pith. sign in

arxiv: 2606.30920 · v1 · pith:CLDH6TKTnew · submitted 2026-06-29 · ⚛️ physics.ao-ph · cs.LG

Conditional Tropical Cyclogenesis Rates via Rare-Event Sampling in a Neural Weather Emulator

John S. Schreck , William Chapman , Charlie Becker , David John Gagne II This is my paper

Pith reviewed 2026-07-01 00:50 UTC · model grok-4.3

classification ⚛️ physics.ao-ph cs.LG
keywords tropical cyclogenesisforward flux samplingneural weather emulatorrare-event samplingAtlantic hurricanesconditional ratesSDL-WXFormer
0
0 comments X

The pith

Forward flux sampling with a neural emulator estimates tropical cyclogenesis rates across three orders of magnitude from 98 initial conditions.

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

The paper couples Forward Flux Sampling, a rare-event technique, to a neural weather emulator to compute conditional rates at which tropical cyclones reach hurricane strength without changing the model's dynamics. Direct ensemble runs cannot resolve the wide variability in these rates because the events are too rare, but the sampling method breaks the path into an initial flux and a chain of conditional crossing probabilities across pressure interfaces. When run on 98 Atlantic initial conditions from August to October 2022, the approach produces rates that differ by nearly 1000 times and follow the observed seasonal pattern. A direct comparison to independent sampling on the same conditions yields rates that agree within 15 percent on average. The method also delivers computational speedups that grow larger in environments where genesis is suppressed.

Core claim

Forward Flux Sampling decomposes the rare disturbance-to-mature-cyclone path into a flux through an initial interface pressure and a product of conditional crossing probabilities across four intermediate interface pressures. Applied to the 1-degree SDL-WXFormer neural emulator on 98 Atlantic basin initial conditions spanning 21 August to 8 October 2022, the method resolves conditional genesis rates spanning nearly three orders of magnitude while reproducing a seasonal cycle qualitatively consistent with observations. Self-consistency checks against independent direct-sampling rates give a mean ratio of 1.03 with standard deviation 0.15, and computational enhancement factors range from 3X to

What carries the argument

Forward Flux Sampling, which decomposes rare intensification paths into an initial flux and successive conditional crossing probabilities across four pressure interfaces.

If this is right

  • Genesis rates span nearly three orders of magnitude across the sampled initial conditions.
  • The computed rates capture a seasonal cycle qualitatively consistent with observations.
  • Self-consistency checks against direct sampling produce a mean ratio of 1.03 plus or minus 0.15.
  • Computational speedups range from 3X in active environments to 140X in suppressed ones.
  • Case studies identify different rate-limiting steps in different environments, such as initial organization for Earl and final intensification for Ian.

Where Pith is reading between the lines

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

  • The same sampling framework could be applied to other rare atmospheric transitions once suitable emulators exist.
  • Seasonal or subseasonal forecast systems might incorporate these conditional rates to improve probabilistic guidance.
  • Finer-resolution emulators would allow the method to diagnose physical mechanisms at scales closer to observed storm structure.

Load-bearing premise

The neural emulator must reproduce atmospheric dynamics and variability for rare intensification trajectories without introducing systematic bias in the crossing probabilities.

What would settle it

Running the same 98 initial conditions through a high-resolution physics-based model with direct sampling and finding large systematic differences from the FFS rates would indicate bias in the emulator.

Figures

Figures reproduced from arXiv: 2606.30920 by Charlie Becker, David John Gagne II, John S. Schreck, William Chapman.

