AI-based separation of turbulence from coherent background flows in decaying hydrodynamic turbulence

Elena S. Volnova; Ji-Hoon Ha

REVIEW 3 major objections 5 minor 25 references

AI-based separation of turbulence from coherent background flows in decaying hydrodynamic turbulence

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A neural network trained only on static synthetic images can separate turbulence from a coherent background in evolving two-dimensional Navier-Stokes flows, without a spectral cutoff, until nonlinear mixing makes the split physically ambigu

desk verdict A useful proof-of-concept undermined by an abstract promise that isn't kept and a simulation metric that partly reuses the training loss. read the letter →

arxiv 2601.18163 v2 pith:MM3MVRF3 submitted 2026-01-26 physics.flu-dyn astro-ph.IM

Ji-Hoon Ha , Elena S. Volnova This is my paper · ORCID

classification physics.flu-dynastro-ph.IM

keywords turbulence-backgroundseparationdecayingturbulencetwo-dimensionalNavier-StokesneuralnetworkpowerspectrumdeeplearningdiagnosticKolmogorovscale

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

The paper tries to show that a neural network taught to separate a synthetic sinusoidal background from synthetic Kolmogorov-spectrum noise can do the same job on snapshots of a self-consistently evolving two-dimensional Navier-Stokes flow, without retraining and without a user-chosen spectral cutoff. It argues that the recovered turbulent field matches the energy decay and inertial-range spectrum of a Fourier-filtered reference during early and intermediate stages, and that the later breakdown reflects the physics of the flow rather than a failure of the model. The point of caring: astrophysical observations and simulations often need to separate small-scale fluctuations from large-scale coherent motion, and standard scale-filtering methods require subjective cutoffs; an AI diagnostic that learns the separation would be applicable where those assumptions are questionable. The study is framed not as a new separation algorithm but as a diagnostic probe of when 'turbulence minus background' is a meaningful notion at all.

What carries the argument

The carrying object is the composite training loss: the MSE term ensures pixel-level fidelity; the spectral term forces the radial power spectrum of the output to match the target turbulence over the band k >= k_inj to k = 0.35, encoding the physical definition of turbulence as small-scale fluctuations; the leakage penalty suppresses the projection of the output onto the known background, teaching the network what 'coherent background' means. The visual-transformer encoder with U-Net-style multi-scale decoder is the means of applying that definition to unseen data. The decaying Navier-Stokes testbed, with injection scale L k_inj / 2pi = 8 and viscosity 2e-3, provides the stress test: no forc

What would settle it

Apply the same trained network to snapshots whose initial turbulence amplitude or background RMS lies outside the training ranges, or whose injection wavenumber differs from L k_inj / 2pi = 8; if early-stage reconstruction error jumps discontinuously, the claimed generalization is narrow. A cheaper check: replace the AI output with a hard Fourier high-pass filter at the same cutoff k_inj and compute the same E_turb(t) curves; if the two are indistinguishable at every time, the learned separation has collapsed to an implicit spectral filter.

Watch

Extended reading notes

Core claim

The central claim is that a hybrid visual-transformer/U-Net trained only on static linear superpositions of a low-wavenumber sinusoidal background and an isotropic k^-5/3 random-phase turbulent field generalizes to time-evolving 2D incompressible Navier-Stokes snapshots in which the two components interact and mix. Trained with a composite objective—pixel MSE plus a radial power-spectrum matching term over the band k >= k_inj plus a penalty on the projection of the output onto the known background—the model removes large-scale coherent power, preserves inertial-range scaling, and tracks the turbulent-band energy decay throughout t <= 0.5. The authors report qualitative and quantitative agree

Load-bearing premise

The evaluation assumes that the training objective's definition of turbulence—the wavenumber band k >= k_inj enforced by the spectral-matching and leakage losses—is also the correct ground truth for the simulation, even though in a nonlinear evolving flow there is no independent true turbulence field once scales mix.

