The CMIP5 version of EC-Earth simulates slightly too cold global mean sea surface temperatures (Sterl et al. 2012) and a cold bias of about 2 K in the Arctic for present day conditions (Koenigk et al. 2013). This cold bias is slightly less pronounced in version 3.0.1 used in the present study (Table 1). In our EXP2000, EXP2015 and EXP2030 simulations, the global mean sea surface temperature (SST) reaches, averaged over the last 80 years, 18.18, 18.32 and 18.55 °C, respectively (Table 1). This compares to a SST of 18.57 °C in HADISST and 18.51 °C in ERA-interim for the period 1980–2013.
Table 1 Mean values and standard deviations (cursive) for the years 21–100 of annual mean Arctic T2m, global mean SST, September and March Arctic ice extent and volume and annual mean max AMOC in EXP2000, EXP2015, EXP2030 and estimates for the recent past (1980–2013, 2004–2013 for RAPID-MOCCHA)
As for the global mean SST, Arctic mean 2 m air temperature (T2m) in EXP2030 (Fig. 1a) agrees best to the 1980–2013 period in the reanalysis data. However, due to sparse observations in the Arctic, uncertainties in reanalyses data are considerable as well (Jakobson et al. 2012). Jakobson et al. (2012) compared different reanalysis data sets in the Arctic and concluded that ERA-interim performs best. However, there is a tendency for a warm bias of locally up to 2 K in ERA-interim below 400 m.
The warming in the Arctic (Table 1) is amplified compared to the global mean T2m values (0.21 K in EXP2015 and 0.58 K in EXP2030) by a factor of about three and reaches 1.78 K in EXP2030. This compares to an observed Arctic warming of about 2 K since 1980. As for global mean values, the warming is stronger between EXP2030 and EXP2015 compared to the difference between EXP2015 and EXP2000. This agrees well with findings from Gregory et al. (2002) and Mahlstein and Knutti (2012).
All three simulations exhibit large decadal variations in a number of variables (Fig. 1). However, the mean changes of the variables in Fig. 1 are all significant at the 95 % significance level. To calculate the significance, we used a two-sided student t test and we calculated decorrelation-times to estimate the number of degrees of freedom (von Storch and Zwiers 1999).
Ocean
The spatial distribution of sea surface temperature (SST) and its changes in EXP2015 and EXP2030 compared to EXP2000 in mid and high northern latitudes are presented in Fig. 2a–c. The main discrepancy from observed present day SST is a cold bias of several K in the sub-polar gyre. This cold bias is a typical problem of coarse resolution ocean models (Large and Danabasoglu 2006; Eden et al. 2004) and is due to a North Atlantic Current that is too zonal and misplaced to the south (see discussion in Sterl et al. 2012). The SST change in EXP2015 is dominated by a strong cooling in an area south of Greenland. In the Labrador Sea and the northeastern North Atlantic, surface temperature rises significantly, with up to 2 K northeast of Iceland. The North Pacific SST increases with up to 0.5 K. In EXP2030, the amplitude of the anomalies grows but the pattern remains similar. Meanwhile, the cooling south of Greenland gets even more pronounced. The changes of SST strongly affect the turbulent surface heat fluxes between the ocean and the atmosphere (sum of latent and sensible heat fluxes, QTLA, positive means heat flux into the ocean, Fig. 3). For instance, over the cooling area in the North Atlantic, significantly positive QTLA anomalies occur, which means that less heat is released to the atmosphere. East of this area, negative QTLA anomalies occur, reflecting increased SST. In addition to the effects of changes in SST, changes in the atmospheric temperature might contribute to modified vertical temperature gradients and changes in QTLA.
