Introduction

Antarctic sea ice contributes to the formation of Antarctic Bottom Water1 that travels over the global oceans and interacts with overlying water masses affecting global climate. The Antarctic sea ice also influences regional climate through induced changes in the heat, momentum and freshwater fluxes across different timescales2. The sea ice variability is not only driven by atmosphere and ocean but also affected by ice shelf variability through changes in freshwater flux3 which induce sea level variations over the global oceans4. Therefore, accurate understanding and prediction of the Antarctic sea ice variability is necessary for future predictions of the global and regional climate.

Previous studies have reported that the Antarctic sea ice shows a slightly increasing trend, particularly in the Ross Sea, in the past few decades5,6,7. Several processes such as the atmospheric circulation changes linked to the stratospheric ozone depletion8,9,10,11, the ocean warming and ice shelf melt12,13 have been proposed to explain the increasing sea ice trend. However, extreme sea ice events have recently been observed with record high winter maximum extent (occurring in September) in 2012–2014, followed by unprecedented decline with record low summer minimum extent (occurring in February/March) in 2016–20197,14. It remains unclear whether the recent sea ice decrease reflects a part of low-frequency variability such as decadal variability.

Several studies have examined physical processes underlying decadal sea ice variability in the Antarctic seas. For example, the decadal sea ice variability in the Pacific sector shows its close link to the atmospheric circulation variability of the Amundsen Sea Low15. Between 2000 and 2014, the sea ice extent in the Ross Sea has significantly increased under the influence of the deepened Amundsen Sea Low10 induced by the Interdecadal Pacific Oscillation (IPO16). The cold air advection from the south by the anomalous low-pressure system is suggested to cause the sea ice increase in the Ross Sea. In addition, ocean-sea ice interaction involving the heat buildup below the mixed layer and the stronger vertical stratification may contribute to the sea ice increase17.

On the other hand, the decadal sea ice variability in the Atlantic sector can be generated through the local surface wind variability18,19 and the local atmosphere-ocean-sea ice interaction20,21,22. The local surface wind variability tends to be associated with the Southern Annular Mode (SAM23). During a positive phase of the SAM, westerly winds strengthen and induce anomalous northward Ekman transport. This acts to enhance the upwelling of warm water from the subsurface ocean, leading to sea ice decrease24. The relationship between the positive SAM and the sea ice decrease in the Weddell Sea is consistent with a previous study25 on the seasonal-interannual timescales, although the relationship is stronger in cold seasons (austral winter-spring) than in warm seasons (austral summer-autumn). In the absence of the SAM influence, decadal decrease of sea ice cover in the Weddell Sea enhances evaporation from the ocean surface and increases the salinity. This results in the strengthening of the Weddell Gyre and larger advection of cold water from the outside of the Weddell Sea, leading to sea ice increase22. The decadal variability of the Weddell Gyre may also have links to a low-frequency variability in the Southern Ocean that is induced by synoptic-scale weather forcing26 and ocean mesoscale eddy-mean flow interaction27. Both the vertical and horizontal ocean advection processes play key roles in inducing the sea ice increase in the Weddell Sea22.

Efforts have been made to demonstrate skillful predictions of seasonal sea ice variability of the Antarctic sea ice. For example, experiments using a coupled general circulation model (CGCM) show that sea ice edge location around Antarctica is predictable when the model’s sea ice, ocean and land are initialized with the late 20th century climate simulated by the model28. The Antarctic sea ice predictability vanishes during austral summer, but re-emerges during the following growing seasons of the Antarctic sea ice. This is mainly due to summertime ocean heat content anomalies that remain below the mixed layer. The winter reappearance of sea surface temperature (SST) anomalies through mixed-layer variations is a source of the wintertime sea ice predictability up to 3 years in some CGCMs29. Seasonal hindcast experiments also show skillful predictions of the wintertime sea ice variability ahead of 11 months in some of the Antarctic seas30. Although the Antarctic sea ice is less sensitive to changes in the sea ice parameterizations for the elastic-viscous-plastic rheology in a sea ice model31, accurate initializations of sea ice concentration (SIC) and thickness help improve sea ice prediction skills in some of the Antarctic seas32,33,34.

