The AdriE ensemble is based on a set of implementations of the ROMS modelling system over the Adriatic Sea, with an open boundary at the Strait of Otranto (Fig. a). All simulations are forced with 0.11° resolution, 6-hourly atmospheric fields (total cloud fraction, relative humidity, sea level air pressure, precipitation, long- and shortwave radiation, near-surface air temperature, and near-surface wind velocity components) from the SMHI-RCA4 regional climate model (RCM) driven by six global climate models (GCMs) from the CMIP5 initiative and listed in Table . SMHI-RCA4 has been used as a reference model in consideration of its extensive record of regional-scale climate projections and its performances in reproducing relevant climate processes (which have been found well representative of the CORDEX multi-model ensemble variability; see ), as well as of the availability of the relevant variables with the necessary sub-daily temporal resolution for a comparatively large number of simulations corresponding to different GCMs. The climate ensemble consists of six free-running climate simulations bracketing the period 1987–2099, thus including part of the “historical” (1987–2005, based on the observed radiative forcing) period and the future “scenario” (2006–2099, based on projected values of radiative forcing) under the severe RCP8.5 pathway. In this work, the comparison between end-of-century and recent past conditions is carried out by comparing climate runs in the scenario (SCE, 2070–2099) and control (CTR, 1987–2016) periods. This set of simulations is flanked by an “evaluation” run (EV hereinafter) in which SMHI-RCA4 was driven by the ERA-Interim reanalysis throughout the 1987–2010 period and therefore synchronised, unlike the climate simulations, with the observed day-to-day climate variability. This allows the isolation of the RCM–ROMS modelling chain and the assessment of its skills under an approximation of the best available information at the boundary (ideally the “perfect boundary conditions”; see ), and it permits a one-to-one comparison between modelled and observed records (i.e. modelled fields for each day at each location should match as closely as possible the observations collected on the same days and in the same locations). Conversely, since GCMs are free-running models not constrained, once initialised, to the observed state of the climate system components, the analysis of the climate runs allows the assessment of the skills of the whole modelling chain only in terms of aggregated statistical properties: as a consequence, the methodology for assessing reanalysis-driven evaluation runs does not fully apply to GCM-driven climate simulations.

Table summarises the main features of the runs performed. To characterise the largest spectrum of uncertainty, we opted for maximising the number of GCMs since these exert a larger influence than RCMs on climate change signal variability . The use of different models entails a broad spectrum of uncertainty sources (different formulations, numerical schemes, dynamical features, etc.), while different realisations of the same model (“r” index) address different, but equally realistic, initial conditions.

The horizontal discretisation follows the orthogonal curvilinear grid used for the wave climate projections described by , composed of 75 × 180 nodes with resolution ranging from approximately 2 km in the northern regions and progressively coarsening to nearly 10 km in the southeasternmost end of the domain, whereas the water column is discretised into 15 terrain-following σ levels with increasingly larger thickness towards the bottom. Six subdomains were identified consistently with the work by (namely Northern Adriatic, Kvarner Bay, Middle Adriatic, Western Coast, Dalmatian Islands, and Deep Adriatic; the Adriatic–Ionian subdomain lies mostly outside of the model domain and was not considered in this study) to facilitate a direct comparison, based on their thermohaline properties and circulation features (Fig. b). An additional subdomain (Strait of Otranto) was also introduced in the southeastern five-cell buffer in order to verify the model performance close to the open boundary. The bathymetry was reinterpolated from finer-resolution information adopted in previous works and smoothed as described by in order to prevent the appearance of artifact horizontal pressure gradients and consequent circulation features.

