Pith. sign in

REVIEW 2 major objections 4 minor 53 references

Using Diffusion Models to do Data Assimilation

T0 review · 2 major / 4 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 Diffusion models trained on long weather time series sample a different posterior from ensemble data assimilation, and matching it requires retraining the denoiser at every cycle.

desk verdict A useful taxonomy of diffusion DA posteriors, but the 'requires retraining each cycle' claim is overstated: a fixed denoiser conditioned on the forecast mean suffices in the linear Gaussian case. read the letter →

arxiv 2506.02249 v2 pith:YKAUXO3H submitted 2025-06-02 physics.ao-ph physics.data-an

Daniel Hodyss , Matthias Morzfeld This is my paper
classification physics.ao-phphysics.data-an
keywords diffusionmodelsdataassimilationBayesianposteriorcyclingpriorclimatologicalensembleKalmanfilterTweedie'sformulascore-basedgenerativemodeling
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 asks whether a diffusion model can replace an entire data assimilation system, and answers that the replacement is not neutral: the training data silently choose which Bayesian posterior the system samples. In a linear Gaussian model, the authors show that a diffusion model trained once on a long time series samples a posterior with a fixed climatological prior, whereas ensemble DA samples a posterior whose prior cycles with past observations. Sampling the ensemble-DA posterior instead requires regenerating the training dataset and retraining the denoiser at every DA cycle. A middle option, feeding a forecast in as an extra pseudo-observation, keeps the fixed prior but lands between the two in accuracy. The upshot is that the common practice of training once on a long past dataset is not equivalent to conventional DA, and the difference appears even before any neural-network training error is considered.

What carries the argument

The load-bearing object is the denoiser $D(v,t)$, defined as the conditional expectation $\mathbb{E}_{x\sim p(x|v,y)}[x]$; Tweedie's formula, the identity $\nabla\log p_t(v|y)=(\mathbb{E}[x|v]-v)/t^2$, ties this denoiser to the score function that drives the reverse diffusion SDE. In this paper the denoiser is computed analytically in a linear Gaussian setting, which the authors treat as equivalent to assuming a well-trained neural network. Because the training set for $D$ is exactly the prior in the Bayesian update, the choice of training set, long climatological time series, regenerated ensemble forecasts, or climatology plus a forecast pseudo-observation, selects which posterior the reverse process samples. The cycling prior is the mechanism that lets ensemble DA carry information across cycles, and the paper's central technical result is that a diffusion model with a cycling prior is a diffusion model whose training set is regenerated each cycle, hence one that must be retrained each cycle.

What would settle it

For the paper's scalar linear-Gaussian test problem, run a diffusion DA system trained once on climatological data through many assimilation cycles without retraining and measure the ensemble variance of an observed variable: the paper predicts it converges to $r/(1+r)$ (the climatological posterior), not to $\alpha r/(\alpha+r)$ (the cycling posterior), so a measured match to the latter would refute the central claim.

Watch

Extended reading notes

Core claim

The central discovery is that diffusion DA systems are distinguished by the prior hidden in their training set, and that this prior determines the posterior they can sample. Using a linear stochastic dynamical system with Gaussian errors, the paper computes the exact denoiser analytically, treating it as the limit of a perfectly trained neural network, and shows three configurations. A system trained once on climatology samples $p(x_k|y_k)$ with the fixed prior $p(x_k)$; a system retrained at each cycle with forecast ensemble members represents the cycling prior $p(x_k|y_1,\dots,y_{k-1})$ and reproduces the ensemble Kalman filter posterior; and a system that appends a forecast to the observation vector samples a posterior with an extended likelihood while keeping the climatological prior. The cycling-prior system requires retraining the denoiser at every cycle because the prior is equivalent to the training dataset and must be regenerated from the latest posterior. The paper also shows that a widely used approximate denoiser, which avoids retraining for new observation operators, produces a gain biased toward the observations and a posterior variance that is too small.

Load-bearing premise

The analysis assumes that the paper's analytic denoiser is exactly what a well-trained neural network would produce; if real networks fall short of that optimum, the posterior the system actually samples and the requirement to retrain each cycle could look different in practice.

