Does This Fly?

A neural wind tunnel for wing sections

Loading model

Wing section

NACA 2412

2.0 %

How arched the section is. More camber, more lift at zero angle.

40 %

Where the highest point of the arch sits, from the front edge.

12 %

Maximum thickness, in percent of the chord.

Flight conditions

4.0°

Tilt of the section relative to the oncoming air. Beyond 15° the model extrapolates beyond its training data.

50 m/s

c = 1.00 m · Re = 3.2 M

Pressure coefficient Cp
−30+1
low pressure, suctionhigh pressure
Pressure along the chord
Downloading the neural model

The section is ready to shape. Flow predictions start as soon as the model is loaded.

Results

Lift Cl
—

How hard the section pushes air down to hold itself up.

Drag Cd
—

Air resistance of the section. Lower is better.

Lift-to-drag L/D
—

Efficiency: units of lift per unit of drag. Gliders reach 40 to 60 for the whole aircraft.

—

Lift vs angle of attack

Confidence —
    What this model does not see
    • An infinitely long wing section: no wingtip vortices, no 3D effects. A real wing makes less lift and more drag.
    • Trained on steady RANS simulations (k-ω SST turbulence model): unsteady stall, buffeting and laminar bubbles are not represented.
    • Incompressible air, Mach below 0.3.
    • Stall angles are predicted by a model trained up to 15°. They have not been validated against wind-tunnel data.

    About Does This Fly?

    You shape a 2D wing section and choose its angle of attack and speed. A graph neural network (MeshGraphNets type), running entirely in your browser, predicts the airflow on a mesh of about 17,000 points: velocity, pressure and skin friction. Lift and drag come from integrating pressure and friction over the surface.

    The model was trained on 800 simulations from the AirfRANS dataset (NACA 4- and 5-digit sections, Reynolds 2 to 6 million, angles of attack from −5° to 15°). On unseen sections its median error is 1.2 % on lift and 7.3 % on drag.

    Colours: blue is pressure below the free stream (suction), red is pressure above it. The red line on the surface marks where the flow separates (negative skin friction).

    Data: AirfRANS (Bonnet et al., NeurIPS 2022), ODbL v1.0. Inference: ONNX Runtime Web. Mesher: Triangle (J. R. Shewchuk). 3D: three.js. Fonts: Inter, JetBrains Mono (OFL).