Climate Model Tuning with Online Synchronization-Based Parameter Estimation

Frank M. Selten; Jordan Seneca; Suzanne Bintanja

arxiv: 2510.06180 · v2 · submitted 2025-10-07 · 🌊 nlin.CD · cs.LG· physics.ao-ph

Climate Model Tuning with Online Synchronization-Based Parameter Estimation

Jordan Seneca , Suzanne Bintanja , Frank M. Selten This is my paper

Pith reviewed 2026-05-18 09:16 UTC · model grok-4.3

classification 🌊 nlin.CD cs.LGphysics.ao-ph

keywords climate model tuningsupermodelingparameter estimationsynchronizationadaptive modelingbias reductiondynamical systemsensemble methods

0 comments

The pith

Adaptive supermodeling tunes internal parameters of coupled climate models to match perfect-model performance.

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

The paper introduces adaptive supermodeling to solve the high cost of tuning climate models that have many variables and require long runs. It tests three methods in sequence: direct optimization of a single model's internal parameters, classical supermodeling that optimizes coupling weights between two models, and a new adaptive version that tunes the internal parameters themselves while the models stay coupled. In a test case built to expose weaknesses in the first two approaches, the adaptive method reaches accuracy comparable to a perfect model. This matters because it shows a route to bias reduction that works with short integration times rather than exhaustive long simulations.

Core claim

Adaptive supermodeling couples multiple climate models and continuously adjusts their internal parameters using synchronization-based estimation performed on short timescales, allowing the combined system to reach performance levels similar to a perfect model even in cases constructed to defeat direct parameter optimization and classical weight-based supermodeling.

What carries the argument

Adaptive supermodeling, which uses online synchronization-based parameter estimation to tune the internal parameters of supermodel members during short coupled integrations.

If this is right

Tuning can be performed on short timescales while still delivering accurate long-term climate statistics.
The method succeeds in cases where optimizing a single model's parameters or optimizing coupling weights alone fails.
Models tuned this way can later be integrated separately without losing the bias-reduction benefit.
Synchronization-based updates offer a practical way to handle the combination of high state dimension and long required integration times.

Where Pith is reading between the lines

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

The same synchronization-driven tuning could be tested on other high-dimensional dynamical systems that need long-term statistical accuracy.
Real-time parameter adjustment during a simulation might become feasible if the short-timescale updates prove robust.
The approach suggests a general strategy for bridging short-term dynamical matching to long-term statistical fidelity in ensemble modeling.

Load-bearing premise

Short synchronization-based updates will keep producing stable, unbiased long-term climate statistics once the tuned models run independently for decades or centuries.

What would settle it

Run the adaptively tuned models independently for multi-decadal or centennial periods and check whether their time-averaged statistics remain close to those of the perfect model without systematic drift or bias growth.

Figures

Figures reproduced from arXiv: 2510.06180 by Frank M. Selten, Jordan Seneca, Suzanne Bintanja.

**Figure 2.** Figure 2: fig. 2. The starting and final values of the parameters [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Figure 3: One timestep forecast synchronization er [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: Difference between truth and model climatology at 500 hPa before (top) and after (bottom) training. The error in the climatology of the trained model does not exceed the 0.5 m/s level. RMSE ⟨u⟩ RMSE σu Model 0 0.53±0.04 0.219±0.006 Model α 3.14±0.03 1.57±0.03 Model α ′ 0.59±0.06 0.246±0.007 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Shortest distance between p 0 and p S by setting w according to eq. (31). w to instead get p S as close to p 0 as possible. This is achieved by setting w = (p B − p A )(p 0 − p A ) |p B − p A| 2 , (31) as represented in fig. 5. Two cases where w ∈ [0,1] and w ∈/ [0,1] are explored in the next section. 4.2 Experiment Three models, 1, 2 and 3, are defined, the parameters of which differing from p 0 are shown… view at source ↗

**Figure 6.** Figure 6: SUMO weight with training started at t = 100 days. effective parameter values of the SUMO are closer to those of the truth model than any of the member models. The final values of the weights and effective parameter values are reported in table 4. SUMO12 SUMO13 τ S h [days] 3.92±0.12 5.88±0.05 τ S r [days] 57.0±1.3 25.60±0.09 w 0.617±0.016 −0.560±0.004 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

**Figure 7.** Figure 7: Error of climatological zonal winds at the [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

