Introduction

Antarctic sea ice (Fig. 1a) contributes to Earth’s climate through the exchange of heat, momentum, freshwater, and gases between the atmosphere and ocean. Antarctic sea ice extent (SIE) and its anomaly (Fig. 1b) show a slightly increasing trend in the past few decades1,2, but abruptly declines after 20163,4. The increasing SIE trend can be caused by various processes: (1) deepening of the Amundsen Sea Low5 driven by the Interdecadal Pacific Oscillation (IPO)6 and possibly linked to the Atlantic Meridional Overturning Circulation7, (2) intensification of westerlies associated with a positive trend of Southern Annular Mode (SAM) arising from stratospheric ozone depletion8,9 and increasing greenhouse gases10, although the role of ozone is debated11, (3) increased freshwater from basal melting of Antarctic ice shelves12, although the role of freshwater is also debated13,14,15, (4) enhanced upper-ocean stratification due to brine rejection associated with sea ice increase16, and (5) weakening of Southern Ocean deep convection17. The sea ice decrease in 2016 could have been caused by more poleward warm air advection accompanied by a negative phase of the SAM3 and atmospheric teleconnection from the tropics18. Since then, the persistence of low sea ice could have been related to the warmer upper Southern Ocean19,20,21. However, due to short observation records, it remains unclear what the relative roles were of natural variability and possibly climate change.

Fig. 1: Observed Antarctic sea ice during 1979–2022.
Fig. 1: Observed Antarctic sea ice during 1979–2022.
Full size image

a Annual mean of Antarctic sea ice concentration (color in %) during 1979–2022 from NSIDC. b Time series of monthly Antarctic SIE anomaly (black in 106 km2) during 1979–2022 from NSIDC. Here we defined the monthly anomaly as a deviation from the monthly climatology during 1979–2022. A dashed line indicates a zero line.

Efforts have been made to reconstruct long-term Antarctic SIE using paleoclimate records22,23,24. A recent study25 reconstructed the total Antarctic SIE using climate indices since 1905 and identified a multidecadal regime shift in the middle 20th century. Another study26 analyzed marine core records in the past 2000 years, suggesting that multidecadal modulation of El Niño-Southern Oscillation (ENSO) and the SAM may affect local sea surface temperature (SST) and sea ice at decadal-to-centennial timescales. These studies provide observational evidence for the multidecadal variability of Antarctic sea ice, but the physical processes have not fully been investigated.

Low-frequency variability in the Southern Ocean also plays a role. Multi-centennial simulations of coupled climate models27,28,29 demonstrate that deep convection predominantly occurs in the Pacific and Atlantic sectors and exhibits multidecadal-centennial variability in association with subsurface heat build-up. When deep convection is weak, the heat accumulates in the subsurface while the upper ocean is cool and sea ice increases. During active convection, on the other hand, the ocean temperature vertically homogenizes, warming the surface ocean and reducing the sea ice. Given these relationships, the increasing SIE trend over 1979–2012 can be reproduced in coupled model simulations initialized from an active phase of the deep convection17. However, most of coupled models tend to simulate stronger deep convection compared to the observation30. The detailed mechanisms on the link between the deep convection and the sea ice variability are not fully understood.

To fill a gap in current understanding of Antarctic sea ice variability, this study attempts to elucidate physical mechanisms on the multidecadal variability of Antarctic sea ice using reconstructed data and coupled model simulations. For this purpose, this study performs a statistical analysis of the 300-yr Antarctic sea ice data assimilation reconstruction and compares that analysis against two 3000-yr coupled model simulations with different atmospheric resolutions, and against other available coupled model simulations. We further discuss a role of external forcings in the Antarctic sea ice multidecadal variability by comparing two coupled model simulations with and without external forcings. All of these model experiments allow us to elucidate the sensitivity of the simulated multidecadal variability of Antarctic sea ice to the model resolutions, physics, and external forcings.

Results

Reconstructed Antarctic SIE anomalies (Fig. 2a; see Methods) over the past 300 years show a significant multidecadal variability with a period longer than 30 years (Fig. 2b). In particular, the multidecadal variability with a period of 80–100 years is pronounced during the analysis period (Fig. 2c), although the variability with a period of 40–50 years is also significant in the earlier period. Since the multidecadal variability with a period of 80–100 years is not significant in the underlying model simulation used for the sea ice reconstruction (Supplementary Fig. 1a, b), the multidecadal variability in the sea ice reconstruction is generated by the data assimilation approach constrained to the paleoclimate proxies24,31. However, it is difficult to verify the existence of multidecadal variability from the reconstructed data only, due to the short observation record. This motivates us to analyze prolonged model simulations for further verification.

