1 Introduction

The large-scale ocean circulation transports heat, nutrients, freshwater, and carbon around the globe. In the Northern Hemisphere alone, 1.5–2 PW (1 PW = 10\(^{15}\) W) of heat is transported meridionally in the ocean (Ganachaud and Wunsch 2003). An essential part of this occurs in the Atlantic Ocean (Trenberth et al. 2001), in the so-called Atlantic Meridional Overturning Circulation (AMOC). AMOC has distinct branches and a complex three dimensional structure, with warm shallow waters entering and sinking in the Labrador and Nordic Seas, recirculating in the Arctic and then exiting the Denmark Straits at depth to supply the North Atlantic Deep Water (NADW) (Johnson et al. 2019; Bower et al. 2019).

AMOC is often represented as a two-dimensional circulation in latitude and depth (Broecker 1991; Wright and Stocker 1991; Schmitz 1996; Cessi 2019). This is as is commonly done in the atmosphere; the three dimensional meridional circulation is made two dimensional following zonal integration of the velocities. The resulting oceanic cells—primarily the AMOC and Antarctic Bottom Water (AABW)—are a simple means of comparing the response across models. But the integration obscures the 3D structure of the flow, with its western boundary currents and horizontal recirculations. The latter, for example, can extend the meridional pathways in both the upper and lower limbs of the AMOC, as fluid parcels may be delayed for years before moving northward. This could greatly increase their net exposure to the atmosphere (Bower et al. 2019).

AMOC is not stationary, but varies on a range of time scales, from weekly to centennial (Bryden et al. 2005; Delworth and Mann 2000; Knight 2005; Jackson et al. 2015; Lopez et al. 2016; Zhang et al. 2019). AMOC is predicted to weaken in the warming climate as well. The observational evidence for this is equivocal, as the records are relatively short; while data from the Rapid Climate Change (RAPID) array at 26.5°N in the North Atlantic indicates some weakening (Häkkinen and Rhines 2004; Bryden et al. 2005; McCarthy et al. 2012; Robson et al. 2013; Smeed 2014, 2018), evidence from basin-wide observations in the North and South Atlantic suggest that there has been no significant change since 2012 (Lozier et al. 2017, 2019; Fu et al. 2020; Li et al. 2021; Caínzos et al. 2022). Predictions of AMOC weakening are based instead on climate model simulations (Cheng et al. 2013; Weaver et al. 2012; Fox-Kemper et al. 2021). These indicate a 24–39% decline by the end of the twentyfirst century, with more weakening with larger increases in greenhouse gas (GHG) emissions (Lin and Zhang 2004; Gregory 2005; Schmittner et al. 2005; Drijfhout and Hazeleger 2007; Weaver et al. 2012; Cheng et al. 2013; Reintges et al. 2017; Menary et al. 2020; Fox-Kemper et al. 2021; Weijer et al. 2020).

Despite consistent indications from model simulations, there is a lack of consensus on the driving mechanisms for the weakening. These have been linked to changes in wind forcing, buoyancy forcing or a combination of the two. Changes in the local zonal wind stress affect the AMOC strength on sub-decadal timescales (Jayne and Marotzke 2001; Hirschi et al. 2003, 2007; Kuhlbrodt et al. 2007; Kanzow et al. 2010; Polo et al. 2014; Yeager 2015). AMOC has also been linked to the wind-driven Ekman transport from the Southern Ocean, which is of a similar magnitude (20 Sv) (Toggweiler and Samuels 1995, 1998; Wunsch 2004; Kuhlbrodt et al. 2007; Cessi 2019). Indeed, there is evidence in some models that a reduction in AMOC could be balanced by increased overturning in the Southern Ocean, due to increased wind forcing there. This “bipolar seesaw" effect has been invoked to explain ice core records from Greenland and Antarctica (Stocker and Johnsen 2003).

Buoyancy forcing, from changes in surface heating or evaporation/precipitation, has been implicated in longer term variations (Yeager and Danabasoglu 2014). Anomalous surface heat fluxes have been shown to increase upper-ocean stratification in the subpolar North Atlantic, affecting deep water formation (Dixon et al. 1999; Schmittner et al. 2005; Gregory 2005; Stouffer et al. 2006; Weaver et al. 2007; Liu et al. 2014, 2017, 2019; Gregory et al. 2016; Maroon et al. 2018). Other studies suggest that subsurface warming weakens overturning (Haskins et al. 2020; Levang and Schmitt 2020a; Bonan et al. 2022).

