Intercomparison of aerosol-cloud-precipitation interactions in stratiform orographic mixed-phase clouds

Anthropogenic aerosols serve as a source of both cloud condensation nuclei (CCN) and ice nuclei (IN) and affect microphysical properties of clouds. Increasing aerosol number concentrations is hypothesized to retard the cloud droplet collision/coalescence and the riming in mixed-phase clouds, thereby decreasing orographic 5 precipitation. This study presents results from a model intercomparison of 2-D simulations of aerosol-cloud-precipitation interactions in stratiform orographic mixed-phase clouds. The sensitivity of orographic precipitation to changes in the aerosol number concentrations is analyzed and compared for various dynamical and thermodynamical 10 situations. Furthermore, the sensitivities of microphysical processes such as collision/coalescence, aggregation and riming to changes in the aerosol number concentrations are evaluated and compared. The participating models are the Consortium for Small-Scale Modeling’s (COSMO) model with bulk-microphysics, the Weather Research and Forecasting (WRF) model 15 with bin-microphysics and the University of Wisconsin modeling system (UWNMS) with a spectral ice-habit prediction microphysics scheme. All models are operated on a cloud-resolving scale with 2 km horizontal grid spacing. The results of the model intercomparison suggest that the sensitivity of orographic precipitation to aerosol modifications varies greatly from case to case and from model 20 to model. Neither a precipitation decrease nor a precipitation increase is found robustly in all simulations. Qualitative robust results can only be found for a subset of the simulations but even then quantitative agreement is scarce. Estimates of the second indirect aerosol effect on orographic precipitation are found to range from –19% to 0% depending on the simulated case and the model. 25 Similarly, riming is shown to decrease in some cases and models whereas it increases in others which implies that a decrease in riming with increasing aerosol load is not a robust result. Furthermore, it is found that neither a decrease in cloud droplet


Introduction
Orographic clouds form as moist air impinges on mountain ranges thereby potentially 10 forming orographic precipitation which is essential for landscape formation, agriculture and the hydrology of watersheds in many regions of the world (e.g., Roe, 2005). Understanding the various dynamical processes leading to orographic precipitation have been objectives of several field experiments such as the Mesoscale Alpine Programme (MAP, Bougeault et al., 2001) and the Improvement of Microphysical Parameterization 15 Through Observational Verification Experiment (IMPROVE-2, Stoelinga et al., 2003). However, the complexity of mixed-phase microphysical processes acting in orographic clouds and the sensitivity of orographic precipitation to changes in microphysics are not fully understood yet (e.g., Rotunno and Houze, 2007) and remain challenging to represent in numerical models of weather and climate on various scales. 20 One of the open questions regarding microphysical effects on orographic clouds is weather aerosol particles can significantly influence the amount and distribution of precipitation from these clouds. Orographic precipitation is hypothesized to be susceptible to aerosols because the time available to form precipitable hydrometeors in rising air parcels is constrained by the flow over the mountain. Increasing the aerosol num- 25 ber concentration is observed to lead to a shift of the cloud droplet spectrum towards smaller sizes (e.g., Twomey et al., 1984;Lowenthal et al., 2004) and, thus, retards the onset of the collision/coalescence process. Microphysical observations in mixed-phase clouds also suggest a decrease in the snowfall rate with increasing aerosol number concentration due to a reduced efficacy of the riming process implied by smaller cloud droplets (Borys et al., 2000(Borys et al., , 2003. Both effects together may modulate the amount and distribution of mixed-phase precipitation over mountainous terrain. However, several in-situ and airborne observations in orographic clouds show that, besides their ability to act as cloud condensation nuclei (CCN), aerosols are important for the initiation of ice in orographic clouds by heterogeneous ice nucleation (Field et al., 2001;Targino et al., 2006;Mertes et al., 2007;Cozic et al., 2008). Measurements from the Cloud and Aerosol Characterization Experiment (CLACE, Choularton et al.,10 2008) indicate the potential of aerosols to influence the partitioning between the liquid and the ice-phase in mixed-phase orographic clouds thereby controlling the microphysical growth regime of hydrometeors (Verheggen et al., 2007). The ability of aerosols to serve as ice nuclei (IN) in various heterogeneous nucleation modes (see Vali, 1985) depends on environmental factors such as temperature and ice-supersaturation but also 15 on a variety of aerosol properties like lattice structure, surface defects, chemical composition, etc. (Pruppacher and Klett, 1997;Baker and Peter, 2008;Hegg and Baker, 2009;Kulkarni and Dobbie, 2010). Based on the results of laboratory experiments, mineral dust and black carbon aerosols are deemed to be good ice nuclei (Roberts and Hallett, 1968;Gorbunov et al., 2001). Ice residual measurements during several field 20 campaigns corroborate the ability of dust (e.g., DeMott et al., 2004;Cziczo et al., 2004) and black carbon aerosols (e.g., Mertes et al., 2007;Cozic et al., 2008) to act as IN which for the latter one also suggests an anthropogenic impact of aerosols on clouds. However, laboratory studies also indicate that the mixing/processing state of aerosols can significantly alter the IN abilities which still is a matter of considerable debate. So 25 far, it is found that coating of particles with sulfuric acid or organics increases the onset supersaturation for heterogeneous nucleation in the deposition mode and lowers the freezing temperatures in the immersion mode. This renders coated, processed or aged substrates of black carbon or mineral dust less efficient for heterogeneous nucleation 10491 when compared with the uncoated samples (e.g., Möhler et al., 2005Möhler et al., , 2008; Kanji and Abbatt, 2006).
Statistical quantifications of aerosols affecting orographic precipitation based on paired rain gauge data are so far controversial and partially contradicting (e.g., Givati and Rosenfeld, 2004;Jirak and Cotton, 2006;Alpert et al., 2008, amongst others). 5 A particular difficulty arising in these studies is to separate the aerosol effect on precipitation from urban/heat island effects or other climatological changes (e.g., changes in circulation patterns, synoptic changes) and it is argued that inferences from these methods can also be misleading (Paldor, 2008;Muhlbauer, 2009). As a consequence, estimates of the effects of aerosols on orographic precipitation from observations regest that the amount of orographic precipitation is sensitive to the available CCN and show decreasing precipitation for increasing CCN (Chaumerliac et al., 1987;Thompson et al., 2004). Similarly, Lynn et al. (2007) find a decrease of orographic precipitation if the background aerosol conditions change from maritime to continental. However, Lynn et al. (2007) also report a sensitivity of the aerosol effect on orographic precipita-20 tion with respect to relative humidity. In a work on the classification of aerosol effects on precipitation Khain et al. (2008) conclude that orographic precipitation is decreased with increasing aerosol number if the environmental conditions at which orographic clouds form are relatively dry. Muhlbauer and Lohmann (2008) find decreasing orographic precipitation with increasing aerosol number concentrations for warm-phase 25 orographic clouds at relative humidities of 80% but the sensitivities crucially depend on the geometry of the mountain and the dynamical flow regime with lower or even reversed sensitivity for blocked flows.
Nevertheless, it is not clear how aerosols can influence precipitation in mixed-phase orographic clouds where the efficient formation of precipitable hydrometeors does not hinge on the warm-phase collision/coalescence process alone. Indeed, Saleeby and Cotton (2009) conduct simulations of winter-time orographic seeder-feeder clouds and found that precipitation tends to be redistributed rather than suppressed even if the rim-5 ing efficiencies in these clouds are decreased. In addition, simulations of mixed-phase orographic clouds by Muhlbauer and Lohmann (2009) suggest that the heterogeneous freezing ability of aerosols (i.e., the heterogeneous freezing mode) can control the sign and the magnitude of the aerosol effect on orographic precipitation. These results highlight the importance of the ice-phase for aerosol-cloud-precipitation interac-10 tion problems. Similarly, Zubler et al. (2010) find a reduction of orographic rainfall with increasing aerosol number concentration but also a tendency towards compensation or even increases in precipitation from snow and graupel at colder temperatures, thus, emphasizing that the sign and magnitude of the aerosol effect on orographic precipitation may strongly depend on the contribution from the ice-phase. 15 However, microphysical sensitivity studies in the past also show large impacts and variability of the results depending on the numerical approach followed in the models to treat microphysical processes (e.g., Morrison and Grabowski, 2007;Li et al., 2009). For example, bulk microphysical parameterisations typically rely on some assumed form of hydrometeor size distributions which introduces additional uncertainties. In 20 contrast, bin-resolving microphysical models explicitly calculate the particle size distributions without assumptions on the shape and slope parameter. Nevertheless, for computing microphysical collection processes (i.e., coalescence, aggregration and riming) both modeling approaches rely on assumptions for the collection efficiencies which are usually derived from laboratory experiments or numerical simulations (e.g., Khain et al., 2000).
Thus, the main goal of this paper is to analyze and inter-compare the sensitivities of different microphysical processes in mixed-phase orographic clouds and orographic precipitation to changes in the ambient aerosol conditions using several state-of-the-art 10493 numerical models and microphysical approaches. Special emphasis is placed on the sensitivity of the riming process.
Thus, we address the following science questions: -Is it possible to get qualitative robust estimates of the sensitivity of orographic precipitation to changing aerosol conditions for dynamically idealised conditions?

