The Mediterranean region (10∘ W, 40∘ E, 30∘ N, 45∘ N; ) is located between the arid and warm northern African climate and the humid and mild European climate . The contrast between them is partly explained by the influence of the surrounding oceans, their interaction with the land surface and the general atmospheric circulation characteristics in the mid-latitudes .

Global warming is not homogeneous, and suggests that the Mediterranean region is a climate change hotspot. Consequently, adaptation to changing climate threats is paramount to countries located around the Mediterranean Sea , which live in a complex and diverse socioeconomic situation and have severe vulnerabilities to climate change and variability . The observed warming in the Mediterranean region during the last decades is expected to continue and grow larger than the global-mean warming . Additionally, total precipitation declines were observed during the late 20th century , and have been projected by different multi-model ensembles for the 21st century . Characteristics of the projected Mediterranean climate change have been linked to thermodynamic sources such as land–ocean warming contrast and lapse rate change in summer , and to dynamical processes such as the changes in upper-tropospheric large-scale flow in winter .

Numerical models are used to estimate future climate change. Accounting for the physical processes and interactions in each climate subsystem (atmosphere, biosphere, cryosphere, hydrosphere and land-surface), global climate models (GCMs) aim to project the state of the future climate system. Model runs over long historical or future periods are driven by natural forcings (i.e. solar irradiance and volcanic aerosols) and anthropogenic emissions that alter greenhouse gas (GHG) concentrations, leading to changes in the radiative forcing . GCMs are developed by a number of institutions who always apply the same physical principles but might use slightly different assumptions. This opens the door to performing the same experiments with multiple GCM outputs, leading to more robust estimates. Modelling uncertainty can be sampled by ensembling various models , while running the same model multiple times (referred to as members), with differing initial conditions , under the same experiment samples' internal variability . To make the results comparable, intercomparison projects, where several models perform standardized experiments, have been organized by the international community . The main community effort is the Coupled Model Intercomparison Project (CMIP). In this study, we consider the two latest CMIP phases, CMIP5 and CMIP6 , and explore their similarities and differences over the Mediterranean region. The almost 10 years between CMIP5 and CMIP6 allowed for improvements in the modelling of certain Earth system processes such as cloud feedbacks, aerosol forcings and aerosol–cloud interactions .

CMIP experiments were performed with a large set of models and therefore show many differences in projected changes due to internal variability and the diverse model designs used by the modelling teams. Weighting single model runs according to their performance in simulating the observed past allows constraining the climate modelling uncertainty and obtaining a potentially more accurate estimate of regional climate change signals. Various studies have used different subsetting/weighting approaches such as emergent constraints , performance-based model subsets and model weighting accounting for performance and independence . The last approach has been used in this study as it additionally considers the interdependencies existing between the models.

This study evaluates and quantifies the Mediterranean climate change hotspot for each season over the 21st century by looking into surface air temperature and precipitation changes in the Mediterranean and how they relate to larger-scale responses. We consider three different emission scenarios in order to assess the impact of anthropogenic emission uncertainties over the Mediterranean climate. The CMIP5 and CMIP6 multi-model ensembles are used to estimate the climate change signal, its uncertainty and to illustrate the differences between the two experiments in the region. Finally, a weighting method is applied to each CMIP ensemble based on the criteria of model performance and independence to obtain more robust projections.

Section  describes the climate models and observational data used, and explains the methods used to quantify climate change and weight the projection members. The climate change hotspot in the Mediterranean and the weighted and unweighted projected changes are presented in Sect. , while these results are discussed in Sect. . Section  concludes and raises questions for further investigation.

Apart from the figures displayed in this section and the Supplement, additional ones generated during the study can be found in a shiny app in the following link https://earth.bsc.es/shiny/medprojections-shiny_app/ (last access: December 2021).

Projections obtained from climate multi-model ensembles contain various sources of uncertainties. Different modelling methods and emission scenarios (e.g. land use, GHG emissions) lead to different results . We use different multi-model ensembles and radiative forcing scenarios to consider as many factors as possible contributing to the uncertainty of the Mediterranean climate change projections. Additionally, a weighting method constraining the projections has been applied to reduce uncertainty in the projections.

We have shown that average Mediterranean temperature changes were larger than the global-mean average during JJA, but close to it during DJF, for all scenarios, time periods and model ensembles. This hotspot is projected to enhance over the 21st century under the scenarios RCP8.5, SSP5-8.5, RCP4.5 and SSP2-4.5, and to diminish from the mid- to long term under the RCP2.6 or SSP1-2.6 scenarios. Interestingly, the multi-model ensemble mean projections of the low emission scenario show a recovery of the precipitation decline towards the end of the century, suggesting that precipitation could be restored to historical values relatively fast in the Mediterranean region if strict mitigation policies are applied. Previous studies also have identified the Mediterranean warming amplification , but it must be stressed that this enhanced warming does not apply to the DJF season.

