Changes in the geometry and strength of the Atlantic meridional overturning circulation during the last glacial (20–50 ka)

. We reconstruct the geometry and strength of the Atlantic meridional overturning circulation during the Heinrich stadial 2 and three Greenland interstadials of the 20– 50 ka period based on the comparison of new and published sedimentary 231 Pa / 230 Th data with simulated sedimentary 231 Pa / 230 Th. We show that the deep Atlantic circulation during these interstadials was very different from that of the Holocene. Northern-sourced waters likely circulated above 2500 m depth, with a ﬂow rate lower than that of the present-day North Atlantic deep water (NADW). Southern-sourced deep waters most probably ﬂowed northwards below 4000 m depth into the North Atlantic basin and then southwards as a return ﬂow between 2500 and 4000 m depth. The ﬂow rate of this southern-sourced deep water was likely larger than that of the modern Antarctic bottom water (AABW). Our results further show that during Heinrich stadial 2, the deep Atlantic was probably directly affected by a southern-sourced water mass below 2500 m depth, while a slow, southward-ﬂowing water mass originating from the North Atlantic likely inﬂuenced depths between 1500 and 2500 m down to the equator.

creased North Atlantic deep-water formation and increased influence of southern-sourced deep waters in the Atlantic during Heinrich stadials (Elliot et al., 2002;McManus et al., 2004;Vidal et al., 1997). Moreover, climate models are able to reproduce the bipolar seesaw pattern characterizing millennial-scale glacial variability through variations of the strength of the Atlantic meridional overturning circulation (AMOC) in response to freshwater forcings (Ganopolski and Rahmstorf, 2001). However, recent studies show that a shallow circulation cell could have been still active during HSs (Bradtmiller et al., 2014;Gherardi et al., 2009;Lynch-Stieglitz et al., 2014;Roche et al., 2014;Wary et al., 2015), suggesting that Greenland temperature millennial-scale variability might be related to more complex changes in Atlantic circulation than simply switching between "on" and "off" circulation modes. A better understanding of the vertical layout and flow rate of the water masses constituting the AMOC during the last glacial is therefore needed to assess the relationship between AMOC and glacial millennial-scale variability.
Sedimentary ( 231 Pa / 230 Th) xs,0 (activity ratio of 231 Pa and 230 Th unsupported by lithogenic and authigenic uranium and corrected from decay to the time of sediment deposition; Pa / Th hereafter) records were first used to assess the AMOC intensity during the last glacial maximum (LGM; Yu et al., 1996). Since then, Pa / Th records have been used in the Atlantic to infer changes in the intensity of the deep ocean circulation during HSs (Böhm et al., 2015;Bradtmiller et al., 2014;Burckel et al., 2015;Gherardi et al., 2005Gherardi et al., , 2009Mc-Manus et al., 2004). The comparison of simulated sedimentary Pa / Th values with core top Pa / Th data has shown that sedimentary Pa / Th reflects circulation intensity in the modern Atlantic Ocean .
However, interpretations of sediment Pa / Th from a single core can be ambiguous because similar values can result from different geometry and overturning strength (Luo et al., 2010). Reconstructing past circulation thus requires combining Pa / Th records from multiple sites over a wide range of latitudes and depths (Gherardi et al., 2009;Lippold et al., 2011Lippold et al., , 2012. In this study, we present new sedimentary Pa / Th data from a deep sediment core recovered from the Brazilian margin and from an intermediate depth core from the midlatitude North Atlantic. We then compare the last glacial Pa / Th records from different water depths and latitudes with Pa / Th values simulated using a simple 2-D box model (Luo et al., 2010) forced by various stream functions. One stream function is derived from present-day geostrophic velocity estimates (Talley et al., 2003) and two others were simulated with the Earth System model iLOVECLIM under different climatic conditions (Roche et al., 2014). Results of this comparison allow us to constrain the geometry and strength of the AMOC during GI-3, GI-8, GI-10 (Rasmussen et al., 2014) and HS2. We chose to focus our study on HS2 and the interstadials surrounding HS2 and HS4 as these periods are asso-ciated with very different ice-sheet volumes (Lambeck and Chappell, 2001).

Sediment cores
Sediment cores MD09-3257 (04 • 14.69 S, 36 • 21.18 W; 2344 m water depth) and MD09-3256Q (03 • 32.81 S, 35 • 23.11 W; 3537 m water depth) were recovered from the Brazilian margin during R/V Marion Dufresne cruise MD173/RETRO3 (Fig. 1). Improved recovery of deep-sea sediments with no or little deformation of sediment layers was achieved during this coring cruise thanks to the systematic use of the CINEMA software (Bourillet et al., 2007;Woerther and Bourillet, 2005). This software computes the amplitude and duration of the elastic recoil of the aramid cable and the piston displacement throughout the coring phase, accounting for the length of the cable (water depth) and total weight of the coring system. The length of the coring cable is indeed of primary importance regarding the deformation rate of the Calypso long piston cores (Bourillet et al., 2007;Skinner and McCave, 2003).
Core GeoB3910 (04 • 14.7 S, 36 • 20.7 W; 2362 m water depth; Jaeschke et al., 2007) was recovered from approximately the same position and depth as core MD09-3257, during the Meteor cruise M34/4. Hereafter, we refer to both GeoB3910 and MD09-3257 as intermediate equatorial cores and to MD09-3256Q as the deep equatorial core. At present, the Brazilian margin at these depths is bathed by the North Atlantic deep water (NADW; Fig. 1). Because the Brazilian margin is affected by western boundary currents (Rhein et al., 1995), these sediment cores are ideally located to observe changes in the strength and extent of the intermediate and deep AMOC water masses (Schott, 2003).
Sediment core 32 • 3.198 W; 2475 m water depth) was recovered from the northern margin of the subtropical gyre (Chapman et al., 2000). Its location in the midlatitude North Atlantic provides information on changes in NADW production rates that could not be deduced from the sole equatorial depth transect.

