The atmospheric component of ROM is REMO. Its dynamic core and discretization in space and time are based on the Europe model of the German Weather Service (Majewski, 1991). The physical parameterizations are taken from the global climate model ECHAM versions 4 and 5 (Roeckner et al., 1996, 2003). The variables that exchange information between REMO and MPIOM via OASIS are 10 m wind velocity, wind stress over water, wind stress over sea ice, liquid precipitation, solid precipitation, net shortwave radiation, total heat flux over water, conductive heat flux and residual heat flux (Fig. 2a). To avoid the largely different extensions of the grid cells close to the poles, REMO uses a rotated grid with the Equator of the rotated system in the middle of the model domain. The horizontal discretization is carried out on the Arakawa C-grid and the hybrid vertical coordinates are defined according to Simmons and Burridge (1981). Our version of REMO does not include an aerosol module. The information about aerosols is based on the climatology from Tanre et al. (1984). Here, the spatial distributions of the optical thickness of land, sea, urban and desert aerosols, as well as well-mixed tropospheric and stratospheric background aerosols, are represented. More information about the parameterizations in the atmospheric component can be found in Sein et al. (2015).

The oceanic component of ROM is the MPIOM developed at the Max Planck Institute for Meteorology (Hamburg, Germany). MPIOM is a free surface, primitive equations ocean model, which uses the Boussinesq and incompressibility approximations. MPIOM is formulated on an orthogonal curvilinear Arakawa C-grid (Arakawa and Lamb, 1977) with variable spatial resolution. This grid allows for the placement of the poles over land, thus removing the numerical singularity associated with the convergence of meridians at the geographical North Pole. An additional advantage of the curvilinear grid is that a higher resolution in the region of interest can be obtained while maintaining a global domain. Using the global ocean model alleviates issues related to ocean open boundary conditions and provides an additional “degree of freedom” in the model setup and tuning, which can help increase the performance of the ocean component within the region of interest. The model parameterizations and setup are described in Sein et al. (2015).

Figure 2a shows the coupling scheme used in ROM. In the region covered by REMO, the atmosphere and the ocean interact while the rest of the global ocean is driven by energy fluxes, momentum and mass from global atmospheric data used as external forcing. In the experiments analyzed here, data from ERA-Interim reanalysis (Dee et al., 2011) and MPI-ESM-LR (low resolution) (Giorgetta et al., 2013) are used to provide lateral boundary conditions to REMO and to force MPIOM outside the coupling region.

The MPIOM grid used in this setup is represented by black lines in Fig. 2b. In the Mediterranean region, the highest horizontal resolution of MPIOM is 7 km (south of the Alboran Sea) while the lowest resolution is 25 km (eastern coasts of the Mediterranean Sea). MPIOM has 40 vertical z levels with increasing layer thickness with depth with the first layer's nominal thickness being 16 m. The spin-up of MPIOM was done according to the procedure described in Sein et al. (2015). In the stand-alone mode, MPIOM is started with climatological temperature and salinity data (Levitus et al., 1998). Subsequently, it is integrated four times through the 1958–2002 period forced by ERA-40. For the coupled runs, the model is started from the final state reached in the last stand-alone run and integrated again, forced two times by ERA-40 and one time by ERA-Interim reanalysis (1979–2012).

The REMO domain covers the North and tropical Atlantic, a large part of Africa, South America, and the Mediterranean region (red line, Fig. 2b) with a resolution of approximately 25 km on a rotated grid and a time step of 120 s. More information about the ROM-coupled system is summarized in Table 1. The HD model (global domain) computes the river discharge at 0.5∘ resolution, and an information exchange takes place every 60 min, while HD interacts with MPIOM and REMO every 24 h (Fig. 2a).

