In this study, the model results of the BalticAPP and CLIMSEA scenario simulations during the historical period were evaluated by calculating the annual and seasonal mean biases during the historical period obtained with RCO-SCOBI simulations and reanalysis data (Liu et al., 2017). Liu et al. (2017) utilised the ensemble optimal interpolation (EnOI) method to integrate profiles of temperature, salinity and the concentrations of oxygen, ammonium, nitrate and phosphate determined by the Swedish environmental monitoring programme into the RCO-SCOBI model. As reanalysis data are available for the period 1971–1999, we limited our bias calculations to 1976–1999, the overlap period between the historical period of the scenario simulations and the reanalysis data. Model data of historical periods of BalticAPP and ECOSUPPORT scenario simulations were evaluated by Saraiva et al. (2019a, b) and Meier et al. (2011b, 2012c, d), respectively.

The mixed-layer depth (MLD) was calculated following de Boyer Montégut et al. (2004), using a threshold value for the difference between the near-surface water temperature at 10 m depth and the temperature at the MLD of ΔT=0.2 ∘C.

Secchi depth (SD) is a measure of water transparency and is calculated from SD = 1.7/k(PAR), where k(PAR) is the coefficient of underwater attenuation of the photosynthetically available radiation (Kratzer et al., 2003). Factors controlling k(PAR) in the RCO-SCOBI model are the concentrations of phytoplankton and detritus. In addition, salinity is used in one of the other Baltic Sea models (MOM-ERGOM) of the ECOSUPPORT scenario simulations as a proxy of the spatiotemporal dynamics of coloured dissolved organic matter or yellow substances.

First, the monthly average of SST was computed from the model output every 48 h. The linear trend was then calculated using the Theil–Sen estimator (Theil, 1950; Sen, 1968). The trend computed with this method was the median of the slopes determined by all pairs of sample points. The advantage of this computationally expensive method is that it is much less sensitive to outliers. The significance of the SST trends was evaluated from a Mann–Kendall non-parametric test with a threshold of 95 %. The SST trends were computed by season and annually. In the latter case, the annual cycle was removed before the linear trend was computed.

Following Kniebusch et al. (2019), we performed a ranking analysis to identify the atmospheric drivers others than air temperature that are most important for the monthly variability of SST in each ESM forcing of the CLIMSEA data set and in the RCP scenarios RCP4.5 and RCP8.5. The SST trend is dominated by the trend in air temperature. Thus, to eliminate the air temperature effect on SST, the difference between the SSTs and a linear regression between the SSTs and surface air temperatures (SATs) was calculated. This was followed by applying a cross-correlation analysis of the residual SSTs to determine the main factor driving the SST trend. For each grid point and variable (i.e. cloudiness, latent heat flux and u–v wind components), the explained variance was calculated and the variable explaining the most variance was identified.

During recent decades, the Baltic Sea region has warmed faster than either the global mean warming (Rutgersson et al., 2015; Kniebusch et al., 2019) or any other coastal sea (Belkin, 2009), making it prone to marine heat waves (MHWs). Indeed, short periods of abnormally high water temperatures have been documented for the Baltic Sea (Suursaar, 2020). MHWs can be defined with reference to the mean climatology (e.g. the 90th, 95th, 98th percentile temperatures) or by temperatures exceeding absolute temperature thresholds, defined with respect to the end-user application (Hobday et al., 2018). In most cases, MHWs are defined by the number of periods, their intensity, their duration and the specific purpose (Hobday et al., 2018). In this study, the focus was on the general impact of climate change and the sensitivity of ecosystem dynamics. Hence, MHWs are defined herein as periods of SST ≥ 20 ∘C lasting for at least 10 consecutive days. For comparison, we showed also MHWs defined as periods of SST exceeding the 95th percentile of the SST distribution also lasting for at least 10 consecutive days.

The climate of the Baltic Sea region varies considerably, due to maritime and continental weather regimes. For the period 1970–1999, the annual mean SST was ∼7.8 ∘C (Fig. 4). The mean seasonal cycle of the SST is pronounced. Thus, every winter, the northern Baltic Sea, including the Bothnian Bay, Bothnian Sea and the eastern Gulf of Finland, is typically covered by sea ice (not shown). Due to its large latitudinal extension, the Baltic Sea is characterised throughout the year by a distinct SST difference between the colder northern and warmer southern sub-basins (Fig. 4). In the southern Baltic Sea, there is also a pronounced west–east temperature gradient, mainly during summer and autumn, which reflects the large-scale cyclonic circulation that transports warmer, more saline southern waters along the eastern coast and colder, less saline northern waters along the western side (see Gröger et al., 2019, their Suppl. Mat. S1; Fig. 4).

On average, during the period 1976–2005, the climate in the CLIMSEA simulations is warmer than the climate according to the reanalysis data (Fig. 4). During spring and summer, the shallow coastal zone of the northern and eastern Baltic Sea is too warm. The spatially averaged biases during winter, spring, summer and autumn and in the annual mean are 0.8, 0.9, 0.8, 1.0 and 0.9 ∘C. The reason for the warm bias is likely a bias of the RCSM. If driven by the ERA-40 reanalysis data (Uppala et al., 2005), RCA4-NEMO systematically overestimates water temperatures and underestimates sea-ice cover in the Baltic Sea for the period 1976–2005 (Gröger et al., 2019, their Suppl. Mat. S1).