Editorial extensions

If this is right

A network can be deployed on simulation or observational snapshots without retraining and without a user-specified cutoff scale, as long as partial scale separation exists.
The turbulent-band energy decay, a physically transparent diagnostic, can be extracted by the AI from mixed fields, matching a spectral-window integration.
The method supplies an explicit temporal boundary (t > 0.5 in this testbed) beyond which turbulence-background separation is not meaningful, useful for designing analyses.
Because it avoids an explicit cutoff, the approach is a candidate for astrophysical data where projected, noisy, or sparsely sampled fields make scale filtering ambiguous.

Reading between the lines

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

If the training loss fixes the definition of turbulence as k >= k_inj, then the AI is not discovering a physically unique decomposition; it is learning and re-applying the training definition. A consequence the authors state in spirit but leave implicit: the same trained model will continue to output 'turbulence' according to its training prior even when the flow no longer contains true turbulence
The success on decaying flows suggests a testable extension: train on superpositions with multiple injection scales and evaluate whether the network can identify a time-dependent effective cutoff in forced three-dimensional turbulence, which would be a stronger claim about learning scale separation rather than memorizing one band.
Since the testbed is two-dimensional and decaying, a natural next test is a comparison of the AI output against wavelet or Hilbert-Huang decompositions on the same snapshots; if the AI tracks the physically relevant inertial-range energy while those methods differ, it would sharpen the claimed advantage over adaptive, non-spectral methods.

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.

Desk Editor's Note

read the letter

The short version: this paper is a plausible proof-of-concept with an overclaim in the abstract. The real question it asks—whether a network trained on static linear superpositions of a smooth background and Kolmogorov turbulence can separate fields in a self-consistently evolving decaying 2D Navier-Stokes simulation without retraining—is worth asking, and the early- and intermediate-stage results look qualitatively convincing. But the abstract's sentence about quantitative comparisons to a Fourier filtering baseline never materializes in the body, and the one quantitative simulation diagnostic, Eturb in Eq. (24), integrates power over k>=k_inj, which is exactly the band the training loss in Eq. (6) forces the network to match. That makes the close agreement in Fig. 8 partially a consistency check with the training objective, not an independent validation.

Credit where it's due: the paper doesn't oversell a new architecture; it explicitly positions the AI model as a diagnostic. Setting up decaying turbulence as a conservative testbed is a good idea, and the discussion of a temporal failure boundary (t>0.5) is physically reasonable. The synthetic test examples (Fig. 4) look clean, and the spectral plots in Fig. 7 do show the expected large-scale suppression. The paper also includes an honest caveat that AI separation is not a unique physical decomposition.

The soft spots are real but fixable. First, the missing baseline: either add the Fourier-filter comparison promised in the abstract or soften the claim. Second, the circularity: because k_inj is baked into both the training loss and the evaluation metric, the Eturb comparison can't distinguish a learned high-pass filter from a physically meaningful separation. You need at least one independent diagnostic—for instance, comparing the recovered small-scale field to a separately resolved run, or checking structure functions or intermittency statistics. Third, there are no error bars, no per-snapshot reconstruction metrics, and no code or trained weights shipped. That makes it hard to judge whether the later-time 'visually plausible' results are actually good. The 'no explicit spectral cutoff' framing is also overstated, since k_inj is still baked in.

Who should read this: anyone working on AI-based decomposition of turbulence in astrophysics or fluid dynamics. It deserves a serious referee—the question is timely and the testbed is well chosen—but it needs substantial revision before I'd trust the quantitative claims. I'd recommend sending it to review with a clear request for the baseline, independent metrics, and code/data release.

Referee Report

3 major / 5 minor

Summary. The paper trains a hybrid Swin Transformer–U-Net on static synthetic images to separate a small-scale turbulent field from a coherent large-scale background, then applies the frozen network to snapshots of freely decaying 2D Navier–Stokes turbulence initialized with a sinusoidal background and a Kolmogorov-like perturbation spectrum. The authors report that the model recovers visually and spectrally plausible turbulent fluctuations at early and intermediate times, preserves inertial-range scaling, and tracks the temporal decay of turbulent-band energy, and the abstract claims quantitative parity with a Fourier-filtering baseline without requiring an explicit spectral cutoff.

