pith. sign in

arxiv: 2606.26389 · v1 · pith:OYYXKFPOnew · submitted 2026-06-24 · ⚛️ physics.ao-ph · cs.AI

Sampling sea state using a diffusion model

Jiarong Wu , Bertrand Chapron , Laure Zanna This is my paper

Pith reviewed 2026-06-26 00:34 UTC · model grok-4.3

classification ⚛️ physics.ao-ph cs.AI
keywords diffusion modelsea statewave forecastingprobabilistic predictiongenerative modelwind forcingWAVEWATCH-III
0
0 comments X

The pith

A diffusion model samples the full distribution of global sea state directly from five days of wind forcing.

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

The paper introduces a generative diffusion model that takes a five-day history of global wind forcing and produces samples from the conditional distribution of sea state. This replaces the step-by-step integration used in traditional spectral wave models. The same model handles not only bulk variables such as significant wave height but also partition variables and derived fields including Stokes drift and mean square slope. Training on thirty years of WAVEWATCH-III hindcast data yields predictions whose skill and ensemble spread match the reference simulations at far lower computational cost. The approach therefore supplies a route to probabilistic wave information that can be coupled into larger earth-system calculations.

Core claim

A diffusion-based generative model conditioned on a five-day history of global wind forcing directly samples the complex conditional distribution of sea state without autoregressive time-stepping, delivering skillful predictions and a calibrated ensemble spread for bulk variables while extending naturally to partition-related variables and derived quantities such as Stokes drift and mean square slope.

What carries the argument

Diffusion-based generative model that conditions on five-day global wind forcing to sample the conditional sea-state distribution in one forward pass.

If this is right

  • Probabilistic sea-state forecasts become feasible at a fraction of the cost of ensemble spectral runs.
  • Sea-state information, including derived quantities, can be supplied to coupled climate or ocean models without prohibitive runtime overhead.
  • Partition variables and Stokes drift become available as direct outputs rather than post-processed additions.
  • Calibrated uncertainty estimates accompany every bulk-variable prediction.

Where Pith is reading between the lines

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

  • The method may allow wave effects to be included in century-scale climate integrations where current wave models are too expensive.
  • Real-time operational systems could generate large ensembles for risk assessment during storms.
  • Similar conditioning on forcing histories might be tested on other slow-response geophysical fields such as sea-ice concentration or ocean mixed-layer depth.

Load-bearing premise

A five-day history of global wind forcing alone is enough for the model to capture the full conditional distribution of sea state without further inputs or iterative stepping.

What would settle it

Ensemble members drawn from the model produce a spread that systematically under- or over-represents the variability seen in independent high-resolution wave simulations for wind histories outside the training distribution.

read the original abstract

Sea state prediction is essential for operational maritime applications and coupled earth system modeling, yet current spectral wave models remain computationally prohibitive for many use cases, including online coupling to climate simulations and making probabilistic (ensemble-based) predictions. While deep learning has recently demonstrated strong performance in weather forecasting, existing AI-based wave models are predominantly deterministic and largely limited to bulk variables such as significant wave height, leaving probabilistic sea state estimation largely unexplored. In this work, we propose a diffusion-based generative model for global sea state estimation that conditions on a relatively long history (5 days) of global wind forcing. This generative model directly samples the complex conditional distribution of sea state without autoregressive time-stepping. Unlike prior approaches, our framework naturally extends beyond bulk variables to estimate partition-related variables and derived quantities, such as Stokes drift and mean square slope. Trained on a 30-year global WAVEWATCH-III hindcast, the model achieves substantial computational acceleration compared with numerical spectral models while delivering skillful predictions and a calibrated ensemble spread for the bulk variables. Our results suggest that diffusion-based sea state sampling offers a promising path toward probabilistic wave forecasting and efficient coupling of sea state information into broader earth system models.

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 / 2 minor

Summary. The manuscript proposes a diffusion-based generative model for global sea state sampling that conditions on a 5-day history of global wind forcing. Trained on a 30-year WAVEWATCH-III hindcast, the model directly samples the conditional distribution of sea state (including partitions and derived quantities such as Stokes drift) without autoregressive stepping, claiming substantial computational acceleration over numerical spectral models along with skillful predictions and calibrated ensemble spread for bulk variables.

