SuperWing: A 3D Transonic Wing Dataset

• SuperWing is a comprehensive open-access dataset featuring 4,239 parameterized wing geometries and 28,856 CFD solutions for 3D transonic swept-wing aerodynamics.
• It employs detailed wing geometry parameterization and steady RANS simulations using the Spalart–Allmaras turbulence model across realistic Mach and angle-of-attack conditions.
• The dataset supports pre-training neural surrogates and foundation-model workflows, enabling zero-shot generalization and multi-objective optimization in aerodynamic design.

SuperWing is a large-scale, open-access dataset of three-dimensional transonic swept-wing aerodynamics designed to support the development of machine-learning surrogates for aerodynamic prediction. The dataset comprises 4,239 parameterized wing geometries and 28,856 steady Reynolds-averaged Navier-Stokes (RANS) flow field solutions, each simulated using the Spalart–Allmaras turbulence model at a fixed Reynolds number of $2 \times 10^7$ over Mach numbers in $[0.75, 0.90]$ and angles of attack in $[2^\circ, 12^\circ]$. SuperWing’s expansive geometric diversity and standardized mesh format are tailored for pre-training neural surrogates, foundation-model workflows, and generalizable aerodynamic modeling tasks (Yang et al., 16 Dec 2025, Yang et al., 20 Apr 2026, Giral et al., 7 May 2026).

1. Purpose, Scope, and Motivation

SuperWing was constructed to address the data scarcity and geometric homogeneity limiting generalizable surrogate modeling in three-dimensional wing aerodynamics. Existing datasets concentrate on minor perturbations of a baseline wing or restricted planform variations, impairing transfer and generalization to new designs. SuperWing systematically spans the design space of transport-class swept wings using well-parameterized, realistic geometry families and broad aerodynamic conditions.

Its primary objectives are:

• To serve as a pre-training corpus for neural surrogates (e.g., ViT, AeroTransformer, AeroJEPA) that can be cheaply fine-tuned on scarce, downstream cases.
• To enable foundation-model approaches in aerodynamic prediction, supporting interactive design, zero-shot generalization to unseen configurations, and multi-objective optimization workflows (Yang et al., 20 Apr 2026).
• To support machine learning research into continuous surrogate modeling, latent-design interfaces, and generalization across broad aerodynamic regimes (Giral et al., 7 May 2026).

2. Wing Geometry Parameterization

SuperWing parameterizes each wing using a combination of global planform and spanwise-varying geometric controls. The core features are:

• Planform Parameters (5 total):
  • Leading-edge sweep, $\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$
  • Aspect ratio, $AR \in [8, 11]$
  • Taper ratio, $TR \in [0.15, 0.40]$
  • Kink location, $\eta_k \in [0.36, 0.42]$ (fraction of semi-span)
  • Root chord adjustment, $\kappa_{\mathrm{root}} \in [0.10, 1.10]$
• Spanwise Distributions:
  • Airfoil shape at each spanwise station is generated from a baseline using the Class-Shape Transformation (CST) method with 20 coefficients (10 per surface), controlling upper and lower shapes:

  $$C_{\text{shape}}(\xi) = \sum_{i=0}^n C_i \, \xi^i (1 - \xi)^{n-i}$$

  where $\xi$ is normalized chordwise position. - Thickness and camber are modulated by B-spline functions with 5 control points per mode. - Dihedral is varied with two spline points (kink, tip). - Geometric twist is set using a 5-point spanwise spline.

• Total Parameterization:

The complete geometry vector consists of 37–54 parameters, depending on dataset version, encompassing planform, airfoil, dihedral, and twist (see details in (Yang et al., 16 Dec 2025, Giral et al., 7 May 2026)). This parameterization permits realistic, non-trivial morphologies—avoiding reliance on single-wing perturbations.

3. CFD Simulation Protocol and Dataset Structure

• CFD Setup:

  • Mach: $[0.75, 0.90]$2
  • Angle of attack: $[0.75, 0.90]$3
• Mesh and Surface Representation:

Each wing is resolved on a high-fidelity structured mesh (typ. $[0.75, 0.90]$4 grid points for learning, $[0.75, 0.90]$532,000 unstructured points for point-cloud experiments). Surface meshes use right-handed coordinates ($[0.75, 0.90]$6 downstream, $[0.75, 0.90]$7 spanwise, $[0.75, 0.90]$8 normal).

• Flow Outputs:
  • Surface pressure coefficient, $[0.75, 0.90]$9 ($[2^\circ, 12^\circ]$0 or as a point cloud).
  • Surface friction vector, $[2^\circ, 12^\circ]$1 ($[2^\circ, 12^\circ]$2), decomposed as streamwise ($[2^\circ, 12^\circ]$3) and spanwise ($[2^\circ, 12^\circ]$4).
  • Integrated aerodynamic coefficients ($[2^\circ, 12^\circ]$5, $[2^\circ, 12^\circ]$6, $[2^\circ, 12^\circ]$7) are also provided per sample.
• Operating Conditions:
  • Each wing is evaluated at up to eight $[2^\circ, 12^\circ]$8 pairs, resulting in mean coverage of 6.8 conditions per geometry.
• Data Organization and Files:
  • coords.npy – $[2^\circ, 12^\circ]$9 coordinates
  • Cp.npy – $\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$0 pressure
  • Cf.npy – $\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$1 friction vector
  • cond.json – Mach, AoA
  • coefs.json – integrated coefficients
  • A global index lists all cases along with geometry parameters (Yang et al., 16 Dec 2025, Yang et al., 20 Apr 2026).

