Nemo-Nordic 2.0: Operational marine forecast model for the Baltic Sea

This paper describes Nemo-Nordic 2.0, an operational marine model for the Baltic Sea. The model is used for both 1 near-real-time forecasts as well as hindcast purposes. It provides estimates of sea surface height, water temperature, salinity and 2 velocity as well as sea ice concentration and thickness. The model is based on the NEMO (Nucleus for European Modelling 3 of the Ocean) circulation model and the previous Nemo-Nordic 1.0 configuration by Hordoir et al. [Geosci. Model Dev., 12, 4 363–386, 2019]. The most notable updates include the switch from NEMO version 3.6 to 4.0, updated model bathymetry and 5 revised bottom friction formulation. The model domain covers the Baltic and the North Seas with approximately 1 nautical 6 mile resolution. Vertical grid resolution has been increased from 3 to 1 m in the surface layer. In addition, the numerical solver 7 configuration has been revised to reduce artificial mixing to improve the representation of inflow events. Sea-ice is modeled 8 with the SI3 model instead of LIM3. The model is validated against sea level, water temperature, and salinity observations, as 9 well as Baltic Sea ice chart data for a two-year hindcast simulation (October 2014 to September 2016). Sea level root mean 10 square deviation (RMSD) is typically within 10 cm throughout the Baltic basin. Seasonal sea surface temperature variation 11 is well captured, although the model exhibits a negative bias of approximately -0.5°C. Salinity RMSD is typically below 1.5 12 g/kg. The model captures the 2014 Major Baltic Inflow event and its propagation to the Gotland Deep. The model assessment 13 demonstrates that Nemo-Nordic 2.0 can reproduce the hydrographic features of the Baltic Sea. 14

The model skill is assessed with respect to variables that are of interest to the users: sea surface height (SSH), water temper-57 ature and salinity, as well as sea ice coverage using observations. Observational data from tide gauges, FerryBox instruments, 58 vertical profiles, as well as ice charts is used. The model configuration and used observation data sets are presented in Section 59 2. Section 3 presents the model assessment metrics, followed by a discussion and conclusions in Section 4.

81
The model configuration was tuned to accurately simulate surface gravity waves and internal gravitational currents. Bottom 82 friction is imposed with implicit nonlinear log-layer parameterization. The bottom drag, C d , was computed from a spatially 83 variable bottom roughness length parameter, z b 0 , Standard deviation (σ m ) and correlation coefficient (R) are given by

201
CRMSD is related to σ m and R through the equation, 203 which can be visualized in a Taylor diagram (Taylor, 2001). In this work, we are normalizing the Taylor diagram by scaling 204 the variables with σ o : The exact vertical reference datum of the circulation model is not well defined. Consequently, SSH bias cannot be reliably 215 evaluated and we therefore assess SSH performance with centralized metrics, i.e. CRMSD and Taylor diagrams. (KielHoltenau, Warnemuende, Travemuende). All of these tide gauges are located inside an estuary mouth or harbor area which 227 the present model cannot resolve.

228
In general, the model reproduces SSH standard deviation well (Fig. 2). The difference is within 2 cm in most cases; the 229 largest deviation at Fredericia is approximately 7 cm. In the Baltic basin, the model has a tendency to overestimate variability.

230
In the Danish Straits, the variability is slightly underestimated at several locations although overestimation also occurs.  and Lübeck (bottom; route is shown in Fig. 1 b). In the beginning of the data set (November 2014) the ferry also visited Gothenburg.
Tide gauge SST metrics are shown in Fig. 4. In general, there is no clear pattern across the domain. The model has a 246 negative bias (typically between -0.4 and -0.9 • C) at almost all stations; the largest bias (exceeding -1.1 • C) occurs at Korsor.

247
The RMSD is below 1.9 • C in all cases. The standard deviation is typically close to the observed value.

248
The Taylor diagram (Fig. 5, left) shows that SST skill is good in general. All locations are within 0.30 NCRMSD, and have bias is less than -0.5 • C for the two ferries ( Fig. 6c and 7c). RMSD is below 0.9 • C, suggesting that SST performance is better 257 in the open sea than at the tide gauge locations. and Lübeck (bottom; route is shown in Fig. 1 b). shorter FinnMaid data set also shows the negative bias during summer 2015 (Fig. 7 c). Although the data coverage is sparse, 261 the comparison therefore suggest that the model has a negative bias during summer months (June-July) whereas the bias is 262 smaller in fall. The magnitude of these deviations is, however, typically below 1.5 • C. radiation. Based on the SST and profile metrics, the model reproduces surface layer temperature well.