5
-Do increases in the aerosol number concentration affect the cloud droplet coalescence and the riming process in mixed-phase orographic clouds? -Do reductions in coalescence and riming from increases in aerosol number concentrations imply a reduction of precipitation from mixed-phase orographic clouds or are there compensating effects such as enhanced aggregation? 10 In order to reduce the complexity of the problem, the simulations are performed within an idealised setup of a 2-D flow over a mountain introduced at the Seventh WMO International Cloud Modeling Workshop (Morrison et al., 2009).
The paper is structured as follows: the participating models and microphysical parameterisations are introduced in Sect. 2. The design and the setup of the model 15 intercomparison is outlined in Sect. 3. In Sect. 4 the results of the different modeling approaches are presented. A discussion of the key findings and conclusions are presented in Sect. 5.

Model descriptions
The model-intercomparison is comprised of three participating models namely the Con-20 sortium for Small-Scale Modeling's (COSMO) model with coupled bulk two-moment aerosol-cloud microphysics COSMO, the Weather Research and Forecasting (WRF) model with bin microphysics WRF and the University of Wisconsin modeling system (UWNMS) with SHIPS microphysics UWNMS. Details of the microphysical parameterisations of each model are given in Table 1. 25

COSMO
The COSMO model is a non-hydrostatic, fully compressible limited-area mesoscale weather prediction model (http://www.cosmo-model.org, Doms and Schättler, 2002;Steppeler et al., 2003). The elastic equations are solved in a split-explicit time-splitting approach (Wicker and Skamarock, 2002) with a two time-level total variation diminish-5 ing (TVD) third order Runge-Kutta scheme in combination with a fifth order horizontal advection scheme. All moisture variables and aerosols are advected by a fourth order positive definite advection scheme after Bott (1989). A Smooth Level Vertical (SLEVE) coordinate system (Schär et al., 2002) is used in the vertical and a Rayleigh damping layer is introduced in the upper parts of the computational domain to minimize reflec-10 tions of vertically propagating gravity waves from the rigid upper model boundary.
Coupled aerosol-and cloud-microphysical processes are parameterized in a twomoment approach for five hydrometeor species (i.e., cloud droplets, rain, ice crystals, snow, graupel) and aerosols with various chemical compositions (i.e., sulfate, black carbon, organic carbon, sea salt, dust). The warm-phase microphysical processes of the 15 scheme include the nucleation of cloud droplets, condensation/evaporation of cloud droplets, autoconversion of cloud droplets to rain (i.e., the collision of cloud droplets leading to rain drops), accretion of cloud droplets by rain (i.e., the collection of cloud droplets by rain drops), self-collection of cloud-droplets and rain, evaporation of rain and the break-up of large rain drops. In the present version of the model the aerosol 20 activation is treated by following the κ-Köhler theory of Petters and Kreidenweis (2007). The ice-phase processes include heterogeneous nucleation, secondary ice nucleation, freezing, melting, vapor deposition, sublimation, riming, aggregation, collection and conversion to graupel (see Table 1 for details). For details of the microphysics parameterisation we refer the interested reader to Seifert and Beheng (2006), Muhlbauer and 25 Lohmann (2008) and references therein.

