Atmospheric Rivers in CMIP5 climate ensembles downscaled with a high resolution regional climate model

Atmospheric rivers (AR) are important drivers of heavy precipitation events in western and central Europe and often associated with intense floods. So far, the ARs response to climate change in Europe has been investigated by global climate models within the CMIP5 framework. However, their spatial resolution between 1 and 3° is too coarse for an adequate assessment of local to regional precipitation patterns. Using a regional climate model with 0.22° resolution we downscale an ensemble of 24 global climate simulations following the greenhouse gas scenarios RCP2.6, RCP4.5, RCP8.5. The performance of the model was tested against ER-I reanalysis data. The downscaled simulation notably better represents small-scale spatial characteristics which is most obvious over the terrain of the Iberian Peninsula where the AR induced precipitation pattern clearly reflect eat-west striking topographical elements resulting in zonal bands of high and low AR impact. Over central Europe the model simulates a less far propagation of ARs toward eastern Europe compared to ERA-I but a higher share of AR forced heavy precipitation events especially Norway where 60% of annual precipitation maxima are related to ARs. We find ARs more frequent and more intense in a future warmer climate especially in the higher emission scenarios whereas the changes are mostly mitigated under the assumption of RCP2.6. They also propagate further inland to eastern Europe in a warmer climate. In the high emission scenario RCP8.5 AR induced precipitation rates increase between 20 and 40% in western central Europe while mean precipitation rates increase by maximal 12%. Over the Iberian Peninsula AR induced precipitation rates slightly decrease around -6% but mean rates decrease around -15%. The result of these changes is an overall increased contribution of ARs to heavy precipitation with greatest impact over Iberia(15-30%). Over Norway average AR precipitation rates decline between -5 to -30 %. These reductions most likely the originate from regional dynamical changes. In fact, over Norway we find ARs originating from >60 °N are reduced by up to 20% while those originating south of 45° N are increased. Also, no clear climate change signal is seen for AR related heavy precipitation and annual maximum precipitation over Norway where the uncertainty of the ensemble is quite large.


Introduction
Atmospheric rivers (ARs) are long and narrow corridors that transport large amounts of moisture from tropical and subtropical origin poleward (e.g. Zhu et al., 1998;Gimeno et al., 2014;Gimeno et al., 2016;Shields et al., 2019). Associated with intense precipitation they can be a strong contributor to the local groundwater inventory in regions prone to droughts and thus play an important role for the local water management (Lavers and Villarini, 2015;Gimeno et al., 2016), for example in dry areas of the Middle East or North Africa . On the other hand AR induced heavy precipitation can likewise cause tremendous economical damage due to flooding (e.g. Gimeno et al., 2016;Payne et al., 2020).
Due to their intense moisture load they play an important role for the global water cycle. It has been estimated that ARs are responsible for >90% of the meridional moisture transport through the midlatitudes (e,g. Gimeno et al., 2016, Gimeno et al., 2018. ARs are associated with extraordinary strong low level winds often positioned at the head of a cold front of extra-tropical storm systems (e.g. Dacre et al., 2015;Gimeno et al., 2016). Accordingly, they are modulated by large scale weather regimes as demonstrated by Pasquier et al. (2019). In the North Atlantic sector the moisture source for ARs originates mainly from the subtropical Atlantic . ARs can occur during the whole year but due to their strong linkage to extra-tropical storm systems they are more frequent during the cold season in the Northern Hemisphere Ramos et al., 2015).
In the North Atlantic and North Pacific Sectors of the World Ocean ARs have been identified as a big risk for heavy precipitation and flooding along the western coasts of California and Europe (e.g. Ralph et al., 2006;Neiman et al., 2011;Ralph and Dettinger, 2012;, Lavers et al., 20122013;Ramos et al. 2015;Gao et al., 2016, Nayak et al., 2016Nayak et al., 2017). In Europe flooding causes large amounts of economic damages which is expected to increase under climate warming (Ashley et al., 2005;Sayers et al., 2015;Alfieri et al., 2016). Therefore, it is essential to investigate mechanisms that rise the risk for floods (e.g. Kouski 2014; Alfieri et al., 2018). For Europe a significant role of ARs in heavy precipitation events and flooding has been demonstrated by previous studies.  analyzed atmospheric reanalysis data sets and found up to 8 out of the 10 annual maximum precipitation events to be related to ARs during the period 1979-2011. Besides their potential to force flooding ARs can contribute significant amounts of precipitation to the annual total precipitation especially in semi-arid regions and thus can be an important source to the local water inventory in regions often threatened by droughts (e.g. Laver and Villarini, 2015;Gimeno et al., 2016). In Europe most damages associated with ARs are located along the western continental margins in particular over the Iberian Peninsula, the United Kingdom (UK) and Scandinavia (Laver et al. 2013, Ramos et al., 2015Whan et al., 2020). However they also can penetrate far inland and produce heavy rainfall events as far east as in Germany and Poland .
Because of the larger water holding capacity of a warmer atmosphere it has been suggested that climate warming will increase the risk for intense flooding (e.g. Held and Soden ,2006).  demonstrated that the intensification of the global water cycle due to climate warming strengthens the mean atmospheric water transports over the North Atlantic by 30 -40%. So far, assessments of AR related flood risk in a future warmer climate are primarily based on climate projections from global climate models (e.g. Warner et al., 2015Gao et al., 2016;Espinoza et al., 2018;Whan et al., 2020).  showed an intensification of AR in terms of frequency and intensity in future climate based on the analysis of five CMIP5 global models.  found a doubling of AR frequency together with an increase of moisture load in the RCP4.5 and RCP8.5 scenarios in six CMIP5 models at the end of the 21 st century compared to the historical period. Gao et al. (2016)  ensemble of 24 CMIP5 global models and found a pronounced increase in the contribution of AR related precipitation to the total annual precipitation in a future warmer climate following the RCP8.5 scenario. Whan et al. (2020) recently used the high resolution version of the CMIP5 EC-Earth model to study the climate change impact on AR induced precipitation over Norway. They indicated that up to 80% of the winter maximum precipitation is associated with ARs. They also found the magnitude of extreme precipitation events to be mainly controlled by AR intensity.

