Investigating the sensitivity to resolving aerosol interactions in downscaling regional model experiments with WRFv3.8.1 over Europe

In this work we present downscaling experiments with the Weather Research and Forecasting model (WRF) to test the sensitivity to resolving aerosol-radiation and aerosol-cloud interactions on simulated regional climate for the EURO-CORDEX domain. The sensitivities mainly focus on the aerosol-radiation interactions (direct and semi-direct effects) with 4 different aerosol optical depth datasets (Tegen, MACv1, MACC, GOCART) being used and changes to the aerosol absorptivity (single 5 scattering albedo) being examined. Moreover part of the sensitivities also investigates aerosol-cloud interactions (indirect effect). Simulations have a resolution of 0.44o and are forced by the ERA-Interim reanalysis. A basic evaluation is performed in the context of seasonal-mean comparisons to ground based (E-OBS) and satellite-based (CMSAF SARAH, CLARA) benchmark observational datasets. Implementation of aerosol-radiation interactions reduces the direct component of the incoming surface solar radiation by 20-30% in all seasons, due to enhanced aerosol scattering. Moreover the aerosol-radiation interac10 tions increase the diffuse component of surface solar radiation in both summer (30-40%) and winter (5-8%) whereas the overall downward solar radiation at the surface is attenuated by 3-8%. The resulting aerosol radiative effect is negative and comprises of the net effect from the combination of the highly negative direct aerosol effect (-17 to –5 W/m2) and the small positive changes in the cloud forcing (+5 W/m2), attributed to the semi-direct effect. The aerosol radiative effect is also stronger in summer (-12 W/m2) than in winter (-2 W/m2). We also show that modeling aerosol-radiation and aerosol-cloud interactions 15 can lead to small changes in cloudiness, mainly regarding low-level clouds, and circulation anomalies in the lower and midtroposphere, which in some cases can be statistically significant. Precipitation is not affected in a consistent pattern by the aerosol implementation and changes do not exceed ± 10% . This result is in contrast to other regional downscaling studies investigating either aerosol-radiation or aerosol-cloud interactions. Temperature, on the other hand, systematically decreases by -0.1 to -0.5oC due to aerosol-radiation interactions with regional changes that can be up to -1.5oC. 20 1 https://doi.org/10.5194/gmd-2019-161 Preprint. Discussion started: 19 August 2019 c © Author(s) 2019. CC BY 4.0 License.


Aerosol data
We use two external aerosol datasets. The first is the Max-Planck-Institute Aerosol Climatology version 1 (MAC-v1) (Kinne et al., 2013). The MAC-v1 is a global climatology of aerosol that has been produced by combining global aerosol models and ground-based measurement by sun-photometer networks. Aerosol optical properties are provided on a global scale at a spatial resolution of 1 degree. Monthly data regarding total, as well as anthropogenic aerosol properties, are available ranging from 5 preindustrial times to the end of 21 st century. We use a part of this climatology that contains the merging of monthly statistics of aerosol optical properties to describe current conditions. The second dataset used is the MACC reanalysis (Inness et al., 2013). Data are provided globally at a horizontal resolution of about 80 km for the troposphere and the stratosphere (Inness et al., 2013). An advantage of the MACC dataset is its daily resolution. A study that tested different climatologies (Mueller and Träger-Chatterjee, 2014), including MAC-v1 and 10 a climatology based on the MACC reanalysis concluded that the MACC climatology leads to the highest accuracy in solar radiation assessments.
The new Thompson aerosol-cloud interacting cloud microphysics scheme has an internal treatment of aerosols. Aerosols are separated into cloud droplet nucleating acting as cloud condensation nuclei (CCNs), and cloud-ice nucleating, acting as ice nuclei (IN). Cloud-droplet nucleating aerosols include sulfates, sea salt and organic carbon. Cloud-ice nucleating aerosols 15 include dust larger than 0.5 µ m. Black carbon is not included. Aerosol initialization and boundary conditions are based on an aerosol climatology constructed from global simulations spanning the period 2001-2007(Colarco et al., 2010 with the use of the Goddard Chemistry Aerosol Radiation and Transport (GOCART) model (Ginoux et al., 2001). The two categories of aerosols are then advected and diffused during the model run. Furthermore, a field representing cloud-droplet nucleating surface aerosol emission flux is introduced to the lowest model level at each time step. Surface emission flux is based on initial 20 aerosol concentrations at the surface and on a constant value of mean surface wind. The aerosols can be allowed to interact with radiation (aer_opt=3), enabling aerosol-radiation interactions in addition to the existing aerosol-cloud interactions, thus providing a complete representation of aerosol interactions.

