An ensemble-hindcast approach is designed to simulate North Atlantic circulation changes during the period 1948–2005 using the Max Plank Institute Ocean Model forced by atmospheric reanalysis data. The experiment seeks to isolate oceanic changes related to the atmospheric forcing history by evaluating the role of initial conditions and masking internal model variability in the ensemble mean results. Characteristics of the complete North Atlantic–Arctic Mediterranean exchange system is described at key sections by time series of exchanges of volume, ice, heat, and liquid freshwater. Volume transports are divided into water masses by properties, and the constructed climatology of exchanges is shown to compare well with available observational estimates of individual branches. In response to the atmospheric forcing history, we find a modest but robust decline of the Atlantic meridional overturning circulation of 0.4 Sv/decade, in total 2–2.5 Sv since 1948. In contrast, overturning exchanges with the North Atlantic and Nordic Seas and with the Nordic Seas and Arctic Ocean show no sign of cessation. A marked increase in the freshwater storage of the Nordic Seas and in particular of the subpolar Atlantic accompany the model decline in Atlantic overturning. The characteristics of the modeled freshening bears similarities with observed patterns of change since the 1950s, but whereas the origin of observed freshening is uncertain, freshening in the model predominantly results from dynamic changes linked to prominent changes in atmospheric circulation. Moreover, documented increase in the drainage of major Arctic-rivers is found to have a negligible effect on the oceanic changes.
 In the Atlantic, the meridional overturning circulation (MOC) is characterized by northward flow of light surface waters accompanied by a transport of heat and salt from the subtropics to the subpolar regions and farther into the Labrador Sea, Nordic Seas, and Arctic Ocean: the latter in common termed the Arctic Mediterranean. This flow of Atlantic Water keeps large parts of the region ice-free, and in combination with a gradual release of heat to the atmosphere it contributes to maintain the relatively mild winters of Europe [Seager et al., 2002]. In particular, during winter months, the cooling leaves the surface waters relatively salty and dense and allows sinking of surface waters to occur at a few locations: in the Greenland Sea, the Labrador Sea, and possibly in the Irminger Sea associated with strong wind events [Pickart et al., 2003]. The sinking water forms the southward flowing branch of the MOC at depth and spreads throughout the World Ocean. Whereas the Labrador Sea is directly connected to the North Atlantic, the Nordic Seas and Arctic Ocean are separated from the North Atlantic by the Greenland–Scotland Ridge. Here southward flow at depth is confined to a very limited number of passages, predominantly the Denmark Strait and the Faroe Bank Channel downstream of the Faroe Shetland Channel [Blindheim and Østerhus, 2005].
 In a future climate forced by increased concentrations of radiative active gasses emitted through human activities, increased atmospheric temperatures are expected to result in an enhanced hydrological cycle, i.e., more evaporation at low latitudes and more precipitation at high latitudes. This increased precipitation in the high-latitude sinking areas will imply a dilution and freshening of the surface waters, which may act to weaken the Atlantic MOC and the associated meridional heat transport [Hansen et al., 2004], as also shown by a number of studies with coupled climate models [e.g., Gregory et al., 2005].
 This scenario finds support in the increasing discharge from Siberian rivers [Peterson et al., 2002] in addition to an increasing number of studies reporting significant freshening of the major Atlantic water masses associated with the lower branch of the Atlantic MOC during the last four decades [Dickson et al., 2002, 2003; Curry et al., 2003; Curry and Mauritzen, 2005]. Also, long-term hydrographic measurements give evidence of a long-term decreasing trend in the interface level of dense water in the Nordic Seas with possible impacts on the southward flow of dense water with origin in the Nordic Seas [Hansen et al., 2001]. However, the sensitivity of the deep overflows to reservoir changes is possibly relatively weak [Curry and Mauritzen, 2005; Wilkenskjeld and Quadfasel, 2005]. In addition, observations of sea surface height indicate a weakening of the subpolar gyre over the past decade, which may not be attributed to local wind stress changes [Häkkinen and Rhines, 2004]. The decline in the subpolar gyre may also be connected to increasing salinities of the inflow branches of Atlantic Water to the Nordic Seas reported in the recent years [Mortensen and Valdimarsson, 1999; Hàtùn et al., 2005]. As seen, observed multidecadal freshening of the deep overturning water masses of the North Atlantic may be early imprints of global warming, and the observed patterns reflect the combined effect of changes in the intensity of the hydrological cycle and the strength of the Atlantic MOC. In fact, some indirect evidence of a possibly dramatic recession of the Atlantic MOC during the last decades has recently been deduced from historical hydrographic data [Bryden et al., 2005].
 The picture is complicated by the fact that the dominant mode of atmospheric variability of the North Atlantic region, the North Atlantic Oscillation (NAO), persisted in its negative phase during the 1960s but has since then systematically changed toward the positive phase with enhanced westerlies across the North Atlantic. As both model studies and theory suggest that the circulation of the North Atlantic and the ocean climate of the Nordic Seas are sensitive to the phase of the NAO (see reviews by Visbeck et al.  and Furevik and Nilsen ), changes in water mass characteristics and circulation observed during the last decades should be seen in light of this shift.
 It is presently not settled whether the shift in NAO is itself connected to the changed radiative forcing of the troposphere because of increased concentrations of greenhouse gasses or should be regarded as a natural climate fluctuation. Recently, Yin  identified a global warming signal with northward-shifted storm tracks that project onto the natural NAO pattern. Earlier, Shindell et al.  argued that the observed trend in NAO could be simulated using known external forcings. However, natural or anthropogenic in origin, it is essential to assess to what extend the known atmospheric evolution including the long-term shift in the NAO can account for the observed changes in ocean climate of the North Atlantic.
1.1. Ocean Hindcast
 Atmospheric imprints on recent changes of the ocean climate can be addressed using ocean hindcast simulations in which past oceanic dynamics and properties are simulated for a given time period using an ocean general circulation model (OGCM) constrained in part by historical atmospheric surface forcings. A number of such studies have been designed to reconstruct and assess the past natural variability and recent trends of the North Atlantic and the Arctic Mediterranean [Haak et al., 2003; Karcher et al., 2003; Nilsen et al., 2003; Zhang et al., 2004; Bentsen et al., 2004; Drange et al., 2005; Marsh et al., 2005; Gerdes et al., 2005]. The reconstruction or hindcast period is given by the availability of atmospheric forcing fields, usually obtained from data sets like the NCEP/NCAR reanalysis product available from 1948 onward [Kalnay et al., 1996; Kistler et al., 2001]. Intercomparison of model results [e.g., Drange et al., 2005] is complicated by major differences in the experimental design between studies, typically in the initial state for the hindcast obtained via a spin-up procedure of the model or in the use of restoring boundary conditions. Model spin-up period can vary from a few decades to a few centuries, largely dictated by the resolution of the OGCM or set by the scope of the study.
 Some restoring of model thermodynamic surface fields toward climatological values is most often needed to prevent model drift, in particular of salinity. Such relaxation is known to distort the free modes of internal ocean variability in the models by introducing an artificial feedback on salinity, and the actual configuration may be important for the result. State-of-the-art OGCM configurations can in some cases produce hindcast simulations without significant drift in upper ocean properties, avoiding the artificial feedback by applying seasonally resolved but annually repeated corrections of the surface salinity [Nilsen et al., 2003].
