1. Introduction

The thermohaline circulation (THC) transports about 1 PW (=10¹⁵ W) of heat to the North Atlantic (Hall and Bryden 1982; Ganachaud and Wunsch 2000, 2003), having a major effect on atmospheric and oceanic temperatures in northern high latitudes (e.g., Manabe and Stouffer 1988; Wood et al. 1999). In addition to these regional effects, the overturning circulation influences climate on a global scale (Winton 2003; Herweijer et al. 2005). Because of the associated heat transport, the thermal equator is shifted northward over the Atlantic and eastern Pacific. The intertropical convergence zone follows this excursion, which leads to altered precipitation patterns compared to a climate state without overturning (Vellinga and Wood 2002; Stouffer et al. 2006). The existence of an interhemispheric overturning circulation yields both a sea surface elevation that is some decimeters lower in the North Atlantic compared to the North Pacific (Levermann et al. 2005) and a cooling of the Southern Hemisphere (Crowley 1992; Stocker 1998). It has strong implications for the marine ecosphere (e.g., Schmittner 2005) and possible teleconnection with the El Niño–Southern Oscillation (ENSO) phenomenon (Timmermann et al. 2005). Combined evidence from proxy data (e.g., McManus et al. 2004) and simulations (e.g., Manabe and Stouffer 1997) suggests that large and abrupt changes in THC have led to major climate shifts during the last glacial period [for a detailed discussion see the reviews by Clark et al. (2002) and Rahmstorf (2002)]. Past cessations of the Atlantic overturning have been associated with massive freshwater release from glaciers to the ocean (Bond et al. 1992).

In light of future climate change, an understanding of the stability properties of the THC is of particular importance. Projections of future climate indicate a crucial role of Greenland meltwater flux for the evolution of the THC (Fichefet et al. 2003). However, some models show a pronounced reduction of overturning that is not due to meltwater inflow. For example, the atmosphere–ocean general circulation model (AOGCM), ECHAM3/LSG, exhibits a complete collapse of deep-water formation for 1% of yearly CO₂ increase and a subsequent stabilization at 1120 ppm (Winguth et al. 2005). Though the model was coupled to a land ice model and therefore included Greenland meltwater flux, the slowdown of overturning was induced mainly by an increase in precipitation in the North Atlantic of ∼0.2 Sv (1 Sv = 10⁶ m³ s⁻¹) over a period of several hundred years. On shorter time scales, most coupled climate models without Greenland meltwater inflow do not show a complete THC collapse. In recent intercomparison projects (Gregory et al. 2005; Petoukhov et al. 2005), this weakening was gradual in all participating models and varied between 10% and 50% of the initial value.

On decadal or shorter time scales, changes in THC will be strongly influenced by changes in North Atlantic buoyancy flux because of their impact on deep-water formation. Saenko et al. (2004) examine the role of northern sea ice cover for the overturning circulation during global warming experiments by altering the thermal diffusion coefficient in their atmospheric energy–moisture balance model, thereby producing varying temperature and sea ice extent in northern high latitudes. Their main conclusion is that the initial climate around the subpolar gyre is crucial for understanding the weakening of the THC under global warming. In their experiments with moderately cold initial temperatures in the north, they observe a weakening of the THC, while it strengthens when northern high latitudes are unrealistically cold and sea ice extends farther south.

Because of a lack of observational data, limitations of currently available climate models, and natural variability, the present strength of the Atlantic meridional overturning circulation is poorly constrained (Ganachaud and Wunsch 2000, 2003; Talley et al. 2003; Zickfeld et al. 2007). The maximum overturning of the participating models of the Coupled Model Intercomparison Project (CMIP; http://www.gfdl.noaa.gov/~kd/CMIP.html) varies between ∼12 and ∼27 Sv and the Third Assessment Report of the Intergovernmental Panel for Climate Change (IPCC TAR; Houghton et al. 2001) gives a range of even ∼10 to ∼30 Sv between the participating models. In view of this uncertainty, the dependence of the sensitivity of the THC to an increase in atmospheric CO₂ on its initial strength is an important issue. Using the Earth System Models of Intermediate Complexity (EMICs) CLIMBER-2 and CLIMBER-3α and results from the CMIP simulations (section 2), we investigate this dependence (section 3) and propose a physical mechanism based on the stabilizing effect of changes in sea ice cover on the THC (section 4). The results are consistent with simulations carried out with a number of AOGCMs in the framework of the IPCC Fourth Assessment Report (IPCC AR4). Section 5 qualitatively discusses the phenomenon in physical terms. Conclusions and a discussion, particularly with respect to a generalization of the mechanism to other models, are given in section 6.

