The Last Interglacial (LIG, ~ 129 to 116 thousand years before present, hereafter ka) was probably the warmest interglacial period of the last 800 ka1. While the concentration of atmospheric CO2 was 10 ppm lower at the peak of the LIG ( ~ 127 ka; 275 ppm) than during the pre-industrial period (PI, 1850 CE)2, the Earth’s orbital parameters were such that the mean insolation north of 60 °N was 47 W m−2 higher between May and July, and 45 W m−2 lower between August and September. At high southern latitudes (south of 60 °S), it was 46 W m−2 higher between August and October, but 36 W m−2 lower between December and February (Fig. S1).

The global mean sea-level (GMSL) at the LIG was likely ~ 3-6 m higher than during PI, although estimates range from 1.2 m to 9.4 m3,4,5. A compilation of sea-level geological records from 42 localities suggested that the LIG peak GMSL was 6.6 to 9.4 m above present-day5. A subsequent reconstruction based on coral U-Th ages agreed with this figure, estimating a range between 5.5 and 9 m GMSL above present3. However, a recent study revised LIG sea-level data across the Bahamas with a suite of glacial isostatic adjustment models to suggest that the LIG GMSL peak was at least 1.2 m but unlikely more than 5.3 m above PI4. While there are still large uncertainties associated with the exact timing of the peak GMSL and potential centennial to millennial-scale sea-level variability during the LIG6, most studies suggest a highstand between ~ 130 and 125 ka5,7.

It is estimated that the mean ocean temperature peaked at 129 ka, leading to a maximum of 0.7 m thermosteric sea-level rise8. While paleo-proxy records and modelling studies suggest a possible Greenland ice-sheet (GrIS) contribution of 0.6 to 5.1 m sea-level equivalent (s.l.e), with a mean contribution of 2 m s.l.e., most studies suggest that the GrIS ice-mass loss occurred after 125 ka7,9,10,11,12. This implies that the Antarctic ice-sheet (AIS) was the major contributor to the early LIG highstand, with possible estimates between 3 and 7.4 m s.l.e5,7,10. A recent study agrees with this range, suggesting the AIS contribution to be of maximum 5.7 m (with 3.6 to 8.7 m central 68% probability)13. An early LIG ice-mass loss from the West AIS is consistent with a Southern Ocean authigenic uranium record, indicating enhanced stratification due to meltwater input14. A retreat of the West AIS in the early LIG has also been suggested by a blue ice record15, while sedimentological and geochemical records suggest a retreat of the Wilkes Subglacial Basin in East Antarctica during the LIG16.

The processes leading to a substantial ice-mass loss from Antarctica at the LIG are, however, poorly constrained. In order to simulate an Antarctic ice volume lower than today, Antarctic ice-sheet models have to be forced by a subsurface warming on the Antarctic shelf on the order of 3 °C17,18. Whether such a large subsurface warming on the continental shelf at the LIG is possible remains to be demonstrated. To address this question, here we analyse the Southern Ocean subsurface temperature response to LIG boundary conditions, as simulated in the Coupled Model Intercomparison Project Phase 6 (CMIP6) – Paleoclimate Modelling Intercomparison Project 4 (PMIP4) lig127k experiment19, performed with the Australian Earth System Model – ACCESS-ESM1.520. We further assess the processes leading to the simulated subsurface warming on the Antarctic shelf.


Subsurface warming around Antarctica

This quasi-equilibrium ACCESS-ESM1.5 lig127k experiment follows the CMIP6-PMIP4 experimental protocol19 and is forced by 127 ka orbital parameters and greenhouse gases, while GrIS and AIS configurations are set at their PI levels (Methods). Compared to PI, the ACCESS-ESM1.5 lig127k simulation features 1.1 °C higher sea-surface temperature (SST) in the Atlantic and Indian sectors of the Southern Ocean (south of 50 °S, 70 °W–160 °E) during austral summer, as well as a ~ 1 °C higher annual mean surface air temperature over Antarctica, in agreement with most paleo-proxy records (Text S1, Fig. 4a, S2a)21,22,23,24,25. Comparison to paleo-proxy data is discussed more in-depth in Text S1.

