Future changes in Antarctic coastal polynyas and bottom water formation simulated by a high-resolution coupled model

Antarctic coastal polynyas produce Dense Shelf Water, a precursor to Antarctic Bottom Waters that supply the global abyssal circulation. Future projections of Dense Shelf Water formation are hindered by unresolved small-scale atmosphere-sea ice-ocean interactions in polynyas. Here, we investigate the future evolution of Antarctic coastal polynyas using a high-resolution ocean-ice-atmosphere model. We ﬁ nd that wintertime sea ice production rates remain active even under elevated atmospheric CO 2 concentrations. Antarctic winter sea ice production rates are sensitive to atmospheric CO 2 concentrations: doubling CO 2 (734 ppm) decreases sea ice production by only 6 – 8%, versus 10 – 30% under CO 2 quadrupling (1468 ppm). While considerable uncertainty remains in future ice-shelf basal melting, which is not accounted for in this study, doubling or quadrupling CO 2 substantially freshens Dense Shelf Water due to increased precipitation. Consequently, doubling CO 2 weakens Dense Shelf Water formation by ~ 75%, while CO 2 quadrupling shuts down Dense Shelf Water formation.

A ntarctic coastal polynyas are areas of open water and thin, newly formed sea ice surrounded by coastline or ice shelves and consolidated thick sea ice 1 .They form where sea ice motion is divergent, typically as a consequence of strong offshore winds 2,3 .Although coastal polynyas occupy a relatively small proportion of the Antarctic sea ice area, they play an important role in regional climate by transporting heat from the ocean to the atmosphere, which affects mesoscale atmospheric motions 4,5 .More importantly, sea surface cooling and sea ice formation with brine rejection lead to Dense Shelf Water (DSW) formation 2,[6][7][8] .DSW can lead to the formation of Antarctic Bottom Water (AABW) when it descends the continental slope and mixes with ambient water masses [7][8][9][10] .AABW production is a key element of the Earth's climate system on centennial to millennial timescales 9,[11][12][13] and serves as an important sink for atmospheric CO 2 and heat [14][15][16][17][18] .Observational studies suggest that AABW volume may have been reduced over the recent decades because of Antarctic coastal freshening and warming [19][20][21] .
Predicting the DSW and AABW responses to future climate change is a challenging task.A recent study suggests that retreating sea ice is likely to substantially weaken open-ocean convection in the Weddell Sea 22 .Predicting changes in Antarctic coastal convection is more challenging because coupled climate models with a relatively high grid resolution 23 , ice-shelf basal melting 24 , and landfast sea ice processes 25 are needed to resolve coastal polynyas and dense water overflows.Most current climate models participating in Coupled Model Intercomparison Project Phase 6 (CMIP6, ref. 26 ) have difficulty in simulating Antarctic coastal polynyas due to their relatively coarse horizontal resolutions 27,28 .In these coarse-resolution models, AABW is usually formed in open-ocean polynyas, particularly in the Weddell Sea, which contrasts with the multi-decadal re-occurrence of such polynyas in nature 27,28 .Recent studies using highresolution ocean models forced by atmospheric boundary conditions [29][30][31] found that the AABW formation declines by ~40% within 50 years in response to meltwater forcings in the highest-emission scenario (RCP8.5 or SSP5-8.5;ref. 32 ).
In a high-resolution atmosphere-ocean coupled climate model, the zonal-mean meridional overturning circulation in the Southern Ocean is reduced by around 50% in response to atmospheric CO 2 doubling 23 (relative to the 1990s CO 2 mixing ratio).While the heat budget analysis over the entire Southern Ocean indicates that the circulation weakening is mainly driven by reduced heat loss to the atmosphere 23 , regional-scale surface heat flux responses remain elusive.Because surface heat fluxes are strongly constrained by sea ice cover in polar oceans 33 , reliable projections of Antarctic sea ice are needed to better predict the responses of DSW and AABW to future climate change.Using a fully coupled high-resolution climate model, this study shows that the wintertime Antarctic coastal sea ice production rates decrease by only 6-8% in doubling CO 2 conditions (corresponding to a 2.5 °C increase of the global average surface air temperature relative to 16 °C; ref. 34 ), and by 10-30% in quadrupling CO 2 conditions (5.1 °C).Despite the continued sea ice production, the freshening of DSW in response to doubling weakens DSW formation by ~75%, whereas quadrupling CO 2 leads to a shutdown of DSW, and thus AABW formation.

