Pith. sign in

REVIEW 5 major objections 6 minor 3 cited by

Atmospheric Predictability Beyond 30 Days with Machine Learning

T0 review · 5 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read This paper claims that optimizing the starting state of a machine-learning weather model can keep deterministic forecasts skillful to roughly 27.5 days, more than double the accepted two-week predictability limit.

desk verdict A careful oracle experiment in ML weather prediction: within GraphCast the skill-extension claim holds, but the leap to a new atmospheric predictability horizon is not secured by the evidence. read the letter →

arxiv 2504.20238 v2 pith:MVM4X5OX submitted 2025-04-28 physics.ao-ph cs.LG

P. Trent Vonich , Gregory J. Hakim This is my paper · ORCID
classification physics.ao-phcs.LG
keywords atmosphericpredictabilitymachinelearningweatherpredictionGraphCastinitialconditionoptimizationgradientdescentHadleycirculationforecastskillhorizonPangu-Weather
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 seeks to overturn the long-standing rule that useful deterministic weather forecasts cannot extend beyond about two weeks. By adjusting the starting atmospheric state with gradient descent through the GraphCast model, the authors reduce 10-day forecast error by 86% on average across 732 twice-daily forecasts from 2020, with anomaly-correlation skill remaining above the practical 0.6 threshold to 27.5 days and statistical significance to 33 days. Because the same optimized states also improve a different model, Pangu-Weather, the paper argues the gains reflect genuine corrections to analysis and initial-condition error, not only GraphCast-specific tuning. If true, the result implies that the practical skill horizon is set more by imperfect initial conditions than by an immovable chaos barrier, and that better analyses could roughly double current lead times.

What carries the argument

The load-bearing mechanism is a fully differentiable forecast model: GraphCast is differentiable with respect to its input state, so one can backpropagate the forecast loss all the way to the initial condition and update that state by gradient descent. This replaces the linear tangent-linear/adjoint machinery of classical four-dimensional variational assimilation with a nonlinear, model-free optimization. The specific procedure that makes it work is quasi-static progressive window expansion: optimization starts on a 2-day forecast window and grows in 3-day increments up to 14 or 32 days, so the optimizer descends a gradually more complex loss landscape rather than attempting the full trajectory at once. The loss being minimized is GraphCast's own weighted mean squared error against the ERA5 verification sequence.

What would settle it

Re-verify the 732 optimized forecasts against independent observations, such as radiosonde reports or satellite-derived winds and temperatures, rather than ERA5. If the optimized forecasts show no systematic error reduction on that independent metric relative to the controls, the claimed predictability extension is an artifact of optimizing toward the verification reanalysis, and the central claim fails.

Watch

Extended reading notes

Core claim

The central discovery, stated sympathetically, is that initial conditions exist for which a deterministic machine-learning forecast stays skillful far beyond the conventional two-week horizon. Optimizing each 2020 initialization against the ERA5 reanalysis over a 14-day window produced no failures: every one of 732 forecasts improved, by at least 77% and up to 91% at ten days. The sample-mean optimal perturbation is spatially coherent and mostly tropical, amounting to an intensification of the Hadley circulation, with magnitudes comparable to typical analysis error. Double-precision optimization over a 32-day window pushes useful skill to about 27.5 days, and transferring the optimized states to Pangu-Weather still yields a 21% mean error reduction peaking near day 4, which the authors read as evidence that the corrections address a blend of analysis error and model bias.

Load-bearing premise

The load-bearing premise is that ERA5 is close enough to the true atmosphere that reducing forecast error measured against ERA5 means reducing error against reality; if the optimized states are mainly correcting GraphCast's learned bias toward its own training data, the experiment demonstrates model fitting rather than a new atmospheric predictability horizon.

Editorial extensions

If this is right

  • If the claim holds, deterministic forecast skill of roughly four weeks is achievable in principle, about twice the accepted two-week intrinsic limit.
  • All 732 optimization cases improved, so the effect is not tied to a special weather regime; it appears to be a general property of the model's phase space.
  • Because the optimized states improve Pangu-Weather as well, part of the correction is portable across models, suggesting it encodes atmospheric-state information rather than pure GraphCast bias.
  • The error growth rate returns to control-like behaviour once the optimization window ends, meaning gains are bought by initial-condition information, not by any long-horizon property of the model.
  • Real-time use would require finding such initial conditions without knowing the future verification, so the result sets a target for data assimilation rather than an immediate operational recipe.

