The ocean model is OPA, which is part of the NEMO3.6 modelling platform . We use the regional CREG025 configuration , which is a regional extraction of the global ORCA025 configuration developed by the Drakkar consortium . It encompasses the Arctic and parts of the North Atlantic down to 27∘ and has 75 vertical levels and a nominal horizontal resolution of 1/4∘(≃ 12 km in the Arctic basin). Initial conditions for the ocean are taken from the World Ocean Atlas 2009 climatology for temperature and salinity. For the lateral open boundaries, monthly climatological conditions (comprising sea surface height, 3-D velocities, temperature, and salinity) are taken from a long ORCA025 simulation performed by the Drakkar group.

The sea ice model is version 2 of the neXtSIM model as recently presented in . It uses the brittle Bingham–Maxwell (BBM) rheology to represent sea ice dynamics. Sea ice thermodynamics are the same as described in Appendix A2 of . In short, the model considers three ice categories: “thick” ice, open water, and newly formed “young” ice. The young ice is made of ice formed from the super-cooling of open water, and this ice category is associated with marginal ice zones (MIZs) and openings in the ice cover (leads and polynyas). This scheme is used to represent the rapid growth of newly formed ice (young ice, frazil, nilas, etc.) from a minimum thickness hmin⁡, set to 5 cm to a maximum thickness hmax⁡ that corresponds to the transition to thicker, consolidated ice. Young ice is redistributed in the “thick” ice category once its thickness exceeds hmax⁡. Increasing hmax⁡ enhances ice growth in the winter. We found that a value of hmax⁡=18 cm gives a reasonable winter sea ice thickness in our simulation (see Sect. ). The source and sink terms from the thermodynamics are computed by applying the zero-layer vertical thermodynamics to the young ice category and that of for the thick ice. We do not use an explicit melt-pond scheme, but the albedo scheme we use (the same as the standard albedo scheme “ccsm3” used in CICE, ) accounts for the effect of melt ponds by reducing the albedo value when the surface temperature of sea ice increases. It is likely that the use of an explicit melt-pond scheme e.g. or more complex representations of processes related to brine (instead of a constant salinity here) or snow would affect the sea ice extent and thickness in our results, but the effect of using another parameterization could only be assessed after a re-tuning of the model (as in ).

The main (advection) model time step is 450 s, with 120 sub-cycles used to solve the dynamics, resulting in a dynamical time step of 6 s. The use of a coupled system has required some changes to the default values of the sea ice parameters that were used for simulations using neXtSIM in its standalone configuration (as in ). These changes are summarized in Table . For the dynamics, our setup is very similar to the one described in , with the exception of small changes in the values of the ice–ocean and ice–atmosphere drag coefficients and a decrease in the value of the scaling parameter for the ridging threshold, P. These changes are needed to ensure that the ice–ocean and ice–atmosphere stresses are properly balanced against the internal ice strength, since this balance is different when neXtSIM is coupled compared to a standalone setup. The stress values are chosen to match the observed large-scale drift and thickness distribution as closely as possible, while still maintaining good deformation patterns and statistics (see Appendix ). We use the ice grounding scheme from to represent landfast ice as in, with a critical thickness parameter for ice grounding of k1=5 see. The initial fields of sea ice thickness and concentration are taken from the same long ORCA025 simulation as was used for the ocean lateral boundaries and climatological conditions.

OPA and neXtSIM are coupled via the OASIS-MCT coupler . The coupling time step is taken to be equal to the ocean model time step (twice the ice model time step) of 900 s. As summarized in Fig. , OPA receives all the required information about surface fluxes (heat and salt) and stresses from neXtSIM. In return, OPA sends back information about properties of the ocean surface to neXtSIM. This includes sea surface temperature, salinity, height, and currents, as well as the absorbed fraction of net solar radiation. NEMO includes a coupling interface to run the component model OPA and an ice model (LIM3, SI3, or CICE) coupled through OASIS. We make use of this coupling interface here, albeit with the minor modifications required to allow for vector orientation on the ocean and ice model grids to be different (as is implicitly assumed in the NEMO code).