Fig. 2: Multidecadal variability of Antarctic SIE anomalies reconstructed during 1700–2000.
Fig. 2: Multidecadal variability of Antarctic SIE anomalies reconstructed during 1700–2000.
Full size image

a Time series of detrended Antarctic SIE anomalies (in million km2) reconstructed via the paleoclimate data assimilation in the past 300 years. Here DL2023 stands for the reconstruction data from Dalaiden et al. (2023)48. b Fourier power spectrum of the detrended Antarctic SIE anomalies. The power density is normalized by its variance (\({\sigma }^{2}\)). A dashed line corresponds to 95% confidence level of the normalized power density using a chi-squared test. c Wavelet power spectrum (color) of the Antarctic SIE anomalies. The power density is normalized by its variance (\({\sigma }^{2}\)). Black contour encloses regions with 90% confidence level of the normalized power density using a chi-squared test.

Antarctic SIE simulated in SPEAR_LO and SPEAR_MED coupled general circulation models32 (see Methods) show distinct multidecadal variations (Supplementary Fig. 2a, b). Compared to SPEAR_MED, SPEAR_LO shows a stronger multidecadal variability in Antarctic SIE. In particular, SPEAR_LO can simulate a slow sea ice increase followed by a rapid sea ice decrease by around 4 million km2 within ‘one year’ (e.g., around 2830 year in Supplementary Fig. 2a), which was observed in 2016 (Fig. 1b). Considering the length of model simulations, the sea ice decline observed in 2016 is an extremely rare event in the model world. A similar rareness of the event in the model world is also reported for the recent sea ice decline in winter 202333.

The detrended Antarctic SIE (Fig. 3a, b) anomalies exhibit significant powers at a period of 80–100 years, consistent with the reconstructed results (Fig. 2c). Fourier power spectra of the detrended Antarctic SIE (Fig. 3c) anomalies show significant peaks with 97-yr period for SPEAR_LO and 92-yr period for SPEAR_MED. Probability density function (PDF) of the standardized Antarctic SIE anomalies (Fig. 3d) shows positive median values of the standardized Antarctic SIE anomalies for SPEAR_LO (0.14) and SPEAR_MED (0.08), respectively. Both models have negative skewness of the PDF with a greater number of events associated with positive SIE anomalies (55.9% for SPEAR_LO and 54.8% for SPEAR_MED) than negative SIE anomalies (44.1% for SPEAR_LO and 45.2% for SPEAR_MED). There are also a greater number of ‘extremely’ negative SIE anomalies than positive SIE anomalies, suggesting a stronger feedback for sea ice decrease than sea ice increase.

Fig. 3: Multidecadal variability of Antarctic SIE in SPEAR_LO and SPEAR_MED control simulations.
Fig. 3: Multidecadal variability of Antarctic SIE in SPEAR_LO and SPEAR_MED control simulations.
Full size image

a Wavelet power spectrum (color) of detrended Antarctic SIE anomalies for the 3000-yr SPEAR_LO CTL simulation. The power density is normalized by its variance (\({\sigma }^{2}\)). Black contour encloses regions with 95% confidence level of the normalized power density using a chi-squared test. b, Same as in (a), but for the 3000-yr SPEAR_MED CTL simulation. c Fourier power spectra of the detrended Antarctic SIE anomalies for the 3000-yr SPEAR_LO (red) and SPEAR_MED (blue) CTL simulations. The power density is normalized by its variance (\({\sigma }^{2}\)). Dashed lines correspond to 95% confidence levels of the normalized power density using a chi-squared test. d Probability distributions (%) of the standardized Antarctic SIE anomalies (\(\sigma\)) for the 3000-yr SPEAR_LO (red) and SPEAR_MED (blue) CTL simulations.

Since the auto-correlations of Antarctic SIE anomalies decrease by around half in 15 years (Supplementary Fig. 3a, b), we investigate generation mechanisms on the multidecadal variability for low and high sea ice events (see Methods) by applying 15-yr lowpass filter to the SIE anomalies (Supplementary Fig. 4a, b). Compared to the short persistence of the observed Antarctic SIE anomalies typically with around one year, the 15-yr persistence of the Antarctic SIE anomalies in the model simulations (Supplementary Fig. 3a, b) is much longer. This is due to a stronger multidecadal variability than an interannual-decadal variability in the model simulations (Fig. 3a–c). This can also be seen in the reconstructed Antarctic SIE anomalies (Fig. 2b, c), although it is hard to verify this feature using the short satellite observation (Fig. 1b). Therefore, we employed the 15-yr lowpass filter to suppress a high-frequency variability and highlight a growth and decay of the multidecadal variability. We have also analyzed the unsmoothed anomalies to identify the timing (i.e., precise year) of the emergence of the multidecadal sea ice variability, as will be discussed later.

