AeroZoo: A Benchmark of Well-Posed Flow Problems for Foundation Models of Aerodynamics
Organizations: NVIDIA, Santa Clara, California, USA.
Abstract
Neural surrogates for computational fluid dynamics (CFD) are trained and ranked on datasets that each vary one design family in one flow domain. Such datasets cannot test the least that an aerodynamic foundation model must do: predict CFD-quality incompressible flow in seconds for new shapes in new enclosures. On DrivAerML, larger models and longer training barely lower Transolver's error; doubling the training cases cuts it by about 12% up to 400 cases, exhausting the dataset. We introduce AeroZoo, a benchmark and training corpus of 3,815 time-mean solutions to the unsteady Reynolds-averaged Navier--Stokes (RANS) equations with a transition model, across seven geometry families, eight enclosure configurations and Reynolds numbers from to . Each sample poses a well-posed PDE problem in the form a solver receives it: the boundary meshes grouped by condition type, the prescribed values and the fluid properties. The 2,816 samples with a stable time mean form the canonical benchmark: two tasks, engineering metrics and 22 frozen splits, disjoint in geometry, that hold out geometry families, enclosures, boundary-condition layouts and Reynolds ranges. On the random split, boundary-condition tokens cut GeoTransolver's median velocity error from 76.9% to 8.4% of the inlet speed in single-seed runs (uniform inlet flow: 11.0%). ISLA, a transformer for boundary-driven PDEs, reaches 5.3% over four seeds, about 4 times the change from a finer mesh. On the same test samples, a held-out geometry family barely changes ISLA's error relative to uniform flow, and a held-out enclosure or boundary-condition layout raises it 2.4 to 9.2 times. The data and the scoring code will be released.
Figures & tables
| Dataset | Families | Enclosures | Operating points | Solver | Samples | Cells per sample |
|---|---|---|---|---|---|---|
| AeroZoo | 7 | 8 | – , continuous | URANS mean, transition | 3,815 | 0.6–24 M |
| DrivAerML ( Ashton et al., 2024c ) | 1 | 1 | one | hybrid RANS–LES | 500 (484 public) | 160 M |
| WindsorML ( Ashton et al., 2024a ) | 1 | 1 | one | wall-modeled LES | 355 | 280–300 M |
| AhmedML ( Ashton et al., 2024b ) | 1 | 1 | one | DDES | 500 | 15–20 M |
| DrivAerNet++ ( Elrefaie et al., 2024 ) | 1 | 1 | one | RANS | 24 M | |
| SHIFT-SUV ( Luminary Cloud, 2025 ) | 1 | – | two scales | DES | – |
| Split | Identifier | Train | Val. | Test |
| Random | full | 2,011 | 290 | 515 |
| Random, train | scarce | 321 | 290 | 515 |
| Random, train | super_scarce | 53 | 290 | 515 |
| Non-box enclosures | family_non_box | 1,850 | 274 | 692 |
| Dual-inlet boxes | profile_dual_inlet | 2,044 | 313 | 459 |
| reynolds | 1,577 | 238 | 1,001 |
| SI units | Per-case scales | ||||||||||
| Errors & ratio | Errors & ratio | ||||||||||
| Model | Given the boundary conditions | ratio | lower than ref. | ratio | lower than ref. | ||||||
| Free-stream reference | yes | unit-invariant (as right) | 0.110 | 0.160 | 0.059 | 1 | – | ||||
| Zero predictor | no | unit-invariant (as right) | 1.005 | 1.007 | 0.059 | – | – | ||||
| Transolver | no | 1.007 | 1.009 | 0.059 | 9.19 | 1% | 0.345 | 0.276 | 0.056 | 1.15 | 41% |
| GeoTransolver (no boundary tokens) | no | 1.004 | 1.005 | 0.062 | 8.88 | 1% | 0.769 | 0.804 | 0.081 | 3.90 | 13% |
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
| Configuration (share) | What its boundaries do | Distances to the faces |
|---|---|---|
| Open box (14.4%) | inlet opposite outlet; the four side faces slip | inlet 0.8–2.2, outlet 2–5, sides 0.8–2.5 |
