Main

There is widespread evidence of pronounced Atlantic Meridional Overturning Circulation (AMOC) variability during glacial times associated with Dansgaard–Oeschger (DO) and Heinrich (H) events1,2,3,4,5. It is now widely accepted that DO events6 can occur as part of internal variability of the ocean–sea ice–atmosphere system7 involving spontaneous transitions between weak and strong AMOC modes under certain boundary conditions8,9,10, with alternating cold (Stadial) and warm (Interstadial, IS) conditions in the North Atlantic (NA). In addition, H events associated with large discharges of icebergs in the NA occurred repeatedly during glacial times11,12,13 and forced a weakening or collapse of the AMOC4,14 through the input of large amounts of freshwater into the NA15. Stadials containing H events are referred to as Heinrich Stadials (HSs).

While DO events have a large effect on the climate over Greenland16 and generally at high northern latitudes in the NA17, some H events also have a large impact on climate at high latitudes in the Southern Hemisphere (SH)18. In particular, it has been noted that HSs and non-Heinrich Stadials (nHSs), which are almost indistinguishable in Greenland ice core records, have a very distinct imprint on Antarctic climate and atmospheric greenhouse gases. Several proxy records hint at qualitative differences between HSs and nHSs that cannot be explained solely by changes in AMOC strength: (1) a large and abrupt CO2 increase by ~10–15 ppm on centennial timescales during HSs as opposed to a steady decrease of CO2 during nHSs19,20,21,22, (2) a large and abrupt warming of 2–3 °C over Antarctica during HSs18,23 and (3) a sudden jump in atmospheric methane concentration during HSs24. Various proxies point to a crucial role played by the Southern Ocean (SO) to explain the peculiar dynamics during HSs, in particular through an increased ventilation of the SO25,26,27,28,29,30, caused by shifts in the westerlies21,31,32,33, enhanced convection and deep water formation28,29,34,35, or a combination of the two processes36.

However, modelling studies so far have failed to realistically reproduce the timing and amplitude of Antarctic temperature and atmospheric CO2 variations in response to freshwater hosing in the NA37,38,39,40,41,42,43,44,45, except for simulations where convection in the SO was enforced to occur by applying an artificial negative freshwater flux at the surface in the SO34. Therefore, a mechanistic understanding of the relation between Northern Hemisphere (NH) and SH climate response to AMOC variations during H events remains elusive. Here we use a fast Earth system model with fully interactive carbon cycle to investigate the Earth system evolution during DO and H events and show that deep convection is triggered in the SO as a response to a strong AMOC weakening following some H events, explaining the observed large and abrupt Antarctic warming and atmospheric CO2 increase.

Simulated millennial-scale climate variability

We use a fast Earth system model46,47 with interactive atmospheric CO2 to simulate millennial-scale glacial climate variability (Methods). The model simulates spontaneous DO events under typical mid-glacial conditions10, with jumps between Stadials and ISs being associated with transitions between two different AMOC states10. We additionally mimic the effect of a H event iceberg discharge through a prescribed plausible input of a freshwater flux of ~0.1 Sv (ref. 15) into the NA (Methods) and investigate the Earth system response to combined DO and H events in the model. The simulated H event is representative for the time interval around HS4 (Methods), which has been particularly extensively studied with many different high-resolution proxy records available for this period of time (Fig. 1).

Fig. 1: Proxy-model comparison of millennial climate variability.
figure 1

a–n, Comparison of simulated (h–n) climate variability with proxy reconstructions (a–g) for a time interval around HS4: strength of the AMOC (ref. 4 in black and ref. 57 in blue) (a and h); Greenland (Grl) temperature16 (b and i); Iberian Margin sea surface temperature (SST)58 (c and j); Antarctic (Ant) temperature averaged over four ice cores23 (d and k); atmospheric CO2 concentration20,21 (e and l); δ13C of CO2 (ref. 49) (f and m); and CH4 concentration24 (g and n). The dashed lines in l and m show results of a model simulation where the land carbon cycle is not active. The dotted line in n shows the CH4 concentration in a simulation with a prescribed constant CO2 of 200 ppm. The shaded areas indicate the HS4 (grey) and a nHS (green). The vertical dashed red lines mark the timing of the jump in CH4 and CO2 in the left column and the onset of convection in the SO in the right column.