In this study, 30-year time series from three different experiments have been analyzed. The first simulation, ROM_P0, was forced by ERA-Interim for the time period 1980–2012 and was used to assess the skill of ROM in reproducing the observed regional climate over the Mediterranean Sea. In order to present an integrated vision of the impact of climate change on the Mediterranean Sea, we dynamically downscaled the MPI-ESM-LR historical simulation, covering the period 1950–2005 (for our analysis we take ROM_P1 from 1976 to 2005), and the climate change projection for 2006–2099 (for our analysis we take ROM_P2 from 2070 to 2099) under the Representative Concentration Pathway 8.5 (RCP8.5) scenario.

The driving model, MPI-ESM, has been used in different configurations for CMIP5 in a series of climate change experiments (Giorgetta et al., 2013). MPI-ESM is composed of ECHAM 6 (Stevens et al., 2013) for atmosphere and MPIOM (Jungclaus et al., 2013) for ocean, as well as JSBACH (Reick et al., 2013) for terrestrial biosphere and HAMOCC (Ilyina et al., 2013) for the ocean's biogeochemistry. The coupling of the atmosphere, ocean and land surface is made possible by the OASIS3 (Valcke, 2013) coupler. MPI-ESM-LR (low resolution) uses T63 (1.9∘) horizontal resolution and 47 hybrid sigma–pressure levels for the atmosphere and a bipolar grid with 1.5∘ horizontal resolution (near the Equator) for the ocean, while the MPI-ESM-MR (mixed resolution) version has the same horizontal resolution in the atmosphere, although it doubles the number of vertical levels in the atmosphere and decreases the horizontal grid spacing of the ocean to 0.4∘ by means of a tripolar grid (Giorgetta et al., 2013).

We have used MPI-ESM-LR to force ROM in experiments ROM_P1 and ROM_P2 because MPI-ESM-LR was used in a wider set of CMIP5 experiments and with more realizations than MPI-ESM-MR (Giorgetta et al., 2013). Both present the same horizontal resolution in the atmosphere, and, although MPI-ESM-MR has a higher vertical resolution mainly in the upper troposphere and lower stratosphere, the main differences in the simulations can be found in the middle atmosphere (Stevens et al, 2013). According to a recent benchmarking exercise of CMIP5 models (Lauer et al., 2017), their overall performance is quite similar. Jungclaus et al. (2013) provided a detailed description and evaluation of the ocean performance of MPI-ESM-LR and MPI-ESM-MR and concluded that both behave similarly in many aspects, although MPI-ESM-LR simulated the Labrador Sea and the North Atlantic more accurately at least in the mean state and its variability.

The ROM-simulated present Mediterranean climate is analyzed in terms of mean state, seasonal cycle and interannual variability of several atmospheric and oceanic variables. For the ROM atmospheric component (REMO), three representative variables were chosen: mean sea level pressure (MSLP), near-surface temperature (T2m) and total precipitation. For the ocean component (MPIOM), sea surface temperature (SST), sea surface salinity (SSS), sea surface height (SSH) and the subsurface current velocity are considered. These fields are compared to gridded data from different sources (interpolated observed data and reanalysis) to evaluate the ROM's ability to simulate the present Mediterranean climate (Table 2).

For MSLP and T2m, we compare the output of ROM with ERA-Interim reanalysis. The ERA-Interim data assimilation system uses a 2006 release of the Integrated Forecasting System (IFC) developed jointly by ECMWF and Météo-France. The spatial resolution of the dataset is approximately 80 km (T255 spectral) on 60 vertical levels from the surface up to 0.1 hPa (Dee et al., 2011); data can be freely accessed at https://www.ecmwf.int/en/research/climate-reanalysis/era-interim (last access: 13 February 2020). Total precipitation is validated against the Tropical Rainfall Measuring Mission (TRMM; Huffman et al., 2014) dataset, a joint mission between NASA and the Japan Aerospace Exploration Agency (JAXA) to study rainfall for weather and climate research.