| Ground-effect box (14.4%) | as the open box, with one side face a no-slip ground | inlet 0.7–2, outlet 1.5–4, sides 0.5–2, ground 0.05–0.3 |
| Confined box (14.4%) | inlet opposite outlet; all four side faces no-slip | inlet 0.3–1, outlet 0.7–2, walls 0.1–0.7 |
| Crossflow box (14.4%) | outlet on a face adjacent to the inlet; one to three other faces no-slip | inlet and outlet 0.5–2.5, sides 0.4–2, no-slip 0.1–0.8 |
| Dual-inlet box (14.4%) | a second inlet, 0.5 to 1.5 times as fast as the first, on a face adjacent to the first; outlet opposite the first; zero to three faces no-slip | inlets 0.5–2, outlet 1–3, sides 0.4–2, no-slip 0.15–1 |
| Opposed sphere (9%) | a smooth or lumpy sphere with inlet and outlet caps on opposite sides | sampled radius |
| Quantity | How it is set |
|---|---|
| Enclosure | one of eight configurations, with the shares of Table 4 |
| Reference length | log-uniform, 0.02 to 20 m. The body is scaled so that its largest extent before rotation is ; in a body-free duct, is the inlet’s hydraulic diameter |
| Inlet speed | log-uniform, 0.25 to 50 m/s, on the primary inlet |
| Reynolds number | one of the five decades from to , in equal shares; log-uniform within it |
| Viscosity | |
| Density | 1.225 kg/m 3 in every case |
| Boundary | Velocity | Pressure | Turbulence fields |
|---|---|---|---|
| inlet | prescribed (Dirichlet) | zero gradient | prescribed |
| outlet | zero gradient; prescribed where flow re-enters | zero (Dirichlet) | as velocity |
| no_slip | zero | zero gradient | wall functions ( , , ); zero gradient ( , ) |
| slip | no penetration, free tangential slip | zero gradient | zero gradient |
| Component | Fixed choice |
|---|---|
| Time integration | second-order backward, adaptive step |
| Courant limit | 10; initial step , at most |
| Pressure–velocity coupling | PIMPLE: one outer, two pressure, one non-orthogonal corrector |
| Momentum convection | bounded limitedLinearV 1 (limited second order) |
| Turbulence convection | bounded upwind (first order) |
| Gradients, interpolation | linear; Laplacian limited corrected 0.33 |
| Geometry family | Open box | Ground-effect box | Confined box | Crossflow box | Dual-inlet box | Opposed sphere | Symmetry hemisphere | Internal duct | Total |
|---|---|---|---|---|---|---|---|---|---|
| Blobs | 229 | 203 | 133 | 206 | 184 | 143 | 153 | 116 | 1,367 |
| Blender bluff | 78 | 70 | 55 | 67 | 51 | 49 | 38 | 31 | 439 |
| Procedural cars | 79 | 57 | 47 | 62 | 58 | 34 | 43 | 36 | 416 |
| Zero-to-CAD | 58 | 64 | 32 | 53 | 39 | 39 | 23 | 25 | 333 |
| Airframe Foundry | 56 | 56 | 55 | 51 | 46 | 40 | 41 | 22 | 367 |
| Streamlined | 65 | 56 | 52 | 56 | 67 | 47 | 45 | 46 | 434 |
| production finer | production wall-resolved | laminar share | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Case | cells R/G/W (M) | order | extrapolated | production | wall- resolved | ||||
| streamlined body | 3.2/10.8/3.4 | 0.7 | 0.2 | 0.9 | 2.3 | 0.5 | 0 | 0 | |
| car in ground effect | 2.8/8.7/3.3 | 0.5 | 1.6 | – | – | 2.7 | 0 | 0 | |
| Airframe Foundry body | 5.9/–/6.0 | – | – | – | – | 3.5 | 0 | 7 | |
| Blender bluff body | 4.4/14.8/4.5 | 5.3 | 1.6 | 2.2 | 9.2 | 0.6 | 0 | 0 | |
| body in an opposed sphere | 3.8/9.6/3.8 | 2.0 | 0.7 | – | – | 8.6 | 0 | 0 | |
| Metric | Task | Error of | Over | Weight | In units of | Defined on |
|---|---|---|---|---|---|---|
| , | interior | velocity, pressure | every interior point | cell volume | , | every sample |
| , , | interior (auxiliary) | , , | every interior point | cell volume | , , | every sample |
| , | interior (auxiliary) | , | every interior point | cell volume | the true field’s root mean square | every sample |