 Because of the potentially long memory of the ocean, hindcast studies must be regarded as a combined initial/boundary value problem. But while the boundary consists of the atmospheric forcing condition, the initial value, i.e., the complete oceanic state on, say, January 1948, is largely unknown. This is a principal problem with the hindcast methods, which must be properly dealt with. Usually, a probable initial ocean state is obtained from spin-up procedures that, as explained, vary significantly in complexity among studies. Most seek to precondition the ocean initial state for a smooth transition to the hindcast experiment where realistic forcing history is applied. Thus, the central idea behind most experimental designs is to reach a quasi-equilibrium state of the model ocean. Among the most straightforward designs are the use of climatological atmospheric forcing fields, possibly with the addition of some synoptic variability. Also used are designs where a single year or sequence of years is repeated a number of times. An example of a more advanced procedure is found by Bentsen et al.  where the spin-up simulation is initially forced by climatological atmospheric fields, followed by repetitions of daily reanalysis fields for a 5-year period central to the subsequent hindcast. It is difficult to evaluate the role of the exact spin-up procedure for the hindcast results, but regardless of the complexity, the central idea of an ocean in quasi-equilibrium can be questioned [e.g., Wunsch and Heimbach, 2006].
 Instead, several studies deal with the problem of the unknown ocean state and time history by applying an ensemble approach to explore the robustness of the model results to initial conditions. Yet, care must be exercised in order to obtain truly independent ocean initial states for the ensemble members considering the large decorrelation time of oceanic characteristics. For instance, if ensemble members are performed end-by-end, i.e., repeating the hindcast simulation using the same atmospheric forcing but using the end of the previous hindcast as initial condition for the next [e.g., Haak et al., 2003; Bentsen et al., 2004; Drange et al., 2005], the series of initial conditions hereby produced share the same forcing history and are therefore strictly speaking not independent. Also, one inevitably forces the ocean with an artificial periodicity (= the length of each hindcast simulation), which creates dependence between ocean states. At best, the computationally attractive end-by-end approach likely gives an underestimated indication of the spread of modeled variables because of internal ocean variability. At worst, it could by design invalidate significant sequences of the hindcast depending on the decorrelation timescale of the ocean and the tendency of exciting internal modes of variability in the particular OGCM configuration.
 The ensemble approach designed and applied in this study focuses on eliminating the role of initial ocean conditions on the reconstructed ocean climate history. In combination with a sufficiently, and unprecedented, large number of ensemble members, robust statistics for the evolution and characteristics of the major exchanges of the North Atlantic are produced. The selected approach builds upon an independent control simulation forced by atmospheric fields from randomly permuted years to ensure that the forcing exhibits a white spectrum without any preferred frequencies. From this, practically independent ocean states can be selected for the ensemble members, as will be described in detail later. The design includes a millennial scale spin-up period in order to reach an overall quasi-equilibrium of the model dynamics and basin scale water mass properties. This facilitates a discussion of interrelated changes in the large-scale Atlantic MOC, water mass properties, and key exchanges with the Nordic Seas and Labrador Sea rarely addressed in comparable hindcast studies.
 In the following section, given is a short description of the ocean model, the experimental design, and the characteristics of spin-up and control simulations. In Section 3, ensemble statistics are presented for the Atlantic MOC and Section 4 introduces the ensemble mean climatology of North Atlantic–Arctic Mediterranean exchanges. Long-term changes are discussed in Section 5 followed by the summary and conclusions in Section 6.
2. Model and Experimental Design
 The ensemble approach and the experimental setup designed and applied in this study are based on a spin-up of the OGCM followed by a control simulation from which independent initial conditions for the ensemble members are picked.
2.1. The OGCM
 The global, coupled ocean, and sea-ice model system applied is the Max Plank Institute Ocean Model version 1 (MPI-OM [Marsland et al., 2003]). MPI-OM is a primitive equation model formulated on z-levels and using a free surface boundary condition with the usual hydrostatic and Boussinesq approximations made. A sea-ice model with viscous plastic rheology (Hibler type) is embedded. The particular model configuration has 30 vertical levels with layer thickness increasing nearly exponentially with depth below the surface layer. Relatively high horizontal resolution for the central sites of North Atlantic deepwater formation is achieved by placing the model North Pole on Greenland. Hereby taking advantage of the orthogonal curvilinear coordinates, the grid resolution varies gradually from 20 km close to Greenland to about 350 km in the tropical Pacific. Thus, the model is on average of relatively coarse resolution. Ocean subgridscale parameterizations include a bottom boundary layer scheme allowing for slope advection, rotated mixing tensors describing the isopycnal/diapycnal nature of ocean mixing, and the Gent and McWilliams parameterization for eddy-induced mixing. Convection in the model is described by enhanced diffusion; for details and description of bulk sea surface flux parameterizations based on atmospheric reanalysis data, see Marsland et al. .
2.2. Control Simulation
 The initial spin-up simulation started from an ocean a rest and with July temperature and salinity properties taken from the Polar Science Center Hydrographic Climatology [Steele et al., 2001]. The model was then integrated forward for 2500 years by applying a randomly permuted yearly time series of NCEP/NCAR reanalysis fields as forcing [Kistler et al., 2001]. This procedure is chosen in order to eliminate any preferred frequencies in the forcing to the largest possible extent. In the permuted forcing series, years are adjoined at Northern Hemisphere midsummer seeking to maintain a consistent forcing history throughout the North Atlantic wintertime, where both wind stress and buoyancy forcing are largest. This is of course at the expense of having a discontinuous forcing history during the Southern Ocean wintertime.
 In addition to bulk formulations of evaporation minus precipitation (E − P), model sea surface salinity (SSS) was forced by observed monthly mean discharge of the world's 50 largest rivers.
 Full three-dimensional Newtonian relaxation of the thermodynamic fields was applied for the first year of the spin-up while the barotropic dynamics adjust to the imposed density field. Hereafter only SSS was weakly relaxed toward climatology for both ice-free and ice-covered regions with an e-folding relaxation timescale of 180 days for a mixed layer depth of 17 m. Partly to prevent large flux corrections in areas of deep ventilation and formation of abyssal water masses and partly to inhibit artificial convection under ice, the e-folding scale was increased linearly with increasing mixed layer depth. This results in a strong seasonal variation of the effective E − P flux by the restoring procedure. The E − P flux correction is practically nonexistent during winter deepening of the mixed layer in large parts of the North Atlantic and Nordic Seas. This is justified by the strong seasonality in melting of sea-ice and drainage from land in the Arctic and Subarctic. No relaxation was used for temperature.
 After about 800 model years of spin-up integration with dynamic restoring as described above, drift in the characteristics of the deep Southern Ocean water mass properties was modest. Other major Atlantic water masses had already reached a quasi-equilibrium here including the intensity of the Atlantic MOC, and year 850 was chosen as the end of the spin-up phase.
 To avoid artificial feedbacks on salinity during the control simulation and the actual hindcasts, a monthly resolved but interannually invariant E − P flux adjustment field was constructed [e.g., Nilsen et al., 2003]. This was done by averaging the effective relaxation of SSS during a 200-yearlong period at the end of the spin-up experiment beyond year 800. For the control simulation in continuation of the spin-up simulation, relaxation of SSS was then replaced by the interannually invariant E − P correction in effect leaving the model ocean largely unconstrained. In all other respects, the control simulation is a continuation of the spin-up phase and continued until model year 2500.