Results
Antarctic sea ice production.In this section, we diagnose sea ice production in Antarctic coastal polynyas in the control simulation, which is forced with a CO 2 concentration level of 367 ppm (see Methods for details).Then, we examine how it changes with increasing CO 2 concentrations.The control experiment simulates localized sea ice production, with rates of up to 10-15 m year -1 along the Antarctic coasts, particularly in the Ross, Weddell, Prydz and Adélie Seas (Fig. 1a).6][37][38][39][40][41][42] , see Methods for details) rate.Figure 1b shows that our model-simulated WMT rates from May to October are mostly positive, implying that the surface watermass loses buoyancy and becomes denser via diabatic processes.In the regions of relatively high sea ice production (see Fig. 1a), the WMT rate is also relatively high.Consistent with a previous study 35 , the WMT rates due to sea ice production explain almost 80-90% of total WMT rates.
This study focuses on the four major coastal regions where the DSW forms: Prydz Bay (known as Cape Darnley in observations), the Adélie Coast, and the Ross and Weddell Seas 6,43,44 .The control simulation shows that the high annual sea ice production rates in these regions are concentrated within ~20 km of the Antarctic coasts (Fig. 1c-f).Our simulations do not include landfast sea ice, which can facilitate sea ice formation in the coastal polynyas [45][46][47] .It is important to note that the technical implementation of landfast sea ice is currently challenging due to the lack of observational data, and as of now, no CMIP6 model incorporates landfast sea ice.
In Prydz Bay, where Cape Darnley Bottom Water forms, the grounded iceberg tongue (icecape) is a key component of DSW formation 48,49 , as it blocks the westward advection of sea ice 50 and thereby promotes sea ice formation.While the discrepancy between the observed and model-simulated sea ice production in Prydz Bay is not large (151 ± 17 km 3 in CESM1 vs. 194 ± 19 km 3 in observations 48 ), the simulated DSW formation rate is smaller than observations (Table 1).Due to the uncertainty in observation-based DSW formation rates, a direct comparison between our model simulations and observations is imperfect.In the Adélie Sea, where the Adélie Land Bottom Water forms, the Mertz Glacier Tongue (MGT) plays a major role in coastal sea ice production 51 by blocking sea ice advection from the east.Simulated sea ice production in CESM1, in which the MGT is not implemented, is less than half of the observed value in the Adélie Sea, although the CESM1-simulated DSW formation rate (0.13 ± 0.08 Sv) overlaps the range of observational estimates (0.1 ~0.5 Sv, Table 1).
In the Weddell Sea polynya (Fig. 1f), simulated sea ice production is four times larger than observations (Table 1).Due to the challenges in estimating sea ice thickness, observed sea ice productions over the Weddell Sea polynya have large uncertainties, leading to discrepancies among several observations 48,52,53 .Although there are discrepancies in regional-scale sea ice productions between CESM1 and observations, the discrepancy reduces in the case when sea ice productions over the four major coastal polynyas are added together.The total sea ice volume productions of the model and observations are about ~580 ± 93 km 3 and ~664 ± 76 km 3 , respectively (Table 1).
In response to CO 2 doubling and quadrupling, the seasonal sea ice areas of all four coastal polynyas show delayed freeze-up in austral autumn, with changes that are most pronounced between March and May (Fig. 2a-d).In March, sea ice area starts to increase in the control simulation (black line in Fig. 2a-d) but the coasts become almost ice-free under CO 2 quadrupling (red line in Fig. 2a-d).The delayed autumn freeze-up is associated with increased upper ocean heat content (~250 m) in summer and autumn (Supplementary Fig. 1), which is consistent with the mechanism underlying delayed freeze-up in the Arctic Ocean 54,55 .In all four polynyas, the upper ocean warming in response to CO 2 doubling and quadrupling maximizes in summer, up to ~3.0 °C relative to the last 30 years of the control simulation, indicative of increased absorption of shortwave radiation associated with decreased summer sea ice cover.In CO 2 quadrupling, 0.5-1.0°C of positive upper ocean temperature anomalies persist into April and May.
