Several authors have suggested that much of the decadal variability in North Atlantic sea surface temperature (SST) can be explained as a local oceanic response to atmospheric variability1,3,13, whereas others have argued that non-local dynamical processes in the ocean must play an important role2,4,14. In the latter group, Hansen and Bezdek4 have recently presented evidence of advection of SST anomalies, especially along the path of the North Atlantic Current (NAC), an extension of the Gulf Stream. The importance of this suggestion is that fluctuations in SST that arise from advection may be predictable. We use analyses of monthly mean SST compiled from ship observations for the period 1945–8915 to investigate the evidence for predictability in North Atlantic SST associated with oceanic advection, and examine whether advection could be an important element in a possible coupled ocean–atmosphere mode of North Atlantic variability7.

Figure 1a shows an estimate of the upper-ocean circulation in the North Atlantic16. Figure 1b shows the correlation between low-frequency fluctuations in local wintertime SST and low-frequency fluctuations in wintertime SST averaged over the region 80–60° W, 31.5–38.5° N (the vicinity of Cape Hatteras) as a function of lag in years. SST anomalies are most likely to express deeper subsurface anomalies in winter, when the oceanic mixed layer is at its deepest. Coherent SST fluctuations appear to propagate from the coast of North America across the Atlantic to the northwest of Scotland, following the strong SST gradients that are associated with the Gulf Stream and NAC.

Figure 1: a, The upper-ocean circulation in the North Atlantic according to ref. 16.
figure 1

The lines indicate the mean flow of waters warmer than 7 °C, and the circles numbers indicate the transport in sverdrups (1 sverdrup ≡ 106 m3 s−1). The main feature is the anticyclonic Subtropical Gyre, shown in solid lines. The cyclonic Subpolar Gyre lies to the north and is shown by dashed lines. The Gulf Stream lies at the western edge of the Subtropical Gyre. It flows polewards along the coast of North America before leaving the coast at Cape Hatteras. In mid-ocean the Gulf Stream splits into a part which recirculates southwards and another part—the North Atlantic Current—which flows northwards, and then northeastwards towards Scotland. b, Correlation between low-frequency fluctuations in local wintertime SST and low-frequency fluctuations in wintertime SST averaged over the region 80–60° W, 31.5–38.5° N (the vicinity of Cape Hatteras, VCH) as a function of lag. The contours pick out the regions where lag-correlation with VCH is maximized. The numbers next to each contour indicate the lag in years. In all cases VCH SST leads local SST. The contour value is 0.8 for lags of 0–8 years and 0.75 for the lag of 9 years. The contours are superimposed on the SST field averaged over all winters between 1945 and 1989 (colour scale in °C). Monthly mean SST data were obtained from the NOAA SMD9415 analysis of ship observations for the period 1945–89 on a 1° × 1° grid. The long-term mean annual cycle was removed from all the data. Winter-mean anomalies for each year are then computed by averaging the SST anomalies for the months November–April, and a five-year running mean is applied to the winter-mean anomalies. Correlations are only computed in the latitude band 25–65° N.

In Fig. 2 we show the propagating SST anomalies that give rise to the correlations seen in Fig. 1b . A series of warm and cold anomalies propagates downstream with an average velocity of 1.7 cm s−1. The along-path length scale of these anomalies is 2,000 km and their amplitude generally lies in the range 0.5–1.0 °C, reaching a maximum between 5,000 and 6,000 km downstream of the tip of Florida at approximately 35° W, 50° N. As noted by Hansen and Bezdek4, this propagation speed is considerably slower than the near-surface currents in the core of the Gulf Stream (>1 m s−1 close to Cape Hatteras17). If, as seems likely, these signals are advective their speed must be determined by the weaker currents either to one side of the core or at some depth below the surface. Subsurface observations close to the path of the Gulf Stream/NAC show evidence of decadal temperature fluctuations, possibly extending below the seasonal thermocline6,13,18, and subsurface propagation in the Subpolar Gyre has also been reported19,20. The existence of a deep signal, away from the regions of maximum shear and surface influence, may help to explain the remarkable coherence of the surface signature.