More often though, AMOC changes are associated with freshwater fluxes from melting sea ice and glaciers (Manabe and Stouffer 1995; Ganopolski and Rahmstorf 2001; Rind et al. 2001; Vellinga and Wood 2002; Knutti et al. 2004; Stouffer et al. 2006; Weijer et al. 2019; He and Clark 2022). Indeed, “hosing" experiments, in which a large volume of freshwater is introduced suddenly in the subpolar gyre, exhibit drastic AMOC weakening. The response however is strongly model dependent (Proshutinsky et al. 2015; Rahmstorf et al. 2015; Bakker 2016).

Changes in surface temperature and salinity often occur in tandem. Enhanced heating at high latitudes reduces surface densities but also melts sea ice, further increasing buoyancy (Cheng et al. 2013; Liu et al. 2019). The resulting increase in surface buoyancy in the subpolar gyre is believed to suppress deepwater formation. It has been suggested that this could tip the system to a new overturning state, following idealized models such as Stommel’s (1961). However the importance of deep convection for the overturning is disputed. Some argue that AMOC is not “pushed" by convection but rather “pulled" by mixing in the ocean interior (Wunsch 2002; Visbeck 2007). Others contend that surface buoyancy forcing itself is the primary driver, by inducing a meridional density gradient (de Verdière 1988; Gjermundsen and LaCasce 2017; Gjermundsen et al. 2018).

Hereafter, we examine AMOC changes in one of the most extreme of the CMIP6 simulations, in which CO2 is suddenly quadrupled. While unrealistic, the abrupt transition facilitates diagnosing the response. We find that wind forcing, both local and in the Southern Ocean, does not change in a consistent manner between the models and is not correlated with AMOC weakening. Rather, AMOC decline is strongly correlated with the density gradient across the separated Gulf Stream, the North Atlantic Current (NAC). The gradient is weakened by freshwater fluxes from melted sea ice which enter the subpolar gyre from the western boundary. The weakened gradient yields a weaker baroclinic shear, following the thermal wind relation. The AMOC, which derives from the zonal and vertical integrals of the meridional velocity, decreases proportionally.

The model simulations are discussed in Sect. 2. AMOC weakening and its dependence on wind- and buoyancy-forcing is examined in Sect. 3. The results are discussed and summarized in Sect. 4.

2 Materials and methods

We employ coupled climate simulations from the Coupled Model Intercomparison Project Phase 6 (CMIP6) project (Eyring et al. 2016). The focus will be on the abrupt-4xCO2 experiment (“4xCO2" hereafter), in which the CO2 concentration is abruptly quadrupled and held fixed thereafter, keeping all other external forcings the same. The results will be compared to those from the pre-industrial control run (“control") with no increase in greenhouse gas emissions. The 4xCO2 run is a standard test case in the CMIP ensemble. For comparison, a 1% increase of CO2 per year would yield a quadrupling in 139 years. We use 35 different GCM simulations for comparison (refer Table 1 in “Appendix”).

The AMOC streamfunction is calculated by integrating the meridional velocity, \(v_o\), zonally between the continental boundaries from \(x_w\) to \(x_e\) at a given latitude and with depth, thus:

$$\begin{aligned} \Psi (y,z) = \int _{z}^{0} \int _{x_w}^{x_e} v_o (x, y, z') \,dx dz' \end{aligned}$$
(1)

The streamfunction is calculated in each of the basins, i.e., Atlantic-Arctic, Indian-Pacific, and globally. Note that while calculating the streamfunction as a function of density yields a better indication of the Lagrangian transport, e.g. (Nycander et al. 2007; Johnson et al. 2019), the z-coordinate version here is more conducive to the present analysis.

When examining the response to surface buoyancy forcing, it is useful to separate the density changes into those due to temperature and to salinity (e.g. Dai 2022):

$$\begin{aligned} \triangle \rho (S,T,P) = \frac{\partial \rho (S,T,P)}{\partial T} \triangle T+ \frac{\partial \rho (S,T,P)}{\partial S} \triangle S \end{aligned}$$
(2)

where \(\rho\) is density at temperature T, salinity S and pressure P. We will be primarily concerned with the surface density, for which \(P=0\). To evaluate the derivatives, we use UNESCO’s equation of state for seawater (Fofonoff and Millard 1983). The derivatives are evaluated from the 50-year-mean of the control run, and the changes in temperature and salinity, \(\triangle T\) and \(\triangle S\), calculated from differences between the 4xCO2 and control runs after a selected time, here 100 years.