The lig127k simulation displays an increase in subsurface ocean temperatures on the Antarctic shelf compared to the PI control simulation (Fig. 1a). As ice shelves can reach ~ 300 m depth and the edge of the Antarctic continental shelf lies at ~ 500 m depth, we focus our analysis on the ocean model depth 385–601 m, where the warming averages 1.3 °C in the Atlantic (70 °W–30 °E) and 1.2 °C in the Indian (30 °E–160 °E) sectors of the Southern Ocean, and extends from the Antarctic coast to 60 °S. In the Ross Sea, the warming is 0.7 °C and also extends to 60 °S, whereas the warming in the Bellinghausen and Amundsen Seas is weaker and closer to the coast. This pattern of regional temperature anomalies is maintained between 200 and 1000 m depth.

Fig. 1: Ocean temperature, salinity and age anomalies.
figure 1

a Ocean temperature, (b) salinity and (c) age anomalies at the LIG compared to PI averaged across 385–601 m depth. Zonally-averaged (d) ocean temperature, (e) salinity, and (f) age anomalies (shading) with density anomalies in kg m−3 overlaid (contours).

The warming extends to the sea floor south of 60 °S, with a mean temperature increase of 1.1 °C (Fig. 1d). The warming averaged across all depths is largest in the Indian sector (+1.4 °C) in particular close to the Antarctic continental slope (Fig. 2b), while ocean temperatures are 0.7 °C higher than PI in the Pacific sector (160 °E to 70 °W) (Fig. 2c). In all three basins, the warming is largest below ~1000 m depth. The oceanic temperature anomaly between LIG and PI south of 60 °S averages +1.2 °C above 1000 m (+1.5 °C below) in both the Atlantic sector and the Indian sector, and +0.3 °C (+0.9 °C below) in the Pacific sector (Fig. 2a–c).

Fig. 2: Ocean temperature anomalies in the Atlantic, Indian and Pacific sectors of the Southern Ocean.
figure 2

Zonally averaged (ac) ocean temperature, (df) salinity at the LIG compared to PI over the (a, d) Atlantic (70 °W to 30 °E), (b, e) Indian (30 °E to 160 °E) and (c, f) Pacific (160 °E to 70 °W) sectors. Contours are potential density anomalies in kg m−3.

This subsurface warming in the Southern Ocean is associated with an increase in the age of the water-masses across all depths and south of 50 °S at the LIG compared to PI (Fig. 1f). The only exception is a canyon near the Ross Sea, below 3000 m depth, which displays negative age anomalies (Fig. 1f). The simulated water-mass age is defined here as the average time since the water in a particular model grid cell has last been in contact with the surface. An 83-year increase in the mean age of the water-masses is simulated south of 60 °S, between 1000 and 3000 m depth. The largest age increase is simulated in the deep and abyssal ocean north of 50 °S, averaging +116 years below 4000 m depth. The strong relationship between changes in temperature and age suggests that the simulated increase in subsurface ocean temperatures is linked to changes in oceanic circulation (Fig. 1d, f).

Changes in deep-ocean convection

A weakening of the lower overturning cell is simulated in our LIG experiment (Fig. 3d–f), with a 4.5 Sv weakening of Antarctic Bottom Water (AABW) transport at 30 °S compared to PI. This weakening of AABW transport is caused by a large reduction in deep-ocean convection at the LIG compared to PI in the Weddell, Lazarev and Ross Seas (Fig. 3a–c). Under PI conditions, deep ocean convection in the Weddell, Lazarev and Ross Seas brings very cold, close to freezing point, water to depth (Fig. S3a, c). Weakened deep-ocean convection at the LIG explains the relative warming and ageing signal observed in those areas (Fig. 1a, c). The shift of convection sites (westward near the Weddell Sea and Lazarev Sea, and eastward in the Ross Sea; Fig. 3a–c) further contributes to the local warming seen in Fig. 1a. The structures of the temperature and age anomalies in the abyssal ocean are also consistent with reduced convection and weaker AABW (Fig. 1d, f).

