Saharan dust event impacts on cloud formation over Western Europe Saharan dust event impacts on cloud formation and radiation over Western Europe

We investigated the impact of mineral dust particles on clouds, radiation and atmospheric state during a strong Saharan dust event over Europe in May 2008, applying a comprehensive online-coupled regional model framework that explicitly treats particle microphysics and chemical composition. Sophisticated parameterizations for aerosol 5 activation and ice nucleation, together with two-moment cloud microphysics are used to calculate the interaction of the di ﬀ erent particles with clouds depending on their physical and chemical properties. The impact of dust on cloud droplet number concentration was found to be low, with just a slight increase in cloud droplet number concentration for both uncoated and 10 coated dust. For temperatures lower than the level of homogeneous freezing, no signif-icant impact of dust on the number and mass concentration of ice crystals was found, though the concentration of frozen dust particles reached up to 100 l − 1 during the ice nucleation events. Mineral dust particles were found to have the largest impact on clouds in a temperature range between freezing level and the level of homogeneous 15 freezing, where they determined the number concentration of ice crystals due to e ﬃ cient heterogeneous freezing of the dust particles and modiﬁed the glaciation of mixed phase clouds. Our simulations show that during the dust events, ice crystals concentrations were increased twofold in this temperature range (compared to if dust interactions are ne- 20 glected). This had a signiﬁcant impact on the cloud optical properties which caused a reduction in the incoming short-wave radiation at the surface up to − 75 W m − 2 in areas with high dust concentrations. Including the direct interaction of number in cloud albedo in the an in cloud droplet radius. These conﬂicting results indicate that the impact on cloud droplet number depends on the individual atmospheric conditions, e.g. and type. A case study performed for an increase in droplet number and in top height with an increasing concentration of mineral dust strong dependency of the dust impact on precipitation on the prescribed aerosol

activation and ice nucleation, together with two-moment cloud microphysics are used to calculate the interaction of the different particles with clouds depending on their physical and chemical properties.
The impact of dust on cloud droplet number concentration was found to be low, with just a slight increase in cloud droplet number concentration for both uncoated and coated dust. For temperatures lower than the level of homogeneous freezing, no significant impact of dust on the number and mass concentration of ice crystals was found, though the concentration of frozen dust particles reached up to 100 l −1 during the ice nucleation events. Mineral dust particles were found to have the largest impact on clouds in a temperature range between freezing level and the level of homogeneous 15 freezing, where they determined the number concentration of ice crystals due to efficient heterogeneous freezing of the dust particles and modified the glaciation of mixed phase clouds. Our simulations show that during the dust events, ice crystals concentrations were increased twofold in this temperature range (compared to if dust interactions are ne-

