Pith. sign in

REVIEW 3 major objections 4 minor 57 references

TIDE: A Physically Diverse 3D Turbulence Benchmark Dataset for Advancing Scientific Machine Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read TIDE introduces a DNS-verified, physically diverse 3D turbulence corpus showing that learned emulators barely beat persistence and stay about twice the error of a spectral solver given the true equations.

desk verdict TIDE is a genuinely useful, carefully documented 3D turbulence benchmark that fills a real gap, though its 'DNS-verified' label is slightly stronger than the evidence warrants because the verification reuses the generating solver. read the letter →

arxiv 2608.04222 v1 pith:4QDUBZZA submitted 2026-08-04 physics.flu-dyn cs.LG

Yilong Dai , Yiming Sun , Yiheng Chen , Shengyu Chen , Peyman Givi , Xiaowei Jia , Runlong Yu This is my paper · ORCID
classification physics.flu-dyncs.LG
keywords 3DturbulencedirectnumericalsimulationscientificmachinelearningbenchmarkdatasetneuraloperatorsphysicalfidelitygeneralizationNavier-Stokes
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

TIDE is a corpus and benchmark purpose-built to let machine learning be tested on three-dimensional incompressible turbulence, the setting where vortex stretching actually lives. The paper's central claim is that current learned forecasting models, trained and evaluated on this corpus under a fixed protocol, barely outperform simply copying the last frame forward, and remain about twice as far from the truth as a classical spectral solver that is handed the Navier-Stokes equations. The paper also claims that a lower pointwise error does not imply a physically faithful rollout: models can rank best by nRMSE while inflating the small-scale enstrophy by an order of magnitude. What makes the claim measurable is the corpus design: 15 configurations on eight controlled physical axes, each with 8-16 independent realizations, verified fields, and a forced-to-decay contrast that isolates whether an operator is conditioned on the external drive. If correct, TIDE turns 'did the model learn the dynamics?' from a slogan into a quantitative comparison.

What carries the argument

The central object is the configuration axis: one incompressible Navier-Stokes system, one pseudo-spectral solver on a $256^{3}$ periodic box, one acceptance standard, with only the physics varied across 15 configurations grouped into forced isotropic, extended physics (rotation, stratification, passive scalar), and free decay. The mechanism that carries the argument is the combination of independent realizations, each seed having its own initial condition and its own forcing sequence, with equation-level verification: divergence residuals at solver precision, momentum residuals measured on frame triplets with the stochastic forcing frozen, a step-halving check confirming second-order time truncation, and a pressure-Poisson consistency check. The forced-to-decay axis is what isolates conditioning: removing the drive changes no term of the governing equations, yet no single snapshot reveals whether a flow is driven or freshly decaying, so a forced-trained operator's persistence-level error on decay is evidence of a missing input rather than of how well the physics was learned.

What would settle it

Take a released forced-isotropic frame triplet and evolve it forward with an independent DNS implementation that differs in dealiasing, projection, and forcing scheme, then compare the ensemble statistics and short-horizon trajectories against TIDE's; if the fields diverge beyond sampling uncertainty while TIDE's own residual gates still pass, the acceptance standard certified self-consistency of one solver rather than fidelity to the incompressible Navier-Stokes equations.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is a negative result made precise: after standardizing training and evaluation on a $256^{3}$ direct-numerical-simulation corpus with independent ensembles, the benchmarked learned operators improve on persistence by at most about 8% in mean forecasting skill and their rollout error is roughly twice that of an equations-informed spectral integrator (nRMSE around 0.426 versus 0.843 persistence mean, with the best learned models near 0.78-0.82). No single architecture leads on all tasks: the attention-based operator is the most consistently positive on forecasting while the spectral convolution operator wins super-resolution on every configuration, and the pressure-recovery task is dominated by the exact Poisson solve that no learned model approaches. The paper further establishes that pointwise accuracy and physical fidelity separate: a model with lower nRMSE than another can have an order-of-magnitude larger enstrophy ratio and a high-band spectral error tens to hundreds of times larger. Finally, forced-to-decay transfer shows a missing conditioning variable: because the stochastic drive cannot be read off a single frame, operators trained under forcing continue to predict driven evolution when the drive is removed, at persistence-level error with in-distribution single-step error.

