On the relationship between Atlantic meridional overturning circulation slowdown and global surface warming

According to established understanding, deep-water formation in the North Atlantic and Southern Ocean keeps the deep ocean cold, counter-acting the downward mixing of heat from the warmer surface waters in the bulk of the world ocean. Therefore, periods of strong Atlantic meridional overturning circulation (AMOC) are expected to coincide with cooling of the deep ocean and warming of the surface waters. It has recently been proposed that this relation may have reversed due to global warming, and that during the past decades a strong AMOC coincides with warming of the deep ocean and relative cooling of the surface, by transporting increasingly warmer waters downward. Here we present multiple lines of evidence, including a statistical evaluation of the observed global mean temperature, ocean heat content, and different AMOC proxies, that lead to the opposite conclusion: even during the current ongoing global temperature rise a strong AMOC warms the surface. The observed weakening of the AMOC has therefore delayed global surface warming rather than enhancing it. Social Media Abstract: The overturning circulation in the Atlantic Ocean has weakened in response to global warming, as predicted by climate models. Since it plays an important role in transporting heat, nutrients and carbon, a slowdown will affect global climate processes and the global mean temperature. Scientists have questioned whether this slowdown has worked to cool or warm global surface temperatures. This study analyses the overturning strength and global mean temperature evolution of the past decades and shows that a slowdown acts to reduce the global mean temperature. This is because a slower overturning means less water sinks into the deep ocean in the subpolar North Atlantic. As the surface waters are cold there, the sinking normally cools the deep ocean and thereby indirectly warms the surface, thus less sinking implies less surface warming and has a cooling effect. For the foreseeable future, this means that the slowing of the overturning will likely continue to slightly reduce the effect of the general warming due to increasing greenhouse gas concentrations.


Introduction
Variations in the Atlantic meridional overturning circulation (AMOC) can change Northern Hemisphere and even global surface temperatures (Knight et al 2005, Stolpe et al 2018. Most model studies have shown that an AMOC decline, in response to increased CO 2 concentrations, weakens the poleward ocean heat transport, increases the ocean heat uptake (Rugenstein et al 2013) and therefore diminishes global warming. Yet a recent study, analysing observations of the last decades (Chen and Tung 2018), challenged this finding and came to the conclusion that a weak AMOC can lead to more rapid surface warming. In the following we show that the observational and reanalyses data analysed in this study are fully consistent with the Original content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence.
Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. previously established understanding, i.e. that the correlation between AMOC strength and global surface warming is (at inter-annual to decadal timescales) negative and that over the last decades a weaker AMOC likely acted to delay global surface warming.
Chen and Tung (2018) base their idea that global surface warming in the next decades may be enhanced by a weaker AMOC on the concept that as a consequence of an AMOC weakening less heat can be carried into the deep ocean via deep convection. This central claim was supported by a visual comparison of the time series for AMOC strength and global surface warming over the last decades (Chen and Tung (2018), figure 3) and rests primarily on the period from 1975 to 1998, during which the AMOC was in a relatively weak state, which coincided with a period of rapid surface warming. They conclude that in recent decades a weakened AMOC warmed the surface by bringing less heat into the deep North Atlantic. Furthermore they argue that this mechanism explains why the trend in ocean heat content in the North Atlantic Ocean went from positive to negative when comparing two time spans of an increasing (1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)) and a decreasing (2005-2016) AMOC, explaining this change in the ocean heat content by a change in the vertical heat transport into the ocean driven by the AMOC: 'Deep convections can now carry more heat downward' (Chen and Tung 2018). This suggested mechanism is unlikely to operate in the northern Atlantic, since convection there is thermally driven: convective mixing results from static instability due to colder water overlying warmer water. It therefore transports heat upwards in the water column, not downward, balancing the oceanic heat uptake occurring over large areas of the ocean by turbulent diffusion (e.g. Winton 1995, Drijfhout 2015. To investigate the relationship between overturning strength and changes in the global mean surface temperature (GMST), we perform a correlation analysis of the two time series. We therefore revisit the analysis of Chen and Tung (2018) and investigate the relationship between the detrended GMST evolution and different indices for the AMOC strength. This correlation analysis shows that the data they present do not support the conclusion that an AMOC weakening currently enhances global surface warming. We furthermore extend the analysis with a second method that accounts for the variability of the radiative forcing as well as feedback processes in the Earth system, yielding very similar results. These results are in agreement with the understanding that a weaker AMOC increases the global ocean heat uptake and therefore has a cooling effect on the global surface temperature. Additional support to this conclusion is provided by the fact that the recent decline of the AMOC (Smeed et al 2014) coincided with an increase in the ocean heat uptake rate.