Purpose of this study
The majority of global models currently employed in the CMIP5 and CMIP6 framework employed provide a resolution typically ranging from 1.4° to 3° allowing a robust analysis of the large scale impact on continental scale precipitation. However, they do not fully resolve small scale characteristics which is necessary to assess the local impact on AR related precipitation patterns. Here, we investigate how far downscaled high resolution regional projections can improve the representation of ARs on the regional scale and thus add value to global assessments. For Europe such regional assessments are still lacking. Here we analyze for the first time a regional climate ensemble for Europe based on a horizontal resolution of 0.22°. We use this ensemble to analyze future changes in AR frequency and AR induced heavy precipitation pattern over Europe and their implication for the local water budget. We likewise aim to assess climate induced changes in the regional pathways of ARs on their journey across Europe. Finally, we assess uncertainties with respect to 3 different climate scenarios (RCP2.6, RCP4.5,RCP8.5) and 9 different global climate models from the CMIP5 suite.
The paper is outlined as followed: Section 2 briefly describes the regional climate model RCA-NEMO as well the AR detection procedure. Section 3 provides a validation RCA historical ensemble with respect to ARs for the present-day period and investigates the added value of high resolution. Section 4 analyzes future changes in AR frequencies and the impact on precipitation under different climate scenarios. Section 5 discusses uncertainties with respect to the choice of the driving global model. Main conclusions are summarized in section 6.

Methods
2.1.The regional climate model RCA-NEMO The regional climate model (Wang et al., 2015;Gröger et al., 2015;Dieterich et al., 2019) consists of the Rossby Center regional atmosphere model RCA (Samuelsson et al., 2011;Kupiainen et al., 2014) version 4 coupled to the Ocean General Circulation model NEMO (Table 1, Nucleus for European Modelling the Ocean NEMO, Madec, 2012). RCA is a hydrostatic atmosphere model which is set up for this study according to the Euro-Cordex domain (Fig. 1). The horizontal resolution is 0.22 degrees on a rotated grid yielding grid cell sizes between 550 -600 km² (Table 1) and the vertical discretization is given by 40 hybrid levels. At the lateral boundaries the model is driven by data from either reanalysis data sets or global climate model output. The forcing data are prescribed at 6-hourly time intervals. The land surface boundary is prescribed by ECOCLIMAP (Champeaux et al., 2005)and used to calculate the land -air mass and energy fluxes. Different to the majority of the EURO-Cordex high resolution ensembles (Jacob et al., 2014), RCA is interactively coupled to the 3D ocean model NEMO. The coupling area comprises the North Sea and the Baltic Sea. Over this region sea ice temperature, sea ice fraction, sea ice albedo, and water temperature is explicitly modeled by NEMO and communicated at 3-h time steps to RCA. Air -sea mass and energy fluxes are then calculated in the atmosphere model and used to drive NEMO which is set up in a resolution of 2 nautical miles (~3.7 km) and 56 vertical varying z* layers. The coupling is managed by the OASIS coupler (Valcke et al., 2003). Outside the coupled domain, i.e. the Mediterranean and the North Atlantic, RCA is driven by reanalyses data in the hindcast case or by global climate model output. Sea ice fields are explicitly modeled by the Louvain-la-Neuve sea ice model LIM3 (Vancoppenolle et al., 2008). Model  RCA-NEMO has been intensively validated and comprehensively described in detail (e.g. Wang et al., 2015;Gröger et al., 2015;Dieterich et al., 2019;Gröger et al., 2019;Gröger etal., 2021a). It has been employed in large ensembles to study the present climate and to simulate the mean response to global climate change by downscaling global climate scenarios in a huge ensemble (Dieterich et al., 2019;Gröger et al., 2019;Gröger et al., 2021a). Gröger et al. (2021a) showed that the RCA-NEMO coupled ensemble is well within the range of the high resolution Euro-Cordex ensemble (Jacob et al., 2014). However, significant differences arise over interactively coupled areas over sea (Gröger et al., 2021a;Gröger et al., 2021b). This applies to both climatic mean changes, as well as climatic extremes (e.g. dry periods cold spells, heat waves etc).