Load-bearing premise

The benchmark's value rests on the assumption that the acceptance procedure, including equation-level residual checks computed with the same pseudo-spectral discretization that generated the data, certifies that the released fields are faithful DNS of the incompressible Navier-Stokes equations rather than merely self-consistent outputs of one solver family.

Editorial extensions

If this is right

  • Any credible learned surrogate for 3D turbulence should be reported against persistence and against an equations-informed spectral solver; the interval between these two references is the unclaimed predictability, at least half of the total on these configurations.
  • Pointwise error alone cannot rank models: a leaderboard should report at least one small-scale physical metric, because the model with lowest nRMSE can have an enstrophy ratio of 51 against a target of 1.0.
  • The forced-to-decay result implies that single-frame surrogate formulations lack a conditioning input for the stochastic drive; adding a short history or an explicit forcing input is the direct remedy the benchmark points to.
  • Because rollout stability is configuration- and seed-dependent, single-seed leaderboard cells are not decisive; replicate seeds set the resolution of model comparisons.
  • Regime shifts whose signature is visible in the input field are mostly coverage problems, while the drive is not visible; generalization tests should be labeled by which category of shift they exercise.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension follows directly: condition an operator on the last two or three frames, or on an estimated injection rate, and measure whether the forced-to-decay error drops from persistence level toward the near-exact decay performance of the equations-informed integrator; TIDE's released frame sequences make this experiment immediate.
  • The separation of accuracy from physical fidelity suggests that small nRMSE improvements, on the order of a few hundredths, are within seed noise on this corpus, so future studies should compare error distributions or multi-seed intervals rather than point estimates.
  • The same missing-conditioning mechanism may explain failures in learned emulators of other stochastic PDEs where the noise realization is invisible in a single observation; the forced/decay contrast is a template for detecting such omitted state variables.
  • The corpus's controlled axes make it possible to train a joint model across regimes with the regime as an input, the defining ingredient of a foundation model for 3D turbulence, giving the field a concrete pretraining target rather than an abstract one.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces TIDE, a 256^3 fp64 DNS corpus and benchmark for 3D incompressible turbulence, comprising 15 configurations across eight controlled physics axes, 8–16 independent realizations per configuration, pressure channels, and an acceptance protocol combining statistical gates with equation-level residuals. The companion benchmark defines five tasks (forecasting, super-resolution, sparse reconstruction, pressure recovery, subgrid-stress closure) with five audited learned baselines and trivial/informed references, plus generalization splits along the controlled axes. Headline findings are that learned models barely outperform persistence and remain about twice as far from the truth as an equations-informed spectral integrator; that lower pointwise error can coincide with badly distorted small-scale dynamics; and that forced-to-decay transfer exposes a missing conditioning input, since the stochastic drive is not part of the model input.

Significance. If the claims hold, TIDE is a valuable shared resource: it is, to my knowledge, the first 3D incompressible turbulence corpus that combines physically diverse configurations with per-configuration ensembles, a certified pressure channel, and a fixed acceptance standard. The paper ships machine-checked acceptance scripts, audited baseline variants with tabulated deviations, reproducible code (per-run costs and seeds), and a measured replicate-seed study, which are concrete strengths. The three-axis evaluation (pointwise error, enstrophy ratio, high-band spectral error) is a genuine methodological contribution, and the forced-to-decay conditioning finding is an interesting, falsifiable open problem. However, the strength of the 'DNS-verified' claim depends on the verification protocol, and that protocol has a circularity that the paper acknowledges but does not resolve; this is the main point requiring revision.

