Projected precipitation and air temperature over Europe using a performance-based selection method of CMIP5 GCMs

General circulation models (GCMs) are able to simulate a large set of features of the atmosphere–ocean system, namely geopotential height, surface pressure, surface air temperature (SAT), wind and moisture parameters, sea surface temperature, sea level change and so forth. SAT and precipitation are widely discussed climate variables. The climatology and variability of precipitation and temperature may be analysed independently or concurrently by employing diverse statistical methods (Zhang et al. 2011; Stankūnavičius et al. 2012; Hao et al. 2013). An increase in the frequency of climate extremes has been found in the last few decades, resulting in a significant loss of life and serious economic impacts on European countries. Owing to the high population density, Europe is a region potentially vulnerable to climatic changes. Therefore, trends in climatic extremes and climate variability, particularly SAT and precipitation, are of great concern to the European community (IPCC 2007; Giorgi & Lionello 2008).

The most advanced tools for the projection of future climate change are up-to-date GCMs (Shukla et al. 2009). Estimates based on their simulation will help in creating adaptive strategies to mitigate climatic change and eventually finding the most appropriate measures for adaptation. The use of coupled ocean–atmosphere model simulations is a significant step forward in studying climate change in vulnerable regions, including Europe, which have been identified as conspicuous ‘hot-spots’ in future climate change projections (Giorgi 2006). Climate change and its manifestation have been frequently examined in regional impact studies (like the PRUDENCE/CORDEX projects) by using regional high-resolution models. Such regional models provide useful information for impact assessments in different regions and require high-performance computing resources. Therefore, many studies consider a smaller set of GCMs as a sensible way to obtain high-resolution projections for future climate change (Pierce et al. 2009; Overland et al. 2011; Bennett et al. 2012; Corney et al. 2013).

The GCMs of the recently initiated fifth phase of the International Coupled Model Intercomparison Project (CMIP5) have finer spatial and temporal resolution, parameterization enhancement and use improved climate change scenarios with an expanded set of climate experiments (Taylor et al. 2012). They can better represent important global processes such as the effect of aerosols, the interaction at the land–ice boundary, stratosphere–troposphere interactions, the carbon cycle, runoff, and bio-chemical interactions between ecosystems and other processes. These diverse processes could lead to different assessments of changes in the SAT and precipitation fields and their statistical significance.

There are certain approaches to choosing a subset of better performing GCMs which lead to less uncertain climate projections. Some of them use the multistaged process of increased resolution with bias adjustment methods (Johnson et al. 2011; Corney et al. 2013). Others are more focussed on analysing information on model errors and the processes that may lead to these errors (Guilyardi et al. 2012). Many of them try to assess the performance of GCMs over a range of metrics, variables, statistics, spatial scales and temporal scales for a region of interest aimed at improving projections (e.g., van Oldenborgh 2005; Smith & Chandler 2010). However, only a few of these studies assess historical trends to reach a ‘better’ performance compared to the above-mentioned results.

The preliminary results of the CMIP5 project have already been discussed (Stoner et al. 2009; Guilyardi et al. 2012; Zappa et al. 2013; Jacob et al. 2013). These results show that most of the El Niño–Southern Oscillation (ENSO) features become more consistent with observational data (Guilyardi et al. 2012). It is important because many authors have shown that ENSO influences climate variability in the European area (Polonsky et al. 1997, 2008; Basharin 2004; Bulić & Kucharski 2012; López-Parages et al. 2014). Also, research papers have argued that the recent warming over Europe is characterized by high spatial and temporal heterogeneity because of intense mid-latitude cyclones generating devastating storms and heavy precipitation (Bengtsson et al. 2006; Polonsky 2008). The impact of atmospheric circulation on future warming has already been considered in the North Atlantic and European regions using the CMIP5 results (Zappa et al. 2013). This analysis showed that the total number of cyclones in the future climate will decrease in winter and summer, but their intensity will increase. Therefore, diversity in the projected mid-latitude circulation may have a direct impact on the quality of the assessment of changes in temperature and precipitation.