Three datasets were used for the evaluation of the SST: ERA-Interim, EN4 and OISST. EN4 was derived by Good et al. (2013), who carried out a 1∘ monthly objective analysis from ocean temperature and salinity bathythermograph profiles (mechanical bathythermograph, MBT; expendable bathythermograph, XBT). The version EN4.1.1 used here includes the improvements of the estimation of MBT's and XBT's downward velocities developed by Gouretski and Reseghetti (2010). The NOAA's daily Optimum Interpolation Sea Surface Temperature version 2 (OISST; Reynolds et al., 2007) combines observations from different platforms (satellites, ships, buoys) on a regular global grid 1/4∘×1/4∘. The OISST dataset offers an accurate representation of the sea surface (Ferster et al., 2018) and is widely used in the evaluation of regional climate models (e.g., L'Hévéder et al., 2013; Akhtar et al., 2019; Cabos et al., 2019).

For SSS, we used the following two datasets: EN4 v.4.1.1 (Good et al., 2013) and MEDSEA_REANALYSIS_PHY_006_009 (Fratianni et al., 2015), implemented by the Copernicus Marine Environment Monitoring Service (CMEMS) with a 1/16∘ horizontal resolution in the Mediterranean.

The potential of ROM to improve the simulation of the regional Mediterranean Sea climate is assessed by comparisons with the MPI-ESM outputs (MPI-ESM-LR and MPI-ESM-MR).

Mean sea level pressure (MSLP) is a good indicator of large-scale circulation which influences near-surface temperature (T2m) and precipitation distributions. Erroneous MSLP gradients lead to an erroneous regional wind circulation and can also have a strong effect on ocean circulation (Sein et al., 2015). Figure 3a and b display the biases of modeled MSLP with respect to ERA-Interim for the boreal winter (defined as December, January and February; DJF) and summer (defined as June, July and August; JJA) in the 1980–2012 period (ROM_P0). ROM_P0 provides a good agreement with ERA-Interim MSLP, showing maximum deviations smaller than 3 hPa over most of the domain for both seasons. The strongest departures can be found in DJF due to an overestimation of the Azores high during the winter months. Those differences could be attributed partly to REMO parameterizations, but a more important role could be played by the deficiencies in the simulated ocean circulation in the North Atlantic, which result in a region of cold SST bias centered east of the Flemish Cap (not shown). Jungclaus et al. (2013) consider this cold bias appearing in MPI-ESM-LR and MPI-ESM-MR to be a persistent feature in state-of-the-art climate models, where the coarse resolution prevents a proper representation of the Gulf Stream separation (Dengg et al., 1996), although they also mention other possible causes. Nonetheless, these relatively small deviations imply a small change in terms of regional wind circulation. During summer months (Fig. 3b) MSLP biases are much smaller over the Mediterranean.

Figure 3c and d show T2m biases for DJF and JJA. For both seasons, the departures are typically below 3.0 ∘C over most of the coupled domain except for the Alps, the Pyrenees, the Atlas Mountains, the Caucasus and the Armenian highlands (Fig. 3c and d). This disagreement can be attributed to differences in the resolution of orographic features. Winter months show the largest T2m biases located close to the Mediterranean coastline.

At first glance, ROM_P0 generally underestimates the simulated cumulative precipitation over most of the Mediterranean region for the both winter and summer seasons. The largest discrepancies for DJF are located over the Black Sea, the Adriatic Sea and the Gulf of Lion (Fig. 3e), where negative anomalies can reach 3 mm d-1. Moreover, it is worth stating that during the same period the total precipitation was overestimated in regions linked to significant topographic reliefs (e.g., the Alps). Some coastal areas also showed positive anomalies that are most likely related to the transport of precipitable water, which is influenced by the simulated evaporation over the ocean (atmosphere–ocean coupling). In the very dry Mediterranean summer season, ROM_P0 shows a clear tendency to underestimate the precipitation (Fig. 3f). Over the ocean, this bias can be related to the cold SST bias common to most of the AORCM simulations of the Mediterranean climate (see Darmaraki et al., 2019). The seasonal mean precipitation is reasonably well simulated by our coupled system throughout most of the Mediterranean basin. However, the systematic errors (up to ±3.5 mm d-1) remain substantial over the region in terms of precipitation.