major comments (3)
  1. [Section 4 / Appendix B (D2, D3)] The momentum residual check D2 is computed on dedicated frame triplets exported with the OU forcing state frozen, not on the released frames; D3 then confirms second-order time truncation on those same triplets. As a result, the released fields' momentum balance is not directly verified—D2 validates the solver's temporal accuracy on auxiliary runs. The abstract and Section 4 currently state that the corpus ships 'equation-level verification,' which overstates what D2 provides. Please either compute momentum residuals on released frames (e.g., by storing forcing realizations or using the frozen-forcing protocol on the released trajectory), or qualify the claim to indicate that D1 (divergence) and D4 (pressure–Poisson) are checked on released frames while D2/D3 certify the solver's time integration on auxiliary triplets.
  2. [Appendix B / Section 4] Every D-group check reuses the same pseudo-spectral operators, dealiasing, projection, and Eswaran–Pope forcing implementation that produced the data. A systematic implementation error would pass both the generator and the verifier, so the verification certifies self-consistency of one solver family rather than DNS fidelity in an absolute sense. The authors disclose the one-solver scope in Appendix K, but the central 'DNS-verified' claim (abstract, Section 1, Table 1) needs either an independent-solver spot check on a subset of released frames or an explicit restatement as 'solver-verified with equation-level residuals computed within the generating discretization.' This is load-bearing because the benchmark's conclusions about learned models are framed as turbulence conclusions, not solver-family conclusions.
  3. [Section 4 / Table 6] The statistical gates carry little independent certification weight at this Reynolds number: the Kolmogorov constant lacks a verdict because the plateau lies in the bottleneck region, and C_epsilon is compared against a low-Re trend. This is disclosed, but it means the certification burden falls almost entirely on the equation-level checks, which strengthens the need for the independent cross-check raised in the previous comment.
minor comments (4)
  1. [Abstract / Introduction] The phrase 'DNS-verified' appears in the abstract and Section 1; consider a precise definition in Section 4 of what 'verified' means relative to the D-group scope, so readers do not infer a stronger guarantee than D1–D4 provide.
  2. [Table 1] The symbols (✓, partial, ✗) are not defined in the caption; consider adding a legend for readability.
  3. [Section 3.3] The statement 'at roughly one Kolmogorov time per frame' is useful, but the exact relation between Δt=0.05T_L and τ_η is not given; please add the measured value for the flagship configuration.
  4. [Table 4 caption] The footnote about median and divergence count is clear, but the table would benefit from explicit divergence counts for all cells, not only the median cells.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark conclusions are measured against external references, and the one-solver verification procedure is a disclosed evidentiary limitation rather than a circular reduction.

full rationale

The paper makes no first-principles derivation whose output is encoded in its inputs. Its central claims are that TIDE is a DNS-verified, physically diverse corpus and that current learned models barely outperform persistence while remaining about twice as error-prone as an equations-informed spectral solver. The second claim is empirical and externally anchored: persistence and the spectral integrator bracket the reported errors, no fitted parameter is renamed as a prediction, and the learned-model results are reproducible from released code and manifests. The first claim rests on the Section 4 and Appendix B D-group verification, which checks divergence, momentum residuals with a step-halving order check, and pressure-Poisson consistency. It is true that these checks reuse the same pseudo-spectral solver family that generated the data, and that D2 uses dedicated frozen-forcing frame triplets rather than the released frames themselves. However, this is a limitation on the independence of the verification, not a circular reduction: the acceptance standard is a fixed, pre-specified protocol with thresholds, it is validated on known-answer and deliberately violating fields, and a candidate release has been rejected by the release-blocking independence test. The paper also explicitly discloses the one-solver scope in Appendix K: 'All data come from one pseudo-spectral solver family on a periodic box, so TIDE is a turbulence-physics corpus rather than a cross-discretization benchmark.' The absence of an independent-solver cross-check is an evidentiary gap in the strength of the 'DNS-verified' label and belongs in a correctness-risk assessment, not in a circularity finding. No load-bearing self-citation, uniqueness import, ansatz smuggling, or renaming of a known result is present.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The ledger contains no invented physical entities. The paper's contribution is a corpus and benchmark; the free parameters are protocol choices (cadence, acceptance thresholds, normalization scales, split weights) rather than fitted physics constants.

