Pith. sign in

REVIEW 5 major objections 5 minor 24 references

Enhancing Near Real Time AI-NWP Hurricane Forecasts: Improving Explainability and Performance Through Physics-Based Models and Land Surface Feedback

T0 review · 5 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read This paper argues that land surface conditions, especially soil moisture, materially alter hurricane tracks near landfall, and that atmosphere-only AI weather models such as Graphcast-operational are missing this process; evidence comes…

desk verdict Useful new case-study data, but the central land-feedback claim rests on one unensembled storm. read the letter →

arxiv 2502.01797 v1 pith:ZCS4OGMA submitted 2025-02-03 physics.ao-ph physics.geo-ph

classification physics.ao-phphysics.geo-ph
keywords hurricanetrackforecastingAI-NWPGraphcast-operationalHWRFxsoilmoisturesensitivityland-atmospherecouplinglandfallinghurricanes
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 argues that the next step for AI-based hurricane forecasting is to couple atmosphere, land, and ocean. It first shows that Graphcast-operational, an atmosphere-only AI model, beat the physics-based HWRFx on five-day track error by roughly 30 to 40 percent for four 2024 hurricanes. It then uses HWRFx sensitivity simulations of Hurricane Beryl to show that changing initial soil moisture from wilting point to field capacity shifts the storm track by hundreds of kilometers near the coast: wetter ground pulls the track west, drier ground east. The paper reads this as evidence that land-atmosphere coupling is a real control on landfalling hurricanes and that AI-NWP models, which currently ignore land, could gain accuracy and explainability by adding it. No fully coupled AI atmosphere-land-ocean model exists yet, and the paper proposes building one.

What carries the argument

The load-bearing machinery is the three-state soil-moisture perturbation experiment in HWRFx: a control run at default soil moisture, a field-capacity run with saturated soils, and a wilting-point run with dry soils, all initialized from the same atmosphere for Hurricane Beryl. The tracks agree over open ocean and diverge near the coast, which is the empirical core of the land-feedback argument. The supporting machinery is Graphcast-operational, a graph-neural-network weather model trained on ERA5 reanalysis, whose five-day track errors for four hurricanes are compared against HWRFx and the IBTrACS best-track database.

What would settle it

Run the same field-capacity, control, and wilting-point experiments for a dozen landfalling hurricanes using a 10-20 member ensemble of atmospheric initial conditions; if the track spread between moisture scenarios is comparable to the ensemble spread within a scenario, the claimed soil-moisture control on storm path is not detectable.

Watch

Extended reading notes

Core claim

The central claim is that land surface conditions, specifically soil moisture, materially alter hurricane tracks near landfall, so atmosphere-only AI-NWP models omit a real physical process. The evidence is a set of HWRFx simulations for Hurricane Beryl: a control run at default soil moisture, a field-capacity run with wetter soil, and a wilting-point run with drier soil. The tracks agree over open ocean and diverge roughly 350 kilometers from the Gulf of Mexico coastline, with the wet run shifted westward and the dry run shifted eastward. The paper also reports that Graphcast-operational reduced five-day track error by approximately 30 to 40 percent compared to HWRFx for hurricanes Beryl, Debby, Francine, and Helene, and interprets this as showing both the current skill of AI models and their missing land feedback.

Load-bearing premise

The track differences between the wet-soil and dry-soil runs of Hurricane Beryl are caused by the prescribed soil moisture change, not by the model's internal variability or spin-up, and one storm's behavior represents landfalling hurricanes generally.

Editorial extensions

If this is right

  • Current atmosphere-only AI-NWP forecasts of landfalling hurricanes may carry track errors that a land-coupled system could reduce.
  • A fully coupled AI atmosphere-land-ocean model would be the first AI-NWP framework able to represent the coast as a physically active surface.
  • Operational hurricane forecasting could use soil-moisture initialization as a lever for track uncertainty, not just intensity.
  • Graphcast-operational's lower track errors over days three to five suggest AI models can complement physics models in mid-latitude transition and rapid intensification phases.
  • Land-atmosphere coupling would make AI forecasts more explainable by linking track deviations to named surface processes like friction and heat exchange.

