Hydrographic data were acquired using the Seaglider SG563, which was deployed from RV Johan Hjort on 19 January 2023 and recovered on 8 June 2023 by RV G.O. Sars, completing 620 dives in total. Seagliders are small, autonomous, remotely piloted, buoyancy-driven vehicles which profile to a maximum depth of 1000 m in a sawtooth pattern . SG563 carried a CT sail measuring conductivity and temperature. Sampling occurred every 10 s (∼1 m vertical resolution at typical vertical speeds of ∼0.1 ms-1) in the upper part of the water column to 300 m and every 15 s (∼1.5 m vertical resolution) below that. SG563 was deployed on the eastern side of the Lofoten Basin and proceeded west, arriving in the vicinity of the LV on 3 March. It completed approximately 200 dives in the vicinity of the LV in the period between 3 March–7 May and then proceeded northeast before being recovered (Fig. ).

The raw data were processed using the University of East Anglia Seaglider toolbox (https://bitbucket.org/bastienqueste/uea-seaglider-toolbox/ commit 025191f, last access 11 January 2023). The Seaglider hydrodynamic flight model was tuned following . Depth-averaged currents (DACs) were calculated from the difference between the glider's flight path found from GPS positions at the beginning and end of each dive and the glider's flight path relative to the water as calculated using dead reckoning and the Seaglider hydrodynamic flight model. The thermal lag of the CT sail was corrected following the methods of . Occasional poor-quality data (e.g. from poor flushing of the conductivity cell when the glider is moving slowly) were flagged and discarded, and outliers in the salinity profiles were removed using a three-run Hampel filter ; together, these account for 3 % of the total data collected. Conservative Temperature and Absolute Salinity were calculated using the thermodynamic equation of sea water , and all references to temperature and salinity from this point onward should be taken to mean Conservative Temperature and Absolute Salinity. The potential density anomaly (σ0) was calculated relative to the surface pressure.

Temperature and salinity were calibrated against available ship conductivity, temperature, and depth (CTD) data managed by the Institute of Marine Research in Bergen, Norway. CTD casts were chosen which were collected within 5 km and 3 d of a glider profile; the top 300 m was excluded to avoid the rapidly changing effects of surface forcing, and data within any pycnoclines were also excluded. These selection criteria were chosen to maximise the number of profiles used for calibration while still ensuring that the matching profiles were sampling the same water masses. In particular, the selection criteria ensured that glider profiles within the LV were matched to CTD profiles within the LV, whereas glider profiles outside the LV were matched to CTD profiles outside the LV since the water column properties inside and outside the LV are significantly different. This led to a total of 13 glider–CTD profile pairs being used for calibration. Salinity was corrected by an offset of 0.021 gkg-1, and temperature was corrected by an offset of 0.061 °C. The glider data are available from the Norwegian Marine Data Centre, including the final, quality-controlled data and the associated raw data files .

Seagliders have a navigation mode which is well suited for piloting inside eddies: steering relative to the DAC of the previous dive. By steering at 90° to the previous dive's DAC, the glider will spiral into the core of the LV, which can be easily identified by pilots based on the deep core of warm Atlantic Water. Once the core is reached, the glider can be commanded to steer at 270° to the previous dive's DAC and will spiral out of the LV. These spiralling tracks, henceforth referred to as vortex realisations, take approximately 4–6 d to complete. This method has been used to study the LV previously .

