Sea-ice evaluation of NEMO-Nordic 1.0: a NEMO–LIM3.6 based ocean–sea ice model setup for the North Sea and Baltic Sea

The Baltic Sea is a seasonally ice covered marginal sea in northern Europe with intense wintertime ship traffic and a sensitive ecosystem. Understanding and modeling the evolution of the sea-ice pack is important for climate effect studies and forecasting purposes. Here we present and evaluate the sea-ice component of a new NEMO–LIM3.6 based ocean–sea ice setup for the North Sea and Baltic Sea region. The setup includes a new depth-based fast ice parametrization for the Baltic Sea. The evaluation focuses on long-term statistics, from a 45-year long hindcast, although short-term daily performance 5 is also briefly evaluated. Different sea-ice metrics such as sea-ice extent, concentration and thickness are compared to the best available observational dataset to identify model biases. Overall the model agrees well with the observations in terms of the long-term mean sea-ice extent and thickness. The variability of the annual maximum Baltic Sea ice extent is well in line with the observations but the 1961–2006 trend is underestimated. Based on the simulated ice thickness distribution we estimate the undeformed and deformed ice thickness and concentration in the Baltic Sea, which compares reasonably well 10 with observations. We conclude that the new North Sea/Baltic Sea ocean–sea ice setup is well suited for further climate studies and sea ice forecasts.


Model Description
The model framework NEMO and the integrated Louvain-la-Neuve sea ice model (LIM) provides possibilities to simulate ocean and sea ice processes on a multitude of time and space scales with applications ranging from global climate simulations to regional forecasts. Here we describe our NEMO-Nordic setup that uses the stable NEMO-LIM3 version 3.6 in a regional 5 configuration covering the North and Baltic seas. Below we briefly describe the different components of the configuration with the specific choices of parameter settings and physics options.

The Ocean Model
The main model domain of NEMO-Nordic covers the English Channel, North Sea and Baltic Sea. In the present study we use a sub-region of the main domain covering only the Baltic Sea and Kattegat to save computational time. This setup has the same 10 physical options, horizontal and vertical resolutions as the larger North Sea/Baltic Sea domain. The only difference is that the open boundary is in Kattegat instead of in the English Channel and the North Sea. The effect of omitting the Skagerrak and North Sea region is very limited for the Baltic Sea ice state as the sea-ice growth and melt is mainly driven by the surface fluxes rather than the advective signal. However, sea-ice can occasionally form along the Swedish west coast in the Skagerrak region and this is obviously not simulated in this setup. 15 NEMO-Nordic's horizontal resolution is 0.055 • in the zonal and 0.033 • in the meridional direction. This amounts to a nominal resolution of 3.7 km (2 nautical miles). Compared to the first baroclinic Rossby radius, which is 2-11 km in the Baltic Sea (Alenius et al., 2003;Osinski et al., 2010), this makes the model operating in an eddy-resolving to eddy-permitting regime.
The vertical resolution is 3 m in the upper layers, down to 60 m, and then gradually increases to 22 m at depths; the vertical discretization uses the z * formulation and the bottom topography is represented by the partial steps approach. The setup utilizes 20 a fully non-linear free surface formulation with a time splitting of barotropic and baroclinic modes to speed up simulation time.
Vertical mixing is represented by the two equation generic length scale formulation (Umlauf and Burchard, 2005). In addition, Laplacian horizontal and isopycnal mixing is used in conjunction with a bottom-boundary layer parametrization (Beckmann and Döscher, 1997).
For further details and evaluation of the NEMO-Nordic ocean model setup the reader is referred to Hordoir et al. (2015).

