Aerosol–cloud–precipitation interactions during a Saharan dust event – A summertime case-study from the Alps

Changes in the ambient aerosol concentration are known to affect the micro-physical properties of clouds. Especially regarding precipitation formation, increasing aerosol concentrations are assumed to delay the precipitation onset, but may increase precipitation rates via convective invigoration and orographic spillover further downstream. In this study, we analyse the effect of increased aerosol concentrations on a heavy precipitation event observed in summer 2017 over northeastern Switzerland, an event which was considerably underestimated by the operational weather forecast model. Preceding the precipitation event, Saharan dust was advected towards the Alps, which could have contributed to increased precipitation rates north of the Alpine ridge. To investigate the potential impact of the increased ambient aerosol concentrations on surface precipitation, we perform a series of sensitivity simulations using the Consortium for Small-scale Modeling (COSMO) model with different microphysical parametrizations and prognostic aerosol perturbations. The results show that the choice of the microphysical parametrization


INTRODUCTION
Precipitation originating from orographic clouds is crucial for the local hydrology, ecology, and climate in Alpine regions (Roe, 2005).Orographic precipitation formation can be influenced by the terrain, the prevailing meteorological conditions such as blocking or turbulence, the atmospheric stability, as well as cloud microphysics (Rotunno and Houze, 2007).However, it is difficult to disentangle these competing factors acting on orographic clouds, and their representation in numerical weather prediction (NWP) models remains challenging.This especially applies for aerosol-cloud-precipitation interaction, despite numerous studies on the topic (Choudhury et al., 2019).Aerosol particles shape microphysical structures and processes in clouds by acting as cloud condensation nuclei (CCN) and/or ice nucleating particles (INPs).Their ultimate impact on precipitation, however, is the result of a complex chain of processes that is briefly discussed in the following.CCN initiate cloud droplet formation by providing surfaces for the atmospheric water vapour to condense on and thereby control the number and size of cloud droplets and the subsequent precipitation formation.In the early cloud stages, high CCN concentrations lead to the formation of a higher number of cloud droplets, and thus enhance the condensation rate and the accompanying latent heat release, as compared to pristine conditions.The resulting increase in buoyancy strengthens the updraught in the cloud, a process that is referred to as convective invigoration (Williams et al., 2002;Khain et al., 2005;Wang, 2005;Tao et al., 2007;Chen et al., Chen et al.).The high cloud droplet number concentration also leads to smaller cloud droplets and an overall narrower cloud droplet size spectrum (Squires and Twomey, 1960), decreasing the efficiency of the collisional growth process.Since growth by collision-coalescence is crucial for the hydrometeors to reach sedimentation size in warm-phase clouds, the formation of precipitation is curtailed and delayed, or can even be suppressed (Borys et al., 2003;Lowenthal et al., 2004;Rosenfeld, 2005;Muhlbauer and Lohmann, 2009;Cheng et al., 2010;Heikenfeld et al., 2019).This slow-down of the sedimentation of larger particles allows for an extended period of diffusional growth of the droplets within the cloud, during which the supplemental latent heat release again adds to the buoyancy (Rosenfeld et al., 2008;van den Heever et al., 2011).However the prolonged diffusional growth produces a higher column loading of condensed water (Khain et al., 2005;Wang, 2005;Martins et al., 2011), which has a dampening effect on buoyancy due to the weight of the additional hydrometeors (Sheffield et al., 2015).This reduction in buoyancy was found to play a minor role in the early stages of a convective cloud, but becomes more important in the dissipating stage (Storer et al., 2014).The temporal delay of the formation of precipitation in a dynamical environment also entails its displacement further downstream, as the hydrometeors move along with the steering flow until they have grown large enough to sediment and fall out.This phenomenon was found in observations as well as model simulations (Givati and Rosenfeld, 2004;Zubler et al., 2011a;Saleeby et al., 2013), and is commonly referred to as the spillover effect.In an orographic setting, this spillover effect implies a decrease in windward precipitation while moisture transport over the mountain ridge is enhanced, which will eventually result in an increase of precipitation on the leeward side (Lynn et al., 2007;Muhlbauer and Lohmann, 2008;Xiao et al., 2014).
