pith. sign in

arxiv: 2605.24009 · v1 · pith:JQJCXSXAnew · submitted 2026-05-19 · ⚛️ physics.ao-ph · cs.LG

Improving Ensemble CAPE Forecasts with a Diffusion Model Incorporating Aerosol Information

Zachary James , Joseph Guinness , Arthur DeGaetano This is my paper

Pith reviewed 2026-06-30 17:23 UTC · model grok-4.3

classification ⚛️ physics.ao-ph cs.LG
keywords CAPEdiffusion modelaerosol optical depthensemble forecastingGFSsevere weathermachine learning
0
0 comments X

The pith

A diffusion model that adds aerosol data to GFS inputs generates ensemble CAPE forecasts outperforming both GFS and GEFS on standard metrics.

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

The authors train a diffusion model to reduce summertime underestimation of convective available potential energy in operational forecasts. The model accepts a GFS CAPE field as primary input and produces an ensemble whose members score lower on root mean square error and better on continuous ranked probability score and Brier score than the raw GFS or GEFS products. Adding aerosol optical depth fields for black carbon, organic carbon, dust, sea salt, and sulfates yields further gains. Classifier-free guidance lets the user trade off ensemble skill against spread. Permutation importance ranks black carbon, organic carbon, and sulfate aerosols as having larger effects on the output than the other two species.

Core claim

The diffusion model takes a GFS CAPE forecast as input and outputs an ensemble that significantly outperforms both GFS and GEFS 6-hour forecasts on root mean square error, continuous ranked probability score, and Brier score. A two-stage training pipeline combines a large historical GFS dataset with a smaller GEFS dataset. Adding aerosol optical depths as extra inputs further improves the forecasts, and black carbon, organic carbon, and sulfate aerosols exert greater influence than sea salt or dust.

What carries the argument

Two-stage diffusion model with classifier-free guidance that ingests GFS CAPE forecasts plus five aerosol optical depth fields and produces calibrated ensemble members.

If this is right

  • The generated ensembles can be fed directly into downstream severe-weather guidance products.
  • Classifier-free guidance provides a tunable knob between forecast accuracy and ensemble dispersion.
  • Permutation importance ranks black carbon, organic carbon, and sulfate aerosols as stronger drivers of CAPE predictions than sea salt or dust.
  • The same input-output structure can accept additional gridded fields without retraining the entire pipeline.

Where Pith is reading between the lines

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

  • The framework could be tested on other convective indices such as lifted index or storm-relative helicity.
  • Operational systems could ingest near-real-time aerosol analyses to produce updated CAPE ensembles each cycle.
  • The two-stage approach may generalize to other forecast variables where a high-resolution deterministic run and a lower-resolution ensemble both exist.

Load-bearing premise

A two-stage training pipeline can combine a large GFS forecast archive with a smaller GEFS archive even though the two systems use different initialization and parameterization schemes that change over time.

What would settle it

Evaluation on a withheld period after a major GFS or GEFS model upgrade in which the diffusion ensemble no longer shows lower RMSE or higher CRPS than the raw GFS or GEFS output.

read the original abstract

Convective available potential energy (CAPE) is an important variable for forecasting severe weather and understanding deep convection and precipitation. The latest versions of the Global Forecast System (GFS) and related Global Ensemble Forecast System (GEFS) have exhibited a bias towards underestimating CAPE values during the summertime. We train an artificial intelligence (AI) diffusion model to improve the skill and uncertainty quantification of afternoon 6-hour lead time ensemble forecasts over the United States. Our model takes a GFS CAPE forecast as input and outputs an ensemble that significantly outperforms both GFS and GEFS 6-hour forecasts on root mean square error, continuous ranked probability score, and Brier score. We propose a two-stage training pipeline to leverage both a larger historical GFS forecast dataset and a smaller historical GEFS dataset, despite the two using initialization and parameterization schemes that vary over time. We also show that classifier-free guidance can be used to control the skill and spread of the forecasts. We then demonstrate the versatility of our framework by adding aerosol optical depths (AODs) of black carbon, organic carbon, dust, sea salt, and sulfates as additional input features. Aerosols can invigorate or suppress convection depending on atmospheric conditions. Our AI models effectively incorporate aerosols to produce improved CAPE forecasts. We interpret the model components by using permutation feature importance to rank the influence of the different AODs and find that black carbon, organic carbon, and sulfate aerosols have a greater impact on the model's CAPE predictions than sea salt and dust aerosols.

