Pith. sign in

REVIEW 3 major objections 6 minor 22 references

A pure machine-learned atmosphere runs stably for decades when coupled to a full-depth ocean, yet produces muted ENSO and a wrong short-wave response to CO2.

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 · grok-4.5

2026-07-13 16:11 UTC pith:EZCQIVMI

load-bearing objection First multi-decadal free coupling of an untuned pure-ML atmosphere to full-depth NEMO is stable and diagnostic, with muted ENSO from weak Bjerknes feedback as the main scientific result. the 3 major comments →

arxiv 2603.28704 v2 pith:EZCQIVMI submitted 2026-03-30 physics.ao-ph

ACE2-NEMO: Coupling an ML atmospheric emulator to a full-depth dynamical ocean model

Bobby Antonio , Kristian Strommen , Pablo Ortega , Hannah M. Christensen This is my paper
classification physics.ao-ph PACS 92.60.Wc92.10.am07.05.Mh
keywords machine-learned atmosphereACE2NEMOcoupled climate modelENSOBjerknes feedbackhybrid Earth-system modelclimate emulator
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.

Machine-learned atmospheric emulators are highly skilled at weather timescales, but it is unclear whether they have learned the feedbacks needed for multi-decadal climate variability and forced response. This paper couples the ACE2 atmospheric emulator, with no retraining, to the full-depth dynamical ocean model NEMO and integrates the hybrid system for seventy years under both fixed 1950s and historical greenhouse-gas forcing. The coupled model remains stable, produces realistic mean states and realistic spatial patterns of fast tropical air-sea coupling, yet the El Niño-Southern Oscillation appears only as very low-amplitude oscillations because the atmospheric wind response to Pacific sea-surface temperature gradients is too weak. The historical runs also diverge after about 1980 because ACE2 overestimates the decline in downward short-wave radiation. The work therefore supplies a stringent out-of-sample test of whether entirely machine-learned atmospheric components can serve as drop-in replacements in next-generation Earth-system models.

Core claim

The first multi-decadal integrations of an unfine-tuned machine-learned atmosphere (ACE2) interactively coupled to a full-depth dynamical ocean (NEMO) are stable for at least seventy years and produce realistic mean states and fast tropical air-sea coupling patterns, yet generate unrealistically low-amplitude ENSO variability because of weak atmospheric Bjerknes feedback and an incorrect short-wave response to CO2 forcing after ~1980.

What carries the argument

The hybrid ACE2-NEMO coupler: shared-storage OASIS exchange of ocean and ice state every six hours, AirSeaFluxCode bulk formulae for missing momentum and evaporation fluxes, and CORE bulk formulae that replace ACE2’s own sea-ice fluxes to suppress drifts in sea-ice volume and sea-surface height.

Load-bearing premise

That replacing ACE2’s own sea-ice fluxes with fixed-coefficient bulk formulae, and starting from ocean initial conditions warmer than the reanalysis the emulator was trained on, does not itself destroy the atmospheric feedbacks the paper diagnoses.

What would settle it

A parallel seventy-year coupled run that uses ACE2’s native sea-ice fluxes (or a version of ACE2 trained with sea-ice temperature) and ocean initial conditions closer to ERA5; if realistic ENSO amplitude and post-1980 short-wave trends then appear, the present diagnosis of weak atmospheric feedback is an artifact of the coupling fixes.

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

If this is right

  • Atmospheric emulators trained only on reanalysis or prescribed-SST data can fail to generate the wind-SST feedbacks required for realistic ENSO even when mean states look good.
  • Hybrid ML-atmosphere / dynamical-ocean models can already be run at roughly one-third the energy cost of a full dynamical climate model, opening the door to larger ensembles and cheaper ocean spin-up.
  • The same coupling framework can be used as a diagnostic test-bed for other atmospheric emulators before they are trusted in free-running climate projections.
  • Separating forced-response training from internal-variability training may be necessary if emulators are to avoid learning spurious associations between CO2 and ENSO.