Predicting decadal sea ice variability in the Antarctic seas remains a big challenge among the climate research community. In a perfect model approach, an intermediate Earth system model demonstrates that sea ice predictability on an interannual timescale is limited to 3 years, and the decadal hindcast experiments initialized with pseudo-observations of surface air temperature outperform those without the initializations35. A recent study36 has demonstrated very limited prediction skills of the decadal sea ice variability in the Antarctic seas using decadal reforecast experiments of eleven models selected from the Coupled Model Intercomparison Project Phase 5 (CMIP5). Most of the models initialized on January 1st or November 1st with direct/indirect ocean and sea ice initializations do not show promising prediction skills of the Antarctic sea ice on a decadal timescale. This is mostly due to model inability to reproduce a slightly positive trend of the Antarctic sea ice extent observed in recent decades. In the case where the linear trend is removed, most of the models show higher prediction skills in the northern Ross Sea and the Weddell Sea. The authors suggested that the atmospheric teleconnection of El Niño-Southern Oscillation (ENSO) may be responsible for the interannual sea ice variability, but they did not provide much evidence of physical processes underlying the skillful predictions of the Antarctic sea ice variability.

To address these issues on the low prediction skills of the decadal sea ice variability in the Antarctic seas, this study attempts to examine impacts of ocean and sea ice initializations on prediction skills of the Antarctic sea ice variability on a decadal timescale using a CGCM. We performed three decadal reforecast experiments (see Methods); the first experiment is the control (CTL) experiment in which the model’s SST is initialized with the observed SST. The second experiment is the sea ice restoring (SIR) experiment in which the model’s SST and SIC are initialized with the observations. Furthermore, we initialized the model’s subsurface ocean temperature and salinity using a three-dimensional variational (3DVAR) data assimilation approach in the 3DVAR experiment. We compare these experiments to identify potential sources of prediction skills for the Antarctic sea ice variability. More details on the datasets and the model experiments are given in the Methods section.

Results

Observed and simulated sea ice variability in the Antarctic seas

Understanding the initial state of the decadal reforecast experiments is essential for discussing the sea ice prediction skills37. Annual mean of the observed SIC during 1982–2018 (Supplementary Fig. 1a) shows higher SIC in the west Antarctic seas than in east Antarctic seas. Annual mean SIC from the initial condition of the CTL experiment (Supplementary Fig. 1b) underestimates the observed SIC. However, annual mean SICs in the SIR and 3DVAR experiments become higher in the west Antarctic seas and get closer to the observation. Since only small differences are found between the SIR and 3DVAR experiments, the improvement of sea ice mean state is mostly due to the SIC initialization.

Standard deviation of the 5-yr mean SIC anomalies observed during 1982–2018 (Supplementary Fig. 2a) shows large variability near the sea ice edges in the west Antarctic seas. Here we used 5-yr mean SIC anomalies to extract sea ice variability with a low frequency beyond a decade. Standard deviation of the 5-yr mean SIC anomalies from the initial conditions of the CTL experiment (Supplementary Fig. 2b) is smaller than that from the observation. The amplitude becomes larger in the SIR and 3DVAR experiments (Supplementary Fig. 2c, d) closer to the observation, but the spatial patterns are similar between the SIR and 3DVAR experiments. Therefore, SIC initialization plays an important role in better representation of decadal sea ice variability in the model experiments.

Decadal sea ice predictability in the Amundsen–Bellingshausen Sea

To quantitatively assess prediction skills of the decadal sea ice variability, we calculated anomaly correlation coefficient (ACC; see Methods) between the observed SIC anomalies and the corresponding SIC anomalies predicted ahead of 1–5 and 6–10 years from the decadal reforecast experiments. The SINTEX-F2 model skillfully predicts the observed decadal sea ice variability in the Amundsen and Bellingshausen Sea when the model’s ocean and sea ice are initialized (Fig. 1). Both experiments with the SST initialization (i.e., CTL experiment) and the SST and SIC initializations (i.e., SIR experiment) show almost no ACC of the 5-yr mean SIC anomalies in the Pacific sector for the lead times of 1–5 years and 6–10 years (Fig. 1a, b, d, e). However, the other experiment with SST, SIC and subsurface ocean initializations (i.e., 3DVAR experiment) shows significantly high ACCs of the 5-yr mean SIC anomalies in the Ross Sea and Amundsen–Bellingshausen Sea for the lead times of 1–5 years (Fig. 1c). The high ACCs are also found in the Amundsen–Bellingshausen Sea for the lead times of 6–10 years (Fig. 1f). This indicates that the subsurface ocean initialization plays an important role in skillful prediction of the decadal sea ice variability in the Amundsen–Bellingshausen Sea. We also note that the subsurface ocean initialization deteriorates the prediction skills in the Atlantic sector, which will be discussed in the following subsection.