WRF
The Advanced Research WRF model version three (Klemp et al., 2007;Skamarock and Klemp, 2008) is used with the new upper gravity wave absorbing layer according to Klemp et al. (2008). A detailed bin microphysics scheme is implemented in WRF and described in Geresdi (1998), Rasmussen et al. (2002) and Geresdi and 5 Rasmussen (2005). The bin schemes use the multi-moment conservation method (Tzivion et al., 1987;Reisin et al., 1996) to insure the conservation of mass and number concentrations over 36 bins. Evolution equations are solved explicitly for the size distributions of cloud droplets, rain, ice crystals, snow and graupel (see Table 1 for details). The warm-phase microphysical processes include the activation of aerosols 10 to cloud droplets, diffusional growth of drops (i.e., condensation/evaporation), collision/coalescence of drops (i.e., autoconversion, accretion, and self-collection) and break-up of drops. Contrary to Rasmussen et al. (2002), aerosol activation and cloud droplet nucleation is treated here by calculating the supersaturation and critical diameter according to Köhler theory (Pruppacher and Klett, 1997;Khain et al., 2000). Hence, 15 40 more bins are added to the microphysics scheme to resolve the aerosol size distribution in the range 0.013-105 µm and the diffusional growth of the wetted aerosols is treated by the method described in Kogan (1991) and Teller and Levin (2006).
Ice-phase microphysical processes include diffusional growth of ice crystals, snow and graupel (i.e., deposition/sublimation), aggregation, riming and melting. Heteroge-20 neous nucleation of ice crystals is considered through freezing of supercooled droplets (immersion freezing, contact freezing) and due to deposition and condensation nucleation. Secondary ice formation includes the Hallett-Mossop ice multiplication process.

UWNMS
The University of Wisconsin non-hydrostatic modeling system (UWNMS) is a regional 25 mesoscale model that is designed for high resolution cloud resolving simulations employing a variable stepped topography (VST) coordinate system (Tripoli, 1992;Tripoli and Smith, 2009). The dynamical core of UWNMS is formulated to conserve vorticity, kinetic energy, and enstrophy with a quasi-compressible closure. A numerical filter is employed to control nonlinear instabilities and spurious numerical noise. For the advection of scalar variables a positive definite nonlinear piecewise parabolic method is used. 5 Cloud-microphysical processes are predicted with the Advanced Microphysics Prediction System (AMPS, Hashino and Tripoli, 2007, 2009a which simulates the detailed spectra of liquid, ice and aerosol particles. The liquid parameterisation is called SLiPS (Spectral Liquid Prediction System) which is a two-moment mass bin model. The size spectrum ranges from 0.1 µm to 0.5 mm radius with 30 bins employed. 10 The ice phase is parameterized in SHIPS which simulates 14 particle property variables over 20 two-moment mass bins. The main aim of SHIPS is to keep track of the growth history of ice particles and evolve the particles properties explicitly in the Eulerian dynamics model. Given the particle property variables, the average habit and type of ice particles is diagnosed for each mass bin. Hence, no categorization of ice 15 particles is necessary. The aerosols are handled by SAPS (Spectral Aerosol Prediction System) which simulates two moments of three lognormal distributions to describe Aitken, accumulation, and coarse modes of CCN and one monodisperse distribution of IN.
The current version of SHIPS includes heterogeneous nucleation, condensa-20 tion/evaporation, deposition/sublimation, coalescence, aggregation, riming, freezing, melting as well as the break-up of rain drops and large snow flakes (see Table 1 for details). The activation of aerosols is treated according to Köhler theory.

10497
3 Design of the model intercomparison

Model setup
All models use a 2-D computational domain with 401 gridpoints in the horizontal and 60 vertical levels. The model grid spacing is 2 km. In order to make the models dynamically as comparable as possible, radiation, turbulence and convection parame-5 terisations are switched off. Furthermore, a free-slip condition is imposed at the lower model boundary. Heterogeneous ice nucleation is treated by following the Meyers et al. (1992) formulation for deposition/condensation nucleation in all models.

Initial conditions
All experiments consider the development of an isolated orographic cloud and oro-10 graphic precipitation for a 2-D mountain flow. 1 A series of sensitivity experiments are conducted by changing aerosol initial conditions, temperature profiles and mountain heights as summarised in Table 2. The initial vertical profiles of temperature and dew-point temperature are shown in Fig. 1 and are given analytically by prescribing a sea-level temperature T SL , sea-level pressure p SL and dry Brunt-Väisälä frequency 15 N d following Clark and Farley (1984). The sea-level temperature is varied between T SL =273 K and T SL =280 K to generate two different thermodynamical conditions. The vertical profile of the relative humidity is given by prescribed constant with height within the first 10 km and increases linearly above to U=40 m s −1 at the model top.
The idealised bell-shaped topography has the form with h 0 denoting the peak mountain height located at x 0 (the middle of the computational domain) and a 0 denotes the half-width of the mountain. In the first set of experiments (linear hydrostatic mountain wave, SIM 1 and SIM 2) the mountain height is h 0 =800 m and the mountain half-width is a 0 =20 km. In the second set of experiments (blocked-flow, SIM 3 and SIM 4) the mountain height is h 0 =3000 m. The physical aerosol properties used to initialise the models are taken from obser-10 vations by Weingartner et al. (1999) and are prescribed vertically constant by means of lognormal aerosol size distributions shown in Fig. 2. The aerosol spectra are representative for mean remote continental winter/summer conditions in the Alps. Since lower aerosol number densities are observed during winter than during summer the mean winter aerosol spectrum is taken as the clean case (CC) and the mean summer 15 aerosol spectrum is taken as the polluted case (PC). Each aerosol spectrum satisfies a superposition of two lognormal size distributions of the form with the aerosol number densities N i , the count median radiir i and the geometric standard deviations σ i specified according to Table 3. Assuming a mean aerosol den-20 sity of ρ=1.5 g cm −3 yields aerosol mass densities of 2.0 µg m −3 during summer and 0.51 µg m −3 during winter. In case models are able to discriminate different aerosol chemical compositions we adopt the results of aerosol mass spectrometry (see Fig. 11 in Cozic et al., 2008 for details) obtained during the CLACE field campaign at the 10499 Jungfraujoch in Switzerland. Otherwise, aerosols are assumed to consist of ammonium sulfate with 100% solubility.