The high resolution ensemble
The above described model was used to downscale a suite of global model climate scenarios taken from the Coupled Model Intercomparison Project phase 5 (CMIP5, Taylor et al., 2012). Table 2 lists the individual realizations and primarily distinguishes the applied scenarios (first row) and individual model configurations (first column). In addition to the used global climate models we also analyze a hindcasst simulation forced by ERA-I reanalysis data to validate the model for the present day climate    The chosen climate scenarios follow the protocol of Representative Concentration Pathways (RCP) used in CMIP5 and comprise three different greenhouse gas assumptions. One low emission scenario assumes vigorous mitigation actions (RCP2.6 van Vuuren et al. 2007Vuuren et al. , 2011 and was developed in regard to limit the global mean temperature to 2 °C warming since the pre-industrial period. It assumes negative emission during the last decade of the 21 st century. RCP4.5 is a moderate scenario where emissions peak at the mid-century (2040) and are kept constant after ~2080 at a value about half of the value at the end of the historical period (Clarke et al. 2007;Thomson et al. 2011). Finally, a totally unmitigated scenario RCP8.5 (Riahi et al. 2007;Riahi et al., 2011) assumes rising emissions up to the end of the century. The three scenarios impose a maximal radiative forcing of 2.6, 4.5, and 8.5 W/m2 compared to pre-industrial conditions.

Detection of atmospheric rivers
The detection procedure of ARs is briefly described below and was performed for 30 year periods at the end of the 20 th century  and at the end of the 21 st century (2070-2099) i.e. for each of the RCP2.6, RCP4.5, and RCP8.5 individually. Only the RCA-ERAI hindcast run was done for 1979 to 2009.
Atmospheric river detection can be basically done using their specific characteristics, namely the extraordinary moisture transport and their elongated shape (width and length scales). A number of studies addressed methods to detect ARs in model simulations and gridded reanalysis data Lavers et al., 2012;Nayak et al., 2014;Nayak and Villarini, 2016;Gao et al., 2015;O'Brian et al., 2020). An overview can be found in Guan and Waliser (2015). We here employ the detection algorithm developed by Laver et al. (2012) and  which has been successfully applied both for hindcast simulations and climate studies (Lavers and Villarini, 2012;. First, the vertically Integrated atmospheric water Vapor Transport (hereafter IVT) is calculated at every model output time step. The vertical integration is done over the models pressure levels from 1000 to 300 hPa that were converted from the models hybrid levels: where g is gravitational acceleration, q is specific humidity kg/kg, u and v the horizontal wind components and dp the pressure level difference of adjacent pressure levels. In the two hydrostatic models' hybrid level space, the IVT is balanced by precipitation minus evaporation.
Then the 85 th percentile IVT is calculated based on IVTs at 12:00 UTC time stamps and along 10 °W longitude (Lavers and Villarini 2012;). Following  this procedure is done for meridional bins of 5° between 35-70°N. The resulting 85th percentile values served then as threshold for the detection of ARs (Fig. 2). After this pre-processing all 6-hour IVT fields are analyzed along 10 °W and the 5 ° latitudinal bins at every single output time step. If the max IVT within an individual bin exceeded the threshold for that bin ( Fig. 2) a westward and eastward search was done starting from 10°W to 30°W and 25 °E. All grid cells where the threshold was exceeded were retained .  (1970-1999, upper left) and RCP climate scenarios  at 10°W used by the algorithm to track ARs..
Next, the resulting fields are further evaluated for spatial and temporal criteria . Hence the axis of a potential AR was determined as maximum along each longitude IVT of the structure. Only those fields were retained with an axis longer than 1500 km and then classified as AR. Furthermore, an AR has to be persistent over a period of at least 18 hours. Figure 3 shows a prominent example for an AR detected in the ERA-I reanalysis (left) and the ERA-I hindcast simulation (right) that demonstrably caused intense rain over Europe . The whole detection procedure is performed separately for the historical and future periods and for each of the ensemble members respectively (Table 2).  Table 3 summarizes the frequency of atmospheric rivers (expressed as total ARs detected in a 30year period and along 10°W) for each run of the ensemble as well as from the ERAI reanalysis data set and the hindcast run. The number of detected ARs in the RCA-ERAI hindcast (322) is nearly identical with that analyzed from the ERAI reanalysis (321) itself which was used to drive the model. This indicates that the number of ARs in RCA is primarily controlled by the lateral boundary conditions. This is not surprising since ARs develop in open ocean regions far outside the model domain and so the potential for alterations by the regional model is quite low. According to this, the number of detectable ARs in the individual RCA historical climate simulations (   Figure 4a compares the moisture transport by ARs over land which indicates the potential to force local heavy precipitation events. We note that the RCA-ERAI run has a lower moisture content over land compared to ERAI (~5%, Figure 4a). This is in line with the models cold bias in air temperature (Gröger et al., 2021a) which favors lower moisture contents. The probability distribution of diagnozed AR durations ( Figure 4b) indicates no systematic differences between the ERA-I reanalysis, the hindcast run or the mean historical climate simulations. For all model realizations about half of the detected AR last for less than one day (Fig. 4). Noteworthy is the lower moisture content of nearly the entire RCA historical ensemble compared to the RCA-ERAI hindcast simulation. Overall, this points to a systematic negative bias in moisture contents over land in the RCA model.