Fig. 3: Winter mixed layer depth (MLD) and meridional overturning stream function.
figure 3

a MLD during austral winter (June to August) in PI, (b) LIG and (c) LIG-PI, and (d) global stream function in PI, (e) LIG and (f) LIG-PI. A paleo-calendar adjustment is applied to the LIG data in (b) and (c) such that its months have the same angular definition as PI as described in Bartlein and Shafer68.

Higher SST and lower sea-surface salinity (SSS), associated with reduced sea-ice formation on the Antarctic shelf, decrease surface water density, thus leading to reduced deep-ocean convection. The lig127k simulation features an annual mean SST increase of 0.7 °C south of 60 °S25. This is associated with an increase in poleward oceanic advective heat transport south of 50 °S in the top 200 m, reaching +0.11 PW at 55 °S (Fig. S4c, d). This contributes to a SST increase south of 60 °S of 0.4 °C in austral winter and 0.8 °C in summer compared to PI (Fig. 4a, b), and to the 41% annual-mean sea-ice area decrease (Fig. 4c, S2b, S5).

Fig. 4: Southern Ocean sea-surface temperature (SST) and salinity (SSS) anomalies in summer, winter, and their annual means.
figure 4

a Austral summer (December to February), (b) austral winter (June to August), and (c) annual mean SST anomalies at the LIG compared to PI; and (d) austral summer, (e) austral winter, and (f) annual mean SSS anomalies at the LIG compared to PI, with 15% sea-ice area contours in (orange) PI, (blue) LIG, and (black) NOAA_OI_v2 dataset for 1982–200174 overlaid. Filled markers are proxy temperatures from the compilation by Chandler and Langebroek23. A paleo-calendar adjustment is applied to the LIG data in (a), (b), (d), and (e) such that its months have the same angular definition as PI as described in Bartlein and Shafer72.

The summer sea-ice area is particularly reduced in the Weddell Sea as well as the Amundsen Sea, whereas in winter there is a large reduction in the Lazarev Sea (Fig. 4a, b). Sea-ice reduction combined with the insolation anomaly leads to an increase in downwelling shortwave radiation over the Southern Ocean throughout most of the year (Fig. S6). The total energy flux into the ocean’s surface (Qnet) is higher at LIG than PI between May and October between 40 °S and 65 °S (Fig. S4a), thus leading to an annual mean positive Qnet anomaly of 0.015 PW (Fig. S4b). Qnet is also higher at the LIG south of 65 °S between Oct and Nov, but the annual mean anomaly is negative there (Fig. S4b).

Due to increased insolation during austral winter and spring (60 °S between days 168 and 317; Fig. S1), the annual mean sea-ice formation is reduced by 40% at the LIG compared to PI (Table 1, Fig. S7), leading to reduced brine rejection. Expectedly, there is also a 36% reduction in the annual mean sea-ice melting south of 60 °S. The net freezing south of 60 °S is 1447 Gt yr−1 lower compared to PI, leading to fresher surface waters around Antarctica (Fig. 4d–f). The decrease in sea-ice area in turn lowers the albedo, thus enhancing the warming and further lowering the sea-ice cover. Reduced seasonality induced by the LIG orbital parameters (Fig. S1) therefore leads to reduced sea-ice formation (Table 1, Fig. S7), which lowers surface salinity in winter (Fig. 4e) and weakens the formation of deep-waters26,27.