Full size image

The model captures the shape of the temperature response over Greenland (Fig. 1b,i), reflecting mainly DO variability associated with transitions in the AMOC state. At the Iberian Margin, the sea surface cooling during the HS is as much as three times larger than during nHSs (Fig. 1c,j). Over Antarctica, temperatures increase rapidly during the middle of the HS, by ~3–4 °C in the model and by ~2–3 °C in proxy data. By contrast, during nHSs, the temperature increase is about three to five times smaller and occurs more gradually over the entire duration of the Stadial (Fig. 1d,k).

The simulated CO2 response to DO variability is generally within ~5 ppm, in agreement with observations (Fig. 1e,l) and previous modelling results44. However, during the HS, the model simulates a rapid increase in atmospheric CO2 by ~15 ppm (Fig. 1e,l), in good agreement with ice core records20,21. In the model, the CO2 response is partly dampened by the land carbon cycle absorbing part of the CO2 released by the ocean (Fig. 1l), while a recent study suggests that substantial increases in biomass burning, which is not represented in our model, could have contributed to the very rapid initial increase and temporary overshoot seen in CO2 ice core data during some HSs48. The increase in CO2 is accompanied by a simulated ~0.2‰ decrease in δ13C–CO2, consistent with ice core data49 (Fig. 1f,m). The model also qualitatively reproduces the response of atmospheric CH4 concentration to DO and H variability (Fig. 1g,n).

Overall, the model reproduces the main features of climate and carbon cycle variability associated with DO and H events, in particular the qualitative differences between HSs and nHSs.

SO convection during HSs

The simulated millenial-scale climate variability is tightly linked to AMOC variations (Fig. 2a). The Atlantic Ocean circulation is fundamentally different during some HSs compared with nHSs, with proxy data supporting a strong AMOC response during HS4 (Fig. 1a,h). nHSs are characterized by a substantially weaker AMOC compared with ISs (Fig. 2g,h), but with convection still active at several locations in the NA (Fig. 2d,e), while the large freshwater input associated with the strong H event 4 leads to a rapid stop of convection and eventual shutdown of the AMOC in the model (Fig. 2f,i). Initially, this results in a widespread cooling in the NA region, but only limited localized warming in the South Atlantic. However, around 1,000 years after the AMOC collapse, deep convection first develops in the Ross Sea (Extended Data Fig. 1) and after some time rapidly spreads around Antarctica (Fig. 2b,f). The enhanced deep convection in the SO also leads to a strengthening of Antarctic bottom water formation (Fig. 2a,k,l). Although there are no direct proxies for SO circulation changes, there is ample evidence for an increased ventilation of the deep SO during past H events. Deep-water cooling29 (Extended Data Fig. 2c), decreased radiocarbon ventilation ages28 (Extended Data Fig. 2f), increased oxygenation of the deep SO26 (Extended Data Fig. 2i) and an increase in primary productivity due to increased nutrient upwelling31 (Extended Data Fig. 3f) all provide support for a more vigorous convection in the SO. Convection in the SO then eventually stops after ~500 years, as soon as the AMOC starts to recover and convection resumes in the NA (Fig. 2a,b and Extended Data Fig. 1).

