An anatomy of Arctic sea ice forecast biases in the seasonal prediction system with EC-Earth

The quality of initial conditions (ICs) in climate predictions controls the level of skill. Both the use of the latest high-quality observations and of the most efficient assimilation method are of paramount importance. Technical challenges make it frequent to assimilate observational information independently in the various model components. Inconsistencies between the ICs obtained for the different model components can cause initialization shocks. In this study, we identify and quantify the contribution of the ICs inconsistency relative to the model inherent bias (in which the Arctic is generally too warm) to the development of sea ice concentration forecast biases in a seasonal prediction system with the EC-Earth general circulation model. We estimate that the ICs inconsistency dominates the development of forecast biases for as long as the first 24 (19) days of the forecasts initialized in May (November), while the development of model inherent bias dominates afterwards. The effect of ICs inconsistency is stronger in the Greenland Sea, in particular in November, and mostly associated to a mismatch between the sea ice and ocean ICs. In both May and November, the ICs inconsistency between the ocean and sea ice leads to sea ice melting, but it happens in November (May) in a context of sea ice expansion (shrinking). The ICs inconsistency tend to postpone (accelerate) the November (May) sea ice freezing (melting). Our findings suggest that the ICs inconsistency might have a larger impact than previously suspected. Detecting and filtering out this signal requires the use of high frequency data.

Automated summary

The quality of initial conditions is crucial in climate predictions. Inconsistencies between different components can cause initialization shocks, impacting the accuracy of forecasts, especially for sea ice concentration. The dominance of ICs inconsistency in forecast biases suggests a larger impact than previously thought, requiring the use of high frequency data for detection and filtering.