The initial version of ICON-CLM was set up by and was thereafter further developed by the Climate limited-area Modelling (CLM) community (https://www.clm-community.eu; last access: 15 July 2025) and within different projects funded by the German Federal Ministry of Education and Research. For the European domain (Fig. ) with a horizontal resolution of about 12.1 km, the setup of ICON-CLM is largely based on the setup used for numerical weather prediction (NWP) by the German Weather Service (DWD). An overview of the set of physical parameterizations used is given by . Compared to the ICON-NWP configuration, ICON-CLM is adapted for climate applications: ICON-CLM uses temporally varying sea surface temperature and sea ice concentration (siconc) fields prescribed from a coarser model or by reanalyses together with the lateral and upper boundary conditions. At the upper boundary, a nudging is applied using a nudging coefficient which is 1 at the model top of 23.5 km and <0.1 in the ninth model layer from the top (14.06 km). To account for temporally varying insolation, the CMIP6 forcing for spectral solar irradiance as provided by is read in. Furthermore, the yearly varying greenhouse-gas forcing of CMIP6 is used.

To make ICON more suitable for climate projections, the model had to be adapted for the use of the transient climatologies of aerosol and ozone. These adaptations were mainly done for the new global ICON-based Earth System Model, ICON-XPP : For ozone, the CMIP6 dataset is used. The aerosol climatologies used originate from ECHAM6 , namely the anthropogenic aerosol climatology MACv2 developed by and , but in the simple-plume version MACv2-SP,, and an extended volcanic aerosol climatology based on . However, the Kinne aerosols provide different input parameters compared to the default aerosol climatology used in ICON for NWP. While different species are available from the aerosol climatology of in the NWP version, which allows to infer properties like hygroscopicity, this distinction of different species is not available for the Kinne aerosols. In NWP, the aerosol properties are used to determine the influence of aerosols on cloud droplet numbers, which are an important quantity in cloud microphysics, with a secondary effect on radiation. Using the Kinne aerosols, the aerosol-cloud coupling was lost and had to be replaced with an alternative treatment. It was decided not to use the scaling factors available from the simple plume scheme , as these depend on assumptions, but to use a climatology of cloud droplet number concentrations (CDNCs) derived from satellite measurements. The non-transient, but monthly varying climatology used here is from and was adjusted with a climatology of ECMWF MODIS CDNC climatology based on.

Another difference to the NWP setup is the use of a soil-type distribution of the Harmonized World Soil Database v2.0 provided by the Food and Agriculture Organization of the United Nations (FAO). As recently in NWP, the urban parameterization TERRA_URB was switched on .

Note that the simulations and the raw output of ICON are done on an icosahedral, irregular grid (R13B5 in ICON notation). However, the raw output is interpolated using the nearest-neighbour approach onto the common EUR-12 grid (Fig. ) and all evaluations for the atmospheric part are, therefore, also done on this rotated longitude-latitude grid.

For EURO-CORDEX, the evaluation simulation for ICON-CLM that will be published on ESGF was generated within the project “Updating the data basis for adaptation to climate change in Germany” (UDAG). Hereafter, this simulation will be called ICON-CLM-UDAG. The latest tuning for ICON-CLM was done in a joint effort of the UDAG project and the CLM community (an article on the tuning of ICON-CLM for CORDEX-CMIP6 is under preparation). For ICON-CLM as well as for ROAM-NBS, an important tuning target was the reduction of a positive surface shortwave radiation bias over land and ocean. In the current version, a tuning of the cloud cover scheme (monthly varying allow_overcast parameter) is applied. One setting that was not modified in ICON-CLM-UDAG compared to NWP, but in our ICON-CLM and ROAM-NBS simulations, was the minimum diffusion coefficient for heat tkhmin. It was increased from 0.6 to 0.8 for the winter months December through February. The effects of this modified setting will be discussed in Sect. 3.2.1.

