Atmospheric Wind Biases: A Challenge for Simulating the Arctic Ocean in Coupled Models?

Many

Circulation in the Arctic Ocean can be described as a two-layered system (e.g., Aagaard, 1989; for a recent comprehensive review of Arctic Ocean circulation dynamics, see Timmermans & Marshall, 2020).The large-scale ocean-surface circulation is anticyclonic and driven by the dominating wind systems centered over the Beaufort Gyre and the Transpolar Drift Zone.Below the surface layer, the flow at intermediate depths is largely confined to narrow boundary currents along the steep slopes of the Eurasian and Canadian Basins (see Figure 1).These boundary currents form a cyclonic flow around the deep Arctic Basin in the opposite direction of the upper-ocean drift.Relatively warm, salty AW circulates around the Arctic in this cyclonic boundary current (Karcher et al., 2007;Rudels et al., 1999;Spall, 2013;Woodgate et al., 2001) at a mean depth of around 400 m.The strength and direction of this Arctic circumpolar boundary current (ACBC) determines the spatial distribution and storage of heat in the Arctic Ocean at this intermediate depth.
With this in mind, we compared Arctic Ocean temperatures in the Alfred Wegener Institute coupled climate model AWI-CM1 (Rackow et al., 2016;Semmler et al., 2020;Sidorenko et al., 2015) to the PHC3.0 climatology (Steele et al., 2001) at 400 m depth (Figures 2a and 2b) as well as the mean vertical temperature profiles in both deep basins (Figures 2c and 2d).The temperature maxima in each profile indicate the location of warm AW in the water column.Generally, the vertical temperature profiles averaged over the Eurasian and Canadian basins show that AWI-CM1 and the observational climatology are in good agreement except for in the Eurasian Basin, in the AW layer between 200 and 600 m depths, where AWI-CM1 exhibits a cold bias.But the basin-average temperature profiles do not yield any information about the spatial distribution and 10.1029/2021JC017565 3 of 15 presumed direction of AW circulation within each basin.Those characteristics are better visualized in the 400 m horizontal temperature slice (Figures 2a and 2b) that shows the large-scale features and pathways of warm AW in the Arctic basin.In choosing to evaluate the horizontal temperature distribution at 400 m depth, we follow the model evaluation example of the Coordinated Ocean-ice Reference Experiments (CORE-II; Ilıcak et al., 2016).For the AWI-CM1 simulation, comparison with PHC3.0 reveals an unrealistic distribution of AW: Warm AW that enters through Fram Strait does not follow the Arctic's deep basin slope cyclonically but rather gets diverted westward toward the Canadian Basin.The result is a cold bias in the eastern Eurasian Basin and a warm bias in the Canadian Basin, which is reminiscent of an anticyclonic circulation pattern.
It was somewhat surprising to find these respective biases after standalone simulations with AWI-CM1's ocean-sea-ice component FES-OM1.4 had shown a very realistic temperature distribution and a cyclonic circumpolar boundary current at horizontal resolutions of 25 and 4.5 km in the Arctic (Wang et al., 2018).
It is worth stressing that AWI-CM1 is not the only model to show deficiencies in simulating the propagation of AW around the Arctic as some recent studies have revealed.Shu et al. (2019), for example, evaluated a number of CMIP5 models and found that 9 out of 41 participating models did not simulate a well-defined AW layer at all.The multi-model mean (MMM) AW layer derived from the other 32 models was too thick and too deep.Furthermore, they found that the interannual variability in AW temperature was much weaker than observed, and none of the models simulated the warming trend observed in the recent decades.A follow-up study of 23 CMIP6 models by Khosravi et al. (2021) shows that the AW layer is still too deep and too thick in most models and the MMM, suggesting that representation of Arctic Ocean hydrography did not visibly improve between CMIP5 and CMIP6.For AWI-CM1, that study shows a simulated average AW core depth for present-day conditions in range with observations; that is, the AW layer is not too deep in AWI-CM1 as it is in many other models.At the same time, the simulated average AW core temperature is lower than observed, and the simulation shows no pronounced warming of the AW for the historical period.Before CMIP6, the problem of incorrect AW simulation was also recognized in standalone ocean models in the CORE-II model intercomparison study (Ilıcak et al., 2016).Suggested solutions for the AW problem have focused on improving the ocean-sea-ice model and include moving to higher-order advection schemes (Holloway et al., 2007;Maqueda & Holloway, 2006), using eddy-topography interaction parameterization ("Neptune parametrization"; Golubeva & Platov, 2007;Holloway & Wang, 2009;Nazarenko et al., 1998), tuning vertical mixing (Zhang & Steele, 2007) and increasing horizontal resolution (Wang et al., 2018).For AWI-CM1, we find that additional challenges may arise when a well-tuned (with respect to AW in the Arctic Ocean) ocean-sea-ice model, here FESOM 1.4, is coupled to an atmospheric model, in this case ECHAM6.3, to form a coupled climate model.
In this paper we aim to understand the processes causing the deterioration of the AW layer in AWI-CM1.If the standalone version of an ocean-sea-ice model can successfully replicate the two-layered circulation in the Arctic Ocean but coupling leads to perturbed AW circulation, then the reason must lie in the changes in ocean-surface forcing after coupling with the atmospheric model.Our investigations narrowed down the potential origins of these biases to a large-scale bias in sea-level pressure and associated wind field as well as a negative bias in sea-ice cover in the coupled model.Both biases had an impact on the surface stress imparted on the ocean.We conducted sensitivity experiments with our standalone ocean-sea-ice model to analyze the influence of biased wind and sea-ice cover over the Arctic.