Each LV realisation aims to characterise the LV's hydrography and dynamics in radial section. First the glider's DACs are used to detect the LV centre by minimising a cost function applied to a rolling window of four consecutive DACs to find the position where DACs are most perpendicular to the vectors joining their positions to the eddy centre . Detected centres are only kept when the glider is reasonably close to the LV's core, where the cost function minimum is properly defined by closed contours. Geographical coordinates of glider sampling points are then transformed into cylindrical coordinates (r,θ,z), neglecting the eddy's ellipticity, with r the radial distance from the LV centre, θ the azimuthal angle, and z the depth. As in , temperature, salinity and DAC data are bin-averaged on a regular grid (3 km in radial distance, 5 m in vertical) and optimally interpolated using correlation scales typical of the LV's radius (Lr=15 km) and of the seasonal thermocline thickness (Lz=15 m). In our analysis, we used the optimally interpolated radial sections. Measurements along the dives of the glider, before this bin-averaging and optimal interpolation, are shown in Fig.  to highlight the lateral resolution of the profiles. The cyclogeostrophic balance is finally solved for azimuthal velocities (see Appendix A in ), as the strong vorticity of the LV implies important effects of the non-linear centrifugal force.

The following LV characteristics are described for each realisation. The radius Rv is defined as the radial distance of the velocity maximum vm. Water properties of the inner core and outside the LV are defined as the mean profiles located at a distance inferior to 2/3Rv and between 3Rv and 5Rv, respectively. LV vorticity is obtained in cylindrical coordinates, as in , using 1ζ(r,z)=1r∂(rvθ)∂r, where vθ(r,z) is the azimuthal velocity, and the Rossby number Ro=ζ/f is the vorticity normalised by f, the Coriolis parameter. PV (m-1s-1) is finally calculated as 2PV(r,z)=1gN2(f+ζ)-∂vθ∂z∂b∂r, with b=-σ0/(ρ0g) the buoyancy, σ0 the potential density anomaly, g=9.81 ms-2 the gravitational acceleration, ρ0=1028 kgm-3 a reference seawater density, and N2=∂b/∂z the buoyancy frequency.

Level-3 SWOT satellite altimetry data were retrieved in their last validated version (v2.0.1, ) during the Cal/Val phase from 29 March 2023 to 10 July 2023 for tracks 5 (descending pass) and 14 (ascending pass) from the AVISO ftp server as these two tracks have a cross-over in the LV region. Fields of noiseless Sea Surface Height Anomaly (SSHA) and Mean Dynamic Topography (MDT) at 2 km resolution were used to infer Absolute Dynamic Topography (ADT) as follows: ADT=SSHA+MDT. Here, we remove the monthly average of the ADT over the study area of 68.5–71.5° N, 0–6° E and refer to what remains as the local ADT anomaly (ADTa). Horizontal velocities (u,v) satisfying the cyclogeostrophic balance were computed using Jaxparrow open-source software . The variational approach of this new method allows one to solve the gradient wind balance, whereas the iterative method previously developed by for larger-scale gridded altimetry does not consistently converge when applied to high-resolution SWOT data. Relative vorticity maps were finally obtained from cyclogeostrophic velocities by calculating 3ζ=∂v∂x-∂u∂y.

In order to refine the characteristics of the LV and the merging eddy during this critical period, we applied an eddy detection method to the SWOT-based high-resolution ADT in April 2023. For each individual SWOT swath sampled during the Cal/Val phase, the maximum of ADT in the study area was considered to be the LV centre position, (xLV,yLV). The parameters (ALV0,ALV,RLV) of an idealised Gaussian eddy, ADTLV(x,y)=ALV04+ALVexp⁡-(x-xLV)2+(y-yLV)2/2RLV2, were then derived from SWOT ADT. ALV0 was set as the median of ADT in the far-field, considered to be the region between 100–150 km away from the LV centre. ALV was set as the maximum of ADT (i.e. at the centre of the LV). Finally, RLV was inferred from a least-square regression between the observed ADT and the Gaussian model in a region less than 45 km from the LV centre. Note that this definition is consistent with a radius of maximum azimuthal velocity. This procedure was repeated on ADT-ADTLV in order to characterise the approaching eddy [(xB,yB),AB0,AB,RB]. As will be discussed below, the approaching eddy is eddy B, whose track is shown in Fig. . To avoid outliers and detection at the swaths' edges, these estimates were not considered in the analysis when less than 75 % of a disc of radius RLV or RB was covered by SWOT data and when the root mean square error between the Gaussian model and the measured ADT was larger than 5 cm in a disc of 100 km radius. The eddies' dynamical intensity was then characterised in terms of maximum Rossby number at the eddy centre by resolving the cyclogeostrophic balance for the Gaussian model: 5Romax⁡=-1+1-4gAi/f2Ri2forRogeomax⁡=-2gAi/f2Ri2>-0.5andRomax⁡=-1otherwise; with i=(LV,B) and with Rogeomax⁡ the maximum Rossby number in geostrophic balance. It is worth noticing that the RLV found with this method is about 30–35 km, which is larger than the glider-inferred Rv but is coherent with previously reported LV radii inferred from satellite data. This larger radius from the satellite data than from the in situ glider observations is most likely due to the Gaussian fit smoothing the finer-scale velocity signal of the core.