During the low sea ice events, negative SIE anomalies start to appear significantly at around 20 years before the peak (Fig. 4a). Positive zonal wind stress and negative wind stress curl anomalies start to appear significantly 20–30 years before the peak of negative SIE anomalies (Fig. 4a), indicating that the wind anomalies lead the sea ice anomalies. Stronger westerlies tend to induce northward Ekman transport thereby cooling the surface and increasing sea ice. However, on longer timescales, the wind anomalies tend to enhance upwelling of subsurface warm water and facilitate sea ice melt from below9,34. This so-called two-timescale oceanic response to stronger westerlies varies among climate models with duration of the initial cooling ranging from 2 to 30 years35. Negative meridional wind stress anomalies may contribute to sea ice decrease through poleward warm air advection3, but net negative (upward) surface heat flux anomalies due to latent heat flux and longwave radiation tend to cool the surface and counteract the sea ice decrease (Fig. 4b). The surface heat flux anomalies represent the response to the sea ice anomaly rather than the driver of the anomaly.

Fig. 4: Atmosphere and ocean variability during low Antarctic sea ice events for SPEAR_LO control simulation.
Fig. 4: Atmosphere and ocean variability during low Antarctic sea ice events for SPEAR_LO control simulation.
Full size image

a Lead/lag composite anomalies of Antarctic SIE (black in 106 km2), Southern Ocean (south of 60°S) zonal (red in 10−3 Pa) and meridional (blue in 10−4 Pa) wind stress, and wind stress curl (light blue in 10−9 Pa m−1) during the low Antarctic sea ice events for the SPEAR_LO CTL simulation. Anomalies are smoothed with a 15-yr running mean filter. Positive (negative) lead/lag years mean that Antarctic SIE leads (lags) the other variables. Thick lines indicate significant anomalies exceeding 95% confidence level using a Student’s t-test. b Same as in (a), but for the net surface heat flux (Total; black in W m−2) and contributions from latent heat flux (red), sensible heat flux (blue), longwave radiation (light blue), and shortwave radiation (orange). Positive values indicate more downward heat flux into the ocean. c Same as in (a), but for Antarctic SIE (black in 106 km2), SAM (red in 10−1 Pa) and IPO (blue in 2 × 10−2 °C) indices. d Same as in (a), but for Antarctic SIE (black in 106 km2), SST (red in 10−1 °C), sea surface salinity (SSS; blue in 10−1 PSU), mixed-layer depth (MLD; purple in 101 m), Southern Ocean deep convection (SODC; light blue in Sv), and the upper 1000-m available potential energy (APE; orange in 101 J m−3). e Same as in (a), but for the upper 200-m ocean temperature tendency (Total; black in 10−1 W m−2) and contributions from horizontal (Adv_xy; red) and vertical (Adv_z; blue) advection, surface boundary forcing (Sbfc; light blue), and mesoscale diffusion and dianeutral mixing (Difmix; orange). f Same as in (e), but for the upper 200-m ocean salinity tendency (Total; black in 10−7 kg m−2 s−1) and contributions from horizontal (Adv_xy; red) and vertical (Adv_z; blue) advection, surface boundary forcing (Sbfc; light blue), and mesoscale diffusion and dianeutral mixing (Difmix; orange).

The wind variability can be induced by large-scale climate variability such as the SAM and IPO (see Methods)6,9. The SAM starts to become significantly positive around 25 years before the peak of negative SIE anomalies (Fig. 4c). This corresponds well to positive zonal wind stress and negative wind stress curl anomalies (Fig. 4a). The SAM also shows strong simultaneous correlations with the zonal wind stress and wind stress curl anomalies (Supplementary Fig. 5a). On the other hand, the IPO exhibits shorter fluctuations (Fig. 4c) and weaker correlations with the wind stress anomalies (Supplementary Fig. 5b) owing to regional differences in the atmospheric teleconnection of the IPO36. Considering a very weak correlation with the IPO (Supplementary Fig. 5a), our results suggest that the multidecadal modulation of the SAM37 plays a greater role in the surface wind variability.