**Figure 8.** Figure 8: Error of climatological zonal winds at the 500 hPa level of SUMO12 and SUMO13, going down. in tendency are now ∂hi ∂w = Fi(Q S , p A )−Fi(Q S , p B ) (32) ∂hi ∂ p A j = (1−w) ∂Fi(Q S , p A j ) ∂ p A j (33) ∂hi ∂ p B j = w ∂Fi(Q S , p B j ) ∂ p B j (34) where p x j is some parameter pj of model x. The update rules for the weight and model parameters are then w˙ = −rw∑ i [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗

**Figure 9.** Figure 9: Training of parameters of ASUMO14. imperfect models with non-trainable parameters was optimized, stabilizing after 25 days, and achieving a global mean RMSE of < 1.5 m/s, compared to > 4.5 m/s for any member model. This included a case where the SUMO members had the same sign parameter biases, which was optimized by finding a negative value for the weight. These results demonstrate the robust performance o… view at source ↗

**Figure 10.** Figure 10: Parameter and tendency space of the different methods used in section 5.2. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗

**Figure 11.** Figure 11: Error in the climatological zonal winds at the 500 hPa level of, going left to right, from the top: [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗

read the original abstract

In climate science, the tuning of climate models is a computationally intensive problem due to the combination of the high-dimensionality of the system state and long integration times. Supermodelling is a technique which has shown the potential for reducing climate model biases by dynamically coupling multiple models together, and training their coupling on a short timescale. Here, we introduce a new approach called \emph{adaptive supermodeling}, where the internal model parameters of the member of a supermodel are tuned. We perform three experiments. We first directly optimize the internal parameters of a climate model. We then optimize the weights between two members of a supermodel in a classical supermodel approach. For a case designed to challenge the two previous methods, we implement adaptive supermodeling, which achieves a performance similar to a perfect model.

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.

Desk Editor's Note private letter to a colleague

Adaptive supermodeling tunes both couplings and internal parameters online during short synchronized runs and claims to match perfect-model performance where direct tuning and classical supermodeling fail, but long-term independent-run validation is missing.

read the letter

The main thing to know is that this paper extends supermodeling by tuning internal model parameters at the same time as the coupling weights, all during short synchronized integrations. In their third experiment, designed to be hard for the first two methods, the adaptive version reportedly reaches performance close to a perfect model. That is the core claim and the practical hook. What the work does reasonably well is frame the real computational bottleneck of climate model tuning and show a logical progression through the three experiments. Direct parameter optimization comes first, then classical weight tuning in a supermodel, and finally the combined adaptive approach. The simultaneous online adjustment during coupling is a clear, if incremental, step beyond the referenced prior supermodeling literature. The idea could matter for anyone trying to reduce bias without running separate long tuning integrations. The soft spots are straightforward and worth naming in proportion. The abstract supplies no quantitative metrics, error bars, or description of the test model, so the performance claim cannot be judged from the given text. More critically, the parameter updates occur only over short coupled runs; nothing shown guarantees that the resulting values produce unbiased statistics once the models run independently over decades or centuries. Synchronization can suppress or compensate for biases that reappear in free runs, and the stress-test note on this point holds up. The paper is aimed at climate model developers and researchers working on ensemble or online adaptation methods. A reader focused on practical tuning algorithms would find the description useful to consider, even if they would want stronger checks before adopting it. It shows clear engagement with the problem and the supermodeling literature, so it qualifies as serious thinking on its own terms. I would send it to peer review rather than desk reject. The method is concrete enough and the underlying issue important enough that referees should see the full details and ask for the missing long-term validation.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces adaptive supermodeling, a technique that tunes the internal parameters of member models within a supermodel using synchronization-based online updates. It describes three experiments: (1) direct optimization of a single climate model's internal parameters, (2) optimization of coupling weights in a classical supermodel, and (3) adaptive supermodeling applied to a case designed to challenge the first two methods. The central claim is that adaptive supermodeling achieves performance comparable to a perfect model in this challenging scenario.

Significance. If the performance claim is supported by quantitative evidence and long-term validation, the method could provide an efficient route to climate model tuning that avoids the computational cost of long integrations by leveraging short-timescale synchronization for parameter adjustment. This would extend existing supermodeling work by making the member models themselves adaptive rather than only their couplings.

major comments (2)

[Abstract] Abstract: the claim that adaptive supermodeling 'achieves a performance similar to a perfect model' is presented without any quantitative metrics, error bars, or description of the test model and evaluation statistics. This absence prevents assessment of whether the result is statistically meaningful or merely qualitative.
[Experiments (likely §3 or §4)] The synchronization-based parameter updates are performed during short coupled integrations. No explicit test is reported showing that the resulting parameter values yield stable, unbiased climate statistics when the tuned models are subsequently run independently over long timescales (decades/centuries) without the synchronizing coupling. Synchronization can suppress or compensate for biases that may reappear in free runs, so this verification is load-bearing for the central claim.

minor comments (2)

[Methods] Provide the explicit form of the synchronization error metric and the parameter-update rule (e.g., as an equation) so that the method can be reproduced.
[Experiments] Clarify the dimensionality and integration length of the test model used in the three experiments.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major point below and indicate the revisions we will incorporate to improve clarity and completeness.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that adaptive supermodeling 'achieves a performance similar to a perfect model' is presented without any quantitative metrics, error bars, or description of the test model and evaluation statistics. This absence prevents assessment of whether the result is statistically meaningful or merely qualitative.