Benthic δ 13 C
The stable carbon isotopic composition (δ 13 C) of the epifaunal benthic foraminifer Cibicides wuellerstorfi has been shown to record the δ 13 C of bottom-water dissolved inor-  (Schlitzer, 2000) and (b) map of the Atlantic Ocean showing the location of the studied sediment cores (see Sect. 2.1 or Table S1 for detailed locations). Cores, for which we provide new Pa / Th or δ 13 C data, are circled in red. Phosphate content follows the structure of the present-day AMOC. White arrows indicate the approximate flow directions of the Antarctic intermediate water (AAIW), Antarctic bottom water (AABW) and North Atlantic deep water (NADW). ganic carbon (DIC) with minor isotopic fractionation (Duplessy et al., 1984;Zahn et al., 1986). Initial DIC isotopic concentration is acquired by a water mass in its formation region by surface productivity (which consumes 12 C therefore increasing dissolved δ 13 C) and temperature-dependent air-sea interactions (Lynch-Stieglitz et al., 1995;Rohling and Cooke, 2003). DIC δ 13 C then evolves as deep water ages, because the constant export of 12 C-enriched biogenic material that is remineralized at depth leads to the decrease of the DIC δ 13 C along the flow path of the water mass. As DIC δ 13 C largely follows water mass structure and circulation in the modern ocean, C. wuellerstorfi δ 13 C has been used to trace water masses, with a decrease in C. wuellerstorfi δ 13 C being interpreted as a decrease in bottom water ventilation, and conversely (e.g., Duplessy et al., 1988). However, the information on bottom water ventilation embedded in C. wuellerstorfi δ 13 C is complicated by the impact of changes in surface water δ 13 C, marine biological productivity and continental biomass changes.
Because LGM δ 13 C values are higher in northern-sourced waters (1.5 ‰) than in southern-sourced waters (< −0.2 ‰; Curry and Oppo, 2005), we interpret a decrease in C. wuellerstorfi δ 13 C values at the equatorial sites as an increase in the time elapsed since the water mass was last in contact with the atmosphere or as an increased influence of nutrient-rich southern-sourced deep waters. Note that because we lack information on past marine productivity changes, we do not account for their potential impact on benthic δ 13 C in the present study.

Sedimentary Pa / Th
In contrast to C. wuellerstorfi δ 13 C, which reflects the nutrient content of bottom waters, sedimentary Pa / Th is a relatively recent tracer that can be used to estimate the renewal rate of water masses occupying the first ∼ 1000 m above the seafloor (Thomas et al., 2006;Luo et al., 2010). This tracer has been successfully used to reconstruct past changes in deep Atlantic circulation intensity (Böhm et al., 2015;Gherardi et al., 2005Gherardi et al., , 2009Guihou et al., 2010Guihou et al., , 2011Hall et al., 2006;Jonkers et al., 2015;Lippold et al., 2011Lippold et al., , 2012Mc-Manus et al., 2004;Negre et al., 2012;Yu et al., 1996). 231 Pa and 230 Th are produced at a constant Pa / Th activity ratio of 0.093 by dissolved uranium, which is homogeneously distributed in the oceans. 230 Th is however much more particle reactive than 231 Pa, as reflected by their respective residence time in the ocean (30-40 years for 230 Th, 200 years for 231 Pa; Francois, 2007). 230 Th is therefore rapidly removed from the water column to the underlying sediment, while 231 Pa can be advected by oceanic currents. High (low) rates of overturning therefore result in high (low) 231 Pa export and hence low (high) sedimentary Pa / Th ratio in the Atlantic. However, affinities of 231 Pa and 230 Th for settling particles depend on the particle type (Chase et al., 2002). For instance, 231 Pa has a high affinity for opal, so that high opal fluxes can result in high sedimentary Pa / Th values even in the presence of lateral advection (Chase et al., 2002).
The origin of sedimentary Pa / Th variability therefore needs to be carefully assessed.
Pa / Th measurements on core MD09-3256Q were performed by isotopic dilution mass spectrometry on a Thermo Finnigan MC-ICP-MS Neptune, following the method of Guihou et al. (2010).
Core SU90-03 sedimentary Pa / Th was measured by isotopic dilution on a single collector, sector field ICP-MS (El-ement2) at the University of British Columbia, following the procedure described by Choi et al. (2001).
For both cores, Pa and Th are corrected from radioactive decay since the time of sediment deposition and from authigenic and lithogenic components using a 238 U / 232 Th ratio of 0.5 ± 0.1 (Fig. S1 in the Supplement; Guihou et al., 2010).

