REVIEW 3 major objections 6 minor 24 references
Advancing Seasonal Prediction of Tropical Cyclone Activity with a Hybrid AI-Physics Climate Model
T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash
Pith's one-line read A hybrid AI-physics climate model, forced only with persisted July 1 ocean anomalies, predicts Northern Hemisphere tropical cyclone activity for July through November with skill comparable to physical climate models.
desk verdict NeuralGCM yields a cheap, plausible seasonal TC hindcast, but the headline r≈0.7 is compromised by training-period overlap; out-of-sample skill is only reported qualitatively. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is NeuralGCM, a differentiable hydrostatic dynamical core whose unresolved subgrid processes, principally convection, are replaced by a learned single-column neural network shared across grid columns. The experimental mechanism is the boundary-forcing simplification borrowed from physical seasonal predictions: sea surface temperature and sea ice follow the climatological seasonal cycle with the July 1 anomaly held fixed, which captures the ocean's slow thermal evolution without an ocean model. Twenty-member ensembles are initialized on July 1 each year from 1990 to 2023, with perturbations to the encoder's learned correction; a vorticity-based tracker converts model output into cyclone tracks. This combination, rather than any new training or model modification, carries the argument: it lets a weather-trained ML model respond to the SST conditions that govern tropical cyclone activity.
What would settle it
Run the same NeuralGCM hindcast protocol with climatological SST and sea ice (no July 1 anomalies) for 1990–2023. If the North Atlantic and Northeast Pacific TC-frequency correlations with observations remain near r=0.7, the claimed seasonal skill is not coming from the model's response to the persisted boundary anomaly; if they collapse, the skill is genuinely forced by the SST information.
Extended reading notes
Core claim
The central claim is that a hybrid ML-physics atmospheric model, given only persisted boundary anomalies from the initialization date plus the climatological seasonal cycle, can skillfully predict the tropical atmosphere and Northern Hemisphere tropical cyclone activity months ahead. On the paper's terms, this establishes that the model has learned enough of the atmospheric response to boundary forcing that simplified persistence forcing is sufficient, at least in TC-active basins. The supporting evidence is the 1990–2023 hindcast correlation (r≈0.7) for TC frequency in the North Atlantic and Northeast Pacific, significant correlations for sub-basin track density (p<0.1) and basin-wide accumulated cyclone energy (p<0.01) in the North Atlantic and North Pacific, and a realistic simulated TC climatology. The authors also position the skill as comparable to earlier physical GCM seasonal prediction studies, while noting that the simplified forcing makes the reported skill a lower-bound estimate.
Load-bearing premise
The load-bearing premise is that holding the July 1 sea-surface temperature and sea-ice anomalies fixed through the climatological cycle is a sufficient boundary forcing, so the reported tropical cyclone skill could largely be inherited from the observed SST anomaly rather than produced by the model's learned dynamics.
Editorial extensions
If this is right
- If the claim holds, a single GPU can produce a 20-member, 5-month seasonal tropical cyclone hindcast in roughly 8 minutes per 100 simulation days, making ensemble seasonal prediction far cheaper than current physical models.
- Skillful TC frequency prediction in the North Atlantic and Northeast Pacific becomes achievable without an ocean model or atmosphere-ocean coupling, so the bottleneck shifts to boundary-condition quality.
- The same setup could be rerun for other initialization months, other basins, or other TC metrics, because the protocol is simple and modular.
- The demonstrated response to persisted SST anomalies gives a benchmark against which future ML climate models that add coupled ocean or land components can be judged.
- Because the model runs far faster than physical GCMs, the hindcast record can be extended and ensemble size increased, tightening skill estimates.
Reading between the lines
- A direct test not performed in the paper would be to replace the persisted SST anomalies with climatological SST in the same protocol; if TC-frequency correlations stay near r=0.7, much of the apparent skill is inherited from the observed June SST anomaly rather than from the model's learned dynamics.
- The paper's assumption that unstable stochastic rollouts are random with respect to TC activity is untested; replacing them with climatology could bias skill upward if instabilities correlate with active seasons.
- The skill in the North Atlantic and Northeast Pacific likely traces to the strong SST-TC relationship in those basins; the same protocol applied to the North Indian Ocean or to Atlantic regions like the Caribbean and Gulf of Mexico should show weaker skill, and indeed the paper reports such regional underprediction.
