pith. sign in

arxiv: 2606.27094 · v1 · pith:XR5AJPRSnew · submitted 2026-06-25 · ⚛️ physics.ao-ph

Learning Climate Variability from Scarce Data with Diffusion Models: A Test Case for ENSO

Lluis Palma , Vincent Verjans , Amanda Duarte , Albert Soret , Markus Donat This is my paper

Pith reviewed 2026-06-26 01:55 UTC · model grok-4.3

classification ⚛️ physics.ao-ph
keywords diffusion modelsENSOclimate variabilityscarce datapre-trainingCMIP6Linear Inverse Modelstropical Pacific SST
0
0 comments X

The pith

Pre-training on climate simulations lets diffusion models recover accurate ENSO patterns from far fewer real observations than direct training needs.

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

The paper examines whether diffusion models can learn the correct modes of tropical Pacific sea surface temperature variability, especially ENSO, despite limited observational records. Tests on synthetic data generated by Linear Inverse Models, which have known low-order structure, show that sufficient training samples allow the models to recover both Gaussian and non-Gaussian features including Eastern versus Central Pacific asymmetry. Real data such as the 700 monthly samples in ERSSTv5 fall short of the roughly 7000 samples needed for convergence, and some diffusion setups fail to capture the structure. Pre-training the model on CMIP6 simulations using a learned embedding, then fine-tuning on the scarce observations, overcomes this shortfall and matches observed statistics more closely than either Gaussian or non-Gaussian LIMs.

Core claim

Pre-training a diffusion model on CMIP6 data with a learned model embedding, followed by fine-tuning on scarce observations, reproduces observed ENSO statistics more faithfully than both Gaussian and non-Gaussian Linear Inverse Models, because the pre-training supplies enough samples for the model to learn the correct low-order structure that direct training on real data alone cannot achieve.

What carries the argument

Diffusion model with learned embedding, pre-trained on CMIP6 simulations and fine-tuned on limited observations, validated against synthetic SST fields from Linear Inverse Models with known structure.

If this is right

  • With enough training samples the diffusion model recovers the correct low-order structure of both Gaussian and non-Gaussian LIMs.
  • The 700 monthly observations in ERSSTv5 are an order of magnitude below the 7000 samples required for convergence without pre-training.
  • Not every diffusion parameterisation succeeds in recovering the correct structure even with abundant data.
  • The pre-training plus fine-tuning pipeline closes the data gap and outperforms standard LIM approaches on observed statistics.

Where Pith is reading between the lines

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

  • The same pre-training strategy could be tested on other climate variables or regions where observational records are similarly short.
  • If the LIM proxy holds, the approach might extend to variability on longer timescales where data scarcity is even more severe.
  • The method supplies a concrete way to check whether diffusion models are capturing physically meaningful modes rather than artifacts of limited samples.

Load-bearing premise

The low-order structure recovered from LIM-generated synthetic fields is a sufficient proxy for the modes present in real tropical Pacific SST so that pre-training benefits carry over to actual observations.

What would settle it

A test showing that the pre-trained and fine-tuned diffusion model still fails to match key observed ENSO statistics, such as the Eastern/Central Pacific asymmetry, in withheld real data from ERSSTv5.

Figures

Figures reproduced from arXiv: 2606.27094 by Albert Soret, Amanda Duarte, Lluis Palma, Markus Donat, Vincent Verjans.