Age model
Over the period 0-34 ka, core MD09-3256Q age model is based on 11 14 C dates measured on planktic foraminifer G. ruber white and converted to calendar age using the Ma-rine13 curve with no additional reservoir age correction (Burckel et al., 2016;Reimer et al., 2013;Fig. S2). During the last glacial, Heinrich stadials were recorded in marine sediment cores from the Brazilian margin as Ti / Ca peaks resulting from increased terrigenous input during periods of increased precipitation associated with southward shifts in the position of the intertropical convergence zone (ITCZ; Jaeschke et al., 2007). Ti / Ca peaks are therefore good stratigraphic markers for correlating sediment cores with neighboring well-dated cores. Therefore, from 34 to 50 ka, core MD09-3256Q was dated by correlation of its Ti / Ca record with that of core GeoB3910 using two tie points corresponding to the Ti / Ca peaks associated with HS4 and HS5. The GeoB3910 age model over the 34 to 50 ka period is based on one 14 C date calibrated using the Marine13 curve and on four speleothem tie points at the onset and end of HS4 and HS5 (Burckel et al., 2015). Core MD09-3256Q age model and sedimentation rates are given in Table S2 and Fig. S3. The age model of core SU90-03 is based on 17 14 C dates measured on various species of planktic foraminifera (Chapman et al., 2000) that were converted to calendar ages using Marine13 calibration curve with no additional reservoir age correction. 14 C dates of all the published Atlantic cores used in this study were converted into calendar ages using the same method (Table S1).

Description of the models
In order to assess the vertical layout and renewal rate of the water masses constituting the AMOC during the last glacial period, the sedimentary Pa / Th data of the studied sediment cores were compared to Pa / Th values simulated with a simple 2-D box model (Luo et al., 2010) forced by different stream functions (Fig. 2b, d, f). Stream functions were generated using the iLOVECLIM coupled climate model, comprising atmosphere, ocean and vegetation components (Roche et al., 2014). A LGM equilibrium state computed using the PMIP-2 protocol was used as background climate (see Roche et al., 2007) for details). Stream functions characterized by a shallow (< 2500 m) northern-sourced overturning cell (shallow overturning stream function) and by a complete shutdown of the AMOC (off-mode stream function) were generated by imposing a 0.16 and 0.35 Sv freshwater forcing in the Labrador Sea, respectively (Roche et al., 2010(Roche et al., , 2014. The freshwater input is added during 300 years on the LGM background state. The stream functions are taken as the mean over the 100-year period of lowest deep-water formation in the North Atlantic, during or right after the period of freshwater forcing. A freshwater input of 0.16 Sv allows the presence of a shallow circulation cell in the Atlantic Ocean, while a freshwater forcing of 0.35 Sv leads to an almost complete shutdown of the AMOC (Roche et al., 2014). Note that the freshwater input values needed to modify the AMOC are strongly model dependent and the important information carried by the model in the present context is the state of the AMOC rather than the freshwater input value. Contrary to shallow overturning and off-mode stream functions, the Holocene stream function was computed using data-based geostrophic velocity estimates (Talley et al., 2003).
Dissolved Pa and Th concentrations in the 2-D box model are controlled by (1) production from U decay in the water column, (2) adsorption and desorption on settling particles and (3) advection by oceanic circulation. Particulate Pa and Th concentrations are controlled by (1) adsorption and desorption from the dissolved pool and (2) removal of sedimentary particles to the seafloor (Luo et al., 2010).
Modeled sedimentary meridional Pa / Th sections generated with the different stream functions are shown in Fig. 2a, c, e. Different water mass circulation intensities and geometries result in different simulated sedimentary Pa / Th (Luo et al., 2010). In the deep Atlantic, increasing circulation intensity above a specific location causes Pa / Th to decrease at that water depth because of the increased Pa export and conversely. Increasing water depth without modifying circulation intensity also causes sedimentary Pa / Th ratio to decrease in the model because of the increased residence time of Pa and resulting higher Pa export, and conversely. Finally, the sedimentary Pa / Th ratio increases along the flow path of any newly formed water mass as low dissolved Pa concentrations of newly formed water masses increase by desorption of Pa from Pa-concentrated settling particles equilibrating with ambient waters (Francois, 2007). Adsorption and desorption rate constants also impact the simulated Pa / Th ratio. These constants were adjusted to reflect the opal belt in the Southern Ocean (Luo et al., 2010). For the Holocene, these constants were also changed to reflect preferential scavenging of Pa by biogenic opal in the northern North Atlantic (Lippold et al., 2012). Depth (m)  Fig. 1). Data gridding was achieved using the Ocean Data View software (Schlitzer, 2015).

Limits of the models
The absence of margins in the simple 2-D Pa / Th model (Luo et al., 2010) prevents it from simulating boundary scavenging, which is the lateral advection of dissolved Pa from open ocean regions characterized by high Pa concentrations to coastal regions of low Pa concentration such as in upwelling zones (Christl et al., 2010). However, as described in the results section below, we verified that our Pa / Th signal is mainly driven by oceanic circulation changes and the importance of diffusive transport is therefore likely negligible here. This simple 2-D Pa / Th model therefore appears adequate for comparison with our Pa / Th data. The vertical resolution of the iLOVECLIM model is depth dependent, with higher resolution (10 to 100 m) in the upper water column than below 1000 m (500 to 700 m). Hence, the uncertainty in the position of the water mass transitions in the stream functions below 1000 m is of 500 to 700 m. However, because sedimentary Pa / Th likely reflects the Pa export in the bottom 1000 m of the water column (Thomas et al., 2006), the model vertical resolution is sufficient to properly simulate sedimentary Pa / Th values. Hence, the relatively low verti-cal resolution of the iLOVECLIM model in the deep ocean does not affect our conclusions.