The Sea Ice Model
LIM3 is a dynamic-thermodynamic sea-ice model with a multi-category ice thickness distribution and multi-layer halothermodynamics (Vancoppenolle et al., 2009;Rousset et al., 2015). The ice dynamics use a modified elastic-viscous-plastic (EVP) rheology (e.g. Bouillon et al., 2009) and accounts for sea-ice deformation processes (ridging and rafting).
The present NEMO-Nordic builds upon version 3.6 of LIM3. Compared to the polar oceans, sea ice 1 in the Baltic Sea is only seasonal, generally thinner, and with a much lower brine content due to the low salinities in the Baltic Sea. Thus, some model parameters need to be adjusted to the Baltic Sea conditions. In Table 1 we show the settings of the physical sea ice parameters that were adjusted in our NEMO-Nordic setup. The settings are also compared to that of a large-scale global ocean configuration ORCA2-LIM3, a configuration that is include in the NEMO-LIM3.6 model system. Below we briefly describe 5 the rationale behind most of these settings.
In NEMO-Nordic the ice thickness distribution is discretized using 5 different categories and the thermodynamic calculations uses 2 vertical layers of ice and 1 layer of snow. When new ice forms in open water it is assumed to have a thickness of 0.01 m. In our setup we neglect all internal halodynamical processes of the sea ice and use a constant bulk salinity of 0.001 g kg −1 , which essentially means that the effect of brine pockets is neglected. This value is 10% of the river salinity used in the 10 model and was chosen for numerical stability reasons. We tested an ice salinity of 0.0 g kg −1 but that yielded an unstable model, likewise with an ice salinity higher than the river water salinity. For numerical stability reasons ice models also require some horizontal diffusion. In our configuration, with a relatively high horizontal resolution, we use a relatively low diffusivity constant of 1.0 m 2 s −1 . LIM3 has a ridging scheme that accounts for the thickness growth due to sea-ice ridging. In this scheme there is a parameter rn_hstar that adjusts the upper bound of the ridged ice thickness. Since ridges are generally thinner 15 in the Baltic Sea compared to polar oceans we lowered rn_hstar to 30.0 m from the default 100.0 m, likewise we lowered the crossover thickness for when sea ice ridged instead of raft to 0.07 m instead of the more common value 0.17 m.
In addition, we have implemented a simple fast ice parametrization since this is not included in the present version of LIM3.
Fast ice is an important feature of the Baltic Sea ice cover and usually occurs in coastal and archipelago regions where the ocean depth is shallow. This is done by simply masking out grid points where the depth is below 15 m in the dynamical components 20 of LIM3 so that the ice remains stationary. The fast ice mask is deactivated if the total ice volume in a cell is below 0.001 m 3 , which means that for extremely low concentrations there can be advection of ice in the fast ice zone. The main effect of the fast ice parametrization is that a region of more or less undeformed ice is formed close to the coasts, and that the ridges are formed outside of this region.

25
On spatial scales of O(km), the scale of most sea-ice models, the ice thickness varies considerably due to both thermodynamical growth and mechanical redistribution of ice. To account for such sub-grid scale ice thickness variations sea ice is described in terms of an ice thickness distribution g(h), following Thorndike et al. (1975). Here g(h)dh gives the areal fraction of ice with a thickness between h and h + dh. From the distribution we can calculate the ice concentration A as 30 the ice, the ice thickness changes and LIM3 accordingly remaps the ice thickness distribution to account for this.
For regional applications, where sea ice usually is thinner than in the polar regions, LIM3 has a new scheme to calculate the ice category bounds (Rousset et al., 2015). Based on an expected domain average ice thickness h the category bounds are fitted to a function (1 − h) α on the interval between 0 and 3h. In NEMO-Nordic the ice thickness distribution is discretized using 5 different categories based on a h = 0.5 m giving the lower bounds: 0.0, 0.25, 0.56, 0.95 and 1.46 m (also shown in Fig 8a).

10
To compare the simulated thicknesses with observations we use two different metrics: which is the mean ice thickness for each grid cell, the discrete counter-part of Eq. 2 (also called cell-averaged thickness); and which is a proxy for the undeformed level ice. The upper bound for the fourth category is 1.46 m which is greater than the 15 maximum thermodynamical ice growth for most conditions of the present Baltic Sea state. In addition, there is a distinct separation in the ice thickness distribution between the first four and the last category, see Fig. 8 (this is discussed more in section 3.3). We interpreted this as representing a separation between the thermodynamically and dynamically grown ice, and thus use the first four ice categories as a proxy for level ice and the fifth category as a proxy for ridged ice. We stress that this is just an approximation as rafting and smaller ridges will also be represented by the model in the lower categories. However, for 20 many applications (e. g. maritime winter traffic) it is the actual thickness rather than the underlying processes that is important.

Forcing and simulation
As atmospheric forcing we use downscaled ERA-40 reanalysis data (Uppala et al., 2006) which, compared to the original ERA-40 reanalysis, features additional regional details which considerably affect the solution of standalone ocean models of the Baltic Sea (Meier et al., 2011). Note that the ERA-40 data set only covers the period up until 2002 and afterwards we use 25 operational analysis from the ECMWF (European Centre for Medium-Range Weather Forecasts) for the downscaling. Using a different data set for the last 4 years could potential impact our results, however, our analysis show no evidence of that.