The impact of interactive atmosphere–ocean coupling in REMO is shown in Fig. 4, presenting the climatology differences between ROM_P0 and stand-alone REMO in the simulations forced by ERA-Interim for MSLP, T2m and precipitation. Over land the simulated fields are less influenced by the coupling and are largely dependent on the details of the atmospheric component. On the other hand, the impact of the coupling can be remote through the large-scale circulation (the signal which comes from the North Atlantic), and the land–sea contrasts account for the local effects. Therefore, we can expect the differences over land to be minimal, except for the regions where the large-scale circulation or the land–sea contrasts are significant. In addition, the ROM model uses an orographic gravity wave drag formulation that improves the representation of the circulation over mountainous regions in REMO.

The winter MSLP over the Atlantic is higher in the coupled run (Fig. 4a), causing an anomalous strong anticyclonic circulation that extends to land and the Mediterranean Sea west of the Balearic Islands. The influence of the large-scale MSLP anomaly cancels the effect of the local, warmer SST, which would create a low-pressure bias here (see Fig. 5, where the SST biases are represented). However, elsewhere over the Mediterranean Sea, where the ROM_P0 SST is colder (warmer) than ERA-Interim, a higher (lower) MSLP is simulated by ROM_P0. In summer (Fig. 4b), the differences in MSLP seem to be determined mainly by the colder SST in ROM_P0, which leads to higher MSLP in the model than in the reanalysis.

The changes in T2m induced by the coupling over the Mediterranean (Fig. 4c and d) seem to be determined mainly by the SST (see also Fig. 5) through the turbulent heat fluxes. In both seasons, the T2m differences induced by the coupling correspond very well with the SST biases with respect to ERA-Interim. However, in winter T2m also seems to be influenced by the transport of Atlantic air carried by the too strong anticyclonic circulation simulated in the Atlantic. Over land the differences in winter T2m are mainly determined by the changes induced in large-scale circulation by the interactive SST in the Atlantic, while in summer the land–sea contrasts seem to be more significant.

The differences between the SST from ERA-Interim and the simulations by ROM_P0 are also reflected in the rainfall simulated by REMO and ROM_P0 (Fig. 4e and f) as shown by the correlation (r=0.63) between winter SST and precipitation biases (including the Black Sea). In winter, the Mediterranean Sea regions where the ROM_P0 SST is warmer have higher precipitation, while colder ROM_P0 SST leads to lower precipitation. The prevalent summer cold SST bias in ROM_P0 leads to weaker precipitation throughout the Mediterranean Sea especially in the northern part.

Figure 9 shows the differences between the SSS modeled by ROM_P0 and the selected datasets averaged for DJF and JJA during the period 1980–2012. All cases show a positive bias over the western basin and Adriatic Sea and negative bias throughout the Levantine Sea and north Aegean Sea. In the northeast Adriatic Sea, by the Po Delta, the largest positive differences occur (3.0 psu), and to the north of the Aegean Sea the largest negative differences (-3.0 psu) are found. Nevertheless, the deviations do not exceed, in absolute value, 0.5 psu in a large part of the domain (Fig. 9). Deficiencies in simulated precipitation are propagated into HD model river discharge, which is reflected in the SSS. ROM-simulated total river runoff into the Mediterranean is smaller than most of the observational estimates (e.g., Struglia et al., 2004; Wang and Polcher, 2019) and lower than other AORCM estimates (see Table 4). The influence of river runoff on SSS is highlighted by the coincidence of the largest SSS biases with locations of large rivers (Po, Nile) and with the Dardanelles, whose net flow is larger than estimates (Sánchez-Gómez et al., 2011).

The ROM_P0 SSS is compared with MPI-ESM-LR and MPI-ESM-MR in Fig. 10. ROM_P0 is always saltier over the whole Mediterranean with a decreasing difference towards the southeast. In general, ROM_P0 SSS is closer to EN4 and CMEMS climatologies than any of the MPI-ESM versions due to the higher horizontal resolution of ROM_P0 in atmosphere and ocean.