Model Simulations
Using the above aerosol options and datasets we performed 7 sensitivity experiments from a control run with no aerosol 25 interactions covering the period 2004-2008.
-The control experiment (CON) does not include aerosol-radiation or aerosol-cloud interactions (aer_opt=0), meaning the simulation is aerosol-insensitive.
The next four experiments, also only account for aerosol-radiation interactions and use the methodology introduced by 30 Ruiz-Arias, Dudhia, and Gueymard (2014) (aer_opt=2): aerosol type.
-ARI_Mv1full uses AOD, single scattering albedo (SSA) and asymmetry factors (ASY) from the MACv1 climatology together with the "rural" aerosol type to parameterize only the Angstrom exponent (AE).
-ARI_MC uses the MACC aerosol optical depth dataset and the "rural" aerosol type.

5
All of these simulations use the Thompson (mp=8) aerosol-cloud interacting cloud microphysics scheme which will be referred to as the Thompson2008 scheme. It must be noted here that implementation of aerosol-radiation interactions in a simulation enables the impact of both the direct and the semi-direct aerosol effect. The single scattering albedo (SSA) at 550nm of the ''rural" type aerosols ranges in our experiments between 0.92 and 0.98 whereas the ''urban" type is much more absorbing with SSA starting as low as 0.6, values that are considered unrealistic (Rodriguez et al., 2013;Tombette et al., 2008;10 Witte et al., 2011). Therefore the ARI_Mv1urban simulation must be considered as an idealized experiment of extremely absorbing aerosols.
Two additional simulations (ACI, ARCI) have been performed using the new Thompson aerosol-cloud interacting cloud microphysics scheme (mp=28), which enables the aerosol indirect effect.
-The ACI simulation does not consider aerosol-radiation interactions.

15
-Simulation ARCI includes both aerosol-radiation and aerosol-cloud interactions with the passing of effective radii from the aerosol-cloud interacting cloud microphysics to the radiation scheme. This simulation presents the most complete physical description of aerosol effects in the simulation ensemble.
All the simulations, aerosol sources and options used are presented in Table 1. The simulations that account for aerosolradiation interactions are symbolized with ARI in their names. The simulation with the Thompson aerosol-cloud interacting 20 scheme that accounts for aerosol-cloud interactions is symbolized as ACI whereas the experiment that accounts both for aerosol-radiation and aerosol-cloud interactions is symbolized as ARCI. The simulations that account only for aerosol-radiation interactions will be referred to as the ARI group of experiments. Finally, for brevity, the Thompson aerosol-cloud interacting scheme is referred to as TE2014 hereafter.