Figure 2: a, SST anomalies (colour scale in °C) as a function of time and distance along the path of the Gulf Stream/North Atlantic Current.
figure 2

Winter (November–April) anomalies only are used, smoothed with a three-year running mean. The path along which the plot is computed (shown in b as a solid black line) was taken from ref. 16 (Fig. 1a). The horizontal axis gives distance in thousands of kilometres measured along the path from the start point at 78.5° W, 24.5° N near the tip of Florida. b, The fraction of the low-frequency variability (root-mean-squared anomalies) in winter-mean sea-level pressure that can be accounted for by a linear response to SST in the storm-formation region (SFR; 82–69° W, 25–35° N, marked with a letter A). Positive (negative) values indicate fluctuations (anti-) correlated with SFR SST. In c, the solid line is a time series of winter-mean SST in region SFR, and the vertical scale is in °C. The dashed line shows winter-mean sea-level pressure in the region 50–30° W, 50–60° N (letter B in panel b), multiplied by −1. The dotted line shows winter-mean sea-level pressure in the region 60–40° W, 15–25° N (letter C in panel b). Both the sea-level pressure times series are renormalized to have approximately the same variance as the time series of SST, and all three time series are detrended and smoothed with a three-year running mean. Before performing the regression for panel b both the SST and sea-level pressure time series are binned into 2-year winter means. The SFR SST time series only is further smoothed with a 3-point running mean. Sea-level pressure at any given location is then regressed on SFR SST, a constant term, and a trend, with the residual modelled as an AR(1) (‘red noise’) process26. Values not significant at the 1.5 standard deviation level are not plotted.

The SST power spectrum averaged along the Gulf Stream/NAC path exhibits a spectral peak at 12–14 years which exceeds the 98th percentile of the red-noise background that is expected if SST fluctuations arise as a local response to atmospheric variability21. In addition to this 12–14-year timescale, the prominence of the significant warm event which began in the mid-1940s and the significant cold event which began in the late 1960s suggests that lower-frequency, multidecadal, fluctuations2 may also be associated with propagation along the Gulf Stream/NAC. Both timescales contribute to the correlations observed in Fig. 1b , and we stress the difficulty of distinguishing, with a short data record, a multidecadal timescale from amplitude modulation of the decadal signals.

The 12–14-year timescale which we have identified with advection along the Gulf Stream/NAC agrees with that which previous authors1,7 have associated with a hypothetical coupled ocean–atmosphere mode. For any such coupled mode to exist, there must be a consistent atmospheric response to SST anomalies. How the atmosphere responds to mid-latitude SST is poorly understood7,8,9,10, but changes in the frequency, intensity or spatial distribution of mid-latitude storms are likely to be important8,10. We may, therefore, expect particular sensitivity to SST anomalies in regions where these storms form. An especially important area is located along the southeastern coast of the USA between Cape Hatteras and Florida22 (which we call the ‘storm-formation region’) where severe wintertime land/sea temperature contrast fuels storm development through baroclinic instability. Figure 2 shows that SST in this region exhibits the decadal-timescale fluctuation, and furthermore that there is a strong correlation between these SSTs and a dipole-like fluctuation in North Atlantic sea-level pressure similar to that discussed in ref. 1. A warming of SST off the US coast, which we would expect to intensify the land/sea contrast and may enhance evaporation, is associated with a strengthening of the mean westerly winds in the mid-Atlantic especially in the region 50–30° W, 30–50° N. Although causality cannot reliably be inferred from correlation alone, this relationship is in line with the theoretical response of the mean flow to an intensified storm track23 and therefore supports a hypothesis that the sea-level pressure dipole is a response to SST in the storm-formation region.

Increased storminess and/or mean winds will cool SSTs through local mixed layer and Ekman layer processes. Wind and net surface heat flux fields (not shown) indicate that cooling of SSTs in the vicinity of the NAC (40–55° N) and in the tropical North Atlantic (5–20° N) is associated with warm SST in the storm-formation region. Thus fluctuations in SST off the southeastern coast of the USA may affect remote regions of the Atlantic both quasi-instantaneously, via this atmospheric teleconnection, and after a delay of some years via advective propagation through the ocean.

