Upper-Ocean Singular Vectors of the North Atlantic Climate with Implications for Linear Predictability and Variability

Abstract

The limits of predictability of the meridional overturning circulation (MOC) and upper-ocean temperatures due to errors in ocean initial conditions and model parametrizations are investigated in an idealized configuration of an ocean general circulation model (GCM). Singular vectors (optimal perturbations) are calculated using the GCM, its tangent linear and adjoint models to determine an upper bound on the predictability of North Atlantic climate. The maximum growth time-scales of MOC and upper-ocean temperature anomalies, excited by the singular vectors, are 18.5 and 13 years respectively and in part explained by the westward propagation of upper-ocean anomalies against the mean flow. As a result of the linear interference of non-orthogonal eigenmodes of the non-normal dynamics, the ocean dynamics are found to actively participate in the significant growth of the anomalies. An initial density perturbation of merely \(0.02 kg m^{−3}\) is found to lead to a 1.7 Sv MOC anomaly after 18.5 years. In addition, Northern Hemisphere upper-ocean temperature perturbations can be amplified by a factor of 2 after 13 years. The growth of upper-ocean temperature and MOC anomalies is slower and weaker when excited by the upper-ocean singular vectors than when the deep ocean is perturbed. This leads to the conclusion that predictability experiments perturbing only the atmospheric initial state may overestimate the predictability time. Interestingly, optimal MOC and upper-ocean temperature excitations are only weakly correlated, thus limiting the utility of SST observations to infer MOC variability. The excitation of anomalies in this model might have a crucial impact on the variability and predictability of Atlantic climate. The limit of predictability of the MOC is found to be different from that of the upper-ocean heat content, emphasizing that errors in ocean initial conditions will affect various measures differently and such uncertainties should be carefully considered in decadal prediction experiments.