We here use the new polynya detection algorithm of , developed with the explicit purpose to detect polynyas in CMIP6 models. As reviewed by e.g. for the Southern Ocean and for the Arctic, there is no consensus in the literature regarding the sea ice concentration or thickness threshold that should be used to detect a polynya, be it for observations or models. Besides, prior to applying that threshold one has to perform a flood-fill to separate polynyas and open ocean e.g.. Flood-fill algorithms change adjacent values to a common value, starting from a “seed” point in the open ocean, effectively detecting the ice edge. This is near-impossible to implement in a multi-model comparison framework in the Arctic owing to the models' different grids, in particular their treatment of the North Pole, their choice of reference longitude, and how their resolution represents the Canadian archipelago. therefore took a new approach and developed, on observations and for CMIP6 models, a neural network-based method that detects patterns, gradients in the sea ice rather than absolute concentration or thickness values, and requires no flood-fill.

We now briefly describe how the algorithm works and re-direct the interested reader to the full publication for the details. As summarised on Fig. , the method consists of four steps:

The daily sea ice concentration (SIC) input image is prepared: All pixels with SIC lower than 15 % are set to 0 to increase the sharpness of the ice edge, and all pixels south of 68° N are cut out. A second version is created by interpolating onto the National Snow and Ice Data Center (NSIDC) Northern Hemisphere Equal-Area Scalable Earth Grid 2.0 .

We leave the settings of unchanged for Unet. For the GMM filtering, following their recommendation, we treat the input of step 3 differently for the observations and the CMIP6 models: For the observations, all daily sea ice concentration values lower than 40 % were set to 0; for the CMIP6 models, all values lower than 15 %. We tested also a 30 % threshold for CMIP6 but found no significant difference in the performance; since running the GMM on all models took nearly two weeks, we did not test other thresholds.

The observed daily NSIDC sea ice concentration from 1 December to 31 March, from 1978 to 2024, at a 25 km resolution, has already been classified at the pixel level as polynya/non polynya by . Rather than re-classifying the sea ice data, we here re-use the observed polynya masks of as reference. The NSIDC sea ice concentration is retrieved from passive microwave satellites, processed with the NASA Team algorithm . For Sect.  we generated a monthly observational reference by first computing the monthly mean of the NSIDC sea ice concentration from their daily data, and then passing these monthly sea ice concentration through the Unet + GMM algorithm.

To compare models to the full observational time series (1978–2024), we use both the historical run that finishes on 31 December 2014 and the Shared Socio-economic Pathway “SSP” climate change run of the Climate Model Intercomparison Project phase 6. Our analysis centres on the 18 models from 10 different model families that had daily sea ice concentration “siconc” available at the time of download (summer 2025), listed in Table . They cover nearly three orders of magnitude in horizontal resolution in the Arctic, from just above 100 km2 per grid cell area for AWI-CM-1-1-MR to nearly 10 000 km2 for IPSL-CM5A2-INCA. Most models are of the order of 1000 to 2000 km2, or about 1° longitude/latitude. Their sea ice and ocean components will be further discussed in the analysis. We acknowledge that real-world Arctic polynyas are preferably detected using sea ice thickness rather than concentration e.g.. However, the daily sea ice thickness in CMIP6 is often not available, and it has been reported to have large errors compared to observations on multiple occasions , so we use the more reliable and widely available daily sea ice concentration. We also use the monthly sea ice concentration for a brief comparison; see Appendix Table  for the ensemble members and SSPs.

When possible, we used up to 10 ensemble members per model. The ensemble spread is the first thing analysed in Sect. . We tried to use only SSP2.4-5 but it was not available for some models. In the absence thereof, we used SSP3.7-0, and SSP5-8.5 when it was the only one available; see Appendix Table . As shown in , using different SSPs should have no impact on our results as the scenarios do not diverge before approximately year 2050 whereas we stop in 2024. We nonetheless redid the analyses with the historical run only and find no significant difference (not shown): the values were obviously different, but the correlation between the two was near-perfect, and most importantly, the relative model biases were unchanged.

For investigating the potential causes of polynya biases, we use the monthly sea ice thickness “sivol” (in m, computed as the total volume of sea ice in m3 divided by grid-cell area in m2) since the daily thickness has large errors, as we described above, or is not available. Note that the variable “sithick” exists but is ambiguously defined as the floe thickness, and is less widely available than “sivol”. For CanESM5 and the EC-Earth3 models, “sivol” is not available; we used “simass” (total mass of sea ice divided by grid-cell area) that we converted to sivol by dividing it by their ice density of 917 kgm-3 . Because of missing files, only 14 models can be used for this analysis; the other four are marked with “no thickness” in Table . We also used the daily surface zonal and meridional wind speeds “uas” and “vas”, respectively; they were not available for IPSL-CM5A2-INCA, as indicated in Table .