2. Models and experiments

To understand the influence of the initial strength of the overturning on its change under global warming, 14 simulations using the EMICs CLIMBER-2 and CLIMBER-3α were carried out, as a complement to the CMIP runs. CLIMBER-2 is a fully coupled climate model composed of a statistical–dynamical atmosphere module and zonally averaged ocean and sea ice models. The sea ice module is based on the one-layer thermodynamic (Semtner 1976) model with simple treatment of advection and diffusion. It calculates the sea ice fraction and thickness based on the sea ice mass balance and a relation between ice thickness and fraction. The ice is freely transported (no ice rheology) by advection and horizontal diffusion. CLIMBER-2 has been used to simulate present day climate (Petoukhov et al. 2000) and the last glacial maximum (Kubatzki et al. 2000), as well as abrupt climatic events of the past (Ganopolski and Rahmstorf 2001) and future scenarios (Rahmstorf and Ganopolski 1999). With its coarse resolution and therefore fast performance, the model is suited for very long integrations including equilibria and large ensemble simulations (e.g., von Deimling et al. 2006).

The EMIC CLIMBER-3α uses the same atmospheric module as CLIMBER-2 with a higher horizontal resolution and explicit topography. The ocean component is based on the Geophysical Fluid Dynamics Laboratory (GFDL) OGCM Modular Ocean Model (MOM-3) with a horizontal resolution of 3.75° × 3.75° and 24 variably spaced vertical levels. The sea ice model was adopted from Fichefet and Maqueda (1997). It is a two-dimensional (latitude–longitude) thermodynamic–dynamic model with one layer of ice and one layer of snow. It takes into account the heat capacity of snow and ice, the thermodynamic effects of the subgrid-scale snow and ice thickness distributions, the storage of latent heat in brine pockets, and the formation of snow ice. The sea ice dynamics allows the sea ice to flow under the effect of winds and currents and to self-interact. The dynamics employs an elasto–visco–plastic rheology formulation, thereby conserving mechanical energy. For a full description of CLIMBER-3α, see Montoya et al. 2005; the sensitivity of the model to vertical diffusivity was investigated by Mignot et al. (2006).

Both of these EMICs took part in the CMIP study together with six AOGCMs and three other EMICs. A complete list of the participating models is given in Table 1 of Stouffer et al. 2006. The CMIP contained a number of idealized experiments designed to investigate the models’ response to a quadrupling of atmospheric CO₂ (Gregory et al. 2005) and an anomalous freshwater flux to the northern Atlantic (Stouffer et al. 2006). In this study we concentrate on the scenario runs with a 1% increase per year of atmospheric CO₂ concentration over 140 yr up to 1120 ppm (i.e., 4 times the preindustrial value of 280 ppm; denoted “standard experiment” hereafter). It should be noted that none of these experiments included meltwater from continental ice, which could become a major source of freshwater if the Greenland Ice Sheet melts rapidly (Alley et al. 2005; Oppenheimer and Alley 2005; Hansen 2005).

The initial strength of the overturning circulation in the participating CMIP models varies between ∼12 and ∼27 Sv. To investigate the importance of this initial scatter, we performed additional experiments. The initial strength of the THC was varied by application of a constant anomalous freshwater flux to the Atlantic Ocean surface both in CLIMBER-2 and CLIMBER-3α. The flux was applied in the North Atlantic between 50° and 70°N and compensated in the Pacific between 45°S and 55°N. With anomalous freshwater fluxes between −0.15 and +0.09 Sv, we obtained 11 initial equilibria with a range of 16.6–26.0 Sv of maximum overturning in CLIMBER-2. The standard value is 21.6 Sv. Anomalous freshwater fluxes of −0.1 and −0.2 Sv in CLIMBER-3α resulted in THC maxima of 13.2 and 16.2 Sv, respectively, compared to the standard value of 11.8 Sv without anomalous freshwater forcing. This way, we cover the entire interval of the CMIP simulations.