Fig. 1: Prediction skills of decadal sea ice concentration (SIC) variability in the Antarctic seas.
figure 1

a Anomaly correlation (ACC) of the 5-yr mean SIC anomalies between the observation and the CTL experiment for the lead times of 1–5 years. Black dots indicate positive ACCs which are statistically significant at 90% confidence level using Student’s t test. b, c Same as in a, but for the ACC from the SIR and 3DVAR experiments. Curved boxes indicate the Weddell Sea (Wed; 60°−0°W, 65°−55°S) and the Amundsen–Bellingshausen Sea (120°−60°W, 70°−60°S) with significantly high ACC. d–f Same as in a–c, but for the ACC from the CTL, SIR, and 3DVAR experiments for the lead times of 6–10 years.

Full size image

To examine how the prediction skills of the sea ice variability evolve on an interannual timescale, we defined the sea ice index as the SIC anomalies averaged in the Amundsen–Bellingshausen Sea (AmB; 120°−60°W, 70°−60°S). The ACC of yearly mean SIC anomalies for the persistence prediction based on the observed SIC anomalies (Fig. 2a) shows a rapid decay of the positive ACC from around 0.4 at one year lead and fluctuates between positive and negative values under 0.3 afterwards. We find a similar tendency for the ACC of the CTL experiment with the SST initialization, although it shows slightly high ACCs above 0.3 at lead times of 7 and 10 years. On the other hand, the ACC of the SIR experiment with the SST and SIC initialization shows higher values than that of the CTL experiment except for the lead times of 9 and 10 years. The ACC of the 3DVAR experiment with SST, SIC and subsurface ocean initializations also shows higher values than that of the SIR experiment for the lead times of 4–6 and 9 years. These results indicate that the model with sea ice and/or ocean initializations can skillfully predict the multi-year sea ice variability up to 6–7 years in advance.

Fig. 2: Prediction skills of decadal sea ice concentration (SIC) variability in the Amundsen–Bellingshausen (AmB) Sea.
figure 2

a Time evolution of the ACCs of the yearly mean SIC anomalies averaged in the Amundsen–Bellingshausen Sea (AmB; 120°−60°W, 70°−60°S) between the observation and the reforecast experiments for each lead year. A black line indicates autocorrelation with the observed SIC anomalies, while red, blue, and green lines show the ACCs from the CTL, SIR, and 3DVAR experiments, respectively. Open circles indicate positive ACCs which are statistically significant at 90% confidence level using Student’s t test. b Time series of the yearly mean SIC anomalies averaged in the AmB region. We calculated the anomalies with respect to the period of 1982–2018. A black line shows the observed SIC anomalies, while colored lines show the SIC anomalies predicted from March 1st 1998 in the CTL (red), SIR (blue), an 3DVAR (green) experiments. Thick colored lines indicate the average of 12 ensemble members with different initial conditions, and dotted colored lines correspond to each ensemble member. c Same as in b, but for the SIC anomalies (in %) predicted from March 1st 2008.

Full size image

The observed SIC anomalies in the AmB region (Fig. 2b) have two high sea ice periods before 1996 and after 2009 and one low sea ice period between 1997 and 2008. During the low sea ice period, both the CTL and SIR experiments predicted from March 1st 1998 show gradual increase of the SIC anomalies which turn into positive SIC anomalies after 2002. On the other hand, the 3DVAR experiment predicts negative SIC anomalies until 2007 except for the slightly positive SIC anomalies in 2003. Compared to the other two experiments, the 3DVAR experiment reasonably captures the low sea ice period observed between 1998 and 2007. Similarly, for the high sea ice period after 2009 (Fig. 2c), the 3DVAR experiment shows larger positive SIC anomalies than the CTL and SIR experiments, which bring 3DVAR experiment closer to the observed SIC anomalies. As such, the 3DVAR experiment best reproduces the amplitude of the decadal sea ice anomalies in the AmB region.

Considering a gradual increase in the subsurface ocean observations below sea ice owing to the implementation of Argo floats from 2000s to 2010s (Supplementary Fig. 3a–c), we further investigate physical processes underlying the skillful prediction focusing on a high sea ice period after 2009. The 10-yr mean of the observed SIC anomalies during 2008–2017 (Fig. 3a) shows significant sea ice increase in the AmB region as well as the other Antarctic seas. Both the CTL and SIR experiments (Fig. 3b, c) capture the positive SIC anomalies in the AmB region, but their amplitudes are much weaker than the observation, as expected from the time series analysis (Fig. 2c). On the other hand, the 3DVAR experiment (Fig. 3d) shows larger amplitudes of the positive SIC anomalies, getting closer to that in the observation (Fig. 3a).

