Record-low Antarctic sea ice in 2023 increased ocean heat loss and storms

Data sources

The surface flux, meteorological variables and sea-ice concentration presented in our study are from the European Centre for Medium-Range Weather Forecasts Reanalysis 5 (ERA5 (ref. ²⁷)) reanalysis. The ERA5 sea-ice data are satellite-observation-derived fields from the EUMETSAT Ocean and Sea Ice Satellite Application Facility (OSI SAF) product²⁷. Ocean surface salinity required for the water-mass-transformation calculation has been obtained from the UK Met Office EN.4.2.2 dataset²⁸.

Determination of box means and anomalies

Area-weighted means of various fields (sea-ice concentration, net heat flux and its components) have been determined for each of the four regions shown in Fig. 1. The weighting is carried out according to the area of individual ERA5 grid cells. Box boundaries are as follows: nWS, northern Weddell Sea (66° S to 59° S, 50° W to 20° E); BS, Bellingshausen Sea (75° S to 65° S, 85° W to 65° W); nwRS, northwestern Ross Sea (68° S to 62.5° S, 152° E to 175° E); EL, north of Enderby Land (66° S to 62.5° S, 40° E to 62° E). The combined four-box means are an average of the four individual box means weighted according to the area of each box. For ice concentration terminology, note, for example, that by use of the term 80% reduction in the main text, we mean a change in the fractional ice concentration of 0.8. All anomalies are determined with respect to the 30-year climatological reference period 1991–2020. Sea-ice-area values reported in the main text have been determined by multiplying the ERA5 sea-ice concentration (which varies in the range 0–100%) in a given grid cell by the grid-cell area and summing over all grid cells within the four boxed regions. For heat flux, the sign convention is for positive/negative heat flux to indicate ocean heat gain/loss to the atmosphere.

Water-mass transformation and formation

Following standard water-mass-transformation theory²⁹, the ERA5 evaporation (E), precipitation (P) and net heat flux into the ocean (Q) are used to estimate the surface-forced transformation (SFT) across an isopycnal, σ,

in which

α is the thermal expansion coefficient, β is the haline contraction coefficient, C_P is the specific heat capacity at 5 m and S is the 5 m salinity. For each month and each isopycnal, σ, the local buoyancy flux (term in square brackets) is integrated over the surface area of the associated density bin, Δσ. Thus, the SFT has a non-zero value for those months when the specified isopycnals outcrop, otherwise it is set to zero. In practice, the calculation is carried out using a discretized version of the above equation by summing over ERA5 grid cells within a particular density interval with a density bin size of Δσ = 0.1 kg m⁻³. Formation is then determined as the difference in transformation values between the upper and lower limits of the density interval.

Robust pattern of heat loss across reanalyses

The spatial patterns of JJ23 net heat-flux anomaly within the sea-ice zone determined from our primary reanalysis (ERA5) and the other leading, regularly updated reanalysis (MERRA-2 (ref. ³⁰)) are shown in Extended Data Fig. 1a,b. The figure highlights the close similarity in the pattern of net heat-flux anomaly in both reanalyses, with strong heat losses in the boxed regions in each case and slightly higher values with MERRA-2. Thus, the pattern is robust to the choice of reanalysis, indicating that the intensity of the increased heat loss in JJ23 dominates any variations that may arise from differences in reanalysis physics and data assimilation. This is important to establish, as reanalyses can vary in their representation of air–sea interaction in the Southern Ocean, particularly the climatological mean heat exchange.

An evaluation of ERA5 against three drifting-buoy measurements in the Weddell Sea ice pack³¹ indicates that the reanalysis air temperature is close to the observations at 0 °C but develops a warm bias as temperatures decrease towards −40 °C. Such a bias would lead to the ERA5 heat-loss values underestimating the true heat loss and may account in part for the stronger losses seen with MERRA-2, but the close pattern agreement between the two reanalyses indicates that it does not strongly influence our conclusions. Note that MERRA-2 was not included in the buoy comparison study, so it is not possible to say whether it is more accurate than ERA5, as it may have a cold bias relative to the buoys and hence overestimate the true heat loss.

As well as ERA5 and MERRA-2, we have also examined the JJ23 net heat-flux anomaly with the NCEP/NCAR reanalysis³², which is another regularly updated reanalysis (Extended Data Fig. 1c). NCEP/NCAR is at a substantially coarser resolution than ERA5 and MERRA-2, as it was developed in the 1990s. Nevertheless, it still shows the main characteristics of the pattern found with the other two reanalyses, that is, increased heat loss in the main ice-decline regions, providing further support that this pattern is robust to the choice of reanalysis. Given the low resolution of the NCEP/NCAR fields, we subsequently focus on comparison of MERRA-2 results with ERA5.

