We use an updated version of the 1/12° regional Arctic-North Atlantic configuration CREG12 Canadian Regional;. CREG12 is based on the ocean modelling platform Nucleus for European Modelling of the Ocean (NEMO) version 4.2.2 and the Sea Ice modelling Integrated Initiative 3 (SI3) sea ice model, with levitating sea ice, five categories of ice, and two layers of snow . The model is run on an ORCA12 seamless regional grid with horizontal resolution ≈3–4 km in the central Arctic . It uses a z* vertical coordinate with 75 levels spaced by 1 m at the surface and 150 m at 1500 m. This relatively fine horizontal grid size allows for an explicit resolution of most of the mesoscale spectrum within the deep basins where the first Rossby radius of deformation R0 is ≈10–15 km, but not over the continental slope and shelf where R0<7 km see also Fig. S1 in the Supplement. Higher resolution simulations of the Arctic Ocean (≈1 km) have shown that the EKE spectrum peaks around 50 km and that more than 80 % (resp. 65 %) of the EKE is contained in scales larger than 10 km resp. 20 km;.Therefore, we argue that 1/12° is a resolution fine enough to represent most of the mesoscale features in the BG and along its margins (but not over the shelves), with an associated computational cost that allows for decadal integrations. The configuration includes a third-order momentum flux formulation, a second-order scheme for tracer advection, with an additional bi-Laplacian viscosity and diffusivity formulation depending on the local velocity, and a turbulence closure scheme for vertical mixing. The representation of tidal mixing effects is included in the comprehensive parameterization of mixing by breaking internal tides and lee waves .

The simulation is initialized in 1979 from the World Ocean Atlas 2009 for temperature and salinity with the ocean at rest and is run until 2020. Sea ice conditions are initialized from the Pan-Arctic Ice Ocean Modeling and Assimilation System PIOMAS;. The ocean and sea ice are forced with hourly atmospheric fields from the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 ERA5,. To compensate for the known warm biases of the sea ice surface temperature of ERA5 e.g., the snow conductivity is set to 0.5 W m⁻¹ K⁻¹, the ice-ocean drag coefficient to 7×10-3, the atmosphere-ocean drag coefficient to 1.2×10-3, and the ice strength to 2×10-4 N m⁻². The open boundary conditions at the Bering Strait and along 27° N in the Atlantic are specified daily from the output of GLORYS12V1, a global reanalysis at 1/12° resolution run from 1993 to 2020 . Prior to 1993, outputs of GLORYS12V1 between 1993 and 2021 are used to build a daily climatology and force the open boundaries of CREG12. At the Bering Strait, meridional velocities are adjusted to constrain the inflow to about 1.1 Sv, matching observation estimates . The river run-off and Greenland melting are specified following . An additional sea surface salinity restoring with piston velocity of 167 mm d⁻¹ is implemented in ice-free regions at monthly frequency using the World Ocean Atlas 2009 . For additional details on the run, the reader is referred to .

In the rest of this paper, the Canadian Basin is defined as the region between 69–85° N and 108–180° W, thus fully encompassing the BG and its surrounding area. For analysis purposes, we define the Beaufort Box (BB), a region in the BG between 73–77° N and 135–152° W (see Fig. ).

We present here a brief evaluation of the model's representation of the hydrography, circulation, and sea ice conditions in the Canadian Basin. For a more in-depth assessment of the model's performance, the reader is referred to and who use similar configurations. In this study, we focus on the period 1995–2020 to let the model equilibrate between 1979 and 1994. Over the period of analysis, the mean September sea ice concentration is comparable to that derived from satellite observations, with small differences on the Eurasian shelf and a low bias in the western Canadian Basin (Fig. a and b). Across the Arctic and along the simulation, the sea ice extent deviates from that derived from satellite observations by around -7 % in September and -16 % in March on average. When compared to the PIOMAS reanalysis , the sea ice thickness is 35 cm thinner in September and 20 cm thinner in March (Figs. S2b and S3b). The interannual variability of sea ice extent is well captured by the model across the 26 years of simulation. A strong decline in sea ice concentration starting around 2000 and persisting in time appears in the model, in agreement with observations (Figs. S2c and S3c). The corresponding location of ice loss is generally well represented despite some biases in the ice concentration along the Eurasian shelf in summer, and high biases along the Yermack plateau and the Greenland eastern shelf in winter (Figs. S2a, b and S3a, b).

