Freshwater content and mean sea-ice thickness
In the reanalyses ORAS4 and ORAP5 we found that much of the Arctic Ocean’s freshwater is stored in the Beaufort Gyre (Fig. 1a, b). According to Serreze et al. (2006), this is a response to the mean Beaufort Sea atmospheric anticyclone that promotes Ekman convergence. Saltier water (Atlantic waters entering the Arctic Ocean) extends from the Norwegian Sea along the eastern side of the Fram Strait northwards into the Nansen and Amundsen Basins until the Laptev Sea. Along the shelves of the Siberian coast and in the Canadian Archipelago, the water is fresh due to the incoming river runoffs. EC-Earth3.1 at low resolution shows saltier conditions in the entire Arctic compared to ORAS4 and ORAP5 (Fig. 1c). The high resolution version of EC-Earth3.1 (Fig. 1d) show saltier Arctic than the coarse resolution, which appears to be the model with lower FWC.
The four CMCC-CM2 simulations show a FWC pattern similar to the reanalyses, the high resolution ORCA025 showing a saltier Beaufort Sea (Fig. 1f, h) than the ORCA1 ocean resolution simulations (Fig. 1e, g). Both CMCC PI simulations (Fig. 1e, f) show saltier conditions over the Central Arctic and the Beaufort Sea compared with CMCC PD simulations (Fig. 1g, h).
The three HadGEM3-GC2 simulations show a similar pattern to ORAS4 and ORAP5 with higher freshwater content compared to the reanalyses. With higher atmospheric resolution (Fig. 1i–k), the FWC agrees better with reanalyses over the Nansen Basin and the Laptev Sea, however the simulations show higher FWC over the Beaufort Sea than the reanalyses.
A persistent wind-induced flow from the Siberian coast across the pole (called Transpolar Drift Stream) to the Fram Strait transports sea ice towards the north coast of Greenland and the Canadian Archipelago and out of the Arctic through the Fram Strait (Kwok and Untersteiner 2011). Every year, approximately 10% of the sea ice area of the Arctic Basin is exported through the Fram Strait into the North Atlantic Ocean (Smedsrud et al. 2017). Similarly, new ice is formed in the Arctic Basin and particularly along the Siberian coast. In concordance to these dynamic and thermodynamic processes, the spatial averages of the reanalyses show a spatial gradient in mean sea ice thickness (MSIT) that progressively decreases from the Canadian Archipelago, Canadian Basin and northern Greenland towards the Eurasian part of the Arctic (Fig. 2a, b). In general, the models reproduce the spatial MSIT distribution in the reanalysis with maximum MSIT over the Canadian Archipelago and in the western Arctic. However, the spatial gradient towards thinner ice at the Siberian coast is weaker in most of the models. The EC-Earth3.1 model shows a strong MSIT reduction from low to high resolution (Fig. 2c, d). While EC-Earth3.1 shows higher MSIT at low resolution, it shows a strong reduction of MSIT at high resolution, showing almost no ice in the North Atlantic Arctic sector. This low MSIT at high resolution might be due to the strong inflow of Atlantic water into the Arctic.
The pre-industrial CMCC-CM2 (CMCC PI) simulations show generally a much higher MSIT than the reanalysis GLORYS and the present day simulations. While with pre-industrial forcing, high resolution leads to a substantial reduction of MSIT (Fig. 2e, f), this is less clear in the present day (CMCC PD) simulations (Fig. 2g, h). Here, we mainly see a different distribution of the MSIT with thinner ice at both the Siberian coast and the Canadian Archipelago. Both CMCC PD show lower MSIT than both reanalyses.
All HadGEM simulations show also lower MSIT than the GLORYS reanalyses, and we see a systematic decrease of MSIT with increasing atmospheric resolution (Fig. 2i–k).
In order to better understand the changes in FWC and MSIT produced by changes in the model resolution or the forcing, we calculated the difference in FWC and MSIT between low and high resolutions (Figs. 3, 4 respectively) for each model. High resolution simulations were interpolated to the low resolution using cubic interpolation. When low and high ocean resolution simulations (ORCA1 minus ORCA025) are compared, we find that low resolution shows larger FWC over the Central Arctic Ocean and lower FWC over the Kara and Laptev Seas compared with high resolution (Fig. 3a–c), EC-Earth3.1 showing the largest difference. In turn, when lower and higher atmosphere resolutions in the HadGEM simulations are compared, we find that lower resolution shows increased FWC over the Kara and Laptev Seas and decreased FWC over the Beaufort Sea. This result is systematic, since lowest and highest resolution show the largest FWC difference (Fig. 3d–f). When same resolution pre-industrial and present day forced CMCC simulations are compared, we find lower FWC in the entire Arctic ocean in the pre-industrial simulations (Fig. 3g, h).
