All summaries · 2026-11-04

Ground-Wing: A Parametric 3D Wing-in-Ground-Effect Dataset for Benchmarking Neural Surrogate Models

Andrea Cenzato, Lukas Karch, Andre Luckow, et al.

arXiv:2611.04571 · abstract · pdf · summarized by meta/muse-spark-1.3-contributor on 2026-11-04

The one-paragraph version

This paper releases Ground-Wing, a public set of about 770 airflow simulations around a race-car style front wing flying very close to the ground, plus a head-to-head test of six modern AI models that try to replace slow fluid simulation with fast prediction. The wing shape is varied systematically — its tilt, twist, length, sweep, airfoil cross-section, and height above the ground — to create 194 distinct wings. Each case was solved with industry-standard time-averaged fluid simulation, and the AI models were asked to predict surface pressure and surface friction, from which total lift and drag are calculated. Transformer-style models won clearly, but all models got noticeably worse when asked to handle extreme drag shapes or airfoil shapes they had never seen, which is exactly the kind of generalization test the authors wanted to create.

The problem they're solving, and why it matters

Designing a front wing for motorsport means running hundreds of computational fluid dynamics simulations — computer programs that step by step solve for how air moves. The underlying equations for viscous air motion, called the Navier-Stokes equations, have no pencil-and-paper solution, so software must iterate on a fine 3D mesh. That is accurate but slow: meshing, watertight CAD cleanup, and solving take people and thousands of CPU hours. Neural surrogates promise a shortcut: learn the mapping from shape directly to airflow, then predict in seconds from a simple point cloud with no mesh. The catch is trust outside the training box. Existing public datasets do not test what makes a front wing hard: ground effect. When a wing sits centimeters above a moving road, the narrow gap acts like a pinched hose — air speeds up and suction grows fast — until at very low gaps the flow separates and downforce collapses, a nonlinear event called choking. No prior wing dataset varies ground clearance as a design variable; road-car datasets keep ride height fixed. Without a benchmark that includes this wall proximity, you cannot know if a surrogate is ready for real front-wing work.

The key idea, in plain words

Build a focused, real-world-like obstacle course specifically for surrogates: one wing family that is simple enough to parametrize cleanly, realistic enough to matter for LMDh prototype racing, and hard enough because of the ground. Then fix the rules of the game — documented shapes, validated solver settings, open file formats, and three predefined train-test splits — and invite six different AI architectures to run it. Think of it like a driving exam with three parts: normal city driving, driving at the icy edges of the envelope, and driving a car body shape you have never seen before. The dataset is the exam track, the splits are the three tests, and the six models show how far the field still has to go.

How it actually works, step by step

The wing and its knobs. The test piece is a single-element wing with rectangular endplates, roughly the size of an LMDh front splitter with 2-meter span. Five continuous knobs are varied: tilt at the center, called centerline angle of attack, extra twist at the tip, length of the chord at the center and at the tip, and how far back the tip is swept. A sixth knob is discrete: which airfoil cross-section is used, chosen from 27 options. Span, endplate shape, and smoothing fillets are held fixed to keep the problem tractable.

Picking which wings to simulate. Random picking leaves clumps and holes. So the authors first fill the five-dimensional knob space evenly with a low-discrepancy Sobol sequence, a quasi-random method designed to avoid gaps, then throw out illegal combos that would stall or break the rules box. Then they bias the survivors toward sensitive regions — places where a small knob tweak causes a big change in lift or drag — using a sensitivity function built from single-knob CFD sweeps fitted with smooth curves. A coverage term prevents collapse onto only the most sensitive corners. This yields 174 wings from 12 base airfoils plus 20 wings from 15 blended or extra airfoils made by averaging coordinates of known high-lift sections.