FESOM1.4 Model Configuration
To study the effect of atmospheric forcing and sea-ice concentration on the AW circulation, we use the Finite Element Sea ice-Ocean Model FES-OM1.4 (Danilov et al., 2015;Wang et al., 2014) in a standalone mode.FES-OM1.4 is the ocean-sea-ice component of the AWI-CM1 model, which is FESOM1.4coupled to ECHAM6.3 (Stevens et al., 2013).The global performance of AWI-CM1 has been evaluated by Sidorenko et al. (2015) and Rackow et al. (2016); the AWI-CM1 contribution to CMIP6 has been evaluated by Semmler et al. (2020).The ocean-sea-ice component, FES-OM1.4,employs unstructured grids that allow for increased grid resolution in an area of interest (e.g., the Arctic) while keeping coarser resolution elsewhere.This, along with excellent scalability characteristics (Koldunov et al., 2019), makes longer, global simulations feasible.At least for AWI-CM1, the problem of incorrect AW circulation and distribution in the coupled set-up is mostly independent of ocean and atmosphere resolution.Our experiments are therefore run on the "baseline" ocean grid which was developed for participation in the CORE-II model intercomparison study (Griffies et al., 2009;Ilıcak et al., 2016;Wang et al., 2016aWang et al., , 2016b)).This ocean grid has a global nominal resolution of 1°; however, the resolution has been refined to about 25 km north of 50°N and to 1/3° at the equator as well as moderately along the coasts.The skill of FESOM1.4 at simulating Arctic hydrography with this baseline mesh has been evaluated comprehensively within the CORE-II model intercomparison study (Ilıcak et al., 2016;Wang et al., 2016b) and by Wang et al. (2018).The latter publication assessed the model's skill in comparison to a higher-resolution mesh, including its ability to reproduce the observed warming of the AW layer in the late twentieth century.Both setups-baseline and high-resolution-simulate this observed warming quite well.Although the ocean circulation itself is not directly evaluated in the CORE-II model intercomparison study, the 400 m temperature layer is shown for each model; and for FESOM1.4,the temperature distribution indicates a realistic pathway of AW into and around the Arctic (Ilıcak et al., 2016).Moreover, a more detailed evaluation of Atlantic Water core temperature (AWCT) and Atlantic Water core depth (AWCD) in simulations on this baseline grid show a temperature pattern indicating the correct circulation direction (Wang et al., 2018).Here, we show that this setup reproduces the vertical temperature structure in the Arctic Ocean very well (Figure 3); for more details on the model performance, we would refer the readers to the aforementioned references.
Comparing AWI-CM1 coupled simulations to atmospheric reanalysis data and satellite observations of sea ice revealed biases in the wind field over the Arctic as well as in the sea-ice cover.We presently consider these as possible candidates for causing biases in the ocean in the coupled setup.Accordingly, we ran sensitivity simulations from 1958 to 1989 and compare differences among simulations for the final 10 years from 1980-1989 (see Section 2.3).Therefore, while we show the biases of the coupled model compared to observational and reanalysis products for the period 1980-1989 please note that the biases are similar for other time periods as well (see Figure S1).The following section briefly describes these biases.

