Pith. sign in

REVIEW 3 major objections 4 minor 78 references

Fourier analysis of the physics of transfer learning for data-driven subgrid-scale models of ocean turbulence

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Spectral mismatch explains why a CNN for ocean subgrid forcing fails to generalize, and retraining one hidden layer corrects it.

desk verdict A useful and largely well-executed extension of spectral analysis to transfer learning for ocean SGS models, but the central mechanistic claim in Fig. 5 is confounded by input-spectrum differences and needs a proper same-input control before it carries weight. read the letter →

arxiv 2504.15487 v1 pith:XNC2YAA2 submitted 2025-04-21 cs.LG nlin.CDphysics.ao-phphysics.geo-ph

Moein Darman , Pedram Hassanzadeh , Laure Zanna , Ashesh Chattopadhyay This is my paper · ORCID
classification cs.LGnlin.CDphysics.ao-phphysics.geo-ph
keywords transferlearningsubgrid-scaleparameterizationquasi-geostrophicturbulenceconvolutionalneuralnetworksactivationspectraFourieranalysisout-of-distributiongeneralizationoceaneddy
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

A 9-layer convolutional neural network trained to predict subgrid forcing in a two-layer quasi-geostrophic ocean turbulence model fails when the dynamics change—from isotropic eddies to anisotropic jets—because the frozen weights and biases underestimate the spectra of out-of-distribution activations as signals pass through the network. Retraining only the first hidden layer with 2–10 percent of target data corrects this underestimation, so the output spectrum approaches the filtered high-resolution reference. This spectral mechanism matters because transfer learning is a standard way to adapt climate and ocean models to new regimes, and the paper offers a concrete offline diagnostic—spectrum RMSE—that exposes generalization failure before expensive online simulation. It also finds that the learned kernels act as low-pass, Gabor, and high-pass filters regardless of flow isotropy, so transfer learning adjusts filter amplitudes and scale preferences rather than learning a new filter vocabulary.

What carries the argument

The central mechanism is the spectral evolution of activations through the CNN. For each layer, the Fourier transform of activation $j$ in layer $\ell$, $\hat{g}^j_\ell$, is a sum over the ReLU-positive mask of the linear pre-activation spectrum, which itself is built from the padded-kernel spectra $\widehat{W}^{\beta,j}_\ell$, the bias term $\hat{b}^j_\ell$ (nonzero only at zero wavenumber), and the previous activation spectrum $\hat{g}^{\beta}_{\ell-1}$. By comparing channel-averaged, meridionally averaged activation spectra between BNN$_{0,0}$ and BNN$_{0,i}$ for each layer, the paper localizes the failure to early layers and shows that re-training layer 2 changes both the kernel spectra and the ReLU-positive regions, propagating an upshift that aligns the output spectrum with FDNS. The learned kernels themselves act as low-pass, high-pass, and Gabor (oriented band-pass) filters, and the analysis tracks how their dominant-wavenumber footprint changes under transfer learning.

What would settle it

Train the same BNN architecture several times from different random seeds and compare out-of-distribution activation spectra: if the underestimation does not recur across seeds, the spectral mechanism is not systematic. Alternatively, retrain a late layer instead of the first hidden layer; if the output spectrum then matches the FDNS reference, the claim that the early-layer spectral bias is the bottleneck is falsified.

Watch

Extended reading notes

Core claim

The paper claims that a CNN trained entirely on Case 0 and applied to Cases 1–3 underestimates the channel-averaged, meridionally averaged activation spectra from the earliest hidden layers onward, and this underestimation compounds toward the output, producing a subgrid-forcing spectrum that falls short of the filtered direct-numerical-simulation (FDNS) reference. Transfer learning with only the first hidden layer ($\ell=2$) retrained on a small fraction of target data (2% and 10%) reverses this: the layer-2 spectra shift upward and the correction propagates through the frozen layers, so transfer-learned network spectra approach the base-network-on-target spectra and the FDNS reference. The Fourier-transformed kernels of layer 2, clustered by k-means, consistently fall into low-pass, Gabor, and high-pass filter families in all four cases; retraining mostly increases the amplitude at unchanged dominant wavenumbers and moves many dominant wavenumbers to lower $\kappa=\sqrt{k_x^2+k_y^2}$, meaning toward larger scales.

Load-bearing premise

The mechanism is inferred from a single trained network: if the underestimation in the shown layer-wise activation spectra is an artifact of one initialization or of the specific channel and meridional averaging, the claim that transfer learning works by correcting frozen-filter spectral bias collapses.

Editorial extensions

If this is right

  • Spectrum RMSE, not RMSE or correlation coefficient, is the offline metric that reveals whether a parameterization will generalize; models that look strong on standard errors can still be spectrally wrong.
  • Re-training only the first hidden layer with 2–10% of target data is enough to correct the spectral gap, so targeted transfer learning can replace full retraining in similar subgrid-scale applications.
  • The filter vocabulary (low-pass, Gabor, high-pass) is universal across isotropic and anisotropic training data, so adaptation between regimes is a matter of adjusting the amplitude and scale preference of existing filters, not of inventing new ones.
  • Online (a posteriori) tests show transfer-learned networks improve kinetic-energy spectra and potential-vorticity PDF tails wherever the base CNN leaves room, although scale-selective dissipation can mask these gains.

