Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Huge ensembles of ML hindcasts show tropical cyclones spread super-diffusively, approaching the ballistic limit of travel through background winds.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 06:13 UTC pith:7AUBZ6YF

load-bearing objection The HENS ensemble is a valuable resource and the spread-error analysis is solid, but the central claim of reproducing anomalous TC diffusion is not supported because Eq. 3 measures absolute displacement, making e≈2 the null expectation for steadily translating storms. the 4 major comments →

arxiv 2607.21954 v1 pith:7AUBZ6YF submitted 2026-07-24 physics.ao-ph

Anomalous Diffusion of Tropical Cyclones Observed in Huge Ensembles of Hindcasts

Abdoul R. Zeba , William D. Collins , Ankur Mahesh , Boris Bonev , Karthik Kashinath , Thorsten Kurth , Michael S. Pritchard , Shashank Subramanian This is my paper
classification physics.ao-ph
keywords tropical cyclonesanomalous diffusionsuper-diffusionensemble forecastingmachine learning weather emulatormean-square displacementtrack predictabilityhindcast ensemble
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper sets out to show that tropical cyclones obey anomalous super-diffusion, not ordinary Brownian motion, and that a massive ensemble of machine-learning-based hindcasts reproduces the same power laws seen in observed storm tracks. Using thousands of counterfactual replays of each of 33 storms from summer 2023, the authors find that the variance of a storm's position grows as a power law in time with exponent e between 1 and 2, often near the ballistic limit e=2. If correct, this gives a per-storm, ensemble-based estimate of track uncertainty growth that observational records cannot provide, and it implies that storm-track uncertainty cones widen faster than a Brownian model would predict. The claim matters because it would establish machine-learning weather emulators as a tool for measuring the intrinsic diffusive behavior of storms and for bounding forecast lead times.

Core claim

The central claim is that HENS reproduces the anomalous super-diffusion previously inferred for real tropical cyclones: the mean-square displacement of a storm's replicated tracks grows as t^e with e>1 for every storm examined, and individual ensemble members cluster between the Brownian limit e=1 and the ballistic limit e=2. The paper computes two versions of mean-square displacement—one measuring spread about the ensemble-mean track and one measuring each member's squared distance from its own starting point—and finds the fitted exponents exceed 1 for all four representative storms. It further shows that the distribution of per-member exponents across thousands of replays of a single storm

What carries the argument

The key object is the huge ensemble of hindcasts (HENS): 7,424 fifteen-day forecasts initialized daily from observed atmospheric states during summer 2023, generated by a spherical Fourier neural operator (SFNO) weather emulator with bred-vector perturbations to the initial conditions. Storm tracks are extracted by a detection/tracking pipeline with adjusted thresholds and matched to real storms via dynamic time warping plus a stable-matching algorithm. The load-bearing quantity is the diffusion exponent e, the power-law slope of mean-square displacement versus time, which is the variable that carries the argument from ensemble spread to the claim of super-diffusion.

Load-bearing premise

The load-bearing premise is that the mean-square displacement computed from each ensemble member's own starting point (Eq. 3) measures the same diffusive process that earlier observational studies characterized using fluctuations of tracks about the shortest path between initiation and termination; if this equivalence fails, the claim that HENS reproduces the observed anomalous-diffusion power laws collapses, even though the ensemble-spread growth itself is real.

What would settle it

Compute the ensemble-relative mean-square displacement (Eq. 1) for every storm after subtracting the ensemble-mean track, and check whether the fitted exponent e' remains significantly above 1; or repeat the individual-member analysis with displacement measured relative to each storm's own shortest path (great-circle line) instead of its starting point. If e' ≈ 1 once the mean drift is removed, the super-diffusion is an artifact of storm translation rather than of turbulent dispersion.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If e is consistently greater than 1 and sometimes near 2, tropical cyclone track uncertainty grows much faster than Brownian motion predicts, so forecast cones should widen more aggressively with lead time.
  • Because the emulator is cheap enough to generate thousands of members, the diffusion exponent can be estimated separately for each storm, something impossible with observational records that yield one exponent per storm.
  • The resemblance between HENS and observed exponent distributions supports the use of machine-learning emulators for studying tropical cyclone dynamics and predictability.
  • Since the paper treats e as an intrinsic property of the storm and its environment, improving initial conditions alone would not reduce super-diffusive spread; ensemble systems must account for the exponent.