In the ECOSUPPORT scenario simulations, there is also a systematic warm bias of the RCAO driven by GCMs at the lateral boundaries, such that winter water temperatures are too warm and sea-ice cover is too low (Meier et al., 2011d, 2012c, d). While these biases occur in all three applied Baltic Sea models (Table 3) forced with RCSM atmospheric surface fields, in the simulations driven by regionalised reanalysis data (ERA-40), the mean biases are smaller (Eilola et al., 2011).

Figure 5 shows the seasonal MLD cycle calculated after de Boyer Montégut et al. (2004). A deeper MLD with pronounced west–east gradients characterises the open ocean. This is related to the predominant southwesterly wind regime, with the larger wind fetches and higher significant wave heights in the eastern Gotland Basin causing wave-induced vertical mixing. Furthermore, a positive sea–atmosphere temperature contrast favours higher wind speeds (“positive winter thermal feedback loop”; Gröger et al., 2015, 2021b). In spring, a weakening wind regime, which reduces heat exchange (with a shift from heat loss to heat gain), together with the increased solar irradiance leads to a thinner MLD in the southern Baltic Sea, while melting sea ice and subsequent thermal convection and wind-induced mixing maintain a MLD > 50 m in the sea's northern part. During summer, when atmosphere–ocean dynamics are weakest, a pronounced thermocline develops, and MLDs are shallowest (the “summer thermal short circuit”; Gröger et al., 2021b). During autumn, the atmosphere cools faster than the Earth's surface, and land masses cool faster than open sea areas. These increased thermal differences result in a stronger large-scale wind regime with a positive feedback on the MLD.

The ensemble model mean in CLIMSEA reproduces these dynamics and the spatial pattern relatively well. During the cold season, however, the MLD is somewhat shallower in the simulation than in the reanalysis data of Liu et al. (2017). This may be the result of air–sea coupling. Gröger et al. (2015, 2021b) demonstrated that the complex thermal air–sea feedbacks in winter are less well represented by stand-alone ocean models than by fully coupled ocean–atmosphere GCMs. This can result in SST biases and too-shallow MLDs (Gröger et al., 2015; Fig. 7a therein; Gröger et al., 2021b). However, the real reasons for the underestimated winter MLD are unknown.

In the literature, MLDs in the ECOSUPPORT scenario simulations have not been analysed.

Baltic Sea MHWs are defined herein as periods of >10 d duration during which (1) the SST is >20 ∘C and (2) the SST exceeds the 95th percentile temperature. The CLIMSEA climate model ensemble mean and the reanalysis data set generated by the same model are compared in Fig. 6 (Liu et al., 2017).

The first MHW index uses a fixed threshold that emphasises the environmental impact of the heat waves. In particular, diazotrophic nitrogen fixation becomes effective at higher temperatures. The spatial pattern of MHWs is strongly related to the simulated SST. Figure 6a shows that MHWs are mostly absent in the open sea of the Baltic Proper and further north in the Gulf of Bothnia, but they are highly abundant in shallow marginal bays such as the Gulf of Finland and Gulf of Riga as well as along the coasts. The MHWs produced by the RCO ensemble mean are generally more frequent and of longer duration than those of the reanalysis data set. Furthermore, the coastal signature of high abundance extends further offshore (Fig. 6a). For the Belt Sea and Bay of Lübeck, this leads to considerable deviations from the reanalysis data set.

The second index is based on a reference climatology, here defined as that of 1976–1999. The number of MHWs (Fig. 6c) correlates negatively with their average duration (Fig. 6d). This is somewhat more pronounced in the reanalysis data set. In general, the patterns obtained with the reanalysis data and the RCO are similar but the amplitude of spatial variance is higher in the former (Fig. 6c), perhaps as it includes small-scale regional observations. In the RCO (Fig. 6d), MHWs in the open sea are of the longest duration, with their interruption likely due to the vertical mixing induced by wind events.

Since MHWs in the Baltic Sea are predominantly a summer phenomenon, the stability of the seasonal thermocline is likely a key element in their dynamics such that processes related to vertical mixing can be considered a benchmark in their simulation by the models. Given that mixing is highly parameterised in current ocean models, the RCO reproduces the spatial patterns of the number and average duration of MHW reasonably well.

In the literature, MHWs in the ECOSUPPORT scenario simulations have not been analysed.

The annual mean sea surface salinity (SSS) distribution shows a large north–south gradient mirroring both the input of freshwater from rivers, mostly located in the northern catchment area, and saltwater inflows from the North Sea (Fig. 7). The SSS drops from about 20 g kg-1 in the Kattegat to <2 g kg-1 in the northern Bothnian Bay and eastern Gulf of Finland. For the period 1970–1999, the annual mean SSS of the Baltic Sea including the Kattegat was ∼7.3 g kg-1. Large inflows of heavy saltwater from the Kattegat occasionally ventilate the bottom water of the Baltic Sea, filling its deeper regions (Fig. 7). As tides are almost absent, mixing is limited such that the water column is characterised by a pronounced vertical gradient in salinity, and consequently also in density, between the sea surface and the bottom.

