Oceanographic regional climate projections for the Baltic Sea 1 until 2100 2

The Baltic Sea, located in northern Europe, is a semi-enclosed, shallow and tide-less sea with seasonal 13 sea-ice cover in its northern sub-basins. Its long water residence time contributes to oxygen depletion in the bottom 14 water of its southern sub-basins. In this study, recently performed scenario simulations for the Baltic Sea including 15 marine biogeochemistry were analysed and compared with earlier published projections. Specifically, dynamical 16 downscaling using a regionally coupled atmosphere-ocean climate model was used to regionalise four global Earth 17 System Models. However, as the regional climate model does not include components representing terrestrial and 18 marine biogeochemistry, an additional catchment and a coupled physical-biogeochemical model for the Baltic Sea 19 were included. Previous scenario simulations and scenarios taking into account the impact of various water levels 20 were examined. According to the projections, compared to the present climate, higher water temperatures, a 21 shallower mixed layer with a sharper thermocline during summer, less sea-ice cover and greater mixing in the 22 northern Baltic Sea during winter can be expected. Both the frequency and the duration of marine heat waves will 23 increase significantly, in particular in the coastal zone of the southern Baltic Sea (except in regions with frequent 24 upwellings). Nonetheless, due to the uncertainties in the projections regarding regional winds, the water cycle and 25 the global sea level rise, robust and statistically significant salinity changes could not be identified. The impact of 26 a changing climate on biogeochemical cycling is predicted to be considerable but still smaller than that of plausible 27 nutrient input changes. Implementing the proposed Baltic Sea Action Plan, a nutrient input abatement plan for the 28 entire catchment area, would result in a significantly improved ecological status of the Baltic Sea, including 29 reductions in the size of the hypoxic area also in a future climate, which in turn would increase the resilience of 30 the Baltic Sea against anticipated climate change. While our findings regarding changes in heat-cycle variables 31 mainly confirm earlier scenario simulations, they differ substantially from earlier projections of salinity and 32 biogeochemical cycles, due to differences in experimental setups and in input scenarios for bioavailable nutrients. 33

northern Baltic Sea during winter can be expected. Both the frequency and the duration of marine heat waves will 23 increase significantly, in particular in the coastal zone of the southern Baltic Sea (except in regions with frequent 24 upwellings). Nonetheless, due to the uncertainties in the projections regarding regional winds, the water cycle and 25 the global sea level rise, robust and statistically significant salinity changes could not be identified. The impact of 26 a changing climate on biogeochemical cycling is predicted to be considerable but still smaller than that of plausible 27 nutrient input changes. Implementing the proposed Baltic Sea Action Plan, a nutrient input abatement plan for the 28 entire catchment area, would result in a significantly improved ecological status of the Baltic Sea, including 29 reductions in the size of the hypoxic area also in a future climate, which in turn would increase the resilience of 30 the Baltic Sea against anticipated climate change. While our findings regarding changes in heat-cycle variables 31 mainly confirm earlier scenario simulations, they differ substantially from earlier projections of salinity and 32 biogeochemical cycles, due to differences in experimental setups and in input scenarios for bioavailable nutrients. 33 Interpolation (EnOI) method to integrate profiles of temperature, salinity and the concentrations of oxygen, 271 ammonium, nitrate and phosphate determined by the Swedish environmental monitoring program into the RCO-272 SCOBI model. As reanalysis data are available for the period 1971-1999, we limited our bias calculations to 1976-273 1999, the overlap period between the historical period of the scenario simulations and the reanalysis data. Model 274 data of historical periods of BalticAPP and ECOSUPPORT scenario simulations were evaluated by Saraiva  First, the monthly average of SST was computed from the model output every 48 h. The linear trend was then 291 calculated using the Theil-Sen estimator (Theil, 1950;Sen, 1968). The trend computed with this method was the 292 median of the slopes determined by all pairs of sample points. The advantage of this computationally expensive 293 method is that it is much less sensitive to outliers. The significance of the SST trends was evaluated from a Mann-294 Kendall non-parametric test with a threshold of 95%. The SST trends were computed by season and annually. In 295 the latter case, the annual cycle was removed before the linear trend was computed. 296 297 Following Kniebusch et al. (2019), we performed a ranking analysis to identify the atmospheric drivers others than 298 air temperature that are most important for the monthly variability of SST in each ESM forcing of the CLIMSEA 299 data set and in the RCP scenarios RCP4.5 and RCP8.5. The SST trend is dominated by the trend in air temperature. 300 Thus, to eliminate the air temperature effect on SST, the difference between the SSTs and a linear regression 301 between the SSTs and surface air temperatures (SATs) was calculated. This was followed by applying a cross-302 correlation analysis of the residual SSTs to determine the main factor driving the SST trend. For each grid point 303 and variable (i.e. cloudiness, latent heat flux and u-v wind components), the explained variance was calculated and 304 the variable explaining the most variance was identified. 305 306