Potential temperature (θ), salinity (S), momentum, and sea level were prescribed as boundary conditions at the Strait of Otranto and derived from the 1/16° horizontal resolution Mediterranean Forecasting System (MFS; see ). Chapman conditions were imposed for free surface; Flather conditions for 2D momentum components; and nudged radiative conditions for 3D momentum components, θ, and S. For the EV run, daily reanalysis values were interpolated on the model grid points throughout the cross section. For the climate runs, climatological monthly values were first computed from the reanalysis fields with reference to the 1987–2017 period. Potential temperature and salinity values were then perturbed with the anomalies computed, with reference to the same period, from Med-CORDEX-derived CMCC-CM profiles in the northeasternmost grid cell of the Ionian Sea. Furthermore, since during the post-processing of the climate runs a new version of MFS was released (1/24° horizontal resolution; see ), an additional evaluation run EV^∗ was carried out to assess the impact of the difference between the two versions on the description of the Adriatic Sea dynamics and the possible implications for the climate projections. Time-averaged potential temperature, salinity, and velocity patterns and their trends are summarised in Fig. . The thermohaline property distribution appears consistent with typical values of the Eastern Intermediate Water (EIW; ), particularly in terms of the component identified in previous literature as Levantine Intermediate Water (LIW, also indicated in the southern Adriatic Sea as “Modified” LIW – MLIW), entering the basin along the eastern continental margin (see for instance , and references therein). The known cyclonic flow across the boundary cross section is weaker in EV than in EV^∗, but this does not affect the capability to reproduce the main circulation patterns in the southern Adriatic Basin, as shown in Sect. . The multidecadal trends for potential temperature in the climate simulations in the CTR period are well bracketed between the values from the EV and EV^∗ runs (namely, the CMEMS reanalyses used as boundary conditions) and consistent (at least for surface values) with the estimates by , while salinity trends appear weaker and less consistent both in the evaluation and in the climate runs.

River mouths were included as 39 point freshwater sources. Climatological values from were used for the main Alpine (Isonzo, Tagliamento, Piave, Brenta, Adige, and Po split into five branches) and Apennine (Reno, Foglia, Metauro, Esino, Musone, Potenza, Chienti, Tronto, Pescara, Sangro, Trigno, Biferno, Cevaro, Ofanto) rivers, as well as for the Drin in the southeastern Balkans, while the freshwater input from the Venice Lagoon drainage basin was taken from . For the rivers and submarine springs along the Croatian coast (Mirna, Raša, Rječina, Bakarac, Crikvenica, Zrmanja, Krka, Jadro, Cetina, Neretva, Ombla, and the Senj and Dubrovnik/Kupari hydropower plants) estimates were taken from , which in turn also relied on for the Italian coast. For the EV and EV^∗ runs, the input from the river Po was computed from the observed discharge at Pontelagoscuro (approximately 80 km upstream; data publicly available at https://simc.arpae.it/dext3r/, last access: 6 June 2025, ), whereas for the climate runs climatological river discharges were rescaled following the spatially averaged modelled rainfall anomaly in the Adriatic catchment (Fig. a). Tides were included only in the EV and EV^∗ runs by imposing 15 tidal constituents from the TPXO dataset on the open boundary at the Strait of Otranto.

The evaluation of the model skills was based on a broad and heterogeneous set of observations for different variables and from different sources (Table ). In order to maximise the comparability of the results, and as far as it was relevant and consistent with the aim of this study, the analysis used the datasets, quantities, and methods from the works by and .

E-OBS v21.0 provided gridded data with 0.1° resolution for daily rainfall (R), sea level pressure (SLP), and near-surface temperature (T) over land. Delivered and periodically updated by the Copernicus Climate Change Service, this dataset combines in situ observations ensembles and is used as a reference for the evaluation of RCMs within the EURO-CORDEX community . Additional rainfall data were also collected from the Tropical Rainfall Measuring Mission (TRMM) gridded dataset with 0.25° resolution over land and sea, based on the combination of microwave–IR data calibrated with rain gauges over land. Gridded near-surface wind information (zonal and meridional components – U, V – and intensity UV) was retrieved from the Cross-Calibrated Multi-Platform wind vector analysis (CCMP) version 2, fitting radiometer and scatterometer data, in situ observations, and ERA-Interim reanalysis into a 6-hourly, 0.25° resolution dataset . Pointwise observations were retrieved with hourly frequency from 312 NOAA stations (NOAA, 2024) for sea level pressure, near-surface temperature, and wind velocity and from the Acqua Alta oceanographic tower (AA, 12°30.55^′ E, 45°18.8^′ N; see Fig. a) for wind velocity.

Sea level information was provided by tide gauge (TG) datasets from 11 stations along the Adriatic coast (Table ). The original time series consist of hourly sea level elevations. The data of Ancona, Marina di Ravenna, Vieste, and Otranto come from ISPRA, Rome (Italian Institute for Environmental Protection and Research; https://www.mareografico.it, last access: 6 June 2025); Venice data come from CPSM, Venice (Tide Monitoring and Forecast Centre; https://www.comune.venezia.it/node/6214, last access: 6 June 2025); Trieste data come from CNR-ISMAR, Trieste ; Bakar data come from the University of Zagreb ; the data of Rovinj, Zadar, Split, and Dubrovnik come from HHI, Split (Croatian Hydrographic Institute). Sea levels at Rovinj, Bakar, Zadar, Split, and Dubrovnik tide gauges were obtained by the mechanical instruments located in a stilling well, with 7 d charts digitised with the Auto-CAD software to obtain hourly sea level values.