Table 1 Freshwater flux into the ocean due to sea-ice formation and melt

The annual-mean surface freshening is particularly large in the Atlantic and Indian sectors of the Southern Ocean, averaging −0.1 psu south of 65 °S (Fig. 2d,e, 4f). The negative SSS anomalies in the Indian sector (Fig. 4f) are situated closer to shore and are associated with increased sea-ice melting (Fig. S7). The freshening in the Pacific sector (-0.07 psu south of 70 °S) is relatively weaker (Fig. 2f). The freshening extends to a depth of ~1000 m but gradually decreases with depth (Fig. 1b, 2d-f). Enhanced stratification leads to positive salinity (+0.09 psu south of 60 °S) and temperature anomalies below 1000 m depth (Fig. 1e), particularly in the Atlantic and Indian sectors of the Southern Ocean (Fig. 2a, b, d, e).

In contrast, positive SSS anomalies are simulated in the seasonal sea-ice zone as well as between 60 °S and 50 °S in the Atlantic and Indian sectors (Fig. 4d-f). These positive SSS anomalies result from reduced sea-ice melting (Fig. S7) as well as stronger evaporation over precipitation (Fig. S8).

In summary, lower SSS due to a reduction of sea-ice formation, an increase in sea-ice melting in the Indian sector ( ~ 65 °S), and an increase in precipitation over Antarctica leading to higher coastal runoff (Text S2, Fig. S9), combined with warmer surface conditions, lower the density in the top ~ 1000 m at the LIG compared to PI. In contrast, the density below 2000 m depth south of 60 °S is higher at the LIG due to an increase in salinity. There is thus a strong increase in stratification in the Southern Ocean at the LIG compared to PI in our simulation.

Advection of warm waters to East Antarctica and along the Antarctic peninsula

While the subsurface warming simulated in the Weddell, Lazarev and Ross Seas can be attributed to reduced deep-ocean convection, the subsurface warming in the Indian sector of the Southern Ocean results from lateral advection of warmer waters.

Due to changes in orbital configuration, the seasonality is smaller in the southern hemisphere (Fig. S1), which affects oceanic and atmospheric circulation and heat transport at high southern latitudes. Between 60 °S and 10 °S, the meridional air temperature difference at 850 mb is 1 °C lower at the LIG than for PI (Fig. S10a). This induces a 2.5 ° equatorward shift and 7% weakening of the southern hemispheric westerlies (Fig. S10b). This wind change amplifies the temperature response as it reduces the northward Ekman transport between 50 °S and 60 °S by 19% annually and by 38% in DJF due to even weaker westerlies in austral summer (Fig. S10c). This contributes to the increase in annual mean southward oceanic heat transport south of 50 °S by up to 0.11 PW, with the strongest increase taking place between December and April (Fig. S4c,d). At the same time, polar easterly coastal winds are strengthened by 10% (Fig. 5d-f). As a result, the transition between easterly and westerly winds shifts equatorward by ~ 2 ° in most regions, except in the Pacific sector, where the boundary remains mostly unchanged.

Fig. 5: Subsurface ocean zonal currents and surface zonal winds over the Southern Ocean.
figure 5

Ocean zonal currents averaged across 385-601 m depth at (a) PI, (b) LIG and (c) LIG-PI, and surface zonal winds at (d) PI, (e) LIG and (f) LIG-PI. Colour shading represents the intensity of the zonal velocity components, while the overlain arrows are the full current/wind vectors which include the meridional components. Contour lines mark the boundaries at which the zonal component are zero at (purple) PI and (green) LIG.

Due to the changes in surface winds and the reduced oceanic meridional density gradient, the Antarctic Circumpolar Current weakens from 168 Sv to 116 Sv in the LIG simulation (Fig. 5c, Text S3), consistent with a recent reconstruction estimating a ~ 20% weaker ACC during the early part of the LIG and until 124 ka28. As the ACC acts as a barrier between the mid and high-latitude waters, the ACC weakening might contribute to the increase in southward oceanic heat transport south of 50 °S.

On the other hand, the stronger easterlies coupled with the reduction in density close to shore strengthen the Antarctic Slope Current by 28% (Fig. 5c, Text S3). As there are large horizontal pressure gradients in the frontal zone over the continental slope, thermohaline effects can intensify the wind-driven flow in the same nominal direction29. A freshening and warming of coastal waters in the upper 1000 m, as seen in our LIG experiment relative to PI (Fig. 2), increases the sea surface height near the coast (Fig. S11) and thus further enhances the horizontal pressure gradient, which ultimately strengthens the westward current30.