Coupled Model Bias in Mean Sea-Level Pressure and Sea Ice
The comparison of mean sea-level pressure (SLP) to the ERA5 reanalysis data set (Hersbach et al., 2020) shows that AWI-CM1 exhibits a bias over the central Arctic.This bias displays a dipole pattern: erroneously low SLP is found over the Canadian Basin whereas SLP is biased high over the Eurasian Basin and over the Barents and Kara seas (Figure 4).The reduced pressure gradient is associated with a shift of the Beaufort High toward the Eurasian Arctic.This dipole bias pattern is not unique to AWI-CM1.The coupled climate model MPI-ESM (Müller et al., 2018), which also uses ECHAM6.3 but with a different ocean-sea-ice model, shows the same bias in pattern and magnitude (not shown).A similar SLP bias pattern also dominated the model mean of other atmospheric and coupled models not including ECHAM, and this particular pattern has been attributed to the truncation of the North Atlantic storm track in atmosphere models (Walsh et al., 2002).
When sea-ice concentration from the coupled AWI-CM1 simulation is compared on one hand to the results from the standalone ocean-sea-ice model FESOM1.4 and on the other hand to a gridded sea-ice concentration data product based on satellite observations of brightness temperature provided by the National Snow and Ice Data Center (NSIDC; Cavalieri et al., 1996), it is apparent that AWI-CM1 overestimates sea-ice concentration in the Greenland Sea and eastern Barents Sea in March, when it is at its maximum extent (Figure 5c).This is similar to the results from FESOM1.4 (Figure 5b).In September, when sea-ice extent is at its minimum, AWI-CM1 underestimates the observed minimum sea-ice concentration in the central Arctic and predicts less sea ice than FESOM1.4forced with CORE-II forcing (Figure 5, bottom row).Only in some shelf regions, including the northeastern Barents Sea and the East Siberian Sea, more sea ice is simulated than observed.

Sensitivity Experiments
Three sensitivity experiments have been devised to determine the impact of the biases in wind and sea-ice cover (both separately and together) on AW circulation at depth (Table 1).A control simulation and the sensitivity experiments are initialized in 1958 using EN4 climatology (Good et al., 2013) and run with CORE-II forcing (Large & Yeager, 2009) for 32 years, until 1990.In sensitivity experiment 1 (WIND), the 6-hourly CORE-II wind forcing north of 67ºN, the Arctic Circle, was replaced with wind from an AWI-CM1 historical  simulation run on the same ocean grid, also started in 1958 from EN4.The 6-hourly AWI-CM1 wind output was interpolated onto the CORE-II forcing grid.In sensitivity experiment 2 (ALBEDO), the ice and snow albedos are lowered to such a degree that sea-ice cover is reduced to values similar to the ones found in the coupled simulation.Sensitivity experiment 3 (WIND + ALB) applied both of these changes together.Freshwater forcing is the same for all four experiments, with monthly varying precipitation and monthly climatological continental runoff.
In all four experiments, the drag coefficients for wind-stress computation over ice and water were adapted to track the stress computation in the atmospheric component of AWI-CM1, ECHAM6.3, as closely as possible.The neutral drag coefficient over ice Cd n,i is set to 1.89 × 10 −3 , the same value as in ECHAM6.3.The neutral drag coefficient over water Cd n,w is usually dependent on wind speed and is computed using a bulk formula.In ECHAM6.3 it is computed using the Charnock equation (Charnock, 1955).For our experiments using FESOM1.4,the neutral drag coefficient over water Cd n,w was set to a constant value of 1.285 × 10 −3 that represents the mean value of applying the Charnock relation to the most common wind-speed values over the Arctic, which are between 1 m/s and 8 m/s.