Impact on precipitation
In a first step we estimate the likelihood for a certain region to be affected by an AR. As a simple index, we count the total number of days during which a land cell was covered by an AR (AR days, Figure 5a). As expected, ARs are most present over the UK and the coastal regions of western Europe. Further inland the AR imprint declines as ARs lose moisture due to rainfall and thus do not further meet the IVT threshold (Fig. 2). Strong moisture loss is also indicated along the Norwegian coast where the landfall of ARs cause strong rain events due to uplift.
Next we quantitatively evaluate the imprint of ARs on extreme precipitation by calculating the portion of yearly maximum of daily precipitation that is caused by ARs. The result is shown in Fig. 5b which shows the percentage of yearly maximum daily precipitation rates forced by ARs. ARs explain up to ~60% of the yearly maxima precipitation rates over southwestern Norway. A strong imprint is likewise seen over western UK and along the European coasts. Besides their potential to force yearly precipitation maxima ARs strongly affect the fraction of heavy precipitation events. Here we investigate the AR contribution of precipitation to the total annual precipitation ( Figure 5d) and to the fraction of heavy precipitation (Figure 5c). The spatial pattern 10 mainly resembles the pattern seen in the AR days (Figure 5a) but further reflects the mean climatic conditions in Europe. In humid regions of central and western Europe, AR related rain contributes up to 40% (Bretagne) to the >95th percentile precipitation. These are also the regions which are frequently affected by ARs (Fig. 5a). In semi-arid regions like the Iberian Peninsula the contribution increases to almost 60% which compares well with results from gridded reanalysis data sets (Lavers and Villarini, 2015). A considerable contribution is likewise seen along the western coast of Italy. Here AR contribution is up to 30 % (Fig. 5d) although ARs do not notably cause annual maximum precipitation rates (Fig. 5b). In the topographically elevated regions of Norway and over the Alps, i.e. regions supporting often strong convective rain events, the influence of ARs is also less pronounced. As expected in more humid regions of eastern Europe and Scandinavia ARs contribute only minor amounts to the annual total. A similar pattern is shown for the contribution to the total annual precipitation (Figure 5d). Maxima contributions are seen over western France and the western Iberian Peninsula where ARs contribute up to 10% to the total precipitation.

Comparison of the RCA ensemble mean with the ERAI hindcast and ERAI reanalysis
In the historical simulations, i.e. when the model is forced by climate output the regional coupled models develop their own weather regimes which can not be expected to be in phase with recorded weather data. Therefore, individual AR incidences can be analyzed only statistically which is done in terms of calculated indices during the climatological historical period (Fig. 5).
We now compare the results of the RCA historical ensemble mean ( Figure 5, left column) with the RCA-ERAI hindcast ( Figure 5, middle column) and the ERAI reanalysis data Figure 5, right column) in which internal natural variability can be expected to reflect the observed weather. Figure 5 demonstrates that for most of the above described indices the RCA historical ensemble mean represents reasonably well the spatial pattern found in the RCA-ERAI simulation. The spatial correlation coefficients between RCA-ERAI and RCA-ENSM are respectively calculated as 0.98 for AR the frequency (Fig. 5a), 0.82 for the percentage of yearly maximum precipitation (Fig. 5b), and respectively 0.92 for contribution to total annual and heavy rain precipitation amount (Fig.  5c,d).
Besides the spatial pattern seen in the AR indices, the detected ARs also undergo a strong seasonal cycle. Figure 6 shows that ARs are most abundant during fall and early winter. A pronounced difference is, however, seen in August where the relative share is about twice as much in the RCA historical ensemble compared to the ERAI hindcast which implies that the AR season starts a bit too early in the RCA ensemble. However, all in all we find that spatial patterns and the seasonal cycle of ARs is well preserved in the models climate mode compared to the ERAI data set and the ERAI hindcast.