Figure 1
Figure 1. Figure 1: Schematic of the Forward Flux Sampling algorithm (after Allen et al. 2006). [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of conditional crossing probabilities [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FFS genesis rates across all 98 initial conditions (August 21 – October 8, 2022) on a [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Seasonal evolution of FFS conditional crossing probabilities [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Log-log scatter of FFS genesis rate k FFS versus independent direct-sampling rate k direct for all 98 initial conditions. Points are coloured by the ratio k FFS/kdirect: green indicates near-unity agreement; red/blue indicate over- or under-estimation. The black dashed line is the 1:1 reference; shaded band spans ±50%. Three case-study ICs are marked with stars (Earl: red; Fiona: orange; Ian: yellow). The … view at source ↗
Figure 6
Figure 6. Figure 6: Geographic distribution of state-B arrival locations (MSLP ≤ 975 hPa) across all 98 initial conditions. Each point marks where a reactive trajectory first reached state B. Colour encodes the IC initialization date from August 21 (light) to October 8 (dark). The 2◦ density contours (background) show the spatial clustering of genesis events across the season. Approximate observed genesis locations for Earl, … view at source ↗
Figure 7
Figure 7. Figure 7: Reactive genesis trajectories for the Earl IC (2022-09-02T00Z). [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FFS shooting trees for the three case-study ICs on a single combined map. Each star ( [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Composite 500 hPa geopotential height (filled, row 1) with 500 hPa wind vectors and [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Genesis committor curve pB(λi) — the probability of eventually reaching state B (≤975 hPa) given arrival at interface λi — averaged across all 98 initial conditions. Blue: FFS AI model mean ±1σ across ICs. Orange: IFS brute-force ensemble mean ±1σ across ICs. The shaded bands show IC-to-IC variability. The annotated values are the all-season FFS means. The grey band highlights the rate-limiting step (larg… view at source ↗
Figure 11
Figure 11. Figure 11: Storm-centred physics along a representative reactive pathway for the Earl IC (2022-09- [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Storm-centred physics composites along reactive pathways for the Earl IC (2022-09- [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: As in Fig. 12 but for the Fiona IC (2022-09-09T12Z). Stronger vorticity and warm [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: As in Fig. 12 but for the Ian IC (2022-09-22T00Z). The [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Convergence of FFS estimates as a function of the number of shooting attempts per [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗
read the original abstract

We couple Forward Flux Sampling (FFS), a non-equilibrium rare-event technique from statistical mechanics, to a neural weather emulator (SDL-WXFormer, 1{\deg} grid spacing) to estimate conditional tropical cyclogenesis rates, or how often a tropical cyclone achieves a hurricane-level central pressure, without modifying model dynamics. Tropical cyclogenesis rates vary by orders of magnitude across regimes, yet direct ensemble sampling cannot resolve this variability at operationally feasible ensemble sizes. FFS decomposes the rare disturbance to mature cyclone intensification path into a flux through an initial interface pressure and a product of conditional crossing probabilities across four intermediate interface pressures. We use the 1{\deg} emulator because FFS requires O(10^4) model trajectories per initial condition, and because the model's calibrated stochastic layers provide the necessary exploratory spread. Applied to 98 Atlantic basin initial conditions spanning 21 August - 8 October 2022, FFS resolves genesis rates spanning nearly three orders of magnitude, capturing a seasonal cycle qualitatively consistent with observations. A self-consistency check comparing FFS rates to independent direct-sampling rates yields a mean ratio of 1.03 +/- 0.15 across all initial conditions. Computational enhancement factors range from 3X (most active environment) to 140X (most suppressed), with a geometric mean of 14X. Three case studies illustrate the physical diagnostics the method provides: the rate-limiting step is initial tropical organization for the Earl environment, uniformly high crossing probabilities for the Fiona precursor environment, and a compound barrier at the final intensification stages for the Ian environment. More efficient emulators would enable application of FFS to finer resolutions.

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

1 major / 0 minor

Summary. The manuscript couples Forward Flux Sampling (FFS) to the 1° SDL-WXFormer neural weather emulator to compute conditional tropical cyclogenesis rates for 98 Atlantic initial conditions (21 Aug–8 Oct 2022). FFS decomposes the intensification path into an initial flux and products of conditional crossing probabilities across four pressure interfaces. Reported results include genesis rates spanning nearly three orders of magnitude, qualitative seasonal-cycle agreement with observations, a self-consistency ratio of 1.03 ± 0.15 versus direct sampling, computational speedups of 3–140× (geometric mean 14×), and case-study diagnostics identifying rate-limiting steps for the Earl, Fiona, and Ian environments.

