psetinek/multi-fidelity-airfoil

Multi-fidelity Airfoil Dataset

This repository contains the airfoil dataset used in Compute-optimal data scaling for neural surrogates via multi-fidelity training, to be presented at NeurIPS 2026. All other datasets used in the publication are cheap to regenerate and we therefore provide the respective simulation source code in the accompanying GitHub repository.

Here, we build on the AirfRANS dataset, a subsonic 2D RANS dataset with varying NACA 4 and 5 digit geometries as well as inlet velocities and angles of attack. To create multiple fidelities, we modify the discretization and meshing stragegy.

Our dataset contains 1,627 samples, each at all 11 different fidelities ranging from about 7,600 to 180,000 mesh nodes.

File Structure

.
├── README.md
├── metadata.csv            one row per (case, level): 17,897 rows, see the column table below
├── 00001/                  one folder per case (sample_id, five digits, 00001 … 02000 with gaps due to filtering)
│   ├── 00.h5               coarsest mesh level
│   ├── 01.h5
│   ├── …
│   └── 10.h5               finest mesh level (the reference for the error columns)
├── 00002/
│   └── …
└── …                       1,627 case folders, 11 files each

The folder name is the sample_id and the file name the fidelity_id of the corresponding row in metadata.csv, so metadata.csv is the index of the dataset: filter it (by case, by level, by cost, by error) and open the files it points to. Sample ids are not contiguous because cases that failed the selection criteria were removed.

Each HDF5 file contains

path	content
data/fields	float64 array (N, 13), one row per mesh node; columns named by channels/
channels/<name>	column indices into data/fields: coords (2), u_in (2, inlet velocity), sdf (1, distance to the airfoil surface), normals (2, surface normals, zero off the surface), on_surface (1), u (2), p (1), wss (2)
mesh/connectivity	int32 (T, 3), triangles
data/edge_index	int64 (2, E), undirected mesh edges
mesh/aero_line_connectivity, mesh/aero_line_length, mesh/aero_line_normals	the airfoil surface as a polyline with segment lengths and normals
mesh/aero_to_internal	index of each surface node in the volume mesh
cond/fidelity_raw	the level's fidelity value in [0, 1]
reference/cd_openfoam, reference/cl_openfoam	force coefficients of this level
attributes metadata_json, coeff_ref_json	the level's run metadata (mesh cells, y⁺ estimates, solver timings, Cd, Cl) and the full OpenFOAM force-coefficient record

Velocities are in m/s, pressure is the kinematic pressure (m²/s²), wall shear stress in m²/s², coordinates in chord lengths with the leading edge at the origin. The mesh is clipped to x ∈ [−2.6, 4.7], y ∈ [−2.0, 1.9] around the airfoil; u_in is the inflow velocity vector, constant over the mesh.

metadata.csv columns:

column	meaning
sample_id, case_id	sample folder and the campaign's case name (NACA code, Re, AoA)
fidelity_id, fidelity_index, fidelity_raw, meshing_fidelity	the level (00 to 10, 0 to 10, 0.0 to 1.0)
n_nodes, dofs	mesh size of the level
reynolds, aoa_deg, u_inf	case parameters; inflow speed in m/s
cd_openfoam, cl_openfoam	force coefficients at the final iteration
rel_l2_error_ux, rel_l2_error_uy, rel_l2_error_p, rel_l2_error_wss, rel_l2_error	relative L2 error against the finest level of the same case, and their mean
simulation_time_raw	solver wall time in seconds on 16 cores
simulation_time	cost in core-seconds: wall time × 16, halved for
cd_rel_amp	relative Cd oscillation over the last 20% of the iterations (the stability criterion)
rel_residual, status, is_clean	solver bookkeeping; all rows are ok and clean

Example Usage

We provide a GitHub repository with a download script, data loaders and a multi-fidelity traning pipeline as an example.

Citation

@inproceedings{...,
  title     = {...},
  author    = {...},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}