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

3 major / 2 minor

Summary. The manuscript presents a diffusion model that takes GFS CAPE forecasts as input and generates ensemble forecasts for 6-hour afternoon lead times over the United States. It claims these ensembles significantly outperform both GFS and GEFS baselines on RMSE, CRPS, and Brier score. A two-stage training pipeline is proposed to combine a large historical GFS dataset with a smaller GEFS dataset despite differing initialization and parameterization schemes. Classifier-free guidance is used to control forecast skill and spread. The model is extended to incorporate aerosol optical depth (AOD) inputs for black carbon, organic carbon, dust, sea salt, and sulfates; permutation feature importance is then applied to rank their influence on CAPE predictions.

Significance. If the reported skill improvements and aerosol effects are robust, the work could contribute to operational severe-weather forecasting by improving ensemble uncertainty quantification and by demonstrating a practical route to include aerosol-convection interactions in AI post-processing. The classifier-free guidance mechanism offers a controllable way to trade off bias and spread that may be reusable in other ensemble post-processing settings.

major comments (3)
  1. [Methods (two-stage training pipeline)] Methods (two-stage training pipeline): The pipeline is asserted to leverage both GFS and GEFS data despite time-varying schemes, yet no domain-adaptation layer, scheme-specific covariates, or time-aware conditioning is described. Without these, measured gains on RMSE/CRPS/Brier (and the subsequent AOD permutation ranking) could arise from unmodeled distribution shift rather than genuine aerosol signal or model architecture.
  2. [Results (aerosol incorporation)] Results (aerosol incorporation): The claim that “AI models effectively incorporate aerosols to produce improved CAPE forecasts” rests on the addition of five AOD fields and a permutation-importance ranking, but the manuscript does not report an ablation that isolates the contribution of the AOD inputs relative to the base diffusion model trained without aerosols.
  3. [Results (performance metrics)] Results (performance metrics): The abstract states significant outperformance on three scores, yet the manuscript must supply the actual numerical deltas, ensemble size, validation-split details, and error bars so that the magnitude and statistical reliability of the claimed improvements can be evaluated.
minor comments (2)
  1. [Abstract] Abstract: the phrase “significantly outperforms” should be accompanied by the key numerical improvements so readers can judge effect size without consulting the full text.
  2. [Notation] Notation: ensure AOD and CAPE are defined at first use in the main text and that all aerosol species are listed consistently between abstract and methods.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough review and valuable comments. We address each of the major comments below and will revise the manuscript accordingly to improve clarity and completeness.

read point-by-point responses

  1. Referee: Methods (two-stage training pipeline): The pipeline is asserted to leverage both GFS and GEFS data despite time-varying schemes, yet no domain-adaptation layer, scheme-specific covariates, or time-aware conditioning is described. Without these, measured gains on RMSE/CRPS/Brier (and the subsequent AOD permutation ranking) could arise from unmodeled distribution shift rather than genuine aerosol signal or model architecture.

    Authors: We agree that additional details on the two-stage training pipeline are needed to clarify how it handles the differences in GFS and GEFS schemes. The pipeline pre-trains the diffusion model on the larger historical GFS dataset and then fine-tunes it on the GEFS dataset to adapt to the specific ensemble characteristics. This sequential approach implicitly addresses distribution shifts through the fine-tuning stage. To address the referee's concern, we will expand the methods section to describe the pipeline in more detail, including data selection criteria and training hyperparameters that help mitigate scheme variations. We maintain that the reported improvements are due to the model architecture rather than unmodeled shifts, but the added description will strengthen this claim. revision: yes

  2. Referee: Results (aerosol incorporation): The claim that “AI models effectively incorporate aerosols to produce improved CAPE forecasts” rests on the addition of five AOD fields and a permutation-importance ranking, but the manuscript does not report an ablation that isolates the contribution of the AOD inputs relative to the base diffusion model trained without aerosols.

    Authors: We acknowledge the value of an ablation study to isolate the contribution of the AOD inputs. While the manuscript shows that adding the five AOD fields leads to improved forecasts and uses permutation importance to rank their effects, a direct comparison to the base model without aerosols is not explicitly reported. We will include this ablation in the revised manuscript, providing performance metrics for the model with and without AOD inputs to quantify the additional gains from aerosol information. revision: yes

  3. Referee: Results (performance metrics): The abstract states significant outperformance on three scores, yet the manuscript must supply the actual numerical deltas, ensemble size, validation-split details, and error bars so that the magnitude and statistical reliability of the claimed improvements can be evaluated.