Figure 1
Figure 1. Figure 1: Top row (a,b,c,d,e): model trained on Linear Inverse Model (LIM) data. Bottom row (f,g,h,i,j): trained on nonlinear-Gaussian LIM (NG-LIM). Note that “true” refers to LIM￾and NG-LIM-generated data in the top and bottom row, respectively. (a,f) Scatter plots of the two leading principal components (PC1 vs PC2) for true data (green) and generated samples (grey/black), with quadratic fit curves highlighting th… view at source ↗
Figure 2
Figure 2. Figure 2: Performance of x-prediction diffusion as a function of training set size (N), eval￾uated on the NG-LIM dataset. (a) Variance-weighted PC pattern correlation. (b) Projected variance ratio. (c) Training (solid) and validation (dashed) loss. (d) Spatial skewness correlation. (e) Memorization ratio (nearest-neighbour distance of generated vs held-out to training set). Shaded bands show ±1 standard deviation ac… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of training strategies for learning the NG-LIM distribution from only N = 720 samples. (a) Variance-weighted PC pattern correlation, (b) spatial skewness correlation, (c) projected variance ratio, and (d) memorisation ratio. Strategies shown: direct training in full space (black), PCA reduction to 100D (blue), CMIP6 pre-training without fine-tuning (purple), CMIP6 pre-training with fine-tuning (… view at source ↗
Figure 4
Figure 4. Figure 4: (a–d) Scatter plots of the two dominant modes of sea surface temperature variabil￾ity (PC1 vs PC2), as generated by: (a) a linear model (LIM), (b) a non-Gaussian linear model (NG-LIM), (c) our diffusion model leveraging CMIP6 simulations and fine-tuned to ERSSTv5, and (d) the observed record (ERSSTv5, 1948–2022). (e–h) Maps of skewness—a measure of warm–cold asymmetry, where positive values (red) indicate … view at source ↗
read the original abstract

Diffusion models are increasingly applied to climate emulation, but whether they capture the correct modes of variability remains unclear, a concern amplified by data scarcity at longer timescales. We investigate this using synthetic tropical Pacific SST fields from Linear Inverse Models (LIMs), whose known low-order structure bypasses the overlapping and confounding modes of real observations. With sufficient training data, our model recovers the correct structure of both Gaussian and non-Gaussian LIMs, including ENSO's Eastern/Central Pacific asymmetry. Yet an ablation study on the number of monthly training samples reveals that the 700 observations in ERSSTv5 fall an order of magnitude short of the 7,000 samples needed for convergence, and that not all diffusion parameterisations recover the correct low-order structure. Pre-training on CMIP6 with a learned model embedding, followed by fine-tuning on scarce observations, closes this gap, reproducing observed statistics more faithfully than both Gaussian and non-Gaussian LIMs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper investigates diffusion models for emulating ENSO variability under data scarcity. Using synthetic tropical Pacific SST from Gaussian and non-Gaussian LIMs as a controlled testbed with known low-order structure, it shows that diffusion models recover the correct modes (including ENSO asymmetry) with sufficient samples (~7000), but 700 ERSSTv5 observations are insufficient. Pre-training on CMIP6 with a learned embedding followed by fine-tuning on scarce observations is shown to close the gap, yielding better reproduction of observed statistics than the LIM baselines.

Significance. If the transfer from LIM proxies to real observations holds, the approach offers a practical route to leverage large CMIP6 ensembles for improving statistical emulation of observed climate modes in data-limited regimes. The controlled LIM testbed is a clear strength, providing a falsifiable check on mode recovery before real-data application.

major comments (2)
  1. [Ablation study and real-data evaluation sections] The central claim that pre-training + fine-tuning reproduces observed ERSSTv5 statistics more faithfully than non-Gaussian LIMs rests on the untested assumption that success on LIM-generated fields implies success when the target distribution includes additional overlapping, non-stationary modes present in real tropical Pacific SST. The ablation study quantifies sample requirements only for the LIM proxy and does not include a direct comparison of embeddings learned from LIMs versus real data.
  2. [Results on ERSSTv5] The outperformance over LIM baselines on observed statistics is load-bearing for the headline result, yet the manuscript provides no quantitative metrics, error bars, or details on the exact statistics (e.g., power spectra, asymmetry measures) used for the ERSSTv5 comparison, making it impossible to assess whether the reported improvement exceeds sampling variability.
minor comments (1)
  1. [Methods] Clarify the precise architecture of the 'learned model embedding' used for pre-training and how it is transferred during fine-tuning.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive review. The comments highlight important aspects of our evaluation strategy and the need for more rigorous quantification on real data. We address each major comment below and indicate planned revisions.

read point-by-point responses

  1. Referee: The central claim that pre-training + fine-tuning reproduces observed ERSSTv5 statistics more faithfully than non-Gaussian LIMs rests on the untested assumption that success on LIM-generated fields implies success when the target distribution includes additional overlapping, non-stationary modes present in real tropical Pacific SST. The ablation study quantifies sample requirements only for the LIM proxy and does not include a direct comparison of embeddings learned from LIMs versus real data.