 No apparent discontinuity in model dynamics could be identified in the transition from dynamic restoring, and for the first 50 years the mean state and variability of the control simulation resembled well the parallel spin-up integration also carried forward in time.
 After this apparently smooth transition to a largely unconstrained ocean, the control simulation is characterized by centennial scale excursions to a more energetic state with intensified convection, primarily in the Labrador Sea (Figure 1a) and a stronger Atlantic MOC at all latitudes south of Greenland (Figure 1b). In the phase of intensifying convection and Atlantic overturning, surface and subsurface salinities of the Nordic Seas and Labrador Sea (not shown) are increasing while temperature below the winter mixed layer shows warming associated with reduced convection. These large-scale settings are reversed in the opposite phase of weakening convection and decreasing Atlantic MOC. In general terms, the nature of this quasi-periodic variability is best explained by alternating build-up and breakdown of subpolar stratification in particular, coupled to the intensity of the Atlantic MOC. Such signatures of internal modes of variability are expected in the now less constrained model ocean.
2.3. North Atlantic Characteristics
 For a qualitative validation of model performance, North Atlantic characteristics and water mass properties at selected diagnostic sections are presented in Figures 2 and 3based on year 0–50 of the control simulation: the quasi-equilibrium state reached after the spin-up phase (Figure 1). In consideration of the coarse model grid and the millennial scale spin-up, reasonable agreement is found with estimated large-scale characteristics of the Atlantic. The intensity of the Atlantic MOC measured at 35°N as the strength of the upper cell of North Atlantic origin is about 12 Sv (Figure 2a) with a broad sinking region north of 50°N. Inflow of Antarctic Bottom Water to the North Atlantic occurs at depths below about 3000 m (Figure 2a), with a cross equator intensity of little less than 4 Sv. Intense deep convection is found in the Greenland Sea reaching the bottom layer of the model (Figure 2b). In the Labrador Sea, convection only occasionally reaches deeper than 2000 m during the first 50 years of the control simulation, and deep convection here is found to be a less robust feature of the model northern hemisphere winter than deep, bottom-reaching convection in the Greenland Sea. Strong mixed layer deepening is also found during a few years in the northern part of the Baffin Bay, and convection to intermediate depth occurs in the Iceland Basin and Irminger Sea along the northern branch of the subpolar gyre (Figures 2b and 2c).
 A somewhat stronger overturning of about 18 Sv was found in a comparable hindcast study also based on MPI-OM [Haak et al., 2003] and of 16 Sv in the reference state for MPI-OM obtained by Marsland et al. . Most likely, this discrepancy relies on less intense water mass transformation associated with deep convection in the Labrador Sea to be attributed to differences in the experimental designs and, in particular, to the different use of flux adjustment in between these and the present study.
 The subtropical gyre is displayed in the barotropic stream function (Figure 2c) with a strength of 32 Sv and aligns with the cyclonic subpolar gyre at about 48°N in the central Atlantic. A clear two-cell structure of the subpolar gyre is simulated with a Labrador Sea and an Iceland Basin center. Recirculation is strongest in the western cell. Some recirculation is seen within the eastern cell in the Irminger Sea linked to the northward flow in the Irminger Current along the western slope of the Reykjanes Ridge (Figures 2b and 3a).
 The model grid only allows for a coarse representation of the Faroe Bank, not the actual islands (Figure 3b). Nevertheless, all three branches of Atlantic surface inflow to the Nordic Seas can be identified from the simulated upper ocean velocity field (Figures 3a and 3c [Østerhus et al., 2005]): the Iceland Branch north of Iceland, the Faroe Branch between the Faroes and Iceland, in the model found close to the Icelandic shelf break turning east upon passage of the Ridge, and finally the Shetland Branch modeled as a broad inflow extending off the Scottish Slope.
 The coarse model bathymetry features two pathways for deep overflow exchanges with the Nordic Seas and the North Atlantic. These roughly represent the Denmark Strait and Faroe Shetland Channel with realistic sill depths of about 600 and 900 m, respectively (Figure 3c [Blindheim and Østerhus, 2005]). The model deep overflow water in the Denmark Strait is found at depths below about 300 m, clearly separated in temperature and salinity from both the northward flow of Atlantic Water and the relatively fresh waters of the East Greenland Current. Model deep overflow in the Faroe Shetland Channel exists at depths below 550 m and thus deeper than the Denmark Strait overflow and easily distinguished in water mass characteristics from the broad Atlantic inflow above. Atlantic Water is also present at the Barents Sea Opening with the core of the inflow resting on the western shelf slope but with Atlantic water dominating most of the section consistent with observations (Figures 3b and 3c [Ingvaldsen et al., 2004]). A branch of Atlantic Water also enters the Fram Strait (Figures 3b and 3c) where it is seen to be strongly cooled and entrained, and as in observations, model Atlantic Water is confined to the eastern part of the section at subsurface to intermediate depths [Fahrbach et al., 2001]. Export from the Arctic Ocean to the Nordic Seas is limited to the Fram Strait consisting of low saline surface outflow of Polar Water in the East Greenland Current on the Greenland Shelf and of bottom intensified deep outflow below the core of Atlantic Water of cold, relatively high saline water produced or modified in the Arctic Ocean (Figure 3c). The circulation of the Labrador Sea is largely horizontal and, at the selected diagnostic section, characterized by a gyre transport of the model Atlantic derived Labrador Sea Water and exchanges of Polar Water in the West Greenland Current and Labrador Current. In addition, deep horizontal circulation of eastern overflow derived waters can be identified from the model water mass properties (Figure 3c).
 By letting the model reach a quasi-equilibrium using only weak restoring on SSS, model salinity drifted toward values about 0.2–0.3 higher than the climatology of the Atlantic. Also, model deep waters, e.g., in the overflows from the Nordic Seas, are to warm by about 2°C. These shortcomings can be identified in the model water mass characteristics at the selected sections shown in Figure 3c. For the North Atlantic, the offset in salinity is nearly uniform with depth and not considered problematic for the simulated dynamics. The warmer than observed overflows may result in reduced baroclinicity in the model overflow systems though the effect is arguably small as density is only weakly dependent on temperature in this range.
2.4. Hindcast Simulations
 The preferred period of internal model variability is estimated from the millennial scale control simulation to be roughly 300 years. In turn, this means that initial conditions for the hindcast ensemble will have to span at least a millennial scale period to ensure that virtually independent model states are equally represented. With this in mind, a total of 26 hindcasts have been performed for the period January 1948 to June 2005, both months included, where initial conditions for the hindcast ensemble are chosen as “1. January” with 50 years separation from the control simulation. While the initial conditions are changed, the atmospheric forcing from the NCEP/NCAR reanalysis is identical for all members of the ensemble. The annually repeated correction of the E − P flux is unchanged from the control simulation, which should allow the ocean to respond to changes in fresh water forcing during the hindcast period. Annually varying observed river discharges from six major rivers with outlet in the Arctic Ocean were applied based on data obtained from the Global Runoff Data Centre (http://grdc.bafg.de/). These data were also used to construct the climatological monthly means also applied in the spin-up and control simulations, but the observed discharge time series include interannual variability and a long-term increasing trend since 1950 [Peterson et al., 2002].