Reading between the lines

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

  • Inference: The layer-localized spectral gap could be used as a selection criterion for which layer to retrain in other architectures: retrain the first layer where the in- versus out-of-distribution activation-spectrum gap appears, rather than a fixed layer.
  • Inference: Because transfer learning shifts many kernel maxima toward lower wavenumbers, regimes dominated by large-scale energy may need only early-layer fine-tuning, while small-scale-dominated regimes may require retraining deeper layers; this is testable but not tested here.
  • Inference: The single-global-maximum analysis ignores kernels with multiple significant spectral peaks; counting all local maxima might reveal whether transfer learning works mainly by amplitude scaling or by spectral redistribution.
  • Inference: The same activation-spectrum diagnostic could be applied to other data-driven parameterizations (such as atmospheric convection or boundary-layer closures) to predict beforehand whether a pretrained model will fail on a warmer climate and how much adaptation data is needed.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 studies transfer learning (TL) for a 9-layer convolutional neural network that predicts subgrid-scale potential vorticity forcing in a two-layer quasi-geostrophic ocean model. Four dynamical regimes are considered: two isotropic eddy configurations and two anisotropic jet configurations. The authors train a base network on Case 0 (BNN0), evaluate its out-of-distribution performance on Cases 1–3, and show that retraining only the first hidden layer on small amounts of target data (TLNN0,i) improves both offline metrics and online kinetic-energy spectra. Using Fourier analysis of kernels and hidden-layer activations, the paper argues that BNN0 fails to generalize because its learned weights and biases systematically underestimate out-of-distribution activation spectra as signals propagate through the network, and that retraining one layer corrects this spectral mismatch. The paper also claims that the learned convolutional kernels act as low-pass, Gabor, and high-pass filters regardless of whether the training data are isotropic or anisotropic.

Significance. If the proposed mechanism is correct, the paper would provide a physically interpretable explanation for when and why TL helps in data-driven subgrid-scale parameterization, potentially guiding layer-selection strategies in geophysical machine learning. The study is strengthened by its combination of offline and online evaluation, by covering both isotropic and anisotropic regimes, and by releasing code. The spectral diagnostics (kernel Fourier magnitudes and layer-wise activation spectra) are a useful and transferable interpretability tool. However, the central causal claim rests on a comparison that conflates input-spectrum differences with weight-induced underestimation, and the kernel-taxonomy claim is supported mainly by a qualitatively chosen clustering analysis. These issues are fixable but currently leave the headline conclusions less secure than the abstract suggests.

major comments (3)
  1. [Section 3.2, Fig. 5 Row a] The claim that BNN0's learned weights 'underestimate the out-of-distribution sample spectra' is not established by the comparison shown in Fig. 5 Row a. That row compares BNN0,0 applied to in-distribution Case 0 inputs with BNN0,i applied to out-of-distribution Case i inputs. Since the first-layer activations are convolutions of the fixed first-layer kernels with the input velocities (Eq. 8 with ell = 1), any difference in input spectra between cases (visible in Fig. 2b-d) appears directly in the layer-1 activation spectra, independent of whether the frozen weights are 'biased' for the target regime. The same gap would occur even if the weights were optimal for the target but the inputs had different spectral content. The appropriate control is a same-input comparison: feed identical target-case samples to BNN0 and to BNNi,i and compare their activation spectra. Fig. 5 Rows b-d provide such a comparison only after retraining (TLNN0,i vs BNNi,i), not for the pre-TL failure that the mechanism claim is about. Without this control, the central explanatory story in the abstract and Section 3.2 is not supported by the displayed evidence.
  2. [Section 3.3, Fig. 6 and Fig. 7] The claim that the learned kernels are 'low-pass, Gabor, and high-pass filters regardless of whether the training data are isotropic or anisotropic' rests on k-means clustering with the number of clusters chosen 'until qualitatively similar patterns are observed' (Fig. 6 caption) and on visual classification of the resulting cluster centers. No quantitative criterion is given for the cluster count, no error bars or seed ensembles are provided for the histograms in Fig. 7, and Section 4 acknowledges that the analysis assumes a single global maximum per kernel. These limitations are real and affect the cross-case generality claim. To make the claim load-bearing, the authors should report a reproducible cluster-selection rule, quantify the classification of filters (e.g., by peak location and bandwidth), and show stability across multiple training runs.
  3. [Section 3.2, Fig. 5; Section 2.3] The manuscript does not report error bars or seed ensembles for the activation spectra, which are channel-averaged and meridionally averaged. Fig. 5 and Fig. 7 appear to be based on a single training run per configuration. Because the paper's mechanism claim is that the spectral underestimation is a systematic property of the learned filters rather than an artifact of one initialization, the authors should either provide ensemble statistics over several independent training runs or explicitly justify why a single run is representative. This is especially important given that the central comparison in Fig. 5 Row a is confounded by input-spectrum differences, as noted above.
minor comments (4)
  1. [Eq. (12), Section 2.4] The symbol ⊛ is overloaded: it denotes spatial convolution in Eqs. (8) and (10), but in Eq. (12) it is used for convolution in the frequency domain. Please introduce a separate notation for frequency-domain convolution. Also, the statement that the Fourier transform 'can still be derived analytically' is misleading, because the resulting expression depends on the input-dependent set {h_j^ell > 0}, which is not known in closed form.
  2. [Eq. (13), Section 2.5] The RMSE formula in Eq. (13) is typeset incorrectly: the square root appears to apply only to the numerator, leaving the denominator outside the root in a way that does not match the intended normalized RMSE. Please fix the equation and clarify the normalization.
  3. [Fig. 5 caption] The caption says 'Rows a to d show spectra of hidden layers for different models and cases,' but it does not specify which networks are compared in each row (e.g., BNN0,0 vs BNN0,i in Row a, TLNN0,i vs BNNi,i in Rows b-d). Please make the row-by-row comparison explicit.
  4. [Abstract and Section 3.2] The phrase 'the learned weights and biases from one dataset underestimate the out-of-distribution sample spectra' presumes the causal conclusion that the paper is trying to establish. Until the same-input control is added, I suggest rewording to something like 'are associated with lower activation spectra on out-of-distribution inputs' in the abstract and in the Section 3.2 mechanism statement.