Changes in the surface density affect the vertical shear of the horizontal velocities via the thermal wind relation. Under the Boussinesq approximation, the relation for the zonal velocity is:

$$\begin{aligned} \frac{\partial u}{\partial z} = \frac{g}{\rho _c f} \frac{\partial \rho }{\partial y} \end{aligned}$$
(3)

where \(\rho _c\) is the reference density of seawater and f is the Coriolis parameter. The eastward thermal wind surface velocities are calculated by integrating the meridional density gradient from the bottom to the surface of the ocean assuming the bottom velocities are zero. This neglects bottom-intensified currents, such as the deep western boundary currents, but yields very realistic surface velocities in the interior, as seen hereafter.

The role of wind forcing is also assessed. We examine both local forcing in the subpolar gyre and remote forcing in the Southern Ocean. The former is gauged using the wind stress curl:

$$\begin{aligned} \overrightarrow{k} \cdot (\nabla \times \overrightarrow{\tau }) = \frac{1}{r_e \cos {\theta }} \bigg (\frac{\partial }{\partial \varphi } \tau ^y-\frac{\partial }{\partial \theta }(\tau ^x \cos {\theta })\bigg ) \end{aligned}$$
(4)

where \(r_e\) is the radius of the Earth, \(\theta\) and \(\varphi\) are latitude and longitude respectively, and \(\tau ^x\) and \(\tau ^y\) are the zonal and meridional components of the surface wind stress. The curl is related to the wind-driven circulation via the Sverdrup relation (Gill 1982). The changes in the curl are averaged over the subpolar gyre.

The wind forcing in the Southern Ocean is assessed via the zonally-integrated eastward wind stress, which is directly linked to the surface Ekman flux (Toggweiler and Samuels 1998; LaCasce and Isachsen 2010):

$$\begin{aligned} M_e = \frac{1}{\rho _c f(\theta )} \int _{-\delta }^0 \int _0^{360} \tau ^x \; r_e cos(\theta ) d \phi dz \end{aligned}$$
(5)

where \(\delta\) is the depth of the Ekman layer and \(\phi\) is the longitude. We integrate the winds at \(\theta =55^{\circ }\)S, a latitude unblocked by continental boundaries.

3 Results

3.1 Overturning circulation changes in CMIP6

The 50-year mean global overturning streamfunction from the pre-industrial control run in the Norwegian Earth System Model, version 2.0 (NorESM2-LM, Seland et al. 2020; “NorESM2" hereafter) is shown in Fig. 1a. The shallower cell, in the upper 1000 m, has a maximum transport of 28 Sv (Sv = \(10^6\) m\(^3\)/s) and extends from the surface to just below 2000 m. The largest portion of this is in the Atlantic, the AMOC, with a maximum transport of 24 Sv. The deeper Antarctic Bottom Water (AABW) cell is also seen, as is the surface-intensified “Deacon Cell" between 40 and 60°S.

The corresponding 50-year mean overturning 100 years after the CO2 quadrupling is shown in panel (b). The MOC has weakened greatly, to about 8.9 Sv, and is shallower. The AABW has also weakened substantially, indicating there is no bipolar seesaw compensation (Broecker 1999). On the other hand the Deacon Cell has actually strengthened somewhat, as discussed below.

Fig. 1
figure 1

The 50-year mean global meridional overturning mass streamfunction (Sv) for a the control run and b the 4xCO2 run after 100 years in NorESM2. c shows the difference between the two. The red (blue) colour shows a clockwise (anti-clockwise) circulation

Full size image

The difference between the two overturning streamfunctions is shown in Fig. 1c. The AABW cell is nearly 10 Sv weaker, but the shallower cell declines even more, by nearly 20 Sv. The maximum difference is centered near 40°N and at a depth of around 1000 m. The location of maximum weakening and the shallowing of the shallow cell are consistent among the models. The observed change is dominated by the decline in AMOC.