 A second ensemble also with 26 members is performed using the same set of initial ocean conditions and the same atmospheric forcing but with observed fluxes of the six major rivers draining into the Arctic Ocean replaced by their climatology.
3. The Atlantic MOC
 By design, the present experiment facilitates a robust assessment of the long-term change in the Atlantic MOC, in particular the role of initial conditions versus atmospheric forcing. Temporal statistics for individual ensemble members and the ensemble mean are presented in Figures 4a and 4b for the Atlantic MOC at 35°N based on the ensemble applying observed Arctic river discharges. Two features are apparent by inspection, an overall decline and a strong interannual variability of more than 1.5 Sv (2σ, Figure 4b). The ensemble spread is about 3 Sv in 1948 reflecting the centennial scale variability of the control simulation (Figure 4b). The initial overshoot of the ensemble deviation relative to the control simulation is due to short-term intraannual fluctuations passed to the set of initial conditions for the hindcast simulations but not included in the estimate of the interannual variability for the control simulation in the figure.
 Throughout the hindcast period, the ensemble spread is gradually reduced but maintains a level well above the interannual level of variability. The decorrelation timescale of the MOC of the control simulation when filtered for centennial quasi-periodic variations suggests a significant oceanic memory of 10–20 years, which is significant compared to the length of the hindcast. This, however, also shows that the role of initial conditions on any decadal or multidecadal variability is weak at the end of the hindcast and that the persistent high spread (Figure 4b) is predominantly due to the internal variations of the model and the centennial scale mode. Overlaid on the long-term decreasing trend in the spread are periods of divergence and convergence between ensemble members on interannual to decadal timescales reaching up to 0.5 Sv (2Δσ, Figure 4b). If the performed ensemble is sufficiently large, divergence can be interpreted as periods where the model circulation is more weakly constrained by the atmospheric forcing (late 1960s to early 1970s and mid to late 1980s), allowing internal modes of variability to evolve. For the Atlantic MOC, these findings illustrate that the internal model variability on scales up to decades is likely equally problematic when hindcasting the evolution as is the unknown initial conditions. This questions the feasibility of producing a realistic hindcast and calls for an ensemble approach.
 The robustness of the modeled trend in the Atlantic MOC is illustrated in Figure 4c showing a histogram of linear trends of the MOC calculated from all ensemble members. It is seen to be shifted toward negative values relative to the similar histogram of trends calculated for random segments from the control simulation of length equal to the hindcast period. For the control simulation dominated by centennial variations, the average gives zero trend illustrating that it is not drifting. For the ensemble, the mean trend is −0.4 ± 0.5 Sv/decade, where the uncertainty is estimated as the width of the distribution based on a normal distribution fit. In the model, the wide range of the oceanic multidecadal response to the atmospheric forcing history is related to the spread of initial conditions (see below) and ultimately to the centennial mode of variability identified in the control simulation. This in turn highlight the inherent problem in assigning the role of, say, the atmospheric forcing on any observed long-term trend, though the model centennial scale mode may not have an analog in nature.
 In the present experiment, it is however possible to isolate in part the forcing related trend for each ensemble member by subtracting the background trend related to the centennial mode of the corresponding segment of the control simulation. The forcing related tendencies show a much more narrow distribution consisting of entirely negative or zero members (Figure 4c). The mean of the corrected distribution is also −0.4 Sv/decade but with some clustering of ensemble members seen. This is explained by different response to the forcing of members initialized on increasing or decreasing branches of the MOC, with the weakest trends of the ensemble members found for strong decreases in the MOC of the control simulation (not shown). In conclusion, the atmospheric forcing history during the period 1948–2005 would in an uncoupled sense have caused a decline of the Atlantic MOC in the order of 0.4 Sv/decade in case no other forcings or internal long-term ocean variability existed.
 Analysis of the ensemble using climatological Arctic river discharges returned quantitatively similar results in all respects. The evolution of the MOC and other key diagnostics of the individual members differed only in details between the two ensembles (not shown) illustrating the weak signal in the discharge history and, at the same time, highlighting the importance of the initial conditions.
 The overall quality of the precipitation field of the NCEP/NCAR reanalysis used in this study has been criticized [Trenberth and Guillemot, 1998; Serreze and Hurst, 2000], but also the alternative, the ERA-40 data, are problematic. Bengtsson et al.  conclude that ERA-40 contains spurious trends in the precipitation fields. This is in agreement with a recent work by Chen and Bosilovich  who found that the P − E is balanced in the NCEP/NCAR reanalysis but severely unbalanced in ERA-40 with a shift around 1972 linked to the assimilation of water vapor data from remote sensing. Based on this, NCEP/NCAR reanalysis data tend to be the better choice. As the results using the observed Arctic River discharge histories clearly illustrate that the impact of relatively subtle changes in freshwater forcing is negligible, it also seems unlikely that the quality of the precipitation fields applied has played a significant role for the modeled trend in the MOC.
4. Ensemble Mean Exchanges
 The hindcast ensemble simulation is analyzed with special emphasis on the North Atlantic–Arctic Mediterranean exchanges. In addition to key diagnostic parameters for the Atlantic circulation (e.g., the Atlantic MOC above), time series of exchanges of volume, ice, heat, and salt fluxes are calculated for the sections shown in Figure 3b. Ensemble mean and time mean results including interannual variability of exchanges are summarized in Table 1 and Figure 5, and at each section, volume fluxes are divided into major branches of exchange by water mass properties. The applied criteria are listed in the table, and for the overflows it differs from classic observation-based definitions based on density. Atlantic inflow and Polar water exchanges are typically defined in a similar but varying fashion. The choice here is internally consistent and serves the purpose of assigning all flow at the sections to one of up to three branches of exchange defined. Total heat and liquid freshwater transports are given in Table 2 for each section. All estimates in the tables are based on the ensemble simulation applying observed Arctic river discharges, but again practically identical results are obtained using climatological Arctic river discharges.
Table 1. Summary of Ensemble Mean Exchanges Between the North Atlantic and the Arctic Mediterranean for 1948–2005a
Exchanges are divided into water masses at each sections by properties. Net, mean northward volume fluxes are positive, R gives the rate of recirculation, and σ the interannual variability calculated from detrended time series with the annual cycle removed. Sections are shown in Figure 3.
T > 5°C and S > 34.8
Deep overflow water
T < 5°C and S > 34.8
S < 34.8
Iceland Faroe Section
T > 5°C
Deep overflow water
T < 5°C
Faroe Shetland Section
T > 5°C
Deep overflow water
T < 5°C
T > 3.5°C and S > 34.8
Deep overflow water
T < 3.5°C and S > 34.8
S < 34.8
Barents Sea Opening
T > 3.5°C
Deep overflow water
T < 3.5°C
No criteria used
Labrador Sea Section
Labrador Sea water
S > 34.8
S < 34.8
Table 2. Heat and Liquid Freshwater Transports Between the North Atlantic and the Arctic Mediterraneana
Northward (N) and southward (S) volume fluxes are given at key sections (Figure 3). Liquid freshwater transport values do not include the sea-ice contribution (Table 1). Observed heat fluxes in the Atlantic inflow from direct current measurements are given for comparison (bold) [Østerhus et al., 2005].