Assessing the impact of ocean resolution, we found that low resolution shows higher MSIT than high resolution in CMCC-PI and EC-Earth (Fig. 4a, c), while CMCC-PD shows lower MSIT with lower ocean resolution than with high resolution (Fig. 4b). For EC-Earth3.1, the MSIT reduction might be a combined effect of high atmospheric and oceanic resolutions. The higher MSIT and FWC in CMCC-PI and EC-Earth3.1 is apparently due to a net positive inflow of freshwater into the Arctic, which will be further discussed in this section. The lower MSIT along the Arctic ocean in the CMCC-PD at low resolution might be due to a stronger ice response to temperature at lower resolution. It is worth noting that the changes in MSIT due to resolution for CMCC-PD are relatively small compared to CMCC-PI and EC-Earth.
Regarding the impact of changes in atmospheric resolution, as for FWC, systematic change of MSIT with higher atmospheric resolution, showing higher MSIT with lower resolution is found (Fig. 4d–f). Higher MSIT in the lower resolution might lead to a slower ocean circulation in the Beaufort Sea, since thicker ice over the Beaufort Sea, would lead into higher ice strength, which has been shown to decrease the ice speed, hindering the transmission of momentum from the atmosphere into the ocean (Rampal et al. 2011; Olason and Notz 2014; Docquier et al. 2017), which in turn weakens the Ekman convergence, and thus the inflow of freshwater from the Bering Strait into the Beaufort Sea. This contributes to lower FWC in Beaufort Sea at lower atmosphere resolution (Fig. 3d–f). In the same way at low resolution and in line with Roy et al. (2015), the winds acting in the Transpolar drift, might have less contact with the liquid water on ice-covered Laptev and Kara Seas, decreasing the freshwater transport from the Laptev and Karas seas, and therefore increasing the amount of FWC over this region of the Arctic ocean.
When same resolution pre-industrial and present day forced simulations are compared, we find that MSIT reduction from pre-industrial to present-day run in CMCC is enhanced at low resolution (Fig. 4g, h). Therefore the high MSIT in pre-industrial forced simulations explains and compensates the low liquid FWC (Fig. 3g, h).
To summarize this sub-section, we have found systematic lower freshwater volume with increased ocean and atmospheric resolutions. Although we use models with the same ocean component (NEMO model), the atmospheric components are different and use different parameterizations, which would impact the ice directly, and consequently the freshwater volume. According to Hodson et al. (2013), the majority of Arctic ice-volume and freshwater uncertainties are related to sub-gridscale parameterizations and model structure variations rather than intrinsic internal variability. Therefore, for this study, the different intensities we obtained in the comparison of simulations at different model resolution could arise from different radiation scheme, and ultimately different temperatures in the atmosphere.
Parameter uncertainty seems to play an important role in the spread of variables. Small changes in the radiation scheme, the convection scheme or any other atmospheric or ocean parameterization, can derive in significant changes in the Arctic climate. For example Roy et al. (2015) showed how different atmosphere-ice-ocean surface layer treatments impact the ice volume, freshwater volume and freshwater exports from the Arctic. Structural uncertainty dominates the transports i.e. the large scale circulation, which affects the Arctic key variables (Hodson et al. 2013), however a deeper analysis of sources of uncertainty besides model resolution is out of the scope of this paper. From our study we can remark that resolution has an impact on the spread of variables, however parameterization changes matter, particularly when comparing results from different models.
