Pith. sign in

REVIEW 4 major objections 4 minor 99 references

The paper demonstrates that sparse ocean observations alone can refine an AI ocean model's reconstruction and forecasting through a two-phase EM loop, cutting reconstruction RMSE by 28% and forecast RMSE by 27%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Using an EM loop with Langevin sampling, a generative state-space model learns ocean dynamics from incomplete satellite observations and improves SST reconstruction and one-step forecasts over its pretrained baseline.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A promising EM-based generative SSM for learning ocean dynamics from sparse observations, but the headline attribution is confounded by training volume—referee it, don't kill it. the 4 major comments →

arxiv 2607.19147 v1 pith:T67RKYN5 submitted 2026-07-21 cs.LG cs.AI

Incomplete Observations Boost Evolutionary Performance in Ocean Modeling

Yangyang Kong , Yutong Jiang , Yanhai Gan , Junyu Dong , Feng Gao , Xiaopei Lin This is my paper
classification cs.LG cs.AI
keywords generative state-space modelexpectation-maximizationsparse observationsocean modelingLangevin dynamicssea surface temperaturesatellite remote sensingself-evolution
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Ocean and weather foundation models are typically trained on complete reanalysis fields, which are expensive to produce and cap modeling skill. This paper argues that a machine-learning ocean model can instead learn and improve directly from sparse, noisy observations such as satellite swaths, by treating the complete ocean state as a hidden variable in a generative state-space model and training it with an expectation–maximization (EM) loop. In the E-step, Langevin dynamics in the latent space reconstruct the hidden fields conditioned on the partial observations; in the M-step, those reconstructions are used as pseudo-ground-truth to update the initial-state generator and the transition network. On a simulated sea-surface-temperature benchmark with 67% of the domain masked and Gaussian noise added, 17 EM iterations reduced reconstruction RMSE by 28.13% and one-step forecast RMSE by 27.09% relative to the pre-trained baseline. The same procedure also produced qualitatively sharper and more temporally coherent reconstructions from real Arctic satellite swaths, supporting the paper's claim that incomplete observations can directly improve the model's representation of ocean-state dynamics.

Core claim

The paper's central discovery is that a generative state-space model can bootstrap a better representation of ocean dynamics from incomplete observations alone, without complete ground-truth fields as supervision. The model factorizes the joint distribution of hidden ocean states and measurements into an initial-state model, a stochastic transition model, and a masked Gaussian emission model. Training alternates between an E-step that samples latent noise variables via Langevin dynamics to reconstruct full sea-surface-temperature fields consistent with the sparse measurements, and an M-step that updates the generator and transition networks using those reconstructions as pseudo-targets. The

What carries the argument

The central object is a continuous-state hidden Markov model whose state is the complete ocean field and whose emission is a masked Gaussian p(ot|st)=N(ot|Mt⊙st, σ²I), with Mt a binary mask for observed pixels. The initial-state distribution is a style-based GAN generator using depthwise-separable convolutions; the transition distribution is a stochastic U-Net with layer-wise Gaussian noise injection; the observation model has no learnable parameters. The EM loop uses latent-space Langevin dynamics in the E-step: instead of sampling states, it samples the full set of latent noise vectors (initial input plus all layer-wise noises), with a Gaussian prior regularizing unobserved regions. The M-

Load-bearing premise

The method assumes that the reconstructed ocean fields the model produces in each round are accurate enough to serve as training targets — if that reconstruction mainly reflects what the model already believes rather than what the sparse measurements show, the iterative loop could lock in the model's mistakes instead of learning from the data.

What would settle it

Run the identical 17-iteration EM protocol but delete the observation-likelihood term from the E-step energy so the Langevin sampler draws purely from the learned prior. If reconstruction and forecast RMSE still improve as much as in the paper's tables, the gains are not attributable to the sparse observations; if the improvements vanish or reverse, the observation term is the active ingredient.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • AI ocean models can be trained or fine-tuned directly on sparse, noisy satellite swaths, removing the need for complete reanalysis fields as the primary training data.
  • On the simulated benchmark with 67% occlusion, the final model cuts reconstruction RMSE by 28.13% and one-step forecast RMSE by 27.09% relative to the pre-trained baseline, implying that incomplete observations alone can refine a pre-trained prior.
  • When the transition model is given perfect ground-truth inputs, it still forecasts better after EM (RMSE 0.7431 vs 0.9263), showing the EM loop improves the learned dynamics, not just the input fields.
  • Because the framework is posed as a general state-space model with an EM update, the same recipe can in principle be applied to longer sequences and to other Earth-system variables, though the paper only demonstrates the two-frame setting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial: The reported gains are against a baseline pre-trained on roughly 800 complete samples. A head-to-head against a model trained on all ~1,500 complete (unmasked) samples would reveal how much of the improvement is EM extracting extra information from the sparse data versus simply more training volume.
  • Editorial: A direct ablation of the observation-consistency term in the E-step — sampling from the prior alone for the same 17 iterations — would isolate whether the gain is truly driven by the sparse observations or by the model's own self-training dynamics.
  • Editorial: If the mechanism extends to longer sequences, running EM on T>1 trajectories (e.g., 5–10 frames) should compound the transition-model gain; with only T=1, the framework does not demonstrate multi-step memory beyond one step.
  • Editorial: The stationary, ergodic assumption used for efficiency may fail over seasonal cycles; an online variant that re-estimates the prior with seasonal conditioning would be a natural stress test for real-world operational forecasting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The manuscript proposes a generative state-space model for learning ocean dynamics directly from sparse, noisy observations. The model couples an initial-state generator (modified StyleGAN2-ADA) and a stochastic U-Net transition model with a masked Gaussian observation model, and is trained by Monte Carlo EM: the E-step reconstructs full states via latent-space Langevin sampling, and the M-step uses those reconstructions as pseudo-ground-truth targets for the generative networks. Experiments on a CMIP6-based OSSE compare a warm-up 'Pre-trained Model' with a 'Final Model' after 17 EM iterations on masked/noisy SST fields, reporting 27–28% RMSE reductions for reconstruction and one-step forecasting (Tables 1–2); a qualitative FY-3D Arctic swath case is also presented. The paper argues that sparse observations can directly improve the model's representation of ocean-state dynamics.

