Current Atlantic Meridional Overturning Circulation weakest in last millennium

Abstract

The Atlantic Meridional Overturning Circulation (AMOC)—one of Earth’s major ocean circulation systems—redistributes heat on our planet and has a major impact on climate. Here, we compare a variety of published proxy records to reconstruct the evolution of the AMOC since about ad 400. A fairly consistent picture of the AMOC emerges: after a long and relatively stable period, there was an initial weakening starting in the nineteenth century, followed by a second, more rapid, decline in the mid-twentieth century, leading to the weakest state of the AMOC occurring in recent decades.

Main

The Atlantic Meridional Overturning Circulation (AMOC) is a major mechanism for heat redistribution on our planet and an important factor in climate variability and change. The AMOC is a sensitive nonlinear system dependent on subtle thermohaline density differences in the ocean, and major AMOC transitions have been implicated, for example, in millennial climate events during the last glacial period¹. There is evidence that the AMOC is slowing down in response to anthropogenic global warming²—as predicted by climate models—and that the AMOC is presently in its weakest state for more than 1,000 years³. As continuous direct measurements of the AMOC started only in 2004⁴, longer-term reconstruction must be based on proxy data. In general, there are three different types of AMOC proxies: (1) reconstructions of surface or subsurface temperature patterns in the Atlantic Ocean that reflect the changes in ocean heat transport associated with the AMOC^3,5; (2) reconstructions of subsurface water mass properties, for example, the advance of the subpolar versus subtropical slope water, that reflect AMOC changes⁶; and (3) evidence for physical changes in deep-sea currents, such as those reflected by changes in sediment grain size³. As all kinds of proxies are limited in their representation of the AMOC (all three can be influenced to some degree by factors in addition to changes in the AMOC), a combination of all three proxy types is needed to provide robust evidence about the evolution of the AMOC.

In this article, using several different and largely independent proxy indicators of the AMOC evolution over the past 100 to nearly 2,000 years, we provide strong evidence that the AMOC decline in the twentieth century is unprecedented and that over the past decades, the AMOC is in its weakest state in over a millennium.

The proxies are taken from various locations in the Atlantic or the surrounding land areas (inset of Fig. 1) and represent either different subsystems associated with the AMOC (such as Labrador Sea density³, the presence of subtropical versus subpolar slope waters along the North American East coast^6,7) or the effect of changes in the Atlantic meridional heat transport associated with the AMOC^2,3,5,8, as well as surface ocean productivity changes that have been related to the AMOC^9,10. The records going the furthest back in time (ad 400) are taken from marine sediments (sortable-silt data³, proxy records of subsurface ocean temperatures³, δ¹⁸O in benthic foraminifera⁷, δ¹⁵N of deep-sea gorgonian corals⁶, relative abundance of certain planktic foraminifera (Turborotalita quinqueloba)¹⁰). The temperature-based AMOC index⁵, however, is based on a Northern Hemisphere land-and-ocean temperature reconstruction that uses a range of terrestrial proxies, including, for example, tree rings and ice-core data¹¹. Data taken from Greenland ice cores (the methanesulfonic acid concentration) furthermore provide an estimate for AMOC-related changes in productivity in the subpolar gyre region⁹. Most of these records extend into the modern era, for which additional AMOC proxies exist that are based on instrumental temperature records^2,8.

Fig. 1: SST-based AMOC reconstructions compared with various proxy reconstructions.

a, The SST-based proxies (light and dark blue) represent the North Atlantic temperature response to changes in the Atlantic meridional heat transport associated with an AMOC slowdown. b–k, It is compared with proxy records of subsurface ocean temperatures (purple) (b), δ¹⁵N data of deep-sea gorgonian corals (magenta) (c,h), mean grain size of sortable-silt data $\overline{ss}$ (shades of green, shown with a 12-year lag to the temperature-based indices³) (d,i), δ¹⁸O data in benthic foraminifera (shades of brown) (e,j), the relative abundance of T. quinqueloba in marine sediment cores (orange-red) (f), methanesulfonic acid concentration in Greenland ice cores (orange) (f,k), both indicators for local/regional marine productivity, and the relative change in Atlantic Ocean heat content versus that in the Southern Ocean (dark magenta) (g). As a reference for the actual change in volume transport, the April 2004–April 2018 linear trend of the RAPID data⁴ (black) is given (g). The map (using the same colour-coding as the time series) gives an overview of the various locations the proxies were taken from (with small markers denoting single sites and large markers denoting the areas with multiple proxy sites). All curves were smoothed with a 20-year (50-year) LOWESS filter for the shorter (longer) time series to make them more comparable. Shading and error bars show the 2σ (95%) confidence intervals of the individual proxies as they were reported and the uncertainty of the AMOC representation of the Caesar et al. (2018) temperature proxy, respectively (Methods). SPG, subpolar gyre; VPDB, Vienna PeeDee Belemnite; ZJ, zetajoules.

