Composition of the CORDEX ensemble
The CORDEX ensemble in this study consists of 22 different combinations of GCM, GCM member and RCM, with eight GCMs and five RCMs in total. Not all combinations of these have been realized, leaving about half of the GCM (-member)/RCM matrix blank (Table 1). The composition is quite heterogeneous, with a slight dominance of EC-EARTH and HadGEM2-ES for the GCMs and CCLM4-8-17 and REMO2015 for the RCMs. The EC-EARTH is the only model with different downscaled members, and fortunately there is even a pair of two members downscaled with the same RCM (RACMO22E). Additionally, there are two simulations using the first member of the CanESM2 ensemble, downscaled with CCLM4-8-17 and REMO2015, giving us insight in the role of CRCM5’s contribution to the signals of the CanESM2-CRCM5-LE. Yet overall, the sampling is relatively random, which makes systematic analysis on the influence of GCM (-member) and RCM on the variability extremely difficult. Nevertheless, it seems valuable to take a look at the variabilities inside the CORDEX ensemble to better assess the capacity of the ensemble for the variance comparison with the CRCM5-LE.
To get an overview of the influence of the components (GCM/RCM) of each simulation, the climate change signals of temperature and precipitation for the far future FUT3 (2070–2099) are displayed in scatterplots for winter (DJF, Fig. 2) and summer (JJA, Fig. 3) for the TOT domain. A general clustering of the simulations sharing the same GCM can be observed, although there are large differences between the GCM cluster extents.
In winter, the differences between RCM simulations, driven by the same global model, range from 0.1 K in CanESM2 to 0.9 K in MIROC5 for temperature and from 3.6% points [pp] in MIROC5 to 5.1 pp in CanESM2 for precipitation (Fig. 2). These ranges are usually similar to the CRCM5-LE extent. The EC-EARTH members r1 and r3 are close again, as well as the two RACMO22E simulations (r1 and r12). The two CanESM2 simulations fit quite well into the CRCM5-LE, although being at the colder end of the cloud.
In summer, the spread of temperature signals of the same GCM downscaled by different RCMs range from 0.2 K in CNRM-CM5 to 1.5 K in HadGEM2-ES, while the spread for precipitation signals ranges from 4.5 pp in CNRM-CM5 to 36 pp in CanESM2 (Fig. 3). These GCM ranges are larger than in winter, due to the higher importance of large scale circulations, and these are mainly driven by the GCM. The RCM CCLM4-8-17 shows the strongest decreases in precipitation regardless the driving GCM—except CNRM-CM5, which generally seems to have a larger influence on the RCM output than other GCMs. The two CanESM2 simulations show large differences and span a larger range, both in temperature and precipitation, than the CRCM5-LE. The combination of the rather warm and dry CanESM2 with the also rather dry CCLM4-8-17 (usually the driest RCM, driven by the same GCM) results in an extreme decrease of precipitation, accompanied by a strong warming signal. The EC-EARTH is the only GCM with different members (1*r1, 1*r3, 4*r12), giving insight into the variability of another multi-member ensemble. The two simulations of member r1 and r3 have colder and less dry signals than the r12 simulations, yet still at the edge of the r12 cluster. The two simulations using the same RCM and only different members of EC-EARTH (r1_RACMO22E and r12_RACMO22E) are very close. The CRCM5-LE cluster is at the very warm and dry end of the CORDEX ensemble, but the MIROC5_CCLM4-8-17, HadGEM2-ES_CCLM4-8-17 and CanESM2_CCLM4-8-17 simulations show similar or even stronger signals.
The GCMs usually dominate the signal of the simulations, but there are some cases where the RCM can significantly impact the resulting signal. These findings for the TOT domain can be found in a similar manner in the subdomains as well, although differences occur of course. For example, the positive winter precipitation signals mostly originate from Northern European subregions like Scandinavia, whereas the Mediterranean shows mostly negative signals (Figs. S2 and S3, SM). On the other hand, the same contrasting signals (mostly positive in SC, mostly negative in MD) result in a more or less balanced signal spread in TOT for summer (Figs. S4 and S5, SM).
In general, this CORDEX ensemble consists of a number of GCMs and RCMs with a wide range of signals. Although it is not a perfectly composed multi-model ensemble (which would be necessary for a real structured framework), the analysis suggests the assumption that its composition represents a fair assumption of the sources of uncertainty in a multi-model ensemble. It is therefore suited for the comparison with the 50 member single model large ensemble.
Comparison of variability in signals of CRCM5-LE and CORDEX
To better assess and quantify the fraction of internal variability in the CORDEX ensemble, we compare the standard deviations of the CRCM5-LE and the CORDEX ensemble. The variability between EURO-CORDEX models is analyzed on the grid point level, while other publications usually only mark areas where models agree on the sign of change and significant changes, without quantifying the uncertainty of the respective ensemble (Jacob et al. 2014; Vautard et al. 2014; Rajczak and Schär 2017).