To conclude with the analysis of the ocean component of ROM, the SSH was analyzed. The time-averaged SSH and horizontal current velocity at 31 m depth simulated by ROM_P0 between 1980 and 2012 are shown in Fig. 11. The 31 m depth level has been chosen to remove the high-frequency variability of the uppermost ocean while retaining a characteristic upper ocean circulation pattern. Furthermore, the choice of this depth makes our results more comparable with previous studies, such as L'Hévéder et al. (2013) and Sevault et al. (2014). It can be clearly seen that Atlantic surface waters enter through the Strait of Gibraltar to the western Mediterranean; after crossing the Alboran Sea, the Atlantic Water flows along the African coast. At the Strait of Sicily, part of the Atlantic Water deflects northward along the coast of the Tyrrhenian Sea while the rest continues flowing to the eastern basin. ROM_P0 reproduces quite clearly the well-known deep water formation sites, especially in the Gulf of Lion, the southern Adriatic Sea and the Levantine Sea (near the islands of Crete and Rhodes), identified by the presence of three cyclonic gyres. These cyclonic gyres concur with negative SSH values, which highlight the sinking of surface waters. The mean SSH closely reproduces the well-established and steady basin- and subbasin-scale circulation pattern (e.g., Bergamasco and Malanotte-Rizzoli, 2010). However, mesoscale structures of circulation, such as the Mersa Matruh and Shikmona anticyclonic gyres, escape the model's horizontal resolution in the eastern basin (ca. 25 km).

A first-order comparison of the model's SSH to the AVISO sea level anomaly (SLA) (SSALTO/DUACS, 2013) can be done by adding only the thermosteric contribution (as a constant resulting from the average over the whole basin) to the dynamic SSH of the model (Sevault et al., 2014). Figure 12 shows the yearly mean and the seasonal cycle of ROM_P0 SSH compared to altimetric data. The modeled SSH shows lower values than those observed (Fig. 12a); however, it represents the behavior of the AVISO SLA time series well. The amplitude of the mean seasonal cycle is 12 cm for the simulation and 14.5 cm for AVISO (Fig. 12b). Therefore, the model is able to reproduce a realistic interannual variability and seasonal cycle.

Finally, a mass balance was carried out to estimate the net transport of water throughout the Strait of Gibraltar and the Dardanelles in order to compare the water flux modeled by ROM with the observations. Table 4 gives the water budget of ROM_P0 averaged over the period 1980–2012. The water loss by evaporation (E) is greater than the gain by precipitation (P) and river runoff (R), generating a deficit of 0.053 Sv in the basin. However, this deficit is partially compensated for by the net water inflow through the Strait of Gibraltar (0.030 Sv) and the Dardanelles, where the inflow (0.132 Sv) exceeds the outflow (0.109 Sv). The ROM_P0 water budget (E-(P+R)) is 0.007 Sv lower compared to the RCSM4 model (Sevault et al., 2014), although a significant part of the difference is due to difference in river runoff.

Figure 13 shows the mean SST and SSS fields for the present climate (1976–2005, ROM_P1) together with the differences with respect to future projections under the RCP8.5 scenario (ROM_P2-ROM_P1). At basin scale, the SST (ROM_P1; Fig. 13a) increases from northwest to southeast over the Mediterranean Sea with the western Mediterranean colder than the eastern, especially in the Gulf of Lion and in the northern Adriatic Sea where the SST minima are located (Fig. 13a). The warmest area is found along the Levantine Sea coast. The averaged Mediterranean SST is 18.6 ∘C, and, by the end of the 21st century under the RCP8.5 scenario, it is expected to have a mean increase of 2.7 ∘C with a projected warming ranging from a maximum of 3.8 ∘C in the Aegean Sea to a minimum of 0.9 ∘C in the Alboran Sea (Fig. 13b).

