Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Generative Lagrangian data assimilation for ocean dynamics under extreme sparsity

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A neural-operator-conditioned diffusion model reconstructs full-resolution ocean fields from 99%-sparse Lagrangian observations, preserving high-wavenumber dynamics that deterministic baselines lose.

desk verdict A credible empirical study of conditional diffusion for extremely sparse reconstruction, but the 'Lagrangian' framing is untested—the synthetic masks are random, not advected, and the real data are satellite tracks. read the letter →

arxiv 2507.06479 v1 pith:3PUVBRQL submitted 2025-07-09 physics.ao-ph cs.AIcs.LGmath.DSnlin.CD

classification physics.ao-phcs.AIcs.LGmath.DSnlin.CD
keywords LagrangiandataassimilationconditionaldiffusionmodelsneuraloperatorsoceanstatereconstructionextremesparsityspectralbiasseasurfaceheightGulfofMexico
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

This paper proposes a two-stage deep-learning method for turning extremely sparse ocean observations into full-resolution snapshots of sea-surface height and velocity. A Fourier Neural Operator first maps the sparse input to a coarse full-field guess, and a denoising diffusion probabilistic model, conditioned on that guess, refines it into a realistic turbulent field. The paper argues this recovers small-scale, high-wavenumber features even when 99% of grid points are missing in synthetic experiments and around 99.9% are missing for real satellite altimetry, outperforming plain UNET and FNO baselines on spectral and physical diagnostics. If true, the method offers a way to initialize ocean forecasts without running expensive ensemble data assimilation through a dynamical model.

What carries the argument

The load-bearing mechanism is the FNO-conditioned DDPM: a denoising diffusion model (a generative model that learns to reverse a gradual noise-adding process) whose UNET-based denoiser is conditioned, at every diffusion step, on the FNO's coarse prediction from the sparse observations, and whose reverse process is initialized from that same prediction plus noise. The FNO supplies global spatial structure and the diffusion process supplies high-wavenumber detail, so the two together overcome the spectral bias that limits either alone.

What would settle it

A decisive test would be to train and evaluate FNO+DDPM on synthetic observations generated by advecting virtual particles with the flow field from GLORYS, producing trajectory-correlated, non-uniform masks with temporal continuity, and compare high-wavenumber spectral error against the paper's random-mask results; if the spectral advantage over FNO shrinks substantially, the claim that the method reconstructs from Lagrangian observations would not hold for true Lagrangian sampling.

Watch

Extended reading notes

Core claim

The central discovery is that conditioning a denoising diffusion probabilistic model on the output of a Fourier Neural Operator allows full-resolution ocean fields to be sampled from a conditional distribution rather than regressed to a point estimate. The FNO maps the extremely sparse input to a coarse but globally coherent prior; the diffusion model then refines this prior, restoring small-scale, high-wavenumber structure that the FNO and UNET alone lose. The paper shows that the resulting spectra track the true Fourier spectrum far into the high-wavenumber range, and that derived physical fields (kinetic energy, strain rate, relative vorticity) remain faithful even at 99% synthetic sparsity and roughly 99.9% satellite sparsity.

Load-bearing premise

The central premise is that removing 99% of grid points at random, independently at each time step, reproduces the difficulty of real Lagrangian observations, whose sampling is trajectory-correlated and spatially non-uniform.

Editorial extensions

If this is right

  • FNO+DDPM and UNET+DDPM reconstructions maintain high-wavenumber spectral energy, so derived fields such as strain rate, relative vorticity, and kinetic energy stay physically consistent under extreme sparsity.
  • The pipeline reconstructs states in one pass without a background model or ensemble, avoiding the cost of ensemble Kalman filtering or variational assimilation.
  • The method works on real satellite altimetry over the Gulf of Mexico with about 99.9% missing data, recovering mesoscale eddies and gradients that deterministic models oversmooth.
  • The evaluation shows that RMSE, CC, and SSIM can reward over-smooth reconstructions, implying that data-driven ocean state estimation should be judged with spectral and physics-based diagnostics.