    Authors: We appreciate this suggestion for improving the presentation of results. The manuscript includes comparative figures, but we will add explicit numerical values for the improvements in RMSE, CRPS, and Brier score, specify the ensemble size (e.g., number of members generated), detail the validation split used (such as the time periods or geographic regions), and include error bars or statistical significance measures. These details will be incorporated into the results section and, where feasible, the abstract. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper describes an empirical machine-learning pipeline (diffusion model trained in two stages on GFS then GEFS data, with optional AOD inputs) whose central claims are performance improvements measured against external, independent baselines (GFS and GEFS forecasts) on standard metrics (RMSE, CRPS, Brier). No equations, fitted parameters, or self-citations are presented that reduce the reported skill gains or aerosol rankings to the inputs by construction; the two-stage training is a methodological choice whose validity can be tested externally rather than a definitional loop. The derivation chain is therefore self-contained against the stated benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no free parameters, axioms, or invented entities are identifiable from the provided text.

pith-pipeline@v0.9.1-grok · 5808 in / 1131 out tokens · 42872 ms · 2026-06-30T17:23:50.720573+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

35 extracted references · 28 canonical work pages · 1 internal anchor

  1. [1]

    Dong, W., Y. Lin, J. S. Wright, Y. Xie, X. Yin, and J. Guo, 2019: Precipitable water and cape dependence of rainfall intensities in china. Climate Dynamics, 52 (5), 3357--3368, doi:10.1007/s00382-018-4327-8

    work page doi:10.1007/s00382-018-4327-8 2019

  2. [2]

    Ferro, C. A. T., 2014: Fair scores for ensemble forecasts. Quarterly Journal of the Royal Meteorological Society, 140 (683), 1917--1923, doi:https://doi.org/10.1002/qj.2270, https://rmets.onlinelibrary.wiley.com/doi/pdf/10.1002/qj.2270

    work page doi:10.1002/qj.2270 2014

  3. [3]

    Abaza, F

    Fortin, V., M. Abaza, F. Anctil, and R. Turcotte, 2014: Why should ensemble spread match the rmse of the ensemble mean? Journal of Hydrometeorology, 15 (4), 1708 -- 1713, doi:10.1175/JHM-D-14-0008.1

    work page doi:10.1175/jhm-d-14-0008.1 2014

  4. [4]

    Ginoux, P., M. Chin, I. Tegen, J. M. Prospero, B. Holben, O. Dubovik, and S.-J. Lin, 2001: Sources and distributions of dust aerosols simulated with the gocart model. Journal of Geophysical Research: Atmospheres, 106 (D17), 20\,255--20\,273, doi:https://doi.org/10.1029/2000JD000053, https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/2000JD000053

    work page doi:10.1029/2000jd000053 2001

  5. [5]

    Goddard Earth Sciences Data and Information Services Center ( GES DISC ), Greenbelt, MD, USA, doi:10.5067/KLICLTZ8EM9D

    GMAO , 2015: MERRA-2 tavg1\_2d\_aer\_nx: 2d, 1-hourly, time-averaged, single-level, assimilation, aerosol diagnostics v5.12.4. Goddard Earth Sciences Data and Information Services Center ( GES DISC ), Greenbelt, MD, USA, doi:10.5067/KLICLTZ8EM9D

    work page doi:10.5067/klicltz8em9d 2015

  6. [6]

    Katzfuss, 2014: Probabilistic forecasting

    Gneiting, T., and M. Katzfuss, 2014: Probabilistic forecasting. Annual Review of Statistics and Its Application, 1 (Volume 1, 2014), 125--151, doi:https://doi.org/10.1146/annurev-statistics-062713-085831

    work page doi:10.1146/annurev-statistics-062713-085831 2014

  7. [7]

    Boucher, 2000: Estimates of the direct and indirect radiative forcing due to tropospheric aerosols: A review

    Haywood, J., and O. Boucher, 2000: Estimates of the direct and indirect radiative forcing due to tropospheric aerosols: A review. Reviews of Geophysics, 38 (4), 513--543, doi:https://doi.org/10.1029/1999RG000078, https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/1999RG000078

    work page doi:10.1029/1999rg000078 2000

  8. [8]

    Jain, and P

    Ho, J., A. Jain, and P. Abbeel, 2020: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., Curran Associates, Inc., Vol. 33, 6840--6851, ://proceedings.neurips.cc/paper_files/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf

    2020

  9. [9]

    Classifier-Free Diffusion Guidance

    Ho, J., and T. Salimans, 2021: Classifier-free diffusion guidance. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, ://arxiv.org/abs/2207.12598

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  10. [10]