For investigating potential consequences of polynya activity on the ocean, we used the monthly full-depth potential temperature “thetao” and practical salinity “so” on the models' native grid. Full depth variables being very heavy, we use only the three models with the most and three models with the least polynya activity (^o next to the model name in Table ). As we will show, there is a large agreement in polynya activity across ensemble members, so we use only the first ensemble member of each model: r1i1p1f1 for ACCESS-CM2, AWI-CM-1-1-MR, CanESM5, IPSL-CM6A-LR, and NESM3, and r1i1p1f2 for MIROC-ES2L.

In line with e.g. the “real” Arctic polynya assessment of or that of polynyas in CMIP6 in Southern Ocean of , our evaluation metric is the yearly/winter mean polynya area. We define winter for polynyas as the period from 1 December to 31 March to reduce the risk of mis-detecting late freeze-up and early melt onset as polynyas. The yearly/winter mean polynya area is then the mean of the daily polynya areas, which are for each day the sum of the grid cell areas of all pixels detected as a polynya. For some diagnostics, we keep the 45-winter time series. For most of the manuscript however, we then compute the 45-winter mean of the mean polynya area. We tested using instead the historical run only (i.e. 1979 to 2014 instead of 1979 to 2024) and found no significant differences, with correlations exceeding 0.99 (not shown); we found the same result if using instead 1982–2014 as did.

Two other diagnostics were used by : the cumulative polynya area, i.e. sum of the daily values instead of the mean (described above), and the total area of each winter, i.e. the sum of the grid cell areas of all pixels detected as a polynya at least once each winter. We computed them both for all CMIP6 models (not shown) and found a near-perfect correlation (> 0.9) between our yearly/winter means and their diagnostics. In particular, the relative biases of the models did not change, i.e. the models with the largest values with one diagnostic were still the largest with another one. We therefore use the yearly mean polynya area only. We define the polynya frequency of each pixel as the number of winters where that pixel is detected as a polynya at least once.

As is common in observational studies e.g., we study the polynya behaviour of different regions of the Arctic separately. We use the same region definition as as it is optimised to reduce noise from incorrectly detected marginal ice zone and is adapted to the same time period as our study, including the emergence of polynyas in north-east Greenland. The coordinates of each region are given in Appendix Table  and shown on the first pan-Arctic map of Sect. .

Unless stated otherwise, all across-model correlations shown are significant at 95 % and conducted on the ensemble-average values. We show linear regressions only if the corresponding regression is significant. For the trend analysis, linear trends (with 95 % significance) are calculated on the individual ensemble members first and the ensemble-average trend is the value presented. As for the yearly means, changing the time period for the trend calculation to 1979–2014 or even 1982–2014 made no significant difference (not shown), and in particular did not change the models' relative biases.

For the beginning of the analysis, where we determine the pan-Arctic behaviour of the models, the sea ice area is computed as the sum of all concentration x grid cell area, using the concentration on the interpolated grid cut at 68° N. Using the same grid, the sea ice extent is computed as the sum of the grid cell areas for all grid cells where the concentration is strictly larger than 15 %. Along with the 1979–2024 mean of the whole-year means we also compute the mean seasonal cycle: for each year, we compute the difference between the yearly maximum and the yearly minimum SIA or SIE, respectively, and take the 45-year mean.

The objective of this study is to try and find a link between polynya biases and model biases across the CMIP6 ensemble, not to conduct a process study in individual models. Therefore, our priority is to find across-model relationships. Only in the absence of such relationship do we resort to time series correlations at the individual model level. Since polynya area and frequency have regularly been linked to wind speed, including its extremes, and air temperature seeand references therein, we use these two parameters in our study.

The daily wind speed is computed from the daily “uas” and “vas”. In each region, we extract the mean of the three variables, but also their 90th percentile as the most extreme value. For the two components of the wind “uas” and “vas”, we have to extract both the 10th percentile (most extreme negative values, i.e. west- or southward) and the 90th percentile (most extreme positive values, i.e. east- or northward).

We use the air temperature biases of . They use only the first ensemble member of each model (r1i1p1f^∗), only the time period 1982–2014 (historical run), and before computation regridded all the models to a common 1° grid. For across-model correlations, we therefore also use the first ensemble member only and the 1982–2014 time period for our polynya areas. Compared to our list of models, the only model missing from their study is AWI-CM-1-1-MR. All their values are biases, i.e. model value minus a common satellite-based reference; see for more information. Since it is a common offset for all models, it has no impact on our across-model correlations.