free parameters (4)
  • frozen normalization velocity scale per configuration = not reported numerically; fit on each configuration's training trajectories
    Inputs to all baselines are scaled by these constants; the shared velocity scale preserves component anisotropy, and held-out frames can fall slightly outside the nominal range (Section 3.3, Appendix K).
  • benchmark cadence = Δt = 0.05 T_L
    Design choice in Section 3.3; sets forecasting difficulty so consecutive frames differ by about 17% relative error. Not fitted to data, but it controls the persistence comparison.
  • acceptance gate thresholds = k_max·η ≥ 1.5; resolved dissipation ≥99.5%; energy drift ≤1%; injection-dissipation closure ≤2%; scale separation…
    Hand-chosen acceptance standard (Section 4, Table 7). One configuration, strong rotation, is admitted with scale separation below threshold and reported rather than gated.
  • train/test quality score weights = equal weights on five trajectory quality terms
    Trajectory ranking for deterministic benchmark splits uses five equally weighted terms (Appendix H); changing weights changes the split.
assumptions (5)
  • domain assumption Incompressible Navier-Stokes equations with exact spectral projection are the correct governing model for the intended turbulent flows.
    Invoked throughout; the dataset is a corpus of solutions to these equations, not a model of real wall-bounded or compressible flows.
  • domain assumption A 256^3 pseudo-spectral discretization with 2/3 dealiasing, RK3 time stepping, and fp64 computation faithfully approximates the true resolved dynamics at the claimed resolution.
    Section 3.2; the acceptance procedure checks self-consistency with this discretization, so solver error is not independently benchmarked against another solver.
  • domain assumption Eswaran-Pope stochastic Ornstein-Uhlenbeck forcing restricted to large scales yields representative statistically steady homogeneous turbulence.
    Section 3.2 and Appendix L; random-phase forcing is chosen after a measured metastability failure of deterministic band forcing.
  • ad hoc to paper The statistical gates and equation-level residual checks (A1-A13, D1-D4) certify that released fields are usable DNS-quality data.
    Section 4; the acceptance standard is defined by the authors and validated on synthetic fields, but it is not an external certification.
  • domain assumption The exogenous OU forcing state cannot be recovered from a single frame, so a forced-trained operator sees a statistically matched but differently driven flow when applied to decay.
    Section 5.1.7 and Appendix N; the G4 interpretation of a missing conditioning variable depends on this non-observability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TIDE: A Physically Diverse 3D Turbulence Benchmark Dataset for Advancing Scientific Machine Learning." pith.science (2026). https://pith.science/paper/4QDUBZZA

@misc{pith2026260804222,
  author       = {Pith},
  title        = {Pith review of: TIDE: A Physically Diverse 3D Turbulence Benchmark Dataset for Advancing Scientific Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4QDUBZZA}},
  note         = {Machine review of arXiv:2608.04222}
}
read the original abstract

Turbulence is a central testbed for machine learning on physical dynamics because its governing laws are known exactly. However, most existing studies remain in 2D, while 3D turbulence has fundamentally different physics and is far more costly to simulate. Existing 3D resources also typically provide only one realization per configuration, making it difficult to distinguish learning the dynamics from fitting the statistics of a single flow. In this paper, we introduce TIDE (Turbulent Incompressible DNS Ensembles), a 256^3 DNS corpus and benchmark for 3D incompressible turbulence, with 15 configurations on eight controlled axes, independent ensembles, pressure fields, and equation-level verification. The benchmark includes five tasks, standardized learned baselines, controlled generalization splits, and physical-fidelity metrics alongside pointwise error. Across the main forecasting configurations, current learned models barely outperform persistence and still make about twice the error of a spectral solver given the true equations. Moreover, lower pointwise error can coincide with severely distorted small-scale dynamics, showing that accuracy alone does not ensure physical fidelity. Generalization results further show that most regime shifts reflect limited training coverage, whereas forced-to-decay transfer exposes a missing conditioning variable: operators trained under forcing continue to predict driven evolution when the external drive is removed. Closing these accuracy, fidelity, and conditioning gaps is the central open problem made measurable by TIDE.