281
The bottom temperature skill is poorer (Fig. 8c). RMSD in the Gotland basin is quite low (below 0.5 • C) while being higher The model reproduces salinity profiles relatively well (Fig. 8d). The deviation is small (RMSD < 0.73 g/kg) in the Gotland 287 basin and Gulf of Finland. NRMSD is large in the Gulf of Bothnia due to the fact that salinity is low in this region as a result of 288 riverine freshwater input. In contrast, in the Danish Straits and Kattegat NRMSD is small while RMSD is relatively high (up 289 to 1.9 g/kg) due to the opposite reason: In this region, salinity regularly varies by more than 10 g/kg.   The observed salinity from the Arkona Buoy is shown in Figure 9

338
During both winters, the model tended to overestimate the ice extent by roughly 25 000 km 2 . In relative numbers, the maximum 339 ice extent was overestimated by 45% and 25% for the two winters, respectively. The ice season also started earlier in the model, 340 especially in November 2014.

341
The spatial distribution of modeled sea ice is compared against FMI ice charts in Figure 12. The shown dates correspond to reproduce. The model has a tendency to form more ice near the coasts (e.g. along the coast of Sweden). The model field also 347 shows larger areas with low (<0.4) sea ice area fraction, which may contribute to the larger extent seen in Figure 11. To assess how well the model is able to reproduce sea level variations under storm conditions, we analyze the Elon and Felix 350 storms that occurred between January 7 and 12, 2015 (Fig. 13). The storms created strong westerly winds in the southern  Fig. 13b,c); as the winds calmed, sea level at Helsinki retracted. The main 355 storm event occurred on January 10, when strong westerly winds pushed water from the North Sea to Kattegat (Fig. 13d).

356
Initially, sea levels rose in the Arkona basin (event 2A in Fig. 13c) but as winds prevailed and moved to the east, sea levels 357 dropped by roughly 1.5 m in 12 h (2B). This led to a sea level increase in the Gulf of Finland and the Bothnian Sea (2B in Fig.   358 13a,b). On January 11, northerly winds pushed water to the south, causing an opposite sea level change (event 3). This was 359 followed by another weaker westerly wind event (event 4).  (Fig. 13a,b) are in good 363 agreement with the observations. In Helsinki, the amplitude tends to be slightly overestimated and there is a phase lag of 1 to 364 2 h. The tendency to underestimate SSH variability in the Danish Straits region, and overestimate it in the Baltic basin agrees 365 well with the results of long-term validation using tide gauge data (Fig. 3).  and temperature in the presented simulation does depend on the initial conditions as well.

395
The modeled sea ice coverage is in good agreement with ice charts, but the model tends to overestimate the overall sea ice 396 extent by 25 to 45%. As ice growth is strongly affected by SST, the overestimation is likely affected by the negative SST bias.

397
Our calibration runs indicate that the sea ice model parameters have only a minor effect on the sea ice extent. Further research 398 is needed to improve the biases in SST and sea ice. Basin, the MBI is relatively well simulated by the model, especially considering that the initial bottom salinity is slightly 402 low. After the inflow, however, the modelled bottom salinity decreases at a semi-constant rate, and much faster than in reality.

403
There could be several reasons for this. First, vertical profile comparison (Figs. 9 and 10) suggest too high vertical mixing.

404
It is possible that the vertical resolution is too coarse in the deeper layers (see discussion below), leading to high numerical 405 diffusion. Second, the fast decline in bottom salinity could be due to an absence of smaller inflows after the MBI, which the 406 model is not able to simulate well enough. Third, the type of vertical discretization employed in the model (z * coordinates) is 407 not so well suited to simulate dense bottom currents over rough topography, which may lead to a spurious vertical circulation (UBS for momentum advection), as well as turbulence closure parameters (lowering the Galperin limit). Thus, horizontal and 418 vertical grid resolution, numerical mixing and turbulence parameterizations all play important roles in simulating the inflow 419 dynamics. MBI dynamics will be studied in more detail in the future. Acknowledgements. This work is supported by the Copernicus Marine Environment Monitoring Service (CMEMS). Observational data was 448 provided by DMI, BSH, SHMI, MSI, LEGMA and FMI. This data was collected and made freely available by the Copernicus project and 449 programmes that contribute to it. Data analysis and visualization were carried out with Matplotlib (Hunter, 2007) and Iris (Met Office, 2010

450
-2020) Python packages. The authors wish to acknowledge CSC -IT Center for Science, Finland, for computational resources.