Fig. 2: The bipolar convection seesaw.
figure 2

NA versus SO climate evolution in model simulations for mid-glacial conditions with internally generated DO events and a prescribed H event. a, Time series of the maximum of the AMOC streamfunction (red), Antarctic bottom water (AABW) formation rate (blue) and freshwater flux (FWF) applied to represent a H event in the NA (grey). b, Time series of the potential energy released by convection (PE conv) in the NA and the SO (south of 55° S). c, Time evolution of the maximum sea ice area in the two hemispheres. The shaded areas in a–c indicate the HS and a nHS. The vertical dashed red lines mark the onset of convection in the SO. d–f, Maximum mixed layer depth for three time intervals marked by the magenta intervals in a representing DO IS (d), nHS (e) and HS (f) conditions. Blue lines indicate the maximum sea ice extent, and the dotted lines mark the parallels at 60° S and 30 and 60° N. EQ, equator. g–l, For the same three time periods, the Atlantic (g–i) and global (j–l) meridional overturning streamfunctions are also shown.

Full size image

The abrupt onset of convection has a pronounced impact on climate at high southern latitudes (Fig. 3b), resulting in a large warming due to heat released from the deep ocean and consequent large retreat of sea ice (Figs. 2c and 3b). The increase in SO ventilation (Fig. 3f) resulting from the enhanced convection and the associated sea ice retreat brings large amounts of carbon that were stored in the deep ocean (Fig. 3h and Extended Data Fig. 2l) in contact with the atmosphere, leading to a rapid release of carbon (Fig. 3d) and a CO2 increase of ~15 ppm within a few centuries (Fig. 1l).

Fig. 3: Climate and ocean biogeochemical changes during a DO cycle and an H event.
figure 3

a–h, Differences in simulated variables during nHSs (a, c, e and g) and HSs after the onset of SO convection (b, d, f and h) relative to IS conditions for annual mean near-surface air temperature (a and b), net carbon flux to the atmosphere (c and d), zonally averaged radiocarbon ventilation age (e and f) and zonally integrated dissolved inorganic carbon content (g and h). The filled circles in a and b represent temperature changes estimated from proxy records16,23,58. The green lines in a–d show the maximum sea ice extent during the IS, while the grey lines represent the maximum sea ice extent during nHS (a and c) and HS (b and d). All ocean fields in e–h are shown separately for the Atlantic Ocean (Atl) and the Indo-Pacific (Indo-Pac) Oceans, with the SO sectors being included.

Full size image

Rhodes et al. (2015)24 suggested that the jump in CH4 seen in ice core records during HS1 reflects the beginning of the H event, and this argument has been extended also to previous H events50. However, our results show that the CH4 jump in the middle of the HSs rather reflects the timing of the onset of convection in the SO, which is delayed by many centuries relative to the start of the actual H event. In our simulation, the CH4 increase following convection onset in the SO is mainly a consequence of larger emissions from the tropics (Extended Data Fig. 3i), resulting from a combination of a general increase in precipitation over SH land (Extended Data Fig. 3c) and an increase in net primary production due to CO2 fertilization (Extended Data Fig. 3f ).

The bipolar convection seesaw

The term bipolar seesaw was originally introduced by Broecker (1998)51, who proposed an anti-phase response of deep sea ventilation in the NA and the SO to explain the Antarctic temperature response during the Younger Dryas. Later, Stocker and Johnsen (2003)52 introduced the thermal bipolar seesaw concept to explain the temperature response in Antarctica resulting from changes in the AMOC during glacial times. In their simple and purely thermodynamic model, the changes in meridional heat transport induced by AMOC changes combine with a heat reservoir in the SO to produce the Antarctic temperature response to AMOC perturbations. This simple model has since been widely invoked, with some refinements, to explain the relation between Greenland and Antarctic temperature evolution38,53,54.

However, the thermal seesaw alone cannot explain the qualitative differences between nHSs and HSs. Skinner et al. (2020)29 suggested that convection in the SO occurred as a response to an AMOC weakening/collapse during HS4 in what they call a bipolar ventilation seesaw55,56, which amplified Antarctic warming and the atmospheric CO2 response. However, they do not relate the SO convection directly to the H event and can therefore not explain why convection would occur only during some HSs, but not during nHSs.