Editorial extensions

If this is right

  • A diffusion DA system trained once on a long historical record does not reproduce ensemble DA; it samples a climatological-prior posterior and is suboptimal when observations are frequent.
  • To match the posterior targeted by the ensemble Kalman filter, a diffusion DA system must regenerate its training ensemble and retrain its denoiser at every assimilation cycle.
  • Adding a forecast as a pseudo-observation restores part of the lost information without retraining, giving accuracy between the climatological and cycling systems.
  • The approximate denoiser that avoids retraining for new observation networks carries a bias: it trusts observations too much and underestimates posterior variance.
  • The difference between cycling and climatological priors shrinks as the time between observations grows, because the cycling prior relaxes toward climatology.

Reading between the lines

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

  • A testable consequence the authors do not pursue is that in strongly non-Gaussian systems, where ensemble Kalman filters are known to struggle, a once-trained diffusion system with a climatological prior could outperform cycled ensemble DA despite being theoretically suboptimal, because it may sample the non-Gaussian posterior more faithfully.
  • The retraining cost of the cycling-prior system could be attacked by fine-tuning or transfer learning applied to the previous cycle's network rather than training from scratch; the authors list these as possibilities but do not test them.
  • The three-way classification likely extends beyond diffusion models to any generative sampler trained on a fixed dataset, since a fixed training set would impose the same climatological prior on normalizing flows or GANs used for data assimilation.
  • In operational settings where observation networks change frequently, the paper implies that the no-retraining advantage of the approximate denoiser is real but bounded, and that adding a forecast pseudo-observation is the natural first fix.
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

2 major / 4 minor

Summary. The paper studies whether diffusion models can replace conventional data assimilation (DA) and argues that different diffusion-based DA systems target different Bayesian posteriors. In a linear Gaussian test problem with analytic denoisers, the authors identify several variants: a denoiser trained on a long climatological time series, which samples the posterior with a climatological prior; an approximate score-based denoiser, which is shown to be biased; a denoiser retrained each cycle on samples from the cycling prior, which samples the same posterior as ensemble DA; and an extended-likelihood system that ingests a forecast as a pseudo-observation. Exact reverse-process SDE solutions and numerical experiments confirm the predicted means and variances for all three main variants. The paper concludes that matching the ensemble-DA posterior requires re-training the denoiser at every DA cycle, which is computationally costly.

Significance. If the results stand, the paper provides a valuable conceptual taxonomy for diffusion-based DA and a clean exact benchmark in the linear Gaussian case. The analytic treatment is careful: Tweedie's formula is derived, the reverse SDEs are solved exactly in the appendices, and the numerical experiments verify the closed-form means and variances for the climatological, cycling, and extended-likelihood posteriors. The distinction between climatological and cycling priors, and the quantification of the approximate-denoiser bias, are useful for practitioners. However, the headline 'requires re-training' conclusion is stronger than the derivation supports: the proof establishes sufficiency of one training strategy, not necessity, and a fixed denoiser conditioned on the forecast mean would sample the same cycling posterior in the linear Gaussian steady state. The extended-likelihood calibration is also self-consistent rather than independently specified. These issues affect the paper's main conclusions but are fixable within the manuscript's scope.

major comments (2)
  1. [4.c.2] The claim that matching ensemble DA 'requires re-training at each DA cycle' is not supported by the derivation. Equation (68) expresses the optimal denoiser as a function of v, y_k, x_f, and the constant steady-state covariance P_f_inf; past observations enter only through the forecast mean x_f. In the linear Gaussian steady state, x_f is a sufficient statistic for y_1, ..., y_{k-1}, so a single fixed denoiser D(v, y_k, x_f, t) trained once on offline triples (x_k, v, y_k, x_f) can be evaluated at the current forecast mean each cycle. The proof in Section 4.c.2 establishes sufficiency of the retraining-based construction, not necessity. The manuscript should reframe the conclusion as describing one implementation route, and the abstract, the closing paragraph of Section 4.c.2, Section 6, and Section 7 item 2 should be revised accordingly. If the authors intend to claim necessity in nonlinear/non-Gaussian settings, a separate argument is required; the linear Gaussian analysis cannot bear that claim.
  2. [5.b] The extended-likelihood parameters a and r_f are determined by a fixed-point iteration that runs the very same extended-likelihood Kalman filter whose posterior variance is later reported. This makes the comparison in Figure 2(b) self-consistent rather than independently calibrated; a different calibration, such as estimating r_f from an independent forecast-error sample or using a=1 with r_f equal to the steady-state forecast error variance, would change the posterior variance in Eq. (83) and could change the ranking of the extended-likelihood system relative to the cycling-prior system. The authors should either derive a and r_f from independent data or include a sensitivity analysis over reasonable parameter choices.
minor comments (4)
  1. [Figure 1] The figure legend contains a typo: 'Climotological prior' should be 'Climatological prior'.
  2. [5.b] The fixed-point iteration for determining a and r_f is described only in prose; a compact algorithm statement or pseudocode would improve reproducibility.
  3. [Section 7 item 4] The statement that the approximate denoiser's variance bias 'is likely to be easily fixed by the commonly available tunable parameters in the reverse diffusion process' is speculative and not demonstrated; the paper should either demonstrate the variance correction or soften this claim.
  4. [Acknowledgments] There is a typo: 'an additional anomymous reviewer' should be 'an additional anonymous reviewer'.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor circularity only: the extended-likelihood parameters are fitted by iterating the same filter they are then used to confirm; the main posterior derivations are self-contained.