To compare the change in AMOC among the models, we calculate the maximum decrease between 20–50°N and 450–3000 m depth. The cell decreases in all models, but to differing degrees. The decline ranges from 5 to 22 Sv (Table 1), corresponding to a 15% to fully an 80% weakening (Fig. 2b). The GISS-E2-1-G simulation exhibits the largest reduction in overturning while the INM-CM4-8 run the least. The NorESM2 run (the orange dot) lies between these, with about a 60% decrease.

Fig. 2
figure 2

a Time series of annual-mean maximum Atlantic Meridional Overturning Circulation strength (AMOC) in 4xCO2 simulation relative to its initial value in CMIP6. The thick black curve shows the multi-model mean. The mean percentage change in maximum AMOC strength in 4xCO2 experiment relative to control run for the period 100–150 years for CMIP6 models is shown in (b). NorESM2 is highlighted in orange in both plots

Full size image

The temporal variation also varies between models (Fig. 2a). While AMOC reaches its minimum after only 20 years in some models, it takes 80 years in others. The AMOC in NorESM2 (orange curve) reaches its minimum around 40 years, recovers slightly and then settles thereafter. Given this, we use the period from 100 to 150 years to calculate averages for the perturbed model state.

3.2 Correlation with changes in wind stress

In most models, the winds change following the increase in CO2. But the changes are inconsistent, increasing in some cases and decreasing in others.

Consider the response in NorESM2. The wind stress curl in the control run is positive over the subpolar gyre and negative over the subtropical gyre (the solid and dashed contours in Fig. 3a). Following the increase in CO2, the curl weakens south of Greenland and over the Irminger and Labrador Seas, but strengthens south of the Gulf Stream (color contours in Fig. 3a). The other models behave similarly. All models exhibit weakening in the subpolar gyre, but the extent varies greatly.

To see whether the changes are correlated with those of AMOC, we averaged the difference in the curl over the subpolar gyre, i.e., 60°W–30°W, 50°N–60°N in all models. We then plotted the change in AMOC against the relative difference in curl (Fig. 3b) and calculated the Pearson correlation coefficient. The two measures are uncorrelated, with a correlation coefficient not significantly different from zero (based on the T test).

Fig. 3
figure 3

a Spatial change in wind stress curl in 4xCO2 simulation in NorESM2. Contours are superimposed to show the mean pre-industrial windstress curl. Mean change in wind stress curl in the subpolar gyre in 4xCO2 experiment relative to control for the period 100–150 years for CMIP6 models is shown in (b). Contours show \(0.8\times 10^{-7}\) N/m\(^3\) intervals. The Pearson correlation coefficient (r) is shown in blue, with the significance level (p) from the T test. The numbers indicate CMIP6 models as listed in Table 1

Full size image
Fig. 4
figure 4

a Spatial change in zonal wind stress in the Southern Ocean in 4xCO2 simulation in NorESM2. Mean change in the zonal wind stress in the Southern Ocean along Antarctic Circumpolar Current (ACC) in 4xCO2 experiment relative to control for the period 100–150 years for CMIP6 models is shown in (b). Contours show \(3\times 10^{-2}\) N/m\(^2\) intervals. The Pearson correlation coefficient (r) is shown in blue, with the significance level (p) from the T test. The numbers indicate CMIP6 models as listed in Table 1

Full size image

As noted, AMOC has also been linked to wind-driven Ekman transport in the Southern Ocean, following (Toggweiler and Samuels 1998). Thus, we also compared changes in the wind stress in the Southern Ocean with those in AMOC. The change in the zonal wind stress in the Southern Ocean in NorESM2 is shown in Fig. 4a. The winds have significantly strengthened in the circumpolar belt, and particularly in the South Pacific south of Australia (see also Chen et al. 2019). This results in a substantial increase in the northward Ekman transport (roughly 17 Sv in NorESM2).

We averaged the change in the zonal wind stress, integrated zonally at 55°S, from all models and compared the result to the change in AMOC (Fig. 4b). In most models the integrated stress increases. The correlation with the change in AMOC however is weak and negative (r = − 0.28, p = 0.14).

Thus, the change in Ekman transport in the Southern Ocean is unrelated to the decrease in AMOC in these simulations. The former instead causes an increase in the Deacon Cell, as seen in Fig. 1c. Indeed, that the Deacon Cell strengthens as AMOC weakens indicates the two are dynamically distinct over the adjustment period.