For ROAM-NBS as well as for UDAG and ICON-CLM, the open source release icon-2024.07 was used. However, the OASIS3-MCT interfaces necessary for the NEMO coupling are not included in the official version as YAC is the only coupler supported by the ICON consortium. Therefore, the modifications of the feature branch containing the OASIS3-MCT interfaces are made available via github.

The ocean component is based on the Geesthacht Coupled cOAstal model SysTem (GCOAST) setup, developed within the coastal model framework of HEREON . In the original GCOAST setup, NEMOv3.6 is used in the regional setup for the North West Shelf and the Baltic Sea . The main feature of the GCOAST setup is the enhanced representation of the coastal processes in the larger European shelf domain (Fig. ), together with North and Baltic Seas, allowing a horizontal resolution of 2 nm and mixed σ–z*-coordinates with 50 vertical layers. This vertical coordinate allows for the representation of the deeper and shallower regions simultaneously and consists of a predominantly σ-coordinate with a hyperbolic transient transition between the top and bottom layers following . While creating the domain file containing the smoothed bathymetry used by NEMO, a slope is determined at which the terrain-following σ-coordinate intersects the sea bed and becomes a pseudo z-coordinate. This transition leads to a smaller bottom-level index in areas with steep slopes, like the North West Shelf. A high resolution in the upper layers (<0.18 m in the top layer and <1.0 m within the upper 16 layers) leads to a good representation of the ocean's interface to the atmosphere. The physical parameterizations and their settings were also chosen as in . In the turbulence parameterization, the Generic Length Scale closure is chosen for vertical diffusion, and a geopotential Laplacian operator is used for the lateral diffusion of the tracers. The iso-level bilaplacian operator is applied within the momentum equations. The three-dimensional eddy-diffusivity is set to 0.01 m² s⁻¹ in the Atlantic and deeper layers, 0.01 m² s⁻¹ in the North Sea, and 0.3 m² s⁻¹ within the Baltic. A constant eddy-viscosity of 2.8×108 m² s⁻¹ is applied throughout the simulations.

A number of changes were made to the original GCOAST version: In the current work, the updated NEMOv4.2.0 with a new sea ice model SI3 instead of NEMOv3.6 was used. Besides the model version, an updated bathymetry from the European Marine Observation and Data Network (EMODnet) framework , which comes with a finer resolution, especially relevant for the coastal representation, has substituted the General Bathymetric Chart of the Oceans (GEBCO)-based bathymetry. Further, manual fine-tuning of the German coasts and Danish straits and a Laplacian smoothing were applied to the EMODnet data. In Fig. , the depths below 200 m are set to dark blue to accentuate the European North West Shelf area. Another modification to the original GCOAST setup was made in the radiation scheme: a three-band RGB radiation scheme instead of a two-band scheme was used. This scheme allows for a differentiated treatment of radiation in the North and Baltic Seas, enabling a better representation of highly stratified zones. The spatial variability in radiation is represented through a two-dimensional climatological field, which captures the mean chlorophyll concentration. This field serves as a prescribed parameter for the RGB radiation scheme. Since neither ROAM-NBS nor NEMO-NBS includes a biogeochemical module, the chlorophyll field is derived from literature and observational datasets: for the Baltic Sea, for the North Sea, and for the Atlantic Ocean. Due to this change, a different radiative penetration of the sea surfaces of the shallower Baltic Sea and the deeper North Sea could be achieved. The state-of-the-art Thermodynamic Equation of Seawater (TEOS-10, https://www.teos-10.org, last access: 26 June 2025) is applied within NEMO-NBS for the calculation of state variables.

The river runoff dataset was provided by the German Bundesanstalt für Gewässerkunde and combines observational data in German national waters with model results of the WaterGAP hydrological model .

In the NEMO stand-alone setup, atmospheric forcing fields of the ERA5 reanalysis in an hourly temporal resolution are used. These comprise wind velocities at 10 m height, air temperature and dew point temperature at 2 m height, mean sea level pressure, downward solar and thermal radiations. The surface turbulence and momentum fluxes are estimated using the bulk formulation as implemented in the Aereobulk package . An additional influence is set by adding atmospheric pressure as an inverse barometer sea surface height to the ocean momentum equation.