Finally for the ocean variables, we limit our investigation to a comparison of the Kara and Laptev seas. We extend the geographical limit of the Kara Sea northward to 86° N compared to the rest of the study (coordinates in Appendix Table ) to capture the potential effect of the polynya not only on the water column below, but in the deep basin offshore via potential dense water overflows and effects on the halocline; the original definition of the Laptev Sea already includes the relevant deep ocean area. We further split each sea into three regions based on their bathymetry

continental shelf: shallower than 500 m;

The bathymetry itself was computed as the depth of the last level with salinity data, since the modelled bathymetry “deptho” was not available for all models at the time of download. We compute the bottom properties as the properties T and S in that last level. Given the large range of depths considered, we use the density anomaly referred to 2000 m σ2, computed following TEOS10 as implemented in the Gibbs SeaWater toolbox . Finally, following , we extract each model's properties of the Atlantic Water core, defined as the depth of the maximum temperature deeper than 200 m.

At the pan-Arctic scale, 10 out of 18 models overestimate the winter polynya area for all their ensemble members (Fig. ). The largest overestimation is from AWI-CM1-1-MR, which is nearly 10 times too high. This is also the model with the highest horizontal resolution by far (Table ), which is consistent with both and who showed that higher resolution in the product leads to larger polynya area detected. The bias does not seem related to the ocean model: both CanESM5 and NESM3 use the same model NEMO (both version NEMO3.4), yet the polynya area in CanESM5 is twice as large as the observations, while that of NESM3 is only a third of the observed area. We will show in Sect.  that this is probably a consequence of their different sea ice models. The intermodel spread is much larger than the spread across ensemble members of each model (Fig. ). This result still holds when looking at the individual regions separately (Appendix Fig. ). The rest of the bias analysis is therefore done with the ensemble mean of each model, unless specified otherwise.

At the pan-Arctic scale, models not only overestimate the 45-year mean polynya area, they also overestimate how often polynyas open (Fig. ). Most models have more polynya activity than the observations, either because low-activity areas extend further north or south (red colours), or because the regions are more active than the observations (yellow colours). In observations, an increase in polynya activity as sea ice area/extent declines has been reported e.g.. At the multi-model level, we find the opposite relationship. We find an across-model correlation between pan-Arctic sea ice area/extent and mean winter polynya area of 0.75/0.84, i.e. the models with most sea ice also have the most polynyas. This is most likely because, with the exception of CanESM5 and the EC-Earth3 models, polynya activity in models is mostly limited to the seasonally ice covered area (between the summer extent in cyan and the winter extent in dark blue, Fig. ), whereas in observations there are polynyas in the inner ice pack (north of the cyan line). The across-model correlation between polynya activity and the model's mean seasonal cycle in SIA/SIE is 0.50/0.53: the larger the seasonally ice-covered region, the more polynyas. This suggests a link with sea ice thickness, which we will investigate in Sect. .

Again, the highest resolution model AWI-CM1-1-MR has the most activity. Note that we visually verified its daily sea ice concentration output and can confirm that these are indeed polynyas. There are however many pixels with a low polynya opening frequency out of the main, predefined regions, for most models: this is noise from the marginal ice zone in the North Atlantic, Barents Sea and on the Pacific side. As discussed both for observations by and for models by , this is unavoidable in the Arctic, even with traditional threshold + flood-fill methods, owing to the complex shape of the coastline and the MIZ itself. To reduce this noise, we conduct the rest of the bias analysis on the nine regions highlighted on Fig. a (see also Appendix Table ).

To quantify the biases in polynya frequency, we compare the percentage of each region where polynyas have opened for more than 25 out of the 45 years (not necessarily continously), in both models and observations (pale orange and pale yellow on Fig. ). In the regions Svalbard (column 1 in Table ), East Siberian (5) and Chukchi (6), a majority of models overestimate locations that have a high opening frequency (10, 12 and 12 models, respectively). The overestimation is most dramatic in the East Siberian Sea where the observational value is 0. Most models show very little activity in the North Water Polynya (NOW, column 8). This region is one of the most active polynya regions in the Arctic but is difficult to model owing to the narrow straits and complex bathymetry e.g.. It is therefore surprising that the models' behaviour in the North Water Polynya region is not related to the model's horizontal resolution. For example, the higher resolution MPI-ESM1-2-HR is lower than observations (3.9 % instead of 6.9 %, column 8 in Table ), but its low resolution counterpart MPI-ESM1-2-LR is higher than the observations (15.0 % instead of 6.9 %).