Reading between the lines

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

  • Beyond the paper: repeating the wet-versus-dry soil experiment as a multi-storm, multi-member ensemble would separate a causal soil-moisture effect from chaotic internal variability; the single Beryl case cannot do that on its own.
  • Beyond the paper: an incremental path to the proposed fully coupled AI system would be to add soil moisture and surface heat fluxes as input channels to an existing graph-based weather model, which is cheaper than full land coupling and directly testable.
  • Beyond the paper: if the reported mechanism is frictional and thermal surface fluxes, then training an AI model on surface flux fields might recover much of the land effect without any new coupling architecture.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper evaluates Google Graphcast-operational forecasts for four 2024 Atlantic hurricanes (Beryl, Debby, Francine, Helene) against IBTrACS and reports roughly 30–40% reductions in 5-day track error relative to HWRFx. To motivate adding land surface coupling to AI-NWP models, the authors run Hurricane WRF experimental (HWRFx) simulations for Hurricane Beryl with three soil-moisture initializations: control, field capacity, and wilting point. They interpret the divergence of these tracks near the Gulf Coast as evidence that land surface conditions significantly alter storm paths, and then propose a conceptual coupled AI atmosphere-land-ocean model as a future direction.

Significance. If the land-feedback result were robust, it would be an important and timely motivation for incorporating land surface processes into AI-based weather prediction, with direct relevance to landfalling hurricane forecasts. The paper also provides a useful, if preliminary, side-by-side comparison of an AI model against a physics-based hurricane model on real 2024 cases. However, the central empirical claim is supported by only one storm, one deterministic simulation per soil-moisture setting, no uncertainty quantification, and no process-level verification of the proposed mechanism. The Graphcast-vs-HWRFx skill comparison is also reported as point estimates without statistical or methodological detail. These limitations currently prevent the abstract's general conclusion from being supported.

major comments (5)
  1. [§3.2, Figure 3] The central claim that land surface conditions significantly alter storm paths rests on three deterministic HWRFx runs for a single storm (Beryl). No ensemble size is reported, no statement confirms that initial and boundary conditions are identical except for soil moisture, and no experiment with unperturbed stochastic variability is described. The three tracks agree over the ocean and diverge only within roughly 350 km of the coast, which is precisely the region where the same model has its largest track error and where the control run fails to reproduce Beryl's observed landfall. Without a measure of unperturbed run-to-run spread, the FC/WP separation could be chaotic internal variability, spin-up effects, or model error growth rather than a soil-moisture response. This single-storm, one-run-per-condition design cannot support the general conclusion stated in the abstract.
  2. [§3.2, 'vortex correction'] The text states that 'after applying vortex correction, the control run has shown similar track as best track but east of the track most of the time,' but the vortex correction is not described. It is not stated what correction was applied, whether it was applied identically to the field-capacity and wilting-point runs, or how it affects the comparison of the three soil-moisture scenarios. Because the control track is the baseline for the reported FC/WP divergence, this missing information is load-bearing for the interpretation.
  3. [§3.2, physical mechanism] The proposed mechanism—wetter soil increases friction and cooling, shifting the track westward, while drier soil reduces friction and shifts the track eastward—is post hoc. The paper does not present any supporting diagnostics such as soil-moisture evolution, surface sensible/latent heat fluxes, boundary-layer height, surface drag, or storm-relative asymmetries. Without these process-level fields, the track shifts cannot be causally attributed to the hypothesized land-atmosphere feedback rather than to other differences in the simulations.
  4. [§3.1] The reported track-error reductions (40%, 35%, 30%, and 37% for Beryl, Debby, Francine, and Helene) are point estimates with no confidence intervals, no measure of forecast-cycle spread, and no description of the evaluation protocol. The manuscript does not state how many initialization times were used per storm, how errors were aggregated across lead times, or whether the same verification times and forecast lengths were used for both models. These details are necessary to assess whether the claimed superiority of Graphcast is statistically meaningful.
  5. [§3.2, generalizability] HWRFx did not accurately reproduce Beryl's landfall in any of the three soil-moisture scenarios, and the text separately notes that near-real-time HWRFx runs initialized five days ahead predicted a westward track while Graphcast forecasted landfall near Houston. This means the sensitivity experiment is conducted in a model state that has substantial track error in the very region where the soil-moisture divergence appears. The paper should address whether the FC/WP track shifts are meaningful for forecasting if the model's unperturbed control is already strongly biased in that region.
minor comments (5)
  1. [Abstract] The abstract refers to 'Google's Graphcast operation' where 'operational' appears to be intended; this typo appears in the main text as well.
  2. [§3 heading] The heading 'Graphcast-operational Hurricane T rack F orecasting Results' contains a typographical error ('T rack' should be 'Track').
  3. [§2.1 vs Data Availability] The text says soil moisture data were from Chen et al. (2023), but the Data Availability statement says soil moisture and atmospheric parameters were sourced from ERA5; the relationship between these datasets should be clarified.
  4. [§3.1, Figures S1–S3] The manuscript refers to supplementary figures for Debby, Francine, and Helene (Figures S1–S3) and for intensity (Figure S4), but these supplements are not included in the posted manuscript. Without them, the multi-storm track-error comparison cannot be independently checked.
  5. [§3.2, Figure 3 caption] The caption's phrase 'diverge significantly as they approach land ( 350 km from the Gulf of Mexico coastline)' should specify the unit ('about 350 km') and should indicate whether the quoted distance is a distance from the coast or from the landfall location.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the land-feedback claim rests on an independent model sensitivity experiment, not on self-citation or fitted inputs.