The main tidal constituents M2, N2, 2N2, S2, K2, K1, O1, Q1, P1, and M4 for both NEMO-NBS as well as the coupled ROAM-NBS are provided at the lateral boundaries using the FES2014 dataset .

Within NEMO-NBS and ROAM-NBS, only the ice thermodynamics of the SI3 sea ice model is applied, using five ice categories and two ice layers.

This setup represents one of the first implementations of NEMOv4.2.0 with SI3 in a fully coupled regional system for the European shelf, North and Baltic Sea, with direct applicability to coastal hazard and climate risk assessments.

ICON and NEMO are coupled via the OASIS3-MCT_5.0 coupler . The interfaces within the ICON code are based on the implementation described by . On the NEMO side, the OASIS3-MCT interfaces are used as provided with NEMOv4.2.0. In our setup, we are using exclusively the approach of the flux coupling , which has the advantage over the bulk coupling that both the atmosphere and the ocean model see the same turbulent fluxes. Additionally, as ICON incorporates a tile approach, the flux coupling ensures the best possible local conservation of energy on non-common grids without using an exchange grid as, for example, ; when variables are sent from the atmosphere at lower horizontal resolution to the ocean via OASIS3-MCT, tiled quantities are used. This approach is very similar to the one used in ICON-XPP . These tiled quantities are representative for the ocean and sea ice fractions, respectively, in each atmospheric model grid cell, and are used by the ocean for the computations for open ocean and sea ice. The respective quantities are solar shortwave radiation, momentum flux, and the so-called non-solar radiation, which is the sum of sensible heat flux, latent heat flux, and long-wave radiation. As the ice fraction is sent from the ocean model to the atmospheric one via OASIS3-MCT, the ice fractions of both models are consistent. In the default NWP and ICON-CLM setup, the sea ice scheme calculates the heat transfer within the sea ice, dependent on the current ice depth and a standard bottom temperature. It cannot generate new sea ice, but the ice thickness can decrease and ice albedo and surface temperature are determined depending on the heat transfer as well as on snow lying on the ice. To make the thermal part more consistent with the NEMO/SI3, which has a more detailed treatment of thermodynamical processes within sea ice, the sea ice albedo, thickness and surface temperature are transferred from the ocean to the atmosphere in our setup. The ocean albedo over water is not sent, as this option is not available in the default coupling setup of NEMOv4.2.0. Only the one over ice is available for the coupling. Over the ocean, ICON uses a formulation by for the direct albedo and the value of 0.06 for diffuse albedo as in ECMWF's IFS model. Additionally, an albedo increase by whitecap cover after is considered.

The complete list of variables exchanged during the coupling in ROAM-NBS is given in Table . For selected variables, a global conservation is applied, which means that the area average is compared before and after horizontal interpolation. By selecting “GLBPOS” in the OASIS3-MCT settings, the difference between the area average on the target minus the source grid is distributed proportionally to the value of the original field.

Because only the respective portions of fluxes for the ocean and sea ice tiles are used, flux contributions from land points are completely excluded when fluxes are sent from the atmosphere to the ocean, which also ensures a consistent exchange of energy and momentum along the coastlines. As suggested by for models that incorporate the tile approach, the ocean fraction of the atmospheric model is adapted to that of the ocean model, which is obtained by interpolating the binary land-sea mask of the ocean model to the atmospheric grid. The land and lake fractions of the atmospheric model are adapted accordingly. For a correct treatment of coastal points, great care was put on identical interpolation methods for the NEMO land-sea mask and the NEMO fields to the ICON grid during coupling, as well as on the consistency of the common land-sea mask and the masks used by OASIS3-MCT. Thus, the flux coupling as included in ROAM-NBS is a good compromise between simplicity and correctness.