Time slice definition
We define three interstadial time slices and one Heinrich stadial time slice (Fig. 3) to compare Pa / Th data measured in Atlantic cores to Pa / Th values simulated with the different stream functions. We focus on HS2, the preceding GI and the GIs bracketing HS4. We did not include HS4 in our study because core MD09-3257 HS4 Pa / Th data are affected by boundary scavenging (see Sect. 3.1, Burckel et al., 2015) and we therefore lack information from an important location of the Atlantic Ocean.
GI-3, GI-8 and GI-10 time slices are defined as the periods of stable sedimentary Pa / Th values in core MD09-3257 associated with the NGRIP GI time intervals (Fig. 3). More specifically, we used as a reference MD09-3257 Pa / Th values bracketing the middle of NGRIP GI time intervals in the GICC05 age scale (Rasmussen et al., 2014). Contiguous Pa / Th values within 1σ uncertainty of the Pa / Th reference value form a plateau of stable Pa / Th values that was used to  (Chapman et al., 2000) and GeoB3910 (Burckel et al., 2015) C. wuellerstorfi δ 13 C; and (c) NGRIP temperature record on the GICC05 timescale (Kindler et al., 2014). In (a) lines pass through average Pa / Th values in case of replicates, while diamonds and squares (MD09-3257) correspond to individual Pa / Th measurements. V29-172 Pa / Th values are represented by two purple diamonds. White squares indicate Pa / Th values not considered in core MD09-3257, as they might not be influenced by oceanic circulation only (Burckel et al., 2015). The red and blue arrows indicate the late Holocene Pa / Th values in cores MD09-3257 (0.065 ± 0.004, Burckel et al., 2015) and MD09-3256Q (0.043 ± 0.002), respectively. Error bars on Pa / Th measurements are given in Fig. S1 and Tables S7 and S8. In (b) thick lines are three-point running averages of the C. wuellerstorfi δ 13 C records; the black arrow indicates present-day NADW δ 13 C value (∼ 1.36 ‰; the Supplement). δ 13 C values are given in Table S9. In (c) numbers indicate the GIs. Red vertical bands represent the GI-3, GI-8 and GI-10 time slices and the blue vertical band the HS2 time slice. define the GI time slices. With this definition, GI time slices represent the periods of stable oceanic circulation associated with each GI. The HS2 time slice was defined in core MD09-3257 as the period of maximum sedimentary Pa / Th after the abrupt rise associated with the onset of HS2. Sedimentary Pa / Th values associated with each time slice and core are given in Table S3. We computed uncertainties on the Pa / Th values associated with each time slice accounting for the uncertainty on individual Pa / Th measurements and uncertainties on the age model (see the Supplement).
Both GI-8 and GI-10 time slices are associated with high temperatures recorded in Greenland ice cores. The GI-3 time slice includes both the period of high Greenland temperatures associated with GI-3 and periods of low temperatures associated with GS-4 and the beginning of GS-3. Given the low Pa / Th and high δ 13 C values in the intermediate equatorial core at that time, we consider that the GI-3 time slice mainly reflects interstadial conditions. However, because temporal resolution of the marine records is too low to clearly distinguish between GI-3 and GS-4, information about the state of the AMOC during GIs derived from this time slice should be considered with caution.

Quantification of the model-data agreement
In order to quantify the agreement between simulated and measured sedimentary Pa / Th, we compute the Euclidean distance, defined as the square root of the sum of squared differences between simulated and measured Pa / Th for each core. Minimum values indicate the best agreement between simulated and measured Pa / Th (Tables S4 and S5). Note that due to the limited number of sedimentary Pa / Th records during MIS3, we can only provide an approximate estimate of water mass boundary positions. Our equatorial transect is however ideally located to record shifts in the position of the transition between southern-and northern-sourced water masses.