full rationale

The paper's central claim—that soil moisture variations significantly alter hurricane tracks—is supported by controlled HWRFx sensitivity simulations in Section 3.2 and Figure 3. The Field Capacity and Wilting Point runs perturb initial soil moisture relative to the control while otherwise using the same model configuration; the resulting tracks are compared with each other and with IBTrACS observations. No output quantity is used to define the perturbation, and no parameter is fitted to the target track or to the reported error reductions. The Graphcast-operational evaluation is an external benchmark against IBTrACS, and the 30-hour and 5-day track errors are straightforward verification metrics. Self-citations to HWRF development literature (e.g., Gopalakrishnan et al. 2011, Alaka et al. 2022, Zhang et al. 2016) provide model provenance and are not used as the evidential basis for the land-atmosphere coupling conclusion. No equation reduces to its input, no fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported from prior work by the same authors. The paper's actual vulnerability is statistical: each soil-moisture scenario is a single deterministic run with no ensemble or uncertainty quantification, so the Figure 3 track divergence might reflect chaotic internal variability rather than the imposed soil-moisture difference. That is a legitimate evidence-quality objection, but it is a correctness and robustness concern, not circular reasoning. The derivation chain is self-contained with respect to its stated inputs and external observations.

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

The paper introduces no new parameters or entities. Its conclusions rest on external model outputs (Graphcast, HWRFx, GFS, IBTrACS) and the assumption that these are faithful representations of the real atmosphere and land surface.

assumptions (4)
  • domain assumption IBTrACS best tracks are accurate enough to serve as ground truth for forecast error evaluation.
    The study uses IBTrACS positions to compute track errors for all models (Section 2.1).
  • domain assumption GFS initial conditions provide valid starting states for both Graphcast and HWRFx.
    Both models are initialized from GFS data (Section 2.1).
  • domain assumption HWRFx faithfully represents land-atmosphere interactions such that soil moisture perturbations produce physically meaningful track responses.
    The central sensitivity result depends on HWRFx credibility (Section 2.2.2 and Section 3.2).
  • domain assumption Graphcast-operational is representative of current AI-NWP models for the purpose of generalizing the paper's recommendation.
    The paper extends findings from a single AI model to all AI-NWP models (Sections 1 and 4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Near Real Time AI-NWP Hurricane Forecasts: Improving Explainability and Performance Through Physics-Based Models and Land Surface Feedback." pith.science (2026). https://pith.science/paper/ZCS4OGMA