One of neXtSIM's particularities is that it uses a Lagrangian moving mesh . This ensures little numerical diffusion, which is a condition required to obtain a good localization of sea ice deformations. However, this particularity makes the interface between neXtSIM and OASIS more complex than a standard coupling interface, as OASIS is not able to accommodate a moving mesh. Therefore, we chose to implement a fixed exchange grid within neXtSIM, which we use to interface with OASIS. For this exchange, neXtSIM interpolates all coupling quantities between the moving mesh and the exchange grid internally, while all communications with OASIS are carried out on the exchange grid (Fig. ). This interpolation is done by averaging exchanged quantities weighted by the area of overlap between triangles of the mesh and quadrangles of the exchange grid in a conserving way. The interpolation weights are recomputed after each Lagrangian regridding.

Heat fluxes between the ocean, ice, and atmosphere are computed using traditional bulk formulae. For ocean–atmosphere exchanges, bulk formulae from OPA have been implemented in neXtSIM. This was done using the AeroBulk library , upon which OPA has relied since version 4. The bulk formulae for atmosphere–ice fluxes are described in . The bulk formulae for ice–ocean fluxes are the same as in the Louvain-La-Neuve Sea Ice Model (LIM) version 3.6 .

The model simulation, hereafter referred to as OPA-nex, starts on 1 January 1995 and runs until 31 December 2018. Atmospheric forcings are taken from the hourly, 1/4∘ horizontal resolution, ERA5 reanalysis . We exclude the first 5 years of the simulation from the analysis to account for model spin-up. To estimate the length of the spin-up, we applied different initial ice conditions on 1 January 1995 and found no sensitivity of our results over the period 2000–2018. All output variables from neXtSIM are interpolated using a conservative scheme from the moving Lagrangian model mesh onto a fixed and regular Eulerian grid and are averaged on a 6-hourly basis.

Our analysis focuses on sea ice properties in the Arctic region (Fig. ), where sea ice deformations in neXtSIM have been evaluated before (for instance in ). We divide the Arctic region into eight sub-domains. The first six are similar to those considered in (corresponding to the Barents, Kara, Laptev, East Siberian, Chukchi, and Beaufort seas). We also consider two additional regions, the eastern and western sectors of the central Arctic (the sub-domains labelled 7 and 8 in Fig. , respectively). The eastern sector is typically covered by first-year ice being advected towards Fram Strait following the Transpolar Drift, while the western sector is mostly covered by multiyear ice, which is generally thicker (> 2 m) and less mobile than the ice present in the eastern sector.

We take sea ice concentration from the climate data record of the EUMETSAT Ocean and Sea Ice Satellite Application Facility OSI-SAF,. To cover the period 2000–2018, we use two different versions of the product: the OSI-450 (1980–2015) and the OSI-430-b (2016–present). Sea ice volume and thickness are evaluated using two independent datasets: outputs from the Pan-Arctic Ice-Ocean Modeling and Assimilation System PIOMAS, and data produced by combining the observations retrieved from the CryoSAT-2 and SMOS satellites, referred to as CS2SMOS version 2.2,. PIOMAS data are the results of coupled ocean–sea ice model simulations with the daily assimilation of satellite sea ice concentration and sea surface temperature. The main interest of the PIOMAS dataset is that it is available for the whole simulated period and has been thoroughly evaluated against ice thickness observations (e.g. ), meaning that some of its biases are known. For this reason, it is regularly used as a reference for large-scale sea ice thickness evolution in the Arctic (e.g. ). For evaluating the sea ice drift, we use the low-resolution OSI-SAF sea ice drift product that provides 2 d integrated sea ice displacement . This product includes information about summer ice drift and the uncertainties associated with the drift retrieval from June 2017 onwards.