We now examine the ocean properties before quantifying the impact of upwelling and horizontal advection. During the low sea ice events, positive SST and SSS anomalies (Fig. 4d) are almost in phase with the negative SIE anomalies. The positive SST and SSS anomalies lag behind positive mixed-layer depth (MLD) and deep convection anomalies by a few years (Fig. 4d). The positive MLD and deep convection anomalies lag behind positive available potential energy (APE; see Methods) anomalies in the upper 1000 m by a few years (Fig. 4d). The positive APE anomalies start to appear around 30 years before the peak of negative SIE anomalies, similar to the positive zonal wind stress and negative wind stress curl anomalies (Fig. 4a). These lags support the physically plausible interpretation that the decrease in the upper-ocean stability causes a deeper mixed layer and stronger convection, which bring more subsurface warm and saline water to the surface. The role of vertical process is further supported by the analysis of the upper 200-m temperature and salinity tendency anomalies (Fig. 4e, f; see Methods). The positive temperature tendency anomalies are mostly explained by vertical advection followed by mesoscale diffusion and dianeutral mixing (Fig. 4e). The warming effect of ocean upwelling is stronger than the cooling effect of the equatorward Ekman transport by the stronger westerly (Fig. 4e). We also find that the positive salinity tendency anomalies are due to vertical advection followed by horizontal advection (Fig. 4f).

Positive and negative temperature anomalies start to appear above and below 200 m around 20 years before the peak of negative SIE anomalies (Fig. 5a). Positive salinity anomalies also start to develop in the subsurface ocean and gradually extend to the surface (Fig. 5b). These anomalies lead to the development of positive density anomalies in the upper 400 m (Fig. 5c), which results in deeper mixed layer and weaker ocean stratification in the upper ocean (Fig. 5d). Both positive and negative temperature anomalies above and below 200 m (Fig. 5a) contribute to stronger ocean stratification (Fig. 5e), while the positive salinity anomalies act to weaker stratification (Fig. 5f). This indicates a positive salinity-convection feedback in which the increased salinity associated with the enhanced vertical advection (Fig. 4f) weakens upper ocean stratification and induces stronger convection, bringing more subsurface warm and saline water to the surface.

Fig. 5: Subsurface ocean variability during low Antarctic sea ice events for SPEAR_LO control simulation.
Fig. 5: Subsurface ocean variability during low Antarctic sea ice events for SPEAR_LO control simulation.
Full size image

a Lead/lag composite anomalies of potential ocean temperature (in 10−1 °C) in the upper 700 m during the low Antarctic sea ice events for the SPEAR_LO CTL simulation. Positive (negative) lead/lag years mean that Antarctic SIE leads (lags) the other variables. Colors indicate significant anomalies exceeding 95% confidence level using a Student’s t test. Black line corresponds to the mixed-layer depth at which the potential density increases by 0.03 kg m−3 from the ocean surface. b Same as in (a), but for salinity (in 10−1 PSU). c Same as in (a), but for potential density (Rho0 in 10−1 kg m−3). d Same as in (a), but for ocean stratification (in 10−4 kg m−4) estimated by minus vertical gradient of potential density (-dRho0/dz). Positive (negative) values indicate stronger (weaker) stratification. e Same as in (d), but for contribution from the potential temperature anomaly to the ocean stratification anomaly (-dRho0_T/dz in 10−4 kg m−4). f Same as in (d), but for contribution from the salinity anomaly to the ocean stratification anomaly (-dRho0_S/dz in 10−4 kg m−4).

One may wonder which of the surface wind and deep convection variations is more important for the upper-ocean variability. The SST and SSS anomalies averaged in the Southern Ocean (south of 60°S) show significantly high correlations and almost in-phase relationships with the deep convection anomalies (Supplementary Fig. 5c, d). The correlations with surface wind stress anomalies are weaker, but their peaks lead the peaks of the deep convection anomalies by 10–20 years. This lagged relationship can also be seen in the lead/lag correlations of the deep convection anomalies with the wind anomalies and the SAM index (Supplementary Fig. 5e). Our model results suggest that the surface wind variability associated with the SAM may trigger the salinity-convection feedback, although the deep convection is a primary driver of the SST and SSS variability.

To verify the triggering role of surface wind variability in the deep convection more clearly, we describe time series of the unsmoothed anomalies composited for the low sea ice events (Supplementary Fig. 6a). The stronger westerlies associated with the positive SAM start to appear significantly at −30 years when the positive SIE anomalies start to decline. On the other hand, the Southern Ocean deep convection anomalies remain significantly negative at −30 year to contribute to the sea ice increase, playing a counteractive role. Here we should keep in mind that the signs of the SAM and deep convection anomalies matter for the rate of changes in the SIE anomalies. The stronger westerlies associated with the positive SAM after −30 year act to weaken the negative deep convection anomalies and turn them into the positive anomalies after −25 year so that the upwelling of subsurface warm water is enhanced to generate the negative SIE anomalies afterwards. This represents a slow oceanic response to the surface wind variability through the Southern Ocean deep convection rather than a fast oceanic response via the Ekman transport9,35. We note that this mechanism is based on our model results, so our models may underestimate the dynamical influence of surface wind on sea ice compared to the observations38,39. It is also difficult to verify the role of the deep convection using observational data owing to short observation records and a lack of subsurface ocean observations.

