AI-Assisted Capture of Missing Physics: PINNs Trained on Wind Tunnel Data on a Single-Board Computer
March 17, 2026
Neural networks that obey the laws of fluid flow, trained not on simulations but on wind-tunnel measurements from 1959, can capture physics that conventional CFD misses. This project did it on a six-watt computer with no graphics card, driven entirely by text messages.
Disclaimer: This work is a technology demonstration. The results have not been independently validated against new experimental campaigns. All experimental data are sourced from publicly available literature. This study should not be used for engineering design decisions without further testing and validation.
The physics simulations miss
A simulation is only ever as good as the physics written into it, and even sophisticated turbulence models leave things out. Where a wing meets a body, real flows form horseshoe vortices and corner separation that common models misjudge. Trailing-edge noise can nudge the point where a boundary layer turns turbulent. Turbulence in the oncoming air and roughness on the surface shift transition and separation in ways most simulations simply omit. And real separated flows develop three-dimensional structures that standard two-equation models cannot resolve.
Train a neural network on simulation results alone and it inherits every one of those blind spots. Its predictions come out smooth, plausible and wrong in exactly the same places as the source.
Old data, new networks
The alternative has been sitting on library shelves for more than sixty years. In 1959, Abbott and Von Doenhoff’s Theory of Wing Sections gathered decades of NACA wind-tunnel measurements of lift, drag and surface pressure into tables anyone can read. Sixty years later, in 2019, Raissi and colleagues introduced physics-informed neural networks, or PINNs, which are trained to satisfy equations such as Navier–Stokes as part of their learning.
This project joins the two. Measured data for the NACA 0014 aerofoil at a Reynolds number of three million, the Navier–Stokes equations as constraints, and neural networks that produce the flow around the wing section. The crucial point is that measurements carry the fingerprints of everything that was really happening in the tunnel, wing–body effects, acoustics, tunnel turbulence and transition included, so a model trained on them absorbs physics a simulation-trained model would never see. And because the networks are trained at the same Reynolds number as the experiments, there is no scaling gap between model and data.
The dataset is small: 17 lift and 17 drag values, 9 moment values, 51 pressure points and a handful of Reynolds-number points, drawn from Abbott and Von Doenhoff, NACA Report 460 from 1933 and Mark Drela’s XFOIL references. With physics doing much of the work, small is enough.
A six-watt laboratory

All of the training ran on a ZimaBoard 2, a low-cost single-board computer that draws about six watts. There was no graphics card: PyTorch ran on four processor threads, and all five models trained in roughly 74 minutes.
Three networks, three questions
The first network, AeroCoeffNet, answers the simplest question: given an angle of attack and a Reynolds number, what are the lift, drag and pitching moment? Alongside the measurements it is taught three facts of aerodynamics: a symmetric aerofoil must behave symmetrically, lift should rise at about 2π per radian at small angles, and drag can never be negative.

The second, CpNet, predicts the pressure along the upper and lower surfaces at any point on the chord and any angle. It learns that the flow must come to rest at the leading edge, leave the trailing edge smoothly, and that the pressure difference between the surfaces must add up to the lift.

The third, FlowFieldNet, is the heart of the project. Given any point in space, it returns the velocity and pressure of the air there, and it is trained to satisfy the Navier–Stokes equations at thousands of points, to bring the air to rest on the aerofoil’s surface, and to match the measured surface pressures wherever they exist. Two ideas make it work at such a high Reynolds number. Curriculum learning starts the network on a thick, slow, easy flow and gradually thins the viscosity to the real value, so it learns the big picture before the fine detail. And the sample points are clustered close to the surface, where the boundary layer is thinnest. Separate models were trained at 0°, 4° and 8°.

Measurements already contain all the physics. The network’s job is to learn it.
Seeing the flow
To move from a wing section to a wing, the 4° solution was stretched into a finite wing 1 m in span and 0.2 m in chord, with two physical corrections: Prandtl’s elliptic loading, which fades the lift to nothing at the tips, and a pair of tip vortices that create the downwash behind. The result can be explored in ParaView, a full 960,000-point volume ready to slice and colour any way you like.



For everyone without ParaView, there is a single web page (open the animation) in which 72 animated streamlines flow around the wing. It switches between speed and pressure colouring, lets you orbit, zoom and pan, and runs in any modern browser with nothing to install.

How good is it?


Lift agrees well across the linear range, up to about ±10°. The largest errors appear near the stall at ±16°, which is expected, because the networks do not model separation or transition. The drag polar reproduces its characteristic bucket shape, and the pressure distributions agree well at all three angles. Four more flow-field models, at 2°, 6°, 10° and 12°, took about fifteen minutes each. Once trained, the networks become instant aerodynamic look-up tools: lift, drag and moment at any angle and Reynolds number, surface pressures anywhere on the chord, and full flow fields at seven angles from 0° to 12°.


Built by conversation
None of the code was written by hand. The whole workflow was driven by plain-language instructions sent through OpenClaw, which relays commands to the ZimaBoard from a Telegram chat and can use different language models behind the scenes, in this case two different LLMs, interchangeably. From the first instruction to animated 3D streamlines took under an hour, with no manual coding, no mesh generation and no choice of turbulence model.
Open and free
A conventional CFD workflow usually rests on commercial software with annual licences, and with meshing and post-processing tools added, the software can cost more than the computer it runs on. Everything here is open source: PyTorch for the networks, NumPy and SciPy for the numerics, Matplotlib and PyVista for figures and meshes, ParaView for 3D exploration, Linux for the operating system and OpenClaw for the conversation. The whole pipeline, from training data to 3D streamlines, costs nothing in licences and can be reproduced by anyone.
Where it falls short
This is a technology demonstration, and its results have not been validated against independent experiments. The curriculum schedule matters a great deal; different rates or starting points can give noticeably different answers. Separation beyond about 12° is poorly predicted, unsteady effects such as vortex shedding are absent, and the extension to three dimensions is an approximation of real wingtip behaviour. The next steps are comparison with RANS solutions, a more robust curriculum, and unsteady capability for dynamic stall.
A shift in philosophy
The headline result is that meaningful aerodynamics at a Reynolds number of three million, matched to the original experiments, can be computed on a six-watt computer with no graphics card in about 74 minutes, using measurements made between the 1930s and the 1950s.
The deeper point is a change of approach. Instead of trying to model every physical mechanism from first principles, and inevitably missing some, we can anchor neural networks in measurements that already contain all the physics: wing–body interactions, acoustics, transition and turbulence, all waiting in the data to be learned. Conventional CFD will remain essential for detailed design and diagnosis. But where fidelity to real-world behaviour matters most, networks informed by experimental data offer a compelling alternative. Good data, sound physics, inexpensive computing and the right tools: that era is only just beginning.