    Authors: We agree that the LIM testbed is a proxy and that success on synthetic data with known structure does not automatically transfer to observations containing additional non-stationary modes. The LIM experiments were designed specifically to provide a falsifiable check on whether the diffusion model can recover the correct low-order modes (including asymmetry) when the ground truth is known, which is impossible with real data. For the ERSSTv5 application, evaluation is performed by direct comparison of generated statistics against those computed from the observations themselves, rather than assuming transfer from LIM. A direct comparison of LIM-learned versus CMIP6-learned embeddings is not feasible because the true underlying modes in observations are unknown; the LIM provides the only controlled ground truth. We will add a dedicated discussion paragraph clarifying the proxy rationale, its limitations, and why embedding comparison on real data is not possible. revision: partial

  2. Referee: The outperformance over LIM baselines on observed statistics is load-bearing for the headline result, yet the manuscript provides no quantitative metrics, error bars, or details on the exact statistics (e.g., power spectra, asymmetry measures) used for the ERSSTv5 comparison, making it impossible to assess whether the reported improvement exceeds sampling variability.

    Authors: We acknowledge that the current manuscript relies on qualitative descriptions of improved reproduction of observed statistics without providing quantitative metrics or uncertainty estimates. In the revised version we will add a new table and accompanying text that reports specific quantitative comparisons (including power spectra, skewness/asymmetry measures, and variance explained by leading EOFs), with error bars obtained via bootstrap resampling of the generated ensembles. We will also explicitly list the statistics used and the procedure for assessing whether differences exceed sampling variability. revision: yes

Circularity Check

0 steps flagged

No significant circularity; central claims rest on external LIM structure and independent datasets

full rationale

The paper trains diffusion models on LIM-generated synthetic SST fields (with known low-order structure), pre-trains on CMIP6, and fine-tunes/evaluates on ERSSTv5 observations, comparing performance to Gaussian and non-Gaussian LIM baselines. No quoted equations or steps reduce a reported result to a fitted parameter on the same data, a self-definition, or a load-bearing self-citation chain. LIMs and CMIP6/ERSSTv5 serve as external references whose structure is not derived from the diffusion model itself. The ablation on sample count and the pre-training benefit are evaluated against these independent benchmarks, keeping the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that LIM-generated fields capture the relevant variability structure and that the number of samples needed for convergence (7,000) is a reliable threshold for real data. No explicit free parameters or invented entities are stated in the abstract.

axioms (1)
  • domain assumption LIMs provide a faithful low-order representation of tropical Pacific SST variability against which diffusion models can be validated
    The entire synthetic-data test bed is built on this premise; if LIMs miss important higher-order or non-stationary features, the recovery results do not transfer.

pith-pipeline@v0.9.1-grok · 5703 in / 1392 out tokens · 18262 ms · 2026-06-26T01:55:29.777793+00:00 · methodology

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

72 extracted references · 63 canonical work pages · 13 internal anchors

  1. [1]

    Conditional Diffusion Models for Downscaling and Bias Correction of

    Aich, Michael and Hess, Philipp and Pan, Baoxiang and Bathiany, Sebastian and Huang, Yu and Boers, Niklas , year = 2026, month = mar, journal =. Conditional Diffusion Models for Downscaling and Bias Correction of. doi:10.5194/gmd-19-1791-2026 , langid =

    work page doi:10.5194/gmd-19-1791-2026 2026

  2. [2]

    and Hosking, J

    Andersson, Tom R. and Hosking, J. Scott and. Seasonal. Nature Communications 2021 12:1 , volume =

    2021

  3. [3]

    doi:10.1029/2023MS004194 , langid =

    Bassetti, Seth and Hutchinson, Brian and Tebaldi, Claudia and Kravitz, Ben , year = 2024, journal =. doi:10.1029/2023MS004194 , langid =

    work page doi:10.1029/2023ms004194 2024

  4. [4]