Where Pith is reading between the lines

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

  • Inference: The high individual-member exponents may partly reflect the storm's net translation rather than anomalous dispersion, because Eq. (3) measures displacement from the storm's own starting point; a storm moving at near-constant speed yields e≈2 by construction. The ensemble-relative metric (Eq. 1) is the cleaner test, and the case would be stronger if e' > 1 held robustly across all storms
  • Inference: All results rest on a single summer (2023); extending HENS to multiple seasons would test whether the exponent distribution is a universal feature of tropical cyclones or specific to that unusually warm period.
  • Inference: A direct falsification check is to compute the ensemble-relative exponent e' for all 33 storms and compare it to the per-member e; disagreement would confirm that ballistic advection inflates the per-member exponents.
  • Inference: If the super-diffusive spread is real, operational centers could use ML-ensemble-derived exponents to recalibrate track cone sizes, but this requires verifying that the emulator's ensemble spread is calibrated for lead times beyond 7 days, where the current verification is thin.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper analyzes tropical cyclone (TC) tracks in a 7,424-member ensemble of 15-day SFNO-based hindcasts for summer 2023. It detects and matches emulated TCs to IBTrACS tracks, then computes two mean-squared-displacement (MSD) metrics: an ensemble-relative spread (Eq. 1) and a per-member absolute displacement (Eq. 3). The authors report power-law exponents in both cases, claim that the exponents exceed 1 and approach the ballistic limit 2, and conclude that HENS reproduces the anomalous super-diffusion of TCs previously inferred by Meuel et al. (2012).

Significance. If the central claim were established, the paper would offer a new per-storm, ensemble-based estimate of TC track uncertainty growth and a striking confirmation of anomalous diffusion using ML-based huge ensembles. The resource itself is impressive: 7,424-member hindcasts with open data and code, plus a credible spread-error calibration in Appendix C. However, the central claim is not supported as stated because Eq. (3) is an absolute-displacement MSD that is dominated by uniform TC translation, making exponents near 2 expected by construction. The comparison to Meuel et al. (2012) is also not like-for-like, since that work used fluctuations about the shortest path between genesis and lysis. These are load-bearing issues that affect the title, abstract, and conclusions.

major comments (4)
  1. [§3.2, Eq. (3)–(4); Table A1; Fig. D1] The per-member MSD, MSD_j(t)=⟨(X_j(t+t')−X_j(t'))²⟩, contains no subtraction of the mean track or of a shortest-path reference. For a storm translating at roughly constant velocity v, this quantity is approximately v²t² and yields e≈2 by construction. The near-universal e_j values clustered between 1.8 and 2.0 in Table A1 and the IBTrACS values in Fig. D1 are exactly the signature of uniform drift, not of anomalous diffusion about a preferred path. The statement in §3.2 that 'e_j appreciably exceed 1 ... implying super-diffusive behavior' is therefore not supported by Eq. (3).
  2. [§3.2 and Appendix D, claim of reproducing Meuel et al. (2012)] The paper claims to reproduce the anomalous-diffusion power laws of Meuel et al. (2012). However, Meuel et al. computed fluctuations of hurricane tracks about the shortest path between genesis and lysis, whereas Eq. (3) measures absolute displacement from an arbitrary starting time. Appendix D recomputes exponents on IBTrACS using the same Eq. (3), so it cannot serve as a validation of the Meuel et al. result. The qualitative resemblance of the CDFs in Fig. 4 to Meuel et al. is not evidence of reproduction because the estimators differ.
  3. [§3.2, Eq. (1)–(2) and Fig. 3; Plain Language Summary] The ensemble-relative MSD'_n(t) in Eq. (1) is a legitimate measure of ensemble dispersion, and the spread–error calibration in Appendix C is a creditable contribution. However, e'_n>1 in Fig. 3 describes growth of ensemble variance about the ensemble-mean track, not the diffusion of individual TC tracks through the background flow. The conclusion in the Plain Language Summary that 'tropical cyclones can move ... in a ballistic manner' is drawn from the per-member estimator (Eq. 3), not from Eq. (1). These two different quantities should be cleanly separated, and only the latter can support the per-track diffusion claim.
  4. [§3.2 and Fig. 2/Fig. 4] The matched replicates for a given storm come from 15-day hindcasts initialized on consecutive days (Fig. 2), so the same synoptic evolution contributes to many ensemble members at overlapping lead times. The per-storm distributions of e_j in Fig. 4 and Table A1 treat all matched replicates as independent samples. No account is given of this dependence, so the reported CDFs and percentile spreads may overstate the robustness of the per-storm exponent distributions. An effective-sample-size calculation or subsampling of non-overlapping windows is needed.
minor comments (5)
  1. [Abstract vs. §3.1/Table A1] The abstract states 34 individual TCs, while Section 3.1 and Table A1 list 33 observed TCs. Please reconcile the count.
  2. [Figure 4 caption] The caption says the MSD is 'from its initial position at its start time t0,' but Eq. (3) averages over t' and does not fix the reference at t0. The caption should be reworded to match the estimator actually used.
  3. [Eq. (1)] The double-angle-bracket notation ⟪...⟫ is never defined. Please define it explicitly as the ensemble-mean operator.
  4. [Appendix C] The caveat in Appendix C that the verification uses a single summer and leaves substantial sampling uncertainty should be reflected in the conclusions, which currently make general claims without this qualification.
  5. [General] There are several typographical and formatting slips (e.g., 'V ariable' in Table 1, 'T racks' in the Figure 1 caption, 'Tempest Extremes' vs. 'TempestExtremes'). A careful proofreading pass is needed.

