Main

Around Antarctica (poleward ~60° S), ocean below 200 m is warmer than the surface ocean1. This inversion in vertical temperature profile is not present north of ~60° S (see ‘Observations’ in Methods) (Extended Data Fig. 1). Southern high-latitude winds drive divergent surface flows that draw water up from below2,3,4,5,6. Most of the water upwells from a depth of 2–3 km along sloping density layers with little heat input or mixing required7. Upon surfacing, the upwelled waters split into two branches8. One flows equatorward (northward ~60° S) where the upwelled water is cooler than the ambient ocean and atmosphere and is heated by the atmosphere before eventual subduction into the mid-latitude ocean interior, ensuring that the Southern Ocean takes a large portion of the excess heat from greenhouse warming. The other branch is located poleward ~60° S where the upwelled water is warmer than the ambient ocean and air temperatures and loses heat to the atmosphere, affecting sea ice formation. Imprints of the importance of the warm upwelling are highlighted in a heat budget analysis, showing that air–sea heat flux cools the ocean and the warm upwelling heats the shelf and surface waters (Extended Data Fig. 2).

Observed warming of the Southern Ocean shows a rich vertical structure around Antarctica. Pan-Antarctic surface water has cooled and Antarctic sea ice has increased during the 1979–2015 period9,10. In a sharp contrast, the Antarctic shelf ocean waters (waters above the seabed with bathymetry shallower than 1,500 m) in the Amundsen and Bellingshausen Seas have warmed substantially10,11,12,13. Consequently, West Antarctic ice shelves and ice sheet have lost substantial mass between 1992 and 2017, particularly from the Pine Island and Thwaites Glacier catchments of the Amundsen Sea Embayment14,15. The warming of shelf ocean water critically affects the pace of Antarctic ice shelf and ice sheet melt11. Under greenhouse warming, heating from warm upwelling intensifies but varies vastly across models16.

A projected increase in El Niño–Southern Oscillation (ENSO) variability is found to slow mid-latitude Southern Ocean warming17. During El Niño, including central-Pacific El Niño18, tropical convective anomalies induce a Pacific South American pattern of anomalies and an equatorward contraction of the Hadley, Ferrell and Polar cells, generating zonally symmetric high-latitude easterly anomalies17,18. During La Niña, the reverse occurs. Because El Niño warm anomalies and the teleconnections are overall greater than those of La Niña cold anomalies19, over several ENSO cycles, there are anticyclonic circulation anomalies over the Amundsen and Bellingshausen Seas but easterly anomalies over much of the southern high latitudes17. Due to this ENSO rectifying effect, models that project a greater increase in ENSO variability systematically generate a weaker westerly poleward intensification over the southern high latitudes. The associated smaller increase in circum-Antarctic upwelling leads to the slower mid-latitude Southern Ocean warming17. However, how the ENSO response affects Antarctic shelf ocean is unknown.

Here we show that a greater increase in future ENSO variability accelerates Antarctic shelf ocean warming but slows surface warming around sea ice edges. Models that simulate a greater increase in ENSO variability systematically generate an accelerated circum-Antarctic shelf ocean warming but slowed surface temperature increases, with important implications.

We examine 31 climate models that participated in Phase 6 of the Coupled Model Intercomparison Project20 (CMIP6) under historical forcings before 2014 and a high-emission scenario (that is, Shared Socioeconomic Pathways (SSP) 5–8.5) (see ‘CMIP6 outputs’ in Methods and Supplementary Table 1). These models simulate the observed feature of the high-latitude temperature structure, which persists into the twenty-first century (Extended Data Fig. 3). In addition, models simulate reasonably well the dominant variability patterns, their link to ENSO through the Pacific South American pattern21,22,23 and the zonally symmetric meridional anomalies17.

Multimodel mean change, defined as the difference between the twenty-first (2000–2099) and the twentieth (1900–1999) century, features a westerly poleward intensification, an increase in the associated negative wind stress curls around Antarctica, a reduction in Antarctic sea ice concentration and zonally averaged Southern Ocean warming with a maximum centred around 45° S (Extended Data Fig. 4); poleward 60° S, as greenhouse warming proceeds, atmospheric cooling of the warm upwelled water weakens. However, the warm upwelling increases such that even by 2100, the ocean poleward 60° S still loses heat to the atmosphere (Extended Data Fig. 5).