15
We use several observational data sets to evaluate the sea ice model. An extensive historical data set, named BASIS, covers the winters 1960/61 to 78/79. This data set contains the, at that time, best available information on the ice concentration, thickness as well as on the dominant ice types. BASIS is based on hand drawn sea ice charts which were provided by the local weather services for shipping. These ice charts were collected and then digitized 1981 in a joint project of the Finnish Institute of Marine Research (today FMI) and SMHI. The original data were hard to access as BASIS-ice was designed for storage on punchcards. 20 Thus, Löptien and Dietze (2014) provided an easier to access version in the free file format NetCDF via www.baltic-ocean.org (or PANGEA doi:10.1594/PANGAEA.832353).
In addition to BASIS, we use modern ice charts, called IceMaps, (which interpolate various observations on sea ice); the Swedish Ice Service of SMHI also provided weekly ice thickness measurements    1965-2006, 1971-2006, 1963-2006 and 1961-2006 respectively. The BSH data consist of high-quality satellite SST data product compiled into a monthly data set covering

Freezing degree days model
To evaluate NEMO-Nordic in the coastal fast ice zone we compare the model to the ice thickness growth based on a simple freezing degree days (FDD) model. In this kind of simple model the ice growth is purely thermodynamical and estimated 5 from the cumulative number of days when the air temperature is below the freezing temperature. In its most simple form the FDD model assumes bare ice, i.e. no snow, and uses a linear air-ice heat flux. The FDD model can be formulated as (see e.g. Leppäranta and Myrberg, 2009) where h is the ice thickness, Θ the number of FDD. Here we use, following Luomaranta et al. (2014), a = 3.0, c = 10.0 and 10 n = 0.5. Θ is calculated as where T f is the freezing temperature and T the daily mean temperature. Note that only days when T < T f in (6) is considered. This type of model only estimates the sea-ice thickness during the growth season until the maximum is reached and not the melting season when sea ice starts to thin. Furthermore, as snow is omitted the thickness estimated from the FDD model 15 represents an upper limit of the pure thermodynamical growth.

Model Evaluation
In this section we evaluate NEMO-Nordic's performance against a set of different observational data sets. The main focus is on the long-term statistics of important sea-ice parameters such as sea-ice concentration, extent and thickness. For the climate system any changes in these sea-ice parameters are crucial and are thus important to evaluate for future climate studies 20 and related studies on e.g. winter navigation and hazard. We also briefly compare single days when the ice cover reached its maximum extent, for 2 extreme winters. This is done to evaluate the model's capability to capture extremes on daily time scales which is important for forecasting purposes. However, we stress that the model data is from a hindcast simulation, which is a totally different mode of operation compared to how sea-ice forecasts are run. Before evaluating the sea-ice parameters we evaluate the simulated the SST.

Sea surface temperature bias
The SSTs reflect the air-sea interaction of heat and biases in the SSTs indicate that either the atmospheric forcing and/or the ocean dynamics could be misrepresented. Any such biases will also affect the growth of sea ice and we therefore briefly evaluate the SST biases in NEMO-Nordic in the following. Here we compare the simulated SST with a satellite derived SST product from BSH and with CTD casts at a few stations in the Baltic Sea and Kattegat. Figure 2 shows the long-term (1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006) wintertime SST biases over the model domain. We caution that in regions which usually are ice-covered these means are heavily weighted to ice-free conditions as the satellite sensor has limited capabilities to estimate under-ice SSTs. As seen in Proper. Another evident feature is that there is a persistent strong negative bias in Kattegat. These results are inline with the study by Gröger et al. (2015) who found that the dynamical downscaling of ERA-40 has a cold bias, due to a cold bias in the prescribed SST, particularly in Kattegat and the Baltic Sea, which in their forced simulation lead to too cold SSTs. The BSH data only cover the last two decades of the simulation and to further evaluate the simulated SSTs we compared NEMO-Nordic 10 with CTD data from four different stations which has long-term monitoring, two in the Baltic Sea (BY15 and BY31) and two in Kattegat (Fladen and Anholt E). Figure 3 compares January-April SSTs from NEMO-Nordic with the CTD data at the different stations. Here simulated daily averages where extracted from the same dates and periods as the individual stations.
Also here it is evident that there is a systematic bias in Kattegat while in the Baltic Sea the observations and NEMO-Nordic SSTs seem to better match. The long-term means for the Fladen, Anholt E, BY15 and BY31 are −0.7, −0.7, 0.2 and −0.1 15 • C, respectively. If we exclude all data prior to 1990 (not shown) the results are almost identical, with only BY15 and Fladen becoming roughly −0.1 colder in NEMO-Nordic. We note that the model suffers from a negative SST bias which is especially pronounced in Kattegat. Further, the CTD data confirm that the bias is persistent in Kattegat over the entire simulation. In the Baltic Sea, on the other, the two data sources disagree on the SST biases for the same period, the BSH data show a negative bias at both stations while the CTD data show a lower smaller negative bias at BY31 and a small positive bias at BY31. We 20 also note that the satellite derived data set yields a larger negative bias for all stations, compared to the CTD data, for the same periods. Whether the basin scale negative bias evident in the BSH data comparison is present in the first three decades can not be determined from these data sources. It is beyond the scope of the present study to further explore the underlying reasons behind the SST biases.