Methodology
We analyze the following variables: temperature at 2m, precipitation, shortwave down welling radiation at the surface (Rsds), direct normalized irradiance at the surface (DNI), diffuse irradiance at the surface (DIF), total cloud fraction (CFRACT) and the wind field at various pressure levels. Besides total column cloud fraction we also examine cloud fraction regarding low (<2.5 km), medium (2.5<z<6 km) and high (>6 km) level clouds. Cloud fraction for each level, as well as for the total column, 5 is calculated using the random overlapping method where the total cloud fraction C rand for two layers is considered as: C rand = c a +c b -c a c b where c a , c b are the cloud fraction in each layer (Hogan and Illingworth, 2000).
We also calculate the following metrics: 1. The radiative effect of aerosol on shortwave radiation at the surface (RE). It is the difference in net shortwave radiation at the surface (netRsds) between an aerosol simulation and the CON experiment. Thus: 2. The direct radiative effect of aerosol on shortwave radiation at the surface under clear-sky conditions (DRE). This is the difference in net clear-sky shortwave radiation at the surface (netCRsds) between an aerosol simulation and the CON experiment. Thus: Since the DRE is calculated under clear-sky conditions it encompasses only the direct aerosol effect and not the semidirect effect.
3. The effect of clouds on shortwave radiation at the surface (SCRE). It is the difference of the net shortwave radiation at 5 the surface (netRsds) and the net clear-sky shortwave radiation at the surface (netCRsds) for a given experiment: 4. In order to assess the impact of the aerosol implementation on the radiative effect of clouds the difference of SCRE (∆SCRE) is calculated between an aerosol experiment and CON. Therefore: When comparing the group of simulations that account for the aerosol-radiation interactions only with CON the calculated ∆SCRE accounts for the semi-direct effect of aerosols.
Regarding all the variables examined, in order to assess the impact of aerosol implementation we always compare the aerosolinteracting simulation to the non-interacting control simulation CON. To assess the impact of the aerosol-radiation interactions 15 and the impact of different aerosol parameterizations, we compare the simulation family ARI, which use the Thompson2008 scheme, to CON. Comparison of the simulation ACI to CON indicates the impact of the Thompson aerosol-cloud interacting cloud microphysics scheme which implements the indirect aerosol effect. Comparison of ARCI to CON indicates the impact of both aerosol-radiation interactions and the Thompson aerosol-cloud interacting coloud microphysics scheme. Finally the only situation when a comparison is not performed against CON is when comparing ARCI to ACI, both using the aerosol-cloud 20 interacting Thompson cloud microphysics. This enables to assess the aerosol direct and semi-direct effect under an environment where the indirect effect is also present. In order to enable grid cell comparisons of the model output against observations we use distance weighted average remapping using the four nearest neighbor values. We always remapped the finer grid onto the coarser. Therefore, all satellite products were remapped onto the WRF 0.44 o grid, whereas temperature and precipitation model output was remapped onto the E-OBS 30 0.44 o rotated grid. Furthermore, simulated temperature has been corrected with respect to the E-OBS elevation, using a temperature lapse rate of 0.65 K/km throughout the domain.
We analyze our data over the whole European domain, which we define as the as the region that consists of the Prudence subregions (Christensen et al., 2007)  3 Results

Aerosol optical depth
The mean seasonal fields of aerosol optical depth (AOD) at 550nm used (or produced in the case of Thompson aerosol-cloud interacting scheme) in our experiments can be seen in Fig. 1 together with the AOD field of CM SAF SEVIRI satellite product 10 for comparison. The fields of both simulations using the Thompson aerosol-cloud interacting scheme are very similar thus only the AOD of ARCI is presented.
The climatology of Tegen has a lower AOD compared to SEVIRI, but follows the latter's seasonal spatial variability. MACv1 has a central European local maximum in JJA, which is not discernible in any other AOD product and overestimates the North African dust component in spring. The MACC reanalysis agrees best with the satellite data. Finally, the Thompson aerosol 15 optical depth has a couple of spurious features: the Eastern European AOD maximum in spring and the higher generally higher AOD in colder months (DJF and SON) compared to all other products.

Evaluation of the Control Simulation
Despite some biases the control simulation (CON) captures the basic features of the European climate, which in turn indicates that the main physical processes are represented with a reasonable degree of fidelity, thus increasing the confidence on the sensitivity results.

