The data used in this study were collected during the MOSAiC expedition . In particular, we use data from the Distributed Network (DN) , which was installed around the Central Observatory (CO), where the Polarstern was anchored to the ice. We restrict our analysis to the winter period from 19 October 2019 to 15 March 2020 (for daily drift locations see Fig. a). The DN consisted of several autonomous ice-tethered systems designed to collect Arctic Ocean properties at different temporal and spatial scales. The DN includes both fixed-depth time series data and vertical profiles. Although fixed-depth sensors provide high temporal resolution data (on the order of minutes) and capture eddy signatures , they are unsuitable for detailed characterisation of individual eddies. This is because fixed-depth data do not capture the full vertical structure of IHEs, which require vertical profiling to resolve their thermohaline and velocity structure. Therefore, this study uses exclusively vertical profile data to analyse the structure and dynamics of wintertime IHEs. We focus on the three instrument deployment locations, termed L-sites, which were positioned at a distance of approximately 12–24 km around the CO (Fig. b). At the L-sites, Ice-Tethered Profilers (ITPs) provided Conductivity, Temperature and Depth (CTD) measurements, and Autonomous Ocean Flux Buoys (AOFBs) equipped with Acoustic Doppler Current Profilers (ADCP) measured horizontal velocity. We also used CTD measurements from the surface to the ocean floor and velocity data from the Polarstern shipboard ADCP. All datasets were used in their publicly released, quality-controlled form from the official MOSAiC data products, including the ITP , the Polarstern sADCP , the AOFB , and the Polarstern CTD . No additional corrections, averaging, or interpolation were applied. Only the velocity data were smoothed with a half-day low-pass filter to reduce high-frequency noise while retaining eddy-scale variability. This was done to remove the short-period noise while preserving temporal variability at time scales expected for (sub)mesoscale eddies, which typically last several days. The MLD was defined as the first depth at which the Brunt–Väisälä frequency anomaly, ΔN, between successive measurements exceeded 3×10-4 s-2. This threshold was selected following a visual inspection of all available profiles, as it reliably captured the transition from the mixed layer to the onset of the halocline during the MOSAiC drift. All thermodynamic variables, including density, N, and derived quantities, were computed using TEOS-10 through the GSW Python toolbox .

The three L-sites were instrumented with ITPs and AOFB-mounted ADCPs, each operating with different profiling intervals and vertical sampling ranges (Table ). Because all platforms drifted with the sea ice, the horizontal spacing between consecutive profiles depended on both drift speed and profiling interval, ranging approximately from 1 to 10 km, with the L3 ITP providing the smallest spacing and L2 the largest. The L1 and L2 sensors remained operational throughout the winter and drifted toward the Fram Strait, whereas the L3 sensors ceased operation on 31 January 2020 following an ice-ridging event. The L2 ADCP did not return usable data due to early technical failure, and when L2 and the Central Observatory were aligned, velocity measurements were supplemented using the Polarstern shipboard ADCP.

In this study, the DN moved with the sea ice at a mean drift speed of 0.11 ms-1, while the underlying ocean current below the mixed layer had an average speed of 0.02 ms-1 (Fig. c). Because ice drift is an order of magnitude faster than ocean currents, the ice-tethered platforms move quickly relative to the ocean features beneath them. The DN geometry also remained stable during the drift (Fig. b), with inter-platform distances changing only slightly and always exceeding the expected diameter of Arctic intrahalocline eddies. This large difference in speeds and the absence of significant deformation or rotation of the array justifies the quasi-synoptic assumption, which means that measurements from the ice-advected platforms can be considered as near-instantaneous snapshots (“frozen fields”) of the slower-evolving ocean eddies . Furthermore, the analysis was restricted to DN trajectories that followed approximately linear paths during each eddy encounter, thereby minimising potential geometric biases arising from the relative motion between the platform and a propagating eddy. This interpretation agrees with observations that Arctic eddies propagate at speeds roughly an order of magnitude slower than the sea ice drift .