Circularity Check

1 steps flagged

Central 'reproduction' of anomalous TC diffusion is built into Eq. 3's absolute-displacement MSD; near-ballistic exponents are definitional for translating tracks.

specific steps
  1. self definitional [Abstract; Section 3.2, Eqs. (3)-(4)]
    "Anomalous diffusion has been inferred for actual TCs from the fluctuations in their tracks from the shortest paths between the initiation and termination of each cyclone. ... Following (Meuel et al., 2012), the MSD for individual ensemble members indexed by j can be computed by MSDj(t)=⟨(X_j(t+t′)−X_j(t′))²⟩ (3) ... The vast majority of individual ensemble members exhibit super-diffusive behavior with e_j > 1."

    Eq. 3 is an absolute-displacement MSD about an arbitrary start time, not the shortest-path fluctuation measure the abstract identifies with the prior observational result. For a track translating at roughly constant velocity, X_j(t+t')−X_j(t') ≈ v t, so Eq. 3 gives v²t² and the fitted exponent e_j in Eq. 4 is ≈2 by construction. The 'super-diffusive' and 'ballistic limit' statements are therefore mathematical consequences of the estimator for any persistently moving object; they do not reproduce Meuel et al.'s fluctuation-based exponents, and the like-for-like comparison collapses.

full rationale

The paper's headline physical claim—that HENS reproduces the anomalous super-diffusion of individual TC tracks found by Meuel et al.—reduces to the absolute-displacement MSD in Eq. 3. A track with persistent velocity automatically yields e≈2, so the near-ballistic exponents in Figure 4 and Table A1 are largely definitional rather than a discovery about TC diffusion. The Abstract itself states that the prior observational inference used fluctuations about the shortest path, so Eq. 3 is not the same estimator; Appendix D repeats the same estimator on IBTrACS and thus cannot serve as an independent confirmation. The ensemble-spread metric (Eq. 1) and the spread–error calibration (Appendix C) are legitimate forecast-uncertainty diagnostics, and the citations to the authors' HENS/SFNO papers are provenance for the dataset rather than a self-citation chain that forces the result. However, because the central 'reproduction' claim is compromised by construction, a partial circularity score of 6 is appropriate.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The analysis rests on standard atmospheric-science assumptions (reanalysis fidelity, best-track quality) and on domain-specific choices that are not independently grounded: the power-law form for MSD, the transfer of detection thresholds from ERA5 to an ML emulator, the equivalence of absolute-displacement MSD to the prior shortest-path fluctuation measure, and the effective independence of overlapping ensemble members. No new physical entities are introduced. The fitted detection thresholds and the 723 km filter are the main free parameters that shape the data entering the exponent calculation.