Variation of heat loss with sea ice

The variation of June–July air–sea heat flux with sea-ice concentration is shown in Extended Data Fig. 2. A clear relationship showing increasing heat loss with ice decline is evident and JJ23 is seen to fall a long way outside the group of all other points, emphasizing the exceptionally strong heat loss and low ice fraction in winter 2023. Typical values for the earlier years are sea-ice fraction in the range 0.6–0.8 and heat loss from −75 to −50 W m⁻², whereas JJ23 has an ice fraction of only 0.4 and increased heat loss at −90 W m⁻². Although marked ice reductions are known to have occurred within the 2016–2022 period, they were mostly in the austral summer sea-ice extent (for example, refs. ^3,9,17) rather than austral winter. The JJ1622 values are seen to be shifted towards the lower end of the distribution for JJ9015 but are not clearly separated from them, unlike JJ23.

Climatological and JJ23 heat flux

To aid interpretation of the JJ23 net heat-flux-anomaly map shown in Fig. 1b, we present mean heat-flux maps for both the climatological case (1991–2020) and for 2023 (Extended Data Fig. 3). The two maps show that, in both cases, the June–July net heat flux ranges from near zero to strongly negative (heat loss approaching −200 W m⁻²) over the sea-ice region, with strongest losses in the outer half towards the ice edge. Comparison of the maps reveals stronger heat loss in the boxed regions in 2023 compared with climatology. Consideration of Fig. 1b reveals that the negative heat-loss anomaly in the Ross Sea box is different in character to the other three boxes, as it is accompanied by a narrow band of positive heat-flux anomaly along the outer ice edge to the northeast of the box. This positive anomaly can be seen through comparison of Fig. 1b and Extended Data Fig. 3 to reflect a reduction in the normally strong heat loss in this band. For the wider Ross Sea region, the positive and negative anomalies in Fig. 1b have the potential to cancel out, but this does not affect the conclusions reached for the northern Ross Sea box as defined in our analysis.

Water-mass formation across reanalyses

The main-text analysis of the surface flux contribution to water-mass formation in the pre-ice-decline and ice-decline periods determined from our primary reanalysis (ERA5) has been repeated using MERRA-2 (Extended Data Fig. 4). Comparison of Extended Data Fig. 4 with Fig. 4 shows the close similarity between the two sets of results, that is, the stronger high-density contribution to June–July water-mass formation in 2023 than both 1990–2015 and 2016–2022 is seen with both reanalyses. The only difference of note is in the Weddell Sea, in which the MERRA-2 JJ23 water-mass-formation anomalies are stronger than ERA5 (note the difference in y-axis range between Extended Data Fig. 4 and Fig. 4a), consistent with the increase in heat loss with MERRA-2 noted in the discussion of Extended Data Fig. 1. Otherwise, Extended Data Fig. 4 is virtually indistinguishable from Fig. 4. This demonstrates the robustness to variations in reanalysis physics and data assimilation of our result that the surface flux contribution to JJ23 water-mass formation has undergone a notable shift to higher density classes. Further insights into changes in the ice-decline regions, including the processes underlying water-mass formation and the balance of changes in ice cover and heat flux in setting the temperature and salinity fields, are anticipated from analysis of ocean-state estimates for 2023 when they become available.

Increase in storminess across reanalyses

The increase in storminess in June–July of 2023 relative to 1990–2015 obtained from ERA5 and reported in the main text (Fig. 5 and associated discussion) has been recalculated using the MERRA-2 reanalysis (Extended Data Fig. 5c,d). Comparison of the ERA5 and MERRA-2 fields for both the JJ23 storminess and the JJ23–JJ9015 storminess difference shows very similar results. In both cases, the June–July storm frequency has increased by up to 7 days per month in 2023 in the sea-ice-decline regions. Thus, our conclusions about the storminess increase are not sensitive to differences in reanalysis physics and data assimilation. Note also that the increased storminess is not confined to only the high sea-ice-decline areas, in particular, there are strong increases to the east of the Ross Sea box. This may indicate broader-scale impacts of the sea-ice decline on the atmospheric circulation that warrant further analysis using atmospheric-model experiments.

The robustness of our choice of primary wind speed metric (number of days above a fixed 10 m s⁻¹ threshold) has been assessed using an alternative index, the 90th percentile of wind speed determined on a grid cell by grid cell basis. The results shown in Extended Data Fig. 5e,f are very similar to those obtained with the fixed threshold. Thus, both the 10 m s⁻¹ threshold and the 90th percentile wind-speed-based metrics indicate an increase in storminess. These metrics-based results motivate further analyses in subsequent research, in particular, the use of storm-tracking algorithms to investigate impacts on trajectories as well as frequency and coupled-model analysis to study the atmospheric response in detail.

