Multivariate surrogate model of sea ice in the Arctic region

The rapid changes occurring in Arctic sea ice influence climate and marine ecosystems, mid-latitude weather on timescales from weeks to months, and human activities, further motivating the need for accurate forecasts. A novel generation of sea ice models based on Elasto-Brittle rheologies, such as neXtSIM (Rampal et al, 2016), successfully represents sea-ice processes, with a remarkable accuracy at the mesoscale, for resolutions of about 10 km. However, these models are computationally expensive, limiting their practical application for long-term forecasting. To address this challenge, we leverage deep learning techniques to build an accurate and computationally affordable surrogate of the physical model.  

Following up from the initial work by Durand et al., 2024 on univariate surrogate of the sea-ice thickness (SIT) in neXtSIM, we present here a multivariate surrogate model designed to emulate simultaneously SIT, sea-ice concentration (SIC), and sea-ice velocities (SIU and SIV) in the Arctic region. As its core, our deterministic neural-network-based surrogate model uses a U-Net architecture, tailored to the sea-ice forecasting problem. The model is trained on reforecast-like data generated from neXtSIM and atmospheric forcings from ERA5, which help the model to better represent advective and thermodynamic processes. The neural network is trained to predict sea-ice fields with a 12-hour lead time, and it can iteratively be applied to extend predictions for up to a year. 

We thoroughly investigate the learning process, providing a detailed analysis of our choice of customized loss function and its optimal parameter values. In particular, we investigate the importance of each predicted variable and perform a feature sensitivity analysis. The forecast skills of our model have been successfully evaluated for lead times of up to one year, using both statistical and physical-dynamical metrics. Our preliminary results indicate that the model demonstrates good prediction capabilities at much lower computational costs than the original physical model. The application of a supervised deep learning approach to sea-ice modeling offers a promising alternative to traditional, computationally intensive methods. The positive results from our model's predictions underscore its potential as a reliable tool for seasonal sea ice forecasting. 

 

Rampal P. et al. “neXtSIM: a new Lagrangian sea ice model”. In: The Cryosphere 10.3 (2016), pp. 1055–1073 

Durand C. et al. “Data-driven surrogate modeling of high-resolution sea-ice thickness in the Arctic”. In: The Cryosphere 18.4 (2024), pp 1791-1815