@misc{pith2026250201797,
  author       = {Pith},
  title        = {Pith review of: Enhancing Near Real Time AI-NWP Hurricane Forecasts: Improving Explainability and Performance Through Physics-Based Models and Land Surface Feedback},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZCS4OGMA}},
  note         = {Machine review of arXiv:2502.01797}
}
read the original abstract

Hurricane track forecasting remains a significant challenge due to the complex interactions between the atmosphere, land, and ocean. Although AI-based numerical weather prediction models, such as Google Graphcast operation, have significantly improved hurricane track forecasts, they currently function as atmosphere-only models, omitting critical land and ocean interactions. To investigate the impact of land feedback, we conducted independent simulations using the physics-based Hurricane WRF experimental model to assess how soil moisture variations influence storm trajectories. Our results show that land surface conditions significantly alter storm paths, demonstrating the importance of land-atmosphere coupling in hurricane prediction. Although recent advances have introduced AI-based atmosphere-ocean coupled models, a fully functional AI-driven atmosphere-land-ocean model does not yet exist. Our findings suggest that AI-NWP models could be further improved by incorporating land surface interactions, improving both forecast accuracy and explainability. Developing a fully coupled AI-based weather model would mark a critical step toward more reliable and physically consistent hurricane forecasting, with direct applications for disaster preparedness and risk mitigation.

Figures

Figures reproduced from arXiv: 2502.01797 by the authors.

Figure 1
Figure 1. Best tracks for hurricanes Beryl, Debby, Francine, and Helene from IBTrACS. 2.2 Models 2.2.1 Graphcast-operational model Graphcast-operational is a state-of-the-art AI-based Numerical Weather Predic￾tion (NWP) model that employs deep learning to produce medium-range weather fore￾casts (Lam et al., 2023). The model, trained on four decades of historical weather data from the ECMWF ERA5 dataset, can generate 10-day fo… view at source ↗
Figure 2
Figure 2. The forecasted tracks for Hurricane Beryl from HWRFx and Graphcast-operational with an initial condition of 07-03-2024 00hrs. Figure (a) compares tracks from HWRFx and Graphcast-operational with the best tracks from IBTrACS. The forecast is for 5 days, valid until 07-08-2024 00hrs. Graphcast-operational closely reproduces the best track, but the landfall tim￾ing differs. Figure (b) shows track errors for both models… view at source ↗
Figure 3
Figure 3. Simulated tracks of Hurricane Beryl under different soil moisture conditions us￾ing HWRFx. The ”Best Track” (dashed blue line) represents the observed hurricane path. The control simulation (CNTL) with default soil moisture is shown in black, while the Field Capac￾ity (FC) and Wilting Point (WP) scenarios are represented in blue and red, respectively. The results indicate a westward shift in the FC storm track and a… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Schematic of a proposed AI-based coupled model incorporating land and ocean processes into an atmospheric AI forecasting system. –9– [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 20 canonical work pages

  1. [1]

    , Zhang, X

    alaka2022high APACrefauthors Alaka, G J. , Zhang, X. \ Gopalakrishnan, S G. APACrefauthors \ 2022 . High-definition hurricanes: Improving forecasts with storm-following nests High-definition hurricanes: Improving forecasts with storm-following nests . Bulletin of the American Meteorological Society 103 3 E680--E703

  2. [2]

    , Hamrick, J B

    battaglia2018relational APACrefauthors Battaglia, P W. , Hamrick, J B. , Bapst, V. , Sanchez-Gonzalez, A. , Zambaldi, V. , Malinowski, M. others APACrefauthors \ 2018 . Relational inductive biases, deep learning, and graph networks Relational inductive biases, deep learning, and graph networks . arXiv preprint arXiv:1806.01261

  3. [3]

    , Thorpe, A

    bauer2015quiet APACrefauthors Bauer, P. , Thorpe, A. \ Brunet, G. APACrefauthors \ 2015 . The quiet revolution of numerical weather prediction The quiet revolution of numerical weather prediction . Nature 525 7567 47--55

  4. [4]

    , Wang, T

    chen2023dynamically APACrefauthors Chen, H. , Wang, T. , Zhang, Y. , Bai, Y. \ Chen, X. APACrefauthors \ 2023 . Dynamically weighted ensemble of geoscientific models via automated machine-learning-based classification Dynamically weighted ensemble of geoscientific models via automated machine-learning-based classification . Geoscientific Model Development...