  1. fitted input called prediction [Section 5.b, 'Diffusion DA implementation details' (extended-likelihood parameter calibration; see also Eq. (96) and Section 5.b.1 item (iii)).]
    "We then fit two functions to this large dataset of forecast-truth pairs. ... Second, we can compute the MSE associated with the forecast, i.e., MSE_f = E(([f_k]_j-[x_k]_j)^2) = (1-a)^2 [x_k]_j^2 + r_f, and subsequently fit a quadratic to obtain another estimate of the parameter a and also an estimate of the parameter r_f. We then repeat this process by re-running the cycling experiment with the new values of a and r_f until the two estimates of a agree to two decimal places."

    The parameters a and r_f are not independent external inputs: they are obtained by cycling the extended-likelihood Kalman filter of Section 4.d.1 and fitting its own forecast-truth pairs and forecast MSE. Because r_f is fit through Eq. (96) so that the modeled MSE matches the observed MSE, and the posterior variance in Eq. (83) is a function of r_f, the later claim that 'posterior variances match time-averaged MSE' is a self-consistency condition of the fitting loop rather than an independent confirmation. The diffusion SDE's equality to that posterior is proven analytically in Eqs. (87)-(90), so the numerical experiment mainly checks the discretization of expressions already known to be equal.

full rationale

Section 4.b.2, 4.c.2, and 4.d.2 derive the three posterior distributions directly from Bayes' rule, the forward SDE (11), and Tweedie's formula, without fitted parameters; the analytic-denoiser assumption ('equivalent to assuming a well-trained neural network') is explicit and is an idealization rather than a circular input. The central claim that matching ensemble DA 'requires re-training the denoiser at each cycle' is a sufficiency/necessity overstatement: Eq. (68) shows the denoiser depends on the past only through x_f and the steady covariance P_f_inf, so a single offline-trained denoiser conditioned on x_f could evaluate the same reverse SDE (70) without per-cycle retraining. That is a correctness/scope gap, not an equation-level circularity, and it does not increase the circularity score. The one genuine, though non-load-bearing, circular step is the extended-likelihood numerical loop, where a and r_f are calibrated by running the very same extended-likelihood Kalman filter until slope and MSE fits agree, making part of the 'confirmation' self-consistent by construction. Overall the central theory is independent of any fit, so the circularity score is low.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The analysis relies on standard probability and diffusion-model theory, plus the idealization that the denoiser is perfectly trained. The extended-likelihood model adds two calibrated parameters (a and r_f) that are fitted to the same system. No new entities are introduced.

free parameters (2)
  • a (extended likelihood slope) = 0.61 for Delta=0.1
    Calibrated by an iterative Kalman filter with extended likelihood on the same system used for evaluation; Section 5.b.
  • r_f (extended likelihood forecast error variance) = 0.34 for Delta=0.1
    Calibrated together with a by fitting the forecast MSE; Section 5.b.
assumptions (5)
  • standard math Bayes' rule and standard probability calculus.
    Used throughout to define posterior distributions (e.g., eq. 1, 27, 78).
  • domain assumption A perfectly trained denoiser recovers the exact conditional expectation (Tweedie's formula).
    Section 4.b.2 states the analytic denoiser 'is equivalent to assuming a well-trained neural network.' This idealization underlies all posterior-matching claims.
  • domain assumption The linear Ornstein-Uhlenbeck model with steady-state cycling prior.
    Section 4.a defines the model; Section 4.c.1 assumes the cycled covariance converges to a steady state P^f_infty for the analysis.
  • domain assumption Extended likelihood has the linear form f_k = a x_k + epsilon_f.
    Section 4.d and Appendix E derive this for the test problem but generalize it as a modeling assumption; the authors state 'We make our assumptions based on what we think is most reasonable at the time of writing.'
  • domain assumption Approximate denoiser score uses the standard DPS-style approximation grad log p_t(y|v) approx grad log p_0(y|D(v,t)).
    Equation (21) in Section 4.b.3; this is the object of study, and the bias results follow from it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Using Diffusion Models to do Data Assimilation." pith.science (2026). https://pith.science/paper/YKAUXO3H

