AeroZoo v1.2

Most 3D aerodynamics datasets vary a single parametric shape family inside one fixed wind tunnel, allowing surrogate models to memorize coordinate-to-field shortcuts instead of solving boundary-value problems or generalizing out of distribution. AeroZoo poses 3,816 distinct 3D incompressible flow problems across Reynolds numbers from ${10}^{3}10^3$ to ${10}^{8}10^8$, solved with unsteady RANS (Langtry–Menter transition SST in OpenFOAM 14) and time-averaged over a fixed convective window.

One geometry, four boundary-value problems: placing the same car at the same reference speed and Reynolds number into four enclosures shifts axial force from C_x = 0.025–0.029 to 0.554 and vertical force from near zero to C_z = +1.066.

Every sample is stored in dimensional SI units as a PhysicsNeMo DomainMesh (a memory-mapped PyTorch TensorDict tree) that bundles the posed boundary-value problem with its time-mean solution across:

• Seven geometry families: Blobs, Blender bluff, Procedural cars, Zero-to-CAD, Airframe Foundry, Streamlined, and Blender streamlined (plus one mascot sample outside every split).
• Eight enclosure configurations across four enclosure families: rectangular boxes with five boundary layouts (open, ground-effect, confined, crossflow, and dual-inlet), opposed inlet/outlet spheres, symmetry hemispheres, and internal ducts (with or without an internal body).
• Three body-wall layouts: all no-slip, two contiguous slip/no-slip bands, or three bands.

v1.2 is the official release, exported 2026-10-08 from a population frozen on 2026-10-02. Authors: Peter Sharpe*, Carmelo Gonzales*, Corey Adams, Mehdi Ataei, and Sanjay Choudhry (NVIDIA; *lead authors).

Links

• This dataset: https://huggingface.co/datasets/nvidia/AeroZoo
• PhysicsNeMo (DomainMesh): https://github.com/NVIDIA/physicsnemo
• The AeroZoo paper: https://arxiv.org/abs/2611.17152

Visual Overview

Every body stores time-averaged pressure on its remeshed boundary triangles (no_slip and slip) and wall shear stress on no_slip, spanning high-genus organic blobs, bluff shapes, procedural cars, CAD parts, aircraft, streamlined bodies, wings, and the mascot:

At matched sample counts, AeroZoo's seven geometry families span 25 to 390 effectively distinct shapes out of 400 under the Vendi diversity score (compared with 13 for DrivAerML and 11 for SHIFT-SUV at 400 shapes, and 7 out of 126 for HiLiftAeroML). Across five decades of Reynolds number, boundary layers transition from fully laminar near $Re\approx {10}^{3}Re\; \backslash approx\; 10^3$ (tracking Blasius flat-plate friction) through transitional turbulent wedges near $Re\approx {10}^{5}Re\; \backslash approx\; 10^5$ to fully turbulent at $Re\ge {10}^{6}Re\; \backslash ge\; 10^6$ (tracking Prandtl turbulent friction):

Alongside external enclosures, 303 released internal-duct samples span 10 passage geometries (161 body-free ducts and 142 ducts containing an internal body):

Contents

Quantity	Value
Released samples	3,816 (3,815 in splits + one mascot outside every split)
Stability labels	2,817 mean_stable / 999 mean_unstable (canonical benchmark uses mean_stable)
Interior query points	23.8 billion total; median 5.0M cells per sample (range 0.6M–24M)
Total storage	1.48 TB in uncompressed tars; median 347 MB per sample
Benchmark splits and tasks	22 frozen train / validation / test splits; two tasks (interior and boundary)

README.md                         this card
LICENSE                           the dataset license
NOTICE                            third-party credits and terms: Zero-to-CAD 100K
                                  (Apache License 2.0) and UIUC Airfoil Coordinates Database
zero-to-cad-provenance.csv        one row per Zero-to-CAD body in a sample: the
                                  part, its source revision, and NVIDIA's changes to it
assets/                           dataset overview and sample-anatomy figures
index.json                        one record per released sample: identity, hashes, stability label,
                                  quality diagnostics, and export lineage
release-v1.2.json                 release manifest: status per planned case (complete / failed /
                                  timed_out), planned and released sizes per split part, and SHA-256
                                  digests of index.json and the split files
splits/cases.json                 planned population: geometry family (`bank`), enclosure family,
                                  box layout (`profile`), body-wall layout, Reynolds decade, source
                                  geometry hash, and split placement per case
splits/splits.json                22 named splits as "<split>_<part>" -> [case_id, ...]
splits/body-free-ducts.json       the 219 internal-duct case IDs posed without an internal body
splits/splits-metadata.json       placement scheme (geometry hash) and design seed (note: `full` is
                                  unstratified despite its description string here)
checksums.sha256                  SHA-256 of every sample tar
SHA256SUMS                        SHA-256 of every data file of the unpacked release (see Verifying a copy)
samples/shard-NNN/<case_id>.tar   one uncompressed tar per sample (shards are administrative only)

Each sample tar unpacks to one directory:

<case_id>/
  domain.pdmsh/                   PhysicsNeMo DomainMesh (memory-mapped TensorDict tree, schema 5)
  domain.pdmsh.validated.json     reload-validation receipt of the artifact
  reexport-lineage.json           validated solver attempt and exporter provenance
  smoke-audit.json                structural audit of the exported artifact
  case-manifest.json              the posed boundary-value problem
  status.json                     execution outcome and attempt history
  result.yaml                     solver summary: termination, cell count, and stability

Loading a sample

Install PhysicsNeMo (>=2.1.0 for DomainMesh, Python >=3.11) and the Hugging Face client:

pip install "nvidia-physicsnemo>=2.1.0" "huggingface_hub>=1.7.1"

Load one sample lazily via memory mapping (tensors are read on first access):

import json, tarfile
from pathlib import Path
from huggingface_hub import hf_hub_download
from physicsnemo.mesh import DomainMesh

repo = "nvidia/AeroZoo"
index = json.loads(Path(hf_hub_download(repo, "index.json", repo_type="dataset")).read_text())
case_id, record = next(iter(index["samples"].items()))
tar_path = hf_hub_download(repo, f"samples/shard-{record['shard']:03d}/{case_id}.tar", repo_type="dataset")
with tarfile.open(tar_path) as tar:
    tar.extractall(Path("aerozoo"), filter="data")
sample = DomainMesh.load(Path("aerozoo") / case_id / "domain.pdmsh")

interior = sample.interior                                    # one point per finite-volume cell centroid
x = interior.points                                           # [N, 3] meters, query coordinates
u_mean = interior.point_data["solution", "velocity"]          # [N, 3] m/s, a target
volume = interior.point_data["_target_quadrature_measure"]    # [N] m^3, loss and metric weight
inlet = sample.boundaries["inlet"]                            # triangle mesh; prescribed/* are inputs
u_in = inlet.cell_data["prescribed", "dirichlet", "velocity"] # [F, 3] m/s
nu = sample.global_data["physics", "kinematic_viscosity"]     # () m^2/s, an input
stable = sample.global_data["diagnostics", "solve", "mean_window", "stability_pass"]  # () bool; True iff mean_stable

To download a specific split in parallel without duplicating tars in the global cache (for example super_scarce_train, 53 canonical samples, ~18 GB):

import json
from pathlib import Path
from huggingface_hub import hf_hub_download, snapshot_download

repo = "nvidia/AeroZoo"
index = json.loads(Path(hf_hub_download(repo, "index.json", repo_type="dataset")).read_text())["samples"]
splits = json.loads(Path(hf_hub_download(repo, "splits/splits.json", repo_type="dataset")).read_text())
patterns = [
    f"samples/shard-{index[c]['shard']:03d}/{c}.tar"
    for c in splits["super_scarce_train"]
    if c in index and index[c]["quality"]["solution_quality"] == "mean_stable"
]
snapshot_download(repo, repo_type="dataset", allow_patterns=patterns, local_dir="aerozoo-raw", max_workers=8)

All physical quantities use dimensional SI units. Nested TensorDict keys are indexed in Python with string tuples (for example cell_data["prescribed", "dirichlet", "velocity"]). A sample contains three top-level groups (example dimensions shown for a 4.55M-cell aircraft sample):

DomainMesh
|-- interior  (Mesh: point cloud of cell centroids, no interior connectivity)
|   |-- points                                              [4550479, 3] float32   m
|   |-- cells                                               [0, 1] int64
|   `-- point_data
|       |-- solution
|       |   |-- velocity                                    [4550479, 3] float32   m/s
|       |   |-- gauge_pressure                              [4550479] float32      Pa
|       |   |-- turbulent_kinetic_energy                    [4550479] float32      m^2/s^2
|       |   |-- specific_dissipation_rate                   [4550479] float32      1/s
|       |   |-- turbulent_kinematic_viscosity               [4550479] float32      m^2/s
|       |   |-- intermittency                               [4550479] float32      1
|       |   `-- momentum_thickness_reynolds_number          [4550479] float32      1
|       `-- _target_quadrature_measure                      [4550479] float32      m^3 (cell volume)
|-- boundaries
|   |-- inlet  (Mesh: triangle faces)
|   |   |-- points                                          [12800, 3] float32     m
|   |   |-- cells                                           [25122, 3] int64
|   |   `-- cell_data
|   |       |-- solution/*                                  [25122, ...] float32   (seven fields, as interior)
|   |       `-- prescribed/dirichlet/{velocity, turbulent_kinetic_energy,
|   |             specific_dissipation_rate, intermittency, momentum_thickness_reynolds_number}
|   |-- outlet  (Mesh: triangle faces)
|   |   |-- points / cells                                  [17289, 3] / [34080, 3]
|   |   `-- cell_data
|   |       |-- solution/*                                  [34080, ...] float32   (seven fields, as interior)
|   |       |-- prescribed/dirichlet/gauge_pressure         [34080] float32        Pa (zero)
|   |       `-- prescribed/reverse_flow/*                   [34080, ...] float32   (five inlet fields)
|   |-- no_slip  (Mesh: body no-slip bands and enclosure walls)
|   |   |-- points / cells                                  [55553, 3] / [110356, 3]
|   |   `-- cell_data
|   |       |-- solution/*                                  [110356, ...] float32  (seven fields, as interior)
|   |       |-- prescribed/dirichlet/velocity               [110356, 3] float32    m/s (zero)
|   |       |-- annotation/body_id                          [110356] int64         (0 enclosure, 1 primary, 2 secondary)
|   |       `-- derived/{wall_shear_stress [F, 3] Pa, wall_y_plus [F]}
|   `-- slip  (Mesh: body slip bands, slip walls, and symmetry planes)
|       |-- points / cells                                  [68480, 3] / [135042, 3]
|       `-- cell_data
|           |-- solution/*                                  [135042, ...] float32  (seven fields, as interior)
|           `-- annotation/body_id                          [135042] int64
`-- global_data
    |-- schema_version                                      () int64
    |-- physics/{density () kg/m^3, kinematic_viscosity () m^2/s, body_acceleration [3] m/s^2, closure_model_id ()}
    |-- reference/{velocity [3], secondary_inlet_velocities [M, 3], force_projection_direction [3],
    |     speed (), length (), dynamic_pressure ()}
    `-- identity/*, solution_definition/*, diagnostics/*    (provenance, averaging window, and stability metrics)