    Atmospheric Environment, 51, 165--174, doi:https://doi.org/10.1016/j.atmosenv.2012.01.021

    Huang, M., and Coauthors, 2012: Sectoral and geographical contributions to summertime continental united states (conus) black carbon spatial distributions. Atmospheric Environment, 51, 165--174, doi:https://doi.org/10.1016/j.atmosenv.2012.01.021

    work page doi:10.1016/j.atmosenv.2012.01.021 2012

  11. [11]

    Jiang, J. H., H. Su, L. Huang, Y. Wang, S. Massie, B. Zhao, A. Omar, and Z. Wang, 2018: Contrasting effects on deep convective clouds by different types of aerosols. Nature Communications, 9 (1), 3874, doi:10.1038/s41467-018-06280-4

    work page doi:10.1038/s41467-018-06280-4 2018

  12. [12]

    Johnson, B. T., J. M. Haywood, and M. K. Hawcroft, 2019: Are changes in atmospheric circulation important for black carbon aerosol impacts on clouds, precipitation, and radiation? Journal of Geophysical Research: Atmospheres, 124 (14), 7930--7950, doi:https://doi.org/10.1029/2019JD030568, https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/2019JD030568

    work page doi:10.1029/2019jd030568 2019

  13. [13]

    S., and J

    Kain, J. S., and J. M. Fritsch, 1993: Convective Parameterization for Mesoscale Models: The Kain-Fritsch Scheme, 165--170. American Meteorological Society, Boston, MA, doi:10.1007/978-1-935704-13-3_16, ://doi.org/10.1007/978-1-935704-13-3_16

    work page doi:10.1007/978-1-935704-13-3_16 1993

  14. [14]

    Diendorfer, and N

    Kaltenböck, R., G. Diendorfer, and N. Dotzek, 2009: Evaluation of thunderstorm indices from ecmwf analyses, lightning data and severe storm reports. Atmospheric Research, 93 (1), 381--396, doi:https://doi.org/10.1016/j.atmosres.2008.11.005

    work page doi:10.1016/j.atmosres.2008.11.005 2009

  15. [15]

    T., and B

    Kiehl, J. T., and B. P. Briegleb, 1993: The relative roles of sulfate aerosols and greenhouse gases in climate forcing. Science, 260 (5106), 311--314, doi:10.1126/science.260.5106.311, https://www.science.org/doi/pdf/10.1126/science.260.5106.311

    work page doi:10.1126/science.260.5106.311 1993

  16. [16]

    Natural Hazards and Earth System Sciences, 7 (2), 327--342, doi:10.5194/nhess-7-327-2007

    Kunz, M., 2007: The skill of convective parameters and indices to predict isolated and severe thunderstorms. Natural Hazards and Earth System Sciences, 7 (2), 327--342, doi:10.5194/nhess-7-327-2007

    work page doi:10.5194/nhess-7-327-2007 2007

  17. [17]

    Learning skillful medium-range global weather forecasting

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

    work page doi:10.1126/science.adi2336 2023

  18. [18]

    Veneziano, and A

    Lepore, C., D. Veneziano, and A. Molini, 2015: Temperature and cape dependence of rainfall extremes in the eastern united states. Geophysical Research Letters, 42 (1), 74--83, doi:https://doi.org/10.1002/2014GL062247, https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1002/2014GL062247

    work page doi:10.1002/2014gl062247 2015

  19. [19]

    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, https://www.science.org/doi/pdf/10.1126/sciadv.adk4489

    work page doi:10.1126/sciadv.adk4489 2024

  20. [20]

    Hutter, 2019: Decoupled weight decay regularization

    Loshchilov, I., and F. Hutter, 2019: Decoupled weight decay regularization. 7th International Conference on Learning Representations, ICLR 2019 , OpenReview.net, New Orleans, LA, USA, ://openreview.net/forum?id=Bkg6RiCqY7

    2019

  21. [21]

    Lu, C., Y. Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, 2022: Dpm-solver: a fast ode solver for diffusion probabilistic model sampling in around 10 steps. Proceedings of the 36th International Conference on Neural Information Processing Systems, Curran Associates Inc., Red Hook, NY, USA, NIPS '22

    2022

  22. [22]

    Lu, C., Y. Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, 2025: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. Machine Intelligence Research, 22 (4), 730–751, doi:10.1007/s11633-025-1562-4

    work page doi:10.1007/s11633-025-1562-4 2025

  23. [23]

    Malloy, K., and M. K. Tippett, 2025: Forecasting u.s. tornado outbreak activity and associated environments in the global ensemble forecast system (gefs). Weather and Forecasting, 40 (4), 593 -- 608, doi:10.1175/WAF-D-24-0138.1

    work page doi:10.1175/waf-d-24-0138.1 2025

  24. [24]