Sea-ice concentration and extent 25
The extent of the sea-ice cover and its concentration are two important parameters that a sea-ice model needs to simulate well both from a climatological and forecasting perspective. Here we show the long-term spatial coverages and the time variability of the total sea-ice extent in the Baltic Sea. The annual maximum Baltic Sea ice extent (MBI) is a widely used metric to describe climate variability in the region, and the first recordings date back to 1720 (e.g. Vihma and Haapala, 2009, and reference therein). To evaluate the MBI we compare 5 NEMO-Nordic with observational estimates from BASIS extended with IceMaps from the Swedish Ice Service. Note that in this study sea-ice extent is calculated as the area where sea-ice concentration is at least 15%; and that we have excluded sea ice in the Skagerrak region in the observational estimates as this region is missing in the present configuration. Figure 6 shows that the model captures the variability quite well. The simulated and observed standard deviations are 78 and 98 km 2 , respectively, and the correlation between the model and observations is 0.85. However, the simulated trend (-10 km 2 /decade)  In the Baltic Sea ice winters are usually classified, following Seinä and Paluso (1996) Figure 7 shows the long-term mean monthly seasonal cycles and standard deviations. Again NEMO-Nordic is well in line with the observational estimates, however, the reduction of sea ice during spring tends to occur earlier in the season in the model. This offset is especially pronounced during the first two decades (not shown).
Overall NEMO-Nordic agrees well with the BASIS data sets in terms of sea-ice concentration and extent. The offsets with a too high sea-ice extent in Kattegat and the overestimation of the total sea-ice extent the in last three decades are both inline 25 with the already discussed SST biases.

Sea-ice thickness, volume and deformation
Sea-ice thickness is another important sea-ice parameter that both reflects the thermodynamical and dynamical evolution of the ice pack. Here we evaluate the sea-ice thickness distribution as well as the long-term mean sea-ice thickness, sea-ice volume and ice concentration in the thickest category, where the latter is used as a proxy for the sea-ice ridge concentration.  To further explore the composition of the different ice categories we calculate the area-integrated ice volume per ice category, integrated over the Bothnian Bay and the entire Baltic Sea, as an average of the January-April 1961-1979 period. From Fig. 12 we can see that a large portion (25-50%) of the ice volume is found in the thickest ice category, both for the Bothnian Bay and the entire Baltic Sea, especially late in the ice season. Compared to the BASIS data set the total ice volume in NEMO-Nordic is (18-36%) higher, particularly in the Bothnian Bay. For the entire Baltic Sea region the total sea-ice volume better matches the observational estimates, here the fraction of thin ice is also greater in the model. Additionally, Fig. 12 clearly shows the redistribution of ice towards thicker categories as the ice season evolves. Figures 9-12 show that, compared to BASIS, NEMO-Nordic generally agrees well on the spatial patterns and area-averaged level ice thickness, but has a greater total ice volume. Observations of ridges are usually quite sparse and BASIS reflects the 5 level ice thickness and volume. As the simulated volumes imply that the deformed ice constitutes a considerable amount of the total volume this primarily explains the differences in total ice volumes both for the Bothnian Bay and the total domain.
Other factors that could impact the difference are biases in the snowfall rate, air temperature and/or wind forcing. As already discussed, Gröger et al. (2015) found that over the Baltic Sea region the air temperature and windstress in the dynamical downscaling of ERA-40 is generally lower and higher, respectively. This implies that these biases in the atmospheric forcing 10 parameters would lead to an overestimation of the total ice volume. If we exclude the volume in the fifth category we have an estimate of the simulated level ice volume which, on the other hand, is lower than the observed total volume. Thus, the mean ice thickness from the thickness distribution, the long-term January-April area-averaged level ice thickness and the area-integrated level ice volume all show that NEMO-Nordic has a lower ice thickness/volume compared to BASIS and EM measurements.