- A natural extension is to feed NeuralGCM predicted SST fields from a coupled or statistical ocean model instead of persisting July 1 anomalies; comparing the two would separate model dynamics from boundary information.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper reports retrospective seasonal predictions of Northern Hemisphere tropical cyclone (TC) activity using the 1.4-degree NeuralGCM hybrid ML-physics atmospheric model. Starting from 1 July initial conditions for 1990–2023, the authors run 20-member ensembles forced with the climatological seasonal cycle of SST and sea ice superimposed with anomalies persisting from the initialization date. They evaluate July–November large-scale fields against ERA5 and TC tracks, counts, and ACE against IBTrACS, using TempestExtreme with adjusted vorticity and duration thresholds. The central claim is that the model reproduces observed TC climatology and achieves statistically significant interannual skill, notably basin-wide TC frequency correlations of about 0.7 in the North Atlantic and Northeast Pacific, comparable to physical GCMs with similar persistence-type boundary forcing. The paper also reports computational costs (~8 minutes per 100 simulation days on one GPU) and compares with SEAS5 in the supplementary material.
Significance. If the central claim holds, this is a useful demonstration that a hybrid ML-physics model can provide a computationally cheap baseline for seasonal TC prediction and motivate coupling such models to ocean and sea-ice components. The paper has several strengths: it evaluates against an external observational dataset (IBTrACS) rather than reanalysis-derived tracks; it provides confidence intervals by resampling; it benchmarks against published physical-model results (Chen and Lin 2013; Zhang et al. 2019) and includes a preliminary SEAS5 comparison; and code and data links are provided. The main caveats are that most evaluation years lie inside the model's training period and that the persistence-SST forcing may be responsible for much of the apparent skill. These issues are addressable with additional quantitative out-of-sample and baseline analyses.
major comments (3)
- [Section 2.1, 2.2, 3.2, Figure 4] The deterministic NeuralGCM was trained on ERA5 1979–2017 and the stochastic version on 1979–2019 (Section 2.1), while the headline correlation r≈0.7 is computed over 1990–2023. This means 28 of 34 deterministic verification years are inside the training window, so the reported skill is not a clean out-of-sample estimate. The Supplementary Materials state that 2018–2023 'also show some skill' but do not report the correlation coefficient or its significance. Please provide explicit out-of-sample skill metrics, such as the correlation for 2018–2023 alone or a leave-period-out analysis, and compare them with in-sample skill. If out-of-sample skill is much lower, the claim of useful seasonal predictions needs to be softened.
- [Section 2.2 and 3.2] Because the boundary forcing is simply the observed July 1 SST and sea-ice anomaly persisted through the season, much of the TC skill may be inherited from the observed SST state rather than from NeuralGCM's learned dynamics. The paper shows environmental skill relative to a persistence baseline in Supplementary Figure 11, but no equivalent persistence baseline is provided for basin-wide TC counts or ACE. Please add a baseline that correlates observed TC activity with the initial June/July SST anomaly alone, or with a simple statistical persistence forecast, and show whether NeuralGCM adds skill beyond it. Without such a baseline, the attribution of r≈0.7 to the model dynamics is not secure.
- [Section 2.3] The tracker's vorticity threshold (4×10^-5 s^-1) and duration threshold (54 h) are explicitly tuned to match IBTrACS climatological counts. This is not a direct fit to interannual skill, but it is an additional tunable choice that could affect the reported correlations. Please report how the interannual correlations vary over a plausible range of these thresholds, and state whether any tuning was performed against the skill metrics themselves. In addition, about 10% of stochastic-physics simulations are flagged as unstable and replaced with climatological fields (Section 2.3, Supplementary Figure 1); please show that the deterministic headline results are insensitive to excluding unstable years, or that the stochastic results are robust to alternative handling of those members.
minor comments (6)
- [Section 3.2, paragraph 2] There is a typo: 'NeruralGCM' should be 'NeuralGCM'.
- [Table 1] The two NeuralGCM rows are not explicitly labeled as deterministic and stochastic in the table; adding labels would make the comparison much clearer.
- [Supplementary Figure 15 caption] The unit 'm-2 s-2' should be 'm^2 s^-2'.
- [Section 2.3] The vorticity threshold '4×10!"!" s!"#' is a rendering error; it should be 4×10^-5 s^-1.
- [Introduction] The notation '10-3 to 10-5' should be formatted as 10^-3 to 10^-5.
- [Figure 4] With 20 individual ensemble members plotted as light blue lines, the ensemble mean is difficult to distinguish; please consider plotting a subset of members or a shaded spread.