Under PI conditions, cold waters sink in the Weddell Sea sector and contribute to deep-ocean convection. Some of these waters are advected towards the Indian sector of the Southern Ocean by the ACC (Fig. 5a, S3a, c). The combination of reduced convection northeast off the Weddell Sea, and a weakened ACC thus decreases the advection of cold waters to the Indian sector. In addition, the westward Antarctic Slope Current transports water from the Ross Sea sector towards the Indian sector. Since there is warming near the Ross Sea region at the LIG relative to PI, this warming signal is carried towards the Indian basin, contributing to the positive temperature anomalies.

Currents in the top ~ 1000 m in the Bellinghausen and Amundsen Seas are weak in our PI simulation, which is in line with observations31 (Fig. 5a), while the ACC and thus the eastward currents across the Drake Passage are sustained. However, at the LIG the equatorward shift of the westerlies weakens the ACC and the strengthening of the easterlies induce a westward current along the Antarctic peninsula and towards the Ross Sea (Fig. 5b). This change in ocean circulation in the Bellinghausen and Amundsen Seas brings relatively warm waters to the Antarctic shelf.


The AIS remains the largest potential contributor to future global sea-level rise. While the annual mean global SST at the LIG was only 0.5 °C higher than during PI21,24, GMSL was most likely 3-6 m higher than PI, with a substantial contribution from the AIS.

Here we have shown that reduced sea-ice formation and associated higher SSTs in the Southern Ocean during the LIG reduce AABW formation and increase stratification at high southern latitudes. Reduced deep-ocean convection then leads to a subsurface warming in the Ross, Lazarev and Weddell Seas. In addition, changes in ocean currents further affect subsurface temperatures. Namely, the strengthening of the Antarctic Slope Current, which is due to a reduction in surface water density close to shore and stronger easterly winds, leads to a subsurface warming in the Indian sector of the Southern Ocean. Moreover, due to the ACC weakening, the eastward advection of cold waters from the Weddell Sea in PI is reduced at the LIG, further warming the Indian sector. Lastly, a westward current along the Antarctic peninsula and towards the Ross Sea is induced in the LIG, which transports relatively warm waters from lower latitudes to the Antarctic shelf.

It should however be noted that instead of being formed by dense water export from the Antarctic continental shelf, AABW is formed via deep ocean convection in the ACCESS-ESM1.5, as is the case for 77% of the CMIP6 models32. In addition, as is the case for most if not all CMIP6 models, the ACCESS-ESM1.5 relies on the Gent-McWilliams (GM) parameterisation to represent the transport associated with mesoscale eddies33. Despite this, the ACCESS-ESM1.5 provides a reasonable representation of the state of the Southern Ocean which includes a reasonable representation of the temperature and density of the Antarctic Shelf Bottom Water34 and of the Southern Ocean sea-ice cover20. While it would be preferable to repeat a similar study with an eddy-rich model, no such simulations currently exist for the LIG.

The simulated ACCESS-ESM1.5 PI Antarctic sea-ice area is in agreement with observations in austral summer, while the simulated austral winter sea-ice area is slightly underestimated. The simulated LIG sea-ice area is well within the spread of other CMIP6-PMIP4 models35. While the simulated Antarctic sea-ice anomalies in the ACCESS-ESM1.5 are the largest of the CMIP6-PMIP4 models, with a ~ 50% smaller LIG winter Antarctic sea-ice area compared to present-day observations (Fig. S5), it is in agreement with a previous study36, which suggested that a 65% reduction in winter Antarctic sea-ice area was consistent with the ice core evidence at 128 ka. This reduced sea-ice cover is associated with higher high-latitude surface air temperatures (Fig. S2a). Lower sea-ice concentration leads to warmer air above as the albedo decreases, which explains the relatively strong warming of surface air between 60 °S and 80 °S in the lig127k ACCESS-ESM1.5 experiment. While this high latitude surface air temperature anomaly is ~ 2 °C higher than the CMIP6-PMIP4 multi-model mean, it is in agreement with the other models that simulate lower Antarctic sea-ice cover at the LIG35, and this warming is still lower than the one estimated from ice-core records (Text S1)35. In agreement with more than 70% of the PMIP4 lig127k simulations, the maximum surface warming is simulated to the east of the Weddell Sea, extending into the Indian sector of the Southern Ocean35. This warming could be due to a decrease in winter sea-ice in most models. This region might be particularly sensitive to warming as it corresponds to the northernmost position of the sea-ice edge today 37.