Fig. 3: Sea ice concentration (SIC) anomalies during a high sea ice period (2008–2017).
figure 3

a 10-yr mean SIC anomalies (color in %) observed during 2008–2017. Black boxes indicate the Amundsen–Bellingshausen Sea (120°−60°W, 70°−60°S) and Weddell Sea (60°−0°W, 65°−55°S). We calculated the anomalies with respect to the period of 1982–2018. b–d Same as in a, but for the 10-yr mean SIC anomalies (in %) predicted from March 1st 2008 in the CTL, SIR, and 3DVAR experiments.

Full size image

We find that the positive SIC anomalies in the AmB region are associated with deepening of the Amundsen Sea Low. 10-yr mean SLP anomalies from the atmospheric reanalysis (Fig. 4a) show an anomalous cyclone over the eastern Ross Sea and the AmB region. The associated southerly anomalies tend to advect cold air from Antarctica and increase sea ice extent in the Ross Sea10 (Fig. 3a), whereas the northerly anomalies act to reduce the sea ice in the AmB region by bringing warm air from the north. However, the 2-m air temperature anomalies in the AmB region show negative values (Fig. 4a). This suggests that other processes (e.g., ocean and sea ice) may play roles in inducing the sea ice increase over the AmB region (Fig. 3a). In fact, the CTL experiment (Fig. 4b) does not well capture this atmospheric circulation variability, but the SIR experiment (Fig. 4c) reproduces the stronger Amundsen Sea Low. The anomalous cyclone is slightly located west compared to the atmospheric reanalysis (Fig. 4a), so that the AmB region receives more warm air from the north. The 3DVAR experiment (Fig. 4d) also reproduces the northerly anomalies associated with the deepened Amundsen Sea Low.

Fig. 4: Atmospheric circulation and temperature anomalies during a high sea ice period (2008–2017).
figure 4

a 10-yr mean sea level pressure (SLP; contour interval of 0.1 hPa), zonal (U10; in m s−1) and meridional (V10; in m s−1) winds at 10 m, and 2-m air temperature (T2m; color in °C) anomalies during 2008–2017 from the ERA5 atmospheric reanalysis. Purple boxes indicate the Amundsen–Bellingshausen Sea (120°−60°W, 70°−60°S) and Weddell Sea (60°−0°W, 65°−55°S). We calculated the anomalies with respect to the period of 1982–2018. b–d Same as in a, but for the 10-yr mean SLP, U10, V10, and T2m anomalies predicted from March 1st 2008 in the CTL, SIR, and 3DVAR experiments.

Full size image

One may wonder if surface heat flux anomalies may also contribute to the sea ice increase in the AmB region. The 10-yr mean net surface heat flux anomalies from the atmospheric reanalysis (Supplementary Fig. 4a) show negative values in the AmB region. This means that more heat is released from the ocean, contributing to surface cooling and increase in sea ice there. However, all of the model experiments (Supplementary Fig. 4b–d), especially 3DVAR experiment, show some positive anomalies near the Antarctica, so other processes such as ocean and sea ice advections may dominate in the model results.

To investigate possible roles of ocean and sea ice advections in decadal sea ice variability over the AmB region, we assessed yearly mean sea ice volume (SIV) tendency anomalies and its components averaged in the AmB region using Eq. (1) (see Methods) for the 3DVAR experiment. The SIV tendency anomalies show positive values on a decadal average from 2008 (Fig. 5a). This is mostly due to positive contributions from the zonal convergence term and residual term in Eq. (1). Since the sea ice zonal velocity anomalies (Fig. 5b) show positive values over a decadal average, the positive anomalies of zonal sea ice convergence (Fig. 5a) are mostly due to anomalous sea ice advection from the west of the AmB region, i.e., the Ross Sea. On the other hand, there is a positive contribution from the residual term that consists of sea ice growth/melt associated with atmosphere and ocean heat fluxes in the vertical. Since the net surface heat flux anomalies in the 3DVAR experiment (Supplementary Fig. 4d) are positive to contribute to warm the ocean and sea ice surface, subsurface ocean processes should play a role in the sea ice increase. In fact, ocean temperature in the mixed layer below the sea ice (Fig. 5c) shows negative anomalies, consistent with the sea ice increase there. The mixed-layer temperature anomalies may also have links with the subsurface ocean temperature anomalies, because subsurface ocean temperature below the mixed layer shows larger negative anomalies from the initial year of 2008.