Reading between the lines

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

  • The ERA5-trained/target circularity is the key caveat: if the optimization is mostly correcting the gap between ERA5 and the true atmosphere, the 27.5-day result is a statement about reanalysis error, not intrinsic predictability; the Pangu-Weather transfer reduces but does not eliminate this worry.
  • A decisive test would optimize against independent observations rather than ERA5 and verify against those same observations; if the skill gain vanishes, the result is an artifact of training-target matching.
  • The small gain (1-2%) from adding the fixed sample-mean perturbation to every control implies the useful corrections are strongly state-dependent; any operational method must solve for them per forecast, which is computationally expensive (about 4 GPU-hours per case here).
  • The same optimization procedure could be applied to coupled atmosphere-ocean or higher-resolution ML models; if the skill horizon extends further there, that would support the paper's suggestion that current limits are analysis-limited, not chaos-limited.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper optimizes GraphCast initial conditions against ERA5 verification over 14- and 32-day windows for 732 initialization times in 2020, reporting a mean 86% reduction in 10-day loss, ACC≥0.6 to 27.5 days, statistical significance to 33 days, and a 21% error reduction when the optimized initial conditions are transferred to Pangu-Weather. The authors interpret the results as evidence that, given accurate initial conditions, deterministic forecast skill can extend far beyond two weeks, challenging the conventional atmospheric predictability limit.

Significance. The study is a valuable demonstration of adjoint-style initial-condition optimization for a differentiable ML weather model, and the large sample and cross-model validation are strengths. If interpreted as an oracle result about what an ML model can achieve when the initial state is optimized against its own verification target, the quantitative claims are internally consistent and reproducible in principle. However, because GraphCast and the verification both come from ERA5 and GraphCast was trained on ERA5, the experiment does not by itself support the paper's atmospheric-predictability interpretation. The paper's own Discussion acknowledges the ambiguity, which is a sign of scientific care, but the title and abstract overstate what the evidence establishes.

major comments (5)
  1. [Section 2 and Eq. (2)] The central 86% ten-day error reduction is computed with the GraphCast training loss against ERA5, the same dataset on which GraphCast was trained. Optimizing the initial condition to minimize this loss is therefore an inversion of the model's learned ERA5-to-ERA5 mapping, not a measurement of predictability relative to the true atmosphere. To support the atmospheric claim, please verify the optimized forecasts against an independent target (e.g., JRA-55 or JRA-3Q) and/or raw observations, or explicitly reframe the claim as a model-relative oracle result.
  2. [Section 3 and Section 5] The paper reports that the mean optimal perturbation intensifies the Hadley circulation, consistent with a known ERA5 divergent-wind bias, and then concedes that 'separating model bias from reanalysis error remains ambiguous.' These statements undermine the inference that the optimized states are closer to the true atmosphere. The Supplementary Figure S4 result (adding the sample-mean perturbation to controls yields only 1-2% improvement) supports the interpretation that the mean perturbation is mostly a bias correction. Please provide a quantitative decomposition of the error reduction into analysis-bias correction, GraphCast model-bias correction, and genuine initial-condition error improvement, or add a limitation statement that the atmospheric-predictability conclusion is not supported.
  3. [Section 4 / Fig. 4] The Pangu-Weather transfer is the main evidence against GraphCast-specific overfitting, but Pangu-Weather was also trained on ERA5, so the transfer does not break the circularity with respect to the verification target. The 21% improvement, peaking at day 4 and with some forecasts worse than the control, is modest. The sentence 'crudely suggesting a 2:1 ratio of model error to initial condition error' is not justified, because the ratio conflates model bias with analysis bias.
  4. [Section 2] The uniform improvement across all 732 cases (minimum 77%) is more consistent with a systematic bias correction than with a state-dependent recoverable predictability signal. Please discuss this diagnostic: if the optimization were finding physically meaningful initial-condition corrections, one would expect more case-to-case variability in the amount of improvement.
  5. [Section 5, Fig. S2] The paper acknowledges that GraphCast forecasts become blurred under multi-day loss minimization, citing Brenowitz et al., Charlton-Perez et al., and Bonavita. The measured 5.8-day error-doubling time and the ACC persistence to 27.5 days may therefore be artifacts of this blurring rather than evidence about atmospheric error growth. Please discuss how blurring affects the interpretation of the error-doubling time and the significance of the ACC at long lead times.
minor comments (6)
  1. [Eq. (2) and Section 6.3] Equation (2) uses typesetting like 'Ttime' and 'G1.0◦'; please clean up the notation and define all symbols consistently in one place.
  2. [Section 6.4] The text says ACC is 'statistically different from the control' but the test computes critical values for a one-tailed t-test that ACC differs from zero; please align the wording with the actual test.
  3. [Abstract and Section 2] The abstract states 'skill lasting beyond 30 days' while the practical-skill threshold (ACC≥0.6) is quoted at 27.5 days and statistical significance at 33 days; please clarify which metric is being reported.
  4. [References and text] Several author names appear with stray spaces or broken formatting (e.g., 'V onich and Hakim' on page 2), and some citations lack complete year or journal information; please fix these formatting errors.
  5. [Section 7] The data-availability statement says optimal initial conditions 'will be made available at the time of publication' but does not link the optimization code; please include the code or an availability statement for it to make the machine-generated results reproducible.
  6. [Figure S4] Figure S4 shows a 1-2% improvement without confidence intervals; please add sampling uncertainty or state explicitly that the effect is within the noise.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline skill numbers are in-window minimization of the ERA5 loss; the atmospheric-predictability conclusion is partly definitional.

