Pith. sign in

REVIEW 4 major objections 5 minor 21 references

ReconMOST: Multi-Layer Sea Temperature Reconstruction with Observations-Guided Diffusion

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A diffusion model pretrained on climate simulations reconstructs global multi-layer ocean temperature from extremely sparse observations, down to 7.5% of grid points, with total MSE 0.633.

desk verdict A plausible new application of guided diffusion to multi-layer ocean temperature reconstruction, but the headline MSE is measured on a self-consistent inpainting proxy, not real observations. read the letter →

arxiv 2506.10391 v1 pith:ULJ6MM7E submitted 2025-06-12 cs.CV

Yuanyi Song , Pumeng Lyu , Ben Fei , Fenghua Ling , Wanli Ouyang , Lei Bai This is my paper · ORCID
classification cs.CV
keywords oceantemperaturereconstructionguideddiffusionsparseobservationsmulti-layerseaDDPMCMIP6EN4generativeprior
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 sets out to show that a diffusion model can reconstruct the full global, multi-layer ocean temperature field from extremely sparse observations, with up to 92.5% of grid points missing. The proposed ReconMOST pipeline first trains an unconditional denoising diffusion model on historical CMIP6 numerical simulations, so the model internalizes physically consistent spatial patterns of temperature across 42 depth levels. At generation time, sparse observation points act as guides: the reverse diffusion trajectory is shifted by gradients of the distance between the current estimate and the observed values, with a Gaussian soft-extension spreading each constraint to nearby grid cells. The paper reports total mean squared error of 0.633 on EN4 analysis data when only 7.5% of grid points are given, and shows accuracy improving steadily as guidance density rises. If correct, this would move machine-learning-based ocean reconstruction from sea-surface and regional cases to global three-dimensional fields, a setting where conventional statistical methods struggle.

What carries the argument

The load-bearing object is a pretrained denoising diffusion probabilistic model (DDPM) with a U-Net backbone that treats the entire 42-layer ocean temperature field as a single image-like volume. During inference the model is never retrained on observations; instead, at each denoising step the predicted clean field $\hat{x}_0$ is compared with the sparse observation vector $y$ through a Euclidean distance, and the gradient of that distance is added to the reverse-sampling mean, scaled by a guidance strength $s$. This gradient field is then smoothed by the soft-extension module, a Gaussian kernel that diffuses each point constraint over neighboring grid cells, following the geographic principle that nearby points covary. Setting the diffusion variance $\Sigma$ to zero in the guided sampling step suppresses noise accumulation. Together the pretrained prior supplies physical plausibility, the gradient supplies fidelity to data, and soft-extension spreads sparse constraints into unobserved regions so local accuracy does not collapse.

What would settle it

Run the reconstruction with guidance taken only from real Argo float profiles and score against withheld Argo profiles and independent ship or glider measurements; if the total MSE rises far above the 0.633 reported for uniform grid-point sampling of EN4, the 92.5%-missing-data claim has not been demonstrated for true observations. An even more direct check is to rebuild the EN4 field with guidance masks that reproduce actual Argo float locations instead of uniform random masks and compare error at unobserved floats.

Watch

Extended reading notes

Core claim

The central claim is that a single unconditional DDPM pretrained on seven CMIP6 historical simulations carries enough physical prior about the ocean temperature field that, during reverse diffusion, sparse in-situ-like observations can guide the generation of a full 42-layer global field. On EN4 objective analysis data with 7.5% of grid points provided as guidance, the method achieves 0.049 MSE at guidance points, 0.680 MSE at unobserved reconstruction points, and 0.633 total MSE; with guidance raised to 40%, total MSE falls to 0.222. The accompanying generalization experiments show the framework transferring to EN4 after pretraining on CMIP6 only, with a single well-chosen CMIP6 model nearly matching the full seven-model ensemble, and ablations trace much of the gain to the soft-extension module, since removing it raises total MSE to 1.376.

Load-bearing premise

The evaluation assumes that randomly sampling 7.5% of grid points from the EN4 objective analysis, and scoring against that same EN4 analysis field, faithfully represents how the method will perform with real Argo and buoy observations, which are irregular, noisy, and not aligned to the analysis grid.