Many future climate projections showed the smallest warming south of Greenland (Stocker et al. 2013) but only very few models simulated a significant cooling as found in our study. Also transient future climate projections with EC-Earth2.3 showed reduced warming but no cooling in this area (Koenigk et al. 2013). Observational based data, however, show a significantly negative temperature trend in this area south of Greenland (Rahmstorf et al. 2015), widely debated as the “Atlantic cold blob”. Rahmstorf et al. (2015) related this cooling to a reduction in the AMOC, especially after 1970. The strong cooling in our model might partly be due to the constant forcing as opposed to the transient forcing in CMIP5. The AMOC (Fig. 1b) and the associated oceanic northward heat transports vary on multi-decadal time scales. Thus, the transient year 2030 climate is affected by ocean water masses that have been formed decades before, in a climate with still high AMOC activity. In our experiments, the use of a constant forcing might increase the cooling effect from changes in the northward oceanic heat fluxes compared to the increased greenhouse gas forcing in transient climate simulations. All three experiments show large variations and a tendency to a weakening of the AMOC in the first two decades, particularly in EXP2030. Thereafter, the AMOC increases again but its average stays about 3 Sv smaller in EXP2030 than in EXP2000; also in EXP2015 a significant reduction occurs (compare Table 1). This reduction in our quasi-equilibrium simulations is almost three times as large as the AMOC reduction between year 2030 and year 2000 in transient CMIP5 projections with EC-Earth, and as large as the change until the second half of the twenty-first century in the CMIP5 simulations (Brodeau and Koenigk 2015). Towards the end of our 100-year simulations, the differences between the three simulations decrease. While part of this reduction is due to relatively low AMOC values in the last 20 years of EXP2000, we cannot rule out that a partial recovery of the AMOC in EXP2015 and EXP2030 contributes as well. Results by Blackport and Kushner (2016) indicated a recovery of an initial AMOC-reduction to Arctic sea ice loss after a few 100 years. Our experiment setup differs substantially from the experiments done by Blackport and Kushner (2016) but our time series are too short to see if a similar AMOC-recovery would take place. A partial recovery of the AMOC would likely lead to a less pronounced cold temperature blob.
The sea surface salinity (SSS) is strongly reduced in the same area of the North Atlantic where the ocean surface is getting cooler (Fig. 2d–f). Again, this is likely the consequence of the reduced transport of warm and salty water masses into this region. Also in the Central Arctic, the surface gets significantly fresher, likely due to increased freshwater input from rivers and increased precipitation (Koenigk et al. 2007, 2013). The ocean circulation stores most of the additional freshwater in the Beaufort Gyre or transports it in the Transpolar Drift Stream towards Fram Strait. An increased freshwater storage in the Beaufort Gyre is in agreement with observations (Giles et al. 2012). The increase of the salinity along some coastlines is likely caused by enhanced mixing due to longer periods with open water. Particularly along the North American coast, the mixed layer depth is increased (not shown). Enhanced sea ice melt in the Arctic might locally affect the surface salinity as well.
A somewhat surprising increase of salinity occurs in the Labrador Sea. This is surprising since the CMIP5 future simulations with EC-Earth (Brodeau and Koenigk 2015) indicated strongly reduced deep water convection in the Labrador Sea. This prevents that the relatively fresh surface layer is mixed with the underlying saltier layers, which would lead to a further reduction of SSS and convective activity. Both EXP2015 and EXP2030 show an increase of up to 300 m of the mixed layer depth (MXLD) in the area in the Labrador Sea where SSS increases. However, Fig. 2g–i show that this increase of MXLD occurs east of the main convection area in the Labrador Sea. In the main convection area, MXLD is reduced as expected.
The increase of MXLD by 300 m is likely not sufficient to substantially contribute to the North Atlantic Deep Water formation. The depth of the maximum AMOC is around 900 m in EC-Earth (not shown) and thus convection depth exceeding 900 m would be needed to feed the deep water path to the south. To further investigate this hypothesis, we calculate the “deep mixed volume” (DMV) for the Labrador Sea. The DMV is an index for the deep convection. It integrates the mixed water masses below a specific depth over a specific region (Brodeau and Koenigk 2015). Here, we use the same depth criterion of 1000 m as done by Brodeau and Koenigk (2015) but we have moved the box for the convective region of the Labrador Sea region used in Brodeau and Koenigk (2015) slightly to the northwest to capture the more realistic placement of the deep convection in EC-Earth3.0.1 compared to EC-Earth2.3. Figure 4a shows less frequent deep convective events in the Labrador Sea in EXP2015 compared to EXP2000 but the amplitude of the events is still similar. In EXP2030, however, both frequency and amplitude are substantially reduced. EXP2030 shows almost no deep convective activity between year 20 and 50 but a resurgence occurs from year 65 to 85. The latter period falls together with the warming period in EXP2030. Brodeau and Koenigk (2015) found a highly significant correlation between decadal variations of the DMV in the Labrador Sea and the AMOC a few years thereafter. Also in our simulations, the DMV is clearly related to the AMOC. Indeed, almost every peak in the AMOC can be explained by deep convection in the Labrador Sea (compare Figs. 1b and 4a, c, e). The correlation between 11-year running mean values of the DMV and the AMOC reaches 0.68, 0.80 and 0.73 in EXP2000, EXP2015 and EXP2030, respectively (Table 2). The correlation is largest when the DMV leads the AMOC by 4 years and the significance exceeds the 95 % significance level in all three simulations (using a student t test and taking the smoothing of the time series into account). A detailed description of the linkages between the DMV and the AMOC and the causes for the DMV variability is given in Brodeau and Koenigk (2015).