Introduction
Aerosol particles are an important part of the atmosphere. They directly modify the planetary radiation budget by scattering and absorption of long and shortwave radiation, and they affect the properties of clouds. Depending on their size and chemical 10 composition, aerosol particles can act as cloud condensation nuclei, CCN, and ice nuclei, IN, and profoundly impacting the cloud microphysical processes and optical properties, hence the hydrological cycle and climate.
The complexity of the interactions of aerosol particles with radiation and the hydrological cycle render the most uncertain factors in climate studies and weather predic- 15 tion.
One of the major contributors to the atmospheric aerosol mass is mineral dust. Even though the emission sources of mineral dust in the atmosphere are mainly desert regions, dust particles can be transported over long distances. Europe and the Mediterranean can be strongly affected by dust outbreaks originating in the Sahara Desert containing clouds over Europe with cloud top temperatures from −10 to −20 • C in air masses that contained Saharan dust. Analyzing a Saharan dust outbreak over the Atlantic in 2004, Min and Li (2010) showed that ice clouds tend to form at higher temperatures when dust particles are present. Global modelling studies demonstrated the important contribution of mineral dust to IN (e.g. Hoose et al., 2008;Lohmann and Diehl, 2006). The impact of mineral dust on the formation of cirrus clouds for temperatures lower than −35 • C depends on the competition of the heterogeneous freezing of mineral dust with the homogeneous freezing of liquid aerosol particles. For conditions with low updraft velocities, the heterogeneous freezing of dust particles is able to inhibit the homogeneous freezing and cause a decrease in ice crystal number concentration 15 of the affected cirrus clouds (Barahona et al., 2010b). Based on satellite remote sensing, Min et al. (2009) showed that mineral dust significantly increases ice crystal number concentration and cloud droplet concentration with the consequence of a decreasing precipitation rate. Similar results were found by Rosenfeld et al. (2001) with a decrease in cloud droplet radius in clouds affected 20 by dust and a suppression of light precipitation. In contrast, Pradelle and Cautenet (2002) found a decrease in cloud albedo in the presence of mineral dust because of an increase in cloud droplet radius. These conflicting results indicate that the impact of mineral dust on cloud droplet number depends on the individual atmospheric conditions, e.g. background aerosol and cloud type. A case study performed for the Eastern 25 Mediterranean by Solomos et al. (2011) showed an increase in droplet number and in cloud top height with an increasing concentration of mineral dust particles, but also a strong dependency of the dust impact on precipitation on the prescribed aerosol background conditions. The impact of dust on the cloud droplet concentration in most studies was hypothesized to be caused by hygroscopic coatings, which make them important CCN due to their large size. With the exception of dust from dry lake beds, freshly emitted dust contains little soluble material . Processing of dust particles in clouds (Levin et al., 1996) and the condensation of acids on the particles (Sullivan et al., 2007) 5 are prime mechanisms thought of creating hygroscopic coatings on dust. Kumar et al. (2011) however pointed out that dust particles do not require deliquescent material to act as CCN in the atmosphere, as the adsorption of water on the particle can induce hygroscopicity equivalent to having a considerable soluble fraction. In a global study, Karydis et al. (2011) pointed out that even uncoated mineral dust can contribute significantly to cloud droplet number close to the source regions and that hygroscopic coating of dust particles can decrease the cloud droplet number in regions which are affected by anthropogenic emissions due to a decrease in water supersaturation.
This study focuses on a major dust event that occurred in May 2008. Its origin was in the Sahara and from there mineral dust particles were transported over the West- 15 ern Mediterranean, covering large areas of Western Europe. During the episode, high aerosol concentrations were observed throughout Europe; ice nuclei concentrations significantly increased (compared to pre-event levels) at Kleiner Feldberg, Germany (Klein et al., 2010). During this time, traditional weather forecast models exhibited poor prediction skill. The German national meteorological service (Deutscher Wet-20 terdienst, DWD) detected a significant bias in their numerical weather forecast of the 2-m temperatures of +1.5 K when the dust outbreaks reached SW-Germany (Damrath, 2010, DWD, personal communication). The operational weather forecast at DWD is performed with the model system COSMO (Consortium for Small-scale Modeling) (Baldauf et al., 2011). Since the impact of aerosol particles on atmospheric state is ac-25 counted for in COSMO by using prescribed profiles for aerosol optical properties and prescribed cloud droplet and ice crystal numbers, extraordinary aerosol conditions like the dust event in May 2008 are not represented in the simulations. This can potentially cause biases in the weather forecast. 10 m-windspeed, and friction velocity, and surface properties, such as sea surface temperature, soil type, and soil moisture (Lundgren, 2010;Stanelle et al., 2010). A detailed description of the treatment of gases, aerosols, and their emissions is given by Vogel et al. (2009).