Heat flux referenced to 0°C (TW = 1012W).
Liquid freshwater transport relative to S = 34.8.
Iceland Faroe Section
Faroe Shetland Section
Barents Sea Opening
Labrador Sea Section
 Caution needs to be exercised when constructing and interpreting the ensemble mean, in particular if the ensemble members show tendency of clustering. To elucidate this issue, a careful analysis as for the MOC above has been performed for the different branches of exchange. The results show that there is little or no tendency of clustering that lends confidence in using the ensemble mean results for the North Atlantic region and that the ensemble mean is to some degree representative for the model physics.
 Interannual to decadal variability in the ensemble mean time series of exchanges is somewhat reduced compared to the individual ensemble members as could be expected as free modes of model ocean variability will ideally be masked when constructing the ensemble mean. Besides, validation against partly unpublished observational data series from the Greenland–Scotland Ridge area collected within the ASOF-E FP5 project MOEN shows a realistic level of variability in the ensemble mean series of Atlantic inflow and deep overflow, in particular for the well-established time series of the eastern deep overflow [Hansen et al., 2001]. This is true on interannual as well as monthly timescales (S. Østerhus et al., personal communication, 2006).
4.1. The Greenland–Scotland Ridge Exchanges
 For all three branches, modeled volume fluxes of Atlantic inflow to the Nordic Seas agree well with direct observational estimates from the last decade [Østerhus et al., 2005]. This includes the modeled Iceland Branch inflow of 0.7 Sv relative to the observed estimate of 0.75 Sv and modeled Faroe Branch and Shetland Branch inflow of 3.8 and 4.2 Sv, respectively, both with an observational-based estimate of 3.8 Sv.
 Model eastern deep overflow is only found in the Faroe Shetland Channel with a mean flux of 3.1 Sv. Consistent with this model result, estimates of the combined eastern overflow suggest a flux of 3 Sv [Hansen and Østerhus, 2000]. This includes a robust estimate of the fairly stable and persistent flow through the Faroe Bank Channel of 1.9 Sv [e.g., Hansen et al., 2001] and less robust estimates of the occasional spills over the Wyville Thomson Ridge, also downstream of the Faroe Shetland Channel. In addition, the transport through a number of narrow gaps in the Iceland Faroe Ridge possibly contributes about 1 Sv. In dramatic cases, overflow water cascading over the Wyville Thomson Ridge can reach 50% of the deep overflow through the Faroe Bank Channel [Sherwin and Turrell, 2005], but a good estimate of the mean value is lacking though it seems unlikely to exceed a few tens of a Sverdrup. The Iceland Faroe Ridge overflow is perhaps the least constrained and most sparsely monitored part of the deep overflow system from the Nordic Seas. The most recent flux estimates range from 0.7 Sv and fairly stable [Perkins et al., 1998] to only 0.1–0.2 Sv at most and with an intermittent nature [Saunders, 1996]. Both estimates are considered uncertain and are based on only 1 year of deep current observations approximately.
 Consistent with the observational data from the Faroe Bank Channel, modeled eastern overflow is persistent, not recirculating, and with an interannual variability of only 0.2 Sv. The possible intermittent nature of the Wyville Thomson Ridge overflow is not represented in the coarse model grid lacking the required details of the topography as well as the mesoscale eddy features of the circulation important for the intermittent overflow. It is also noteworthy that the model features a weak intermittent southward overflow across the Iceland Faroe Ridge of 0.1 Sv with properties of model modified intermediate water of the Nordic Seas: a flux that match the conservative observational estimate [Saunders, 1996].
 A near balance of the net flow across the Greenland–Scotland Ridge is modeled via a deep transport of 2.8 Sv in the Denmark Strait deep overflow and a flux of Polar Water of 2.5 Sv in the East Greenland Current. Both this surface and deep outflow are associated with a significant interannual variability of roughly 0.6 Sv though still persistent. The recirculation rate of surface and deep outflows estimated as the ratio between northward and southward transports is less than 10%. Recent direct measurements of the overflow from a 4-year campaign reveal a high level of interannual variability in the flow, ranging from 3.1 to 3.7 Sv [Macrander et al., 2005]. This contradicts previous estimates showing a rather stable transport on such timescales of 2.7–2.9 Sv [Dickson and Brown, 1994]. Apparently, the model mean overflow compares favorable with the classic transport estimates [Dickson and Brown, 1994] but the modeled variability of 0.6 Sv (2σ) is in the range of variability seen in the 4-year observational record [Macrander et al., 2005].
 At the latitude of the Denmark Strait, estimates of the transport in the East Greenland Current are based on hydrography only or result from budget considerations for the Arctic Mediterranean [Hansen and Østerhus, 2000]. These include values for the Bering Strait inflow and the surface outflows through the Canadian Archipelago. From such considerations, Hansen and Østerhus  give a flux of 1.3 Sv in the East Greenland Current but note that this seems somewhat low compared with the available indirect estimates based on hydrograhic surveys. Such exchange is also significantly lower than the modeled flux of 2.5 Sv, which is, however, comparable to the historical estimates [see Hansen and Østerhus, 2000; Pickart et al., 2005, and references herein).
 Observational constraints also exist on the net heat flux associated with the Atlantic inflow for all three branches at the Greenland–Scotland Ridge [Østerhus et al., 2005]. At the Denmark Strait, modeled recirculation of the defined water masses is small, and modeled northward heat transport across the section of 23 PW is directly comparable to the observational estimate of 22 PW in the Atlantic inflow. Also, at the Iceland Faroe section, intermittent overflow and recirculation of Atlantic Water play a minor role for the net heat transport and the modeled value of 127 PW compares favorably with the observational estimate of 134 PW. Østerhus et al.  give a heat transport in the Shetland Branch of 156 PW that is somewhat lower than the modeled northward heat flux at the Faroe Shetland Section of 175 PW when corrected for recirculation of Atlantic Water (10%). This can be explained by excess flow in this branch of 0.4 Sv relative to the observations. No reliable observational estimates exist for the liquid freshwater transport that compare with the calculated modeled fluxes.
4.2. Exchanges With the Arctic Ocean
 Of the total inflow of Atlantic Water across the Greenland–Scotland Ridge, 40% enters the Arctic Ocean via the Barents Sea Opening and the Fram Strait and carrying about 25% of the heat content at the Greenland–Scotland Ridge. Deep and intermediate outflow from the Arctic Ocean in the Fram Strait of 1.8 Sv is partly compensated by a net poleward flow to the east of 0.4 Sv of water 3–5°C colder than the inflow of Atlantic water and characterized as model intermediate or surface waters of the Nordic Seas. In terms of the volume flux, one quarter of the deep overflows across the Greenland–Scotland Ridge hereby originates from the Arctic Ocean.