The evaluation simulation was run for the coupled model (ROAM-NBS) and for both the atmospheric and ocean components individually (ICON-CLM and NEMO-NBS, respectively); i.e., we are overall evaluating and comparing three simulations. In Sect. , a basic comparison to the evaluation simulation from the UDAG project (called ICON-CLM-UDAG) is additionally shown. According to CORDEX specifications , “the evaluation experiment must cover at least the 1979–2020 period”. Our experiments were started in September 1978 because of the stable thermal mixed layer in the ocean. Starting in January would mean starting in the middle of the winter season when thermal conditions are unstable and the sea ice has already evolved, which has proven disadvantageous for model initialization, especially in coupled mode. To avoid a so-called “cold start”, a spin-up for the ocean component was carried out with NEMO-NBS for the years 1974–1978, starting with an initial field for temperature and salinity. This initial field was, in turn, a combined product from the ORAS5 reanalysis and the Baltic reanalysis product provided by SMHI (personal communication) based on . The restart field for 1 September 1978 from the spin-up simulation with NEMO-NBS was then used to start ROAM-NBS.

For a better comparison of NEMO-NBS and the ocean part of ROAM-NBS, it was decided to use the same observation-based dataset for the runoff for both experiments. ICON-CLM as well as the atmospheric part of ROAM-NBS were started from the ERA5-driven ICON-CLM-UDAG simulation, which had been running since 1940.

For the atmospheric part, ERA5 reanalyses were used as lateral boundary conditions, as prescribed by the CORDEX-CMIP6 experiment protocol . They are available at a 3-hourly resolution and are also used to prescribe SST and sea ice in the uncoupled ICON-CLM simulation. For ROAM-NBS, ERA5 SST and sea ice fields are also used in ocean regions outside of the NEMO-NBS domain. As lateral boundary conditions for the ocean, the temperature, salinity, zonal and meridional ocean velocities, and sea surface height (SSH) fields from the ORAS5 global reanalysis dataset are used. The prescribed fields were used at a monthly resolution, as daily data were not available for the whole simulation period. To ensure that the spatial mean of the modeled SSH fits the observed SSH, the boundary conditions for the SSH from ORAS5 were bias corrected with respect to the SSH at Helgoland from the GESLAv3.0 observational data . This bias was calculated on the basis of a 5-year test simulation with uncorrected boundary conditions.

The evaluation is generally conducted for the years 1979–2020, which is the minimum period required for CORDEX. However, especially for the ocean part, many reference data are available for shorter time periods only, so that the evaluation period had to be adapted in these cases: SST and sea ice data are available from September 1981 and surface salinity from 1993 . Station data are very sparse before 1993. Copernicus reanalyses used for the evaluation of marine heat waves start in 1989. For the atmospheric part, satellite data used for the evaluation of surface radiation CERES; are available from 2001 only. Therefore, the evaluated time periods had to be adapted in these cases; an overview is given in Table . For evaluations in which statistics were calculated from hourly data, shorter time periods were used, partly due to limited data availability as well, partly to reduce the computational costs (see Sects.  and ).

The most crucial variable at the interface of the atmosphere and the ocean is the SST, as it serves as the first indicator of the model's performance and proper coupling. Here, the simulated SST is compared to satellite observations from Copernicus. The Copernicus observations used are available from September 1981 to December 2024. Detailed information on the quality of the observational data is provided in the quality information document of the dataset, which can also be found on the product-specific Copernicus website. Figure a shows the difference between the simulated seasonal mean SST of ROAM-NBS and the observations. During winter (DJF) and spring (MAM), the area in the Northern Baltic is covered by sea ice with varying extent over time. In the Copernicus analyses, the SST is artificially set to -1.8 °C over the regions covered by sea ice. The points where these artificial SST values are found are masked out for the calculation of the mean differences.