5
COSMO includes an efficient bulk cloud microphysics scheme, designed for operational weather forecast (Doms et al., 2005). This scheme cannot treat aerosol-cloud interactions, because only one moment of the size distribution is calculated. Bangert et al. (2011) extended the cloud scheme to a two-moment representation of the cloud droplet size distribution to consider the impact of aerosol particles on warm cloud 10 microphysics. To further improve the clouds microphysics and to enable the simulation of aerosol impacts on ice clouds a comprehensive two-moment bulk microphysical scheme Noppel et al., 2006;Blahak, 2008) was introduced, which replaces the old schemes. This scheme distinguishes six hydrometeor categories (cloud drops, cloud ice, rain, snow, graupel, and hail) and rep-15 resents each particle type by its respective number and mass concentration. A (generalized) gamma size distribution is used for each hydrometeor class, where the so-called shape parameters are held constant during the simulation. For the warm clouds, the scheme considers autoconversion of cloud droplets to rain, accretion of cloud droplets by rain drops, self-collection of cloud and rain droplets, break-up of rain drops, and 20 evaporation of rain drops. Condensational growth of cloud droplets is calculated with a saturation adjustment technique. For the cold clouds, homogeneous and heterogeneous ice nucleation, diffusional growth of ice crystals, freezing of cloud and rain droplets, aggregation, self-collection, riming, conversion to graupel, melting, sublimation, shedding, and Hallett-Mossop ice multiplication are considered. The freezing of 25 cloud and rain drops is calculated with a classical statistical approach based on an empirical relation for the freezing probability as a function of temperature . Note that the freezing of cloud and rain droplets is independent of 31943 Introduction A detailed description of the cloud microphysical processes is given in , and a statistical analysis of the aerosol-cloud interaction for three summer seasons using the microphysics scheme is presented in Seifert et al. (2011). Although COSMO is a non-hydrostatic model and therefore permits the simulation of convection, a parameterization of subgrid-scale deep convection has to be used for coarse model resolutions. We use a modified Tiedtke scheme (Tiedtke, 1989) for horizontal resolutions coarser than 7 km, which considers the convective transport of 10 gases and aerosol particles.

Aerosol activation
The activation of an aerosol particle to a clouds droplet depends on its ability to remain in a stable equilibrium with the ambient water vapor. The water vapor saturation ratio, S eq , of an aerosol particle in equilibrium with surrounding water vapor can be expressed where σ is the surface tension at the particle-gas interface, M w is the molar mass of water, R is the universal gas constant, T is the temperature, ρ w is the density of water, and D is the equivalent particle diameter. The exponential in Eq.
(1) is commonly 20 referred to as the curvature or Kelvin effect. α w is the activity of the water and depends on the physiochemical properties of the particle. In case of a soluble aerosol particle α w can be expressed by the effective mole fraction of water in the solution droplet, which results in the well known Köhler equations. For insoluble particles, e.g. dust, this classical theory cannot be applied, because S eq is now affected by another physical Introduction where A FHH , B FHH are empirical constants, and Θ is the surface coverage (defined as the number of adsorbed layers of water). A FHH characterizes interactions of adsorbed 5 molecules with the aerosol surface and adjacent adsorbate molecules (i.e., those in the first monolayer). B FHH characterizes the attraction between the aerosol surface and the adsorbate in subsequent layers; the smaller the value of B FHH , the greater the distance at which the attractive forces act (Sorjamaa and Laaksonen, 2007). A FHH and B FHH are compound-specific and determined experimentally. Both theories (Köhler 10 and FHH) show a similar behavior for S eq , with a characteristic maximum (the critical supersaturation s c ). For ambient supersaturations greater than s c the particle can get activated and grow to the size of a cloud droplet. In this study we use a comprehensive activation parameterization (Kumar et al., 2009;Barahona et al., 2010a) based on a cloud parcel framework, in which a parcel 15 of air containing an external mixture of Köhler and FHH particles is lifted. At first the s c distribution is calculated for each aerosol. Then the supersaturation equation of the parcel is solved using the bisection method to determine the maximum supersaturation, s max , and the number of activated particles, N * a . As input data the simulated cloud-scale dynamics (updraft, temperature, pressure) and aerosol properties (size distributions 20 and chemical compositions of all aerosol modes) are used.
To account for hygroscopic coating of e.g. dust particles, Kumar et al. (2011) developed a unified activation theory by merging Köhler and FHH theory in their parameterization. For coated dust particles α w is represented by where x w is the mole fraction of water in the droplet and represents the solute effects on water activity. Based on the procedure of Bangert et al. (2011) the activation rate, ACT, of aerosol particles in the model is calculated in different ways, replacing the activation rate in Eq. (17) of . For a newly formed cloud the parameterization is directly applied and N * a /∆t equals the activation rate, where ∆t is the time step used. In case of an already existing cloud the activation rate at the cloud base is calculated 5 on the basis of advection and turbulent diffusion of particles into the cloud base where w is the grid-scale updraft, z is the height above sea level, and K is the turbulent diffusion coefficient.
In-cloud activation is calculated, if the simulated grid-scale supersaturation increases again above the cloud base, e.g. from strong updrafts. In this case the growth of the existing cloud droplets is considered by assuming they act as giant CCN that deplete supersaturation by the approach of Barahona et al. (2010a). The corresponding activation rate is calculated as N * a /∆t. 15 The nucleated number concentration of ice crystals, N * i , is calculated using the parameterization of Barahona andNenes (2008, 2009a,b), which is based on the framework of an ascending Lagrangian air parcel. Competition between homogeneous and heterogeneous freezing is explicitly considered in the calculation of the ice supersaturation, s i . In doing so the dependency of N * i on the conditions of cloud formation 20 (i.e., T , p), updraft velocity, deposition coefficient, and soluble and insoluble aerosol concentrations is explicitly resolved. N * i is given by

