We begin with the DJF multi-model mean PSL responses for each single-forcing experiment and the all-forcings experiment; these are shown in Fig. . (The responses during the other seasons are presented in Fig. S1 in the Supplement.) Grid-points where at least 80 % of the models agree on the response sign are not stippled. Gray contours represent the standard deviation of the intermodel responses (scaled by in-line values and by the line-thickness). A robust anomalous ridge over the Mediterranean develops in hist-GHG and to a lesser degree in the all-forcings (historical) in DJF. More broadly, a positive North Atlantic Oscillation (NAO) response is also seen in both experiments in DJF. The effect of aerosols over Europe and the Mediterranean is somewhat weaker than that of GHGs and there is less agreement across models as to the sign of the response, however aerosols show an opposite signed response to GHGs nearly everywhere. These effects are similar if we average over different centennial-scale periods (Fig. S2).

Figure shows the observed DJF PSL response between the same periods (the observed responses for the other seasons are shown in Fig. S3). A meridional dipole response is evident in DJF, generally resembling the multi-model mean all-forcings and GHG-induced responses in the LESFMIP models (the pattern in the observations in DJF is similar if the latter epoch starts in 1980 or 1970 instead of 1990; not shown). However, the signals differ in magnitude (note the different colorbars for Figs.  and ), which is not surprising since the observed anomalies include both the forced response and also internal variability. In order to assess whether the modeled response is discrepant with the observed PSL response, we contrast in Fig.  the distribution of the DJF PSL response as obtained in the all-forcings LESFMIP experiment for all of the models' ensemble members (color-coded bars) with the observed DJF response (vertical thick black line), for two regions. For the Mediterranean region (Fig. a, solid red box on Fig. ) the observed response lies well within the LESFMIP historical distribution, and 36 % (143/396) of the ensemble members exceed the observed PSL response. If we instead focus on the region to the west of France in which the observed response is stronger (Fig. b, dashed red box on Fig. ), the LESFMIP multi-model mean shows a weaker response than observed, however 8 % (32/396) of the ensemble members nonetheless exceed the observed PSL response. Hence, there is no evidence for a model vs. observations discrepancy over this period.

The DJF multi-model mean precipitation response is shown in Fig. , the responses during the other seasons are presented in Fig. S4. Consistent with the hist-GHG and all-forcings PSL meridional dipole in Fig. , precipitation decreases in the Mediterranean region and increases in Northern Europe in response to GHGs, and to a lesser degree in all-forcings. Note that the overall structure of the precipitation responses are very similar in the all-forcings and hist-GHG experiments. The precipitation signals in the European sector extend westward towards North America, and include increased precipitation in Canada and reduced precipitation in the Caribbean. Aerosols, on the other hand, have an opposite signed response with increased precipitation over North Africa and reduced precipitation over Northern Europe, but there is generally less intermodel agreement than for GHGs, and in the Mediterranean itself the precipitation response is weak. These patterns are similar if we average over different centennial-scale periods (Fig. S5).

The responses in hist-GHG and hist-aer are larger than in any of the other single forcing experiments for both precipitation and PSL. In most regions aerosols and GHGs have opposite signed responses for both PSL and precipitation, and hence the historical all-forcings changes are relatively muted and are less robust across models than for GHG alone. The spread across the different models' means is quantified with the multi-model standard deviation gray contours. This spread is higher across the NA, Europe and the Mediterranean, relative to that across other regions. Intermodel spread in winter precipitation is particularly large near the coastline of North Africa than in other regions for both hist-GHG and hist-aer.

Comparing the simulated precipitation to observations is challenging due to the unreliability and limited nature of observed precipitation data in the early parts of the historical simulation. Nevertheless, more recent precipitation data is more reliable and extensive, and we therefore show in Fig. S6 trends in precipitation for each all-forcings LESFMIP model ensemble mean in the periods considered by . In agreement with , recent trends in precipitation (i.e., 1981–2014) indicate wettening in most of the Mediterranean (and not only there) with the majority of the models indicating a positive precipitation trend. Drying trends in most of Mediterranean in the historical simulations only emerge for start-dates of 1931 or earlier, however many individual models simulate wettening trends over Southern Europe (especially Southern France, Italy, and the Balkans) even for a start date in 1871.

Based on the above comparison of the LESFMIP response to external forcings and the observed PSL centennial difference, it may be suggested that the emergence of the ridging (and drying) signal across the Mediterranean out of the system's internal variability has been delayed by aerosols, which have canceled out a substantial portion of the GHGs-induced drying. That is, GHGs and aerosols have had a strong but opposite impact on historical wintertime climate in the Mediterranean, potentially explaining the lack of emergence in observations .