Significance. If the attribution claim holds, the work is a meaningful step toward reducing reliance on complete reanalysis products and toward learning observation-consistent dynamics from satellite swaths. The paper has strengths worth acknowledging: a clear probabilistic state-space formulation, an explicit EM derivation in Appendix A, a publicly released code repository with license and Dockerfile, a simulated benchmark with independent ground truth not used in the EM loop, and a real-data track. The central numerical improvements are plausible, but the current design does not isolate the effect of incompleteness from additional training volume, and the theoretical EM guarantee is invoked in a regime (M=1, approximate latent-space Langevin sampling) it does not cover. The real-data results are only qualitative. These issues are addressable within the manuscript's scope.

major comments (4)
  1. [§3.2.2, Tables 1–2] The comparison conflates learning from incomplete observations with additional training volume. The Pre-trained Model is a checkpoint after warm-up on ~800 complete fields; the Final Model is the result of 17 additional EM iterations on ~1,500 incomplete samples. Thus Tables 1–2 cannot attribute the 27–28% RMSE improvements to incompleteness; they could equally reflect larger training volume, more optimization steps, or self-distillation. Please add a complete-data control: run the same 17-iteration EM pipeline on ~1,500 complete, noise-free fields (or fine-tune the Pre-trained Model on the same number of complete fields) and report the same metrics for that checkpoint. Without this control, the central claim that sparse observations themselves improve dynamics is not established.
  2. [§2.5, Appendix A (A30.1)] The EM monotonicity argument requires q to be the exact posterior and the M-step expectation to be accurate. The implementation replaces q with a single latent-space Langevin sample (M=1), as the appendix explicitly acknowledges, and uses an approximate score through a pretrained generator. The reconstructed states used as M-step targets are therefore biased by the current model's prior, so the observed gains may be a self-distillation/regularization effect rather than extraction of observation-consistent dynamics. This is a load-bearing assumption. Please provide convergence diagnostics for the Langevin sampler (e.g., energy traces, multiple chains), sensitivity of results to K, η and M, and a check of E-step reconstruction quality against held-out observations.
  3. [Tables 1–2, §3.3.3] All reported metrics are point estimates: a single training run, a single set of 17 EM iterations, and averages over five masks with no variance or significance measure. Given that the central comparison is quantitative, per-mask breakdowns, at least 3 independent training seeds, and standard deviations (or confidence intervals) are needed to confirm that the 27–28% differences are outside the noise of GAN training and Langevin sampling.
  4. [§3.4] The real-world FY-3D track is qualitative only. The claims of structural fidelity, spatiotemporal coherence, and physical plausibility rest on visual inspection, with no ground truth and no quantitative metric. A quantitative proxy is needed—for example, withheld-swath observation likelihood, forecast skill against later orbits, or comparison with a standard interpolation or persistence baseline. As written, the real-data generalization claim is not supported beyond inspection.
minor comments (4)
  1. [§4.4, Acknowledgements] Typo: 'dge' should be 'acknowledge' in both 'we dge certain limitations' and 'We dge the World Climate Research Programme'.
  2. [§5.1? (Fig. 1, §2.5.3)] Figure 1 lacks color bars/axes, making the visual contrast between uniform and Gaussian noise priors difficult to judge; a quantitative measure (e.g., gradient-norm statistics) would strengthen the claim.
  3. [Eq. (5)–(6), §2.5] The relationship between z and s is implicit in the score decomposition. State explicitly that s_t = G(z) and define the composite latent Z (initial noise plus layer-wise injection noise) before Eq. (5), since the entire E-step derivation depends on this mapping.
  4. [Tables 1–2] The s1 RMSE for the Pre-trained Model differs notably between Table 1 (0.7587) and Table 2 (0.9263). The two tasks condition on different information (Table 1 uses o1 observations; Table 2 forecasts from perfect s0 without o1). The text should explicitly note this difference to avoid apparent inconsistency.