Table 2 Correlations calculated from 11-year running mean values
Similar to the Labrador Sea, the MXLD in the Greenland Sea and in the Irminger Sea are strongly reduced. The DMV for the Greenland-Iceland-Norwegian Seas (GIN-Sea, Fig. 4b, d, f) shows a strong reduction in amplitude, particularly in EXP2030, compared to EXP2000 and might contribute to a reduced AMOC as well. However, no clear relation between the variations of the DMV in the GIN-Sea and the AMOC could be found (Figs. 1b, 4) and correlations are not significant. This might be explained by the less direct effect of the GIN-Sea bottom waters on the AMOC compared to the deep water that is formed in the Labrador Sea. During the complicated travel across the overflows towards the North Atlantic, the GIN-Sea bottom waters are more mixed with other water masses, and contribute thus less distinctly to the AMOC than the deep water that is formed in the Labrador Sea.
The ocean volume transports into the Arctic show pronounced decadal variations (Fig. 5, left). Both the inflow into the Arctic through the Barents Sea Opening (BSO) and the outflow through the Fram Strait are slightly increased in EXP2015 and EXP2030. However, the increase is small compared to the amplitude of the decadal variations. The heat transports through BSO and Fram Strait (Fig. 5, right) are substantially enhanced, particularly in EXP2030. The increase in the heat transport through BSO is partly due to the larger volume flux but is predominantly the consequence of warmer temperatures, which also explains the amplified increase in EXP2030. In contrast, the variations of the heat transport through BSO are mainly governed by variations in the volume transport. After year 60, the heat transport through BSO in EXP2030 increases and stays at a higher level with reduced decadal variations. The volume transport through BSO is also high in this period and its variability is small. In EXP2030, this increase in ocean heat transport into the Arctic coincides with the warming of the Arctic after year 60 (Fig. 1). In this period, the increased convective activity in the Labrador Sea and the resulting increase in the AMOC might contribute to the warmer temperatures and enhanced ocean transports through BSO as well. We find significant correlations between 11-year running means of the DMV in the Labrador Sea and the BSO heat transport and T2m in the Arctic (Table 2). Correlations between 11-year running means of the AMOC and the BSO heat transport also reach about 0.8 in all three simulations (Table 2) and the heat transport through BSO is highly correlated with the mean Arctic T2m. This underlines the importance of the ocean heat transport into the Arctic for both Arctic climate variation and change. However, beside the increasing heat transport through BSO after year 60, the heat transport through the Bering Strait is also very high between year 70 and 90 in EXP2030 and might contribute to the Arctic warming after year 60.
Table 2 shows a significantly positive correlation between the AMOC and mean Arctic temperature. A possible explanation is that the AMOC affects the heat transport into the Arctic and thus consequently the Arctic sea ice and air temperature. As such, the reduction of the AMOC should have a dampening effect on the Arctic temperature in EXP2015 and EXP2030. By regressing the 11-year running mean values of the AMOC on the Arctic temperature in our three simulations, and taking the standard deviation of the AMOC into account, the reduction of the AMOC between EXP2030 and EXP2000 would lead to a reduction of the Arctic temperature increase by 0.6–1.1 K. However, this assumes that the processes that link the variability of the AMOC with Arctic T2m are the same as the processes that relate the mean changes in the AMOC with the changes of Arctic T2m.
Sea ice
EC-Earth tends to underestimate the observed seasonal cycle of the sea ice extent (Fig. 6). The ice extent is lower than in NSIDC-data (average over 1980–2013; Cavalieri et al. 1996) in March but, except for EXP2030, higher in September. The ice extent difference between EXP2015 and EXP2000 is substantially smaller than the difference between EXP2030 and EXP2015. The simulated mean reduction between September ice extent in EXP2030 and EXP2000 reaches about 1.3 million km2 in our model simulations and thus accounts for less than half of the observed reduction since 1980.