    Bonnaire, Tony and Urfin, Rapha. Why. doi:10.48550/arXiv.2505.17638 , archiveprefix =. 2505.17638 , primaryclass =

    work page doi:10.48550/arxiv.2505.17638

  5. [5]

    doi:10.48550/arXiv.2306.01984 , archiveprefix =

    Cachay, Salva R. doi:10.48550/arXiv.2306.01984 , archiveprefix =. 2306.01984 , primaryclass =

    work page doi:10.48550/arxiv.2306.01984

  6. [6]

    Elucidated

    Cachay, Salva R. Elucidated. doi:10.48550/arXiv.2506.20024 , archiveprefix =. 2506.20024 , primaryclass =

    work page doi:10.48550/arxiv.2506.20024

  7. [7]

    doi:10.1016/j.patter.2026.101497 , archiveprefix =

    Channing, Georgia and Ghosh, Avijit , year = 2026, month = mar, journal =. doi:10.1016/j.patter.2026.101497 , archiveprefix =. 2509.06580 , primaryclass =

    work page doi:10.1016/j.patter.2026.101497 2026

  8. [8]

    and Wittenberg, Andrew T

    Choi, Kit-Yan and Vecchi, Gabriel A. and Wittenberg, Andrew T. , year = 2013, month = dec, journal =. doi:10.1175/JCLI-D-13-00045.1 , langid =

    work page doi:10.1175/jcli-d-13-00045.1 2013

  9. [9]

    Geophysical Research Letters , volume =

    Global Modes of Climate Variability , author =. Geophysical Research Letters , volume =. doi:10.1002/grl.50386 , langid =

    work page doi:10.1002/grl.50386

  10. [10]

    and Deser, Clara , year = 2014, month = oct, journal =

    DiNezio, Pedro N. and Deser, Clara , year = 2014, month = oct, journal =. Nonlinear. doi:10.1175/JCLI-D-14-00033.1 , langid =

    work page doi:10.1175/jcli-d-14-00033.1 2014

  11. [11]

    Are We Misdiagnosing Ensemble Forecast Reliability?

    Dirkson, Arlan and Buehner, Mark , year = 2026, month = mar, journal =. Are We Misdiagnosing Ensemble Forecast Reliability?. doi:10.1002/qj.70186 , archiveprefix =. 2512.02160 , primaryclass =

    work page doi:10.1002/qj.70186 2026

  12. [12]

    Geoscientific Model Development , volume =

    The. Geoscientific Model Development , volume =. doi:10.5194/gmd-19-2821-2026 , langid =

    work page doi:10.5194/gmd-19-2821-2026 2026

  13. [13]

    A., et al

    Eyring, Veronika and Bony, Sandrine and Meehl, Gerald A. and Senior, Catherine A. and Stevens, Bjorn and Stouffer, Ronald J. and Taylor, Karl E. , year = 2016, month = may, journal =. Overview of the. doi:10.5194/gmd-9-1937-2016 , langid =

    work page doi:10.5194/gmd-9-1937-2016 2016

  14. [14]

    Achieving

    Feng, Jie and Lian, Tao and Liu, Ting and Chen, Dake and Li, Xichen and Tang, Youmin and Gao, Yanqiu and Song, Xunshu , year = 2026, journal =. Achieving. doi:10.1029/2025GL117573 , langid =

    work page doi:10.1029/2025gl117573 2026

  15. [15]

    Generative Emulation of Weather Forecast Ensembles with Diffusion Models , doi =

  16. [16]

    Communications Earth & Environment , volume =

    Training Machine Learning Models on Climate Model Output Yields Skillful Interpretable Seasonal Precipitation Forecasts , author =. Communications Earth & Environment , volume =. doi:10.1038/s43247-021-00225-4 , langid =

    work page doi:10.1038/s43247-021-00225-4

  17. [17]

    Guo, Cheng and Berkhahn, Felix , year = 2016, month = apr, number =. Entity. doi:10.48550/arXiv.1604.06737 , archiveprefix =. 1604.06737 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1604.06737 2016