ROAM-NBS shows a cold bias of locally up to 2 K in the shallower regions of the European North West Shelf and in the Biscaya region during the colder seasons (SON, DJF, MAM). In the Baltic, the bias is around zero during these seasons. In summer, a positive bias can be observed in most parts of the NBS domain apart from the Baltic, where the bias becomes negative (up to -1 K in some years in the spatial average, Fig. d). A persistent warm bias is found near the northern Atlantic boundary, east of Iceland. This temperature bias could originate from the northward heat transport by the Gulf Stream at the surface and a too weak vertical diffusion/downward transport of heat at the model's boundary. The border between the positive and negative differences in all seasons except summer seems to envelope the European North West Shelf with a stronger bathymetry gradient, which could be a hint at an underestimation of Atlantic on-shelf transport into the North Sea. In Fig. b, the difference between the seasonal mean SST of ROAM-NBS and NEMO-NBS is shown. The coupled simulation is generally slightly warmer than the uncoupled one, especially in summer (JJA). The tuning of the cloud cover scheme in ICON-CLM and ROAM-NBS reduced a positive surface shortwave radiation bias over land and ocean, so that it lies between ±10 W m⁻² (see Fig. ) in the NBS region compared to CERES data. This reduction of the radiation bias contributed to a decrease in the positive SST bias in summer. However, it was not possible to reduce it further by a tuning of the atmospheric part. Eddy diffusivity in the eastern Atlantic was parameterized to be an order of magnitude higher than in the western Atlantic, which may have contributed to the cold bias observed near the French and Portuguese coasts.

Time series of the spatial mean seasonal SST biases against the Copernicus observations are shown for different regions in Fig. . The outlines of these regions (Baltic Sea, North Sea, Open Atlantic, and the whole domain) are shown in Fig. . As for the bias maps, points with ice cover in the observations were masked out for the calculation of the biases. For the whole domain, the area-averaged bias is about -0.5 K for both simulations in winter (DJF, Fig. a, upper panel). In summer (JJA, Fig. a, lower panel), the bias is slightly larger for NEMO-NBS (about -0.75 K), but smaller for ROAM-NBS. However, this smaller bias for ROAM-NBS in summer is due to the higher warm bias in the Atlantic (Fig. b, lower panel), combined with a negative bias in the Baltic Sea (Fig. d). The magnitudes of the biases for the Open Atlantic region and the North Sea are similar to those for the whole domain, or even smaller. In these regions during the summer season, ROAM-NBS is warmer than NEMO-NBS by about 0.3 to 0.4 K. In the Baltic Sea region, the SST biases in both simulations fluctuate around zero in winter while reaching -0.75 to -1 K during summer. As the spatial averaging of biases may cancel out positive and negative values, the time series of the RMSE are additionally shown in Fig. . Especially in the Baltic Sea, the RMSE is, with values of about 1.5 K in summer, higher than the absolute values of the mean bias. In the Open Atlantic and the North Sea, it is comparable to or slightly higher than the mean bias, with about 0.75 K in all seasons. The present SST biases in DJF and JJA (Fig. ) as well as the RMSE (see Fig. ) in all seasons and areas remain reasonably stable during the evolving simulation. Thus, no accumulation of errors takes place. The time series for the absolute area-averaged SSTs (see Fig. ) demonstrate that the year-to-year variability as well as warming trends in the North and Baltic Sea are well reproduced in both simulations.

The Baltic Sea is characterized by its strong stratification of temperature as well as salinity, and its correct representation is needed for marine climate applications. To validate the stratification, the Baltic thalweg level 4 dataset is used for temperature and salinity. The interpolated gridded CTD (Conductivity, Temperature, Depth) dataset contains the longitudes and latitudes of the stations where measurements were taken, and these slightly vary with each observation cruise. Therefore, the exact coordinates of the bathymetry profiles along the Baltic thalweg in Figs.  and  may differ from date to date, but the model data is always extracted at the locations given in the observational data. An exemplary observational route is shown in Fig. a, and the distance along the route to the start is plotted on the x axis.

Under typical conditions, the inflow from the North Sea into the Baltic Sea through the Danish Straits is relatively moderate and exhibits a seasonal cycle. Inflow generally occurs as dense, saline water entering the Baltic and is strongest during late winter and early spring due to wind-driven and barotropic forcing . These inflows gradually ventilate the deep basins of the Baltic and maintain the salinity balance . Under normal circumstances, the inflow is steady and predictable, with variations primarily controlled by seasonal winds, sea-level differences between the North Sea and the Baltic Sea, and long-term atmospheric pressure patterns . Significant episodic events in which dense, saline water from the North Sea flows through the Danish Straits into the Baltic Sea, ventilating its deep basins and temporarily raising deep-water salinity and oxygen levels are called MBI events.