Figures

Figures reproduced from arXiv: 2608.04222 by the authors.

Figure 1
Figure 1. Vortex structure of three regimes at an early and a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the machine-learning pipeline for tur [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The taxonomy of TIDE: three regime families, de [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Predicted vs. true energy spectra at rollout step 20 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Energy spectra of all 15 configurations (forcing band shaded, [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 9
Figure 9. Figure 9: Passive scalar 𝜃 (𝑆𝑐=1) at four instants [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 6
Figure 6. Figure 6: Forced HIT, 𝑢𝑥 mid-plane slices: 𝑅𝑒𝜆 ∈ {55, 70, 86} and 𝑘𝑓 =4 (shared colormap) [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Remaining forced configurations: 𝑘𝑓 =3, 𝜏=1, helical [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Rotating turbulence, 𝜔𝑧 on a vertical slice at four times: strong (top) and moderate (bottom) rotation. being starved of steps; a reproducer changing batch size or resolu￾tion should account for the resulting change in optimizer steps per epoch. How each architecture m…
Figure 13
Figure 13. Figure 13: Zero-shot transfer onto the 𝑅𝑒𝜆86 test set, sources ordered by physical distance. The trivial-to-informed interval. Section 6 reports the bracket in aggregate; per configuration the equations-informed integrator reaches nRMSE 0.410, 0.403, 0.488, 0.411, 0.405, and 0.4…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 39 canonical work pages

  1. [1]

    Guido Boffetta and Robert E Ecke. 2012. Two-dimensional turbulence.Annual review of fluid mechanics44, 1 (2012), 427–451

    work page 2012

  2. [2]

    Wouter JT Bos, Liang Shao, and J-P Bertoglio. 2007. Spectral imbalance and the normalized dissipation rate of turbulence.Physics of fluids19, 4 (2007)

    work page 2007

  3. [3]

    Marc E Brachet, Daniel I Meiron, Steven A Orszag, Bernhard G Nickel, Rudolf H Morf, and Uriel Frisch. 1983. Small-scale structure of the Taylor–Green vortex. Journal of Fluid Mechanics130 (1983), 411–452

    work page 1983

  4. [4]

    Steven L Brunton, Bernd R Noack, and Petros Koumoutsakos. 2020. Machine learning for fluid mechanics.Annual review of fluid mechanics52, 1 (2020), 477–508

    work page 2020

  5. [5]

    Junho Choi, Teng-Yuan Chang, Namjung Kim, and Youngjoon Hong. 2025. A data free neural operator enabling fast inference of 2D and 3D Navier Stokes equations.arXiv preprint arXiv:2510.23936(2025)

    work page arXiv 2025

  6. [6]

    Wai Tong Chung, Bassem Akoush, Pushan Sharma, Alex Tamkin, Ki Sung Jung, Jacqueline Chen, Jack Guo, Davy Brouzet, Mohsen Talei, Bruno Savard, et al

    work page

  7. [7]

    Yilong Dai, Shengyu Chen, Xiaowei Jia, Peyman Givi, and Runlong Yu. 2026. PEST: Physics-Enhanced Swin Transformer for 3D Turbulence Simulation. arXiv:2602.10150 [physics.flu-dyn] https://arxiv.org/abs/2602.10150

    arXiv 2026

  8. [8]

    Yilong Dai, Shengyu Chen, Xiaowei Jia, and Runlong Yu. 2026. Flow Learn- ers for PDEs: Toward a Physics-to-Physics Paradigm for Scientific Computing. arXiv:2604.07366 [cs.LG] https://arxiv.org/abs/2604.07366

    work page Pith review arXiv 2026

Show all 57 references
  1. [9]

    Yilong Dai, Shengyu Chen, Ziyi Wang, Xiaowei Jia, Yiqun Xie, Vipin Ku- mar, and Runlong Yu. 2026. Learning PDE Solvers with Physics and Data: A Unifying View of Physics-Informed Neural Networks and Neural Operators. arXiv:2601.14517 [cs.LG] https://arxiv.org/abs/2601.14517