Circularity Check

1 steps flagged · score 6.0 of 10

Spectral-mechanism claim reduces to input-spectrum differences via Eq. (12); other empirical contributions are independent.

  1. self definitional [Abstract; Section 3.2, Fig. 5 Row a; cf. Eq. (12)]
    "When BNN 0 is applied to out-of-distribution inputs, it underestimates the channel-averaged, meridionally averaged activation spectrum relative to BNN 0,0, whose weights are tuned for in-distribution data. This underestimation begins in early layers and propagates through the network, ultimately resulting in a mismatch with the FDNS output spectrum."

    Fig. 5 Row a compares BNN0 on in-distribution Case 0 inputs with the same BNN0 on out-of-distribution inputs. Because the network and weights are identical, Eq. (12) shows the first-layer activation spectrum is, up to ReLU/bias terms, the product of the fixed weight spectrum and the input spectrum: ĝ1 = Σ Ŵ1 ⊙ ĝ0. Therefore the cross-case difference in layer-1 activations is exactly Ŵ1 ⊙ (ĝ0_OOD − ĝ0_Case0), i.e., the input-spectrum gap. The paper labels this gap as an underestimation caused by the learned weights and biases, but by construction the gap is entirely determined by the input spectra and would occur for any fixed first-layer weights.

full rationale

The only substantial circular step is the spectral-underestimation explanation of transfer-learning failure. Section 3.2 and the Abstract assert that learned weights and biases 'underestimate' out-of-distribution sample spectra, but the supporting comparison (Fig. 5 Row a) varies the inputs while holding the network fixed, so Eq. (12) makes the observed layer-1 underestimation identical to the input-spectrum difference. This affects the paper's central mechanistic claim, but not the independent empirical content: transfer learning does improve offline RMSE/spectrum RMSE and online kinetic-energy spectra, and the kernel-maxima statistics are computed from the trained networks themselves. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and the self-citations (Subel et al. 2023, Guan et al. 2022a) supply architecture and methods rather than the load-bearing result. The score of 6 reflects partial circularity in the main explanatory narrative, while recognizing that substantial parts of the study remain self-contained.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities and no fitted free constants. The ledger records the hand-chosen k-means cluster count, the QG and filtering assumptions that define the target, the architecture-specificity caveat, and the single-maximum simplification that underpins the kernel histograms.

free parameters (1)
  • number of k-means clusters = not reported; varied per case
    Used in Section 3.3 and Fig. 6 to classify kernel spectra into filter types; chosen by increasing until qualitatively similar patterns emerge, so part of the filter-taxonomy claim is sensitive to this hand choice.
assumptions (4)
  • domain assumption Two-layer quasi-geostrophic model with beta-plane and scale-selective dissipation faithfully represents the mesoscale ocean turbulence regimes studied.
    The entire data-generation pipeline (Section 2.2) uses pyqg QG simulations as ground truth; if QG misses relevant physics, the subgrid mapping is not representative.
  • domain assumption Filtering and coarse-graining (Gaussian filter, Eq. 7) with the assumption that coarsened high-resolution data resembles low-resolution data yields the correct subgrid forcing target.
    Section 2.2.3 defines the target via filtered DNS; the analysis inherits that definition.
  • domain assumption The chosen CNN architecture (9 layers, 5x5 kernels, ReLU) and MSE training produce parameterizations whose spectral behavior is representative of data-driven SGS models generally.
    The paper states findings may be specific to this architecture (Section 4), so generality rests on this assumption.
  • ad hoc to paper Each kernel's Fourier magnitude has a single global maximum; multi-lobed kernels are treated by their largest peak.
    Section 3.3 and the Discussion explicitly acknowledge this simplification, which underpins the histograms and the claim that TL shifts maxima toward low wavenumbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fourier analysis of the physics of transfer learning for data-driven subgrid-scale models of ocean turbulence." pith.science (2026). https://pith.science/paper/XNC2YAA2

@misc{pith2026250415487,
  author       = {Pith},
  title        = {Pith review of: Fourier analysis of the physics of transfer learning for data-driven subgrid-scale models of ocean turbulence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XNC2YAA2}},
  note         = {Machine review of arXiv:2504.15487}
}
read the original abstract

Transfer learning (TL) is a powerful tool for enhancing the performance of neural networks (NNs) in applications such as weather and climate prediction and turbulence modeling. TL enables models to generalize to out-of-distribution data with minimal training data from the new system. In this study, we employ a 9-layer convolutional NN to predict the subgrid forcing in a two-layer ocean quasi-geostrophic system and examine which metrics best describe its performance and generalizability to unseen dynamical regimes. Fourier analysis of the NN kernels reveals that they learn low-pass, Gabor, and high-pass filters, regardless of whether the training data are isotropic or anisotropic. By analyzing the activation spectra, we identify why NNs fail to generalize without TL and how TL can overcome these limitations: the learned weights and biases from one dataset underestimate the out-of-distribution sample spectra as they pass through the network, leading to an underestimation of output spectra. By re-training only one layer with data from the target system, this underestimation is corrected, enabling the NN to produce predictions that match the target spectra. These findings are broadly applicable to data-driven parameterization of dynamical systems.