Consistent with the upper ocean temperature anomalies (Supplementary Fig. 1), the freeze-up is delayed until around April in CO 2 doubling and until May in CO 2 quadrupling (Fig. 2a-d).However, sea ice area in both the CO 2 doubling and CO 2 quadrupling simulations almost reaches the level of the control simulation by mid-winter due to the seasonal northward growth of sea ice.In the Prydz Bay and Weddell Sea (Fig. 2a, d), the delayed autumn freeze-up is followed by a rapid increase in sea ice area, which recovers to match the sea ice area of the control simulation in CO 2 doubling, but the level of recovery is limited to 85-90% of the control simulation in CO 2 quadrupling, primarily associated with sea surface warming along the northern margins of the Antarctic seas.Consistent with the sea ice area changes, the sea ice production starts about half a month delayed for the CO 2 doubling and a month delayed for the CO 2 quadrupling (Fig. 2e-h).In winter (June-September), during which sea ice production rates are highest, the CO 2 doubling leads to a moderate decline in sea ice production, by only ~6-8%.However, the CO 2 quadrupling leads to 10-30% decrease in sea ice production (Fig. 2i-l), highlighting the sensitivity of winter sea ice production to greenhouse gas emission scenarios.Surface heat fluxes.Changes in sea ice production rates are closely associated with changes in net surface heat flux.In response to CO 2 doubling, the wintertime surface air temperatures over the Prydz, Adélie, Ross, and Weddell Seas increase by 1.9 to 5.2 °C (Table 2), which can suppress surface heat loss to the atmosphere 23 .In response to increased surface air temperatures, wintertime (June-September) sea ice production rates in coastal polynyas are decreased in CO 2 doubling by 6-8% (Fig. 2i-l).Because the formation of coastal polynyas strongly depends on katabatic and synoptic winds 44 , the strengthening of surface winds in response to global warming can enhance sea ice production.Although our model simulates a deepening of Antarctic low-pressure around the coasts in response to CO 2 doubling and quadrupling (Supplementary Fig. 2) (consistent with Southern Annular Mode strengthening under global warming [56][57][58] ), we find that these pressure changes are not accompanied by substantial changes in surface wind strength over the Antarctic coasts (Supplementary Fig. 2).Another possible explanation for the modest reduction of sea ice formation under a CO 2 doubling is that reduced sea ice concentration (SIC) over the Antarctic coasts can increase heat loss to the atmosphere, partially offsetting the increase in surface air temperatures in warmer climates.Indeed, Fig. 3 shows that daily net surface heat flux is negatively correlated with SIC over the four polynyas.Here, net surface heat flux (NetFlux " ) includes all the heat and radiative fluxes over both sea ice and open water: where, SHF " and LHF " denote upward sensible and latent heat fluxes, respectively, LW " and LW # represent upward and downward longwave radiative fluxes, respectively, and SW " and SW # represent upward and downward shortwave radiative fluxes, respectively.The wintertime average SIC is very high because open waters in coastal polynyas freeze-up rapidly 50 : a slight reduction of SIC, down to 80-90%, is followed by significant heat loss to the atmosphere in observations 59 .Consistent with observations, our model simulation shows that a decrease in SIC from 100% to 80% can effectively increase upward net surface heat flux to 40-90 W m -2 in all four polynya regions (Fig. 3a-d).