In the following sub-section we focus on the freshwater and ice transport from the Arctic. When we analyze the Arctic’s liquid freshwater mean inflow and outflow and its variability (Fig. 5; Table 2), we observe that for ORAS4 the largest export of liquid freshwater occurs via the Baffin Bay with a mean outflow of 2280 km3 year−1 through North Baffin, comparable with observations (3200 ± 320 km3 year−1 according to Serreze et al. 2006). For ORAP5, in turn, the outflow across North Baffin is of about 800 km3 year−1 which is the second highest export after the Fram strait. We find that all models show a larger agreement with ORAP5 reanalysis on freshwater export across North Baffin Bay (with less than 1550 km3 year−1). We find that low ocean resolution simulations show higher freshwater export across North Baffin Bay than the high resolution simulations. Models show liquid freshwater exports through Fram Strait similar to observations (2700 ± 530 km3 year−1), and therefore higher exports than ORAS4 (~ 1030 km3 year−1) and ORAP5 (~ 1465 km3 year−1), which underestimate the observations. Particularly CMCC-CM2 and HadGEM show the largest freshwater export across Fram Strait (> 2200 km3 year−1), and therefore the best agreement with observations. Freshwater export across the Barents Sea is small (90 ± 90 km3 year−1 from observations) and there is a spread in simulated freshwater export across models, from about − 1450 km3 year−1 in EC-Earth to around 600 km3 year−1 in CMCC-ORCA1-PI. Consistent with previous works, we find that the largest liquid freshwater inflow occurs through the Bering Strait supplying around 2840 km3 year−1 in ORAS4 and ORAP5, quite similar to observation estimates (2400 ± 300 km3 year−1). Coarse ocean resolution simulations tend to show larger freshwater inflow across Bering Strait than high ocean resolution models, which causes a positive total net inflow of liquid freshwater into the Arctic (red line is always positive in Fig. 5, see Table 2) in EC-Earth3.1 and CMCC-PI. On the other hand, increased atmospheric resolution tends to increase the modelled freshwater export from the Arctic (Fig. 5i–k; Table 2).
Consistently with previous studies (e.g. Vinje et al. 1998; Vinje 2001; Aagaard and Carmack 1989; Schmith and Hansen 2003), in GLORYS2V1 and GLORYS2V4 the largest sea ice exports take place through Fram Strait and amount to 1860 km3 year−1 in average (Fig. 6a, b), which is slightly lower than observation estimates (from 2400 to 3200 km3 year−1). The Fram Strait is properly reproduced by all models as the main gateway with the largest Arctic ice exports (see Table 3). The CMCC-PI ORCA1 model shows the largest ice export with 4720 ± 1580 km3 year−1 (Fig. 6e) and high resolution EC-Earth3.1 the smallest with ice export in average around 267 km3 year−1 (Fig. 6d). Low ocean resolution simulations show larger ice export across Fram Strait than high resolution simulations (Fig. 6c–h) although the difference is less pronounced in the CMCC-PD simulation (Table 3). Consistent with MSIT (Fig. 3a-c), models that exhibit higher MSIT over the Arctic at low resolution, have larger ice export from the Arctic.
Mean sea ice thickness decreases with increased atmospheric resolution in the HadGEM model (Fig. 4d–f), the ice transport across Fram Strait is also sensitive to changes in atmospheric resolution, and shows smaller values with increased resolution (Fig. 6i–k; Table 3). Higher roughness length at higher atmospheric resolution, due to better representation of the ice surface, increases the momentum transfer from the sea ice to the ocean slowing down the sea ice drift in the Beaufort Gyre. According to Roy et al. (2015), the slower drift reduces sea ice convergence into the center of the Beaufort Gyre, thereby increasing the concentration around the margins of the Beaufort Gyre. The reduction in lead fraction reduces the sea ice growth. This results in a net decrease in the Arctic sea ice volume in higher atmospheric resolutions.
The amount of ice export is comparable to the liquid freshwater export; however the ice export depends more on atmospheric dynamics due to its confinement to the sea surface. Ice motion is largely wind driven, although it is also influenced by ocean currents and ice internal stress, and the amount of ice transport from the Arctic across the Fram Strait has been found to be highly related with the SLP gradient across the Fram Strait. According to Kwok and Rothrock (1999), the SLP gradient explains around 80% of the variance of the winter ice area flux across Fram Strait. Therefore - in order to analyse differences in the representation of the relation between the SLP gradient and the Fram Strait ice transport across simulations, we calculated correlations between ice transport across the Fram Strait and SLP in every grid point of the northern Hemisphere (Fig. 7; Table 4). The spatial correlation pattern between GLORYS2v1 reanalysis ice transport time series and ERA-Interim SLP, show negative correlation in the Eurasian Arctic sector. The most intense negative correlation (r < − 0.8) is located over Svalbard and the Barents Sea (Fig. 7a). The correlation weakens towards the west across the Fram Strait. Dynamically this implies that a negative anomaly of SLP gradient over the Barents Sea would enhance the southward wind component across the Fram Strait, which as a consequence would increase the ice transport towards the south.