Marine heat waves 307
During recent decades, the Baltic Sea region has warmed faster than either the global mean warming (Rutgersson 308 et al., 2015;Kniebusch et al., 2019) or any other coastal sea (Belkin, 2009), making it prone to marine heat waves 309 (MHWs). Indeed, short periods of abnormally high water temperatures have been documented for the Baltic Sea 310 (Suursaar, 2020). MHWs can be defined with reference to the mean climatology (e.g. the 90th, 95th, 98th percentile 311 temperature) or by temperatures exceeding absolute temperature thresholds, defined with respect to the end-user 312 application (Hobday et al., 2018). In most cases, MHWs are defined by the number of periods, their intensity, their 313 duration and the specific purpose (Hobday et al., 2018). In this study, the focus was on the general impact of 314 climate change and the sensitivity of ecosystem dynamics. Hence, MHWs are defined herein as periods of SST ≥ 315 20°C lasting for at least 10 consecutive days. For comparison, we showed also MHWs defined as periods of SST 316 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 321 period 1970-1999, the annual mean SST was ~7.8°C (Fig. 4). The mean seasonal cycle of the SST is pronounced. 322 Thus, every winter, the northern Baltic Sea, including the Bothnian Bay, Bothnian Sea and the eastern Gulf of 323 Finland, is typically covered by sea ice (not shown). Due to its large latitudinal extension, the Baltic Sea is 324 characterised throughout the year by a distinct SST difference between the colder northern and warmer southern 325 sub-basins (Fig. 4). In the southern Baltic Sea, there is also a pronounced west-east temperature gradient, mainly 326 during summer and autumn, which reflects the large-scale cyclonic circulation that transports warmer, more saline 327 southern waters along the eastern coast and colder, less saline northern waters along the western side (see Gröger  In the ECOSUPPORT scenario simulations, there is also a systematic warm bias of the RCAO driven by GCMs 339 at the lateral boundaries, such that winter water temperatures are too warm and sea-ice cover is too low (Meier et 340 al., 2011d, d;2012c, d). While these biases occur in all three applied Baltic Sea models (Table 3)   The first MHW index uses a fixed threshold that emphasises the environmental impact of the heat waves. In 371 particular, diazotrophic nitrogen fixation becomes effective at higher temperatures. The spatial pattern of MHWs 372 is strongly related to the simulated SST. Figure 6a shows that MHWs are mostly absent in the open sea of the 373 Baltic proper and further north in the Gulf of Bothnia, but they are highly abundant in shallow marginal bays such 374 as the Gulf of Finland and Gulf of Riga as well as along the coasts. The MHWs produced by the RCO ensemble 375 mean are generally more frequent and of longer duration than those of the reanalysis data set. Furthermore, the 376 coastal signature of high abundance extends further offshore (Fig. 6a). For the Belt Sea and Bay of Lübeck, this 377 leads to considerable deviations from the reanalysis data set. 378

379
The second index is based on a reference climatology, here defined as that of 1976-1999. The number of MHWs 380 ( Fig. 6c) correlates negatively with their average duration (Fig. 6d). This is somewhat more pronounced in the 381 reanalysis data set. In general, the patterns obtained with the reanalysis data and the RCO are similar but the 382 amplitude of spatial variance is higher in the former (Fig. 6c), as it includes small-scale regional observations. In 383 the RCO (Fig. 6d), MHWs in the open sea are of the longest duration, with their interruption likely due to the 384 vertical mixing induced by wind events. 385 386 Since MHWs in the Baltic Sea are predominantly a summer phenomenon, the stability of the seasonal thermocline 387 is likely a key element in their dynamics such that processes related to vertical mixing can be considered a 388 benchmark in their simulation by the models. Given that mixing is highly parameterised in current ocean models, 389 the RCO reproduces the spatial patterns of the number and average duration of MHW reasonably well. 390