Ice nucleation
where s i,max is the maximum supersaturation that develops in the cirrus, s hom is the homogeneous freezing threshold (Koop et al., 2000), and N hom and N het are the number Introduction  (2008) is used for N het (s i ), which is derived from several field campaign data sets. It provides the contribution of individual aerosol species (dust, black carbon, and organics) and freezing mechanisms (e.g. immersion and deposition) to N het (  The homogeneous contribution to N * i is given by where N 0 is the number concentration of the supercooled liquid droplet population and f c is the droplet freezing fraction for cirrus clouds formed in situ ( Barahona and Nenes, 2008). N 0 equals the sum of the number concentration of the soluble aerosol modes 20 (Table 2) which are assumed to deliquesce during ice cloud formation. The nucleation rate of ice crystals, NUC, is calculated following  Eq. (34) by where N i is the number concentration of ice crystals before the nucleation pulse.

Subgrid-scale updrafts
Though regional models are able to resolve individual cloud systems, they cannot explicitly capture the updraft velocities which control nucleation of ice and droplets. Therefore parameterizations of the subgrid-scale vertical velocity must be applied to address 5 this issue. Sub-grid scale vertical velocities, w , are described with a Gaussian probability distribution function, P w (w ). The mean of P w (w ) is set equal to the grid scale updraft, w, and the standard deviation, σ w , is calculated as the square root of the turbulent kinetic energy (Morales and Nenes, 2010). A weighted mean of the activated parti-10 cles/nucleated ice crystals, N * x , is calculated by numerically calculating the integral

Radiation
The radiative fluxes are calculated with the GRAALS (Ritter and Geleyn, 1992) radiation scheme of COSMO for eight spectral bands. To consider the impact of varying 15 aerosol and clouds the necessary optical properties, as they are the extinction coefficient, single scattering albedo, and asymmetry factor, have to be calculated. The optical properties of the aerosol particles are calculated as a function of the size distributions, the chemical composition, as well as the soot and water content with the parameterizations of Vogel et al. (2009), Lundgren (2010, and Stanelle et al. (2010). 20 To consider the varying droplet and ice crystal size in the radiation scheme the parameterizations of Hu and Stamnes (1993) and Edwards et al. (2007) are applied to calculate the optical properties of the cloud droplets and ice crystals. Here the effective radii of clouds droplets, r c , and ice crystals, r i , are calculated as the ratio of the third to second moment of the respective size distribution. The adaption of the parameterizations for the eight spectral bands of GRAALS is carried out following Zubler et al. (2011). Precipitating hydrometeors of snow, graupel, and rain, are not considered in the radiation calculations, which may introduce a bias in the calculated radiation fields (Waliser et al., 2011).