    2026

  2. [10]

    Yilong Dai, Yiming Sun, Yiheng Chen, Shengyu Chen, Xiaowei Jia, and Run- long Yu. 2026. FlowRefiner: Flow Matching-Based Iterative Refinement for 3D Turbulent Flow Simulation. arXiv:2604.17149 [physics.flu-dyn] https: //arxiv.org/abs/2604.17149

    2026 arXiv

  3. [11]

    Yilong Dai, Yiming Sun, Yiheng Chen, Ziyi Wang, Shengyu Chen, Xiaowei Jia, and Runlong Yu. 2026. Physics-Preserving Latent Compression for Zero-Shot Resolution Transfer in 3D Turbulence. arXiv:2606.21781 [physics.flu-dyn] https: //arxiv.org/abs/2606.21781

    2026 arXiv

  4. [12]

    Pan Du, Meet Hemant Parikh, Xiantao Fan, Xin-Yang Liu, and Jian-Xun Wang

  5. [13]

    Karthik Duraisamy, Gianluca Iaccarino, and Heng Xiao. 2019. Turbulence model- ing in the age of data.Annual review of fluid mechanics51, 1 (2019), 357–377

    2019

  6. [14]

    Vinayak Eswaran and Stephen B Pope. 1988. An examination of forcing in direct numerical simulations of turbulence.Computers & Fluids16, 3 (1988), 257–278

    1988

  7. [15]

    Kai Fukami, Koji Fukagata, and Kunihiko Taira. 2019. Super-resolution recon- struction of turbulent flows with machine learning.Journal of Fluid Mechanics 870 (2019), 106–120

    2019

  8. [16]

    Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. 2018. Datasheets for datasets. arXiv preprint arXiv:1803.09010(2018)

    2018 arXiv

  9. [17]

    Massimo Germano, Ugo Piomelli, Parviz Moin, and William H Cabot. 1991. A dynamic subgrid-scale eddy viscosity model.Physics of fluids a: Fluid dynamics 3, 7 (1991), 1760–1765

    1991

  10. [18]

    Zhongkai Hao, Chang Su, Songming Liu, Julius Berner, Chengyang Ying, Hang Su, Anima Anandkumar, Jian Song, and Jun Zhu. 2024. Dpot: Auto-regressive denoising operator transformer for large-scale pde pre-training.arXiv preprint arXiv:2403.03542(2024)

    2024 arXiv

  11. [19]

    Maximilian Herde, Bogdan Raonić, Tobias Rohner, Roger Käppeli, Roberto Moli- naro, Emmanuel De Bezenac, and Siddhartha Mishra. 2024. Poseidon: Efficient foundation models for pdes.Advances in Neural Information Processing Systems 37 (2024), 72525–72624

    2024

  12. [20]

    Sergio Hoyas and Javier Jiménez. 2006. Scaling of the velocity fluctuations in turbulent channels up to Re𝜏= 2003.Physics of fluids18, 1 (2006)

    2006

  13. [21]

    Yuchi Jiang, Zhijie Li, Yunpeng Wang, Huiyu Yang, and Jianchun Wang. 2025. An implicit adaptive Fourier neural operator for long-term predictions of three- dimensional turbulence.arXiv preprint arXiv:2501.12740(2025)

    2025

  14. [22]

    Yuchi Jiang, Yunpeng Wang, Huiyu Yang, and Jianchun Wang. 2025. Integrating Fourier Neural Operator with Diffusion Model for Autoregressive Predictions of Three-dimensional Turbulence.arXiv preprint arXiv:2512.12628(2025)

    2025 arXiv

  15. [23]

    Junhyuk Kim and Changhoon Lee. 2020. Deep unsupervised learning of turbu- lence for inflow generation at various Reynolds numbers.J. Comput. Phys.406 (2020), 109216

    2020

  16. [24]