Reading between the lines

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

  • Editorial inference: The synthetic experiments use independently redrawn random masks, whereas real drifters and floats trace temporally coherent, trajectory-correlated paths; the method's performance on true Lagrangian trajectory sampling is therefore not directly measured.
  • Editorial inference: The same recipe of a cheap learned prior plus a diffusion refiner could transfer to other sparse inverse problems, such as reconstructing subsurface fields from Argo floats or using SWOT swath altimetry.
  • Editorial inference: Because the model is trained on reanalysis labels (GLORYS, CNAPS), it inherits the dynamical priors of those products; removing that dependence would require training directly on raw observations.
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 / 5 minor

Summary. The paper proposes a generative data-assimilation framework for reconstructing high-resolution ocean surface fields from extremely sparse observations. The method trains a Fourier Neural Operator (FNO) or UNET as a deterministic regressor from sparse inputs to full fields, then conditions a denoising diffusion probabilistic model (DDPM) on that regressor output; FNO+DDPM and UNET+DDPM are compared against the two regressors. Experiments cover three systems: a 256x256 beta-plane Kolmogorov-flow DNS with 99% random per-sample sparsity, GLORYS Gulf of Mexico SSH/SSU/SSV fields with the same synthetic 99% sparsity, and real satellite altimetry sea-surface height with the CNAPS reanalysis as target. Skill is reported via RMSE, correlation, SSIM, Fourier spectra, kinetic energy, strain rate, and relative vorticity. The paper concludes that the conditional diffusion models outperform the deterministic baselines in preserving small-scale, high-wavenumber dynamics, and positions the approach as a background-model-free alternative to traditional data assimilation.

Significance. If the central claim holds, the work is practically valuable: a one-shot, purely data-driven reconstruction of turbulent ocean fields at 99% sparsity that preserves spectral energy better than deterministic regression would be useful for initialization of data-driven forecast models and for rapid reconstruction from altimetry or future SWOT-like observations. The study has genuine strengths: it evaluates on a canonical turbulent benchmark, on a realistic regional reanalysis, and on real altimetry data; it presents public code; and it goes beyond pointwise metrics by assessing Fourier spectra and derived dynamical quantities such as strain and vorticity. The main weakness is that the experiments do not actually use Lagrangian trajectory observations, despite the title and abstract making Lagrangian data assimilation the central claim; this, together with the partly circular real-data validation, means the headline result is narrower than stated.