Compared to the low sea ice events, the high sea ice events tend to develop slowly with smaller amplitudes (Supplementary Figs. 7, 8), consistent with the observed behavior of Antarctic sea ice in the satellite period (Fig. 1b). We cannot explain the development of positive SIE anomalies by positive (downward) surface heat flux anomalies (Supplementary Fig. 7b). Rather, the positive SIE anomalies are strongly associated with negative zonal wind stress and positive wind stress curl anomalies linked to multidecadal modulation of negative SAM (Supplementary Fig. 7a, c). Composite of the unsmoothed anomalies (Supplementary Fig. 6b) also shows that negative zonal wind stress anomalies associated with the negative SAM start to appear at −40 year, around 10 years before Southern Ocean deep convection anomalies start to become negative with the emergence of the positive Antarctic SIE anomalies at −30 year. Our model simulations demonstrate a triggering role of the surface wind variability associated with SAM in the deep convection variability responsible for the sea ice variability. On the other hand, the 15-year low-pass-filtered salinity anomalies (Supplementary Figs. 7d, 8b) tend to show smaller amplitudes with shallower structures compared to the low sea ice events (Figs. 4d and 5b). The deep convection and stratification anomalies also show weaker anomalies (Supplementary Figs. 7d, 8d). This suggests that the high sea ice events may involve a weaker salinity-convection feedback because of a weaker contribution from the subsurface warm and saline ocean due to shallower mixed-layer (Supplementary Fig. 7d). The mixed-layer has a limited potential in shoaling because of the presence of the ocean surface with the strong stratification. This may be responsible for the asymmetric feature of the high sea ice events with smaller anomalies.

In comparison to SPEAR_LO (Fig. 4), SPEAR_MED (Supplementary Fig. 9) tends to simulate weaker negative SIE anomalies during the low sea ice events. The ocean temperature, salinity, and density anomalies (Supplementary Fig. 10) have shallower and weaker structures. However, amplitudes of the positive zonal wind stress and negative wind stress curl anomalies associated with the multidecadal modulation of positive SAM (Supplementary Fig. 9a, c) are similar to those in SPEAR_LO (Fig. 4a, b). We also find a similar influence of the positive zonal wind stress anomalies associated with the positive SAM in triggering the positive anomalies of the Southern Ocean deep convection in SPEAR_MED (Supplementary Fig. 11). Therefore, the model differences may be due to differences in the ocean background mean states29. We note that both models reach almost equilibrium states at around 2000 year (Supplementary Fig. 12a, b). SPEAR_MED tends to show lower temperature and salinity in the subsurface ocean than SPEAR_LO (Supplementary Fig. 13a, b) indicating a weaker influence of subsurface warm and saline on the sea ice variability, although differences in the ocean density and stratification are small (Supplementary Fig. 13c, d). This may be related to weaker poleward advection of warm and saline Circumpolar Deep Water associated with weaker Deacon Cell (Supplementary Fig. 13e) at around 60–50°S owing to weaker westerly (Supplementary Fig. 13f).

Discussion

We have demonstrated a possible existence of multidecadal variability in Antarctic sea ice using the long-term sea ice reconstruction and model simulations. This is the first attempt to apply the paleoclimate data assimilation approach for examining the multidecadal sea ice variability (Fig. 2). We have also identified comprehensive physical processes that control the multidecadal sea ice variability in quantitative ways: the surface wind variability associated with the SAM triggers the Southern Ocean deep convection variability through changes in upwelling of subsurface warm and saline water and the resultant salinity-convection feedback plays a key role in generating the SST and sea ice variability (Figs. 4 and 5), which was not thoroughly investigated in the previous modeling studies17,19,20,21. The salinity-convection feedback operates stronger for the low sea ice events because of the mixed layer having a large potential in deepening, which allows the low sea ice events to evolve faster with stronger amplitude than the high sea ice events (Fig. 3d). This is consistent with the current observation of slow sea ice increase followed by abrupt sea ice decrease after 2016 (Fig. 1b).