391
In the literature, MHWs in the ECOSUPPORT scenario simulations have not been analysed. 392

Salinity 393
The annual mean sea surface salinity (SSS) distribution shows a large north-south gradient mirroring both the 394 input of freshwater from rivers, mostly located in the northern catchment area, and saltwater inflows from the 395 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 396 and eastern Gulf of Finland. For the period 1970-1999, the annual mean SSS of the Baltic Sea including the 397 Kattegat was ~7.3 g kg −1 . Large inflows of heavy saltwater from the Kattegat occasionally ventilate the bottom 398 water of the Baltic Sea, filling its deeper regions (Fig. 7). As tides are almost absent, mixing is limited such that 399 the water column is characterised by a pronounced vertical gradient in salinity, and consequently also in density, 400 between the sea surface and the bottom. 401

402
Probably due to differences in the data of the hydrological model (E-HYPE) compared to observations, SSS in the 403 coastal zone and the Kattegat is on average lower in the CLIMSEA climate models than in the reanalysis data of 404 Liu et al. (2017) (Fig. 7). The spatially averaged, annual mean bias is −0.4 g kg −1 . Bottom salinities in the Belt 405 Sea, Great Belt area and the Gotland Basin (especially in the northwestern part) are considerably higher and in the 406 Bornholm Basin considerably lower in the climate models than in the reanalysis data (Fig. 7). The spatially 407 averaged, annual mean bias is +0.3 g kg −1 . Hence, vertical stratification in the Belt Sea, Great Belt area and the 408 Gotland Basin is also larger in the climate models than in the reanalysis data, because the difference between 409 surface and bottom salinities is a good proxy for vertical stratification.

Sea level 416
Due to the seasonal cycle in wind speed, with wind directions predominantly from the southwest, the sea level in 417 the Baltic Sea varies considerably throughout the year, with the highest levels (~40 cm), measured relative to the 418 Kattegat, occurring during winter at the northern coasts of the Bothnian Bay and at the eastern coasts of the Gulf 419 of Finland (Fig. 7). For the period 1976-1999, the annual mean sea level in the Nordic height system 1960 (NH60) 420 as determined by Ekman and Mäkinen (1996) was ~16 cm, with a horizontal north-south difference of ~35 cm 421 (not shown). This sea level slope was explained by the lighter brackish water in the northeastern Baltic Sea than 422 in the Kattegat and by wind coming from the southwesterly direction, which pushes the water to the north and east 423 (Meier et al., 2004a). 424

425
The differences in the mean sea level between the CLIMSEA climate models and the reanalysis data are small 426 ( Fig. 7) and the spatially averaged, winter mean bias is only +0.6 cm. Sea levels in some parts of the coastal zone 427 such as the western Bothnian Sea are higher in the climate models than in the reanalysis data, probably due to 428 lower salinities. The negative sea level bias in the eastern Gotland Basin suggests an intensified, basin-wide 429 cyclonic gyre. The seasonal cycle of the ensemble mean sea level is relatively well simulated, but with an 430 overestimated sea level in early spring and an underestimated sea level in summer at all investigated tide gauge 431 locations compared to observations and to a hindcast simulation driven by regionalised ERA40 data (Fig. 8). Likewise, winter surface nitrate concentrations in the simulations are close to those in the reanalysis data but in 466 coastal regions they differ due to differences in the inputs from large rivers (Fig. 9). This is exemplified by the 467 Gulf of Riga and the eastern Gulf of Finland, where the large differences between them are due to inputs from the 468 Neva River. Spatially averaged biases during winter, spring, summer, autumn and in the annual mean are rather 469 small but systematic: −1.1, −1.3, −0.5, −0.7 and −0.9 mmol N m −3 respectively. and hydrological data. The authors found a satisfactory agreement, with the results mainly within the uncertainty 508 range of the observations. However, simulated monthly mean nitrogen fixation during 1999-2008 showed a 509 prolonged peak period in July and August whereas according to observations the peak was mostly confined to 510 July. It should be noted that the RCO-SCOBI version used in the scenario simulations discussed here (e.g., Saraiva 511 et al., 2019a) does not contain a CLC model. 512