In general, not all climate models involved in CMIP (particularly newly designed GCMs) have enough capacity for accurate description of the regional changes and their spatial–temporal variability in retrospective simulations (Polonsky et al. 2011; Handorf & Dethloff 2012). From the entire set of available models it is worth choosing a few that describe the current climate more precisely. Errors in simulating natural climate variability, climatic tendencies and extreme climatic values can cause multi-model spread to become too large, resulting in projections that are too uncertain to be useful for decision-makers. Therefore, the main task of this study is to assess future climate changes in the SAT and precipitation fields by the end of the 21st century based on a selected ensemble of models from the CMIP5 project that are capable of reproducing the current European climate more realistically.

The ‘Data and methods’ section includes a description of the data and methods used in this study. Then the research results are presented, while the discussion and conclusions are provided in the final section. The ‘Results and analysis’ section has two subsections. The first one deals with the historical climate CMIP5 simulations in order to select better models. Based on the obtained results, the description of temperature and precipitation change assessment according to two CMIP5 climate scenarios are presented in the second subsection.

The results of simulations from 12 well-established GCMs included in the CMIP5 were used in this study. Spatial resolution of the models typically varies from 0.9 to 2.8 degrees on a regular latitude–longitude grid. The output from these models is stored at the Program for Climate Model Diagnosis and Intercomparison and is publicly accessible (http://www-pcmdi.llnl.gov). The list of the models used in this study is given in Table 1. Earlier versions of many of these models were reasonably well validated in the previous phase of the CMIP3 project. Data for verification were taken from the Climatic Research Unit of the University of East Anglia (further CRU3.1), constructed on the basis of in situ observations of climatic variables, and interpolated into regular 0.5 × 0.5 degree latitude–longitude grid space (Mitchell & Jones 2005; Harris et al. 2014).

At the very beginning, the CMIP5 historical climate simulations of SAT were analysed to select the models that best reproduce the observed tendencies and variability over Europe, affected by interactions between mid-latitude and tropical processes (Polonsky et al. 2008; Stankūnavičius et al. 2012). Supposedly, such a model selection procedure can reduce the area of inconsistency in the model responses and prevent diverging projections (Tebaldi et al. 2005; Deser et al. 2012; Knutti & Sedlacek 2013). At the subsequent stage, the projections of the SAT and precipitation fields in the selected climate models were examined according to the RCP4.5 and RCP8.5 scenarios for the 21st century. The historical simulations of precipitation fields were not analysed in this study because even the main re-analysis projects (NCEP, ERA40, JRA25) show a wide range of uncertainties in regional precipitation fields (Bosilovich et al. 2008; Basharin 2009).

The decision about the significance of all further statistical values (mean, linear trends and the others) was made at the 95% confidence level, according to Student's t-test, therefore only significant changes estimated with p > 0.95 were highlighted. All statistical characteristics and indices were performed using standard algorithms.

The selection of the GCMs was done in three steps.

At the first stage, the linear trends of SAT for each model were estimated for the period 1950–2005 and the results were compared with the CRU data for the same period. The second half of the 20th century was chosen for the comparison because the observations became more numerous over a larger part of Europe. Preference was given to those climate models that have had more scenario runs.

In particular, the simulations of the Australian (CSIRO-MK36) and Russian (INMCM4) models (Figure 1(a) and 1(b)) demonstrate significant negative linear SAT trends with the highest magnitude in the northeastern part of Europe up to −3 °K/100 years and −10 °K/100 years, respectively. Meanwhile, the Norwegian model has an area with insignificant negative trends (Figure 1(c)), primarily in the eastern part of Europe – from 0 to −1.5 °K/100 years. As the negative trends are insignificant, the Norwegian (NorESM1) model was also included in the analysis for consideration of its simulated variability at the further stage of the research (see the results in Table 2). Thus, the comparison of the results of the simulation of climate models showed that the SAT linear trends for the period 1950–2005 are satisfactorily reproduced in sign and magnitude in the following models: CNRM-CM5, HadGEM2ES, GFDL-CM3, CanESM2, IPSL-CM5A, MIROC5, MPI-ESM, MRI-CGCM3, CSIRO-ACCESS1, BCC-CSM1.1 and NCAR-CCSM4. The 11 selected models also show satisfactory results for the linear trends in temperature over Europe for the summer season. Therefore, these models were examined in more detail in the later stages of the research.