We now consider differences across the models in the response to forcings. In Table  we specify for each model, for each single-forcing experiment and for the all-forcings experiment, the ensemble-mean precipitation response (in percent; values are rounded to the closest integer) averaged over the Mediterranean region. For each grid-cell within this region the response in calculated by subtracting and dividing wintertime precipitation averaged over the time periods pr1990-2014-pr1851-1920pr1851-1920. In parenthesis are the averaged precipitation responses within that region but southward of 40° N. Bold values mark that at least 80 % of the ensemble members agree on the sign of the area-averaged precipitation response. Italicized values mark that at least half of the ensemble members agree on the sign (but less than 80 %).

Under the GHG forcing, the ensemble means of all models show a reduction in precipitation within the Mediterranean region between the historical periods, which vary in the range -3 % to -10 % between the models but with a large agreement in the overall trend. The drying is more pronounced south of 40° N (within the Mediterranean region) and reaches -15 %. Under hist-aer, some models indicate drying while others indicate wetting or negligible change; the responses vary in the range -8 % to 5 %. Southward of 40° N (within the Mediterranean), the responses tend to be more positive. Notably, the precipitation responses under the other single-forcings are smaller than the responses under GHG and aerosols, and indicate less agreement among models. Overall, the all-forcings (historical) experiment show a trend similar to that indicated by hist-GHG. In addition, it is evident that the single-forcings responses are not linearly additive compared to the all-forcings response.

As evident by the multi-model means (Sect. ), the responses to GHGs and aerosols are more significant than the responses to any of the other three forcings on these centennial timescales. Similarly, based on Table , the most pronounced precipitation responses in the single-model means are evident in the hist-GHG and hist-aer experiments. Therefore, we focus on these two in the rest of our paper.

There is a large intermodel spread in both the PSL and precipitation responses to GHGs and aerosols. This is demonstrated in Figs.  and , which show the DJF ensemble mean responses for each model separately in the hist-GHG and hist-aer experiments. We also show the zonal wind at 700 hPa (U700) and 2-meter temperature (T2m) fields due to their relevance to the indices discussed in Sect. . The comparable figure for all-forcings is in Fig. S7.

The response of U700 to GHGs in the North Atlantic differs qualitatively across the models. About half show an intensification of U700 across the UK and southern Scandinavia, but others show an opposite response. The ensemble means of all of the models, except that of CanESM5, show an increase in PSL across the Mediterranean region, but with variable intensity. About half of the models show a positive PSL response also across Europe and the NA. All of the models show a positive near-surface temperature response, except over the NA south of Greenland, hereafter referred to as the “warming hole” region . Four of the models show a cooling in this region of the NA (CESM2, CMCC-CM2-SR5, MIROC6, and MPI-ESM1-2-LR), and several more show a local minimum in warming (GISS-E2-1-G, NorESM2-LM). It is worth noting that the ensemble mean of the CanESM5 model shows the most significant warming response, in particular across the Atlantic, Africa, and Eurasia; this model is known for its particularly high climate sensitivity . Over most of the region presented in Figs.  and , the responses in hist-aer are opposite in sign relative to these in hist-GHG. The main exception is seen for HadGEM3-GC31-LL which shows a positive PSL response across almost the entire field shown, that is even larger than that in hist-GHG. Similar patterns are seen if we average over different centennial-scale periods (Figs. S8 and S9).

In this section we examine what large-scale climate or circulation response might explain the spread in the regional wintertime response of PSL and precipitation to identical forcing over the historical period. The first possible explanation is that each model's response is influenced by its time-mean bias, and we therefore show the time-mean PSL biases for each model's ensemble mean for the historical experiment in Fig. S10. The biases range between -10 and 10 mb across the NA, Europe and Mediterranean and are very similar, in each model, across the three temporal periods. This indicates that the models responses are substantially smaller then their bias. Most of the models show a meridional dipole bias pattern: negative across northern Europe and positive southward and across the NA. A somewhat different pattern is observed for GISS-E2-1-G, HadGEM3-GC31-LL, MIROC6, and MPI-ESM1-2-LR. Crucially, there does not appear to be a clear relationship between each model's bias and its PSL response. For example, although the PSL biases in the top 4 rows of Fig. S10 are very similar, their PSL responses (in particular for CanESM5) differ substantially (Fig. ). Our various attempts to use mean-state biases to “predict” the forced response in a given model were not successful (details excluded for brevity), which suggests that differences across the models in processes unrelated to the large-scale time-mean dynamical state are more important factors in explaining intermodel spread in the response.