Circularity Check

0 steps flagged

No significant circularity: EM bootstrap is explicit and validated against external ground truth.

full rationale

The paper's training loop is self-referential in the generic EM sense—the E-step uses the current model to impute states and the M-step trains on those imputations—but this is not a definitional circularity: the optimization objective is the marginal observation likelihood pθ(o) (Eq. 1, Appendix A), and the E-step samples an auxiliary posterior q=pθ(n)(s|o) (Eq. A9) rather than defining the target as the model's own output. The paper even names the potential loop: “a dilemma of circular dependency between ‘state reconstruction’ and ‘model learning’,” and treats it with standard EM (Dempster et al., 1977). The central empirical claims are checked against simulated CMIP6 ground-truth fields that are not used as EM training targets (Tables 1–2), so the reported RMSE/MAE/SSIM gains are not forced by construction. No self-citation chain is load-bearing: the derivation cites standard EM/MCEM references, and the method's key claims do not reduce to an overlapping-author citation. The absence of a control that holds training volume/compute fixed between the Pre-trained and Final checkpoints is an experimental-attribution gap, not an equation-level reduction of the prediction to its inputs. The theoretical guarantee in Appendix A assumes q equals the exact posterior and M≥1 samples, whereas the implementation uses M=1 approximate Langevin samples; this is a fidelity/approximation concern, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method's central claim depends on a handful of hand-set hyperparameters and domain assumptions. The largest burden is that latent Langevin sampling with one Monte Carlo sample is treated as exact posterior inference, and that Markovian length-two dynamics captures the relevant ocean evolution.

free parameters (4)
  • observation noise variance σ² = not stated
    Weights the data-fidelity term in the E-step energy (Eq. 3, Algorithm 1 line 9); too large ignores observations, too small overfits noise.
  • latent prior weight Λ = {λ_z} = not stated
    Balances observation likelihood against the Gaussian latent prior in Algorithm 1 line 9; hand-set, affects reconstruction fidelity.
  • loss-balancing weights λ_pl, λ_R1, λ = not stated
    Weights for path length, R1, and L1 penalties in Eqs. 7–9; chosen empirically, not reported.
  • Langevin sampling hyperparameters K and η = not stated
    Number of Langevin steps and step size in Algorithm 1; both control E-step posterior approximation quality.
axioms (4)
  • domain assumption Ocean-state evolution is stationary, ergodic, and first-order Markovian; length-two sequences T=1 suffice for learning dynamics.
    Stated in the abstract and §2.1/§3.2.3; if daily/1.72-hour SST evolution is not Markovian or not stationary at the scales studied, the EM objective is misspecified.
  • ad hoc to paper Latent-space Langevin sampling approximates the true posterior pθ(s|o).
    Invoked in §2.5; the paper provides no calibration check that the generator's latent manifold covers the posterior, and convergence is assumed.
  • ad hoc to paper A single Monte Carlo sample (M=1) suffices for the E-step expectation.
    Appendix A3 notes 'satisfactory results can be achieved even when M = 1' without an analysis of estimator variance; this weakens the theoretical EM guarantee.
  • domain assumption Observations are independent masked pixel measurements with additive Gaussian noise of known variance.
    Eq. 3 defines p(ot|st)=N(ot|Mt⊙st, σ²I); correlated satellite errors or unknown masks would violate the likelihood used in the E-step.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Incomplete Observations Boost Evolutionary Performance in Ocean Modeling." pith.science (2026). https://pith.science/paper/T67RKYN5

@misc{pith2026260719147,
  author       = {Pith},
  title        = {Pith review of: Incomplete Observations Boost Evolutionary Performance in Ocean Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T67RKYN5}},
  note         = {Machine review of arXiv:2607.19147}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Data-driven methods have revolutionized ocean modeling, yet current approaches rely heavily on complete reanalysis datasets, imposing computational constraints and limiting model performance to that of the training data. Here, we present a generative state-space model and an optimization framework that enable learning directly from sparse and noisy observations. The model is essentially a hidden Markov model with a continuous state space, where oceanic physical quantities are treated as hidden states and measurements as observations, enabling a unified representation of ocean fields and observational data. Both the initial-state and state-transition modules are implemented as neural networks to capture the complexity and temporal evolution of ocean states, while the emission module is formulated as a masked Gaussian distribution. To train the model from sparse observations, we derive an optimization framework based on the expectation-maximization (EM) algorithm. The framework alternately reconstructs high-fidelity ocean fields via Langevin dynamics and optimizes deep neural networks to capture temporal evolution. Theoretical analysis shows that the framework maximizes the likelihood of observations under the generative model. For efficiency, we assume that ocean-state evolution follows a stationary, ergodic, and Markovian stochastic process and adopt only length-two state sequences during optimization. Experiments on CMIP6 simulation data and FY-3D satellite data demonstrate high-fidelity reconstruction and accurate prediction, showing that sparse observations can directly improve the model's representation of ocean-state dynamics. This work offers a scalable pathway for next-generation Earth system models to learn directly from sparse, incomplete real-world observations.

Figures

Figures reproduced from arXiv: 2607.19147 by Feng Gao, Junyu Dong, Xiaopei Lin, Yangyang Kong, Yanhai Gan, Yutong Jiang.

Figure 1
Figure 1. Figure 1: Comparison of spatial continuity under different noise priors. a, Reconstruc [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic of the generative state-space modeling framework. a, Probabilistic graphi￾cal model illustrating the hidden Markov process. Latent complete physical states (S) evolve tem￾porally and emit partial observations (O) via the observation model. b, c, Neural architectures for the generator components. b, The Initial State Model (pθ(s0)) utilizes a StyleGAN2-ADA backbone with modulated depthwise separab… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on representative simulation samples. Columns from [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Robust spatiotemporal reconstruction from sparse satellite swaths. The figure com￾pares raw observations, pre-trained baseline, and final model results across consecutive time steps for three samples. The framework ensures temporal continuity and recovers fine-scale structures. –17– [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

99 extracted references · 6 canonical work pages

  1. [1]