Probably due to differences in the data of the hydrological model (E-HYPE) compared to observations, SSS in the coastal zone and the Kattegat is on average lower in the CLIMSEA climate models than in the reanalysis data of Liu et al. (2017) (Fig. 7). The spatially averaged, annual mean bias is -0.4 g kg-1. Bottom salinities in the Belt Sea, Great Belt area and the Gotland Basin (especially in the northwestern part) are considerably higher and in the Bornholm Basin considerably lower in the climate models than in the reanalysis data (Fig. 7). The spatially averaged, annual mean bias is +0.3 g kg-1. Hence, vertical stratification in the Belt Sea, Great Belt area and the Gotland Basin is also larger in the climate models than in the reanalysis data, because the difference between surface and bottom salinities is a good proxy for vertical stratification.

In the ECOSUPPORT scenario simulations, SSS is overestimated in the entire Baltic Sea, in particular in its northern and eastern regions (Meier et al., 2011c, 2012c). In both, the ensemble mean bottom salinity and vertical stratification are also overestimated, while the bottom salinity in the eastern Gotland Basin is well reproduced (Meier et al., 2012c).

Due to the seasonal cycle in wind speed, with wind directions predominantly from the southwest, the sea level in the Baltic Sea varies considerably throughout the year, with the highest levels (∼40 cm), measured relative to the Kattegat, occurring during winter at the northern coasts of the Bothnian Bay and at the eastern coasts of the Gulf of Finland (Fig. 7). For the period 1976–1999, the annual mean sea level in the Nordic height system 1960 (NH60) as determined by Ekman and Mäkinen (1996) was ∼16 cm, with a horizontal north–south difference of ∼35 cm (not shown). This sea level slope was explained by the lighter brackish water in the northeastern Baltic Sea than in the Kattegat and by wind coming from the southwesterly direction, which pushes the water to the north and east (Meier et al., 2004a).

The differences in the mean sea level between the CLIMSEA climate models and the reanalysis data are small (Fig. 7), and the spatially averaged winter mean bias is only +0.6 cm. Sea levels in some parts of the coastal zone such as the western Bothnian Sea are higher in the climate models than in the reanalysis data, probably due to lower salinities. The negative sea level bias in the eastern Gotland Basin suggests an intensified, basin-wide cyclonic gyre. The seasonal cycle of the ensemble mean sea level is relatively well simulated, but with an overestimated sea level in early spring and an underestimated sea level in summer at all investigated tide gauge locations compared to observations and to a hindcast simulation driven by regionalised ERA-40 data (Fig. 8).

In the ECOSUPPORT scenario simulations, sea levels were not systematically analysed. In one of the three models (RCO-SCOBI), seasonal mean biases were comparable to the biases in the CLIMSEA scenario simulations (Meier et al., 2011a).

Since the 1950s, nutrient inputs into the Baltic Sea have increased due to population growth and intensified fertiliser use in agriculture (Gustafsson et al., 2012; Fig. 3). Nutrient inputs reached their peak in the 1980s but have steadily declined following the implementation of nutrient input abatement strategies. Nonetheless, since the 1960s, the bottom water of the Baltic Sea below the permanent halocline has been characterised by oxygen depletion and large-scale hypoxia (Fig. 9).

Consistent with the stratification biases in the deeper sub-basins of the Baltic Sea, summer bottom oxygen concentrations in the Bornholm Basin are higher and those in the Gotland Basin lower in the CLIMSEA/BalticAPP climate simulations than in the reanalysis data of Liu et al. (2017) (Fig. 9). The stronger vertical stratification, especially at the halocline depth, hampers vertical fluxes of oxygen, causing prolonged residence times and lower bottom oxygen concentrations. Spatially averaged biases during winter, spring, summer and autumn and in the annual mean are small but systematic: -0.6, -0.7, -0.7, -0.5 and -0.6 mL L-1, respectively.

In the ECOSUPPORT scenarios, the ensemble mean deep-water oxygen concentration in the eastern Gotland Basin is slightly higher (but within the range of natural variability) and that in the Gulf of Finland significantly lower than determined from observations (Meier et al., 2011b, 2012d).

Nutrient (i.e. phosphorus and nitrogen) content in the surface layer during winter is a good indicator of the intensity of the following spring bloom. Sea surface mean winter concentrations of phosphate and nitrate are highest in the coastal zone, in particular close to the mouths of the large rivers in the southern Baltic Sea that transport elevated inputs of nutrients into the sea (Fig. 9).

For the historical period of 1976–1999, winter surface phosphate concentrations according to the climate simulations are relatively close to those of the reanalysis data (Fig. 9). The concentrations differ substantially only in those coastal regions influenced by large rivers, such as those affected by discharges of the Oder, Vistula and Pärnu rivers. Spatially averaged biases are largest during summer and autumn, with an average bias in summer of +0.2 mmol P m-3.