Significance. If the central claims hold with rigorous validation, the approach could enable efficient probabilistic wave forecasting and coupling of sea-state information into earth-system models where traditional spectral models are computationally prohibitive.

major comments (2)
  1. [Abstract and Results] Abstract and Results: The abstract asserts 'skillful predictions and a calibrated ensemble spread' without supplying quantitative metrics (skill scores, error bars, validation splits, or baseline comparisons). This information is load-bearing for the headline claim and must be provided with explicit numbers and statistical tests.
  2. [Methods] Methods: The model relies on conditioning solely on a 5-day wind history without autoregressive stepping or additional inputs. No ablation on history length, no verification against longer-memory regimes (e.g., swell propagation), and no comparison of generated ensemble statistics to the numerical model's conditional distribution are described; this directly affects whether the sampled distribution is correctly specified.
minor comments (2)
  1. Notation for partition variables and derived quantities (Stokes drift, mean square slope) should be defined explicitly on first use and cross-referenced to the training dataset variables.
  2. Figure captions should include the exact validation period, ensemble size, and any baseline models shown for direct comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive comments. We address each major point below and agree that revisions are needed to strengthen the presentation of quantitative results and methodological details.

read point-by-point responses

  1. Referee: [Abstract and Results] Abstract and Results: The abstract asserts 'skillful predictions and a calibrated ensemble spread' without supplying quantitative metrics (skill scores, error bars, validation splits, or baseline comparisons). This information is load-bearing for the headline claim and must be provided with explicit numbers and statistical tests.

    Authors: We agree that the abstract would be strengthened by including explicit quantitative metrics. While the results section reports skill scores (e.g., RMSE and correlation for significant wave height), ensemble calibration diagnostics, validation splits, and baseline comparisons against the WAVEWATCH-III hindcast, these details are not summarized numerically in the abstract. We will revise the abstract to incorporate key metrics with error bars and note the statistical validation approach. revision: yes

  2. Referee: [Methods] Methods: The model relies on conditioning solely on a 5-day wind history without autoregressive stepping or additional inputs. No ablation on history length, no verification against longer-memory regimes (e.g., swell propagation), and no comparison of generated ensemble statistics to the numerical model's conditional distribution are described; this directly affects whether the sampled distribution is correctly specified.

    Authors: The 5-day history was selected to cover typical global swell propagation timescales. We acknowledge that an ablation on history length is absent and will add this analysis. We will also include verification for longer-memory swell regimes by examining cases with distant wind sources and add direct comparisons of generated ensemble statistics (means, variances, and partition distributions) against the numerical model's conditional distribution to confirm correct specification. revision: yes

Circularity Check

0 steps flagged

No significant circularity; model trained and evaluated on external hindcast

full rationale

The paper trains a diffusion-based generative model on an external 30-year WAVEWATCH-III hindcast dataset, conditions explicitly on 5-day global wind forcing, and reports predictive skill plus ensemble calibration on held-out data. No equations or claims reduce the sampled sea-state distribution to fitted parameters defined by the same inputs, no self-citation chains justify uniqueness or ansatzes, and no renaming of known results occurs. The central claim rests on standard supervised training and evaluation against an independent numerical model, making the derivation self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the representativeness of the 30-year WAVEWATCH-III hindcast for real-world conditional distributions and on the unstated assumption that standard diffusion training objectives suffice for high-dimensional wave spectra.

free parameters (1)
  • conditioning history length
    The choice of 5 days is presented as relatively long but is a modeling decision that affects what the conditional distribution can capture.
axioms (1)
  • domain assumption Diffusion models can faithfully sample complex, high-dimensional conditional distributions arising from physical wave dynamics.
    Invoked implicitly when claiming the model samples the true conditional distribution of sea state.

pith-pipeline@v0.9.1-grok · 5731 in / 1171 out tokens · 17411 ms · 2026-06-26T00:34:39.806354+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

40 extracted references

  1. [1]

    Janssen, P.The Interaction of Ocean Waves and Wind1 edn (Cambridge University Press, 2004). 17

    2004

  2. [2]