We use Niño3.4 (5° S–5° N, 170° W–120° W) sea surface temperature (SST) index in the ENSO matured season of December, January, February (DJF) to describe ENSO (see ‘ENSO response’ in Methods). Monthly SST anomalies over 1900–2099 are constructed with reference to the climatology over the 1900–1999 period and quadratically detrended. Zonally averaged shelf ocean temperatures are constructed by shifting the latitude16,24 relative to the 1,000 m isobath at each longitude before zonal averaging (see ‘Zonal mean of shelf ocean temperatures’ in Methods). To examine inter-model relationships, changes are scaled by the increase of global mean SST in each model (last column, Supplementary Table 1) to remove impact from climate sensitivity.

A greater increase in ENSO variability is systematically linked to a faster shelf ocean warming but to a slower surface warming around the sea ice edge (Fig. 1a) that in turn slows sea ice melt (Fig. 1b). The linkage is seen in an inter-model correlation of ENSO changes with sea ice changes averaged over the 60° S–70° S circumpolar band (r = 0.70) (Fig. 1c) and with averaged shelf ocean warming over 400–1,500 m near the shelf break (r = 0.63) (area confined by black thick line, Fig. 1c), or horizontally (Fig. 1d). The averaged shelf ocean warming reflects warming at various depth ranges including the upper 200–400 m (Extended Data Fig. 6). A large inter-model spread exists in the projected changes, but the inter-model correlation amplitude suggests that the inter-model differences in ENSO changes account for approximately 40% (r2) of the spread in the shelf ocean warming, and approximately 45% of the spread in sea ice reduction.

Fig. 1: Enhanced ENSO slows sea ice melt but hastens shelf ocean warming by suppressing warm upwelling.
figure 1

a–c, Inter-model correlation of changes in ENSO variability (oC oC−1 of global warming) with zonally averaged shelf water temperature (oC oC−1 of global warming) (a), sea ice concentration (SIC) changes (% oC−1 of global warming) (b), and an averaged Antarctic shelf water warming (black box in a) and sea ice changes averaged over the 60o S–70o S latitude band (c). d, Inter-model correlation of the averaged Antarctic shelf water warming with ENSO variability changes, both scaled by global warming in each model. In a, the latitude is shifted relative to the 1,000 m isobath (green curve in d) at each longitude before zonal averaging (see ‘Zonal mean of shelf ocean temperatures’ in Methods). In c, linear fits are displayed together with the correlation coefficient r and P value; dotted areas in a, b and d indicate statistical significance above the 95% confidence level determined by a two-sided Student’s t-test. e,f, Projected change in heating from upwelling (oC s−1 oC−1 of global warming) averaged over the top five models with greatest ENSO variability increase (e), and over the bottom five models with smallest ENSO variability change (f). Dotted areas in f indicate where the differences between e and f are statistically significant above the 95% confidence level as determined by a two-sided Student’s t-test.

Full size image

A total of 23 out of 31 (74%) models generate an increased ENSO amplitude with a multimodel mean increase of ~11.4%, consistent with a recent finding25. The overall increase in ENSO variability accelerates Antarctic shelf ocean warming but slows sea ice reduction. Below we show that the impact of ENSO change is effected via the circum-Antarctic warm upwelling.

Based on a dynamically consistent ocean–atmosphere reanalysis over the 1979–2018 period, poleward 60° S, grid-point upwelling values at 100 m can be approximated by wind-driven Ekman upwelling values with a correlation of 0.86. Thus, we use wind-driven Ekman upwelling and modelled grid-point vertical temperatures at 100 m to calculate changes in heating from warm upwelling, as in previous studies26.

In models with a larger increase in ENSO amplitude, ENSO-induced easterly anomalies are greater over much of Antarctica. The difference in the ENSO-rectified winds is seen in an inter-model relationship, or in a comparison between the top five models that produce the largest ENSO increase and the bottom five models that simulate the smallest ENSO change (Extended Data Fig. 7). In the top five models, as the rectified easterlies offset the westerly poleward intensification, the increase in circum-Antarctic warm upwelling is rather small (Fig. 1e), except in the Amundsen and Bellingshausen Seas. In the bottom five models, the increase in circum-Antarctic warm upwelling is large and little affected by ENSO change. The difference between the two groups of models is statistically significant (stippled areas in Fig. 1f) but not exactly zonally symmetric, suggesting that the change due to the Pacific South American teleconnection modifies the zonal-symmetric pattern. In models with a greater ENSO increase, because relatively less of the shelf ocean warm water is transported to the surface, shelf ocean warming accelerates but surface warming slows. Through this process, a greater increase in ENSO variability leads to a slower sea ice reduction despite a larger heating from the atmosphere (Extended Data Fig. 8).