5
In the simulation CON winter temperatures are mostly underestimated (-0.5 o C domain average, land only), with higher cold biases over Scandinavia, the Mediterranean and the Alps (-1 o C) as indicated in the upper panel of (Fig. 2). Winter cold biases especially over northern Europe are common in many EURO-CORDEX simulations (Kotlarski et al., 2014). In this study winter biases are reduced in comparison to previous WRF exercises in EURO-CORDEX hindcast experiments (Katragkou et al., 2015). Since many of these WRF studies implement the Noah land surface model, we contend that the use of the CLM 10 land surface model in this study is a factor for the reduced cold bias. Especially since northern Europe is largely covered with snow during winter and the treatment of the snowpack by the land scheme is of particular importance. Also summer features a cold bias over most of the domain (-0.5 o C domain average) with a tendency for minor warm biases in south and Eastern Europe. This bias pattern -cold in the north and warm in the south -has been detected in other RCM simulations over Europe such as RCA4, CCLM4, HIRHAM (Kotlarski et al., 2014).

Precipitation
Winter precipitation is overestimated throughout the domain (43% domain average), with pronounced biases existing over central (+50% ) and especially over Eastern Europe, locally exceeding 100% (Fig. 2). Wet biases during DJF in Eastern Europe are common in WRF simulations (Katragkou et al., 2015;García-Díez et al., 2015;Mooney et al., 2013). The current parameterization (CON) seems to amplify the commonly simulated wet bias in the eastern part of Europe during winter. In 20 summer biases are smaller and mostly dry (-3% domain average), which is not very typical for WRF, with most subregions presenting underestimation around -20 to -30% . However, areas with high positive relative biases are seen at the southern parts of Europe, where precipitation amounts are very small during the warm months which amplifies the relative biases. The above winter/summer bias patterns are seen in both cloud microphysics schemes used, the Thompson in CON and Thompson aerosol-interacting. An additional simulation conducted using the WDM6 cloud microphysics (not shown) yielded very similar 25 results regarding precipitation bias indicating that the cloud microphysics scheme is not the main cause of precipitation bias.

Cloud fraction
Cloud fraction is overestimated in winter at 0.17 (+35% ). The relative increase is more pronounced over the Iberian Peninsula (+60% ) (Fig. 2, 3 rd panel). In summer, the average overestimation is lower (0.08 or 12% ) but there is a zonal pattern with ∼ 30% overestimation in northern Europe and a 10 % underestimation in the Mediterranean region. However, relative biases have 30 to be interpreted with caution in southern Europe during summertime because of the small cloud fraction amount. For both seasons similar spatial patterns, including the bias magnitudes, have been observed in other WRF simulations (Katragkou et al., 2015;García-Díez et al., 2015). In the study of Katragkou et al. (2015), the WRF simulations that had a higher cloud fraction overestimation over the northern part of the domain were the ones implementing the Grell-Devenyi cumulus parameterization.
The Grell-Freitas scheme used in this study is similar to the Grell-Devenyi scheme, consequently cloud overestimation in our case could be to some extent linked to the cumulus parameterization selection, especially during summer.

Shortwave radiation to the surface and direct normalized irradiance
Shortwave downwelling radiation at the surface (Rsds) averaged for the entire European domain is underestimated for both winter and summer. In winter Rsds is in general slightly underestimated (-4% average), with some subdomains like Mid-Europe, France and British Isles reaching -20 to -40% . In summer the domain averaged Rsds underestimation is approximately -8% . Larger negative biases are seen in the north and decrease in intensity as we move to the south following quite closely the 10 cloud fraction bias pattern. The cloud and Rsds bias patterns are spatially correlated, as expected.
The bias pattern of direct normalized irradiance (DNI) is similar to that of Rsds but intensified. The underestimation in winter is around 13% whereas for summer the dual pattern of underestimation to the north (-20% ) and overestimation to the south (20-30% ) is even more pronounced.
In general, the aerosol-interacting simulations, implementing aerosol-radiation and/or aerosol-cloud interactions and the 15 Thompson aerosol-cloud interacting cloud microphysics, present a similar behavior to the control simulation CON, regarding the biases of the main variables described above. This indicates that aerosol representation, despite its considerable impact seen in the next chapter, is not the main source of bias in our simulations. Moreover, aerosol introduction, despite making the representation of physical processes in the model more complete, often does not lead to bias improvements. Furthermore the improvement of bias does not necessarily mean that the aerosol representation is correct, since model biases can be the result of 20 error compensation between errors in the aerosol representation and errors induced by other physical mechanisms. Zubler et al. (2011) in an RCM study (aerosol-climatology paper) reached similar conclusions, stating that the overestimation of aerosol optical depth was responsible for masking strong biases in the simulated cloud fraction. Figure S1 in the supplement presents the basic biases for simulation ARI_T with the Tegen climatology.