Lag between heat loss and storms

We have carried out further analysis of the relationship between heat loss and storminess using daily time series of the net turbulent heat flux within the JJ23 period (Extended Data Fig. 6). For the analysis, a sub-region (65° S to 59° S, 25° W to 0° E) of the northern Weddell Sea has been selected (Extended Data Fig. 6a), which has a strong increase in storm number. The 95% significance level contour encloses much of the sub-region, that is, there is robust coverage of the sub-region by increased storm number. Cumulative turbulent heat flux and storm occurrence time series for 1–10 June 2023 are shown in Extended Data Fig. 6b, with an extended version for the whole of JJ23 in Extended Data Fig. 6c. A delayed increase in storm number with respect to the increase in heat flux on a timescale of approximately 3 days is evident in Extended Data Fig. 6b. This result is consistent with a scenario in which the increased heat supply to the atmosphere results in an increase in storm frequency. By the end of July, the number of storm days is close to 30, nearly 50% higher than the value of 20 seen for climatology (Extended Data Fig. 6c).

The analysis has been repeated for the whole northern Weddell Sea box (largest box in Fig. 1b), which contains regions of both strong and weak heat-loss increase. In this case, a longer delay in storm increase is observed for the whole box of about 20 days (Extended Data Fig. 7). This is to be expected, as the rapid 3-day signal seen in the sub-region will be extended by the inclusion of the remainder of the region, which includes areas of weaker increase in heat loss that are likely to take longer to experience a storm response or may not respond at all.

We have repeated the northern Weddell Sea analysis for the northern Ross Sea box, shown in Fig. 5 and reproduced at larger scale in Extended Data Fig. 8. This is another region of strong increase in the storminess index in JJ23 (Extended Data Fig. 8a). When averaged over the box, the time series again show a rapid response of storminess that, in the northern Ross Sea, lags the heat-flux increase by 4–5 days (Extended Data Fig. 8b,c).

Further work using model-based experiments to explain the underlying causal processes responsible for the lagged response of storminess to increasing ocean heat loss is desirable. Such a study would require careful consideration of model experiment design, including whether a coupled ocean–atmosphere model is necessary rather than a forced atmospheric model and the need to carry out a sufficiently large ensemble of model control and perturbed simulations to be able to draw robust conclusions. We would hope to see such a study stimulated by the results of our paper

In summary, we find evidence for a short-timescale (3–5-day lag) increase in storm number following the increase in heat loss in both the northern Weddell Sea sub-region and the northern Ross Sea (Extended Data Figs. 6b and 8b). A rapid response on this timescale is consistent with a causal relationship in which the heat loss drives the storm increase. Note that this result for within-season variation stands in contrast to drivers of multidecadal trends in which variations in the wind field and storminess have been found to control the sea-ice extent^33,34 and so, potentially, the heat loss.

Variation of M and storms with heat loss

The variation of June–July water-mass formation (M) at densities >27.6 kg m⁻³ summed over the four ice-decline regions with air–sea heat flux is shown in Extended Data Fig. 9a. The amount of water formed in this density range in JJ23 (4.1 Sv) is nearly twice as large as that found in the strongest of the earlier years (2.2 Sv), indicating the exceptional water-mass-formation conditions that took place in 2023. In many of the early years, little or no water was formed at densities >27.6 kg m⁻³ and, in some years, small negative values indicate a net lightening of water in this density range, rather than the strong densification seen in 2023.

The variation of June–July number of storms averaged over the four ice-decline regions with air–sea heat flux is shown in Extended Data Fig. 9b. The number of storms in JJ23 exceeds that in June–July of all other years, although the separation is less clear cut than for dense water formation. The JJ23 value is 11.6 days compared with 9.1 ± 1.0 days for 1990–2015, that is, the 2023 value is significantly different at the 95% level.

Annual heat-loss anomalies

The main focus of our study is the extreme winter heat loss in JJ23, but we have also carried out some more analysis for the year as a whole. A map of the annual 2023 turbulent heat-flux anomalies is shown in Extended Data Fig. 10. The flux anomalies are typically less than 10 W m⁻², although stronger negative values in the range 20–30 W m⁻² are evident in the Bellingshausen Sea, northwestern Ross Sea and Enderby Land regions. The northern Weddell Sea tends to exhibit rather weak anomalies over the whole year, in contrast to the strong winter anomalies described in the main text, and we plan further investigation of the reasons for this difference in subsequent work. We have also calculated the annual mean values for the different curves in Fig. 3 and find that the 2023 annual mean value is not statistically different from climatology.