East Atlantic Pattern Drives Multidecadal Atlantic Meridional Overturning Circulation Variability During the Last Glacial Maximum

The variability of the Atlantic meridional overturning circulation (AMOC) and its governing processes during the Last Glacial Maximum (LGM) is investigated in the Kiel Climate Model. Under LGM conditions, multidecadal AMOC variability is mainly forced by the surface heat flux variability linked to the East Atlantic pattern (EAP). In contrast, the multidecadal AMOC variability under preindustrial conditions is mainly driven by the surface heat flux variability associated with the North Atlantic Oscillation. Stand‐alone atmosphere model experiments show that relative to preindustrial conditions, the change in AMOC forcing under LGM conditions is tightly linked to the differences in topography. Plain Language Summary The Last Glacial Maximum (LGM, about 21,000 years before present) was a period when extensive continental ice sheets covered parts of North America and northern Europe. A climate model simulates enhanced multidecadal variability of the Atlantic meridional overturning circulation (AMOC), a major oceanic current system with global climatic impacts, during the LGM relative to modern times. The main driver of multidecadal AMOC variability under LGM conditions is the surface heat flux variability associated with the East Atlantic pattern, a mode of internal atmospheric variability. Under modern conditions, the main driver of the multidecadal AMOC variability is the heat flux variability associated with the North Atlantic Oscillation, another mode of internal atmospheric variability. The difference in AMOC forcing is basically due to the change in topography linked to the extensive ice sheets during the LGM.


Introduction
During the Last Glacial Maximum (LGM) about 21,000 years before present, global ice volume reached its maximum (Mix et al., 2001). Concurrent with expanded Antarctic and Greenland ice sheets, there were extensive continental ice sheets over parts of North America and northern Europe during that time. Coupled climate models and atmosphere models have shown that the ice sheets, along with much lower concentrations of atmospheric greenhouse gases (e.g., Bereiter et al., 2015), force large perturbations in the mean atmospheric circulation and its internal variability over the North Atlantic sector, especially during boreal winter (e.g., Justino & Peltier, 2005;Pausata et al., 2009;Pausata et al., 2011). There are two prominent modes of internal atmospheric variability over this region: the North Atlantic Oscillation (NAO) and the East Atlantic pattern (EAP). The NAO is the leading mode of atmospheric circulation variability over the North Atlantic/European sector in boreal winter, characterized by a meridional dipole structure with a low-pressure center over Iceland and high-pressure center over the Azores (e.g., Hurrell, 1995). The second most energetic mode is the EAP. The EAP is structurally similar to the NAO and consists of a north-south dipole, with the anomaly centers of the EAP located southeastward to the approximate nodal lines of the NAO pattern (Barnston & Livezey, 1987). The NAO and EAP drive climate variations via anomalous surface winds and heat fluxes (Cayan, 1992).
The North Atlantic Ocean plays a crucial role in the global climate system through a basin-wide thermohaline circulation, the Atlantic meridional overturning circulation (AMOC), which is driven by deep water formation at high northern latitudes. The AMOC is characterized by the northward flow of upper-layer warm waters by the Gulf Stream and its extensions and its southward return flow of cold waters by the deep western boundary current. The AMOC exhibits variability on a wide range of time scales, from monthly to centennial ( (Folland et al., 1986(Folland et al., , 2001, Atlantic hurricanes (Wang & Lee, 2009), and summer climate over North America and Western Europe (Sutton & Hodson, 2005).
Climate models simulate pronounced multidecadal AMOC variability throughout the last millennium and the Holocene (Wei & Lohmann, 2012;Zanchettin et al., 2013). In many climate models (Danabasoglu et al., 2012;Delworth et al., 2016;Kim et al., 2018), the multidecadal AMOC variability under present-day conditions is strongly linked to NAO-related changes in surface heat flux over the subpolar North Atlantic. A few modeling studies suggest that the multidecadal AMOC variability could be also closely related to the EAP (Msadek & Frankignoul, 2009;Ruprich-Robert & Cassou, 2015). However, the multidecadal AMOC variability during the LGM and its governing physical processes so far only have received little attention. Under the LGM boundary conditions, large perturbations in mean atmospheric circulation and its internal variability, massively extended sea ice and shifted deep convection sites over the North Atlantic Sector are simulated by climate models. These changes can essentially affect the response of subpolar North Atlantic deep convection to the atmospheric forcing and subsequently the AMOC. A major goal of this study is to identify possible differences in the multidecadal AMOC variability between the LGM and modern times. In particular, we are interested in the atmospheric forcing of multidecadal AMOC variability.
