPU-accelerated high-fidelity explicit, wall-modeled LES (WMLES) approach on solution-adapted grids containing between 300M and 500M cells. By making this data publicly available, the aim is to accelerate the research and development of AI surrogate modeling within the aerospace industry by addressing the lack of high-fidelity, scale-resolving CFD training data for complex 3D airframes in the high-lift regime.

CFD Solver

The simulations were performed using the "Fidelity Charles" flow-solver, an explicit, unstructured, finite-volume solver for the compressible Navier-Stokes equations.

• Accuracy: 2nd-order accurate in space and 3rd-order accurate in time.
• Meshing: The domain was discretized using "Fidelity Stitch", an unstructured mesh generator based on Voronoi diagrams.
• Turbulence Modeling: To capture complex, unsteady flow phenomena such as massive flow separation, a wall-modeled LES using the Dynamic Smagorinsky subgrid-scale model and an equilibrium wall model were employed.

How to Cite This Dataset

In order to cite the use of this dataset, please cite the corresponding preprint paper (https://arxiv.org/abs/2605.19565):

@article{ashton2026hiliftaeroml,
  title={{HiLiftAeroML: High-Fidelity Computational Fluid Dynamics Dataset for High-Lift Aircraft Aerodynamics}},
  year={2026},
  journal = {arxiv.org}, url={https://arxiv.org/abs/2605.19565}, 
  author={Ashton, Neil and Clark, Adam and Heidt, Liam and Ivey, Christopher and Bose, Sanjeeb and Agrawal, Rahul and Goc, Konrad and Ranade, Rishi and Adams, Corey and Sharpe, Peter and Nidhan, Sheel and Akkurt, Semit and Leibovici, Daniel and Kossaifi, Jean}
}

Dataset Structure and Files

Each folder (e.g., geo_LHCi_AoA_j where i is the geometry ID from 001 to 180 and j is the angle of attack) corresponds to a different geometry and angle of attack. Inside each run folder, you will find:

• boundary_geo_LHCi_AoA_j.vtu.tgz: Time-averaged flow quantities on the surface boundary (approx. 13 GB compressed, 21 GB unzipped).
• volume_geo_LHCi_AoA_j.vtu.tgz: Time-averaged flow quantities within the domain volume (approx. 23 GB compressed, 86 GB unzipped).
• geo_LHCi_AoA_j.stl: Surface mesh (tris) of the geometry (approx. 197 MB).
• geo_LHCi_AoA_j.stp: Surface geometry definition (approx. 48 MB).
• force_mom_geo_LHCi_AoA_j.csv: Time-averaged drag, lift, moment, and pressure/viscous coefficients.
• geo_values_geo_LHCi_AoA_j.csv: Reference geometry values used to define the geometry via the DoE method.
• ref_values_geo_LHCi_AoA_j.csv: Reference values such as area, Q, and AoA.
• img_wss_LHCi_AoA_j.png: Visual rendering of the wall shear stress/skin-friction on the aircraft surface.
• plot_CD_geo_LHCi_AoA_j.png: Convergence plot of the Drag Coefficient over time.
• plot_CL_geo_LHCi_AoA_j.png: Convergence plot of the Lift Coefficient over time.
• plot_CM_geo_LHCi_AoA_j.png: Convergence plot of the Pitching Moment Coefficient over time.

In addition to the files per run folder, there are also general summary files in the root directory:

• geo_values_all.csv: Reference geometry values used to define the geometry via the DoE method for all runs.
• force_mom_all.csv: Time-averaged drag, lift, moment, and pressure/viscous coefficients for all runs.
• splits/: Deterministic train/validation/test split manifest, documentation, and generation code.

Recommended Dataset Splits

For reproducible machine-learning evaluation, HiLiftAeroML provides 14 deterministic split families in splits/manifest.json. Case IDs match the top-level geo_LHC..._AoA_... directories.

Evaluation goal	Suggested splits	Training population
Standard baseline and case-level data efficiency	full → medium → scarce → super_scarce	1260 → 510 → 210 → 35 cases
Unseen-geometry generalization and data efficiency	geometry → geometry_medium → geometry_scarce → geometry_super_scarce	126 → 51 → 21 → 4 complete geometries
Fixed-AoA unseen-geometry evaluation	single_aoa_4, single_aoa_12, single_aoa_22	126 train / 18 validation / 36 test geometries
Out-of-distribution evaluation	aoa, deflection, stall	High-AoA, high-deflection, or post-stall extrapolation

The levels within each data-efficiency ladder are nested and share the same validation and test sets, enabling direct comparisons across training-set sizes. Use validation data for model selection and reserve test data for final evaluation. See splits/README.md for the complete definitions, counts, and selection guidance, or splits/generate_splits.py for the generation and validation logic.