Sensitivities
In this section we explore the impact of the aerosol-radiation and aerosol-cloud interactions implementation on each variable separately.

Temperature
Accounting for the aerosol-radiation interactions only (ARI group) leads to surface cooling (Fig. S2), as expected due to the 5 lower radiation levels reaching the ground. Domain averaged changes compared to CON are negative and generally between -0.1 to -0.3 o C with the largest impact seen during summer and autumn (   The temperature decrease is not constrained to the surface but is also detected at higher levels, with decreasing intensity at 10 higher altitudes, usually reaching 850hPa. In the case of autumn over the Balkans and the Black Sea a decrease of -0.2 o C can be seen almost up to 400hPa. In summer ARI_Mv1urban is the only simulation from the ARI group that presents a large area of statistically significant warming at the surface, seen in summer over parts of the Alps, the Iberian Peninsula, Italy and the Balkans, coinciding with a decrease in total cloud fraction (CFRACT). This warming can be attributed to the highly absorbing "urban" type aerosols that 15 warm the atmosphere by absorbing solar radiation but also can affect temperature through circulation and cloud cover amount changes (Fig. S3). This temperature increase clearly affects the surface but also reaches higher levels up to 200hPa. The aerosol absorptivity, expressed through the SSA, can have a strong effect on the signal of the temperature changes presented. Warming of near surface temperature, including the pattern described above during summer (with slightly smaller warming), has also been described by other studies (Huszar et al., 2012;Zanis, 2009) that implemented much more realistic and less absorbing aerosols compared to ARI_Mv1urban.
Contrary to the ARI group, simulation ACI using the Thompson aerosol-cloud interacting cloud microphysics and accounting for indirect effects only results in a domain averaged temperature increase (0.1 to 0.2 o C) compared to CON for all seasons 5 except autumn. When the aerosol-radiation interactions are also enabled in ARCI, temperature decreases compared to ACI by -0.1 to -0.2 o C regarding domain averages. This is roughly the same amount of cooling seen when the aerosol radiation interactions are enabled in control CON.

Precipitation
Aerosol related domain averaged changes of precipitation are small in all experiments (± 0.08mm/day), at most up to ± 10% 10 in relative values (Table 3). ARI_Mv1urban again has a more intense impact with a relative decrease of -12% in JJA and MAM (-0.2m/day). All the other ARI experiments have no specific tendency of precipitation change throughout the year. In general winter is the season which is least impacted by aerosol implementations. The study of Nabat et al. (2015) using a coupled atmospheric-ocean model showed a decrease in precipitation over Europe. This decrease was attributed to the aerosol induced cooling of SST that led to decrease latent heat fluxes consequently decreasing atmospheric humidity and cloud cover. Therefore 15 the use of prescribed SST in the current study can be seen as a limitation and could particularly affect precipitation results.
The small domain averages are to an extent a product of sign compensation since the spatial pattern of precipitation differences from control is not homogenous but consists of small areas with increases and decreases scattered around the domain (Fig. S4). Precipitation changes at a grid scale level in several cases can exceed ± 50% , however this effect can probably be attributed to internal model variability and not to aerosol implementation.