Intercomparison strategy
In order to compare the sensitivities of microphysical processes in mixed-phase orographic clouds and orographic precipitation to changes in the aerosol initial conditions 5 among the models the following metrics are used: The amount of precipitation from rain, snow and graupel as well as the orographic precipitation distribution are compared for all models after 10 h of simulation. Estimates for the total domain precipitation (TP), the spillover factor (SP) and the drying ratio (DR) are presented. Throughout this study, SP is defined as the leeward precipitation frac-10 tion (i.e., the ratio of leeward precipitation to total precipitation) according to Jiang and Smith (2003). DR is defined as the ratio of horizontally integrated total precipitation flux at the surface to the vertically integrated water vapor influx (e.g., Smith et al., 2003). SP and DR provide integral measures for orographic precipitation and do not depend on the choice of reference points at the topography. Both measures may be com-15 pared against observations from field campaigns, estimates from rain gauge networks and isotopic analyses of stream water or sap water on a climatological timescale (e.g., Smith et al., 2005). Cloud-microphysical fields (e.g., cloud droplet number concentration, liquid water content) are compared after 10 h of simulation. Averaged values of liquid water path (LWP) and ice water path (IWP) are compared among the models as 20 a function of time. Statistics of the dynamics of the individual models are computed for the computational domain excluding relaxation and damping zones. Furthermore, microphysical conversion rates for collision/coalescence, aggregation and riming are analysed and compared to get a better insight into aerosol-cloud-precipitation interactions on a process based level.

Simulations of orographic precipitation without flow blocking
For the first set of simulations a 2-D moist unblocked flow over a mountain is considered (i.e., a setup where linear theory predicts a hydrostatic mountain wave) for a warm case (SIM 1 CC, SIM 1 PC) and a cold case (SIM 2 CC, SIM 2 PC) (see Table 2 for details 5 on the setup).

Dynamics comparison
In order to show that the models are simulating the same dynamical state, the statistics of the wind velocities and the specific humidity are compared in Fig. 3. All models are able to simulate the hydrostatic mountain wave with an upstream region of flow deceleration and a downstream region of flow acceleration (not shown). All simulations start with a maximum specific humidity of approximately 7.3 g m −3 at the lowest model level according to the sounding shown in Fig. 1. The maximum values of specific humidity stay fairly constant throughout the simulations for all models except for UWNMS which develops a slight increase in the maximum specific humidity. This increase in maximum 15 specific humidity is caused by stronger evaporation processes on the upstream side of the mountain at the levels close to the surface. The integrated water vapor decreases over time according to the individual model's precipitation amount. For example, the domain integrated water vapor stays highest in WRF throughout the whole simulation because WRF is the model with the lowest 20 total precipitation rates as will be discussed later. Note that the slight differences in the domain integrated water vapor initially originate from differences in the model sampling volume. Since each model applies a different numerical setup (e.g., depth of the Rayleigh damping layer, width of the lateral relaxation zone) the model sampling volume varies slightly among the models. The maximum and domain integrated statistics 25 are obtained for the entire model domain excluding relaxation and damping zones.

10501
The horizontal wind speeds are prescribed at 15 m s −1 at the initial time and increase throughout the simulation to roughly 26 m s −1 in the downdraught region on the leeward side of the mountain after 10 h simulation. The differences in maximum horizontal wind velocities among the models are within a range of 3 m s −1 . However, in some models these maximum wind speeds originate from a region of wave breaking between 10 km 5 and 12 km and are not related to the downdraught region on the leeward side of the mountain. In the latter region, the agreement among the models is reasonable with maximum wind speeds of approximately 23 m s −1 in COSMO, 24 m s −1 in UWNMS and 27 m s −1 in WRF. According to Durran and Klemp (1983), one would expect that the remaining water vapor in the model simulations would lead to a damping of the gravity 10 wave amplitude and to lower horizontal wind speeds especially in the WRF simulation. Since the opposite is the case, the differences among the models in terms of horizontal wind speeds are likely caused by differences in either the numerical time integration scheme or the dynamical setup (e.g., depth of the Rayleigh damping layer, damping coefficients, upper model boundary condition) and not by the remaining moisture con-15 tent on the leeward side of the mountain. Nevertheless, the maximum deviations of the mean horizontal wind velocities are below 0.5 m s −1 among the models.
The vertical wind speeds are generally in good agreement with the maximum wind speeds increasing throughout the simulation from approximately 0.5 m s −1 after initial spin-up to approximately 0.8-1.0 m s −1 after 10 h. The maximum spread among the 20 models does not exceed 0.2 m s −1 in terms of maximum vertical velocities and is well below 1.0 cm s −1 for the mean vertical wind velocities.
In the simulation SIM 2 CC the initial sea level temperature is decreased in comparison to the simulation SIM 1 CC which leads to lower values of the maximum specific humidity of approximately 4.7 g m −3 . However, the horizontal and vertical wind velocities change only marginally in response to the decreased specific humidity and are comparable to the wind velocities of simulation SIM 1 CC (not shown).
In the simulations with increased aerosol number concentrations SIM 1 PC and SIM 2 PC, respectively, the horizontal and vertical wind velocities change slightly due to dynamical feedbacks of aerosol-cloud-precipitation interactions (not shown) but these differences are much smaller than the differences between the individual model simulations and, thus, are not investigated further.

Clouds and precipitation comparison
Despite the reasonable agreement of the dynamical states of the model considerable 5 differences exist among the models in terms of cloud structure, cloud microphysics and precipitation. All models simulate a low-level orographic cloud attached to the upstream side of the mountain and an upper-level orographic wave cloud in a region of wave breaking on the leeward side of the mountain. Figures 4 and 5 show the cloud fields for the simulation SIM 1 CC for all models averaged over the 10 h simulation 10 time. The depth of upstream low-level cloud varies among the models between approximately 3 km and 4 km after 10 h. Considerable discrepancies exist especially in the partitioning of the liquid and the ice in the low-level orographic cloud. For example, COSMO and UWNMS both simulate cloud layers with sustained mixed-phase conditions sandwiched by a partly supercooled liquid cloud at lower levels and a pure ice 15 cloud aloft. These mixed-phase cloud layers are deeper in COSMO than in UWNMS. In contrast, in WRF the entire cloud consists of liquid with cloud water mixing ratios up to approximately 0.6 g m −3 . Due to the lack of ice in the WRF simulation, the WRF model produces also deeper cloud water fields than the other models (