The largest overestimation of the frequent-opening locations is from AWI-CM-1-1-MR in the Beaufort Sea, with a value 48 times that of the observations. In fact, this model largely overestimates the area with frequent polynyas in all regions except Franz Josef Land, where it underestimates it by nearly half (8.0 % instead of 13.1 %). MRI-ESM2-0 is the only other model that is quite consistent in overestimating the frequent-opening areas: it does so in all regions except Franz Josef Land, Kara Sea, and the North Water polynya. It is more common for CMIP6 models to underestimate the frequent-opening areas: seven of the models do so in at least 7 out 9 regions (ACCESS-CM2, EC-Earth3-AerChem, EC-Earth3-Veg, IPSL-CM5A2-INCA, MIROC6, MPI-ESM-1-2-HAM, NESM3), and 13 in at least 6 (same + EC-Earth3-CC, EC-Earth3-Veg-LR, MIROC-ES2L, MPI-ESM1-2-LR, NorESM2-LM, NorESM2-MM). In summary, even though Fig.  suggests that models open polynyas too often, what they do is to open with a low frequency at too many locations. The modelled polynyas' spatial distribution is mostly confined to the marginal seas and rarely form in the central ice pack.

Finally, winter Arctic polynya area has been increasing in observations e.g., with in particular finding positive trends since satellite records began in all regions except Svalbard. Do the CMIP6 models reproduce these trends? In most regions, even in Svalbard, a majority of CMIP6 models exhibit a positive trend in polynya area over 1979–2024 (Fig. ). The trend is however inaccurate when compared to the observed one, but there is no consistency in whether the models over- or underestimate it (compare symbols to the horizontal black line). We find rather that there often is a relationship between the 45-year mean bias in polynya area and the bias in trend of the polynya area (Fig. ), with five regions exhibiting large significant positive correlations between the two, albeit with three different behaviours. In Franz Josef Land (region 2, R = 0.81) and the Laptev Sea (region 4, R = 0.84), most models have a low bias in both 45-year mean and trend, except for CanESM5 (d) and IPSL-CM6A-LR (j) that have a large bias in both. In the Kara (3, 0.67) and Beaufort (7, 0.77) seas, all models overestimate the trend, but the trend is greater for larger polynya areas. In the North Water polynya (8, 0.91) in contrast all models underestimate the trend, but likewise the trend is larger for larger polynya areas. Note that AWI-CM-1-1-MR is excluded, as it is a clear outlier; when it is included (see Appendix Fig. ), all regions except the MIZ-dominated Svalbard have an extremely strong correlation, exceeding 0.9 in five regions. Note that although found, in observations, an increased performance of the detection algorithm as SIA/SIE decreased, we find no across-model correlation between the trends in SIA/SIE and either the trend or mean values of the polynya area.

In summary, there is no across-model agreement on the polynya area trend, rather different clusterings in the different regions. We see no link with the model resolution in any of the regions (shading on Fig. ), but there appears to be some clustering by sea ice model. The LIM models (stars on Fig. ) in particular are often found together, most visibly by the Pacific inflow regions of the East Siberian (all of them are to the right) and Beaufort seas (all of them are to the bottom centre/left). We will investigate why that might be in Sect. , but first, since the number of models with daily sea ice is way less than that with monthly sea ice, let us investigate the impact of using monthly CMIP6 data for polynya detection.

There are nearly twice as many CMIP6 models with monthly sea ice concentration than with daily sea ice concentration (35 vs 18). It is therefore tempting to try and detect polynyas in CMIP6 models using monthly data, although polynyas usually last but a few days . We can immediately see that both areas and frequency are largely underestimated when using monthly output (Fig. ), including for the monthly-version of the observations where only a handful of pixels have a polynya 5 to 10 out of the 45 years, but none more often than that (i.e. only red pixels on Fig. a). The same holds for the CMIP6 models, with most models having very few polynya pixels at the pan-Arctic scale, and nearly no model having frequencies exceeding 15 years (i.e. another colour than red) anywhere.

The relationship with the sea ice area/extent and its seasonal cycle have also been lost, probably because of the extreme behaviours in the monthly-only models. Without the need for any calculation, we can flag several models as having clearly inaccurate sea ice concentration, models that are therefore not trustworthy in terms of their polynya representation. For example, CIESM and the two CMCC models have virtually no summer sea ice on average over 1979–2024 (cyan contours around the Canadian archipelago on Fig. ). CAMS-CSM1-0, CAS-ESM2-0, and the CNRM models in contrast have an immense winter ice cover (dark blue contours), including the entire Greenland Sea all the way to Iceland. CAS-ESM2-0 also has one of the most extreme seasonal cycles (distance between the cyan and dark blue lines), while CESM2 has virtually none. Should the reader consider using monthly output for a polynya analysis, we strongly recommend that they reject these models with unrealistic sea ice cover.