The simulated ice volume loss is quite large (from 15 to 9.9 million km3) but also smaller than estimates for the real ice volume reduction. However, sea ice extent and volume in the Arctic (Fig. 6) show pronounced decadal-scale variations in our simulations. These variations could mask or enhance human-induced trends at interannual to decadal scales. This finding is in agreement with a recent study by Swart et al. (2015) analyzing the internal sea ice variability in CMIP5 models.
Not only the ice reduction but also the simulated T2m increase in the Arctic in EXP2030 (1.78 K) is smaller than in ERA-interim since 1980 (about 2 K). Still, it seems that Arctic sea ice in our model might respond less sensitively to global warming than in reality.
Sea ice extent in EXP2030 is reduced after year 60, which is consistent with the warm Arctic temperature after year 60. As discussed in Sect. 3.1, the ocean heat transports into the Arctic through both the Barents Sea Opening and Bering Strait are high after year 60. Increased ocean heat fluxes through these sections are likely to reduce the Arctic sea ice extent as shown by previous observational (Schlichtholz 2011, Woodgate et al. 2010) and modelling studies (Koenigk and Brodeau 2014). We find a strongly negative correlation, exceeding −0.8 in all three experiments (Table 2), between the low-pass filtered (11-year running means) heat transport through BSO and the Arctic sea ice extent in September. The relation between the heat transport through the Bering Strait and sea ice extent seems to be more variable, the correlations vary substantially between the three simulations (Table 2).
Similar to the ice extent, the amplitude of the annual cycle of ice volume seems to be slightly underestimated in our EC-Earth simulations. However, note that ice thickness observations are still extremely uncertain and that the reference data from the Pan-Arctic Ice Ocean Modeling and Assimilation System (PIOMAS, Zhang and Rothrock 2003) used here should not be considered as the truth. In contrast to sea ice extent and air temperature, sea ice volume (Fig. 6c, d) is already considerably reduced between EXP2015 and EXP2000, and no clear acceleration until 2030 can be seen. This might indicate that the sea ice first has to be thinned down to a critical thickness before the ice extent is substantially reduced. This is partly in contrast to results from Holland et al. (2006), who linked the rate of summer ice retreat mainly to the interplay of simulated natural variability and forced changes. In agreement to our study, Holland et al. (2006) found an important influence of the ocean heat transport into the Arctic on sea ice variability and decline.
The ice export out of the Arctic constitutes an important linkage between the Arctic and lower latitudes by its influence on the deep water formation in the North Atlantic, particularly in the Labrador Sea, and consequently on the global ocean circulation (Jungclaus et al. 2005; Holland et al. 2001). Observationally based estimates suggested exports of about 0.7–1 Sv (Vinje et al. 1998, Kwok and Rothrock 1999, Schmith and Hansen 2003), which fits well to the simulated values of EC-Earth: in EXP2000, the ice export through Fram Strait reaches about 0.8 Sv in average, 0.7 Sv in EXP2015 and drops below 0.5 Sv in the last 30 years of EXP2030 (Fig. 6e, f). This reduced ice export in EXP2030 goes along with a thinning of the ice (not shown), which overcomes a potential increase in velocity, and is in line with future projections (Koenigk et al. 2007, 2013; Vavrus et al. 2012). Together with the drop in the mean ice export, the interannual to decadal variations are strongly decreased in the last 30 years of EXP2030. Figure 7 shows the spatial distribution of sea ice concentration in EXP2000 and the changes in EXP2015 and EXP2030. EXP2000 simulates too much ice in the Greenland Sea, and slightly too little ice along the rest of the ice edges in March compared to the OSISAF satellite data product (Eastwood et al. 2010). Generally, the ice concentration is too high along the ice edges in September. In EXP2015, in March, the ice concentration is mainly decreased in the Barents Sea, north of Iceland and in the Sea of Okhotsk. A slight increase can be seen in the Greenland Sea south of Svalbard in both EXP2015 and EXP2030. This increase might be related to northerly wind anomalies in this area (Fig. 8a, c). The changes in sea ice concentration strongly affect QTLA with a dipole-like pattern (Fig. 3); the strongest turbulent heat loss to the atmosphere moves northward together with the ice edge in EXP2015 and EXP2030.
In September, sea ice decreases mainly in the Barents Sea and in the Beaufort Sea along the North America coast in EXP2015. In EXP2030, we see a more general retreat of sea ice along all ice edges with largest reductions in the Barents and Kara Seas. Compared to the sea ice change in transient CMIP5 climate projections with EC-Earth 2.3 (Koenigk et al. 2013), the response is somewhat less focused on the Barents Sea region.
