Multicentennial variability of the Atlantic meridional overturning circulation and its climatic influence in a 4000 year simulation of the GFDL CM2.1 climate model


Corresponding author: T. L. Delworth, Geophysical Fluid Dynamics Laboratory, NOAA, 201 Forrestal Rd., Princeton, NJ 08542, USA. (


[1] We investigate decadal to multicentennial variability of Northern Hemisphere surface air temperature in a 4000-year control simulation of the GFDL CM2.1 climate model. Spectral analysis shows the presence of a distinct multicentennial timescale of temperature variability. The associated spatial pattern is broad, covering the entire Northern Hemisphere extratropics, but with enhanced amplitude in the Atlantic and Arctic sectors. This variability appears to be driven by interhemispheric fluctuations in oceanic heat transport associated with the Atlantic Meridional Overturning Circulation (AMOC). The AMOC variability is associated with century-scale propagation of salinity anomalies from the Southern Ocean to the subpolar North Atlantic, with out of phase transport variations between the upper ocean and deeper layers of the Atlantic. When positive (negative) upper ocean salinity anomalies reach the subpolar North Atlantic they strengthen (weaken) the AMOC by modulating upper ocean density and vertical stratification. The large-scale warming also appears to be enhanced by reductions in surface albedo associated with reduced sea-ice and low-level cloudiness, thereby increasing the absorption of shortwave radiation and amplifying the warming from AMOC changes. We speculate that such multicentennial variations in the AMOC could contribute to long-time scale climate fluctuations in the observed paleo record. This could arise purely as internal variability of the climate system, or through radiatively-induced changes to atmospheric circulation patterns, such as the NAO, that would in turn influence the AMOC.

1. Introduction

[2] A crucial goal in climate research is to better understand the mechanisms of observed climate variability and change, leading to an ability to assess the relative roles of internal variability and radiative forcing variations in explaining observed changes in the climate system. It is difficult to obtain an estimate of internal variability purely from the observational record, since observed climate variations arise both from internal variability and from changing radiative forcing agents, such as solar irradiance and volcanic eruptions. The analysis of instrumental and proxy climate records has revealed important characteristics of decadal to millennial climate variability [see, e.g., Delworth and Mann, 2000; Sicre et al., 2008; Jones et al., 2009; Mann et al., 2009], but identifying the underlying cause of these variations is difficult. To provide a further perspective on low frequency climate variations that could arise solely from internal variability, we can use the output from long control simulations of climate models as a surrogate for the climate system. The analysis of these simulations is extremely useful in characterizing the internal variability in the climate system in the absence of external forcing changes. Such perspectives are limited by the fidelity of the climate model.

[3] Here we analyze the output from a 4000-year control simulation of a climate model (GFDL CM2.1 [Delworth et al., 2006]) to assess decadal to multicentennial simulated climate variability that arises purely from internal variability of the coupled climate system. A number of papers have shown that the CM2.1 model provides a very credible perspective on internal variability of the global climate system (see, e.g., Reichler and Kim [2008] for overall climate simulation; Russell et al. [2006] for Southern Ocean analysis; Schneider et al. [2007] for Atlantic analysis), thereby demonstrating that this model is an appropriate tool for such a study. In addition, the length of this control simulation allows a robust assessment of decadal to multicentennial scale variability in this model.

2. Model Description and Experimental Design

[4] The model used is described fully in Delworth et al. [2006]. The coupled model consists of atmosphere, ocean, land, and sea ice component models. The horizontal resolution of the atmospheric model is 2.5° longitude by 2.0° latitude, with 24 levels in the vertical. The horizontal resolution of the ocean model is 1° in the extratropics, with meridional grid-spacing in the Tropics gradually reducing to 1/3° near the Equator. The ocean model has 50 levels in the vertical, with 22 evenly spaced levels over the top 220 m. The model does not employ flux adjustments. For further information and model output see

[5] We make use of a 4000-year control experiment in which atmospheric constituents are held constant at 1860 conditions. Output from this integration is used to provide a statistical description of unforced, internal variability in the model.

3. Temperature and Precipitation Signals of Multicentennial Variability

[6] We show in Figure 1 the time series of Northern Hemisphere extratropical surface air temperature (NHESAT), averaged over the domain 20°N–90°N, as well as its spectrum. The model is still drifting over the first several centuries of the simulation, after which it comes slowly into a statistical steady state. In subsequent analyses we will remove much of this drift through the use of filtering that removes most of the variations on time scales longer than 700 years (although spectral analysis is performed on unfiltered data).

Figure 1.