  1. fitted input called prediction [Section 2 (Forecast Performance), with Eq. (2) and Sec. 6.2]
    "Each initial condition is optimized per the GraphCast loss function (Sec. 6.3) to reduce cumulative global forecast error over a 14-day window, yielding a set of 732 optimized forecasts ... When the loss is measured at ten days, the single-precision mean (green) shown in Fig. 1 displays an 86% reduction in error compared to the control mean (black)."

    The 10-day error is exactly Eq. (2), the weighted MSE against ERA5 that the optimizer minimizes over a 14-day window. Sec. 6.2 specifies the loop: 'Calculate the forecast loss function by verifying against ERA5 at every step during inference.' Thus the quoted 86% improvement is the value of the objective function being minimized, not an independent prediction. The optimizer is guaranteed to reduce this same loss; reporting it as forecast skill is reporting a fit, so the '86% error reduction' reduces by construction.

  2. self definitional [Discussion and Conclusion (Section 5)]
    "In contrast, our study defines the limit for a single deterministic forecast as the time beyond which adjustments to the initial condition no longer reduce the error when verified against ERA5."

    The paper redefines the 'predictability limit' as the horizon over which IC adjustments can reduce model error against ERA5, then uses that fitted horizon (ACC to 27.5-33 days from 32-day-optimized forecasts) as evidence of atmospheric predictability beyond two weeks. Because ERA5 is both GraphCast's training target and the optimization target, this limit is a property of the fitting setup, not an independently measured atmospheric property. The 'skill beyond 30 days' is largely inside the 32-day optimization window, so the central conclusion is tied to the objective by definition rather than to the real atmosphere.

full rationale

The paper's headline improvement (86% at 10 days) is in-sample: the optimizer minimizes the GraphCast training loss (Eq. 2) against ERA5 over a 14-day window, and the 10-day loss is inside that window. Likewise, the 32-day optimized forecasts are optimized against ERA5 through day 32, so practical skill to 27.5 days and significance to 33 days are mostly within the objective window. This is a fitted input called a prediction, and the paper itself concedes the key ambiguity: 'Since GraphCast was trained on ERA5, separating model bias from reanalysis error remains ambiguous.' There is some independent content: the Pangu-Weather cross-model transfer is not directly optimized, and the 14-day-optimized forecasts retain some out-of-window improvement beyond day 14. However, Pangu-Weather shares the same ERA5 training target and the transfer improvement is modest and peaks near day 4. The out-of-window and cross-model results prevent a fully circular score, but the central atmospheric-predictability claim is substantially supported by in-window optimization against the verification dataset, so partial circularity is present. The score of 6 reflects that several of the main 'predictions' reduce by construction, while the cross-model experiment provides partial, though not fully independent, support.