A sample stores more tensors than a benchmarked model is permitted to take as input; see Tasks for the input contract.

Tasks

AeroZoo supports general physics-AI training and evaluation, with two benchmark tasks that map a posed boundary-value problem to its time-averaged solution.

Inputs, Targets, and Withheld Keys

Only data that define the governing boundary-value problem enter a model:

Role	Interior task	Boundary task
Allowed inputs	The four boundary triangle meshes (points and cells, whose mesh names identify each condition), cell_data["prescribed", ...], global_data["physics", "kinematic_viscosity"], density (1.225 kg/m³ everywhere, to work in pascals), and body_acceleration (zero)	Same inputs (no interior points)
Query coordinates	interior.points (cell centroids)	Boundary face centroids, computed from points and cells
Predicted targets	interior.point_data["solution", ...] (the seven mean fields; predicting velocity and pressure alone is also valid when stated)	cell_data["solution", ...] on boundaries and derived/wall_shear_stress on no_slip, excluding Dirichlet-fixed fields (inlet velocity and turbulence, outlet pressure, no_slip velocity) and slip velocity on symmetry hemispheres
Loss and metric weights (withheld)	interior.point_data["_target_quadrature_measure"] (cell volume)	Boundary triangle face areas
Withheld from model inputs	solution/*, derived/*, _target_quadrature_measure, annotation/body_id, and all of global_data except kinematic_viscosity, density, and body_acceleration	Same withheld keys, plus interior

Why reference scales and body annotations are withheld. Keys under global_data["reference", ...] (reference length, speed, dynamic pressure, inlet vectors, and force direction) define scoring units rather than the PDE. Renaming which length is called the reference leaves the physical flow unchanged, so feeding reference/length or a Reynolds number built on it leaks convention into the inputs. Likewise, the PDE applies the same wall boundary condition regardless of annotation/body_id (0 for enclosure walls, 1 or 2 for bodies); only losses and post-processing metrics use body_id. Nor does every problem have a single free-stream vector (for example a 90° duct bend or a dual-inlet box). Models that normalize coordinates or condition on a length, speed, or Reynolds scale derive it from one sample's boundary inputs and pass that derived scale (or Reynolds number) alongside rescaled coordinates.

Benchmark Rules

• Canonical subset (mean_stable): Train, validate, and test on mean_stable samples only. In the remaining 999 mean_unstable samples, averages over the two halves of the window differed beyond the stability tolerances (30% of crossflow boxes and 34% of symmetry hemispheres pass, and predicting the inlet velocity everywhere yields median ${\epsilon }_U=0.110\backslash varepsilon\_U\; =\; 0.110$ on full's canonical test set versus 0.184 across all released test samples). Scores over all released samples, or models trained with mean_unstable samples included, can be reported beside the canonical result when labeled.
• Body-only test scoring on body-property splits: The seven geometry-family splits (bank_*) and layout place body-free ducts according to a body or wall layout drawn from the case seed and never posed. Canonical test sets for those eight splits exclude body-free ducts (12 of 322 in bank_cars, 38 of 930 in layout). Training and validation parts, and all other splits, retain them.
• Single-sample input functions only: Any deterministic function of one sample's allowed inputs (such as a boundary-derived length, speed, or Reynolds number) is permitted. Normalization constants or other statistics fitted across test samples are prohibited.
• Validation for selection only: Select models and hyperparameters on the validation part, then fit scored weights on the training part alone without refitting on train + val (super_scarce has 53 canonical training samples and 290 validation samples). Hyperparameters chosen on full's validation part can be reused across other splits when stated in the report (every validation case of one split belongs to the test set of another).