The role of the changing circum-Antarctic warm upwelling in affecting shelf ocean warming and sea ice melt is confirmed in a multivariate empirical orthogonal function27 (MV-EOF) analysis on modelled changes, all scaled by the increase in global mean SST in each model. The first principal pattern depicts their dominant covarying changes, regardless of how the changes are driven, and the associated first principal component (PC1) indicates the inter-model differences. The process discussed above emerges as a smaller increase in Antarctic warm upwelling generates a faster Antarctic shelf ocean warming, which would favour a faster ice shelf and ice sheet melt, but a slower warming in the surface ocean that is conducive to a slower sea ice reduction (Fig. 2a–c). The role of the projected ENSO changes is reflected in an inter-model correlation of r = 0.68 between ENSO changes and MV-EOF PC1 (Fig. 2d).

Fig. 2: Reduced warm upwelling accelerates shelf ocean warming but slows surface warming.
figure 2

a–c, Shown are the first principal pattern of a multivariate EOF (MV-EOF) analysis of inter-model differences in projected changes in the domain poleward of 55o S from the twentieth century to the twenty-first century of heating of the surface layer from warm upwelling (oC s−1 oC−1 of global warming) (a), zonally averaged shelf water temperatures (oC oC−1 of global warming) (b) and SIC (% oC−1 of global warming) in the domain south of 55o S (c). d, Inter-model relationship between the associated principal component (PC1, s.d.) and ENSO variability change (oC oC−1 of global warming); a linear fit is displayed together with the correlation coefficient r and the P value. Dotted areas in a–c indicate regression coefficients that are statistically significant above the 95% confidence level as determined by a two-sided Student’s t-test.

Full size image

In summary, the projected twenty-first century enhancement in ENSO variability probably accelerates Antarctic shelf ocean warming and ice shelf and ice sheet melt, but slows surface warming around sea ice edge and sea ice reduction. This impact from projected increase in ENSO variability is exerted by moderating the high-latitude poleward westerly intensification, ultimately reducing the increase in circum-Antarctic warm upwelling. Thus, implications of the projected increase in ENSO variability transcend risks of extreme weathers28 and extend into impacts on changing Antarctic sea ice and Antarctic ice shelf and ice sheet.

Methods

Observations

We used ocean temperature data from Argo gridded data1 since 2004 to examine the vertical temperature structures and their changes over time (Extended Data Fig. 2), and SSTs from the Hadley Centre Sea Ice and SST dataset (HadISST) (ref. 29) to examine the observed ENSO property. Monthly fields were used and, where needed, anomalies were calculated with reference to the monthly climatology over the full period (1979–2018) and quadratically detrended.

We used surface heat fluxes and winds from six products for a multiproduct average, and these products included the National Centre for Environmental Prediction (NCEP)/National Centre for Atmospheric Research (NCAR) reanalysis 1 (1.9° latitude and 1.875° longitude) (ref. 30), the fifth generation European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis for the global climate and weather (ERA5; 0.25° grid) (ref. 31), the Japanese 55 year Reanalysis (JRA-55; 1.25° grid) (ref. 32), the Modern-Era Retrospective Analysis for Research and Applications Version 2 (MERRA2; 0.5° latitude and 0.625° longitude) (ref. 33), and the combination of NCEP Climate Forecast System Reanalysis (CFSR; 1979–2010; ~0.31° grid) (ref. 34) and the NCEP Climate Forecast System Version 2 (CFSV2; 2011–2020; ~0.2° grid) (ref. 35). Each field was re-gridded onto a common 1° × 1° resolution before averaging across products. Together with ECMWF three-dimensional monthly fields of Ocean Reanalysis System 5 (ORAS5; 0.25° grid) (ref. 36) over the 1979–2018 period, the multiproduct averaged winds were used to reveal the imprints of two Southern Ocean upwelling pathways on temperature as shown in Extended Data Fig. 2a–c.