    Dmitrii Kochkov, Jamie A Smith, Ayya Alieva, Qing Wang, Michael P Bren- ner, and Stephan Hoyer. 2021. Machine learning–accelerated computational fluid dynamics.Proceedings of the National Academy of Sciences118, 21 (2021), e2101784118

    2021

  17. [25]

    Felix Koehler, Simon Niedermayr, Rüdiger Westermann, and Nils Thuerey. 2024. APEBench: A benchmark for autoregressive neural emulators of PDEs.Advances in Neural Information Processing Systems37 (2024), 120252–120310

    2024

  18. [26]

    Georg Kohl, Liwei Chen, and Nils Thuerey. 2024. Benchmarking autoregressive conditional diffusion models for turbulent flow simulation. InICML 2024 AI for Science Workshop

    2024

  19. [27]

    Jean Kossaifi, Nikola Kovachki, Kamyar Azizzadenesheli, and Anima Anandku- mar. 2023. Multi-grid tensorized fourier neural operator for high-resolution pdes. arXiv preprint arXiv:2310.00120(2023)

    2023 arXiv

  20. [28]

    Myoungkyu Lee and Robert D Moser. 2015. Direct numerical simulation of turbulent channel flow up to Reˆ sub [tau]ˆ[approximately equal to] 5200.Journal of Fluid Mechanics774 (2015), 395

    2015

  21. [29]

    Yi Li, Eric Perlman, Minping Wan, Yunke Yang, Charles Meneveau, Randal Burns, Shiyi Chen, Alexander Szalay, and Gregory Eyink. 2008. A public turbulence database cluster and applications to study Lagrangian evolution of velocity increments in turbulence.Journal of Turbulence9 ...

    2008

  22. [30]

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. 2020. Fourier neural oper- ator for parametric partial differential equations.arXiv preprint arXiv:2010.08895 (2020)

    2020 arXiv

  23. [31]

    Zijie Li, Dule Shu, and Amir Barati Farimani. 2023. Scalable transformer for pde surrogate modeling.Advances in Neural Information Processing Systems36 (2023), 28010–28039

    2023

  24. [32]

    Julia Ling, Andrew Kurzawski, and Jeremy Templeton. 2016. Reynolds averaged turbulence modelling using deep neural networks with embedded invariance. Journal of Fluid Mechanics807 (2016), 155–166

    2016

  25. [33]

    Phillip Lippe, Bastiaan S Veeling, Paris Perdikaris, Richard E Turner, and Johannes Brandstetter. 2023. Pde-refiner: Achieving accurate long rollouts with neural pde solvers. InThirty-seventh Conference on Neural Information Processing Systems

    2023

  26. [34]

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karni- adakis. 2021. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature machine intelligence3, 3 (2021), 218–229

    2021

  27. [35]

    Yining Luo, Yingfa Chen, and Zhen Zhang. 2023. CFDBench: A large-scale benchmark for machine learning methods in fluid dynamics.arXiv preprint arXiv:2310.05963(2023)

    2023 arXiv

  28. [36]

    Michael McCabe, Bruno Régaldo-Saint Blancard, Liam Holden Parker, Ruben Ohana, Miles Cranmer, Alberto Bietti, Michael Eickenberg, Siavash Golkar, Ger- aud Krawezik, Francois Lanusse, et al . 2023. Multiple physics pretraining for physical surrogate models.arXiv preprint arXiv:...

    2023 arXiv

  29. [37]

    Charles Meneveau and Joseph Katz. 2000. Scale-invariance and turbulence models for large-eddy simulation.Annual Review of Fluid Mechanics32, 1 (2000), 1–32

    2000

  30. [38]

    Ruben Ohana, Michael McCabe, and others (Polymathic AI). 2024. The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning. Advances in Neural Information Processing Systems (NeurIPS) Datasets and Bench- marks Track(2024)

    2024

  31. [39]

    Stephen B Pope. 2001. Turbulent flows.Measurement Science and Technology12, 11 (2001), 2020–2021

    2001

  32. [40]

    Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. 2018. Physics- informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational physics378, C (2018)