How to Download

HiLiftAeroML is publicly available on Hugging Face. There is no Hugging Face access charge, but the dataset is very large: the full repository is about 65.9 TB as stored on Hugging Face and substantially larger after extracting the .tgz archives. Please make sure you have sufficient storage and bandwidth before downloading large subsets.

We recommend using the Hugging Face CLI rather than git clone for selective downloads, resumable local downloads, dry-run previews, and Xet-backed large-file support.

Prerequisites

pip install -U huggingface_hub hf_xet

Authentication is optional for this public dataset, but strongly recommended for large downloads to avoid unauthenticated rate limits:

hf auth login

Example 1: Preview or Download the Full Dataset

Preview the full download first:

hf download nvidia/HiLiftAeroML \
  --repo-type dataset \
  --local-dir ./hiliftaeroml_data \
  --dry-run

Remove --dry-run to download the full dataset:

hf download nvidia/HiLiftAeroML \
  --repo-type dataset \
  --local-dir ./hiliftaeroml_data

Example 2: Download Selected CSV Metadata and Split Files

The following command downloads only force/moment CSVs, geometry CSVs, reference CSVs, and split files. Keep the glob patterns in quotes so your shell does not expand them locally.

Preview first:

hf download nvidia/HiLiftAeroML \
  --repo-type dataset \
  --local-dir ./hiliftaeroml_data \
  --include "geo_LHC*_AoA_*/force_mom_*.csv" \
  --include "geo_LHC*_AoA_*/geo_values_*.csv" \
  --include "geo_LHC*_AoA_*/ref_values_*.csv" \
  --include "force_mom_all.csv" \
  --include "geo_values_all.csv" \
  --include "splits/**" \
  --dry-run

Remove --dry-run to download those files.

Example 3: Download Specific Runs with a Bash Script

This script loops over geometry IDs 1-180 and AoA values 4-22 in steps of 2. By default it runs in preview mode. Set DRY_RUN=0 to download the files.

#!/usr/bin/env bash
set -Eeuo pipefail

REPO="nvidia/HiLiftAeroML"
LOCAL_DIR="./hiliftaeroml_data"
DRY_RUN="${DRY_RUN:-1}"

HF_ARGS=()
if [[ "$DRY_RUN" == "1" ]]; then
  HF_ARGS+=(--dry-run)
fi

for i in {1..180}; do
  XX=$(printf "%03d" "$i")

  for YY in $(seq 4 2 22); do
    ID="geo_LHC${XX}_AoA_${YY}"
    echo "Processing: $ID"

    FILES=(
      "$ID/boundary_${ID}.vtu.tgz"
      "$ID/${ID}.stl"
      "$ID/${ID}.stp"
      "$ID/force_mom_${ID}.csv"
      "$ID/geo_values_${ID}.csv"
      "$ID/ref_values_${ID}.csv"
      "$ID/volume_${ID}.vtu.tgz"
    )

    hf download "$REPO" "${FILES[@]}" \
      --repo-type dataset \
      --local-dir "$LOCAL_DIR" \
      "${HF_ARGS[@]}"
  done
done

To actually download after previewing:

DRY_RUN=0 bash download_hiliftaeroml.sh

Credits

The dataset was created through a collaboration between NVIDIA, Cadence Design Systems, and The Boeing Company.

Computing resources were generously provided by:

• Cadence
• NVIDIA
• Texas Advanced Computing Center (TACC) at The University of Texas at Austin
• Swiss National Supercomputing Centre (CSCS)
• Oak Ridge Leadership Computing Facility (ORNL)

License

This dataset is provided under the permissive open-source CC-BY-4.0 license. Users are free to share and adapt the material for any purpose, even commercially, provided appropriate credit is given to the original authors.

Version History

• 06/07/2026 - updated the splits to include a couple of extra ones (geometry_medium, scarce, super_scare and medium)

• 26/05/2026 - IDs 0-43 for angles 16-22 had incomplete force_mom (didn't have std dev etc). Corrected. Please download forces again if you took files before this date

• 20/05/2026 - dataset fully uploaded and arxiv preprint made available