The so far described mechanisms solely affect rain formation in the warm-phase cloud processes.However, as the cloud grows to higher altitudes, it eventually enters the freezing regime, where the impact of aerosol particles on ice and mixed-phase processes has to be considered as well.On the one hand, INPs play an important role by facilitating the formation of ice crystals at temperatures above the homogeneous freezing point (Ladino Moreno et al., 2013;Kanji et al., 2017;Paramonov et al., 2020).Ice crystals then grow rapidly through deposition of water vapour for a wide range of updraughts.At updraught velocities between approximately 0 and 2 m⋅s −1 (Korolev, 2007), ice crystals grow at the expense of evaporating cloud droplets through the Wegener-Bergeron-Findeisen process (WBF; Wegener, 1911;Bergeron, 1935;Findeisen, 1938).Once the ice particles are large enough, they start to sediment and grow further by collision with other ice particles (accretion) or upon collision with supercooled cloud droplets (riming) (Harimaya, 1975;Lowenthal et al., 2011).In an orographic setting, the impinging flow on the mountain forces moist air to ascend, providing a constant source of moisture for liquid drops and ice particles to grow.Under conditions of strong orographic forcing, mixed-phase clouds (MPCs) can be persistent (e.g., Lohmann et al., 2016).Suppose warm mixed-phase orographic clouds (cloud-top temperatures > −20 • C) are limited in ice particle number concentration.In this case, an enhancement in INPs in the liquid-rich environment will favour the riming process over deposition in the growth of ice particles, reducing the importance of the WBF process (Fan et al., 2017).Drastically enhancing the INPs results in riming and deposition becoming equally important.If cloud tops are colder (temperatures < −20 • C), and the INPs are enhanced, more ice crystals form, which limits the available cloud liquid water.In this case, depositional growth becomes more efficient than riming at low cloud temperatures (Fan et al., 2017), enhancing the importance of the WBF process.Dedekind et al. (2021) showed similar results where an enhancement in the ice crystal number concentrations significantly reduces cloud liquid, favouring the growth of ice particles by deposition over riming at low cloud temperatures.However, the redistribution of the growth processes lead to insignificant changes in the regional precipitation amount (Glassmeier and Lohmann, 2018).Because the ice phase is highly efficient in precipitation formation, an enhancement in INPs can cause an early glaciation of a cloud and change the onset of precipitation.Fan et al. (2017) found that the increase in the INPs intensifies the growth of snow precipitation, thereby enhancing precipitation on the windward slope of the Sierra Nevada.As a consequence, fewer hydrometeors will be transported to higher levels and beyond the mountain ridge (Lohmann and Hoose, 2009;Ault et al., 2011;Creamean et al., 2013;Fan et al., 2017).This effect can reinforce itself, as the reduced upward transport of cloud water leads to less latent heat release in the higher levels of the cloud, which eventually weakens the convection and inhibits the formation of deeper clouds (Khain et al., 2005;Koren et al., 2005;Wang, 2005;Ten Hoeve et al., 2012).
On the other hand, the temperature and corresponding height at which the cloud droplets start to freeze are also determined by the cloud droplet's size and chemical composition which is partly dependent on the ambient CCN concentration.Thus, indirectly, CCN play a crucial role not only for warm-phase precipitation processes, but also for the ice phase processes (Altaratz et al., 2014).If the droplets are small (due to a high CCN concentration), their freezing temperature is comparatively low, leading to ice formation at higher levels in the cloud, where the released latent heat increases the buoyancy at these high levels.This in turn enhances the updraught and invigorates convection, thereby contributing to the thermodynamic effect (Khain et al., 2005;Koren et al., 2005;Wang, 2005).As smaller droplets form smaller ice particles as they freeze, the fall velocities of the crystals are reduced, prolonging their residence time at high altitudes.While this favours the WBF process, subsequent mixed-phase precipitation formation processes are delayed (Tao et al., 2012;Saleeby et al., 2013).Analogous to the slow-down of the warm-rain processes described above, this residence time effect also contributes to the spillover effect (Saleeby et al., 2009;Zubler et al., 2011a;Letcher and Cotton, 2014;Xiao et al., 2015;Fan et al., 2016).In both of these cases, where aerosols can modify the onset and location of the precipitation event, the altered precipitation patterns may also trigger atmospheric instability.Sedimenting hydrometeors cool the sub-cloud air when ice particles melt or sublimate or when cloud droplets evaporate.This results in the formation of cold pools.Cold-pool dynamics in turn drive the formation of gravity currents which enhances updraught velocities and the vertical advection of moisture along the cold pools' leading edges, favouring the triggering of convection (Schlemmer and Hohenegger, 2014;Haerter and Schlemmer, 2018).Such feedbacks are essential for the evolution of shallow and deep convective clouds, promoting enhanced precipitation.In summary, the net effect of the aerosol concentration on convection, moisture transport and eventually precipitation is the sum of complex interdependent processes that are hard to disentangle.While additional CCN tend to delay precipitation and invigorate the updraught, INPs can counteract those effects, promote early precipitation and weaken the convection.In the European Alps, a prominent source of aerosols is Saharan dust (Coen et al., 2004).Dust particles are known to particularly affect the INP number concentration (Chou et al., 2011;Lacher et al., 2018a) and deposition nucleation (Weger et al., 2018) and hence determine the cloud ice phase and the associated aerosol-induced response in precipitation (Zubler et al., 2011b, and interactions as outlined above).