Metrics and Reporting

Metrics are evaluated per test sample in that case's reference units so flows spanning a 200-fold range in inlet speed contribute on a common dimensionless scale:

Metric	Task	Definition
${\epsilon }_U,{\epsilon }_p\backslash varepsilon\_U,\backslash ;\backslash varepsilon\_p$	interior	Headline metrics: volume-weighted RMS errors of mean velocity and gauge pressure over the fluid, in units of $U_{\mathrm{r}\mathrm{e}\mathrm{f}}U\_\{\backslash mathrm\{ref\}\}$ and $q_{\mathrm{r}\mathrm{e}\mathrm{f}}q\_\{\backslash mathrm\{ref\}\}$
${\epsilon }_U^{\mathcal{N}},{\epsilon }_p^{\mathcal{N}}\backslash varepsilon\_U^\{\backslash mathcal\{N\}\},\backslash ;\backslash varepsilon\_p^\{\backslash mathcal\{N\}\}$	interior	Same volume-weighted RMS errors within one $L_{\mathrm{r}\mathrm{e}\mathrm{f}}L\_\{\backslash mathrm\{ref\}\}$ of a body face centroid (samples with a body)
$r_U,r_{p}r\_U,\backslash ;r\_p$	interior	Unweighted relative $L_{2}L\_2$ errors $\parallel \widehat{\mathbf{y}}-\mathbf{y}{\parallel }_2\mathrm{/}\parallel \mathbf{y}{\parallel }_2\backslash lVert\backslash hat\{\backslash mathbf\{y\}\}\; -\; \backslash mathbf\{y\}\backslash rVert\_2\; /\; \backslash lVert\backslash mathbf\{y\}\backslash rVert\_2$ across interior points, for comparison with unweighted ML baselines
${\epsilon }_k,{\epsilon }_{\omega },{\epsilon }_{{\nu }_t},{\epsilon }_{\gamma },{\epsilon }_{{\widetilde{\mathrm{R}\mathrm{e}}}_{\theta t}}\backslash varepsilon\_k,\backslash ;\backslash varepsilon\_\backslash omega,\backslash ;\backslash varepsilon\_\{\backslash nu\_t\},\backslash ;\backslash varepsilon\_\backslash gamma,\backslash ;\backslash varepsilon\_\{\backslash widetilde\{\backslash mathrm\{Re\}\}\_\{\backslash theta\; t\}\}$	interior	Auxiliary: volume-weighted RMS errors of the five turbulence fields, normalized by $U_{\mathrm{r}\mathrm{e}\mathrm{f}}^{2}U\_\{\backslash mathrm\{ref\}\}^2$, $U_{\mathrm{r}\mathrm{e}\mathrm{f}}\mathrm{/}{L}_{\mathrm{r}\mathrm{e}\mathrm{f}}U\_\{\backslash mathrm\{ref\}\}/L\_\{\backslash mathrm\{ref\}\}$, $U_{\mathrm{r}\mathrm{e}\mathrm{f}}{L}_{\mathrm{r}\mathrm{e}\mathrm{f}}U\_\{\backslash mathrm\{ref\}\}L\_\{\backslash mathrm\{ref\}\}$ for $k,\omega ,{\nu }_{t}k,\; \backslash omega,\; \backslash nu\_t$, and by the target field's own volume-weighted RMS for $\gamma ,{\widetilde{\mathrm{R}\mathrm{e}}}_{\theta t}\backslash gamma,\; \backslash widetilde\{\backslash mathrm\{Re\}\}\_\{\backslash theta\; t\}$
${\epsilon }_{C_F}\backslash varepsilon\_\{C\_F\}$	boundary	Error of the body force coefficient along the primary inlet direction ${\mathbf{e}}_F\backslash mathbf\{e\}\_F$, divided by the larger of $C_{F}C\_F$'s magnitude and 0.05
${\epsilon }_{\mathbf{C}}\backslash varepsilon\_\{\backslash mathbf\{C\}\}$	boundary	Norm of the error of the full body force coefficient vector $\mathbf{C}\backslash mathbf\{C\}$, divided by $\max (\parallel \mathbf{C}{\parallel }_2,0.05)\backslash max(\backslash lVert\backslash mathbf\{C\}\backslash rVert\_2,\backslash ,0.05)$
${\epsilon }_{C_p},{\epsilon }_{\tau }\backslash varepsilon\_\{C\_p\},\backslash ;\backslash varepsilon\_\backslash tau$	boundary	Area-weighted RMS errors of pressure on all body faces (\(\mathcal{B}\)) and wall shear stress on no-slip body faces (\(\mathcal{W}\)), in units of $q_{\mathrm{r}\mathrm{e}\mathrm{f}}q\_\{\backslash mathrm\{ref\}\}$
${\epsilon }_K\backslash varepsilon\_K$	boundary	Error of an internal duct's mass-flow-weighted total-pressure loss coefficient $KK$, divided by $\max (K,0.05)\backslash max(K,\backslash ,0.05)$ (a non-finite value or predicted outlet mass flux at most 1% of inlet mass flux counts as 1)

Specifically, velocity error, body force coefficient, and axial force error are defined as

where $V_{i}V\_i$ is cell $ii$'s volume, $\mathcal{B}\backslash mathcal\{B\}$ is the set of body faces (annotation/body_id >= 1 on no_slip and slip), $\mathcal{W}\backslash mathcal\{W\}$ is the no-slip subset of $\mathcal{B}\backslash mathcal\{B\}$, $A_{j}A\_j$ and ${\mathbf{n}}_j\backslash mathbf\{n\}\_j$ are face $jj$'s area and outward normal (pointing out of the fluid), and ${\mathbf{e}}_F\backslash mathbf\{e\}\_F$ is reference/force_projection_direction. Note that $\mathbf{C}\backslash mathbf\{C\}$ uses the reference area $L_{\mathrm{r}\mathrm{e}\mathrm{f}}^{2}L\_\{\backslash mathrm\{ref\}\}^2$ rather than frontal area.