For each dataset, we process OPA-nex output in order to compare them with observations in a consistent way. First, we integrate in time the 6-hourly OPA-nex output to obtain an output with the same time frequency as the observations (i.e. daily for sea ice concentration, 2 d displacement for the drift, and weekly for the thickness/volume in CS2SMOS). We then interpolate OPA-nex outputs onto the same grid as the observations. There is an additional step for the ice thickness and volume: as PIOMAS is only available monthly, we do a monthly average of OPA-nex outputs and CS2SMOS data and compare the three datasets with PIOMAS data interpolated on the CS2SMOS 25 km grid.

We also evaluate the ice volume export through Fram Strait, as it is an important contributor to the Arctic sea ice mass balance . We use the time series produced by , which covers the period 1992–2014, overlapping with most of our simulation. This dataset is based on sea ice thickness measurements derived from one to four upward-looking sonar measurements installed on moorings across the strait (from which the section of ice thickness is extrapolated), and sea ice drift is retrieved from the JPL sea ice motion dataset . We also compare OPA-nex results with the time series of , spanning 2003–2009, that is obtained by combining sea ice thickness from ICESat altimeter observations and sea ice area and drift retrieved from AMSR-E 89 GHz. To compare our export to these datasets in a consistent way, we first estimate the sea ice transport (i.e. motion vectors) at the output frequency of OPA-nex (6 h), and then we interpolate these transports onto the same grid as . The gate used for the computation is located at ≃79∘ N (see Fig. ).

We also compare the simulated sea ice area flux through Fram Strait to the time series estimated by . They combined a blended historical and modern record of sea ice concentration from the National Snow and Ice Data Center NSIDC, with sea level pressure observations across Fram Strait to retrieve the sea ice area flux over the period 1935–2015. Based on the 6-hourly sea ice drift and concentration outputs from OPA-nex, we estimate a time series of the sea ice area flux across the same section at 79∘ N (see Fig. ).

In order to evaluate the ice mass balance and its spatio-temporal variations, we make use of the ice volume change dataset from . The originality of this dataset is that it separates the dynamic contribution (i.e. the import and export of ice in each region) from the thermodynamic contribution (i.e. the net sea ice growth in winter) to ice volume change in the freezing season (November to March) for six Arctic seas (Barents, Kara, Laptev, East Siberian, Chukchi, and Beaufort; see Fig. ). To estimate these contributions, combine sea ice volume and motion information to retrieve the dynamic volume change and then subtract it from the total volume change to estimate the net sea ice growth. also compare their estimates with outputs from two models, PIOMAS and NAOSIM . In OPA-nex, we compute the dynamic volume changes in each region from the transports across the gates shown in Fig. . Thermodynamic volume changes are directly output from the model.

We first evaluate the large-scale properties of the simulated sea ice. Our computations of bias, RMSE, and integrated ice-edge error IIEE, are done in a similar way to (Sect. 4.1). The evaluation of small-scale dynamics of sea ice in the coupled neXtSIM–OPA setup provided no qualitative differences in sea ice deformations compared to a standalone setup (see Fig.  and ).

We start our evaluation with the sea ice extent (Fig. ). To quantify the agreement between OPA-nex and the OSI-SAF data over the study domain, we compute the integrated ice-edge error (IIEE), a metric used to evaluate the quality of predicted ice extent that accounts for both errors in the absolute extent and misplacement of ice . The IIEE in our study domain is almost zero in winter (December to May), mostly because we limit our analysis to the Arctic Basin, which is fully covered in ice in those months. If we extend our analysis to the whole model domain (that also includes most of the North Atlantic, Hudson Bay, and the Baltic Sea, but not the Pacific side), the modelled sea ice extent remains consistent with observations over the winter. The IIEE remains low (0.62 M km2 in average in March) and almost constant over the winter, with a small tendency of the model to underestimate the maximum extent. The IIEE increases in summer and peaks in September (1.7 M km2 on average for this month), mostly due to misplacement of the summer minimum extent (as the absolute value of the extent is generally well estimated, with the exception of 2016 and 2017). Therefore, OPA-nex successfully captures the seasonal cycle of the ice extent, and its interannual variability, particularly in the summer, with the exception of 2016 and 2017 where the ice loss is overestimated.

