1,1,1,1- 1NERSC - Nansen Environmental and Remote Sensing Center, Bergen, Norway
- 2Development Centre for Weather Forecasting, Norwegian Meteorological Institute, Oslo, Norway
We present IceCastNet, a data-driven model for forecasting Arctic sea ice state (including concentration, thickness, and drift) up to 10 days ahead. The model is trained on satellite-derived products: OSI-SAF for sea ice concentration and drift, and CS2SMOS for sea ice thickness. During the melt season, when CS2SMOS data are unavailable, the dataset is supplemented with the TOPAZ reanalysis. IceCastNet also uses meteorological forecasts (10-meter wind and 2-meter air temperature) from ECMWF as forcing.
The architecture is based on a graph-transformer design, similar to that used in the Artificial Intelligence/Integrated Forecasting System (AIFS), and implemented within ECMWF’s Anemoi framework. IceCastNet achieves skills comparable to, and in some cases better than, established baselines such as the operational TOPAZ system, particularly for sea ice concentration. These improvements appear to stem from reduced biases in initial conditions and lower forecast error. This is assessed by comparing IceCastNet outputs with debiased TOPAZ forecasts and independent sea ice concentration products, including SAR-derived estimates from the Danish Meteorological Institute (ASIP) and ice charts produced by experts at the U.S. National Ice Center.
The inference time for a 10-day forecast with IceCastNet of about 10 seconds is approximately two orders of magnitude shorter than that of physics-based systems such as TOPAZ. This substantial reduction in computational cost makes IceCastNet a computationally efficient alternative, although IceCastNet only provides the observed variables. Moreover, since it relies exclusively on operational, near-real-time data, IceCastNet is well-suited for integration into operational sea ice forecasting workflows.
The spatial resolution of IceCastNet forecasts follows that of the training data. We also show that applying a super-resolution procedure trained on high-resolution sea ice simulations from the model neXtSIM can enhance the resolution of IceCastNet outputs.
References:
CS2SMOS: https://doi.org/10.5194/tc-11-1607-2017
OSI-SAF: https://osi-saf.eumetsat.int/
AIFS: https://arxiv.org/abs/2406.01465
TOPAZ: https://doi.org/10.48670/moi-00001
ASIP: https://ocean.dmi.dk/asip/
neXtSIM: https://egusphere.copernicus.org/preprints/2025/egusphere-2024-3521/
How to cite: Brajard, J., Edel, L., Palerme, C., Korosov, A., and Bertino, L.: End-to-end forecast of the Arctic sea ice initialised directly from observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6436, https://doi.org/10.5194/egusphere-egu26-6436, 2026.