Editorial extensions

If this is right

  • Pretraining on climate-model simulations alone gives a reusable physical prior: new observational networks can be assimilated at inference time without retraining the diffusion model.
  • Reconstruction accuracy scales smoothly with observation density, so investments in denser buoy or Argo coverage translate directly into lower global error.
  • The same guided-generation recipe could be applied to other ocean variables such as salinity, currents, or biogeochemical tracers by retraining the prior on corresponding simulation data.
  • One well-chosen CMIP6 model can nearly match the full seven-model ensemble on EN4, meaning the pretraining cost could be cut substantially without much loss in reconstruction quality.
  • The method's ability to fill large coherent gaps suggests a path toward global subsurface reanalysis products where traditional optimal interpolation degrades.

Reading between the lines

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

  • The decisive unknown is distribution shift: real Argo data are noisy and irregular, while EN4 is a smooth objective analysis; a natural next test is to withhold actual Argo profiles and score against them rather than against the analysis grid.
  • Because the guidance mechanism is a form of inference-time data assimilation, a direct comparison with optimal interpolation or an ensemble Kalman filter under identical sparse masks would isolate how much of the gain comes from the diffusion prior versus the gradient scheme.
  • The April and August local errors noted in the temporal visualization suggest the purely spatial soft-extension misses seasonal vertical structure; a temporal or isopycnal-following extension may reduce those errors and is a natural next step.
  • The framework leaves $s$, kernel size, and observation density as manually tuned hyperparameters; an adaptive schedule relating them to local observation density could make the method more reliable in regions like the Southern Ocean where buoy coverage is heterogeneous.
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

4 major / 5 minor

Summary. The paper proposes ReconMOST, a guided diffusion framework for reconstructing global, multi-layer ocean temperature fields from sparse observations. A DDPM is pre-trained on CMIP6 historical simulations; at inference, sparse observations are used as gradient guidance during the reverse diffusion process, with a Gaussian soft-extension to spread the influence of observed points. Experiments on CMIP6 and EN4 analysis report guidance MSE of 0.049, reconstruction MSE of 0.680, and total MSE of 0.633 with a 7.5% guidance rate (92.5% missing data), together with ablations on guidance strength, soft-extension size, sampling variance, and observation density. The paper claims superior generalization to real-world sparse observation scenarios.

Significance. If the evaluation were independent, the method would be a meaningful contribution: it extends diffusion-based reconstruction from sea-surface or regional settings to a global 42-depth field, releases source code, reports systematic ablations, and shows intuitive scaling behavior with observation density. However, the current evaluation protocol cannot support the headline generalization claim. The training-mode subset is selected by test-set performance, the key hyperparameters are tuned on the same test set, and the guidance points are sampled from the same smooth EN4 analysis field used as ground truth. These issues make the reported MSE numbers optimistically biased and not a measurement of reconstruction skill under realistic sparse-observation conditions.