Likewise, winter surface nitrate concentrations in the simulations are close to those in the reanalysis data but in coastal regions they differ due to differences in the inputs from large rivers (Fig. 9). This is exemplified by the Gulf of Riga and the eastern Gulf of Finland, where the large differences between them are due to inputs from the Neva River. Spatially averaged biases during winter, spring, summer, autumn and in the annual mean are rather small but systematic: -1.1, -1.3, -0.5, -0.7 and -0.9 mmol N m-3, respectively.

In the ECOSUPPORT scenario simulations, the simulated profiles of phosphate, nitrate and ammonium are within the range of observations for 1978–2007, except in the case of phosphate in the Gulf of Finland (Meier et al., 2012d). According to hindcast simulations, the biases in the coupled physical–biogeochemical models of the Baltic Sea relative to the standard deviations of observations are larger for the northern Baltic Sea than for the Baltic Proper (Eilola et al., 2011).

During the period 1976–1999, dense phytoplankton blooms were confined to the coastal zone, i.e. the area with the highest nutrient concentrations (Fig. 10). Water transparency, measured by Secchi depth, is lower in the Baltic Sea than in the open ocean (Fleming-Lehtinen and Laamanen, 2012), and for the period 1970–1999 the annual mean Secchi depth averaged for the entire Baltic Sea, including the Kattegat, was only ∼6.6 m. The Secchi depth is also much smaller in the coastal zone than in the open Baltic Sea (Fig. 10), and in the northern Baltic Sea than in the Gotland Basin, attributable to yellow substances originating from land (Fleming-Lehtinen and Laamanen, 2012).

Due to nutrient concentration biases, the annual mean surface phytoplankton concentrations of the simulations are close to those of the reanalysis data of Liu et al. (2017) but they deviate in coastal regions (Fig. 10). Spatially averaged biases during winter, spring, summer and autumn and in the annual mean are relatively small: +0.02, -0.1, -0.009, +0.06 and -0.008 mg chlorophyll (Chl) m-3, respectively. Note that the reanalysis data of Liu et al. (2017) assimilate nutrient and oxygen concentrations but not chlorophyll data.

Similar results are found for the mean biases in the simulated Secchi depths (Fig. 10). In climate simulations, Secchi depths are systematically deeper in the regions south of Gotland and at the entrance to the Gulf of Finland (northeastern Gotland Basin) than elsewhere in the Baltic Sea. Spatially averaged biases during winter, spring, summer and autumn and in the annual mean are +0.2, +0.4, +0.06, +0.1 and +0.2 m, respectively.

Compared to the Secchi depth data from HELCOM (HELCOM, 2013a; their Table 4.3) and Savchuk et al. (2006; their Table 3), the CLIMSEA climate simulations under- and overestimate the Secchi depth in the southwestern and northern Baltic Sea, respectively, while in the Gotland Basin the model results well fit the observations (Meier et al., 2019a).

In the ECOSUPPORT scenario simulations, Secchi depth was not compared with observations.

An evaluation of biogeochemical fluxes, such as primary production and nitrogen fixation, is difficult because observations are lacking. An exception is the study by Hieronymus et al. (2021), in which historical simulations with RCO-SCOBI were compared with in situ observations of nitrogen fixation. The RCO-SCOBI model includes a cyanobacteria life cycle (CLC) model (Hense and Beckmann, 2006, 2010) driven by reconstructed atmospheric and hydrological data. The authors found a satisfactory agreement, with the results mainly within the uncertainty range of the observations. However, simulated monthly mean nitrogen fixation during 1999–2008 showed a prolonged peak period in July and August, whereas according to observations the peak was mostly confined to July. It should be noted that the RCO-SCOBI version used in the scenario simulations discussed here (e.g. Saraiva et al., 2019a) does not contain a CLC model.

In Figs. 11–12 and Table 7, annual and seasonal mean SST changes between 1976–2005 and 2069–2098 in RCO-SCOBI are depicted and quantified respectively. The maximum seasonal warming signal propagates between winter and summer from the Gulf of Finland via the Bothnian Sea into the Bothnian Bay (Fig. 11). Maximum warming occurs during summer in the Bothnian Sea and Bothnian Bay. The seasonal patterns of RCP4.5 and RCP8.5 are similar although warming is greater in the latter. As SLR has almost no impact on SST changes, BalticAPP and CLIMSEA scenario simulations yield similar results (not shown). The warming level according to ECOSUPPORT is between that predicted by CLIMSEA/BalticAPP RCP4.5 and RCP8.5 because the GHG emissions of the A1B scenario, which forces the ECOSUPPORT ensemble

One of the scenario simulations of ECOSUPPORT is driven by the A2 scenario, which due to higher GHG emissions is generally warmer than the A1B scenario. However, this particular simulation of the ECHAM5/MPIOM GCM is exceptional and at the end of the 21st century the temperature is not much warmer than that obtained with the corresponding run based on the same model under the A1B scenario.