3.3 Correlation with changes in surface buoyancy

The surface density also changes in response to the forcing, and in a consistent way among the models. The change in the 50-year mean surface density from NorESM2 is shown in the left panel of Fig. 5. The surface density decreases dramatically north of the separated Gulf Stream (which we refer to as the NAC hereafter, for brevity) (Fig. 5). The greatest decreases are in the Labrador Sea, near Fram Strait and along a path extending from the east coast of North America to the tip of Greenland. The changes in the subtropical North Atlantic are weaker, with a slight increase over much of the region.

Fig. 5
figure 5

Change in mean a surface density in 4xCO2 experiment for NorESM2 relative to control simulation, due to b surface temperature and c surface salinity for the period 100–150 years. Contours show 1 Kg/m\(^3\) intervals. The box denotes the area to the north (solid blue) and south (dotted black) of NAC where the density difference is calculated in Sect. 3.3

Full size image

The changes in the subpolar gyre are due largely to salinity (Fig. 5c). The surface waters are fresher in the western subpolar gyre, extending south from Greenland and along the path of the North Atlantic Current. They are also fresher in the Labrador Sea and in the northern Nordic Seas. The salinity-induced change in the subtropical gyre on the other hand is positive, due to increased evaporation.

At the same time, the subtropical gyre is warming (Fig. 5b). The resulting decrease in density nearly balances the increase due to evaporation, yielding relatively small changes in total density (see also Levang and Schmitt 2020b). In the northern subpolar gyre, there is cooling south of Greenland and Iceland. This is the so-called “cold blob" or “warming hole” (Drijfhout et al. 2012; Rahmstorf et al. 2015; Liu et al. 2020; Li et al. 2022; Ma et al. 2020). But the resulting increase in density is overwhelmed by the freshwater fluxes (Fig. 5c) so that the overall density change is negative.

Similar changes are seen in the other models. To test the relation with AMOC, we calculated the density difference between the subpolar and subtropical gyres, in the regions (50°W–40°W, 40°N–50°N) and (50°W–40°W, 30°N–40°N) (Fig. 5). This roughly captures the meridional density gradient across the NAC. We decomposed the density differences into temperature and density contributions and calculated the mean values in the perturbed experiment during the period 100–150 years. Those values were then differenced with the corresponding values from the control run and compared to the decrease in AMOC between the runs during the same period.

AMOC weakening is correlated with the change in surface density, with a correlation coefficient of r = 0.531 (p = 0.004) (Fig. 6a). Models which exhibit the largest decrease in the meridional density gradient exhibit the greatest weakening. The salinity-induced density change is similarly correlated (r = 0.525, p = 0.004) (Fig. 6c). The temperature-induced change on the other hand is insignificantly anti-correlated with AMOC change (r = − 0.125, p = 0.52). The correlations are negatively affected by a few outliers, in particular the models of NASA-GISS and MRI. Removing these two increases the correlation with both the surface density change (r = 0.75, p = 0.00005) and the salinity-induced change (r = 0.733, p = 0.0001) while leaving the relation to the temperature-induced gradient essentially unchanged.

Fig. 6
figure 6

Change in maximum AMOC strength for the period 100–150 years in the 4xCO2 experiment relative to the control simulation plotted against a the change in the surface density difference across the NAC, b the corresponding density change due to temperature and c the density change due salinity. The Pearson correlation coefficients (r) and significance levels (p) from the T test are inserted at the bottom right. As in Fig. 3, the numbers indicate CMIP6 models as listed in Table 1

Full size image
Fig. 7
figure 7

Changes in surface density (Kg/m\(^3\)) (a), ocean currents (cm/s) at the surface (b), and Atlantic meridional overturning streamfunction (Sv) (c). Contours show 1 Kg/m\(^3\) and 10 Sv intervals in (a) and (c)k, respectively. The change is calculated by subtracting the control experiment 50-year means for each of the variable from the respective five-year means from the 4xCO2 experiment. Year 1 is the year when \(CO_2\) is abruptly quadrupled

Full size image

The sequence of density changes in the North Atlantic from NorESM2 are shown in 5 year time averages in the Fig. 7a. Of particular interest is the tongue of light water which extends from Atlantic Canada into the interior, just north of the NAC. This is apparent even in the first 5 years, and intensifies thereafter.