AWI-CM-1-1-MR and IPSL-CM5A2-INCA are not available as monthly output, and EC-Earth3-Veg had no ensemble member in common for the daily and monthly output at the time of download. We therefore momentarily limit our study to the 15 remaining models and compare the mean polynya area obtained when using their daily or monthly output. Note that we used only the ensemble members common to both time resolutions, so the polynya area values presented on Fig.  differ slightly from those on Fig.  and in Appendix Table .

We find a very strong significant across-model correlation in all regions between polynya area from monthly and that from daily values (Fig. ). In four regions, as expected, the monthly value is much smaller than daily: half as small in the Beaufort Sea (region 7), 0.4 times in the North Water polynya (region 8), 5 times smaller in the Laptev Sea (region 4), and up to 10 times smaller in the North East Water polynya (region 9). In four more regions, the regression shows a nearly one-to-one relationship between the monthly and daily results (0.82 in Svalbard, 0.93 in Franz Josef Land, 0.72 in the Kara Sea, and 0.86 in the East Siberian Sea). Because detection in these regions is impacted by noise from the MIZ, this agreement suggests that monthly and daily results are similarly impacted. In the Chukchi Sea (region 6), the monthly results are larger than the daily results for 6 of the models, which toggles the regression slope above 1. That is the only such region. Visual inspection (not shown) reveals that the daily sea ice there is very “binary” in these models: small areas are clearly open (concentration of 15 % or less) next to a very sharp transition to the consolidated ice pack (concentration above 80 %). So even though this transitional zone varies a lot from day to day, this results in a large region with sea ice concentration around 40 % as monthly mean, which the algorithm detects as a large polynya. The same happens to MRI-ESM2-0 (model p) in the Beaufort Gyre (not shown): It experiences a very late freeze up in December 2015 and 2017, during which the daily ice edge moves a lot from daily to daily images. The daily polynya area is not affected, since these regions are filtered out as open ocean for the daily results, but the resulting monthly mean sea ice concentration is low over a large region and erroneously interpreted as many polynya pixels for the monthly result.

Note that standard threshold methods would be affected by these behaviours as well, unless the authors select a high sea ice concentration as threshold, which then further reduces their ability to remove the marginal ice zone. This could be fixed in our method by fine-tuning the thresholds and settings of either Unet or the GMM per model and/or region, but we want to keep the same values for all models, at the pan-Arctic scale. One could also use the daily sea ice concentration to generate the MIZ filter and apply it to the monthly classification result, although the computing time saved would be unsignificant – if daily output are available, we recommend their consistent use. A dedicated study of these specific models in the Chukchi and Beaufort seas is beyond the scope of this paper. Note also that even though the observations are not included in these correlation and regression calculations, they fall close to the regression line but have systematically lower results from monthly than daily values, as is to be expected.

In summary, one can mix and match monthly and daily results if one is careful, since the relationship between the two differs depending on the Arctic region. We recommend against doing that and would rather encourage the widespread release of daily sea ice data, as polynyas are short-lived events that open and close on daily to sub-daily time scales. For that reason, in the next section, we use only the daily outputs. We would also like to re-iterate that sea ice concentration alone is not optimum for detecting polynyas in the Arctic and would welcome a widespread release of daily sea ice thickness output.

One expects a negative correlation between sea ice thickness and polynya area in winter, since sea ice thickness is commonly used to detect polynyas e.g.,. For each model and each region, we correlate the time series of their region-average, season-average sea ice thickness with the time series of the same region's DJFM polynya area. We indeed find negative correlations between winter thickness and winter polynya area in most regions for most models (Appendix Table ): the thinner the winter sea ice, the larger the polynyas that same winter. The occasional positive correlation in the Svalbard region in our opinion is an indirect one, which rather indicates that by definition where there is sea ice (an exception in that region, the norm in all the others), polynyas can open. We also find many significant negative correlations with the previous summer (JJA) and autumn (SON) ice thicknesses. However, since there is a strong autocorrelation between different seasonal sea ice thicknesses (around 0.5 in all regions and models, not shown), and bearing in mind the link between sea ice seasonal cycle and polynya activity of Fig. , these many correlations suggest rather that models that tend to have thin ice in general have large polynya activity.