Where Pith is reading between the lines

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

  • The muted ENSO spectrum is more diagnostic of missing atmospheric stochasticity or of training-data length than of any fundamental limit of machine learning, because the spatial coupling patterns themselves remain realistic.
  • If later ACE versions that emit momentum fluxes natively still show the same short-wave bias, the problem is intrinsic to the learned radiative physics rather than an artifact of the flux calculator.
  • The framework immediately suggests a practical path to accelerated coupled spin-up: use the cheap hybrid model to bring the deep ocean close to equilibrium, then switch to a full dynamical atmosphere for the production run.

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

3 major / 6 minor

Summary. The manuscript presents ACE2-NEMO, a hybrid system in which the ACE2 machine-learned atmospheric emulator is interactively coupled, without fine-tuning, to the full-depth NEMO ocean (ORCA1, 75 levels). It reports 70-year fixed-1950s control and 1950–2020 historical ensembles and compares them with EC-Earth3/3P runs that share the same ocean component. The central claims are carefully scoped: the hybrid system is multi-decadal stable with broadly realistic mean states and surface fluxes; tropical Pacific precipitation, SLP and wind regressions onto Niño 3.4 resemble EC-Earth3P and ERA5; yet Niño 3.4 amplitude and spectral power are unrealistically weak, which the authors attribute to a weak Bjerknes wind response to the basin SST gradient (Fig. 4) and related feedback differences; and the historical run tracks EC-Earth3P surface warming only until ~1980, after which global SST cools, linked to an excessive decline in ocean downward short-wave radiation (Fig. 6d).

Significance. If the reported behaviour holds, this is a genuine first multi-decadal free coupling of an untuned pure ML atmosphere to a full-depth dynamical ocean, and a useful stress test of whether weather-trained atmospheric emulators have learned the feedbacks needed for coupled variability and forced response. The work is concrete: shared-storage OASIS + AirSeaFluxCode infrastructure is documented, code and processed data are released, and diagnostics are compared against independent EC-Earth3P and ERA5 rather than against a fitted target. That combination of novelty, reproducibility, and falsifiable coupled diagnostics makes the paper a valuable contribution to hybrid Earth-system modelling, even though the ENSO and post-1980 SW results are negative findings.