Here we propose that, to explain the observed millennial-scale glacial climate variability, the bipolar thermal seesaw should be complemented by the concept of a bipolar convective seesaw. The former always operates in response to AMOC changes via changes in interhemispheric heat transport. The latter operates only during some HSs, when the AMOC weakening is large enough to substantially affect salinity in the deep SO. An AMOC weakening generally leads to a salinity pile-up in the upper South Atlantic and a decrease in salinity in the deep Atlantic (Fig. 4b,c). In the case of an AMOC collapse following a H event, the response is generally more pronounced compared with the weak AMOC state during nHSs and the salinity decrease extends well into the deep SO (Fig. 4c). The resulting general salinity-induced destratification of the SO (Fig. 4h, Extended Data Figs. 4 and 5) creates the conditions for convection to start in the Ross Sea, an area that was already very close to convective instability (Extended Data Fig. 5a). Once convection starts somewhere, salty water from the deep ocean is mixed with low-salinity surface water, increasing sea surface salinity (Fig. 4d,h and Extended Data Fig. 4). This positive salinity anomaly then spreads horizontally, resulting in an increase in surface density in neighbouring areas, eventually cascading into large-scale convection. A positive feedback between deep convection and surface salinity enables the rapid expansion of convection around Antarctica. Temperature changes play only a secondary role for changes in stratification around Antarctica (Extended Data Figs. 4 and 5).

Fig. 4: Bipolar convection seesaw mechanism.
figure 4

a IS state, showing Atlantic Ocean and SO (south of 50 °S, dotted line) salinity, global overturning streamfunction (black; solid lines for clockwise and dashed lines for anti-clockwise circulation, drawn at steps of 8 Sv), zonally averaged sea ice concentration in the NA and SO, zonally averaged maximum mixed layer depth (MLD) (green line), northward Atlantic heat transport at 30° N (red arrow) and global southward heat transport at 60° S (blue arrow). b, The same as a but for nHSs, showing salinity differences relative to the IS. The dashed green line and the dotted sea ice concentration show the reference IS state. c–e, The same as b but for the HS before (c), during (d) and after (e) convection in the SO. f, Potential energy (PE) released by convection, the same as Fig. 2b. g, Northward Atlantic heat transport at 30° N (red) and global southward heat transport at 60° S (blue). h, Average sea surface salinity and deep ocean salinity (below 1,000 m) in the SO (south of 60° S). The numbered time intervals marked in grey in f–h show the time slices represented in a–e.

Full size image

Convection in the SO is then sustained by an increased southward ocean heat transport in the SO (Fig. 4g), caused by the stronger SO overturning circulation associated with the enhanced Antarctic deep water formation (Fig. 2l). While the sea surface is efficiently cooled by the atmosphere, the subsurface warming resulting from the increased heat transport keeps convection going. As soon as the H event ends, the AMOC slowly starts to recover (Fig. 2a), increasing also the northward ocean heat transport in the Atlantic and consequently decreasing the southward heat transport in the SH (Fig. 4g) until convection can no longer be sustained in the SO (Fig. 4e and Extended Data Fig. 1). The AMOC recovery is therefore the ultimate cause of the convection stop in the SO. This is further confirmed by additional simulations in which the Heinrich event is extended in duration, leading to a longer persistence of the AMOC off state (Extended Data Fig. 6c). Convection in the SO always remains active until the eventual recovery of the AMOC, keeping CO2 levels and Antarctic temperature high throughout this time period (Extended Data Fig. 6f,i). The convective bipolar seesaw is thus fully controlled by AMOC dynamics. The quick resumption of the AMOC after the end of the H event occurs because the AMOC off state is not stable under the considered boundary conditions (Extended Data Fig. 7), leading to a spontaneous AMOC recovery once the freshwater forcing is removed.