In warmer climates, grid cells with winter SIC lower than 60% occupy a larger proportion of the coastal areas (Fig. 3e-l), which may be expected to further increase the area-averaged surface temperatures.In response to CO 2 doubling, the sensitivity of the  ocean-to-atmosphere heat flux associated with decreasing SIC slightly weakens.In all four polynyas, the regression slopes in CO 2 doubling are about 8-16% smaller than the Control simulation (Fig. 3e-h).The sensitivity of the ocean-toatmosphere heat flux to SIC further weakens in CO 2 quadrupling, being about 20-30% smaller than the Control simulation (Fig. 3i-l).These sensitivity changes might partly explain the differing responses of the wintertime sea ice production rates to different CO 2 increases, with little change for a medium-emission scenario and 10-30% reduction in a high-emission scenario (Fig. 2i-l).Figure 4 indicates that the increased surface sensible heat flux (SHF " ) in response to decreasing SIC explains ~60-70% of the net surface heat flux sensitivity to SIC, i.e., the regression slopes in Fig. 4 are ~60-70% of those in Fig. 3. To aid conceptual understanding of the changing regression slopes in Fig. 4, we propose a simple 1D model for surface sensible heat flux over coastal polynyas.Assuming that sensible heat loss over the sea ice-covered fraction of the water is negligible (as the sea ice skin temperature is very close to the air temperature), the wintertime surface sensible heat flux can be re-written as a function of SIC utilizing a bulk formula: where U 10 , T S , and T a are wind speed at 10 m height, surface temperature, and surface air temperature, respectively.T f is freezing temperature of seawater, which is set to −1.8 °C.ρ a is surface air density (1.275 kgm À3 ) and c p is the specific heat capacity of air (1005 J kg À1 K À1 ).C D is turbulent transfer coefficient over a partially sea ice-covered area, which is around 0.0015 (ref. 60).Derivation of the above Eq.(2A) by SIC (%) leads to: In the control simulation, for example, the wintertime T a areaaveraged over the Ross Sea is around −27 °C and T a increases to around −20 °C in the CO 2 quadrupling.In terms of Eq. (2B), increasing T a from −27 °C to −20 °C decreases the sensitivity of SHF " to SIC by ~28%, which is comparable to our modelsimulated changes in the sensitivity (black line in Fig. 4k).In general, Eq. (2B) explains the sensitivity of net surface heat flux to SIC reasonably well: Similar to the black solid line (regression line) in Fig. 4, the slope of the pink dotted line also decreases with increased CO 2 simulation.
One of the limitations of this 1D linear model is that the sensitivities of net surface heat flux and SHF to SIC are not linear but slightly quadratic.Because the coastal polynyas are often driven by anomalously strong wind-driven ice flux divergence, there might be a correlation between U 10 and SIC in Eq. (2A).While adding this negative correlation changes the linear relationship of Eq. (2A) into a quadratic relationship, the winddriven formation of coastal polynyas occurs sporadically.At each grid cell, there is no statistically significant relationship between daily U 10 and SIC in our model simulation, although the areaaveraged U 10 and SIC show a clear negative correlation (Supplementary Fig. 3).Another limitation of our 1D model is that sea ice thickness variation is not accounted for.While the formation of coastal polynyas is accompanied by ice thinning, net surface heat fluxes are far more sensitive to sea ice concentration than sea ice thickness (Supplementary Fig. 4).This is consistent with our conceptual 1D model that the majority of net surface heat flux can be estimated as a function of sea ice concentration.
Water-mass transformation and offshore transport.The active wintertime sea ice production in the coastal polynyas in warmer climates may be expected to provide favorable conditions for the maintenance of DSW production, despite the continuing sea surface freshening.The formation of dense water masses in coastal polynyas can be assessed by the WMT framework (see Methods for details), which relates surface density distributions and buoyancy fluxes to rates at which waters change their density at the interface between ocean and atmosphere, and between the ocean and sea ice.
Figure 5 shows the wintertime WMT rates as a function of potential density referenced to 1000 dB.In the control simulation, the formation of dense waters in Prydz Bay (consisting of Cape Darnley and Mackenzie Bay) amounts to 1.2 Sv (as indicated by the peak of the solid black line in Fig. 5a).In the Adélie Sea, the formation rate is ~0.5 Sv (Fig. 5b), somewhat smaller than those of other polynyas and but in agreement with an observation-Fig.5 Water-mass transformation rates in Control and warmer climate simulations.Thirty-year climatological winter season (June-September) watermass transformation rates driven by total surface flux in the a Prydz, b Adélie, c Ross, and d Weddell polynyas in the (black) Control, (orange) 2×CO 2 , and (red) 4×CO 2 simulations.The water-mass transformation rates are calculated over the polynya regions from the Control simulation (green lines in Fig. 1c-f).
based estimate 61 .Sea ice production is larger in the Ross Sea than in the Weddell Sea, which is generally consistent with observations 53 .However, the WMT rate associated with high surface density is higher in the Weddell Sea than in the Ross Sea (compare Fig. 5c, d).This implies that the WMT rate is not necessarily proportional to the sea ice production rate, but rather surface density distribution also plays a modulating role.Consistent with the shrinking areas of high sea ice production in warmer climates, the maximum WMT rates decrease by around 0.2-0.6Sv under CO 2 doubling and by 0.4-2.0Sv under CO 2 quadrupling in the four coastal polynyas.