Daily sea surface temperature (SST) observations at 0.0417° resolution have been retrieved from the L4 Optimal Interpolation (L4OI) Mediterranean Advanced Very High Resolution Radiometer (AVHRR) SST analysis dataset . A continuous multi-decadal time series for daily in situ sea water temperature at 2 m depth at the Trieste harbour station (T2TS, 13°45^′ E, 45°39^′ N) was adopted from . The observational reference for basin-wide characterisation of the thermohaline properties of the water masses was given by the CTD profile collection from different survey campaigns described by and partially available at 10.5281/zenodo.5707773 .

Due to the absence of the Ionian Sea dynamics in the model implementation and the consequent impossibility of properly reproducing the Adriatic–Ionian Bimodal Oscillating System (BiOS; see ) signal, the analysis of the sea level variability cannot replicate the approach by . Likewise, the absence of a very high resolution along the Croatian coast prevents the possibility of relying on the ADCP dataset used in that work for the validation of modelled circulation. Therefore, the scope of the assessment of these quantities in the present work is to verify that the dynamical properties of the basin are compatible with the observations and with the well-known basin-scale circulation features.

The projection of oceanographic processes within a relatively small regional basin and in changing climate conditions is subject to different sources of uncertainty and different levels, from the evolution of the global climate to how this signal propagates through different scales and how the adopted numerical description impacts the final results. Extensively tackling all the possible sources of error requires the combination of different techniques and in principle an enormous effort, but the main elements of uncertainty can still be circumscribed at an affordable cost. First, evaluating the performance of the RCM–ROMS modelling chain by pursuing the “perfect boundary conditions” approach driving the RCM with a reanalysis allows one to depurate the assessment from the intrinsic errors of the driving GCM. Nonetheless, since this kind of information is obviously not available for the future, the result of this operation is not automatically telling of the capability of the whole modelling chain (GCM–RCM–ROMS in this case) to actually reproduce the future climate. This aspect can be addressed by comparing modelled and observed statistics in the CTR period providing an aggregated description of the climate variability, under the assumption that the skills exhibited by a climate modelling system under reconstructed radiative forcings (as this is what ultimately drives GCMs in historical simulations) are representative of what can be obtained under projected conditions. Finally, the use of an ensemble approach can provide some degree of information about the uncertainty associated with different modelling strategies. The results of the evaluation (EV) run and of the atmospheric fields used as a forcing under “perfect boundary conditions” are presented and discussed in Sect. , whereas the overall climate model skills and uncertainty are addressed in Sect. , alongside the projected future variations for some relevant quantities and processes.

If the focus on the EV run allows the investigation of the model skills in the presence of the best available if not actually “perfect” information, the analysis of the climate ensemble in the recent past provides, besides of course a terms of reference for comparison against the future figures, a further element to complement the evaluation by verifying to what extent the observed climate statistics in the recent past are well reproduced also in a GCM-driven condition. In this direction, and again with reference to the three subdomains considered in Fig. , the climate normals of the monthly values of 1st, 50th, and 99th percentiles of SST in different datasets are shown in Fig.  including also statistics from observed data and the EV run. As a recurring pattern, the climate ensemble statistics appear to satisfactorily match the observations in the winter months while underestimating and overestimating respectively summer and autumn values, most likely as a result of the description of the fluxes of heat along the upper layers of the water column associated with the excess mixing described in Sect. . This behaviour is more evident in the Northern Adriatic, where the maximum mismatch reaches 2.7 °C for the 99th percentile in June. The ensemble spread is generally narrow (around 1 °C) from October to April both in CTR and SCE conditions and significantly larger (> 4 °C) in summer, particularly in the SCE datasets, in some cases obscuring the statistical significance of the future change signal. An examination of uncertainty partitioning through the different modelling chain steps lies beyond the scope of the present study. However, speculations can be made about the well-known large uncertainty characterising GCMs in reproducing crucial mid-latitude summer season dynamics like blocking atmospheric patterns and their response to a warmer climate . Nevertheless, another source of ensemble variability could be given by the non-linear ingestion of the GCM large-scale signal into the local-scale forcing in the nested simulations (SMHI-RCA4 and ROMS) given the non-linear ingestion of the GCM large-scale signal.