These results are based on a single model with different atmospheric resolutions, and the processes might be model-dependent. To further discuss the model dependency, we analyze other CMIP6 models40 (see Methods) available for more than 1000 years. Antarctic SIE anomalies show multidecadal variability with various periods (Figs. 6 and 7). Among the CMIP6 models, CanESM541 (Fig. 6a) shows a similar period (80–100 years) with the SPEAR models (Fig. 3a–c), while other models represent a wide range of periods (30–90 years). The difference in the periods may be related to the background ocean mean state29, which is represented by the relationship with annual mean mixed-layer depth (Fig. 7f). The models with deeper mixed-layer tend to have a higher frequency and vice versa. This is probably due to the more frequent occurrence of deep convection in the Southern Ocean because of weaker ocean stratification in the models. It should be noted that the magnitude and frequency of deep convection in most of CMIP6 models30 may be so high that the models tend to simulate longer persistence of sea ice anomalies compared to the observation (Fig. 1b). Composite anomalies of atmosphere and ocean variables during the low sea ice years of CanESM5 (Supplementary Fig. 14) show similar results with SPEAR models (Fig. 4, Supplementary Fig. 9), but the positive deep convection and SSS anomalies lead positive SST and negative SIE anomalies as well as positive zonal wind stress and negative wind stress curl anomalies. This supports the primary role of the deep convection in the Southern Ocean, but there is no evidence for the triggering role of surface wind variability in CanESM5.

Fig. 6: Multidecadal variability of Antarctic SIE in CMIP6 model control simulations.
Fig. 6: Multidecadal variability of Antarctic SIE in CMIP6 model control simulations.
Full size image

a Wavelet power spectrum (color) of detrended Antarctic SIE anomalies for the 1000-yr CanESM5 CTL simulation. The power density is normalized by its variance (\({\sigma }^{2}\)). Black contour encloses regions with 95% confidence level of the normalized power density using a chi-squared test. b–e Same as in (a), but for the 1200-yr CESM2, 2000-yr IPSL-CM6A-LR, 1000-yr MPI-ESM-1-2-LR, and 1200-yr INM-CM5-0 simulations, respectively.

Fig. 7: Multidecadal variability of Antarctic SIE in the CMIP6 model control simulations and its relationship with mixed-layer depth (MLD).
Fig. 7: Multidecadal variability of Antarctic SIE in the CMIP6 model control simulations and its relationship with mixed-layer depth (MLD).
Full size image

a Fourier power spectrum (solid) of detrended Antarctic SIE anomalies for the 1000-yr CanESM5 CTL simulation. The power density is normalized by its variance (\({\sigma }^{2}\)). A dotted line indicates 95% confidence level of the normalized power density using a chi-squared test. b–e Same as in (a), but for the 1200-yr CESM2, 2000-yr IPSL-CM6A-LR, 1000-yr MPI-ESM-1-2-LR, and 1200-yr INM-CM5-0 simulations, respectively. f Relationship between the annual mean MLD (in m) and the period (in years) of multidecadal SIE variability for the CMIP6 models and SPEAR_LO/MED models. Due to the limited availability of subsurface ocean variables, we removed INM-CM5-0 from the panel. The black solid line indicates a linear regression using the least square method and the correlation values is shown in the panel.

We have focused on the natural variability of Antarctic sea ice by analyzing the control simulations only. However, the Antarctic sea ice can be influenced by both internal and forced variability in the real world. To further discuss the relative role of the external forcings, we analyze ensemble mean of 13 CESM1_LME42 (see Methods) simulations with full external forcings during 850-2005 CE and compare the results with each member of CESM1_LME simulations as well as a single CESM1_CTL42 simulation (see Methods). Both wavelet and Fourier power spectra of detrended Antarctic SIE anomalies show that the multidecadal variability gets weaker and disappears for the ensemble mean of CESM1_LME simulations (Fig. 8a, c). However, some of the members show a significant multidecadal variability with a period of 80–100 years (Supplementary Fig. 15). We also find a distinct multidecadal variability with a period of 72.3 years for CESM1_CTL simulation (Fig. 8b, c). These comparison results indicate that the forced variability (Fig. 8a) cannot generate the multidecadal variability simulated in some of the members (Supplementary Fig. 15). On the other hand, in the real world, the internally generated multidecadal variability appears to be so large not to be suppressed by the forced variability. This is also evident in the distinct multidecadal variability (Fig. 2) seen in the CESM1_LME simulation when it assimilates paleoclimate proxy data that include both the internal and forced variability of the real world. While there is some evidence of multidecadal variability with a period of 80–100 years in the underlying model simulations used for the sea ice reconstruction (Supplementary Fig. 15), the data assimilation method is such that the multidecadal variability in the sea ice reconstruction is generated only by the data assimilation approach constrained to the paleoclimate proxies24,31 and not the underlying model. Therefore, the CESM1_LME simulation may have a bias in simulating the internal variability smaller than the forced variability, in contrast to the real world. We need further in-depth analysis for the role of external forcings in the multidecadal sea ice variability and future change of Antarctic sea ice, but this is beyond the scope of this study and will be explored in a separate study.