free parameters (3)
  • TempestExtremes detection thresholds = ≥6 grid points, latitude 5°S–36.25°N, wind speed >10 m/s, MSL pressure increase ≥170 Pa over 5.5° GCD, (z300-z500) decre
    These thresholds were iteratively adjusted to maximize hit rate and minimize false alarms against IBTrACS for summer 2023 (Section 2.5, Table 2). They are fitted to the same season used in the main analysis, so they encode the target observations into the detection pipeline.
  • Initial-distance matching filter = 723 km
    Section 2.6, step 3: matches with initial distance less than 723 km are kept. This threshold was derived from lead-time-0 distances on ERA5 and directly selects which HENS storm tracks enter the exponent statistics.
  • Per-storm diffusion exponents e_j and e'_n = Table A1: medians range 1.13–1.99; e' fitted per storm in Figure 3
    The exponents are fit by least squares to the MSD data (Eqs. 2 and 4). They are the reported results, not external inputs, but the central claim depends on their values and they carry no error bars.
axioms (6)
  • domain assumption The HENS/SFNO ensemble members provide realistic counterfactual tropical cyclone tracks.
    The whole analysis assumes the ML emulator's ensemble spread represents plausible TC behavior. The spread-error ratio in Appendix C is the only validation, and it is based on a single summer with wide confidence intervals.
  • domain assumption The mean-squared displacement of TC positions follows a power law in time.
    Eqs. (2) and (4) posit MSD ∝ t^e without derivation; this is a modeling hypothesis imported from prior diffusion studies (Meuel et al., 2012), not derived from TC dynamics.
  • domain assumption The variance of ensemble positions (Eq. 1) and the single-track MSD (Eq. 3) measure the same diffusive process as the fluctuation-from-shortest-path estimator of prior observational work.
    Section 3.2 compares HENS exponents to prior observational exponents, but Eq. 3 computes absolute displacement from the storm origin rather than fluctuations about the shortest path; this equivalence is assumed, not established.
  • domain assumption TempestExtremes thresholds calibrated on ERA5 transfer to HENS tracks.
    Section 2.5 asserts that because HENS initial conditions approximate ERA5, the detection framework calibrated on ERA5 is directly applicable to HENS output. This is reasonable but unverified for later lead times.
  • domain assumption Matched HENS ensemble members are approximately independent counterfactual recreations.
    Section 3.2 and Figure 2 treat thousands of matched members as independent replicates, but members come from overlapping 15-day hindcasts initialized on successive days, so many share forecast period and model state; the effective sample size is smaller than the raw count.
  • domain assumption IBTrACS best-track positions and ERA5 reanalyses are adequate ground truth for TC tracks.
    IBTrACS aggregates agency best tracks and ERA5 is used as a proxy for observations; both have known uncertainties and irregular sampling (e.g., four storms excluded in Appendix D), but the paper treats them as reference truth.

pith-pipeline@v1.3.0-alltime-deepseek · 16900 in / 11899 out tokens · 120900 ms · 2026-08-01T06:13:15.076914+00:00 · methodology

0 comments
read the original abstract

We examine whether tropical cyclones (TCs) obey ordinary Brownian or anomalous diffusion using a huge ensemble (HENS) of hindcasts for summer 2023. Anomalous diffusion has been inferred for actual TCs from the fluctuations in their tracks from the shortest paths between the initiation and termination of each cyclone. We reproduce the same anomalous diffusion power laws connecting spatial position and time using HENS. In addition, we show that the variance in the position of a single TC across HENS since initiation follows a scaling law with time that, in some cases, corresponds to ballistic motion of the TC through the background atmospheric flow. This determination was enabled by the exceptional statistics determined from thousands of plausible yet counterfactual recreations of 34 individual TCs. HENS consists of 7424 15-day hindcasts initiated from observed atmospheric conditions each day from June 1, 2023 to August 31, 2023 using the ECMWF ERA5 meteorological reanalysis. The hindcasts were generated using NVIDIA's Spherical Fourier Neural Operator (SFNO) machine-learning-based weather and climate emulator. We identify tropical cyclones in HENS using a variant of the Tempest Extremes detection and tracking frameworks for TCs with adjustments to the disposable parameters to minimize the numbers of false positives and negatives relative to the International Best Track Archive for Climate Stewardship (IBTrACS) records for TCs observed in summer 2023. We conclude with the implications of our findings for the predictability of TC tracks and landfall locations on lead times of days to weeks.