As the annual mean insolation at high latitudes increases with obliquity 38, numerical simulations have shown that higher obliquity weakens the meridional SST gradient in the Southern Ocean39,40 and leads to a lower sea-ice cover. With obliquity at lig127k being higher than today (24.04° vs 23.46°), our simulated warming in the Southern Ocean and reduced sea-ice cover are in line with these findings. Furthermore, our simulated equatorward shift and weakening of the southern hemispheric westerlies under low precession at the LIG is consistent with the study of Persch et al.41, who found that high precession leads to a strengthening and poleward shift of the westerlies.

While our lig127k simulation displays freshening and increased stratification in the Southern Ocean mainly due to increased sea-ice melting, potential freshwater flux from ice shelf and iceberg melt or enhanced basal melt from Antarctica are not included here. Previous studies have shown that enhanced freshwater flux in the Southern Ocean, such as one that would result from a partial disintegration of the AIS, would cool the surface ocean by 1 to 2 °C42,43,44 and lead to a sea-ice advance45,46, thus reducing the Southern Ocean surface warming. However, the meltwater input would also enhance stratification and reduce AABW formation, thus increasing the subsurface warming42,44,47.

AIS modelling studies have suggested that a ~ 3 °C subsurface warming compared to PI on the Antarctic shelf is required to trigger a partial collapse of the WAIS and of some basins of the EAIS during warm interglacials16,17,18. However, some ice sheet modeling work has shown that a sustained retreat of the WAIS can occur over the coming centuries even with present-day melt rates48,49. Our LIG simulation exhibits a subsurface warming on the Antarctic shelf, averaging +0.8 °C at 385-601 m depth south of 60 °S. Across the same depths, the Atlantic sector has the highest warming (+1.3 °C mean, +2.4 °C maximum), followed by the Indian sector (+1.2 °C mean, +2.2 °C maximum), indicating that the LIG subsurface warming on the Antarctic shelf could have triggered substantial ice-mass loss from the AIS. AABW formation is responsible for drawing cold, salty, dense waters into the deep and abyssal ocean during winter sea-ice formation, and is sensitive to changes in upper ocean density50,51. Weaker AABW formation can thus lead to a subsurface warming around Antarctica, potentially giving rise to increased ice-flux from the AIS, which would further weaken AABW formation, thus creating a positive feedback42,52,53.

Such a subsurface warming is simulated under full LIG conditions, and is associated with a weaker AABW but stronger North Atlantic Deep Water formation compared to PI. This complements previous studies, which have suggested that the rapid AIS disintegration during the penultimate deglaciation and early part of the LIG was a result of the long Atlantic Meridional Overturning Circulation shutdown that occurred during Heinrich stadial 1154,55. While these studies suggest that a loss of AIS mass beyond its PI level could have been the result of the transient deglacial climate response, here we show that increased ice-flux from the AIS could also result from increases in surface air, sea surface and subsurface temperature that arise from equilibrium interglacial conditions. This is consistent with Southern Ocean sediment records suggesting decreased oxygenation of deep waters at 127 ka, implying a reduction of AABW formation during peak LIG conditions14. There is also evidence of ice-mass loss from the Wilkes Subglacial Basin during the maximum high southern latitudes warmth at the onset of the LIG, though it might not have manifested from LIG equilibrium conditions due to its early timing16.

