Predictability of prominent modes of North Pacific subsurface temperature in CCSM3

We investigate decadal time-scale predictability of the North Pacific resulting from information residing in the initial state of forecasts. We do this by assessing how quickly this information is lost due to sensitivity of the system to uncertainties in the initial state through examination of a) the growth of small initial perturbations in three 40-member ensemble integrations of Community Climate Model version 3 (CCSM3) and b) properties of intrinsic variability in a 700 year control run of this model. Using relative entropy, we measure information that remains in forecasts at various ranges by comparing the distribution of predicted states to the climatological distribution of states. In this way we determine inherent limits on the ability to predict the evolution of the climate system.

So that the effects of weather noise are minimized we consider layer mean temperatures in the upper 300m, and to focus on signals that may be especially predictable we focus on the leading EOFs of intrinsic variability. The leading pattern of interannual variability (EOF1) in CCSM3 has a similar structure to the Pacific Decadal Oscillation (PDO) observed in nature. Though it has pronounced spectral peaks on decadal time-scales, the ensemble experiments indicate its predictability is less than six years. However, this pattern is one phase of a propagating mode of the system, which is captured when EOF1 and EOF2 are considered together. Physically, this mode corresponds to an eastward propagation of heat content anomalies in midlatitudes. The combined mode retains predictability for more than a decade. This predictability results from information contained in both the mean and the standard deviation of predicted distributions. Highly predictable events tend to be those with large initial projection onto EOF1 as this results in a mean signal that persists and evolves into EOF2, a pattern that is especially insensitive to small perturbations.

Further calculations demonstrate that the mode represented by EOF1 and EOF2 is not the most predictable component of North Pacific variability; other structures are less susceptible to initial perturbations. But this pair makes an important contribution to North Pacific predictability because it represents a large fraction of system variability and because it has a large influence on surface ocean conditions.

To generalize the conclusions drawn from the three ensemble experiments, we investigate predictability properties implied by the behavior of the control experiment. We do this by approximating the behavior of the control with a stochastic model known as a linear inverse model. From this exercise we conclude that the predictability limits found in the ensemble experiments are likely to hold more generally.