Sedimentary Pa / Th data
Cores MD09-3257, MD09-3256Q and SU90-03 Pa / Th measurements were centered on HS2 and HS4 (Burckel et al., 2016;Fig. 3). No Pa / Th values were measured within HS2 in core MD09-3257 because of the presence of turbidite layers (Burckel et al., 2015). During HS4 and before HS2, the sedimentary Pa / Th ratio of core MD09-3257 rises above the production ratio of 0.093, indicating the absence of Pa export. Pa / Th variability associated with GS and GI is observed, with high Pa / Th values occurring during GS and low Pa / Th values during GI. Pa / Th variations in core MD09-3256Q are more muted (Fig. 3). The main Pa / Th variation in core MD09-3256Q occurs during HS4, when Pa / Th values rise from ∼ 0.06 to ∼ 0.08. This increase in MD09-3256Q Pa / Th also corresponds, within dating uncertainties, with the largest Pa / Th change from ∼ 0.03 to ∼ 0.05 in core SU90-03.
Before interpreting our Pa / Th records in terms of ocean circulation changes, we need to assess whether varying lithogenic or opal fluxes impacted the scavenging intensities of Pa and Th. To do so, we use the preserved opal and 232 Th fluxes as tracers for past opal and terrigenous fluxes, respec-tively (Anderson et al., 2006;Lippold et al., 2012). Core MD09-3257 Pa / Th data are mainly influenced by oceanic circulation, except during the high lithogenic flux period associated with HS4 ( 232 Th flux > 12 dpm cm −2 ky −1 , Fig. 3a, white squares; Burckel et al., 2015). In core MD09-3256Q, opal fluxes do not covary with the Pa / Th ratio and are very low (0.01-0.02 g cm −2 kyr −1 , Burckel et al., 2016; Table S6; Fig. S4a). In the Atlantic, the lowest opal flux value observed to influence the sedimentary Pa / Th ratio is 0.2 g cm −2 kyr −1 (Lippold et al., 2012). Hence, given the much lower opal fluxes recorded in core MD09-3256Q and their lack of correlation with sedimentary Pa / Th, we conclude that biogenic silica had no (or very little) influence on Pa / Th variability. Similarly, we find no correlation between 232 Th fluxes and Pa / Th values in this core (P value = 0.48, n = 22; Fig. S5a). We can therefore safely assume that the Pa / Th variability recorded in our equatorial cores is mainly driven by changes in oceanic circulation intensity.
Core SU90-03 opal fluxes are low (< 0.1 g cm −2 kyr −1 , Table S6) and do not show any correlation with Pa / Th data (P value = 0.52, n = 10; Fig. S4b). In contrast, 232 Th fluxes could be correlated to the Pa / Th signal (P value = 0.03, n = 16, Fig. S5b). This correlation is only driven by the highest Pa / Th value, and removing this single value results in the disappearance of the correlation (P value = 0.31, n = 15). However, we chose to keep this value as SU90-03 232 Th flux is low (< ∼ 1.5 dpm cm −2 kyr −1 ) and its Pa / Th signal remains low (0.03-0.05) on the entire 20-50 ka period, indicating a constant significant Pa export through water mass advection. Oceanic circulation is therefore the main process explaining SU90-03 Pa / Th data.
The published Pa / Th values of other sediment cores used in this study have been shown to be mainly driven by oceanic circulation intensity (Bradtmiller et al., 2014).
3.2 C. wuellerstorfi δ13C data C. wuellerstorfi δ 13 C of both equatorial cores shows millennial-scale variability, with lower δ 13 C values occurring during HSs (Burckel et al., 2016). GS and GI are also recorded in the δ 13 C record of the intermediate core by low and high δ 13 C values, respectively. In the deep core, the low sedimentation rate induces a low temporal resolution and a smoothing of the δ 13 C signal that may have erased δ 13 C decreases associated with short GSs.
δ 13 C values during the LGM are 0.24 ± 0.07 and 0.66 ± 0.06 ‰ in the deep and intermediate equatorial core, respectively (Fig. 3). Given that the late Holocene δ 13 C measured in the equatorial cores are both close to 1.35 ‰ (see the Supplement), these LGM δ 13 C values are much lower than what would be expected from the ∼ 0.3 ‰ glacialinterglacial change in mean ocean δ 13 C in response to a reduced continental biosphere during glacial periods . The low δ 13 C values observed during the LGM and some of the GSs could thus indicate either a slow-down of the deep-water circulation, or an increased influence of southern-sourced water masses at both sites during the glacial with respect to the Holocene.

Ocean circulation signals
In cores MD09-3257/GeoB3910, SU90-03 and ODP1063, GI time slices are generally characterized by lower Pa / Th and higher δ 13 C values than those characterizing GS periods (Fig. 3). The only exception is GI-10 as this period is associated with a transition from relatively low to high Pa / Th values associated with HS4 in core SU90-03. Conversely, the HS2 time slice is associated with high Pa / Th values and low δ 13 C values in these cores, except for core SU90-03 that exhibits Pa / Th and δ 13 C values similar to GI time slices.
Core MD09-3256Q Pa / Th values are below 0.07 in all the studied time slices, with minor variability even between GI and HS2 time slices. Its δ 13 C record is systematically 0.4-0.5 ‰ below that of core MD09-3257, even though the δ 13 C difference between the two cores is reduced during HS2. Core MD09-3256Q therefore appears to be constantly bathed by a nutrient-rich water mass exporting dissolved protactinium.
There is only one Pa / Th value within our GI time slices in core MD02-2594 (GI-3, Fig. 3). However, two other Pa / Th values can be attributed to both GI-8 and GI-10 as they lie within these time slices considering age model uncertainties. These low MD02-2594 Pa / Th values indicate that, during GI time slices, protactinium was exported away from the intermediate-depth South Atlantic Ocean.
In addition to the above Pa / Th records, we use one Pa / Th value from core V29-172. This value lies within the GI-3 time slice and is rather low (0.04), similarly to the shallower North Atlantic core SU90-03. This supports the existence of Pa export between 1500 and 3500 m depth in the North Atlantic during GI-3.

Discussion
Time slice sedimentary Pa / Th data from the six selected Atlantic sediment cores are compared, when available, to the sedimentary Pa / Th pattern simulated in response to the different stream functions (Figs. 4-6). We focus in particular on the equatorial cores and describe their modeled and measured sedimentary Pa / Th by referring to the vertical Pa / Th gradient between 2300 and 3500 m (i.e., the water depths of the sediment cores). Indeed, vertical Pa / Th gradients are useful indicators of the vertical layout of water masses. In what follows, we will see how each vertical gradient can be interpreted.