major comments (3)
  1. [Sec. 4.1.4, Eqs. (2)–(6); S1, S5] Sec. 4.1.4, Eqs. (2)–(6) and Supplementary S5/S1: The load-bearing ENSO and forced-response claims rest on atmospheric feedbacks diagnosed from free coupled runs, yet ACE2’s own ice fluxes are replaced by CORE bulk formulae with fixed C_ice = 1.4×10^{-3}, coastal heat-flux masking, and nn_fwb=1 freshwater conservation, and the ocean IC is ~0.5–1 K warmer than ERA5. The paper shows that these choices remove SSH/SIvol drift (Fig. S5) and that mean open-ocean fluxes look reasonable (Fig. S2), but it does not quantify whether the altered polar energy/freshwater balance or the initial tropical spin-up (Fig. S1) shift the tropical mean state enough to weaken the Bjerknes slope in Fig. 4a. A short sensitivity (e.g. alternate C_ice, or a nudged/closer-to-ERA5 IC) or at least a residual open-ocean heat-budget table would make the attribution of muted ENSO to ACE2’s learned wind feedback more secu
  2. [Sec. 2.2; Fig. 3a; Fig. 4] Sec. 2.2 and Fig. 4 (left column): The explanation that weak wind response to the east–west SST gradient is the primary cause of low-amplitude Niño 3.4 is consistent with the maps shown, but the analysis stops at regression slopes. A standard Bjerknes feedback index (or wind–SST lagged regressions and a simple recharge-oscillator growth-rate estimate) would test whether the diagnosed wind deficit is large enough, relative to the heat-flux damping in the right column of Fig. 4, to account for the near-red-noise spectrum in Fig. 3a. Without that, the causal link remains plausible but not fully closed.
  3. [Sec. 2.3; Fig. 6] Sec. 2.3 and Fig. 6d: The post-1980 SST cooling is attributed to an excessive decline in net/downward short-wave over ice-free ocean. ACE2 has no aerosol forcing and does not output cloud fraction or optical depth, so the physical pathway (implicit cloud feedback vs missing aerosol vs other) cannot be diagnosed from the presented fields. The manuscript already notes aerosols as a possible factor; this needs to be elevated to a clear limitation on the forced-response claim, and any available TOA or clear-sky SW diagnostics should be added if they exist in the ACE2 output stream.
minor comments (6)
  1. [Abstract; Introduction] Abstract and Introduction claim “to our knowledge, the first multi-decadal integrations of a machine-learned atmosphere interacting with a full-depth dynamical ocean.” Duncan et al. (2025) is correctly distinguished later as ML–ML with multi-day ocean steps; a single clarifying phrase in the abstract would prevent misreading.
  2. [Sec. 2.2; Fig. 3a] Fig. 3a: Welch spectra use concatenated ensemble members and segment length 256. State the effective degrees of freedom and whether pre-whitening or linear detrend was applied, especially for ACE2-NEMO-hist where the long-period peak is linked to the secular SST cooling.
  3. [Fig. 4] Fig. 4 colour scales and units (×10^7 m^2 s^{-1} K^{-1} for U10 vs d(SST)/dx) are hard to compare across panels; a common colour bar range for control/hist/ECE3P/ERA5 within each column would help.
  4. [Sec. 4.1.3] Sec. 4.1.3: Solid precipitation is set equal to total precipitation for T2m ≤ 273 K over sea. Cite the ERA5-based threshold more precisely and note sensitivity if any was tested.
  5. [Sec. 4.1.5] Energy proxy in Sec. 4.1.5 (kJ per simulated day from node-seconds and TDP) is useful but approximate; label it clearly as an order-of-magnitude proxy rather than a measured energy cost.
  6. [Throughout; Fig. 2; Supplement] Typos / notation: “Ni˜ no” encoding is inconsistent in places; “T otal HF” in Fig. 2d; “Sec. S6” is referenced for SLP/wind regressions but the supplement heading is S6—ensure cross-references match the compiled PDF.

Circularity Check

0 steps flagged

No circularity: empirical free-running hybrid simulations compared to independent EC-Earth/ERA5 benchmarks; no quantity is defined from a fit that is then re-presented as a prediction.

full rationale

The paper’s central claims are scoped empirical outcomes of multi-decadal free integrations of an untuned ACE2 atmosphere coupled to full-depth NEMO (stability, mean-state realism, fast tropical air-sea regression patterns, muted ENSO amplitude attributed to weak Bjerknes wind feedback, and post-1980 short-wave cooling). These are obtained by running the hybrid model under fixed-1950s and historical CO2 forcings and comparing diagnostics (Niño-3.4 spectra, pointwise regressions of precipitation/SLP/U10/heat fluxes, global temperature and flux time series) against independent EC-Earth3P/EC-Earth3 historical and control runs that share the same ocean component plus ERA5. No parameter is fitted to a subset of the target diagnostics and then re-used as a “prediction”; the CORE bulk-formulae replacement of ACE2 ice fluxes and the warmer initial ocean state are openly described engineering choices whose effects are diagnosed (and shown to be remote from the tropical-Pacific feedbacks that drive the ENSO conclusion). Self-citations supply the ACE2 checkpoint, NEMO configuration files, and ocean restarts; they do not import a uniqueness theorem or ansatz that forces the reported behaviour. The derivation chain is therefore ordinary numerical experimentation against external benchmarks and contains no self-definitional, fitted-input-as-prediction, or load-bearing self-citation circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claims rest on standard climate-modelling assumptions plus a small set of engineering choices needed to keep the hybrid system stable. No new physical entities are postulated; free parameters are the conventional bulk-transfer coefficients and a few ad-hoc numerical fixes.