We examine how oceanic advection and atmospheric teleconnections may combine to determine the decadal cycle in SST by regressing SST and sea-level pressure at various lags onto SST in the storm-formation region (Fig. 3). The analysis is designed (see Fig. 3 legend) to focus on the decadal timescale. The regression map at lag zero shows that warm SST in the storm formation region is associated with cold SST in the NAC region and in the tropical North Atlantic. These cold SSTs arise largely as the local SST response to heat-flux anomalies1. We suggest that these heat flux anomalies arise as part of the atmospheric response to SST in the storm-formation region. At lags of 2 and 4 years, consistent with the reduced amplitude of SST anomalies in this region, the sea-level pressure dipole is much weaker than at lag 0. The SST regression map evolves in a manner broadly consistent with our expectations of advection (Fig. 1a). The warm anomaly in the storm-formation region moves eastwards and then appears to split into two branches over the Mid-Atlantic Ridge, in good agreement with the bifurcation of the Gulf Stream indicated in Fig. 1a . The cold anomaly in the North Atlantic moves eastwards along the path of the NAC. At a lag of 4 years we see the development of cold anomalies in the Gulf of Mexico which, at a lag of 6 years, spread into the storm-formation region. The development of cold anomalies in the latter region stimulates, according to our hypothesis, an atmospheric response of the opposite sign to that with which we began. This atmospheric response forces warm anomalies in the NAC region and in the tropical North Atlantic, developing a pattern which is 180° out of phase with that seen at lag zero.

Figure 3: The fraction of low-frequency local SST variability (grey-scale) and low-frequency local sea-level pressure variability (contours accounted for by a lagged version of the low-frequency SST variability in the storm-formation region, SFR; see Fig. 2 legend.
figure 3

SST in region SFR leads local SST, and local sea-level pressure, by the number of years indicated on each panel. The regression analysis is identical to that described in Fig. 2 legend but with winter-mean SST replacing winter-mean sea-level pressure for the grey-scale maps, and with varying lag. The contour levels shown for the sea-level pressure regression are ±0.5 (thin lines) and ±0.65 (thick lines). Dashed lines indicate negative correlations. In c, the fraction of sea-level pressure variability accounted for fails to reach the 0.5 level, so no contours appear.

The explanation for a preferred timescale is likely to involve feedbacks to SST in the storm-formation region, mediated by the quasi-instantaneous atmospheric response and by subsequent oceanic processes that introduce a lag. The effects of the atmospheric fluctuations on the oceanic mixed layer in the tropical Atlantic, and in the vicinity of the Gulf Stream front (where ventilation may be affected), could be fed back to the storm-formation region by advection. Alternatively, fluctuations in the wind-stress curl over the Subtropical Gyre could generate fluctuations in the strength of the gyre circulation24. These fluctuations could also influence SST in the storm-formation region, but in this case the dominant timescale is associated with the propagation of baroclinic Rossby waves. The fact that anomalies arise in the Gulf of Mexico before their arrival in the storm-formation region is consistent with both mechanisms. In reality it is likely that both advection and wave propagation influence the timescale7; external forcing might also play a role25.

In the NAC region and in the tropical North Atlantic, interference between the instantaneous atmospheric influence and the delayed advective influence of SST in the storm-formation region could also favour the emergence of a preferred timescale31, tending to suppress (amplify) variability on timescales that are odd (even) multiples of the 6-year (Fig. 1b) advection time from the storm-formation region to these regions. Both the instantaneous and delayed signals will have contributed to the correlations seen in Fig. 1b . We note that our account of the decadal variability in tropical North Atlantic SST26 is an alternative to the hypothesis that this region is under the control of processes local to the tropical Atlantic27,28.

By exploiting the memory inherent in the decadal fluctuations, our results suggest it may be possible to predict a significant fraction of the low-frequency variability in North Atlantic SST and sea-level pressure several years in advance. Although further work will be needed before a serious attempt at forecasting can be made, particularly investigation with models of the mechanisms we have discussed, we consider briefly the outlook for the decadal fluctuations. We have examined the most recent evolution of North Atlantic SST in the NOAA NCEP analyses29 and of North Atlantic sea-level pressure in the NCEP reanalyses30. SST in the storm-formation region reached a maximum in the late 1980s and has since been falling. In line with our expectations of the atmospheric response, sea-level pressure in the region marked C in Fig. 2b peaked at the end of the 1980s and has since been falling, whereas sea-level pressure in the region marked B reached a minimum at the same time and has since been rising. Also consistent with our understanding, SST in the mid-Atlantic at 50° N reached a minimum in the late 1980s and has since been rising. Similar, if somewhat less clear, behaviour is seen in tropical North Atlantic SST. Based on the apparent period of 12–14 years, and to the extent that other influences (for example, El Niño and multidecadal fluctuations) allow, we expect all these trends to continue into the next century before reversing.