To test the robustness of the physical mechanism proposed in this study (section 4), we use simulations that were carried out in the framework of the IPCC AR4. In particular, we analyze simulations of AOGCMs, for which overturning strength and sea ice concentration for the standard experiment is provided (www-pcmdi.llnl.gov/ipcc/about_ipcc.php; sea ice data are not available for the CMIP.). Table 1 gives the simulations used here.

3. THC sensitivity to atmospheric CO₂ increase: Dependence on initial overturning strength

Figure 1 shows the time evolution of the overturning strength in CLIMBER-2 (black) and CLIMBER-3α (colors) runs with different initial THC strength for the standard experiment. The CO₂ concentration is shown as the dashed curve. As a consequence of increased CO₂ forcing, all simulations show a rapid reduction of the THC. When the atmospheric CO₂ concentration is kept constant, the THC gradually recovers. After the final CO₂ concentration is reached (140 yr after the onset of the scenario), the interval of initial volume transport M₀ ∈ [11.8, 26.0] Sv reduces to values M₁₄₀ ∈ [9.6, 17.4] Sv and becomes M₂₀₀₀ ∈ [13.6, 20.4] Sv in the final equilibrium (2000 yr after the start of the scenario). The shrinking of the interval with time from 14.2 via 7.8 Sv to 6.8 Sv reflects the fact that the CO₂ sensitivity of the THC depends on its initial strength M₀. Figure 2 gives the THC weakening ΔM_t = M₀ − M_t as a function of M₀, after t = 70 yr (top panel), t = 140 yr (middle panel), and for the final equilibrium ΔM₂₀₀₀ (bottom panel). For each of these time slices, the results of the simulations with CLIMBER-2 (black dots) and CLIMBER-3α (colored dots) exhibit a quasi-linear relationship of ΔM_t with the initial THC strength M₀ (correlation coefficients of the linear regression r ∈ [0.97, 0.99]). The other CMIP simulations (red crosses) and the IPCC AR4 simulations (blue crosses) scatter around these lines. Linear regression yields a correlation coefficient r = 0.63 for a fit of the CMIP runs alone and r = 0.60 for the CMIP and IPCC AR4 experiments. The correlation coefficient increases to r = 0.72 when all available simulations in the middle panel of Fig. 2 are taken into account. Simulations that were carried out in the framework of the EMIC Intercomparison Project (EMIP; Petoukhov et al. 2005) exhibit a similar scatter around the CLIMBER simulations (top panel, green crosses). Here the 1% CO₂ increase scenario was integrated only up to a doubling of the preindustrial concentration, 560 ppm.

Note that in the course of the scenario, the overturning eventually becomes stronger in the two CLIMBER-3α simulations with weakest initial THC. Consequently, the line formed by the CLIMBER equilibrium simulations (Fig. 2, bottom panel) crosses the x axes ΔM_t = 0 Sv for finite initial overturning strength M⁽⁰⁾₂₀₀₀ = ∼15 Sv. For the two earlier time slices, the transition toward a strengthening occurs for M⁽⁰⁾₁₄₀ = ∼11 Sv and M⁽⁰⁾₇₀ = ∼8 Sv (Fig. 2, middle and top panel). Note also that the linear behavior of ΔM_t with M₀ occurs throughout the entire scenario and is not a specific feature of the selected time slices. For warming surface air temperature (SAT), that is, in the course of the CO₂ scenario, the slope of the curves increases (cf. top and middle panel of Fig. 2). After 140 yr (i.e., when CO₂ is fixed to 1120 ppm), the slope slightly declines and stabilizes at 0.51 (full evolution with time not shown).