We used the oceanic circulation fields from ORAS5 (ref. 36) over the 1979–2018 period to conduct a full heat budget analysis over the upper 100 m poleward 60° S (Extended Data Fig. 2d). The heat budget is described by

$$\frac{{\partial T_\mathrm{m}}}{{\partial t}} = - V\frac{{\partial T}}{{\partial y}} - W\frac{{\partial T}}{{\partial z}} + \frac{{\mathrm{NHF}}}{{\rho _0C_PH}} + R$$
(1)

where Tm is average temperature over the upper 100 m; \(- V\frac{{\partial T}}{{\partial y}}\) is meridional advection averaged over the upper 100 m; \(- W\frac{{\partial T}}{{\partial z}}\) is vertical advection across the 100 m depth; \(\frac{{\mathrm{NHF}}}{{\rho _0C_PH}}\) is atmospheric heating/cooling from net heat flux (NHF) averaged over the six analysis products, in which ρ0 is the density of seawater taken as 1,025 kg m−3, CP the heat capacity of seawater taken as 4,200 J kg−1 K−1, H is the ocean upper layer depth taken as 100 m and R is the residual which includes zonal advection, mixing and diffusive terms. The vertical gradient \(\frac{{\partial T}}{{\partial z}}\) was calculated as the difference between averages over the 70–100 m and 100–130 m depth. The regional average poleward 60° S over the period shows that atmospheric cooling (referred to as Q = \(\frac{{\mathrm{NHF}}}{{\rho _0C_PH}}\)) and cooling due to meridional advection (referred to as Fv) are largely balanced by heating from warm upwelling (referred to as Fw).

A comparison of oceanic currents with those inferred from wind-driven Ekman flows shows that upwelling across the 100 m depth and meridional currents across 60° S are highly correlated with the wind-driven Ekman upwelling \(W\mathrm{e} = \left( {\nabla \times \frac{\tau }{{\rho f}}} \right)\) and Ekman velocity \(V\mathrm{e} = - \frac{{\tau _x}}{{Hf\rho }}\), respectively, with a correlation of 0.86 and 0.95. Here, τx is the grid-point zonal wind stress; f is the Coriolis parameter calculated as 2Ω sin(θ), where Ω is the rotation rate of the planet and θ is latitude (f is negative in the Southern Hemisphere). As such, we examined heating associated with Ekman flows in climate models.

CMIP6 outputs

We took monthly outputs of ocean and atmosphere fields from 31 CMIP6 models in which data are available for ocean temperature, surface wind stress, heat flux (latent heat flux from Korea Institute of Ocean Science and Technology Earth System Model was unavailable) and Antarctic sea ice concentration. These models were forced under historical forcings (1850–2014) and the SSP5–8.5 (2015–2100) emission scenario20 (see Supplementary Table 1 for details). Before data analysis, outputs of each model were re-gridded onto a common 1° × 1° resolution. For ocean temperatures, the common vertical levels from the surface to 5,000 m were on a 5 m increment.

The projected change in heating from warm upwelling was calculated as the difference between averages over the 2000–2099 (twenty-first century) and the 1900–1999 (twentieth century) periods. Change in heating from warm upwelling was calculated as \(- \Delta \left( {\nabla \times \frac{\tau }{{\rho f}}} \right)\frac{{d\bar T}}{{dz}}\) between the twenty-first and twentieth centuries; \(\frac{{d\bar T}}{{dz}}\) is the grid-point climatological vertical temperature gradient. The gradient was calculated as the difference between temperature averaged over 50–100 m and that averaged over 100–150 m. Using averages over different depth ranges leads to similar results.

ENSO response

We used the conventional Niño3.4 index for ENSO, which is defined as the SST anomaly in the central-eastern equatorial Pacific (5° S–5° N, 170° W–120° W). Monthly SST anomalies were constructed with reference to the climatology over the 1900–1999 period and quadratically detrended. We calculated the change in ENSO variability as the difference between the twenty-first and twentieth centuries in the standard deviation of the Niño3.4 index during the ENSO peak season (DJF). Similarly, the change was scaled by °C−1 of global mean SST warming.

Zonal mean of shelf ocean temperatures

Conventional zonal average of ocean temperatures could lead to averages over the open ocean and near-shelf areas, rather than the shelf ocean only. To overcome this issue, we used an approach employed in previous studies16,24 in which the latitude is shifted relative to the 1,000 m isobath at each longitude before zonal averaging. That is, the 1,000 m isobath was used as a reference point to composite the temperature offshore and onshore from this point around the Antarctic shelf37. This allowed the Antarctic shelf to be clearly seen in adjusted latitudes south of the 1,000 m isobath (that is, negative x axis), and the bottom shelf and open ocean to be averaged north of the 1,000 m isobath (that is, positive x axis).