10503
Despite the reasonable agreement in the cloud droplet number concentrations there exists remarkable variability among the models with respect to the mass mixing ratios of the different hydrometeor types (Fig. 5). COSMO and UWNMS both show contributions from cloud liquid and cloud ice at lower levels but values are considerably higher in UWNMS than in COSMO. 5 The differences in the liquid to ice mass partitioning in the orographic cloud also affect the development of precipitable hydrometeors. For example, WRF shows mainly contributions from rain and snow with negligible small contributions from graupel. In COSMO and UWNMS precipitation shows contributions from snow generated in the upper regions of the cloud as well as graupel and rain in the lower levels of the clouds. 10 These differences suggest that the contributions to rain that originate from the collision/coalescence process of clouds droplets are stronger in WRF than in COSMO and UWNMS where contributions from frozen hydrometeors are dominant. These frozen hydrometeors are generated aloft, start melting below the freezing level and eventually lead to rain. Due to the diminishing role of the ice-phase in the WRF simulation, the 15 collision/coalescence process is more important for the rain formation in WRF than in the other models. Furthermore, COSMO and UWNMS produce considerable amounts of snow from aggregation of ice crystals and graupel is produced by riming in regions of sufficient supercooled liquid water. These differences in the precipitable hydrometeors fields have implications for the amount and distribution of orographic precipitation as 20 will be discussed later.
The upper-level wave clouds are similar among the models but cloud fields are varying regarding the ice water content and the spatial extent of the cloud. The maximum ice water mixing ratios agree reasonably between COSMO and UWNMS but are slightly smaller in WRF. The horizontal cloud extent of the upper-level wave cloud 25 is largest in UWNMS. These discrepancies may be attributable to the production of smaller ice crystals and their inherent lower terminal fall velocities in UWNMS and to the different ice aggregation efficiencies used in the models. For example, in the COSMO simulation snow develops in the upper-level cloud whereas there is no snow in the WRF and UWNMS simulations in this region. However, the differences in the upper-level cloud are not of further interest here because the contribution of this cloud to the orographic precipitation in the models is negligible small. Figure 6 shows the time series of liquid water path (LWP) and ice water path (IWP) in the models averaged over the mountain slopes (i.e., the horizontal distances from -40 5 to 40 km) for simulation SIM 1 CC. The time series confirms that WRF is lacking ice compared to the other models and that liquid water is dominant during the simulation. In contrast, COSMO and UWNMS both show lower LWP but higher IWP and, thus, sustain more ice in the low-level orographic cloud. Figure 7 shows the cloud droplet and ice crystal number concentrations for the 10 simulation with increased aerosol number concentration SIM 1 PC averaged over the 10 h simulation time. In SIM 1 PC the cloud droplet number concentrations are increased in all models but the variability among the models is larger than in the simulation SIM 1 CC with cloud droplet number concentrations ranging from approximately 150 cm −3 (WRF) to almost 440 cm −3 (COSMO). The mean cloud droplet size averaged 15 over the 10 h simulation decreases relative to the clean case simulation SIM 1 CC in all models. In contrast, there is only little change in the ice crystal number concentrations compared to the clean case simulation SIM 1 CC. However, the result that the ice crystal number concentrations are not much affected in the simulations may be related to the oversimplified treatment of heterogeneous ice nucleation and in particular to the 20 fact that the ice initiation is independent from the aerosol chemical properties. The differences in the LWP and the IWP between the low aerosol simulation SIM 1 CC and the high aerosol simulation SIM 1 PC are shown in Fig. 8. Despite the considerable variability of the microphysical properties of the simulated orographic cloud all models agree qualitatively on the result that on average the LWP is increased 25 with increasing aerosol number concentrations but there is little quantitative agreement.
Regarding the IWP all models show slight changes throughout the 10 h simulation but on average changes in IWP are generally small and positive for WRF but negative for COSMO and UWNMS.

10505
The intercomparison of precipitation is shown in Fig. 9 for the simulation SIM 1 CC. The orographic precipitation is in good agreement in COSMO and UWNMS regarding the peak and the shape of the precipitation distribution as well as the total precipitation in the domain (see Table 4). In all models orographic precipitation maxima are found on the upslope side of the mountain with slight displacements of the peaks towards 5 the top of the mountain in the WRF simulation. However, the amount of peak and total precipitation simulated by WRF is much smaller than in the other models which implies that the precipitation regime simulated by WRF is less efficient than in both of the other models. Furthermore, there is generally little agreement among the models on how different hydrometeor categories contribute to the overall precipitation distribution. For example, precipitation at the surface originates almost entirely from rain in the WRF simulation whereas the other models also show contributions from graupel (and snow in the case of UWNMS). Most of the total precipitation in COSMO is made up of rain with some but little contribution from graupel whereas in UWNMS the contribution from graupel is much stronger. Moreover, the fallout of graupel is faster in UWNMS and 15 occurs already on the upslope side of the mountain whereas in COSMO the precipitation distribution of graupel is centered around the mountain peak. The reasons for the relative shift in the precipitation from graupel may be related to the different terminal fall velocities of graupel used in the models but may also be due to the varying strength of the simulated ice-phase or the thresholds used to convert rimed snow flakes to grau-20 pel. However, the bottom line of the intercomparison of simulation SIM 1 is that the presence of an ice-phase in the upper part of the low-level orographic cloud and the sustained mixed-phase conditions lead to a more efficient precipitation regime (i.e., a seeder-feeder cloud type) in COSMO and UWNMS with more upslope precipitation and considerable contributions from graupel. Increasing aerosol number concentrations in 25 simulation SIM 1 PC leads qualitatively to a reduction of total domain precipitation in all models but the sensitivities vary greatly among the models. WRF depicts the highest sensitivity of precipitation with respect to changing aerosol mainly due to a strong reduction in rain formation and little contributions from ice-phase hydrometeors. The sensitivity of the warm rain production is weaker in the case of COSMO and UWNMS. The contribution to rain in the simulations are dominated by the melting of frozen hydrometeors formed aloft. However, relative to COSMO and UWNMS the rain formation in WRF is more dependent on the warm-phase collision/coalescence process. Furthermore, COSMO and UWNMS show increases in graupel which partly compensate 5 for the loss caused by the reduction of rain. Hence, the high sensitivity of WRF to changes in aerosol number may be explained by the diminishing role of the ice-phase and the lack of graupel formation in this specific simulation. Overall, WRF shows the highest sensitivity with a total precipitation reduction of approximately 19% compared to the clean case simulation but also shows the smallest total precipitation amounts. 10 COSMO and UWNMS both simulate much more precipitation and depict much weaker sensitivities with respect to changes in aerosol number concentration of 11% and 1%, respectively (see Table 4 for details). All models except UWNMS tend to simulate higher values of SP meaning that relative to the low aerosol simulation more precipitation is falling on the leeward side of the mountain in the high aerosol simulation (see 15   Table 5 for details). DR tends to decrease with increasing aerosol number concentration and the relative magnitude of the changes are ranked according to the decrease in total precipitation in the individual model simulations. Hence, relative changes in DR are more pronounced in WRF than in COSMO whereas in UWNMS DR is virtually unchanged (see Table 6 for details).