Changes in the atmosphere
Atmospheric circulation
Observations and recent studies suggested substantial changes in the large scale atmospheric circulation and linked these changes to the recent observed Arctic sea ice reduction (a review is given in Vihma 2014). Our simulations indicate, despite the strong changes in QTLA (Fig. 3), no dramatically changing atmospheric circulation between the three climate states (Fig. 8, left). Still, some areas with a significant SLP response occur in EXP2015 and EXP2030 compared to EXP2000. In winter, significantly increased SLP is found over the North Pacific and southern Europe, and reduced SLP from northern Canada towards eastern Siberia. The response pattern is similar in EXP2015 and EXP2030 but the amplitude of the response is larger in EXP2030. Over the North Pacific, the change pattern in winter agrees with the observed positive trend in the last decades. In contrast to this, the simulated SLP-changes over the North Atlantic-Eurasian area do not agree with the observed trend pattern, which is dominated by positive SLP trends extending from the Nordic Seas across large parts of northern Asia, and negative trends over the North Atlantic. Our results agree largely with results by Barnes and Polvani (2015) who found that the projected response of the circulation in CMIP5 models is either in the opposite direction to the observed one, or the spread among the models is too large to discern any robust response. It is unclear if the CMIP5 models are not able to reproduce the observed trends or if the observed trends are not robust. Most recent studies analyzing the linkage between sea ice reduction and atmospheric circulation trends were based on observational data of a time length of roughly 30 years. In order to investigate the robustness of the atmospheric changes between EXP2030, EXP2015 and EXP2000 for 30-year periods, our three 100-year simulations are subsampled into different 30-year periods. To keep it simple, we use 30-year running mean differences of SLP and T2m. Interestingly, it turns out that the SLP-differences between EXP2000, EXP2015 and EXP2030 vary strongly between 30-year periods; we find both positive and negative NAO-like atmospheric responses to 2015 and 2030 greenhouse gas concentrations in different 30-year periods. Figure 9 shows an example of possible SLP and T2m differences between EXP2015 and EXP2000 and EXP2030 and EXP2000 in winter. Especially, the EXP2015-EXP2000 difference is dominated by a strong negative NAO pattern (Fig. 9a). This leads to a cooling over a large area from northern and eastern Europe across Asia, with a maximum temperature decline of almost 2 K. This cooling is of similar amplitude as the observed one in the last decades (Cohen et al. 2012). A comparison with Fig. 8 clearly shows that 30-year periods, and thus any random period of time, which represents the length of the observational period, can deviate substantially from the entire 80-year period. Thus, one might speculate that even the observed temperature and SLP trends and their relation to the recent sea ice reduction (Vihma 2014; Mori et al. 2015) might not be robust. Even the 2030-year climate shows areas over Europe and Asia with a cooling compared to EXP2000 in selected 30-year periods, while the 80-year mean shows a substantial warming in the same region. Similarly to our results, Blackport and Kushner (2016) found large variations in the atmospheric circulation response to summer sea ice loss in different 50-year simulations and no indication for mid-latitude winter cooling. Also findings by Screen et al. (2014) indicated that very long integrations are needed to see robust atmospheric circulation changes in response to sea ice loss.
In summer, the SLP response to enhanced greenhouse gas forcing in EXP2015 and EXP2030 is generally relatively weak (below 1 hPa almost everywhere), although quite large areas with significant changes of SLP can be seen: increased SLP occurs over the North Atlantic, parts of Europe and Asia, and decreased SLP occurs west of Greenland and over the North Pacific (Fig. 8e, g).
To investigate possible changes in the variability of the winter atmospheric circulation in EXP2015 and EXP2030 compared to EXP2000, we calculated Empirical Orthogonal Functions (EOF) of the winter (DJF) averaged SLP between 30°N and 90°N (Fig. 10). The first EOF shows a similar large scale pattern representing the Arctic. Oscillation and explaining roughly 30 % of the total variance in all three simulations. However, we see a weakening of variations in the North Atlantic area, and the area with the strongest signal, around Iceland, extends further to the north in EXP2015 and moves into the Central Arctic in EXP2030. This extension into the Arctic is significant but the change over the Iceland-Nordic Seas area is not significant. Also the centre of action over the North Pacific moves slightly northward, which leads to a substantial increase of the SLP-gradient in the area of the Bering Strait.