15
The sea ice in the coastal regions around the Baltic Sea usually consist of (land) fast ice which is immobile. These areas are important platforms for both human activity and marine wild life. Here the ice pack is dominated by thermodynamical growth driven by the changes in surface air temperature. In the initial stages the wind conditions also affect the formation and extent of the fast ice. The fast ice zone also affects the atmosphere-ocean momentum interaction as the wind stress is damped out. To evaluate the model performance we first compare simulated cell-averaged ice thickness with long-term weekly ice thickness 20 observations at two coastal stations, outside Ratan and Järnäs (see Fig. 1 for the locations). Then we proceed with comparing simulated cell-averaged ice thickness outside Luleå and Kemi with ice growth estimates based on a FDD model. To further evaluate the ice growth in the fast ice zone we compare the simulated ice thickness outside Luleå and Kemitwo important ports in the Bothnian Bay -with a FDD ice growth model (see Eq. 5). Here we use the air temperature from the downscaled ERA-40 at the coastal sites to calculate ice thickness estimates for the 1962-2005 period. We chose the same forcing as in the hindcast simulation to eliminate any uncertainties related to the forcing, selecting nearby air temperature observations would presumably yield more realistic results. As seen in Fig. 14 NEMO-Nordic agrees very well with both the mean thickness and variability during the growth season outside Luleå and Kemi. This means that NEMO-Nordic's ice growth at these locations is dominated by the thermodynamical growth. The slightly later onset of ice formation is presumably due to ocean dynamics which can transport heat from nearby regions and maintain the sea surface temperatures above freezing for 5 a few days to weeks longer than estimated by the FDD model. It is somewhat surprising to find that NEMO-Nordic matches the FDD model so well as NEMO-Nordic includes a snow cover that, at the Kemi location, varies with an annual maximum between 17-31 cm, for the investigated period. To further evaluate the simulated snow cover, the monthly mean snow cover at Kemi, calculated for the period 1979-1990 (see Fig. 15), was compared to the results reported by Saloranta (2000). Compared from the E-OBS gridded dataset (Haylock et al., 2008). Here they find a somewhat higher maximum ice thickness of 98 cm, for the same period, which they attribute to the lack of snow cover in their FDD model. 15 Here we used the cell-averaged thickness in the comparison with observations and a FDD model. Using the proxy level ice thickness (not shown) yields a slight underestimation at the Luleå, Kemi and Ratan sites and a larger underestimation at the more southern Järnäs site. For the latter two, which are just within the fast ice mask, it is evident that there is a signal of ridged ice late in the season affecting the seasonal cycles. Just as the ice breaks up and starts melting, small concentrations of relatively thick ice are usually advected into the fast ice zone, and thus strongly influence the variability. Our results also 20 demonstrate that it is not straightforward to compare simulated grid point values with point measurements, mainly due to the difference in scales. The observations represent the ice thickness for one point in space, while the cell-averaged ice thickness in a model grid cell represents the mean of the thickness distribution on the scale of ∼4 km. Another contributing factor to the difference in the seasonal cycles is that the measurements rely on a stable ice cover. When the ice starts to form or break up it is much harder to go out and measure the thickness of the ice, thus the observations at the coastal stations Ratan and Järnäs 25 underestimate the length of ice season.