Microphysics comparison
Comparing the microphysical processes among the models elucidates why the models disagree on the contributions of precipitation from rain and graupel and sheds some light on why the sensitivities with respect to aerosols are different. Figure 10 shows the amount of liquid and ice mass density that is converted to 25 precipitable hydrometeors through the microphysical processes collision/coalescence, aggregation and riming. In Sim 1 COSMO and UWNMS both agree on the result that riming is the dominant process that leads to precipitation in the simulation and that 10507 the riming rates are increased if the aerosol number concentrations are increased. As a consequence, the amount of precipitation from graupel is increased with increasing aerosol number concentrations in both models. In contrast, in the WRF simulation the contribution from riming and aggregation is comparable and riming is decreased if the aerosol number concentrations are increased. Furthermore, aggregation is increased 5 in the simulations with high aerosol number concentrations in COSMO and WRF but is slightly decreased in the UWNMS simulation. Moreover, both models show a comparably weak contribution to precipitation from the collision/coalescence process but indicate a decrease in the collision/coalescence if the aerosol number concentrations are increased. Note that the values for collision/coalescence in the WRF simulations 10 are missing. However, the decrease in collision/coalescence is rather subtle in the case of UWNMS because the reduction in the cloud droplet coalescence is almost completely offset by the accretion of cloud droplets by rain. Thus, the accretion by rain may be an important microphysical process with a potential to reduce the overall sensitivity of the cloud droplet collision/coalescence process in mixed-phase clouds if 15 rain drops can be produced by melting.

Sensitivity studies with respect to temperature
In the second set of simulations (SIM 2 CC and SIM 2 PC) the same dynamical regime is simulated with a colder surface temperature in the initial thermodynamical profile (see Fig. 1 and Table 2 for details). The dynamical states of the models are only 20 marginally affected by this temperature change and the inter-model comparison is similar to the previous set of simulations. As in the previous set of simulations, all models show on average an increase in LWP in the simulations with higher aerosol number concentrations but only little change in the IWP. Figure 11 shows the comparison of precipitation for the set of simulations with colder 25 surface temperature SIM 2 CC and SIM 2 PC. The variability among the models in the orographic precipitation distribution is larger than in the warm set of simulations with little agreement regarding the shape of the precipitation distribution or the amount of precipitation. Total precipitation is highest in UWNMS and lowest in WRF. However, the spatial precipitation distribution is more narrow for COSMO than for the other models which leads to the highest peak precipitation rates. Most of the precipitation reaches the surface in form of snow and graupel and none of the models produces any significant amount of rain. As in the previous case, there is little agreement on the 5 partitioning of precipitation to snow or graupel which in turn explains the discrepancies in the simulated precipitation distribution. These discrepancies are also reflected in the microphysical comparison of the models (see Fig. 12). In COSMO and UWNMS aggregation is the prevalent microphysical process and most of the precipitation is from snow flakes. On the contrary, in the WRF simulation riming dominants aggregation but 10 the rimed snow flakes are not converted to graupel as in the other models. Contributions from the collision/coalescence process are negligible in all models. Riming is decreased in COSMO and WRF but is slightly increased in UWNMS. The sensitivity of orographic precipitation to changes in the aerosol number concentrations as simulated by the models is generally weaker than in the previous simulations with warmer temperatures (see Table 4 for details). In COSMO and WRF total precipitation is decreased by 4% and 5%, respectively, whereas in UWNMS total precipitation budgets are virtually unchanged (a subtle increase). COSMO and WRF both show higher values of SP in the high aerosol simulation meaning that some of the precipitation (i.e., snow in this particular case) is redistributed towards the leeward side of the 20 mountain. The DR is only little affected according to the change in the precipitation budgets of the models.

Simulations of orographic precipitation with flow blocking
In the next set of simulations (SIM 3, SIM 4) the height of the mountain is increased while the thermodynamical initial conditions are kept constant. This generates a flow 25 regime with upstream flow blocking and severe downslope winds on the leeward side of the mountain.  Figure 13 shows a comparison of the time evolution of the specific humidity and wind speeds in the models. Generally, the differences in maximum horizontal and vertical wind speeds are much higher in this case than in the unblocked flow case because the downslope winds on the leeward side of the mountain and the regions of wave break- 5 ing are very turbulent. The largest differences in maximum and minimum wind speeds occur in the regions of downslope winds on the leeward side of the mountain below ridge height and in the zones of wave breaking aloft. The strength of the downslope winds and the breaking waves are to a large extent controlled by the amount of precipitation on the upstream side of the mountain and the water vapor budgets of the models.

Dynamics comparison
Nevertheless, the winds on the upstream side of the mountain are in relatively good agreement and these are the ones that are important for cloud formation. All models develop a blocked air layer on the upstream side of the mountain with a depth of approximately 2800 m. In this layer the horizontal wind speeds are decelerated to roughly 10 m s −1 . The UWNMS model develops slightly stronger blocking than COSMO and 15 WRF with a stagnant air layer at low levels and even slight reverse circulations. These differences may be attributable to differences in the numerics of the model or the VST coordinate system used in UWNMS.