major comments (3)
  1. [Section 2.1 (Systems 1 and 2)] The synthetic observation operator used for the two main training/evaluation systems is not Lagrangian. The text states that 99% of the spatial grid points are randomly removed 'for every sample', i.e., an independent Bernoulli mask that is re-drawn at each time step. Real Lagrangian observations (drifters, floats) have positions that are advected by the flow, are temporally coherent, cluster in dynamically active regions, and sample states along trajectories, not independently random grid points. Because the title, abstract, and introduction claim 'Lagrangian data assimilation', the central capability of the method is not measured by Systems 1 and 2. I recommend either adding experiments with flow-advected synthetic particle trajectories (e.g., integrating tracer positions with the DNS/GLORYS velocity fields) or explicitly reframing the contribution as sparse random-point and satellite-track reconstruction, leaving genuine Lagrangian validation to future work.
  2. [Section 3.3 and Abstract] The real-data validation has two problems. First, the sparsity level is inconsistent: the Abstract claims 99.9% sparsity for real satellite observations, while Section 3.3 and Figure 8 state approximately 99.7%. Second, the validation is partly circular: the CNAPS target fields were produced by assimilating the same DUACS satellite altimetry observations into ROMS, so System 3 demonstrates reconstruction of a reanalysis that already contains the input observations, not an independent ground-truth check. The authors should state this circularity explicitly, correct the sparsity number, and soften the claim that real-data performance validates Lagrangian reconstruction.
  3. [Section 4 (Conclusion and Discussion)] The conclusion states that the framework is 'a scalable, accurate, and computationally efficient alternative to traditional ocean data assimilation systems', but no traditional DA baseline (EnKF, 3D-Var, or optimal interpolation) is included in any experiment. The comparisons in Section 3 are only against UNET and FNO regressors. The comparative claim against DA systems is therefore unsupported by the presented evidence and should either be backed by a baseline experiment or removed from the conclusion.
minor comments (5)
  1. [Section 3.2] The text before Figure 7 contains a typo: 'FNO+DDMP' should read 'FNO+DDPM'.
  2. [Section 3.1 / Figure 4] The caption and text say the kinetic energy comparison is 'averaged over 3000 samples', but Section 2.1 describes only 7000 training samples and 100 independent test samples; please specify whether these 3000 samples are from training, validation, or an extra held-out set.
  3. [Section 2.1 / Section 3.2] For System 2, the paper does not specify the exact temporal locations of the 50 test samples beyond 'outside the training set'; please state the test period explicitly, since consecutive days from a single season would have strong autocorrelation and would change the interpretation of the reported standard deviations.
  4. [Section 3.3] Please state explicitly how the missing satellite grid cells are encoded as model input (e.g., zeros with a separate mask channel), because the model must distinguish 'unobserved' from 'observed value of zero' in the sparse SSH fields.
  5. [Equations (5)-(9)] Equation (7) defines the initial noisy sample u_T using the FNO prediction as the mean; please clarify in the text whether this is the actual initialization used at inference time for all reverse-pass steps or only a conceptual prior, and whether any correction is applied for the fact that the reverse chain in Eq. (8) is normally started from pure noise.

Circularity Check

1 steps flagged · score 3.0 of 10

Partial circularity in the real-data validation: the CNAPS target is generated from the same satellite altimetry observations used as model input, while the synthetic Systems 1 and 2 remain independent.

  1. fitted input called prediction [Section 2.1, System 3 (and Abstract, Section 3.3)]
    "we validate our proposed framework on a third system using real satellite altimetry observations and a corresponding high-resolution reanalysis dataset, CNAPS [41], which was produced by assimilating these observations into the Regional Ocean Modeling System (ROMS) ... the model is trained to reconstruct full-resolution SSH fields from these sparse observations using the CNAPS reanalysis as the supervisory target."

    The System 3 ground truth is not an independent observation of the ocean state: CNAPS is the output of a ROMS-based data assimilation system whose input includes the same sparse satellite SSH observations fed to the model. Training a model to map sparse observations to CNAPS, and then evaluating reconstruction skill against CNAPS, measures how well the model reproduces the assimilation product rather than how well it recovers an independently measured state. The target is therefore partly a function of the model input, so the real-data validation is partially forced by construction. The synthetic Systems 1 and 2 use external DNS and GLORYS targets and are not affected by this circularity.

full rationale

The core derivation is largely self-contained: in Systems 1 and 2, the models are trained on synthetic sparse inputs (random 99% removal) against full-resolution DNS and GLORYS targets that are not derived from the inputs, so those benchmark results are independent and not circular. The FNO-conditioned DDPM architecture is a straightforward combination of published components and does not smuggle in the target through a self-citation or ansatz. The only identified circular step is confined to System 3, where the CNAPS reanalysis used as the ground truth was itself produced by assimilating the very satellite altimetry observations used as model input; this weakens the claim of validation on real satellite data but does not invalidate the synthetic demonstrations. Self-citations to the authors' prior work on spectral bias and OceanNet are contextual and not load-bearing for the main reconstruction result. Overall circularity burden is low to moderate, warranting a score of 3 rather than 0.

