SimpleWing / README.md
yunplus's picture
Update README.md
5fb9950 verified
|
Raw History Blame Contribute Delete
6.28 kB
metadata
pretty_name: Single-Segment Transonic Swept Wings
tags:
  - aerodynamics
  - computational-fluid-dynamics
  - scientific-ml
  - transfer-learning
  - 3d
  - arxiv:2409.12711

Single-Segment Transonic Swept Wings

This dataset supports physics-embedded transfer learning from 2D airfoils to 3D swept wings. (paper). We provide three linked sets of aerodynamic data for this workflow.

  • 2D airfoil CFD provides 22,104 supercritical airfoil cases for pretraining the airfoil model.
  • 3D wing CFD provides surface geometry and aerodynamic coefficients for 1,842 single-segment swept-wing cases, which serve as reference data for the transfer model.

The CFD surface data come from Reynolds-averaged Navier–Stokes (RANS) simulations with CFL3D and the SST turbulence model.

  • Vortex lattice method (VLM) provides a low-fidelity spanwise distribution of sectional lift for each 3D wing case. These estimates help determine the effective operating conditions for the corresponding 2D sections.

It's part of floGen collection.

Files

All seven files are NumPy `.npy``. Shapes below follow the on-disk axis order.

File Shape Contents
airfoilndata.npy (22104, 3, 321) Airfoil surface profiles. Axes are case, channel, and contour point. Channels are the normalized surface ordinate y/c, pressure coefficient Cp, and stored signed skin-friction channel.
airfoilnindex.npy (22104, 29) Airfoil case metadata, including geometry and condition identifiers, freestream Mach number, angle of attack, and CST geometry coefficients.
wingdata.npy (1842, 6, 61, 321) Wing surface geometry and CFD coefficients. Axes are case, channel, spanwise station, and contour point. Channels are x, y, z, Cp, scaled streamwise skin friction, and scaled spanwise skin friction.
wingindex.npy (1842, 35) Wing IDs, operating conditions, planform and section parameters, and integrated lift and drag coefficients.
wing101roottip.npy (1842, 3, 2, 321) Root and tip geometry in the wing coordinate frame. Axes are case, coordinate (x, y, z), section (root, tip), and contour point.
winggeomroottip.npy (1842, 2, 3, 321) Root and tip geometry in sectional coordinates. Axes are case, section (root, tip), coordinate, and contour point. The in-plane axes are aligned with the local chord and divided by its length. The third coordinate retains the wing-frame spanwise position.
cldistribution_vlm.npy (1842, 101) Low-fidelity spanwise sectional lift coefficient cl from VLM, with 101 values per wing case.

Wing cases

All five wing arrays use the same case order. Use wingindex[i, 0] to recover the case ID associated with row i. The CFD surface tensor runs from root to tip along its 61-station axis. The 101-value VLM distribution uses its own spanwise sampling.

The wing surface file stores 200 × Cf_streamwise in channel 4 and 250 × Cf_spanwise in channel 5. Divide the stored values by those factors to recover the dimensionless friction coefficients. The wing geometry uses a unit root chord. The sampled wing freestream Mach numbers are approximately 0.72–0.85, and the angles of attack approximately 1–6°.

wingindex.npy columns (zero-based)

Columns Description
0 Wing case ID
1–2 Angle of attack (degrees), freestream Mach number
3–7 Leading-edge sweep (degrees), dihedral (degrees), aspect ratio, taper ratio, stored twist magnitude (degrees)
8–10 Tip-to-root thickness ratio, reference area, root relative thickness
11–20 Upper-surface class–shape transformation (CST) coefficients
21–30 Lower-surface CST coefficients
31–32 Lift and drag coefficients reported by the CFD solver
33–34 Lift and drag coefficients reconstructed from the surface distributions

Airfoil cases

airfoilndata[i] and airfoilnindex[i] describe the same 2D case. There can be multiple operating conditions for a single airfoil geometry. The two initial metadata columns contain the airfoil geometry ID and the condition ID within that geometry. Columns 2 and 3 contain freestream Mach number and angle of attack in degrees. Columns 6–15 and 16–25 are the upper- and lower-surface CST coefficients, respectively. The remaining metadata columns and the airfoil skin-friction scale are unspecified in this release documentation.

The airfoil files contain 1,842 geometry IDs with 12 cases per ID. The condition IDs retain their original values and can have gaps. The 2D and 3D index files share 1,841 IDs. To link the two groups, join on column 0 and account for the unmatched ID in each group. The arrays are provided without predefined split files. For evaluation on unseen airfoil shapes, group 2D cases by geometry ID when constructing a split.

Load a sample

Run this example from the dataset repository root. Memory mapping allows you to inspect one case at a time.

from pathlib import Path

import numpy as np

data_dir = Path('data')


def load(name):
    return np.load(data_dir / name, mmap_mode='r', allow_pickle=False)


wing = load('wingdata.npy')
wing_index = load('wingindex.npy')
vlm = load('cldistribution_vlm.npy')
airfoil = load('airfoilndata.npy')
airfoil_index = load('airfoilnindex.npy')

i = 0
wing_id = int(wing_index[i, 0])
x, y, z, cp, cf_streamwise_scaled, cf_spanwise_scaled = wing[i]
cf_streamwise = cf_streamwise_scaled / 200
cf_spanwise = cf_spanwise_scaled / 250
sectional_cl_vlm = vlm[i]

j = 0
airfoil_id, condition_id = airfoil_index[j, :2].astype(int)
airfoil_y, airfoil_cp, airfoil_cf_stored = airfoil[j]

Citation

@article{yang2025rapid,
  author  = {Yang, Yunjia and Li, Runze and Zhang, Yufei and Lu, Lu and Chen, Haixin},
  title   = {Rapid Aerodynamic Prediction of Swept Wings via Physics-Embedded Transfer Learning},
  journal = {AIAA Journal},
  year    = {2025},
  volume  = {63},
  number  = {6},
  pages   = {2545--2559},
  doi     = {10.2514/1.J064576}
}

Journal article