 Modeled recirculation of Atlantic water at the Fram Strait section is small, but the interannual variability of 0.6 Sv is significant compared to the mean inflow of only 1.3 Sv. For the deep outflow, recirculation makes up nearly 50% of the net flow that is also associated with large interannual variability. The most detailed observations in the strait to date suggest that a considerable amount of Atlantic Water is returned southward in a modified form [Fahrbach et al., 2001; Schauer et al., 2004; Walczowski et al., 2005]. When correcting for recirculation, Schauer et al.  find that the net flow of Atlantic derived water can be either northward (1.6 Sv) or southward (−0.75 Sv) in individual years. These observations reveal a strong variability consistent with the model results but are otherwise hard to compare against, partly because of somewhat different definitions of water masses. Modeled Atlantic inflow of 1.3 Sv does, however, compare well with the measurements reported from the Svalbard Branch farther to the north showing a weakly varying inflow of 1.5–1.7 Sv [Schauer et al., 2004]. The model climatological value of the Atlantic inflow of 1.3 Sv is also in the range of the recent transport estimates by Walczowski et al. . Based on direct current measurements, the authors derive transport values between 1.2 and 3.5 Sv when using a definition of Atlantic Water comparable to the one used here.
 Atlantic inflow to the Barents Sea is not persistent on monthly timescales and in observations generally characterized by large interannual variations around a mean value of 1.5 Sv [Ingvaldsen et al., 2004], somewhat weaker than the modeled mean inflow of 2.2 Sv. Modeled Atlantic inflow is found also to be strongly variable in intensity but at odds with the observations, reversals are not seen.
 Observed yearly averaged heat advected with northward currents in the Fram Strait varies on an annual scale from 32 to 55 TW with a mean resembling modeled northward transport of 40 TW [Schauer et al., 2004]. However, a net observed southward heat flux between 16 and 41 TW exceeds the modeled value of 20 TW, which may be explained by limited recirculation of Atlantic Water in the model.
 Deep and intermediate outflow from the Arctic Ocean in the Fram Strait are strongly recirculating reflecting a strong barotropic nature of the flow in the part of the strait extending off the Greenland Shelf [Fahrbach et al., 2001]. This also explains the relatively high level of interannual variability seen for Atlantic Water and deep outflow in the model of 0.3 and 0.5 Sv, respectively, where also the strong recirculation seen in observations agrees qualitatively with the model results for the deep outflow. Schauer et al.  deduce a net southward outflow of 2.3–2.4 Sv in the Fram Strait when excluding Atlantic Water types and Polar water that, being only about 0.5 Sv higher, is consistent with the modeled deep outflow. In contrast, Fahrbach et al.  give a net flow to the south of about 2–6 Sv and in turn describe a much more intense exchange system than modeled.
 The Canadian Arctic Archipelago and the Fram Strait are the dominant pathways for surface outflow of the Arctic Oceans to the Atlantic. Transport of Polar Water in the East Greenland Current (1.4 Sv) is persistent in the model though slightly higher than the reported flux of 1.0 Sv over 3 years [Schauer et al., 2004]. Export of 1.0 Sv of Polar Water to the west is within the range of the existing estimates ranging from 0.7 to 2.5 Sv, where the lower value may well be underestimated [Prinsenberg and Hamilton, 2005].
4.3. Labrador Sea
 Circulation of the Labrador Sea is characterized by a 20-Sv gyre transport of the model, Atlantic-derived Labrador Sea Water and bounded by exchanges of Polar Water in the upper layers. Deep horizontal circulation of eastern overflow derived waters can be identified in the model water mass properties (Figure 3c) but are not diagnosed separately from the circulation of model Labrador Sea Water.
 The net flow of Polar Waters is 1.7 Sv southward, in part representing the flux of Polar Water entering through the Canadian Archipelago. Some entrainment of Atlantic water occurs north of the section leaving a net flow of 0.5 Sv northward of model Labrador Sea Water. The West Greenland Current carrying entrained Polar Water, modified by surface forcing and melting of sea-ice, is roughly doubled in strength from the Denmark Strait. When joining with the Polar Water exported through the Canadian Archipelago and returning southward at the western branch of the section in the Labrador Current, the Polar Water makes up a quarter of the gyre circulation. The resulting net southward freshwater transport of 91 mSv is in close agreement with the observed contribution from Polar Water and sea-ice in the Davis Strait [Cuny et al., 2005]. The modeled interannual variability of both the gyre transport of Labrador Sea Water and especially the Polar Waters is large and comparable to the net southward transport at the section.
5. Long-Term Changes
 Ensemble mean, time mean fluxes discussed in detail above, and thus the relative distribution between individual branches of exchange are found to compare favorably with available observational estimates from various different time periods. Also, interannual variations are realistically captured for some of the exchange branches where long time series are available (see Section 4).
 From here, it is not straightforward to infer that also decadal variability and any long-term tendencies of ensemble mean exchanges represent a robust response to the forcing with connection to changes occurring in Nature. However, partial validation on both monthly to interannual timescales as well as the long-term mean is taken as strong corroborating evidence. To further strengthen this line of argumentation, Section 5.2 presents a qualitative discussion of modeled changes in hydrography and characteristic patterns of observed changes during the last four decades.
 Time series of region-wise convective activity and exchanges with the Arctic Mediterranean and the Labrador Sea are compared in Figure 6 with the evolution of the Atlantic MOC. Ideally, the ensemble approach should eliminate any trend component from internal ocean dynamics, leaving any long-term trend to be related to changes in the forcing. For the region of interest, the forcing changes are best represented by the time history of the NAO. Characteristic for the hindcast period 1948–2005 is the long-term sustained shift in the state of the NAO in the 1970s from predominantly negative indices in the 1960s to strong positive NAO years in the 1990s ([Hurrell et al., 2001], Figure 6g).
5.1. North Atlantic Exchanges
 Inspection of Figure 6 suggests that the long-term negative trend in the MOC at 35°N (Figure 6a) discussed in Section 3 (Figure 4) to be linked to changes in the intensity of deepwater formation by convection in the Nordic Seas and the Labrador Sea (Figure 6b). Modeled history of convection in the Nordic Seas shows a marked decrease since the early 1970s where a maximum depth is reached. This evolution compares well with known NAO-induced changes of the Nordic Seas ocean climate [Dickson et al., 1996]. For the Labrador Sea, the modeled long-term decline in convective depth is apparently at odds with the established view describing a close coupling to the NAO where the Labrador Sea evolved from tightly capped in the 1960s to a state of deep-reaching convection by the early 1990s [Dickson et al., 1996]. However, modeled decadal variations do reflect changes in the NAO, including a decreasing tendency until 1970, a marked deepening during the early 1970s, and some recovery during the long high NAO period of the 1980s and early 1990s. Such direct coupling between the Labrador Sea convection and the NAO has previously been established in a model context [Gerdes et al., 2005]. Recently, this simple picture has been questioned based on analysis of a coupled model simulation [Oka et al., 2007]. Here it is shown that a significant part of the variability in the Labrador Sea convection is uncorrelated with the NAO but is instead driven by variations in the export of ice through the Denmark Strait.