    , Schiller, A

    Bell2015OM APACrefauthors Bell, M J. , Schiller, A. , Le Traon , P Y. , Smith, N R. , Dombrowsky, E. \ Wilmer-Becker, K. APACrefauthors \ 2015 . An introduction to GODAE OceanView An introduction to GODAE OceanView . Journal of Operational Oceanography 8 S1 s2--s11 . APACrefURL https://doi.org/10.1080/1755876X.2015.1022041 APACrefURL APACrefDOI doi:10.108...

    arXiv 2015

  2. [2]

    APACrefauthors \ 2002

    Bennett2002 APACrefauthors Bennett, A F. APACrefauthors \ 2002 . Inverse Modeling of the Ocean and Atmosphere Inverse modeling of the ocean and atmosphere . Cambridge University Press . APACrefDOI doi:10.1017/CBO9780511535895 APACrefDOI

    work page doi:10.1017/cbo9780511535895 2002

  3. [3]

    , Xie, L

    Bi2023Nature APACrefauthors Bi, K. , Xie, L. , Zhang, H. , Chen, X. , Gu, X. \ Tian, Q. APACrefauthors \ 2023 . Accurate medium-range global weather forecasting with 3D neural networks Accurate medium-range global weather forecasting with 3d neural networks . Nature 619 533--538 . APACrefDOI doi:10.1038/s41586-023-06185-3 APACrefDOI

    doi:10.1038/s41586-023-06185-3 2023

  4. [4]

    , Brajard, J

    Bocquet2019 APACrefauthors Bocquet, M. , Brajard, J. , Carrassi, A. \ Bertino, L. APACrefauthors \ 2019 . Data assimilation as a learning tool to infer ordinary differential equation representations of dynamical models Data assimilation as a learning tool to infer ordinary differential equation representations of dynamical models . Nonlinear Processes in ...

    doi:10.5194/npg-26-143-2019 2019

  5. [5]

    , Carrassi, A

    Brajard2020JCS APACrefauthors Brajard, J. , Carrassi, A. , Bocquet, M. \ Bertino, L. APACrefauthors \ 2020 . Combining data assimilation and machine learning to emulate a dynamical model from sparse and noisy observations: A case study with the Lorenz 96 model Combining data assimilation and machine learning to emulate a dynamical model from sparse and no...

    arXiv 2020

  6. [7]

    APACrefauthors \ 2017

    Chollet2017 APACrefauthors Chollet, F. APACrefauthors \ 2017 . Xception: Deep Learning with Depthwise Separable Convolutions Xception: Deep learning with depthwise separable convolutions . Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Proceedings of the ieee conference on computer vision and pattern recognition (cvpr...

    doi:10.1109/cvpr.2017.195 2017

  7. [8]

    , Laird, N M

    Dempster1977EM APACrefauthors Dempster, A P. , Laird, N M. \ Rubin, D B. APACrefauthors \ 1977 . Maximum likelihood from incomplete data via the EM algorithm Maximum likelihood from incomplete data via the EM algorithm . Journal of the Royal Statistical Society: Series B (Methodological) 39 1 1--22 . APACrefDOI doi:10.1111/j.2517-6161.1977.tb01600.x APACrefDOI

    arXiv 1977

  8. [9]

    APACrefauthors \ 2009

    Evensen2009 APACrefauthors Evensen, G. APACrefauthors \ 2009 . Data Assimilation: The Ensemble Kalman Filter Data assimilation: The ensemble kalman filter . Springer Science & Business Media . APACrefDOI doi:10.1007/978-3-642-03711-5 APACrefDOI

    doi:10.1007/978-3-642-03711-5 2009

  9. [10]

    , Beauchamp, L

    Fablet2021 APACrefauthors Fablet, R. , Beauchamp, L. , Drumetz, L. \ Rousseau, F. APACrefauthors \ 2021 . Joint interpolation and representation learning for irregular sampled satellite-derived geophysical fields Joint interpolation and representation learning for irregular sampled satellite-derived geophysical fields . Frontiers in Applied Mathematics an...

    arXiv 2021

  10. [11]

    , Tan, Z

    Feng2025JGRMLC APACrefauthors Feng, D. , Tan, Z. , Lin, Z. , Xu, D. , Yu, C W. , He, Q. \ . APACrefauthors \ 2025 . A Comparative Study of Physics-Informed and Data-Driven Neural Networks for Compound Flood Simulation at River-Ocean Interfaces: A Case Study of Hurricane Irene A comparative study of physics-informed and data-driven neural networks for comp...

    2025

  11. [12]

    , Campin, J M

    Forget2015ECCO APACrefauthors Forget, G. , Campin, J M. , Heimbach, P. , Hill, C N. , Ponte, R M. \ Wunsch, C. APACrefauthors \ 2015 . ECCO version 4: an integrated framework for non-linear inverse modeling and global ocean state estimation Ecco version 4: an integrated framework for non-linear inverse modeling and global ocean state estimation . Geoscien...

    work page doi:10.5194/gmd-8-3071-2015 2015

  12. [13]

    \ Roweis, S T

    Ghahramani1999 APACrefauthors Ghahramani, Z. \ Roweis, S T. APACrefauthors \ 1998 . Learning nonlinear dynamical systems using an EM algorithm Learning nonlinear dynamical systems using an em algorithm . Advances in Neural Information Processing Systems (NeurIPS) Advances in neural information processing systems (neurips) \ ( 11). APACrefURL https://proce...