free parameters (3)
  • C_ice heat/momentum transfer coefficient = 1.4e-3
    Fixed at 1.4e-3 in the CORE bulk formulae used over ice (Eqs. 2–4); standard value but chosen by hand rather than learned or re-tuned.
  • Solid-precipitation temperature threshold = 273 K
    Heuristic: all precipitation is solid when T2m ≤ 273 K over sea points; chosen from ERA5 inspection, not optimised.
  • Coastal heat-flux mask = 1 grid cell, 10 % ice
    Heat fluxes zeroed within one grid cell of coast when ice fraction > 10 % to prevent SSH crashes; ad-hoc numerical fix.
axioms (4)
  • domain assumption ACE2 fluxes (except momentum, evaporation and solid precipitation) remain physically usable when the ocean state is out-of-sample relative to ERA5 training.
    Stated as the core experimental premise in Sec. 1 and 4.1; never proved, only tested a posteriori by stability and mean-state diagnostics.
  • ad hoc to paper Bulk CORE formulae correctly replace ACE2’s own ice fluxes without destroying the atmospheric feedbacks under study.
    Introduced in Sec. 4.1.4 after large drifts were observed; the paper treats the replacement as a technical necessity rather than a validated physical equivalence.
  • domain assumption Three lagged ensemble members suffice to characterise multi-decadal variability and forced response.
    Standard small-ensemble practice in climate modelling; used throughout Sec. 2 without formal power analysis.
  • domain assumption NEMO ORCA1 + 75 levels with EC-Earth3P namelists is an adequate dynamical ocean for the comparison.
    Inherited from the PRIMAVERA/EC-Earth configuration (Sec. 4.1.2).
invented entities (1)
  • ACE2-NEMO hybrid coupler (shared-storage OASIS + AirSeaFluxCode router) independent evidence
    purpose: Bridge GPU-based ACE2 and multi-CPU NEMO that cannot share an MPI communicator on conventional HPC.
    Engineering construct required for the experiment; no claim of new physics.

pith-pipeline@v1.1.0-grok45 · 31313 in / 3079 out tokens · 32300 ms · 2026-07-13T16:11:20.656886+00:00 · methodology

0 comments
read the original abstract

Understanding how fast atmospheric variability shapes slow climate variability and sensitivity remains a central challenge in Earth-system science. Recent advances in machine-learned (ML) atmospheric models have demonstrated remarkable skill on weather timescales, but their emergent behaviour in a fully coupled climate system remains largely unexplored. We present early results from a new hybrid modelling framework, in which the ACE2 ML atmospheric emulator is interactively coupled to the NEMO ocean model. We report on a set of 70-year coupled simulations (1950-2020 historical forcing and fixed-1950s control). These experiments represent, to our knowledge, the first multi-decadal integrations of a machine-learned atmosphere interacting with a full-depth dynamical ocean. Several historical and fixed-1950s control simulations from the fully dynamic global coupled climate model EC-Earth, which has the same ocean component used in ACE2-NEMO, are also considered for comparison. We assess the behaviour of the coupled system, with particular focus on low-frequency tropical variability and the climate response to greenhouse-gas forcing. Analysis of potentially emergent El Ni\~{n}o-like variability reveals realistic fast timescale air-sea coupling in the tropical Pacific, but the temporal variability is unrealistic, with very low amplitude oscillations; this appears to be due to weak atmospheric feedback in the tropical Pacific. The response to CO2 forcing shows initial agreement with EC-Earth3P, but deviates due to reduced downward short-wave radiation in ACE2. These results provide a unique test of physical realism for atmospheric emulators, and evaluate the possible role of entirely machine-learned components in next-generation Earth system models.