Significance. If the emulator’s stochastic layers reproduce the variability required for rare intensification trajectories, the work supplies an efficient, non-intrusive method for resolving order-of-magnitude variability in genesis rates that direct ensembles cannot access. The internal self-consistency check and the physical diagnostics extracted from the case studies constitute clear methodological strengths. The approach is extensible to other rare atmospheric events provided emulator fidelity for those trajectories can be established.

major comments (1)
  1. [Abstract] Abstract (final two paragraphs) and the self-consistency paragraph: the reported three-order-of-magnitude rate variability and qualitative seasonal agreement rest on the assumption that the 1° emulator’s crossing probabilities match real atmospheric statistics. The only quantitative validation supplied is the internal ratio 1.03 ± 0.15 against direct sampling within the same emulator; no comparison of the 98 computed rates to observed 2022 genesis outcomes, reanalysis, or higher-resolution models is presented, leaving the central claim load-bearing on an untested fidelity assumption explicitly flagged by the authors’ own remark on the need for finer-resolution emulators.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and for recognizing the methodological strengths of the FFS-emulator coupling. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract (final two paragraphs) and the self-consistency paragraph: the reported three-order-of-magnitude rate variability and qualitative seasonal agreement rest on the assumption that the 1° emulator’s crossing probabilities match real atmospheric statistics. The only quantitative validation supplied is the internal ratio 1.03 ± 0.15 against direct sampling within the same emulator; no comparison of the 98 computed rates to observed 2022 genesis outcomes, reanalysis, or higher-resolution models is presented, leaving the central claim load-bearing on an untested fidelity assumption explicitly flagged by the authors’ own remark on the need for finer-resolution emulators.

    Authors: We agree that the only quantitative validation is the internal self-consistency ratio and that no direct comparison of the 98 rates to 2022 observations, reanalysis, or higher-resolution models is provided. The manuscript is a methodological demonstration that the reported rates and variability are those generated by the 1° SDL-WXFormer; the qualitative seasonal-cycle statement is based on the overall pattern of higher rates during more active periods within the emulator. We will revise the abstract (final two paragraphs) and the self-consistency section to state explicitly that the results characterize the emulator’s statistics, that external fidelity remains to be established with finer-resolution models (as already noted in the text), and that the central claims concern the method’s performance rather than direct reproduction of observed genesis outcomes. This revision will be incorporated in the next version. revision: yes

Circularity Check

0 steps flagged

No significant circularity; rates obtained from direct sampling on emulator

full rationale

The derivation computes conditional genesis rates by applying FFS to O(10^4) trajectories generated by the SDL-WXFormer emulator and validates the method via an independent direct-sampling comparison (mean ratio 1.03 +/- 0.15). This internal check confirms numerical consistency of the rare-event algorithm inside the fixed emulator but does not reduce the reported rates to a fitted parameter, self-definition, or self-citation chain. No equations or steps in the abstract equate the output to its inputs by construction, and the emulator itself is treated as an external black-box model whose dynamics are not redefined within the paper.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the fidelity of the neural emulator for rare paths and the standard applicability of FFS decomposition; no free parameters or invented entities are described in the abstract.

axioms (2)
  • domain assumption The neural weather emulator accurately captures the dynamics and stochastic variability needed for rare-event intensification paths.
    The method requires O(10^4) trajectories per initial condition and relies on the emulator's stochastic layers to provide exploratory spread.
  • standard math Forward Flux Sampling correctly factors the rare-event probability into an initial flux and a product of conditional crossing probabilities across interfaces.
    Standard non-equilibrium statistical mechanics technique invoked without modification to model dynamics.

pith-pipeline@v0.9.1-grok · 5841 in / 1387 out tokens · 41459 ms · 2026-07-01T00:50:02.341907+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 11 canonical work pages · 2 internal anchors

  1. [1]

    Allen, R. J. and Warren, P. B. and ten Wolde, P. R. , title =. Physical Review Letters , volume =. 2005 , doi =

    2005

  2. [2]

    Simulating rare events in equilibrium or nonequilibrium stochastic systems

    Allen, Rosalind J and Frenkel, Daan and ten Wolde, Pieter Rein. Simulating rare events in equilibrium or nonequilibrium stochastic systems. J. Chem. Phys. doi:10.1063/1.2140273

    work page doi:10.1063/1.2140273

  3. [3]

    Allen, R. J. and Valeriani, C. and ten Wolde, P. R. , title =. J. Phys.: Condens. Matter , volume =. 2009 , doi =

    2009

  4. [4]

    and Allen, R

    Valeriani, C. and Allen, R. J. and Morelli, M. J. and Frenkel, D. and ten Wolde, P. R. , title =. J. Chem. Phys. , volume =. 2007 , doi =