To identify eddies, we follow the methodology suggested by and . Eddies were first recognised in the ITP profiles by visually detecting coherent vertical displacements of isopycnals across several consecutive casts, reflecting the eddy's spatial structure as the drifting platform crosses it. In the case of the anticyclonic IHEs, the only type of eddies detected in our dataset, the upper part of the eddy shows a convex upward doming of isopycnals, whereas the lower part exhibits a concave downward displacement, producing opposite slopes above and below the core. In a second step, we analyse the velocity profiles measured by the ADCPs and look for the characteristic eddy velocity anomaly, with two local maxima in horizontal speed, one on each side of the isopycnal displacement centre. Figure  illustrates this two-step identification using the hydrographic displacement and the associated speed anomalies. E8 (Fig. a) at L3 was sampled with the smallest horizontal spacing between profiles, whereas E9 (Fig. b) at L2 had the largest spacing. These examples show how eddy-like structures are recognised from the combined ITP–ADCP signal before applying dynamical consistency tests.

The last step in confirming that an eddy-like structure is in fact a rotating eddy is to verify that the velocity field is dynamically consistent with coherent rotation rather than with other features such as meanders or frontal intrusions, which have been documented in the central and marginal Arctic Ocean e.g.. As a first diagnostic, we require that the azimuthal (cross-track) velocity vθ exhibit a reversal in sign across the centre of the isopycnal displacement, indicating opposite flow directions on the two flanks of the feature and ensuring that the profiler crossed through or very near the eddy core. As a second diagnostic, we test whether vθ increases approximately linearly with radius within the core, consistent with the solid-body rotation expected in mesoscale eddies ; the radius of maximum velocity then marks the edge of the core . Because these dynamical diagnostics confirm coherent rotation, we also accept eddies sampled with fewer than the 4 ITP profiles required by . In our dataset, the combined ITP–ADCP observations can confirm the presence of an eddy even when only two consecutive profiles show the isopycnal displacement.

To compute the cross-track component vθ and its along-track counterpart vr, the velocity profiles were rotated by the angle θ between consecutive drift-track segments: 1vr=ucos⁡(θ)+vsin⁡(θ)

2vθ=-usin⁡(θ)+vcos⁡(θ)3θ=arctan⁡y2-y1x2-x1 where (x1,y1) is the position of the first measurement and (x2,y2) of the second measurement. vθ, also termed the swirl velocity, provides a sense of the rotation of the fluid (i.e., the tangential velocity component within a swirling flow).

In the theoretical Rankine vortex , the azimuthal velocity increases linearly with the distance to the centre, having the maximum value of the velocity at Rmax⁡; the distance between the location of the absolute smallest azimuthal velocity (centre of the eddy) and the maximum azimuthal velocity (Vmax⁡) (Fig. , red line). A method for computing the azimuthal velocity vθ of a theoretical eddy as a function of radius (Eq. ) follows the analytical velocity model introduced initially by , and later applied explicitly to eddy-velocity fitting by :

4vθ(r)=rVmax⁡Rmax⁡,for r≤Rmax⁡Vmax⁡exp⁡-r-Rmax⁡λ,for r>Rmax⁡ where λ is a damping coefficient that indicates decay. Equation () assumes an inner part of the eddy that rotates like a solid body (r≤Rmax⁡) and an outer part (r>Rmax⁡) where the velocity decays rapidly at the e folding scale λ, which is typically about 13Rmax⁡. Here, we are focusing on the inner part (i.e., the core of the eddy). Furthermore, the dynamics outside the limit of Rmax⁡ are out of the scope of this study.

The azimuthal velocity (vθ) calculated using Eq. () was compared to the theoretical Rankine vortex (Eq. ) to assess whether the eddy cores exhibit solid-body rotation. In both E8 (15 January) and E9 (12 February) (Fig. ), the observed azimuthal velocity profiles measured by the ADCP closely follow the theoretical shape. The inner region displays solid-body rotation, while the outer region shows a rapid velocity decay, consistent with the Rankine vortex structure. For this comparison, the Rankine model is scaled using the maximum azimuthal velocity directly measured from the ADCP data along the drift trajectory, ensuring that the comparison reflects only the observations and the analytical model introduced in this subsection.