Comparison with the Holocene simulation
A large vertical Pa / Th gradient between the two equatorial core sites is simulated by the model in the presence of a southward-flowing northern-sourced deep water mass such as NADW in the Holocene simulation (Fig. 4). This is due to the fact that the sedimentary Pa / Th ratio decreases with depth within a single water mass of uniform flow rate. This effect is intensified in the case of the Holocene stream function, as the flow rate of NADW is not uniform but stronger between 2500 and 3500 m and weaker between 1300 and 2300 m (acquisition depths of sedimentary Pa / Th for the deep and intermediate cores, respectively). Hence, Pa export at 3500 m is more intense than at 2300 m, thereby signifi-

Latitude (degrees)
HS2 data GI-3 data GI-8 data GI-10 data Shallow overturning streamfunction cantly increasing the vertical Pa / Th gradient between the two cores. Interstadial Pa / Th data in the deep equatorial and North Atlantic cores are consistent with simulated Pa / Th values obtained with the Holocene stream function (Fig. 4b-d). However, interstadial sedimentary Pa / Th values in the equatorial core at intermediate depth are lower than predicted by the Pa / Th model forced with the Holocene stream function. The vertical Pa / Th gradient between our equatorial cores during interstadials is small, which is in contradiction with the large vertical Pa / Th gradient simulated with the Holocene stream function. Moreover, data from Southern Ocean core MD02-2594 are systematically between 0.045 and 0.050 during GI (Fig. 3) and are in conflict with the high Holocene Pa / Th value (∼ 0.09) simulated at this core site (Fig. 4).
High δ 13 C values in the intermediate equatorial core during MIS3 interstadials suggest that northern-sourced deep waters influenced the equatorial Atlantic at 2300 m depth (Fig. 3b). However, the lower δ 13 C values of the deep equatorial core imply that, unlike in the present-day Atlantic, nutrient-rich southern-sourced deep waters were present at 3500 m depth in the equatorial West Atlantic.
Therefore, both Pa / Th and δ 13 C data indicate that the geometry and strength of the AMOC during the studied GI were different from those of the Holocene.

Comparison with the off-mode simulation
Pa / Th values simulated with the off-mode stream function exhibit a small vertical Pa / Th gradient between the two equatorial cores (Fig. 5). However, the sedimentary Pa / Th values measured during interstadials are much lower than the simulated values at both equatorial sites and in the North Atlantic Ocean. These low measured Pa / Th values imply a significant export of Pa by oceanic circulation and therefore exclude the possibility of an almost halted deep Atlantic circulation above 3500 m depth (Fig. 5b-d).
In addition, while in the off-mode stream function no significant deep convection occurs in the high-latitude North Atlantic, the high δ 13 C values of core SU90-03 and MD09-3257 and low δ 13 C values of core MD02-2594 (< 0.5 ‰, Negre et al., 2010), indicate that northern-sourced waters were present at ∼ 2500 m in the North and equatorial Atlantic (Fig. 3).
Hence, it is highly unlikely that the off-mode stream function depicts the deep Atlantic circulation during the studied GI.

Comparison with the shallow overturning simulation
The shallow overturning stream function also induces a small vertical Pa / Th gradient between the equatorial records (Fig. 6). In contrast to the simulation obtained with the off-mode stream function, this small vertical Pa / Th gradient is associated with significant lateral export of Pa at the depth of both the intermediate and deep equatorial cores, as well as at the Bermuda Rise and Southern Ocean cores, in agreement with Pa / Th data (Fig. 6b-d). Such a small vertical Pa / Th gradient is simulated in the case of two water masses overlying each other and flowing in opposite directions ( Fig. S6; Lippold et al., 2012). Indeed, in the shallow overturning stream function, northern-sourced waters affect the depth of the intermediate equatorial core (above 2500 m). Below ∼ 4000 m, northward-flowing, southern-sourced waters are active and can lead to a return flow (between 2500 and 4000 m depth) that influences the depth of the deep equatorial core (3500 m; Fig. 2e, f). This circulation scheme results in decreasing or invariant lateral export of Pa with depth, which in turns causes sedimentary Pa / Th to increase or to be constant with depth at the equator. However, the agreement between simulated and measured Pa / Th in the Atlantic Ocean cores is lower (i.e., larger Euclidean distances, Table S4) with the shallow overturning stream function than with the Holocene stream function. The better model-data agreement obtained with the Holocene stream function is driven by the Pa / Th data of the intermediate and deep midlatitude North Atlantic cores (SU90-03 and V29-172), as the deep convection of NADW induces low modeled Pa / Th values south of 50 • N. In the shallow overturning stream function, the region of deep-water formation is shifted southward (Fig. 2). In the Pa / Th model, dissolved protactinium and thorium concentrations are therefore vertically homogenized between 40 and 60 • N (against 60-70 • N in the case of the Holocene stream function), preventing the presence of low sedimentary Pa / Th values north of ∼ 40 • N. A southward shift in the deep convection zone during the last glacial has been observed in earlier studies (e.g., Vidal et al., 1997), but the simulated southward shift in the present shallow overturning stream function could be overestimated. Assuming a more northerly position of the region of deepwater formation, the Pa / Th observed in the North Atlantic cores would agree with the shallow overturning stream function. Indeed, if we remove both midlatitude North Atlantic cores from the computation of the sum of squared residuals, we find that the shallow overturning stream function best explains the Pa / Th data observed during GIs (Table S5).
Interstadial benthic δ 13 C values at the equator indicate the presence of (1) a northern-sourced water mass at the intermediate core site and (2) a southern-sourced water mass at the deep core site (Fig. 3b). Hence, benthic δ 13 C data support the existence of two water masses overlying each other and flowing in opposite directions as in the shallow overturning stream function. Moreover, the southern-sourced water mass likely affected core MD02-2594 on its way towards the deep equator site as reflected by its low benthic δ 13 C (< 0.5 ‰, Negre et al., 2010).
Based on our definition of the interstadial time slices, we assume that the GI-3 time slice reflects interstadial conditions. However, because GI-3 seen in Greenland ice cores is of relatively short duration, the Pa / Th signal of the studied sediment cores might not reflect full interstadial circulation conditions. Nonetheless, we consider it unlikely that the Pa / Th of GI-3 reflects stadial conditions. Indeed, core MD09-3257 sedimentary Pa / Th values observed during GI-3 are similar to those recorded during the GI-8 and GI-10 time slices that correspond to strict interstadials (Fig. 3). There are several studies discussing a potential link between Dansgaard-Oeschger events and Atlantic circulation changes (Gottschalk et al., 2015; see Boyle, 2000 for a review), but to our knowledge, there has been no study about the geometry of the AMOC water masses during these periods. Combining the information provided by sedimentary Pa / Th and benthic foraminiferal δ 13 C data, we reach the following conclusions concerning the Atlantic circulation below ∼ 1300 m during the studied GIs. A southwardflowing, northern-sourced water mass likely circulated above ∼ 2500 m, while southern-sourced deep water circulated northwards below ∼ 4000 m and southwards as a return flow between ∼ 2500 and 4000 m depth (Figs. 2f, 7a). Moreover, our data indicate that the geometry and state of the AMOC appear similar for GI-3, GI-8 and GI-10, despite the different ice sheet volumes characterizing the periods encompassing HS2 and HS4, respectively.