3 Simulation setup
A nesting approach is used for the simulations. At first a simulation for a domain D0 covering Northwest Africa and Western Europe was performed. This domain enables the explicit simulation of the dust emissions over Africa and the long-range transport of the dust particles to Europe (Fig. 1). All gases and aerosol particles are simulated 10 (see Sect. 2), but are not allowed to interact with clouds and radiation. Calculations of the cloud microphysical processes are carried out with the one-moment cloud scheme, which includes cloud water, rain, cloud ice, snow, and graupel. The grid size is 0.  REF the interaction of dust particles with clouds and radiation was not allowed. To investigate the impact of the dust particles due to their interaction with clouds we carried out simulation C which includes the impact of the dust particles on cloud formation via activation of dust particles and heterogeneous nucleation of cloud ice on the dust particles. In simulation C dust particles are assumed to not have a hygroscopic coating. 10 Therefore the FHH adsorption theory is used for the activation of the dust particles using A FHH = 2.5, and B FHH = 1.2 . Aged dust particles can potentially have hygroscopic coatings which decreases the supersaturation needed for their activation. For this simulation C * is carried out, assuming that the dust particles are coated by 10 % ammonium sulfate (in volume). In this case the activation of the dust particles is calculated with the unified approach of Kumar et al. (2011). In simulation CR, the direct interaction of the dust particles with long and shortwave radiation is considered additionally to their impact on cloud formation (with the conditions of simulation C). during these days. Due to steady southerly winds over the Mediterranean dust particles were transported efficiently from the emission sources in the Sahara to Central Europe. Figure 2 shows the measured and simulated aerosol mass concentration of particles with a diameter below 10 µm (PM 10

Dust plume
The main dust plume affecting Europe originated in Algeria where steady moderate 15 winds caused an effective emission of mineral dust into the atmosphere. Due to south easterly winds, the dust particle were transported to the North where they were lifted into the middle to upper troposphere at the Atlas Mountains. Figure 4 presents the simulated daily mean dust number concentrations, N d , for the nested model domain during the period of the dust event. N d is calculated for three 20 different temperature ranges, each of them favorable for a different process of aerosol cloud interaction. In the lowest layer, which reaches from the ground up to the freezing level, liquid clouds can form, with N d being highest, up to more than 50 cm The upper layer, which reaches from the level of homogeneous freezing up to 15 km height and hence is favorable of combined homogeneous and heterogeneous freezing, has mean dust concentrations in the order of 1 cm −3 over Europe.

Dust impact on cloud droplet and ice crystal number concentration
Owing to their size and surface properties dust particles can act as CCN. High dust con-5 centrations can therefore modify the cloud droplet size distribution and consequently the microphysical processes in the warm cloud phase. Dust particles are found to be very efficient ice crystal nuclei in several lab and field studies Möhler et al., 2006). Ice nucleation due to immersion freezing of dust particles starts at temperature around 263 K with a freezing fraction on the order of 0.1 %. At temperature 10 around 253 K, the freezing fraction of dust particles was found to be in the order of 1 % (Phillips et al., 2008). At temperatures below 235 K ice crystal nucleation due to homogeneous freezing of aerosol droplets sets in. At this temperature range heterogeneous freezing of the dust particles has to compete with homogeneous freezing for water vapor. Despite the fact that the coated dust particles can be activated at lower supersaturations than the uncoated particles, the domain wide joint histogram of cloud droplets in simulation C * shows only slightly higher N c in comparison with simulation C (Fig. 5).

5
One reason is that even though much more dust particles get activated in simulation C * in comparison to simulation C the number is still low in comparison to the total number of activated particles for most of the grid points (Fig. 6). But we also want to point out here, that there are more cloudy grid points in simulation C * where N * d contributes to 10 % or more of N *  to roughly 1 % of N d which nucleate to ice crystals in the considered temperature range. Additionally, the distribution of N i is broadening from a narrow distribution of N i in case REF centered at low N i in the order of 15 l −1 to a wide distribution centered at N i in the order of 60 l −1 . This can be explained be the more variable distribution of N d in comparison to the distribution of soot particle IN in case REF.