Figures

Figures reproduced from arXiv: 2607.21954 by Abdoul R. Zeba, Ankur Mahesh, Boris Bonev, Karthik Kashinath, Michael S. Pritchard, Shashank Subramanian, Thorsten Kurth, William D. Collins.

Figure 1
Figure 1. Figure 1: Tracks of observed and named tropical cyclones initiated between 1 June and 31 August, 2023. Tracks are color-coded by the Saffir-Simpson scale (Saffir, 1973; Simpson, 1974). The inset depicts 100 of the 7424 counterfactual tracks for Hurricane Dora generated with HENS. These challenges motivate the application of entirely new methodological approaches – such as those based on artificial intelligence and m… view at source ↗
Figure 2
Figure 2. Figure 2: The numbers of matched replicates for each of the four representative TCs (§3.1) versus the initialization dates for the 15-day HENS hindcasts from summer of 2023. The ensembles of 15-day hindcasts are identified by the dates shown on the x-axis when each is initialized. The numbers are plotted between the formation and dissipation of the corresponding representative TC. The maximum possible number of matc… view at source ↗
Figure 3
Figure 3. Figure 3: For several of these TCs, in particular Biparjoy, a power law in time is indeed [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: Log-log plots of the variance in centroid positions of the HENS TCs versus time t since time of initialization t0. Each red point represents a 6-hourly calculation of MSD′ n (t) (Eq. 1). Green dashed lines show least-squares fits to Eq. 2. Solid blue lines depict Eq. 2 with e ′ = 1. The x axis refers to the log10 of the number of time steps, or 6-hour intervals, in the ML model rollout. –12– [PITH_FULL_IM… view at source ↗
Figure 4
Figure 4. Figure 4: Histograms (labeled “H”) and cumulative probability distributions (labeled “C”) of the track diffusion exponent e for each of the selected TCs (§3.1). The approximate power-law growth with time of the mean-square displacement of each TC from its initial position at its start time t0 (Eq. 3) is quantified by e (Eq. 4). Note the different x axis scale used for each TC. –13– [PITH_FULL_IMAGE:figures/full_fig… view at source ↗

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

45 extracted references · 21 canonical work pages · 1 internal anchor

  1. [1]

    Bernido and M

    Christopher C. Bernido and M. Victoria Carpio-Bernido and Matthew G.O. Escobido , keywords =. Physics Letters A , volume = 378, number = 30, pages =. doi:https://doi.org/10.1016/j.physleta.2014.06.003 , url =

    work page doi:10.1016/j.physleta.2014.06.003 2014

  2. [2]

    and Carpio-Bernido, M.V

    Bernido, C.C. and Carpio-Bernido, M.V. and Escobido, M.G.O. , year = 2015, month = jan, pages =. doi:10.1016/j.physleta.2014.10.043 , number = 3, journal =

    work page doi:10.1016/j.physleta.2014.10.043 2015

  3. [3]

    and Fraedrich, K

    Blender, R. and Fraedrich, K. and Lunkeit, F. , year = 1997, month = apr, pages =. doi:10.1002/qj.49712353910 , number = 539, journal =

    work page doi:10.1002/qj.49712353910 1997

  4. [4]

    doi:10.48550/ARXIV.2306.03838 , url =

    Bonev, Boris and Kurth, Thorsten and Hundt, Christian and Pathak, Jaideep and Baust, Maximilian and Kashinath, Karthik and Anandkumar, Anima , keywords =. doi:10.48550/ARXIV.2306.03838 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.03838

  5. [5]

    doi:10.1038/s41612-024-00568-7 , number = 1, journal =

    doi:10.1038/s41612-024-00568-7

  6. [6]