major comments (4)
  1. [Sec. 4.4, Table 2] The Top1/Top3/Top5 training-mode selections are ranked by their MSE on the EN4 test set itself, and the reported 'best' Top5 total MSE of 0.648 is then computed on that same test set. This is post-hoc selection on the evaluation data, so the headline generalization result is partly a fitted quantity. The authors should use a held-out time period or a nested validation split to obtain an unbiased estimate of the generalization error.
  2. [Sec. 4.6, Table 4] The guidance strength s, the choice of Sigma, and the soft-extension kernel size are ablated on the same EN4 test set used for the final evaluation; the S-5 configuration is selected because it yields total MSE 0.633, and the abstract's headline MSE values are exactly this selected configuration. Tuning hyperparameters on the test set inflates the reported accuracy. The authors should fix the hyperparameters on a validation set before reporting final test-set performance, or report results across a range of settings without selecting the best one for the headline.
  3. [Sec. 4.2, Sec. 4.4] The evaluation constructs guidance points by randomly sampling 7.5% of grid cells from the EN4 analysis field and uses the same EN4 analysis as ground truth. EN4 is a smooth objective-analysis product, so this protocol measures performance on an inpainting task over a smooth field, not on reconstruction from realistic Argo/XBT/buoy observations with irregular spatial coverage, depth-dependent sampling, measurement noise, and biases. Consequently, the 92.5% missing-data claim is demonstrated only for this synthetic mask. In addition, the reported MSE-g of 0.049 at guidance points is not a prediction error, because those point values are the conditioning information used during reconstruction; MSE-g should not be included as evidence of reconstruction accuracy.
  4. [Sec. 4.2] The assertion that a 7.5% uniform random guidance rate 'matches real-world buoy distribution densities' is not supported by any analysis of actual Argo or buoy observation geometry. A uniform 7.5% mask across all 42 depth layers does not reproduce the depth-dependent and geographically irregular sampling of real ocean profile networks. The authors should either provide a quantitative comparison with real observation distributions or temper the real-world generalization claim accordingly.
minor comments (5)
  1. [Table 4] In the S-5 row, '50.049' should read '0.049'; the missing space and leading zero make the table harder to read.
  2. [Appendix E, Table 5] The text and table caption refer to 'Top5 in Table 5' but should refer to Table 2; the U-Net comparison also uses a 'pre-train guided rate' that is not clearly defined for an unconditional diffusion pretraining phase.
  3. [Fig. 7] The caption of Figure 7 is a placeholder ('Enter Caption') and should be replaced with a descriptive caption.
  4. [Introduction] There is a typo in the first paragraph: 'V on Schuckmann et al.' should read 'von Schuckmann et al.'.
  5. [Sec. 3.2.2] The discussion of why no vertical extension is applied is useful, but the uniform layer-wise sampling of guidance points should be clarified: is the 7.5% rate applied per depth layer or as a fraction of the total grid across all layers? This matters for interpreting how the experiments relate to real depth-dependent Argo sampling.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline MSE is partly circular: guidance-point error is a fitting residual by construction, and the best model/hyperparameters are selected on the EN4 test set whose MSE is then reported as generalization.

  1. fitted input called prediction [Sec. 3.2.1 (Eqs. 6-7), Sec. 4.2, Table 2 caption]
    "At each reconstructing(denoising) step from xt to xt−1, we steer the distribution of xt−1 towards the actual sea temperature distribution associated with the observed values y. This is achieved by computing the discrepancy between the reconstructed ˆx0 from xt and the observations y, and using the resulting gradient to adjust the reconstructing direction of xt ... The final performance is evaluated by averaging the MSE over four reconstruction results with ground truth temperature. ..."

    The observed values y are the conditioning inputs of the reverse diffusion. The gradient guidance explicitly minimizes L(ˆx0(xt), y) at those points, so MSE-g is the residual of the fitting objective itself, not an independent prediction of unseen data. Folding this quantity into the reported 'total MSE 0.633' makes the headline partly a measure of how well the output matches the very inputs it was given. The unobserved-point MSE-r is a genuine prediction; the circularity is limited to MSE-g and total MSE.

  2. fitted input called prediction [Sec. 4.4, Table 2 and Table 4 (S-5)]
    "we selected the Top1, Top3, and Top5 modes (ranked by MSE performance), along with the complete seven-mode dataset, to retrain DDPM models. This strategy balances generalization capacity with training data volume to capture more diverse distribution patterns. The results in Table 2 show that the Top5 model achieves the best performance of 0.648 ... S-5 4 0 7.5%(3188) 5 0.049 0.680 0.633"

    The mode ranking used to choose Top5 is computed from the EN4 analysis test set itself ('We applied the seven models ... to the EN4 analysis dataset, using global average MSE as the evaluation metric'), and the final S-5 configuration is the minimum total MSE in the ablation table on that same test set. The reported 0.633 is therefore a minimum over configurations evaluated on the test data, not an independent estimate of generalization.

full rationale

The pre-training and the unobserved-point reconstruction are independent content: a DDPM is trained on CMIP6 fields and MSE-r measures error at points that were not provided as guidance, so the core inpainting claim is not definitionally forced. No load-bearing self-citation chain is present: citations to Fei et al. and Tu et al. support standard guided-diffusion implementation choices, not a uniqueness or equivalence argument. The EN4-analysis-as-ground-truth proxy is a benchmark-validity limitation rather than a circularity under the stated rules, so it is noted but not scored here. The circularity score of 6 reflects two concrete reductions: (1) MSE-g is by construction the residual of the gradient-fitted guidance objective and is included in the headline total MSE; (2) the Top5 model choice and the S-5 hyperparameters are selected by their MSE on the EN4 test set, and that same test-set MSE is then presented as superior generalization. Both are partial: the reported MSE-r at unobserved points remains a genuine prediction, preventing a higher score.

