Coupled Data Assimilation with AWI-CM and PDAF

Sea surface temperature difference between model simulations and observations for the assimilation run and the free run after 10 and 30 days.

Assimilation of satellite SST data into a coupled ocean-atmosphere model AWI-CM with parallel data assimilation framework PDAF

Data assimilation (DA) combines the real-world observations with model predictions to provide a better estimate of the state for the system. Until now it has already been widely applied in the investigation of the ocean, the atmosphere, and so on. A coupled ocean-atmosphere model simulation the ocean, atmosphere and land surface component in one framework. DA for such a coupled system is challenging and can be divided into two groups: weakly coupled and strongly coupled data assimilation. Traditionally, DA analyses the ocean and the atmosphere separately and no cross-covariances are presented between the two compartments. Alternatively, strongly coupled DA is able to transfer information between the ocean and the atmosphere by using the cross-covariances. What we are currently doing and testing is the weakly coupled DA which applied DA for the ocean only. The coupled ocean-atmosphere model AWI-CM 1.4 consists of the ocean model FESOM including a sea-ice model, and the atmosphere model ECHAM6 including a land surface model JSBACH. The two compartments are coupled through the coupler OASIS3-MCT, which exchanges fluxes between the two compartments. The parallel data assimilation framework PDAF is used in this work. PDAF is a software for ensemble forecasting and sequential data assimilation. It provides fully implemented and parallelized assimilation methods.

A global model domain is used in this study. For the atmosphere model ECHAM6 a resolution of T63 with 47 layers was adopted and the ocean model FESOM used unstructured grids with varying resolution between 30km and 160km. For the data assimilation experiments, LESTKF were applied by assimilating the satellite sea surface temperature (SST) into the ocean variables with a daily update frequency. The ensemble size was set up to 46. The simulation/assimilation period was one year. The computation time was 12 hours on the supercomputer HLRN. The satellite SST observations have a high spatial resolution of 0.1oX0.1o and a temporal resolution of one day. It covers most of the global regions except the polar regions. Observations at the sea-ice edge were excluded by removing the SST observations where the model prediction had sea-ice and no DA was done for the sea-ice grid points.

Results were evaluated using the root mean square error (RMSE) of the SST by comparing the difference between simulation and observations. for scenarios with and without DA after 10 and 30 days. The reduction of the RMSE was up to 50% for DA scenarios after 30 days. The current issue we have during the DA experiments were the extremely low ocean temperatures in deeper layers (up to -8oC), which can be due to the effect of the large deviation of the SST at the initial time step and the analysis step into deeper layers by the cross-covariance. A possible solution can be reducing the vertical assimilation effect by implementing vertical localization.

In summary, in this study we implemented weakly coupled DA by assimilation of SST into ocean variables. This can be used as a prototype implementation for other coupled models. A next step would be the assimilation of subsurface temperature and salinity observations into the ocean variables and the atmosphere observations into atmosphere variables. A final step can be the strongly coupled DA for the coupled model, which assimilates the ocean and atmosphere observations into both the ocean and the atmosphere variables directly using the cross-covariances.