Figures

Figures reproduced from arXiv: 2504.15487 by the authors.

Figure 1
Figure 1. Row a displays the schematic of the CNN and inputs and outputs in physical space. Each TLNN is initialized with the weights of BNN0 , and only the first hidden layer (ℓ = 2) is re-trained using a smaller percentage of data. The inputs are the meridional and zonal velocities of the upper and lower levels, and the output is the subgrid forcing for each level. Row b shows the inputs and outputs of the CNN in spectral s… view at source ↗
Figure 2
Figure 2. Comparative analysis of the base system and three target configurations using 10 years [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. A priori evaluation of CNN parameterization. Panel a: CC, RMSE, and spectrum RMSE for upper and lower levels, comparing BNNi,i, BNN0,i, and TLNN0,i with different re-training data percentages across three target cases. Panel b: Ratio of output spectrum to FDNS spectrum for each case 18 [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: A posteriori evaluation of CNN parameterization. Panels a, c, and e: Kinetic energy spectra from 10-year simulations using BNN0,i and TLNN0,i across different cases. Panels b, d, and f : PDFs of potential vorticity at the upper level for the same simulations 19 [PITH_…
Figure 5
Figure 5. Figure 5: Channel-averaged, meridionally averaged spectra of hidden layer activations across different networks. [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Cluster centers of filter spectra obtained by applying the [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Schematic and analysis of kernel changes in spectral space. [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

78 extracted references · 39 canonical work pages

  1. [1]

    Abernathey, Rochanotes , A

    R. Abernathey, Rochanotes , A. Ross, M. Jansen, Ziwei Li , F. J. Poulin, N. C. Constantinou, Anirban Sinha , Dhruv Balwada , SalahKouhen , S. Jones, C. B. Rocha, C. L. P. Wolfe, Chuizheng Meng , H. Van Kemenade, J. Bourbeau, J. Penn, J. Busecke, M. Bueti, and , Tobias . pyqg/pyqg: v0.7.2, 2022. URL https://zenodo.org/record/6563667

    work page arXiv 2022

  2. [2]

    Arakawa and V

    A. Arakawa and V. R. Lamb. Computational Design of the Basic Dynamical Processes of the UCLA General Circulation Model, page 173–265. Elsevier, 1977. doi:10.1016/b978-0-12-460817-7.50009-4. URL http://dx.doi.org/10.1016/B978-0-12-460817-7.50009-4

    doi:10.1016/b978-0-12-460817-7.50009-4 1977

  3. [3]

    Bracco, J

    A. Bracco, J. Brajard, H. A. Dijkstra, P. Hassanzadeh, C. Lessig, and C. Monteleoni. Machine learning for the physics of climate. Nature Reviews Physics, 7 0 (1): 0 6–20, Nov. 2024. ISSN 2522-5820. doi:10.1038/s42254-024-00776-3. URL http://dx.doi.org/10.1038/s42254-024-00776-3

    work page doi:10.1038/s42254-024-00776-3 2024

  4. [4]

    Bracco, J

    A. Bracco, J. Brajard, H. A. Dijkstra, P. Hassanzadeh, C. Lessig, and C. Monteleoni. Machine learning for the physics of climate. Nature Reviews Physics, 7 0 (1): 0 6--20, 2025

    work page 2025

  5. [5]

    Carati, S

    D. Carati, S. Ghosal, and P. Moin. On the representation of backscatter in dynamic localization models. Physics of Fluids, 7 0 (3): 0 606–616, Mar. 1995. ISSN 1089-7666. doi:10.1063/1.868585. URL http://dx.doi.org/10.1063/1.868585

    work page doi:10.1063/1.868585 1995

  6. [6]

    Chattopadhyay and P

    A. Chattopadhyay and P. Hassanzadeh. Long-term instabilities of deep learning-based digital twins of the climate system: The cause and a solution. arXiv preprint arXiv:2304.07029, 2023

    arXiv 2023

  7. [7]

    Chattopadhyay, A

    A. Chattopadhyay, A. Subel, and P. Hassanzadeh. Data-driven super-parameterization using deep learning: Experimentation with multiscale L orenz 96 systems and transfer learning. Journal of Advances in Modeling Earth Systems, 12 0 (11): 0 e2020MS002084, 2020

    work page 2020

  8. [8]

    Chattopadhyay, M

    A. Chattopadhyay, M. Gray, T. Wu, A. B. Lowe, and R. He. Oceannet: A principled neural operator-based digital twin for regional oceans. Scientific Reports, 14 0 (1): 0 21181, 2024

    work page 2024

Show all 78 references
  1. [9]

    X. Chen, C. Gong, Q. Wan, L. Deng, Y. Wan, Y. Liu, B. Chen, and J. Liu. Transfer learning for deep neural network-based partial differential equations solving. Advances in Aerodynamics, 3 0 (1), Dec. 2021. ISSN 2524-6992. doi:10.1186/s42774-021-00094-7. URL http://dx.doi.org/1...