4. Statistical Diversity and Dataset Splits

• Diversity and Intrinsic Dimension:
  • 4,239 unique planform/airfoil/twist/dihedral combinations, with continuous uniform sampling across parameter ranges.
  • Principal component analysis (PCA) indicates 5 modes explain 99% of shape variance, 11 for 99.9%—demonstrating dimensional regularity yet sufficient complexity for generalization studies.
  • No explicit data augmentation applied; diversity arises solely from the parameter sampling.
• Splitting and Coverage:
  • Standard split: 90% for training, 10% for held-out evaluation with splits by geometry (no cross-leakage).
  • Some studies use 80/10/10 splits (train/val/test) by geometry.
  • Validation sets are extracted from training during model selection.

5. Machine Learning Benchmarks and Generalization Tests

SuperWing supports rigorous benchmarking for high-capacity surrogates:

• Architectures:
  • U-Net, Vision Transformer (ViT), Transolver, Point-cloud transformers (AeroJEPA).
• Input Modalities:
  • Structured mesh (coords, operating conditions).
  • Unstructured surface point cloud (for point-cloud networks).
• Target Outputs:
  • Surface fields $\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$2, $\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$3, $\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$4 and integrated coefficients.
• Training Details:
  • Mean squared error loss on fields, Adam/AdamW optimizers, batch sizes 4–32; standard normalization/centering as preprocessing.
• Representative Benchmark Results (Yang et al., 16 Dec 2025):

Model	$\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$5 err (%)	$\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$6 (%)	$\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$7 (%)	$\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$8 ($\Lambda_{\mathrm{LE}} \in [25^\circ, 40^\circ]$9)	$AR \in [8, 11]$0 ($AR \in [8, 11]$1)	Time (h)	Mem (GB)
U-Net	1.101	0.642	0.698	23.41	14.78	17.1	5.55
ViT	0.329	0.245	0.281	2.76	2.48	12.6	5.78
Transolver	0.359	0.271	0.310	2.71	2.53	37.9	16.42

ViT exhibits the best surface and global aerodynamic accuracy (e.g., 2.48 drag-count error) with low memory and wall time.

• Zero-Shot Generalization:
  • ViT pretrained on SuperWing matches CRM and DLR-F6 benchmark wings without fine-tuning, reproducing surface shocks and global $AR \in [8, 11]$2, $AR \in [8, 11]$3 curves.
  • Out-of-distribution (OOD) predicted error in drag-counts remains comparable to in-domain test error (2–4 counts).
• Latent-Space Surrogates:
  • AeroJEPA learns geometry- and condition-encoded latent predictions, supporting scalable field reconstruction and smooth latent optimization (Giral et al., 7 May 2026).

6. Practical Use and Integration

• Loading and Preprocessing:
  • Mesh coordinates are normalized by root chord length ($AR \in [8, 11]$4), centered at the origin.
  • Surface field outputs are standardized to $AR \in [8, 11]$5 based on training set extrema.
  • Provided Python loaders (PyTorch Dataset, etc.) enable direct ingestion into training loops.

$AR \in [8, 11]$7

• Augmentation:
  • No baseline augmentation; optional noise, twist, or dihedral perturbations can be layered for regularization if desired.
  • Structured mesh shape ensures simple batching—collation is row-wise stacking.
• Foundation-model and Interactive Scenarios:
  • Pretrained models can be fine-tuned on new domains using $AR \in [8, 11]$61,000 cases.
  • Interactive design tools and downstream co-optimization frameworks leverage standardized format and open-access pre-trained weights (Yang et al., 20 Apr 2026).

7. Impact and Research Directions

SuperWing enables research in:

• Pretraining large-scale, generalizable 3D aerodynamic surrogates for design automation and iterative optimization.
• Studying latent representations that support controllable interpolation, design-variable probing, and "CAD-free" gradient-based optimization (Giral et al., 7 May 2026).
• Benchmarking scalable architectures and zero-shot transfer, with evidence that ViT and AeroJEPA capture regime-crossing generalization.
• Supporting workflows for generative aerodynamic shape design, surrogate-assisted co-optimization, and rapid evaluation in architecture screening and conceptual design.
• Future directions include extension to unsteady flows, multi-point flight envelopes, and integration with multi-modality structural models (Yang et al., 16 Dec 2025, Giral et al., 7 May 2026).

SuperWing’s expansive geometric coverage and rigorously standardized data modalities position it as a reference benchmark for aerodynamic surrogate model research as well as practical tools for data-driven wing design.