    1998

  13. [14]

    \ Malanotte-Rizzoli, P

    Ghil1991 APACrefauthors Ghil, M. \ Malanotte-Rizzoli, P. APACrefauthors \ 1991 . Data assimilation in meteorology and oceanography Data assimilation in meteorology and oceanography . Advances in Geophysics 33 141--266 . APACrefDOI doi:10.1016/S0065-2687(08)60442-2 APACrefDOI

    doi:10.1016/s0065-2687(08)60442-2 1991

  14. [15]

    , Pouget-Abadie, J

    goodfellow2014generative APACrefauthors Goodfellow, I J. , Pouget-Abadie, J. , Mirza, M. , Xu, B. , Warde-Farley, D. , Ozair, S. Bengio, Y. APACrefauthors \ 2014 . Generative adversarial nets Generative adversarial nets . Advances in Neural Information Processing Systems (NeurIPS) Advances in neural information processing systems (neurips) \ ( 27). APACre...

    2014

  15. [16]

    , Danabasoglu, G

    Griffies2016GMD APACrefauthors Griffies, S M. , Danabasoglu, G. , Durack, P J. \ . APACrefauthors \ 2016 . OMIP contribution to CMIP6: experimental and diagnostic protocol for the physical component of the Ocean Model Intercomparison Project Omip contribution to cmip6: experimental and diagnostic protocol for the physical component of the ocean model inte...

    work page doi:10.5194/gmd-9-3231-2016 2016

  16. [17]

    , Bell, B

    Hersbach2020QJRMS APACrefauthors Hersbach, H. , Bell, B. , Berrisford, P. , Hirahara, S. , Hor \'a nyi, A. , Mu \ n oz-Sabater, J. \ . APACrefauthors \ 2020 . The ERA5 global reanalysis The era5 global reanalysis . Quarterly Journal of the Royal Meteorological Society 146 730 1999--2049 . APACrefDOI doi:10.1002/qj.3803 APACrefDOI

    doi:10.1002/qj.3803 2020

  17. [18]

    , Zhu, J Y

    isola2017image APACrefauthors Isola, P. , Zhu, J Y. , Zhou, T. \ Efros, A A. APACrefauthors \ 2017 . Image-to-image translation with conditional adversarial networks Image-to-image translation with conditional adversarial networks . Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Proceedings of the ieee conference on c...

    doi:10.1109/cvpr.2017.632 2017

  18. [19]

    APACrefauthors \ 2003

    Kalnay2003 APACrefauthors Kalnay, E. APACrefauthors \ 2003 . Atmospheric modeling, data assimilation and predictability Atmospheric modeling, data assimilation and predictability . Cambridge University Press . APACrefDOI doi:10.1017/CBO9780511802270 APACrefDOI

    doi:10.1017/cbo9780511802270 2003

  19. [20]

    , Aittala, M

    Karras2020ada APACrefauthors Karras, T. , Aittala, M. , Laine, S. , H \"a rk \"o nen, E. , Hellsten, J. , Lehtinen, J. \ Aila, T. APACrefauthors \ 2020 . Training generative adversarial networks with limited data Training generative adversarial networks with limited data . Advances in Neural Information Processing Systems (NeurIPS) Advances in neural info...

    2020

  20. [21]

    , Laine, S

    StyleGAN APACrefauthors Karras, T. , Laine, S. \ Aila, T. APACrefauthors \ 2019 . A style-based generator architecture for generative adversarial networks A style-based generator architecture for generative adversarial networks . Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Proceedings of the ieee/cvf conference...

    arXiv 2019

  21. [22]

    , Laine, S

    karras2020analyzing APACrefauthors Karras, T. , Laine, S. , Aittala, M. , Hellsten, J. , Lehtinen, J. \ Aila, T. APACrefauthors \ 2020 . Analyzing and improving the image quality of StyleGAN Analyzing and improving the image quality of StyleGAN . Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Proceedings of the ie...

    arXiv 2020

  22. [23]

    \ Welling, M

    kingma2013auto APACrefauthors Kingma, D P. \ Welling, M. APACrefauthors \ 2013 . Auto-encoding variational B ayes Auto-encoding variational B ayes . arXiv preprint arXiv:1312.6114 . APACrefURL https://arxiv.org/abs/1312.6114 APACrefURL

    Pith/arXiv arXiv 2013

  23. [24]

    , Romera-Paredes, B

    Kohl2018ProbUNet APACrefauthors Kohl, S A A. , Romera-Paredes, B. , Meyer, C. , De Fauw, J. , Ledsam, J R. , Maier-Hein, K H. \ . APACrefauthors \ 2018 . A Probabilistic U-Net for Segmentation of Ambiguous Images A probabilistic U-Net for segmentation of ambiguous images . Advances in Neural Information Processing Systems (NeurIPS) Advances in neural info...