Significance. If established, the result would be useful: a model trained on idealized linear superpositions generalizing to nonlinearly evolving flows, without retraining, would provide a practical diagnostic for turbulence–background separation in simulations and observations. The paper also frames an important caveat—that the separation becomes intrinsically ambiguous at late times—and uses decaying turbulence as a conservative testbed. The synthetic-test portion, with visual and spectral agreement, is a reasonable proof of concept, and the explicit statement that AI separation is not a unique physical decomposition is honest. However, the quantitative claims in the abstract are not backed by the body, and the simulation-stage evaluation is partly self-referential, so the significance currently rests on unverified assertions.

major comments (3)

[Abstract and §3.7] The abstract states that 'quantitative comparisons with a Fourier filtering baseline show that the AI-based approach achieves comparable reconstruction accuracy.' No Fourier-filtered reference field, no RMSE/correlation metric, no comparison table, and no statistical test appear anywhere in Section 3 or elsewhere. The only quantitative simulation diagnostic, Eq. (24), measures turbulent-band energy of the AI output against the input, not against a Fourier-filter baseline. This central quantitative claim is therefore unsupported as written and must either be substantiated with actual baseline comparisons or removed/qualified.
[§2.3, Eq. (6) and §3.7, Eq. (24)] The simulation evaluation is partly circular. Eq. (6) trains the network to match the power spectral density over the band k_min=k_inj to k_max=0.35, and Eq. (7) penalizes projection onto the known background. Eq. (24) then defines the recovered turbulent energy by integrating the input and output spectra over exactly the band k≥k_inj. Consequently, the close agreement in Fig. 8 is to a significant degree a consistency check with the training objective: any network that learned a high-pass filter on the k≥k_inj band would satisfy this diagnostic. Since the nonlinear simulation has no true turbulence field once scales mix, the paper needs an independent validation target—for example, a synthetic test with a known ground truth that is not generated by the same k_inj and linear-superposition prescription, or comparison to a Fourier filter with an explicitly chosen cutoff.
[§2.4, Figs. 4–5] The synthetic-test evaluation is presented only through representative images and averaged spectra, without quantitative error metrics, confidence intervals, or per-sample scatter. Given that the paper's core claim is 'quantitative comparisons' and 'comparable reconstruction accuracy,' even the synthetic benchmark should report metrics such as normalized RMSE, correlation, spectral error as a function of k, and variability across the 128 test samples. Without these, the reader cannot assess whether the model's performance is statistically robust or whether the claimed parity with Fourier filtering is meaningful.

minor comments (5)

[§3.7, Eq. (24)] The normalization of E_turb uses '2πkΔk' with P(k) presumably radially averaged; please clarify whether P(k) is the 1D spectral density already integrated over angles, so that the extra 2πk factor is appropriate, and define Δk explicitly.
[Fig. 6] The top row is described as 'original simulation fields' but it is unclear which field is displayed (vorticity, velocity magnitude, or streamfunction). Please state the plotted quantity in the caption or text.
[§2.2] The notation '562, 282, 142, 72' for feature-map resolutions is confusing; use '56², 28², 14², 7²' or write '56×56, 28×28, ...' for clarity.
[§2.3, Eq. (6)] The spectral matching band is stated as k_min=k_inj and k_max=0.35 cycles/pixel. Please specify how k_inj is defined on the discrete FFT grid (exact bin index vs. continuous wavenumber) and whether the band is inclusive.
[§2.4] The training description says '40 epochs' and 'batch size 32' but does not report training time, learning-rate schedule, or the validation loss at convergence. A brief statement of these would help reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

Simulation energy diagnostic reuses the exact spectral band baked into the training loss, so the reported quantitative agreement is partially by construction.

fitted input called prediction [Section 2.3, Eq. (6); Section 3.7, Eq. (24)]
"L_spec = <(log10 P_{\hat I}(k) − log10 P_I(k))^2>_{kmin≤k≤kmax}, where we set k_min = k_inj and k_max = 0.35 (units: cycles per pixel). ... E_turb(t)= sum_{k≥k_inj} P(k,t) 2π k Δk, with L k_inj / 2π = 8."