Simultaneously, the velocities in the NAC weaken (Fig. 7b). During the first 5 years, the velocity decreases south of the NAC and strengthens to the north. Subsequently the weakening in the core of the NAC becomes evident. During the years 6–10, the decrease in velocities stretches back to the Florida Straits. The weakening intensifies after this, suggesting a near collapse of the NAC. Note too that the cyclonic flow in the eastern subpolar gyre and the southward flow east of Greenland both intensify.

Fig. 8
figure 8

Change in the mean zonal velocities from the vertically-integrated thermal wind (a). The change in the surface currents from the model are shown in (b) for comparison. Both fields are from NorESM2 from the period 16–20 years after the quadrupling of CO2. Contours show 4 cm/s intervals

Full size image

The AMOC changes mirror those in the surface velocities (Fig. 7c). There is a weak adjustment during the first 5 years, except at the surface near the equator. The cell begins weakens dramatically subsequently, from years 5–10 and centered at 40°N. From years 10–20, the weakening intensifies and spreads with latitude, connecting to the weakened cell in the equatorial band.

The three fields are related. The tongue of freshwater which extends into the interior decreases the meridional density gradient across the NAC. This implies a weakening of its zonal velocity shear, \(\partial u/\partial z\), from the thermal wind relation (Eq. 3) and hence a decrease in the surface velocities. To illustrate this, we integrate the thermal wind relation from the ocean bottom, assuming no flow there, and compare the resulting zonal velocity differences with those obtained in the model. The two fields, from the period 16–20 years, are nearly identical (Fig. 8). Thus the reduction in the meridional density gradient across the NAC directly reduces its transport.

Initially the freshwater tongue weakens the density gradient on its southern side but strengthens the gradient to the north. This is the reason for the dipolar response during the first 5 years, with the speeds increasing north of the separated Gulf Stream and decreasing to the south (Year 1–5 of Fig. 7b). As the lighter water spreads into the subpolar gyre, the dominant effect is the weakened gradient to the south, decreasing the NAC transport.

Fig. 9
figure 9

The zonally-integrated meridional transport at 40°N as a function of depth (left panel) in NorESM2, and the corresponding overturning streamfunction (right panel). The transports have been converted to Kg/s, using the ocean density. The streamfunction curves are cumulative sums of the left curves, integrating downward from the surface. These are compared to the model’s own streamfunction curves, plotted as the dashed curves in the right panel

Full size image

The weakened zonal transport implies a decreased meridional flow towards the subpolar gyre, and this directly impacts the AMOC cell. This is reflected in the overturning streamfunction at 40°N (Fig. 9), obtained by integrating the meridional velocity zonally and vertically, as in equation 3. The meridional mass transport changes most dramatically at the surface, even reversing sign in the 4xCO2 run. But the subsurface transport weakens as well, both northward at 500 m depth and southward below 1000 m.

As the velocities are integrated downward, the changes in the upper 500 m are mirrored deeper down. Note the overturning streamfunction goes to zero at depth, so that the southward flow, which is primarily in a deep western boundary current, is weakening simultaneously. This is due to incompressibility; if only the inflow were decreasing, there would be a mass deficit at high latitudes.

Thus as the density gradient decreases at 40°N, the NAC transport declines, and the AMOC streamfunction mirrors this. As the velocity deficit spreads to the Florida Current, which is shallower, the AMOC weakening spreads southwards. And as the inflow to the subpolar gyre weakens, AMOC changes there as well.

3.4 Freshwater sources

The central element in AMOC weakening in these simulations is then the spreading of freshwater from the western boundary. The freshwater fluxes come from land runoff, precipitation and melting sea ice.

Fig. 10
figure 10

a Freshwater flux components for the 150 years after increasing CO2, from runoff (dotted blue), sea ice (orange) and precipitation minus evaporation (green dash-dot). The sum is shown by the dotted red curve. b The total freshwater flux (red dots) plotted with the decrease in AMOC (black dash-dot). The freshwater fluxes have been smoothed with a 5 year running average for clarity. The first 2 years show annual values

Full size image

Time series of the fluxes north of 60°N in NorESM2 during the 150 years after the increase in CO2 are plotted in Fig. 10a. Initially, the largest components are from sea ice (in orange) and runoff (dotted blue), while the contribution from precipitation minus evaporation (dotted green) is less than a third as large. During the first 10–20 years, the sea ice flux declines precipitously, to approximately the same level as P–E. This is due to the widespread loss of sea ice under the warming, primarily in the Arctic. In contrast, the runoff increases somewhat and then levels off, and P–E is nearly constant. The net effect is that the total freshwater flux decreases during the first 20 years and fluctuates around a constant value thereafter.