Data
To ensure that differences in the results between this study and that of Chen and Tung (2018) are not due to differences in the underlying data we used the same global mean temperature time series, the same AMOC proxies and the same time series for the ocean heat content (OHC). The latter is extended by the improved OHC estimates by Cheng et al (2017).
2.1. Global mean surface temperature, forcing time series and feedback parameter We start by considering the temperature evolution in the light of the global energy balance. For this purpose we use the median of the HadCRUT4.6 data (Morice et al 2012) that provides an estimate of the global mean surface temperature anomaly since 1850 with respect to . For the changes in the radiative forcing we use a time series that combines the known individual forcing data sets (greenhouse gases, ozone, solar irradiance, land use, snow albedo, orbital parameters, direct and indirect effect of tropospheric aerosols and volcanic aerosols) that are used to drive the CMIP5 historical simulations and represents all changes in both the natural and anthropogenic forcing from 1850 until 2012 (Miller et al 2014). For our analysis we need to consider the temperature difference ΔT relative to the preindustrial equilibrium state. Therefore, both GMST and forcing anomaly are given relative to the year 1850. Since we also account for the feedback response of the Earth system to an initial temperature change, we need to estimate the strength of this response, i.e. the feedback parameter λ. As a best estimate we chose a feedback parameter of λ=2.3±0.7 W K −1 m −2 that was determined for the period 1979-2008 across an ensemble of 19 AGCMs (Gregory and Andrews 2016). However, there is a growing understanding that the Earth's climate sensitivity is time-scale dependent, as the different climate feedbacks act on different time scales. An ensemble study based on the analysis of a modified energy balance model, constrained by observations and the outputs from the CMIP5 models, give a feedback parameter of 1.9±0.3 W K −1 m −2 for intra-annual scales that decreases to around 1.5±0.3 W K −1 m −2 and 1.3±0.3 W K −1 m −2 on response time scales of 10 and 100 years, respectively (Goodwin 2018). We therefore tested the correlation between AMOC changes and surface warming with the following values for the feedback parameter: λ = 1.3, 1.5, 1.9, 2.3, 3.0 W K −1 m −2 as well as for a counterfactual case where we neglect all radiatively forced surface temperature changes.

AMOC indices and ocean heat content
Due to the lack of long-term AMOC measurements, the evolution of the AMOC over the last century has to be reconstructed from proxy data. To cover the uncertainty of the reconstruction we base our analysis on three different AMOC indices. The first two indices are based on the upper (0-1500 m) subpolar ocean salinity with salinity values taken either from ISHII and Scripps (A ISHIIS+Scripps ) or from EN4 (A EN4 ) data (Chen and Tung 2018). The third is the sea surface temperature (SST) based index as defined by Caesar et al (2018) based on the HadISST data (A HadISST ). To determine the trend in the ocean heat uptake (OHU) over the last decade, the ocean heat content based on the ISHII and Scripps datasets (Chen and Tung 2018) is used. It is a time series of the 0-1500 m OHC from mid-2000 until mid-2014. We further analyse the OHC distribution over the different ocean basins using the improved OHC estimates by Cheng et al (2017) that are given until a depth of 2000 m.

