On 12 November 2026, NVIDIA researchers Peter Sharpe, Carmelo Gonzales, Corey Adams, Mehdi Ataei, and Sanjay Choudhry released AeroZoo, an open benchmark and pretraining corpus evaluating whether neural surrogates for computational fluid dynamics (CFD) can generalize across diverse geometries, enclosure boundaries, and operating conditions.
The Limits of Single-Family CFD Datasets
Deep learning surrogates for fluid dynamics aim to predict complex velocity and pressure fields in seconds rather than the hours or days required by numerical solvers. However, most existing 3D CFD corpora—such as DrivAerML, AhmedML, WindsorML, and ShapeNet-Car—sweep morphs of a single design family inside a fixed virtual wind tunnel. While valuable for single-product engineering optimization, these corpora cannot test whether an aerodynamic model generalizes out-of-distribution to unfamiliar shapes or altered physical boundaries.
To demonstrate this limitation, the authors analyzed Transolver scaling on DrivAerML. Expanding model capacity from 20 million to 354 million parameters altered surface pressure test error by just 2.6%. Quadrupling training compute reduced error by approximately 4%, while doubling training samples yielded diminishing returns of roughly 12% per step before exhausting the 400 available cases. Single-family datasets hit an irreducible error floor governed by simulation averaging noise rather than neural network capacity, failing to serve as an adequate foundation model testbed.
A Benchmark of Well-Posed Boundary-Value Problems
AeroZoo addresses this bottleneck by framing fluid flow as a well-posed partial differential equation (PDE) problem. Rather than treating neural prediction as a pure shape-to-field regression, each sample provides the exact boundary meshes grouped by condition type (inlet, outlet, no-slip wall, and slip wall), face-wise prescribed values, density, and kinematic viscosity. The corpus comprises 3,815 time-mean solutions (23.8 billion computational cells, 1.48 TB) generated in OpenFOAM 14 using the unsteady Reynolds-averaged Navier–Stokes (URANS) equations coupled with the Langtry–Menter transition SST model.
The dataset spans seven distinct geometry families (including procedural cars, aircraft from Airframe Foundry, streamlined bodies, and CAD models sampled from the public Zero-to-CAD collection), eight enclosure types (such as open, confined, ground-effect, crossflow, and dual-inlet boxes, as well as internal ducts), physical scales from 2 centimeters to 20 meters, and Reynolds numbers spanning five orders of magnitude from 10³ to 10⁸. The canonical benchmark isolates 2,816 mean-stable simulations across 22 frozen, geometry-disjoint splits, evaluating models on interior flow fields (velocity, gauge pressure, and turbulence quantities) and boundary wall shear.
Surrogate Baselines and Generalization Findings
The researchers benchmarked standard and boundary-conditioned neural surrogates against a baseline free-stream reference that predicts the mean inlet state everywhere. On the canonical test split, providing boundary-condition tokens reduced GeoTransolver’s median relative velocity error from 76.9% down to 8.4% of the inlet velocity, outperforming the reference on 82% of test cases. ISLA, a coordinate-free transformer tailored for boundary-driven PDEs, reached a median velocity error of 5.3%, outperforming the free-stream reference on 98% of cases.
| Model | Parameters | Boundary Tokens | Median Velocity Error (εU) | Near-Field Error (εNU) | Cases Outperforming Reference |
|---|---|---|---|---|---|
| Free-Stream Reference | — | Yes | 11.0% | 16.0% | — |
| Transolver | 9.0M | No | 34.5% | 27.6% | 41% |
| GeoTransolver | 27.6M | No | 76.9% | 80.4% | 13% |
| GeoTransolver | 27.6M | Yes | 8.4% | 12.4% | 82% |
| ISLA | 67.5M | Yes | 5.3% | 9.1% | 98% |
Crucially, the paper presents a four-link error stack-up comparing surrogate inaccuracy against solver numerical sensitivities. Across stable simulations, solver averaging noise accounted for just 0.13% median velocity variation, cluster reruns varied by 0.0002%, and a 1.5× mesh refinement produced a 1.2% shift. In-distribution surrogate error (5.3% for ISLA) remains about four times larger than mesh refinement sensitivity.
When evaluated on held-out splits, surrogate error remained relatively stable when encountering unseen geometry families like blobs (6.2%) or CAD parts (10.9%). However, when faced with unseen enclosure geometries or altered boundary-condition layouts, model error surged 2.4-fold to 9.2-fold relative to uniform flow, reaching 65% median error on crossflow layouts and 134% on dual-inlet enclosures. While surrogates generalized across geometric variations, they struggled when physical flow constraints shifted.
Implications for Scientific Machine Learning
For scientific machine learning groups and university engineering departments developing physics foundation models, AeroZoo establishes a new standard for dataset curation and evaluation rigor. First, it demonstrates that neural surrogates cannot be evaluated purely as geometric encoders; boundary conditions and domain enclosures are active PDE inputs that dictate global flow topology. Second, the study highlights the necessity of systematic code verification and numerical uncertainty budgets, ensuring that surrogate benchmark targets are not confounded by meshing resolution or solver averaging noise.
The benchmark also models best practices in data governance by providing comprehensive datasheets for datasets that document provenance, failure cases, and explicit geometry-split boundaries. As research institutions construct domain-specific foundation models, moving beyond single-car or single-wing benchmarks toward diverse corpora will be essential for conducting reproducible AI experiments and replacing narrow synthetic benchmark evaluations with physically grounded challenges.
Sources
- Sharpe, P., Gonzales, C., Adams, C., Ataei, M., and Choudhry, S., ‘AeroZoo: A Benchmark of Well-Posed Flow Problems for Foundation Models of Aerodynamics’, arXiv:2611.17152v1 (2026-11-12): https://arxiv.org/abs/2611.17152
- arXiv HTML Full Text, ‘AeroZoo: Well-Posed Flow Problems for Foundation Models of Aerodynamics’ (2026-11-12): https://arxiv.org/html/2611.17152v1
- NVIDIA PhysicsNeMo Framework Repository (2026-11-12): https://github.com/NVIDIA/physicsnemo
- AeroZoo Dataset on Hugging Face (2026-11-12): https://huggingface.co/datasets/nvidia/AeroZoo