Second, realistic simulation of the European multidecadal variability by GCMs is important because of its profound influence on the projected climate values. The decadal variability is closely related to the multidecadal variability due to nonlinearities in the ocean–atmosphere system (Alvarez-Garcia et al. 2011; Li et al. 2013). For example, a non-stationary relationship between the North Atlantic oscillation (NAO) and the Atlantic multidecadal oscillation has already been recognized in reports of the Intergovernmental Panel on Climate Change (IPCC 2007). That is why the realistically reconstructed decadal climate variability has to be a prerequisite for better reproduction of lower time scales, which is important for assessing climate changes.

Therefore, at the second stage, the reproduced interannual to decadal variability was analysed using the ensemble mean of every model. So for this reason, the leading spatial–temporal mode of SAT variability was determined over the European region, where high natural variability takes place. In winter it represents the well-known NAO pattern revealed in all seasons of the year (Barnston & Livezey 1987; Basharin 2004; Stankūnavičius et al. 2012 and others). Its spatial–temporal pattern determines the climate anomalies on interannual as well as on decadal time scales over the European region. For identification of patterns, empirical orthogonal function (EOF) decomposition can be an effective method for separating natural variability from the noise (Barnston & Livezey 1987; Kim & Wu 1999). Certainly there are other, more advanced techniques able to identify and compare low-frequency variability in GCM simulations and observations, for example presented in Ghil et al. (2002) and Jamison & Kravtsov (2010). However, the EOF analysis remains a quite simple and effective technique often applied in recognizing natural climate variability.

It is widely believed within the CMIP3 project that the leading spatial mode of climate variability of each model ensemble mean is very similar to the one received by using observation-based data (Handorf & Dethloff 2012). It represents an NAO-like meridional dipole pattern. This result was confirmed by our analysis. At the same time, the temporal variation of the leading modes within the CMIP3 project was simulated poorly (Polonsky et al. 2011). Therefore, the contingency between the temporal behaviour of the leading SAT mode for each model and the corresponding mode based on CRU data was calculated for the CMIP5 models selected previously (Table 2). The correlation analysis shows that seven models are partly capable of reproducing the interannual to decadal time scale of the SAT leading mode at a statistically significant level, suggesting that the simulation of this temporal variability within the CMIP5 project has taken a step forward compared to the results obtained within CMIP3 (Polonsky et al. 2011; Handorf & Dethloff 2012). It points to the constructive enhancement (like the finer spatial resolution/parameterization and others) of GCMs within CMIP5. Correlations show that most of the analysed models are able to reproduce the large-scale temporal SAT variability over the European region at a significant level. Therefore, the analysis of historical simulation of the models within the CMIP5 project reveals that CNRM-CM5, HadGEM2ES, GFDL-CM3, CanESM2, IPSL-CM5A, MIROC5 and MPI-ESM models sufficiently reproduce historical tendencies and natural variability at least over the Eurasian continent, and could give a more accurate simulation of future climate change compared to other models.

Third, the subsequent SAT variance of the historical climate simulations throughout the multi-model ensemble was analysed at the final stage of the current research. For this purpose, the maximum SAT variance of the multi-model ensemble (1950–2005) over Europe was calculated step-by-step for the seven models (CNRM-CM5, HadGEM2ES, GFDL-CM3, CanESM2, IPSL-CM5A, MIROC5, MPI-ESM-LR) based on a stepwise addition technique (Troccoli et al. 2008). As a result, this procedure shows that included IPSL ensemble members cause a statistically significant increase (at the 95% confidence level) of the maximum SAT variance of the multi-model ensemble from 20.25 °K² to 44.89 °K² over the European region (35–65N, 10E–60W). This indicates that simulation of the IPSL model could lead to a significantly larger spread of the multi-model ensemble and more uncertain climate assessment. Therefore, the IPSL model has been excluded from the final multi-model ensemble.

Thus, the analysis of historical SAT simulations within the CMIP5 project reveals that the use of the subset of climate models, namely CNRM-CM5, HadGEM2ES, GFDL-CM3, CanESM2, MIROC5 and MPI-ESM-LR, appears to be optimal for a more accurate assessment of future climate change over Europe.