Robustness of the bipolar convection seesaw

The bipolar convection seesaw is a robust feature of our model. It does not depend on the timing of the H event during the DO cycles (Extended Data Fig. 6a,d,g) and is relatively insensitive to model parameters (Extended Data Fig. 8). It also works under very different boundary conditions in terms of CO2 and ice sheets (Extended Data Fig. 9), although the amount of freshwater flux needed to weaken the AMOC sufficiently to trigger convection in the SO varies with changing boundary conditions. Changes in the SH westerlies are not needed to explain the Earth system response to H events in our model, as shown by a simulation with fixed wind stress (Extended Data Fig. 10).

H events occurred under different conditions in terms of orbital configuration and greenhouse gases concentrations, and reconstructions also indicate that the magnitude of ice discharge varied substantially across different H events15. The response seen in many proxies is large during some HSs, such as HS4 and HS5, but is less apparent during, for example, HS2 and HS3. Our results suggest that conditions during HS4 and HS5 were generally more favourable for a strong bipolar convection seesaw to operate (Fig. 5), with the response being additionally modulated by the magnitude of the H events and the amplitude of the induced AMOC weakening (Fig. 5 and Extended Data Fig. 6b,e,h). In our model, a very weak or collapsed AMOC is a necessary condition to produce a sufficient decrease in deep ocean salinity, which subsequently triggers convection in the SO, while it is clear that the weak AMOC state associated with nHSs (~10 Sv) does not have a sufficiently large impact on the stratification of the SO to trigger convection there.

Fig. 5: The bipolar seesaw during different HSs.
figure 5

a–p, Dependence of the simulated AMOC (a, e, j and m), potential energy released by convection (PE conv) in the SO (b, f, j and n), Antarctic temperature (c, g, k and o) and atmospheric CO2 (d, h, l and p) to different boundary conditions (greenhouse gases concentration and orbital configuration) corresponding to the occurrence of different HSs during the last glacial period (HS5, HS4, HS3 and HS2, from left to right), assuming a range of amplitudes of the H events (different colours). The dotted lines in the top panels (a, e, i and m) show the freshwater hosing flux (FWF) applied to the NA to mimic the different H events.

Full size image

The robustness of the proposed bipolar convection seesaw with respect to different climate and boundary conditions suggests that it could also operate in response to a possible future collapse of the AMOC under global warming, with potentially large impacts on regional and global climate and the Antarctic ice sheet. We suggest that SO convection should be considered as a tipping element in the Earth system, with the bipolar convection seesaw forming a previously unidentified potential cross-hemispheric link between different tipping elements.

Methods

Earth system model

We use the CLIMBER-X Earth system model46,47, including the frictional-geostrophic three-dimensional ocean model GOLDSTEIN59,60 with 23 vertical layers, the semi-empirical statistical-dynamical atmosphere model SESAM46, the dynamic-thermodynamic sea ice model SISIM46, the land surface model with interactive vegetation PALADYN61 and the ocean biogeochemistry model HAMOCC662,63,64. The comprehensive carbon cycle in the model allows one to interactively compute the atmospheric CO2 evolution. Here we use the ’closed’ carbon cycle set-up, in which marine sediments and chemical weathering on land are ignored. In this set-up, carbon is conserved in the atmosphere–ocean–land system, which is a reasonable assumption on millennial scales. All components of the model have a horizontal resolution of 5° × 5°. Ice sheets are prescribed and the net freshwater flux from ice sheets is zero. The model is described in detail in ref. 46 and ref. 47 and in general shows performances that are comparable with state-of-the-art CMIP6 models under different forcings and boundary conditions. Notably, the model has been shown to reproduce DO-like variability under mid-glacial conditions10, and the stability of the AMOC in the model has been thoroughly explored65.