Despite the different locations, timescales and processes represented by these proxies, they provide a consistent picture of the AMOC evolution since about ad 400: before the nineteenth century, the AMOC was relatively stable. A decline in the AMOC, beginning during the nineteenth century, is evident in all the proxy records (Fig. 1 left panel). Around 1960 there began a phase of particularly rapid decline that is found in several, largely independent proxies. A short-lived recovery is evident in the 1990s before a return to decline from the middle of the first decade of the 2000s (Fig. 1 right panel). All indices additionally show multi-decadal variability, albeit with different amplitudes and frequencies making it questionable whether this is driven mainly by the AMOC. Some of the differences probably relate to the large range in temporal resolution in the proxies (from annual to 50-year binning), while others are probably due to complicating factors, such as non-AMOC-related influences on a proxy system (for example, changes in trophic structure of coral’s food source in δ¹⁵N, local fluctuations in circulation impacting single-site palaeoceanographic reconstructions or other controls on subpolar heat content¹²). An additional factor may be that different components of the AMOC respond on different timescales. While the strength of the AMOC, typically measured at 26° N, has been shown to be correlated with the multi-decadal variability of North Atlantic sea surface temperature (SST)¹³ (suggesting that a large part of this variability in the temperature-based proxies is due to AMOC changes), changes in the deep ocean appear to occur on a different timescale. Therefore, it is unsurprising that for the larger part of the last millennium, the multi-decadal variabilities in the proxies differ.

The strength of this multi-proxy comparison lies in tracing the centennial and longer AMOC evolution. To test whether the reduction in AMOC strength that is seen in all proxy records is significant, a change-point model is fitted to each time series and the data means before and after the change points are compared (Methods). Assuming, in the first approximation, only a single change point, the model finds a significant reduction in the mean in all but one proxy record (see Table 1). The timing of the change point varies in the different proxy series (also due to the different lengths of the time series) but can be sorted into two clusters: one change occurring in the second half of the nineteenth century and a second change occurring in the 1960s. To test the significance of differences between different time periods, we divided each time series into 50-year intervals (30-year intervals for the Cheng et al. (2017) data given that the length of the time series is only 64 years and 100-year intervals for the Spooner et al. (2020) data given the coarse resolution of this time series), going backward from the present, and we estimated the means and data uncertainty for each of these intervals. The mean of any 50- (30-, 100-) year interval is assumed to be significantly lower when its uncertainty range does not overlap with the uncertainty range of the mean of any other interval. The results show that in 9 of the 11 proxy series, the most-recent 50- (30-, 100-) year mean value is significantly lower than any other before (see Table 1). In addition, the high-resolution proxies suggest a progressive AMOC decline within that most-recent interval.

Table 1 Results of the change-point and significance testing of the various proxies used to reconstruct the evolution of the AMOC

Together, these data consistently show that the modern AMOC slowdown is unprecedented in over a thousand years. Improved understanding of this slowdown is urgently needed. The next step is to resolve which components and pathways of the AMOC have altered, how, and why—no small feat, and requiring a community effort that combines observational, modelling and palaeoclimatological approaches.

Methods

Uncertainties

The uncertainty range represents, in all but one case, the 2σ confidence interval for the individual proxy reconstruction, that is, for (1) the proxy-based surface temperature reconstruction (validated against independent instrumental temperature data)⁵, (2) the subsurface temperature dipole in the Atlantic based on the published uncertainties for age assignment and temperature reconstructions³, (3) the δ¹⁵N record based on a mixed effect linear model based on year and specimen colony⁶, (4) the sortable-silt data that are shown with their full (reduced) procedural error³, (5) the δ¹⁸O data based on analytical reproducibility determined by replicate measurements of internal standard carbonate material⁷, (6) the abundance of T. quinqueloba with the uncertainty estimated using a binomial approach¹⁰ and (7) the marine productivity in the subpolar gyre based on a bootstrapping method⁹. As an upper bound for the 2σ confidence interval of the relative change in Atlantic Ocean heat content versus that in the Southern Ocean, the confidence intervals for the individual ocean heat content time series, considering among others instrumental errors, methodological choices and data gaps, were simply added⁸.

For the temperature-based AMOC proxy², the uncertainty in converting these proxy data to an AMOC slowdown is given, not the uncertainty of the temperature data itself. This is based on the relationship between the relative temperature change in the subpolar North Atlantic and AMOC variability in the Coupled Model Intercomparison Model phase 5 model ensemble.

Although only this last uncertainty interval considers the spread in the proxy that is unrelated to the AMOC, the other proxies have all been related to AMOC variability. Moreover, given that the proxies were taken from multiple locations across the Northern Hemisphere and the only inferred common driver for them all is AMOC, combined, they provide strong evidence for a centennial decline related to the AMOC.

Statistical significance

To determine whether there has been a significant change in the proxy time series that would indicate a reduction in AMOC strength, we tested each proxy record for a single significant change in the mean of the time series. Using a Bayesian framework, we fit a model that assumes that the data fluctuate around a constant mean, allowing for a single change in the mean at some point in time. The approach takes both the data uncertainty and the data variability into account. Once the model finds the timing of the change, we compare the means before and after the change point to check whether the difference is significant (we assume significance when the 95% Bayesian credible interval of the difference between the means does not contain a zero value).

To test whether the AMOC is at its weakest in over a millennium, we applied a similar framework that fixed change points at 50- (30-, 100-) year intervals. Starting in the present, the mean and 95% uncertainty interval for each 50- (30-, 100-) year interval was estimated (taking data uncertainty and the number of data points in each interval into account).