We argue that the different results obtained from CMIP5 and CMIP6 for the Mediterranean hotspot and the unweighted projections are largely due to the global response from each multi-model ensemble. Figures , S2 and S4 show how the regional changes relative to the larger scale are similar for both CMIPs, indicating that CMIP6 is not producing a regional enhancement of climate change, but it rather follows a larger global change. This behaviour is most evident in JJA than in DJF, as the relative changes with respect to larger scales are more similar for the two multi-model ensembles. To further support this statement, we look at the spatial distribution of changes within the Mediterranean region in Figs. , , S3, S8 and S9. The figures generally agree on the spatial distribution of changes even if the magnitudes differ. Therefore, we can argue that the main difference in TAS and PR output from the older (CMIP5) and newer generation (CMIP6) multi-model ensembles is an enhancement of the global change, while its relation with the Mediterranean region response has been maintained. The work of arrives at a similar conclusion for the European region.

The drivers of the projected Mediterranean climate change has been studied by , and . They have found that the mechanisms projected to drive the Mediterranean climate are large-scale upper-tropospheric flow changes (PR in DJF), reduction in the regional land–sea temperature gradient (PR in DJF and JJA) and changes in the north–south lapse rate contrast (PR in JJA, TAS in DJF and JJA). While these drivers have been deeply studied for CMIP5, affirming that the same mechanisms remain valid for the CMIP6 ensemble would be speculative.

Consistent with basic radiative forcing theory , temperature projections have shown that the warming over the 21st century is larger when stronger radiative forcing scenarios are applied. There is confidence in a precipitation decline for the high emission scenario over the whole Mediterranean region in JJA and only in the south during DJF. Conclusions should be drawn carefully from precipitation as there is a large inter-model spread. For other seasons and scenarios, precipitation declines are projected, although results are uncertain due to the large spread and low significance and robustness over most of the region. Regarding HighResMIP, the HR near-term precipitation and temperature changes generally fall within the CMIP6 ensemble distribution and no clear improvement could be seen from the increased resolution in the historical trends, probably due to the small number of HighResMIP models available for the assessment, and the focus on larger-scale changes and temporal resolutions.

The largest source of uncertainty in determining the warming and precipitation changes over the mid- and long-term periods is the emission scenario (as seen in Fig. ). To illustrate scenario uncertainty, let us take the range between the 5th and 95th percentile of the low (high) and high (low) emission scenario distributions for temperature (precipitation) changes. CMIP6 shows a range from 1 to 9 ∘C warming and -62 % to 19 % precipitation long-term changes in JJA. CMIP5 ranges from 0.1 to 7.5 ∘C warming and -54 % to 18 %. This broad spectrum of possible futures has various possible associated outcomes. The inter-model spread grows at faster rates throughout the 21st century with higher radiative forcing, in part due to the differing climate sensitivities of the models inside the ensemble (see Fig. S6); i.e. the differences between a low and a high climate sensitivity model will become amplified with larger radiative forcing.

The implications of an 8.5 Wm-2 increase in radiative forcing from preindustrial times by the end of the century could pose severe strains on human health, due to heat-related illnesses and altered transmission of infectious diseases ; food security due to crop pests and diseases and productivity declines in many countries whose economies depend on agriculture ; water insecurity due to droughts and changing rainfall patterns in vulnerable regions . Note that the three climate change-induced impacts defined above are closely intertwined and may increase existing scarcities.

In face of the very pessimistic future projected by the high emission scenario, some studies argue that 8.5 Wm-2 forcing is highly unlikely as it is based on an expansion of coal use throughout the 21st century instead of on a reduction . In the context of energy transition and decreasing demand for coal, the high emission scenario has often been criticized . Nevertheless, studies on the carbon cycle discuss that CO2 feedbacks might be underestimated in the GHG concentration scenarios , and thus we have considered keeping the 8.5 scenarios as an extreme yet possible future.

The CMIP6 ensemble is known to have models with notably higher climate sensitivity than CMIP5; i.e. radiative forcing generates stronger changes and at a faster rate . The higher sensitivity could be due to model design or the definition of the radiative forcing scenario. Even if SSP and RCP scenarios are labelled after the radiative forcing (in Wm-2) by the end of the century, the transient GHG concentrations are different . Wyser et al. (2020) suggests that running the same model with equal 2100 GHG concentrations from SSP and RCP (2.6, 4.5 and 8.5 Wm-2) leads to larger temperature changes when forcing the model with the former. It has been argued that improvements in the formulation of clouds and aerosols in CMIP6 are major contributors to larger climate sensitivities with respect to CMIP5 . Even if there is higher sensitivity to radiative forcing in some CMIP6 models, this behaviour is not reproduced by all of them, resulting in a larger inter-model spread compared to CMIP5.