The only quantitative simulation metric, E_turb, is defined by integrating power over k≥k_inj, which is exactly the band kmin=k_inj enforced by the spectral training loss in Eq. (6), using synthetic targets constructed from the same injection scale. Any network that satisfies Eq. (6) will approximately reproduce the input power in that band, so the close agreement in Fig. 8 between E_turb(input) and E_turb(AI) is largely a consistency check with the training objective. The leakage penalty in Eq. (7) additionally forces the output to be orthogonal to the known background, so the 'recovered turbulence' is defined by the training losses, not by an independent ground truth in the nonlinear simulation.
renaming known result [Abstract; Section 2.3, Eq. (6)]
"Quantitative comparisons with a Fourier filtering baseline show that the AI-based approach achieves comparable reconstruction accuracy while not requiring an explicit spectral cutoff scale. [Eq. (6):] we set k_min = k_inj and k_max = 0.35."

The 'no explicit spectral cutoff' claim holds only at inference time; during training the spectral-matching loss explicitly fixes kmin=k_inj and the leakage penalty projects against the known background. The network is therefore effectively a learned high-pass/band-pass filter whose cutoff is the training injection scale, so presenting it as not requiring scale separation renames the spectral-filtering assumption embedded in the training data and loss. Moreover, the promised Fourier-filtering baseline comparison does not appear anywhere in the body, and Eq. (24) reuses the same k≥k_inj band, so the abstract's quantitative parity claim lacks independent support.

full rationale

The paper involves no self-citation chain, and the training/generalization setup is not circular in itself: the network is trained on static synthetic superpositions and then applied to snapshots of a nonlinear Navier–Stokes simulation, which are genuinely different objects. The circularity is concentrated in the evaluation. The quantitative energy diagnostic in Eq. (24) integrates P(k) over k≥k_inj, the same band used to define the spectral training loss in Eq. (6) and to construct the synthetic turbulence targets, while Eq. (7) forces orthogonality to the known background. Consequently, the close spectral and energy agreement shown in Figs. 7 and 8 is partly a verification that the network implements its training objective on new inputs, not an independent confirmation that the extracted field is the physically correct turbulent component. This is genuine but partial circularity: the early-time visual separation and the suppression of large-scale power are nontrivial and are not guaranteed solely by L_spec, so the central claim retains some independent content. Score 6 reflects one or more predictions that reduce by construction, without the entire derivation being definitionally forced.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper contributes a trained network and a test protocol rather than a new physical theory. The load-bearing choices are the hand-set loss weights, the injection-wavenumber band, and the assumption that matching the power spectrum in that band is equivalent to physical separation. These are not derived from the simulation data and constitute the main circularity burden.

free parameters (6)

Loss weights λ_MSE, λ_spec, λ_leak = 1, 5e-3, 5e-2
Chosen by hand in Eq. (4) to balance pixel fidelity, spectral matching, and leakage suppression; no sensitivity analysis is provided.
Training injection scale ℓ_inj = L/8
Fixed in synthetic training data (§2.1) and reused as k_inj in the simulation (§3.6); defines the scale separation the model learns and the evaluation band.
Spectral matching band k_min/k_max = k_inj to 0.35 cycles/pixel
Chosen in Eq. (6); k_max=0.35 is set by hand and determines how much of the spectrum the network is trained to reproduce.
Turbulence-to-background ratio (turb_ratio) = 0.2
Initial turbulent RMS relative to background RMS in the simulation (§3.6); this single value sets the difficulty of the separation test.
Kinematic viscosity ν = 2e-3
Simulation parameter (§3.6) controlling small-scale dissipation; affects how quickly the turbulent band decays and mixes with the background.
Training data amplitude and RMS ranges = A∈[0.3,0.8], turbulence RMS∈[0.05,0.25]
Randomized ranges in the synthetic data generator (§2.1); these define the domain shift between training and simulation.