When comparing the low versus high ocean resolution simulations with reanalysis, we find that the low resolution simulations tend to reproduce the extension of the negative correlation over the Barents Sea and the eastern Arctic slightly better than high ocean resolution simulations (Fig. 7b–g). High ocean resolution simulations show a smaller region of strong negative correlation east of Svalbard with less extension towards the eastern Arctic Ocean. However all of them show a zonal correlation gradient across the Fram Strait.
When comparing the correlation patterns at different atmospheric resolutions in HadGEM, we find that although all simulations show negative correlations over the eastern Arctic, the correlation intensities are lower than in the reanalyses (Fig. 7h–j). We also observe that the increase in atmospheric resolution does not show a systematic impact on the intensity of the correlations since the intermediate resolution shows the most intense correlation patterns (Fig. 7i) and more similar spatial pattern to the reanalysis, compared to lower (Fig. 7h) and higher resolution (Fig. 7j) simulations. Despite the smaller negative correlation east of Svalbard in HadGEM, we still find that the SLP-gradient across Fram Strait plays a governing role for the ice export. However, in this case, the SLP gradient does also depend on high pressure anomalies over Greenland as much as on low pressure over the Barents Sea area. Table 4 summarizes these results with correlation of ice export from the Arctic across the Fram Strait and the SLP gradient across the Atlantic Sector of the Arctic Ocean.
For the ice transport into the Arctic across the Bering Strait, all models show positive correlation with SLP over the Eastern Siberia and Laptev Seas and negative correlation over Alaska and the North Pacific, reaching the Bering Sea (not shown). Therefore an intensified Aleutian Low pressure would enhance the movement of ice towards the Arctic Ocean. In turn, for the ice export from the Arctic across the Baffin Bay, all models show a strong negative correlation over the Davis Strait and over Greenland (not shown), therefore a low pressure over Greenland and the Davis Strait would promote the northerly winds across the Baffin Bay enhancing the ice export from the Arctic into the North Atlantic.
Impact of ice transport from the Arctic on the North Atlantic convection
Previous works have shown that convection in the North Atlantic Ocean is highly sensitive to changes in the freshwater balance. An example of this is the Great Salinity Anomaly in the early 1970s. Häkkinen (1999) showed with an idealized study how freshwater pulses in the East Greenland Current cause a reduction in convection in the Labrador Sea. Koenigk et al. (2006) showed that the freshwater exported through the Fram Strait propagates in the East Greenland Current to the south, reaches Labrador Sea in about 1–2 years, and produces a negative salinity anomaly in the Labrador Sea, that strongly reduces the convection. Ensemble simulations by Mikolajewicz et al. (2005) indicated a risk for a decadal-long shutdown of the Labrador Sea convection after huge Fram Strait ice exports.
Here, we investigate only the HadGEM simulations since the other simulations are too short for comparing decadal scale variations. In order to analyse the impact of the atmospheric resolution on the convection in the Labrador Sea and its relationship with freshwater export across the Fram Strait, we compared an index of deep convection in the Labrador Sea (Brodeau and Koenigk 2016) with the previous Fram Strait freshwater export. The convection index is based on March mixed layer depth and takes both vertical and horizontal extensions of the convection areas into account. We used 10-year running mean for both time-series to filter out the high frequencies and to focus on the long term relationship between the freshwater transport across the Fram Strait and the convection in Labrador Sea. We find a systematic behaviour showing a resolution dependence of the relationship; the correlations between ice export and convection were for the low resolution r = 0.31, for the intermediate resolution r = 0.52, and for the highest resolution r = 0.57 (Fig. 8) when the ice transport leads the convection with 1 year for all resolutions.
The deep convection is reduced with higher resolution in the HadGEM model (Fig. 9). In general, all three model simulations show higher convection in the Labrador Sea compared to ARGO-mixed layers (Holte et al. 2010) and particularly, the low resolution simulation shows deep convection every year. Thus, the convection is less sensitive to changes in the upper surface density due to freshwater inflow in the low resolution simulation compared to the high resolution simulation. The reason for the lower convection in high resolution is not entirely clear. While the turbulent surface heat fluxes are very similar in all three model setups (not shown) and also liquid and solid freshwater exports through Fram Strait agree well (Figs. 5, 6), the freshwater export through the Davis Strait is slightly larger in the high-resolution simulations (Fig. 5) and could be the reason for slightly increased vertical stratification in the Labrador Sea at high resolution.