Here we use the Kiel Climate Model (KCM; Park et al., 2009) to investigate the multidecadal AMOC variability during the LGM and compare it to that simulated under preindustrial conditions. The paper is organized as follows: Section 2 describes the climate model and the experimental setup. In section 3, the multidecadal AMOC variability simulated under LGM conditions and its driving factors are explored. Summary and discussion are presented in section 4 and conclude the paper.

Model, Experimental Setup, and Method
The KCM (Park et al., 2009) is a fully coupled atmosphere-ocean-sea ice general circulation model. In the version applied here, the atmospheric component ECHAM5 (Roeckner et al., 2003) is used with a horizontal resolution of T42 (2.8°× 2.8°) and 19 vertical levels reaching up to 10 hPa. The ocean-sea ice component NEMO (Madec, 2008) is run on a 2°Mercator mesh and with 31 vertical levels. Enhanced meridional resolution of 0.5°is used in the equatorial region. The two models are coupled with the OASIS3 coupler (Valcke, 2006). Two simulations are performed: one 5,300-year-long preindustrial control run (PI) that is initialized with the Levitus climatology of temperature and salinity. PI employs the modern land-sea mask, orography, and continental ice sheet configuration. In the other simulation for the glacial period (GLAC), the boundary conditions are implemented in accordance with the Paleoclimate Modelling Intercomparison Project phase 4 (PMIP4) protocol for LGM experiments (Kageyama et al., 2017). The ice sheet configuration is an average of three different reconstructions following the PMIP3 LGM experiments (Braconnot et al., 2012). The greenhouse gas concentrations and orbital parameters are listed in Table 1. To account for the 116-m drop of the mean sea level, GLAC is initialized with the Levitus temperature climatology and Levitus salinity climatology to which 1 psu has been added. GLAC is integrated for 4,650 years to reach quasi-equilibrium ( Figure 1a). Monthly output from the last 1,000 years of the two integrations is used for analysis.
To extract the dominant modes of atmosphere-ocean variability over the North Atlantic, widely used empirical orthogonal functions (EOFs) analysis is applied here (Chapter 13.5 in Jolliffe, 2002). EOF analysis decomposes a multivariate data set into eigenvectors (EOFs) for spatial structure. The time evolution of each EOF, the principal components (PCs), are obtained by projection of the original data on the corresponding EOF. The EOFs are orthogonal and sorted by the amount of explained variance in a descending order.

Mean State and AMOC
The JFM (January-February-March) sea surface temperature (SST) averaged over the North Atlantic (40-60°N) in PI and GLAC are 7.3°C and 3.0°C, respectively (Figures 1b and 1c). The colder SSTs in GLAC favor a substantial increase of sea ice extent. In PI, sea ice covers the Baffin Bay, the Labrador Sea, and western Greenland-Iceland-Norwegian (GIN) Sea (contours in Figures 1d and 1e). Deep convection (as indicated by JFM mixed layer depth) in PI is simulated basically east of the sea ice margin. In GLAC, one rather large deep convection site is simulated in the eastern North Atlantic south of Iceland (color shading in Figures 1d and 1e). There is a stronger low sea level pressure (SLP) center over the North Atlantic in GLAC relative to PI (Figures 1f and 1g), which goes along with enhanced negative surface heat fluxes (i.e., oceanic heat loss) over the convection site in the eastern North Atlantic (Figures 1h and 1i). The enhanced oceanic heat loss in GLAC is likely due to the intensified north-easterlies implied by the SLP changes. As a result, deep convection is stimulated south of Iceland.