To evaluate the model simulation's capability to depict the inflow of saline water into the Baltic, thalweg profiles before the MBI 2014 (November 2014) and after the MBI event (February and March 2015) are displayed for temperature (Fig. ) and salinity (Fig. ).

As can be observed in Fig. , the temperature thalweg of the coupled ROAM-NBS and ocean stand-alone NEMO-NBS simulation only subtly differs. Before the MBI event, a temperature bias in the deep basins already existed between both model simulations and the observational data. The surface temperature obtained with NEMO-NBS in November 2014 highly coincides with the observational data for 500 km < distance < 1000 km, whereas surface temperatures obtained with ROAM-NBS are closer to observational data for 0 km < distance < 500 km. Also in the Bornholm basin (300 km < distance < 450 km), ROAM-NBS temperature results are closer to observational data in November 2014. After the MBI, temperatures > 6.0 °C can be observed in the Bornholm basin and Slupsk Furrow (450 km < distance < 550 km), and partially at the Gotland Basin bottom (∼800 km). ROAM-NBS and NEMO-NBS both capture this, with NEMO-NBS slightly better representing the bottom signal, though it shows a stronger cold bias in upper layers (>500 km).

In Fig. , the salinity along the thalweg is displayed around the MBI event 2014. In both models, the MBI-related salinity transport is visible but underestimated, particularly in deep basin penetration (Fig. ). The inflow event can partly be depicted for the Fehmarn Belt, Darss Sill, and Arkona Basin regions. The transport of salty water into the deeper basins is discontinued, resulting in an underestimation of salinity for both NEMO-NBS and ROAM-NBS. Nevertheless, the salinity is already underestimated before the MBI by approximately 2 g kg⁻¹ in both simulations. The flow from the Slupsk Furrow towards the Gotland Basin is underestimated in both model results. Overall, the temperature and salinity stratification characteristics of the Baltic are depicted well by both NEMO-NBS and ROAM-NBS for the MBI event, although a cold and fresh bias in the deep basins is present for both models.

During 9–11 January 2015, two consecutive winter storms crossed northern Europe ELON and FELIX, and caused much damage and a severe storm surge in the North Sea. Storm ELON passed on 9 January and dissolved afterwards. For 10 January 2015, 12:00 UTC, the center of storm FELIX with a minimum pressure of less than 960 hPa is simulated by ROAM-NBS directly to the west of the Norwegian coast (Fig. a). The location and minimum pressure agree well with the one reported by the DWD surface analysis . Accordingly, ROAM-NBS simulates strong westerly winds (due to surface friction, near-surface winds are rotated towards the low pressure center compared to the geostrophic wind) on 10 January with a maximum of up to 24 m s⁻¹ (about 86 km h⁻¹) to the northwest of Helgoland (Fig. b). Observed winds on Helgoland show maxima of up to 20 m s⁻¹ on 9 and 10 January and slightly weaker maxima on 8 and 11 January (Fig. c), indicating that storm conditions prevailed over several days. The near-surface wind speed is not as well matched with the observations by ROAM-NBS as by ERA5 (Fig. c), but the maximum on 10 January is well reproduced by both.