    2007

  5. [5]

    Schreck, J. S. and Ouldridge, T. E. and Romano, F. and. Nucleic Acids Res. , volume =. 2015 , doi =

    2015

  6. [6]

    Bolhuis, P. G. and Chandler, D. and Dellago, C. and Geissler, P. L. , title =. Annu. Rev. Phys. Chem. , volume =. 2002 , doi =

    2002

  7. [7]

    and Parrinello, M

    Laio, A. and Parrinello, M. , title =. Proc. Natl. Acad. Sci. , volume =

  8. [8]

    Khoo , author J

    Khoo, Yuehaw and Lu, Jianfeng and Ying, Lexing. Solving for high-dimensional committor functions using artificial neural networks. Res. Math. Sci. doi:10.1007/s40687-018-0160-2

    work page doi:10.1007/s40687-018-0160-2

  9. [9]

    Adaptive multilevel splitting for rare event analysis , journal =

    C. Adaptive multilevel splitting for rare event analysis , journal =. 2007 , doi =

    2007

  10. [10]

    and Wouters, J

    Ragone, F. and Wouters, J. and Bouchet, F. , title =. Proc. Natl. Acad. Sci. , volume =. 2018 , doi =

    2018

  11. [11]

    Webber, R. J. and Plotkin, D. A. and O'Neill, M. E. and Abbot, D. S. and Weare, J. , title =. Chaos , volume =. 2019 , doi =

    2019

  12. [12]

    Bringing statistics to storylines: Rare event sampling for sudden, transient extreme events

    Finkel, Justin and O'Gorman, Paul A. Bringing statistics to storylines: Rare event sampling for sudden, transient extreme events. J. Adv. Model. Earth Syst. doi:10.1029/2024ms004264

    work page doi:10.1029/2024ms004264

  13. [13]

    Plotkin, D. A. and Webber, R. J. and O'Neill, M. E. and Weare, J. and Abbot, D. S. , title =. J. Adv. Model. Earth Syst. , volume =. 2019 , doi =

    2019

  14. [14]

    A large deviation theory-based analysis of heat waves and cold spells in a simplified model of the general circulation of the atmosphere , journal =

    G. A large deviation theory-based analysis of heat waves and cold spells in a simplified model of the general circulation of the atmosphere , journal =. 2019 , doi =

    2019

  15. [15]

    , title =

    Touchette, H. , title =. Phys. Rep. , volume =. 2009 , doi =

    2009

  16. [16]

    Learning skillful medium-range global weather forecasting

    Lam, Remi and Sanchez-Gonzalez, Alvaro and Willson, Matthew and Wirnsberger, Peter and Fortunato, Meire and Alet, Ferran and Ravuri, Suman and Ewalds, Timo and Eaton-Rosen, Zach and Hu, Weihua and Merose, Alexander and Hoyer, Stephan and Holland, George and Vinyals, Oriol and Stott, Jacklynn and Pritzel, Alexander and Mohamed, Shakir and Battaglia, Peter....

    work page doi:10.1126/science.adi2336

  17. [17]

    Accurate medium-range global weather forecasting with 3D neural networks.Nature, 619(7970):533– 538, July 2023

    Bi, Kaifeng and Xie, Lingxi and Zhang, Hengheng and Chen, Xin and Gu, Xiaotao and Tian, Qi. Accurate medium-range global weather forecasting with 3D neural networks. Nature. doi:10.1038/s41586-023-06185-3

    work page doi:10.1038/s41586-023-06185-3

  18. [18]

    FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators

    Pathak, Jaideep and Subramanian, Shashank and Harrington, Peter and Raja, Sanjeev and Chattopadhyay, Ashesh and Mardani, Morteza and Kurth, Thorsten and Hall, David and Li, Zongyi and Azizzadenesheli, Kamyar and Hassanzadeh, Pedram and Kashinath, Karthik and Anandkumar, Animashree. FourCastNet : A Global Data-driven High-resolution Weather Model using Ada...