Several methods have been used to determine the centre of eddies in the open ocean. However, most need a surface expression of the eddy to obtain the horizontal velocity field . In our study region, the presence of thick sea ice prevents the use of satellite-derived velocity fields, so the eddy centre must be inferred directly from the in situ ADCP observations. To estimate the centre of the eddy and obtain an accurate approximation of its radius, we applied the Maximum Swirl Velocity (MSV) method as described by . This method assumes that an eddy is axisymmetric, with all the momentum associated with the azimuthal component of the velocity. Hence, its centre is defined as the reference point in a cylindrical coordinate system that maximises the measured azimuthal velocity vθ among the available data points, while the radial component of the velocity vr is vanishing V=vθ2+vr2=vθ.

To find the centre, we followed the approximation of , who proposed testing the MSV method over a gridded search area, as follows. First, we defined an area of 2Rmax⁡×2Rmax⁡ around the location where the minimum velocity inside the eddy was measured, dividing it into a 100 m resolution grid. We then used every point of the grid as a theoretical centre, decomposing all of the ADCP velocities, as in Eqs. ()–(), into tangential and radial components relative to each candidate centre. As vθ has opposite signs for cyclones and anticyclones, it is easier to determine the centre of the eddy by finding the location where vr is minimal, thereby minimising the cost function J : 5J=12N∑n=1NvrnVn2 where N is the number of ADCP measurements used and V is the speed. To perform this minimisation, we computed the value of J at every grid point and selected as the eddy centre the point where J reached its minimum. Once the centre is detected, we recalculated the radius Rm as the distance between the theoretical centre and Vmax⁡ (as shown in Fig. ).

We apply this eddy detection method to eddies E8 and E9 (Fig. ). The centre of the eddy E8, calculated by minimising the cost function J, lies approximately 5 km from the point of minimum velocity in the transect, and its radius is estimated at 8.4 km. In contrast, the radius obtained by measuring the distance between the locations of minimum and maximum velocity along the same transect is considerably smaller, about 4.6 km. For eddy E9, the L2 transect crossed nearly through its centre, and the difference between the two radius estimates is minimal (5.9–6.1 km).

The Rossby radius of deformation is a fundamental scale in geophysical fluid dynamics that characterises the horizontal extent over which baroclinic processes, such as eddies, are influenced by the Earth's rotation. It represents the length scale at which the restoring force due to stratification (buoyancy) is balanced by the Coriolis force (rotation), and is thus a critical parameter in controlling the dynamics of mesoscale structures . To constrain the local scale of these mesoscale processes, we calculated the first mode of the baroclinic Rossby radius of deformation using the approximation of (Eq. ): 6∂∂zf2N2∂Fm∂z=-1LR,m2Fm, where N, the Brunt–Väisälä frequency, is derived from the CTD vertical cast. f is the Coriolis parameter (f=1.45×10-4 s-1), Fm is the dynamic mode eigenfunction and LR,m is the Rossby radius of the mth baroclinic mode. To solve this equation, we apply a flat-bottom boundary condition (dFmdz=0 at z=0,-H), appropriate for our study region within the deep Amundsen Basin. Although the Amundsen Basin is limited by the Lomonosov and Gakkel Ridges, our study area lies in the central interior of the basin, where the upper-ocean stratification and water-mass structure are known to be horizontally uniform and largely independent of ridge-controlled dynamics

In Fig. , we show the spatial distribution of LR,1 in our study area. Stations located close to each other sometimes show different LR,1 values, likely due to local variations in the vertical stratification of the water column. These differences reflect the sensitivity of the method to small-scale changes in water column stability, which are captured in the CTD profiles. The mean value of LR,1 in the study area is 6.93 km, which justifies our use of the term (sub)mesoscale throughout the paper, as several eddy structures observed fall near or below this threshold.

We observed nine eddies, which, for the purpose of explanation, we label E1 to E9. We start this section by analysing in detail two representative anticyclonic eddies: E8, observed on 15 January, and E9, observed on 12 February (see Tables  and . These examples are used throughout the Methods section to illustrate our detection and characterisation approach. The E8 eddy (Figs.  and , upper panels) was captured by the L3 buoys with 12 ITP profiles and 10 ADCP profiles within the solid-body rotation region (core). The MSV method (see Sect. ) revealed a radius of 8.4 km. The E9 eddy (Figs.  and , lower panels), captured by the L2 buoy with only 2 ITP profiles and 9 ADCP profiles, had a radius of 6.14 km. We selected these two eddies because they were the largest and most energetic detected along the drift track, with strong azimuthal velocities and clear hydrographic signatures representative of wintertime IHEs. Although the CTD profiles only coarsely resolved E9, the ADCP data did resolve it well, and it suggests the eddy was crossed almost through its centre (Fig. ).