In terms of which multi-model ensemble performs better, some studies argue that the CMIP6 ensemble shows improvements in simulating the climate of historical references in China , Turkey , the Tibetan plateau and the global mean . Nevertheless, as no performance studies have been made specifically for the Mediterranean region, we cannot speculate as to which ensemble performs better. Therefore, this would be a topic of interest for further study.

Assessing the weighted temperature ensemble, we found that the CMIP6 distribution shifts to lower changes, meaning that models showing larger TAS changes have been down-weighted, reducing the differences between CMIP6 and CMIP5 experiment medians and means. To find the reason behind this shift we plotted the ensemble warming distribution for the long term after applying only the performance weights (numerator of equation ) and compared it to the raw and fully weighted ensembles (see Fig. S12). We found that the independence weights are those shifting the CMIP6 ensemble to lower warmings rather than the performance. In this regard, CMIP5's median is unaltered by independence and its effect can only be seen in inter-model spread changes. JJA performance weights shift CMIP5 and CMIP6 to larger warmings, suggesting that a number of the members projecting larger changes do a better job at representing the historical climate. A last remark that can be extracted from Fig. S12 is that both independence and performance weighting play an important role, which changes between seasons and ensembles. Therefore, there is not a straightforward interpretation of the general behaviour of the weights.

This study aims to analyse the projected temperature and precipitation changes by the CMIP5 and CMIP6 multi-model ensembles in the Mediterranean region. Different scenarios and seasons have been assessed to tackle the uncertainties inherent to ensemble projections. To complement the traditional information provided, a weighting method that accounts for historical performance and inter-independence of the models has been applied to offer an alternative view of the temperature projections.

The Mediterranean is a climate-change hotspot due to amplified warming and drying when compared to the large-scale climate behaviour. The amplified warming of the Mediterranean is found in JJA and not in DJF. Comparing the Mediterranean hotspot in CMIP5 and CMIP6, we found that the ratio of warming amplification is similar for both multi-model means, meaning that no enhanced regional warming is projected by the CMIP6 ensemble, but it is rather the consequence of a globally warmer ensemble.

Conclusions must be drawn carefully from multi-model ensembles as the single models perform very differently and might share dependencies with each other. Model agreement gives high confidence in significant and robust warming affecting the entire Mediterranean region throughout the 21st century caused by anthropogenic emissions. The Balkan Peninsula during DJF and the Balkan and Iberian peninsulas during JJA are expected to be the most affected regions. Precipitation changes are less robust and significant and show greater spatial heterogeneity than the warming. Significant and robust declines in precipitation are expected to affect the Mediterranean in JJA and the southern part in winter by the end of the 21st century if high emission scenarios are considered. The warming combined with a precipitation decline could put the whole region under strain, especially the south, which has fewer resources to adapt to the changing climate. The biggest source of uncertainty to determine the magnitude of TAS and PR changes is the emission scenario, which will depend on the future policies and measures for mitigation followed. Considering three scenarios, the long-term projected warming (given by the 50 % inter-model spread) spans 1.83–8.49 ∘C according to CMIP6 and 1.22–6.63 ∘C according to CMIP5 in JJA. For precipitation, the decline ranges from -49 % to -16 % in CMIP6 and from -47 % to -22 % in CMIP5. It has also been concluded that part of the increasing warming inter-model spread with time is related to the wide range of ECS values among the ensemble members.

A weighting method has been applied to reduce the uncertainty caused by models that poorly represent key aspects of the historical climate and by the high dependence of the results provided by families of models (that might be overrepresented in the multi-model ensemble). Based on the constrained projections, we conclude that CMIP6 overestimates warming in the Mediterranean and its 25th to 50th percentile inter-model spread. The shift to lower warming seen by the weighted CMIP6 ensemble is driven by the independence weighting. CMIP5 slightly underestimates warming and generally overestimates the IQR inter-model spread. The weighted projections are relevant because they help to reconcile the conclusions extracted from the last two CMIP phases, reducing future climate change uncertainties. The fact that CMIP6's 90 % spread range is unaltered shows that the climate uncertainty might have been underestimated in previous, less physically advanced, CMIP exercises, which displayed smaller inter-model spread when constrained.

Further work is required for the weighting method to identify the most relevant diagnostics that best assess historical precipitation model performance. As spatial heterogeneities can be seen in the Mediterranean region, we suggest considering subregions for the Mediterranean in order to extract more user-relevant information from the constrained projections. Furthermore, it would be of great interest for the community to update studies on the physical mechanisms and the performance of the CMIP6 multi-model ensemble in the Mediterranean region.