The Sea Surface Height (SSH) patterns are comparable in the model and observations (Fig. c and d), indicating that the surface geostrophic circulation of the model correctly reproduces the circulation of the BG (anticyclonic) and of the Nansen Basin (cyclonic). Within the BG, CREG12 successfully represents the vertical distribution of temperature extrema associated with the three main water masses present in this region (Fig. e and f), namely the sPW (temperature maximum at 100 m), the winter Pacific Waters (wPW; temperature minimum at 200 m), and the AW (temperature maximum at 550 m). Small biases in the magnitude of the temperature extrema themselves (warm bias for the wPW, and cold bias for the sPW and AW) are noted. Despite a high salinity bias at the surface in CREG12, the modelled stratification displays the so-called “bowl shape” of the BG, visible through the tilted isopycnals along the edges of the gyre, although slightly weaker in the northernmost side of the BG in CREG12. The overall fresh water content, referenced to 34.8 g kg⁻¹, shows a strong increase between 2003–2009 in the Canadian Basin as documented from the Beaufort Gyre Exploration Project followed by a plateau (Fig. S4).

Overall, the model offers a realistic representation of the main circulation features, with the anticyclonic BG extending down to ≈250 m and intensifying along the Chukchi shelf break (see the map of Mean Kinetic Energy (MKE) in Fig. S5). The cyclonic boundary current within the AW layer is found around the BG at 500 m with a returning branch of weaker intensity along the Canadian Archipelago (Fig. S5c). Upper outflows through the Canadian Archipelago are similar to observation-based derived circulation Fig. S5, see also. Climatologies of EKE computed relative to monthly means show large values along the shelf break and along topographic features such as the Northwind ridge, both at the surface (not shown) and within the pycnocline (Fig. a). In contrast, the deep basin is more quiescent, with EKE one to two orders of magnitude lower than on the shelves (Fig. a and b). The shelf-deep basin contrast in EKE magnitude is a typical feature of the mooring-based estimates . Yet, the intensity of EKE is about one order of magnitude smaller in our model than that derived from observations , as documented previously in . The MKE, which captures the location of the main currents, is of similar order of magnitude as in observations , with discrepancies being partly attributed to the difference in the exact location of the main currents between models and observations (Fig. S5). Finally, the vertical structure of the total kinetic energy is similar to that derived from the Beaufort Gyre Exploration Project Moorings compare Fig. S6 with for instance Fig. A1 from with sub-surface intensified structures between 30–200 m, and deeper (although weaker) structures between 400–2000 m, as evidenced in observations by .

We perform an offline detection and tracking of mesoscale eddies within the Canadian Basin over 1995–2020. The mesoscale eddies we capture span a broad range of mesoscale rotating features, from the evanescent vortices quickly dissipated by sea ice to the more persistent features that may eventually evolve into materially coherent vortices. The eddy population detected thus includes parts of the “turbulent soup” that is expected to develop at the surface in response to the atmospheric and ice forcings, and should be captured by the model. Though short-lived, these features, which are characteristic of the surface ocean, deserve an investigation as they allow one to examine the energy dissipation exerted by sea ice and contribute to the dynamical equilibrium of the basin. In the following, we focus on features with characteristic sizes from R0≈10 km to 2πR0≈60 km defining the mesoscale, e.g.. To identify eddies, we use the eddytools python package documented in . Eddies are detected using the Okubo–Weiss parameter OW;, which measures the relative importance of shear and strain to vorticity in the velocity field (Fig. a) : 1OW=∂xu-∂yv2+∂xv+∂yu2-∂xv-∂yu2 where u, v denote the velocities along the x and y directions of the grid, locally orthogonal. The resulting OW field (Fig. b) is compared to the local OW standard deviation (σOW(x,y)) averaged over the full time period (Fig. c). σOW is computed over a Lσ×Lσ box, Lσ being chosen small enough to capture the regional differences between e.g. the centre of the gyre and the boundary currents, but large enough so that σOW is not impacted by individual eddies. Eddies have to meet the following condition to be retained in the census: 2OW(x,y,t)<-ασOW(x,y), where α is a threshold value typically chosen between 0.2 and 0.5 . As we aim to detect any vortex-like feature that may develop in the Canadian Basin, including those that are not materially coherent, we choose an Eulerian over a Lagrangian approach for detection. The OW-method is based on velocities (u, v) and thus preferable over SSH-based methods for detection in sea ice-covered areas where SSH-based detections are known to miss objects that do not have a surface expression. Additionally, the OW-method has the advantage of being computationally efficient and thus seems well-suited for a detection run for several decades. A comparison of the OW-based detection with those from u, v-based and SSH-based was performed by see their Fig. S3. They show that the OW-based method detects higher numbers of eddies compared to the other methods, mostly due to its capability to detect weak eddies, i.e. eddies with small rotational velocities and SSH anomaly. This detection bias towards weak eddies is commented in Sect. .