We then examine sea ice volume (Fig. a). Agreement with PIOMAS is generally good, although sea ice volume in OPA-nex is generally lower than in PIOMAS in the early 2000s. After 2008, the agreement becomes very good in both winter and summer, and the two models show a similar interannual variability. A lot of factors could explain the discrepancies between OPA-nex and PIOMAS (differences in atmospheric forcings, differences in the dynamics and thermodynamics of the models, and the use of data assimilation in PIOMAS), and it is difficult to attribute these differences to one of these factors or the other. OPA-nex agrees well with CS2SMOS, with an average RMSE of 0.34 m and an average bias of 0.03 m for the whole period when observations are available (from October to April each year from 2011). This is also true for the ice thickness distribution during the ice growth season (Fig. b, c, d). Biases in the distribution in OPA-nex compared to CS2SMOS are quite similar to those in PIOMAS and also found in most sea ice models: thick ice is not as thick as the observations in the central Arctic, while thin ice is often too thick, particularly on the western side of the Arctic Basin .

The simulated drift generally shows a good agreement with the OSI-SAF data (Fig. ), with a low negative bias (-0.35 km d-1 on average from 2010 to 2018) and a low RMSE (3.82 km d-1) for the freezing season (October to April), when most of the data are available. OPA-nex also captures both seasonal and interannual variability (Fig. a) and the day-to-day variability (Fig. b). Uncertainty and drift estimates in the summer only start in June 2017, explaining our choice of zooming in on the year 2018 in Fig. b. From Fig. b, modelled summer ice drift is overestimated in OPA-nex compared to OSI-SAF but remains within the larger uncertainties of observations during the melting season. Importantly, the variability remains well captured year round.

We quantify the contribution of each source and sink of sea ice over time in the domain of interest (Fig. ) for the whole study period (Fig. a, b). As in Sect. , we partition the sources and sinks of sea ice into dynamic and thermodynamic processes. The dynamic processes are sea ice transport through Fram Strait (in green) and the sum of ice transport through all the other external boundaries of the domain (Fig. ). The thermodynamic processes are the same as in Fig. . This way of presenting the mass balance is similar to what have done for sea ice components of climate models that are part of the latest Coupled Model Intercomparison Project (CMIP6). The only difference is that the process we call “growth of young ice” includes both the ice volume of frazil ice production and the ice volume corresponding to the growth of this frazil ice until it enters the consolidated thicker ice category, which occurs when the thickness of the young ice exceeds hmax⁡ (see Sect.  for details). This definition of young ice is broadly similar to the one of the World Meteorological Organization (forming ice thinner than 30 cm, ).

We find that the interannual variability of the net mass balance in our domain is controlled by both thermodynamics and the export through Fram Strait, with these two terms being of similar amplitude (Figs. a and ). Sea ice transport through other gates is negligible compared to the export through Fram Strait (hence, they are almost not visible in Fig. a, b). Previous reports suggest that Fram Strait represents ≃ 90 % of the net sea ice export of the Arctic, the second main source of export being through Davis Strait south of our domain . In our case, the contributions from all gates other than Fram Strait almost cancel out, being slightly positive (ca. +1×103 km3 over 2000–2018 against around -30×103 km3 through Fram Strait). This is likely because (i) the Canadian Arctic Archipelago is not included in our analysis (and we therefore miss all the ice that forms there and is then exported through Davis Strait) and (ii) 12 km is too coarse to resolve the outflow through Nares Strait, leading to an underestimated export through this narrow gate (only ≃ 1 km3 yr-1 in the model, while observations suggest an average up to ≃ 190 km3 yr-1 over 2017–2019, ). If we consider the seasonal cycle of Arctic sea ice volume, sea ice export only plays a minor role in the variations in the ice volume, which are mostly driven by sea ice thermodynamics (Fig. b). We note that this seasonal cycle is very similar to the multi-model mean seasonal cycle presented in (see their Fig. 4a), even though the framework (fully coupled climate models, 1960–1989 climatology) is different.

