Abstract
MELTED - MachinE Learning for arcTic ice prEDiction. The primary goal is to develop a hybrid physics–data driven models for sea ice prediction. In this approach, machine learning is used to learn model error parameterizations by leveraging data assimilation outputs (analyses), which serve as an online, ML-based, state-dependent bias correction to the physical model core. Risultati Attesi: "1. Develpment of an hybrid physics-ML model based on the 1D column-physics model Icepack, using observing system simulation experiments (OSSEs). 1.1 Detailed investigation of the effect of different parametric model errors on forecasts and correspondent ML bias corrections. 1.2 Evaluation of the hybrid models' performance on long lead-time forecasts and assessment of its robustness under perturbed forcing conditions. 1.4 Execution of novel transfer learning experiments, examining in detail the capabilities and conditions allowing for an ML model optimized to bias-correct a given physical model to be successfully applied to another physical model. 2. Training of a 2D convolutional neural network to predict sea ice concentration and thickness analysis increments from the CMCC Global Ocean Physical Reanalysis System, based on the NEMO-CICE coupled ocean-sea ice model. 2.1 Investigation of transfer learning strategies to effectively use a dataset that is not statistically homogeneous, due to changes in the assimilated sea ice observations. 2.2 Development of a NEMO-CICE-NN hybrid model."
Results achieved
: The University of Bologna (Unibo) unit carried out a comprehensive assessment of the state of the art in machine-learning applications to sea-ice modelling and forecasting. The review covered fully data-driven forecasting systems, hybrid physics–machine-learning approaches, data-assimilation applications, model-error correction, parameter estimation and remote-sensing methods. This activity identified the most promising methodological directions for the subsequent project work. Building on this background, a hybrid framework combining the Icepack sea-ice column model with neural-network-based error correction was developed and evaluated. The neural networks were trained to estimate state-dependent forecast errors associated with uncertainties in snow thermodynamics and radiative properties and were coupled online with the physical model through sequential corrections. The resulting hybrid system proved stable and physically consistent over long forecasts and consistently outperformed corrections based on climatology. The ability of the machine-learning corrections to adapt to updates of the underlying physical model was also investigated. Transfer-learning experiments showed that fine-tuning can outperform retraining from scratch when only limited data from the updated model are available. A diagnostic criterion was proposed to determine when fine-tuning is likely to be advantageous, providing practical guidance for reducing the computational and data requirements associated with model updates. A feature-importance analysis identified sea-ice concentration, ice volume and thermodynamic variables, particularly ice-layer enthalpies, as the most informative predictors of model error. Atmospheric forcing variables were found to contribute comparatively little when supplied as instantaneous inputs. This analysis enabled the identification of a reduced and more interpretable set of predictors for future hybrid implementations. The results reported above, encompassing the development and analysis of the hybrid Icepack framework, led to the publication of the article Hybrid physics–data-driven modeling for sea ice thermodynamics and transfer learning in the Quarterly Journal of the Royal Meteorological Society (DOI: 10.1002/qj.70200). Finally, in collaboration with CMCC, the methodology was extended to the two-dimensional C-GLORS reanalysis system. A U-Net architecture using partial convolutions was developed to learn sea-ice concentration and thickness analysis increments. The proposed architecture produced physically consistent corrections, avoided spurious increments in the open ocean, improved performance near coastlines and outperformed climatological corrections in terms of both absolute error and spatial correlation. This work established the basis for the future integration of machine-learning-based model-error corrections into an ocean–sea-ice data-assimilation cycle.Project details
Unibo Team Leader: Natale Alberto Carrassi
Unibo involved Department/s:
Dipartimento di Fisica e Astronomia "Augusto Righi"
Coordinator:
Politecnico di MILANO(Italy)
Total Unibo Contribution: Euro (EUR) 81.486,00
Project Duration in months: 24
Start Date:
28/09/2023
End Date:
28/02/2026