Authors: We agree that the abstract is concise and omits quantitative details. The body of the manuscript (Sections 3 and 4) reports specific metrics, including time-averaged RMS errors with standard deviations across ensemble members and direct comparisons to the perfect-model reference for the low-order climate model used. To address the concern, we will revise the abstract to include key quantitative results, a brief description of the test model, and the primary evaluation statistics. revision: yes
Referee: [Experiments (likely §3 or §4)] The synchronization-based parameter updates are performed during short coupled integrations. No explicit test is reported showing that the resulting parameter values yield stable, unbiased climate statistics when the tuned models are subsequently run independently over long timescales (decades/centuries) without the synchronizing coupling. Synchronization can suppress or compensate for biases that may reappear in free runs, so this verification is load-bearing for the central claim.

Authors: This is a substantive point. The current experiments focus on performance during the adaptive synchronization phase for the designed challenging case. We will add explicit long-term free-run validation: after parameter tuning, we will integrate the tuned member models independently for extended periods (hundreds of model years) and report climate statistics (means, variances, and spectra) compared to the perfect model. This will confirm that the tuned parameters produce stable, unbiased behavior without the synchronizing coupling. revision: yes

Circularity Check

0 steps flagged

No circularity: experimental results from short-time tuning do not reduce to input by construction

full rationale

The paper reports three numerical experiments on parameter optimization and adaptive supermodeling using synchronization-based updates over short integration times. The headline result (adaptive supermodeling matching perfect-model performance in a designed challenge case) is presented as an empirical outcome rather than a closed-form derivation. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the provided abstract or description that would make the reported performance equivalent to the tuning inputs by construction. The work is self-contained as an experimental demonstration evaluated against independent long-term statistics benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the untested assumption that short-timescale synchronization updates remain valid for long climate integrations; no free parameters, axioms, or invented entities are explicitly listed in the provided abstract.

pith-pipeline@v0.9.0 · 5664 in / 1092 out tokens · 31607 ms · 2026-05-18T09:16:03.065217+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

˙pj =−r ∑i ∂L0/∂ei ∂hi/∂pj with L0≡(Q−Q0)² and hi the tendency error

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

E., Leung, L

Barthel Sorensen, B., Charalampopoulos, A., Zhang, S., Harrop, B. E., Leung, L. R., and Sapsis, T. P.: A Non-Intrusive Machine Learning Framework for Debiasing Long-Time Coarse Resolution Climate Simulations and Quantifying Rare Events Statistics, Journal of Advances in Modeling Earth Systems, 16, e2023MS004 122, doi:https://doi.org/10.1029/2023MS004122, ...

work page doi:10.1029/2023ms004122
[2]

Hourdin, F., Mauritsen, T., Gettelman, A., Golaz, J.- C., Balaji, V ., Duan, Q., Folini, D., Ji, D., Klocke, D., Qian, Y ., Rauser, F., Rio, C., Tomassini, L., Watanabe, M., and Williamson, D.: The Art and Science of Climate Model Tuning, Bulletin of the American Meteorological Society, 98, 589 – 602, doi:10.1175/BAMS-D-15-00135.1, URLhttps: //journals.am...

work page doi:10.1175/bams-d-15-00135.1
[3]

com/doi/abs/10.1029/2022MS003367, e2022MS003367 2022MS003367,

Lguensat, R., Deshayes, J., Durand, H., and Balaji, V .: Semi-Automatic Tuning of Cou- pled Climate Models With Multiple Intrinsic Timescales: Lessons Learned From the Lorenz96 Model, Journal of Advances in Modeling Earth Systems, 15, e2022MS003 367, doi: https://doi.org/10.1029/2022MS003367, URL https://agupubs.onlinelibrary.wiley. com/doi/abs/10.1029/20...

work page doi:10.1029/2022ms003367
[4]

com/doi/abs/10.1029/2012MS000154,

Mauritsen, T., Stevens, B., Roeckner, E., Crueger, T., Esch, M., Giorgetta, M., Haak, H., Jungclaus, J., Klocke, D., Matei, D., Mikolajewicz, U., Notz, D., Pincus, R., Schmidt, H., and Tomassini, L.: Tuning the climate of a global model, Journal of Advances in Modeling Earth Systems, 4, doi:https://doi.org/10.1029/2012MS000154, URL https://agupubs.onlinel...