assumptions (4)

standard math The incompressible 2D Navier-Stokes equations in vorticity-streamfunction form, solved with a pseudo-spectral method and 2/3 dealiasing, adequately describe the testbed flow.
Invoked in §3.1–3.4; the entire simulation relies on this standard numerical model.
domain assumption A unique, physically meaningful turbulence-background separation exists for t≤0.5 and not beyond.
Stated in §3.7 and §4; the authors define the analysis window by this assumption, but no ground-truth separation is available to confirm it.
ad hoc to paper Linear superposition of independently generated background and turbulence is a valid training distribution for learning separation in nonlinearly evolving flows.
This is the central inductive leap: §2.1 builds training data by linear superposition, and the generalization claim in §4 depends on this transfer without retraining.
domain assumption Radially averaged power spectra and energy integrated over k≥k_inj adequately characterize the turbulent component.
Used in both training (Eq. (6)) and evaluation (Eq. (24)); this assumption makes the validation partly circular because it presumes the band that defines turbulence.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AI-based separation of turbulence from coherent background flows in decaying hydrodynamic turbulence." pith.science (2026). https://pith.science/paper/MM3MVRF3

@misc{pith2026260118163,
  author       = {Pith},
  title        = {Pith review of: AI-based separation of turbulence from coherent background flows in decaying hydrodynamic turbulence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MM3MVRF3}},
  note         = {Machine review of arXiv:2601.18163}
}

read the original abstract

Separating turbulent fluctuations from coherent large-scale background flows is a longstanding challenge in the analysis of numerical simulations and astronomical observations. Traditional approaches commonly rely on decomposition-based techniques such as Fourier or wavelet filtering, which assume that a meaningful separation can be achieved through scale selection. In realistic flows, however, coherent motions and turbulence often overlap across a broad range of scales and interact nonlinearly, making a unique separation inherently ambiguous. In this work, we investigate the robustness of an AI-based turbulence-background separation approach using two-dimensional incompressible Navier-Stokes simulations of decaying hydrodynamic turbulence. The simulations are initialized with a coherent background flow and divergence-free turbulent perturbations with a Kolmogorov-like spectrum and evolve without external forcing, providing a controlled physical testbed. A neural network trained exclusively on static synthetic images is applied to simulation snapshots at different evolutionary stages. The model recovers turbulent fluctuations during early and intermediate stages when partial scale separation is present. At later stages, where nonlinear interactions increasingly mix coherent and turbulent structures, the separation becomes less distinct; nevertheless, the recovered fields remain visually and spectrally consistent with the expected turbulent behavior. Quantitative comparisons with a Fourier filtering baseline show that the AI-based approach achieves comparable reconstruction accuracy while not requiring an explicit spectral cutoff scale. These results suggest that AI models trained on static data can provide a flexible diagnostic tool for turbulence-background separation in time-evolving flows, with potential applications to astrophysical datasets.

Figures

Figures reproduced from arXiv: 2601.18163 by the authors.

**Figure 1.** Representative examples of the synthetic training data used for the AI-based turbulence–background separation. The top row shows input images [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 3.** Evolution of the loss components during training. The total loss (solid [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Example results of the turbulence–background separation on independent synthetic test samples. From left to right, each column shows the true turbulent [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Figures from the paper (4 more)

**Figure 5.** Figure 5: Comparison of the radially averaged power spectra for the true tur [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Application of the trained AI model to snapshots from a two-dimensional incompressible Navier–Stokes simulation of decaying hydrodynamic turbulence. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Power spectra of the original simulation fields and the AI-extracted [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: Time evolution of the turbulent-band energy inferred from the input [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 3 linked inside Pith

[1]

M. P. Brenner, J. D. Eldredge, and J. B. Freund, ``Perspective on machine learning for advancing fluid mechanics,'' Phys. Rev. Fluids 4(10), 100501 (2019)

2019
[2]