Water temperature 514
Annual and seasonal mean changes 515 In Figures 11 and 12 and Table 7, annual and seasonal mean SST changes between 1976-2005 and 2069-2098 in 516 RCO-SCOBI are depicted and quantified respectively. The maximum seasonal warming signal propagates 517 between winter and summer from the Gulf of Finland via the Bothnian Sea into the Bothnian Bay (Fig. 11). 518 Maximum warming occurs during summer in the Bothnian Sea and Bothnian Bay. The seasonal patterns of RCP4.5 519 and RCP8.5 are similar although warming is greater in the latter. As SLR has almost no impact on SST changes, 520 BalticAPP and CLIMSEA scenario simulations yield similar results (not shown). The warming level according to 521 ECOSUPPORT is between that predicted by CLIMSEA/BalticAPP RCP4.5 and RCP8.5 because the GHG 522 emissions of the A1B scenario, which forces the ECOSUPPORT ensemble 2 , are between those of the RCP4. Baltic proper and in the Bothnian Bay, the second most important variable is cloudiness. This is also the case in 574 the Bothnian Sea under the RCP4.5 scenario. However, in RCP8.5 the second most important variable at this 575 location is the latent heat flux. The difference is perhaps due to the absence of sea ice, and therefore the amplified 576 air-sea exchange, under RCP8.5. 577 578 In the vertical, temperature trends are larger in the surface layer than in the winter water of the Baltic Sea above 579 the halocline, thus causing a more intense seasonal thermocline (see Section 3.2.2). Surface layer trends are largest 580 in spring and summer (not shown). Elevated trends also characterise deep water, due to the influence of saltwater 581 inflows that will be warmer in a future climate because they originate from the shallow entrance area and occur 582 mainly in winter. Hence, in sub-basins that are sporadically ventilated by lateral saltwater inflows, such as the 583 Bornholm Basin and the Gotland Basin, the deep water below the halocline will warm more than the overlaying 584 Bay favours a widespread deepening of their MLDs, likely caused by wind-induced mixing. In spring, the most 590 pronounced feature is a strong shallowing of the MLD in the Bothnian Sea, probably attributable to the radiative 591 fluxes that warm the surface layer and to less thermal convection . During the historical 592 period, water temperatures in this area were between 2.0 and 3.0°C (Fig. 4). Thus, in the future, surface water 593 warming between 1.6 and 2.4°C (Fig. 11)  MHWs can also be analysed by calculating them with respect to the 95th percentile temperature of the historical 618 reference climate (Fig. 19). For the historical climate, the average duration of MHWs in most regions is < 20-30 619 days, although in the southern Baltic Sea, especially west of the Baltic proper, MHWs are more frequent. However, 620 the climate change signal is characterised by MHWs that are both more frequent and of longer duration. In RCP4.5, 621 MHWs in the Baltic Sea occur at least every year. The strongest increase in frequency is near the coasts, but the 622 average duration increases less than in the open sea (Fig. 19). This is probably related to repeated cold-water 623 entrainments from the open sea that interrupt warm periods because of the larger variability in the coastal zone 624 than in the open sea. In addition, with their lower heat storage capacity, shallow areas are more sensitive to cold 625 weather events and the associated oceanic heat loss. 626

Salinity 627
In the CLIMSEA ensemble, salinity changes are not robust, i.e. the ensemble spread is larger than the signal (Meier 628 et al., 2021). Under both RCP4.5 and RCP8.5, the ensemble mean salinity change is small because the impact on 629 salinity of the projected increase in total river runoff from the entire catchment (Fig. 3) is approximately 630 compensated by the impact of larger saltwater inflows due to the projected SLR (Table 8) (Table 9). In BalticAPP and CLIMSEA scenario simulations, sea level changes are small (Fig. 12, Table 8) whereas in 651 ECOSUPPORT scenario simulations they are larger, particularly in spring, because one member of the multi-652 model ensemble considers Archimedes' principle (not shown). Note that the sea level changes shown in Figure 12  653 consider only changing river runoff, changing wind, and melting sea ice as affecting the sea level according to 654 Archimedes' principle (only in the ECOSUPPORT ensemble); as neither the global mean SLR nor land uplift is 655 included, they have to be added (e.g. Meier, 2006;Meier et al., 2004a). 656

657
In CLIMSEA, there are no statistically significant seasonal changes in the SLR (Fig. 20). In both GHG 658 concentration scenarios, the largest changes are only about ±5 cm. According to these results, systematic changes 659 in the regional wind field (