The detection is implemented using daily averaged model output in the Canadian Basin and is run for each vertical level of the model independently above 1200 m (which represents a lower bound of the AW layer), totalling 49 levels. No 3D representation of eddies is attempted here, as connecting the results between the vertical layers is not trivial and would require a substantial development of the detection and tracking algorithm. A brief evaluation of the vertical structure of eddies is however proposed in Sect. . Note that because σOW is computed independently for each depth level, the minimum OW used to identify an eddy also varies with depth. In other words, at depths of intense mesoscale activity, the OW an individual eddy needs to overpass to be identified as such is higher. For each eddy, we estimate its radius with R=area/π even though the eddies might be elliptical in shape. We set the smallest eddies that the algorithm detects to occupy 20 grid points, which correspond to equivalent circular eddies with a minimum of 5 grid points across the diameter. A 5-grid point diameter circular structure corresponds to an eddy of about 7.5 to 10 km radius, depending locally on the grid size of the model, which is the lower bound of R0 over the deep part of the Canadian Basin see. Over the shelf, where R0 is smaller (≈2–5 km), we only detect the largest of the mesoscale rotating features. Statistics presented here include all detected features, but remain valid when filtering out eddies on the shelf, as the vast majority of eddies are detected over the continental slope and within the basins (not shown).

Sensitivity tests for α show that the vertical distribution of the metrics investigated (the mean eddy radius, duration, polarity rC/T i.e. the ratio of cyclones to total number of eddies, and a proxy for the vorticity |Ω|, see Sect. ), is robust to changes of α from 0.1 to 0.5. Yet, we note changes in the total number of detected features with slightly larger, weaker, and shorter eddies for smaller α (Fig. S7). For our analyses, we choose α=0.3 as an intermediate value. The box length Lσ over which to compute σOW needs to be tuned to the main spatial scales of dynamical regimes in the basin. In other words, Lσ should be small enough to capture the jet-like circulation along the Alaskan and Chukchi slopes that are about 200–300 km large, and large enough to allow statistically relevant values of σOW. We found that Lσ=200 km is a reasonable value to resolve the different dynamical regimes within the Canadian Basin. Similar to α, sensitivity tests indicate that changing Lσ within the range [50, 400] km does not modify the vertical distribution of the mean eddy radius, duration, vorticity, and polarity, although the total number of detected features varies (not shown). Overall, modifications of α and Lσ impact the precise definition of particular eddies, but not the statistical properties of the eddy field.

Eddies are tracked over consecutive days using three main criteria: (i) their speed of propagation, (ii) their polarity, and (iii) their radius R (Fig. c). For each eddy detected on day t, if an eddy with a similar radius (within [0.5R, 1.5R]) and the same polarity lies within a search radius Rs on day t+1, it is chosen as a continuation of the track. In case there is more than one eddy matching these criteria, the one with the centre located the closest to the original eddy's centre is chosen. Results do not appear sensitive to the choice of a search radius Rs∈[15,53] km, and we choose a search radius Rs=22 km corresponding to a propagation speed of 25 cm s⁻¹, which is approximately the speed of the fastest simulated current within the domain, located along the Chukchi slope in summer. Location of eddies is obtained from their centre of mass – thus, the distance travelled by a given eddy between two consecutive days is possibly smaller than the grid resolution if an eddy is very slow. The radius, location, and grid points occupied by each individual eddy are detected every day. We assume eddies to be born (generated) the first time they are detected and to die (dissipated) the last time they are detected. However, the algorithm may occasionally lose track of eddies, leading to a given eddy being counted as two successive eddies of similar properties. This interruption in the tracking generally occurs with weak features that are not well developed and thus not detected as eddies over consecutive days. We remove these eddies by filtering out any eddy that does not persist over at least two consecutive days. While this filtering does not fully overcome the interruption in the eddy detection – in particular if one eddy splits into two different ones, or equivalently if two eddies merge – it removes most of the issue and enables us to focus more on well-developed eddies. Still, the majority of eddies we detect are relatively weak and have a duration shorter than their turnaround time scale, defined as the time it takes for a water parcel to do a full revolution, τ=2π/|Ω| i.e. an approximation of the expression suggested bythat is τ=2π/ζrms, where ζrms is the root-mean-square vorticity. A discussion of the characteristics of the more vigorous and persistent eddies is proposed in Sect. .