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

The paper introduces no new physical entities such as particles, forces, dimensions, or conserved quantities. The central claim rests on standard mathematical machinery (DDPM, FNO, UNET) and on domain assumptions about how synthetic observations are generated and what constitutes valid ground truth. The most consequential assumptions are that random masks simulate Lagrangian sampling and that reanalysis products built from the same observations can serve as independent validation targets.

free parameters (5)
  • FNO width = 20
    Architecture capacity choice; controls the expressivity of the neural operator conditioning the diffusion model.
  • Number of FNO Fourier layers = 4
    Architecture depth choice; affects how many spectral modes the operator retains.
  • Diffusion steps T = 1000
    Standard DDPM setting; trades sampling cost against generation quality.
  • Learning rate for FNO training = 0.001
    Adam optimizer setting reported in Section 2.2.2; standard choice, not data-fitted.
  • DDPM beta schedule endpoints (beta_start, beta_end) = not reported
    Noise schedule parameters in Eq. (9) are not given numerically, so the exact diffusion dynamics cannot be reproduced without the code.
assumptions (4)
  • domain assumption Random per-snapshot removal of 99% of grid points is analogous to Lagrangian observations.
    Used in Section 2.1, Systems 1 and 2. Real Lagrangian instruments are trajectory-correlated and non-uniform, so this equivalence is a strong, untested modeling assumption.
  • domain assumption CNAPS reanalysis is a valid ground truth for reconstruction skill.
    Used in Section 2.1, System 3. CNAPS is produced by assimilating the same satellite altimetry observations used as model input, so target and input are not fully independent.
  • standard math Standard DDPM forward and reverse processes (Eqs. 5-9) are correct for conditional reconstruction.
    The paper relies on the established denoising diffusion probabilistic model framework of Ho et al.; no new derivation is attempted.
  • domain assumption Fourier power spectrum is a sufficient diagnostic of physical fidelity.
    The paper's core support for DDPM superiority rests on spectral matching (Figures 3d, 6d-f, 9d), while pointwise metrics RMSE and SSIM do not favor the diffusion models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Lagrangian data assimilation for ocean dynamics under extreme sparsity." pith.science (2026). https://pith.science/paper/3PUVBRQL

@misc{pith2026250706479,
  author       = {Pith},
  title        = {Pith review of: Generative Lagrangian data assimilation for ocean dynamics under extreme sparsity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3PUVBRQL}},
  note         = {Machine review of arXiv:2507.06479}
}
abstract

Reconstructing ocean dynamics from observational data is fundamentally limited by the sparse, irregular, and Lagrangian nature of spatial sampling, particularly in subsurface and remote regions. This sparsity poses significant challenges for forecasting key phenomena such as eddy shedding and rogue waves. Traditional data assimilation methods and deep learning models often struggle to recover mesoscale turbulence under such constraints. We leverage a deep learning framework that combines neural operators with denoising diffusion probabilistic models (DDPMs) to reconstruct high-resolution ocean states from extremely sparse Lagrangian observations. By conditioning the generative model on neural operator outputs, the framework accurately captures small-scale, high-wavenumber dynamics even at $99\%$ sparsity (for synthetic data) and $99.9\%$ sparsity (for real satellite observations). We validate our method on benchmark systems, synthetic float observations, and real satellite data, demonstrating robust performance under severe spatial sampling limitations as compared to other deep learning baselines.

Figures

Figures reproduced from arXiv: 2507.06479 by the authors.

Figure 1
Figure 1. Schematic of the FNO-conditioned diffusion framework (FNO+DDPM). The [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Vorticity reconstructions under 99% sparsity. Each column shows a time [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Reconstruction performance across 100 test samples using (a) RMSE, (b) [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Average kinetic energy (KE) fields and absolute error maps over 3000 samples. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Reconstructed fields over the Gulf of Mexico on January 1, 2013: (a) SSH, [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Quantitative comparison over 50 test samples: (a) RMSE, (b) CC, (c) SSIM, [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Derived diagnostics: (a) strain rate σ and (b) relative vorticityζ—computed from reconstructed SSU and SSV fields. Ground truth shows sharp eddy and filament structures near the Loop Current. FNO captures smooth patterns but lacks detail; UNET delineates eddies better …
Figure 8
Figure 8. Figure 8: Three snapshots of satellite-observed sea surface height (SSH) over the Gulf [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Reconstruction performance averaged over 50 days: (a) RMSE, (b) correlation [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. GPU Performance of an Entropy-Stable Discontinuous Galerkin Euler Solver with Non-Conservative Terms