Free-stream reference (\(\varepsilon_U^{\mathrm{fs}}\)). Predicting the area-weighted mean prescribed inlet velocity everywhere (with zero gauge pressure and zero wall shear) requires no training and achieves a median ${\epsilon }_U\backslash varepsilon\_U$ of 0.110 on full's canonical test set, ranging from 0.03 in opposed spheres to 1.5 in dual-inlet boxes. Because a split's raw median depends strongly on its enclosure mix, every interior-task evaluation also reports three sample-paired comparisons using each test sample's own ${\epsilon }_U^{\mathrm{f}\mathrm{s}}\backslash varepsilon\_U^\{\backslash mathrm\{fs\}\}$:

• the median of ${\epsilon }_U\mathrm{/}{\epsilon }_U^{\mathrm{f}\mathrm{s}}\backslash varepsilon\_U\; /\; \backslash varepsilon\_U^\{\backslash mathrm\{fs\}\}$ across test samples,
• the fraction of test samples with , and
• the median ${\epsilon }_U\backslash varepsilon\_U$ within each of the eight enclosure configurations (profile for rectangular boxes and family otherwise in splits/cases.json).

Reporting checklist. For every benchmark result, state:

1. Task (interior or boundary), split name, release version (v1.2), and index.json SHA-256 (release-v1.2.json records it).
2. Whether evaluation and training used the canonical mean_stable subset, and whether hyperparameters were reused from full's validation part.
3. Number of random seeds and exact train and test sample counts.
4. Mean and median of each applicable metric across defined test samples (plus shares within 0.02 and 0.05 for ${\epsilon }_{C_F}\backslash varepsilon\_\{C\_F\}$ and ${\epsilon }_{\mathbf{C}}\backslash varepsilon\_\{\backslash mathbf\{C\}\}$), noting whether the five auxiliary turbulence fields were predicted.
5. For interior results, the three free-stream comparisons above.

Splits

The 22 frozen splits in splits/splits.json map "<split>_<part>" to lists of case_ids. Split placement is a deterministic hash of each case's primary geometry, so train, validation, and test share no primary body, and adding cases in a future release moves no existing case. (The second body that 3% of cases include is a scaled copy of that case's own primary body, so no geometry is shared across train, validation, and test.) Validation sets are drawn from the training distribution rather than held-out regimes. Of 4,360 planned split cases in release-v1.2.json, 3,815 completed (2,816 mean_stable), 208 failed mesh or solver checks, and 337 timed out when generation stopped. Uncompleted cases keep their split assignments and are never replaced.

Canonical (mean_stable) / all-released sample counts per split (canonical test counts for the seven bank_* splits and layout exclude body-free ducts):

Split	Test regime	Train (canonical / all)	Val (canonical / all)	Test (canonical / all)
Random
full	Unstratified 70 / 10 / 20 baseline split	2,011 / 2,713	290 / 386	515 / 716
scarce	Data efficiency: one-sixth of full train	321 / 440	290 / 386	515 / 716
super_scarce	Data efficiency: one-thirty-sixth of full train	53 / 74	290 / 386	515 / 716
Held-out enclosure
profile_open_opposed	Open box	1,943 / 2,775	293 / 405	580 / 635
profile_ground_effect	Ground-effect box	2,011 / 2,837	294 / 400	511 / 578
profile_confined_opposed	Confined box	2,114 / 2,963	301 / 418	401 / 434
profile_crossflow	Crossflow box	2,298 / 2,820	345 / 417	173 / 578
profile_dual_inlet	Dual-inlet box	2,044 / 2,875	313 / 424	459 / 516
family_opposed_sphere	Opposed sphere	2,158 / 3,017	302 / 418	356 / 380
family_opposed_hemisphere	Symmetry hemisphere	2,330 / 2,982	352 / 442	134 / 391
family_duct	Internal duct	2,266 / 3,055	348 / 457	202 / 303
family_non_box	Spheres, hemispheres, and ducts (train on boxes)	1,850 / 2,390	274 / 351	692 / 1,074
Held-out geometry family
bank_blobs	Blobs	1,662 / 2,156	237 / 292	872 / 1,367
bank_blender_bluff	Blender bluff	2,149 / 2,935	331 / 441	326 / 439
bank_cars	Procedural cars	2,173 / 2,965	321 / 434	310 / 416
bank_zero_to_cad	Zero-to-CAD	2,250 / 3,040	332 / 442	224 / 333
bank_airframes	Airframe Foundry	2,198 / 3,018	321 / 430	288 / 367
bank_streamlined	Streamlined	2,132 / 2,940	332 / 441	337 / 434
bank_blender_streamlined	Blender streamlined	2,148 / 2,938	310 / 418	346 / 459
Reynolds range and body-wall layout
reynolds	Extrapolation: train , test $Re\ge {10}^{6}Re\; \backslash ge\; 10^6$	1,577 / 2,119	238 / 311	1,001 / 1,385
reynolds_interior	Interpolation: decade held out	1,935 / 2,614	292 / 380	589 / 821
layout	Three-band bodies (train on all-no-slip and two-band)	1,647 / 2,233	239 / 311	892 / 1,271

full, scarce, and super_scarce share identical validation and test sets. Every other split holds out one regime as its test set, so test-set sizes and difficulty vary across splits; compare models on the same split, starting with full and using the held-out splits to measure extrapolation.