Four new climate scenarios of greenhouse gas emissions (representative concentration pathway (RCP)) emerged in the CMIP5 project (Moss et al. 2010; Meinshausen et al. 2011). Only two of the most discussed future scenarios are generally considered: RCP4.5 and RCP8.5 (Knutti & Sedlacek 2013; Jacob et al. 2013; Dirmeyer et al. 2013). Moreover, these climate scenarios include a much broader set of numbers of GCMs runs for all climate models in comparison with the other two scenarios: RCP2.6 and RCP6.0. (http://www-pcmdi.llnl.gov). Their amount ranges from 1 to 5 members for each climate model. The RCP4.5 and RCP8.5 scenarios suppose a gradual increase in external radiative forcing of 4.5 Wm⁻² and 8.5 Wm⁻² by the end of the 21st century. The RCP4.5 scenario provides for an increase in carbon emissions up to 10 Gt/year (as well as other greenhouse gases) in the atmosphere by the middle of the 21st century, and a reduction to almost the current level by the end of the century. The worst case scenario – RCP8.5 provides for a quasi-exponential increase in greenhouse gas emissions by the end of the 21st century, which exceeds the present level of CO₂ by about 2.5 times. The further analysis is conducted within these selected scenarios.

Thus, based on the previously selected CMIP5 models, the plausible changes in SAT and precipitation over Europe by the end of the 21st century were obtained (Figures 2 and 3). According to the RCP4.5 scenario, in the winter season (Figure 2), the maximum warming magnitude – 5–5.5 °C is expected in the northeastern part of Europe by the end of the 21st century. Typical changes in SAT over Europe vary between 2 and 4 °C. In the summer season, the expected temperature increase varies from 5 °C in southern Europe to 2.5–3 °C in the north. Precipitation changes are not significant in the winter all over Europe, while in summer a precipitation decrease is expected over much of Europe (on average 20–40%), with maximum values in Spain, Greece and Turkey (60–70%). The exception is the northern part of Europe, where changes are not significant in the summer season.

The RCP8.5 scenario (Figure 3), of course, demonstrates a greater magnitude of warming because of stronger radiative forcing. In winter, the maximum temperature rise (up to 9 °C) is expected in the furthermost northeastern part of Europe, while the typical values of the SAT increment around Europe vary around 3–7 °C. In the summer season, the maximum SAT change (also up to 9 °C) will occur in southern Europe and the Balkans, while the least change is expected in northwestern Europe, of about 3–5 °C. However, estimated changes in precipitation in winter are not statistically significant over most of Europe, except for the northern and central parts of the European territory of Russia, where expected changes slightly exceed 20%. In summer, a precipitation decrease is expected over larger Europe, while the maximum decrease (up to 80%) appears over some parts of Spain, Greece and Turkey.

The estimated change of SAT and precipitation over Europe using the presented GCMs selection method is mostly similar to the Fourth Assessment Report of the IPCC (IPCC 2007) and the preliminary CMIP5 results (Cattiaux et al. 2013; Knutti & Sedlacek 2013). However, the structure of the temperature change coincides less closely with the shape of continents, as in Knutti & Sedlacek's (2013) paper, quite possibly due to poor resolution of one of the models, CanESM2, included in the selected multi-model ensemble. There are also differences in the magnitude of changes: the winter temperature change over northern Eurasia is smaller by 1–2 °C than those given by Knutti & Sedlacek (2013). This value of winter maximal SAT warming over northern Europe might be crucial in determining the rate of the Arctic Sea ice melting and shrinking (Jun et al. 2013).