The second EOF, which explains 15 % of the variance in EXP2000 and about 20 % in EXP2015 and EXP2030, shows a large signal over the northern North Pacific in all three simulations. However, it differs distinctly and significantly over the North Atlantic-European area: while EXP2000 has poles south of Iceland and over southwestern Europe, EXP2015 and EXP2030 both show a dipole with a large signal over northeastern Europe/Barents Sea region and southwestern Europe. This Barents Sea-southwestern Europe pattern can be found in EOF4 of EXP2000 but its importance is obviously growing in EXP2015 and EXP2030. Such a pattern favors northerly and easterly wind anomalies and thus cold winter conditions over parts of Asia and Eastern Europe, and resembles the observed atmospheric pattern in the cold winter 2010/2011.
Air temperature
The winter 2 m air temperature (Fig. 8, right) is already in EXP2015 significantly increased in most mid and high latitude regions compared to EXP2000. A region with a significant reduction of T2m occurs over the North Atlantic, which is related to the cooling in ocean surface temperature (Fig. 2) and to the reduced QTLA into the atmosphere (Fig. 3). The amplitude and extension of this cooling area is somewhat smaller for T2m compared to SST. This can be explained by the limited size of the Atlantic cold blob and the advection of warmer air masses from the surrounding areas with warmer SST. No significant T2m change occurs in an area extending from eastern Europe towards central Asia in EXP2015. The largest warming takes place over the Barents Sea with up to 3 K, otherwise the warming stays below 1.5 K over the continents and does not exceed 0.5 K over the ocean. In EXP2030, the warming in winter is strongly intensified and reaches more than 2 K in the entire Arctic (up to 5 K in the Barents Sea and Hudson Bay) and 1–2 K over the mid and high-latitude continents. Over the North Atlantic, instead, the cooling further amplifies and reaches −1 to −2 K southeast of Greenland. The strong warming in the Barents Sea is a consequence of the retreat of sea ice and enhanced surface heat fluxes. In the Hudson Bay, the sea ice concentration does not change but the ocean heat loss to the atmosphere increases. This might be due to reduced ice thickness in winter (not shown). Furthermore, the SLP-changes (Fig. 8a, c) indicate anomalous southerly winds that advect warmer air masses into the Hudson Bay area.
In summer, the warming is more evenly distributed with about 0–1 K in EXP2015 and 0.5–1.5 K in EXP2030. Again, a significant cooling occurs over the North Atlantic south of Greenland.
In winter, the vertical temperature distribution in the Arctic atmosphere is characterized by a strong near-surface inversion in the high Arctic in EXP2000 (Fig. 11). Warmest temperatures are typically found in heights between 850 and 900 hPa. At the surface, the temperature is up to 6 K colder, because of the continuous loss of heat from the surface through emission of infrared radiation. In addition, the insulating properties of the sea ice and snow prevent the heat exchange between the relatively warm ocean beneath the ice and the cold atmosphere. The strength and the spatial pattern of the inversion in EXP2000 (not shown) agree well with results based on the ERA-interim reanalysis data (Medeiros et al. 2011).
In EXP2015 and EXP2030, the strongest warming signal occurs near the surface, which tends to reduce the winter inversion strength and thus the atmospheric stability in the Arctic. Compared to lower latitudes, the temperature amplification decreases with height and disappears above 800 hPa in EXP2015. In EXP2030, the amplification extends far more up (up to 500 hPa) and is largest near the pole, while in EXP2015 it is more constrained to 70–85°N.
In summer, the Arctic atmosphere is relatively uniformly warmed with 0–1 K in EXP2015 and up to 1.5 K in EXP2030 (not shown).
The zonal mean distribution of the specific humidity follows closely the temperature distribution in EXP2000 (Fig. 11d). The changes in EXP2015 and EXP2030 show a much stronger increase of specific humidity at the surface at all latitudes, but contrarily to temperature, no amplification at high latitudes. The reason for this is the exponentially growing capacity of warmer air to uptake water vapor.