Resolve any canonical split part directly from the top-level metadata files before downloading sample tars:

import json
from pathlib import Path
from huggingface_hub import hf_hub_download

repo = "nvidia/AeroZoo"
fetch = lambda name: json.loads(Path(hf_hub_download(repo, name, repo_type="dataset")).read_text())
splits = fetch("splits/splits.json")
release = fetch("release-v1.2.json")
body_free = set(fetch("splits/body-free-ducts.json")["case_ids"])
quality = {c: r.get("solution_quality") for c, r in release["cases"].items()}

def canonical(part: str) -> list[str]:
    split, kind = part.rsplit("_", 1)
    chosen = [c for c in splits[part] if quality[c] == "mean_stable"]
    if kind == "test" and (split.startswith("bank_") or split == "layout"):
        chosen = [c for c in chosen if c not in body_free]
    return chosen

train_ids, test_ids = canonical("bank_cars_train"), canonical("bank_cars_test")

An unpacked sample records the same body-free indicator locally: geometry.body_count == 0 in case-manifest.json, and zero faces with annotation/body_id >= 1 on no_slip and slip.

What to know before training

• Time-averaged URANS targets: Stored fields are time means over an 80-convective-time window (following 15 convective times of spin-up) of unsteady RANS with the Langtry–Menter transition SST closure in OpenFOAM 14. Internal ducts that did not settle in the first window ran one fresh 80-convective-time window. mean_stable marks samples whose two 40-convective-time half-window averages agree within 0.02 of $U_{\mathrm{r}\mathrm{e}\mathrm{f}}U\_\{\backslash mathrm\{ref\}\}$ in velocity, 0.05 of $q_{\mathrm{r}\mathrm{e}\mathrm{f}}q\_\{\backslash mathrm\{ref\}\}$ in pressure, and 0.02 times $\max (\mid C_F\mid ,0.05)\backslash max(\backslash lvert\; C\_F\backslash rvert,\backslash ,0.05)$ in axial force coefficient.
• High-Reynolds attrition: Of the 337 planned cases still running when generation stopped, 230 lie in the ${10}^{7}10^7$ to ${10}^{8}10^8$ Reynolds decade (release-v1.2.json records every planned case's status).
• Sign and pressure conventions: derived/wall_shear_stress follows OpenFOAM's convention (traction exerted by the wall on the fluid, giving the minus sign in $\mathbf{C}\backslash mathbf\{C\}$). gauge_pressure is in pascals relative to the outlet (0 Pa), and boundary face normals point out of the fluid domain.
• Enclosed volumes: Compute fluid or body volumes from a sample's boundary triangle meshes (for example via the divergence theorem over boundaries/*). Those match the sum of cell volumes across all released samples, including three Zero-to-CAD cases (550, 2273, and 3538) whose upstream CAD surfaces self-intersected before exterior meshing.

Geometry sources

Six of the seven geometry families (plus the EasterEggs mascot meshes) were created in-house at NVIDIA. Two of the 11 airfoil sections used in Blender streamlined wings (Eppler E387 and SD7003-085-88) come from coordinate files in the UIUC Airfoil Coordinates Database (M. S. Selig, UIUC), and the remaining nine sections were evaluated from the analytic NACA equations. The Zero-to-CAD family is a farthest-point sample of 500 parts drawn across three bluffness strata from the 52,408 watertight, single-body parts of Autodesk AI Lab's Zero-to-CAD 100K dataset (ADSKAILab/Zero-To-CAD-100k); 320 of them appear in released samples, and the family's other 13 released samples are body-free ducts. That dataset is released under the Apache License 2.0, which applies to those parts. LICENSE and NOTICE at the top of this repository state the terms, attributions, and modifications, and zero-to-cad-provenance.csv maps each Zero-to-CAD sample body to its upstream part and transform.

License

GOVERNING DOWNLOAD TERMS: Use of the dataset is governed by the Creative Commons Attribution 4.0 International License (CC BY 4.0), and includes 500 parts selected from Autodesk AI Lab's Zero-to-CAD 100k dataset which is governed by Apache License, Version 2.0.