  18. [18]

    and Jolliffe, I

    Hannachi, A. and Jolliffe, I. T. and Stephenson, D. B. , year = 2007, journal =. Empirical Orthogonal Functions and Related Techniques in Atmospheric Science:. doi:10.1002/joc.1499 , langid =

    work page doi:10.1002/joc.1499 2007

  19. [19]

    Held, I. M. and Guo, H. and Adcroft, A. and Dunne, J. P. and Horowitz, L. W. and Krasting, J. and Shevliakova, E. and Winton, M. and Zhao, M. and Bushuk, M. and Wittenberg, A. T. and Wyman, B. and Xiang, B. and Zhang, R. and Anderson, W. and Balaji, V. and Donner, L. and Dunne, K. and Durachta, J. and Gauthier, P. P. G. and Ginoux, P. and Golaz, J.-C. and...

    work page doi:10.1029/2019ms001829 2019

  20. [20]

    Denoising Diffusion Probabilistic Models

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , year = 2020, month = dec, number =. Denoising. doi:10.48550/arXiv.2006.11239 , archiveprefix =. 2006.11239 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2006.11239 2020

  21. [21]

    and Freeman, Eric and Lawrimore, Jay and Liu, Wei and Peterson, Thomas C

    Huang, Boyin and Banzon, Viva F. and Freeman, Eric and Lawrimore, Jay and Liu, Wei and Peterson, Thomas C. and Smith, Thomas M. and Thorne, Peter W. and Woodruff, Scott D. and Zhang, Huai-Min , year = 2015, month = feb, journal =. Extended. doi:10.1175/JCLI-D-14-00006.1 , langid =

    work page doi:10.1175/jcli-d-14-00006.1 2015

  22. [22]

    and Banzon, Viva F

    Huang, Boyin and Thorne, Peter W. and Banzon, Viva F. and Boyer, Tim and Chepurin, Gennady and Lawrimore, Jay H. and Menne, Matthew J. and Smith, Thomas M. and Vose, Russell S. and Zhang, Huai-Min , year = 2017, month = oct, journal =. Extended. doi:10.1175/JCLI-D-16-0836.1 , langid =

    work page doi:10.1175/jcli-d-16-0836.1 2017

  23. [23]

    Technical

    Immorlano, Francesco and Tavares, Elijah and Draxler, Felix and Smyth, Padhraic and Gentine, Pierre and Mandt, Stephan , year = 2025, month = nov, number =. Technical. doi:10.48550/arXiv.2511.22970 , archiveprefix =. 2511.22970 , primaryclass =

    work page doi:10.48550/arxiv.2511.22970 2025

  24. [24]

    Proceedings of the National Academy of Sciences , volume =

    Transferring Climate Change Physical Knowledge , author =. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.2413503122 , langid =

    work page doi:10.1073/pnas.2413503122

  25. [25]

    and Stockdale, Timothy N

    Johnson, Stephanie J. and Stockdale, Timothy N. and Ferranti, Laura and Balmaseda, Magdalena A. and Molteni, Franco and Magnusson, Linus and Tietsche, Steffen and Decremer, Damien and Weisheimer, Antje and Balsamo, Gianpaolo and Keeley, Sarah P.E. and Mogensen, Kristian and Zuo, Hao and. Geoscientific Model Development , volume =

  26. [26]

    Elucidating the Design Space of Diffusion-Based Generative Models

    Karras, Tero and Aittala, Miika and Aila, Timo and Laine, Samuli , year = 2022, month = oct, number =. Elucidating the. doi:10.48550/arXiv.2206.00364 , archiveprefix =. 2206.00364 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2206.00364 2022

  27. [27]

    npj Climate and Atmospheric Science , volume =

    Skilful Global Seasonal Predictions from a Machine Learning Weather Model Trained on Reanalysis Data , author =. npj Climate and Atmospheric Science , volume =. doi:10.1038/s41612-025-01198-3 , langid =

    work page doi:10.1038/s41612-025-01198-3

  28. [28]

    URL https://www.nature.com/articles/ s41586-024-07744-y

    Neural General Circulation Models for Weather and Climate , author =. Nature , volume =. doi:10.1038/s41586-024-07744-y , langid =

    work page doi:10.1038/s41586-024-07744-y

  29. [29]