Sea ice production slightly exceeds melt in the domain (with the exception of 2016, when the net ice production becomes negative; Figs. a and ). This is because we have excluded the domain south of Fram Strait, where a large part of the melting occurs (Fig. d). The yearly amount of ice growth in the domain is closely linked to the amount of melt (Fig. a). This is most likely because strong melt events lead to large areas of open water and thinner ice at the end of the summer, enhancing the refreezing in the next autumn and winter . We do not find any trend in sea ice growth nor melt using this domain, and large changes in the total ice volume (as in 2002, 2012, 2014, or 2016) are mostly associated with the interannual variability of the balance between melt and growth (Fig. ).

Overall, interannual variations in the net volume change associated with thermodynamic processes are mostly due to variations in the basal growth of thick ice and basal and surface melt (Fig. a). Basal and surface melt contribute about equally to the yearly ice melt in the domain, while the basal melt dominates south of Fram Strait (outside the study domain), likely because sea ice encounters warmer surface waters in the Greenland Sea . Young ice growth accounts for about half of the yearly ice production. This proportion is sensitive to the choice of minimum and maximum thickness for the young ice in our three-category thermodynamics scheme in OPA-nex (Sect. ), as is the case in most models . Young ice growth variability is weaker than that of basal growth and is mostly controlled by two drivers: the ice extent at the end of the summer (positive anomalies are found in, e.g. 2008, 2013, or 2017, which are years following low-extent anomalies), and the amount of openings (leads or polynyas) that are present in pack ice. Ice production due to the flooding of snow is negligible for the area and time period discussed in this analysis. This contribution may, however, be underestimated by the mass-conserving snow-ice formation scheme used . The ice volume loss over time in the domain is clearly visible in Fig. . This loss is qualitatively similar to the winter volume evolution reported by using ice age to estimate the ice volume. We find a statistically significant (p≃0.01) trend of -280 km3 per year over 2000–2008, which is within the range of sea ice volume trends (from both models and observations) discussed in (between ≃-200 and -400 km3), but no significant trend for the period 2009–2018 (also as reported in ). The comparable orders of magnitude of the ice exported and the net change in ice volume due to thermodynamics reinforce the importance of the dynamic contribution to the mass balance that was suggested in Fig. . Net ice production peaks in 2013–2014, when it dominates the net export by a factor of ≃2, resulting in an increase in the ice volume in the domain. Loss of sea ice volume in the domain mostly occurs in years of low (or negative) net ice production (such as 2002, 2012, and 2016). The yearly net sea ice export varies very little in comparison to the net sea ice production.

We now estimate the contribution of leads and polynyas to the winter ice mass balance. This estimate is based on the simulated ice formation in open water and ice growth in the young-ice category (see Sect. ). In winter and in pack ice, such ice growth will only take place where the ice has been recently diverging, because young ice quickly grows thick enough to be transferred to the “old ice” category (a few days at most). In the absence of divergence, the domain would be fully covered by old ice. The following analysis could be carried out with any sea ice model with multiple ice thickness categories. However, the amount of ice produced in openings (i.e. leads and polynyas) in pack ice and its localization are very likely to be strongly impacted by the ability of the model to reproduce the small-scale sea ice dynamics. This is because the highest values of divergence rates (and deformation rates in general) in Arctic pack ice are very localized (Fig. a, b), which would not be the case if the ice cover was homogeneous e.g.. For instance, found that in neXtSIM at 10 km resolution, 50 % of the divergence in the central Arctic was associated with only 5 %–10 % of the surface area in the domain used for the analysis (this surface ratio would be 50 % in the case of a homogeneous ice cover). Divergent ice motion therefore results primarily in the formation of localized leads in the central pack or of polynyas near the coast. An underestimate of divergence rates, which “standard” sea ice models run at resolutions coarser than 5 km tend to do , would therefore imply a subsequent underestimation of ice production in winter if there is not a sufficient parameterization to represent the effect of leads. This parameterization can be done using, for instance, a minimum value for the lead fraction in each grid cell, resulting in a more uniform distribution of lead growth over the domain (as this can be done in the LIM3 model, ). The importance of resolving leads versus using parameterizations to represent the ice growth in leads in numerical models has not been assessed to our knowledge. This would likely require a model comparison between a model that captures divergence rates well and another one using a parameterization for leads, which is out of the scope of this study. Instead, we focus on estimating the importance of ice production in leads in our simulation, as this has not been estimated at a pan-Arctic scale before. The advantage of using neXtSIM in our analysis is that its ability to reproduce small-scale sea ice dynamics has been thoroughly evaluated before seeand Appendix . In addition, it has been shown that the model is able to capture rates of divergence consistent with observations and relevant statistics of the observed lead fraction in the central Arctic at spatial resolutions similar to the one used here and Fig. .