A typical error for the amplitude estimate (Ai) was estimated by examining the distribution of the background ADT signal (Ai0) over the merging period, leading to an error of 2 cm. Error in the radius estimate could be related to SWOT gridded product resolution; hence, we chose an error of 2 km for Ri. These error estimates were then propagated to Romax using its mathematical definition.

Eddy tracks were taken from the altimetric Mesoscale Eddy Trajectories Atlas (META3.2exp NRT, henceforth META3.2), which is produced by SSALTO/DUACS and distributed by AVISO+ (https://www.aviso.altimetry.fr/, last access: 14 August 2024) with support from CNES (France), in collaboration with IMEDEA (Spain) . This dataset uses data from previous-generation altimetric satellites such as Jason, Sentinel, and TOPEX/Poseidon and does not incorporate SWOT data. We show only those tracks where at least part of the track lies within the region 68.5–71.5° N, 0–10° E in the time period of 1 January–30 June 2023 (Fig. ). Two of these eddy tracks will be discussed in this paper: eddy A, which was tracked from 17 February to 12 March and is shown as a thick blue line in Fig. , and eddy B, which was tracked from 21 March to 4 April and is shown as a thick orange-red line in Fig. . Using these data allows us to extend the period in which we have data on eddy movements before the start of the SWOT Cal/Val phase.

Surface forcing data were taken from the ECMWF ERA5 reanalysis and are shown in Fig. . ERA5 has a temporal resolution of 1 h and a horizontal resolution of 31 km. Each of the surface forcing variables was averaged over the area 69.5–70° N, 2.5–4.1° E (i.e. over 21 grid points), covering the typical variability in the location of the LV. This also agrees well with the vortex centre locations found in our data (see below). Mean net surface heat flux (Qnet, in W m-2) was then calculated as the sum of the short-wave, long-wave, sensible and latent heat fluxes. The mean net surface freshwater flux (FWnet, in kgm-2) was calculated as the mean total precipitation rate minus the mean evaporation rate.

To assess the sensitivity of the derived variables (e.g. azimuthal velocity maximum, relative vorticity, Rossby number) to various sources of uncertainty, we conducted a series of Monte Carlo simulations , each isolating a specific source of error. For the measurement uncertainties of temperature and salinity, we generated 1000 sets (i.e. temperature and salinity) of perturbed data fields by adding normally distributed errors, with a prescribed standard deviation, to each measurement point. Each set of perturbed fields is then analysed by applying transformation into cylindrical coordinates, binning, optimal interpolation, and calculations as required to obtain the derived variables, exactly as in Sect. , giving 1000 estimates of each derived variable. The measurement uncertainty of the DACs is treated in the same way. Assessing the errors introduced by the assumption of axial symmetry involved applying scaling factors to the radial distances from the centre of the LV to profiles and DAC measurements and again generating 1000 realisations. These scaling factors were inferred from the vortex ellipticity seen in the SWOT data. Further details can be found in Appendix . We report the mean, standard deviation and 95 % confidence intervals of the resulting distributions obtained from the 1000 estimates for each source of uncertainty.