Fig. 5: Atmospheric circulation and temperature anomalies during a high sea ice period (2008–2017).
figure 5

a Time series of sea ice volume (SIV) tendency anomalies (Total; black in 10−11 m3 s−1) and its zonal (Znl; red), meridional (Mrd; blue) and residual (Rsd; green) components (see Eq. (1) in Methods) averaged in the Amundsen–Bellingshausen Sea (AmB; 120°−60°W, 70°−60°S) in the 3DVAR experiment predicted from March 1st 2008. We calculated the anomalies with respect to the period of 1982–2018. b Same as in a, but for sea ice zonal velocity (Ice U; in 10−2 m s−1) anomalies in the AmB region. c Time series of the upper 700-m ocean temperature (T; in °C) anomalies predicted from March 1st 2008 in the 3DVAR experiment. A black line indicates the mixed layer depth defined by the depth at which the potential density increases by 0.03 kg m−3 from the ocean surface. d Same as in c, but for the upper 700-m zonal current (U; in 10−2 m s−1) anomalies as a function of depth. Positive values indicate anomalous eastward current.

Full size image

The negative temperature anomalies are associated with positive zonal current anomalies in the AmB region (Fig. 5d). This indicates that the Antarctic Circumpolar Current becomes stronger during the high sea ice period. Since the Antarctic Circumpolar Current flows from the northwest to southeast in the AmB region, anomalous advection of relatively cold water from the Ross Sea contributes to anomalous cooling of the subsurface ocean temperature (Fig. 5c). We find that meridional current anomalies (Supplementary Fig. 5a) are negative to warm the subsurface ocean, while the vertical current anomalies (Supplementary Fig. 5b) are positive on a decadal average to warm the mixed layer through upwelling of warm water from the subsurface ocean. None of these ocean current anomalies can explain colder ocean below sea ice. Therefore, we conclude that the strengthened Antarctic Circumpolar Current in the 3DVAR experiment is mostly responsible for the skillful prediction of decadal sea ice variability in the AmB region.

Decadal sea ice predictability in the Weddell Sea

Prediction skills of 5-yr mean SIC anomalies in the northern Weddell Sea (60°−0°W, 65°−55°S) for lead times of 1–5 and 6–10 years are high in all the model experiments (Fig. 1). The high ACCs in the CTL experiment (Fig. 1a, d) indicate that some of the sea ice prediction skills in the northern Weddell Sea benefit from the SST initialization. The SIR experiment (Fig. 1b, e) shows higher ACC than the CTL experiment, indicating that the sea ice initialization adds positive values to prediction skills in the northern Weddell Sea. However, the prediction skills in the 3DVAR experiment (Fig. 1c, f) become lower than those in the SIR experiment. The initialization of subsurface ocean may be insufficient due to a smaller number of subsurface ocean observations in the Weddell Sea than the Pacific sector (Supplementary Fig. 3a–d), which may affect the quality of the initial ocean state in the model experiments.

In fact, the objectively analyzed ocean temperature anomalies in the upper 700 m of the northern Weddell Sea show slightly negative values before 1990, positive values around 1995, then negative values after 2000 (Supplementary Fig. 6a). All of the model experiments during the initialization (Supplementary Fig. 6b–d) show negative temperature anomalies before 1990 and positive anomalies around 1995. However, the CTL experiment (Supplementary Fig. 6b) fails in reproducing the negative anomalies after 2000. Although the 3DVAR experiment (Supplementary Fig. 6d) shows large positive anomalies around 2000, it captures negative temperature anomalies after 2005, consistent with the objective analysis, when the subsurface ocean observations substantially increase owing to the Argo floats (Supplementary Fig. 3d). These unrealistic simulations of ocean temperature anomalies in the CTL and 3DVAR experiments during the initialization may deteriorate the prediction skills in the northern Weddell Sea, as compared to the SIR experiment.

