Datasets:

ydai17
/

TIDE

license: cc-by-4.0
task_categories:
  - time-series-forecasting
  - image-to-image
tags:
  - turbulence
  - computational-fluid-dynamics
  - direct-numerical-simulation
  - navier-stokes
  - neural-operator
  - scientific-machine-learning
  - physics
arxiv: 2608.04222
pretty_name: 'TIDE: Turbulent Incompressible DNS Ensembles'
size_categories:
  - n>1T

TIDE: Turbulent Incompressible DNS Ensembles

A physically diverse 3D turbulence corpus: 15 configurations of the same incompressible Navier–Stokes system along eight physics axes, each shipping 8–16 fully independent realizations at 256³ in fp64 (134 trajectories, ~2.6 TB), released only after passing a fixed acceptance standard of statistical gates and equation-level residual checks.

Code (solver, acceptance referee, benchmark): https://github.com/Dyloong1/TIDE-dataset-benchmark
Citable DOI record (datasheet, code snapshot, split manifest): https://doi.org/10.5281/zenodo.21589489
Paper: arXiv:2608.04222 — TIDE: A Physically Diverse 3D Turbulence Benchmark Dataset for Advancing Scientific Machine Learning (under submission, ACM SIGKDD Datasets & Benchmarks Track)

All 15 configurations are live. Each ships as its own repository (linked in the table below). This index repository carries the datasheet, the split manifest, and the per-configuration manifests.

Structure

TIDE is one incompressible NS system: one solver, one grid, one acceptance standard, and only the physics varies.

Family	What is done to the dynamics	Axes
Forced isotropic	equations untouched, flow held statistically steady by stochastic injection	Reynolds number, forcing scale, forcing memory, helicity
Extended physics	exactly one ingredient added (a term in the momentum balance or a transported field)	rotation, stratification, passive scalar
Free decay	the drive removed	initial state (four controlled families)

Configurations

Each configuration lives in its own repository (so you can download exactly what you need). Sizes are the on-disk zarr size.

Configuration	Repository	Size	Family
Re_lambda 86 (flagship)	`ydai17/TIDE-ou_relam90_256_fp64`	478 GB	forced
k_f = 3	`ydai17/TIDE-ou_robust_kf3_256_fp64`	237 GB	forced
k_f = 4	`ydai17/TIDE-ou_robust_kf4_256_fp64`	238 GB	forced
tau = 1	`ydai17/TIDE-ou_robust_tau1_256_fp64`	295 GB	forced
rotating (strong)	`ydai17/TIDE-rotating_ro0p2_256_fp64`	222 GB	extended
rotating (moderate)	`ydai17/TIDE-rotating_ro0p2_v2_256_fp64`	231 GB	extended
passive scalar	`ydai17/TIDE-scalar_sc1_256_fp64`	269 GB	extended
decay (hot-start)	`ydai17/TIDE-decay_hotstart_re86`	80 GB	decay
decay (Saffman)	`ydai17/TIDE-decay_saffman_v2`	79 GB	decay
decay (Batchelor)	`ydai17/TIDE-decay_batchelor_v2`	80 GB	decay
Re_lambda 70	`ydai17/TIDE-ou_relam70_256_fp64`	473 GB	forced
Re_lambda 55	`ydai17/TIDE-ou_relam50_256_fp64`	320 GB	forced
helical	`ydai17/TIDE-helical_re86_retune2_256_fp64`	236 GB	forced
stratified	`ydai17/TIDE-stratified_reb40_256_fp64`	294 GB	extended
decay (ABC)	`ydai17/TIDE-abc_turb_full_256_fp64`	94 GB	decay

What a frame contains

Each configuration is one chunked zarr store, one frame per chunk (zstd):

<CASE>.zarr/
  u          [N, 3, 256, 256, 256]  fp32   velocity
  p          [N, 256, 256, 256]     fp32   pressure (spectrally solved, certified)
  theta / b  [N, 256, 256, 256]     fp32   passive scalar / buoyancy (5-channel cases)
  t          [N]                           physical time of each frame
  k_max_eta  [N]                           per-frame resolution margin
  seed       [N]                           trajectory index

Fields are computed in fp64 and stored in fp32; frames are exported every 0.05 T_L (about one Kolmogorov time). Every released frame is individually Class I (k_max·eta >= 1.5).

Usage

pip install -U huggingface_hub zarr
huggingface-cli download ydai17/TIDE-ou_relam90_256_fp64 --repo-type dataset \
    --local-dir ./tide-data/corpus

import zarr
z = zarr.open("./tide-data/corpus/ou_relam90_256_fp64.zarr", mode="r")
u = z["u"][0]        # (3, 256, 256, 256) velocity of the first frame
print(z["t"][:5], z["k_max_eta"][:5])

The benchmark harness reads these stores directly; see the code repository for the training and evaluation protocol, the acceptance referee, and the released result rows behind every number in the paper.

Files in this index repository

DATASHEET.md — datasheet for the dataset
benchmark_slice.json — the deterministic train/val/test manifest (3 train, 1 validation, 3 test trajectories per configuration, chosen by a model-independent quality score)
manifests/ — per-configuration manifests (seed and frame counts, channels)

License and citation

Data under CC-BY-4.0. Please cite the paper (see the code repository's CITATION.cff) and the DOI record above.

Citation

@misc{dai2026tide,
  title={TIDE: A Physically Diverse 3D Turbulence Benchmark Dataset for Advancing Scientific Machine Learning},
  author={Yilong Dai and Yiming Sun and Yiheng Chen and Shengyu Chen and Peyman Givi and Xiaowei Jia and Runlong Yu},
  year={2026},
  eprint={2608.04222},
  archivePrefix={arXiv},
  primaryClass={physics.flu-dyn},
  url={https://arxiv.org/abs/2608.04222},
}