For each detected and tracked eddy, we extract its intensity, which we define as the absolute vorticity amplitude of the eddy, i.e., the absolute value of the difference between the vorticity in the centre of the eddy and the average vorticity along the edge of the eddy (|Ω|). We equivalently report its relative intensity |Ω|/f(λ) where f is the Coriolis parameter computed as a function of latitude λ. By spatially averaging over the eddy area, we also extract its absolute salinity (S), conservative temperature (T), and the mean sea ice concentration A and thickness h above each eddy for each day.

Eddy properties are computed and tracked along the eddy pathway. Except where mentioned, properties are extracted at the eddy generation time. The properties of the eddy environment are defined by spatially averaging over a box that we take to be n=3 times larger than the eddy dimensions in x and y directions (thus not of identical size along both directions) and from which we remove the eddy area. We note ΔX=Xieddy-Xienv the anomaly of property X at the time of eddy generation i. If two eddies develop next to each other, they will become each other's environment as we do not use a 2D eddy mask. In the interior of the basin, the gradients of density that may generate eddies are generally small, and so is the density anomaly of each eddy. To increase the robustness of the quantification of these anomalies, we choose to report only on the strongest anomalies. We define such significant anomalies in the following way. We first define the deviation δX from the noise of the environment using the standard deviation of X across the environment (i.e., excluding the eddy area), σXenv: 3δX=ΔX++σXenvifΔX<0-σXenvifΔX>0. Then, the anomaly ΔX is said to be significant if it is of the same sign as δX, that is, if the anomaly is larger in absolute value than the standard deviation over the area.

Finally, the normalized amplitude of the seasonal cycle is defined for each property X as: 4SCX=Xmax-XminX‾, where the maximum, minimum, and mean are taken along the seasonal cycle.

Over 1995–2020, and on average along the vertical, we detect and track about 6000 eddies per year in the Canadian Basin. This large number contrasts with the very few vortices detected from in situ observations below the ice O(10) eddies per year,. However, it is closer to the numbers reported from satellite observations in the MIZ or the open ocean up to O(1000) eddies per year,. Most of the eddies detected in the model have a radius similar to the Rossby radius of deformation (R‾=12.1 km), are short lived with an average duration dt‾=10 d and do not travel far with an average distance travelled D‾=11.1 km (Fig. a–c). Of all eddies detected, 49 % are cyclones. Cyclones and anticyclones have a similar intensity (|Ω‾|=4.6×10-6 s⁻¹ corresponding to a relative intensity of |Ω|/f‾=0.03; Fig. b). The eddy intensity, lifetime, and distance travelled have a standard deviation of the same order of magnitude as the mean. In particular, the distribution of the distance travelled shows three peaks in the distribution (Fig. b): a first one corresponding to quasi-stationary eddies, and two secondary ones centred around 4 and 8 km. Only 15 % of the eddies show significant temperature and salinity anomalies with respect to their environment (Fig. e and f). The narrow and short tail of the statistical distribution of ΔS indicates that the overwhelming majority of detected eddies have properties close to the mean, while the wider distribution of ΔT indicates relatively large temperature anomalies for a significant portion of the eddy population (see box whiskers on Fig.  and the 10th and 90th percentiles in Table ). Interestingly, eddies with properties at the tail of the distributions do not represent a distinct population of eddies. For instance, larger eddies do not live systematically longer (see Fig. S8).