  5. [5]

    , Franklin, J

    gall2013hurricane APACrefauthors Gall, R. , Franklin, J. , Marks, F. , Rappaport, E N. \ Toepfer, F. APACrefauthors \ 2013 . The hurricane forecast improvement project The hurricane forecast improvement project . Bulletin of the American Meteorological Society 94 3 329--343

  6. [6]

    , Marks, F

    gopalakrishnan2011experimental APACrefauthors Gopalakrishnan, S G. , Marks, F. , Zhang, X. , Bao, J W. , Yeh, K S. \ Atlas, R. APACrefauthors \ 2011 . The experimental HWRF system: A study on the influence of horizontal resolution on the structure and intensity changes in tropical cyclones using an idealized framework The experimental hwrf system: A study...

  7. [7]

    , Bates, G T

    hamill2013noaa APACrefauthors Hamill, T M. , Bates, G T. , Whitaker, J S. , Murray, D R. , Fiorino, M. , Galarneau, T J. Lapenta, W. APACrefauthors \ 2013 . NOAA's second-generation global medium-range ensemble reforecast dataset Noaa's second-generation global medium-range ensemble reforecast dataset . Bulletin of the American Meteorological Society 94 1...

  8. [8]

    , Bell, B

    hersbach2020era5 APACrefauthors Hersbach, H. , Bell, B. , Berrisford, P. , Hirahara, S. , Hor \'a nyi, A. , Mu \ n oz-Sabater, J. others APACrefauthors \ 2020 . The ERA5 global reanalysis The era5 global reanalysis . Quarterly Journal of the Royal Meteorological Society 146 730 1999--2049

Show all 24 references
  1. [9]

    , Neumann, C J

    jarvinen1984tropical APACrefauthors Jarvinen, B R. , Neumann, C J. \ Davis, M A. APACrefauthors \ 1984 . A tropical cyclone data tape for the North Atlantic Basin, 1886-1983: Contents, limitations, and uses A tropical cyclone data tape for the north atlantic basin, 1886-1983: ...

  2. [10]

    APACrefauthors \ 2018

    kepert2018boundary APACrefauthors Kepert, J D. APACrefauthors \ 2018 . The boundary layer dynamics of tropical cyclone rainbands The boundary layer dynamics of tropical cyclone rainbands . Journal of the Atmospheric Sciences 75 11 3777--3795

  3. [11]

    , Wood, K M

    klotzbach2022trends APACrefauthors Klotzbach, P J. , Wood, K M. , Schreck III, C J. , Bowen, S G. , Patricola, C M. \ Bell, M M. APACrefauthors \ 2022 . Trends in global tropical cyclone activity: 1990--2021 Trends in global tropical cyclone activity: 1990--2021 . Geophysical ...

  4. [12]

    , Dirmeyer, P A

    koster2004regions APACrefauthors Koster, R D. , Dirmeyer, P A. , Guo, Z. , Bonan, G. , Chan, E. , Cox, P. others APACrefauthors \ 2004 . Regions of strong coupling between soil moisture and precipitation Regions of strong coupling between soil moisture and precipitation . Scie...

  5. [13]

    , Knapp, K R

    kruk2010technique APACrefauthors Kruk, M C. , Knapp, K R. \ Levinson, D H. APACrefauthors \ 2010 . A technique for combining global tropical cyclone best track data A technique for combining global tropical cyclone best track data . Journal of Atmospheric and Oceanic Technolog...

  6. [14]

    , Sanchez-Gonzalez, A

    lam2023learning APACrefauthors Lam, R. , Sanchez-Gonzalez, A. , Willson, M. , Wirnsberger, P. , Fortunato, M. , Alet, F. others APACrefauthors \ 2023 . Learning skillful medium-range global weather forecasting Learning skillful medium-range global weather forecasting . Science...