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

The paper introduces no new physical entities. Its parameters are optimization hyperparameters. The crucial conceptual imports are the assumptions that model error growth mirrors atmospheric error growth and that ERA5 is a valid verification target; both are plausible but unproven, and the paper's own Discussion acknowledges the ERA5/model-bias ambiguity.

free parameters (4)
  • learning rate = 10^-3, reduced for windows > 14 days
    Chosen by hand from previous work; affects optimization convergence but not the conceptual claim.
  • optimization window schedule = 2 days initial, then 3-day increments
    The paper says the sizes are 'arbitrary' and 'based on empirical testing' (Section 6.2); the 14-day and 32-day limits are set by precision and GPU memory, not physics.
  • epochs per window = not stated
    The number of gradient epochs per window step is not specified, so the exact optimization trajectory is underdetermined.
  • negative humidity clipping = clipped to zero
    Ad hoc handling of non-negative variables; the paper states it has negligible impact.
assumptions (4)
  • domain assumption GraphCast is an adequate surrogate for the atmosphere's error growth in predictability studies
    The paper interprets model error growth as atmospheric error growth, despite noting the model is deterministic, blurs under multi-day loss minimization, and becomes unstable beyond 45 days (Discussion; Karlbauer et al. 2024).
  • domain assumption ERA5 is a faithful proxy for the true atmospheric state
    Both optimization and verification use ERA5; Section 3 notes ERA5 has a weaker divergent wind component (citing Li et al. 2024), so the optimized states may correct reanalysis bias rather than atmospheric error.
  • domain assumption The GraphCast training loss is a valid measure of atmospheric forecast skill
    The weighted mean squared error (Eq. 2), designed for model training, is used as the optimization objective and the headline error metric.
  • domain assumption Gradient descent finds a meaningful optimum rather than an adversarial direction
    The paper argues against the adversarial interpretation in Section 5, but this premise is required for the atmospheric interpretation; the optimizer's robustness to starting point is tested on one case only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Atmospheric Predictability Beyond 30 Days with Machine Learning." pith.science (2026). https://pith.science/paper/MVM4X5OX

@misc{pith2026250420238,
  author       = {Pith},
  title        = {Pith review of: Atmospheric Predictability Beyond 30 Days with Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MVM4X5OX}},
  note         = {Machine review of arXiv:2504.20238}
}
read the original abstract

Atmospheric predictability research has long held that rapid error growth at small spatial scales imposes an intrinsic limit of roughly two weeks on deterministic weather forecast skill. We challenge this limit using GraphCast, a machine-learning weather model, by optimizing initial conditions for twice-daily forecasts spanning 2020. This approach yields an average error reduction of 86% at ten days relative to control forecasts from reanalysis initial conditions, with skill lasting beyond 30 days. Mean optimal initial-condition perturbations reveal large-scale, spatially coherent corrections primarily reflecting an intensification of the Hadley circulation. Forecasts using GraphCast-optimal initial conditions in the Pangu-Weather model achieve a 21% error reduction, peaking at four days, indicating that analysis corrections reflect adjustments that target both model and analysis error. These results demonstrate the existence of initial conditions producing skillful deterministic forecasts far beyond two weeks. Whether such initial conditions can be identified in real-time for improving operational weather forecasts remains a topic of future research.

Figures

Figures reproduced from arXiv: 2504.20238 by the authors.

Figure 1
Figure 1. Weighted mean squared error as detailed in Eq. 2 for all 732 control forecasts (black), 14-day optimized [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Sample-mean optimal perturbations averaged over 732 cases for (a) 850 hPa temperature, (b) 700 hPa specific [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. ERA5 2020 climatological anomaly and optimal perturbation autocorrelation as a function of lag for [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Mean, median, and interquartile range (25th–75th percentiles) of the relative change in mean squared error [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. DLESyM-Ocean: A Deep Learning Probabilistic Global Model for Simulating Present-Day Upper Ocean and Sea Ice

    physics.ao-ph 2026-08 conditional novelty 6.0 of 10

    DLESyM-Ocean is a probabilistic deep learning ocean-sea ice emulator trained with a patch energy score loss that stays stable over multi-year simulations.

  2. Weather Emulators at the Frontier of Heat Extremes Predictability

    physics.ao-ph 2026-07 accept novelty 6.0 of 10

    At 10–15 day leads, AI weather emulators can match or beat dynamical models on global temperature skill but under-represent heat-extreme intensity and lose to IFS on recall.

  3. Rigorous uncertainty quantification of probabilistic AI weather forecasts with conformal prediction

    physics.ao-ph 2026-06 unverdicted novelty 6.0 of 10

    Online conformal prediction post-processing guarantees calibrated uncertainty coverage for GenCast, NeuralGCM, and AIFS-ENS forecasts of temperature and precipitation including extremes.