The storm events in January occur shortly after the Major Baltic inflow event in December 2014. The non-detided SSH results are compared against GESLAv3.0 observational data described in . Stations in the German Bight (Helgoland), Skagerrak (Helgeroa), and Baltic Sea (Travemünde) are chosen to discuss the model's instant behavior exemplarily. Associated time series of the sea level minus the yearly mean sea level are presented in Figs. d and . At the station Helgoland, the warning level of 2.5 m for a severe storm surge in the North Sea is exceeded on 11 January in the early morning hours. A small time shift can be observed compared to the wind speed maximum on 10 January. The maximum sea level on 11 January is slightly better represented by the coupled model than NEMO-NBS, although wind speed maxima are comparable in ROAM-NBS and ERA5. Further, the coupled model better coincides with the displayed maxima, which fits the results obtained in the scatter plot at Cuxhaven (Fig. ). The better representation of the SSH maxima in ROAM-NBS can mainly be attributed to the differences in the treatment of surface boundary conditions, especially the calculation of the wind stress by the surface momentum fluxes (see Sect. ).

In the Skagerrak, again, the maximum sea level for storm events ELON and FELIX in January 2015 is better represented by the coupled model, where some of the minimums in the time series are better represented by NEMO-NBS.

At station Travemünde, no high sea level maximum was present in January 2015. Higher amplitudes around the maximum sea level (4 January) as well as lower sea level values around the sea level minimum (3 January) can be observed for ROAM-NBS in comparison to NEMO-NBS.

Marine heatwaves (MHWs) are discrete periods of anomalously high SSTs. Following the widely used definition by , MHWs are identified as periods of at least five consecutive days during which temperatures exceed the 90th percentile of a baseline climatology. To detect MHWs in data and model output, we apply the open-source Python package developed by to SST data.

To evaluate model performance in terms of reproducing extreme events, we compare three standard MHW metrics – the annual number of events, the maximum intensity, and the total number of MHW days per year – across four datasets: the two model configurations, in-situ observations, and reanalysis data from Copernicus at two locations. The locations are chosen based on the availability of long-term (>30 years) observational data. One station is in the western Baltic Sea (station Leuchtturm Kiel, 10.27° E, 54.4° N) and one in the German Bight (station UFS Deutsche Bucht, 7.45° E, 54.17° N); see also Fig. a for their locations. All MHW metrics are computed relative to each dataset's own climatology, using the common baseline period from 1993 to 2020. Figure  in the Appendix compares the seasonal climatology and corresponding 90th percentile threshold across datasets for the station Leuchtturm Kiel. While the model shows a small cold bias in this region, the MHW detection is not affected by this, as it is performed relative to each dataset's individual climatology. Figure  presents a comparative analysis of annual MHW metrics at Leuchtturm Kiel derived from in-situ observations (black), reanalysis data (orange), and the two NEMO model configurations: NEMO-NBS (blue) and ROAM-NBS (green), spanning the period 1989–2020. Figure a shows the annual number of MHW events, Fig. b illustrates the maximum intensity of MHWs (in °C) in each year, and Fig. c displays the total number of MHW days per year. The same MHW evaluation is presented in Fig.  for the location UFS Deutsche Bucht.

Overall, all model configurations capture the inter-annual variability in MHW characteristics reasonably well at both locations. The NEMO-NBS (blue) generally aligns a little more closely with the observational data in terms of event frequency, maximum intensity, and duration, particularly in recent years. At Leuchtturm Kiel, the NEMO-NBS simulation has a slightly higher Pearson correlation coefficient r with the observed events (NEMO-NBS: 0.85, ROAM-NBS: 0.84), intensity (NEMO-NBS: 0.57, ROAM-NBS: 0.30, i.e. not significant) and MHW days (NEMO-NBS: 0.97, ROAM-NBS: 0.92). However, discrepancies are observed in certain years where model simulations either overestimate or underestimate the magnitude and extent of MHWs. For example, at Leuchtturm Kiel, in 2014 and 2018, both NEMO-NBS and ROAM-NBS tend to overestimate MHW metrics relative to observations, particularly in terms of total days and maximum intensity.

Despite some variability, the model simulations demonstrate skill in reproducing the temporal patterns and intensities of MHWs observed in the region, supporting their application for understanding past and projecting future marine heatwave conditions.

Overall, the evaluation of variability and extreme events shows that both NEMO-NBS and ROAM-NBS can generally reproduce but underestimate the Major Baltic inflow event, that they are able to represent storm surge events, and capture MHWs.