Results of Sensitivity Experiments
First comes the assessment of the effects of changing the wind forcing and lowering the albedo on mean sea-ice concentration in March and September (Figure 6).In winter, replacing the wind forcing over the Arctic Circle with wind from the coupled model results in increased sea-ice concentration along the ice edge in the Greenland Sea and eastern Barents Sea.Spatially more confined reductions in sea-ice concentration can be seen in Davis Strait and in the central Barents Sea.In March, the reduced albedo has almost no effect on sea-ice extent but some effect on sea-ice thickness (not shown).Therefore, the wind replacement governs the combined effect of changed wind forcing and reduced albedo in terms of sea-ice concentration  in March.The differences in sea-ice concentration between the experiments shown for the month of March (Figures 6a-6d) are representative of each month from November to April, when shortwave radiation is low and sea ice is more compact.
In September (Figures 6e-6h), the wind replacement leads to decreased sea-ice concentration north of the Canadian Archipelago and in the Beaufort, Chukchi, East Siberian and Kara seas but to increased seaice concentration in the Laptev Sea, in the central Arctic along the transpolar drift route, in the northern Barents Sea and along eastern Greenland.Reduced albedo leads to a decrease in sea-ice concentration everywhere.In experiment WIND + ALB, which combines wind replacement and albedo reduction, sea-ice concentration is mostly lower than in the control simulation, except along the transpolar drift route and east of Svalbard, which is associated with the impact of winds.The differences for the month of September (Figures 6e-6h) show the maximum effects of the wind replacement and albedo reduction in the melt season, and this is representative of situations from May to October.Mean differences in sea-ice concentration are depicted for each individual month in Figure S2.
Next to be investigated was the effect of the replaced wind and albedo reduction on ocean surface conditions.The difference in surface stress between the sensitivity experiment with replaced wind forcing and the control run in the central Arctic (Figure 7a) is generally acting from the Canadian Arctic Archipelago toward the Siberian side.In the Barents Sea and the Nordic Seas, the difference in surface stress is directed westward.The difference in surface stress due to reduced sea-ice cover is relatively small, showing a westward component in the Eurasian Basin (Figure 7b, note different scale).Differences due to the replaced wind dominate the sensitivity experiment WIND + ALB (Figure 7c).
The differences in surface stress lead to differences in Ekman transport.In the experiments with replaced wind forcing, the transport anomaly in the central Arctic is from the Canadian Basin toward the Eurasian  [1980][1981][1982][1983][1984][1985][1986][1987][1988][1989].Note that the scale for the experiment ALBEDO is doubled.
side.In the Nordic Seas, the difference in Ekman transport is directed toward Fram Strait and the Barents Sea Opening; in the northern Barents Sea, the difference in transport is directed toward the central Arctic (Figures 7d and 7f).In the albedo experiment, the differences in Ekman transport are relatively small and are directed from the Eurasian Basin toward the central Arctic.North of Fram Strait, there is a small area with Ekman convergence (Figure 7e).
In the Arctic Ocean, the differences in Ekman transports lead to a redistribution of low salinity ocean surface water.Halosteric height differences, i.e., the component of change in sea-surface height induced by differences in freshwater content in the water column, are shown in Figures 8a-8c.Regions of Ekman-transport-driven freshwater convergence (or divergence) show an increase (or decrease) in halosteric height.
In both the wind replacement and the albedo reduction experiments, two regions of increased halosteric height emerge: the Canadian Basin, where freshwater is shifted from the boundaries toward the center; and the Eurasian Basin, where increases in halosteric height are visible north of Fram Strait.Decreases in halosteric height are visible north of the Barents and Kara seas and in the Nordic Seas.Overall, the halosteric height differences are smaller in the albedo scenario.But the sea-ice reduction amplifies the effect of wind perturbation in the third sensitivity experiment, in which both perturbations were applied.
The differences in total sea surface height between the sensitivity runs and the control run are shown in ).The increase in sea surface height, then, is not due to salinity changes but rather due to warmer temperatures (Figures 9e and 9g, Nordic Sea region), that is, to thermosteric height changes.
These differences in sea surface height lead to differences in surface geostrophic velocity.In the WIND experiment, the velocity anomaly is anticyclonic over the Arctic basin and south of Fram Strait; in the Nordic Seas, the velocity anomaly is anticyclonic too (Figure 8g).Just north of Fram Strait, the difference in geostrophic surface current points westward.In the ALBEDO experiment, there is also an anticyclonic surface geostrophic current anomaly, this one mostly confined to the Canadian Basin (Figure 8h).In the Eurasian Basin, there is a weak cyclonic anomaly in surface geostrophic velocity.The effect of the wind replacement is amplified again by sea-ice decline in the experiment in which wind and albedo are considered together (Figure 8i).
Finally, the effect of applying the wind from the coupled model and reducing sea ice cover on the horizontal temperature distribution at 400 m depth is evaluated (Figure 9).Compared to the control run, applying the wind from the coupled model leads to colder temperatures both north of Fram Strait and along the cyclonic boundary current pathway.Warmer temperatures can be seen along the Greenland and Canadian slope and in the Greenland Sea (Figure 9e).
In the ALBEDO experiment, a slight cold bias has developed along the deep boundary current pathway; it is most pronounced at the intersection of the Eurasian Basin, the Lomonosov Ridge and the Canadian Basin, and also in the Greenland Sea (Figure 9f).When wind replacement and albedo reduction are combined, the cold biases grow more pronounced (Figure 9g); in fact, here the warm AW is no longer visible in the boundary current beyond the Barents Sea (Figure 9d).Instead, there is a cold pool in the Eurasian Basin and warmer temperatures north of Greenland and the Canadian Arctic Archipelago as well as along the southern boundary of the Canadian Basin; together, these imply an anticyclonic reversal of the AW flow (Figure 9d).The biased temperature distribution shown in Figure 9d is remarkably similar to the one in the coupled set-up (Figure 2b), and it is thereby with greater confidence that biases in the coupled system can be explained by analyzing the sensitivity experiments in an ocean-only configuration.The results are averaged over the last 10 years of the simulations (1980)(1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989).