    math.NA 2026-05 unverdicted novelty 4.0 of 10

    GPU port of entropy-stable DG Euler solver with non-conservative buoyancy terms reaches nearly 70% of 64-bit peak on A100 volume kernels, delivers 10x speedup and 13x better energy efficiency versus CPU, and preserves...

Reference graph

Works this paper leans on

53 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [1]

    These initial conditions are obtained from satellite altimeters, ocean buoys, drifters, etc

    Introduction Forecasting multi-scale chaotic dynamical systems, e.g., the Earth’s ocean, requires a dynamical model of the system’s equations and accurate initial conditions. These initial conditions are obtained from satellite altimeters, ocean buoys, drifters, etc. [1, 2, 3]. All of these measuring devices that observe the different variables of the oce...

  2. [2]

    Methods 2.1. Systems In this section, we discuss the three different oceanic systems: a canonical high- resolution fluid flow simulation, regional high-resolution reanalysis data, and real satellite observations, for which we have validated our proposed model. System 1: We consider the 2D forced Kolmogorov flow on a β-plane at Re = 10,000. The governing e...

  3. [3]

    Results In this section, we demonstrate the performance of our proposed DDPM+FNO and DDPM+UNET models as compared to traditional UNET and FNO models in terms of statistical metrics such as RMSE, CC, SSIM, as well as physics-based metrics such as the Fourier spectrum of the reconstructed variables, kinetic energy, strain rate, and relative vorticity. It mu...

  4. [4]

    Conclusion and Discussion In this study, we introduced a novel generative modeling framework for reconstructing oceanic surface fields from highly sparse and non-stationary Lagrangian satellite observations. By conditioning denoising diffusion probabilistic models (DDPMs) on Fourier Neural Operator (FNO) predictions, we demonstrated significant improvemen...

  5. [5]

    NA and LJ conducted the research and developed the computational models with equal contributions

    Acknowledgement AC and RH designed the research. NA and LJ conducted the research and developed the computational models with equal contributions. TW provided the satellite observations. NA and AC wrote the manuscript. All authors analyzed the results and edited the manuscript. AC and LJ were supported by the National Science Foundation (grant no. 2425667...

  6. [6]

    Data and Code The computational models used in this study are publicly available in https: //github.com/TACS-UCSC/GenDA-Lagrangian

  7. [7]

    2019 Frontiers in Marine Science 6 90

    Moore A M, Martin M J, Akella S, Arango H G, Balmaseda M, Bertino L, Ciavatta S, Cornuelle B, Cummings J, Frolov S et al. 2019 Frontiers in Marine Science 6 90

  8. [8]

    Dickey T D 2003 Journal of Marine Systems 40 5–48

Show all 53 references
  1. [9]

    Kubota M, Iwasaka N, Kizu S, Konda M and Kutsuwada K 2002 Journal of oceanography 58 213–225

  2. [10]

    Molcard A, Griffa A and ¨Ozg¨ okmen T M 2005Journal of Atmospheric and Oceanic Technology 22 70–83 Generative Lagrangian Data Assimilation 22

  3. [11]

    Salman H, Kuznetsov L, Jones C and Ide K 2006 Monthly Weather Review 134 1081–1101

  4. [12]

    Kuznetsov L, Ide K and Jones C K 2003 Monthly Weather Review 131 2247–2260

  5. [13]

    Apte A, Jones C K and Stuart A 2008 Tellus A: Dynamic Meteorology and Oceanography 60 336–347

  6. [14]