Phytoplankton concentrations 746
Annual mean changes in surface phytoplankton concentration (expressed as chlorophyll concentration) follow the 747 changes in nutrient concentrations and are confined to the productive zone along the coasts (Fig. 24). In 748 ECOSUPPORT projections, annual mean Secchi depths decrease in all scenario simulations (see Fig. 25 and Table  749 11). In the BalticAPP projections, the area-averaged Secchi depths generally increase, except in the combined 750 RCP8.5 and BAU scenario (Table 11) between BalticAPP and CLIMSEA ensembles and between the CLIMSEA ensemble mean and high SLR 754 scenarios) have only a minor impact on water transparency (Table 11). The overwhelming driver of the changes 755 in the Secchi depth are nutrient input scenarios (illustrated by the differences between ECOSUPPORT and 756 BalticAPP/CLIMSEA ensembles and highlighted by, in some cases, contradictory signs in the changes). 757

Biogeochemical fluxes 758
In CLIMSEA under the BSAP, primary production and nitrogen fixation are projected to considerably decrease in 759 a future climate (Fig. 23). According to this scenario, the interannual variability declines. Under REF, nitrogen 760 fixation is projected to slightly decrease until ~2050, as a delayed response to nutrient input reductions, and then 761 to increase towards the end of the century, likely in response to increased nutrient inputs and warming. At the end 762 of the century, primary production and nitrogen fixation will be at the same level as under current conditions. The 763 impact of warming is larger under high than under low nutrient conditions (Saraiva et al., 2019b). 764

Relation to large-scale atmospheric circulation 765
The dominant large-scale atmospheric pattern controlling the climate in the Baltic Sea region during winter is the 766 North Atlantic Oscillation (NAO; Hurrell, 1995). However, its influence is not stationary but depends on other 767 modes of variability, such as the Atlantic Multidecadal Oscillation (AMO; Börgel et al., 2020). During the past 768 climate, the relationship between the NAO index and regional climate variables in the Baltic Sea region, such as 769 SST, changed over time (Vihma and Haapala, 2009 The various scenario simulation sets have in common that plausible nutrient input changes have a bigger impact 809 on changes in biogeochemical variables, such as nutrient, phytoplankton and oxygen concentrations, than of either 810 the projected changes in climate, such as warming, or changes in vertical stratification. The latter would be caused 811 by increased freshwater inputs, SLR or changes in regional wind fields, assuming RCP4.5 or RCP8.5 scenarios. feedback. This implies that, with increasing warming, SST trends in the northern Baltic Sea will become larger 842 than those in the southern Baltic Sea. Accordingly, in contrast to the present climate, in which mean SSTs 843 considerably decline from south to north, in a future climate the north-south temperature gradient will weaken. 844

845
In contrast to previous scenario simulations, recent scenario simulations considered the impact of the global mean 846 SLR on Baltic Sea salinity, which for the ensemble mean salinity would more or less completely compensate for 847 the effects of the projected increasing river runoff. However, as future changes in all three drivers of salinity (wind, 848 runoff and SLR) are highly uncertain, the spread in the salinity projections of the various ESMs is larger than any 849 signal. 850

851
In agreement with earlier assessments, we conclude that SLR has a greater potential to increase surge levels in the 852 Baltic Sea than does changing wind speed or direction. For the latter, there have been no statistically significant 853 changes during the 21 st century thus far. 854

855
In agreement with earlier studies, changes in nutrient input according to the BSAP or REF scenarios will have a 856 larger impact on biogeochemical cycling in the Baltic Sea than will a changing climate driven by RCP4.5 or 857 RCP8.5 scenarios. Furthermore, the impact of climate change will be more pronounced under higher than under 858 lower nutrient conditions. Hence, without further nutrient input reductions, as suggested by the BSAP, 859 eutrophication and oxygen depletion will worsen. However, the response determined in recent studies differs   (SLR). In the reference simulation, the mean salinity is 7.42 g kg −1 . In method 1, the increase in the water level 1143 was added to the first vertical grid box of the RCO-SCOBI model, while in method 2 the increase in the water 1144 level was evenly divided between the first and second grid boxes. 1145 SLR (in m) −0.24 +0.5 +1.0 +1.5 (method 1) +1.5 (method 2) Salinity (in g kg −1 ) -0.35 +0.71 +1.41 +2.10 +2.03 1146 Table 10. As in Table 8