Origin of every input:

Input	Origin	Terms
Blobs, Blender bluff, Blender streamlined, Streamlined, Procedural cars, Airframe Foundry geometry families	generated in-house at NVIDIA	the dataset license
Two of the 11 airfoil sections of the Blender streamlined wings (Eppler E387, SD7003-085-88)	coordinate files from the UIUC Airfoil Coordinates Database (M. S. Selig, UIUC); the other nine were computed from the published NACA equations	the dataset license; the coordinate files are not included (credited in NOTICE)
Zero-to-CAD parts (500 selected; 320 appear in released samples)	Autodesk AI Lab, Zero-to-CAD 100K	Apache License 2.0; see NOTICE and zero-to-cad-provenance.csv
Two mascot meshes (unicorn, jackalope; the EasterEggs family, outside every split, one released sample)	modeled in-house at NVIDIA by Carmelo Gonzales	the dataset license
All flow solutions, meshes, the stability label and metadata	computed by NVIDIA	the dataset license

Verifying a copy

SHA256SUMS lists the SHA-256 of every data file of the release as stored on NVIDIA's clusters: the top-level files alongside each sample tar unpacked into its shard directory. It was added on 2026-10-16 with no change to any data file. To reconstruct that layout from a download and verify every file:

for t in samples/shard-*/*.tar; do
  s=$(basename "$(dirname "$t")"); mkdir -p "$s"; tar -xf "$t" -C "$s"
done
grep -v '  README.md$' SHA256SUMS | sha256sum --quiet -c -

The second command prints nothing when every listed file matches (it skips the cluster-tree README.md, and extra repository files such as LICENSE, NOTICE, zero-to-cad-provenance.csv, assets/, or splits/body-free-ducts.json are ignored by sha256sum -c). Running sha256sum SHA256SUMS prints a single digest covering the entire release: f63dc064ca4c8dc5e03d6668c4e60efdb197d4b6cd0d965ec3fe83a2ba77bb38.

Provenance

Generated with OpenFOAM 14 (mesh policy snappyhexmesh_external_v14 for every case, ducts included, and solver policy openfoam14-external-transition-sst-mean-co10-w80-duct-two-stage-v10, artifact schema 5) on NVIDIA's compute clusters (8.1 million CPU core-hours of solver time); population frozen 2026-10-02 and exported 2026-10-08. The index.json SHA-256 recorded in release-v1.2.json is c8e9508e8a9d8291….

Citation

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

Corresponding contact: Peter Sharpe (psharpe@nvidia.com).

NVIDIA dataset card

Dataset Description

AeroZoo is a dataset of computed fluid flows for physics-AI research. It holds 3,816 samples. Each sample is one solved flow problem: one body (or none, in some ducts) inside its own enclosure, at its own size, speed and Reynolds number. The Reynolds number ranges from 10³ to 10⁸, which covers laminar and turbulent flow. No two planned samples share the same primary geometry.

Each sample stores three things:

• the problem that was posed: the boundary surfaces, the conditions set on them, and the fluid's viscosity;
• the solution: seven time-averaged flow fields (velocity, pressure and five turbulence-model fields) at the center of every cell of the solver's mesh;
• the volume of every cell, used to weight a training loss.

All samples were computed by NVIDIA with the solver OpenFOAM 14. The solver ran unsteady Reynolds-averaged Navier–Stokes (URANS) simulations with the Langtry–Menter transition SST turbulence model. Each stored field is a time average over a fixed window, not an instantaneous snapshot.

The bodies come from seven geometry families: Blobs, Blender bluff, Blender streamlined, Streamlined, Procedural cars, Airframe Foundry and Zero-to-CAD. They sit in eight enclosure configurations: a rectangular box with five boundary layouts (open, ground-effect, confined, crossflow and dual-inlet), an opposed sphere, a symmetry hemisphere and an internal duct. Each body uses one of three body-wall layouts (all no-slip, two slip/no-slip bands, three bands).

The dataset is meant for training and testing machine-learning models that predict fluid flow ("surrogate models"). It ships with 22 fixed train / validation / test splits and two benchmark tasks with defined metrics.

This dataset is ready for commercial or non-commercial uses.

Dataset Owner(s)

NVIDIA Corporation.

Dataset Creation Date

2026-10-08 (v1.2 export). The simulations behind the released samples completed between 2026-09-13 and 2026-10-02, when generation was stopped and the population was frozen.

Versioning

v1.2 (2026-10-08 export).

Intended Usage

AeroZoo is intended for researchers and engineers who build machine-learning models of fluid flow. Intended uses are:

• training and evaluating surrogate models of steady or time-averaged incompressible flow;
• pretraining such models on many unrelated flows;
• measuring how well a model generalizes to unseen geometry families, enclosures, boundary layouts and Reynolds numbers, using the 22 frozen splits;
• comparing models on the two benchmark tasks. The interior task predicts the mean fields at every cell center. The boundary task predicts the mean fields on every boundary face and the wall shear stress on no-slip walls. The canonical benchmark uses only mean_stable samples.

What the dataset does not support:

• Forecasting a flow over time. Only time averages are stored.
• Compressible, heated, multiphase or non-Newtonian flow; moving or rotating walls; porous media.
• Use as measurements of real flows. Each sample solves one turbulence model's equations on one mesh. Each solution is converged at the recipe's placement of the first wall cell, not independently of it: on meshes 1.5 times finer per direction the force changes by a median of 3.7% (at most 19%), and a first cell at y⁺ ≈ 1 changes it by a median of 6.9% (at most 36%) and can move transition.