S. L. Brunton, B. R. Noack, and P. Koumoutsakos, ``Machine Learning for Fluid Mechanics,'' Annu. Rev. Fluid Mech. 52, 477--508 (2020)

2020
[3]

Cho, ``A Technique for Removing Large-scale Variations in Regularly and Irregularly Spaced Data,'' Astrophys

J. Cho, ``A Technique for Removing Large-scale Variations in Regularly and Irregularly Spaced Data,'' Astrophys. J. 874(1), 75 (2019)

2019
[4]

Dmitruk, D

P. Dmitruk, D. G\'omez, A. Costa, and S. Ponce Dawson, ``Asymptotic states of decaying turbulence in two-dimensional incompressible flows,'' Phys. Rev. E 54(3), 2555--2563 (1996)

1996
[5]

J. M. Girart, R. Rao, and D. P. Marrone, ``Magnetic Fields in the Formation of Sun-Like Stars,'' Science 313(5788), 812--814 (2006)

2006
[6]

Granero-Belinchon, S

C. Granero-Belinchon, S. G. Roux, and N. B. Garnier, ``Multiscale and anisotropic characterization of images based on complexity: An application to turbulence,'' Physica D 459, 134027 (2024)

2024
[7]

S. A. Grebenev, W. Forman, C. Jones, and S. Murray, ``Wavelet Transform Analysis of the Small-Scale X-Ray Structure of the Cluster Abell 1367,'' Astrophys. J. 445, 607 (1995)

1995
[8]

R. H. Hildebrand, L. Kirby, J. L. Dotson, M. Houde, and J. E. Vaillancourt, ``Dispersion of Magnetic Fields in Molecular Clouds. I,'' Astrophys. J. 696(1), 567 (2009)

2009

Show all 25 references

[9]

Houde, J

M. Houde, J. E. Vaillancourt, R. H. Hildebrand, S. Chitsazzadeh, and L. Kirby, ``Dispersion of Magnetic Fields in Molecular Clouds. II.,'' Astrophys. J. 706(2), 1504 (2009)

2009
[10]

N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.-C. Yen, C. C. Tung, and H. H. Liu, ``The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis,'' Proc. R. Soc. A 454(1971), 903--995 (1998)

1971
[11]

N. E. Huang, Z. Shen, and S. R. Long, ``A New View of Nonlinear Water Waves: The Hilbert Spectrum,'' Annu. Rev. Fluid Mech. 31, 417--457 (1999)

1999
[12]

Imara and L

N. Imara and L. Blitz, ``Angular Momentum in Giant Molecular Clouds. I. The Milky Way,'' Astrophys. J. 732(2), 78 (2011)

2011
[13]

Kochkov, J

D. Kochkov, J. A. Smith, A. Alieva, Q. Wang, M. P. Brenner, and S. Hoyer, ``Machine learning--accelerated computational fluid dynamics,'' Proc. Natl. Acad. Sci. USA 118(21), e2101784118 (2021)

2021
[14]

R. H. Kraichnan and D. Montgomery, ``Two-dimensional turbulence,'' Rep. Prog. Phys. 43(5), 547 (1980)

1980
[15]

Li, ``Magnetic Fields in Molecular Clouds---Observation and Interpretation,'' Galaxies 9(2), 41 (2021)

H.-B. Li, ``Magnetic Fields in Molecular Clouds---Observation and Interpretation,'' Galaxies 9(2), 41 (2021)

2021
[16]

J. Ling, A. Kurzawski, and J. Templeton, ``Reynolds averaged turbulence modelling using deep neural networks with embedded invariance,'' J. Fluid Mech. 807, 155--166 (2016)

2016
[17]

Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, ``Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,'' arXiv:2103.14030 (2021)

2021 arXiv
[18]

Loshchilov and F

I. Loshchilov and F. Hutter, ``Decoupled Weight Decay Regularization,'' arXiv:1711.05101 (2019)

2019 arXiv
[19]

W. H. Matthaeus, W. T. Stribling, D. Martinez, S. Oughton, and D. Montgomery, ``Decaying, two-dimensional, Navier--Stokes turbulence at very long times,'' Physica D 51(1), 531--538 (1991)