A stronger AMOC is simulated in GLAC relative to PI. AMOC strength (defined as the maximum overturning stream function at 30°N) amounts to 16.8 Sv (1 Sv= 10 6 m 3 /s) in GLAC as compared to 12.7 Sv in PI (Figures S1a and S1b). The GLAC simulation by the KCM is consistent with most coupled climate models (Figures S1c-S1l; except CCSM4 in PMIP3, Figures S1c and S1d), also featuring enhanced and deeper AMOC in LGM simulations (Muglia & Schmittner, 2015). Paleoclimate reconstructions, however, clearly indicate that the AMOC was shallower during the LGM but do not agree on the AMOC strength (Böhm et al., 2015;Lippold et al., 2012;McManus et al., 2004). This apparent inconsistency between the simulated AMOC and reconstructions seems to be linked to the accuracy of the control state in climate models (Weber et al., 2007). Models with too shallow overturning do not simulate further shoaling during the LGM. The compensating effects of glacial ice sheets and glacial CO 2 concentrations on the AMOC strength, insufficient sea ice formation and export in the Southern Ocean, and incomplete deep-ocean equilibrium may also account for the discrepancies among the climate models (Klockmann et al., 2016;Marzocchi & Jansen, 2017).
The AMOC in the KCM exhibits significant multidecadal variability during the LGM ( Figure S2). The standard deviation of the AMOC index derived from band-pass-filtered (11 to 90 years) data amounts to 0.53 Sv in GLAC as compared to 0.35 Sv in PI.

Atmospheric Forcing of Multidecadal AMOC Variability
EOF analysis is performed on JFM SLP anomalies (Figure 2). The corresponding normalized PCs are shown in Figure S3. The leading EOF in PI and GLAC is the NAO, which accounts for 53.2% and 35.2% of the variance in PI and GLAC, respectively. The patterns, however, differ considerably. While in PI the NAO covers the whole North Atlantic sector (Figure 2a), the pattern in GLAC only is well developed east of about 50°W (Figure 2b). Similar NAO patterns in LGM simulations are seen in Pausata et al. (2009) and four out of nine PMIP3 models ( Figure S4). The second most energetic EOF (EOF2), which accounts for 12.0% and 20.1% of total variance in PI and GLAC, respectively, is the EAP. The EAP in both integrations features the largest loadings over the eastern North Atlantic with center around the British Isles (Figures 2c and 2d). Eight out of nine PMIP3 models also exhibit an increase of the variance explained by EOF2 in LGM simulations, with four models exhibiting an EAP pattern similar to that in the KCM ( Figure S5).
To investigate the relation between the AMOC and atmospheric variability, the meridional overturning stream function is regressed on the normalized SLP PCs with different time lags. Figures 3a and 3b show the 9-year lag regression of the overturning stream function on the NAO index (SLP-PC1) in PI and GLAC (refer to Figure S6 for time lags ranging from 0 to 10 years). The lagged overturning response in PI is significant between 10°S and 40°N, with the maximum regression coefficient of 0.13 Sv/σ NAO (σ NAO is the standard deviation of SLP-PC1) at around 20°W and at 1-to 2-km depth (Figure 3a). This result indicates that the AMOC variability is tightly linked to the NAO under preindustrial conditions. Such a tight relationship is not observed in GLAC (Figure 3b). With the exception of the upper northern tropical Atlantic (from the equator to 10°N), there is no significant response of the overturning to the NAO in GLAC.
We now turn to the influence of the EAP on the AMOC. Previous studies suggested that the EAP should be included as an additional driver for the multidecadal AMOC variability (e.g., Msadek & Frankignoul, 2009;Ruprich-Robert & Cassou, 2015). This is supported by PI in which the EAP index (SLP-PC2) drives significant variations in the overturning when leading by 10 years (Figure 3c; refer to Figure S7 for lags from 0 to 10 years). The influence of the EAP, however, is spatially restricted to the North Atlantic between 20°N and 45°N and the upper 2 km. The impact of the EAP on the AMOC is strongly intensified in GLAC and extends from 45°N to the equator (Figure 3d). The maximum regression coefficient amounts to 0.11 Sv/ σ EAP at 40°N and 2 km. In region 45-20°N, the response extends through the entire water column. Thus, the EAP is the mode of SLP variability that primarily drives the multidecadal AMOC variability in GLAC.
As reported by Biastoch et al. (2008), interannual to decadal variability of the AMOC in the midlatitude North Atlantic can be regarded as linear superposition of decadal and longer-term thermohaline forcing variability and high-frequency wind variability. Therefore, the JFM Q net is regressed on SLP-PC1 (NAO) in PI and on SLP-PC2 (EAP) in GLAC (Figures 4a and 4b). When Q net leads the AMOC by 10 years (Figures 4c and 4d), significant negative regression coefficients (color shading), that is, enhanced oceanic heat loss is observed over the convection sites (contours) in both PI and GLAC. In these regions, ocean density is enhanced and deep water formation intensified as suggested by the mixed layer depth (Figures 4e and  4f). The Q net pattern associated with the NAO (EAP) in Figure 4a (Figure 4b) and the Q net pattern leading the AMOC by a decade shown in Figure 4c (Figure 4d) share similarities over the deep convection sites. This suggests that the Q net anomalies associated with the NAO (EAP) drive the AMOC variations in PI (GLAC). In contrast, the Q net changes in response to AMOC variations occur in the midlatitudes south of the convection sites (for PI see Figure 15 in Park et al., 2016).