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

22 extracted references · 3 canonical work pages

  1. [1]

    N. D. Brenowitz, B. Henn, S. Clark, A. Kwa, J. McGibbon, W. A. Perkins, O. Watt-Meyer, and C. S. Bretherton. Machine learning climate model dynamics: Offline versus online performance. In NeurIPS 2020 Workshop on Tackling Climate Change with Machine Learning,

    2020

  2. [2]

    URLhttps: //www.climatechange.ai/papers/neurips2020/50. N. D. Brenowitz, T. Ge, A. Subramaniam, P. Manshausen, A. Gupta, D. M. Hall, M. Mardani, A. Vahdat, K. Kashinath, and M. S. Pritchard. Climate in a bottle: Towards a generative foundation model for the kilometer-scale global atmosphere.arXiv:2505.06474,

    Pith/arXiv arXiv

  3. [3]

    arXiv:2504.06007 [physics]. H. M. Christensen, J. Berner, D. R. Coleman, and T. Palmer. Stochastic parameterization and El Ni˜ no–southern oscillation.Journal of Climate, 30(1):17–38,

    Pith/arXiv arXiv

  4. [4]

    D¨ oscher, M

    R. D¨ oscher, M. Acosta, A. Alessandri, P. Anthoni, A. Arneth, T. Arsouze, T. Bergmann, R. Bernadello, S. Bousetta, L.-P. Caron, et al. The EC-Earth3 Earth system model for the Coupled Model Inter- comparison Project 6.Geoscientific Model Development Discussions, 2021:1–90,

    2021

  5. [5]

    EC-Earth-Consortium

    arXiv:2509.12490 [physics]. EC-Earth-Consortium. EC-Earth-Consortium EC-Earth3 model output prepared for CMIP6 CMIP historical, Version 20201215,

    arXiv

  6. [6]

    URLhttps://doi.org/10.22033/ESGF/CMIP6.4700. A. El Aouni, Q. Gaudel, C. Regnier, S. Van Gennip, O. Le Galloudec, M. Drevillon, Y. Drillet, and J.-M. Lellouche. Glonet: Mercator’s end-to-end neural global ocean forecasting system.Journal of Geophysical Research: Machine Learning and Computation, 2(3):e2025JH000686,

    work page doi:10.22033/esgf/cmip6.4700

  7. [7]

    H. Guan, T. Arcomano, A. Chattopadhyay, and R. Maulik. Lucie-3d: A three-dimensional climate emulator for forced responses.arXiv preprint arXiv:2509.02061,

    Pith/arXiv arXiv

  8. [8]

    doi: 10.5194/gmd-13-3507-2020. K. Hasselmann. Stochastic climate models Part I. Theory.Tellus, 28(6):473–485,

    work page doi:10.5194/gmd-13-3507-2020 2020

  9. [9]

    doi: 10.1111/ j.2153-3490.1976.tb00696.x. H. Hersbach, B. Bell, P. Berrisford, S. Hirahara, A. Hor´ anyi, J. Mu˜ noz-Sabater, J. Nicolas, C. Peubey, R. Radu, D. Schepers, et al. The ERA5 global reanalysis.Quarterly journal of the royal meteoro- logical society, 146(730):1999–2049,

    arXiv 1976

  10. [10]

    15 R. M. Hoesly, S. J. Smith, L. Feng, Z. Klimont, G. Janssens-Maenhout, T. Pitkanen, J. J. Seibert, L. Vu, R. J. Andres, R. M. Bolt, et al. Historical (1750–2014) anthropogenic emissions of reactive gases and aerosols from the Community Emissions Data System (CEDS).Geoscientific Model Development, 11(1):369–408,

    2014

  11. [11]

    Large and S

    W. Large and S. Yeager. Diurnal to decadal global forcing for ocean and sea-ice models: the data sets and flux climatologies (no. ncar/tn-460+ str).University Corporation for Atmospheric Research: NCAR, 2044:111,

    2044

  12. [12]

    Z. Li, R. Ding, J. Mao, Z. Ren, and J. Li. On the westward shift and strengthening of the atmosphere- to-ocean Bjerknes feedback in the tropical pacific since 2000.Geophysical Research Letters, 51(12): e2024GL109874,