The decrease in freshwater fluxes occurs simultaneously with the decline in AMOC (Fig. 10b). The flux (smoothed with a 5 year moving filter) decreases during the first 20 years, as does AMOC. When the fluxes reach their new equilibrium thereafter, the AMOC curve flattens out. Thus the large release of freshwater primarily from sea ice melting and coincides with the AMOC weakening.

Fig. 11
figure 11

Change in sea surface salinity in NorESM2 for the first 20 years after CO2 is quadrupled for NorESM2-LM. The change is calculated by subtracting the control experiment 50-year mean from the respective annual means from 4xCO2 experiment

Full size image

The 20 year adjustment period in part reflects the time required for freshwater to exit the Nordic and Labrador Seas and enter the subpolar gyre. This occurs in stages, as shown in Fig. 11. Even at year 2 (panel a), anomalous freshening is seen in the subpolar gyre. This can be traced to freshwater from the Gulf of St. Lawrence and, to a lesser extent, the Labrador Sea. By year 4, the northern Labrador Sea and the region east of Greenland are significantly fresher, and the signal is established north of the NAC by year 6. The connection with the Labrador Sea is clear by year 10, intensified along the western boundary. The connection with the Nordic Seas is evident later, after year 12. All sources feed the western boundary current and enter the interior north of the NAC.

The evolution is similar in the other models, although the exact route taken by the freshwater varies. In some cases, freshwater surges directly south from the Nordic Seas, flushing the northern subpolar gyre. But more typically the water moves in the western boundary current.

The diversity of weakened AMOC states seen in Fig. 2 reflects differing amounts of sea ice melt. In models with substantial Arctic sea ice melt, the AMOC is weakened the most. Increased meltwater production freshens the subpolar gyre to a greater extent, weakening the density contrast across the Gulf Stream and hence the meridional flow of the NAC.

4 Discussion

The models in the CMIP6 ensemble exhibit consistent weakening of AMOC when CO2 is abruptly quadrupled, with a multi-model average of 52%. The increase in CO2 triggers extensive Arctic warming, which induces widespread melting of sea ice. This results in a large freshwater fluxes to the ocean during the first decade. The freshwater flows south along the eastern coast of North America and then eastward, north of the Gulf Stream. This decreases the density of the near-surface waters in the subpolar gyre. At the same time, the density changes little in the subtropical gyre, as surface warming is largely balanced by increased evaporation. Thus the density gradient across the North Atlantic Current is decreased, due to the fresher subpolar gyre. This weakens the vertical shear in the NAC and consequently its transport.

The CMIP6 models vary greatly in the extent of AMOC weakening, from 15 to 80%. This spread reflects the different freshwater production in the models. As the summer sea ice extent is reduced to zero in nearly all the models, the difference likely reflects the range of ice present initially. The exact route the freshwater takes when entering the subpolar gyre is also important. This causes deviations in the timing of AMOC weakening, from roughly 20 to 80 years (Fig. 2a). Changes in local (subpolar) and remote winds were uncorrelated with AMOC weakening in these models. That is not to say wind forcing is unimportant; sub-decadal variability, which is pronounced, owes much to wind forcing. But the response to the drastic forcing in this scenario is buoyancy-driven in these simulations.

Dai (2022) noted the connection between Arctic amplification and a weakening AMOC, in simulations in which CO2 was increased by 1% per year. That study linked the decrease in overturning to the meridional density gradient, as argued here, but also to reduced mixed layer depths, as claimed in other studies. In contrast, de Boer et al. (2010) contended that AMOC strength and the meridional density gradient were poorly correlated; the gradient in their case was between the equator and North Atlantic however, rather than across the NAC. The authors also found a positive correlation with the wind stress in the Southern Ocean, which was absent in the present simulations. Levang and Schmitt (2020a) concluded that salinity was less important for AMOC weakening in CMIP5 simulations than temperature. Their focus was on the zonal density gradient, at depths greater than 1000 m. They noted too that the increase in density due to evaporation was offset by warming. But as seen here, that applies to the subtropical gyre; changes in the subpolar gyre are dominated by salinity, and this is what correlates with the weakened NAC in CMIP6. And Bonan et al. (2022) concluded that subsurface temperature changes in the deep water formation regions were responsible for AMOC weakening, in millennial length simulations under quadrupled CO2. Salinity changes in the same regions preceded AMOC recovery in some models. We did not examine whether such changes occur in the present simulations.