    Nature Reviews Physics4(12), 761–769 (2022) arXiv:2204.01467 [cs.CY]

    On Scientific Understanding with Artificial Intelligence , author =. Nature Reviews Physics , volume =. doi:10.1038/s42254-022-00518-3 , langid =

    work page doi:10.1038/s42254-022-00518-3

  30. [30]

    Lai, Chieh-Hsin and Song, Yang and Kim, Dongjun and Mitsufuji, Yuki and Ermon, Stefano , year = 2025, month = oct, number =. The. doi:10.48550/arXiv.2510.21890 , archiveprefix =. 2510.21890 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.21890 2025

  31. [31]

    Science , volume =

    Learning Skillful Medium-Range Global Weather Forecasting , author =. Science , volume =

  32. [32]

    Liang, Jin and Yang, Xiu-Qun and Sun, De-Zheng , year = 2012, month = nov, journal =. The. doi:10.1175/JCLI-D-11-00490.1 , langid =

    work page doi:10.1175/jcli-d-11-00490.1 2012

  33. [33]

    Li, Tianhong and He, Kaiming , year = 2025, month = nov, number =. Back to. doi:10.48550/arXiv.2511.13720 , archiveprefix =. 2511.13720 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2511.13720 2025

  34. [34]

    Lipman, Yaron and Chen, Ricky T. Q. and. Flow. doi:10.48550/arXiv.2210.02747 , archiveprefix =. 2210.02747 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2210.02747

  35. [35]

    Liu, Xingchao and Gong, Chengyue and Liu, Qiang , year = 2022, month = sep, number =. Flow. doi:10.48550/arXiv.2209.03003 , archiveprefix =. 2209.03003 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2209.03003 2022

  36. [36]

    Nonlinear

    Liu, Yi and Cai, Wenju and Lin, Xiaopei and Li, Ziguang and Zhang, Ying , year = 2023, month = sep, journal =. Nonlinear. doi:10.1038/s41467-023-41551-9 , langid =

    work page doi:10.1038/s41467-023-41551-9 2023

  37. [37]

    Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models

    Lu, Cheng and Song, Yang , year = 2024, month = oct, number =. Simplifying,. doi:10.48550/arXiv.2410.11081 , archiveprefix =. 2410.11081 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410.11081 2024

  38. [38]

    Geophysical Research Letters , volume =

    A. Geophysical Research Letters , volume =. doi:10.1029/2025GL118649 , langid =

    work page doi:10.1029/2025gl118649

  39. [39]

    Geophysical Research Letters , volume =

    Observed. Geophysical Research Letters , volume =. doi:10.1029/2019GL082922 , langid =

    work page doi:10.1029/2019gl082922

  40. [40]

    Artificial Intelligence for Climate Prediction of Extremes:

    Materia, Stefano and Garc. Artificial Intelligence for Climate Prediction of Extremes:. WIREs Climate Change , volume =. doi:10.1002/wcc.914 , langid =

    work page doi:10.1002/wcc.914

  41. [41]

    and Zebiak, Stephen E

    McPhaden, Michael J. and Zebiak, Stephen E. and Glantz, Michael H. , year = 2006, month = dec, journal =

    2006

  42. [42]

    Meehan, Casey and Chaudhuri, Kamalika and Dasgupta, Sanjoy , year = 2020, month = apr, number =. A. doi:10.48550/arXiv.2004.05675 , archiveprefix =. 2004.05675 , primaryclass =

    work page doi:10.48550/arxiv.2004.05675 2020

  43. [43]

    Physical Review Fluids , volume =

    Probabilistic Forecasts of Extreme Heatwaves Using Convolutional Neural Networks in a Regime of Lack of Data , author =. Physical Review Fluids , volume =

  44. [44]

    and Battisti, David S

    Naylor, Rosamond L. and Battisti, David S. and Vimont, Daniel J. and Falcon, Walter P. and Burke, Marshall B. , year = 2007, month = may, journal =. Assessing Risks of Climate Variability and Climate Change for

    2007

  45. [45]

