Hybrid physics-data-driven modeling for sea ice thermodynamics and transfer learning

Alberto Carrassi; Dorotea Iovino; Giovanni De Cillis; Julien Brajard; Laurent Bertino; Marc Bocquet; Matteo Broccoli; Tobias Sebastian Finn

arxiv: 2601.23190 · v2 · submitted 2026-01-30 · ⚛️ physics.ao-ph

Hybrid physics-data-driven modeling for sea ice thermodynamics and transfer learning

Giovanni De Cillis , Alberto Carrassi , Julien Brajard , Laurent Bertino , Matteo Broccoli , Dorotea Iovino , Tobias Sebastian Finn , Marc Bocquet This is my paper

Pith reviewed 2026-05-16 09:15 UTC · model grok-4.3

classification ⚛️ physics.ao-ph

keywords sea icehybrid modelingneural networkstransfer learningforecast errorsthermodynamicsradiative properties

0 comments

The pith

Neural networks trained on model-state errors deliver more stable and accurate long-lead sea-ice forecasts than fixed climatological corrections.

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

The paper tests a hybrid approach in which a neural network learns to predict and correct forecast errors inside a sea-ice column model, using the current thermodynamic state as input. The authors generate training data by perturbing snow thermodynamics and radiative properties, train separate networks for each configuration, and compare the hybrids to a benchmark that applies climatological error estimates. The neural-network versions remain stable, tolerate errors in initial conditions and atmospheric forcing, and produce better forecasts at long lead times. Separate experiments show that networks pretrained on one model setup can be adapted to a different setup through transfer learning, avoiding the need for full retraining whenever the physics changes. Feature analysis reveals that ice-layer enthalpies drive most of the predictive skill while atmospheric forcing contributes little.

Core claim

A neural-network component that acts as a state-dependent parameterization of forecast errors produces hybrid sea-ice models that are stable over long lead times, robust to initial-condition and forcing errors, and consistently superior to climatology-based error estimates; pretrained networks can be successfully transferred to new physical configurations when their direct performance exceeds a trivial baseline.

What carries the argument

The neural network that maps the sea-ice model's thermodynamic state (primarily ice-layer enthalpies) to corrections for snow and radiative forecast errors.

If this is right

Long-lead sea-ice forecasts can be run without progressive drift when error correction depends on the current ice state rather than fixed statistics.
Model updates in snow or radiative physics require only transfer learning of the existing network instead of retraining from scratch.
Error correction can be focused on ice enthalpy profiles, reducing the data volume needed from atmospheric inputs.
The hybrid system tolerates realistic errors in initial conditions and forcing without becoming unstable.

Where Pith is reading between the lines

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

Operational centers could adopt the hybrid approach to maintain forecast skill after routine physics updates without rebuilding training datasets each time.
The same state-dependent correction idea could be tested on other column models such as ocean mixed-layer or snow-on-land schemes where thermodynamic errors are also state-sensitive.
Prioritizing accurate ice-temperature observations would likely yield larger gains in forecast skill than adding more atmospheric variables.

Load-bearing premise

The chosen perturbations in snow thermodynamics and sea-ice radiative properties generate training errors that adequately represent the range of forecast errors seen under real atmospheric forcing.

What would settle it

An independent test set of real forecasts with varied atmospheric forcing in which the hybrid model shows no accuracy gain or loses stability over the climatological benchmark would disprove the central performance claim.

read the original abstract

This study explores a physics-data driven hybrid approach for sea-ice column physics models, in which a machine learning (ML) component acts as a state-dependent parameterization of forecast errors. We examine how perturbations in snow thermodynamics and sea-ice radiative properties affect forecast errors, and train dedicated neural networks (NNs) for each model configuration. The performance of the hybrid models is evaluated for long lead-time forecasts and compared against a benchmark system based on climatological forecast-error estimates. The NN-based hybrids prove to be stable, robust to initial condition and atmospheric forcing errors, and consistently outperform their climatology-based counterpart. To derive guiding principles for efficiently handling possible physical model updates, we perform transfer learning experiments to test whether pretrained NNs optimized for one model configuration can be successfully adapted to another. Results indicate that direct evaluation of pretrained networks on the target task provides useful insights into their adaptability, recommending transfer learning whenever performance exceeds a trivial baseline. Finally, a feature-importance analysis shows that atmospheric forcing inputs have negligible influence on NN predictive skill, while ice-layer enthalpies play a key role in achieving satisfactory performance.

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

The paper shows a workable state-dependent NN correction for sea-ice thermodynamics that transfers across configs, but the robustness claims rest on thin quantitative evidence.

read the letter

The main thing here is a hybrid setup where neural nets learn state-dependent corrections to forecast errors in a sea-ice column model, trained on perturbations to snow thermodynamics and radiative properties, then tested for transfer learning when the base physics changes. They compare against a climatology benchmark and report stability plus better long-lead performance, plus a feature analysis that downplays atmospheric forcing inputs in favor of ice enthalpies. That combination of targeted error parameterization and transfer experiments is new for this subfield and gives some practical guidance on reusing networks after model updates. The stability result under varied initial conditions is a clear positive. The soft spots are the missing numbers: the abstract gives no RMSE values, training volumes, cross-validation scheme, or significance tests, so the size of the improvement and its reliability stay hard to judge. The perturbation ensemble used for training may also under-sample the range of real atmospheric forcing variability, which weakens the generalization claim even if it does not break the central argument. This is for people building hybrid corrections in polar modeling or operational sea-ice forecasting. It has enough substance and relevance to merit a serious referee, though revisions will need to add the quantitative validation and broader forcing tests. I would send it to review.

Referee Report

3 major / 2 minor

Summary. The manuscript presents a hybrid physics-data-driven approach for sea ice column thermodynamics in which neural networks serve as state-dependent corrections to forecast errors induced by targeted perturbations in snow thermodynamics and sea-ice radiative properties. Dedicated NNs are trained for each model configuration, evaluated on long-lead-time forecasts, and compared to a climatology-based error benchmark. The study further examines transfer learning across configurations and performs feature-importance analysis, concluding that the hybrids are stable and robust to initial-condition and forcing errors while outperforming the climatological baseline, with ice-layer enthalpies dominating predictive skill.

Significance. If the performance claims are substantiated with quantitative metrics, the work offers a practical route to improving sea-ice forecast skill without sacrificing physical consistency. The transfer-learning results could reduce the cost of incorporating model updates, and the feature-importance finding that atmospheric forcing inputs contribute negligibly while enthalpies are critical provides mechanistic insight into error sources. Demonstrated robustness to perturbations supports potential operational deployment, provided generalization beyond the training ensemble is confirmed.

major comments (3)

[Results section] Results section: the abstract and main text assert 'consistent outperformance' and 'robustness' to initial-condition and atmospheric-forcing errors, yet no quantitative error metrics (RMSE, bias, anomaly correlation), training-set sizes, cross-validation scheme, or statistical significance tests are reported, leaving the central performance claims only moderately supported.
[Methods section] Methods section describing the perturbation ensemble: no quantitative coverage metric (range or distribution of surface heat fluxes, wind stresses, or precipitation rates realized in the training perturbations) is provided relative to operational reanalysis, so the claim that the hybrids remain stable under arbitrary forcing cannot be assessed.
[Transfer learning experiments] Transfer-learning subsection: the recommendation to apply transfer learning 'whenever performance exceeds a trivial baseline' is not accompanied by explicit numerical thresholds, adaptation protocols, or ablation results, limiting the practical guidance offered for model updates.

minor comments (2)

[Abstract] Abstract: inclusion of at least one representative quantitative skill score would strengthen the summary of results.
[Figures] Figure captions and legends: ensure all panels are clearly labeled with units and that NN versus climatology comparisons are visually distinguishable.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their insightful comments, which have helped us improve the clarity and rigor of our manuscript. We address each major comment below and indicate the revisions we plan to make.

read point-by-point responses

Referee: [Results section] Results section: the abstract and main text assert 'consistent outperformance' and 'robustness' to initial-condition and atmospheric-forcing errors, yet no quantitative error metrics (RMSE, bias, anomaly correlation), training-set sizes, cross-validation scheme, or statistical significance tests are reported, leaving the central performance claims only moderately supported.

Authors: We agree that including explicit quantitative metrics would strengthen the presentation of our results. In the revised manuscript, we will add a new table summarizing RMSE, bias, and anomaly correlation for key variables (e.g., ice thickness, snow depth, surface temperature) at various lead times, comparing the hybrid models to the climatological benchmark. Training set sizes will be specified (approximately 10,000 samples per configuration from the perturbation ensemble), and we will describe the 5-fold cross-validation scheme used during training. Statistical significance of the outperformance will be evaluated using Wilcoxon signed-rank tests, with results reported in the text and table. These additions will provide the quantitative support requested. revision: yes
Referee: [Methods section] Methods section describing the perturbation ensemble: no quantitative coverage metric (range or distribution of surface heat fluxes, wind stresses, or precipitation rates realized in the training perturbations) is provided relative to operational reanalysis, so the claim that the hybrids remain stable under arbitrary forcing cannot be assessed.

Authors: We appreciate this observation. While our stability claims are specific to the range of perturbations applied in the ensemble (which were designed to represent plausible variations in snow thermodynamics and radiative properties), we acknowledge the value of contextualizing these against operational data. In the revision, we will include a supplementary figure or table detailing the min/max and distribution statistics of the perturbed surface heat fluxes, wind stresses, and precipitation rates. These will be compared to corresponding statistics from ERA5 reanalysis over the Arctic region for the same period. We will also clarify in the text that the robustness is demonstrated for perturbations within the sampled range, not for completely arbitrary forcing outside this range. revision: yes
Referee: [Transfer learning experiments] Transfer-learning subsection: the recommendation to apply transfer learning 'whenever performance exceeds a trivial baseline' is not accompanied by explicit numerical thresholds, adaptation protocols, or ablation results, limiting the practical guidance offered for model updates.

Authors: We thank the referee for highlighting the need for more precise guidance. In the revised version, we will specify a numerical threshold: transfer learning is recommended if the pretrained network achieves at least 70% of the performance (in terms of RMSE reduction) of a model trained from scratch on the target configuration. We will describe the adaptation protocol, including the number of fine-tuning epochs (typically 20-50) and learning rate schedule (initial rate of 1e-4 for the output layers). Additionally, we will include ablation results comparing three approaches: (i) direct application of pretrained NN, (ii) fine-tuning, and (iii) training from scratch, with performance metrics for each. This will offer clearer practical advice for handling model updates. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical training on external forecast errors yields independent performance claims

full rationale

The paper generates forecast errors from independent physics-model integrations under targeted perturbations to snow thermodynamics and radiative properties, then trains NNs to predict those errors as a state-dependent correction. Evaluation compares hybrid forecasts against a separate climatology benchmark on long-lead runs, with transfer-learning tests checking adaptability across configurations. No equation or claim reduces by construction to a fitted parameter renamed as prediction, no self-citation chain justifies a uniqueness result, and no ansatz is smuggled via prior work. The central robustness and outperformance statements rest on direct numerical comparison to external data rather than tautological re-use of the training distribution itself.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that forecast errors are learnable from state variables and that the tested perturbations represent realistic error distributions. No new physical entities are introduced.

free parameters (1)

Neural network weights and biases
Fitted during supervised training to minimize the difference between predicted and observed forecast errors.

axioms (1)

domain assumption Forecast errors in sea-ice column models are state-dependent functions that can be approximated by neural networks
Invoked when the ML component is introduced as a parameterization of forecast errors.

pith-pipeline@v0.9.0 · 5514 in / 1214 out tokens · 26007 ms · 2026-05-16T09:15:34.310285+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.

Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We examine how perturbations in snow thermodynamics and sea-ice radiative properties affect forecast errors, and train dedicated neural networks (NNs) for each model configuration.
Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The NN-based hybrids prove to be stable, robust to initial condition and atmospheric forcing errors

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.