The shift of dense water formation to lighter density classes as CO 2 increases implies that it becomes difficult for DSW to transform into AABW (ref. 31). Figure 6 quantifies how much water-mass cascades down the continental slope and contributes to AABW in the control and increased CO 2 simulations.In the control simulation, shelf waters denser than σ 1 ¼ 32:38 (the density at which the transport is maximum) flow offshore across the 1000-m isobath in the amount of 6.7 ± 2.2 Sv (black line in Fig. 6b), which agrees well with observational estimates, 8.1 ± 2.6 Sv of AABW precursor waters 11,61 .Note that our simulated DSW outflow is somewhat smaller than that of other modeling studies using a high-resolution ocean-sea ice model 7,30 , which simulates ~10.6 Sv of DSW outflow.However, ocean-sea ice models forced by atmospheric surface boundary conditions are known to overestimate coastal ice production and DSW outflows, because the lack of atmosphere-ocean coupling can overestimate sensible and latent heat fluxes from ocean to the atmosphere 4,38,62 .
In the CO 2 doubling and quadrupling simulations, net offshore transports in higher density classes (denser than σ 1 ¼ 32:38) rapidly decrease.They reduce to ~1.8 Sv under CO 2 doubling and ~0.03 Sv in CO 2 quadrupling (orange and red lines in Fig. 6b), suggesting CO 2 doubling decreases the export of newly formed DSW by around 75%, whereas CO 2 quadrupling leads to a shutdown of DSW export, and thus of AABW formation.These results are generally consistent with the changes in surface water-mass transformation rate in warmer climates (Fig. 5), showing that DSW denser than σ 1 ¼ 32:38 does not form in the major coastal polynyas in the CO 2 quadrupling simulation.
Moreover, there is a clear shift of dense water formation and net offshore transport to lighter density classes as CO 2 increases, primarily because of decreasing sea surface salinity in coastal polynyas (Supplementary Fig. 5).Upper ocean freshening has multiple potential contributors, such as increased precipitation, increased freshwater discharge from the Antarctic continent, and melting ice shelves.In our CESM1 simulation, the freshwater discharge from Antarctica is calculated through "snow-capping" (see Method section), in which excess snow beyond 1.0 meter is discharged to the nearest ocean.In the CESM1 simulation, precipitation (mainly snowfall) increases by around 20-60% over the Antarctic continent and the Antarctic oceans in warmer climates (Fig. 7), which is consistent with previous modeling studies 63,64 .In the CMIP5 and CMIP6 ensembles, the highest-emission scenarios predict that Antarctic precipitation may increase approximately by 44-66% (ref. 64), which is comparable to our high-resolution model response to CO 2 quadrupling.Consistent with our high-resolution model simulation (Fig. 7d, e), the increased precipitations in the CMIP5 and CMIP6 are found to be more prominent over the coastal margins than in the Antarctic continental interiors 64 .
The increased snowfall is a key cause of sea surface freshening over the coastal polynyas, while meltwater from the adjacent Antarctic ice shelves is represented in an overly simplistic way via increased runoff from the Antarctic continent.The increased precipitation in the Weddell Sea coasts corresponds to 10 and 20 Gt year -1 of freshwater input in CO 2 doubling and quadrupling, respectively.These are far smaller than over 100 Gt year −1 of freshwater input over the southern Weddell Sea that has been estimated to result from increased ice-shelf melting in warmer climates 31 .Previous studies 31,37,39,65 have shown that future ice-shelf melting decreases sea surface salinity by around 0.1-0.3PSU, although this estimate remains subject to significant uncertainties 31 .When considering the contribution of ice-shelf melting alongside precipitation changes, sea surface freshening is expected to intensify further.

Conclusions
Using a fully coupled high-resolution climate model that can resolve coastal polynyas, this study investigates how Antarctic coastal sea ice production and DSW formation respond to increasing greenhouse gas concentrations.Our model simulations suggest that delayed autumn freeze-up associated with increased upper ocean heat content substantially decrease the annual-mean sea ice area and sea ice production rates over Antarctic polynyas.However, the wintertime sea ice production rates decrease by only 6-8% in response to CO 2 doubling and 10-30% in response to CO 2 quadrupling.Modest sea ice reduction in warmer climates is mainly because the reduction in upward heat flux due to surface air warming is offset by decreasing sea ice cover, which effectively increases the mean surface temperature of the ocean/sea ice area.While the wintertime sea ice production in the coastal polynyas declines modestly in response to CO 2 doubling, the resulting density of newly formed DSW lightens due to surface freshening associated with increased precipitation rates.In response to CO 2 quadrupling, the wintertime sea ice production rates decrease further because surface air warming decreases the rate at which heat is lost to the atmosphere per unit of open-water area.Moreover, increased precipitation decreases surface salinity by ~0.7 PSU, further increasing the possibility of a shutdown of DSW formation.