| , | interior | velocity, pressure | points within of a body face centroid | cell volume | , | samples with a body |
| , | interior | velocity, pressure | every interior point | none | the true field’s norm | every sample |
| , | boundary | force along ; force vector | the body’s faces | face area | ; | samples with a body |
| Split ( split_id ) | Question | Train | Validation | Test | ||||||
| canon. | released | planned | canon. | released | planned | canon. | released | planned | ||
| Random | ||||||||||
| full | random split of every case, about 70/10/20% | 2,011 | 2,713 | 3,104 | 290 | 386 | 440 | 515 | 716 | 816 |
| scarce | as full , with of its training part | 321 | 440 | 506 | 290 | 386 | 440 | 515 | 716 | 816 |
| super_scarce | as full , with of its training part | 53 | 74 | 84 | 290 | 386 | 440 | 515 | 716 | 816 |
| Held-out enclosure | ||||||||||
| Model setup | Inputs | Given the boundary conditions | Parameters |
|---|---|---|---|
| Transolver ( Wu et al., 2024 ) | the query points, each with the signed distance to the no-slip walls and its gradient | no | 9.0 M |
| GeoTransolver ( Adams et al., 2025 ) , no boundary tokens | the same point features; its geometry input is the interior query points | no | 27.6 M |
| GeoTransolver, boundary tokens | every sampled boundary face as a geometry token with its centroid, a one-hot condition type and the prescribed values (22 channels); the same point features | yes | 27.6 M |
| ISLA ( Sharpe, 2026b ) | all four boundary meshes with face condition types and prescribed values, plus interior query coordinates (divides by per-case scales inside the model) | yes | 67.5 M |
| Statement | Evidence | Where |
|---|---|---|
| At 400 cases the error changes by 2.6% from 20 to 354 million parameters | at equal budget ( , 16 cases per step): 0.0416, 0.0412, 0.0407 and 0.0418 for 20, 35, 105 and 354 million parameters | Figure 13 b |
| Four times the training compute lowers it by 3.8 to 4.1% | from to at 16 cases per step: 0.0416 to 0.0399 and 0.0412 to 0.0397 | Figure 13 c |
| Each doubling of cases lowers it by about 12% | 11% from 100 to 200 cases and 12% from 200 to 400 (mean of the 20- and 35-million-parameter models) | Figure 13 a |
| With unlimited cases the scaling floor is 21% below the best error | (0.0247–0.0365); best error 0.0393 | Table 22 ; Figure 13 a |
| Target averaging noise sets a model-independent lower bound of 0.0079 to 0.025 | at the central time scale, 30% of | Appendix H (Results) |
| Appendix only: single-case overfitting reaches 0.28 times (8% of ) | ; 0.0118 in strict float32 (0.38 times) | Figure 13 d |
| Fit | Fitted to | (95%) | Note |
|---|---|---|---|
| (used) | the 12 converged cells of the 20- and 35-million-parameter models | 0.0309 (0.0247–0.0365) | ; residual 2.1% |
| the same without | 0.0345 | ||
| Joint law | all 15 converged cells | 0.0318 ( to 0.0356) | at its bound (3.0): the size term is not identified |
| Joint law | every cell, converged or not | 0.0338 | |
| Law in , and case draws | the end points of all 57 runs at | 0.0327 | floor at for unlimited compute and model size: 0.0380 |
| Dataset | Families | Enclosures | Operating points | Solver | Samples | Cells per sample |
|---|---|---|---|---|---|---|
| AirfRANS ( Bonnet et al., 2022 ) | 1 | 1 | – , incidence | RANS | 0.25–0.3 M | |
| UniFoil ( Kanchi et al., 2025 ) | 1 | 1 | wide and Mach | RANS, transition | – | |
| CFDBench ( Luo et al., 2023 ) | 4 | 4 | varied | laminar or RANS | 739 | – |
PDE Surrogate Modeling
PDE surrogate modeling builds fast approximations of partial differential equation solutions for simulation, prediction, and design.
See all 361 papers