Decadal Prediction Research at NOAA/GFDL

This paper is concerned with whether climate projections for the next several decades may be enhanced when models are initialized from the observed state of the climate system. Climate projections are started from an arbitrary point in time of a control integration. They are forced by specifying concentrations of greenhouse gases. The model is then integrated into the present and future. However, additional predictability may be obtained due to the low frequency variability of the climate system, especially from the ocean. In order to estimate the state of the climate system an initialization procedure that combines observations with the model needs to be in place. The GFDL Ensemble Coupled Data Assimilation estimates the temporally evolving probability distribution of climate states under an observational constraint. From this system a reanalysis from 1971-2010 is produced and is used to generate initial conditions. Using GFDL's CM2.1 CGCM 10 member ensembles of hindcasts and forecasts starting in January for every year 1971-2009 are run for 10 years.The predictions use the RCP4.5 scenario. Contrasting the uninitialized predictions with the initialized ones gives us an estimate of the role the internal variability may have in additional predictability of the climate system on decadal time scales. Our focus will be mainly on the prediction skill of the ocean. These predictions must be considered in the presence of model error, prediction uncertainty, projection uncertainty, and observational uncertainty.