Simulating the air. Each wing is placed above a simulated wind-tunnel floor with a moving belt patch underneath to mimic a rolling road. All runs use the same incompressible time-averaged approach, called Reynolds-Averaged Navier-Stokes with a realizable turbulence model for approximating gustiness, run for 1000 iterations in Siemens STAR-CCM+. The mesh has about 28 million cells, finest near the surface. A mesh study showed this is within 1% of a much finer 77-million-cell mesh, and adding many more near-wall layers changed lift and drag by only 0.6%, justifying wall functions. Four flat ride heights are simulated per wing. A separate sweep confirmed all heights sit in the smooth, well-behaved regime, not in choked flow. Total cost was about 38,300 core-hours. Each sample ships with geometry, surface fields, volume fields, and a spreadsheet of parameters and forces.

Testing the AIs. Six models span the major families: DoMINO for convolution on point clouds, X-MeshGraphNet for message passing on graphs, AB-UPT, GeoTransolver and GAOT for transformers, and AeroJEPA for joint-embedding prediction where the model first learns a semantic latent code then decodes fields. All are trained to predict only two surface quantities — pressure coefficient, a normalized number saying how much the local pressure pushes compared to the free stream, and wall shear stress vector, the local friction force per area from air rubbing the skin — from which lift coefficient, a normalized measure of total downforce, and drag coefficient, a normalized measure of total resistance, are integrated. Three splits test interpolation, drag extremes, and unseen airfoil shapes. Error is reported both as percentage error on lift and drag and as normalized field errors, plus whether the model gets the direction of change right between two designs.

What they showed — results, honestly, with limitations

On the standard stratified test, transformer-style models led clearly and in the same order for both lift and drag with one small swap. The best model, AB-UPT, averaged 0.85% error on lift and 0.89% on drag with near-perfect trend correlation. Next came GeoTransolver at 2.33% and 3.15%, then AeroJEPA at 2.83% and 4.38%. GAOT, DoMINO, and graph-based X-MeshGraphNet trailed at roughly 5-9% on lift and 6-14% on drag, with heavier outlier tails. Field errors mirrored this ranking, and errors concentrated in high-gradient spots like the suction peak and endplate junctions rather than spreading evenly. Generalization was the hard part. Retrained on drag-extreme splits, every model's lift error grew 1.5 to 3.4 times. Tested on unseen blended airfoils without retraining, drag degraded sharply — for example GeoTransolver's drag error jumped to about 19% — while lift held up better. In practical terms, AB-UPT predicted the correct up-or-down direction of a design change 99% of the time, while the worst model managed only about 91% for lift and 85% for drag. Limitations are stated plainly: single wing alone, no multi-element flap interactions, no rake or roll, fixed span and endplates, constant cross-section along span, linear sweep and twist only, time-averaged rather than eddy-resolving turbulence, and six missing runs from processing errors. The sensitivity sweeps and mesh studies used a proprietary BMW airfoil that cannot be shared. The paper does not evaluate volume-field accuracy or training energy cost.

Why this might matter to me

For spatial reasoning in vision-language models and world models, this is a clean stress test of proximity reasoning: the same wing at 75 mm versus 30 mm above ground behaves nonlinearly, not as a linear offset. A model that merely memorizes shape without representing the gap-as-a-channel will fail, which is why DoMINO needed the moving belt explicitly added to distinguish ride heights. That lesson transfers to any world model that must predict how confinement changes physics. For text-to-CAD and parametric geometry, the paper is a worked example of how to turn knobs into stable CAD via CATIA, enforce legality-box and stall constraints, and sample densely where outputs are sensitive. That sampling logic — evenly cover, then over-sample where gradients are large — is reusable for generating training data for generative CAD. For 3D generation and scene representation, the results hint at which encodings carry shape best. Anchor-based attention that summarizes geometry with a few thousand anchor points generalized best; graph construction that leaked information across thin wing sections failed qualitatively. The interpolated-airfoil split is essentially an out-of-distribution generation test: blended but familiar shapes still broke drag prediction, suggesting future generative and surrogate hybrids will need explicit airfoil priors or physics-informed losses, not just larger point clouds.

Terms worth knowing