Discussion
The sensitivity experiments performed with the forced ocean-sea-ice model show that biases like the ones in AWI-CM1, which affect surface pressure and wind field over the Arctic, paired with underestimation of the sea-ice cover, together can result in an unrealistic temperature distribution at 400 m depth.In such a case, the warm AW is no longer visible in the boundary current beyond the Barents Sea; instead, a cold pool emerges in the eastern Eurasian Basin.And at the same time, warmer temperatures appear both along the western boundary of the Arctic Ocean and in the Greenland Sea south of Fram Strait.These changes imply that some of the AW that does enter the Arctic Ocean is directed westwards and circulates anticyclonically around the Canadian Basin; the same changes also imply that less of the warm AW propagates from the Nordic Seas into the Arctic Ocean.
To investigate the westward direction and anticyclonic circulation of a portion of the AW around the Canadian Basin, we computed the depth-average topostrophy between 300 m and 3,000 m.The concept of topostrophy was introduced as part of the Arctic Ocean Model Intercomparison Project (AOMIP, Proshutinsky et al., 2001) by Holloway et al. (2007) in order to compare the simulated circulation between models.The velocity vector field is reduced to a scalar quantity τ that characterizes the tendency of a current to follow topographic slopes: where V is velocity, D is the depth gradient, and z is the unit vertical vector.A positive value indicates a current with shallower water to the right (northern hemisphere).In the case of the Arctic Circumpolar Boundary Current, positive topostrophy indicates a cyclonic flow direction.The control simulation shows a mostly cyclonic flow all around the Arctic Ocean Basin except for in a small area north of Greenland (Figure 10a).With wind forcing from the coupled model, the circulation in the Canadian Basin turns anticyclonic (Figure 10b), a shift that is most prominent on the East Siberian side.In the ALBEDO experiment, the direction of the circulation over continental slopes is barely affected; only north of the East Siberian Sea does the current have a slightly higher anticyclonic tendency (Figure 10c).In the experiment WIND + ALB, the circulation in most of the Canadian Basin is clearly anticyclonic whereas the cyclonic circulation in the Eurasian Basin seems strengthened (Figure 10d).
The mean net volume transport through the Barents Sea Opening is higher in the WIND (2.4 Sv) and WIND + ALB (2.7 SV) experiments than in the CTL (2.1 Sv) and ALB (2.3 Sv) experiments.These changes are associated with a reduced SSH in the northern Barents Sea (Figure 8).The AW loses most of its heat in the Barents Sea region before entering the Arctic deep basin (Smedsrud et al., 2013), so the increased AW transport through the BSO strengthens the cyclonic boundary circulation in the eastern Eurasian Basin (Figure 10), but it rather feeds more cold water to the ACBC (Figure 9d).This effect is most prominent in the WIND + ALB experiment, which has the strongest increase in BSO inflow.
Looking at the topostrophy in the Nordic Seas (Figure 10), we see that both of the simulations that feature replaced wind forcing show the West Spitsbergen Current (WSC) and the East Greenland Current (EGC) as weaker than they are in the control simulation.2018) investigated the impact of the Greenland Sea Gyre (GSG) circulation on AW temperature variability at Fram Strait based on ocean reanalysis data and found that cold anomalies at Fram Strait are related to an atmospheric pattern that shows a high-pressure anomaly centered over Svalbard, northerly winds over the western Greenland Sea, and southerly wind along the Norwegian coast (Chatterjee et al., 2018, see their Figure 3c).Such a wind pattern, which resembles the model bias in AWI-CM1, leads to Ekman convergence in the GSG and positive sea surface height anomalies which weaken the cyclonic GSG circulation.Figure 8d shows the difference in SSH between the WIND experiment and the control run as a positive height difference in the Nordic Seas.A composite (not shown) of SSH differences for the months with the largest negative net heat transport differences (Figure 10b, starred months show the 10 th percentile) exhibits even higher SSH differences.The increased SSH in the Nordic Seas reduces AW transport to the Nordic Seas across the Iceland-Scotland Ridge.Therefore, the wind anomaly both reduces AW transport toward the Arctic Ocean and increases the amount of AW entering the Arctic through the BSO.As a consequence, there is both a reduction in AW transport in the West Spitsbergen Current at Fram Strait as well as a reduction in AW recirculation in the EGC.