    2021 doi

  2. [10]

    Desai, M

    S. Desai, M. Mattheakis, H. Joy, P. Protopapas, and S. Roberts. One-shot transfer learning of physics-informed neural networks, 2021. URL https://arxiv.org/abs/2110.11286

    2021 arXiv

  3. [11]

    Dipankar, B

    A. Dipankar, B. Stevens, R. Heinze, C. Moseley, G. Z\" a ngl, M. Giorgetta, and S. Brdar. Large eddy simulation using the general circulation model <scp>icon</scp>. Journal of Advances in Modeling Earth Systems, 7 0 (3): 0 963–986, July 2015. ISSN 1942-2466. doi:10.1002/2015ms...

    2015 doi

  4. [12]

    J. A. Domaradzki, R. W. Metcalfe, R. S. Rogallo, and J. J. Riley. Analysis of subgrid-scale eddy viscosity with use of results from direct numerical simulations. Physical Review Letters, 58 0 (6): 0 547–550, Feb. 1987. ISSN 0031-9007. doi:10.1103/physrevlett.58.547. URL http:/...

    1987 doi

  5. [13]

    D. G. Fox and S. A. Orszag. Pseudospectral approximation to two-dimensional turbulence. Journal of Computational Physics, 11 0 (4): 0 612--619, 1973

    1973

  6. [14]

    R. O. Fox. Large-eddy-simulation tools for multiphase flows. Annual Review of Fluid Mechanics, 44 0 (1): 0 47–76, Jan. 2012. ISSN 1545-4479. doi:10.1146/annurev-fluid-120710-101118. URL http://dx.doi.org/10.1146/annurev-fluid-120710-101118

    2012 doi

  7. [15]

    Fox-Kemper and D

    B. Fox-Kemper and D. Menemenlis. Can large eddy simulation techniques improve mesoscale rich ocean models?, page 319–337. American Geophysical Union, 2008. doi:10.1029/177gm19. URL http://dx.doi.org/10.1029/177GM19

    2008 doi

  8. [16]

    Gallet and R

    B. Gallet and R. Ferrari. A quantitative scaling theory for meridional heat transport in planetary atmospheres and oceans. AGU Advances, 2 0 (3): 0 e2020AV000362, 2021

    2021

  9. [17]

    Y. Gao, K. C. Cheung, and M. K. Ng. Svd-pinns: Transfer learning of physics-informed neural networks via singular value decomposition. In 2022 IEEE Symposium Series on Computational Intelligence (SSCI). IEEE, Dec. 2022. doi:10.1109/ssci51031.2022.10022281. URL http://dx.doi.or...

    2022

  10. [18]

    Germano, U

    M. Germano, U. Piomelli, P. Moin, and W. H. Cabot. A dynamic subgrid-scale eddy viscosity model. Physics of Fluids A: Fluid Dynamics, 3 0 (7): 0 1760–1765, July 1991. ISSN 0899-8213. doi:10.1063/1.857955. URL http://dx.doi.org/10.1063/1.857955

    1991 doi

  11. [19]

    Goswami, C

    S. Goswami, C. Anitescu, S. Chakraborty, and T. Rabczuk. Transfer learning enhanced physics informed neural network for phase-field modeling of fracture. Theoretical and Applied Fracture Mechanics, 106: 0 102447, Apr. 2020. ISSN 0167-8442. doi:10.1016/j.tafmec.2019.102447. URL...

    2020

  12. [20]

    Goswami, K

    S. Goswami, K. Kontolati, M. D. Shields, and G. E. Karniadakis. Deep transfer operator learning for partial differential equations under conditional shift. Nature Machine Intelligence, 4 0 (12): 0 1155–1164, Dec. 2022. ISSN 2522-5839. doi:10.1038/s42256-022-00569-2. URL http:/...

    2022 doi

  13. [21]

    M. A. Gray, A. Chattopadhyay, T. Wu, A. Lowe, and R. He. Long-term prediction of the gulf stream meander using oceannet: a principled neural operator-based digital twin. EGUsphere, 2024: 0 1--23, 2024

    2024

  14. [22]

    I. Grooms. Backscatter in energetically-constrained leith parameterizations. Ocean Modelling, 186: 0 102265, Dec. 2023 a . ISSN 1463-5003. doi:10.1016/j.ocemod.2023.102265. URL http://dx.doi.org/10.1016/j.ocemod.2023.102265

    2023

  15. [23]

    I. Grooms. Backscatter in energetically-constrained leith parameterizations. Ocean Modelling, 186: 0 102265, 2023 b

    2023

  16. [24]

    Y. Guan, A. Chattopadhyay, A. Subel, and P. Hassanzadeh. Stable a posteriori LES of 2D turbulence using convolutional neural networks: B ackscattering analysis and generalization to higher R e via transfer learning. Journal of Computational Physics, 458: 0 111090, 2022a

  17. [25]

    Y. Guan, A. Subel, A. Chattopadhyay, and P. Hassanzadeh. Learning physics-constrained subgrid-scale closures in the small-data regime for stable and accurate LES . Physica D: Nonlinear Phenomena, 443: 0 133568, 2023

    2023

  18. [26]

    Y. Guan, P. Hassanzadeh, T. Schneider, O. Dunbar, D. Z. Huang, J. Wu, and I. Lopez-Gomez. Online learning of eddy-viscosity and backscattering closures for geophysical turbulence using ensemble kalman inversion, 2024. URL https://arxiv.org/abs/2409.04985

    2024 arXiv

  19. [27]

    H. Guo, X. Zhuang, P. Chen, N. Alajlan, and T. Rabczuk. Analysis of three-dimensional potential problems in non-homogeneous media with physics-informed deep collocation method using material transfer learning and sensitivity analysis. Engineering with Computers, 38 0 (6): 0 54...

