On emulating Sea Ice in the Finite Element Sea Ice-Ocean Model (FESOM)

In recent years, Artificial Intelligence (AI) has been a game-changer in climate modeling, providing innovative and adaptable approaches while improving accuracy and computational efficiency. For instance, hybrid models can preserve the robustness of physical modeling while emulating components that are computationally expensive or poorly represented. The relatively two-dimensional and localized nature of sea ice makes it an ideal candidate for AI-based emulation, offering a solution to the significant computational burden it imposes on ocean models. Here, we present a sea ice emulator for the Finite Element Sea Ice-Ocean Model (FESOM), capable of predicting the evolution of sea ice thickness (SIT), concentration (SIC), and drift (SID) on timescales ranging from weeks to months.

First, an adaptive U-Net-based model is trained to predict sea ice state (SIT, SIC, SID) increments at one (or multiple) lead times ahead, using corresponding atmospheric forcing and past and current sea ice states. The model is driven by multi-decadal series of daily to sub-daily atmospheric forcing and 2D sea ice and ocean outputs from FESOM, which have been preprocessed and re-interpolated onto a regular grid. To ensure scalability, training sequences are divided into chunks, managed by a custom mapper that balances their usage during training and supports compatibility with multi-GPU configurations. The model is trained by minimizing a penalized mean square error loss function, with an adaptive learning rate controlled via a dedicated scheduler, until convergence. The quality and accuracy of the training process are systematically assessed prior to inference.

Emulation of sea ice can then be performed using recursive inference of the trained models for rollouts spanning from some weeks to a year. Subsequent sea ice states are occasionally clipped into their physical range in order to prevent non-physical behaviors. Rolling predictions can be eventually generated daily or weekly along the test sequences, similar to operational forecasting.

Apart from SIT, SIC, and SID maps, metrics including Integrated Ice Edge Error, root mean square error, mean ice thickness, and Sea Ice Extent are implemented to evaluate the quality of the prediction in comparison to the actual FESOM outputs and some predefined baselines. The emulator demonstrates robust predictions up to 100 days, while still maintaining a realistic representation of various sea ice states beyond this time. Both training and inference are scalable and have been deployed on GPUs, although rolling predictions can be run on a single CPU without incurring prohibitive costs. Computation times for both steps will be estimated, along with the time required for a standard FESOM simulation including sea ice, to assess the potential gain in computational efficiency.