Because we set out to understand the influence of large-scale wind biases in a coupled model, we did not, for the sensitivity experiments, separate the wind bias into local (over the central Arctic) and remote (over the Nordic Seas) components.Nonetheless, the results discussed above are consistent with previous model studies on the role of local and remote wind forcing for the ACBC.Lique et al. (2015) and Lique and Johnson (2015) studied the effects of local and remote wind forcing on AW circulation at depth; they found that remote wind forcing over the Nordic Seas and the Barents Sea can cause a rapid, direct response in the AW circulation in the Arctic Ocean through a change in AW inflow.At the same time, they found that local wind forcing over the Canadian Basin results in slower changes that are "filtered" through surface circulation, which in turn modulates the deeper AW circulation.Here, a stronger anticyclonic wind forces a strong, deep Beaufort Gyre in the same direction such that no deep counterflow can develop.
Although the bias in sea-ice cover does not affect net heat transport through Fram Strait (Figures 11c), the sensitivity experiment with reduced albedo shows that reduced sea-ice cover does influence the surface distribution of freshwater (Figure 8b).Though reduced sea-ice cover has less of an effect than the wind bias has, it too contributes to a strengthening of the anticyclonic surface current in the western Arctic (Figure 8h).This effect of sea-ice decline on surface dynamics in the Arctic Ocean was also apparent in Wang et al. (2019).Spall (2013), based on idealized model simulations, had already shown that when the iceocean stress was removed completely, anticyclonic circulation in the Canadian Basin is lost and eddy fluxes from the boundary are enhanced, which indicates that the instability of the boundary current is suppressed by ice cover.Ideally, this potential consequence of sea-ice decline, and any related feedback should be investigated further using skillful coupled models to project future Arctic conditions.This study has investigated effects of a SLP and wind bias that are specific to the coupled climate model AWI-CM1.It is worth stressing that this particular bias is not unique to AWI-CM1: Another CMIP6 model, MPI-ESM, which also employs ECHAM6.3 as its atmospheric component but which has a different ocean-sea-ice component, shows a SLP bias nearly identical both in shape and magnitude.Furthermore, such a SLP bias has been found in other atmospheric models (Koenigk et al., 2013).Indeed, it is the dominant feature in multi-model means of SLP bias in model intercomparison studies of coupled and uncoupled models (Walsh et al., 2002).It has been suggested that the truncation of the North Atlantic storm track in the models, which prevents Atlantic cyclones from moving further north and into the region of the Norwegian Sea, Barents Sea, and Kara Sea, is what causes this bias.It has also been suggested that biased surface winds in the Arctic can adversely affect sea-ice transport, the resulting distribution of sea-ice concentration and thickness, and the export of ice and freshwater into the North Atlantic (Chapman & Walsh, 2007;Walsh et al., 2002).Our study shows that this wind bias pattern can additionally affect simulated circulation in the deep Arctic Ocean by imposing an anticyclonic surface circulation anomaly that then imprints on the deeper ocean circulation.Our results also reveal that a negative sea-ice bias could amplify this issue.Walsh et al. (2002) recommended that efforts to ameliorate the SLP bias in atmospheric models should focus on representing the topography across northern Asia and Greenland, or, more specifically, on topographic parameterizations which could affect the exchange of mass between Asia and the Arctic.They noted that the resolution of the atmospheric model may in part determine the magnitude of the bias.For AWI-CM1, we tried tuning some parameters related to model topography (gk_wake, gk_drag, and gk_lift) but without obtaining a significant reduction in the SLP or AW circulation bias.Mu et al. (2020) showed that the assimilation of sea surface temperature (SST) into AWI-CM1 led to a more realistic atmospheric circulation and also reversed the erroneous direction of the deep boundary current carrying AW after 8 simulation years with data assimilation.This reversal of the current into a cyclonic flow was attributed to the improvement of atmospheric states over the ice-free area that can further propagate to the whole Arctic dynamically.