Fig. 8: Multidecadal variability of Antarctic SIE in the CESM1 control and last millennium ensemble simulations.
Fig. 8: Multidecadal variability of Antarctic SIE in the CESM1 control and last millennium ensemble simulations.
Full size image

a Wavelet power spectrum (color) of detrended Antarctic SIE anomalies during 850-2005 CE for the ensemble mean of 13 CESM1_LME simulations with full radiative forcings. The power density is normalized by its variance (\({\sigma }^{2}\)). Black contour encloses regions with 95% confidence level of the normalized power density using a chi-squared test. b Same as in (a), but for a single CESM1_CTL simulation. c Fourier power spectrum of detrended Antarctic SIE anomalies for CESM1_LME (blue solid line) and CESM1_CTL (red solid line) simulations. The power density is normalized by its variance (\({\sigma }^{2}\)). A dotted line indicates 95% confidence level of the normalized power density using a chi-squared test.

This study elucidates physical mechanisms on the multidecadal variability of Antarctic sea ice, but the simulated amplitude and period are dependent on the model resolutions and physics. This is consistent with a previous study29 that demonstrates the dependence of Southern Ocean multidecadal variability on the ocean background mean states. An increase in the atmospheric resolution leads to a smaller amplitude of the multidecadal variability due to colder and fresher mean states of the Southern Ocean. However, the Southern Ocean involves rich mesoscale eddies with a horizontal scale of around 100 km, which are not resolved in the models. The mesoscale eddies can affect the convection and restratification through diapycnal and lateral mixing43. They also delay and intensify the Southern Ocean warming/cooling associated with the SAM through changes in the vertical advection44. Resolving the mesoscale eddies enhances the poleward heat transport, warms the ocean surface, and reduces the sea ice mean state45. Given these effects, we need to increase ocean resolution, at least to 0.1°46. A recent study47 demonstrates in the 350-yr simulation that the increase in ocean resolution from 1° to 0.1° allows to simulate the multidecadal variability of Antarctic sea ice, although the period is 40–50 years shorter than our model simulations. The periodicity may also be subject to the ocean background mean state29 depending on the model resolutions and physics, so further concerted efforts using longer model simulations with higher resolutions are underway.

Methods

Reconstructed sea ice data

We used yearly SIE reconstruction via paleoclimate data assimilation48 covering the 1700–2000 CE49. The model simulations used for the data assimilation comprise three simulations of the the isotope-enabled Community Earth System Model version 1 (iCESM150). The model simulations are compared against available observations for each year via a statistical proxy system model (PSM51) that assimilates the model variables to the observed quantity. The paleoclimate records used for the data assimilation include 46 snow accumulation records52, one additional snow accumulation record53, 33 stable isotopes oxygen ratio records54, 17 sodium flux records55, and 12 tree-ring records in the Southern Hemisphere56,57. We analyzed the SIE anomalies relative to the 1979–2000 period and detrended the SIE anomalies by removing the linear trend using the least squares method. Then, we performed Morlet wavelet58 and Fourier spectral analyses to the time series of the detrended SIE anomalies to identify the multidecadal variability.

Coupled general circulation model experiments

We employed two coupled general circulation models recently developed at NOAA/GFDL, called Seamless System for Prediction and Earth System Research Low-Resolution (SPEAR_LO32) and Medium-Resolution (SPEAR_MED32). Both models consist of the AM4 atmospheric and LM4 land surface components59,60 and the MOM6 ocean and SIS2 sea ice components61. The atmospheric component of SPEAR_LO (SPEAR_MED) has a horizontal resolution of approximately 100 km (50 km) and 33 vertical levels with the model top at 1 hPa. The ocean and sea ice components have a nominal horizontal resolution of 1° with a gradual increase to 1/3° in the meridional direction toward the tropics. The ocean model has 75 layers in the vertical which include 30 layers in the top 100 m with a finer resolution. It uses a hybrid vertical coordinate which is based on a function of height in the upper ocean (a z-layer coordinate), transitioning to isopycnal layers in the interior ocean. Using SPEAR_LO (SPEAR_MED), we performed 5000-yr (3000-yr) control (CTL) simulations forced with atmospheric composition fixed at preindustrial era levels. To make a fair comparison with SPEAR_MED, we analyzed yearly output of the first 3000-yr simulation for SPEAR_LO. Here we defined the simulated climate anomalies by removing the mean value and linear trend (model drift) by the least squares method. Details on the CTL simulation, historical simulation and future projection of the Southern Ocean multidecadal variability in these models are also described in recent papers29,62.