We do find an across-model relationship between 45-year mean sea ice thickness and 45-year mean polynya activity, but not everywhere nor in all regions (Table ). In the Kara, Laptev and Beaufort seas and the North Water polynya, the correlation is strongest with the summer sea ice thickness; in the Chukchi sea, it is strongest with the winter thickness; in Franz Josef Land, it is nearly equally strong with both seasons. Since few models had correlations between their time series of autumn thickness and polynya area, it is not surprising that we find across-model correlations between their 45-year means only in the Kara and NOW regions, and these are of lesser magnitude than for the other seasons (-0.66 in Kara, compared to -0.91 with summer and -0.77 in winter). Many of the correlations depend on whether the AWI-CM-1-1-MR model is included, as exemplified on Appendix Fig. : in the Kara Sea where it has an average thickness, including its very high polynya area breaks the correlations; in the Laptev Sea where it has a low thickness, it makes them.

In observations, identified a regime shift in sea ice when the sea ice draft passes below the 1.3 m threshold, which corresponds to sea ice becoming thinner than approximately 1.5 m. This 1.5 m thickness threshold has also been seen in models: argued that it is because it matches the sea ice concentration threshold at which the free drift parameterisation is switched on or off (approximately 80 %), while in some sea ice models at 1.5 m thickness the sea ice albedo parameterisation changes . Looking only at the regions with a significant across-model correlation between thickness and polynya activity, we do see a suggestion of larger polynya activity for models with mean thickness below 1.5 m (Appendix Fig. , 1.5 m is the black horizontal line). This is however not a clean cut: in all regions, some models exhibit polynya activity even at thicknesses larger than 2 m while others still have no polynya below 1 m thickness. The thickness bias alone cannot explain the polynya activity bias.

To summarise, (1) there is no consistent across-model relationship between polynya area and sea ice thickness biases, not even in winter; and (2) there seems to be some clustering of the biases by sea ice model (Fig. ). A clue that the polynya bias is not linked to one specific, physical process, but rather caused by a deep model-dependent bias, e.g. in its treatment of thermodynamics, would be an across-model link between polynya bias and the model's equilibrium climate sensitivity (ECS). We do find a significant correlation of 0.76 between the ECS of and the pan-Arctic polynya area, with the caveat that there are only 9 models in common in our studies. That is, this link may not be robust, but it is encouraging. Let us therefore investigate more closely the impact of the sea ice component of the models. Two sea ice models are over-represented in our study: the Los Alamos sea ice model CICE and the Louvain-La-Neuve sea ice model LIM , with the latter typically simulating larger polynya areas (see Sect. ). Many parameters can be relevant for polynyas, sometimes with opposite results: found for example that, in ocean reanalyses, having several thickness categories (as CICE does) resulted in a thinner ice, but so did having a low “P∗” ice strength (as LIM does). One crucial point is that the two models have completely different thermodynamics treatments and surface processes included, with CICE having long included a scheme for melt ponds. As a result, CICE has a rather realistic sea ice growth/melt when compared to observations, whereas in LIM both growth and melt are faster and with a larger amplitude than expected . This would explain LIM's large and frequent polynyas.

The dynamics treatments also differ between LIM and CICE. Winter Arctic polynyas typically form in response to strong winds , making an accurate representation of wind forcing on sea ice crucial for simulating polynya openings. In sea ice models, it is the rheological formulation and roughness (or drag) that control how efficiently wind forcing is transmitted into sea ice motion. Ocean reanalyses with a low air–ice drag coefficient (such as LIM) have a slower sea ice drift , i.e. should be harder to open by wind divergence. There is also an ice–ocean drag coefficient; ocean reanalyses with a low ice–ocean drag coefficient (such as CICE) have a thinner sea ice , i.e. should be easier to open by wind divergence. Besides, the sea ice models use different rheologies: elastic-viscous-plastic for CICE and LIM3 but viscous-plastic for LIM2. This could impact sea ice deformation on short time scales and its sensitivity to wind stress. All these hypotheses however assume that models, as in reality, form Arctic polynyas in response to winds rather than thermodynamics forcing; we verify this assumption in the next subsection.

There is no across-model correlation, even when excluding the polynya area outlier AWI-CM-1-MR, between the 45-year mean wind characteristics and the polynya area in any region. However, most models have strong significant correlations between the time series of the polynya area and that of the wind speed and/or the wind components (Appendix Tables  to ). Correlations with the wind speed, both its mean and extreme, are widespread in the Eurasian Arctic, from Svalbard to the East Siberian sea, whereby stronger wind speeds are associated with larger polynya areas. This is in agreement with e.g. .