The SST increase between SCE and CTR is generally evenly distributed throughout the year and among the different statistics, ranging in most cases, for ensemble means, between +2.8 and +3.2 °C, except for higher percentiles in summer, in which the increase ranges between +3.2 and +3.8 °C in most of the basin (Fig. ). While spatial patterns are relatively uniform in spring and summer, north–south and coast–offshore gradients are visible in autumn and winter, in agreement with the patterns in seasonal temperature change shown for the first time in the Mediterranean region by . Locally higher values appear for 10th percentile SSTs along the southwestern coast in winter, while conversely warming seems to be generally milder in the Northern Adriatic. Sea surface salinity (SSS; Fig. ) is projected to increase, at different rates, throughout the basin with the only major exception of the northeastern coast in winter, in which very small or statistically non-significant variations are envisaged as a consequence of the increase in river runoff in this season counteracting the generalised salinisation trend. Conversely, in the same region and more broadly in the northern and western basin, the highest increase in S is projected in spring and even more in summer, when river runoff is expected to decrease. The picture is less clear for net surface heat fluxes (Fig. ), in which a tendency to some decrease (up to -40 W m⁻², although with an increase in summer heating in the north) is predominant in patchy patterns characterised by a north–south gradient throughout the basin.

Specific applications can benefit from a thorough comparison of the cumulative distributions of SST in the CTR and SCE periods. In particular, with reference to thermal extremes, any difference in the variations of the 90th and 99th percentiles or of the 1st and 10th percentiles is associated with a change in the shape of the tails of the statistical distributions for the two periods and could be a possible source of variability respectively in MHW and CS statistics. Figure  represents the daily SST cumulative distributions for the ensemble (again considering an 11 d sliding window) in different sub-basins. For low-temperature extremes, narrower distribution tails can be found in winter in NA and KB (0.5 °C) and to a lesser extent in WC (0.3 °C) and Middle Adriatic (MA; 0.2 °C); slightly broader distribution tails can be found in late spring and early summer (NA, KB, MA) and autumn (0.3 to 0.5 °C, all subdomains). Conversely, for higher temperatures, projections show moderately narrower distributions in late summer in NA, MA, and WC (0.2 to 0.4 °C) and broader tails in late winter (0.3 to 0.5 °C) and autumn (0.2 to 0.4 °C) throughout the whole domain.

A more comprehensive overview of how climate change affects the statistics of the thermohaline properties of the Adriatic Sea (as well as of the subdomains considered in this study) along the whole water column can be drawn from the quantile–quantile plots depicted in Fig. .