Conclusions
Atlantification might be an important driver in amplified Arctic sea-ice melting and Arctic Ocean warming (Årthun et al., 2019) as well as in regional differences in sea-ice loss (Årthun et al., 2021).Using global coupled climate models to study the evolution of Atlantification and any associated feedback in the Arctic climate system requires realistic simulations of AW inflow and circulation through the Arctic Ocean.A skill assessment of the AW layer representations in the CMIP5 (Shu et al., 2019) and CMIP6 (Khosravi et al., 2021) models shows that, in many of the models that do simulate a distinct Arctic AW layer, this layer is too thick, too deep, or does not show the observed warming trend.Biases like this are commonly related to insufficient resolution, too much mixing in the ocean component, and unrealistic Atlantic-Arctic Ocean exchanges.
Our study shows that even if all impediments to realistically simulating AW are addressed in the ocean model (e.g., by implementing sufficient resolution in the horizontal and vertical, Arctic gateways-resolving grid resolution, faithful bathymetry representation, the right amount of mixing), new obstacles may arise when the ocean model is coupled to an atmosphere model with its own shortcomings in the Arctic.AWI-CM1, like many other coupled climate models, exhibits a bias in SLP over the Arctic Ocean, with higher pressures over the Eurasian Basin and the Barents and Kara Seas and lower SLP over the Canadian Basin.In the Arctic Ocean, the bias in wind stress (which is related to the biased surface pressure gradient) leads to differences in Ekman transport, freshwater distribution, and halosteric height that strengthen the anticyclonic surface circulation in the Canadian Basin to the point of reversing the deep counterflow.This effect is visualized by the negative topostrophy (Figures 10b and 10d) and warm bias (Figures 9e and 9g) in the Canadian Basin.An underestimation of sea-ice concentration such as is seen in AWI-CM1 can amplify the described processes locally.In the Nordic Seas, an anticyclonic wind bias increases the SSH and weakens the cyclonic gyre circulation, which leads to reduced volume and heat (a proxy for AW) transport into the Arctic Ocean through Fram Strait.At the same time, the wind bias increases AW transport through the BSO.And since the BSO feeds cold water to the deep basin, the overall effect is a cold bias in the Eurasian Basin (Figures 9e and 9g).
The problem of biased SLP over the Arctic cannot be overcome easily.Efforts as suggested by Walsh et al. (2002) to tune parameters related to model topography did not lead to a significant reduction of the SLP bias or the circulation bias in our practice.SST assimilation seems to rectify the circulation in AWI-CM1, but this constraint cannot of course be applied to the future scenario simulations.
Currently, a new version (version 3) of AWI-CM is in development.For its atmospheric component, ECHAM is replaced with OpenIFS (OIFS, Roberts et al., 2018).Preliminary results show that the SLP bias pattern over the Arctic is present in OIFS, too, but it is of smaller magnitude than in ECHAM.The temperature distribution at 400 m implies a cyclonic circumpolar circulation in simulations with AWI-CM3.This new model version has yet to be released and will be described later, and separately, after the initial model tuning process.
The Arctic is a hotspot in global warming, but there is still a large inter-model spread as well as model uncertainty in CMIP6 projections of the surface warming-especially in the Arctic (Cai et al., 2021).Detailed investigations of model biases are needed to improve the simulations and reduce sources of model uncertainty.Simulating ocean heat transport into and within the Arctic more faithfully will help to understand and better predict the course of change in the Arctic.