4. The role of initial THC strength and northern sea ice cover

In this section we propose a physical mechanism that leads to the qualitative behavior seen in Fig. 2. In the CMIP, the standard experiment was complemented by runs, where freshwater fluxes to the ocean in the control and scenario simulation were switched (Gregory et al. 2005; cf. Fig. 3). This experimental setup provides a linear approach to determine the comparative influence of changes in freshwater and heat forcing during the scenario. Gregory et al. (2005) showed that in the CMIP simulations, the changes in heat flux dominate the weakening of the THC in all models (see their Fig. 4). The freshwater fluxes include evaporation, precipitation, and river runoff as well as fluxes from production and the melting of sea ice. The dominance of heat flux changes is particularly pronounced in CLIMBER-2 (Fig. 3), while in other models freshwater fluxes may have a stronger influence. Focusing here on the ensemble of CLIMBER-2 simulations, we will therefore propose a purely heat flux–related mechanism, while keeping in mind that the influence of freshwater flux plays a secondary, but possibly nonnegligible role in other models. Limitations of a generalization of the mechanism to other models and the real climate system will be discussed in section 6.

a. Primary effects of CO₂ increase on THC strength

In boreal winter, the atmosphere in high northern latitudes is much colder than the freezing point. Thus, the area of extreme oceanic heat loss follows closely the ice-free area A, while heat loss is relatively weak underneath sea ice. This is illustrated in Fig. 4, where heat flux from atmosphere to ocean and sea ice fraction are plotted as functions of latitude for different initial THC strength in CLIMBER-2. Shown are zonal and annual mean at the beginning of the scenario (t = 0) for the North Atlantic. The sea ice edge and the area of intense oceanic heat loss retreat northward with increasing M₀ due to stronger northward heat transport by the overturning. In contrast, the southern border of negative heat flux (∼20°N) is independent of the initial THC strength M₀.

The atmosphere–ocean heat flux is composed of latent, radiative, and turbulent heat fluxes, which all depend on the ice-free area A and temperature difference between ocean and atmosphere ΔT ≡ T_o − T_a. To first order, the heat loss of the ocean H is thus proportional to the product of the two quantities:

View Expanded

where α_H is a positive proportionality constant. Note that other factors such as shortwave radiation may influence H. We restrict our discussion to situations where Eq. (1) holds. A reduction in oceanic heat loss H decreases the oceanic surface density in the northern convection sites and leads to a weakening of North Atlantic Deep Water (NADW) formation and THC strength. We will now successively examine the effect of changes in temperature and sea ice cover on H in the course of the standard experiment.

Let us first consider the effect of temperature changes. Figure 5 shows the time series of the difference between oceanic and atmospheric temperature in January averaged over the entire North Atlantic between 20° and 80°N. The lines represent different CLIMBER-2 runs with varying initial THC strength. Note that for the sake of clarity we restrict our discussion to January (i.e., to the time of intense deep convection in the north), while similar though less pronounced results are obtained for averages over boreal winter or the entire year. The temperature difference ΔT ≡ T_o − T_a decreases with time in all simulations, mainly because of atmospheric warming (not shown). This reduces oceanic heat loss [Eq. (1)] and therefore NADW formation. A reduced overturning in turn transports less heat northward and thereby further reduces ΔT. This positive feedback loop (hereafter “temperature effect”) is schematized in Fig. 6: an increase in atmospheric CO₂ reduces ΔT, which reduces heat loss and THC strength and thereby ΔT.

Changes in near-surface temperature also affect the oceanic heat loss indirectly through the melting of sea ice. This effect is represented as the lower feedback loop in Fig. 6 (hereafter “sea ice effect”). Sea ice retreat due to atmospheric warming will expose a larger area of the ocean to the atmosphere (Fig. 7), which is significantly colder than sea ice in high northern latitudes. The ocean thus loses more heat (Fig. 4), which favors the formation of NADW and tends to counteract the direct temperature effect (i.e., it tends to stabilize the circulation by limiting the reduction in overturning). The time series of ice-free area between 60° and 80°N shows this increase due to rising atmospheric temperatures (Fig. 7). These are the two primary effects of increasing CO₂ concentration on the THC: reduction due to warming of the atmosphere and stabilization due to retreating sea ice.

Note that atmospheric temperature does not increase linearly with time in the standard experiment and neither does the ice-free area. The transient relationship between atmospheric warming and sea ice cover, however, is quasi-linear (cf. Fig. 8a). Such model behavior is consistent with simulations of more complex models, which seems to indicate a linear relationship between surface warming and sea ice extent in transient experiments (e.g., Rind et al. 1995; Gregory et al. 2002; Holland and Bitz 2003). We will now investigate the influence of the initial equilibrium on this transient response.