The impact of leads and polynyas on winter ice production is visible in Fig. a and is clearly linked to the growth of young ice (Fig. b). We note that the spatial patterns in Fig. a, b look similar to maps of observed ice divergence or lead fraction distribution described in previous studies (e.g. ). The imprint of leads on ice production is particularly visible in the Beaufort Sea, with long linear features orthogonal to the coast. Their presence in a 18-year-long climatology demonstrates their strong impact on sea ice production and the likely recurrence of these features year after year. Lead-type features are also visible in the central Arctic when looking at the contribution of openings to the total growth (Fig. b). This is likely because the thick ice covering the central Arctic limits the amount of ocean heat loss, and hence basal growth, that can occur, meaning that local openings in this thick ice cover significantly contribute to the total ice growth there. We also note that coastal areas are places of intense production of newly formed ice in winter, likely due to the recurrent opening of coastal polynyas. Before quantifying the impact of leads and polynyas to winter ice production, we assess the limitations of associating the growth of young ice with these types of features. From Fig. c, we can differentiate two phases in ice growth within the freezing season. The first period (October to December) is when ice production is dominated by the growth of young ice from open water as the open water refreezes. Young ice growth occurs in open-water areas in a homogeneous way until the whole Arctic Basin is ice covered. The second phase is from January to March, when the contributions of basal growth and young ice growth reach an equilibrium. This corresponds to a plateau in the contribution of the growth of young ice to total growth, visible every year in Fig. c. Sea ice concentrations in the domain are then very close to 100 % everywhere, which means that the young ice is mostly (if not totally) produced from openings due to divergence in pack ice . There are, however, regions in our domain that remain covered with thin ice in winter or even include open-water areas (e.g. the Barents Sea). To avoid the inclusion of young ice production in MIZs in our analysis, we remove the southern part of the Chukchi Sea, as well as the Barents and Kara seas, from the domain we consider (black contour, Fig. a, b).

We now quantify the contribution of ice growth due to openings within pack ice over the total ice growth. This contribution adds up to ≃ 25 % to 35 % of the ice growth in the domain (black line in Fig. e). This corresponds to an annual winter ice volume production of ≃ 270 to 380 km3 (black line in Fig. d). We find that young ice growth is increasing over the studied period with a significant (i.e. with a p-test result lower than 0.05) positive trend of +74.7 km3 per decade. This is also true for its contribution to total growth, with a positive trend of +4.3 % per decade. Over the same period, the basal and total growth show no significant trends (not shown).

As mentioned above, ice growth from leads and polynyas is particularly substantial in coastal areas. It is therefore interesting to distinguish between coastal areas, where both leads and polynyas can occur, from the interior Arctic Basin, where most openings correspond to leads. We reproduce our analysis, this time excluding the regions shallower than 300 m (cyan dashed contour in Fig. a, b). We find that the interior Arctic basin accounts for about half of the young ice growth from openings in pack ice in winter (grey lines in Fig. d, e). We still find significant positive trends in young ice growth (+29.4 km3 per decade) and its contribution to total ice growth. This means that while ice production in leads in the interior Arctic is a significant contributor in the study domain, it is the coastal areas that are most important to the modelled increase in ice growth in leads and polynyas over 2000–2018 (Fig. d).