Clouds and precipitation comparison
Similar to the previous case with unblocked flow, the cloud fields simulated by the 20 models in the blocked flow case correspond qualitatively to the classical picture of a seeder-feeder cloud system with a liquid cloud at lower levels and an ice cloud aloft. Figure 14 shows the comparison of different hydrometeor fields in all three simulations of case SIM 3 CC averaged over the 10 h simulation time. The simulated cloud is roughly 6.5 km deep in all models but there are discrepancies in the partitioning of the 25 cloud water between the liquid and the ice-phase. In COSMO most of the cloud consists of supercooled liquid water down to temperatures of -20 • C whereas, for example, in UWNMS the liquid-phase is confined to levels below approximately 1 km and above the preexisting ice crystals grow by vapor deposition on the expense of the liquid-phase (i.e., the Bergeron-Findeisen process). Ice crystals are predominantly found between 2 km and 7 km but are rapidly aggregated to snow flakes in all models. The snow flakes then sediment into the lower layers where they start riming to graupel which enhances 5 the orographic precipitation greatly. Most of the precipitation originates from snow and graupel with some contributions from rain that forms below the freezing level. In the simulations SIM 4 the cloud fields are qualitatively similar to case SIM 3 (not shown) except that the ice part of the cloud starts at much lower elevations and the contributions from the liquid-phase are small. The orographic precipitation distributions are 10 shown in Fig. 15 for the case with warmer temperatures (SIM 3) and in Fig. 16 for the case with colder temperatures (SIM 4), respectively. Despite the fact the dynamical differences in the individual model simulations are generally larger in the case with flow blocking than in the unblocked flow case the amount of total precipitation is in better agreement with some but overall less variability among the models. All models agree on the qualitative shape of the orographic precipitation distribution (Fig. 15). There is good correspondence of the spatial distribution of rain whereas for snow and graupel the location of the peaks vary presumably due to the different fall velocities for snow and graupel used in the models. Furthermore, there are discrepancies regarding the contributions from snow and graupel to the total precipitation whereas there is little vari-20 ation in terms of rain. COSMO is the model with most precipitation whereas UWNMS is the model with least precipitation. Similar results are found in the simulations SIM 4 with the distinction that contributions from rain are negligible in all models (Fig. 16). There is a good agreement in terms of total precipitation and the result that most of the precipitation is due to snow. Again, there is some discrepancy in the precipitation from 25 graupel particles.
The sensitivity of orographic precipitation with respect to changes in aerosol number concentration is much smaller and almost negligible for the blocked flow case in both the warm and the cold subset of the simulations. All models show a tendency towards 10511 reduced total orographic precipitation but the sensitivities do not exceed 3% in any of the simulations (see Table 4 for details). Some of the simulations show an increase of SP in the polluted case which points to a slight redistribution of precipitation towards the leeward side of the mountain. The DR shows only a weak signal in most of the simulations (see Tables 5 and 6 for details.)

Microphysics comparison
Figures 17 and 18 show the microphysical rates for the simulations SIM 3 and SIM 4, respectively. Two (COSMO and WRF) out of three models suggest that riming is the prevalent microphysical process for producing precipitation from the cloud whereas in the UWNMS model aggregation and riming are equally important. The collision/coalescence process contributes only little to the overall precipitation in all models. Riming is found to decrease with increasing aerosol number concentrations in COSMO and UWNMS but is increased in WRF. Furthermore, all models show an increase in aggregation with increasing aerosol concentrations. The collision/coalescence process is only little affected because the decrease in the coalescence of cloud droplets is largely 15 compensated by the accretion of cloud droplets by rain that is produced from melting ice hydrometeors. Similar results are found in the set of simulations with colder temperature SIM 4. However, in this case all models suggest that aggregation is the dominant microphysical process followed by riming and a negligible small role of collision/coalescence. 20