The present study supports the relevance of “hosing experiments", in which freshwater is introduced at high latitudes (Mecking et al. 2016; Jackson et al. 2016; Haskins et al. 2018; Jackson et al. 2022). In these, freshwater is usually added uniformly, typically over the entire subpolar gyre. In the North Atlantic hosing model intercomparison project (NAHosMIP), 0.3 Sv of freshwater was added uniformly north of 50°N for 50–100 years, and this led to AMOC weakening in all the models. The present simulations differ because the freshwater in most cases follows the western boundary before entering the subpolar gyre north of the Gulf Stream separation. If applied uniformly, the impact on the meridional density gradient could differ. A more realistic approach might be to apply hosing solely in the Labrador Sea, for example.

The connection between freshwater production at high latitudes and AMOC weakening has been made before (see Weijer et al. 2022 and references therein). But the reason usually given is that freshwater “caps the convective regions", preventing deep water formation. We posit instead that freshwater weakens the inflow to the high latitude regions, by altering the meridional density gradient. This view is consistent with the study of Gjermundsen et al. (2018), in which the AMOC and AABW cells were successively weakened in an ocean GCM simply by altering the surface density gradient. Consistently, the extent of AMOC weakening in the CMIP6 models depends on the volume of freshwater introduced to the subpolar gyre, which explains the large spread of AMOC weakening among the models. Such a connection has not been made previously with respect to the convective hypothesis; indeed, it is unclear how different degrees of “capping" could reduce AMOC by different amounts.

Fig. 12
figure 12

50-year-mean vertical ocean mass transport at 1000 m in piControl (a) and 4xCO2 (b) 100 years after after the quadrupling of CO2 from NorESM2. Dotted black contours show mean of the maximum ocean mixed layer depth define by sigma T for the same period for the two experiments. The contour interval is 300 m in (a), but it is reduced to 60 m in (b) for clarity. Gray contours show 2000 m isobaths the model bathymetry

Full size image

The focus on the convective regions began, arguably, with the idealized models of Stommel and Arons (1959), which examined the abyssal circulation resulting from specified deep water sources in different locations. In more realistic settings, dense water is formed in boundary currents rather than in the centers of the gyres in the Labrador, Greenland and Norwegian Seas, where convection is typically observed (Mauritzen 1996; Spall and Pickart 2001). This is the case in the present simulations as well. The mixed layer depths in the Labrador and Irminger Seas decrease as AMOC weakens (Fig. 12), due primarily to warmer air temperatures. But here too the sinking occurs primarily along the boundaries (left panel), adjacent to the regions with the deepest mixed layers. Indeed, the mixed layer depths also decrease in the Nordic Seas, but the exchange across the Greenland-Scotland ridge is very modest in these models, implying the convection in the Nordic Seas is local and seasonal. Downwelling thus appears to be part of the larger scale circulation stemming from the surface meridional density gradient, as suggested by Gjermundsen and LaCasce (2017), Gjermundsen et al. (2018).

How long does the weakened state persist? In the present 4xCO2 simulations, AMOC remains weak in all models for at least 150 years. In contrast, in half the simulations in the NAHosMIP study, AMOC recovered after the freshwater hosing ceased (Jackson et al. 2022). Following the present arguments, the weakened state should persist as long as the meridional density gradient remains weak. If the excess freshwater in the subpolar gyre were exported, the gradient could presumably strengthen again (Liu et al. 2014). This could trigger AMOC recovery, like that described by Bonan et al. (2022). The NAC is a barrier to mixing, hindering such export, but this presumably occurs over longer time scales. We are currently examining this, using longer simulations.

There are numerous caveats about the CMIP6 models. Dense water formation occurs primarily south of Greenland, rather than in the Nordic Seas, for one. The models also employ a range of different sea ice models, which will impact the freshwater input to the ocean. Nevertheless, the link between AMOC and the meridional density gradient across the NAC is dynamically plausible and consistent with previous studies with more idealized ocean models. The results also suggest a straightforward way to model future changes in AMOC, by monitoring salinity north of the NAC.