In the other regions, fewer models have correlations between the winter wind and polynya time series. In the Chukchi and Beaufort seas and in the North East Water polynya, the models that have such correlation tend to have a negative correlation with mean U and the 10th percentile of U, and positive with mean V and the 90th percentile of V 90th or negative with the 10th percentile of V (Appendix Tables  and ), i.e. the stronger the northwestward winds, the larger the polynya area. The correlation with the northward winds is in agreement with real-world polynya observations; in particular showed that the few observed winter polynya events in NEW were entirely driven by northward (southerly) winds mechanically redistributing the thick ice. The correlation with the westward winds both for NEW and the Pacific side seems counterintuitive, but in observations found that enhanced westward winds lead to a stronger inflow of Pacific Water through Bering Strait, i.e. a larger supply of upper-ocean heat that can melt the sea ice in the Chukchi and Beaufort regions. found that some CMIP6 models do not represent any Pacific Water; this may explain why the correlation is not more widespread among our models. Similarly, westward winds would impact the models' (re)circulation of Atlantic Water north of Fram Strait, which may explain the occasional model's correlation with the NEW polynya area. Finally, as in the real world, the North Water Polynya is associated with southward winds, albeit only seen for the high horizontal resolution models (Appendix Table ). The other models most likely are too coarse to properly represent the Nares Strait geography or because the resolution of the atmosphere component directly impacts the representation of the most extreme winds , crucial for Greenland polynyas . explicitly states that a resolution of at least 9 km in the atmosphere is required; for comparison, the high resolution version of the CMIP6 models that participated in HighResMIP is of the order of 25 km .

In contrast, we find strong across-model correlations between air temperature and polynya area in all regions except the Laptev Sea and North East Water polynya (Table ). Correlations with the same winter are to be expected, since e.g. found a similar relationships in observations in the Arctic, which they explained as warm air thinning the ice, preconditioning the region for polynya opening. This is consistent with the strong across-model negative correlation between air temperature and sea ice thickness in winter, and that between thickness and polynya area discussed in the previous section (Table ). We however find hardly any across-model relationship in summer, even though summer thickness and polynya area were strongly associated, which further suggests that a generally thin summer ice is not a driver but a consequence of large polynya activity in a model.

We rather find large across-model relationships between autumn air temperature and both winter polynya area and autumn sea ice thickness throughout the Arctic (Table , middle), with mean autumn air temperature being the only one correlated with mean polynya area in Svalbard (R = 0.60), the Beaufort Sea (R = 0.42) and the North Water Polynya (R = 0.49). In all regions, the absolute value of the correlation between autumn air temperature and autumn sea ice thickness exceeds 0.7. Interestingly, this is similar to what found in observations for open water polynyas in the Southern Ocean: the subsequent opening of the polynya was detectable already in the autumn as a thin ice anomaly, which pre-conditioned the region for opening with the winds acting as final trigger.

To summarise, as in observations, we find significant relationships in each model between polynya area and wind speed or direction. The most important wind component depends on the location, but also on the horizontal resolution of the model in narrow locations such as the Canadian archipelago. We also find a large contribution of the model's overall temperature bias via its impact on sea ice thickness, with warmer models having thinner ice in autumn. Determining whether individual models are actively melting sea ice in autumn or simply delaying its growth is beyond the scope of this study. The outcome is the same: a sea ice in which it is easier to open polynyas in the subsequent winter. Across-model correlations between polynya area and sea ice thickness year-round, along with the autocorrelation of sea ice, perpetuate the polynya cycle, resulting in some models being more polynya-prone than others. We finish this manuscript with an analysis of the potential impacts of this diffence in overall polynya behaviour on the ocean.

We contrast the three models with most polynya activity (red on Figs.  and ) to the three with least activity (blue). We limit our analysis to the Kara and Laptev seas, two regions that are regularly cited in observational and modelling studies as being highly productive in terms of sea ice formation and brine rejection, which impact dense water formation and deep ocean ventilation e.g.. In the Kara Sea, the mean temperature at the surface of the shelf break is colder in the models with least activity: all three are below -1 °C (blue on Fig. ), whereas the surface of the most active models is between -0.8 to +1.2 °C. That they are not at freezing temperature, whether because of warm water upwelling or warm air, suggests that brine rejection will be hindered. This surface temperature difference is also consistent with our earlier finding of thinner ice and stronger seasonal cycle being associated with higher polynya activity, although whether the upper ocean is systematically warmer because sea ice melts in summer, or sea ice melts in summer because the ocean is warmer, we cannot tell from these data. It may also be a coincidence in this subsample of models, since the other depth levels, properties and regions show no consistent behaviour. In fact, considering the temperature on the shelf, the lines of the least (blue) and most active (red) models cross already at 40 m depth. And since even at the surface, the salinity biases are inconsistent between these models on the continental shelf, so is their density (Fig. , left). In particular, the models with most polynya activity do not have denser water at the bottom of the shelf than the models with least activity: of the six models, the densest is MIROC-ES2L, among the least active models (36.88 kgm-3, Appendix Table ). Therefore, it comes as no surprise that there is no consistency at the shelf break, in any of the properties, nor in the deep basin. Although it seems at first glance that the Atlantic Water core is deeper off-shelf in the least active models (blue in Fig. , right), we find that the core depth varies from 358 to 876 m in the most active models, and 667 to 917 m in the least active, i.e. their ranges overlap (Appendix Table ).