In the CLIMSEA/BalticAPP RCSM projections, the annual mean SST changes in the Baltic Sea driven by four ESMs, i.e. MPI-ESM-LR, EC-EARTH, IPSL-CM5A-MR, HadGEM2-ES, under the RCP8.5 scenario are +2.3, +3.7, +3.5 and +4.7 ∘C respectively (Gröger et al., 2019). Thus, the ensemble mean change is +3.5 ∘C. The corresponding ensemble mean change in the RCO-SCOBI scenario simulations is smaller, +2.9∘. Different MLDs, vertical stratification and sea-ice cover in the two ocean models, RCO-SCOBI and NEMO, may explain the different responses. Indeed, a comparison of the MLD between the two models reveals a shallower MLD in the RCSM than in RCO-SCOBI (not shown), which argues for a higher sensitivity of the RCSM to climate warming.

While the spatial patterns of the SST changes in the scenario simulations of ECOSUPPORT (e.g. Meier et al., 2012c) and CLIMSEA (e.g. Saraiva et al., 2019b) are similar, the uncertainties due to the applied global (Meier et al., 2011a) or regional (Meier et al., 2012b) model are in some cases considerable. Of note is the summer ensemble range of the various GCMs (Meier et al., 2011a). The strong dependence on forcing is seen by comparing the different warming levels in the RCP4.5 and RCP8.5 scenarios shown in Fig. 12.

Since SLR and nutrient input scenarios have a negligible impact on SST changes, only the RCP4.5 and RCP8.5 scenarios in CLIMSEA/BalticAPP are compared. The multi-model mean of the annual mean SST trends averaged over the Baltic Sea is ∼0.18 and ∼0.35 ∘C per decade in the RCP4.5 and RCP8.5 scenarios respectively (Fig. 13f and k). At the Baltic Sea scale, seasonal SST trends based on annual values vary only slightly (±0.01 ∘ C per decade in both scenarios). However, at the sub-basin scale, seasonal variations are much stronger, reaching ±0.05 ∘C per decade in the northern Baltic Sea, with a maximum in summer (Fig. 13). This summer maximum can be explained by the projected decline in sea-ice cover in summer, as occurred during the period 1850–2008 (Kniebusch et al., 2019).

As seen in Fig. 14, the relative SST trends indicate faster warming of the northern than the southern Baltic Sea (0.02 and 0.04 ∘C per decade in the RCP4.5 and RCP8.5 scenarios respectively), with the largest trends, calculated over the entire period 2006–2099, reaching ∼0.24 and ∼0.45 ∘C per decade in RCP4.5 and RCP8.5, respectively. However, a calculation of the SST trends by 30-year slice periods every 10 years over the entire period shows that annual SST trends are variable over time (not shown). The natural variability appears to modulate these trends, with successive periods of increasing and decreasing SST trends over a period of about 30 years. For example, in the RCP8.5 scenario, SST trends gradually increase over the first 50 years of the period, reaching a maximum of 0.5 ∘C per decade between 2046 and 2075, before declining slightly from 2060 onwards. As in the RCP4.5 scenario, this is a result of the pronounced natural variability in this scenario. Despite the robustness of the spatial pattern of the SST trends (p<0.05 everywhere), an analysis of SST trends for the four ESM forcings reveals an important dependency of those trends on atmospheric forcings, with a spread of ±0.06 ∘C per decade from the multi-model mean of both scenarios (not shown).

At an annual timescale, the variability in the air temperature, through the sensible heat fluxes, is the main driver of the Baltic Sea's SST (Kniebusch et al., 2019), illustrated here by the high variance of SST explained by air temperature (between 0.85 and 0.95, Fig. 15). The minimum of variance explained is located in the Bothnian Bay, where the sea-ice cover isolates seawater from the air in winter.

The processes responsible for the SST trends were analysed using a rank analysis of atmospheric variables (i.e. latent heat fluxes, cloud cover and u-v wind components) following Kniebusch et al. (2019; Fig. 16). The second parameter (after SAT) explaining the variability in the SST differs according to the location and ESM. Nevertheless, in all ESMs and in both RCP scenarios, zonal and meridional wind components are the variables that best correlate with SST along most of the coastal areas, probably because of upwelling. In the open sea of the Baltic Proper and in the Bothnian Bay, the second most important variable is cloudiness. This is also the case in the Bothnian Sea under the RCP4.5 scenario. However, in RCP8.5, the second most important variable at this location is the latent heat flux. The difference is perhaps due to the absence of sea ice and therefore the amplified air–sea exchange, under RCP8.5.

In the vertical, temperature trends are larger in the surface layer than in the winter water of the Baltic Sea above the halocline, thus causing a more intense seasonal thermocline (see Sect. 3.2.2). Surface layer trends are largest in spring and summer (not shown). Elevated trends also characterise deep water, due to the influence of saltwater inflows that will be warmer in a future climate because they originate from the shallow entrance area and occur mainly in winter. Hence, in sub-basins that are sporadically ventilated by lateral saltwater inflows, such as the Bornholm Basin and the Gotland Basin, the deep water below the halocline will warm more than the overlaying intermediate layer water.

In the literature, trends in ECOSUPPORT scenario simulations have not been analysed.

Figure 17 shows the changes in the MLD. During winter, reduced sea-ice cover in the Bothnian Sea and Bothnian Bay favours a widespread deepening of their MLDs, likely caused by wind-induced mixing. In spring, the most pronounced feature is a strong shallowing of the MLD in the Bothnian Sea, probably attributable to the radiative fluxes that warm the surface layer and to less thermal convection (Hordoir and Meier, 2012). During the historical period, water temperatures in this area were between 2.0 and 3.0 ∘C (Fig. 4). Thus, in the future, surface water warming between 1.6 and 2.4 ∘C (Fig. 11) may hamper thermal convective mixing.