@misc{pith2026250602249,
  author       = {Pith},
  title        = {Pith review of: Using Diffusion Models to do Data Assimilation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YKAUXO3H}},
  note         = {Machine review of arXiv:2506.02249}
}
read the original abstract

The recent surge in machine learning (ML) methods for geophysical modeling has raised the question of how these methods might be applied to data assimilation (DA). We focus on diffusion modeling (a form of generative artificial intelligence) for systems that can perform the entire DA process, rather than on ML-based tools used within a conventional DA system. We identify at least three distinct types of diffusion-based DA systems and show that they differ in the posterior distribution they target for sampling. These posterior distributions correspond to different priors and/or likelihoods, which in turn result in unique training datasets, computational requirements, and state estimate qualities. Our analysis further shows that a diffusion DA system designed to target the same posterior distribution as current ensemble DA algorithms requires re-training at each DA cycle, which is computationally costly. We discuss the implications of these findings for the use of diffusion modeling in DA.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 32 canonical work pages

  1. [1]

    Sanz-Alonso, and W

    Adrian, M., D. Sanz-Alonso, and W. R., 2025: Data assimilation with machine learning surrogate models: A case study with F our C ast N et. Artificial Intelligence for Earth System

    work page 2025

  2. [2]

    Nature, doi:https://doi.org/10.1038/s41586-025-08897-0

    Allen, A., and Coauthors, 2025: End-to-end data-driven weather prediction. Nature, doi:https://doi.org/10.1038/s41586-025-08897-0

    doi:10.1038/s41586-025-08897-0 2025

  3. [3]

    D., 1982: Reverse-time diffusion equation models

    Anderson, B. D., 1982: Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12 (3), 313--326, doi:https://doi.org/10.1016/0304-4149(82)90051-5, ://www.sciencedirect.com/science/article/pii/0304414982900515

    arXiv 1982

  4. [4]

    Monthly weather review, 129 (12), 2884--2903

    Anderson, J., 2001: An ensemble adjustment K alman filter for data assimilation. Monthly weather review, 129 (12), 2884--2903

    work page 2001

  5. [5]

    Zhang, and G

    Bao, F., Z. Zhang, and G. Zhang, 2024: A score-based filter for nonlinear data assimilation. Journal of Computational Physics, 514, 113\,207, doi:https://doi.org/10.1016/j.jcp.2024.113207, ://www.sciencedirect.com/science/article/pii/S002199912400456X

    arXiv 2024

  6. [6]

    Stanczuk, C.-B

    Batzolis, G., J. Stanczuk, C.-B. Schönlieb, and C. Etmann, 2021: Conditional image generation with score-based diffusion models. ://arxiv.org/abs/2111.13606, 2111.13606

    arXiv 2021

  7. [7]

    Bickel, and B

    Bengtsson, T., P. Bickel, and B. Li, 2008: Curse of dimensionality revisited: T he collapse of importance sampling in very large scale systems. IMS Collections: Probability and Statistics: Essays in Honor of David A. Freedman, 2, 316--334

    work page 2008

  8. [8]

    Bengtsson, and J

    Bickel, P., T. Bengtsson, and J. Anderson, 2008: Sharp failure rates for the bootstrap particle filter in high dimensions. Pushing the Limits of Contemporary Statistics: Contributions in Honor of Jayanta K. Ghosh, 3, 318--329

    work page 2008

Show all 53 references
  1. [9]

    Isaksen, and E

    Bonavita, M., L. Isaksen, and E. H\'olm, 2012: On the use of EDA background error variances in the ECMWF 4D-Var . Quarterly Journal of the Royal Meteorological Society, 138 (667), 1540--1559, doi:https://doi.org/10.1002/qj.1899