    2022 doi

  20. [28]

    Haghighat, M

    E. Haghighat, M. Raissi, A. Moure, H. Gomez, and R. Juanes. A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics. Computer Methods in Applied Mechanics and Engineering, 379: 0 113741, June 2021. ISSN 0045-7825. doi:10.1016/j.cma.20...

    2021

  21. [29]

    Hallberg

    R. Hallberg. Using a resolution function to regulate parameterizations of oceanic mesoscale eddy effects. Ocean Modelling, 72: 0 92--103, 2013

    2013

  22. [30]

    J. M. Hanna, J. V. Aguado, S. Comas-Cardona, R. Askri, and D. Borzacchiello. Residual-based adaptivity for two-phase flow simulation in porous media using physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 396: 0 115100, June 2022. ISSN 00...

    2022

  23. [31]

    Hornik, M

    K. Hornik, M. Stinchcombe, and H. White. Multilayer feedforward networks are universal approximators. Neural Networks, 2 0 (5): 0 359–366, Jan. 1989. ISSN 0893-6080. doi:10.1016/0893-6080(89)90020-8. URL http://dx.doi.org/10.1016/0893-6080(89)90020-8

    1989 doi

  24. [32]

    Jakhar, Y

    K. Jakhar, Y. Guan, R. Mojgani, A. Chattopadhyay, and P. Hassanzadeh. Learning closed-form equations for subgrid-scale closures from high-fidelity data: Promises and challenges, 2023. URL https://arxiv.org/abs/2306.05014

    2023 arXiv

  25. [33]

    M. F. Jansen and I. M. Held. Parameterizing subgrid-scale eddy effects using energetically consistent backscatter. Ocean Modelling, 80: 0 36–48, Aug. 2014. ISSN 1463-5003. doi:10.1016/j.ocemod.2014.06.002. URL http://dx.doi.org/10.1016/j.ocemod.2014.06.002

    2014 doi

  26. [34]

    M. F. Jansen, A. Adcroft, S. Khani, and H. Kong. Toward an energetically consistent, resolution aware parameterization of ocean mesoscale eddies. Journal of Advances in Modeling Earth Systems, 11 0 (8): 0 2844–2860, Aug. 2019. ISSN 1942-2466. doi:10.1029/2019ms001750. URL http...

    2019 doi

  27. [35]

    J. Kent, C. Jablonowski, J. Thuburn, and N. Wood. An energy‐conserving restoration scheme for the shallow‐water equations. Quarterly Journal of the Royal Meteorological Society, 142 0 (695): 0 1100–1110, Jan. 2016. ISSN 1477-870X. doi:10.1002/qj.2713. URL http://dx.doi.org/10....

    2016 doi

  28. [36]

    R. M. Kerr, J. A. Domaradzki, and G. Barbier. Small-scale properties of nonlinear interactions and subgrid-scale energy transfer in isotropic turbulence. Physics of Fluids, 8 0 (1): 0 197–208, Jan. 1996. ISSN 1089-7666. doi:10.1063/1.868827. URL http://dx.doi.org/10.1063/1.868827

    1996 doi

  29. [37]

    Khani and M

    S. Khani and M. L. Waite. Backscatter in stratified turbulence. European Journal of Mechanics - B/Fluids, 60: 0 1–12, Nov. 2016. ISSN 0997-7546. doi:10.1016/j.euromechflu.2016.06.012. URL http://dx.doi.org/10.1016/j.euromechflu.2016.06.012

    2016 doi

  30. [38]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  31. [39]

    Knaepen and P

    B. Knaepen and P. Moin. Large-eddy simulation of conductive flows at low magnetic reynolds number. Physics of Fluids, 16 0 (5): 0 1255–1261, May 2004. ISSN 1089-7666. doi:10.1063/1.1651484. URL http://dx.doi.org/10.1063/1.1651484

    2004 doi

  32. [40]

    Krueger, E

    D. Krueger, E. Caballero, J.-H. Jacobsen, A. Zhang, J. Binas, D. Zhang, R. L. Priol, and A. Courville. Out-of-distribution generalization via risk extrapolation (rex), 2020. URL https://arxiv.org/abs/2003.00688

    2020 arXiv

  33. [41]

    P. R. Larraondo, L. J. Renzullo, I. Inza, and J. A. Lozano. A data-driven approach to precipitation parameterizations using convolutional encoder-decoder neural networks, 2019. URL https://arxiv.org/abs/1903.10274

    2019 arXiv

  34. [42]

    D. C. Leslie and G. L. Quarini. The application of turbulence theory to the formulation of subgrid modelling procedures. Journal of Fluid Mechanics, 91 0 (01): 0 65, Mar. 1979. ISSN 1469-7645. doi:10.1017/s0022112079000045. URL http://dx.doi.org/10.1017/S0022112079000045

    1979 doi

  35. [43]

    Z. Li, H. Zheng, N. Kovachki, D. Jin, H. Chen, B. Liu, K. Azizzadenesheli, and A. Anandkumar. Physics-informed neural operator for learning partial differential equations, 2021. URL https://arxiv.org/abs/2111.03794

    2021 arXiv

  36. [44]

    T. Lund. On the use of discrete filters for large eddy simulation. Annual Research Briefs, pages 83--95, 1997

    1997

  37. [45]

    Lupin-Jimenez, M

    L. Lupin-Jimenez, M. Darman, S. Hazarika, T. Wu, M. Gray, R. He, A. Wong, and A. Chattopadhyay. Simultaneous emulation and downscaling with physically-consistent deep learning-based regional ocean emulators. arXiv preprint arXiv:2501.05058, 2025

