Cashen/optiwing-airfoil-2d-full-v0

OptiWing Airfoil 2D — full

Everything: every optimizer iteration with the complete surface-field set, the raw section coordinates at each iteration, the sectional force breakdown, and adjoint sensitivities. Built with adjoint sensitivities.

This is the 2D OptiWing aerodynamic shape-optimization dataset, formatted to match the EngiBench airfoil convention (one row per optimization case). The three OptiWing-2D datasets share case_id, so a row in one joins to the same case in the others. All three are built from a single raw resampling pass, so their geometry and fields are identical where they overlap.

Geometry is chord-normalized (divided by the true chord c = max(x) − min(x)) and vertically shifted so the trailing-edge midpoint sits at y/c = 0: coords columns are (x/c, y/c) on a shared 192-point cosine arc grid (arc ∈ [0, 1], ordered TE → upper → LE → lower → TE). The 2D area is the analogue of the 3D paper's volume constraint; area_ratio is the achieved area / initial area at the optimum.

Splits

split	examples
train	878
val	49
test	107

Splits are at the case level and shared with the OptiWing 3D companion data, so paired 2D/3D cases land in the same split.

Features

field	type	description
case_id	int	optimization case id (shared across all three datasets)
mach	float	freestream Mach number (sampled input condition)
reynolds	float	Reynolds number (sampled input condition)
temperature	float	freestream static temperature [K] (sampled input condition)
cl_target	float	target lift coefficient (input condition / constraint)
area_ratio_min	float	minimum allowable area ratio (input constraint)
area_initial	float	chord-normalized area of the initial section (shoelace)

Per case the dataset additionally carries:

• designs: a list over every optimizer iteration present in the archive. Each entry has iter, stage (init / intermediate / opt), failed, angle_of_attack, the per-iteration objectives/constraints (cd, cl, cl_con_violation, area_ratio), chord, the raw section coordinates coords (x/c, y/c) and arc, the sectional force breakdown (cl_sec, cd_sec, cd_pressure_sec, cd_friction_sec), and the complete per-node surface-field set: cp, pressure, density, temperature, velocity_x, velocity_y, velocity_z, cf_x, cf_y, cf_z, yplus, stanton, force_drag, force_lift, sepsensor, sepsensor_ks, sepsensor_ks_area (192 each).
• adjoint_sensitivities (present when built with adjoint): per-iteration objective sensitivities w.r.t. the FFD shape design variables (sens_obj_wrt_shape), w.r.t. angle of attack (sens_obj_wrt_alpha), and the objective-gradient norm (grad_obj_norm). These are the solver's native adjoint outputs (w.r.t. the FFD parameterization); the geometry itself is given as raw coordinates in designs.

The geometry trajectory is the raw resampled coordinates at every iteration — not FFD cage values. velocity_z / cf_z are ≈0 in 2D (no spanwise flow) and kept only for schema parity.

Loading

from datasets import load_dataset
ds = load_dataset("Cashen/optiwing-airfoil-2d-full-v0")
print(ds["train"][0].keys())

Citation

This dataset does not yet have a companion paper or DOI. Until one exists, please cite the dataset directly by its Hugging Face URL and cite the prior-version works in "Built upon".

@misc{optiwing_airfoil_2d_2026_full,
  title        = {OptiWing Airfoil 2D (full): a 2D aerodynamic shape-optimization dataset},
  author       = {Cashen Diniz and Mark Fuge},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/Cashen/optiwing-airfoil-2d-full-v0},
  note         = {Version v0}
}

Built upon

This is a new dataset (wider parameter bounds, full surface fields, and complete optimization trajectories), but its lineage and format derive from prior work — please cite these as well:

@misc{diniz2025optiwing3d,
  title         = {OptiWing3D: A Diverse Dataset of Optimized Wing Designs},
  author        = {Diniz, Cashen and Fuge, Mark D.},
  year          = {2025},
  eprint        = {2512.12867},
  archivePrefix = {arXiv},
  doi           = {10.48550/arXiv.2512.12867},
  url           = {https://arxiv.org/abs/2512.12867}
}

@inproceedings{felten2025engibench,
  title     = {EngiBench: A Framework for Data-Driven Engineering Design Research},
  author    = {Felten, Florian and Apaza, Gabriel and Br\"{a}unlich, Gerhard and
               Diniz, Cashen and Dong, Xuliang and Drake, Arthur and Habibi, Milad and
               Hoffman, Nathaniel J. and Keeler, Matthew and Massoudi, Soheyl and
               VanGessel, Francis G. and Fuge, Mark},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS),
               Datasets and Benchmarks Track},
  year      = {2025},
  eprint    = {2508.00831},
  archivePrefix = {arXiv},
  url       = {https://arxiv.org/abs/2508.00831}
}

The EngiBench airfoil format this dataset mirrors is airfoil_v0: https://huggingface.co/datasets/IDEALLab/airfoil_v0.

License

cc-by-nc-sa-4.0