1991
[20]

Nakamura, K

T. Nakamura, K. Fukami, K. Hasegawa, Y. Nabae, and K. Fukagata, ``Convolutional neural network and long short-term memory based reduced order surrogate for minimal turbulent channel flow,'' Phys. Fluids 33(2), 025116 (2021)

2021
[21]

Pattle, D

K. Pattle, D. Ward-Thompson, D. Berry, J. Hatchell, H.-R. Chen, A. Pon, P. M. Koch, W. Kwon, J. Kim, P. Bastien, J. Cho, S. Coud \'e , J. Di Francesco, G. Fuller, R. S. Furuya, S. F. Graves, D. Johnstone, J. Kirk, J. Kwon, C. W. Lee, B. C. Matthews, J. C. Mottram, H. Parsons, ...

2017
[22]

Pattle, L

K. Pattle, L. Fissel, M. Tahani, T. Liu, and E. Ntormousi, ``Magnetic fields in star formation: from clouds to cores,'' arXiv:2203.11179 (2022)

2022 arXiv
[23]

Ronneberger, P

O. Ronneberger, P. Fischer, and T. Brox, ``U-Net: Convolutional Networks for Biomedical Image Segmentation,'' in Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015, pp. 234--241 (2015)

2015
[24]

Seta and C

A. Seta and C. Federrath, ``Structure functions with higher-order stencils as a probe to separate small- and large-scale magnetic fields,'' Mon. Not. R. Astron. Soc. 533(2), 1875--1886 (2024)

2024
[25]

Sioulas, T

N. Sioulas, T. Zikopoulos, C. Shi, M. Velli, T. A. Bowen, A. Mallet, B. D. G. Chandran, L. Sorriso-Valvo, M. M. Martinovi \'c , S. S. Cerri, A. Verdini, N. Davis, and C. Dunn, ``Higher-order Analysis of Three-dimensional Anisotropy in Imbalanced Alfv \'e nic Turbulence,'' Astr...

2025

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

M. P. Brenner, J. D. Eldredge, and J. B. Freund, ``Perspective on machine learning for advancing fluid mechanics,'' Phys. Rev. Fluids 4(10), 100501 (2019)

2019

[2] [2]

S. L. Brunton, B. R. Noack, and P. Koumoutsakos, ``Machine Learning for Fluid Mechanics,'' Annu. Rev. Fluid Mech. 52, 477--508 (2020)

2020

[3] [3]

Cho, ``A Technique for Removing Large-scale Variations in Regularly and Irregularly Spaced Data,'' Astrophys

J. Cho, ``A Technique for Removing Large-scale Variations in Regularly and Irregularly Spaced Data,'' Astrophys. J. 874(1), 75 (2019)

2019

[4] [4]

Dmitruk, D

P. Dmitruk, D. G\'omez, A. Costa, and S. Ponce Dawson, ``Asymptotic states of decaying turbulence in two-dimensional incompressible flows,'' Phys. Rev. E 54(3), 2555--2563 (1996)

1996

[5] [5]

J. M. Girart, R. Rao, and D. P. Marrone, ``Magnetic Fields in the Formation of Sun-Like Stars,'' Science 313(5788), 812--814 (2006)

2006

[6] [6]

Granero-Belinchon, S

C. Granero-Belinchon, S. G. Roux, and N. B. Garnier, ``Multiscale and anisotropic characterization of images based on complexity: An application to turbulence,'' Physica D 459, 134027 (2024)

2024

[7] [7]

S. A. Grebenev, W. Forman, C. Jones, and S. Murray, ``Wavelet Transform Analysis of the Small-Scale X-Ray Structure of the Cluster Abell 1367,'' Astrophys. J. 445, 607 (1995)

1995

[8] [8]

R. H. Hildebrand, L. Kirby, J. L. Dotson, M. Houde, and J. E. Vaillancourt, ``Dispersion of Magnetic Fields in Molecular Clouds. I,'' Astrophys. J. 696(1), 567 (2009)

2009