    Climate Dynamics , volume =

    An Empirical Model of Tropical Ocean Dynamics , author =. Climate Dynamics , volume =. doi:10.1007/s00382-011-1034-0 , langid =

    work page doi:10.1007/s00382-011-1034-0

  46. [46]

    and Grover, Aditya , year = 2023, month = dec, number =

    Nguyen, Tung and Brandstetter, Johannes and Kapoor, Ashish and Gupta, Jayesh K. and Grover, Aditya , year = 2023, month = dec, number =. 2301.10343 , publisher =

    arXiv 2023

  47. [47]

    npj Climate and Atmospheric Science , volume =

    Data-Driven Seasonal Climate Predictions via Variational Inference and Transformers , author =. npj Climate and Atmospheric Science , volume =. doi:10.1038/s41612-026-01320-z , langid =

    work page doi:10.1038/s41612-026-01320-z

  48. [48]

    and Goncalves, Andr

    Pan, Baoxiang and Anderson, Gemma J. and Goncalves, Andr. Improving. Journal of Advances in Modeling Earth Systems , volume =. doi:10.1029/2021MS002766 , archiveprefix =. 2010.14610 , publisher =

    work page doi:10.1029/2021ms002766 2010

  49. [49]

    FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators

    Pathak, Jaideep and Subramanian, Shashank and Harrington, Peter and Raja, Sanjeev and Chattopadhyay, Ashesh and Mardani, Morteza and Kurth, Thorsten and Hall, David and Li, Zongyi and Azizzadenesheli, Kamyar and Hassanzadeh, Pedram and Kashinath, Karthik and Anandkumar, Animashree , year = 2022, month = feb, number =. doi:10.48550/arXiv.2202.11214 , archi...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2202.11214 2022

  50. [50]

    Penland, C. A. Journal of Climate , volume =. doi:10.1175/1520-0442(1994)007<1352:ABCFSN>2.0.CO;2 , langid =

    work page doi:10.1175/1520-0442(1994)007 1994

  51. [51]

    Penland, C. The. Journal of Climate , volume =. doi:10.1175/1520-0442(1995)008<1999:TOGOTS>2.0.CO;2 , langid =

    work page doi:10.1175/1520-0442(1995)008 1995

  52. [52]

    Penland, Cecile , year = 1989, month = oct, journal =. Random. doi:10.1175/1520-0493(1989)117<2165:RFAFUP>2.0.CO;2 , langid =

    work page doi:10.1175/1520-0493(1989)117 1989

  53. [53]

    and El-Kadi, Andrew and Masters, Dominic and Ewalds, Timo and Stott, Jacklynn and Mohamed, Shakir and Battaglia, Peter and Lam, Remi and Willson, Matthew , date =

    Price, Ilan and. doi:10.48550/arXiv.2312.15796 , archiveprefix =. 2312.15796 , publisher =

    work page doi:10.48550/arxiv.2312.15796

  54. [54]

    Multi-Year Simulations at Kilometre Scale with the

    Rackow, Thomas and. Multi-Year Simulations at Kilometre Scale with the. Geoscientific Model Development , volume =. doi:10.5194/gmd-18-33-2025 , langid =

    work page doi:10.5194/gmd-18-33-2025 2025

  55. [55]

    and Barnes, Elizabeth A

    Rader, Jamin K. and Barnes, Elizabeth A. , year = 2023, journal =. Optimizing. doi:10.1029/2023GL104983 , langid =

    work page doi:10.1029/2023gl104983 2023

  56. [56]

    Rasp, Stephan and Lerch, Sebastian , year = 2018, month = nov, journal =. Neural. doi:10.1175/MWR-D-18-0187.1 , langid =

    work page doi:10.1175/mwr-d-18-0187.1 2018

  57. [57]

    Stephan Rasp, Stephan Hoyer, Aravind Merose, Johannes Langguth, Sebastian Deiser, et al

    Rasp, Stephan and Hoyer, Stephan and Merose, Alexander and Langmore, Ian and Battaglia, Peter and Russell, Tyler and. Journal of Advances in Modeling Earth Systems , volume =. doi:10.1029/2023MS004019 , langid =

    work page doi:10.1029/2023ms004019

  58. [58]