Figures 4 and 5 show the standard deviation of signals of CRCM5-LE and CORDEX for all three future periods and the respective SDR values for winter and summer temperature over Europe (respective figures for SON and MAM: Figs. S6 and S7, SM). The CRCM5-LE variability shows a large scale gradient from lower values in the West to higher values in the (North-) East, which might be associated with increasing continental climate as seen in Köppen-Geiger climate classifications (Beck et al. 2018). In winter CORDEX shows a similar, yet not so clear gradient, whereas in summer higher values seem to be found in southern parts of Europe. In both seasons, the most obvious gradient in CORDEX appears in mountainous regions like the Alps, the Pyrenees and the Scandinavian Mountains. Additionally, the variability increases from FUT1 to FUT3, especially in winter. The CRCM5-LE in contrast shows rather small variability in these mountainous regions.
The SDR mainly lies well below 1 in both seasons for most of Europe. This result suggests that for temperature, only a small part of the variability in the CORDEX ensemble can be explained by internal variability. For most parts, the SDR is smaller in summer than in winter. This is a result of two opposing effects: On the one hand, the overall CRCM5-LE variability is higher in winter; on the other hand, the CORDEX variability is smaller in winter. This mainly affects the British Islands, Scandinavia and other parts of Northern Europe, where SDR values can even exceed 1, especially in the early and mid-future. In these cases, the internal variability estimated from CRCM5-LE is larger than the CORDEX multi-model variability. While this result would be rather unexpected in a systematic framework, the current CORDEX imbalanced composition could lead to an underestimation of either (or both) RCM and GCM contributions to the total ensemble spread. In this context, it is not clear whether CRCM5-LE over- or underestimates the average internal variability of the CORDEX models. It is also not clear to which extent the true internal variability of the CORDEX ensemble is fully sampled by the available simulations.
A two-sample F-Test reveals the grid points with similar variances in both ensembles (Figs. 4 and 5, lowest rows). Empirical analysis shows that these are generally grid points with SDR values between 0.7 and 1.5 (similar to findings of Deser et al. 2012b). The share of grid points with similar variance decreases in both seasons for further future periods, with generally higher values in winter. Interesting to note is how the British Isles and parts of Norway fail the test for FUT1 because of the high variance in CRCM5-LE, showing similar variances for FUT2 and FUT3 with increasing CORDEX variability and relatively stable CRCM5-LE variability.
The precipitation signals are calculated for each member individually in percent change, so the standard deviation over these members is also expressed in percent change. In winter, the general patterns of CRCM5-LE and CORDEX are quite similar with higher variability in southern parts of Europe (Fig. 6). Northern Africa shows the largest standard deviations of relative changes, because the absolute sums of precipitation are relatively small here. A remarkable band of high variability stretches from the southeastern parts of Spain over coastal France to the Italian Alps in both ensembles. The variability in CRCM5-LE on the Iberian Peninsula is higher than in CORDEX, leading to high SDR values in FUT1. Other high SDR values can be found in Northern Africa and some mainly coastal areas in FUT1 and FUT2, whereas in FUT3 most of Europe shows SDR values around 1 and below.
Some differences in the variability of signals can be observed in summer between the ensembles, despite a general increasing West–East and North–South gradient. The variability in CORDEX is larger in all future periods and almost all of Europe (Fig. 7). Topography does not seem to be a significant factor in CORDEX, while CRCM5-LE displays especially low variability in mountainous regions, as already seen for temperature in winter. This results in SDR values around 1 in FUT1, decreasing to values well below 0.5 for FUT3. For FUT1 in summer, the number of grid points with similar variability is comparable to the number found in the map for winter (Fig. 6), but decreases significantly until FUT3 in contrast to the winter season. The respective figures for SON and MAM can be found in the SM, Figs. S8 and S9.
The pattern correlations as function of the time horizon between the standard deviation maps of both ensembles show two different behaviors for temperature and precipitation (Fig. 8). Correlations are generally high for precipitation as well as for summer and fall temperatures. For all precipitation seasons, a decrease of correlation can be observed, which fulfills the expectation of a decreasing contribution of internal variability on the overall variability in the further future.
Temperature seasons show a remarkable behavior. The pattern correlation increases in all seasons during the first half of the twenty-first century, and remains more or less stable in tas-DJF and tas-MAM, while dropping significantly for tas-JJA and tas-SON afterwards. The patterns of temperature thus do not seem to follow the expectation of decreasing correlation with time, and even increase for winter and spring. Further research is needed in this direction, especially if these results can be reproduced with other initial condition ensembles in the future.
To give an overview, Fig. 9 shows the regionally averaged SDR grid point values for all seasons and future periods for the subregions from Fig. 1. In general, the contribution of internal variability is much higher for precipitation than for temperature, and it decreases significantly the further the future period lies ahead. Annual and summer temperature/precipitation and fall temperatures have significantly lower SDR values than the other seasons. The share of boxes with at least 2/3 of the grid points confirming the F-Test on equal variances is especially high for temperatures in FUT1 (12 boxes) and precipitation in FUT1 (32) and FUT2 (15). Ratios above 1 mainly appear in FUT1 for spring temperatures and fall precipitation. The threshold of 2/3 is chosen on the basis of similar existing concepts like robustness of change as a function of the numbers of climate models agreeing in the sign of change (Jacob et al. 2014).