    & Hemer, M

    Cavaleri, L., Fox-Kemper, B. & Hemer, M. Wind Waves in the Coupled Climate System.Bulletin of the American Meteorological Society93, 1651–1661 (2012)

    2012

  3. [3]

    E.et al.Propagation of ocean swell across the Pacific.Philosoph- ical Transactions of the Royal Society of London

    Snodgrass, F. E.et al.Propagation of ocean swell across the Pacific.Philosoph- ical Transactions of the Royal Society of London. Series A, Mathematical and Physical Sciences259, 431–497 (1966)

    1966

  4. [4]

    Kitaigorodskii, S. A. Application of the theory of similarity to the analysis of wind-generated wave motion as a stochastic process.Izv., Geophys. Ser. Acad. Sci., USSR1, 105–117 (1962)

    1962

  5. [5]

    & Skiba, E

    Dulov, V., Kudryavtsev, V. & Skiba, E. On fetch- and duration-limited wind wave growth: Data and parametric model.Ocean Modelling153, 101676 (2020)

    2020

  6. [6]

    & Chapron, B

    Yurovskaya, M., Kudryavtsev, V. & Chapron, B. A self-similar description of the wave fields generated by tropical cyclones.Ocean Modelling183(2023)

    2023

  7. [7]

    (WW3DG), T. W. I. D. G. Wave Watch Manual (2016)

    2016

  8. [8]

    & Chapron, B

    Hell, M., Fox-Kemper, B. & Chapron, B. A Particle-in-Cell Wave Model for Efficient Sea-State Estimates in Earth System Models—PiCLES.Journal of Advances in Modeling Earth Systems17, e2025MS005221 (2025)

    2025

  9. [9]

    & Chapron, B

    Kudryavtsev, V., Yurovskaya, M. & Chapron, B. 2D Parametric Model for Surface Wave Development Under Varying Wind Field in Space and Time.Journal of Geophysical Research: Oceans126, e2020JC016915 (2021)

    2021

  10. [10]

    Kochkov, D.et al.Neural General Circulation Models for Weather and Climate (2024)

    2024

  11. [11]

    Bodnar, C.et al.A Foundation Model for the Earth System (2024)

    2024

  12. [12]

    & Jiang, H

    Wang, X. & Jiang, H. Physics-guided deep learning for skillful wind-wave modeling.Science Advances10, eadr3559 (2024)

    2024

  13. [13]

    Wang, X.et al.Data-driven rolling model for global wave height.Geoscientific Model Development18, 5101–5114 (2025)

    2025

  14. [14]

    Zhang, Z.et al.Ocean Wave Forecasting With Deep Learning as Alternative to Conventional Models.Journal of Advances in Modeling Earth Systems17, e2025MS005285 (2025)

    2025

  15. [15]

    Hahner, S.et al.Representing the Surface Ocean in ECMWF’s data-driven forecasting system AIFS (2026)

    2026

  16. [16]

    & Ardhuin, F

    Rascle, N. & Ardhuin, F. A global wave parameter database for geophysical appli- cations. Part 2: Model validation with improved source term parameterization. Ocean Modelling70, 174–188 (2013). 18

    2013

  17. [17]

    & Dodet, G

    Alday, M., Accensi, M., Ardhuin, F. & Dodet, G. A global wave parameter database for geophysical applications. Part 3: Improved forcing and spectral resolution.Ocean Modelling166, 101848 (2021)

    2021

  18. [18]

    & Palmer, T

    Leutbecher, M. & Palmer, T. Ensemble forecasting.Journal of Computational Physics227, 3515–3539 (2008)

    2008

  19. [19]

    Price, I.et al.Probabilistic weather forecasting with machine learning.Nature 637, 84–90 (2025)

    2025

  20. [20]

    Lang, S.et al.AIFS-CRPS: ensemble forecasting using a model trained with a loss function based on the continuous ranked probability score.npj Artificial Intelligence2, 18 (2026)

    2026

  21. [21]

    Song, Y.et al.Score-Based Generative Modeling through Stochastic Differential Equations (2021)

    2021

  22. [22]

    & Abbeel, P

    Ho, J., Jain, A. & Abbeel, P. Denoising Diffusion Probabilistic Models (2020)

    2020

  23. [23]