(a) Time series of Northern Hemisphere extratropical surface air temperature (NHESAT; averaged over 20°N–90°N). Thin black line is annual mean, while the thick red line shows results after applying a 100-year low pass filter. (b) Spectra of annual mean surface air temperature averaged over 20°N–90°N (black line) and 20°S–20°N (blue line). Periods are listed along the bottom in years. In order to estimate the significance of the spectral peaks, a first order Markov process was estimated for each time series based on their respective lag-one autocorrelations, and then 95% confidence intervals around those spectra were calculated; the thin solid (dashed) red line indicates the 95% confidence level for the spectrum of the 20°N–90°N (20°S–20°N ) temperature time series.

[7] Spectral analysis of NHESAT, shown by the black line in Figure 1b, shows a generally red spectrum, with greatest variance at longer time scales. There are several time scales at which the spectral values rise above a red noise background. One peak occurs at approximately 30 years, while additional peaks occur on centennial to multicentennial time scales, with largest values over a broad time scale from approximately 200 to 500 years. For comparison purposes we also include in Figure 1b the spectrum of simulated surface air temperature averaged over the deep tropics (20°S–20°N, shown in blue). A comparison of the extratropics to the tropics is striking. The tropical spectrum is clearly dominated by ENSO, with dominant variance on timescales of 3 to 7 years, and weak variability at multidecadal to centennial time scales. These results suggest that internal climate variability on interdecadal to multicentennial time scales originates in the extratropics, and not the tropics.

[8] We focus the rest of our analyses on the multicentennial variability. We first compute linear correlations between the time series of annual mean SAT at each grid point and the time series of NHESAT. In order to isolate the multicentennial variability, we apply a 200 to 700 year bandpass filter to the model output before computing the regression coefficients. The resultant maps of correlation coefficients are shown in Figure 2a. The correlation coefficients are generally larger at higher latitudes, where albedo feedbacks are strong, and in the Atlantic sector. There is a seasonal dependence (not shown), with larger correlations over the Arctic and Eurasia in the winter season. As is the case with temperature changes due to increasing greenhouse gases, the Arctic signal is weakest in summer (not shown), as the near-surface temperature is constrained to be close to the freezing point of water.

Figure 2.

(a) Correlation of surface air temperature at each grid point with the time series of Northern Hemisphere mean surface air temperature (20°N–90°N). All time series were band-pass filtered to retain time scales between 200 and 700 years. Areas without stippling are statistically significant at the 95% level (using a two-sided Students t test, assuming 20 degrees of freedom in each time series; this is estimated from a 4000 year time series subjected to a 200 year low-pass filter). (b) Same as Figure 2a but for annual mean precipitation.

[9] Shown in Figure 2b are the correlation coefficients between annual mean precipitation at each grid point and the time series of NHESAT. There is enhanced precipitation in parts of the Arctic and North Atlantic associated with the warmer temperatures. There is also a tendency for a northward shift of the Intertropical Convergence Zone (ITCZ) and associated rainfall, such as over the Sahel of Africa. Conversely, northern regions of South America have reduced rainfall, also associated with the anomalous northward movement of the ITCZ. These changes are consistent with paleorecords of changes in the South American monsoon in relation to Atlantic ocean temperature [Bird et al., 2011].

4. Oceanic Drivers of Multicentennial Variability

[10] We now turn to the question of what drives this multicentennial variability. In the absence of externally imposed radiative forcing changes, we look to sources of internal variability in the climate system. We first examine the spectra of meridional ocean heat transport at 20°N, shown in Figure 3a for the Atlantic and the Indian-Pacific sectors separately. The differences in these spectra are striking. For the Indian-Pacific sector the ocean heat transport fluctuations are concentrated on relatively short time scales, less than a decade, likely associated with changes in wind-induced gyre circulations. This spectrum suggests that Indian-Pacific ocean heat transport changes do not contribute to the multidecadal to multicentennial fluctuations of NHESAT.

Figure 3.

(a) Spectra of ocean heat transport at 20°N in Atlantic basin (black) and Indo-Pacific (blue). Periods are listed along the top and bottom of the spectrum. The spectrum is plotted on a log-log axis. In order to estimate the significance of the spectral peaks, a first order Markov process was estimated for each time series based on their respective lag-one autocorrelations, and then 95% confidence intervals around those spectra were calculated; the thin solid (dashed) red line indicates the 95% confidence level for the spectrum of the Indo-Pacific (Atlantic) heat transport time series. (b) Same as Figure 3a, but for spectra of the Atlantic Meridional Overturning Index at 20°N (blue) and at 33°S (black). The index at each latitude is defined as the maximum value in the vertical of a meridional overturning streamfunction. The 95% significance level for the spectra is shown as the thin solid (dashed) red line for the AMOC at 20°N (33°S).