To characterise the eddies, we used profiles of conservative temperature (Θ), absolute salinity (SAC), density (σθ), buoyancy frequency (N), and azimuthal velocity (vθ) (Fig. ). We selected the central profile, where the isopycnal displacement was greatest, based on ITP data (Fig. a and b, red dashed line). The N profiles were used to identify the isopycnal surfaces that bound each eddy vertically. We define the upper limit of the eddy (Figs.  and , red line) as the isopycnal coincident with the depth of the first peak in N, and the lower limit (Figs.  and , cyan line) as the isopycnal coincident with the depth of the second peak in N, and the core-centre is defined as the depth where N reaches a minimum between these two limits. Our definition of the core-centre depth differs from that used by , who used the level of minimum Θ. That criterion was not applicable to the eddies observed in this study, as no clear Θ minimum was present. The eddy thickness is thus given by the depth difference between its upper and lower boundaries. The eddies observed are IHEs, located near the base of the mixed layer and interacting with the upper halocline. As they translate, they uplift the mixed layer, making it thinner. The eddy E8 yielded a decrease in the MLD from 41 m to 24 m from the eddy edge to the eddy centre, which is similar to that resulting from the eddy E9, with a decrease of 20 m depth. At the upper boundary, the isopycnal was displaced upwards 14 m in both eddies; at the lower boundary, it was displaced downwards by 23 m in E8 and 19 m in E9. In E8, the eddy's upper boundary is located at depth with vθ≈0 (Fig. d), indicating good agreement between the CTD and ADCP data. In E9, the eddy's upper boundary, as determined using the N profiles, does not align with the depth with vθ≈0, likely due to the coarser temporal resolution of the CTD data. The maximum azimuthal velocity of the eddy was 0.25 ms-1 and 0.28 ms-1 in E8 and E9, respectively.

In the period from 19 October 2019 to 15 March 2020, we detected nine well-developed anticyclonic eddies in the central part of the Amundsen Basin (Fig. a, Tables  and ). These eddies are consistent with our criteria, showing evident isopycnal displacement and solid-body rotation. At the L3 site, we sampled five eddies at high horizontal resolution, enabled by the high-frequency sampling of both ITP and ADCP instruments at this location (Fig. , diamond markers), and their presence is evident in both the isopycnal displacement and the large subsurface azimuthal velocity. The ITP at the L2 site had a more complex profiling schedule (Table ), with a greater horizontal distance between consecutive profiles, making the identification of eddies by isopycnal displacement alone more challenging. However, when analysing the ADCP data, we detected one eddy in December and one in February (Fig. , square markers). We do not have hydrographic data for the inner part of the eddy in December (E7), and only two profiles are available for the eddy in February (E9). At the L1 site, we detected two eddies, one in October (E3) and one in November (E4) (Fig. , inverted triangle markers). There is a one-month gap in eddy detections in the DN (17 December 2019–15 January 2020). However, there were no obvious changes in the drift speed of the DN during that period (Fig. c, mean speed of ∼ 0.1 ms-1). Hence, the absence of eddies is likely unrelated to the temporal and spatial resolution of the measurements.