work page doi:10.1029/2012ms000154
[5]

M., Sexton, D

Murphy, J. M., Sexton, D. M. H., Barnett, D. N., Jones, G. S., Webb, M. J., Collins, M., and Stainforth, D. A.: Quantification of modelling uncertainties in a large ensemble of climate change simulations, Nature, 430, 768–772, doi: 10.1038/nature02771, URLhttps://doi.org/ 10.1038/nature02771,

work page doi:10.1038/nature02771
[6]

22541/essoar.173238717.76577266/v1,

Nath, P., Moss, H., Shuckburgh, E., and Webb, M.: RAIN: Reinforcement Algorithms for Improving Numerical Weather and Climate Models, doi:10. 22541/essoar.173238717.76577266/v1,

work page arXiv
[7]

Oliver, S., Cartis, C., Kriest, I., Tett, S. F. B., and Khatiwala, S.: A derivative-free optimisation method for global ocean biogeochemical models, Geoscientific Model Development, 15, 3537– 3554, doi:10.5194/gmd-15-3537-2022, URL https://gmd.copernicus.org/articles/ 15/3537/2022/,

work page doi:10.5194/gmd-15-3537-2022 2022
[8]

Parlitz, U.: Estimating Model Parameters from Time Series by Autosynchronization, Phys. Rev. Lett., 76, 1232–1235, doi:10.1103/PhysRevLett. 18 76.1232, URLhttps://link.aps.org/doi/ 10.1103/PhysRevLett.76.1232,

work page doi:10.1103/physrevlett
[9]

Prosperino, D., Ma, H., and R ¨ath, C.: A generalized method for estimating parameters of chaotic sys- tems using synchronization with modern optimiz- ers, Journal of Physics: Complexity, 6, 015 012, doi:10.1088/2632-072X/adaa46, URLhttps:// doi.org/10.1088/2632-072X/adaa46,

work page doi:10.1088/2632-072x/adaa46
[10]

copernicus.org/articles/10/789/2019/,

Schevenhoven, F., Selten, F., Carrassi, A., and Keenlyside, N.: Improving weather and cli- mate predictions by training of supermodels, Earth System Dynamics, 10, 789–807, doi: 10.5194/esd-10-789-2019, URLhttps://esd. copernicus.org/articles/10/789/2019/,

work page doi:10.5194/esd-10-789-2019 2019
[11]

van Noije, T., Bergman, T., Le Sager, P., O’Donnell, D., Makkonen, R., Goncalves-Ageitos, M., Doscher, R., Fladrich, U., von Hardenberg, J., Keskinen, J.-P., Korhonen, H., Laakso, A., Myriokefalitakis, S., Ollinaho, P., Perez Garcia- Pando, C., Reerink, T., Schrodner, R., Wyser, K., and Yang, S.: EC-Earth3-AerChem: a global climate model with interactive ...

work page doi:10.5194/gmd-14-5637-2021 2021
[12]

Estimat- ing Model Parameters from Time Series by Autosynchronization

Yu, D. and Wu, A.: Comment on “Estimat- ing Model Parameters from Time Series by Autosynchronization”, Phys. Rev. Lett., 94, 219 401, doi:10.1103/PhysRevLett.94.219401, URLhttps://link.aps.org/doi/10.1103/ PhysRevLett.94.219401,

work page doi:10.1103/physrevlett.94.219401

[1] [1]

E., Leung, L

Barthel Sorensen, B., Charalampopoulos, A., Zhang, S., Harrop, B. E., Leung, L. R., and Sapsis, T. P.: A Non-Intrusive Machine Learning Framework for Debiasing Long-Time Coarse Resolution Climate Simulations and Quantifying Rare Events Statistics, Journal of Advances in Modeling Earth Systems, 16, e2023MS004 122, doi:https://doi.org/10.1029/2023MS004122, ...

work page doi:10.1029/2023ms004122

[2] [2]

Hourdin, F., Mauritsen, T., Gettelman, A., Golaz, J.- C., Balaji, V ., Duan, Q., Folini, D., Ji, D., Klocke, D., Qian, Y ., Rauser, F., Rio, C., Tomassini, L., Watanabe, M., and Williamson, D.: The Art and Science of Climate Model Tuning, Bulletin of the American Meteorological Society, 98, 589 – 602, doi:10.1175/BAMS-D-15-00135.1, URLhttps: //journals.am...

work page doi:10.1175/bams-d-15-00135.1

[3] [3]

com/doi/abs/10.1029/2022MS003367, e2022MS003367 2022MS003367,

Lguensat, R., Deshayes, J., Durand, H., and Balaji, V .: Semi-Automatic Tuning of Cou- pled Climate Models With Multiple Intrinsic Timescales: Lessons Learned From the Lorenz96 Model, Journal of Advances in Modeling Earth Systems, 15, e2022MS003 367, doi: https://doi.org/10.1029/2022MS003367, URL https://agupubs.onlinelibrary.wiley. com/doi/abs/10.1029/20...

work page doi:10.1029/2022ms003367

[4] [4]

com/doi/abs/10.1029/2012MS000154,

Mauritsen, T., Stevens, B., Roeckner, E., Crueger, T., Esch, M., Giorgetta, M., Haak, H., Jungclaus, J., Klocke, D., Matei, D., Mikolajewicz, U., Notz, D., Pincus, R., Schmidt, H., and Tomassini, L.: Tuning the climate of a global model, Journal of Advances in Modeling Earth Systems, 4, doi:https://doi.org/10.1029/2012MS000154, URL https://agupubs.onlinel...

work page doi:10.1029/2012ms000154

[5] [5]

M., Sexton, D

Murphy, J. M., Sexton, D. M. H., Barnett, D. N., Jones, G. S., Webb, M. J., Collins, M., and Stainforth, D. A.: Quantification of modelling uncertainties in a large ensemble of climate change simulations, Nature, 430, 768–772, doi: 10.1038/nature02771, URLhttps://doi.org/ 10.1038/nature02771,

work page doi:10.1038/nature02771

[6] [6]

22541/essoar.173238717.76577266/v1,

Nath, P., Moss, H., Shuckburgh, E., and Webb, M.: RAIN: Reinforcement Algorithms for Improving Numerical Weather and Climate Models, doi:10. 22541/essoar.173238717.76577266/v1,

work page arXiv

[7] [7]

Oliver, S., Cartis, C., Kriest, I., Tett, S. F. B., and Khatiwala, S.: A derivative-free optimisation method for global ocean biogeochemical models, Geoscientific Model Development, 15, 3537– 3554, doi:10.5194/gmd-15-3537-2022, URL https://gmd.copernicus.org/articles/ 15/3537/2022/,

work page doi:10.5194/gmd-15-3537-2022 2022

[8] [8]

Parlitz, U.: Estimating Model Parameters from Time Series by Autosynchronization, Phys. Rev. Lett., 76, 1232–1235, doi:10.1103/PhysRevLett. 18 76.1232, URLhttps://link.aps.org/doi/ 10.1103/PhysRevLett.76.1232,

work page doi:10.1103/physrevlett

[9] [9]

Prosperino, D., Ma, H., and R ¨ath, C.: A generalized method for estimating parameters of chaotic sys- tems using synchronization with modern optimiz- ers, Journal of Physics: Complexity, 6, 015 012, doi:10.1088/2632-072X/adaa46, URLhttps:// doi.org/10.1088/2632-072X/adaa46,

work page doi:10.1088/2632-072x/adaa46

[10] [10]

copernicus.org/articles/10/789/2019/,

Schevenhoven, F., Selten, F., Carrassi, A., and Keenlyside, N.: Improving weather and cli- mate predictions by training of supermodels, Earth System Dynamics, 10, 789–807, doi: 10.5194/esd-10-789-2019, URLhttps://esd. copernicus.org/articles/10/789/2019/,

work page doi:10.5194/esd-10-789-2019 2019

[11] [11]

van Noije, T., Bergman, T., Le Sager, P., O’Donnell, D., Makkonen, R., Goncalves-Ageitos, M., Doscher, R., Fladrich, U., von Hardenberg, J., Keskinen, J.-P., Korhonen, H., Laakso, A., Myriokefalitakis, S., Ollinaho, P., Perez Garcia- Pando, C., Reerink, T., Schrodner, R., Wyser, K., and Yang, S.: EC-Earth3-AerChem: a global climate model with interactive ...

work page doi:10.5194/gmd-14-5637-2021 2021

[12] [12]

Estimat- ing Model Parameters from Time Series by Autosynchronization

Yu, D. and Wu, A.: Comment on “Estimat- ing Model Parameters from Time Series by Autosynchronization”, Phys. Rev. Lett., 94, 219 401, doi:10.1103/PhysRevLett.94.219401, URLhttps://link.aps.org/doi/10.1103/ PhysRevLett.94.219401,

work page doi:10.1103/physrevlett.94.219401