    2012 doi

  2. [10]

    Mourneau, and C

    Buehner, M., J. Mourneau, and C. Charette, 2013: Four-dimensional ensemble-variational data assimilation for global deterministic weather prediction. Nonlin. Processes Geophys., 20, 669--682

    2013

  3. [11]

    Burgers, G., P. V. Leeuwen, and G. Evensen, 1998: Analysis scheme in the ensemble K alman filter. Monthly weather review, 126 (6), 1719--1724

    1998

  4. [12]

    Chung, H., J. Kim, M. T. Mccann, M. L. Klasky, and J. C. Ye, 2023: Diffusion posterior sampling for general noisy inverse problems. The Eleventh International Conference on Learning Representations, ://arxiv.org/abs/2209.14687

    2023 arXiv

  5. [13]

    Ding, X., Y. Wang, K. Zhang, and Z. J. Wang, 2025: CCDM : Continuous conditional diffusion models for image generation. ://arxiv.org/abs/2405.03546, 2405.03546

    2025 arXiv

  6. [14]

    Journal of the American Statistical Association, 106(496), 1602--1614

    Efron, B., 2011: Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496), 1602--1614

    2011

  7. [15]

    Journal of Geophysical Research: Oceans, 99 (C5), 10\,143--10\,162, doi:10.1029/94JC00572

    Evensen, G., 1994: Sequential data assimilation with a nonlinear quasi-geostrophic model using M onte C arlo methods to forecast error statistics. Journal of Geophysical Research: Oceans, 99 (C5), 10\,143--10\,162, doi:10.1029/94JC00572

    1994 doi

  8. [16]

    Evensen, G., and Coauthors, 2009: Data assimilation: the ensemble K alman filter , Vol. 2. Springer

    2009

  9. [17]

    Gharamti, M. E., K. Raeder, J. Anderson, and X. Wang, 2019: Comparing adaptive prior and posterior inflation for ensemble filters using an atmospheric general circulation model. Monthly Weather Review, 147 (7), 2535 -- 2553, doi:10.1175/MWR-D-18-0389.1, ://journals.ametsoc.org...

    2019 doi

  10. [18]

    M., and C

    Hamill, T. M., and C. Snyder, 2000: A hybrid ensemble K alman filter-3 D variational analysis scheme. Monthly Weather Review, 128, 2905--2919

    2000

  11. [19]

    Jain, and P

    Ho, J., A. Jain, and P. Abbeel, 2020: Denoising diffusion probabilistic models. 34th Conference on Neural Information Processing Systems (NeurIPS 2020), ://proceedings.neurips.cc/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf

    2020

  12. [20]

    Bichop, and M

    Hodyss, D., C. Bichop, and M. Morzfeld, 2016 a : To what extent is your data assimilation scheme designed to find the posterior mean, the posterior mode or something else? Tellus A, 68 (30625), doi:http://dx.doi.org/10.3402/tellusa.v68.30625

    2016 doi

  13. [21]

    Hodyss, D., W. F. Campbell, and J. S. Whitaker, 2016 b : Observation-dependent posterior inflation for the ensemble K alman filter. Monthly Weather Review, 144, 2667--2684

    2016

  14. [22]

    Morzfeld, 2023: How sampling errors in covariance estimates cause bias in the K alman gain and impact ensemble data assimilation

    Hodyss, D., and M. Morzfeld, 2023: How sampling errors in covariance estimates cause bias in the K alman gain and impact ensemble data assimilation. Monthly Weather Review, 151 (9), 2413 -- 2426, doi:10.1175/MWR-D-23-0029.1, ://journals.ametsoc.org/view/journals/mwre/151/9/MWR...

    2023 doi

  15. [23]

    Gianinazzi, Y

    Huang, L., L. Gianinazzi, Y. Yu, P. D. Dueben, and T. Hoefler, 2024: Diffda: a diffusion model for weather-scale data assimilation. ://arxiv.org/abs/2401.05932, 2401.05932

    2024 arXiv

  16. [24]

    Cambridge University Press

    Kalnay, E., 2002: Atmospheric Modeling, Data Assimilation and Predictability. Cambridge University Press

    2002

  17. [25]

    Aittala, T

    Karras, T., M. Aittala, T. Aila, and S. Laine, 2022: Elucidating the design space of diffusion-based generative models. 36th Conference on Neural Information Processing Systems, 35, 26\,565–26\,577