  7. [15]

    , Dueben, P D

    rasp2020weatherbench APACrefauthors Rasp, S. , Dueben, P D. , Scher, S. , Weyn, J A. , Mouatadid, S. \ Thuerey, N. APACrefauthors \ 2020 . WeatherBench: a benchmark data set for data-driven weather forecasting Weatherbench: a benchmark data set for data-driven weather forecast...

  8. [16]

    , Moorthi, S

    saha2014ncep APACrefauthors Saha, S. , Moorthi, S. , Wu, X. , Wang, J. , Nadiga, S. , Tripp, P. others APACrefauthors \ 2014 . The NCEP climate forecast system version 2 The ncep climate forecast system version 2 . Journal of climate 27 6 2185--2208

  9. [17]

    , Dirmeyer, P A

    santanello2018land APACrefauthors Santanello Jr, J A. , Dirmeyer, P A. , Ferguson, C R. , Findell, K L. , Tawfik, A B. , Berg, A. others APACrefauthors \ 2018 . Land--atmosphere interactions: The LoCo perspective Land--atmosphere interactions: The loco perspective . Bulletin o...

  10. [18]

    , Armas, F

    santini2024hurricane APACrefauthors Santini, M. , Armas, F. , Carioli, A. , Ceccato, P. , Coecke, S. , De, G L. others APACrefauthors \ 2024 . Hurricane Beryl--JRC Emergency Report Hurricane beryl--jrc emergency report

  11. [19]

    \ Snyder, C

    torn2012uncertainty APACrefauthors Torn, R D. \ Snyder, C. APACrefauthors \ 2012 . Uncertainty of tropical cyclone best-track information Uncertainty of tropical cyclone best-track information . Weather and Forecasting 27 3 715--729

  12. [20]

    , Pritchard, M S

    wang2024coupled APACrefauthors Wang, C. , Pritchard, M S. , Brenowitz, N. , Cohen, Y. , Bonev, B. , Kurth, T. Pathak, J. APACrefauthors \ 2024 . Coupled Ocean-Atmosphere Dynamics in a Machine Learning Earth System Model Coupled ocean-atmosphere dynamics in a machine learning e...

  13. [21]

    , Durran, D R

    weyn2020improving APACrefauthors Weyn, J A. , Durran, D R. \ Caruana, R. APACrefauthors \ 2020 . Improving data-driven global weather prediction using deep convolutional neural networks on a cubed sphere Improving data-driven global weather prediction using deep convolutional ...

  14. [22]

    , Lindzen, R S

    zhang2016increasing APACrefauthors Zhang, B. , Lindzen, R S. , Tallapragada, V. , Weng, F. , Liu, Q. , Sippel, J A. Bender, M A. APACrefauthors \ 2016 . Increasing vertical resolution in US models to improve track forecasts of Hurricane Joaquin with HWRF as an example Increasi...

  15. [23]

    apacite url apacite =6pt Acknowledgments. 6pt 1sp \@dates Received \@recvdate\@empty\@rcvaccrule \@recvdate \@revisedate\@empty ; revised \@revisedate; \@accptdate\@empty \@revisedate\@empty; accepted \@accptdate \@pubdate\@empty. ; published \@pubdate. -2pt \@authaddrs @list\...

  16. [24]

    gc" journal option for G-Cubed Nov 3, 2003 M Kelly, fixed noindent in subsubsubsection titles and for all sections in rog option Oct 2, 2003 M Kelly, added

    \@ifstar \@figbox \@figbox \@figbox#1#2#3 to !#1! #3 [#1][c] !#2!#3 \@tempdima#2 \@tempdima by2 \@tempdima by- \@tempdima by- \@height\@tempdima\@depth\@tempdima\@width @ to @ #3 Bib ??? ??? ??? =0 =0 = @figure=0 @table=0 #1 --#1 -24pt -2ex #1 0= #1 to 0 #1 I NDEX T ERMS: #1 #...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.