We now describe the mean properties of the detected eddies (Fig.  and Tables  and ). The dynamical nature of the eddies can be characterised through the interplay between four key parameters: the eddy radii estimated from the MSV (Rm = 6.09 ± 1.4 km), the first mode of the baroclinic Rossby radius of deformation (LR,1 = 6.9 ± 1.3 km), the maximum azimuthal velocity (Vmax⁡ = 0.14 ± 0.07 ms-1), and the Rossby number ( Ro = 0.32 ± 0.14). The condition Rm<LR,1 places these eddies in the (sub)mesoscale regime, indicating a transitional dynamical scale. Consistent with this interpretation, the Burger number (Bu) computed for the nine eddies (0.4–2.9) further show that most lie in a submesoscale-to-mesoscale transitional regime (Bu≈1), while a minority (Bu>1) exhibit more compact structures where cyclogeostrophic effects may become relevant. We computed Ro using the cylindrical approximation Ro=2UfR , where U is Vmax⁡, f is the Coriolis parameter and R is the radius. Similarly, the Burger number was computed as Bu=(RLR,1)2, where LR,1 is the first baroclinic Rossby deformation radius. This yields Ro (0.16<Ro<0.62) that is consistent with the quasi-geostrophic balance, although the upper range allows for curvature effects e.g.,. It is interesting that the radii would be underestimated by ≈ 1.7 km if we did not perform the MSV correction (Fig. b), which would indicate that the eddies appear closer to the submesoscale regime than they actually are. In the centre of the eddies, the MLD becomes shallower on average by 11.5 ± 5.73 m, ranging from 33.6 to 22.12 m depth. The eddy thickness differs by 23 to 80 m, having an average thickness of 59.75±18.1 m. The depth of the eddy centre (Dc) were found at 59.3±13.6 m depth with an average temperature of Θc = -1.773±0.04 °C, salinity of SAc = 33.954±0.26 and a potential density of σθc = 27.49±0.24 kgm-3. These values correspond to the range of surface waters in the Amundsen Basin, but then, if we look at the anomalies against the mean values of the profiles at ± 30 km around the eddy, we find small yet significant anomalies (ΔΘc = -0.082°C, ΔSAc = -0.189 gkg-1, Δσθc = -0.15 kgm-3), that we will use later to discuss the possible origin of the eddies.

In the Arctic Ocean, density is primarily driven by salinity changes (Figs. b and b) rather than temperature due to the well-developed halocline, as cold waters remain close to the freezing point, minimising thermal effects . This dependency is evident in the core-centre properties, with density anomalies linearly associated with salinity anomalies (Fig. a). However, the relationship between Vmax⁡ and the eddy thickness does not follow a single linear trend (Fig. c), where we can distinguish two different groups. The first one with lower Vmax⁡ and Rossby numbers Ro≈0.2, where the relative vorticity term (2Vθ/r) represents only about 20 % of the planetary vorticity (f), and the second group with lower Ro (≈ 0.3–0.6), where relative vorticity accounts for 30 %–60 % of f. This separation in Ro is consistent with the natural segmentation apparent in the Ro distribution itself (Fig. b), which qualitatively corresponds to what a non-parametric clustering approach would identify. Higher Ro eddies require stronger velocity adjustments to balance changes in thickness, and for these cases, the Coriolis force alone becomes insufficient to balance the radial pressure gradient. The centrifugal contribution therefore becomes relevant, indicating that those eddies are closer to the cyclogeostrophic regime in which both Coriolis and centrifugal forces balance the pressure gradient, consistent with the findings of . In contrast, low Ro eddies can accommodate thickness variations with comparatively small changes in Vmax⁡, reflecting different modes of potential vorticity adjustment . This behaviour is also consistent with their thermohaline structure: the more energetic eddies not only tend to be thicker, but also exhibit larger core-centre anomalies. We find that most of the eddies with larger thickness are located deeper in the water column (Fig. d).

No eddies were detected after 12 February, the date of the last confirmed detection. Although the DN continued drifting, the velocity measurements became progressively limited due to the sequential failure of the AOFBs, first at L1 (27 February) and later at the Central Observatory (19 March), leaving the sADCP as the only velocity source, which did not reveal any coherent eddy signatures. Meanwhile, the ITPs at L1 and L2 remained operational but did not register additional eddies. By mid-March, after the drift crossed the Gakkel Ridge and transitioned to the Nansen Basin, the mixed layer had deepened markedly (exceeding 150 m), which likely inhibited the detection of IHEs within the 200 m vertical range of the remaining ITPs.