Method and results
First, we study the relationship between AMOC strength and the forcing corrected global surface warming with a correlation analysis over the whole length of the time series for which all data is available . The forcing correction is done in two different ways: on the one side by just removing the long-term warming signal (either by removing the linear trend or by removing a nonlinear trend as done by Chen and Tung (2018)) and on the other side by using a simple equation for the global mean energy balance. This will answer the question whether the opposing course of the two variables between 1975-1998, as identified by Chen and Tung (2018) (their figure 3(b)), is also valid during other time periods. While this correlation analysis will not suffice to determine the contribution of different processes to global temperature changes, it is sufficient to identify whether periods of a weaker AMOC over the last decades had a distinct cooling, warming or close to no effect on the global surface temperature.
Second, we investigate the trend reversal in the ocean heat content in the North Atlantic Ocean from positive, during a time period of increasing AMOC strength (2000)(2001)(2002)(2003)(2004), to negative, during a time period of a decreasing AMOC (2005AMOC ( -2016, considering the role of the AMOC in transporting heat horizontally from the Southern Ocean into the Atlantic.
3.1. Energy budget and the influence on the vertical ocean heat transport Global mean surface temperature changes-i.e. the lower atmosphere and upper ocean, which are wellmixed and thermally tightly coupled-are forced by radiative forcing from the top and heat exchange with the deep ocean below (Trenberth et al 2010, Brown et al 2014: m r a d o c e a n

( ) /
Here, T is the global mean surface temperature, c m is the effective heat capacity of the system (dominated by the ocean mixed layer), Q rad the radiative forcing and Q ocean the vertical heat transport across the bottom of the ocean mixed layer e.g. through diffusion (fluxes are positive downward) (Brown et al 2014). Δ indicates differences to a previous equilibrium state (e.g. preindustrial). The term λ ΔT represents the equilibrium response ΔT of the surface temperature to the forcing anomaly, which depends on the climate feedback parameter λ. The equation holds for the global mean temperature, therefore horizontal transport processes play no role. Solved for ΔQ ocean this equilibrium is: Since we are looking at temperature changes at multidecadal timescales we can assume that the mixed layer is close to equilibrium and thus neglect the transient term on the left hand side in equation (1). This term would lead to some delay of the surface temperature response to forcing changes, yet empirical correlation shows that the lag of the global surface temperature response to a change in the radiative forcing, e.g. the 11 year solar cycle, is of the order of a month (Foster and Rahmstorf 2011), so for our purposes this lag is not significant. With given time series for ΔT and ΔQ rad (both with respect to the preindustrial equilibrium state of 1850) and the different estimates for the feedback parameter λ we can now use equation (2) to test how AMOC variations (represented by the AMOC indices) correlate with the part of surface temperature changes that are not directly radiatively forced (i.e. the right hand side of the equation (2)).
The correlation values (figure 1) are positive (with r=0.49, 0.57 or 0.22 depending on the AMOC proxy) with particularly warm GMST anomalies coinciding with a strong AMOC. This is in direct contradiction to the idea that a strong AMOC acts to cool the surface and in full agreement with the established understanding of the AMOC's role in vertical heat transport (Drijfhout 2015). We use the smoothed time series to determine the influence of the AMOC since the short-term fluctuations in the ocean heat uptake in the North Atlantic are dominated by atmospheric variability (Gulev et al 2013).
While the exact values of the correlation coefficients depend on the choice of the feedback parameter and the AMOC index, they are positive in all cases (i.e. between 0.01 and 0.65, see table 1) and therefore do not support the hypothesis that a weak AMOC enhances surface warming by decreasing the ocean heat uptake (in that case the correlation coefficients would be negative). The fact that most of the correlation values are not significant at the 5% level (this was tested using amplitude-adjusted Fourier transform (AAFT) surrogates (Donges et al 2015)) is also irrelevant for deducing that an AMOC weakening does not enhance surface warming, as it is sufficient to show that the coefficients are not negative. As can be seen in figure 1 there is no apparent lag between the adjusted surface warming and the AMOC strength, consistent with our assumption that the ocean's mixed layer and the atmosphere are responding to the changes in forcing within a year. This was verified with a lag-correlation analysis that showed no significant time lag between the two.
Our analysis takes the role of radiative forcing in affecting GMST as a given and looks at any additional effect of the AMOC. Alternatively, the internal climate variability can be estimated by removing the warming trend from the original time series. This is the approach taken by Chen and Tung (2018) who removed a nonlinear secular trend (their figure 3) that is very similar to the 100 year linear trend. In the case that we remove either a linear warming trend or a nonlinear trend (using the exact same data as Chen and Tung (2018)) we get even larger positive correlation values with r = 0.62, 0.42, 0.65 for the linear warming trend removed and r = 0.57, 0.39, 0.61 when removing the nonlinear trend (for A ISHIIS+Scripps, A EN4 and A HadISST , see right columns of tables 1 and 2).
Even if we consider only the period after 1975, on which Chen and Tung rested their argument, we find mostly positive correlation values (see table 2). While certain combinations of λ and AMOC proxy yield a negative correlation (especially for smaller values for the feedback parameter), the correlation between AMOC strength and GMST variability is still positive when the radiative forcing is taken into account by removing the linear or nonlinear trend from the data (right columns of table 2).
These results are consistent with several model studies which likewise found a positive correlation with no lag between the AMOC strength and global as well as northern hemisphere temperature (e.g. Knight et al 2005, Maroon et al 2018. It is also in alignment with the fact that the decline of the AMOC over the last decade (Smeed et al 2014), for which direct AMOC measurements exist, coincided with an increase in the rate of ocean heat uptake (figure 2), not a decrease.