So what factors can explain the spread in the responses? We answer this by analyzing the multi-model correlations of the Mediterranean region PSL responses and the responses of a group of climatic or circulation measures. Note that this analysis in not designed to determine mechanisms for the Mediterranean drying, rather to assess whether differences in the manifestation of these mechanisms across models might account for intermodel spread in the drying. The left (right) columns in the following figures refer to results for the hist-GHG (hist-aer) experiment.

The upper panels of Fig.  show scatter plots of the DJF PSL response versus the precipitation response. Each data point (color-coded dots) corresponds to an ensemble member of a specific model. The corresponding linear regression equation, the correlation coefficient, and its p-value, calculated based on all >240 individual ensemble members, are indicated in red text outside the red brackets. Stars mark the ensemble-mean responses of each model, and the correlation and p-value when using these ensemble-mean responses only are indicated in the square brackets.

Figure demonstrates the expected strong relationship between PSL and precipitation under both GHG and aer forcings, with a correlation coefficient of about -0.85 and a reduction of precipitation by about 7.5 mm d⁻¹ for every 1hPa increase in PSL. The fit is similar if we use the ensemble-mean response in each model only. However, there is large inter- and intra-model spread in the Mediterranean precipitation and PSL responses to the single forcings. The spread in the PSL response covers the range ∼(-2.5) to 3 mb under the GHG forcing and ∼(-2) to 3.5 mb under aer. Still, ∼ 78 % of the hist-GHG simulations show a positive PSL response and ∼ 74 % show a reduction of precipitation. Similar spread is indicated in the hist-aer experiment with ∼ 52 % of the simulations showing a negative PSL response and ∼ 32 % a precipitation increase. The dominant role of uncertainty in circulation responses for uncertainty in precipitation changes in this region is consistent with previous work . The rest of this section is devoted to understanding this intermodel and across-ensemble spread.

The middle row in Fig.  shows the PSL response vs. that of the land-ocean near-surface temperature difference (see Sect.  for definition). In both the hist-GHG and hist-aer experiments, a more pronounced relative land warming is associated with troughing over the Mediterranean. This is in contrast with the results of who found that a relative Mediterranean cooling (that is, a relative land warming) is associated with ridging and drying. It is worth noting that the CanESM5 model contributes substantially to this relationship: if excluded, the correlation coefficients reduce to -0.34 and -0.35 for hist-GHG and hist-aer, respectively, but these are nonetheless opposite to the relationship predicted by . Thus, under both single forcings experiments there seems to be a relationship between the PSL response and the response of the land-ocean temperature contrast, but it appears to be opposite to that suggested by previous work. We return to this in the discussion.

Next, we consider the relationship between the responses of Mediterranean PSL and of the NA Ocean temperature in the “warming hole” region defined in Sect. ; the corresponding scatter plots are shown in the bottom row of Fig. . For both single forcing experiments, a stronger NA warming hole is associated with a stronger Mediterranean ridge, which is consistent with results of previous studies . The correlation between the two responses is higher under the hist-aer experiment than under hist-GHG. In this case, excluding the CanESM5 model reduces the correlation coefficients to -0.43 and -0.71 for hist-GHG and hist-aer, respectively, but this is nonetheless highly statistically significant (p-values of 10-12 and 10-33 for hist-GHG and hist-aer, respectively).

In Fig.  we show the correlations between the PSL response and the responses of three large-scale climatic measures: global-mean near-surface temperature (top), stratospheric polar vortex (middle), and polar amplification of near-surface temperature (bottom). In both hist-GHG and hist-aer, a larger global mean warming is associated with troughing (and wettening, r=0.52) over the Mediterranean. This result is opposite to naive expectations: if the thermodynamic response was dominant, enhanced global warming should be associated with more sub-tropical drying . It is worth mentioning in this context that for the northern hemisphere the global-mean surface temperature response was found to be correlated only with the DJF Hadley cell edge response (but not with the DJF responses of the midlatitude jet or the subtropical dry zone edge) . Moreover, reported an absence of significant correlation between the global-mean surface warming and the mean Mediterranean precipitation response across the CMIP5 models. The negative correlation between the global-mean surface temperature response and that of PSL reported here is even larger if we focus on the trend dictated by the ensemble mean responses of the models, however, in most of the cases, the correlation is near-zero if we focus on the spread across each individual model's ensemble (not shown). The net effect is that intermodel spread in the precipitation response in the Mediterranean is not dictated by spread in thermodynamic processes in agreement with who used individual members from individual CMIP models.