To examine how far and skillfully these model experiments predict the year-to-year sea ice variability in the Weddell Sea, we calculated ACC of yearly mean SIC anomalies averaged in the northern Weddell Sea (Fig. 6). The ACC for the persistence prediction based on the observed SIC anomalies (Fig. 6a) shows a rapid decay of prediction skills from lead year 1 onward. The CTL experiment with the SST initialization outperforms the persistence prediction in the first two lead years. The prediction skills decay rapidly for the lead times of 3–6 years, but gradually re-emerge after 7 year lead, reaching a significant peak at 9 year lead. On the other hand, the SIR experiment shows higher prediction skills in the first four lead years than the CTL experiment. After a few years decay, the prediction skills re-emerge to become significantly high in 8–9 year lead. We do not find such high prediction skills in the 3DVAR experiment except for 1–2 year lead. Therefore, the SIR experiment can skillfully predict the year-to-year sea ice variability up to 8–9 years in advance, which contributes to significantly high ACC at lead times of 6–10 years (Fig. 1e).

Fig. 6: Prediction skills of decadal sea ice concentration (SIC) variability in the Weddell (Wed) Sea.
figure 6

a Same as in Fig. 2a, but for the ACC of the SIC anomalies in the Weddell Sea (Wed; 60°−0°W, 65°−55°S) as a function of lead years. b Same as in Fig. 2b, but for the time series of the SIC anomalies in the Wed region. A black line shows the observed SIC anomalies, while colored lines show the SIC anomalies predicted from March 1st 1999 in the CTL (red), SIR (blue), an 3DVAR (green) experiments. We calculated the anomalies with respect to the period of 1982–2018. c Same as in b, but for the SIC anomalies (in %) predicted from March 1st 2008.

Full size image

Time series of the observed SIC anomalies in the northern Weddell Sea show two high sea ice periods before 1997 and after 2007 and one low sea ice period between 1998 and 2006 (Fig. 6b). Both the CTL and 3DVAR predictions from March 1st 1999 show rapid increases of the SIC anomalies which turn into the positive anomalies after 2003 and 2004. However, the SIR experiment shows a gradual decay of the negative SIC anomalies, closer to the observed SIC anomalies in the low sea ice period. We find a similar but opposite tendency for the high sea ice period after 2008 (Fig. 6c). On a decadal average, the SIR experiment shows higher amplitude of the positive SIC anomalies than the CTL and 3DVAR experiments.

The observed SIC anomalies averaged over the low sea ice period of 2008–2017 show negative values in the entire Weddell Sea (Fig. 3a). All of the model experiments capture the negative SIC anomalies, but the amplitude of the negative SIC anomalies in the northern Weddell Sea are larger in the SIR experiment (Fig. 3c) than in the CTL and 3DVAR experiments (Fig. 3b, d). This is well consistent with earlier results from the time series analysis (Fig. 6c). The atmospheric circulation anomalies from the atmospheric reanalysis (Fig. 4a) show eastward extension of the deepened Amundsen Sea Low which induces northwesterly wind anomalies in the northern Weddell Sea. This advects more warm air from the north and play a counteractive role for the sea ice increase there. The CTL and 3DVAR experiments (Fig. 4b, d) do not capture the northwesterly wind anomalies, but the SIR experiment (Fig. 4c) well reproduces the eastward extension of the low-pressure anomaly and the associated northwesterly wind anomalies. Furthermore, the net surface heat flux anomalies from the atmospheric reanalysis (Supplementary Fig. 4a) and all the model experiments (Supplementary Fig. 4b–d) show positive values in the northern Weddell Sea, preventing the sea ice increase there. Therefore, the increase of sea ice in the northern Weddell Sea appears to have links with ocean and sea ice advection processes rather than the atmospheric variability22.

During the high sea ice period after 2008, tendency anomalies of annual mean SIV in this region show large interannual variations (Fig. 7a), but on a decadal average, the anomalies show positive values to increase sea ice. Both the zonal and meridional sea ice convergence anomalies (Fig. 7a) show an in-phase relationship with the SIV tendency anomalies, while the residual term including the atmospheric and ocean heat fluxes is out of phase and counteracts to the sea ice increase. Zonal and meridional sea ice velocity anomalies (Fig. 7b) show positive values on a decadal average, indicating more sea ice advection from the southwestern Weddell Sea where large sea ice volume exists. The 10-yr mean ocean currents and density anomalies in the upper 50 m (Fig. 7c) show anomalous clockwise circulation in the southern Weddell Sea, which is associated with anomalous increase of ocean density in the middle of Weddell Sea. The density increase is mostly due to an increase of salinity, probably linked to anomalous evaporation from ocean surface during the low sea ice period before 2007. This indicates that the Weddell Gyre anomalously strengthens to export more sea ice northeastward in the basin. Therefore, accurate simulation of the upper ocean and sea ice interaction is crucial for skillful prediction of decadal sea ice variability in the northern Weddell Sea22.