Basin shift or the influence on the horizontal ocean heat transport
Since the Argo era, ocean heat content measurements have increased in quality and extent in particular considering the deep ocean (Cheng et al 2017). The data show that there is a large shift in the regional ocean heat content between the period 2000-2004 and Smoothing) filter fits a regression curve to a scatterplot using weighted local linear regressions depending on the smoothing span, in this case 10 years (Cleveland 1979). The correlation coefficients r were calculated with the smoothed time series. (To remove any correlations due to common trends the time series were first linearly detrended.) Table 1. Results of the sensitivity analysis of the correlation values considering the uncertainties of the feedback parameter λ. The correlation values were calculated for the whole time period . Values that are significant at the 5%-level are shown in boldface.  figure 3) as identified by Chen and Tung (2018).
As there have only been very few ARGO measurements prior to 2005 (especially in the Southern Ocean) it is uncertain how accurate the magnitude of this shift is. However, the fact that the Southern Ocean has been experiencing the greatest warming of all oceans since 1998 is robust (Cheng et al 2017). While Chen and Tung (2018) explain this shift with a change in the vertical heat transport into the ocean driven by the AMOC, we show that it can largely be explained within the established understanding that AMOC variations cause a change in the horizontal heat transport in the Atlantic.
As the AMOC carries relatively warm, saline water from the low-latitudes and Southern Ocean to the polar North Atlantic and returns cold, deep water southwards, it accounts for about 90% of the maximum meridional heat transport in the Atlantic of about 1.3 PW (Johns et al 2011, Xu et al 2016 occurring in the subtropics. AMOC variations lead therefore to a change in the meridional heat transport. A  statistical analysis of expendable bathythermograph (XBT) data suggests that a 1 Sv weaker AMOC leads to a decrease of 0.04±0.02 PW in the associated meridional heat transport at 35°S either due to less import of warm surface waters like the Agulhas Current, a reduced export of colder subsurface waters or a combination of the two (Garzoli et al 2013).
Between the periods 2000-2004 and 2005-2014 the AMOC decreased by about 1.5 Sv (Caesar et al 2018). Neglecting short term variability, a constant AMOC strength can be taken for the years 2000-2004, followed by a linear decline of 1.5 Sv until the year 2014. This means that the mean strength of the AMOC in the years 2005-2014 was 0.75 Sv weaker than in the years before. This leads to a cumulative change of 0.75 Sv · 10 yr · 0.04 PW/Sv=9.5 ZJ in the meridional heat transport over the duration of these 10 years. Thus, the AMOC decline is estimated to cause a shift of about 9.5 ZJ of heat from the Atlantic Ocean to the Southern Ocean over this period. Taking the reduced heat transport due to the weaker AMOC into account, the changes in the division in ocean heat uptake between the two basins can largely be explained in terms of horizontal transport rather than surface heat uptake change (see figure 3).