Circularity Check
No significant circularity: the reported TC prediction skill is evaluated against external observations and is not fitted by construction.
full rationale
The paper's derivation chain is not circular. NeuralGCM is a pretrained hybrid ML-physics model from Kochkov et al. (2024), trained on ERA5 weather forecasts over short time windows, not on seasonal TC counts or on the IBTrACS target. The authors configure boundary forcings by persisting July 1 SST and sea ice anomalies, following established physical-model practice, and then track TCs in the resulting hindcasts with thresholds tuned only to match climatological TC counts, not the interannual correlations that constitute the headline skill. The evaluation target (IBTrACS) is external to the training data, and the reported r values for TC frequency and ACE are not parameters fitted to the observations. The cited prior work on NeuralGCM is code-reproduced and externally published; the overlap of one coauthor is not load-bearing. The training-period overlap (1990–2017 lies inside the ERA5 training window) is a legitimate validation and generalization concern, but it is not a definitional equivalence: the model was not trained to predict the evaluated seasonal TC metrics, so the skill estimate is not forced by construction. The supplementary statement that 2018–2023 'also show some skill' is a reporting limitation, not evidence that the headline correlation was manufactured from the training label. Overall, the central claim has independent evidentiary content, and no circular step can be exhibited from the paper's own equations or fitting choices.
Assumptions & free parameters
free parameters (5)
- TC tracker vorticity threshold =
4e-5 s^-1
- TC storm duration threshold =
54 hours
- Ensemble perturbation correlation length =
1000 km
- Stability flag threshold =
2x initial zonal standard deviation
- MDR latitude and longitude bounds =
10N-25N; NA 90W-20W, EP 140W-100W, WP 110E-160E
assumptions (6)
- domain assumption ERA5 provides sufficiently accurate initial conditions, boundary conditions, and verification fields for this hindcast evaluation.
- domain assumption The pre-trained deterministic and stochastic NeuralGCM versions generalize from short-window weather training to 5-month forced rollouts.
- domain assumption Persisting July 1 SST and sea ice anomalies through the climatological seasonal cycle is a reasonable approximation of boundary evolution during July-November.
- ad hoc to paper Flagged unstable simulations, replaced by climatological fields, fail for reasons unrelated to the TC quantities being scored.
- domain assumption The vorticity-based TempestExtreme tracker with lowered thresholds identifies storms whose counts and tracks are comparable to IBTrACS.
- domain assumption Detrending with linear least squares removes only the secular trend and leaves predictable interannual variability intact.
Cite this review
Pith. "Pith review of Advancing Seasonal Prediction of Tropical Cyclone Activity with a Hybrid AI-Physics Climate Model." pith.science (2026). https://pith.science/paper/TZYPUKKZ
@misc{pith2026250501455,
author = {Pith},
title = {Pith review of: Advancing Seasonal Prediction of Tropical Cyclone Activity with a Hybrid AI-Physics Climate Model},
year = {2026},
howpublished = {\url{https://pith.science/paper/TZYPUKKZ}},
note = {Machine review of arXiv:2505.01455}
}
read the original abstract
Machine learning (ML) models are successful with weather forecasting and have shown progress in climate simulations, yet leveraging them for useful climate predictions needs exploration. Here we show this feasibility using Neural General Circulation Model (NeuralGCM), a hybrid ML-physics atmospheric model developed by Google, for seasonal predictions of large-scale atmospheric variability and Northern Hemisphere tropical cyclone (TC) activity. Inspired by physical model studies, we simplify boundary conditions, assuming sea surface temperature (SST) and sea ice follow their climatological cycle but persist anomalies present at the initialization time. With such forcings, NeuralGCM can generate 100 simulation days in ~8 minutes with a single Graphics Processing Unit (GPU), while simulating realistic atmospheric circulation and TC climatology patterns. This configuration yields useful seasonal predictions (July to November) for the tropical atmosphere and various TC activity metrics. Notably, the predicted and observed TC frequency in the North Atlantic and East Pacific basins are significantly correlated during 1990 to 2023 (r=~0.7), suggesting prediction skill comparable to existing physical GCMs. Despite challenges associated with model resolution and simplified boundary forcings, the model-predicted interannual variations demonstrate significant correlations with the observation, including the sub-basin TC tracks (p<0.1) and basin-wide accumulated cyclone energy (p<0.01) of the North Atlantic and North Pacific basins. These findings highlight the promise of leveraging ML models with physical insights to model TC risks and deliver seamless weather-climate predictions.