    2025 arXiv

  38. [46]

    P. J. Mason and D. J. Thomson. Stochastic backscatter in large-eddy simulations of boundary layers. Journal of Fluid Mechanics, 242: 0 51–78, Sept. 1992. ISSN 1469-7645. doi:10.1017/s0022112092002271. URL http://dx.doi.org/10.1017/S0022112092002271

    1992 doi

  39. [47]

    Mathieu, M

    M. Mathieu, M. Henaff, and Y. LeCun. Fast training of convolutional networks through ffts. arXiv preprint arXiv:1312.5851, 2013

    2013 arXiv

  40. [48]

    Mattheakis, H

    M. Mattheakis, H. Joy, and P. Protopapas. Unsupervised reservoir computing for solving ordinary differential equations, 2021. URL https://arxiv.org/abs/2108.11417

    2021 arXiv

  41. [49]

    Meneveau and J

    C. Meneveau and J. Katz. Scale-invariance and turbulence models for large-eddy simulation. Annual Review of Fluid Mechanics, 32 0 (1): 0 1–32, Jan. 2000. ISSN 1545-4479. doi:10.1146/annurev.fluid.32.1.1. URL http://dx.doi.org/10.1146/annurev.fluid.32.1.1

    doi

  42. [50]

    S. A. Orszag. On the elimination of aliasing in finite-difference schemes by filtering high-wavenumber components. Journal of Atmospheric Sciences, 28 0 (6): 0 1074--1074, 1971

    1971

  43. [51]

    H. A. Pahlavan, P. Hassanzadeh, and M. J. Alexander. Explainable offline‐online training of neural networks for parameterizations: A 1d gravity wave‐qbo testbed in the small‐data regime. Geophysical Research Letters, 51 0 (2), Jan. 2024 a . ISSN 1944-8007. doi:10.1029/2023gl10...

    2024 doi

  44. [52]

    H. A. Pahlavan, P. Hassanzadeh, and M. J. Alexander. Explainable offline-online training of neural networks for parameterizations: A 1d gravity wave-qbo testbed in the small-data regime. Geophysical Research Letters, 51 0 (2): 0 e2023GL106324, 2024 b

    2024

  45. [53]

    Piomelli

    U. Piomelli. Large-eddy simulation: achievements and challenges. Progress in Aerospace Sciences, 35 0 (4): 0 335–362, May 1999. ISSN 0376-0421. doi:10.1016/s0376-0421(98)00014-1. URL http://dx.doi.org/10.1016/S0376-0421(98)00014-1

    1999 doi

  46. [54]

    S. Pope. Turbulent flows. Cambridge university press, 2000

    2000

  47. [55]

    Porta Mana and L

    P. Porta Mana and L. Zanna. Toward a stochastic parameterization of ocean mesoscale eddies. Ocean Modelling, 79: 0 1–20, July 2014. ISSN 1463-5003. doi:10.1016/j.ocemod.2014.04.002. URL http://dx.doi.org/10.1016/j.ocemod.2014.04.002

    2014 doi

  48. [56]

    K. G. Pressel, S. Mishra, T. Schneider, C. M. Kaul, and Z. Tan. Numerics and subgrid‐scale modeling in large eddy simulations of stratocumulus clouds. Journal of Advances in Modeling Earth Systems, 9 0 (2): 0 1342–1365, June 2017. ISSN 1942-2466. doi:10.1002/2016ms000778. URL ...

    2017 doi

  49. [57]

    S. Rasp, M. S. Pritchard, and P. Gentine. Deep learning to represent subgrid processes in climate models. Proceedings of the National Academy of Sciences, 115 0 (39): 0 9684–9689, Sept. 2018. ISSN 1091-6490. doi:10.1073/pnas.1810286115. URL http://dx.doi.org/10.1073/pnas.1810286115

    2018 doi

  50. [58]

    A. Ross, Z. Li, P. Perezhogin, C. Fernandez-Granda, and L. Zanna. Benchmarking of machine learning ocean subgrid parameterizations in an idealized model. Journal of Advances in Modeling Earth Systems, 15 0 (1): 0 e2022MS003258, 2023

    2023

  51. [59]

    P. Sagaut. Large eddy simulation for incompressible flows: an introduction. Springer Science & Business Media, 2005

    2005

  52. [60]

    Sagaut, M

    P. Sagaut, M. Terracol, and S. Deck. Multiscale and multiresolution approaches in turbulence-LES, DES and Hybrid RANS/LES Methods: Applications and Guidelines. World Scientific, 2013

    2013

  53. [61]

    Sarlak, C

    H. Sarlak, C. Meneveau, and J. Sørensen. Role of subgrid-scale modeling in large eddy simulation of wind turbine wake interactions. Renewable Energy, 77: 0 386–399, May 2015. ISSN 0960-1481. doi:10.1016/j.renene.2014.12.036. URL http://dx.doi.org/10.1016/j.renene.2014.12.036

    2015 doi

  54. [62]

    Schneider, S

    T. Schneider, S. Lan, A. Stuart, and J. Teixeira. Earth system modeling 2.0: A blueprint for models that learn from observations and targeted high‐resolution simulations. Geophysical Research Letters, 44 0 (24), Dec. 2017. ISSN 1944-8007. doi:10.1002/2017gl076101. URL http://d...