Experiments

The main model experiment is designed to simulate realistic millenial-scale climate variability during glacial times, that is, Marine Isotope Stage 3, and more specifically a time interval around HS4. The boundary conditions for this simulation include mid-glacial ice sheets from the GLAC-1D reconstruction66, a CH4 concentration of 425 ppb, a N2O concentration of 230 ppb and the orbital configuration of 40 kyr before present (BP). Similarly to ref. 10, we apply noise in the surface freshwater flux in the northern NA so that the model produces robust internal DO cycles. We first perform a model spin-up of 10,000 years with a prescribed atmospheric CO2 concentration of 200 ppm, which is representative for the HS4 interval, to allow the climate and carbon cycle to reach an (oscillating) equilibrium state. Starting from that, we then run a 10,000-year-long simulation with interactive CO2. In this simulation, we introduce a plausible H event starting in simulation year 2,200, during a DO IS phase. The prescribed temporal shape of the freshwater flux associated with the H iceberg discharge event is derived qualitatively using results of different ice sheet model simulations67,68,69,70, with a peak freshwater flux of 0.13 Sv, followed by a gradual decline until the event ends around 1,200 years later (Fig. 2a). Spatially, the freshwater flux is added uniformly to the ice-rafted-debries belt in the NA, between 40° N and 60° N and between 10° W and 70° W. No compensation of the freshwater flux is applied, and the average ocean salinity decreases by ~0.1 psu by the end of the H event.

To separate the effect of changes in wind stress over the ocean on the model response to DO and H events, we have run an additional simulation in which the seasonal wind stress fields are kept constant at their initial (nHS) state.

A simulation where only the ocean carbon cycle is allowed to contribute to the atmospheric CO2 evolution, while the land carbon fluxes are ignored, was additionally performed to isolate the contribution of ocean carbon cycle processes to atmospheric CO2.

To quantify the effect of CO2 fertilization on the increase in CH4 emissions following the onset of SO convection, we performed an additional simulation with the same boundary conditions as the reference model run, but in which atmospheric CO2 is prescribed at a constant value of 200 ppm.

In addition to the reference model simulation, we performed a number of simulations to explore the sensitivity of the results to the amplitude (×0.25, ×0.5, ×0.75 and ×1.5), timing (H event starting at the beginning and in the middle of a Stadial) and duration of the H event (idealized constant 0.1 Sv freshwater flux for 1,000, 2,000, 3,000 and 4,000 years).

We also run an ensemble of simulations with perturbed parameters to assess the robustness of the results with respect to changes in ocean model parameters (minimum and maximum diapycnal diffusivities, Gent–McWilliams diffusivity and the maximum slope of the isopycnals). To avoid having to run a spin-up of the carbon cycle for each of the ensemble members, these experiments are run with a prescribed constant CO2 of 200 ppm, starting from the same initial condition as the reference run, and the H event is applied starting from the year 3,000.

To explore the dependence of the results on the background climate conditions, the set of simulations for HS4 boundary conditions with different amplitudes of the H event were extended also to other time intervals during MIS3 corresponding to the times when HS5, HS3 and HS2 occured. For HS5, we used CO2 = 210 ppm, CH4 = 450 ppb, N2O = 240 ppb and an orbital configuration of 49 kyr BP. For HS3, we used CO2 = 190 ppm, CH4 = 400 ppb, N2O = 215 ppb and an orbital configuration of 31 kyr BP. For HS2, we used CO2 = 185 ppm, CH4 = 375 ppb, N2O = 200 ppb and an orbital configuration of 25 kyr BP. The orbital parameters are from ref. 71.

To further investigate the robustness of our results to a wider range of boundary conditions, we performed additional simulations with present-day and last glacial maximum ice sheets and with different constant CO2 concentrations (180, 220 and 280 ppm). For this more idealized set of simulations, we used present-day orbital parameters. These simulations are run for 10,000 years, and the H event is applied starting in the year 5,000, to give the system enough time to equilibrate with the different boundary conditions. The initial condition for these experiments is a preindustrial equilibrium state. We also repeated these simulations with the amplitude of the H event halved and doubled.