To verify that the simulated warming trend remains stable and is not affected by the strong ROM SST bias, comparisons for DJF and JJA have been performed separately (see Supplement). The comparable warming is appreciable in both seasons with a larger SST in the eastern basin. The influence of the seasonal cycle is limited to the location of the minima and maxima.

As shown in Fig. 13c, the surface of the eastern Mediterranean is saltier than the western Mediterranean, reaching 39.0 psu at the Levantine Sea. The western basin presents lower salinities (< 38.3 psu) influenced by the inflow of less saline Atlantic Water through the Strait of Gibraltar (36.6 psu) along the African coasts up to the Ionian Sea. Another source of freshwater is located at the Dardanelles strait where the Black Sea outflow has salinities lower than 35 psu. The averaged Mediterranean SSS is 38.0 psu, while under the RCP8.5 projection it will experience a mean increase of 0.2 psu. The differences between the mean SSS projection and the present climate show a dipolar structure through the Mediterranean Sea (Fig. 13d). Under the RCP8.5 scenario, the western Mediterranean is expected to increase slightly in fresh water (from -0.5 to -1.0 psu), while the eastern will become saltier. It is precisely in the north of the Aegean Sea where the largest SSS increases (4.0 psu) are found.

MPI-ESM-LR and MPI-ESM-MR projections under the RCP8.5 scenario by the end of the 21st century are slightly warmer than those of ROM over most of the Mediterranean Sea. Namely, the projected mean SST increases are 2.8 and 2.9 ∘C for MPI-ESM-LR and MPI-ESM-MR, respectively (Table 6). Compared to ROM, both MPI-ESMs show a tendency to shift the largest warming to the west, more prominently in MPI-ESM-MR, with a local minimum extending over the eastern basin (Fig. 14a and b). It is also remarkable that MPI-ESM-MR identifies the maximum warming in the Adriatic Sea and the northern Aegean Sea in the Dardanelles water outlet (Fig. 14b).

The mean SSS increase projected by ROM for the 2070–2099 period compared to 1976–2005 under RCP8.5 (ROM_P2-ROM_P1) is larger than for any of the MPI-ESMs (Table 4), but the salinity change dipolar spatial pattern is roughly the same in all three projections (Figs. 13d and 14c and d).

Figure 15 shows the mean temporal evolution of temperature and salinity anomalies in the water column over the western and eastern Mediterranean throughout the 21st century according to the ROM projection for the RCP8.5 scenario. To calculate these anomalies in a given region, we first average horizontally, over the area indicated in the Fig. 15 insets, the temperature and salinity in each MPIOM level for the present time period (1976–2005) and the RCP8.5 projection period (2006–2099). The anomalies are defined as the difference between the time series for the RCP8.5 scenario (2006–2099) and the time mean for the present climate period (ROM_P1). The Mediterranean Sea shows a gradual increase in its temperature throughout the entire water column (Fig. 15a and c), which is most pronounced in surface layers. The warming accelerates in the second half of the century, with a very clear warming signal in the upper 500 m of the eastern Mediterranean. This warming signal propagates at intermediate depths (200–500 m, corresponding to the equilibrium depth of Levantine Intermediate Water (LIW); e.g., Menna and Poulain, 2010) into the western basin. By the end of the 21st century, the eastern basin is expected to experience a surface temperature increase of up to 3.8 ∘C and the western up to 3 ∘C. At 1000 m depth the water temperature will increase by 0.6 ∘C for both basins, which is a very notable warming at these depths.

The time evolution of mean salinity anomalies displays different patterns throughout the Mediterranean Sea. During the 21st century, the upper layer (0–100 m) of the western Mediterranean is projected to freshen (-0.5 psu) while the deeper layers tend to get saltier by up to 0.5 psu. However, the eastern Mediterranean will increase its salinity by up to 0.5 psu in the entire water column. It is interesting to note that both temperature and salinity increases in the western Mediterranean at intermediate depths are delayed compared to the eastern Mediterranean.