Estimation of the AMOC intensity over the interstadial time slices
Our Holocene equatorial Pa / Th values are in reasonable agreement with previously published data from the Brazilian margin  and with Pa / Th values simulated with a 2-and 3-fold increased Holocene stream function (Fig. S7). At present, increasing the Holocene stream function is indeed necessary to improve the agreement be-tween simulated and measured equatorial Pa / Th values . This increase was proposed to account for the absence of west-east difference in circulation strength in the 2-D Pa / Th model, which reflects a zonally averaged circulation. Moreover, the width of the Atlantic basin is the shortest at the equator, while it is assumed constant in the model. Both these effects could cause the flow speed at the equator to be underestimated and therefore the simulated Pa / Th ratio to be overestimated. We performed sensitivity tests of the Pa / Th model to varying flow rates by multiplying the shallow overturning stream function by a factor of 1, 2 and 3. Our results indicate that factors of 1 and 2 best reproduce the sedimentary Pa / Th ratio in the Atlantic during GI time slices (Table S5).
When using the shallow overturning stream function, the flow of both the northern and southern-sourced water masses at the equator is of 5-10 Sv. As this stream function amplified by a factor of 1 and 2 best agrees with our interstadial Pa / Th data, we assume that the water-mass flow rates provided by these shallow overturning stream functions (5-20 Sv) depict well the oceanic circulation strength during the studied interstadials. The modern flow rates of NADW (northern-sourced water mass) and AABW (southern-sourced water mass) are ∼ 27 ± 7 and ∼ 3 Sv at 4.5 • S, respectively (Lux et al., 2001). We conclude that during GI-3, GI-8 and GI-10, the flow rate of the southern-sourced deep water was likely larger than present-day AABW and that the flow rate of the northernsourced deep water may have been smaller than present-day NADW.

Comparison with the shallow overturning simulation
Pa / Th data from the HS2 time slice display a large vertical Pa / Th gradient between the depths of our equatorial cores (Figs. 4a, 5a or 6a). The gradient results from the high Pa / Th value at intermediate depth, which indicates that the core was likely overlain by sluggish waters. Pa / Th data in the equatorial cores are therefore incompatible with the low vertical Pa / Th gradient simulated by the shallow overturning stream function (Fig. 6, Table S4). Moreover, the decrease in benthic δ 13 C values in both equatorial cores during HS2 is incompatible with the active northern-sourced circulation cell simulated above 2500 m with the shallow overturning stream function (Fig. 3, b).
Hence, both circulation and ventilation proxies indicate that there was no intense northern-sourced water flow between 1300 and 2300 m depth at the equator during HS2.

Comparison with the Holocene simulation
Large vertical Pa / Th gradients are simulated in the model when a single water mass affects both equatorial cores. Pa / Th values simulated with the Holocene stream function therefore best fit Pa / Th data during HS2 (Fig. 4a, Table S4). However, the very low benthic δ 13 C values measured in both equatorial cores during HSs (Fig. 3b) exclude the presence of an active northern-sourced deep-water mass in the intermediate and deep equatorial Atlantic at that time.
Hence, it is unlikely that the Holocene stream function depicts the geometry and strength of the AMOC during HS2.

