Data assimilation into sea ice models designed for climate studies has started about 15 years ago. In most of the studies conducted so far, it is assumed that the improvement brought by the assimilation is straightforward. However, some studies suggest this might not be true. In order to elucidate this question and to find an appropriate way to further assimilate sea ice concentration and velocity observations into large-scale sea ice models, we analyze here results from a number of twin experiments (i.e. experiments in which the assimilated data are model outputs) carried out with first a simplified model of the Arctic sea ice pack then with NEMO-LIM2, a primitive equation ocean general circulation model coupled to LIM (Louvain-la-Neuve sea ice model). Our objective is to determine to what degree the assimilation of ice velocity and/or concentration data improves the global performance of the model and, more specifically, reduces the error in the computed ice thickness. A simple scheme is used, and outputs from a control run and from perturbed experiments without and with data assimilation are thoroughly compared. Our results indicate that, under certain conditions depending on the assimilation weights and the type of model error, the assimilation of ice velocity data enhances the model performance. The assimilation of ice concentration data also helps in improving the model results, but it has to be handled with care because of the strong link between ice concentration and ice thickness. Therefore, we show that one should conserve the ice thickness (not the ice volume) when ice concentration data are assimilated into the model. We also demonstrate that one should assimilate sea ice concentration and velocity data simultaneously. Finally, we give some concrete keys in order to choose which observational data set to assimilate.

Data assimilation into sea ice models designed for climate studies has started about 15 years ago. In most of the studies conducted so far, it is assumed that the improvement brought by the assimilation is straightforward. However, some studies suggest this might not be true. In order to elucidate this question and to find an appropriate way to further assimilate sea ice concentration and velocity observations into large-scale sea ice models, we analyze here results from a number of twin experiments (i.e. experiments in which the assimilated data are model outputs) carried out with first a simplified model of the Arctic sea ice pack then with NEMO-LIM2, a primitive equation ocean general circulation model coupled to LIM (Louvain-la-Neuve sea ice model). Our objective is to determine to what degree the assimilation of ice velocity and/or concentration data improves the global performance of the model and, more specifically, reduces the error in the computed ice thickness. A simple scheme is used, and outputs from a control run and from perturbed experiments without and with data assimilation are thoroughly compared. Our results indicate that, under certain conditions depending on the assimilation weights and the type of model error, the assimilation of ice velocity data enhances the model performance. The assimilation of ice concentration data also helps in improving the model results, but it has to be handled with care because of the strong link between ice concentration and ice thickness. Therefore, we show that one should conserve the ice thickness (not the ice volume) when ice concentration data are assimilated into the model. We also demonstrate that one should assimilate sea ice concentration and velocity data simultaneously. Finally, we give some concrete keys in order to choose which observational data set to assimilate.