    & Louppe, G

    Rozet, F. & Louppe, G. Score-based data assimilation.Advances in Neural Information Processing Systems36(2023)

    2023

  24. [24]

    D.et al.Climate in a Bottle: Towards a Generative Foundation Model for the Kilometer-Scale Global Atmosphere (2025)

    Brenowitz, N. D.et al.Climate in a Bottle: Towards a Generative Foundation Model for the Kilometer-Scale Global Atmosphere (2025)

    2025

  25. [25]

    A., Hegermiller, C., Teicheira, C

    Houghton, I. A., Hegermiller, C., Teicheira, C. & Smit, P. B. Operational Assim- ilation of Spectral Wave Data From the Sofar Spotter Network.Geophysical Research Letters49, e2022GL098973 (2022)

    2022

  26. [26]

    Badulin, S. I. A physical model of sea wave period from altimeter data.Journal of Geophysical Research: Oceans119, 856–869 (2014)

    2014

  27. [27]

    Local balance in the air-sea boundary processes.Journal of Oceanography28, 109–120 (1972)

    Toba, Y. Local balance in the air-sea boundary processes.Journal of Oceanography28, 109–120 (1972)

    1972

  28. [28]

    Hanson, J. L. & Phillips, O. M. Automated Analysis of Ocean Surface Directional Wave Spectra.Journal of Atmospheric and Oceanic Technology18, 277–293 (2001)

    2001

  29. [29]

    Journal of Geophysical Research: Oceans117(2012)

    Cavaleri, L.et al.Rogue waves in crossing seas: The Louis Majesty accident. Journal of Geophysical Research: Oceans117(2012)

    2012

  30. [30]

    & Benetazzo, A

    Cavaleri, L., Barbariol, F. & Benetazzo, A. Wind–Wave Modeling: Where We Are, Where to Go.Journal of Marine Science and Engineering8, 260 (2020)

    2020

  31. [31]

    S.et al.inTransfer Across the Air-Sea Interface(eds Liss, P

    Garbe, C. S.et al.inTransfer Across the Air-Sea Interface(eds Liss, P. S. & Johnson, M. T.)Ocean-Atmosphere Interactions of Gases and Particles55–112 (Springer, Berlin, Heidelberg, 2014). 19

    2014

  32. [32]

    An inconvenient sea truth: spread, steepness, and skewness of surface slopes.Annual Review of Marine Science1, 377–415 (2009)

    Munk, W. An inconvenient sea truth: spread, steepness, and skewness of surface slopes.Annual Review of Marine Science1, 377–415 (2009)

    2009

  33. [33]

    L., Tracy, B

    Hanson, J. L., Tracy, B. A., Tolman, H. L. & Scott, R. D. Pacific hindcast per- formance of three numerical wave models.Journal of Atmospheric and Oceanic Technology(2009)

    2009

  34. [34]

    & Fox-Kemper, B

    Webb, A. & Fox-Kemper, B. Wave spectral moments and Stokes drift estimation. Ocean Modelling40, 273–288 (2011)

    2011

  35. [35]

    & Yuan, C

    Permenter, F. & Yuan, C. Interpreting and Improving Diffusion Models from an Optimization Perspective.Proceedings of Machine Learning Research235, 40461–40483 (2024)

    2024

  36. [36]

    & Laine, S

    Karras, T., Aittala, M., Aila, T. & Laine, S. Elucidating the Design Space of Diffusion-Based Generative Models (2022)

    2022

  37. [37]

    Noise schedules considered harmful (2024)

    Dieleman, S. Noise schedules considered harmful (2024)

    2024

  38. [38]

    Brier, G. W. Verification of forecasts expressed in terms of probability.Monthly Weather Review(1950)

    1950

  39. [39]

    & Hasselmann, K

    Hasselmann, S., Lionello, P. & Hasselmann, K. An optimal interpolation scheme for the assimilation of spectral wave data.Journal of Geophysical Research: Oceans102, 15823–15836 (1997)

    1997

  40. [40]

    & Chapron, B

    Collard, F., Ardhuin, F. & Chapron, B. Monitoring and analysis of ocean swell fields from space: New methods for routine observations.Journal of Geophysical Research: Oceans114(2009). 20

    2009