5
The more efficient ice nucleation in the temperature range between 273 to 235 K from heterogeneous freezing of the dust particles has an impact on the glaciation of mixed phase clouds. Though the freezing of cloud droplets in the realized simulations is independent of the simulated aerosol particles, the enhanced ice nucleation involving the dust particles is able to modify the cloud droplet distribution for temperature below 10 263 K (Fig. 5). The total number of grid points with m c > 0 is significantly reduced in case C with respect to case REF. In particular, grid points with low m c and N c are strongly decreased in number. This is due to the Bergeron-Findeisen process, where ice crystals grow at the expense of the existing water droplets due to the difference in saturation ratio with respect to water and ice. The total number of grid points containing 15 cloud water decreased by 61 % in simulation C with respect to simulation REF in the temperature range between 263 to 235 K. For atmospheric layers above the level of homogeneous freezing (T < 235 K), the joint histograms of m i and N i are very similar in both simulations. At this temperature, the heterogeneous freezing of the dust particles has to compete with the homoge-20 neous freezing of the aerosol droplets for water vapor during the ice nucleation. An increasing number of heterogeneous freezing particles, N het , will at first decrease the maximum ice supersaturation during the nucleation event and therefore decrease the total number of nucleated ice crystals because homogeneous freezing is less effective. The value for which N het completely inhibits homogeneous freezing is defined as 25 N lim (see Sect. 2.3). For N het greater than N lim , heterogeneous freezing is the sole contributor to ice crystal number concentration and will increase with increasing N d (Barahona and Nenes, 2009a). In Fig. 8  grid points where N het exeeds N lim for conditions with N lim lower than 60 l −1 . In the latter simulation, N het is generally much higher due to the heterogeneous freezing of the dust particles, but rarely exceeds 100 l −1 . Therefore N het is mostly below N lim and is not able to significantly impact N i . For relatively few conditions (with N lim lower than 100 l −1 ) N het is able to exceed N lim .

Dust impact on cloud properties and radiation
In the previous section we focused on the net impact of the dust particles on cloud droplet and ice crystal number concentration for the whole model domain during the simulation period and found systematically higher values in ice crystal number concentration for atmospheric layers between freezing level and the level of homogeneous 10 freezing in case C in comparison to case REF. This has numerous consequences on cloud properties. We already discussed the more effective glaciation of mixed phase clouds in simulation C, which has potential consequences on the vertical distribution of the latent heat release and the dynamic development of the clouds. The higher number concentrations in N i (but similar values for m i ) in both simulations causes difference 15 in the size of the ice crystals and therefore has consequences on the sedimentation velocity of the ice crystals and on the optical properties of the ice clouds. Short wave radiation is scattered more efficiently by smaller ice crystals, which results in optically thicker clouds. In Fig. 9 the difference in the net surface shortwave radiation flux, F sw , around noon at the 26 to 29 May and the difference in the net surface long wave ra- The temporal evolution of the median effective radii of cloud droplets, r c , and ice crystals, r i , calculated for domain D1, in simulation C and REF is shown in Fig. 10. For r c slightly lower values occur in case C with maximum differences to case REF mostly lower than 1 µm. In contrast, r i is significantly lower in simulation C on the order of 10-25 µm. The difference in r i scales with the time evolution of the magnitude of r i .