The results from this error analysis are detailed in Tables  and in Appendix . The errors from the CT sail measurement uncertainty are trivially small, but those from the DAC measurement uncertainty and from the assumption of axial symmetry are significant. When stating errors within the body of this paper, we give a conservative error estimate using the largest standard deviation in each case. For vortex realisation 3, the largest standard deviation for the maximum azimuthal velocity is derived from the DAC measurement uncertainty. In all other cases, the largest standard deviation is found from the assumption of axial symmetry, specifically from assuming that the glider's transect was aligned with the major axis of the LV ellipse. Thus the errors stated below may be considered to be the “worst case” and are likely to be an overestimate of the true uncertainty.

In realisation 1, (Fig. , first column) the LV displays a well-mixed winter core with weakly stratified waters extending to 1000 m. Core profiles of temperature, salinity and PV are vertically uniform, and the density is well-mixed throughout the water column to 1000 m (Fig. ). The core of the LV (radial distance of about 10–15 km in Fig. ) is surrounded by a rim of increased azimuthal velocity, reaching a maximum of 0.4(±0.05) ms-1 (Fig. g). Low PV is seen throughout the core from the surface to 1000 m (Fig. p). The shape of the low-PV area is lens-like but extends further out of the core between depths of approximately 50–500 m. The relative vorticity of the core is strongly anticyclonic, reaching -0.57(±0.12)f (Fig. ). Note this is not a double-core vertical structure, such as many previous observations have seen e.g., but instead shows the result of winter mixing to at least 1000 m, similar to the earlier observations of and . In March, prior to realisation 1, net surface heat losses were large (-400 to -650 Wm-2), and wind stress exceeded 1 Nm-2 in several episodes, suggesting strong convection and atmospheric forcing (Fig. ).

Outside the LV, profiles are typical of the Lofoten Basin, with a thin (approximately 20 m), fresh and cold surface layer and somewhat warmer and saltier waters beneath, which then gradually cool and freshen towards 1000 m (Fig. ). The change in properties is almost density-compensating from the surface to 100 m (Figs.  and ), and even to 400 m the density changes are small.

The LV is clearly visible as a persistent, approximately circular area of elevated ADTa (Fig. ) and strongly anticyclonic vorticity (Fig. ), centred between 69.75–70° N, 2.5–4° E (consistent with the LV positions in the literature), throughout the SWOT Cal/Val phase. The positions of the LV from META3.2 (grey triangles with black outlines in Fig. ) are close to the positions of maximum ADT in the SWOT data (grey dots with black outlines in Fig. , henceforth SWOT-derived centre positions), with an average separation of 6.0 km in April 2023. On seven days (14–17 and 21–23 April 2023), positions for the centre of the LV were also found from the glider data, as described in Sect. . On these days, the average distance between the LV centre positions from META3.2 and the glider-derived LV centre positions was 4.8 km, and the average distance between the SWOT-derived and glider-derived LV centre positions was 4.9 km. At first glance, this might suggest that META3.2 gives comparable estimates of the centre position of the LV to the SWOT-derived centre positions. However, it is worth noting that, firstly, the sample size is small; secondly, the META3.2 positions are daily estimates, the SWOT-derived positions come from the time of a particular pass of the SWOT altimeter, and the glider-derived positions are based on data from four dives covering a period of approximately 20 h, and thus they do not all represent equivalent time periods; thirdly, the glider-estimated centre positions rely on an assumption of axial symmetry, but there are some slight distortions from circular in the ADTa contours (Fig. ) which will cause some slight alterations in the glider-estimated centre positions. Given these provisos, it seems reasonable to conclude that the META3.2, SWOT-derived, and glider-estimated centre positions are all within a few kilometres of each other, but we cannot say that one position estimate is better than the others during the limited period considered.