Fig. 7: Sea ice volume (SIV) tendency anomalies and near-surface ocean conditions in the Weddell (Wed) Sea.
figure 7

a Time series of sea ice volume (SIV) tendency anomalies (Total; black in 10−11 m3 s−1) and its zonal (Znl; red), meridional (Mrd; blue) and residual (Rsd; green) components (see Eq. (1) in Methods) in the Weddell (Wed; 60°−0°W, 65°−55°S) Sea for the SIR experiment predicted from March 1st 2008. We calculated the anomalies with respect to the period of 1982–2018. b Same as in a, but for sea ice zonal and meridional current velocity (Ice U and V; in 10−2 m s−1) anomalies in the Wed region. c 10-yr mean of the upper 50-m potential density, zonal and meridional current velocity anomalies predicted from March 1st 2008 in the SIR experiment. A black box indicates the Wed region. d Same as in c, but for the upper 50-m ocean salinity (S; in 10−1 PSU) anomalies.

Full size image

Discussion and conclusions

This study has revealed the impacts of ocean and sea ice initializations on the prediction skills of decadal sea ice variability in the west Antarctic seas using a CGCM. Initialization of the model’s subsurface ocean temperature and salinity with observations improves prediction skills of decadal sea ice variability mostly in the Amundsen–Bellingshausen Sea. This allows skillful prediction of the year-to-year sea ice variability up to 6 years, which are twice as long as the lead times of skillful prediction for the pan-Antarctic sea ice variability35 and the eastern Pacific sea ice variability36. The significant improvement of sea ice prediction in the Amundsen–Bellingshausen Sea arises from a better representation of subsurface ocean circulation variability, involving the decadal variability in the Antarctic Circumpolar Current. Since the poleward shift of the Antarctic Circumpolar Current is observed during 1992–2011 in the eastern Indian sector38,39, decadal changes in the intensity of the Antarctic Circumpolar Current in the Ross Sea can influence the subsurface ocean temperature variability in the Amundsen–Bellingshausen Sea through zonal advection of cold water from the west.

However, the subsurface ocean initialization does not much improve the sea ice prediction skills in the northern Weddell Sea. This is partly due to a smaller number of subsurface ocean observations as compared to the Pacific sector (Supplementary Fig. 3a–d), leading to an inaccurate representation of the initial ocean temperature anomalies in the model experiments (Supplementary Fig. 6d). Nevertheless, the initialization of the SIC is found to be most effective for skillful prediction of the decadal sea ice variability in the northern Weddell Sea. The sea ice initialization affects the intensity of the Weddell Gyre through changes in surface evaporation and hence ocean salinity and density22. The horizontal advection of sea ice anomalies through changes in the Weddell Gyre is responsible for generating the decadal sea ice variability in the Weddell Sea.

This study provides evidence that ocean and sea ice initializations improve prediction skills of decadal sea ice variability in the west Antarctic seas. However, the prediction skills remain low in the Ross Sea and the east Antarctic seas for the lead times of 6–10 years (Fig. 1d–f). Decadal sea ice variability in these seas is more influenced by the forced atmospheric variability of the IPO10 or decadal modulation of ENSO40 and by the intrinsic atmospheric variability of the SAM40. We need to further improve model simulation of the atmospheric variability, for example, with the inclusion of atmospheric initialization schemes which are not adopted in this study. Furthermore, a recent study34 has reported that the sea ice prediction during austral summer significantly improves with the sea ice thickness (SIT) initialization during austral winter, as compared to the sea ice prediction with the SIC initialization only. Since the SIT has much longer memory than the SIC, both initializations of the SIC and SIT may help improve skillful sea ice prediction on a decadal timescale. None of the CMIP5 models36 have so far adopted the SIT initialization scheme. Therefore, decadal sea ice prediction with the SIT initialization is one of the promising ways forward to further improve the sea ice prediction skills. Further research along this line is now underway.

Methods

Observation data and reanalysis product

We used monthly SIC from the OISSTv2 observation data41. The SIC data has a horizontal resolution of 1° × 1° and covers the period of 1982–2019. We employed monthly reanalysis product of sea level pressure (SLP), air temperature at 2 m, and zonal and meridional winds at 10 m from the ERA542. To have a consistency with the ocean dataset, we adopted the atmospheric product interpolated on the same horizontal resolution of 1° × 1° during the same analysis period. For all of these datasets, we removed monthly climatology and a linear trend using the least squares method to calculate monthly anomalies.