Effect of regionalization
The above described spatial characteristics is also well reflected in the ERAI reanalysis data (Fig.  5, left column). However, the lower spatial resolution of this data set (approx. factor of 10, Table 1) destroys much of spatial variation which demonstrates the effect of the downscaling by RCA. This is more visible in the precipitation related indices (Fig. 5b-d) than in the AR frequencies (Fig. 5a). This is expected as precipitation patterns are modulated by stochastic processes and further modulated by topography while the integrated atmospheric moisture transports (used to detect and track ARs) are less affected by small scale patterns.
The added value of regionalization is demonstrated by the comparison between the high resolution RCA-ERAI run with a resolution of 0.22° and the corresponding ERAI-data set with a resolution of 0.75 °. First, the downscaled RCA-ERAI run shows notably higher AR frequencies and larger contribution to the total precipitation (and >95 th percentile precipitation) along the coasts compared to the original ERAI data set (Fig. 5a,c,d). In turn, AR frequencies in the distal parts of eastern Europe are higher in the reanalysis data than in the ERA-I hindcast run. This implies ARs tend to penetrate less far inland after landfalling in the downscaled run. Apart from this, the downscaling effect is most pronounced in regions with elevated topography as well as in in semi aride climate zones of southern Europe. Over Iberia the RCA-ERAI run clearly resolves the distinct effect the of the west -east striking topographic features seen in the contributions to the precipitation budgets.
They occur as small WSW-ENE striking bands of alternating high and lower AR precipitation following the topographic elements build up by the Sistema Central Plateau, the Sierra Morena mountains, and the Penibaetic orogenic system (Fig. 5c,d). Furthermore, the local maxima of AR contributions seen along Italy are by far less pronounced in the coarser ERAI reanalysis (Fig. 5c,d).
Apart from this, we also find a notably lower amount of annual precipitation maxima related to ARs 12  . 5b). Over Norway the difference can be as large as 15 %. Here, the lower resolution likely smooths heavy precipitation events.
Another slight but noteworthy RCA-ERAI vs ERAI difference is seen in the seasonal distribution of ARs. In the RCA-ERAI hindcast the fall maximum is already registered in September while in ERAI it is recorded for October. The most likely explanation for this is a temperature bias introduced by RCA. Gröger et al., (2021a) showed that RCA has a negative temperature bias against the observational E-OBS data set. This means that air masses enter the lateral model boundary with a thermodynamically too high moisture load. The adaption, i.e. the loss of moisture when the air masses are confronted which the mean thermodynamics state of the inner RCA-domain, probably takes longer for fast moving air masses like within ARs. This would favor the detection of too many ARs.
Finally, we note that the global CMIP5 models used to drive RCA have an even lower resolution ranging between 1.4° (CNRM) and 3° (CAN) which is notably lower than the ERAI reanalysis data set. This implies that the added value of regionalization in the future scenarios can be expected of even greater importance. Figure 7 summarizes the relative change in average moisture transport by ARs according to the different greenhouse gas scenarios and for each of the downscaled global models. Consistently, ARs become more intense i.e. they have a higher moisture load in a warmer climate. The intensity at the end of the century increases on average by 9% (RCP2.6), 13% (RCP4.5), and 24% (RCP8.5) which is more or less in line with the corresponding increases of IVT thresholds (Fig. 2).

General response of AR frequency and intensity
Besides intensity, also the frequencies of detected ARs increases (RCP2.6=8.6%; RCP4.5=18.6%, RCP8.5=24.1%, Table 3) which are roughly proportional to the increase in intensity. However, not only the frequency of ARs increases but also the spread of the individual realizations at the end of the century. The relative change in ensemble spread (Table 3, 2nd row) increases even more than the ensemble average (RCP2.6=16.5%, RCP4.5=23.6%, RCP8.5=56.0%). This highlights the large uncertainty with respect to the chosen global model used for downscaling. Some advanced approaches for weighted model averaging were developed to reduce this type of uncertainties and have been tested for the case of AR over the US (Massoud et al., 2019;Massoud et al., 2020, Wootten et al., 2020.