At the beginning of the SWOT Cal/Val phase, a smaller anticyclonic eddy is also visible to the east of the LV (Figs.  and ). This eddy will be termed eddy B from now on. Eddy B gradually approached the LV over the period from 21 March to 4 April, as shown by its track from the META3.2 data (thick orange-red track in Fig. ). Eddy B is a region of very high anticyclonic vorticity where the cyclostrophic component is significant compared to other areas (Fig. ). Eddy B gradually elongates (7–8 April) and becomes attached to the southeastern side of the LV (9–12 April), following an approximately circular path around the LV centre (Figs.  and ). Indeed, the evolution of the centre positions of eddy B and the LV (Fig. ) suggests that it is not simply that eddy B travels around the LV but that the LV and eddy B co-rotate, as was previously found in idealised numerical studies . This elongation, attachment and co-rotation are similar to the merging process described by . There are no valid detections of eddy B in the SWOT data on 12–13 April, but between 11 April and 14 April, it seems to have accelerated considerably and wrapped around the LV. Drifting along a circle at about 50 km from the LV centre in three days, eddy B would have travelled at speeds comparable to the LV swirling velocities of about 0.6 ms-1 during this period. It is possible that some of eddy B is ejected to the west around 12–13 April, but we cannot be certain as that area is not covered by the glider observations. Eddy B begins to merge into the LV on 13–14 April as the distance between the two centre positions drops below about 50 km. However, it is still visible as a distortion (on the northern side of the LV at a distance of approximately 30 km from the LV centre) in the otherwise fairly circular ADTa contours around the LV (Figs.  and ) and as a region of strongly anticyclonic vorticity (Fig. ). The glider was then approaching the LV for realisation 3 (12–15 April), and thus on 13 April it recorded temperatures in the uppermost 100 m which are slightly elevated compared to the waters surrounding the LV. On 14 April the glider passed through eddy B and recorded noticeably warmer temperatures in the uppermost 100 m, up to 6.2 °C near the surface and nearly 6 °C for the 0–100 m average. Eddy B is also seen clearly as less dense water at a given depth at around 30 km from the centre of the LV (Fig. ) and displays high eddy kinetic energy (EKE) and higher PV than the LV (Fig. k and q). Finally, on 15 April, the glider reached the core of the LV, which was slightly warmer (5.6 °C near the surface) than the waters surrounding the LV (4.9 °C near the surface) but not as warm as eddy B, which suggests that eddy B had not yet merged into the core of the LV (Fig. b).

The Gaussian fits to the LV and eddy-B ADT, described in Sect.  and illustrated in Fig. , show the evolution of various properties of both the LV and eddy B from 3 to 25 April (Fig. ). Prior to the merger, the LV and eddy B have similar amplitudes (ALV≃AB≃0.25 m), but thanks to its smaller radius, eddy B has a significantly stronger vorticity signature, with Romax≃-1 vs. -0.5 for the LV if one considers cyclogeostrophy (or geostrophy only: Romax≃-0.6 for eddy B vs. -0.35 for the LV). The error in these Ro estimates computed from the SWOT data is estimated to be 0.07. As the distance between the LV and eddy B decreases, the LV's amplitude and radius increase, while eddy B's amplitude and radius decrease, suggesting that the interaction between the two vortices results in an exchange of mass from eddy B to the LV. The rapid change in these properties should be treated with some caution, however: as eddy B and the LV become more elongated due to their mutual interaction, a circular Gaussian fit becomes less representative. The absorption of eddy B by the LV, and potential ejection of a low-PV parcel from the LV, could also explain the transitional increase in the LV radius (RLV) observed between 15–20 April, indicating the final stage of the merger. Eddy B's amplitude slowly decreases from 3 April as the two approach each other, but then it drops more rapidly from 7 April onwards as the distance between the LV and eddy B becomes smaller than approximately 80 km, testifying to a stronger interaction between them. The merging process seems to lead to the dissipation of eddy B despite an apparently more dynamical initial Rossby number at the surface (Fig. c and Fig. ). In terms of Rossby number, eddy B quickly loses its strong signature as the interaction with the LV begins, even before the actual merger with the LV core. It is also worth noting that the LV's cyclogeostrophic Rossby number seems to increase in absolute value from about -0.4 to -0.5 between 3–23 April.