So far, we have presented the statistics of the eddy properties aggregated over the whole 1995–2020 period and over all depth levels above 1200 m, hence accounting for the same eddy several times if that eddy spans several depth levels. We observe some vertical coherency when looking at a few structures individually, in particular for structures spanning the pycnocline between 85–225 m, surface intensified eddies, or eddies spanning the whole water column below 200 m (Fig. ). This vertical structure is similar to the vertical structure obtained from observations or predicted from baroclinic instability estimate in the BG . This coherency reflects in the statistics of the characteristics of eddies with significant differences across depth (see the coloured plain lines and ◂, ▸ and + in Fig. ). Within the top 1200 m, the total number of eddies generated at each model depth level remains roughly constant between the surface and 85 m, and below 225 m, but increases by two-thirds between 85 and 225 m (Fig. a). We also note important transitions in the ratio of anticyclones versus cyclones, radius, and eddy durations around these depths, suggesting different mechanisms of formation and dissipation. On average across the basin and over the 26 years, these transition depths correspond to the depth at which the sPW (∼85 m), and the wPW (∼225 m, see Fig. g and h) are found, forming together the pycnocline layer.

Based on the evolution of the statistical properties with depth, together with the observation of the coherent structures with finite depth extent, we thus define three layers: the upper layer (0–85 m), the pycnocline layer (85–225 m), and the AW layer (225–1200 m). Next, we describe the eddy properties and discuss their formation and dissipation within each of these three layers. The results presented in the following are robust to the exact definition of the layer boundaries (±2 model depth levels).

Within the top 85 m, about 6000 eddies are detected every year. The properties of these eddies show a marked seasonal cycle (Fig. ), mainly following the seasonal cycle of the sea ice cover. From winter to summer, the number of eddies increases by a factor of 10 with a minimum in April, just before the onset of sea ice melting, and a maximum in September when sea ice is at its minimum (Fig. a). The Mixed Layer (ML) depth – computed as the depth at which the potential density has increased by 0.01 kg m⁻³ compared to the potential density at 1 m – decreases from 35 m in January–May to 3 m in July, and increases from August to December, when it recovers 30 m (see also the stratification on Fig. h). The change in stratification associated with the ML depth delimits different regimes of variations for the mean radius, polarity, and intensity. Within the ML, the averaged radius of eddies increases from 12 km in early summer to 14 km in fall, while below the ML, the averaged radius of eddies barely changes (increases from 12 to 12.5 km; Fig. b). In winter, a dominance of anticyclones is found at the surface (0–10 m) and of cyclones just below (10–40 m, i.e., to the base of the ML), while at the very surface, eddies are essentially anticyclonic year-long (Fig. d). Below the base of the ML, the proportion between cyclones and anticyclones remains more equally distributed throughout the year, with about 55 % anticyclones. Within the top 85 m, the most intense eddies are found during the stratifying and de-stratifying periods, corresponding to the onsets of sea ice growth (October and November) and to the melt season (May–July), respectively. In winter, intense eddies are also found at the base of the ML (Fig. c). Eddies persist longer in summer (7–8 d) than in winter when their lifetime is reduced by about half (Fig. e). Lifetimes likely influence the distance travelled, with eddies propagating over 15–16 km in summer and 7–8 km in winter on average (Fig. f). Lifetime for the vast majority of eddies (85 %) is significantly shorter than the theoretical mean turnaround time. Thus, part of the detected eddies are likely not fully developed and belong to the “turbulent soup” that is generated in response to the surface density gradients and gets quickly dissipated by sea ice in winter. These eddy lifetimes are similar to the characteristic times of spin-down through sea ice dissipation when sea ice is taken as the main drag e.g. ≲4 d,. We come back to this point in Sect. .

Eddy properties present large spatial variations in the surface layer of the Canadian Basin (Fig. ). In particular, there is a strong contrast between the continental slope and the deep basin, with up to 10 times larger density of eddy population over the slope (≈150 km wide, Fig. a). The greater generation of eddies over the slope peaks in October, when sea ice extent is close to its minimum and winds start to increase (see Fig. S9). While the production of eddies in the deep basin remains low on average, it becomes similar to the production over the slope when sea ice concentration drops below ≈80 %, that is, between July and November, depending on the latitude (see Fig. S9a). Simultaneously, eddy lifetime increases from evanescent (dt=1–3 d) below the pack in winter to about 15 d in summer on average (Fig. S9b). Thus, over the domain, eddy lifetime is mainly enhanced where the ice concentration is lower than 15 % (not shown).