Spatial changes
Next we analyze spatial pattern changes in future AR imprint. First, AR frequencies are addressed. Figure 8a demonstrates an overall increase in the frequency of AR days which is strongest over in the southwestern sector of the Atlantic (Biscay) and adjacent land areas with maximum response over western France and the southern UK. This pattern is more or less consistent across the RCP scenarios but differs in strengths ranging from ~+20% to more than 200% (Fig. 8a) We now investigate how the increased AR frequencies influence precipitation patterns. In a first step the impact on the AR forced yearly maximum precipitation is addressed (Fig. 8b). The most robust change is the strong increase over the western central part of Europe which extends from western France along the coast of Belgium, the Netherlands northern Germany and Denmark up to the southern coast of Norway. Further spots of stronger AR impact are also visible along the northwestern tip of the Iberian Peninsula and the southern part of the UK. This general response is by far strongest in the unmitigated RCP8.5 scenario. In the moderate scenario RCP4.5 the changes are less pronounced in eastern central Europe (Germany, Denmark). In the mitigation scenario RCP2.6 notable robust changes are restricted to a small area in NW France (Bretagne, Normandy). Over southern Scandinavia where the highest values are found during the historical period (see Fig. 5b) no robust changes are detected.
Next we address the number of AR forced heavy precipitation events. Figure 8c shows the number of events nearly everywhere increases. The strongest response is seen over southern Scandinavia where the relative increases exceed 300%. Over the western European continent and the UK the average increases are in the range between 75-150% pointing to a roughly doubling risk for flooding.
Apart from this very strong increases (>300%) are noticed over eastern Europe (Figure 8c). Although in this region absolute AR occurrences are rather low, the strong relative increases indicate a further eastward propagation of AR under the warmer future climate compare to present day.
14 Figure 7: Relative change in moisture transport within ARs at the end of the century  compared to the historical period  The higher AR frequencies and moisture loads have also consequences for the local precipitation budget. Figures 8 displays the changes in the contribution of ARs to the heavy precipitation fraction (Fig. 8d) and the total annual precipitation (Fig. 8e). Most pronounced changes are seen in regions where already under historical climate condition large contributions are seen (Fig. 5c,d). Strongest increases are primarily seen over Iberia with changes up to +30% to the heavy rain precipitation and up to 20% along the French coast (Bay of Biscay) compared to the historical period.
At least for the moderate and high greenhouse gas scenarios RCP4.5 and RCP8.5 the contributionanomalies are robustly positive (Fig. 8d,e). The positive anomalies imply that in these regions the AR induced precipitation rates increase stronger than mean precipitation rates. The different response of mean precipitation and AR induced precipitation to climate change is depicted in Figure 9. The climatological mean change (Figure 9a) shows the typical change in precipitation over Europe pointing to dryer conditions over southern Europe and wetter conditions over northern Europe (e.g. Jacobs et al., 2014; Kjellström et al., 2018;Teichmann et al., 2018;Gröger et al., 2020a;Christensen et al., 2021). Hence, mean precipitation rates increase only slightly by up to maximal 12 % over central Europe or even decrease over southern Europe. By contrast, AR induced precipitation increases between ~25 -40% over central Europe (Fig. 9b). In fact AR induced precipitation ranges are in most regions equally high as mean winter precipitation changes (not shown), i.e. the season where mean changes are highest. Also decreasing AR precipitation is found in southern Europe but the reductions are less strong compared to mean rates. The only exception is Norway where in fact AR induced precipitation decreases while mean rates increase. This likewise explains the low response of annual maximum changes (Fig. 8b). 16