While our model simulation reproduces the coastal sea ice production in the major polynyas well, the sensitivity of the sea ice production may depend on the sea ice model and parameter choices.Therefore, further studies with other high-resolution coupled models will be needed to establish a measure of uncertainty.We also acknowledge the limitations of our model configuration in this study, particularly the lack of ice-shelf melting that can cause further sea surface freshening around the Antarctic coasts.However, the impact of future ice-shelf melting has considerable uncertainty, not only in the amount of melting but also in the resulting sea surface freshening.Recent modeling studies 29,31,66 suggest that the accelerated iceshelf melting can potentially lead to the shutdown of AABW when additional freshwater forcing is prescribed to the coastal sea surface or when ice-shelf melting is calculated dynamically.Although our CO 2 quadrupling simulation predicts substantial DSW freshening and a near-complete shutdown of AABW formation, we note that the low bias in AABW production in CESM1 (ref. 67) may make the model artificially susceptible to a shutdown of AABW formation.Despite this limitation, our study provides a more realistic and comprehensive perspective on coastal sea ice production and DSW formation changes using a high-resolution coupled model, which simulates air-sea interactions in coastal polynyas.

Methods
High-resolution CESM coupled model simulations.The Community Earth System Model version 1.2.2 (CESM1.2.2) consists of coupled atmosphere (CAM5; Community Atmosphere Model version 5; ref. 68 ), ocean (POP2; Parallel Ocean Program version 2; ref. 69 ), land (CLM4; Community Land Model version 4; ref. 70 ), and sea ice (CICE4; Community Ice CodE version 4; ref. 71 ) components.We used the highest resolution "ne120_t12", in which CAM5 and CLM4 use a 0.25°(~25 km) and POP2 and CICE4 use a 0.1°(2.5 ~11 km) horizontal resolution.CAM5 has 30 vertical levels with model top at ~2 hPa and POP2 has 62 levels with a maximum depth of 6000 m.Our high-resolution model does not directly interact with Antarctic ice sheets/shelves, discharge from Antarctica is treated with snow-capping parameterization.The snow-capping parametrization is activated whenever the snow depth in a grid cell exceeds a prescribed maximum value, set to 1 m in this study.Upon activation, the parameterization redistributes the excessive snow (ice flux) the ocean via a nearest-neighbor algorithm, spreading it using a Gaussian distribution over a radius of 300 km.If the ocean is warm enough, this ice flux is melted in the ocean immediately.These simplifications impose limitations on the potential evolution of excessive snowfall into glaciers.No overflow parameterization is employed in this study.We carried out three simulations with different CO 2 concentrations: (i) 367 ppm CO 2 concentration (Control simulation), (ii) CO 2 doubling (2×CO 2 , 734 ppm), and (iii) CO 2 quadrupling (4×CO 2 , 1468 ppm).All other greenhouse gas and aerosol concentrations were the same as in the control simulation.The Control simulation was initialized from a quasi-equilibrated climate state 72 and then integrated for another 140 years.The abrupt CO 2 doubling and quadrupling experiments were branched off from year 71 of the Control experiment and then integrated for 100 years each (see ref. 34 for details).Ref. 30 found a decline in AABW formation in response to meltwater forcings in RCP8.5 scenarios.Similarly, ref. 31 and ref. 29 obtained similar results to ref. 30 but used meltwater forcing with the SSP5-8.5 emission scenario.Both RCP8.5 and SSP5-8.5 scenarios project a CO 2 concentration increase up to 1100 ppm by 2100.The CO 2 quadrupling scenario, in this study, has a slightly higher CO 2 concentration than RCP8.5 or SSP5-8.5.