    Pith/arXiv arXiv 2018

  24. [25]

    , Sanchez-Gonzalez, A

    Lam2023Science APACrefauthors Lam, R. , Sanchez-Gonzalez, A. , Willson, M. , Wirnsberger, P. , Fortunato, M. , Alet, F. \ . APACrefauthors \ 2023 . Learning skillful medium-range global weather forecasting Learning skillful medium-range global weather forecasting . Science 382 6677 1416--1421 . APACrefDOI doi:10.1126/science.adi2336 APACrefDOI

    doi:10.1126/science.adi2336 2023

  25. [26]

    APACrefauthors \ 1986

    Lorenc1986 APACrefauthors Lorenc, A C. APACrefauthors \ 1986 . Analysis methods for numerical weather prediction Analysis methods for numerical weather prediction . Quarterly Journal of the Royal Meteorological Society 112 474 1177--1194 . APACrefDOI doi:10.1002/qj.49711247414 APACrefDOI

    doi:10.1002/qj.49711247414 1986

  26. [27]

    , Hannun, A Y

    maas2013rectifier APACrefauthors Maas, A L. , Hannun, A Y. \ Ng, A Y. APACrefauthors \ 2013 . Rectifier nonlinearities improve neural network acoustic models Rectifier nonlinearities improve neural network acoustic models . Proc. icml Proc. icml \ ( 30, 3). APACrefURL https://ai.stanford.edu/ amaas/papers/relu_hybrid_icml2013_final.pdf APACrefURL

    2013

  27. [29]

    APACrefauthors \ 2016

    McWilliams2016 APACrefauthors McWilliams, J C. APACrefauthors \ 2016 . Submesoscale currents in the ocean Submesoscale currents in the ocean . Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 472 2191 20160117 . APACrefDOI doi:10.1098/rspa.2016.0117 APACrefDOI

    arXiv 2016

  28. [31]

    \ Hinton, G E

    nair2010rectified APACrefauthors Nair, V. \ Hinton, G E. APACrefauthors \ 2010 . Rectified linear units improve restricted boltzmann machines Rectified linear units improve restricted boltzmann machines . Proceedings of the 27th International Conference on Machine Learning (ICML) Proceedings of the 27th international conference on machine learning (icml) ...

    2010

  29. [33]

    , Han, T

    Latent_EBM APACrefauthors Pang, B. , Han, T. , Nijkamp, E. , Zhu, S C. \ Wu, Y N. APACrefauthors \ 2021 . Learning Latent Space Energy-Based Models via MCMC Learning latent space energy-based models via MCMC . Proceedings of the 38th International Conference on Machine Learning (ICML) Proceedings of the 38th international conference on machine learning (i...

    2021

  30. [35]

    , Fillion, A

    Peyron2021 APACrefauthors Peyron, M. , Fillion, A. , G \"u rol, S. , Marchais, V. , Gratton, S. , Boudier, P. \ Goret, G. APACrefauthors \ 2021 09 . Latent space data assimilation by using deep learning Latent space data assimilation by using deep learning . Quarterly Journal of the Royal Meteorological Society 147 740 3759--3777 . APACrefURL https://doi....

    doi:10.1002/qj.4153 2021

  31. [36]

    , Perdikaris, P

    Raissi2019 APACrefauthors Raissi, M. , Perdikaris, P. \ Karniadakis, G E. APACrefauthors \ 2019 . Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involv...

    doi:10.1016/j.jcp.2018.10.045 2019

  32. [37]

    , Seeger, M W

    Deep_SSM APACrefauthors Rangapuram, S S. , Seeger, M W. , Gasthaus, J. , Stella, L. , Wang, Y. \ Januschowski, T. APACrefauthors \ 2018 . Deep state space models for time series forecasting Deep state space models for time series forecasting . Advances in Neural Information Processing Systems (NeurIPS) Advances in neural information processing systems (ne...

    2018

  33. [38]

    , Lenc, K

    Ravuri2021Nature APACrefauthors Ravuri, S. , Lenc, K. , Willson, M. , Kangin, D. , Lam, R. , Mirowski, P. \ . APACrefauthors \ 2021 . Skilful precipitation nowcasting using deep generative models of radar Skilful precipitation nowcasting using deep generative models of radar . Nature 597 672--677 . APACrefDOI doi:10.1038/s41586-021-03854-z APACrefDOI

    doi:10.1038/s41586-021-03854-z 2021

  34. [39]

    , Camps-Valls, G

    Reichstein2019Nature APACrefauthors Reichstein, M. , Camps-Valls, G. , Stevens, B. , Jung, M. , Denzler, J. , Carvalhais, N. \ Prabhat. APACrefauthors \ 2019 . Deep learning and process understanding for data-driven Earth system science Deep learning and process understanding for data-driven earth system science . Nature 566 195--204 . APACrefDOI doi:10.1...

    doi:10.1038/s41586-019-0912-1 2019

  35. [40]

    , Fischer, P

    Ronneberger2015 APACrefauthors Ronneberger, O. , Fischer, P. \ Brox, T. APACrefauthors \ 2015 . U-net: Convolutional networks for biomedical image segmentation U-net: Convolutional networks for biomedical image segmentation . Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) Proceedings of t...

    doi:10.1007/978- 2015

  36. [42]

    , Bathiany, S

    Runge2019NatComms APACrefauthors Runge, J. , Bathiany, S. , Bollt, E. , Camps-Valls, G. \ . APACrefauthors \ 2019 . Inferring causation from time series in Earth system sciences Inferring causation from time series in earth system sciences . Nature Communications 10 1 2553 . APACrefDOI doi:10.1038/s41467-019-10105-3 APACrefDOI

    doi:10.1038/s41467-019-10105-3 2019

  37. [43]

    salimans2022progressive APACrefauthors Salimans, T. \ Ho, J. APACrefauthors \ 2022 . Progressive distillation for fast sampling of diffusion models Progressive distillation for fast sampling of diffusion models . International Conference on Learning Representations (ICLR). International conference on learning representations (iclr). APACrefURL https://ope...