    2017 doi

  55. [63]

    Shevchenko and P

    I. Shevchenko and P. Berloff. On a minimum set of equations for parameterisations in comprehensive ocean circulation models. Ocean Modelling, 168: 0 101913, Dec. 2021. ISSN 1463-5003. doi:10.1016/j.ocemod.2021.101913. URL http://dx.doi.org/10.1016/j.ocemod.2021.101913

    2021

  56. [64]

    V. Shinde. Proper orthogonal decomposition assisted subfilter-scale model of turbulence for large eddy simulation. Physical Review Fluids, 5 0 (1), Jan. 2020. ISSN 2469-990X. doi:10.1103/physrevfluids.5.014605. URL http://dx.doi.org/10.1103/PhysRevFluids.5.014605

    2020 doi

  57. [65]

    General circulation experiments with the primitive equations: I

    Smagorinsky. General circulation experiments with the primitive equations: I. the basic experiment*. Monthly Weather Review, 91 0 (3): 0 99–164, Mar. 1963. ISSN 1520-0493. doi:10.1175/1520-0493(1963)091<0099:gcewtp>2.3.co;2. URL http://dx.doi.org/10.1175/1520-0493(1963)091<009...

    1963 doi

  58. [66]

    R. J. Stevens, L. A. Martínez-Tossas, and C. Meneveau. Comparison of wind farm large eddy simulations using actuator disk and actuator line models with wind tunnel experiments. Renewable Energy, 116: 0 470–478, Feb. 2018. ISSN 0960-1481. doi:10.1016/j.renene.2017.08.072. URL h...

    2018 doi

  59. [67]

    Subel, A

    A. Subel, A. Chattopadhyay, Y. Guan, and P. Hassanzadeh. Data-driven subgrid-scale modeling of forced B urgers turbulence using deep learning with generalization to higher R eynolds numbers via transfer learning. Physics of Fluids, 33 0 (3): 0 031702, 2021

    2021

  60. [68]

    Subel, Y

    A. Subel, Y. Guan, A. Chattopadhyay, and P. Hassanzadeh. Explaining the physics of transfer learning in data-driven turbulence modeling. PNAS Nexus, page pgad015, 2023

    2023

  61. [69]

    Subramanian, P

    S. Subramanian, P. Harrington, K. Keutzer, W. Bhimji, D. Morozov, M. Mahoney, and A. Gholami. Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. 2023. doi:10.48550/ARXIV.2306.00258. URL https://arxiv.org/abs/2306.00258

    2023 doi arXiv

  62. [70]

    Y. Q. Sun, H. A. Pahlavan, A. Chattopadhyay, P. Hassanzadeh, S. W. Lubis, M. J. Alexander, E. Gerber, A. Sheshadri, and Y. Guan. Data imbalance, uncertainty quantification, and generalization via transfer learning in data-driven parameterizations: Lessons from the emulation of...

    2023 arXiv

  63. [71]

    Z. Tan, T. Schneider, J. Teixeira, and K. G. Pressel. Large‐eddy simulation of subtropical cloud‐topped boundary layers: 2. cloud response to climate change. Journal of Advances in Modeling Earth Systems, 9 0 (1): 0 19–38, Jan. 2017. ISSN 1942-2466. doi:10.1002/2016ms000804. U...

    2017 doi

  64. [72]

    Thuburn, J

    J. Thuburn, J. Kent, and N. Wood. Cascades, backscatter and conservation in numerical models of two‐dimensional turbulence. Quarterly Journal of the Royal Meteorological Society, 140 0 (679): 0 626–638, June 2013. ISSN 1477-870X. doi:10.1002/qj.2166. URL http://dx.doi.org/10.1...

    2013 doi

  65. [73]

    C. Xu, B. T. Cao, Y. Yuan, and G. Meschke. Transfer learning based physics-informed neural networks for solving inverse problems in engineering structures under different loading scenarios. Computer Methods in Applied Mechanics and Engineering, 405: 0 115852, Feb. 2023. ISSN 0...

    2023

  66. [74]

    W. Xu, Y. Lu, and L. Wang. Transfer learning enhanced deeponet for long-time prediction of evolution equations, 2022. URL https://arxiv.org/abs/2212.04663

    2022 arXiv

  67. [75]

    Yosinski, J

    J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. How transferable are features in deep neural networks? 2014. doi:10.48550/ARXIV.1411.1792. URL https://arxiv.org/abs/1411.1792

    2014 doi arXiv

  68. [76]

    Zanna and T

    L. Zanna and T. Bolton. Data‐driven equation discovery of ocean mesoscale closures. Geophysical Research Letters, 47 0 (17), Aug. 2020. ISSN 1944-8007. doi:10.1029/2020gl088376. URL http://dx.doi.org/10.1029/2020GL088376

    2020 doi

  69. [77]

    Y. Zhou. Eddy damping, backscatter, and subgrid stresses in subgrid modeling of turbulence. Physical Review A, 43 0 (12): 0 7049–7052, June 1991. ISSN 1094-1622. doi:10.1103/physreva.43.7049. URL http://dx.doi.org/10.1103/PhysRevA.43.7049

    1991 doi

  70. [78]

    Zhuang, Z

    F. Zhuang, Z. Qi, K. Duan, D. Xi, Y. Zhu, H. Zhu, H. Xiong, and Q. He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109 0 (1): 0 43–76, Jan. 2021. ISSN 1558-2256. doi:10.1109/jproc.2020.3004555. URL http://dx.doi.org/10.1109/JPROC.2020.3004555

    2021

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.