AeroZoo: A Benchmark of Well-Posed Flow Problems for Foundation Models of Aerodynamics

Peter Sharpe, Carmelo Gonzales, Corey Adams, Mehdi Ataei, Sanjay ChoudhryarXiv 2026 · arXiv

TL;DR

AeroZoo is a massive new benchmark for aerodynamic foundation models, featuring 3,815 high-fidelity CFD solutions across diverse shapes, enclosures, and boundary conditions. It shifts the paradigm from simple "shape-to-flow" mapping to learning the actual PDE solution operator by explicitly including boundary conditions as model inputs, revealing that current surrogates struggle far more with new environments than with new shapes.

What Problem Does This Paper Address?

Most existing 3D datasets for neural Computational Fluid Dynamics (CFD) are "one-trick ponies." They typically vary a single design family (like a specific sedan model) within a fixed wind tunnel environment. While useful for optimizing a specific product, these datasets fail to test the true goal of an aerodynamic foundation model: the ability to predict flow for any shape in any environment.

The authors argue that scaling models on limited datasets hits a "bottleneck" where adding more parameters doesn't help because the geometry diversity is exhausted. Furthermore, many current models are "boundary-blind"—they learn a map from shape to flow field without knowing the boundary conditions (BCs). This means they can't distinguish between a car in an open field, a car in a tunnel, or a car near a second inlet, even though the physics in these cases are vastly different.

Figure 1. One car in four enclosures yields four distinct flows at the same inlet speed and Reynolds number: (a) open box, (b) ground- effect box, (c) confined box and (d) dual-inlet box (center-plane speed with streamlines; insets show full domain boundaries; x along primary inlet, z vertical). Within one car length of the body, the three single-inlet flows differ by 3 to 7% of Uref (more than mesh refinement; Section 6), no-slip tunnel walls in (c) raise drag by 13%, and the second inlet in (d) changes near-body velocity by 152 to 177% and raises Cx from 0.026 to 0.55.

Methods

The core innovation is the formulation of well-posed flow problems. In AeroZoo, each sample is treated as a complete boundary-value problem. The model receives:

  1. Boundary Meshes: Grouped by condition type (inlet, outlet, no-slip wall, slip wall).
  2. Prescribed Values: Face-wise values for velocity, pressure, and turbulence.
  3. Fluid Properties: Kinematic viscosity and density.

The authors introduce boundary tokens to encode this information, allowing models like GeoTransolver and ISLA (an Integrated Spectral-Laplacian Attention transformer) to "see" the environment. They also utilize a DomainMesh schema that separates PDE inputs from scoring scales, ensuring models learn the underlying physics rather than memorizing dataset-specific constants.

Figure 3. An annotated DomainMesh sample showing how PDE inputs and query coordinates are separated from withheld scoring metadata.

Datasets

The paper introduces AeroZoo, which consists of:

  • 3,815 time-mean solutions to the unsteady Reynolds-averaged Navier–Stokes (URANS) equations.
  • 7 Geometry Families: Including procedural cars, aircraft (Airframe Foundry), CAD parts (Zero-to-CAD), and "blobs."
  • 8 Enclosure Configurations: Ranging from standard boxes to internal ducts and symmetry hemispheres.
  • Reynolds Numbers: Spanning five decades ($10^3$ to $10^8$).
  • 22 Frozen Splits: Designed to test out-of-distribution generalization by holding out specific geometry families or enclosure types.
Figure 2. The zoo: every tile is a distinct geometry posed as its own boundary-value problem. Columns show the eight enclosure configurations (sketched above by boundary type; opposed spheres may be smooth or lumpy) and rows the seven geometry families; each tile displays the first mean_stable sample in seed order (with a body for ducts), cut along the primary inlet velocity over its full fluid domain on its own scale (colorbar linear on 0–1 and 1–2.5).

Key Insights and Ideas

  • Boundary Conditions > Geometry: Surprisingly, holding out a geometry family barely impacts a model's error. However, holding out a new enclosure or boundary layout increases error by 2.4 to 9.2 times. This suggests that learning the "environment" is the current frontier for foundation models.
  • The "Boundary Token" Breakthrough: Simply adding boundary tokens to GeoTransolver cut its median velocity error from a staggering 76.9% down to 8.4%. Without these tokens, the model is essentially guessing the wind direction and speed.
  • Scaling Diminishing Returns: On the automotive-only DrivAerML dataset, quadrupling compute only lowers error by 4%. AeroZoo provides the diversity needed to break through this plateau.
  • The Error Stack-up: The authors provide a rigorous "error stack-up" analysis, showing that surrogate errors (around 5.3%) are still significantly larger than the numerical sensitivities of the CFD solver itself (around 1.2%). This proves there is still a massive "generalization gap" to close.

Key Results

  • ISLA Performance: The ISLA model achieved the best results, with a median velocity error of 5.3% of the inlet speed.
  • Generalization: ISLA beat the "uniform flow" baseline on 98% of samples.
  • Enclosure Sensitivity: Models performed best in "open" environments (like opposed spheres) where the far-field flow is simple, but struggled in complex "dual-inlet" or "crossflow" boxes where the boundary conditions dictate complex internal flow structures.
  • Efficiency: These neural surrogates can predict flow fields in seconds, compared to the hours required for traditional CFD solvers.

Limitations and Future Work

  • Incompressible Flow: The current benchmark is limited to incompressible, isothermal flow. Future versions will likely need to address compressible flow and heat transfer.
  • Time-Mean Solutions: AeroZoo focuses on time-mean solutions of URANS. While practical for engineering, it doesn't capture the full transient dynamics of highly unsteady flows.
  • Solver Fidelity: While URANS is an industrial standard, it is less accurate than costlier methods like Large Eddy Simulation (LES). The authors suggest that as surrogates improve, the benchmark will need to move toward these higher-fidelity targets.
  • Task Breadth: The current focus is on velocity and pressure; future work could expand to more complex engineering metrics like conjugate heat transfer or solid mechanics within the same unified representation.

Summary updated

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 $10^3$ to $10^8$. 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.
BibTeX
@misc{sharpe2026aerozoo,
  author = {Peter Sharpe and Carmelo Gonzales and Corey Adams and Mehdi Ataei and Sanjay Choudhry},
  title = {{AeroZoo: A Benchmark of Well-Posed Flow Problems for Foundation Models of Aerodynamics}},
  year = {2026},
  eprint = {2611.17152},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17152},
}
Figure