While future warming is primarily driven by increased CO2, the climatic differences between PI and LIG as simulated here are due to variations in orbital configuration. The reduction of Antarctic sea-ice and increased heat transport into the Southern Ocean simulated in this study might be specific to the radiative forcing at the LIG, but this serves as an example that considerable subsurface warming on the Antarctic shelf can be achieved by a decline in sea-ice and the associated reduction in deep-ocean convection.

A subsurface ocean warming has contributed to the melting and thinning of West Antarctic ice shelves in recent decades56,57. While Southern Ocean sea-ice extent has increased slowly at the end of the 20th century and until 2016, it has abruptly decreased since then58, hinting at an accelerated warming of the Southern Ocean. In addition, modern observations indicate the surface Southern Ocean has freshened since the 1950s due to increased melting of AIS, thus weakening deep ocean convection and AABW formation50. AABW has warmed over recent decades, leading to higher temperatures in the deep Southern Ocean and global abyssal waters59,60. A Southern Ocean sea-ice decline over the coming century could further increase ocean stratification around Antarctica, thus accelerating AIS mass loss and global sea-level rise.


The results presented in this study are from the Australian Community Climate and Earth System Simulator (ACCESS-ESM1.5)20, which is a fully coupled atmosphere-land-ocean-sea-ice-carbon-cycle model.

The ACCESS-ESM1.5 includes the UK Met Office Unified Model version 7.3 (UM7.3)61,62 as its atmospheric component with a resolution of 1.875° longitude by 1.25° latitude and 38 vertical levels. The land surface model, Community Atmosphere Biosphere Land Exchange model (CABLE2.4)63, is coupled to UM7.3. It is configured with 13 plant functional types (10 vegetated and 3 non-vegetated types), and their spatial distribution is fixed such that the vegetation is static, but the leaf area index is interactive in the model.

The ocean component of the ACCESS-ESM1.5 is the NOAA/GFDL Modular Ocean Model version 5 (MOM5)64, which has a nominal resolution of 1° with 50 vertical grid levels, and has higher resolution near the equator (0.33°) and in the Southern Ocean ( ~ 0.4° at 70 °S). It relies on a flow-dependent GM parameterisation33 to represent the transport associated with mesoscale eddies as described in Kiss et al.65. The ocean configuration is similar to the ACCESS1.0 and ACCESS1.366. The LANL CICE4.1 model67 simulates sea-ice with five thickness classes, and has the same horizontal grid as the ocean model. Ocean biogeochemistry is simulated by the Whole Ocean Model of Biogeochemistry And Trophic-dynamics (WOMBAT)68,69. The OASIS-MCT70 couples the ocean and sea ice to the atmosphere. A brief description of the model’s performances under present-day conditions is included in Text S1.

The initial conditions of the lig127k experiment are based on a PI simulation (1850 CE; piControl)20, which follows the CMIP6 protocol71. Orbital parameters are set at the year 1850 and the greenhouse gas concentrations are the following: CO2 at 284.3 ppm, CH4 at 808.2 ppb, and N2O at 273.0 ppb. The piControl run also serves as a reference in this study.

The lig127k experiment follows the CMIP6-PMIP4 protocol with appropriate 127 ka orbital parameters and greenhouse gas concentrations (i.e., CO2 at 275 ppm, CH4 at 685 ppb and N2O at 255 ppb)19,25. Continental ice sheets and vegetation cover are constant and are the same as prescribed in piControl. The lig127k experiment was integrated for 650 years, and 1000 years for piControl. Results from their last 100 years are used here. The quasi-equilibrium state and drifts of lig127k are discussed in Text S4. A paleo-calendar adjustment was performed on all the seasonal data presented in this study (Figs. 3b, c, 4, S5, and S10b-c) using the program described in Bartlein and Shafer72, except for data associated with energy flux (Fig. S1, S4, and S6).