Over the slope, eddies have, on average, a positive temperature anomaly (≈0.3 °C) with the exception of some anomalously cold eddies forming over the Chukchi shelf. Where the mean September sea ice concentration is higher than 15 %, eddies do not have a temperature anomaly, aligning with 's detection of a majority of vortices with no significant temperature anomaly. We also find a contrast in radius between the eastern and western sides of the gyre, especially off-shore the Chukchi shelf break and above Northwind ridge, where eddies are found to be about 60 % larger (Fig. b). These eddies are carried within the intense anticyclonic BG circulation (Fig. S5) and therefore travel up to 40 km throughout their lifetime, a distance much larger than the averaged distance travelled of 8 km at the centre of the basin.

Anticyclones are predominant over the centre of the gyre, while over the slope, a greater proportion of cyclones are found (Fig. d). Vorticity anomalies within the BG do not indicate any preference for the generation of anticyclones (not shown). One hypothesis that explains this cyclone/anticyclone asymmetry, which has been formulated for the Mediterranean Sea and more generally in other contexts of turbulent flows, is that anticyclones are more persistent than cyclones that tend to split into smaller objects, leading to anticyclones being more systematically identified in eddy censuses . further suggests the role of sea ice in preferentially dissipating small cyclones. Whether this applies to the BG is worth future investigation. We speculate that such a filtering mechanism might mainly apply in the centre of the gyre, where mean currents are negligible, and the turbulent field can freely develop, while strong mean currents that generate and dissipate eddies are likely to be the dominant factor in determining eddy polarity near the gyre's edges.

Up to 300 km off the shelf, high density in the eddy population is accompanied by intensities in the eddy field up to one order of magnitude higher than in the deep basin (Fig. c). This is visible all along the shelf break of the domain, from the Chukchi shelf break to the Canadian Arctic Archipelago shelf break. The most intense eddies are found at the mouth of the McKenzie River and at Pt. Barrow (not shown), being respectively fresher and saltier than their environment (Fig. e and f). Along the Alaskan and Chukchi slopes, on both sides of Pt. Barrow, eddies with positive salinity anomalies are detected in the inner part of the slope, while eddies with negative anomalies are detected in the outer part. This pattern illustrates the penetration of the Pacific Waters from Pt. Barrow along the baroclinically unstable Alaskan coastal and Chukchi Slope currents , and supports observations of the penetration of eddies associated with a salty anomaly into the BG at Pt. Barrow in the submesoscale range. Additionally, eddies associated with a fresh anomaly found along the outer part of the slope confirm the role of fresh water input from McKenzie River in generating instabilities that develop into eddies propagating downstream along the anticyclonic circulation .

Over the 26 years of simulation, there are about 30 % more eddies detected in the pycnocline layer (∼85–225 m) than in the upper layer (9000 on average in the pycnocline layer vs. 6000 eddies on average in the upper layer per year; Fig. a). Eddies detected within the pycnocline layer are evenly distributed between cyclones and anticyclones (Fig. d) and are smaller and weaker than in the upper layer on average (mean radius decreases from 12.4 to 11.6 km and intensity decreases from 7.4×10-6 s⁻¹ in the upper layer to 3.3×10-6 s⁻¹, Fig. b and c). Although weaker, eddies in the pycnocline layer last about 6 d longer than in the upper layer, likely due to the absence of ice or air drag to dissipate eddies through friction (Fig. e). Despite this increased longevity, the mean distance travelled by eddies in the pycnocline layer only increases by about 1.1 km compared to the upper layer (Fig. f), presumably because of the weaker background flow advecting these eddies. Therefore, of the eddies generated over the slope, only the strongest and longest-lived eddies may be able to travel far enough to reach the gyre and could thus participate in the transport of heat, salt, and nutrients from the continental shelf to the deep basins (see Sect. ).

In the pycnocline layer, eddy characteristics show a weaker seasonality compared to the upper layer. Quantitatively, the normalized amplitude of the seasonal cycle in eddy number diminishes with depth, from SCN=2.3 at 30 m to SCN=0.5 at 150 m (see Sect.  for a definition of SCX). The other properties also show a decreased normalized amplitude of their seasonal cycle compared to the upper layer, by ≈60 % for the radius, 50 % for the intensity and distance travelled, and by 40 % for the lifetime (refer to Table  for detailed modifications of the seasonal cycles). This damped seasonality is expected as the pycnocline shields eddies from dissipation by sea ice.