The current analysis of the changes in the precipitation field suggests that lack of precipitation in the summer season is very likely in most parts of central, southern Europe and the Anatolia peninsula (20–40%), as in previous studies (Giorgi & Lionello 2008; Knutti & Sedlacek 2013). Essentially, such a decrease in precipitation agrees with simulations of the climate models showing a future reduction in the number of Mediterranean cyclones (Bengtsson et al. 2006; Raible et al. 2010), which could increase the susceptibility of the region to droughts. At the same time, our results show that in winter there is no significant precipitation change and drying tendency over southern Europe (Figures 2 and 3) that contradicts the IPCC report (IPCC 2007). It may be caused by the well-known fact that this region is characterized by an intense spatial–temporal variability of storm tracks, while such variability is still insufficiently reproduced (see the correlations in Table 2) by climate models (Polonsky et al. 2011; Handorf & Dethloff 2012). Also, the lack of reliable and consistent precipitation estimations found for GCM simulations might result from model deficiencies in the representation of organized convective systems (Toreti et al. 2013). Therefore, this result has to be interpreted as a rough regional climate assessment.

It is worth mentioning that spatial patterns of changes in both precipitation and SAT for two profoundly different emission scenarios are very similar and differ mostly in magnitude, particularly in summer (compare Figures 2 and 3 for July). This is quite striking indication linear response to external forcing to a certain extent.

The physical processes underlying these projected changes have been widely discussed (Rowell & Jones 2006; Giorgi & Lionello 2008; Guilyardi et al. 2012). These studies show that the high summer drying over southern Europe is the result of seasonal migration of the atmospheric centres of action in the North Atlantic measured by the NAO index, probably having a bimodal structure in a warming climate (Coppola et al. 2005). It is associated with an increasingly anticyclonic circulation over the southern part of Europe and a northward shift of the mid-latitude storm track, which reaches its maximum in summer and minimum in winter. Rowell & Jones (2006) highlighted two main mechanisms of summer drying which are: (1) the spring decrease in soil moisture leading to low summer convection and (2) the large sea–land contrast leading to reduced relative humidity and precipitation over the continent. A reduction in the summer convection in a warming climate is recognized in Kendon et al. (2010) and Efimov & Anisimov (2012). It reveals the significant reduction of the convective part in total precipitation. The comparative analysis presented in this study, as well as in the studies listed above, confirm that the European region (particularly its southern part) is among the most vulnerable regions to future global climate change.

In this study, relatively simple comparisons of selected CMIP5 climate models have been done without any in-depth analysis of variation related to model resolution and representation of physical process in the CMIP5 models. The current and related studies (Guilyardi et al. 2012; Zappa et al. 2013; Jacob et al. 2013) have shown that the majority of the climate models within the CMIP5 project reproduce historical climate and its statistical characteristics better than the models in the previous phase (CMIP3). Together with these studies, our results also show that the simulation of temporal large-scale SAT variability within the CMIP5 project has been improved compared to the results obtained within CMIP3 (Polonsky et al. 2011; Handorf & Dethloff 2012). The reason for such improvement is a better description of processes in the selected CMIP5 models leading to a better representation of SST annual cycle amplitude, reduction of the heat flux RMS error and others (Guilyardi et al. 2012). The poor correlations between the leading SAT mode and CRU data of the non-selected climate models (non-bold values and correspondent name of the models in the Table 2) do not necessarily mean they are unable to simulate low-frequency climate variability. These models most likely need more ensemble members (Deser et al. 2012).

Despite constant enhancement, climate models are still uncertain and imperfect to some extent, because they can never fully describe the ocean–atmosphere–biosphere system that they attempt to specify (McWilliams 2007; Knutti et al. 2008; Kundzewicz & Stakhiv 2010; Guilyardi et al. 2012; Dirmeyer et al. 2013). The underestimation or overestimation of changes in temperature, precipitation, soil moisture and other parameters, would result in an inadequate description of the climate extreme conditions and their variability (on different time scales). However, the worst features of the up-to-date climate model simulations are the inadequacy of historical climate trends. In this regard, the continuous development of the coupled ocean–atmosphere system models, their resolution, the implementation of more high-performance computing technology for data transfer and processing hopefully will improve the accuracy of future climate projections (Shukla et al. 2009).

We would like to thank the modelling groups for making their simulations available for analysis, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) for collecting and archiving the CMIP5 model output, and the WCRP's Working Group on Coupled Modelling (WGCM) for making the WCRP CMIP5 multi-model data set available for users. Special thanks to E. Swenson for the initial revision of the manuscript and remarks. We are very grateful to the reviewers for the constructive and helpful criticism of this paper.