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

The method introduces no new physical entities. It relies on a set of tuned hyperparameters (s, kernel size, Sigma, guided_rate) and on domain assumptions about spatial autocorrelation, the validity of EN4 as ground truth, and the transferability of CMIP6 priors to the real ocean.

free parameters (4)
  • guidance strength s = 5 (ablation over 3,4,5,6)
    Controls strength of the observation gradient. Chosen in Sec. 4.6 to minimize total MSE on the EN4 test set.
  • soft-extension kernel size = 5 (ablation over 3,5,7,9)
    Controls spatial diffusion of observation gradients. Chosen in Sec. 4.6 to minimize total MSE on the EN4 test set.
  • diffusion sampling variance Sigma = 0 in final configuration
    Deterministic reverse sampling chosen by ablation; improves accuracy by avoiding noise accumulation.
  • guided_rate (fraction of observed grid points) = 0.075
    Set to approximate real-world buoy density; not fitted to error but affects all reported results.
assumptions (4)
  • domain assumption Tobler's first law of geography: nearby spatial values are more strongly correlated
    Invoked in Sec. 3.2.2 to justify Soft-Extension of observation gradients to neighboring grid points.
  • domain assumption EN4 objective analysis can serve as ground truth for ocean temperature reconstruction
    Sec. 4.1 and 4.2 use EN4 analysis fields as ground truth and sample guidance points from the same field, assuming this represents real-world sparse observations.
  • domain assumption Pre-training on CMIP6 historical simulations provides a physically consistent prior transferable to real ocean states
    Sec. 3.1 assumes that distributions learned from numerical model simulations capture real ocean physics.
  • standard math Low curvature of log p(y|x_t) compared to Sigma^-1, enabling Taylor expansion of the conditioner
    Appendix B.2 uses this common assumption from the guided diffusion literature (Dhariwal and Nichol, 2021).

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReconMOST: Multi-Layer Sea Temperature Reconstruction with Observations-Guided Diffusion." pith.science (2026). https://pith.science/paper/ULJ6MM7E

@misc{pith2026250610391,
  author       = {Pith},
  title        = {Pith review of: ReconMOST: Multi-Layer Sea Temperature Reconstruction with Observations-Guided Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ULJ6MM7E}},
  note         = {Machine review of arXiv:2506.10391}
}
read the original abstract

Accurate reconstruction of ocean is essential for reflecting global climate dynamics and supporting marine meteorological research. Conventional methods face challenges due to sparse data, algorithmic complexity, and high computational costs, while increasing usage of machine learning (ML) method remains limited to reconstruction problems at the sea surface and local regions, struggling with issues like cloud occlusion. To address these limitations, this paper proposes ReconMOST, a data-driven guided diffusion model framework for multi-layer sea temperature reconstruction. Specifically, we first pre-train an unconditional diffusion model using a large collection of historical numerical simulation data, enabling the model to attain physically consistent distribution patterns of ocean temperature fields. During the generation phase, sparse yet high-accuracy in-situ observational data are utilized as guidance points for the reverse diffusion process, generating accurate reconstruction results. Importantly, in regions lacking direct observational data, the physically consistent spatial distribution patterns learned during pre-training enable implicitly guided and physically plausible reconstructions. Our method extends ML-based SST reconstruction to a global, multi-layer setting, handling over 92.5% missing data while maintaining reconstruction accuracy, spatial resolution, and superior generalization capability. We pre-train our model on CMIP6 numerical simulation data and conduct guided reconstruction experiments on CMIP6 and EN4 analysis data. The results of mean squared error (MSE) values achieve 0.049 on guidance, 0.680 on reconstruction, and 0.633 on total, respectively, demonstrating the effectiveness and robustness of the proposed framework. Our source code is available at https://github.com/norsheep/ReconMOST.

Figures

Figures reproduced from arXiv: 2506.10391 by the authors.