 In contrast to the Atlantic MOC (Figure 6a), overturning exchanges with the Nordic Seas are found to be relatively stable in terms of the total Atlantic inflow (Figure 6d) and combined deep overflow (Figure 6e). Direct observations of the three branches from the last decade of Atlantic inflow do not reveal a significant correlation with the NAO on interannual timescales [Hansen et al., 2003]. Coupling on decadal timescales is, however, suggested by model results, where increasing inflow has been found during the strongly positive NAO years of the 1990s [Karcher et al., 2003; Zhang et al., 2004]. Note that the present result is different from the findings of both Nilsen et al.  and Zhang et al.  showing respectively a reduced and intensified Atlantic inflow associated with the recent shift in the NAO, whereas it is unclear to what extent the transport time series given by Drange et al.  represents Atlantic water. As a caveat, it should be pointed out that direct correlations calculated for the full hindcast between the individual exchange branches and the NAO are generally weak if excluding the long-term trends and even including any trends; correlations are not very significant considering the limited length of the hindcast and the autocorrelation of the time series. This is an intrinsic problem and should be kept in mind.
 Farther north, combined Fram Strait and Barents Sea Atlantic inflow to the Arctic Ocean (Figure 6c) tend to better reflect a decreasing trend in the early decades of low NAO and an increase in the recent decades exceeding the early decrease. Decadal changes are mainly found in the inflow to the Arctic through the Barents Sea opening and partly balanced by an increasing deep outflow in the Fram Strait (individual time series are not shown). This anomalous response may be expected as a result of an intensified Norwegian Atlantic Current in the positive NAO phase [Hansen et al., 2003] but is apparently not manifested by a long-term increasing inflow across the Greenland–Scotland Ridge in the model. Drifter data tend to support both the modest response of the Atlantic inflow at the Greenland–Scotland Ridge to NAO changes and the generally stronger boundary currents within the Nordic Seas during high NAO conditions that impact on the exchanges with the Arctic Ocean [Jakobsen et al., 2003].
 In closer detail, a clear imprint of the NAO on the tendencies of the North Atlantic meridional exchange can, however, be identified for all branches of the Atlantic inflow to the Nordic Seas in the latter segment of the hindcast from the 1970s to the present. This is illustrated in Figure 7 comparing a cumulated NAO index reflecting periods of persistent forcing with annual tendencies of net exchanges. Since the late 1960s, the tendencies of both the Atlantic inflow to the Nordic Seas and Arctic Ocean closely follows the constructed index with a delay of about 1–2 years, with the longest response in the north. In the 1950s into the 1960s, the evolution of the inflow to the Arctic Ocean still mirrors the inflow to the Nordic Seas, delayed about 1 year, but decoupled from the NAO history. It is most likely that the lack of correlation in the early part of the period simply reflects the inability of the station-based NAO index to capture the sea level pressure variability in the northern part of the region in this period as, in the 1950s and 1960s, the characteristic pressure patterns associated with the NAO of the Nordic Seas was positioned farther to the west (for recent reviews, see Visbeck et al.  and Furevik and Nilsen ). Alternatively, the lack of correlation prior to the 1970s is simply a result of the long autocorrelation timescale identified in the model, for the MOC of at least 10–20 years (see Section 3).
 Consistently, since the mid-1970s, also decadal tendencies in the Atlantic MOC at 35°N mirror the constructed index (Figure 7) whereas any apparent correlation between the NAO and the direct strength of the MOC tends to break down also in the mid-1990s (consult Figure 6). In turn, this suggests a 1- to 4-year response time of the MOC to the forcing. Variations of the inflow to the Nordic Seas lag the MOC by about 1 year and the Arctic Ocean by about 3 years. This propose a baroclinic adjustment in the Nordic Seas rather than a fast barotropic mode consistent with a traveling time of about 3 years of first mode planetary waves to cross the Nordic Seas. In the earlier part of the hindcast period, the MOC is decoupled from the other exchanges as well as the constructed NAO index. It is unclear whether this can also be linked to changes in the characteristics of the NAO discussed above or simply reflects a somewhat longer memory of the Atlantic MOC relative to the exchange branches in question.
 The apparent decoupling in the model of long-term variations of the Atlantic MOC (i.e., the multidecadal declining trend) and Nordic Seas exchanges is already suggested from the centennial variations in the control simulation (Figure 1). Here a strong response is seen in the MOC linked to the changes in (Labrador Sea) convection. Again, relative variations in the overturning exchange with the Nordic Seas are small. This may suggest a weak role of the volume exchange with the Nordic Seas on the strength of the MOC during the period.
5.2. North Atlantic Freshening
 In contrast to the relatively stable overturning exchange across the Greenland–Scotland Ridge, modeled water mass properties of the Atlantic inflow and deep overflows of the Nordic Seas show significant concurrent changes (Figures 8a and 8b). The model results suggest that the salinity of the Atlantic inflow never recovers from the freshening associated with the Great Salinity Anomaly of the 1970s [e.g., Belkin et al., 1998]. The anomaly of the 1970s is well represented in the model and can be traced back to an excess freshwater pulse from the Arctic in the late 1960s of relatively short duration [see also Haak et al., 2003]. In the present study, pulses of freshwater are seen at the Fram Strait in 1968/1982/1989 and 1995 (Figure 8d). Despite being highly variable, no significant trend is found in the southward liquid freshwater transport associated with Polar Waters and sea-ice at either side of Greenland (Figure 8d). Instead, a sustained freshening of the Atlantic inflow is modeled (Figure 8a), which is eventually communicated to the deep overflows at the Greenland–Scotland Ridge by deepwater ventilation (Figure 8b). The freshening of the Atlantic inflow represents a large-scale shift in the circulation of the subpolar North Atlantic around the 1970s affecting the properties of the inflow [Hàtùn et al., 2005].
 This view is supported by the modeled gradual increase in the circulation of Polar Water in the Labrador Sea gyre (Figure 8c) unaccompanied by any steady increase of the freshwater transport of the source waters as described above (Figure 8d). The increase partly occurs on the expense of model Labrador Sea Water (Figure 6f) and consequently manifested in a marked increase in the amount of liquid freshwater circulated within the gyre (Figure 8d). The most prominent changes are seen to follow about a year after excess freshwater outflow from the Arctic in the late 1960s, late 1980s, and middle 1990s. This finding is consistent with the conclusion reached by Curry and Mauritzen  when analyzing the temporal evolution of freshwater storage in the subpolar region based on historic hydrographic data. Note also that three low convection events in the Labrador Sea follow immediately after each of these pulses partly explaining the modeled long-term decline in convection here [Oka et al., 2007].
 As speculated, these modeled changes represent a long-term freshening of the subpolar gyre and Nordic Seas. To illustrate this, difference maps of model salinity distribution and barotropic stream function are constructed as mean 1980–2005 minus 1948–1975 fields (Figure 9). The selection of periods is guided by the discussion of the time series above and the well-known changes in the NAO including an eastward shift in the spatial pattern in the 1970s and the pronounced shift occurring from the weak westerlies in the 1960s to the strong westerlies in the 1990s [e.g., Furevik and Nilsen, 2005].
 Changes in the barotropic stream function reflect a general intensification of the southern branch of the subpolar gyre (Figure 9c). This stronger zonal flow at the gyre–gyre boundary derives from a stronger meridional baroclinic density gradient and contributes to an intensified cyclonic circulation, in particular of the eastern subpolar gyre. The decline in the circulation of Atlantic derived water across the Labrador Sea section (Figure 6f) only partly compensated by increased circulation of polar water (Figure 8c) is seen to be linked to a displacement toward the southeast of the center of the western cell.