Reference graph

Works this paper leans on

50 extracted references · 26 canonical work pages · cited by 3 Pith papers

  1. [1]

    Edward N. Lorenz. The predictability of a flow which possesses many scales of motion. Tellus, 21 0 (3): 0 289--307, 1969. doi:10.1111/j.2153-3490.1969.tb00444.x. URL http://eaps4.mit.edu/research/Lorenz/publications.htm

    arXiv 1969

  2. [2]

    J. G. Charney, R. G. Fleagle, V. E. Lally, H. Riehl, and D. Q. Wark. The feasibility of a global observation and analysis experiment. Bulletin of the American Meteorological Society, 47: 0 200--220, 1966

    work page 1966

  3. [3]

    The dual nature of chaos and order in the atmosphere

    Bo-Wen Shen, Roger Pielke, Xubin Zeng, Jialin Cui, Sara Faghih-Naini, Wei Paxson, Amit Kesarkar, Xiping Zeng, and Robert Atlas. The dual nature of chaos and order in the atmosphere. Atmosphere, 13 0 (11): 0 1892, 2022. ISSN 2073-4433. doi:10.3390/atmos13111892. URL https://www.mdpi.com/2073-4433/13/11/1892

    work page doi:10.3390/atmos13111892 2022

  4. [4]

    Pielke, Xubin Zeng, and Xiping Zeng

    Bo-Wen Shen, Roger A. Pielke, Xubin Zeng, and Xiping Zeng. Lorenz's view on the predictability limit of the atmosphere. Encyclopedia, 3 0 (3): 0 887--899, 2023. ISSN 2673-8392. doi:10.3390/encyclopedia3030063. URL https://www.mdpi.com/2673-8392/3/3/63

    work page doi:10.3390/encyclopedia3030063 2023

  5. [5]

    Pielke, Sr., Xubin Zeng, and Xiping Zeng

    Bo-Wen Shen, Roger A. Pielke, Sr., Xubin Zeng, and Xiping Zeng. Exploring the origin of the two-week predictability limit: A revisit of lorenz’s predictability studies in the 1960s. Atmosphere, 15 0 (7): 0 837, 2024. doi:10.3390/atmos15070837. URL https://doi.org/10.3390/atmos15070837

    work page doi:10.3390/atmos15070837 2024

  6. [6]

    Predictability: A problem partly solved

    Edward N Lorenz. Predictability: A problem partly solved. In Proc. Seminar on predictability, volume 1. Reading, 1996

    1996

  7. [7]

    What is the predictability limit of midlatitude weather? Journal of the Atmospheric Sciences, 76 0 (4): 0 1077--1091, 2019

    Fuqing Zhang, Y Qiang Sun, Linus Magnusson, Roberto Buizza, Shian-Jiann Lin, Jan-Huey Chen, and Kerry Emanuel. What is the predictability limit of midlatitude weather? Journal of the Atmospheric Sciences, 76 0 (4): 0 1077--1091, 2019

    work page 2019

  8. [8]

    Insights into atmospheric predictability through global convection-permitting model simulations

    Falko Judt. Insights into atmospheric predictability through global convection-permitting model simulations. Journal of the Atmospheric Sciences, 75 0 (5): 0 1477 -- 1497, 2018. doi:10.1175/JAS-D-17-0343.1. URL https://journals.ametsoc.org/view/journals/atsc/75/5/jas-d-17-0343.1.xml

    doi:10.1175/jas-d-17-0343.1 2018

Show all 50 references
  1. [9]

    Atmospheric predictability of the tropics, middle latitudes, and polar regions explored through global storm-resolving simulations

    Falko Judt. Atmospheric predictability of the tropics, middle latitudes, and polar regions explored through global storm-resolving simulations. Journal of the Atmospheric Sciences, 77 0 (1): 0 257 -- 276, 2020. doi:10.1175/JAS-D-19-0116.1. URL https://journals.ametsoc.org/view...