CMIP6 models have 10 times smaller polynya areas in the Laptev Sea than in the Kara Sea, and they are more sporadic (see Sect. ). These differences appear to have no impact on the ocean profiles (Fig. ). We find again that the only consistent clustering is in the temperature at the very surface of the continental shelf, with the least active models being colder than the most active ones (blue, below -1 °C; red, -0.8 to 0 °C). It is worth noting again that in the most active models, these temperatures are above freezing, i.e. not compatible with sea ice production and brine rejection. The bias order is not maintained with depth and there is no consistent bias in salinity, so the density at the bottom of the shelf does not significantly differ between the three most polynya-active and three least active models; again, the densest model is in fact MIROC-ES2L, among the least active models (35.97 kgm-3, Appendix Table ). We find no consistency in the Atlantic Water core or bottom properties, either at the shelf break or in the deep basin (Fig. ).

In particular in the Laptev Sea, where polynyas are simulated less frequently, it may be that taking the 45-year average masks the effect of polynyas on the water column. We therefore correlate the time series of polynya area and ocean properties, considering both the yearly-mean properties and the different seasons to account for the time the dense water may need to travel from the shelf to the deep basin. In observations, this takes 3 to 6 months . In the two active models CanESM5 and IPSL-CM6A-LR, this takes up to one year, with maximum correlations found between polynya activity and properties at the bottom of the deep ocean in both Kara and Laptev in the subsequent autumn (Table ). In AWI-CM-1-1-MR, correlations are rather with the yearly value or that of the same winter; it is not impossible that in that very high resolution model, jets of dense water reach the deep basin shortly after their formation. The active models tend to agree on the sign of the correlations, with more polynya activity being associated with warmer and fresher deep basins. AWI-CM-1-1-MR and CanESM5 also agree that more polynya activity is associated with warmer and less-dense waters at the bottom of the shelf and at the bottom of the shelf break, in both seas. This is unexpected since polynya activity is supposed to generate dense water . Even when polynyas have been observed to result in warmer deep waters, it was in conjunction with high salinities . Models do not systematically have a correlation between polynya activity and the bottom salinity there, but when they do, this correlation is negative, i.e. more polynya activity is surprisingly associated with fresher waters (Table , bottom S). This would be consistent with polynyas opening by melting the ice, i.e. sensible heat polynyas, rather opening dynamically in response to strong winds, i.e. latent heat polynyas . Note that the third of the most active polynya models, IPSL-CM6A-LR, has no such correlation between polynya activity and bottom properties (Table ), even though it uses the same ocean and sea ice model as CanESM5, albeit a newer version of it. In the “real” Arctic, there are no purely sensible heat polynyas, but there are hybrid ones that switch modes while open from latent heat to sensible heat; after the switch, dense water production has been observed to stop . Models simulating sensible heat polynyas would be consistent with the lack of connection between polynya activity and bottom properties.

Finally, both most and least active models have negative correlations at the Kara Sea shelf break between the Atlantic Water core depth and polynya activity, i.e. the more polynya activity the shallower the core (Table , AW core depth). In the real Eurasian Arctic, the dense water cascade entrains Atlantic Water as it flows off the shelf there, which results in a cooling and deepening of the Atlantic Water core e.g.. found that CMIP6 models systematically had too deep an Atlantic Water core, often sitting below the shelf break depth, and that this bias was the result of a biased Atlantic Water inflow through Fram Strait that was not corrected by ventilation in the Arctic, since there was virtually no ventilation. It may therefore be that polynya activity momentarily corrects that advected Atlantic Water bias. However, the lack of consistency between polynya activity and both temperature and salinity, most likely a result of how models open their polynya, suggests that polynyas cannot correct the other ocean biases.