[11] In contrast, the spectrum for Atlantic ocean heat transport is much redder, and contains two distinct spectral peaks, one at approximately 15–30 years, and a second at approximately 200–500 years. These two peaks are consistent with peaks in the spectrum of NHESAT, thus suggesting that fluctuations in Atlantic oceanic heat transport are closely linked to NHESAT variability on interdecadal to multicentennial time scales, and may drive those fluctuations. Regression analyses (Figure S1 in the auxiliary material) reveal that poleward ocean heat transport in the Atlantic leads the maximum NHESAT by approximately 20–25 years.

[12] The spatial pattern of the ocean heat transport changes associated with the multicentennial variability (not shown) has substantial anomalies in the Atlantic basin. The range of the variations in heat transport in the Atlantic is approximately 0.04 PW (1 PW = 1015 W). These changes are only a few percent of the total mean Atlantic ocean heat transport (of order 1 PW), but are sustained over centennial time scales.

[13] To further examine what aspects of Atlantic ocean heat transport are crucial for the NHESAT variability, we show in Figure 3b the spectra of time series of the Atlantic Meridional Overturning Circulation (AMOC) at two different latitudes. Both spectra show a clear peak at 200–500 years, consistent with the NHESAT spectra, thereby suggesting that AMOC variations drive the NHESAT variability.

[14] What factors give rise to the multicentennial AMOC variability? Shown in Figure 4 are the linear regression coefficients of ocean density versus the AMOC, averaged over a region encompassing the Labrador Sea, Irminger Sea, and Greenland-Icelandic-Norwegian (GIN) Seas, all regions of deepwater formation. Consistent with previous studies [see, e.g., Danabasoglu, 2008, and references therein] a strengthened AMOC is associated with positive density anomalies over regions of deepwater formation in the North Atlantic. The positive upper ocean density anomalies weaken the oceanic stratification, leading to enhanced convection and deep water formation; this in turn “spins up” the AMOC. We next decompose the density changes into temperature and salinity induced density components, also shown in Figure 4. Salinity variations dominate the overall density fluctuations and thus drive the AMOC variations, while temperature variations are out of phase with the AMOC related density variations. The temperature-induced density variations appear to be a response to the AMOC changes; a strong AMOC transports more heat into the high latitudes of the North Atlantic, leading to warmer water and reduced upper ocean density. Since the temperature-induced negative density anomalies around the time of the AMOC maximum would suppress convection by increasing stratification of the water column, these temperature changes would weaken the AMOC and are therefore a negative feedback on the AMOC variations.

Figure 4.

Regression coefficients of various quantities versus the AMOC. The x-axis is time in years, representing the leads or lags relative to the maximum AMOC. Negative (positive) values on the x-axis denote times before (after) a maximum of the AMOC (occurring at lag 0). Black curve denotes ocean density regression coefficients averaged vertically over the upper 300 meters, and in the horizontal from 70°W to 10°E, and from 50°N to 80°N. The blue curve indicates regressions for the component of density attributable to salt variations, and the red curve indicates the component of density attributable to temperature variations. The black dashed vertical line indicates the time of AMOC maximum.

[15] What is the origin of the salinity anomalies? Shown in Figure 5a are plots of surface salinity anomalies zonally integrated across the Atlantic basin. This shows that a salinity signal propagates from south to north in the Atlantic, with a propagation time of approximately 100–150 years. When the positive salinity signal reaches the subpolar gyre in the North Atlantic, convection is enhanced (not shown), and the AMOC strengthens. Shown in the bottom panel are salinity anomalies at 2500 m. integrated zonally across the Atlantic. Here we see a southward propagation of a salinity signal – strong convection in the North Atlantic creates a positive salinity anomaly at depth, which then propagates southward. These results suggest that meridional propagation of salinity anomalies in the Atlantic (with opposite propagation directions in the upper and deeper Atlantic) is an essential part of the mechanisms governing the multicentennial AMOC variations. These propagation time scales appear to set the overall time scale of the AMOC variations.

Figure 5.

Zonal integral of the regression coefficients of salinity in the Atlantic versus the AMOC time series. Units are PSU m Sv−1. The y-axis is latitude, and the x-axis is lag with respect to the time of maximum AMOC. Negative (positive) values on the time axis indicate periods before (after) a maximum AMOC (occurring at lag 0). (a) Values at sea surface. The contours moving from lower left to upper right imply a northward propagation of a salinity signal. (b) Values at 2500 m depth.