5. Qualitative discussion of the sign of the slope

Purely in the line of reasoning of section 4 (i.e., ignoring the possible role of freshwater fluxes), we can further discuss the robustness of the described mechanism. To better understand the sign of the slopes in Fig. 2, let us reconsider the schematic of Fig. 6. In section 4 we argue that the dependence on the initial strength of the THC enters through the initial ice-free area, which influences the rate of sea ice retreat under global warming. More initial sea ice cover retreats more strongly. The sign of the slope is then determined by the competition between the direct effect of the retreating sea ice on the heat loss and the indirect effect via the albedo change and the resulting change in temperature difference between ocean and atmosphere. This can be expressed mathematically in terms of derivatives, following the two respective paths in Fig. 6. In this view, the sign of the slope is determined by the dependence of the heat loss H on the ice-free area A, that is, by the sign of

View Expanded

with ΔT = T_o − T_a. The first term on the right-hand side describes the direct effect of reduced sea ice area on the heat loss; the second term gives the indirect effect through the change of surface albedo. Here α = [A × α_O + (A_O − A) × α_I]/A_O is the surface albedo averaged over the North Atlantic region of oceanic heat loss with a_I and α_O being the albedo of sea ice and ocean, respectively. A_O is the entire area of ocean heat loss in the North Atlantic irrespective of the sea ice cover. In the framework of section 4, that is the Atlantic area between 20° and 80°N. For

View Expanded

the full derivative in Eq. (2) is positive and so is the slope in Fig. 2. If we assume that Eq. (1) approximates the oceanic heat loss, we can simplify this relation to

View Expanded

We write here in order to emphasize that the temperature difference between ocean and atmosphere is averaged over the North Atlantic area of oceanic heat loss. Equation (4) is written for a specific time. To make a statement about the slope of the curves in Fig. 2, the relation has to hold throughout the simulation. Since A cannot be larger than A_O and (α_I − α_O) cannot be larger than 1, we can say that relation (4) holds when

View Expanded

This means that a strong initial THC weakens more under global warming as long as the temperature difference between ocean and atmosphere averaged over the North Atlantic area of oceanic heat loss is larger than the increase of this temperature difference due to a change of surface albedo from that of sea ice to that of the ocean. In this case the direct effect of ice melting dominates over the indirect albedo effect.

Note that the quantity ∂ΔT/∂α is highly model dependent and involves a number of complicated physical processes. The slope in Fig. 2 indicates that in CLIMBER-2 it is bounded by 4.8 K, the lower bound of in the simulations (Fig. 5). Indeed, experiments with CLIMBER-2 have shown that surface air temperature increases by only about 3 K when the albedo of sea ice is set artificially to that of the ocean (A. Ganopolski 2005, personal communication), consistent with the observed behavior.

6. Discussion and conclusions

Based on a number of model simulations with the EMIC CLIMBER-2 and an idealized CO₂-increase scenario (1% yearly increase up to 1120 ppm), we propose a sea ice–related process that tends to stabilize the THC under atmospheric warming. This mechanism is weaker for initially stronger THC and is identified to dominate the dependence of THC changes on the initial THC strength in CLIMBER-2.

The CMIP intercomparison project showed that a stronger initial THC weakens more under global warming. Experiments with CLIMBER-2 and CLIMBER-3α exhibit a linear relation between THC weakening and initial strength (section 3). The participating models of the CMIP, EMIP, and IPCC AR4 intercomparison projects scatter around this line formed by the CLIMBER simulations. For the CMIP experiments it was shown that the THC weakening is dominated by changes in heat flux as opposed to freshwater fluxes (Gregory et al. 2005). This effect is specifically pronounced in CLIMBER-2. We therefore use additional experiments with CLIMBER-2 to examine heat flux–related processes that are relevant under warming and investigate their dependence on the initial THC strength.

The THC weakening is primarily due to a warming of atmospheric temperatures (Fig. 11) and the associated reduction in oceanic heat loss in the North Atlantic. For CLIMBER-2, this effect is shown to be weaker for initially strong THC because of the ice albedo effect (Fig. 9). It thus cannot explain the described dependence that an initially stronger THC weakens more. A second important effect that occurs under warming is the melting of sea ice in the North Atlantic and Arctic Ocean. This leads to an increase in oceanic heat loss due to a larger ice-free area (i.e., main area, over which the ocean loses heat to the atmosphere). This effect tends to increase deep-water formation in the Nordic Seas and strengthen the THC. It is weaker for an initially stronger THC and was identified to dominate the overall dependence of the THC weakening on its initial strength in CLIMBER-2. We argue that this is the case as long as the temperature difference between ocean and atmosphere in the North Atlantic ΔT is larger than the potential change in ΔT due to a change in surface albedo from its oceanic value to its sea ice value (section 5).