    2020 doi

  2. [10]

    Estimating the intrinsic limit of predictability using a stochastic convection scheme

    Tobias Selz. Estimating the intrinsic limit of predictability using a stochastic convection scheme. Journal of the Atmospheric Sciences, 76 0 (3): 0 757 -- 765, 2019. doi:10.1175/JAS-D-17-0373.1. URL https://journals.ametsoc.org/view/journals/atsc/76/3/jas-d-17-0373.1.xml

    2019 doi

  3. [11]

    The forecast skill horizon

    Roberto Buizza and Martin Leutbecher. The forecast skill horizon. Quarterly Journal of the Royal Meteorological Society, 141 0 (693): 0 3366--3382, 2015

    2015

  4. [12]

    Predictability Limit of the 2021 Pacific Northwest Heatwave from Deep-Learning Sensitivity Analysis , September 2024

    Peter Vonich and Gregory Hakim. Predictability Limit of the 2021 Pacific Northwest Heatwave from Deep-Learning Sensitivity Analysis , September 2024. URL https://doi.org/10.5281/zenodo.13694959

    2021 doi

  5. [13]

    Evaluation of physical processes in an idealized extratropical cyclone using adjoint sensitivity

    Rolf H Langland, Russell L Elsberry, and Ronald M Errico. Evaluation of physical processes in an idealized extratropical cyclone using adjoint sensitivity. Quarterly Journal of the Royal Meteorological Society, 121 0 (526): 0 1349--1386, 1995

    1995

  6. [14]

    Langland, Melvyn A

    Rolf H. Langland, Melvyn A. Shapiro, and Ronald Gelaro. Initial condition sensitivity and error growth in forecasts of the 25 january 2000 east coast snowstorm. Monthly Weather Review, 130 0 (4): 0 957 -- 974, 2002. doi:10.1175/1520-0493(2002)130<0957:ICSAEG>2.0.CO;2. URL http...

    2002 doi

  7. [15]

    Doyle, Carolyn A

    James D. Doyle, Carolyn A. Reynolds, Clark Amerault, and Jonathan Moskaitis. Adjoint sensitivity and predictability of tropical cyclogenesis. Journal of the Atmospheric Sciences, 69 0 (12): 0 3535 -- 3557, 2012. doi:10.1175/JAS-D-12-0110.1. URL https://journals.ametsoc.org/vie...

    2012 doi

  8. [16]

    Doyle, Clark Amerault, Carolyn A

    James D. Doyle, Clark Amerault, Carolyn A. Reynolds, and P. Alex Reinecke. Initial condition sensitivity and predictability of a severe extratropical cyclone using a moist adjoint. Monthly Weather Review, 142 0 (1): 0 320 -- 342, 2014. doi:10.1175/MWR-D-13-00201.1. URL https:/...

    2014 doi

  9. [17]

    Adjoint sensitivity analysis of high-impact extratropical cyclones

    James D Doyle, Carolyn A Reynolds, and Clark Amerault. Adjoint sensitivity analysis of high-impact extratropical cyclones. Monthly Weather Review, 147 0 (12): 0 4511--4532, 2019

    2019

  10. [18]

    Lloveras, James D

    Daniel J. Lloveras, James D. Doyle, and Dale R. Durran. Can observation targeting be a wild goose chase? an adjoint-sensitivity study of a u.s. east coast cyclone forecast bust. Journal of the Atmospheric Sciences, 82 0 (2): 0 343 -- 360, 2025. doi:10.1175/JAS-D-24-0044.1. URL...

    2025 doi

  11. [19]

    Doyle, Carolyn A

    Jorge Baño-Medina, Agniv Sengupta, James D. Doyle, Carolyn A. Reynolds, Duncan Watson-Parris, and Luca Delle Monache. Are ai weather models learning atmospheric physics? a sensitivity analysis of cyclone xynthia. npj Climate and Atmospheric Science, 8 0 (1): 0 92, 2025. doi:10...

    2025 doi

  12. [20]

    Learning skillful medium-range global weather forecasting

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382 0 (6677): 0 1416--1421, 2023

    2023

  13. [21]

    The 2021 western north america heat wave among the most extreme events ever recorded globally

    Vikki Thompson, Alan T Kennedy-Asser, Emily Vosper, YT Eunice Lo, Chris Huntingford, Oliver Andrews, Matthew Collins, Gabrielle C Hegerl, and Dann Mitchell. The 2021 western north america heat wave among the most extreme events ever recorded globally. Science Advances, 8 0 (18...

    2021

  14. [22]

    Heatwave attribution based on reliable operational weather forecasts

    Nicholas J Leach, Christopher D Roberts, Matthias Aengenheyster, Daniel Heathcote, Dann M Mitchell, Vikki Thompson, Tim Palmer, Antje Weisheimer, and Myles R Allen. Heatwave attribution based on reliable operational weather forecasts. Nature Communications, 15 0 (1): 0 4530, 2024

    2024

  15. [23]

    The era5 global reanalysis

    Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, Andr \'a s Hor \'a nyi, Joaqu \' n Mu \ n oz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146 0 (730): ...