According to the middle row of Fig. , a stronger stratospheric polar vortex is associated with ridging over the Mediterranean. This effect is consistent with previous work which associate weak stratospheric polar vortex events with negative NAO periods and a corresponding increase in Mediterranean precipitation . According to Figure , this relation is even stronger based on the ensemble-means than based on the individual ensemble members, and is stronger in hist-aer than in hist-GHG.

Next, we turn to the polar amplification of near-surface temperature (bottom row of Fig. ). Before discussing its relationship with the PSL response, we first clarify the sign of this metric for the two single forcing experiments. Under GHG (aer) forcing, both the pole and tropics warm (cool), but with a larger temperature change in the pole than the tropics. Hence, since the metric is defined as the ratio of the temperature changes, its sign is generally positive in both experiments. However, the slope of the relationship between the PSL response and polar amplification differ in sign between the hist-GHG and hist-aer. This sign difference originates from opposite dynamical responses in the two experiments. For hist-aer, as the pole cools more than the tropics (i.e., as the polar amplification index gets more positive; not shown here), a poleward shift of the jet might be expected and, as a result, a stronger ridge over the Mediterranean. Similar reasoning for the more pronounced polar warming in hist-GHG would explain the troughing trend: polar amplification leads to an equatorward jet shift and a trough over the Mediterranean. Hence, it is to be expected that the signs of the slope in the bottom row should be opposite, as indeed is evident in the bottom row of Fig. . Still, there is a substantial difference between the magnitude of the correlation coefficients for the two experiments, with a more profound relation for the hist-aer. Moreover, in this case too, the correlation between polar amplification and Mediterranean PSL is much more pronounced in the multi-model data and is much less evident in each individual model's large ensemble.

In Fig.  we analyze the relationship between the PSL response and that of the northern edge of the tropical belt using three metrics: zero-crossing latitude of the meridional stream function at 500 hPa, latitude of the subtropical jet, and latitude of the eddy-driven jet (first three rows from the top, respectively). For the hist-GHG (hist-aer) experiment, about 92 % (72 %) of the ensemble members show a poleward (equatorward) shift of the tropical belt edge according to the zero-crossing of the stream function metric. However, the correlation between this zonally averaged shift and the Mediterranean PSL response does not exceed 0.20. Even more striking is that the relationship in hist-GHG is opposite to naive expectations: members with a stronger poleward shift of the zero-crossing of ψ at 500 hPa, which corresponds to a stronger Hadley cell expansion, actually tend to have more of a trough in the Mediterranean . However, when considering each model ensemble separately (not shown), eight of the ten individual models indicate an opposite relation (with correlations between 0.26 and 0.67) which is more in line with expectations (a negative correlation is only found for two models, CESM2 and CMCC-CM2-SR5). For hist-aer, three of the ten models show a negative correlation.

According to the subtropical jet metric, ∼ 52 % of the hist-GHG and hist-aer ensemble members show a poleward and equatorward shift, respectively. There is a lack of a relationship between the PSL response and the latitude of the subtropical jet response, which is also supported by the responses of the individual models. In contrast to the two latter metrics, there is a strong relationship between the PSL response and that of the latitude of the eddy-driven jet (third row of Fig. ), both in hist-GHG and hist-aer: members with a stronger poleward jet shift in the zonal mean tend to have a stronger ridge. All of the individual models' ensembles support a positive correlation (although a single model in each of the hist-GHG and hist-aer experiments has a correlation that is smaller than 0.25).

In the bottom row of Fig. , we show the correlation between the PSL response with that of the averaged U700 in the Northern NA region where there is a high intermodel spread (region marked by green rectangle in Figs.  and ). The relationship between the Mediterranean ridge and the near-surface jet is even stronger than that indicated for the zonal mean eddy-driven jet latitude response. Specifically, a stronger U700 between Iceland and the British Isles, which corresponds to a poleward shift of the Atlantic jet, is highly correlated with Mediterranean ridging both in hist-GHG and hist-aer. This indicates that intermodel spread in the Mediterranean and in the Atlantic are linked. It is worth noting that this relation is somewhat reminiscent to that documented by between the 850 hPa wind over northern Africa and the Mediterranean precipitation, presumably through the Mediterranean ridge. And indeed, the highest U700 anomalies over the North Atlantic in Figs.  and are accompanied by the lowest U700 anomalies over northern parts of Africa.