Dataset Characterization

Data Collection Method

• Synthetic

Every sample was generated by computer simulation; nothing was measured or collected from people. For each case, the generator drew a geometry, its size (2 cm to 20 m), inlet speed (0.25 to 50 m/s) and Reynolds number (10³ to 10⁸) from a seed and a frozen recipe. It then built a mesh with OpenFOAM's snappyHexMesh (mesh policy snappyhexmesh_external_v14). The solver ran OpenFOAM 14 with the Langtry–Menter transition SST model (solver policy openfoam14-external-transition-sst-mean-co10-w80-duct-two-stage-v10). Each solve ran 15 convective times to let the start-up wash out, then averaged every field over the next 80 convective times. An internal duct whose first window did not settle averaged one fresh 80-convective-time window; there was no further extension. (A convective time is the reference length, the body's largest extent, divided by the inlet speed.) The exporter converted each result to a PhysicsNeMo DomainMesh, reloaded it and checked it before release.

The posed problem of every case was fixed before any mesh existed. Cases that failed a mesh or solver check (208) or were still running when generation stopped (337) are listed in the release manifest and are not replaced. The runs used NVIDIA's compute clusters. The released samples took 8.1 million CPU core-hours of solver time.

Geometry origin: six of the seven geometry families were generated in-house at NVIDIA. The seventh, Zero-to-CAD, is a sample of 500 parts from Autodesk AI Lab's Zero-to-CAD 100K, chosen from the 52,408 parts that pass watertightness, single-body and shape checks. Two of the 11 airfoil sections of the Blender streamlined wings (Eppler E387 and SD7003-085-88) come from coordinate files of the UIUC Airfoil Coordinates Database; the other nine were computed from the published NACA equations. The two mascot meshes were modeled in-house at NVIDIA.

The dataset contains no personal data.

Labeling Method

• Automated

The dataset has one label: the stability label, mean_stable or mean_unstable. It says whether the time average had settled. The software compares the averages over the two halves of the 80-convective-time window, with three fixed tests: velocity change at most 0.02 of the inlet speed, pressure change at most 0.05 of the reference dynamic pressure, and a change in the body-force coefficient of at most 0.02 times its mean value (or times 0.05, if that is larger). A sample is mean_stable if it passes all three. The thresholds were fixed before the studies that tested them and were never adjusted. All three measured changes are stored, so users can apply another rule. No person labeled any sample.

Dataset Format

• Modality: 3D physics simulation data: point clouds (interior cell centers with field values) and triangle meshes (boundary surfaces), with scalar metadata. Units are SI throughout.
• Format: one uncompressed .tar file per sample. Each unpacks to a PhysicsNeMo DomainMesh (domain.pdmsh, a memory-mapped TensorDict tree; artifact schema 5) and small JSON / YAML files. Floating-point tensors are float32. Top-level catalog, split and checksum files are JSON and text.
• Structure of one sample (domain.pdmsh):
  • interior: one point per mesh cell, at the cell center. Each point holds 13 float32 values: 3 coordinates, the seven mean fields (9 components) and the cell volume. There is no interior connectivity.
  • boundaries: four triangle meshes, inlet, outlet, no_slip and slip, grouped by boundary condition. They carry the prescribed values, the mean fields on every face, and wall shear stress and y⁺ on no_slip.
  • global_data: fluid properties, reference scales, how the mean was defined, identity hashes and quality diagnostics, including the stability label.
• Files beside the samples: index.json (one record per released sample), release-v1.2.json (status of every planned case), splits/cases.json, splits/splits.json (22 named splits), splits/splits-metadata.json, checksums.sha256 and SHA256SUMS.

Dataset Quantification

• Record count: 3,816 samples: 3,815 of the 4,360 planned cases plus one mascot outside every split (release-v1.2.json). Of these, 2,817 are mean_stable and 999 mean_unstable.
• Feature count: per interior point, 13 values (3 coordinates, 9 mean-field components, 1 cell volume). Per sample, four boundary meshes and the global_data scalars (artifact schema 5).
• Size of the samples: 23.8 billion interior points in total; median 5.0 million cells per sample, range 0.6 to 24 million.
• Total data storage: 1.48 TB: 1,480,746,807,800 bytes of sample artifacts, median 347 MB per sample. The Hugging Face repository stores 1,482,788,671,963 bytes in all (Hub API, 2026-10-30).

Planned and released cases per geometry family:

Geometry family	Planned cases	Released samples
Blobs	1,600	1,367
Blender bluff	500	439
Blender streamlined	500	459
Streamlined	480	434
Procedural cars	480	416
Zero-to-CAD	400	333
Airframe Foundry	400	367
Total	4,360	3,815
EasterEggs mascots (outside every split)	2	1

Reference(s)

• AeroZoo on Hugging Face: https://huggingface.co/datasets/nvidia/AeroZoo
• NVIDIA PhysicsNeMo (the DomainMesh format): https://github.com/NVIDIA/physicsnemo
• OpenFOAM 14: https://openfoam.org/version/14/
• Zero-to-CAD 100K (source of the Zero-to-CAD geometry family): https://huggingface.co/datasets/ADSKAILab/Zero-To-CAD-100k
• Ataei, Askari, Rahimi Malekshan and Jayaraman, "Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data", 2026, https://arxiv.org/abs/2604.24479
• The AeroZoo paper: https://arxiv.org/abs/2611.17152

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.