Discussion
The statistical evaluation of the observed global mean temperature, ocean heat content, and different AMOC proxies, presented in this study, yields a positive correlation between GMST changes and changes in AMOC strength. This supports the understanding that the deep water formation related to the AMOC transports cold surface water downwards and that the recent weakening of the AMOC has therefore delayed global surface warming.
Even though we find this positive correlation between AMOC strength and global mean temperatures, indicating a cooling effect of an AMOC slowdown on GMST, the relatively moderate correlation values suggest that, at least for the considered time period  and timescale (intra-annual to decadal), AMOC variability is not dominant in explaining changes in ocean heat uptake. This is not surprising as there are other processes that cause variations in the ocean heat uptake. For example England et al (2014) showed that a pronounced strengthening of the Pacific trade winds over the last two decades cooled the tropical Pacific and significantly increased the ocean heat uptake which can at least partly explain the slowdown in the observed surface warming since 2001. At the same time there are other mechanisms than changes in ocean heat uptake through which AMOC variability influences global surface temperatures. The reduced meridional heat transport following an AMOC weakening leads to a cooling in the Northern Hemisphere and a warming in the Southern Hemisphere. Although this just changes the distribution of heat on the planet, climate feedbacks that selectively amplify the cooling response in the Northern Hemisphere can lead to a decrease in the global mean temperature (Drijfhout 2015). One example of such a feedback is that a cooling in the subpolar North Atlantic can increase the sea-ice cover and thus lead to a further decrease in GMST through enhanced reflection of solar radiation (Drijfhout 2018). Another feedback is that a strong decline in AMOC strength can enhance the meridional SST gradients in the North Atlantic, leading to stronger Northern Hemisphere storm tracks, as shown in a model simulation of an AMOC shutdown (Jackson et al 2015). Stronger storm tracks allow for a greater lower cloud coverage in the high latitudes (Trossman et al 2016). Thus, the enhanced shortwave cloud feedback cools the surface (Rose et al 2014).
Additionally, we would like to stress the importance of differentiating between the relationship between AMOC and GMST when considering the response of the global mean temperature to an AMOC change (where a weaker AMOC cools the surface) and the forced response of the AMOC to changes in GMST (where a warming leads to a weaker AMOC) as discussed in Maroon et al (2018). Anthropogenic warming will very likely lead to a weakening of the AMOC which among other things due to the reduced ocean heat release associated with a weaker deep convection will dampen the original warming signal and therefore acts as a negative feedback. Figure 1 suggests that this cooling response is of the order of 0.1°C per Sverdrup.
In summary, we find that the observed changes in AMOC and global mean surface temperature over the last decades are fully consistent with the established understanding that a strong AMOC cools the deep ocean, and that this continues to be the case under the current situation of global warming. The observed, recent weakening of the AMOC has therefore delayed global surface warming rather than enhancing it. Acknowledgments LC, SR and GF all receive funding from the Potsdam Institute for Climate Impact Research, Germany.

Data availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.