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Schreck, J. S. and Chapman, W. E. and Becker, C. and Gagne, D. J. and Kimpara, D. and Cherukuru, N. and Berner, J. and Mayer, K. J. and Sobhani, N. , title =. arXiv:2512.18815 , year =

    work page arXiv

  20. [20]

    and Coauthors , title =

    Buizza, R. and Coauthors , title =. Bull. Amer. Meteor. Soc. , volume =

  21. [21]

    2023 , journal =

    ECMWF , title =. 2023 , journal =

    2023

  22. [22]

    and Bell, B

    Hersbach, H. and Bell, B. and Berrisford, P. and Hirahara, S. and Hor. The. Quart. J. Roy. Meteor. Soc. , volume =. 2020 , doi =

    2020

  23. [23]

    Frank, N. L. , title =. Mon. Wea. Rev. , volume =. 1970 , doi =

    1970

  24. [24]

    Gray, W. M. , title =. Mon. Wea. Rev. , volume =. 1968 , doi =

    1968

  25. [25]

    Landsea, C. W. , title =. Climatic Change , volume =. 1998 , doi =

    1998

  26. [26]

    Halperin, D. J. and Fuelberg, H. E. and Hart, R. E. and Cossuth, J. H. and Sura, P. and Pasch, R. J. , title =. Wea. Forecasting , volume =. 2013 , doi =

    2013

  27. [27]

    Hart, R. E. , title =. Mon. Wea. Rev. , volume =. 2003 , doi =

    2003

  28. [28]

    Titley, D. W. and Elsberry, R. L. , title =. Mon. Wea. Rev. , volume =. 2000 , doi =

    2000

  29. [29]

    Knaff, J. A. and Zehr, R. M. , title =. Wea. Forecasting , volume =. 2007 , doi =

    2007

  30. [30]

    Lorenz, E. N. , title =. Tellus , volume =. 1969 , doi =

    1969

  31. [31]

    and Marinescu, P

    Heikenfeld, M. and Marinescu, P. J. and Christensen, M. and Watson-Parris, D. and Senf, F. and van den Heever, S. C. and Stier, P. , TITLE =. Geoscientific Model Development , VOLUME =. 2019 , NUMBER =

    2019

  32. [32]

    and Emanuel, K

    Bister, M. and Emanuel, K. A. , title =. Meteor.\ Atmos.\ Phys. , volume =. 1998 , doi =

    1998

  33. [33]

    Camargo, S. J. and Emanuel, K. A. and Sobel, A. H. , title =. J.\ Climate , volume =. 2007 , doi =

    2007

  34. [34]

    WeatherBench 2: A benchmark for the next generation of data‐driven global weather models

    Rasp, Stephan and Hoyer, Stephan and Merose, Alexander and Langmore, Ian and Battaglia, Peter and Russell, Tyler and Sanchez-Gonzalez, Alvaro and Yang, Vivian and Carver, Rob and Agrawal, Shreya and Chantry, Matthew and Ben Bouallegue, Zied and Dueben, Peter and Bromberg, Carla and Sisk, Jared and Barrington, Luke and Bell, Aaron and Sha, Fei. WeatherBenc...

    work page doi:10.1029/2023ms004019

  35. [35]

    and Haji-Akbari, A

    Hussain, S. and Haji-Akbari, A. , title =. J.\ Chem.\ Phys. , volume =. 2020 , doi =

    2020

  36. [36]

    Lancelin, A

    Lancelin, A. and Wikner, A. and Dubus, L. and. arXiv preprint arXiv:2510.27066 , year =. 2510.27066 , archivePrefix =

    work page arXiv

  37. [37]

    and Chen, Y

    Mu, B. and Chen, Y. and Yuan, S. and Qin, B. and Guo, H. , title =. arXiv preprint arXiv:2512.12545 , year =. 2512.12545 , archivePrefix =

    work page arXiv

  38. [38]

    and Gerber, E

    Finkel, J. and Gerber, E. P. and Abbot, D. S. and Weare, J. , title =. AGU Adv. , volume =. 2023 , doi =

    2023

  39. [39]

    and Fischer, E

    Zhang, Z. and Fischer, E. and Zscheischler, J. and Engelke, S. , title =. Sci.\ Adv. , year =

  40. [40]

    Towards accurate extreme event likelihoods from diffusion model climate emulators

    Manshausen, P. and Brenowitz, N. and Berner, J. and Kashinath, K. and Pritchard, M. , title =. arXiv preprint arXiv:2605.03802 , year =. 2605.03802 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv