Hybrid physics-data-driven modeling for sea ice thermodynamics and transfer learning
G. De Cillis, A. Carrassi, J. Brajard, L. Bertino, M. Broccoli, D. Iovino, T.S. Finn, M. Bocquet · 2026
Read Hybrid physics-data-driven modeling for sea ice thermodynamics and transfer learning by G. De Cillis, A. Carrassi, J. Brajard, L. Bertino, M. Broccoli, D. Iovino, T.S. Finn, M. Bocquet (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-23190.txt.
Introduction
G. De Cillis and seven co-authors blend physics models with machine learning for sea ice forecasts. Their hybrid treats the learning component as a state-dependent fix for forecast errors in a sea-ice column model. Snow thermodynamics and the ice's radiative properties are key error sources.
They train dedicated neural networks for each model configuration and test transfer to new ones. Transfer learning matters because Arctic conditions vary by season and place. Better sea-ice forecasts serve shipping, climate science, and northern communities. State-dependent parameterization is the key mechanism, letting the learned correction change with the ice and snow conditions.
Catalog record
| CREATOR | G. De Cillis, A. Carrassi, J. Brajard, L. Bertino, M. Broccoli, D. Iovino, T.S. Finn, M. Bocquet |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 73 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-23190
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Hybrid physics-data-driven modeling for sea ice thermodynamics and transfer learning G. De Cillis, A. Carrassi, J. Brajard, L. Bertino, M. Broccoli, D. Iovino, T.S. Finn, M. Bocquet 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. Keywords — sea ice, forecast error, hybrid models, neural networks, transfer learning 1 Introduction Over the past 15 years, machine learning (ML) has experienced exceptional growth, driven by remarkable progress in computational power and the pivotal introduction of GPU-accelerated training for deep neural networks as demonstrated with the breakthrough in image classification by 37 with AlexNet. Since then, ML applications have permeated a wide range of scientific and technological domains, among which numerical weather and climate modeling has emerged as a prominent example, enabled by the extension of computer vision techniques to the geosciences. Recent advancements have enabled the development of global, fully data-driven, ML-based weather prediction (MLWP) systems ( 43 ; 38 ; 1 ; 13 ; 4 ) . Trained primarily on ERA5 reanalysis data ( 30 ) , these models have shown comparable and sometimes superior skills with respect to ECMWF’s deterministic operational forecasting system, IFS-HRES, on short-to-medium range forecasts ( 45 ) . The main advantage of these models is their exceptional computational speed and efficiency, compared to classical physics-based models. However, they could remain limited by their lack of physical consistency and realism ( 6 ) , which are crucial for meteorology in complex terrain, extreme event, coupled phenomena, as well as for seasonal and climate predictions. Additionally, the majority of them rely on high resolution physics-based models and data assimilation (DA) for their training and initialization. Another pragmatic approach, more conservative, is represented by hybrid modeling that combines ML with traditional physics-based models, in an attempt to extract the best of both while mitigating their individual weaknesses. This combination is often achieved through the replacement of physics-based or empirical parameterizations with data-driven ML components, with the aim of improving accuracy or reducing computational cost ( 15 ; 41 ; 46 ; 5 ) . In the context of sea ice modeling, 17 trained Neural Networks (NNs) to emulate an advanced physics-based melt pond parameterization within the ice column model Icepack ( 32 ) , while 31 focused on wave-induced ice floe fracturing through a neural-network emulation of a computationally expensive super-parameterization. Despite continuous progress, numerical weather and climate model predictions inevitably remain affected by systematic errors, due to discretization errors, imperfect subgrid-scale parameterizations or inaccurate boundary conditions, initial conditions and forcings. Therefore, instead of focusing on specific process parameterizations, another research branch on hybrid modeling targeted the model error, developing ML-based model error parameterizations that complement physically derived models during execution ( 52 ) . Similar to MLWP models, these hybrid models for model error correction leverage DA that, along with the increasing volume and quality of observational data, plays a fundamental role in geoscience, by providing optimal estimates ( analyses ) of the state of the earth system ( 10 ) . In particular, analysis increments (differences between analysis and forecast states) are exploited to infer state-dependent systematic model errors ( 11 ) . Key contributions along these lines include the work of 7 and 22 , who trained NNs as additive integrated correction term between two forecast times, i.e. in the resolvent of geophysical models, learning from analysis increments, with particular focus on the role of observation noise and sparseness. 21 further investigated the potential of tendency correction and introduced an online learning strategy, in which both the model states and the model parameter are estimated at the same time as soon as new observations become available. Similar approaches to hybrid modeling have also been applied to more realistic scenarios, such as the work of 53 , who trained a ML model to emulate nudging tendency terms within a weather forecast system; 20 who learned a neural network correction to the Integrated Forecasting System of the ECMWF both offline and online; 18 , who developed ML-based corrections for ocean mixed-layer temperature biases in a global climate model; and 12 , who improved multidecadal atmospheric predictions through online bias correction trained on nudging increments. This type of hybrid modeling has also been explored for sea ice. For example, 23 employed convolutional U-nets to parametrize model error due to unresolved subgrid scale of sea-ice dynamics within an unstructured finite element model with a Maxwell elasto-brittle rheology. More recently, 28 successfully trained Convolutional Neural Networks (CNNs) to correct systematic errors of a global coupled ocean-sea ice model, demonstrating improvements achieved by iterating DA and ML steps ( 2 ) . Building on this work, 27 developed a hybrid global coupled ice–atmosphere–ocean climate model, highlighting the importance of fully coupled training for generalization in free forecasts. 29 applied ML for online error correction within an earth system model for seasonal arctic sea ice prediction and compared it with an offline approach where ML post-processes and calibrates model predictions after the simulation. In the present study, we build upon previous works to develop state-dependent, ML-based resolvent corrections for the sea-ice column model, Icepack. Model errors are introduced by perturbing parameters related to snow thermodynamics and radiative properties, while synthetic observations are generated using a reference configuration. Neural networks are trained to predict systematic forecast errors, conditioned on the forecast model state. The online execution of the hybrid models, therefore, combines consecutive forecast steps with NN-based corrections, in a similar fashion as sequential DA algorithms. This hybrid framework enables a detailed assessment of how specific parametric model errors influence forecasts and the corresponding ML bias corrections. Insights into the relative importance of predictors for estimating model error, and indications for a shared minimal model, are obtained through recursive feature elimination analyses applied across all perturbed configurations. An important, yet often overlooked, challenge in the operational deployment of model-error-correcting hybrid systems is the impact of model updates (i.e., when a new model enters into play, with e.g. new parameterizations, numerical schemes, etc.), which can cause abrupt shifts in error statistics. Unlike gradual changes induced by climate variability, which can be addressed with online learning strategies, model version updates require dedicated measures. The most straightforward approach is to generate a