    2000

  13. [13]

    Publisher: Zenodo

    doi: 10.5281/zenodo.8167700. Publisher: Zenodo. S. Manabe and R. T. Wetherald. Thermal equilibrium of the atmosphere with a given distribution of relative humidity.Journal of the Atmospheric Sciences, 24(3):241–259,

    work page doi:10.5281/zenodo.8167700

  14. [14]

    B. I. Moat, D. Smeed, D. Rayner, W. E. Johns, R. H. Smith, D. L. Volkov, S. Elipot, T. Petit, J. B. Kajtar, M. O. Baringer, et al. Atlantic meridional overturning circulation observed by the RAPID-MOCHA-WBTS (RAPID-Meridional Overturning Circulation and Heatflux Array-Western Boundary Time Series) array at 26N from 2004 to 2024 (v2024. 1a)

    2004

  15. [15]

    A100 80GB PCIe GPU Product Brief, date accessed 26-02-2026,

    NVIDIA. A100 80GB PCIe GPU Product Brief, date accessed 26-02-2026,

    2026

  16. [16]

    arXiv:2202.11214 [physics]. I. Price, A. Sanchez-Gonzalez, F. Alet, T. R. Andersson, A. El-Kadi, D. Masters, T. Ewalds, J. Stott, S. Mohamed, P. Battaglia, et al. Probabilistic weather forecasting with machine learning.Nature, 637(8044):84–90,

    Pith/arXiv arXiv

  17. [17]

    arXiv:2402.04342 [physics]. A. Timmermann, S.-I. An, J.-S. Kug, F.-F. Jin, W. Cai, A. Capotondi, K. M. Cobb, M. Lengaigne, M. J. McPhaden, M. F. Stuecker, et al. El Ni˜ no–southern oscillation complexity.Nature, 559(7715): 535–545,

    Pith/arXiv arXiv

  18. [18]

    arXiv:2402.02995 [physics]. O. Watt-Meyer, B. Henn, J. McGibbon, S. K. Clark, A. Kwa, W. A. Perkins, E. Wu, L. Harris, and C. S. Bretherton. ACE2: accurately learning subseasonal to decadal atmospheric variability and forced responses.npj Climate and Atmospheric Science, 8(1):205,

    Pith/arXiv arXiv

  19. [19]

    arXiv:2510.02415 [physics]. G. Zhang, M. Rao, J. Yuval, and M. Zhao. Advancing seasonal prediction of tropical cyclone activity with a hybrid AI-physics climate model.Environmental Research Letters, 20(9):094031,

    Pith/arXiv arXiv

  20. [20]

    Supplementary materials S1 Sea surface temperature behaviour In this section, we investigate the changes in sea surface temperature during the control and historical runs

    17 1951 1952 1953 1954 1955 1956 Year 291 292 293 294Sea surface temperature [K] (a) 1951 1952 1953 1954 1955 1956 Year 110.0 107.5 105.0 102.5 100.0 97.5 Latent heat flux [W/m2] (b) 1951 1952 1953 1954 1955 1956 Year 2.7 2.8 2.9 3.0 3.1Precipitation [mm/day] (c) 1951 1952 1953 1954 1955 1956 Year 24 25 26 27Total water path [mm] (d) ACE2-NEMO-control ECE...

    1951

  21. [21]

    These are calculated by fitting a linear regression model for each grid cell separately, for the first ensemble member of ACE2-NEMO-control only. The drift in SST (panels a and b) show similar magnitudes between ACE2-NEMO-control and ECE3P-control, but with different patterns; for ACE2-NEMO-control the pattern is similar to the patterns of ENSO heating an...

    2018

  22. [22]

    The 100-500m band shows a consistent increase throughout, which no changes in trend around 1980 that might suggest significant amounts of heat being transferred to the deeper ocean. There is a slight change in trend in the 500-1000m band around the year 2000, suggesting that there is increased heat being passed to the deeper ocean and so transfer of heat ...

    1980