The study of eddies under sea ice prompts the question of whether the same eddy has been sampled several times. The answer to this question is not trivial; all the different measurements need to be assessed to constrain it. First, the assumption of quasi-synoptic conditions adopted in Sect.  implies that the eddies cannot move fast enough to pass several sites within the time frame of a single observation period. Second, the eddy size provides an additional constraint: most IHEs observed by in the Amundsen Basin had radii of approximately 5 km, consistent with our estimates. Given the spatial separation among the DN sites, L1–L3 ≈ 32–35 km, L1–CO ≈ 10–17 km, and L3–CO ≈ 22–24 km, it is therefore unlikely that the same eddy would be sampled at more than one location, except in the few cases where the drift geometry brought two platforms over the same region within a short time interval. Third, the mean background flow of the Transpolar Drift, about 0.02 ms-1, advects the IHEs at approximately the same speed , further limiting the distance an eddy can travel between consecutive profiles. Although the DN rotated during the drift, its overall configuration and relative distances remained effectively constant throughout the study period (Fig. b). Therefore, rotation does not alter the spatial separation between sites nor create conditions under which a single eddy could be sampled simultaneously at different locations. In the specific cases of E7 and E9, the CO ADCP detected an eddy shortly before the L2 ITP sampled a similar signal. Because the DN drifted northeastward–southeastward, both platforms consecutively passed over the same region, separated by approximately 9–14 km, allowing us to conclude that L2 and CO sampled the same eddy. These are the only two eddies detected at L2, and since L2 does not provide velocity measurements, such events represent the only situations in which L2 can be meaningfully compared with CO. Moreover, during the periods when E7 and E9 were observed, no eddy signatures were detected at L1 or L3, further limiting the usefulness of L2 for the multi-site duplicate-detection analysis presented below. We therefore examine possible duplicate detections only for the configurations in which velocity and hydrographic data allow meaningful cross-comparison.

Between 29 October and 29 November 2019, five eddies were observed at different L sites within a relatively short time window, raising the possibility that some of these detections correspond to the same eddy sampled at various stages of its path. Although the distances between sites such as L1 and L3 exceed 30 km – well beyond the radius of the eddies in the area of ∼ 6 km – , the temporal coincidence warrants a closer investigation into whether some of these eddies could have drifted between nearby sites, particularly those closer together such as L2 and CO. For instance, L1 and L3 detected an eddy within two days (E3 on October 31 and E2 on 29 October; Fig. c and a, green and lime circles). Both events fall within the smaller-radius range but exhibit different core characteristics. The sADCP from Polarstern (Fig. b, right panel) shows no velocity signature associated with either eddy during the period marked by the green dashed line. Therefore, the two eddies appear dynamically independent, and the CO site did not intersect the azimuthal circulation of either feature during its drift. From 15 to 19 November, a storm affected the ice drift, increasing the speed up to 0.4 ms-1 (Fig. c) and changing the drifting direction several times (Fig. , left panels). As a result, the DN platform sampled some sites more than once. Of the three eddies encountered on 4 November (E4), 17 November (E5), and 24 November (E6) (Fig. , pink, lilac and purple circles), E4 stands out as a well-formed eddy with a strong azimuthal velocity of Vmax⁡ = 0.20 ms-1 (Fig. c, right panel). The periphery of this eddy was also observed at the CO site on November 7 (Fig. b, right panel, dashed pink line), as confirmed by the drift trajectory of CO passing near the core's edge. Site L3 recorded two eddies within a week, with centres separated by 6 km. This suggests that both detections correspond to the same eddy, which would have a translation speed of approximately 0.01 ms-1 during that period. The thermohaline and kinematic properties support this interpretation (Tables  and ): the differences between E5 and E6 are minimal (order 0.01 in Θ and SA) and are consistent with the instruments not sampling the same cross-section of the eddy, which also explains the moderate difference in estimated radius (7.3–5.2 km). The Θ-SA structure (Fig. c) confirms that both features share nearly identical core water masses, indicating no appreciable modification of the eddy over the one-week interval. This is expected, as intrahalocline eddies can persist for extended periods. report a lifespan of at least 21 months for Eurasian Basin IHEs, and therefore their thermohaline structure is not expected to change substantially on weekly timescales. Although the CO site passed near the region where the eddy was located, no clear azimuthal-velocity anomaly was detected (Fig. b, purple line), likely due to an earlier partial crossing or an incomplete intersection with the eddy core.