    2018

  33. [41]

    Maziar Raissi, Alireza Yazdani, and George Em Karniadakis. 2020. Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations.Science 367, 6481 (2020), 1026–1030

    2020

  34. [42]

    Pu Ren, N Benjamin Erichson, Junyi Guo, Shashank Subramanian, Omer San, Zarija Lukic, and Michael W Mahoney. 2023. Superbench: A super-resolution benchmark dataset for scientific machine learning.arXiv preprint arXiv:2306.14070 (2023)

    2023 arXiv

  35. [43]

    Joseph Smagorinsky. 1963. General circulation experiments with the primitive equations: I. The basic experiment.Monthly weather review91, 3 (1963), 99–164

    1963

  36. [44]

    Katepalli R Sreenivasan. 1998. An update on the energy dissipation rate in isotropic turbulence.Physics of Fluids10, 2 (1998), 528–529

    1998

  37. [45]

    Kimberly Stachenfeld, Drummond B Fielding, Dmitrii Kochkov, Miles Cranmer, Tobias Pfaff, Jonathan Godwin, Can Cui, Shirley Ho, Peter Battaglia, and Al- varo Sanchez-Gonzalez. 2021. Learned coarse models for efficient turbulence simulation.arXiv preprint arXiv:2112.15275(2021)

    2021 arXiv

  38. [46]

    Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Dan MacKinlay, Francesco Alesiani, Dirk Pflüger, and Mathias Niepert. 2023. PDEBench: An extensive benchmark for sci-entific machine learning. InICLR 2023 Workshop on Physics for Machine Learning

    2023

  39. [47]

    Kiwon Um, Robert Brand, Philipp Holl, Nils Thuerey, et al. 2020. Solver-in-the- loop: Learning from differentiable physics to interact with iterative pde-solvers. arXiv preprint arXiv:2007.00016(2020)

    2020 arXiv

  40. [48]

    Rui Wang, Karthik Kashinath, Mustafa Mustafa, Adrian Albert, and Rose Yu

  41. [49]

    Gege Wen, Zongyi Li, Kamyar Azizzadenesheli, Anima Anandkumar, and Sally M Benson. 2022. U-FNO—An enhanced Fourier neural operator-based deep-learning model for multiphase flow.Advances in Water Resources163 (2022), 104180

    2022

  42. [50]

    Hao Wu, Yuan Gao, Fan Xu, Fan Zhang, Qingsong Wen, Kun Wang, Xiaomeng Huang, and Xian Wu. 2025. Differential-Integral Neural Operator for Long-Term Turbulence Forecasting.arXiv preprint arXiv:2509.21196(2025)

    2025 arXiv

  43. [51]

    Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long

  44. [52]

    PK Yeung and Ye Zhou. 1997. Universality of the Kolmogorov constant in nu- merical simulations of turbulence.Physical Review E56, 2 (1997), 1746. A Per-Configuration DNS Acceptance Tables Table 7 lists every acceptance gate with its threshold and the mea- sured envelope over t...

    1997

  45. [55]

    Transolver: A fast transformer solver for pdes on general geometries.arXiv preprint arXiv:2402.02366(2024)

    2024 arXiv

  46. [57]

    spectral× multiscale

    is a safeguard against divergence. The learning rate is identical across baselines by design (Appendix L records why a lower rate was rejected), and all models train from the same fixed random seed governing initialization and shuffling, so leaderboard gaps do not carry initia...

    2027

  47. [2020]

    InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining

    Towards physics-informed deep learning for turbulent flow prediction. InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 1457–1466. KDD 2027 D&B, 2027, (under submission) Yilong Dai, Yiming Sun, Yiheng Chen, Shengyu Chen, Peyman...

    2027

  48. [2023]

    InThirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    Turbulence in focus: Benchmarking scaling behavior of 3D volumetric super-resolution with BLASTNet 2.0 data. InThirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track

  49. [2024]

    Conditional neural field latent diffusion model for generating spatiotempo- ral turbulence.Nature Communications15, 1 (2024), 10416

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.