This study focuses on the Antarctic winter season, from May to October, during which the majority of sea ice production occurs.We examine the last 30 years of each simulation, discarding the first ~70 years of each simulation, to better focus on the equilibrium responses.Unless otherwise stated, all diagnostics used in the main text are drawn from the last 30 years of each simulation.Supplementary Fig. 6 illustrates how the seasonal Antarctic sea ice concentration is simulated in the high-resolution CESM1 with 367 ppm of CO 2 concentration.Our high-resolution CESM1-simulated Antarctic sea ice extent underestimates by ~40% compared to observed estimates; this underestimation appears in most climate models participated in CMIP5 (refs. 73,74).6][77][78][79] ), our high-resolution CESM1-simulated sea ice extent is sensitive to CO 2 increases, indicating annual-mean sea ice extent decreases by ~33% in CO 2 doubling and ~50% in CO 2 quadrupling.Our model simulation reproduces the seasonal variations in surface potential density referenced to 1000 dB, consistent with observations (Supplementary Fig. 7).As for seafloor temperature and salinity, our control simulation shows slightly warm and fresh bias near the key coastal polynyas.However, these biases are far smaller than CMIP6 multi-model averages (Supplementary Fig. 8).Simulated sea ice volume production integrated over the major four coastal polynyas in the control simulation is ~90% of the observed estimates (Table 1).
Surface flux-driven water-mass transformation.6][37][38][39][40][41][42] ).WMT is defined as the volume flux of a water-mass consumed or produced by buoyancy forcing in each density class.Buoyancy forcing primarily occurs due to diapycnal mixing in the ocean interior and surface buoyancy fluxes, which is an exchange of buoyancy between the ocean and atmosphere or between the ocean and ice sheet/ice shelves.Our model configuration does not include the flux exchanging process between the ocean and ice sheet/ice shelves.This study focuses on the transformation rate induced by surface heat and freshwater fluxes between the ocean and atmosphere/sea ice.The transformation across the density surface is diagnosed from surface heat and freshwater buoyancy fluxes: where variables in Eq. ( 3) are defined in Table 3.The WMT is calculated numerically by discretizing potential density referenced to 1000 dB, σ k , in 225 unevenly spaced bins from 1027.22 kg m −3 to 1032.68 kg m −3 .The bin spacing, σ kþ1 À σ k , varies from 0.01 kg m −3 at low densities to 0.002 kg m −3 at high densities.
Offshore transport.To quantify how much water cascades down the continental slope, we calculate a net offshore transport across the 1000 m isobath around the Antarctic continental shelf.Following recent studies 7,30 , the density characteristics of waters exchanged between the continental shelf and the open ocean are analyzed by binning daily averaged volume transports across the 1000 m isobath into potential density classes before taking an annual-mean and integrating time-averaged fluxes from high to low density.
where variables in Eq. ( 4) are defined in Supplementary Table 1.Similar to the WMT rate, the offshore transport is calculated in potential density space reference to 1000 dB, σ k , in 225 unevenly spaced bins from 1027.22 kg m −3 to 1032.68 kg m −3 .The bin spacing, σ kþ1 À σ k , varies from 0.01 kg m −3 at low densities to 0.002 kg m −3 at high densities.
Validation.Ocean temperature and salinity from the World Ocean Atlas 2018 (WOA18; refs. 80,81) over the period 1995-2014 were utilized as a reference.To validate the CESM-simulated sea ice extent and concentration, we used the satellite-observed sea ice extent and concentration provided by the National Snow and Ice Data Center (NSIDC; refs. 82,83).

Fig. 1
Fig. 1 Antarctic sea ice production and surface water-mass transformation.Climatological-mean winter (May-October) a sea ice production (m yr −1 ) and b regional surface water-mass transformation rate (m s −1 ) associated with sea ice formation, at potential density levels greater than 31.2σ 1 (kg m −3 ), simulated by the high-resolution coupled climate model with control simulation.Here, the regional surface water-mass transformation rate is divided by the area of each grid cell.(Bottom; c-f) Same as a, except for close-up views of c Prydz, d Adélie, e Ross, and f Weddell polynyas.Solid green lines bound the region in which sea ice production exceeds 4 m yr -1 .Black lines in each figure represent the 1000 m isobath.Pink outlines in a indicate four major coastal polynyas that this study focuses on.