Figure 1 .
Figure 1.Map of the Arctic Ocean.Depicted are the two deep basins (Canadian and Eurasian Basin) divided by the Lomonosov Ridge, the shelf seas and neighboring seas.A schematic view of the dominant features of Arctic Ocean circulation shows the anticyclonic Beaufort Gyre System, the Transpolar Drift Stream (in green) and the subsurface cyclonic flow of AW around the Arctic Basin (in pink); south of Fram Strait, the Greenland Sea Gyre (GSG; in green); and the West Spitsbergen Current (WSC, orange) and the East Greenland Current (EGC, light blue).Sections where transports through Fram Strait and the southern Barents Sea Opening are computed are shown in black.The dashed red line (S1) shows a section along 130°W/70°E across the deep Arctic Ocean.

Figure 2 .
Figure 2. Horizontal distribution of potential temperature [°C] at 400 m depth from the PHC3.0 climatology (a) and the mean potential temperature (1980-1989) at 400 m depth from a coupled AWI-CM1 LR historical simulation (b).This LR simulation was run on an ocean grid with 25 km resolution north of 50°N and at T63 atmospheric resolution.A similarly biased temperature pattern is also evident in AWI-CM1 runs with higher oceanic or atmospheric resolution.Mean vertical temperature profiles over the Eurasian Basin (c) and Canadian Basin (d) are shown for the PHC3.0 climatology and the AWI-CM1 LR simulation.The temperature maxima indicate the average core depth of Atlantic Water.The thin gray line indicates the 400 m depth shown in the horizontal sections in (a and b).

Figure 3 .
Figure 3. Potential temperature along a section across the deep Arctic Ocean from 130ºW/70ºN to 70ºE/81ºN (see section S1 on the map in Figure 1) in the PHC3.0 climatology (a) and mean potential temperature (1980-1989) in a FESOM1.4simulation with CORE2 forcing (b).

Figure 5 .
Figure 5. Mean March (a) and September (e) sea-ice concentration (SIC, [%]) from the National Snow and Ice Data Center (NSIDC) for the period 1980-1989, the difference in SIC between FESOM1.4 and NSIDC (b and f), the difference in SIC between AWI-CM1 and NSIDC (c and g), and the difference in SIC between AWI-CM1 and FESOM1.4(d and h).In March, both FESOM1.4 and AWI-CM1 overestimate SIC in the Greenland Sea and Barents Sea.In September, AWI-CM1 simulates less SIC than observed and less than the standalone model FESOM1.4.

Figure 6 .
Figure 6.Simulated sea-ice concentration (SIC, [%]) in the control run for March (a) and September (e), and differences in SIC to control run for experiment WIND (b and f), Albedo (c and g) and Wind + Alb (d,h), all averaged over the last 10 years of simulation (1980-1989).

Figure 8 .
Figure 8. Halosteric height difference [cm] (a-c), sea surface height difference [cm] (d-f) and differences in geostrophic surface current velocity [cm/s] (g-i) for the three sensitivity experiments compared to the control simulation, averaged over the last 10 years of simulation (1980-1989).
Figures 8d-8f.The comparison of halosteric height differences (Figures 7a-7c) to sea surface height differences (Figures 8d-8f) shows similar patterns for the central Arctic Ocean, in Baffin Bay and in the Barents and Kara seas.The resulting indication is that sea surface height changes are driven by freshwater distribution changes in those regions.For the experiments WIND and WIND + ALB, however, the halosteric height in the Nordic Seas has decreased while sea surface height has increased (compare Figures 8a-8c and Figures 8c-8f
The monthly difference in net heat transport into the Arctic between the WIND experiment and CTL is shown in Figures 11b.Most of the time, monthly net heat transport through Fram Strait in the control simulation is larger; this can be explained by the wind bias over the Nordic Seas (Figure12).Chatterjee et al. (

Figure 11 .
Figure 11.Time series of monthly net heat transport [TW] through Fram Strait for all experiments (a), difference in heat transport [TW] between experiment "Wind" and "CTL," with months below the 10th percentile marked with stars (b), cumulated net Fram Strait heat transport [TW] for all experiments (c).

Figure 12 .
Figure 12.Bias in MSLP [hPa] and wind in AWI-CM1 compared to ERA5 with focus on the Nordic Seas.

Table 1
Overview of Sensitivity Experiments