    IFS Documentation CY41R2 , number = 5, omonth = 2016, publisher =

    ECMWF , title =. IFS Documentation CY41R2 , number = 5, omonth = 2016, publisher =. doi:10.21957/4btqaug2x , language =

    work page doi:10.21957/4btqaug2x 2016

  7. [7]

    doi:10.1038/s41586-024-07512-y , number = 8019, journal =

    Esper, Jan and Torbenson, Max and B\". doi:10.1038/s41586-024-07512-y , number = 8019, journal =

    work page doi:10.1038/s41586-024-07512-y

  8. [8]

    doi:10.1038/s41467-023-40112-4 , number = 1, journal =

    doi:10.1038/s41467-023-40112-4

  9. [9]

    and Smith, Chris and Walsh, Tristram and Lamb, William F

    Forster, Piers M. and Smith, Chris and Walsh, Tristram and Lamb, William F. and Lamboll, Robin and Hall, Bradley and Hauser, Mathias and Ribes, Aurélien and Rosen, Debbie and Gillett, Nathan P. and Palmer, Matthew D. and Rogelj, Joeri and von Schuckmann, Karina and Trewin, Blair and Allen, Myles and Andrew, Robbie and Betts, Richard A. and Borger, Alex an...

    2024

  10. [10]

    and Abaza, M

    Fortin, V. and Abaza, M. and Anctil, F. and Turcotte, R. , year = 2014, month =. doi:10.1175/jhm-d-14-0008.1 , number = 4, journal =

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

  11. [11]

    Knapp and Carl J

    Jennifer Gahtana and Kenneth R. Knapp and Carl J. Schreck III and Howard J. Diamond and James P. Kossin and Michael C. Kruk , title =

  12. [12]

    doi:10.5061/DRYAD.2RBNZS80N , url =

    Mahesh, Ankur and Collins, William and Bonev, Boris and Brenowitz, Noah and Cohen, Yair and Harrington, Peter and Kashinath, Karthik and Kurth, Thorsten and North, Joshua and O'Brien, Travis and Pritchard, Mike and Pruitt, David and Risser, Mark and Subramanian, Shashank and Willard, Jared , keywords =. doi:10.5061/DRYAD.2RBNZS80N , url =

    work page doi:10.5061/dryad.2rbnzs80n

  13. [13]

    Hall and J

    D. Hall and J. Pathak and K. Kashinath and M. Pritchard and P. Messmer and F. Hariri and N. Brenowitz and Y. Cohen and A. Anandkumar and S. Posey , title =

  14. [14]

    and Bates, Gary T

    Hamill, Thomas M. and Bates, Gary T. and Whitaker, Jeffrey S. and Murray, Donald R. and Fiorino, Michael and Galarneau, Thomas J. and Zhu, Yuejian and Lapenta, William , year = 2013, month = Oct, pages =. doi:10.1175/bams-d-12-00014.1 , number = 10, journal =

    work page doi:10.1175/bams-d-12-00014.1 2013

  15. [15]

    and Taylor, Katherine L

    Hegerl, Gabriele C. and Taylor, Katherine L. , year = 2024, pmonth = jul, pages =. doi:10.1038/d41586-024-02057-6 , number = 8019, journal =

    work page doi:10.1038/d41586-024-02057-6 2024

  16. [16]

    doi:10.1002/qj.3803 , number = 730, journal =

    Hersbach, Hans and Bell, Bill and Berrisford, Paul and Hirahara, Shoji and Hor\'. doi:10.1002/qj.3803 , number = 730, journal =

    doi:10.1002/qj.3803

  17. [17]

    and Ghil, Michael , year = 2001, month = aug, pages =

    Huber, Matthew and McWilliams, James C. and Ghil, Michael , year = 2001, month = aug, pages =. doi:10.1175/1520-0469(2001)058<2377:acotdi>2.0.co;2 , number = 16, journal =

    doi:10.1175/1520-0469(2001)058 2001

  18. [18]

