andreacenzato/Ground-Wing

Ground Wing

Summary

The reference paper presents a CFD dataset of a highly parametrized 3D wing operating in ground effect, designed for benchmarking neural surrogate models for aerodynamic prediction, using a wing shape and flow conditions typically encountered in motorsport applications. Six neural network architectures representing distinct paradigms --- convolutional (DoMINO), graph-based (X-MeshGraphNet), transformer-based (AB-UPT, GeoTransolver, GAOT), and JEPA-based (AeroJEPA) --- are evaluated on their ability to predict surface pressure coefficient and wall shear stress distributions, from which integral drag and lift coefficients are recovered. The dataset comprises 770 RANS simulations of 194 unique wing configurations, sampled across 27 airfoil profile variants and 4 distances from the ground, spanning variations in wing sweep, chord, taper, angle of attack, and tip twist. The dataset is intended to be used as a challenging benchmark for surrogate CFD models, due to the uniqueness of the ground effect being absent from other wing and airfoil datasets.

CFD Setup

All simulations are incompressible RANS, with the Realizable K-epsilon turbulence model with wall functions, ran in STAR-CCM+ v2310.01. Air density is set to $\rho =1.18858kg\mathrm{/}{m}^3\backslash rho\; =\; 1.18858\; kg/m^3$, with inlet velocity ${\mathbf{U}}_{\mathbf{\infty }}=50m\mathrm{/}s\backslash mathbf\{U\_\{\backslash infty\}\}\; =\; 50m/s$.

Files

CFD solutions are stored within the `runs' directory.

Each simulation is identified by one four-digit run id which identifies a unique geometry, and one four-digit ride height, which identifies the distance from the ground of a specific sample. The resulting name convention is the following: sample 0123 at a distance from the ground of 75mm will be named run_0123_7575.

For each simulation, three files are distributed:

• A surface-only .stl geometry with no field data, provided for reference and for direct use by geometry-only models.
• A surface .vtp file with the cell-centered fields.
• A volumetric .vtu file with fields in the point data.

The surface meshes contain approximately 1.2M surface cells, while the volume meshes contain approximately 28M cells.

The contents of each file are described in the table below

Paper Symbol	Name in the files	Unit	In the .vtp	In the .vtu
$C_{p}C\_p$	MeanPressureCoefficient	-	✅	✅
$C_{pT}C\_\{pT\}$	MeanTotalPressureCoefficient	-	✅	✅
${\mathbf{w}}_{\mathbf{s}}\backslash mathbf\{w\_s\}$	MeanWallShearStress	Pa	✅	❌
$\mathbf{U}\backslash mathbf\{U\}$	MeanVelocity	m/s	❌	✅
$\mathbf{u}\backslash mathbf\{u\}$	DimensionlessMeanVelocity	-	❌	✅

Alongside the raw CFD outputs, additional descriptive and analysis files are provided.

Dataset Composition

These files are stored within the dataset_composition directory. They describe the contents of the dataset.

• airfoil_list.csv: maps each profile index P to its airfoil name.
• parameter_naming.md: maps the database parameter names (P, CA, OA, CS, OS, AF) to the paper symbols.
• run_db.csv: one row per simulation with its geometry parameters, ride heights, solver residuals and integral force coefficients. Also contains runs that are not part of the dataset (e.g. mesh convergence studies).
• samples_stratified.csv: geometry parameters of the primary-sampled geometries (profiles 1–12).
• samples_interpolated.csv: geometry parameters of the additional and interpolated-profile geometries (profiles 13–27).
• split_stratified.yaml: train/validation/test run IDs of the main benchmark split.
• split_interpolated.yaml: same train/validation as the stratified split; the test set holds the interpolated-profile geometries.
• split_outliers.yaml: train/validation/test run IDs of the split whose validation and test sets are partly filled with drag-coefficient outliers.

Model Configurations

These files are stored within the model_configs directory. They include training configurations of the six benchmarked models as YAML files. All details are given in the referenced paper.

Scripts

These files are stored within the scripts directory. They represent the code used to generate the dataset, and to analyze and plot the benchmarking results.

• airfoils/xfoil_interpolation.py: generates interpolated airfoil profiles with XFOIL.
• airfoils/xfoil_plotting.py: plots the original and interpolated airfoil profiles.
• samples/sample_generation.py: Sobol-based, sensitivity-weighted sampling of the geometry parameters.
• samples/split_dataset.py: builds the train/validation/test split files.
• plotting/inference_io.py: shared utilities for reading the split files and the inference results.
• plotting/surface_metrics.py: computes the surface field errors and integral force coefficients from the prediction meshes.
• plotting/coefficient_error.py: produces the benchmark tables and figures (run after surface_metrics.py).
• plotting/dataset_description.py: produces the dataset description figures.
• plotting/field_screenshots.py: renders surface-field screenshots of the predictions.
• plotting/outputs/: output directory for the generated figures and tables.

Citation

To use this dataset in a publication, cite the following paper, which contains the full description of the dataset.

@misc{cenzato2026groundwing,
      title={Ground-Wing: A Parametric 3D Wing-in-Ground-Effect Dataset for Benchmarking Neural Surrogate Models}, 
      author={Andrea Cenzato and Lukas Karch and Andre Luckow and Yoran Pichon and Francesco Pasqua and Abouzar Ghasemi and Rishikesh Ranade and Ira Shokar and Julia Kowalski},
      year={2026},
      eprint={2611.04571},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2611.04571}, 
}

Revision History

• 2026/11/04: Initial Publication