Coupled general circulation model (CGCM)

To perform decadal reforecast experiments, we employed a CGCM called the SINTEX-F243,44, developed at JAMSTEC under the international research collaboration between EU and Japan. The SINTEX-F2 consists of the ECHAM5 atmosphere model45, the NEMO3 ocean model46 including the LIM2 sea ice model47, and the OASIS3 coupler48. The atmospheric model has a horizontal resolution of T106 (around 1.1° × 1.1°) and 31 vertical levels based on a hybrid sigma-pressure coordinate. The ocean and sea ice models have a horizontal resolution of 0.5° (e.g., 50 km horizontal grid box width at 40°S) on a tripolar global grid (ORCA0549), and 31 vertical levels with 10 layers in the first 100 m. The atmosphere and ocean/sea ice are coupled by means of the OASIS3 coupler without any flux correction.

We first initialized the oceanic component of the SINTEX-F2 model with the mean temperature and salinity of World Ocean Database50, starting from the rest state. Then, we spun up the SINTEX-F2 model with monthly climatology of the observed SST from 1950 to 1981. After that, we conducted three decadal reforecast experiments with/without ocean and sea ice initializations. In the control (CTL) experiment, we initialized the model’s SST with the observed SST until the end of every February, then freely integrated the model over 10 years from every March 1st of 1982–2009. This generates model output covering the period from 1982 to 2018. To estimate the model uncertainties arising from the initial conditions, we generated 12 ensemble members by using two SST data (OISSTv241 and its high-resolution data51), three negative feedback parameters for the SST initialization (−2400, −1200, and −800 W m−2 K−1) and two ocean vertical mixing schemes52. Here we switched on and off the vertical mixing scheme by elevating the background vertical mixing induced by the small vertical mixing from the surface down to the 20 °C isotherm52. Similarly, we conducted the sea ice restoring (SIR) experiment in which the model’s SST and SIC are initialized with the observed SST and SIC data until the end of every February, then the model is freely integrated over 10 years from every March 1st of 1982–2009. Here we adopted the relaxation timescale of 5 days to initialize the model’s SIC, following a previous study using the SINTEX-F2 model33. Furthermore, we conducted a third experiment in which the model’s SST, SIC and subsurface ocean temperature and salinity are initialized until the end of every February using a three-dimensional variational (3DVAR53,54) data assimilation approach, then freely integrated over 10 years from every March 1st of 1982–2009 (3DVAR experiment). During the initialization, we used the EN4 subsurface ocean profiles55 and corrected the model’s subsurface ocean temperature and salinity with a 10-day assimilation window at the end of every month. For all of these three experiments, we calculated monthly anomalies by removing lead-time dependent monthly climatology and linear trend as the model bias (drift).

Anomaly correlation (ACC) skill score

For quantitative assessment of decadal sea ice prediction skills, we adopted the ACC skill score using the observed SIC anomalies and the ensemble mean SIC anomalies predicted at lead times of 1–10 years. For example, we calculated the ACC between the observed SIC anomalies and the corresponding SIC anomalies predicted ahead of 1–5 years and 6–10 years from decadal reforecast experiments (Fig. 1). To detect a low-frequency variability beyond a decade, we used the 5-yr mean SIC anomalies from each initialization year of 1982 to 2009 (i.e., 28 samples). To describe time evolution of prediction skills on an interannual timescale, we calculated the ACC between the observed SIC anomalies and the corresponding SIC anomalies predicted at each lead time of 1–10 years over the west Antarctic seas (Figs. 2a and 6a). We note that the latter assessment explains the model capability of predicting the year-to-year sea ice variability, while the former assessment describes prediction skills of the decadal sea ice variability which is commonly used in decadal prediction community56.

Sea ice volume (SIV) tendency

To investigate physical processes underlying decadal sea ice variability, we defined the SIV by multiplying the sea ice area (\(A\)) with the sea ice thickness (\(h\)) and evaluated the SIV tendency as below:

$$\frac{\partial {Ah}}{\partial t}=-\frac{\partial }{\partial x}\left(u\bullet {Ah}\right)-\frac{\partial }{\partial y}\left(v\bullet {Ah}\right)+{Res},$$
(1)

where \(u\) and \(v\) indicate zonal and meridional sea ice velocities, respectively. The first two terms in the right-hand side of Eq. (1) are zonal and meridional SIV convergence/divergence, while the last term in the Eq. (1) indicates the residual term including heat flux contributions from the atmosphere and the ocean and other sea ice internal dynamics.