    , year = 2024, month = nov, pages =

    Jarvis, Andrew and Forster, Piers M. , year = 2024, month = nov, pages =. doi:10.1038/s41561-024-01580-5 , number = 12, journal =

    work page doi:10.1038/s41561-024-01580-5 2024

  19. [19]

    and Kelley, Douglas I

    Jones, Matthew W. and Kelley, Douglas I. and Burton, Chantelle A. and Di Giuseppe, Francesca and Barbosa, Maria Lucia F. and Brambleby, Esther and Hartley, Andrew J. and Lombardi, Anna and Mataveli, Guilherme and McNorton, Joe R. and Spuler, Fiona R. and Wessel, Jakob B. and Abatzoglou, John T. and Anderson, Liana O. and Andela, Niels and Archibald, Sally...

    2024

  20. [20]

    and Kruk, Michael C

    Knapp, Kenneth R. and Kruk, Michael C. and Levinson, David H. and Diamond, Howard J. and Neumann, Charles J. , year = 2010, omonth = mar, pages =. doi:10.1175/2009bams2755.1 , number = 3, journal =

    doi:10.1175/2009bams2755.1 2010

  21. [21]

    FourCastNet: Accelerating Global High-Resolution Weather Forecasting using Adaptive Fourier Neural Operators

    Kurth, Thorsten and Subramanian, Shashank and Harrington, Peter and Pathak, Jaideep and Mardani, Morteza and Hall, David and Miele, Andrea and Kashinath, Karthik and Anandkumar, Animashree , keywords =. doi:10.48550/ARXIV.2208.05419 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2208.05419

  22. [22]

    and Cangialosi, John P

    Landsea, Christopher W. and Cangialosi, John P. , year = 2018, omonth = nov, pages =. doi:10.1175/bams-d-17-0136.1 , number = 11, journal =

    work page doi:10.1175/bams-d-17-0136.1 2018

  23. [23]

    doi:10.48550/ARXIV.2010.08895 , url =

    Li, Zongyi and Kovachki, Nikola and Azizzadenesheli, Kamyar and Liu, Burigede and Bhattacharya, Kaushik and Stuart, Andrew and Anandkumar, Anima , keywords =. doi:10.48550/ARXIV.2010.08895 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2010.08895 2010

  24. [24]

    and Collins, W

    Mahesh, A. and Collins, W. D. and Bonev, B. and Brenowitz, N. and Cohen, Y. and Elms, J. and Harrington, P. and Kashinath, K. and Kurth, T. and North, J. and O'Brien, T. and Pritchard, M. and Pruitt, D. and Risser, M. and Subramanian, S. and Willard, J. , TITLE =. Geoscientific Model Development , VOLUME = 18, YEAR = 2025, NUMBER = 17, PAGES =

    2025

  25. [25]

    Mahesh, A. and D. Collins, W. and Bonev, B. and Brenowitz, N. and Cohen, Y. and Harrington, P. and Kashinath, K. and Kurth, T. and North, J. and O'Brien, T. A. and Pritchard, M. and Pruitt, D. and Risser, M. and Subramanian, S. and Willard, J. , TITLE =. Geoscientific Model Development , VOLUME = 18, YEAR = 2025, NUMBER = 17, PAGES =

    2025

  26. [26]

    and Bianchette, T

    McCloskey, T. and Bianchette, T. and Liu, K. , title =. American Journal of Climate Change , year = 2013, volume = 2, pages =

    2013

  27. [27]

    doi:10.1038/srep00446 , PDF =

    Meuel, Tinihau and Prado, Ga. doi:10.1038/srep00446 , PDF =

    work page doi:10.1038/srep00446

  28. [28]

    and Tasaka, Y

    Noto, D. and Tasaka, Y. and Yanagisawa, T. and Murai, Y. , year = 2019, pages =. doi:10.1017/jfm.2019.313 , journal =

    work page doi:10.1017/jfm.2019.313 2019

  29. [29]

    doi:10.48550/ARXIV.2202.11214 , url =

    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 , keywords =. doi:10.48550/ARXIV.2202.11214 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2202.11214

  30. [30]

    Jaideep Pathak and Shashank Subramanian and Peter Harrington and Sanjeev Raja and Ashesh Chattopadhyay and Morteza Mardani and Thorsten Kurth and David Hall and Zongyi Li and Kamyar Azizzadenesheli and Pedram Hassanzadeh and Karthik Kashinath and Animashree Anandkumar , year = 2022, eprint =