    2022

  18. [26]

    D., and R

    Keller, J. D., and R. Potthast, 2024: AI -based data assimilation: L earning the functional of analysis estimation. ://arxiv.org/abs/2406.00390, 2406.00390

    2024 arXiv

  19. [27]

    Nature, 632, 1060--1066

    Kochkov, D., and Coauthors, 2024: Neural general circulation models for weather and climate. Nature, 632, 1060--1066

    2024

  20. [28]

    Kuhl, D. D., T. E. Rosmond, C. H. Bishop, J. McLay, and N. L. Baker, 2013: Comparison of hybrid ensemble/4dvar and 4dvar within the navdas-ar data assimilation framework. Monthly Weather REview, 141, 2740--2758

    2013

  21. [29]

    Science, 382 (6677), 1416--1421, doi:10.1126/science.adi2336, ://www.science.org/doi/abs/10.1126/science.adi2336, https://www.science.org/doi/pdf/10.1126/science.adi2336

    Lam, R., and Coauthors, 2023: Learning skillful medium-range global weather forecasting. Science, 382 (6677), 1416--1421, doi:10.1126/science.adi2336, ://www.science.org/doi/abs/10.1126/science.adi2336, https://www.science.org/doi/pdf/10.1126/science.adi2336

    2023 doi

  22. [30]

    Carver, I

    Li, L., R. Carver, I. Lopez-Gomez, F. Sha, and J. Anderson, 2024: Generative emulation of weather forecast ensembles with diffusion models. Science Advances, 10 (13), eadk4489, doi:10.1126/sciadv.adk4489, ://www.science.org/doi/abs/10.1126/sciadv.adk4489, https://www.science.o...

    2024 doi

  23. [31]

    Dong, and P

    Li, Z., B. Dong, and P. Zhang, 2025: State-observation augmented diffusion model for nonlinear assimilation with unknown dynamics. ://arxiv.org/abs/2407.21314, 2407.21314

    2025 arXiv

  24. [32]

    Quarterly Journal of the Royal Meteorological Society, 129 (595), 3183--3203, doi:10.1256/qj.02.132

    Lorenc, A., 2003: The potential of the ensemble K alman filter for NWP -- a comparison with 4 D - V ar. Quarterly Journal of the Royal Meteorological Society, 129 (595), 3183--3203, doi:10.1256/qj.02.132

    2003 doi

  25. [33]

    ://arxiv.org/abs/2406.16947, 2406.16947

    Manshausen, P., and Coauthors, 2024: Generative data assimilation of sparse weather station observations at kilometer scales. ://arxiv.org/abs/2406.16947, 2406.16947

    2024 arXiv

  26. [34]

    Commun Earth Environ, 6, doi:https://doi.org/10.1038/s43247-025-02042-5

    Mardani, M., and Coauthors, 2025: Residual corrective diffusion modeling for km-scale atmospheric downscaling. Commun Earth Environ, 6, doi:https://doi.org/10.1038/s43247-025-02042-5

    2025 doi

  27. [35]

    Hodyss, 2019: Gaussian approximations in filters and smoothers for data assimilation

    Morzfeld, M., and D. Hodyss, 2019: Gaussian approximations in filters and smoothers for data assimilation. Tellus A: Dynamic Meteorology and Oceanography, 71 (1), 1600\,344, doi:10.1080/16000870.2019.1600344

    2019

  28. [36]

    Hodyss, 2023: A theory for why even simple covariance localization is so useful in ensemble data assimilation

    Morzfeld, M., and D. Hodyss, 2023: A theory for why even simple covariance localization is so useful in ensemble data assimilation. Monthly Weather Review, 151, 717--736

    2023

  29. [37]

    F., and J

    Parrish, D. F., and J. C. Derber, 1992: The national meteorological center's spectral statistical-interpolation analysis system. Monthly Weather Review

    1992

  30. [38]

    Parthasarathy, V. B., A. Zafar, A. Khan, and A. Shahid, 2024: The ultimate guide to fine-tuning llms from basics to breakthroughs: An exhaustive review of technologies, research, best practices, applied research challenges and opportunities. ://arxiv.org/abs/2408.13296, 2408.13296

    2024 arXiv

  31. [39]