    2022

  38. [44]

    , Wunsch, C

    Sonnewald2019 APACrefauthors Sonnewald, M. , Wunsch, C. \ Heimbach, P. APACrefauthors \ 2019 . Unsupervised learning reveals geography of global ocean dynamical regions Unsupervised learning reveals geography of global ocean dynamical regions . Earth and Space Science 6 5 784--794 . APACrefDOI doi:10.1029/2018EA000519 APACrefDOI

    work page doi:10.1029/2018ea000519 2019

  39. [45]

    , Bovik, A C

    Wang2004 APACrefauthors Wang, Z. , Bovik, A C. , Sheikh, H R. \ Simoncelli, E P. APACrefauthors \ 2004 . Image quality assessment: from error visibility to structural similarity Image quality assessment: from error visibility to structural similarity . IEEE Transactions on Image Processing 13 4 600--612 . APACrefDOI doi:10.1109/TIP.2003.819861 APACrefDOI

    arXiv 2004

  40. [46]

    \ Tanner, M A

    MCEM_1990 APACrefauthors Wei, G C G. \ Tanner, M A. APACrefauthors \ 1990 . A Monte Carlo implementation of the EM algorithm and the poor man's data augmentation algorithms A monte carlo implementation of the em algorithm and the poor man's data augmentation algorithms . Journal of the American Statistical Association 85 411 699--704 . APACrefURL https://...

    arXiv 1990

  41. [47]

    \ Teh, Y W

    Welling2011 APACrefauthors Welling, M. \ Teh, Y W. APACrefauthors \ 2011 . Bayesian learning via stochastic gradient Langevin dynamics Bayesian learning via stochastic gradient langevin dynamics . Proceedings of the 28th International Conference on Machine Learning (ICML) Proceedings of the 28th international conference on machine learning (icml) \ ( \ 68...

    2011

  42. [48]

    APACrefauthors \ 2011

    Wilks2011 APACrefauthors Wilks, D S. APACrefauthors \ 2011 . Statistical Methods in the Atmospheric Sciences Statistical methods in the atmospheric sciences \ ( 3rd \ , 100). Amsterdam; Boston Academic Press . APACrefURL https://linkinghub.elsevier.com/retrieve/isbn/9780123850225 APACrefURL

    arXiv 2011

  43. [49]

    Wu2019BCC APACrefauthors Wu, T. , Lu, Y. , Fang, Y. , Xin, X. , Li, L. , Li, W. others APACrefauthors \ 2019 . The Beijing Climate Center Climate System Model (BCC-CSM): the main progress from CMIP5 to CMIP6 The beijing climate center climate system model (bcc-csm): the main progress from cmip5 to cmip6 . Geoscientific Model Development 12 4 1573--1600 . ...

    work page doi:10.5194/gmd-12-1573-2019 2019

  44. [50]

    , Zhou, Y

    Zhang2025JGRMLC APACrefauthors Zhang, X. , Zhou, Y. , Ning, C. \ Li, X. APACrefauthors \ 2025 . Physically Guided GAN-Based Super Resolution of Buoy Observations for Internal Solitary Waves Physically guided gan-based super resolution of buoy observations for internal solitary waves . Journal of Geophysical Research: Machine Learning and Computation 2 4 ....

    work page doi:10.1029/2025jh000951 2025

  45. [51]

    Nature , volume=

    Deep learning and process understanding for data-driven Earth system science , author=. Nature , volume=. 2019 , doi=

    2019

  46. [52]

    Nature , volume=

    Accurate medium-range global weather forecasting with 3D neural networks , author=. Nature , volume=. 2023 , doi=

    2023

  47. [53]

    Science , volume=

    Learning skillful medium-range global weather forecasting , author=. Science , volume=. 2023 , doi=

    2023

  48. [54]

    arXiv preprint arXiv:2304.02948 , year=

    FengWu: Pushing the Skillful Global Medium-range Weather Forecast out to 10 Days , author=. arXiv preprint arXiv:2304.02948 , year=

    arXiv

  49. [55]

    arXiv preprint arXiv:2202.11214 , year=

    FourCastNet: A Global Data-driven High-resolution Forecasting Model using Adaptive Fourier Neural Operators , author=. arXiv preprint arXiv:2202.11214 , year=

    Pith/arXiv arXiv

  50. [56]

    Geoscientific Model Development , volume=

    ECCO version 4: an integrated framework for non-linear inverse modeling and global ocean state estimation , author=. Geoscientific Model Development , volume=. 2015 , publisher=

    2015

  51. [57]

    Earth and Space Science , volume=

    Unsupervised learning reveals geography of global ocean dynamical regions , author=. Earth and Space Science , volume=. 2019 , publisher=

    2019

  52. [58]

    arXiv preprint arXiv:2301.10343 , year=

    ClimaX: A foundation model for weather and climate , author=. arXiv preprint arXiv:2301.10343 , year=

    Pith/arXiv arXiv

  53. [59]

    Quarterly Journal of the Royal Meteorological Society , volume=

    The ERA5 global reanalysis , author=. Quarterly Journal of the Royal Meteorological Society , volume=. 2020 , doi=

    2020

  54. [60]

    Wiley Interdisciplinary Reviews: Climate Change , volume=

    Data assimilation in the geosciences: An overview of methods, issues, and perspectives , author=. Wiley Interdisciplinary Reviews: Climate Change , volume=. 2018 , doi=