5
Small r i indicate that ice clouds occur mainly at greater heights, where the difference in r i is much smaller due to the controlling influence of homogeneous freezing. Figure 11 shows the temporal evolution of domain averaged cloud properties and radiation fluxes in simulation C and REF. The temporal evolution of averaged vertical integrated cloud water content, LWP, and ice water content, IWP, shows values of sim-10 ilar magnitude in simulation C and REF with positive and negative difference of a few percent. Whereas LWP shows most of the time slightly lower values and IWP slightly higher values in case C. Differences in LWP and IWP are mainly caused by differences in the dynamics of the cloud systems. The average in-cloud vertical velocity, w cloud , shows slightly higher values in case C during the first three days of the dust event. 15 This can be explained by more effective glaciation of mixed phase clouds in case C and the additional release of latent heat. On the last day w cloud is lower in simulation C which is also reflected in a lower LWP. In general, differences in dynamics can be driven by different processes, e.g. changes in radiation and consequently temperature caused by modified cloud properties, and are hard to relate directly with a specific 20 individual process.
In both simulations, the average hourly precipitation rate, PR is almost identical. Difference in PR scale mostly with the differences in LWP and IWP, e.g. during the last two days of the dust event were the largest differences occur. Despite the almost identical PR, the local differences can be large due to spatial shifts in the distribution 25 of precipitation.
The domain average F sw shows systematic differences between simulation C and REF with lower values in simulation C. The maximum F sw is reduced significantly in case C by 10 to 20 W m −2 . The largest difference in F sw occurs on 27 May with  (Fig. 9). The temporal evolution of the average F lw shows just small differences between simulation C and REF with a maximum of The difference in the net radiation fluxes at the surface have an impact on the simulated temperatures. The domain averaged daily temperature maximum of the temperature in a height of 2 m above ground, T 2 m is up to 0.3 K lower in case C than in case REF. T 2 m is lower in case C until the daily temperature minimum is reached in 10 the early morning, where T 2 m in both simulations converge again. Despite the small difference in the domain average T 2 m , the differences are locally much larger. This will be discussed in Sect. 4.5.

Direct dust impact on radiation
Until now we investigated only the impact of dust on the atmospheric state due to the 15 interaction of dust particles with cloud microphysical and optical properties. During the dust event the dust particles interact also directly with the radiation fields of the atmosphere due to scattering and absorption of short and longwave radiation. Results of simulation CR, which includes additionally to the interaction with the clouds also the direct interaction of dust particles with radiation, are shown together with the results of 20 simulation C and REF in Fig. 11. The domain averaged dust optical depth at a wavelength of 550 nm, τ dust , increases during the dust event up to 0.36 at the 27 May. During the last two days of the dust event τ dust decreases again to 0. is in the order of 3 to 10 W m −2 higher in case CR in comparison to the cases C and REF.
The direct impact of the dust particles in simulation CR causes lower daily maxima of T 2 m when compared to simulation C. The domain averaged T 2 m is 0.5 K lower in CR in comparison to REF at the 27 May. During the night T 2 m converge in the different 5 simulations and get even slightly higher values of T 2 m in case CR in comparson to C due to the higher values of F lw . The additional temperature decrease during the day in CR in comparison to C (from the direct radiative forcing of dust) is comparable in magnitude to the decrease in C in comparison to REF (by the interaction of dust with the cloud properties alone). Whether the direct impact on the radiation or the 10 interaction with the clouds has the larger impact on the domain averaged T 2 m varies from day to day depending on e.g. the spatial distribution of clouds and dust particles.

Impact on T 2 m
In Fig. 12 the impact of the dust on average T 2 m is depicted for interaction with cloud properties only (C minus REF) and for direct interaction with radiation together with the 15 interaction with cloud properties (CR minus REF) for the whole model domain. The T 2 m for case C at noon is systematically lower over the continent by −0.2 to −1 K. Isolated areas where T 2 m in case C is higher than in case REF can be attributed to small spatial and temporal differences in the distribution of the clouds. The difference in T 2 m at midnight is lower in comparison to noon time. The T 2 m at midnight in case C is 20 lower by −0.1 to −0.4 K over the continent. The maximum difference is in the eastern part of France. On average over the whole dust event (26-29 May) T 2 m in case C is systematically lower than in case REF by −0.2 to −0.5 K over the continent. The maximum differences in T 2 m are in the areas of the maximum average dust concentrations (Fig. 4). 25 The average difference in T 2 m from combined interaction of dust with clouds and radiation ( noon time of up to −1.8 K in the areas of the maximum dust concentrations. During midnight, the average difference in T 2 m is only negative in the areas with a high negative difference during daytime. In all other areas the difference is slightly positive with values up to +1 K locally. The average difference in T 2 m between the simulations CR and REF calculated for the whole period of the dust event shows systematically negative values in the order of −0.2 K for the whole continent with maxima up to −1 K locally. As mentioned before, an analysis of the weather forecast results for SW-Germany showed too high T 2 m (by up to +2 K) in comparison with station observations in the afternoon during the dust event. In the early morning hours, the predicted T 2 m converged with the observations. In Fig. 12  This simulation results indicate that the bias in the operational weather forecast can be attributed to the missing interaction of the dust particles with clouds and radiation. Although the observed bias of T 2 m in the weather forecast is two times larger than the simulated difference in T 2 m the temporal evolution is identical. The differences in the 20 magnitude can be explained by uncertainties in the cloud cover and in the exact position of the dust plume; e.g. the differences in T 2 m in Eastern France reach magnitudes comparable to the observed bias.
Both processes (the interaction with clouds and the direct interaction with the radiation) contribute in a similar magnitude to the simulated differences in T 2 m . Whether the direct impact on the radiation or the interaction with the clouds has the bigger impact on T 2 m varies depending on e.g. the spatial distribution of clouds and dust particles.