Fig. 3
Fig. 3 Relationship between sea ice concentration and net surface heat flux.The probability distribution (frequency of occurrence) of the wintertime (June-September) daily net surface heat flux (NetFlux, W m -2 ; ordinate) as a function of daily sea ice concentration (SIC,%; abscissa), calculated for each grid cell over the Prydz, Adélie, Ross, and Weddell Sea polynyas from the (top; a-d) Control, (middle; e-h) 2×CO 2 , and (bottom; i-l) 4×CO 2 simulations.Solid black lines in each panel denote linear regression slopes from the Control, 2×CO 2 , and 4×CO 2 simulations.The regression slopes (s; W m −2 % −1 ) are calculated with grid cells of SIC larger than 15%.Positive heat flux implies heat flux from the ocean to the atmosphere.

Fig. 4
Fig. 4 Relationship between sea ice concentration and surface sensible heat flux.The probability distribution (frequency of occurrence) of the wintertime (June-September) daily surface sensible heat flux (SHF, W m −2 ; ordinate) as a function of daily sea ice concentration (SIC,%; abscissa), calculated for each grid cell over the Prydz, Adélie, Ross, and Weddell Sea polynyas from the (top; a-d) Control, (middle; e-h) 2×CO 2 , and (bottom; i-l) 4×CO 2 simulations.Solid black lines in each panel denote linear regression slopes from the Control, 2×CO 2 , and 4×CO 2 simulations.Dotted pink lines represent the sensitivity of surface sensible heat flux to sea ice concentration, estimated by Eq. (2B).The linear regression slopes (s; W m −2 % −1 ) between SIC and SHF are calculated with grid cells of SIC larger than 15%.Positive heat flux implies heat flux from the ocean to the atmosphere.

Fig. 6
Fig. 6 Changes in net offshore transport.a Bathymetry in the abyss (m) from the Control simulation.Pink line in a represents the 1000 m isobath.Pink vectors denote offshore transport.b Net offshore transport across the 1000 m isobath, cumulatively summed in σ 1 space from high to low density from (black) Control, (orange) 2×CO 2 , (red) 4×CO 2 simulations.The gray vertical line indicates the density threshold of DSW in the Control simulation.Positive and negative values mean offshore and onshore transport, respectively.

Fig. 7
Fig. 7 Antarctic snowfall changes.(Upper; a-c) Climatological snow water equivalent (mm day −1 ) in the winter (June-September) for a Control, b 2×CO 2 , and c 4×CO 2 simulations.(Lower; d, e) Difference of climatological snow water equivalent d between 2×CO 2 and Control, and e between 4×CO 2 and Control simulations.d, e statistically significant values (p < 0.05) calculated from two-sample t test are stippled.

1 S 3 A
Specific heat of seawater (3994) J kg −1 K −1 β Haline coefficient of contraction kg m −3 psu −1 F net Downward surface freshwater flux kg m −2 s −Sea Horizonal ocean surface area of interest m 2

Table 1
3limatological accumulated volume of sea ice production (km3) and DSW formation rate (Sv) during the winter season, along with annual standard deviations, in the Prydz, Adélie, Ross, and Weddell Sea polynyas from observed estimates and CESM1 control simulation.
Fig.2Seasonal cycle of sea ice area and production in Control and increased CO 2 simulations.Climatological monthly-mean (top; a-d) sea ice area (10 3 km 2 ) and (middle; e-h) sea ice production (10 km 3 ) over the Prydz, Adélie, Ross, and Weddell polynyas simulated by (black) Control, (orange) 2×CO 2 , and (red) 4×CO 2 experiments.The abscissa is the time (month).Sea ice production rates are calculated within the coastal polynya boundaries from the Control simulation (solid green lines Fig.1c-f).(bottom; i-l) The areally-averaged winter sea ice productions (accumulated from June to September) for the i Prydz, j Adélie, k Ross, and l Weddell polynyas, simulated by the (black) Control, (orange) 2×CO 2 , and (red) 4×CO 2 experiments.Shown in each box plot are the median (middle solid line), the 25th and 75th percentile (lower and upper hinge), and the 5th and 95th percentiles (whiskers).Changes in median sea ice production rates relative to the Control simulation are given in blue text.

Table 2
Climatological winter season (June-September) surface air temperatures (°C), averaged over the Prydz, Adélie, Ross, and Weddell polynyas (as indicated by the area shown in Fig.1c-f) simulated by the Control, 2×CO 2 , and 4×CO 2 experiments.
The values in parentheses in the 2×CO 2 and 4×CO 2 columns represent the surface temperature changes relative to the Control experiment.