[16] What generates the salinity anomalies that propagate northward in the upper layers of the North Atlantic? One possibility involves an atmospheric response to the AMOC that creates a remote salinity signal that then propagates to the North Atlantic. Vellinga and Wu [2004] report such an atmospheric feedback associated with AMOC fluctuations in HADCM3, where a positive AMOC anomaly induces a northward shift of the ITCZ, leading to northward propagation of fresh water anomalies eventually leading to a weakening of the AMOC. However, in the current simulation the propagation of salinity anomalies occurs through the entire latitudinal extent of the Atlantic, and so that precise mechanism is not likely to be responsible for the AMOC multicentennial variability seen in CM2.1. Additional numerical experiments using CM2.1 are currently underway that artificially enhance the AMOC to evaluate whether coupled feedbacks give rise to negative salinity anomalies in the Southern Ocean, and therefore a delayed weakening of the AMOC. Preliminary analyses (not shown) suggest that such feedbacks are weak in this model and are not a major factor in the mechanism of this variability, but more work is needed to robustly address this point.

[17] From the available analyses, the following picture seems to emerge. Positive upper ocean salinity anomalies in the high latitudes of the North Atlantic drive a strengthened AMOC. In turn the strong AMOC is associated with anomalous northward upper layer flow throughout the latitudinal extent of the Atlantic, drawing on fresher waters of Southern Ocean origin. This freshening signal eventually reaches the high latitudes of the North Atlantic, where it suppresses convection and weakens the AMOC. In turn, a weakened AMOC reduces the supply of freshwater from the Southern Ocean to the North Atlantic, eventually leading to increased salinity in the high latitudes of the North Atlantic and a strengthening of the AMOC. It is the propagation time scale of these fresh water signals through the entire latitudinal extent of the Atlantic that appears to set the pace for the multicentennial AMOC variations.

5. Summary and Discussion

[18] We have analyzed the output of a 4000-year control integration of the GFDL CM2.1 model, and have shown the existence of a distinct multi-centennial pattern of Northern Hemisphere Extratropical Surface Air Temperature (NHESAT) variability with a timescale of 200–500 years. This variability appears to be driven by multi-centennial variations in interhemispheric ocean heat transport associated with the Atlantic Meridional Overturning Circulation (AMOC). The multi-centennial AMOC variations are driven by salinity anomalies that propagate from the Southern Ocean to convective sites in the high latitudes of the North Atlantic. When positive salinity anomalies reach the subpolar North Atlantic they increase near-surface density and strengthen the AMOC. The time scale for the variability is related to the propagation time of the salinity anomalies.

[19] Multi-centennial variability of the AMOC has also been seen in an independent climate model [Park and Latif, 2008], although the variability in that model is more focused on the Southern Hemisphere. A recent comparison of centennial scale AMOC variability in three coupled models [Menary et al., 2012] shows the importance of propagating salinity anomalies in modulating the AMOC. The salinity propagation in those simulations, however, occurred primarily north of the Equator in association with latitudinal movement of the ITCZ, whereas the salinity propagation in the current simulation extends over the full latitudinal range of the Atlantic. This greater domain of propagation likely plays a role in the longer timescale of variability seen in the CM2.1 simulation.

[20] While the model employed for this study is generally regarded as a state of the art model, there are nevertheless deficiencies. For example, the horizontal resolution does not resolve oceanic mesoscale eddies or other smaller scale ocean processes, such as boundary currents that can play a significant role in large-scale ocean heat transport. Thus, one should be cautious about the details of the simulation of climate variability from such a model.

[21] Nevertheless, what is significant about this finding is that it demonstrates that the climate system may be capable of generating coherent hemispheric-scale climate variations on centennial and longer time scales through modulating the interhemispheric transport of heat and salt in the ocean. This has potentially significant implications for understanding past climate variations. For example, this simulated multicentennial variability is associated with sustained surface air temperature variations of several tenths of a degree over northern Europe. Could such variations have contributed to observed phenomena such as the Medieval Warm Period [Helama et al., 2009; Trouet et al., 2009]? Along these lines it is important to note that a number of studies using proxy records from the Atlantic region have demonstrated pronounced multidecadal to multicentennial scale variability, possibly a signature of internal variability of the climate system (the spectrum in Figure 3c of Sicre et al. [2008] shows enhanced variance on multi-centennial time scales in reconstructed SSTs north of Iceland). It is also possible that such a pattern of internal climate variability could interact with changing radiative forcing, such as from solar irradiance changes or volcanic eruptions [Ottera et al., 2010]. Such a link is very plausible in terms of the links between such radiative forcing changes and the Northern Hemisphere Annular Mode [Stenchikov et al., 2006], which in turn is linked to the AMOC. It is important to stress, however, that the amplitude of hemispheric-scale temperature variations associated with the multicentennial variability identified in this model (approximately 0.1–0.3 degrees per century) is an order of magnitude smaller than the simulated temperature changes in response to projected changes in greenhouse gases over the late 20th and 21st centuries (approximately 0.1–0.3 degrees per decade).


[22] We thank two anonymous reviewers for their very helpful comments.

[23] The Editor thanks two anonymous reviewers for assisting in the evaluation of this paper.