Fast and slow responses of Southern Ocean sea surface temperature to SAM in coupled climate models

Abstract

We investigate how sea surface temperatures (SSTs) around Antarctica respond to the Southern Annular Mode (SAM) on multiple timescales. To that end we examine the relationship between SAM and SST within unperturbed preindustrial control simulations of coupled general circulation models (GCMs) included in the Climate Modeling Intercomparison Project phase 5 (CMIP5). We develop a technique to extract the response of the Southern Ocean SST (55°S–70°S) to a hypothetical step increase in the SAM index. We demonstrate that in many GCMs, the expected SST step response function is nonmonotonic in time. Following a shift to a positive SAM anomaly, an initial cooling regime can transition into surface warming around Antarctica. However, there are large differences across the CMIP5 ensemble. In some models the step response function never changes sign and cooling persists, while in other GCMs the SST anomaly crosses over from negative to positive values only 3 years after a step increase in the SAM. This intermodel diversity can be related to differences in the models’ climatological thermal ocean stratification in the region of seasonal sea ice around Antarctica. Exploiting this relationship, we use observational data for the time-mean meridional and vertical temperature gradients to constrain the real Southern Ocean response to SAM on fast and slow timescales.

Appendix: Verification of the methodology

We test our methodology from Sect. 2 in order to ascertain its reliability. Our verification procedure involves applying the regression algorithm to systems with a known prescribed step response function. The latter is convolved with a randomly generated order 1 autoregressive timeseries (AR(1)) that is 1000 years long and resembles a SAM forcing. The result of the convolution is our synthetic SST response, which is strongly diluted with a different AR(1) process characterized by longer memory. We choose parameters for the AR(1) models such that their autocorrelations resemble those of SAM and SO SST timeseries in the CMIP5 GCMs (for instance, Fig. 10a, c). We conduct multiple verification tests with different choices of AR(1) parameters. We also vary the signal to noise ratio in our synthetic SST. Figure 10b and d show examples from two different tests.

Fig. 10

Application of the regression algorithm to systems with a known prescribed step response function. On the top row we show a test case where we assume long memory in our SAM and SST signals. The SST signal is diluted such that 60 % of the variance is noise. In a on the left, we show the lagged autocorrelations of SAM and SST in CCSM4 (gray dashed curves) and our synthetic artificially generated signals (solid black curves). In b we show applications of the regression algorithm. The thick black curve is the true prescribed step response function. The thin gray curves and the vertical bars denote the estimated step response function $SST_{Step}(t)$ and the uncertainties $\sigma _{SSTstep}(t)$ produced by applying our regression algorithm. The two gray curves in panel b result from analyzing separate realizations in which we use the same prescribed step response and AR timeseries with the same statistical properties (illustrated in a) but different random values. On the bottom row we show a test case where we assume shorter memory in the SAM and SST signals, but the SST signal is diluted with more noise, such that the forced response contributes only 20 % of the total variance. c, d Are analogous to a, b

Within every test we generate an ensemble of multiple synthetic SAM and SST signals with the same statistical properties but different random values. We apply our algorithm separately to each realization in the same fashion as our analysis of CMIP5 control simulations. The verification tests confirm the validity of our method for estimating step response functions.