Daily sea-ice extent and thickness for two extreme winters
To evaluate the model performance on shorter time scales we now briefly compare two single days from the hindcast with observational data from IceMap. We chose the day of the MBI for a mild winter (1995), and an extremely severe winter (1987), see e.g. Fig. 6. In Fig. 16 we show the extent and level ice thickness for NEMO-Nordic and IceMap, where we chose 30 the date of MBI in the IceMap data set for both sources. As seen in Fig. 16 the extent of NEMO-Nordic's ice cover agrees well with the IceMap ice cover. For the mild winter the MBI from IceMap was 64·10 3 km 2 and occurred on the 16th of February.
For NEMO-Nordic the total ice extent for the same date was 69·10 3 km 2 , but the seasonal maximum was somewhat larger (95·10 3 km 2 ) and reached already five days earlier. For the extremely severe winter the MBI from IceMap was 369·10 3 km 2 and reached on the 5th of March. Here the NEMO-Nordic total ice extent was 353·10 3 km 2 , while the maximum simulated extent of 377·10 3 km 2 was reached nine days later. We note that there is an offset for NEMO-Nordic both in the total size of the sea-ice extent and the time when it occurs for these two cases. The IceMap data is updated roughly two times per week, which could partly explain the offset. When the model is run in forecast mode data assimilation of SST and sea-ice concentration will also likely improve the performance of the model, in terms of timing of the MBI furthermore. 4 Summary and conclusions 15 We have presented the ice component of a new NEMO-LIM3.6 based configuration of the Baltic Sea. The model system is intended to be used for both climate studies and short-term forecasting in the Baltic Sea region. To adapt NEMO-Nordic to the Baltic Sea a number of parameterizations were tuned to the brackish Baltic Sea conditions. Compared to, for instance, polar regions this means that we had to tune the model to the Baltic sea ice which is only seasonal, thinner and has a much lower brine content. In addition, we implemented a simple fast ice parametrization which is based on the ocean depth. 20 In the present study we evaluated the performance of the model by comparing results from a 45-year long hindcast simulation, forced by data from a downscaled ERA-40 simulation, with several observational data sets. Most of our metrics are based on long-term changes in standard sea-ice parameters. However, we also briefly show how the model performs on a daily time scale by comparing daily means of the sea-ice state during the day of maximum extent for an extremely severe and mild winter.
Our results show that the sea-ice concentration and extent are generally very well simulated, when compared to the BASIS 25 data set over the 1961-1979, although there is a bias in Kattegat, which might be related to a cold temperature bias in the air temperature forcing and the proximity to the open boundary. The downward trend of the MBI, over the 1961-2006 period, is much lower in NEMO-Nordic compared to the observational estimate. This is partly related to the general overestimation of the ice cover in the Kattegat region, which is also evident in the estimated SST biases.
The sea-ice thickness overall also agrees well with the observational data set. For the investigated period (1961)(1962)(1963)(1964)(1965)(1966)(1967)(1968)(1969)(1970)(1971)(1972)(1973)(1974)(1975)(1976)(1977)(1978)(1979) the ice thickness relative to the observational data sets. This is somewhat contradictory to what the negative SST biases imply. Here we speculate that a too thick snow cover could be a contributing factor. Our limited snow cover analysis show that at the Kemi site the snow cover seems to be inline with the reported observations by Saloranta (2000), however, this only represents one model point and further investigations of large-scale snow cover biases are needed to rule out a too thick snow cover as a cause for the slight underestimation of the ice thickness. The ridged ice concentration is also generally in line with the observations, 5 and the fast ice parametrization yields a more realistic distribution of the ridges. We caution that the observations, which are digitized hand drawn ice charts based on ship observations and various other sources, should be interpreted with some caution, as it is difficult to accurately estimate the ice thickness and deformation using that methodology. For future evaluations more objective methods should be preferred.
For the demonstrated extremely severe and mild cases of day of MBI, the total extent and spatial distributions are well in 10 line with the observational estimate, although there is an offset in total extent and when the day occurs in the model. Here the estimated level ice thickness is generally overestimated by the model for the extremely severe case while the mild case agrees better with the observational data. However, the observations which are weekly ice charts are very patchy and only represent the large-scale features of the ice. In addition, data assimilation of SST and sea-ice concentration will presumably improve the model skill further. 15 Furthermore, based on data from a few recent years the simulated ice thickness distribution shows a more bimodal distribution compared to the observational data set. Here better fine tuning of the ice category bounds, ridging parameters and more ice categories could possibly improve our model. However, we find that with the present ice category configurations the thickest ice category can be used as a proxy for ridge ice concentration and the lower four as proxy for level ice thickness and coverage.
Finally, we have implemented a very simple fast ice parametrization, which is fixed based only on the ocean depth. For 20 climate studies and forecasting purposes a more sophisticated parametrization is needed to capture long-term and seasonal changes in the fast ice zone. Modeling of the fast ice zone has received relatively little attention but recent studies (Lemieux et al., 2015;Olason, 2016) have suggested new ways to parametrize the fast ice zone, which could be feasible for a Baltic Sea ice model.     1960 1965 1970 1975 1980 1985 1990 1995 2000 2005    Kemi snow thickness 1979-1990 mod mean mod std