20
A common area of significant precipitation increase in all experiments is seen over the Black Sea in autumn, where a significant CFRACT increase and cyclonic anomaly in the wind field at 850 and 500hPa is present. This characteristic cyclonic anomaly ( Fig. S5) is seen in all ARI group simulations but also to a lesser extent in simulations ACI and ARCI (not shown).
There is no clear spatial correlation between changes in cloud amount and changes in precipitation. Over the Black Sea in autumn increase in precipitation coincided with increase in CFRACT. In another case the extended CFRACT decrease seen 25 in ACI over large part of Central and Northern Europe in summer is accompanied by significant precipitation decrease over a much smaller area over Central Europe.
ARI_Mv1urban exhibits the largest and the spatially most extensive impact on precipitation. During summer and spring large areas of precipitation decrease are seen over Central-Southern Europe and the Balkans coinciding spatially with CFRACT decrease (see section 3.3.6). Clearly, the warming of the mid troposphere (Fig. S3) due to the highly absorbing nature of the 30 aerosols in ARI_Mv1urban stabilizes the atmosphere leading to both precipitation suppression and cloud dissolution.

Clear sky radiation at the surface and DRE
Accounting for the aerosol radiation interactions leads to statistically significant reductions in clear sky downwelling shortwave radiation to the surface (Crsds). Crsds decreases by -5 to -8% (domain average), depending on the simulation, during all seasons. Larger reductions of ∼ 14 % are found in the ARI_Mv1urban simulation. Figure 3 shows the DRE at the surface quantified as the difference of netCrsds between the experiment and CON. The 5 domain averaged DRE when aerosol-radiation interactions are enabled is very similar despite the different aerosol datasets, for all ARI simulations being around -4 to -5W/m 2 in winter and -14 to -17W/m 2 in summer (Table 4). ARI_Mv1urban shows twice the reduction than other aerosol treatments due to the considerably more absorbing nature of "urban" type aerosols.
When implementing aerosol-radiation interactions in an environment where the indirect effect is also present (ARCI minus ACI) the DRE has also comparable values to the ARI experiments. Spatially the DRE correlates very well with the AOD field 10 of each simulation, with the AOD maxima coinciding with the Crsds minima for each experiment. Correlation coefficients for the ARI group range between -0.8 and -0.98. Interestingly, correlations for the simulation ARCI with aerosol-radiationcloud interactions are smaller (around -0.7 to -0.8) for most seasons and especially during winter (-0.43). Since the Thompson aerosol-cloud interacting scheme produces an evolving AOD field, using the mean AOD pattern might not be sufficient in a correlation analysis, thus partially explaining the decrease in correlation coefficient described above.

Radiation at the surface and RE
For the ARI simulations, shortwave downwelling radiation at the surface (Rsds) shows significant attenuation almost all over the domain throughout the year. Domain averaged Rsds attenuation lies in the range -3 to -8% for all seasons, quite similar with the decrease seen in clear-sky Rsds (Table 5). ARI_Mv1urban is again an exception with higher attenuation around -12 to -16% .

5
The change in the net shortwave radiation at the surface, constitutes the radiative forcing (RE) of aerosol (Fig. 4) and comprises of the direct radiative forcing (DRE) and the forcing due to changes in cloud amount and properties (∆SCRE).
Accounting of aerosol-radiation interactions only, leads to a negative RE of -2W/m 2 in winter and -11 to -13 W/m 2 in summer (-7W/m 2 annual average) with ARI_Mv1urban roughly doubling these values (Table 4). Compared to other studies our results present in general a smaller radiative effect over Europe. Nabat et al. (2015) showed an annual average RE of -10 W/m 2 . The 10 study of Huszar et al. (2012) calculated a similar to our study RE during summer (-12 to -15 W/m 2 ) but a considerably larger effect (-7 W/m 2 ) in winter whereas the RegCM3 study of Zanis (2009) for the year 2000 presented a higher summer radiative effect (-16 W/m 2 ).