Influence of potential dynamical changes
In the mid to high latitudes it has been found, that climate changes in mean and extreme precipitation takes place at similar magnitudes and mainly reflect thermodynamical processes (i.e. increasing water carrying capacity with warming air, e.g. Emori and Brown, 2005). Exceptions from this occur only when large scale circulation changes with further impact on moisture transport are involved (Emori and Brown, 2005). Based on global CMIP5 models Gao et al. (2016) suggest that thermodynamics is also the main driver AR related precipitation changes but may be modulated by dynamical circulation changes due to changes in the Jet position.
In order to investigate potential dynamical changes we now elaborate more about the pathway of ARs on the their way from the open ocean to the continent. It is not feasible to analyze the movement of every single AR over time. Instead, we here perform a simple analysis to determine the AR source region for every AR. Hence we check for every AR incident registered on land at which latitude the AR entered the European sector along 10°W meridian. We here consider the latitudinal bins south of 45°N, 45-60°N, and north of 60°N. From this statistic we calculated for every land cell the percentage share of AR incidents for every latitudinal band (so that the sum of all bins at every land point sums up to 100%). The analysis is done for high emission greenhouse gas scenario RCP8.5 compared to the historical period. Figure 10 shows the change in the relative contribution for the latitudes south of 45 °N (Fig. 10a) and north of 60 °N (Fig. 10b). Overall, the RCA ensemble clearly shows a relative increase of those ARs originating south of 45 °N degree (Fig. 10a). Most prominent increases are seen over the Alpine region and Scandinavia whereas western and central Europe are less affected. Between 45 and 60 °N the changes are everywhere below 5% (not shown) indicating a more or less unchanged contribution within this latitudinal band.
By contrast, AR contributions from >60°N are diminished (Fig. 10b). Over Scandinavia ARs from >60°N contribute around 70-90 % to the total AR events during the historical period (not shown). This contribution is reduced by ~20% in RCP8.5 (Fig.10b) and thus likely explains the aforementioned 17 Figure 10: Analysis of AR origin in the RCA ensemble. a) Relative contribution of regional AR occurences originating from the Atlantic north of 60 °N. b) same as a but for ARs originating from south of 45°N. Shown is the change for RCP8.5 (2070RCP8.5 ( -2099RCP8.5 ( minus 1970RCP8.5 ( -1999. dereased AR precipitation rates (Fig. 9b). Hence, for Scandinavia this implies a profound southward shift in the origin of ARs that cross Scandinavia. In turn, this shift suggests that the moisture has to be transported over a longer distance across the central continent compared to those ARs originating from >60°N.
In summary, we can conclude that ARs from the southern Atlantic sectors are more present over most land regions in a warmer climate. By contrast, ARs arriving from the northern sectors of the Atlantic are relative diminished over land (Fig. 10b).

Uncertainties with respect to the global driving CMIP5 models
We now investigate individual ensemble members for a selected set of indices, namely, the changes in AR days frequency, the annual maximum precipitation, and the contribution of ARs to the yearly total heavy precipitation. The response of AR forced annual maximum precipitation events is shown in Figure 11b. No clear consistent response is registered over the topographic elevated region of western Norway i.e. the region which is most impacted by ARs in the historical climate (Fig. 5b). Some models show distinct locations with decreased AR impact over Norway (e.g. RCA-MPI RCA-GFDL, RCA-IPSL) which is probably linked to the aforementioned decrease of ARs arising from >60°N. The most coherent change across the realizations is the more or less strong increase over the western France which in some realizations extends further to the east. However, also in this region local ensemble variability is pronounced and in RCA-ECE even a decrease is seen. The pronounced ensemble variability is not unexpected because the yearly maxima represent the most heavy precipitation events over a certain region.
Uncertainties with respect to the contribution to the heavy precipitation budget is highest in the region of largest changes (Fig. 11c). High inter-model variance is found for the southern tip of Iberia indicated by maximal standard deviations. In this region the contribution can either be reduced as in RCA-HAD or increased as high as +50% as in RCA-CNRM. The same change pattern is likewise seen in Italy but with a lower magnitude. Higher uncertainties are likewise found for central Europe In summary we can conclude that the ensemble members robustly agree on a stronger future AR impact on heavy precipitation over western central Europe with a hot spot over NW France and further extension eastward depending on the realization.

Differences to global projections
Our results generally agree with results from global CMIP5 models (e.g. Gao et al., Lavers et al., 2016; that ARs become more frequent and intense. With respect to previous studies we note that the climate change effect on AR frequency strongly depends on the chosen reference period. Studies that employ the 85 th percentile threshold derived from the historical period likewise for the future period often find a doubling of AR frequency (e.g. . However, we here calculated separate 85 th percentile thresholds distinctively for historical and future periods. This was done because the 85th percentile threshold as suggested by Lavers et al. (2012) represents the median IVT value within observed ARs during the historical period (Neiman et al., 2008). Consequently, the AR frequency increase found in this study is lower (as IVT thresholds for the future period are higher, Fig. 2) than in the aforementioned studies but still amount to +20-30% increase across the models.
Main differences to global projections arise over Norway and the Iberian Peninsula, i.e. two hotspots of AR impact in Europe. Over Iberia the distribution of AR related heavy precipitation is clearly modulated by topographic structures, like the Sistema Central Plateau, the Sierra Morena mountains, and the Penibaetic orogenic system. These valleys and ridges lead to zonal bands of high and low increases of AR precipitation over Iberia in the future. Furthermore, our regional ensemble does not indicate a robust climate change signal over Norway. Global CMIP5 models indicate for this region an increase of 10 to 20% in the AR contribution to heavy precipitation events in RCP8.5 (Gao et al.,2016, Fig. 9. therein). In contrast to this, the downscaled ensemble indicates only a weak and not robust response over Norway. The change in the AR precentage of annual maximum preciptation can be negative or positive depending on the global model and thus does not indicate a clear signal (Fig  11b). Likewise, the contribution to heavy precipitation is in all regional ensemble members below 5% with the exception of the southwestern top of Norway (Fig 11c).
Our finding that ARs from south of 45°N have an increased presence over Europe in the future high emission scenario points to larger scale atmospheric circulation changes. Such change could be related to changes in the low level Jet stream (Gao et al., 2016) and/or indicate systematic changes in regional weather systems (Pasquier et a., 2018). At least in the southern Hemisphere a poleward shift of ARs has been reported for the recent decades (Ma et al., 2020). The authors suggested internal variations of basin scale sea surface temperatures as likely reasons.