    1999

  16. [24]

    Accurate medium-range global weather forecasting with 3D neural networks

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3D neural networks. Nature, pages 1--6, 2023

    2023

  17. [25]

    Durran, Raul A

    Matthias Karlbauer, Nathaniel Cresswell-Clay, Dale R. Durran, Raul A. Moreno, Thorsten Kurth, Boris Bonev, Noah Brenowitz, and Martin V. Butz. Advancing parsimonious deep learning weather prediction using the healpix mesh. Journal of Advances in Modeling Earth Systems, 16 0 (8...

    2024 doi

  18. [26]

    Weatherbench 2: A benchmark for the next generation of data-driven global weather models

    Stephan Rasp, Stephan Hoyer, Alexander Merose, Ian Langmore, Peter Battaglia, Tyler Russel, Alvaro Sanchez-Gonzalez, Vivian Yang, Rob Carver, Shreya Agrawal, et al. Weatherbench 2: A benchmark for the next generation of data-driven global weather models. arXiv preprint arXiv:2...

    2023 arXiv

  19. [27]

    A historical perspective of the la niña event in 2020/2021

    Xiaofan Li, Zeng-Zhen Hu, Yu-heng Tseng, Yunyun Liu, and Ping Liang. A historical perspective of the la niña event in 2020/2021. Journal of Geophysical Research: Atmospheres, 127 0 (7), 2022. doi:10.1029/2021jd035546

    2020 doi

  20. [28]

    Estimates of global analysis error from the global weather experiment observational network

    Roger Daley and Thomas Mayer. Estimates of global analysis error from the global weather experiment observational network. Monthly Weather Review, 114 0 (9): 0 1642 -- 1653, 1986. doi:10.1175/1520-0493(1986)114<1642:EOGAEF>2.0.CO;2. URL https://journals.ametsoc.org/view/journa...

    1986 doi

  21. [29]

    Gregory J. Hakim. Vertical structure of midlatitude analysis and forecast errors. Monthly Weather Review, 133 0 (3): 0 567 -- 578, 2005. doi:10.1175/MWR-2882.1. URL https://journals.ametsoc.org/view/journals/mwre/133/3/mwr-2882.1.xml

    2005 doi

  22. [30]

    Estimation of analysis and forecast error variances

    Marcelo Peña and Zoltan Toth. Estimation of analysis and forecast error variances. Tellus A: Dynamic Meteorology and Oceanography, 66 0 (1): 0 21767, 2014. doi:10.3402/tellusa.v66.21767. URL https://doi.org/10.3402/tellusa.v66.21767

    2014 doi

  23. [31]

    Exploring the differences in kinetic energy spectra between the ncep fnl and era5 datasets

    Zongheng Li, Jun Peng, Lifeng Zhang, and Jiping Guan. Exploring the differences in kinetic energy spectra between the ncep fnl and era5 datasets. Journal of the Atmospheric Sciences, 81 0 (2): 0 363--380, February 2024. doi:10.1175/JAS-D-23-0043.1. URL https://doi.org/10.1175/...

    2024 doi

  24. [32]

    What if? numerical weather prediction at the crossroads

    Peter Bauer. What if? numerical weather prediction at the crossroads. Journal of the European Meteorological Society, 1: 0 100002, 2024. ISSN 2950-6301. doi:https://doi.org/10.1016/j.jemets.2024.100002. URL https://www.sciencedirect.com/science/article/pii/S2950630124000024

    2024

  25. [33]

    Intriguing properties of neural networks, 2014

    Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks, 2014. URL https://arxiv.org/abs/1312.6199

    2014 arXiv

  26. [34]

    Deepfool: a simple and accurate method to fool deep neural networks

    Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. CVPR, 11 2016

    2016

  27. [35]

    Exploring the space of adversarial images

    Pedro Tabacof and Eduardo Valle. Exploring the space of adversarial images. CoRR, abs/1510.05328, 2015. URL http://arxiv.org/abs/1510.05328

    2015 arXiv

  28. [36]

    Goodfellow, Jonathon Shlens, and Christian Szegedy

    Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015. URL https://arxiv.org/abs/1412.6572

    2015 arXiv

  29. [37]

    Adversarial examples are not bugs, they are features, 2019

    Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features, 2019. URL https://arxiv.org/abs/1905.02175

    2019 arXiv

  30. [38]