Figure 1
Figure 1. Difference between traditional methods and our ReconMOST. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The framework of ReconMOST: In the training phase, we pre-train an unconditional diffusion model using a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Demonstration of Soft-Extension Given the extreme sparsity of current buoy observations, di￾rectly applying observation-derived gradients results in insuf￾ficient constraints for unobserved regions. Inspired by Tobler’s first law of geography, which states that all spatial variables are related, with nearby values being more strongly correlated [To￾bler, 1970, Huang et al., 2024], we introduce a Soft-Extension strat… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of SST reconstruction of EN4 dataset in ground truth, reconstruction results, and difference: [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison of increasing observation density with no soft extension, separately 7.5%, 10%, 20%, 30%, 40% [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Comparison of complete method(right) and method with no soft extension at dense observation density(left), [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Enter Caption 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 13 canonical work pages

  1. [5]

    The last three columns are, respectively, average MSE of upper ten levels on guidance points (MSE-g), reconstruction points (MSE-r), and all complete global points (total MSE)

    Guided rate during pre-training and number of training steps are displayed. The last three columns are, respectively, average MSE of upper ten levels on guidance points (MSE-g), reconstruction points (MSE-r), and all complete global points (total MSE). pre-train guided rate guided rate # training steps MSE-g MSE-r total MSE 7.5% 10% 200,000 7.975 0.870 7....

    work page 2020

  2. [6]

    GetMesh: A Controllable Model for High-quality Mesh Generation and Manipulation

    Zhaoyang Lyu, Ben Fei, Jinyi Wang, Xudong Xu, Ya Zhang, Weidong Yang, and Bo Dai. Getmesh: A controllable model for high-quality mesh generation and manipulation.arXiv preprint arXiv:2403.11990,

    work page Pith review arXiv

  3. [7]

    Guided diffusion model for adversarial purification

    Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided diffusion model for adversarial purification. arXiv preprint arXiv:2205.14969,

    arXiv

  4. [9]

    The era5 atmospheric reanalysis

    Hans Hersbach. The era5 atmospheric reanalysis. InAGU fall meeting abstracts, volume 2016, pages NG33D–01,

    work page 2016

  5. [11]

    Weather prediction with diffusion guided by realistic forecast processes.arXiv preprint arXiv:2402.06666,

    Zhanxiang Hua, Yutong He, Chengqian Ma, and Alexandra Anderson-Frey. Weather prediction with diffusion guided by realistic forecast processes.arXiv preprint arXiv:2402.06666,

    arXiv

  6. [12]

    Deterministic guidance diffusion model for probabilistic weather forecasting.arXiv preprint arXiv:2312.02819,

    Donggeun Yoon, Minseok Seo, Doyi Kim, Yeji Choi, and Donghyeon Cho. Deterministic guidance diffusion model for probabilistic weather forecasting.arXiv preprint arXiv:2312.02819,

    arXiv

  7. [13]

    Accessed: 2025-04-21

    URL https://oceanexplorer.noaa.gov/facts/temp-vary.html. Accessed: 2025-04-21. National Oceanic and Atmospheric Administration. Ocean temperature patterns are strongly controlled by ocean forces on a decadal timescale, 2024a. URL https://www.climate.gov/news-features/feed/ ocean-temperature-patterns-are-strongly-controlled-ocean-forces-decadal . Accessed:...

    work page 2025

  8. [16]

    Overview of the coupled model intercomparison project phase 6 (cmip6) experimental design and organization

    Veronika Eyring, Sandrine Bony, Gerald A Meehl, Catherine A Senior, Bjorn Stevens, Ronald J Stouffer, and Karl E Taylor. Overview of the coupled model intercomparison project phase 6 (cmip6) experimental design and organization. Geoscientific Model Development, 9(5):1937–1958,

    work page 1937

Show all 21 references
  1. [17]

    URL https://agupubs.onlinelibrary

    doi:https://doi.org/10.1002/2013JC009067. URL https://agupubs.onlinelibrary. wiley.com/doi/abs/10.1002/2013JC009067. Gustavo J Goni, Janet Sprintall, Francis Bringas, Lijing Cheng, Mauro Cirano, Shenfu Dong, Ricardo Domingues, Marlos Goes, Hosmay Lopez, Rosemary Morrow, et al....

    doi

  2. [19]

    DDPM introduces a latent variable sequence {x1,

    11 arXivTemplateA PREPRINT A Derivation of Denoising Diffusion Probabilistic Models Consider a data distribution q(x0) over continuous data x0 ∈R d. DDPM introduces a latent variable sequence {x1, . . . , xT }by defining a Markov chain that gradually adds Gaussian noise to the...