The impact of the aerosol-radiation interactions on shortwave radiation weakens when implementing it in an environment
where the indirect effect is also present. For example, comparing ARCI with ACI leads to similar behavior as the one from 15 the ARI group, however Rsds attenuation is slightly smaller and RE less negative (-1W/m 2 in DJF, -11W/m 2 in JJA). This is because the change in cloud forcing ∆SCRE (ARCI-ACI) is positive and increased compared to the ARI experiments (eg ACI-CON) (Table 4).
Interestingly, when implementing both the aerosol-radiation interactions and the Thompson aerosol-cloud interacting cloud microphysics with the indirect effect (ARCI) and comparing against the control simulation, domain Rsds averages are increased 20 in the range of 2 to 10% with RE becoming positive, despite the widespread decrease in Crsds. In general, simulations using the Thompson aerosol-cloud interacting scheme, including ACI, present slightly larger positive cloud forcing at the surface, partially attributed to lower cloud fraction values compared to control (section 3.3.6), which leads to an extended Rsds increase and positive RE. As a consequence, ACI that implements the indirect effect only shows domain averaged Rsds increase. In ARCI the domain averaged RE compared to control is almost zero due to the compensation between the negative DRE (direct 25 effect), and positive ∆SCRE. RE includes both positive and negative values, with the positive ones being more intense in the northern and western part of the domain during summer and spring (Fig. 4).
When implementing aerosol-radiation interactions only, the spatial correlation between radiative forcing RE (calculated as a difference from CON) and the AOD field is high (-0.6 to -0.9) and considerably decreases when the indirect effect is included (-0.2 to -0.4). This is due to the cloud fraction changes, attributed not only to aerosol effects but also to the different cloud 30 microphysics scheme used (TE2014).
It is important to note that aerosol optical properties besides AOD can have a severe impact on seasonal radiation amounts.
For example, simulations ARI_Mv1, ARI_Mv1full and ARI_Mv1urban all use the MACv1 AOD data but parameterize the other aerosol optical properties differently. ARI_Mv1 and ARI_Mv1full have similar single scattering albedo (SSA) values (0.92 to 0.98) which leads to similar results in domain averaged Rsds decrease. Aer2urban however has considerably more absorbing aerosols (SSA starting from 0.6) leading to an almost doubled impact on Rsds attenuation. This impact is widespread over the domain with the overall distribution of Rsds decrease being clearly shifted towards more negative values (Fig. S6). Alexandri et al. (2015) also stressed the importance of secondary aerosol parameters such as SSA in simulating solar radiation in regional climate simulations.

Direct and Diffuse radiation
The ARI group of experiments attenuate the direct normalized irradiance (DNI) more severely than Rsds. Compared to control, domain averaged differences are around -30% for all seasons (Table 5). At grid scale level attenuation can exceed -50% (   Changes in total cloud fraction (CFRACT) compared to CON due to aerosol implementation are shown in Fig. 5. In general, regardless of the type of aerosol implementation changes are quite small. Therefore, domain averaged differences from CON do not exceed ± 0.01, (scale of 0 to 1). This partially happens because cloudiness increases and decreases in parts of the 5 domain. However, the absolute differences from CON are still quite small with a range of 0.01 to 0.03. Smallest impacts are seen in winter where cloudiness is mainly affected by synoptic phenomena. In relative values domain changes are around 1-2% for winter and up to 3-4% (6% for ARI_Mv1urban) during summer. In ARI_Mv1urban CFRACT changes exceed in some cases 0.15. The aerosol-radiation interactions has a minor impact on CFRACT. The two simulations using the aerosol-cloud interacting cloud microphysics have lower cloud fraction amounts throughout the year compared to all other simulations using 10 the Thompson2008 scheme. This is probably connected to the fact that the above two simulations also present smaller liquid water path (LWP) values.
Some areas show statistically significant differences in CFRACT which follow the pattern of temperature changes. Cloud fraction increase occurs in areas with temperature decrease whereas decreases in cloud cover are related to areas with atmospheric warming. The most pronounced CFRACT increases occur above the Black Sea and eastern Balkans in autumn (includ-15 ing parts of North Africa and Central-Eastern Mediterranean in some cases). These changes are present in all the simulations (Fig. 5). They are probably related to the previously mentioned cyclonic anomaly in the wind field (both 850 and 500hPa) over the Black Sea region (Fig. S5). The introduction of aerosol-radiation interactions reduces radiation at the surface, thus decreasing temperature. This in turn leads to the formation of a cyclonic anomaly that produces larger cloud fraction amounts which in turn further decreases radiation levels hence decreasing temperature, leading to a feedback mechanism. Temperature at the surface (around -1 o C) is clearly decreased due to the combined effect of aerosols and increased cloud amount. Negative temperature changes extend to higher levels (-0.6 o C at 850hPa and -0.2 o C at 400hPa). A composite plot showing the ARI_T case using the Tegen climatology is presented in the supplement (Fig. S8).
The impact on cloudiness is more pronounced in ARI_Mv1urban as a result of extreme absorbing aerosols. Significant changes in CFRACT are found in extended parts of the domain for all seasons except winter. This highlights the importance of 5 introducing aerosol optical properties (e.g. SSA) in RCM simulations, as they can affect the thermodynamics of the lower and mid atmosphere (Fig. S3).
The patterns of significant changes in total cloud fraction in our simulations are dominated by changes in low clouds, which are mostly affected (Fig. S9). Medium level cloud changes are less pronounced in amplitude and area extent, whereas higher clouds are least impacted by changes in aerosol treatments. This is to be expected, since our aerosol concentrations are located