 Parallel with the decreasing convection in the Nordic Seas, the cyclonic circulation of the Greenland Sea and the Iceland Sea is reduced notably. In the Norwegian Sea, weak anomalous cyclonic circulation gives way for the enhanced inflow of Atlantic Water to the Arctic Ocean via the Barents Sea (see Figure 6c for the total inflow). This enhanced export of Atlantic water out of the Nordic Seas in combination with modest changes in the inflow across the Greenland–Scotland Ridge and a steady flux of polar water in the Fram Strait leaves the Polar Water in the eastern Greenland Sea and Iceland Sea less entrained by Atlantic Water. At the surface, this results in a strong freshening of about 0.3–0.4 psu in the East Greenland Current system south of the Fram Strait. Similar strong surface freshening is found in large parts of the subpolar gyre, in particular in the northern part of the Labrador Sea associated with the displacement of the western cell, with the intensification of the southern branch of the gyre, and with an anomalous closed cyclonic circulation in the Iceland Basin. At depth, the strongest freshening is seen south of the Greenland–Scotland Ridge associated with the anomalous barotropic circulation in the Iceland Basin, most clearly seen below the sill depths of deep overflows (Figure 9e).
 Modeled changes of the properties of the deep overflows reflect the largely uniform freshening of the Nordic Seas (0.015 and 0.012 per decade for the eastern and western deep overflow, respectively; Figure 8b), which compare favorably with the observed rate of freshening of about 0.010–0.015 per decade since the mid-1970s [see Dickson et al., 2002, and references herein]. Also, enhanced freshening in the Iceland Basin (Figure 9d) is indicated by these observations, where a rate in excess of 0.03 per decade is seen at intermediate depths southeast of Iceland. In fact, this is by far the largest signal in the observations analyzed. The stronger subpolar gyre freshening is also in qualitative agreement with the study of Curry and Mauritzen  describing the dilution of both the subpolar basins and Nordic Seas since the 1950s using depth integrated salinity changes. A distinct difference between the Nordic Seas and the subpolar basins is found with the subpolar storage of freshwater about threefold of the storage of the Nordic Seas. This is in qualitative agreement with the pattern of freshening modeled though the storage is not explicitly diagnosed.
 On large scales, modeled dynamic changes act to isolate the water masses of the subpolar gyre from subtropical waters. In consequence, a salinification of the subtropical waters accompanies the northern freshening (Figures 9c and 9d). This pattern is also an expected manifestation of the reduced strength of the Atlantic MOC. A similar structure of change of the Atlantic exists in hydrographic observations over the past decades [Curry et al., 2003]. Since the 1950s, upper and thermocline waters of the subtropical North Atlantic became more saline by up to +0.4 in the upper ocean. A strong decrease with depth is seen in data where the signal of salinification is absent below thermocline depths. This span and depth variation match well the modeled changes (Figures 9c–9e), suggesting that the observed freshening may well in part be explained by a modest reduction of the Atlantic MOC and in part by dynamic changes of the subpolar gyre. In turn, this emphasizes the potential role of local dynamic changes on the observed signal of freshening relative to changes in surface forcing and source water composition.
6. Summary and Conclusions
 The constructed ensemble mean climatology of exchanges at key sections of the Arctic Mediterranean (Figure 5) is shown to compare well with available observational estimates of the individual branches. This experiment hereby add to the limited number of model studies with such focus, where the most complete is possibly the model–model and the model–data intercomparison study by Drange et al. . The ensemble approach used in the present study is novel and the number of water masses addressed in the flux calculations goes beyond existing studies. In combination with a multimillennial spin-up of the model, this facilitated a unique view of the interrelated changes of the Nordic Seas exchanges, the Atlantic MOC, and the characteristics of North Atlantic salinity changes during the last decades. Despite the relatively coarse model resolution, the quality of the model climatology of exchanges is perhaps among the most consistent available for the Nordic Seas.
 Since the 1960s, decadal variations of the Atlantic MOC, the Atlantic inflow to the Nordic Seas, as well as to the Arctic Ocean are shown to respond systematically to variations in the NAO (Figure 7) though correlations on interannual to decadal scales are not very significant given the limited length of the hindcast. The long-term tendency of the overturning exchanges with the Nordic Seas across the Greenland–Scotland Ridge is, however, relatively weak. Consistently, modeled freshening of the Nordic Seas in the recent decades is partly communicated via the freshening of the Atlantic inflow and partly results from enhanced inflow of Atlantic Water to the Arctic Ocean.
 Characteristics of modeled freshening in the subpolar basin and Nordic Seas bear many similarities with observed changes since the 1950s. Whereas freshening in the hindcast simulations is predominantly due to dynamical changes of ocean circulation linked to the NAO, observed changes are possibly also influenced by excess meltwater from glacier ice [e.g., Curry et al., 2003; Peterson et al., 2006].
 In the model, the freshening shows the expected pattern of change resulting from a reduction in the MOC and a longer residence time of water masses in the northern North Atlantic. This yields an effective enhancement of the subpolar freshwater storage as observed. The intensified circulation at the gyre–gyre boundary between the subtropical and the subpolar gyre in the later part of the hindcast period of strong westerlies cannot alone account for the widespread accumulation of freshwater but is an expected feature of high NAO conditions and recently highlighted as an important dynamic system for modifying the properties of the Atlantic inflow to the Nordic Seas [Hàtùn et al., 2005].
 The modeled trend in the Atlantic MOC of −0.4 Sv per decade at 35°N or about 2–2.5 Sv during the reanalysis period (Figure 4) is much less dramatic than the 50% reduction since 1957 (8 Sv) deduced from historical hydrographic data at 25°N [Bryden et al., 2005]. In major contrast to the findings here of nearly unchanged meridional overturning exchanges with the Nordic Seas, the analysis of hydrographic data attributed the changes in the MOC to the eastern overflow system, i.e., the Denmark Strait and Faroe Shetland Channel deep overflows. The method applied by Bryden et al.  is associated with large, unknown uncertainty and the robustness of the result is questionable (for a discussion, see Wunsch and Heimbach ). Other recent estimates of change of the Atlantic MOC span from model hindcast result showing increasing overturning [Bentsen et al., 2004; Marsh et al., 2005] to weakly decreasing overturning during the last decade based on a data-constrained model system [Wunsch and Heimbach, 2006]. In contrast to the present result, it has also been shown in a coupled model context that the subpolar freshening, in particular in the Labrador Sea, can be explained by increasing convection here coupled to an increasing Atlantic overturning [Wu et al., 2004].
 Low-frequency variability of the Atlantic MOC, such as the Atlantic Multidecadal Oscillation (AMO) [e.g., Delworth and Mann, 2000; M. Latif et al., 2004], could a priori be responsible for the observed or diagnosed trends in the marine climate of the North Atlantic since the 1950s. In the present experimental setup, it is however possible to isolate in part the forcing related trend from any trends relating to internal ocean dynamics, such as the AMO. In turn, the present study suggests that the gross characteristics of the recent changes in the North Atlantic can be explained with the known forcing history in terms of the atmospheric reanalysis products alone. Model results also show that the effect of increasing discharge of Arctic rivers is insignificant in comparison with the dynamic changes induced by the variations of the atmospheric forcing.
 This work was supported by the European Commission as part of the MOEN (Meridional Overturning Exchange with the Nordic Seas) contract EVK2-CT-2002-000141.