    2018

  55. [61]

    Combining data assimilation and machine learning to emulate a dynamical model from sparse and noisy observations: A case study with the

    Brajard, Julien and Carrassi, Alberto and Bocquet, Marc and Bertino, Laurent , journal=. Combining data assimilation and machine learning to emulate a dynamical model from sparse and noisy observations: A case study with the. 2020 , doi=

    2020

  56. [62]

    Journal of Geophysical Research: Machine Learning and Computation , volume=

    Physically Guided GAN-Based Super Resolution of Buoy Observations for Internal Solitary Waves , author=. Journal of Geophysical Research: Machine Learning and Computation , volume=. 2025 , publisher=

    2025

  57. [63]

    Journal of Geophysical Research: Machine Learning and Computation , volume=

    Simultaneous Emulation and Downscaling With Physically Consistent Deep Learning-Based Regional Ocean Emulators , author=. Journal of Geophysical Research: Machine Learning and Computation , volume=. 2025 , publisher=

    2025

  58. [64]

    Journal of Geophysical Research: Machine Learning and Computation , volume=

    A Comparative Study of Physics-Informed and Data-Driven Neural Networks for Compound Flood Simulation at River-Ocean Interfaces: A Case Study of Hurricane Irene , author=. Journal of Geophysical Research: Machine Learning and Computation , volume=. 2025 , publisher=

    2025

  59. [65]

    Journal of Advances in Modeling Earth Systems , volume=

    Benchmarking of Machine Learning Ocean Subgrid Parameterizations in an Idealized Model , author=. Journal of Advances in Modeling Earth Systems , volume=. 2023 , publisher=. doi:10.1029/2022MS003258 , url=

    doi:10.1029/2022ms003258 2023

  60. [66]

    Nature Reviews Physics , volume=

    Physics-informed machine learning , author=. Nature Reviews Physics , volume=. 2021 , doi=

    2021

  61. [67]

    Nature , volume=

    Skilful precipitation nowcasting using deep generative models of radar , author=. Nature , volume=. 2021 , doi=

    2021

  62. [68]

    Nature Communications , volume=

    Physics-informed learning of governing equations from scarce data , author=. Nature Communications , volume=. 2021 , doi=

    2021

  63. [69]

    Nature Communications , volume=

    Inferring causation from time series in Earth system sciences , author=. Nature Communications , volume=. 2019 , doi=

    2019

  64. [70]

    Advances in Neural Information Processing Systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in Neural Information Processing Systems , volume=. 2020 , url=

    2020

  65. [71]

    arXiv preprint arXiv:2309.15214 , year=

    Generative Residual Diffusion Modeling for Super-Resolution and Forecasting of High-Resolution Weather and Climate , author=. arXiv preprint arXiv:2309.15214 , year=

    Pith/arXiv arXiv

  66. [72]

    Advances in Geophysics , volume=

    Data assimilation in meteorology and oceanography , author=. Advances in Geophysics , volume=. 1991 , publisher=

    1991

  67. [73]

    Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) , pages=

    U-net: Convolutional networks for biomedical image segmentation , author=. Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) , pages=. 2015 , organization=

    2015

  68. [74]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Deep state space models for time series forecasting , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=. 2018 , url=

    2018

  69. [75]

    Journal of the American Statistical Association , volume =

    A Monte Carlo implementation of the EM algorithm and the poor man's data augmentation algorithms , author =. Journal of the American Statistical Association , volume =. 1990 , doi =

    1990

  70. [76]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Training generative adversarial networks with limited data , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=. 2020 , url=

    2020

  71. [77]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Image-to-image translation with conditional adversarial networks , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2017 , doi=

    2017

  72. [78]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Generative adversarial nets , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=. 2014 , url=

    2014

  73. [79]

    arXiv preprint arXiv:1411.1784 , year=

    Conditional generative adversarial nets , author=. arXiv preprint arXiv:1411.1784 , year=

    Pith/arXiv arXiv

  74. [80]

    Geoscientific Model Development , volume=

    OMIP contribution to CMIP6: experimental and diagnostic protocol for the physical component of the Ocean Model Intercomparison Project , author=. Geoscientific Model Development , volume=. 2016 , doi=

    2016

  75. [81]

    Geoscientific Model Development , volume=

    The Beijing Climate Center Climate System Model (BCC-CSM): the main progress from CMIP5 to CMIP6 , author=. Geoscientific Model Development , volume=. 2019 , doi=

    2019

  76. [82]

    IEEE Transactions on Image Processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE Transactions on Image Processing , volume=. 2004 , doi=

    2004

  77. [83]

    Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=

    Submesoscale currents in the ocean , author=. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=. 2016 , doi=

    2016

  78. [84]

    Auto-encoding variational

    Kingma, Diederik P and Welling, Max , journal=. Auto-encoding variational. 2013 , url=

    2013

  79. [85]

    International Conference on Learning Representations (ICLR) , year=

    Progressive distillation for fast sampling of diffusion models , author=. International Conference on Learning Representations (ICLR) , year=

  80. [86]

    Learning Latent Space Energy-Based Models via

    Pang, Bo and Han, Tian and Nijkamp, Erik and Zhu, Song-Chun and Wu, Ying Nian , booktitle=. Learning Latent Space Energy-Based Models via. 2021 , series=

    2021

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.