b) Cloud forcing
We showed that accounting for the aerosol-radiation interactions does not systematically change CFRACT, whereas the use of the Thompson aerosol-cloud interacting cloud microphysics lowers the cloudiness. Of particular interest is the impact of aerosol on the ability of clouds to interact with radiation. To study this effect we calculate the aerosol related change in the cloud forcing regarding shortwave radiation at the surface (∆SCRE) (Fig. 6). The domain averaged change in the cloud forcing 5 is positive in all experiments (Table 4). Thus, the introduction of aerosol-radiation and/or aerosol-cloud interactions leads to more transparent clouds enabling larger amounts of radiation reaching the surface. For ARI simulations ∆SCRE represents the impact of semi-direct aerosol effect on radiation, which is positive with annuals averages around 3 to 4 W/m 2 is largest during spring (5-7 W/m 2 ). Nabat et al. (2015) had calculated a larger annually averaged semi-direct effect around 5 to 6 W/m 2 . This effect is counteracting the direct radiative effect (DRE) of aerosol that is clearly negative. The semi-direct effect accounts for 10 60% of the direct aerosol effect on radiation (DRE) during winter, 45% during spring and around 20-35% during summer and autumn. Consequently, the impact of semi-direct effect on radiation is considerable, and plays an important role in the overall impact of aerosol-radiation interaction implementation in the model.
When introducing aerosol-radiation interaction in ACI where the indirect effect is also present, the ∆SCRE (ARCI-ACI) represents the combined effect of the semi-direct and indirect effect changes on cloud forcing. In this case ∆SCRE is slightly 15 larger (4-7 W/m 2 ) compared to the effect in the ARI group, and its relative importance is also increased, amounting up to 80% of the DRE in winter and 65% in autumn.
In ARCI the change in cloud forcing is attributed to the semi-direct and indirect effects as well as the change in the cloud microphysics scheme. In this case the ∆SCRE is further increased to 6 to 13 W/m 2 . This results in ∆SCRE being equal in magnitude to the DRE, increasing its relative importance in the overall radiative effect (RE) of aerosol.

20
The changes in cloud forcing can be attributed to changes in cloud amount but also to cloud properties and characteristics.
There are areas with considerable ∆SCRE increase which is accompanied by non-statistical cloudiness change. This behavior is typically seen over the northeastern part of the domain especially over the sea and is more pronounced in the simulations using the TE2014 scheme. In these cases, the positive change in cloud forcing can be attributed to changes in cloud properties that result in decreased cloud optical depth.