As a consequence of the warming, the oceanic heat loss exhibits a decrease in North Atlantic regions that are not affected by sea ice cover (20°–60°N). On the other hand, it increases in regions of strong sea ice retreat (60°–80°N). The latter is more pronounced for initially larger sea ice cover, which is associated with a weak initial THC.

An important question is whether the mechanism found in CLIMBER-2 is relevant to other models. There are some indications. Saenko et al. (2004) already argued that northern sea ice cover is crucial for the behavior of the THC under warming. Even though there is no direct relation between THC weakening and initial strength in their simulations because of their specific experimental setup, their experiments are consistent with the mechanism presented here. In particular, they find a strengthening of the THC under warming in an experiment with unrealistically large initial sea ice cover. They show that in this case, the net buoyancy flux into the ocean further decreases in the course of the experiment, which was caused mainly by thermal fluxes. This would be consistent with a strong increase in ice-free area although this is not shown in their study.

Furthermore, the scattering of the CMIP, EMIP, and IPCC AR4 simulations around the CLIMBER experiments indicates that there is a general tendency for a stronger initial THC to weaken more under warming. The presented mechanism shows this qualitative behavior, while a mere change in atmospheric temperature without additional heat loss in melting regions would lead to the inverse dependence (ice albedo feedback). Note further that some of the scatter in Fig. 2 is likely due to internal variability in some of the AOGCMs of the CMIP and IPCC AR4 intercomparisons. Third, the mechanism discussed in this study is based on heat flux changes under warming, which were shown to dominate THC changes in all CMIP simulations. We concede that this dominance was particularly pronounced in CLIMBER-2 and therefore the proposed mechanism might be stronger in CLIMBER-2 compared to other models. But it is consistent with the findings of the CMIP, EMIP, and IPCC AR4 simulations presented. Unfortunately, no sea ice data were available from the CMIP intercomparison. Models from the IPCC AR4, however, show that a stronger weakening of the THC is indeed associated with less sea ice melt (Fig. 10). This dependence is also found in CLIMBER-2 and is central to the mechanism. This is a fourth indication that the mechanism might be relevant to other models.

The above, however, is not sufficient to claim that the sea ice process is dominant or even at play in other models. There are a number of strong limitations to a generalization of the results. First of all there is the possible influence of changes in freshwater fluxes. Though Gregory et al. (2005) showed that changes in freshwater fluxes are not dominant in the CMIP models, they are not negligible in all models and might be of the same order as the heat flux changes in some experiments. This point is specifically important when including melting of the Greenland ice sheet.

Other possible limitations result directly from the zonally averaged oceanic component of CLIMBER-2. As a consequence, CLIMBER-2 does not distinguish deep-water formation in the Labrador Sea and Nordic Seas. Possible switches between these convection sites that might occur under warming are therefore not captured. Furthermore, changes in ice-free areas are quantitatively different in CLIMBER-2 compared to OGCMs (Fig. 10), which might lead to an overestimation of the described mechanism. CLIMBER-3α, however, does not suffer the shortcomings of a zonally averaged ocean and shows the same qualitative behavior of the overturning weakening under warming. The statistical–dynamical atmosphere component of both CLIMBER models might induce some inherent limitations to atmospheric transport processes (including the lack of atmospheric variability). This might have implications for atmospheric heat and freshwater transport, oceanic mixing processes, and the relative importance of changes in heat and freshwater forcing to the ocean under warming. The representation of sea ice physics could play a role for the mechanism in different models and in the real ocean.

Independent of the quantitative applicability of the CLIMBER-2 results for other models, the physical arguments presented have the potential to qualitatively explain the observed phenomenon and play a role for the THC stability under warming. A better and more quantitative understanding of the mechanism could lead to improved projections of future THC evolution, since changes in sea ice cover and surface fluxes are better constraints than the deep volume fluxes associated with the THC.