We investigate physical mechanisms using the multidecadal periods of high and low sea ice. These periods are identified by performing a 15-yr lowpass filter onto Antarctic SIE anomalies and defining high and low sea ice events as years in which there is a local maximum and minimum of the filtered SIE anomalies that exceeds or falls below one standard deviation from the mean (Supplementary Fig. 4a, b). Here we used a 15-yr lowpass filter because the auto-correlations of Antarctic SAT and SIE anomalies decrease by around half in this period (Supplementary Fig. 3a, b). This leads to 9 (11) high sea ice events and 22 (14) low sea ice events for SPEAR_LO (SPEAR_MED). The number of low sea ice events is higher than that of high sea ice events. This is consistent with the negative skewness of Antarctic SIE anomalies, for which there are a greater number of ‘extremely’ negative SIE anomalies than positive SIE anomalies (Fig. 3d).

Other coupled general circulation model experiments

For comparison with the SPEAR models, we employ a series of pre-industrial control simulations from the sixth Coupled Model Intercomparison Project (CMIP6)40. We selected five CMIP6 models based on the length of simulations exceeding 1000 years, which allows us to obtain a sufficient number of events on the multidecadal variability for composite analysis. These are the 1000-yr simulations from CanESM541 and MPI-ESM-1-2-LR63, the 1200-yr simulations from CESM264 and INM-CM5-065, and the 2000-yr simulation from IPSL-CM6A-LR66, respectively. CanESM541 includes NEMO3.4 ocean model with nominal 1° horizontal resolution, MPI-ESM-1-2-LR63 involves MPIOM1.6 ocean model with approximate 150 km resolution, CESM264 has POP2 ocean model with nominal 1° resolution, INM-CM5-065 includes INM-OM5 with nominal 50 km resolution, IPSL-CM6A-LR66 has NEMO3.6 with nominal 1° resolution. As in SPEAR models, we used the yearly output of these model simulations and calculated the anomalies by removing the annual mean and linear trend using by the least squares method.

To further explore the relative roles of internal variability and external forcings, we analyze the CESM167 Last Millennium Ensemble (CESM1_LME)42 during 850-2005 CE and compare the results with the CESM1 control (CESM1_CTL)42 simulation. CESM167 has POP2 ocean model with nominal 1° resolution. All CESM1_LME simulations start from the year 850 of CESM1_CTL simulation with small random roundoff differences in the air temperature field. We use 13 ensemble members of CESM1_LME simulations with full external forcings and calculate the ensemble mean to describe the forced response. We used the monthly output of the model simulations and calculated the anomalies by removing the annual mean and linear trend using by the least squares method.

Ocean and climate indicators

To calculate Antarctic SIE, we use daily SIC from NSIDC, which is based on Nimbus-7 SMMR and DMSP SSM/I-SSMIS Passive Microwave Data, Version 268. The SIC data has a 25 km × 25 km horizontal resolution and covers a period of 1979–2022. We defined the SIE as the total area where the SIC exceeds 15% and calculated the monthly SIE anomaly as deviation from the monthly climatology. To examine a possible link with large-scale climate variability, we calculated the SAM index based on Antarctic Oscillation index69, which is defined as a difference in the zonally averaged sea level pressure between 40°S and 65°S and has a strong control on the strength of the Southern Ocean zonal wind stress. We also derived the IPO index70, which is defined as the 13-yr running mean differences in the SST anomalies between the tropics (10°S-10°N, 170°E-90°W) and the two subtropics (25°N-45°N, 140°E-145°W; 50°S-15°S, 150°E-160°W). For the ocean variables, we used the mixed-layer depth at which the potential density increases by 0.03 kg m−3 from the ocean surface. We also estimated the strength of the Southern Ocean deep convection by the absolute value of the minimum meridional overturning streamfunction south of 60°S corresponding to the Antarctic Bottom Water cell17. To evaluate the ocean stability, we calculated the available potential energy (APE71) in the upper 1000-m of Southern Ocean (south of 60°S) using the following equation.

$${APE}=\frac{1}{H}{\iint \!\!\! \int}_{-H}^{0}\left(\rho -\bar{\rho }\right){gzdxdydz}$$
(1)

Here, \(H\) is the water depth (set to be 1000 m), \(\rho\) is the in-situ density, and \(\bar{\rho }\) is the in-situ density vertically averaged over the water depth. To further examine the balance of temperature and salinity tendency anomalies, we analyzed the model output of horizontal and vertical advection, parameterized mesoscale diffusion and dianeutral mixing, and surface boundary forcings at each grid.