    2020

  3. [20]

    These values correspond to upper and lower bounds for the true reverse step variance

    find that instead of learning Σθ(xt, t), they can fix it to a constant, choosing either βtI or ˜βtI. These values correspond to upper and lower bounds for the true reverse step variance. 12 arXivTemplateA PREPRINT B Observation Guidance Here, we supplement the exploitation of ...

    2021

  4. [1970]

    Diffda: a diffusion model for weather-scale data assimilation.arXiv preprint arXiv:2401.05932,

    Langwen Huang, Lukas Gianinazzi, Yuejiang Yu, Peter D Dueben, and Torsten Hoefler. Diffda: a diffusion model for weather-scale data assimilation.arXiv preprint arXiv:2401.05932,

    arXiv

  5. [1992]

    Satellite observations guided diffusion model for accurate meteorological states at arbitrary resolution.arXiv preprint arXiv:2502.07814,

    Siwei Tu, Ben Fei, Weidong Yang, Fenghua Ling, Hao Chen, Zili Liu, Kun Chen, Hang Fan, Wanli Ouyang, and Lei Bai. Satellite observations guided diffusion model for accurate meteorological states at arbitrary resolution.arXiv preprint arXiv:2502.07814,

    arXiv

  6. [2012]

    Predicting basin and landfalling hurricane numbers from sea surface temperature.arXiv preprint physics/0701170,

    Stephen Jewson, Roman Binter, Shree Khare, Kechi Nzerem, and Adam O’Shay. Predicting basin and landfalling hurricane numbers from sea surface temperature.arXiv preprint physics/0701170,

    arXiv

  7. [2013]

    A reanalysis of north pacific sea surface temperatures using state-space techniques: The pdo in a new light.arXiv preprint arXiv:1204.2486,

    Roy Mendelssohn and CIndy Bessey. A reanalysis of north pacific sea surface temperatures using state-space techniques: The pdo in a new light.arXiv preprint arXiv:1204.2486,

    arXiv

  8. [2015]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598,

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598,

    arXiv

  9. [2018]

    Generating unseen nonlinear evolution in sea surface temperature using a deep learning-based latent space data assimilation framework.arXiv preprint arXiv:2412.13477,

    Qingyu Zheng, Guijun Han, Wei Li, Lige Cao, Gongfu Zhou, Haowen Wu, Qi Shao, Ru Wang, Xiaobo Wu, Xudong Cui, et al. Generating unseen nonlinear evolution in sea surface temperature using a deep learning-based latent space data assimilation framework.arXiv preprint arXiv:2412.13477,

    arXiv

  10. [2020]

    doi:10.1145/3400066

    ISSN 2157-6904. doi:10.1145/3400066. URLhttps://doi.org/10.1145/3400066. Rachel Furner, Peter Haynes, Dave Munday, Brooks Paige, Daniel C Jones, and Emily Shuckburgh. A sensitivity analysis of a regression model of ocean temperature.Environmental Data Science, 1:e11,

    doi

  11. [2022]

    Sea surface temperature prediction and reconstruction using patch- level neural network representations

    Said Ouala, Cédric Herzet, and Ronan Fablet. Sea surface temperature prediction and reconstruction using patch- level neural network representations. InIGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 5628–5631. IEEE,

    2018

  12. [2024]

    Extended reconstruction of global sea surface temperatures based on coads data (1854–1997).Journal of Climate, 16(10):1495–1510,

    Thomas M Smith and Richard W Reynolds. Extended reconstruction of global sea surface temperatures based on coads data (1854–1997).Journal of Climate, 16(10):1495–1510,

    1997

  13. [2025]

    Context-guided diffusion for out-of-distribution molecular and protein design.arXiv preprint arXiv:2407.11942,

    Leo Klarner, Tim GJ Rudner, Garrett M Morris, Charlotte M Deane, and Yee Whye Teh. Context-guided diffusion for out-of-distribution molecular and protein design.arXiv preprint arXiv:2407.11942,

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.