North Atlantic thermohaline circulation predictability in a coupled ocean-atmosphere model

Abstract

Predictability of the North Atlantic thermohaline circulation (THC) variability as simulated in the GFDL coupled ocean-atmosphere general circulation model is established for a set of ensemble experiments. The ensembles consist of identical oceanic initial conditions underneath a model atmosphere chosen randomly from the model climatology. This experimental design is based on the separation in time scales present in the model which motivates the assumption that the predictability deduced from these ensembles provides an upper limit to the model's THC predictability. The climatology is taken from a multi-century model integration whose THC variability has power concentrated at the 40-60 year time scale. A linear stochastic perspective is shown to be generally consistent with the ensemble statistics. The linear theory suggests a natural measure of ensemble predictability as the time at which the ensemble variance becomes a subjectively defined fraction (0.5 used here) of the climatological variance. It is furthermore of interest to distinguish predictability of the rapidly de-correlating portion of the model's THC from the longer time correlated portion. The rapidly de-correlating portion shows predictability for $\approx 1.5$ years. The slower portion shows predictability for $\approx 5-7$ years. The linear stochastic framework provides for the straightforward construction of an optimal forecast of the model's THC variability and an understanding of why the optimal forecast is useless beyond a particular limit.