TurtLES: A Large-Scale Benchmark for Turbulent 3D Neural PDE Surrogates

NeurIPS 2026 · Evaluations and Datasets Track

TurtLES is a high-fidelity computational fluid dynamics dataset, containing 400 trajectories of three-dimensional turbulent cylinder wakes in the shear-layer transition regime. The cases differ in cross-section shape, Reynolds number and angle of attack.
TurtLES provides a new benchmark for neural PDE surrogates, combining (i) physically realistic turbulence with an active energy cascade, (ii) dense irregular meshes and (iii) long-horizon dynamics.

---

Dataset details

Each case is a 3D turbulent flow past a cylinder with a super-elliptical cross-section, i.e. all points (x, y) satisfying |x/a|ⁿ + |y/b|ⁿ = 1, with the thickness fixed at D = 2b = 1. Four parameters are varied:

Parameter	Range	Sampling
Shape exponent n	1, 2, 5, ∞ (diamond, ellipse, rounded rectangle, rectangle)	discrete, 100 cases each
Half-length a/D	1/6 – 3/2	log-uniform
Reynolds number Re	1000 – 3900	uniform
Angle of attack α	0° – 45°	uniform

The cases are distributed into 320 train, 40 validation and 40 test trajectories, where each split follows the same parameter distribution as the full dataset. The full case listing and splits are available in cases.csv and the dataset viewer. A separate reference case (Re1000_alpha0_a0p5_n2, a circular cylinder) is not part of any split.

For a visual overview of the parameter space and further details on how the data were generated and validated, see the paper.

---

Case structure

turtles/
├── data/
│   ├── train/                                          320 cases
│   ├── val/                                            40 cases
│   ├── test/                                           40 cases
│   │   └── Re1089_alpha12_a1p302_n2/
│   │       ├── Re1089_alpha12_a1p302_n2_b0_99.h5       frames 0–99
│   │       ├── Re1089_alpha12_a1p302_n2_b100_199.h5    frames 100–199
│   │       ├── Re1089_alpha12_a1p302_n2_b200_299.h5    frames 200–299
│   │       ├── Re1089_alpha12_a1p302_n2_b300_399.h5    frames 300–399
│   │       ├── Re1089_alpha12_a1p302_n2_coords.h5      x, y, z of every point
│   │       ├── Re1089_alpha12_a1p302_n2_template.vtu   mesh template 
│   │       └── Re1089_alpha12_a1p302_n2_spectra.npz    velocity spectra 
│   └── reference/
│       └── Re1000_alpha0_a0p5_n2/                      reference case (circular cylinder)
├── catalogue/                                          case catalogue shown in the dataset viewer
├── cases.csv                                           all cases with their parameters and split
├── download_turtles.py                                 download script (see Download instructions)
├── h5_to_vtu.py                                        transforms .h5 to .vtu data (see below)
└── compute_spectra.py                                  computes the velocity spectra (see below)

Case names encode the parameters: Re1089_alpha12_a1p302_n2 is Re = 1089, α = 12°, a/D = 1.302 and n = 2 (p stands for the decimal point, ninf for n = ∞).
Inside each .h5 file, every snapshot is a group named sol_t<timestep> (e.g. sol_t346 in _b300_399.h5) that contains the fields u, v, w, p and rho. Every field is a one-dimensional array in which the i-th entry is the field value at the i-th point in _coords.h5, which holds the x, y and z coordinates.

Two optional files can be downloaded with each case:

• _template.vtu: the case's mesh without solution data. h5_to_vtu.py can be used to fill this template mesh with chosen ground truth/prediction data for visualisation in ParaView, VisIt or PyVista, or for post-processing computations that require the mesh connectivity, such as gradients, vorticity or interpolation to other grids.
• _spectra.npz: power spectra of u and v at three probes behind the body. Details of how the spectra are computed can be found in the paper or in compute_spectra.py, which can also compute the same spectra from your own data. Note that a trajectory covers only a few vortex-shedding periods, so these are finite-window spectra, not converged power spectral densities.

---

mini-TurtLES

Standardised subsets that reduce the dataset by geometry (G4–G1, fewer shape families) and time (T100–T400, first 100–400 snapshots). They are named TurtLES-G{1-4}T{100,200,300,400}, and TurtLES-full = TurtLES-G4T400. The tiers are explained in the appendix of the paper.

Tier	Families n	T100	T200	T300	T400
G4	1, 2, 5, ∞	3.5 TB	7.0 TB	10.5 TB	14 TB
G3	1, 2, ∞	2.6 TB	5.1 TB	7.7 TB	10.2 TB
G2	2, ∞	1.8 TB	3.6 TB	5.3 TB	7.1 TB
G1	2	0.8 TB	1.6 TB	2.4 TB	3.3 TB

TurtLES-debug (about 5 GB) holds the first 100 snapshots of the reference case. Download a subset with, for example, python download_turtles.py TurtLES-G1T100, and please report the subset name with your results.

---

Download instructions

We provide download_turtles.py, which can be used to download the full dataset, any mini-TurtLES subset, or single cases.

The downloaded data keeps the repository layout (data/<split>/<case_id>/). Files that are already present are skipped. Moving up a tier (e.g. from TurtLES-G1T100 to TurtLES-G1T200) therefore only fetches the difference. A download interrupted with Ctrl+C or a lost connection continues where it stopped when you run the same command again. Unfinished files never appear in data/.

Installation

pip install -U huggingface_hub
hf auth login    # Downloads also work without logging in, but logged-in downloads get higher rate limits.
hf download turtlesprojectml/turtles download_turtles.py --repo-type dataset --local-dir .           

Usage

python download_turtles.py --list                       # lists all subsets with their exact sizes
python download_turtles.py TurtLES-debug                # downloads 100 snapshots of the reference case
python download_turtles.py TurtLES-G1T100               # downloads the mini-TurtLES subset G1T100
python download_turtles.py TurtLES-full                 # downloads the full dataset (14 TB)
python download_turtles.py Re1089_alpha12_a1p302_n2     # downloads a single case (all 400 frames)

Option	Effect
--dir PATH	Specify a PATH to download to (default: current directory)
--split test	Download only the specified splits, e.g. --split test or --split train,val
--with-vtu	Also download each case's *_template.vtu
--with-spectra	Also download each case's *_spectra.npz
--yes	Start without asking for confirmation (for batch jobs)

---

Code and baselines

Our GitHub repository includes baseline implementations and evaluation scripts to facilitate training of neural surrogates on TurtLES.

---

Citation

If you find our work useful or use our dataset, please consider citing it:

@inproceedings{platenburg2026turtles,
  title     = {{TurtLES}: A Large-Scale Benchmark for Turbulent {3D} Neural {PDE} Surrogates},
  author    = {Platenburg, Jarno and Kassa{\"\i} Koupa{\"\i}, Armand and Cinnella, Paola and Gallinari, Patrick},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://arxiv.org/abs/2611.09567}
}