Summary and Conclusions
A high resolution regional climate model with a resolution of 0.22° was applied to investigate the impact of ARs on Europe. The added value of regionalization was analyzed by a hindcast run to downscale the ERA-Interim reanalysis data set with a resolution 0.75°. The added value is most obvious in topographic elevated areas and in the semi arid climate zone of southern Europe (Iberia, Italy). In the central and southern part of the Iberian Peninsula the contribution to the regional precipitation budget is strongly modulated by E-W striking topographic signatures. This feature is not seen in ERA-I reanalysis data which shows on the contrary N-S striking gradients with highest precipitation in the west. Generally the AR imprint on analyzed indices in the ERAI data set is lower near the western European coasts but more visible in the distal parts of eastern Europe. This indicates that ARs penetrate less far inland in the downscaled simulation compared to ERAI.
The regional climate model is further used to investigate the dynamics of ARs in present and future climate. For this in an ensemble of in total 34 simulations was carried to downscale global climate model scenarios from the CMIP5 suite with a coarse resolution between 1.4 -3°. The models were used to downscale three greenhouse gas scenarios (RCP2.6, RCP4.5, and RCP8.5).
The historical simulations from the regional climate model ensemble are in good agreement with global ERAI reanalysis data set and an ERA-I simulation hindcast run. The regional climate ensemble shows strongest AR impact along near-coastal regions explaining up to 60% of yearly maximum precipitation rates in regions with orographic uplift (e.g. Norway). In regions with coastlines exposed to the North Atlantic but unaffected of orographic uplift AR related rain events constitute a significant contribution to the total annual yearly rain rate. Here the impact is largest in semi-arid regions of the Iberian Peninsula and along western Italy.
Our results show ARs become more frequent and have a higher moisture load in a future warmer climate (e.g. Warner et al., 2015;Gao et al., 2016;Shields and Kiehl, 2016;Shields et al., 2019;Massoud et al., 2019;Massoud et al., 2020;Whan et al., 2020). The potential of ARs to force annual maximum precipitation events increases most prominently over western France (Brittany) and northernmost Spain by up to 20% (RCP8.5). No robust ensemble response was found over Norway.
Our regional high resolution model allows a spatially more accurate calculation of AR contributions to the local water budget than possible with global earth system models. Over the Iberian Peninsula and western France the contribution of AR forced rain to the total annual precipitation increases by up to 10% as well as it does with respect to the total annual precipitation that falls as heavy precipitation (up to +30%). Furthermore, the well known dryer mean climate conditions in southern Europe (e.g. Jacob et al., 2014;Kjellström et al., 2018;Gröger et al., 2021a;Christensen et al., 2021) at the end of the century favors a more important role of ARs in the future climate as AR precipitation rates increase stronger than mean precipitation rates. The increased AR contribution found for the Iberian Peninsula under contemporaneous climate drying is likely to influence the ground water recharge which is essential for maintaining ecosystem services and supporting agriculture in this area (e.g. Martos-Rosillo et al., 2015).
In particular for northern Europe we find a more southern position in the origin of ARs in the future warmer climate compared to the historical period. In turn, this likely affects the path of ARs before arriving Scandinavia as moisture travels over a longer distance over land. This leads to locally decreased precipitation rates over Norway in RCP8.5.
Our study clearly demonstrates a larger imprint of ARs in Europe on the regional scale and a more dominant role of forcing heavy precipitation events with potential risk for flooding under the higher greenhouse gas scenarios RCP4.5 and RCP8.5. However, under the assumption of the greenhouse gas scenario RCP2.6 most of the climate induced changes are not robust which points to the potential benefit of climate mitigation actions. Finally, the present assessment of AR dynamics in regional ensembles for Europe must be considered as a first step since the realized horizontal resolution is still coarse (24 km) and not explicitly resolves convection. The future trends in regional high resolution modelling will allow resolutions of only a few kilometers and convection permitting models will be applied (e.g. Giorgi, 2019;Jacob et al., 2020). This will allow a more thorough investigation of the processes mediating the response of ARs to climate change and their pathway across Europe.