To investigate the influence of ice sheet geometry and associated changes in land-sea distribution on the internal atmospheric variability in more detail, four sensitivity experiments are carried out with the ECHAM5 atmosphere model in stand-alone mode (Table S1). The results are presented for JFM in terms of the leading EOFs. When the atmosphere model is forced by GLAC (PI) ice sheets and either preindustrial or GLAC SST and sea ice conditions, the NAO (EOF1) and EAP (EOF2) patterns resemble those in GLAC (PI; Figures S8a-S8d). For example, an eastward shifted NAO pattern is simulated in the experiments forced by GLAC ice sheets, irrespective of whether preindustrial or GLAC SST and sea ice conditions are used or not ( Figures S8b and S8d). The stand-alone experiments with the atmosphere model suggest that the ice sheet configuration is the most important factor determining the modes of atmospheric variability.

Summary and Discussion
This study investigates the multidecadal AMOC variability during the LGM and its driving factors by means of a version of the KCM. Our analysis confirms previous results (e.g., Barrier et al., 2014;Biastoch et al., 2008;Delworth et al., 2016;Kim et al., 2018) that under modern climate conditions multidecadal AMOC variations are mainly driven by surface heat flux (Q net ) anomalies associated with the NAO, with Q net anomalies related to the EAP as an additional but much weaker driver. The impact of the NAO on the AMOC is virtually absent in GLAC, and the EAP-related Q net anomalies become the dominant driving force for the multidecadal AMOC variability. This change in forcing can be largely attributed to the presence of extensive continental ice sheets and related changes in land-sea distribution, as confirmed by dedicated experiments with the atmospheric component run in stand-alone mode ( Figure S8).
The multidecadal AMOC variability strengthens when the model is run under LGM conditions (GLAC) as compared to a reference run performed under preindustrial conditions (PI). Further research is needed to understand the increase in multidecadal AMOC variability. One reason could be the extensive deep convection site in GLAC, which is located south of Iceland. This is the region where the EAP imposes great influence on the AMOC through Ekman transport and especially Q net (Barrier et al., 2014;Msadek & Frankignoul, 2009). In PI, on the other hand, the deep convection site in the Labrador Sea is displaced to the east to the region South of Greenland. This model bias may reduce the sensitivity of the AMOC to NAO-related Q net variability (Park et al., 2016).
Another question is the role of the oceanic mean state. Ocean-only simulations suggest that buoyancy fluxes play a dominant role in driving the Labrador Sea Water formation and the subsequent AMOC changes (Barrier et al., 2014). Deep water formation and advection, however, strongly depend on the vertical and horizontal density gradients, thus affecting the thermohaline circulation (Steffen & D'Asaro, 2002). The implication from these studies is that the much-enhanced multidecadal AMOC variability simulated by the KCM in GLAC may be, at least partly, due to the different oceanic mean states.
In summary, simulations with the Kiel Climate Model suggest that the forcing for multidecadal AMOC variability changes under LGM conditions relative to modern times, with the EAP becoming the most important driver of the multidecadal AMOC variability during the LGM. The suggestions of the two anonymous referees greatly improved an earlier version of the manuscript. This study was supported by the Integrated School of Ocean Sciences (ISOS) at the Excellence Cluster "The Future Ocean" at Kiel University sponsored by the German Science Foundation (DFG). We thank Stefan Hagemann and Christian Stepanek for their assistance with setting up the simulations. The integrations with the Kiel Climate Model (KCM) were conducted at the Computing Center of Kiel University (CAU). The KCM results presented in this study are available at Pangaea (https://doi.pangaea.de/10.1594/ PANGAEA.899572). The PMIP3 data output is available at this site (https:// esgf-data.dkrz.de/search/cmip5-dkrz/). This is a contribution to the PalMod project funded by the German Ministry of Education and Research.