    Bennett A F 2005 Inverse modeling of the ocean and atmosphere (Cambridge University Press)

  7. [15]

    Molcard A, Piterbarg L I, Griffa A, ¨Ozg¨ okmen T M and Mariano A J 2003Journal of Geophysical Research: Oceans 108

  8. [16]

    Oke P R, Sakov P and Schulz E 2009 Dynamics of Atmospheres and Oceans 48 121–142

  9. [17]

    Schulz-Stellenfleth J and Stanev E 2010 Ocean modelling 33 270–282

  10. [18]

    Martin M, Hines A and Bell M 2007 Quarterly Journal of the Royal Meteorological Society: A journal of the atmospheric sciences, applied meteorology and physical oceanography 133 981– 995

  11. [19]

    Lermusiaux P F 2006 Journal of Computational Physics 217 176–199

  12. [20]

    Smith G C, Roy F, Mann P, Dupont F, Brasnett B, Lemieux J F, Laroche S and B´ elair S 2014 Quarterly Journal of the Royal Meteorological Society 140 881–894

  13. [21]

    Edwards C A, Moore A M, Hoteit I and Cornuelle B D 2015 Annual review of marine science 7 21–42

  14. [22]

    Evensen G 2003 Ocean dynamics 53 343–367

  15. [23]

    Oke P R and Sakov P 2008 Journal of Atmospheric and Oceanic Technology 25 1004–1017

  16. [24]

    Oke P R, Brassington G B, Griffin D A and Schiller A 2010 Australian Meteorological and Oceanographic Journal 59 67–76

  17. [25]

    Xie J and Zhu J 2010 Ocean Modelling 33 283–298

  18. [26]

    Oke P R, Sakov P and Corney S P 2007 Ocean Dynamics 57 32–45

  19. [27]

    2022 arXiv preprint arXiv:2202.11214

    Pathak J, Subramanian S, Harrington P, Raja S, Chattopadhyay A, Mardani M, Kurth T, Hall D, Li Z, Azizzadenesheli K et al. 2022 arXiv preprint arXiv:2202.11214

  20. [28]

    2022 arXiv preprint arXiv:2212.12794

    Lam R, Sanchez-Gonzalez A, Willson M, Wirnsberger P, Fortunato M, Pritzel A, Ravuri S, Ewalds T, Alet F, Eaton-Rosen Z et al. 2022 arXiv preprint arXiv:2212.12794

  21. [29]

    Guan H, Arcomano T, Chattopadhyay A and Maulik R 2024 arXiv preprint arXiv:2405.16297

  22. [30]

    2023 arXiv preprint arXiv:2310.02074

    Watt-Meyer O, Dresdner G, McGibbon J, Clark S K, Henn B, Duncan J, Brenowitz N D, Kashinath K, Pritchard M S, Bonev B et al. 2023 arXiv preprint arXiv:2310.02074

  23. [31]

    Watt-Meyer O, Henn B, McGibbon J, Clark S K, Kwa A, Perkins W A, Wu E, Harris L and Bretherton C S 2025 npj Climate and Atmospheric Science 8 1–15

  24. [32]

    Chattopadhyay A, Gray M, Wu T, Lowe A B and He R 2024 Scientific Reports 14 21181 URL https://www.nature.com/articles/s41598-024-21181-0

  25. [33]

    Lowe A B, Gray M, Chattopadhyay A, Wu T and He R 2025 Artificial Intelligence for the Earth Systems

  26. [34]

    Gray M, Chattopadhyay A, Wu T, Lowe A and He R 2025 Ocean Science 21 1065–1080

  27. [35]

    Dheeshjith S, Subel A, Adcroft A, Busecke J, Fernandez-Granda C, Gupta S and Zanna L 2025 Geophysical Research Letters 52 e2024GL114318

  28. [36]