The changes during summer are less pronounced. In contrast to winter, there is an overall shallowing in the entire Baltic Sea. This is in agreement with a shallower, more intense thermocline in warming scenarios, as suggested by Gröger et al. (2019), and it is a common feature among the projections, because the changes in wind speed are small (Christensen et al., 2021). Autumn is primarily characterised by a prolongation of the thermal stratification, leading to an overall shallower MLD than during the historical period.

While Hordoir et al. (2018, 2019) speculated that these changes in thermocline depth during summer will impact the vertical overturning circulation, the meridional overturning circulation in the Baltic Proper does not show a clear signal but rather a northward expansion of the main overturning cell (Gröger et al., 2019). Indeed, the effect is expected to be small (Placke et al., 2021).

The number of MHWs within climatological 30-year time slices is shown in Fig. 18. Under historical climate conditions, MHWs are virtually absent in open ocean areas. They are most frequent in shallow regions and more abundant along the eastern (Baltic states) than the western (Swedish) coasts, which may reflect the greater frequency of coastal upwelling events along the western than the eastern coasts of the Baltic Sea. Even under the RCP4.5 scenario, wide areas of the Baltic Proper are affected by MHWs roughly once a year. The strongest response is projected for the high-emission RCP8.5 scenario and specifically in marginal basins such as the Gulf of Riga and the Gulf of Finland, where in the future MHWs will occur two or three times per year. Not only the frequency but also the average duration of the MHWs will increase with climate warming. Under RCP8.5, MHWs of ∼20 d duration will occur even in the open Gulf of Bothnia (Fig. 18). This increase in MHWs is likewise linked to an increased frequency of tropical nights in the Baltic Sea (Meier et al., 2019a; Gröger et al., 2021b).

MHWs can also be analysed by calculating them with respect to the 95th percentile temperature of the historical reference climate (Fig. 19). For the historical climate, the average duration of MHWs in most regions is <20–30 d, although in the southern Baltic Sea, especially west of the Baltic Proper, MHWs are more frequent. However, the climate change signal is characterised by MHWs that are both more frequent and of longer duration. In RCP4.5, MHWs in the Baltic Sea occur at least every year. The strongest increase in frequency is near the coasts, but the average duration increases less than in the open sea (Fig. 19). This is probably related to repeated cold-water entrainments from the open sea that interrupt warm periods because of the larger variability in the coastal zone than in the open sea. In addition, with their lower heat storage capacity, shallow areas are more sensitive to cold weather events and the associated oceanic heat loss.

In the CLIMSEA ensemble, salinity changes are not robust; i.e. the ensemble spread is larger than the signal (Meier et al., 2021). Under both RCP4.5 and RCP8.5, the ensemble mean salinity change is small because the impact on salinity of the projected increase in total river runoff from the entire catchment (Fig. 3) is approximately compensated by the impact of larger saltwater inflows due to the projected SLR (Table 8). Hence, compared to previous studies such as those by Meier et al. (2011b; ECOSUPPORT) and Saraiva et al. (2019a; BalticAPP; Fig. 12), the ensemble mean salinity changes in CLIMSEA are much smaller (Table 8). In idealised sensitivity experiments performed with the RCO-SCOBI model for the period 1850–2008 (Meier et al., 2017, 2019d), the change in the average Baltic Sea salinity (1988–2007) increased linearly with SLR and at a rate of ∼1.4 g kg-1 m-1 (Table 9).

Following global sea level changes, SLR in the Baltic Sea will accelerate (Hünicke et al., 2015; Church et al., 2013; Bamber et al., 2019; Oppenheimer et al., 2019; Weisse and Hünicke, 2019), albeit at a somewhat slower rate than the global mean because of the remote impact of the melting Antarctic ice sheet (Grinsted, 2015). Changes in SLR in the North Atlantic (and the Baltic Sea) will be larger in response to the melting of the Antarctic ice sheet than to the melting of Greenland, due to gravitational effects. For a mid-range scenario, SLR in the Baltic Sea is projected to be ∼87 % of the global mean (Pellikka et al., 2020). Land uplift will partly, or even more than fully, compensate for the eustatic SLR, in particular in the northern Baltic Sea (e.g. Hill et al., 2010). In RCP2.6 and RCP8.5, the global mean sea level in 2100 is 43 and 84 cm higher than during the period 1986–2005 (Oppenheimer et al., 2019). For these two scenarios, likely ranges are 29–59 and 61–110 cm, respectively. Bamber et al. (2019) assessed ice sheet dynamics in detail and subsequently estimated global-median SLRs in 2100 of 69 and 111 cm for low- and high-case scenarios, respectively. Likely ranges according to the authors were 49–98 and 79–174 cm and very likely ranges 36–126 and 62–238 cm.