We note that most of the increase in the contribution of ice growth in leads and polynyas to total ice growth takes place from 2000 to 2010, with a significant positive trend of about +7 % over this decade (whether the shelves are excluded or not). From 2008 to 2018, we find no significant trend for ice growth in leads and polynyas. Using 2018 as the last year of the analysis, the positive trend in the contribution of ice growth in leads and polynyas to total ice growth is only significant when starting prior to 2005 (included) and prior to 2002 if we exclude the shelves. This highlights the important contribution of the earlier years of the analysis (the period from 2000 to 2008 in particular) to this increasing trend in ice production.

To better understand this evolution of ice growth in leads and polynyas, we now examine each region included in our analysis (Fig. a–f). Although ice growth in leads and polynyas strongly varies from one region to another (Fig. a), its contribution to total growth is similar across the subdomains, around 30 % if we account for coastal areas and about 25 % when we only consider the interior of the basin (Fig. e). However, there are two regions where the behaviour differs from the other regions. The contribution is lower in the western central Arctic (≃ 24 % including coastal areas), likely due to thick sea ice, which is not very mobile (Fig. g, i). In contrast, the eastern central Arctic is characterized by a large contribution of leads and polynyas to sea ice production (≃ 44 %; ≃ 39 % when excluding the shelves), likely due to the large cracks that regularly form as the ice undergoes high internal stresses while exiting through the narrow Fram Strait (as visible in Fig. a and in lead frequency maps in ). Increases in winter ice volume growth due to leads and polynyas (and their contribution to total growth) occur in almost every region over 2000–2018 (Fig. b, f) but are only significant in the regions along the Eurasian coast (Laptev, East Siberian and Chukchi seas) and in the interior of the Chukchi Sea. The largest trends are found in the Laptev and East Siberian seas, as well as in the interior of the Chukchi Sea. Total ice volume growth shows no significant trend anywhere (Fig. d).

We now try to relate this increase in the ice growth in leads and polynyas to two other sea ice quantities that are related to sea ice deformation: sea ice drift and thickness (Fig. g–j). Sea ice drift speed is found to increase over time, with positive significant trends varying between +10 % to +20 % per decade depending on the region (Fig. h). Spatial variability in sea ice drift trend magnitudes follows that of ice growth in leads and polynyas, which suggests a close relationship between the two. As for the ice growth in leads and polynyas, most of this increase in ice drift speed occurs over the period 2000–2008, and there is no significant trend for any region for the period 2008–2018 (not shown). In the case of sea ice thickness, the relationship with ice growth in leads and polynyas is less clear. The thinning trend is significant almost everywhere (Fig. j), but the spatial distribution of the trend magnitude does not reflect the one of ice growth in leads and polynyas. We note that sea ice drift and thickness are not independent, and the increase in ice drift speed is most likely driven by the thinning of the ice and the associated reduction in the ice strength . Similar to what was found in previous studies , we find no or very little trend in the wind speed, with magnitudes that are too low to explain the increase in ice drift (not shown).

Our interpretation of the results is that the increasing trend in the ice drift velocity is associated with higher divergence rates, enhancing winter ice production in openings in the ice pack. This feedback has been suggested before. , for instance, has hypothesized that it could contribute to the resilience of sea ice in the Arctic. This enhanced winter ice production is particularly intense close to the coast and in regions with thin ice, such as the Laptev and Siberian seas (Fig. b, f). These regions are also associated with rather low average ice drift speed in winter, most likely due to the presence of landfast ice. Therefore, landfast ice and the leads and polynyas that form along it likely play an important role in the production of ice in winter along the Eurasian coast.