While the projected potential temperature increase in the colder (and deeper) regions of the Deep Adriatic is confined below 1.5 °C, the statistics throughout the different basins reflect the pattern presented in Fig. , with variations mostly clustered around +3 °C. Projected ensemble average salinity increase is mostly encompassed between 0.3 and 0.4, with larger variations on the higher end of the distribution (i.e. for S≥39). In terms of σθ variations, this results in a generalised tendency towards a decrease between 0.4 and 0.5 kg m⁻³ in most of the basin (with the larger decrease corresponding to smaller values), with the exception of the deeper regions of the Deep Adriatic, presently characterised by σθ around 29.2 kg m⁻³ facing a decrease of approximately 0.2 kg m⁻³. Most notably, Fig.  shows that the variability of the results within the ensemble is generally larger than the variability across the subdomains: since the evidence from the EV run (Sect. ) supports a good degree of confidence in the model capability of reproducing the internal dynamics of the Adriatic Sea, this result gives an important account of the relative weight of the GCM–RCM modelling chain in the ocean climate projections at the basin scale. Figure summarises the seasonal modulation of the variation of thermohaline (median) quantities along the water column, again with a focus on the relevant basins for dense water formation and deep ventilation. NA and KB are characterised by similar results, in both qualitative and quantitative terms. In NA (KB), the ensemble-mean increase in median θ values ranges between +2.5 (+2.6) and +2.7 (+2.8) °C in winter and spring and between +2.9 (+2.9) and +3.1 (+3.2) °C in summer, with intermediate variations between +2.7 and +2.8 (+2.8, nearly uniform) °C in autumn. S exhibits a more pronounced vertical variability. For depths smaller than 40 m, increases range between +0.28 (+0.37) in winter and +0.71 (+0.49) in summer, whereas for higher depths the range of the increase lies between +0.34 and +0.43 (+0.32 and +0.43). Median σθ is thus projected to decrease by -0.27 (-0.29) kg m⁻³ in winter, when dense water formation typically takes place, and between -0.27 (-0.36) and -0.42 (-0.40) kg m⁻³, with the largest values for depths larger than 40 m, in autumn. Variations in spring and summer, respectively in the range -0.34 (-0.36) to -0.32 (-0.34) and -0.55 (-0.57) to -0.41 (-0.46) kg m⁻³, albeit with relevant values, are less significant for the thermohaline circulation in the basin. In the Deep Adriatic (DA), the seasonal modulation of climate change on median profiles is mostly visible for h≤200 m; for larger depths, thermohaline quantities vary gradually and, with negligible inter-seasonal differences, up to uniform values for h≥800 m. In the upper layer, variations range between +2.5 and +3.2 °C for θ and between +0.26 and +0.37 for S, again with smaller variations in winter and spring and larger variations in summer and autumn. σθ variations range between -0.40 kg m⁻³ in spring and -0.71 kg m⁻³ in summer. Below the upper layer, the θ increase varies from +2.8 °C for h=200 m to +1.3 °C for h≥800 m, the S increase varies from +0.21 to +0.17, and σθ varies between -0.44 and -0.15 kg m⁻³. Also considered in the light of Fig. , these results indicate that the Adriatic Sea should remain a cooling pool for the Mediterranean Sea, whereas its characteristic behaviour as a dilution basin is expected to be significantly reduced, particularly in summer when the highest salinity increase is expected. In fact, the rate of increase prescribed as a boundary condition (see Fig. ), consistent with other studies (e.g. , for future projections), is not sufficient to explain such comparatively higher values, which appear mostly controlled by the decrease in summerly river runoff.

Before focusing on thermal extremes such as MHWs and CSs, it is worth recalling that the definition of these events is intrinsically associated with some definition of impact, in most cases in the framework of the discourse on climate change. In this direction, the choice of the baseline period as a reference for the computation of the threshold percentiles implies an important assumption of the system on which MHWs and CSs are supposed to act as stressors. More precisely, defining these events in a future scenario with reference to a past climatology implicitly assumes that the target system has limited resilience to warmer conditions (as could be the typical case for human civilisations). In turn, defining these events with reference to a future climatology is compatible with the assumption that the system can adapt to the change in the ordinary conditions and is only (or mostly) vulnerable to significant deviations to those conditions. In the present study, taking as reference thresholds climatological values from the CTR period yields the simple, though important, result that end-of-century conditions under the RCP8.5 climate scenario are persistently corresponding to MHW. This is consistent with the evidence from the recent past, in which the increase in MHW in the Adriatic Sea and eastern Mediterranean has been estimated as high as 100 d yr⁻¹ in the 1982–2020 period . The other hypothesis is considered in the plots in Fig. , in which MHWs and CSs are defined for CTR and SCE, as well as for the EV run and for observations in the control period, with reference to the climatologies for the corresponding periods. Under this approach, modelled differences between SCE and CTR conditions (expressed as monthly mean cumulative intensity of the events) appear generally minor and in any case only occasionally statistically significant. These slight variations appear qualitatively consistent with the changes in the statistical distributions presented in Fig. . The weaker correlation in summer months suggests that the variation in that period could be also associated with the duration of the events rather than with their maximum intensity. In any case, this would suggest that climate change impacts on ecosystems could be mainly controlled by the warming trend of the ordinary conditions, with only a secondary contribution from the change in the characteristics of extreme events. Nonetheless, it is worth noting that the RCM–ROMS modelling chain exhibits some shortcomings in properly capturing the cumulative intensity of extreme events between late spring and early summer (mostly in May and June). While the mismatch is generally fairly small in the case of the EV run (although with the caveat that the reference period is slightly different), this is more evident in the case of the CTR run, thus revealing, alongside the limitations already pointed out in the model performance (e.g. the excess in vertical mixing), some limitations in the GCM–RCM capability to reproduce the extremes during the warm season. Importantly, there is increasing evidence of a generalised tendency of GCMs and RCMs to underestimate temperature extremes historical trends, mostly due to the neglection of aerosol changes (see for instance ). Therefore, this specific result should be treated with some caution.