    2022

  31. [31]

    doi:10.1073/pnas.1712645115 , number = 1, journal =

    Ragone, Francesco and Wouters, Jeroen and Bouchet, Freddy , year = 2017, omonth = dec, pages =. doi:10.1073/pnas.1712645115 , number = 1, journal =

    doi:10.1073/pnas.1712645115 2017

  32. [32]

    Journal of Statistical Physics , publisher =

    Ragone, Francesco and Bouchet, Freddy , year = 2019, omonth = nov, pages =. Journal of Statistical Physics , publisher =. doi:10.1007/s10955-019-02429-7 , number =

    work page doi:10.1007/s10955-019-02429-7 2019

  33. [33]

    doi:10.1098/rspa.1926.0043 , number = 756, journal =

    Lewis Richardson , title =. doi:10.1098/rspa.1926.0043 , number = 756, journal =

    arXiv 1926

  34. [34]

    doi:10.1093/biosci/biad080 , number = 12, journal =

    Ripple, William J and Wolf, Christopher and Gregg, Jillian W and Rockstr\". doi:10.1093/biosci/biad080 , number = 12, journal =

    work page doi:10.1093/biosci/biad080

  35. [35]

    , journal =

    Saffir, Herbert S. , journal =

  36. [36]

    doi:10.1038/d41586-024-00816-z , number = 8004, journal =

    Schmidt, Gavin , year = 2024, omonth = mar, pages =. doi:10.1038/d41586-024-00816-z , number = 8004, journal =

    work page doi:10.1038/d41586-024-00816-z 2024

  37. [37]

    , journal =

    Simpson, Robert H. , journal =

  38. [38]

    doi:10.5061/DRYAD.TTDZ08MBT , url =

    Zeba, Abdoul Rahin and Collins, William and Mahesh, Ankur , keywords =. doi:10.5061/DRYAD.TTDZ08MBT , url =

    work page doi:10.5061/dryad.ttdz08mbt

  39. [39]

    doi:10.5281/zenodo.21158620 , url =

    Zeba, Abdoul Rahin and Collins, William and Mahesh, Ankur , keywords =. doi:10.5281/zenodo.21158620 , url =

    work page doi:10.5281/zenodo.21158620

  40. [40]

    Dunstone and Adam A

    Vikki Thompson and Nick J. Dunstone and Adam A. Scaife and Doug M. Smith and Julia M. Slingo and Simon Brown and Stephen E. Belcher , title =. doi:10.1038/s41467-017-00275-3 , url =

    work page doi:10.1038/s41467-017-00275-3

  41. [41]

    Ullrich, P. A. and Zarzycki, C. M. , TITLE =. Geoscientific Model Development , VOLUME = 10, YEAR = 2017, NUMBER = 3, PAGES =

    2017

  42. [42]

    Ullrich, P. A. and Zarzycki, C. M. and McClenny, E. E. and Pinheiro, M. C. and Stansfield, A. M. and Reed, K. A. , TITLE =. Geoscientific Model Development , VOLUME = 14, YEAR = 2021, NUMBER = 8, PAGES =

    2021

  43. [43]

    doi:10.1175/bams-d-20-0308.1 , number = 2, journal =

    Yu, Hui and Chen, Guomin and Zhou, Cong and Wong, Wai Kin and Yang, Mengqi and Xu, Yinglong and Chen, Peiyan and Wan, Rijin and Hu, Xinrong , year = 2022, omonth = feb, pages =. doi:10.1175/bams-d-20-0308.1 , number = 2, journal =

    work page doi:10.1175/bams-d-20-0308.1 2022

  44. [44]

    Zehnder, J. A. , title =

  45. [45]

    doi:10.1175/bams-d-19-0166.1 , number = 12, journal =

    Zhou, Feifan and Toth, Zoltan , year = 2020, omonth = dec, pages =. doi:10.1175/bams-d-19-0166.1 , number = 12, journal =

    work page doi:10.1175/bams-d-19-0166.1 2020