Conclusions
We investigated the impact of mineral dust particles on clouds, radiation, and atmospheric state during a strong Saharan dust event over Europe in May 2008, applying a comprehensive online-coupled regional model framework that explicitly treats particle microphysics and chemical composition. Sophisticated parameterizations for aerosol 5 activation and ice nucleation, together with two-moment cloud microphysics are used to calculate the interaction of the different particles with clouds depending on their physical and chemical properties. It is shown that the model framework is able to reproduce the measured aerosol mass concentrations during the dust event reasonably well and capture the aerosol 10 background before the arrival of the dust, as well as the spatial distributions of the clouds in comparison with satellite measurements. Dust particles act as CCN and IN in the atmosphere and can interact with clouds in several ways depending on the atmospheric conditions and cloud type. For temperatures lower than the level of homogeneous freezing the dust particles 15 have to compete with the homogeneous freezing of liquid aerosol particles for water vapor during the ice nucleation. Though the concentration of frozen dust particles is up to 100 l −1 during the ice nucleation events we found no significant impact on the number and mass concentration of ice crystals in this temperature range. The impact of the dust on cloud droplet number concentration was found to be low, 20 with just a slight increase in cloud droplet number concentration for both uncoated and coated dust. The number of activated dust particles was found to be in the order of 1 to 20 cm −3 , with higher numbers in the case of the coated dust particles. The additional activation of dust particles caused lower maximum supersaturations during the activation process, especially in the case of the coated dust particles. 25 Mineral dust particles are found to have the largest impact on clouds in a temperature range between freezing level and the level of homogeneous freezing, where they determine the number concentration of ice crystals due to efficient heterogeneous freezing of the dust particles and modify the glaciation of mixed phase clouds. Our simulations show that during the dust events ice crystals concentrations were increased twofold in this temperature range in comparison to a reference simulations which neglects the interaction of the dust with the atmosphere. As a consequence the liquid water fraction was reduced significantly with 60 % less grid points containing cloud water in the temperature range between 263 K to 235 K.

5
The strong increase in ice crystal number concentration has an influence on the optical properties of the ice clouds during the time of the dust event due to a significant decrease of effective radii of the ice crystals in the order of up to −25 µm on average. It was shown that depending on the number concentration of dust available for ice nucleation the incoming short-wave radiation at the surface was reduced up to −75 W m −2 at N d = 100 cm −3 . Consequently, a reduction in surface temperature in the order of up to 1 K over Europe during the day time arises from aerosol-cloud interactions. In the morning hours, the surface temperatures converge again between the different simulations. On average over the whole period of the dust event a reduction in surface temperature in the order of −0.2 K to −0.4 K was found in the eastern part of France 15 and Western Germany.
The simulated aerosol optical thickness of the dust particles was found to be in the order of 0.2 to 0.5 on average during the dust event over Europe. This caused an additional reduction in the incoming short-wave radiation which was found to be in the order of −80 W m −2 on 27 May. On the other days of the event the reduction was found Introduction On average the overall impact of the dust caused a reduction in surface temperature in the order of −0.2 to −0.5 K for most parts of France, Germany, and Italy during the dust event. The maximum difference in surface temperature was found in the East of France, the Benelux, and Western Germany with up to −1 K.