Although the glider passed through eddy B on 14 April and reached the LV core on 15 April, eddy B had not yet reached the LV core by 15 April: the glider was moving faster than the eddy was merging. Hence, the core profiles during realisation 3 are actually profiles from just before eddy B merged in. The LV core is already more stratified than during realisation 1, with warmer temperatures near the surface and a much wider range in density than in realisation 1 (Fig. ). It has been previously reported that, during winters, a weakened PV gradient in the upper part of the LV between the core and the waters outside the LV reduces the dynamical barrier and facilitates the intrusion of warm waters from the vortex rim . While one could speculate that the core and outer temperatures of realisation 1 (Fig. a and d) might mix laterally to generate the core temperatures of realisation 3 (and similarly for PV, Fig. c and f), the salinity in the core in realisation 3 is fresher than both the core and outer salinities of realisation 1 and thus cannot be formed from a mixture of the two. This strongly suggests that another water mass has contributed to the evolution of the core structure. We speculate that another eddy (eddy A) has already merged into the LV in March for the following reasons:

From 14 March (when the LV core was observed at the end of realisation 1) to 15 April (when the LV core was observed at the end of realisation 3), the average net surface heat flux over the mean position of the LV was -107 Wm-2 (Fig. a). A negative surface heat flux cannot explain an increase in temperature in the LV core, and so there must have been another source of warmer water.

To assess the properties of eddy B, observed in April, we use a simple definition to distinguish it from surrounding waters: water with temperatures which are at least 0.2 °C warmer than the temperature of the LV core at the same depth are taken to be within eddy B. This is shown as an orange-red contour in Fig. . The properties of eddy B are calculated as an average in each depth bin over the waters enclosed by this contour and are shown in orange-red in Fig. . By the time of realisation 5 (18–22 April), the LV core profiles to 400 m are similar to those observed in eddy B a week earlier (Fig. ). Although found that incoming eddies tended to be lighter than the LV and thus stacked on top of it during mergers to create a double-core vertical structure, we do not observe that here. This is because eddy B is not significantly lighter than the LV: the probable merger event in March has already reduced the density of the upper layers of the LV. In the second column of Fig. , a magenta contour marks the minimum potential density observed in the LV core during this realisation (σ0=27.6) and is equivalent to the red contour in 's Fig. 14. The example of an incoming eddy shown in 's Fig. 14 is lighter than the LV to a depth of approximately 600 m, but in realisation 3, eddy B is only lighter than the lightest water in the LV core to a depth of 40 m. Eddy B therefore simply mixes with the LV rather than stacking on top of it, as also happens to eddies of the same density as the LV in the work of . However, eddy B's signature in temperature extends to approximately 400 m depth, and thus it is lighter than the LV core at the same depth, as seen by the deepening of the isopycnals in eddy B compared to the LV core (Fig. , second column). Water in the LV core is therefore pushed downward during the merger, as seen by the lowered isopycnals in the LV core in realisation 5 compared to in realisation 3 (Fig. ). We therefore see the vortex intensification and change in relative vorticity as predicted by : in realisation 5, after the merger, the LV's vorticity is even more strongly anticyclonic than in realisation 1, reaching -0.79(±0.19)f (Fig. o).

As with the probable merger event in March, we consider whether surface forcing could have brought about the changes in the core properties observed in April. It has not been possible to determine this for salinity because the net change in core salinity was extremely small in the period from 15 to 22 April (i.e. between realisations 3 and 5); a net freshening of only 0.002 gkg-1 over the full water column observed. The average net surface freshwater flux was 9.3×10-6 kgm-2s-1, which is also small, but in combination with the eddy merger, lowering of isopycnals, and possible leakage from the area outside the vortex, it was not possible to separate the different effects. The situation is simpler for temperature because both eddy B and the average net surface heat flux acted to warm the LV. However, the average net surface heat flux of 23 Wm-2 is an order of magnitude too small to bring about the observed temperature change in the LV core; hence, the effect of eddy B merging is dominant.