Arctic-mid latitude linkages
A number of studies suggested a possible link between the observed sea ice reduction and the large scale atmospheric circulation and mid-latitude air temperature (Jaiser et al. 2013; Inoue et al. 2012; Petoukhov and Semenov 2010; Hopsch et al. 2012; Overland et al. 2011; Peings and Magnusdottir 2014; Rinke et al. 2013; Koenigk et al. 2016). Most of these studies used either the difference between the last decade (with little ice) and the previous two (with much ice), or they used detrended time series to assess the atmospheric response to sea ice variability and trend. One common problem for all these studies is the fact that observational time series are very short. Moreover, they are derived from a climate in transition, which means that both external and internal forcing have likely changed during the three decades of observations.
Here, we investigate the relationship between sea ice anomalies in autumn and SLP and T2m in the following winter, in our three quasi equilibrium simulations, using correlation analysis. This analysis is performed based on variations of sea ice area in eight different Arctic regions as defined in Koenigk et al. (2016) (Northern Hemisphere, Barents-Kara Seas, Greenland Sea, Labrador Sea-Baffin Bay, Laptev-East Siberian Seas, Chukchi-Bering Seas, Beaufort Sea, Central Arctic). The general result is that the correlation between autumn sea ice and winter SLP is weak in all three simulations, and that air temperature shows the strongest response in the location of the ice anomaly and its surroundings. In the following, we will therefore only shortly discuss the response to November ice anomalies in the Barents and Kara Seas area (BAKA, Fig. 12) since previous studies found that the atmospheric response is largest to November sea ice anomalies in the BAKA region. The correlation between November ice area in the BAKA region and SLP in the following winter is below 0.3 everywhere in EXP2000. A small local response in the Kara Sea region can be seen, which is significant at the 95 % level. The T2m response in EXP2000 shows somewhat higher correlations: a significantly negative correlation from the Nordic Seas across the BAKA area (here correlation is maximal and exceeds −0.6) and Siberia to the North Pacific. This means that less ice in the BAKA area leads to a warming. The explanation for this local response is relatively simple: little sea ice in November in the BAKA area leads to reduced sea ice extent in winter—due to the persistence of the ice anomaly—and consequently to warmer temperatures in the Barents Sea. Significantly negative T2m anomalies occur also over southern North America, and positive anomalies over the western North Pacific. These anomalies seem to be related to SLP anomalies, which are not significant but have the potential to advect cold and warm air masses in the regions of the T2m anomalies. Very small positive correlations can also be seen over Central Asia.
In EXP2015, the correlation between November ice and winter SLP is still weak, although some slightly larger areas with significant correlations are found over eastern Siberia and central Europe. The strongest negative correlations between ice and T2m occur in the Barents Sea and its surroundings. Furthermore, we see more wide-spread negative correlations extending from Florida across the North Atlantic towards southern Europe, and negative correlations over the northeastern North Pacific. In EXP2030, the SLP response is similar as in EXP2015 over eastern Asia, but in addition, a significantly negative correlation occurs over the subtropical North Atlantic and over the western North Pacific. Significantly negative temperature correlations extend now all the way from the Caribbean across the North Atlantic, following the North Atlantic Current into the Arctic and further into the North Pacific and eastern Asia. Over a small area of the North Atlantic that spreads from eastern Canada to the south of Greenland, a positive correlation is found.
We performed the same correlation analysis with detrended data (not shown). The results are generally similar to the results from the analysis using the raw data. However, significantly positive correlations between sea ice and T2m occur over the North Atlantic subpolar gyre (up to r = 0.32) and weaker negative correlations along the North Atlantic Current compared to the correlations of the raw data.
None of our three simulations reproduces the suggested relation between reduced sea ice in the Barents Sea area and a trend towards more negative NAO winter atmospheric conditions. However, if we, instead of using the entire 80-year period, perform the same correlation analysis for different 30-year periods, we find large variations in the relation between sea ice and SLP (Fig. 13). Note, that correlations exceeding ∓ 0.38 are significant at the 95 % significance level, based on a two-sided t test and assuming 30 degrees of freedom and a normal distribution. For EXP2000 and EXP2030, 30-year periods exist with a NAO-like response pattern to reduced sea ice area in the BAKA area in November. The correlation coefficients reach similar amplitudes as the observed correlations (compare Koenigk et al. 2016, Fig. 4). Thus, EC-Earth is able to simulate a SLP response, which is similar to the observed response. Interestingly, neighboring 30-year periods show a NAO + like response of the winter SLP to ice reductions in the BAKA area. This shows that the response in our model is not robust over time and it indicates that 30-year periods are very short to make robust statements about a possible relationship between sea ice variations and mid-latitude climate variations.