    Jacob Mitchell Springer, Melanie Mitchell, and Garrett T. Kenyon. Adversarial perturbations are not so weird: Entanglement of robust and non-robust features in neural network classifiers. ArXiv, abs/2102.05110, 2021. URL https://api.semanticscholar.org/CorpusID:231861678

    2021 arXiv

  31. [39]

    Brenowitz, Yair Cohen, Jaideep Pathak, Ankur Mahesh, Boris Bonev, Thorsten Kurth, Dale R

    Noah D. Brenowitz, Yair Cohen, Jaideep Pathak, Ankur Mahesh, Boris Bonev, Thorsten Kurth, Dale R. Durran, Peter Harrington, and Michael S. Pritchard. A practical probabilistic benchmark for ai weather models. arXiv preprint arXiv:2401.15305, 2401.15305v1, jan 2024. URL https:/...

    2024 arXiv

  32. [40]

    Charlton-Perez, Helen F

    Andrew J. Charlton-Perez, Helen F. Dacre, Simon Driscoll, Suzanne L. Gray, Ben Harvey, Natalie J. Harvey, Kieran M. R. Hunt, Robert W. Lee, Ranjini Swaminathan, Remy Vandaele, and Ambrogio Volonté. Do ai models produce better weather forecasts than physics-based models? a quan...

    2024 doi

  33. [41]

    On some limitations of current machine learning weather prediction models

    Massimo Bonavita. On some limitations of current machine learning weather prediction models. Geophysical Research Letters, 51 0 (12): 0 e2023GL107377, 2024. doi:https://doi.org/10.1029/2023GL107377. URL https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2023GL107377. e202...

    2024 doi

  34. [42]

    Espinosa, Raul Moreno, and Matthias Karlbauer

    Nathaniel Cresswell-Clay, Bowen Liu, Dale Durran, Zihui Liu, Zachary I. Espinosa, Raul Moreno, and Matthias Karlbauer. A deep learning earth system model for efficient simulation of the observed climate, 2025. URL https://arxiv.org/abs/2409.16247

    2025

  35. [43]

    Vaillancourt, Jing Yang, and Ayrton Zadra

    Syed Zahid Husain, Leo Separovic, Jean-François Caron, Rabah Aider, Mark Buehner, Stéphane Chamberland, Ervig Lapalme, Ron McTaggart-Cowan, Christopher Subich, Paul A. Vaillancourt, Jing Yang, and Ayrton Zadra. Leveraging data-driven weather models for improving numerical weat...

    2024 arXiv

  36. [44]

    Hersbach, B

    H. Hersbach, B. Bell, P. Berrisford, S. Hirahara, A. Horányi, J. Muñoz‐Sabater, J. Nicolas, C. Peubey, R. Radu, D. Schepers, A. Simmons, C. Soci, S. Abdalla, X. Abellan, G. Balsamo, P. Bechtold, G. Biavati, J. Bidlot, M. Bonavita, G. De Chiara, P. Dahlgren, D. Dee, M. Diamanta...

    1940

  37. [45]

    What is an adjoint model? Bulletin of the American Meteorological Society, 78 0 (11): 0 2577--2592, 1997

    Ronald M Errico. What is an adjoint model? Bulletin of the American Meteorological Society, 78 0 (11): 0 2577--2592, 1997

    1997

  38. [46]

    JAX : composable transformations of P ython+ N um P y programs, 2018

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake Vander P las, Skye Wanderman- M ilne, and Qiao Zhang. JAX : composable transformations of P ython+ N um P y programs, 2018. URL http://github.com/...

    2018

  39. [47]

    Carlos A. L. Pires, Robert Vautard, and O. Talagrand. On extending the limits of variational assimilation in nonlinear chaotic systems. Tellus A, 48: 0 96--121, 1996. URL https://api.semanticscholar.org/CorpusID:122300156

    1996

  40. [48]

    Four-dimensional variational assimilation and predictability in a quasi-geostrophic model

    Kyle Swanson, Robert Vautard, and Carlos Pires. Four-dimensional variational assimilation and predictability in a quasi-geostrophic model. Tellus A: Dynamic Meteorology and Oceanography, Jan 1998. doi:10.3402/tellusa.v50i4.14540

    1998 doi

  41. [49]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. 2017. URL https://arxiv.org/abs/1412.6980

    2017 arXiv

  42. [50]

    Daniel S. Wilks. Statistical Methods in the Atmospheric Sciences. Academic Press, 3 edition, 2011. ISBN 9780123850225

    2011

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.