In BalticAPP and CLIMSEA scenario simulations, sea level changes are small (Fig. 12, Table 8), whereas in ECOSUPPORT scenario simulations they are larger, particularly in spring, because one member of the multi-model ensemble considers Archimedes' principle (not shown). Note that the sea level changes shown in Fig. 12 consider only changing river runoff, changing wind and melting sea ice as affecting the sea level according to Archimedes' principle (only in the ECOSUPPORT ensemble); as neither the global mean SLR nor land uplift is included, they have to be added (e.g. Meier, 2006; Meier et al., 2004a).

In CLIMSEA, there are no statistically significant seasonal changes in the SLR (Fig. 20). In both GHG concentration scenarios, the largest changes are only about ±5 cm. According to these results, systematic changes in the regional wind field (Christensen et al., 2021) and non-linear effects are negligible. Instead, in the projections, the mean absolute sea level in the Baltic Sea simply follows the mean sea level in the North Atlantic. However, the spatially inhomogeneous isostatic adjustment will considerably alter patterns of sea level changes relative to the sea floor.

In response to the global mean SLR, the sea level extremes in the Baltic Sea that are rare today will become more common in the future (e.g. Hieronymus and Kalén, 2020). However, changes in sea level extremes relative to the mean sea level will not be statistically significant because wind velocities are projected to remain unchanged (Christensen et al., 2021). The exceptions are areas with sea-ice decline since they are linked to a decrease in atmospheric stability accompanied by increased wind velocities, the result of increases in temperature and turbulent fluxes (Meier et al., 2011c). These increases will mostly translate as changes from calm to light wind conditions as the stable atmospheric boundary layer becomes less stable. For stronger wind conditions related to high sea level extremes, the impact of stratification effects on mixing is small. In addition, open water areas after sea-ice loss have a smaller surface roughness than ice-covered areas, with the reduced surface friction leading to an increase in wind velocities.

As sea level extremes also depend on the path of low-pressure systems over the Baltic Sea area (Lehmann et al., 2011; Suursaar and Sooäär, 2007), which in a future climate do not show systematic changes (Christensen et al., 2021), changes in sea level extremes relative to the mean sea level are not expected. In addition, a large internal variability at low frequencies prevents the detection of climate-warming-related changes in sea level extremes (Lang and Mikolajewicz, 2019).

Projected changes in bottom oxygen concentrations differ considerably between ECOSUPPORT and BalticAPP/CLIMSEA scenario simulations, as illustrated for summer (Figs. 21 and 22, Table 10), whereas the differences between BalticAPP (SLR = 0 cm) and CLIMSEA (SLR > 0 cm) scenarios are smaller (Meier et al., 2021). The differences between the ECOSUPPORT and BalticAPP ensembles mainly reflect the different experimental setups of the simulations and the different nutrient input scenarios (Meier et al., 2018a). While in shallow regions without a pronounced halocline future bottom oxygen concentrations decrease in all scenario simulations, due to the lower oxygen saturation concentrations, in deeper offshore regions with a halocline, changes in bottom oxygen concentration depend largely on the nutrient input scenario (Figs. 21 and 22). In ECOSUPPORT scenario simulations, the future bottom oxygen concentration decreases significantly in all scenarios except the BSAP, where in deeper regions it changes only slightly on average (see Meier et al., 2011b). By contrast, in the BalticAPP projections, under the BSAP bottom oxygen concentrations in deeper regions increase considerably, regardless of the degree of warming (see Saraiva et al., 2019a; Meier et al., 2011b). Under RCP4.5, bottom oxygen concentrations increase even under the nutrient inputs of REF and WORST, whereas RCP8.5 predicts slight reductions in the Bothnian Sea and southwestern Baltic Sea, in particular under WORST. Similar results were calculated for the CLIMSEA ensemble (Meier et al., 2021).

Most of the differences in the oxygen concentration changes between the ECOSUPPORT and BalticAPP/CLIMSEA ensembles can be explained as follows. In ECOSUPPORT, changes in nutrient input relative to the historical period 1961–2006, including the observed nutrient inputs averaged from the period 1995–2002, were applied (Gustafsson et al., 2011; Meier et al., 2011b). For the historical period 1980–2002, these inputs were lower than in BalticAPP/CLIMSEA scenario simulations because in the latter the observed monthly nutrient inputs, including the pronounced decline from the peak in the 1980s until the much lower recent values, were used as the forcing (Meier et al., 2018a). Furthermore, in ECOSUPPORT, future nutrient inputs under the BSAP scenario were calculated as relative changes, resulting in higher future inputs than in BalticAPP/CLIMSEA, in which absolute values of the BSAP were applied.

Hence, the reductions between future and historical nutrient inputs are smaller in ECOSUPPORT under the BSAP than in BalticAPP/CLIMSEA (Table 6) and result in a smaller response of biogeochemical cycling. We argue that the more realistic historical simulation, including a spin-up since 1850, based on observed or reconstructed nutrient inputs as used in the BalticAPP and CLIMSEA ensembles result in a model response that is more realistic than that of the ECOSUPPORT scenario simulations.

In ECOSUPPORT, the hypoxic area is projected to increase under REF and BAU nutrient input scenarios (Meier et al., 2011b). Only under BSAP is there a slight decrease compared to the early 2000s.