Unlike temperate seas, the generation and trajectory of eddies cannot be remotely observed beneath Arctic sea ice. Although the western Nansen Basin shows stronger eddy kinetic energy than the interior Eurasian Basin, weaker and less frequent eddies have also been observed in the central Arctic . Literature shows that most of the efforts to categorise Arctic eddies have focused on the differences in the thermohaline properties of their cores. Based on this, eddies have been classified into Canadian water and Eurasian water eddies. In turn, this classification is divided into shallow (< 80 m) and mid-depth (> 80 m) core-centre depth, respectively e.g.,. The eight eddies found in this study are shallow Eurasian water eddies, containing saltier waters that are less close to the freezing point than those studied by and therefore depart from the temperature–salinity relationship reported in that study (Fig. b).

The Θ-SA diagram in Fig.  shows three different characteristic shapes: (i) fluctuant temperature with a smooth “wedge” shape in E8 and E9 (Fig. a) located approximately within the range of -1.74±0.03 and 33.62±0.1 gkg-1, (ii) a smoother curve in October (Fig. b) and (iii) a prominent “wedge” shape in November (Fig. c) around -1.81±2 °C and 34±0.3 gkg-1. By “wedge shape,” we refer to a T–S structure in which temperature decreases toward a local minimum at the eddy core-centre, and then increases again as salinity continues to rise, forming a characteristic concave shape in the diagram. The smooth curve in Fig. b is the typical Θ-SA diagram observed in the surface Amundsen Basin water, with the temperature minimum just above the thermocline . It results from advective-convective processes . Following the formation of the winter mixed layer, fresher water originating from the Russian shelves and transported via the Transpolar Drift reaches the freezing point and becomes dense enough to convectively mix with the existing mixed layer . This process generates the cold halocline layer, a key feature of the Nansen Basin surface structure. In summer, meltwater from sea ice accumulates at the surface, stratifying above the cold halocline layer. As freezing resumes in early winter, this freshwater cools to the freezing point and begins to convect into the halocline, forming the distinctive Θ–SA “wedge” shape.

Figure c shows the Θ-SA diagram arising from a convective cold halocline, resulting from the stratification of summer sea ice meltwater and the remnants of a winter mixed layer . It has the particularity of a prominent wedge, typical of surface conditions in the Nansen Basin during late autumn, when the water column is actively adjusting to the changing surface freshwater input and atmospheric cooling. A similar process occurred inside the eddy (Fig. a), but the refreezing and convection during winter altered the upper part, making the wedge smoother than in the early winter season . Comparing the Θ-SA diagram of the eddies with the surrounding water, we find that E1 and E2 have trapped similar water masses, suggesting these eddies were likely formed in the same region. The Θ-SA diagrams of the other eddies have a wedge shape consistent with the typical processes occurring at the surface of the Nansen Basin, which is not seen in the surrounding waters at the time of the observations.

Water mass analyses provide valuable insight into the stratification and convective processes that likely precondition the upper ocean before eddy formation and may help infer their possible region of origin. In particular, the presence of cold, fresh anomalies and a sharpened halocline in the eddy cores suggests that local convection during winter, possibly associated with lead refreezing, played a role. Additionally, the geographic location of the eddies – well within the Transpolar Drift path – indicates that they may have formed upstream, in regions influenced by freshwater input from the Siberian shelves. This supports the hypothesis that baroclinic instability, facilitated by strong vertical stratification and preconditioning from prior surface forcing (e.g., convection in leads), is a plausible generation mechanism . Another hypothesis, supported by observations and modelling, suggests that baroclinic instabilities – largely independent from surface conditions due to the persistent stratification – could be the dominant generation mechanism throughout the year . Comparable generation pathways have been documented elsewhere in the Arctic, where jets flowing along topographic gradients or strong shelf–basin density fronts trigger instabilities that form subsurface eddies and intermediate layers . Although the central Amundsen Basin lacks the intense boundary-current jets present near the Siberian margin, these studies illustrate how localised shear and preconditioning can seed baroclinic instabilities in strongly stratified Arctic environments, and would favour the hypothesis that some of the eddies we observed formed in the basin near the Siberian continental slope. These mechanisms are not mutually exclusive: thermohaline convection in leads may precondition the water column, creating vertical shear and density structures that enable baroclinic instability. Thus, eddy generation may result from a combination of surface-driven convection and deeper baroclinic adjustment, even in the basin interior.