    3rd ed., ://christophm.github.io/interpretable-ml-book

    Molnar, C., 2025: Interpretable Machine Learning. 3rd ed., ://christophm.github.io/interpretable-ml-book

    2025

  25. [25]

    ://www.emc.ncep.noaa.gov/emc/pages/numerical_forecast_systems/gefs.php, accessed: 2026-03-30, NOAA/NWS/NCEP Environmental Modeling Center

    NOAA Environmental Modeling Center , 2024: Global ensemble forecast system ( GEFS ). ://www.emc.ncep.noaa.gov/emc/pages/numerical_forecast_systems/gefs.php, accessed: 2026-03-30, NOAA/NWS/NCEP Environmental Modeling Center

    2024

  26. [26]

    Robins, Andrea Rotnitzky, and Lue Ping Zhao

    Politis, D. N., and J. P. Romano, 1994: The stationary bootstrap. Journal of the American Statistical Association, 89 (428), 1303--1313, doi:10.1080/01621459.1994.10476870, https://doi.org/10.1080/01621459.1994.10476870

    work page doi:10.1080/01621459.1994.10476870 1994

  27. [27]

    Nature, 637 (8044), 84--90, doi:10.1038/s41586-024-08252-9

    Price, I., and Coauthors, 2025: Probabilistic weather forecasting with machine learning. Nature, 637 (8044), 84--90, doi:10.1038/s41586-024-08252-9

    work page doi:10.1038/s41586-024-08252-9 2025

  28. [28]

    A., D.-M

    Randall, D. A., D.-M. Pan, P. Ding, and D. G. Cripe, 1997: Quasi-Equilibrium, 359--385. Springer Netherlands, Dordrecht, doi:10.1007/978-94-015-8828-7_14, ://doi.org/10.1007/978-94-015-8828-7_14

    work page doi:10.1007/978-94-015-8828-7_14 1997

  29. [29]

    Sims, J., D. Koch, H. Tolman, and D. Achuthavarier, 2021: Engaging the forecast community in UFS model development. ://www.ufs.epic.noaa.gov/2021/07/forecasters-workshop/, accessed: 2026-03-28, Unified Forecast System / NOAA Earth Prediction Innovation Center

    2021

  30. [30]

    Meng, and S

    Song, J., C. Meng, and S. Ermon, 2021: Denoising diffusion implicit models. International Conference on Learning Representations, ://openreview.net/forum?id=St1giarCHLP

    2021

  31. [31]

    Heinzeller, L

    Sun, X., D. Heinzeller, L. Bernardet, L. Pan, W. Li, D. Turner, and J. Brown, 2024: A case study investigating the low summertime cape behavior in the global forecast system. Weather and Forecasting, 39 (1), 3 -- 17, doi:10.1175/WAF-D-22-0208.1

    work page doi:10.1175/waf-d-22-0208.1 2024

  32. [32]

    Tang, I. N., A. C. Tridico, and K. H. Fung, 1997: Thermodynamic and optical properties of sea salt aerosols. Journal of Geophysical Research: Atmospheres, 102 (D19), 23\,269--23\,275, doi:https://doi.org/10.1029/97JD01806, https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/97JD01806

    work page doi:10.1029/97jd01806 1997

  33. [33]

    Atmospheric Environment (1967), 8 (12), 1251--1256, doi:https://doi.org/10.1016/0004-6981(74)90004-3

    Twomey, S., 1974: Pollution and the planetary albedo. Atmospheric Environment (1967), 8 (12), 1251--1256, doi:https://doi.org/10.1016/0004-6981(74)90004-3

    work page doi:10.1016/0004-6981(74)90004-3 1974

  34. [34]

    Khalizov, M

    Wang, Y., A. Khalizov, M. Levy, and R. Zhang, 2013: New directions: Light absorbing aerosols and their atmospheric impacts. Atmospheric Environment, 81, 713--715, doi:https://doi.org/10.1016/j.atmosenv.2013.09.034

    work page doi:10.1016/j.atmosenv.2013.09.034 2013

  35. [35]

    Ye, B., A. D. D. Genio, and K. K.-W. Lo, 1998: Cape variations in the current climate and in a climate change. Journal of Climate, 11 (8), 1997 -- 2015, doi:10.1175/1520-0442(1998)011<1997:CVITHCC>2.0.CO;2

    work page doi:10.1175/1520-0442(1998)011 1998