Reference graph
Works this paper leans on
-
[1]
Advancing Seasonal Prediction of Tropical Cyclone Activity with a Hybrid AI-Physics Climate Model Gan Zhang1*, Megha Rao1, Janni Yuval2, Ming Zhao3 1 Department of Climate, Meteorology, and Atmospheric Sciences, University of Illinois at Urbana-Champaign 2 Google Research 3 Geophysical Fluid Dynamics Laboratory, National Oceanic and Atmospheric Administra...
work page 1990
-
[2]
Introduction Machine learning (ML) models recently made breakthroughs in weather forecasting (e.g., Bi et al 2023, Keisler 2022, Kochkov et al 2024, Lam et al 2023, Pathak et al 2022, Price et al 2023). Trained with atmospheric reanalysis or physical model data, these ML models delivered successful forecasts up to two weeks of lead time with skills compar...
work page 2023
-
[3]
1 Model Skill with Large-scale Atmospheric Environment The NeuralGCM hindcasts with simplified boundary forcings simulate the atmospheric climate and seasonal cycle realistically (Figure 1). The July-November means of the 500-hPa geopotential height of the NeuralGCM and the ERA5 show consistent climate patterns. Their differences are the smallest in the t...
work page 2020
-
[4]
Summary and Discussion This study conducts experimental seasonal predictions with the newly available NeuralGCM and simplified boundary forcings. Inspired by earlier studies with physical GCMs, the hindcast experiments focus on July to November which account for most TC activity in the Northern Hemisphere. The NeuralGCM hindcasts of 1990–2023 can simulate...
work page 1990
-
[9]
associated with unrealistic convection and stratosphere features. These waves mostly appear in the tropics and violate the weak gradient constraint of the real-world atmosphere (Charney 1963, Sobel and Bretherton 2000). While fixes are being explored, this study proceeds by labeling the simulations with spurious waves using a check of tropical variability...
work page 1963
-
[10]
The tracking uses the vorticity-based method and does not impose any wind speed thresholds
to track TCs in our retrospective prediction experiments. The tracking uses the vorticity-based method and does not impose any wind speed thresholds. We follow most parameter choices of the TC tracker used by Kochkov et al. (2024) who tuned the parameters such that the TC counts of the ERA5 at 0.25-degree resolution match the values at 1.4 degrees. To bet...
work page 2024
-
[13]
Table 1 shows a direct, like-for-like comparison of prediction skill for TC frequency
in those two basins . Table 1 shows a direct, like-for-like comparison of prediction skill for TC frequency. When model performance is ranked for the North Atlantic and the Northeast Pacific, the NeuralGCM hindcasts with deterministic physics are comparable to or better than at least one of the examined physical models. The prediction skill is associated ...
work page 2019
-
[15]
and other aspects of the Earth system (e.g., Yeager et al 2022). Such development can be accomplished by coupling NeuralGCM with statistical models, ML emulators, or other hybrid models of the ocean and other Earth system components. Contributing to the rapidly evolving field of the ML-based climate modeling, this study demonstrates a practical applicatio...
arXiv 2022
Show all 24 references
-
[17]
The ECMWF seasonal forecasts (SEAS5) are produced every month with a 51-member ensemble at a horizontal resolution of ~36 km
and the observation for a set of available variables (Supplementary Figures 8–10) to facilitate preliminary comparisons of prediction skill. The ECMWF seasonal forecasts (SEAS5) are produced every month with a 51-member ensemble at a horizontal resolution of ~36 km. The SEAS5 ...
2019
-
[19]
In both seasons, NeuralGCM manages to predict the large-scale patterns in the tropics, which have strong associations with TC frequency (Supplementary Figure 14)
to examine potential causes (Supplementary Figure 15). In both seasons, NeuralGCM manages to predict the large-scale patterns in the tropics, which have strong associations with TC frequency (Supplementary Figure 14). The prediction in the extratropics show some notable disagr...
2005
-
[20]
and decays in later months. The skill decay across individual months is not monotonic (e.g., Supplementary Figure 16b), suggesting a larger number of predictions that sample more initialization time are needed to robustly evaluate the subseasonal prediction skill. Consistent w...
2022
-
[21]
Supplementary Figure 8 Skill metrics of the ECWMF SEAS5 for the 500-hPa geopotential (m2 s-2)
The input variables were detrended using a linear least-squared fitting. Supplementary Figure 8 Skill metrics of the ECWMF SEAS5 for the 500-hPa geopotential (m2 s-2). The data of SEAS5 (version 5.1) of 1990-2019 are regridded to the NeuralGCM grid. The analyses only consider ...