Comparison with the off-mode simulation
In contrast, Pa / Th values simulated with the off-mode stream function could reconcile the large Pa / Th vertical gradient and the low δ 13 C of equatorial cores (Fig. 5a). In the current off-mode simulation by the iLOVECLIM model, the vertical extent of the southern-sourced water mass is not large enough to influence the equatorial core at 3500 m depth, resulting in an apparent low vertical Pa / Th gradient (Fig. 2c,  d). However, the model was run for a short period of time, preventing a full response of the southern-sourced deep waters to varying climatic and oceanographic conditions (Roche et al., 2014). Hence, the vertical extent of the southernsourced water mass could be larger, thereby inducing a low sedimentary Pa / Th ratio in the deepest core of the Brazilian margin. Pa / Th data from the Bermuda Rise and South Atlantic cores are consistent with simulated Pa / Th values from the off-mode stream function. However, the Pa / Th data from the intermediate North Atlantic core are not in agreement with the modeled Pa / Th, as there is no deepwater formation in the high-latitude North Atlantic in the offmode stream function.
The very low benthic δ 13 C values measured in the deep equatorial core during HS2 are consistent with the simulation using the off-mode stream function, which shows a strong influence of southern-sourced water masses in the deep Atlantic. This is further supported by the increased ε Nd in the Bermuda Rise core which indicates an increased influence of southern-sourced water masses at 4500 m in the North Atlantic Ocean during HS2 (Gutjahr and Lippold, 2011).
The off-mode simulation could therefore explain both Pa / Th and δ 13 C values in the deep Atlantic and intermediate equatorial Atlantic, but neither this stream function nor the others are able to explain the Pa / Th record of the North Atlantic core at ∼ 2500 m depth.

Geometry and strength of the AMOC during HS2
The low Pa / Th and high benthic δ 13 C observed in core SU90-03 indicates that deep convection was still active in the high-latitude North Atlantic during HS2. However, the resulting water mass was probably slow enough so that its nutrient content significantly raised on its way towards the equator either through turbulent and diffusive mixing with the underlying water mass or through the degradation of 12 Crich organic matter sinking from surface waters. This would explain the low δ 13 C values measured in the intermediate equatorial core. Similarly, a long transit time of the water mass from the North Atlantic deep convection sites to the Brazilian margin could explain the high Pa / Th values measured in the intermediate equatorial core, as dissolved Pa was allowed sufficient time to equilibrate through reversible exchange with particulate matter. This progressive equilibration could have led to the sedimentary Pa / Th latitudinal gradient observed at ∼ 2500 m depth, from low sedimentary Pa / Th values in the midlatitude North Atlantic (SU90-03) to high values in the equatorial Atlantic (MD09-3257; Fig. 5a).
Considering that both Pa / Th and benthic δ 13 C values suggest an increased influence of southern-sourced waters at the equatorial and Bermuda Rise deep sites, the off-mode stream function best depicts the geometry of the AMOC during HS2 below 2500 m, i.e., southern-sourced deep waters likely filled the deep Atlantic Ocean (Fig. 7b). The influence of the southern-sourced water mass likely extended vertically above 3500 m depth, probably at least up to 2500 m depth, as indicated by the low Pa / Th value of the deep equatorial core. However, given the discrepancy between the offmode simulation and observed Pa / Th values in the North Atlantic intermediate core, it is difficult to assess the exact latitudinal and vertical extent of the southern-sourced waters. Above 2500 m, a weak water flow originating from northern deep convection sites probably influenced the Atlantic, perhaps down to equatorial latitudes (Fig. 7b).
The geometry of the Atlantic deep water masses inferred from our sedimentary Pa / Th and benthic δ 13 C records during HS2 is consistent with previous modeling experiments (e.g., Ganopolski and Rahmstorf, 2001) and with water mass ventilation (Elliot et al., 2002;Vidal et al., 1997;Zahn et al., 1997) and circulation intensity (Gherardi et al., 2005;Mc-Manus et al., 2004) proxies indicating that deep water circulation slowed down during Heinrich stadials. Furthermore, our data indicate that a northern-sourced water mass was active above 2500 m during HS2, as inferred for HS1 (Gherardi et al., 2009;Roche et al., 2014) and HS2 (Lynch-Stieglitz et al., 2014;Wary et al., 2015) in previous studies. However, quantifying the intensity of the AMOC upper circulation cell during this period remains difficult since there is at present no numerical simulation in reasonably good agreement with both the circulation and ventilation proxies measured in the Atlantic.

Conclusions
We have shown that both the geometry and strength of the AMOC during three interstadials of the last glacial period (i.e., the GI-3, GI-8 and GI-10 intervals) was markedly different from those of the modern AMOC. Our data suggest that a northern-sourced water mass circulated above 2500 m depth with a flow rate ranging between 5 and 20 Sv, which is lower than the intensity of present-day NADW. Below 4000 m, a southern-sourced deep water mass likely flowed northward with an intensity of 5-20 Sv, which is larger than the modern AABW flow rate. Between 2500 and 4000 m depth, the southern-sourced deep water likely circulated southwards as a return flow.
Our data also show that the geometry of the AMOC likely changed at the onset of HS2 and that the deep Atlantic below 2500 m was then probably dominated by a single southernsourced water mass that can be traced up to 35 • N at 4500 m. This water mass probably affected the equatorial Atlantic between 2500 and 3500 m depth. A slow, southward-flowing water likely circulated between 1500 and 2500 m in the North Atlantic, but its presence at the equator remains difficult to assess.

Data availability
Data related to this article are available as a supplement file and on PANGAEA (doi:10.1594/PANGAEA.865401).
The Supplement related to this article is available online at doi:10.5194/cp-12-2061-2016-supplement. for their constructive comments which helped to clarify this manuscript. This is LSCE contribution 5757.