In CLIMSEA under REF, the hypoxic area is projected to decrease slightly until about 2050, as a delayed response to nutrient input reductions, and then increase again towards the end of the century, presumably in response to increased nutrient inputs and warming (Fig. 23). Larger hypoxic areas are calculated under RCP8.5 than under RCP4.5. Under BSAP, the hypoxic area is projected to considerably decrease. At the end of the century, the size of the hypoxic area is expected to be 22 %–78 % smaller than the average size during the period 1976–2005. This range represents the results of the various scenario simulations.

In accordance with previous studies, such as Saraiva et al. (2019b) and Meier et al. (2021), the impact of warming (reduced oxygen solubility, increased internal nutrient cycling, increased riverine inputs) and of increasing stratification (decreased ventilation) will be an amplified depletion of oxygen that enlarges the hypoxia area in the Baltic Sea and partially counteracts nutrient input abatement strategies such as the BSAP. However, in all available scenarios, the impact of climate change is smaller than the impact of nutrient input changes.

While in ECOSUPPORT scenario simulations of future climate the projected surface phosphate concentrations in winter increase under all three nutrient input scenarios (except in the Gulf of Finland in BSAP), in BalticAPP projections the surface phosphate concentrations in winter decrease almost everywhere (except in the Oder Bight and adjacent areas in REF and WORST) (not shown). In contrast to the nearly ubiquitous changes in the surface phosphate concentration, larger nitrate concentration changes are usually confined to the coastal zone and differ in their signs. In ECOSUPPORT projections, the increases in winter surface nitrate concentrations in REF and BAU are largest in the Gulf of Riga, the eastern Gulf of Finland and along the eastern coasts of the Baltic Proper (not shown). In BalticAPP projections, the increases in winter surface nitrate concentrations in REF and WORST are largest in the Bothnian Bay and the Odra Bight, while in the Gulf of Riga and the Vistula lagoon nitrate concentrations decrease. Overall, the differences in surface nutrient concentrations between the two sets of scenario simulations are considerable (not shown) and can be explained by the large differences in nutrient inputs from land. Thus, while the projected changes in inputs in ECOSUPPORT refer to the average inputs during 1995–2002, in BalticAPP scenario simulations the observed historical changes include a decline in nutrient inputs since the 1980s (Meier et al., 2018a).

Annual mean changes in surface phytoplankton concentration (expressed as chlorophyll concentration) follow the changes in nutrient concentrations and are confined to the productive zone along the coasts (Fig. 24). In ECOSUPPORT projections, annual mean Secchi depths decrease in all scenario simulations (see Fig. 25 and Table 11). In the BalticAPP projections, the area-averaged Secchi depths generally increase, except in the combined RCP8.5 and BAU scenario (Table 11), indicating a general improvement of the water quality in future compared to the present climate. The most striking changes occur in the BSAP scenario, in which the Secchi depth increases by up to 2 m in the coastal zone of the eastern Baltic Proper. Changes in stratification (illustrated by the differences between BalticAPP and CLIMSEA ensembles and between the CLIMSEA ensemble mean and high SLR scenarios) have only a minor impact on water transparency (Table 11). The overwhelming driver of the changes in the Secchi depth is the nutrient input scenario (illustrated by the differences between ECOSUPPORT and BalticAPP/CLIMSEA ensembles and highlighted by, in some cases, contradictory signs in the changes).

In CLIMSEA under the BSAP, primary production and nitrogen fixation are projected to considerably decrease in a future climate (Fig. 23). According to this scenario, the interannual variability declines. Under REF, nitrogen fixation is projected to slightly decrease until ∼2050, as a delayed response to nutrient input reductions, and then to increase towards the end of the century, likely in response to increased nutrient inputs and warming. At the end of the century, primary production and nitrogen fixation will be at the same level as under current conditions. The impact of warming is larger under high than under low nutrient conditions (Saraiva et al., 2019b).

The dominant large-scale atmospheric pattern controlling the climate in the Baltic Sea region during winter is the North Atlantic Oscillation (NAO; Hurrell, 1995). However, its influence is not stationary but depends on other modes of variability, such as the Atlantic Multidecadal Oscillation (AMO; Börgel et al., 2020). During the past climate, the relationship between the NAO index and regional climate variables in the Baltic Sea region, such as SST, changed over time (Vihma and Haapala, 2009; Omstedt and Chen, 2001; Hünicke and Zorita, 2006; Chen and Hellström, 1999; Meier and Kauker, 2002; Beranová and Huth, 2008).

Figure 26 shows the calculated ensemble mean winter (December–February) NAO index for the period 2006–2100. For the RCP4.5 emission scenario, the NAO shows high interannual variability. Following a wavelet analysis, the calculated NAO index exhibits decadal variability, which differs for every model (not shown). A comparison of RCP4.5 with the high-emission scenario RCP8.5 shows that the spread of the ensemble increases with increasing GHG concentrations. Figure 26 also depicts the running correlation between the NAO index and the area-averaged SST. The correlation remains positive but it is not constant in time. Also evident from a comparison of RCP4.5 and RCP8.5 is that there are no systematic changes in the two emission scenarios, although for RCP8.5 the ensemble spread is slightly larger.