For most eddy characteristics (radius, intensity, duration, distance, and polarity), the spatial distribution within the pycnocline layer is generally similar to that of the upper layer (compare Fig.  with Fig. ) but persists throughout the year due to the absence of seasonal variability. The similarity in spatial distribution with the upper layer is expected as the anticyclonic circulation that dominates most of the region investigated extends down to the pycnocline associated with the wPW see Fig. S5, see also. However, the spatial distribution of the density in eddy population is notably different between the pycnocline and the upper layers along the southern edge of the BG (Fig. a). There, a strong reduction in the density of the eddy population is found compared to the shelf and deep basin (Fig. a). This reduced density compared to the upper layer occurs despite the eddies being relatively intense, long-lived, and travelling relatively far along the anticyclonic flow (Fig. c, e and f). We suggest that the inner part of this local reduction in eddy generation is linked to a stabilizing effect of the continental slope. The growth of instabilities is known to be hampered over regions where the ratio of the continental slope to the isopycnal slopes is greater than 1 , as is the case for the slope of the BG in the model not shown, see. However, this reduction occurs up to 250 km away from the shelfbreak. There, we observe diminished background potential vorticity gradients not shown, see Fig. 9 from associated with diminished baroclinic instabilities, which offer an alternative explanation for the extended area with muted eddy activity.

Below the pycnocline layer and down to 1200 m, within the AW layer, the total number of eddies over the 26 years of simulation decreases by 37 % compared to the pycnocline layer (from on average 9000 per year in the pycnocline layer to 5500 eddies per year in the AW layer; Fig. a). This decrease is reduced to 20 % if one compares eddy density, as the area where eddies developed is reduced due to the bathymetry. Because the layer is located below the pycnocline, the seasonal variability in eddy properties is almost completely shut down (Fig. ). In that layer, we find eddies similar in radius (≈11.8 km, Fig. b) but weaker in intensity than in the pycnocline layer (1.5×10-6 s⁻¹ compared to 3.3×10-6 s⁻¹, Fig. c). The distance travelled by eddies decreases from 12.2 to 7 km (Fig. f), and the polarity remains equally shared between cyclones and anticyclones (Fig. d). We note that the averaged lifetime of eddies is longer than in the pycnocline layer (14 d compared to 11 d) but remains small despite the few processes that could dissipate eddies at this depth. This relatively short lifetime may point to the fact that most of the eddies detected in this layer do not fully develop (if we consider their turnaround time). Of all the eddies detected in that layer, only 6 % persist longer than their turnaround time. These long-lasting eddies may live up to 150 d (99th percentile), surpassing all the maximum durations detected in the other layers and matching estimates of weeks to years made from observations although these estimates are largely uncertain as ITPs and moorings only capture a portion of the eddy trajectory. We refer the reader to Sect.  for a discussion on long-lasting eddies.

Within the AW layer, eddy properties show different patterns compared to the layers above. Significant differences are expected given that the mean circulation of that layer departs strongly from that above (Fig. S5). In particular, eddies are predominantly generated over the continental slope along the path of the AW cyclonic boundary current (Fig. a). A smaller density of eddies is generated in the rest of the domain, with some hot spots of high density in eddy population located close to the northern boundary of our domain. The latter corresponds to short-lived eddies, and we discuss more extensively the “turbulent soup” formed by short-lived eddies in Sect. . Along the shelfbreak and boundary currents, the eddy intensity is larger by up to one order of magnitude compared to the rest of the domain, where eddies are notably weaker (Fig. c). No clear spatial pattern in polarity arises at the scale of the basin, except along the shelf breaks of the Chukchi Sea and Canadian Archipelago, where anticyclones tend to dominate in the inshore part of the current while cyclones dominate in the offshore part of the current (Fig. d). Off the western flank of Northwind ridge are found the largest (up to 20 km), farthest-reaching (up to 40 km) and longest-lived (up to 60 d) eddies (Fig. b, e and f). The EKE is one order of magnitude larger in that area than within the deep basin and displays hotspots in the form of large structures detaching from the cyclonic boundary current that hugs Northwind ridge (not shown). In this region, there are large uncertainties in the literature on the exact path of the AW, as AW are thought to intermittently detach from the slope-intensified cyclonic boundary current, or alternatively flow along double boundary currents . We suggest here some instabilities in the cyclonic boundary currents associated with the generation of large eddies.