Abstract
In the context of decadal climate predictions, a climate-mode initialization method is being tested by which ocean ORAS4 reanalysis is projected onto dominant modes of variability of the Earth System Model from the Max Planck Institute for Meteorology (MPI-ESM). The method aims to improve the prediction skill of the model by filtering out dynamically unbalanced noise during the initialization step. Used climate modes are calculated as statistical 3-D modes based on the bivariate empirical orthogonal function (EOF) analysis applied to temperature and salinity anomalies from an ensemble of historical simulations from the MPI-ESM. The climate-mode initialization method shows improved surface temperature skill, particularly over the tropical Pacific Ocean at seasonal-to-interannual timescales associated with the improved zonal momentum balance. There, the new initialization somewhat outperforms the surface temperature skill of the anomaly initialization also for lead years 2–5. In other parts of the world ocean, both initialization methods currently are equivalent in skill. However, only 44% of variance in the original ORAS4 reconstruction remains after the projection on model modes, suggesting that the ORAS4 modes are not fully compatible with the model modes. Moreover, we cannot dismiss the possibility that model modes are not sufficiently sampled with the data set underlying the EOF analysis. The full potential of the climate-mode initialization method for future decadal prediction systems therefore still needs to be quantified based on improved modal representation.
Highlights
How to optimally use available climate observations to generate initial conditions for climate predictions and how to best insert them into a coupled climate model with minimum loss of prediction skill remain active areas of research with many open questions to be answered (Balmaseda and Anderson 2009; Meehl et al 2009, 2014; Kirtman et al 2013; Boer et al 2016; Penny et al 2017)
The reanalysis variability that is not compatible with the climate modes from the model is filtered out, retaining the ocean states that serve as initial conditions for ensembles of decadal predictions
The value of 40% of variance explained in the reconstruction is rather low and raises the questions: will the remaining signal be sufficient for initializing decadal hindcasts, and why does the filtering process eliminate so much of ORAS4 variance? By fitting a logarithmic curve to the variance explained plotted versus number of empirical orthogonal function (EOF) considered, we find that using more members of historical simulations for the EOF analysis could increase the value of explained variance in the reconstructed data (Fig. 1b)
Summary
How to optimally use available climate observations to generate initial conditions for climate predictions and how to best insert them into a coupled climate model with minimum loss of prediction skill remain active areas of research with many open questions to be answered (Balmaseda and Anderson 2009; Meehl et al 2009, 2014; Kirtman et al 2013; Boer et al 2016; Penny et al 2017). From a theoretical point of view, it appears obvious that only a dynamically consistent assimilation approach for generating initial conditions applied to the same coupled model that is being used to perform the predictions can lead to a best prediction skill through a reduced initial model adjustment shock. A refinement of the anomaly initialization for decadal predictions was proposed recently (Volpi et al 2017), where the initial states were weighted with the ratio between the modeled and the observed variability to avoid initialization which goes beyond of the range of the model variability Still, all these initialization methods remain suboptimal when dealing with non-stationary errors including initial shocks (Goddard et al 2013; Magnusson et al 2013). The ocean reanalysis anomalies are projected onto a truncated set of the EOF-modes In this mapping step, the reanalysis variability that is not compatible with the climate modes from the model is filtered out, retaining the ocean states that serve as initial conditions for ensembles of decadal predictions. With a reference approach based on anomaly initialization. time slice 1958–2005 and 15 members of historical-simulaA discussion and concluding remarks are given in Sect. 4. tions ensemble (see Table 1)
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