    , year = 2020, month = dec, journal =

    de Regt, Henk W. , year = 2020, month = dec, journal =. Understanding,. doi:10.1086/710520 , langid =

    work page doi:10.1086/710520 2020

  59. [59]

    Progressive Distillation for Fast Sampling of Diffusion Models

    Salimans, Tim and Ho, Jonathan , year = 2022, month = jun, number =. Progressive. doi:10.48550/arXiv.2202.00512 , archiveprefix =. 2202.00512 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2202.00512 2022

  60. [60]

    Schloer, Jakob and Newman, Matthew and Thuemmel, Jannik and Capotondi, Antonietta and Goswami, Bedartha , year = 2025, month = apr, number =. A. doi:10.48550/arXiv.2412.03743 , archiveprefix =. 2412.03743 , primaryclass =

    work page doi:10.48550/arxiv.2412.03743 2025

  61. [61]

    and Gagne, David John , year = 2025, month = may, journal =

    Sha, Yingkai and Sobash, Ryan A. and Gagne, David John , year = 2025, month = may, journal =. Improving. doi:10.1175/AIES-D-24-0063.1 , langid =

    work page doi:10.1175/aies-d-24-0063.1 2025

  62. [62]

    Deep Unsupervised Learning using Nonequilibrium Thermodynamics

    Deep. doi:10.48550/arXiv.1503.03585 , archiveprefix =. 1503.03585 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1503.03585

  63. [63]

    Generative Modeling by Estimating Gradients of the Data Distribution

    Song, Yang and Ermon, Stefano , year = 2020, month = oct, number =. Generative. doi:10.48550/arXiv.1907.05600 , archiveprefix =. 1907.05600 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1907.05600 2020

  64. [64]

    Song, Yang and. Score-. doi:10.48550/arXiv.2011.13456 , archiveprefix =. 2011.13456 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2011.13456 2011

  65. [65]

    doi:10.1029/2025MS005282 , langid =

    Springenberg, Maximilian and Otero, Noelia and Xue, Yuxin and Ma, Jackie , year = 2026, journal =. doi:10.1029/2025MS005282 , langid =

    work page doi:10.1029/2025ms005282 2026

  66. [66]

    Stock, Jason and Arcomano, Troy and Kotamarthi, Rao , year = 2025, month = sep, number =. Swift:. doi:10.48550/arXiv.2509.25631 , archiveprefix =. 2509.25631 , primaryclass =

    work page doi:10.48550/arxiv.2509.25631 2025

  67. [67]

    A Fresh Perspective on

    Thual, Sulian and Dewitte, Boris , year = 2025, month = nov, journal =. A Fresh Perspective on. doi:10.1038/s41612-025-01233-3 , langid =

    work page doi:10.1038/s41612-025-01233-3 2025

  68. [68]

    and Chen, Nan and Stechmann, Samuel N

    Thual, Sulian and Majda, Andrew J. and Chen, Nan and Stechmann, Samuel N. , year = 2016, month = sep, journal =. Simple Stochastic Model for. doi:10.1073/pnas.1612002113 , langid =

    work page doi:10.1073/pnas.1612002113 2016

  69. [69]

    npj Climate and Atmospheric Science , publisher =

    Large Potential of Performance-Based Model Weighting to Improve Decadal Climate Forecast Skill , author =. npj Climate and Atmospheric Science , publisher =. doi:10.1038/s41612-026-01397-6 , langid =

    work page doi:10.1038/s41612-026-01397-6

  70. [70]

    Wang, Chunzai and Deser, Clara and Yu, Jin-Yi and DiNezio, Pedro and Clement, Amy , editor =. El. Coral. doi:10.1007/978-94-017-7499-4_4 , langid =

    work page doi:10.1007/978-94-017-7499-4_4

  71. [71]

    doi:10.1038/s41612-025-01090-0 , langid =

    npj Climate and Atmospheric Science , volume =. doi:10.1038/s41612-025-01090-0 , langid =

    work page doi:10.1038/s41612-025-01090-0

  72. [72]

    , year = 2000, journal =

    Yeo, In-Kwon and Johnson, Richard A. , year = 2000, journal =. A. 2673623 , eprinttype =

    2000