    2023 arXiv preprint arXiv:2312.12462

    Chen K, Bai L, Ling F, Ye P, Chen T, Luo J J, Chen H, Xiao Y, Chen K, Han T et al. 2023 arXiv preprint arXiv:2312.12462

  29. [37]

    2024 arXiv preprint arXiv:2407.15586

    McNally A, Lessig C, Lean P, Boucher E, Alexe M, Pinnington E, Chantry M, Lang S, Burrows C, Chrust M et al. 2024 arXiv preprint arXiv:2407.15586

  30. [38]

    2025 Nature 641 1172–1179

    Allen A, Markou S, Tebbutt W, Requeima J, Bruinsma W P, Andersson T R, Herzog M, Lane N D, Chantry M, Hosking J S et al. 2025 Nature 641 1172–1179

  31. [39]

    Chattopadhyay A, Nabizadeh E, Bach E and Hassanzadeh P 2022arXiv preprint arXiv:2206.04811

  32. [40]

    Chattopadhyay A, Mustafa M, Hassanzadeh P, Bach E and Kashinath K 2022 Geoscientific Model Development 15 2221–2237

  33. [41]

    Maulik R, Rao V, Wang J, Mengaldo G, Constantinescu E, Lusch B, Balaprakash P, Foster I and Kotamarthi R 2022 Geoscientific Model Development 15 3433–3445

  34. [42]

    Martin S A, Manucharyan G E and Klein P 2024 EarthArXiv Under review at Journal of Advances Generative Lagrangian Data Assimilation 23 in Modeling Earth Systems (JAMES) URL https://doi.org/10.31223/X5WD6H

  35. [44]

    Chattopadhyay A and Hassanzadeh P 2023 arXiv preprint arXiv:2304.07029

  36. [45]

    Maltrud M and Vallis G 1991 Journal of Fluid Mechanics 228 321–342

  37. [46]

    Garric G, Parent L, Greiner E, Dr´ evillon M, Hamon M, Lellouche J, R´ egnier C, Desportes C, Le Galloudec O, Bricaud C et al. 2018 Performance and quality assessment of the global ocean eddy-permitting physical reanalysis glorys2v4 Proceedings of the Eight EuroGOOS Internatio...

  38. [47]

    He R, Wu T, Mao S, Zong H, Zambon J, Warrillow J, Dorton J and Hernandez D 2025 arXiv preprint arXiv:2503.06907

  39. [48]

    Storkey D, Blockley E, Furner R, Guiavarc’h C, Lea D, Martin M, Barciela R, Hines A, Hyder P and Siddorn J 2010 Journal of operational oceanography 3 3–15

  40. [49]

    Ronneberger O, Fischer P and Brox T 2015 U-net: Convolutional networks for biomedical image segmentation Medical Image Computing and Computer-Assisted Intervention (MICCAI) (Springer) pp 234–241 URL http://lmb.informatik.uni-freiburg.de/Publications/2015/ RFB15a/

  41. [50]

    Oommen V, Bora A, Zhang Z and Karniadakis G E 2024 arXiv preprint arXiv:2409.08477 URL https://arxiv.org/abs/2409.08477

  42. [51]

    Li Z, Kovachki N, Azizzadenesheli K, Liu B, Stuart A and Anandkumar A 2020 arXiv preprint arXiv:2010.08895 URL https://arxiv.org/abs/2010.08895

  43. [52]

    Lai C Y, Hassanzadeh P, Sheshadri A, Sonnewald M, Ferrari R and Balaji V 2023 Annual Review of Fluid Mechanics To appear URL https://arxiv.org/abs/2301.01341

  44. [53]

    Neha F, Bhati D, Shukla D K, Dalvi S M, Mantzou N and Shubbar S 2024 arXiv preprint arXiv:2412.02242 URL https://arxiv.org/abs/2412.02242

  45. [54]

    Lupin-Jimenez L, Darman M, Hazarika S, Wu T, Gray M, He R, Wong A and Chattopadhyay A 2025 arXiv preprint arXiv:2501.05058

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.