    ://arxiv.org/abs/2408.10958, 2408.10958

    Pathak, J., and Coauthors, 2024: Kilometer-scale convection allowing model emulation using generative diffusion modeling. ://arxiv.org/abs/2408.10958, 2408.10958

    2024 arXiv

  32. [40]

    Poterjoy, J., and F. Zhang, 2015: Systematic comparison of four-dimensional data assimilation methods with and without the tangent linear model using hybrid background error covariance: E4DVar versus 4DEnVar . Monthly Weather Review, 143 (5), 1601--1621

    2015

  33. [41]

    Nature, 637, 84--90

    Price, I., and Coauthors, 2025: Probabilistic weather forecasting with machine learning. Nature, 637, 84--90

    2025

  34. [42]

    Nathaniel, S

    Qu, Y., J. Nathaniel, S. Li, and P. Gentine, 2024: Deep Generative Data Assimilation in Multimodal Setting . 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), IEEE Computer Society, Los Alamitos, CA, USA, 449--459, doi:10.1109/CVPRW63382.20...

    2024

  35. [43]

    Louppe, 2023: Score-based data assimilation

    Rozet, F., and G. Louppe, 2023: Score-based data assimilation. Thirty-seventh Conference on Neural Information Processing Systems, ://openreview.net/forum?id=VUvLSnMZdX

    2023

  36. [44]

    Bengtsson, P

    Snyder, C., T. Bengtsson, P. Bickel, and J. Anderson, 2008: Obstacles to high-dimensional particle filtering. Monthly Weather Review, 136 (12), 4629--4640

    2008

  37. [45]

    Bengtsson, and M

    Snyder, C., T. Bengtsson, and M. Morzfeld, 2015: Performance bounds for particle filters using the optimal proposal. Monthly Weather Review, 143, 4750--4761

    2015

  38. [46]

    Weiss, N

    Sohl-Dickstein, J., E. Weiss, N. Maheswaranathan, and S. Ganguli, 2015: Deep unsupervised learning using nonequilibrium thermodynamics. Proceedings of the 32nd International Conference on Machine Learning, F. Bach, and D. Blei, Eds., PMLR, Lille, France, Proceedings of Machine...

    2015

  39. [47]

    Courtier, 1987: Variational assimilation of meteorological observations with the adjoint vorticity equation

    Talagrand, O., and P. Courtier, 1987: Variational assimilation of meteorological observations with the adjoint vorticity equation. i: T heory. Quarterly Journal of the Royal Meteorological Society, 113 (478), 1311--1328

    1987

  40. [48]

    Anderson, C

    Tippett, M., J. Anderson, C. Bishop, T. Hamill, and J. Whitaker, 2003: Ensemble square root filters. Monthly Weather Review, 131 (7), 1485 -- 1490, doi:10.1175/1520-0493(2003)131<1485:ESRF>2.0.CO;2

    2003 doi

  41. [49]

    J., 2020: A consistent interpretation of the stochastic version of the ensemble kalman filter

    van Leeuwen, P. J., 2020: A consistent interpretation of the stochastic version of the ensemble kalman filter. Quarterly Journal of the Royal Meteorological Society, 146 (731), 2815--2825, doi:https://doi.org/10.1002/qj.3819, ://rmets.onlinelibrary.wiley.com/doi/abs/10.1002/qj...

    2020 doi

  42. [50]

    S., and T

    Whitaker, J. S., and T. M. Hamill, 2012: Evaluating methods to account for system errors in ensemble data assimilation. Mon. Wea. Rev., 140, 3078--3089

    2012

  43. [51]

    Chen, J.-P

    Zhan, Z., D. Chen, J.-P. Mei, Z. Zhao, J. Chen, C. Chen, S. Lyu, and C. Wang, 2024: Conditional image synthesis with diffusion models: A survey. CoRR, abs/2409.19365, ://doi.org/10.48550/arXiv.2409.19365

    2024 doi arXiv

  44. [52]

    Zhang, and J

    Zhang, F., M. Zhang, and J. A. Hansen, 2009: Coupling ensemble K alman filter with four-dimensional variational data assimilation. Advances in Atmospheric Sciences, 26, 1--8

    2009

  45. [53]

    Zhuang, F., Z. Qi, K. Duan, D. Xi, Y. Zhu, H. Zhu, H. Xiong, and Q. He, 2020: A comprehensive survey on transfer learning. ://arxiv.org/abs/1911.02685, 1911.02685

    2020 arXiv

Pith tools

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