Enhancing Skill of Initialized Decadal Predictions Using a Dynamic Model of Drift

Since variability of present-day climate on the interannual to decadal timescale (near term) is influenced by both internal (natural) variability and external forcing, predictions on this timescale are controlled by both initial condition (IC) predictability and boundary condition predictability. Consequently, initialized decadal prediction experiments (cf. CMIPx) aim to augment the external-forcing related predictability realized in uninitialized projections with natural-variability related predictability by appropriate observation-based initialization. However, and notwithstanding the considerable effort expended in finding such “good” initial states, a striking feature of current, state-of-the-art, initialized decadal hindcasts is their tendency to quickly drift away from the intialized state. A primary reason for such drift is related to the fact that the climate system loses memory of its IC and transitions from a regime of IC-related (equivalently natural-variability related) predictability to forcing-related predictability. When such a transition occurs in a model that is biased, initialization-drift is exacerbated and the initialized prediction trajectories appear to almost jump away from the initialized state. This behavior is observed in numerous models of the CMIP class. Given this state of affairs, post-processing of initialized predictions is essential before they can be skilful.

After developing a dynamical model of initialization drift and validating it, we show that a post-processing methodology based on the model is able to significantly out-perform state-of-the-art techniques of post-processing in terms of enhancing skill of initialized predictions. We also show that the new methodology provides further crucial insights into issues related to initialized-predictions.