1990
-
[22]
The other settings are identical to Figure
Supplementary Figure 12 Predictions and observations of the basin-wide TC counts (1990-2023) and the accumulated cyclone energy (ACE; unit: 10-4 knot2) using the NeuralGCM with stochastic model physics. The other settings are identical to Figure
1990
-
[23]
The blue dots show the individual data points
Supplementary Figure 13 (a) The relationship between the surface pressure (hPa) and the 850-hPa maximum wind speed (m s-1) in the NeuralGCM hindcasts with the deterministic model physics. The blue dots show the individual data points. The red contours show the kernel density e...
1990
-
[24]
Supplementary Figure 16 Monthly tropical cyclone counts and prediction skill (1990–2023)
(e–h), (i–l), and (m–p), are the same as (a–d), but for the anomalies of surface pressure (Pa), 1000-hPa temperature (K), and the vertical shear of 200-hPa and 850 hPa zonal wind (m s-1). Supplementary Figure 16 Monthly tropical cyclone counts and prediction skill (1990–2023)....
2024
-
[1267]
Model Dev
(Geneva: WMO) Yeager S G, Rosenbloom N, Glanville A A, Wu X, Simpson I, Li H, Molina M J, Krumhardt K, Mogen S, Lindsay K, Lombardozzi D, Wieder W, Kim W M, Richter J H, Long M, Danabasoglu G, Bailey D, Holland M, Lovenduski N, Strand W G and King T 2022 The Seasonal-to-Multiy...
2024 arXiv
-
[1998]
The exploratory work with atmospheric GCMs served as a stepping stone for the ensuing development of more advanced prediction systems (e.g., Delworth et al 2020, Vecchi et al 2014)
and TC activity (e.g., Gray 1984). The exploratory work with atmospheric GCMs served as a stepping stone for the ensuing development of more advanced prediction systems (e.g., Delworth et al 2020, Vecchi et al 2014). Inspired by the recent NeuralGCM development and the previou...
1984
-
[2017]
To facilitate comparisons with the NeuralGCM hindcasts, we randomly select 20 members of the SEAS5 ensemble initialized on July 1 of 1990-2019. Our preliminary comparison suggests that SEAS5 delivers more skillful predictions of the environmental variables (Supplementary Figur...
1990
-
[2019]
These models, such as the ECMWF seasonal forecasts (SEAS5) (Johnson et al 2019), are physical models with higher spatial resolutions of the atmosphere (e.g., 36-km grid spacing)
for reference. These models, such as the ECMWF seasonal forecasts (SEAS5) (Johnson et al 2019), are physical models with higher spatial resolutions of the atmosphere (e.g., 36-km grid spacing). We briefly discuss the performance of models in Section 3 and present additional an...
2019
-
[2020]
The gridded ERA5 is generated by a numeric weather forecast model that follows physical laws and ingests multi-sourced observational data (e.g., weather station and satellite data)
serves as the primary data for the training, configuration, and validation of NeuralGCM simulations. The gridded ERA5 is generated by a numeric weather forecast model that follows physical laws and ingests multi-sourced observational data (e.g., weather station and satellite d...
2024
-
[2021]
This TC focus also helps us leverage proven concepts and knowledge in physical model development (Chen and Lin 2013, Zhao et al 2010)
and often remain challenging for physical GCMs to simulate (Roberts et al 2020). This TC focus also helps us leverage proven concepts and knowledge in physical model development (Chen and Lin 2013, Zhao et al 2010). Overall, this effort establishes a performance baseline for f...
2020
-
[2022]
When calculating the anomalies of SST and sea ice at the initialized time, we use the daily climatology of 1991–2020 that is resampled using the monthly data
during July–November. When calculating the anomalies of SST and sea ice at the initialized time, we use the daily climatology of 1991–2020 that is resampled using the monthly data. To ensure the consistency among variables and the configurations described by Kochkov et al. (20...
2024
-
[2023]
Inspired by the seasonal prediction experiments by Zhao et al
We also generated additional simulations (e.g., 1979–1989) to facilitate comparisons with previous TC studies that used physical models (Supplementary Materials). Inspired by the seasonal prediction experiments by Zhao et al. (2010) and Chen & Lin (2013), we use the climatolog...
2010
-
[2024]
and advancing fundamental understanding of atmospheric predictability (e.g., Vonich and Hakim 2024). Similar to the early development of NWP and climate models (Phillips 1956), the success of ML models in weather forecasting also inspired researchers to explore their potential...
2024
Reviewed August 16, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.