Discussion and conclusions
A model intercomparison of aerosol-cloud-precipitation interactions in stratiform orographic mixed-phase clouds is conducted and the sensitivities of orographic precipitation to changes in aerosol number concentrations are analyzed and compared among the models. Furthermore, the sensitivity of microphysical processes such as colli-past topography are performed for two different flow regimes namely a linear mountain wave flow regime without significant orographic flow blocking and a flow regime with strong orographic flow blocking. Additional sensitivity studies are conducted with respect to temperature and the mode of heterogeneous ice nucleation. The simulations are initialized with aerosol size distributions typical and representative for remote-5 continental conditions in the Swiss Alpine region.
The simulated cloud types resemble orographic seeder-feeder clouds with sustained mixed-phase conditions sandwiched between layers of partly supercooled liquid below and layers of ice aloft. These types of clouds are a well known prototype of orographic clouds at mid-latitudes and are climatologically important for explaining observed oro-10 graphic precipitation amounts especially over small mountain ranges (e.g., Browning et al., 1974;Smith, 1989).
The results of the model intercomparison suggest that the sensitivity of orographic precipitation to changes in the aerosol number concentrations varies from case to case but also from model to model. While a case dependent sensitivity of aerosol-cloud- 15 precipitation is generally found in many other studies (e.g., Levin and Cotton, 2009;Khain, 2009) the variability among different models and microphysical approaches for the same case is less recognized. In most of the simulations a decrease of orographic precipitation with increasing aerosol number concentrations is found whereas in some others orographic precipitation is either only marginally affected or slightly increased. 20 Thus, neither a precipitation decrease nor a precipitation increase is found robustly in all simulations. Qualitative robust results can only be found for a small subset of the simulations. However, even for this small subset of simulations quantitative estimates of the aerosol sensitivity of precipitation varies greatly among the models.
Estimates for the indirect aerosol effect on orographic precipitation range from -19% 25 to 0% depending on the simulated case and the model. Nevertheless, a tendency towards smaller precipitation sensitivities is found for the cases with high mountain range or low temperatures. Orographic precipitation is found most susceptible to changes in aerosol number concentrations in the case of a small mountain range and warm 10513 temperatures. These sensitivities can be understood qualitatively from two different points of view. From a microphysical point of view it may be argued that both factors, increasing the height of the mountain range or decreasing the ambient temperature, strengthen the role of the ice-phase in the simulations thereby reducing the aerosol susceptibility of the cloud. On the other hand, similar sensitivities are also found in 5 simulations of warm-phase orographic clouds with no contribution from the ice-phase (Muhlbauer and Lohmann, 2008). Thus, from a dynamical point of view, orographic clouds are found less susceptible to aerosol modifications in the blocked flow case because air parcels in these clouds do not obey the classical picture of a flow over the mountain that constrains the available time for the initiation of precipitation. Hence, one 10 may expect an indirect aerosol effect on precipitation to be less pronounced or even inverted in regions with high terrain (e.g., the Alps or Rocky Mountains) or in regions where ice-phase microphysics play climatologically an important role for orographic precipitation.
There is an agreement among the models that on average the LWP is increased 15 with increasing aerosol number concentrations. In some cases an increase in the IWP is also found but this feature is less robust and not supported by all models. Most of the simulations are not particularly sensitive to the aerosol modification because the cloud droplet collision/coalescence process is overall only slightly affected in the simulated orographic clouds. The reasons are twofold: firstly, in the simulated cases with 20 cold temperatures the contributions from the coalescence process to the orographic precipitation are found to be negligible small and, thus, changes in this particular microphysical process are not important for the precipitation formation. Secondly, in the simulated cases with warm temperatures rain drops are mainly produced by melting of ice hydrometeors generated aloft and not by the coalescence of the cloud droplets. 25 Thus, the efficient production of rain drops in mixed-phase clouds does not hinge on the cloud droplet coalescence process alone. These rain drops fall then through layers of higher liquid water content and collect cloud droplets efficiently. It is noted that small or inverse sensitivities of precipitation to changes in aerosol number concentrations are also reported in cases where rain drops are generated in a cloud efficiently by other means such as giant CCN (e.g., Feingold et al., 1999;Rosenfeld et al., 2002) or due to vertically decreasing aerosol number concentrations that support rain drop formation through coalescence in the upper regions of a cloud (e.g., Muhlbauer and Lohmann, 2008). These mechanisms can be operational and effective even without any aid from 5 the ice-phase. The simulations conducted in this study also suggest that a decrease in the collision/coalescence process does not necessarily imply a decrease in precipitation. The reason is that riming can be increased if more cloud liquid water is available for collection in the case of high aerosol number concentrations. It is emphasised that riming 10 is found to decrease as well as increase in the simulations depending on the case and the model. Hence, there is no robust conclusion whether riming is increased or decreased with increasing aerosol number concentrations in the simulated orographic clouds. Furthermore, a simulated decrease is riming does not imply a decrease in precipitation because the decrease in riming can be compensated by an increase in 15 aggregation.
The disagreement among the models regarding the sensitivity of microphysical processes and the eventual macrophysical outcome in terms of precipitation is linked to the uncertainties in representing microphysical processes especially in mixed-phase and ice-clouds. Some of the persistent discrepancies found in all simulations of the 20 model intercomparison are related to the partitioning of cloud water between the liquid and the ice-phase, the simulation of sustained mixed-phase conditions and the partitioning of precipitation between snow and graupel. Some of these problems reflect the incomplete and limited physical understanding of cloud microphysics (e.g., initiation of ice in mixed-phase clouds by heterogeneous ice nucleation), the fact that many 25 microphysical processes are not very well constrained by observations (e.g., uncertainty of the ice aggregation efficiencies) and the artificial treatment of microphysical processes in numerical models (e.g., the treatment of rimed snow and the conversion to graupel). The latter problem might be mitigated by avoiding artificial hydrometeor 10515 categorisation in cloud-microphysics schemes as suggested by Hashino and Tripoli (2007) or Morrison and Grabowski (2008). However, more observations from laboratory studies and field campaigns are needed to improve the current understanding of aerosol-cloud-precipitation interactions and to better constrain cloud-microphysical parameterisations. 5 Acknowledgements. The first author is grateful to Axel Seifert for sharing his cloudmicrophysics code. We thank the World Meteorological Organization (WMO) for sponsoring the 7th WMO Cloud Modeling Workshop held in Cozumel, Mexico, at July 14-17, 2008 and all the participants for fruitful discussions, suggestions and comments. We acknowledge the following institutions for providing computing time: The European Center for Medium-Range 10 Weather Forecasts (ECMWF) within the special project Cloud Aerosol Interactions, the German Weather Service (DWD), the National Science Foundation (NSF) and the National Center for Atmospheric Research (NCAR). preparation, 2009b. 10497 Hegg, D. A. and Baker, M. B.: Nucleation in the atmosphere, Rep. Prog. Phys., 72, 056801, doi:10.1088/0034-4885/72/5/056801, 2009 An improved approach to calculating terminal fall velocities of plate-like crystals and graupel, J. Atmos. Sci., 44, 1088Sci., 44, -1099Sci., 44, , 1987 Hobbs, P., Easter, R., and Fraser, A.: A theoretical study of the flow of air and fallout of solid precipitation over mountainous terrain: Part II. Microphysics, J. Atmos. Sci., 30, 813-823, 1973. 10492 Jiang, Q. and Smith, R.: Cloud Timescales and Orographic Precipitation, J. Atmos. Sci., 60, 1543-1559. 10500  Seifert andBeheng (2006) Pitter (1977) Hallgren and Hosler (1960) Riming efficiencies Lew et al. (1986), Mitchell (1990) Lin et al. (1983 Hashino and Tripoli (2007) Terminal fall speeds ice Hobbs (1974) Bohm (1992) Terminal fall speeds snow Locatelli and Hobbs (1974) Passarelli and Srivastava (1979) Bohm (1992 Terminal fall speeds graupel Seifert andBeheng (2006) Bohm (1992) 10525 Table 2.
Parameter specifications for the idealised set of simulations. All simulations are conducted with a sea-level pressure p SL =1000 hPa, a dry Brunt-Väisälä frequency of N d =0.011 s −1 and a mountain half-width of a 0 =20 km. Aerosol conditions are prescribed according to a clean case (CC) or a polluted case (PC).     Table 2 for details on the setup). tribution is based a mean aerosol spectrum for winter and summer conditions, respectively. The winter aerosol spectrum is taken as the clean case (CC, solid) and the summer aerosol spectrum is taken as the polluted case (PC, dashed).   (QNI) for COSMO, WRF and UWNMS averaged over 10 h simulation. The contour lines show the temperature (red) and the potential temperature (gray). Units are cm −3 for cloud droplets and l −1 for ice crystals, • C for temperature and K for potential temperature. Only part of the computational domain is shown.