To elucidate these complex aerosol-cloudprecipitation interactions, we investigate the aerosol effect on extreme precipitation based on a case-study from the Swiss Alps, with focus on the northeastern part of Switzerland.In the summer of 2017, starting on 31 August, intense and continuous rain lasting for three days resulted in severe flooding and caused substantial financial damage (Andres and Badoux, 2018).The severity of this event was underestimated by the Swiss Federal Office of Meteorology and Climatology (MeteoSwiss), whose NWP Consortium for Small-scale Modeling (COSMO) model, COSMO-1, particularly failed to forecast the precipitation in the northeastern part of Switzerland.While the synoptic setting during this time was prone to produce a large amount of precipitation, additional small-scale processes which are not sufficiently represented in COSMO-1 might have contributed to the heavy precipitation event .In fact, during the same time, the aerosol concentration in the south of Switzerland was unusually high due to a concomitant Saharan dust plume.We hypothesize that this increase in aerosol load could have added to the high precipitation rate as well as its underestimation by COSMO-1, which by design does not account for aerosol-cloud interaction.Using a comprehensive set of observations in combination with a series of model simulations, we assess the impact of elevated aerosol concentrations on the extreme precipitation event over Switzerland.Our model simulations include different microphysical parametrizations as well as varying prognostic aerosol perturbations.

MODEL SET-UP
Simulations are conducted with the regional weather and climate model COSMO (Steppeler et al., 2003).COSMO has been used for simulating (orographic) MPCs in previous studies (Lohmann et al., 2016;Henneberg et al., 2017;Possner et al., 2017;Eirund et al., 2019a;2019b;Dedekind et al., 2021).The simulations are performed on a rotated latitude-longitude grid with 0.01 Our simulation set-up is based on the hypothesis that aerosol-driven microphysical perturbations as a result of the Saharan dust outbreak account for the high precipitation rates north of the Alps.As operational forecasting centres like MeteoSwiss use a one-moment instead of a two-moment cloud microphysics scheme, any aerosol-cloud-precipitation interactions are neglected due to the missing hydrometeor number and size information in the scheme.Yet, previous studies have identified improvements in the simulated cloud properties (Milbrandt et al., 2010;Igel et al., 2015;Senf et al., 2020), precipitation patterns (Milbrandt et al., 2010;Igel et al., 2015) as well as aerosol-cloud-precipitation interactions in cirrus clouds over Germany during a Saharan dust event (Weger et al., 2018) in model simulations including a two-moment as compared to a one-moment scheme.
Thus, our control (CTRL) simulation is supposed to represent a more realistic model scenario and was performed with the two-moment cloud microphysics scheme of Seifert and Beheng (2006).The scheme solves prognostic equations for the mass mixing ratios and number concentrations of the six hydrometeor classes cloud water (QC), rain (QR), cloud ice (QI), snow (QS), graupel (QG), and hail (QH).The two-moment microphysics scheme has been expanded following Possner et al. (2017) and Eirund et al. (2019b) to include cloud droplet activation based on the ambient CCN concentration according to Köhler theory (K'´ohler, 1936;Nenes and Seinfeld, 2003).In the absence of aerosol sources, the background CCN concentrations are held constant to prevent a loss of aerosols over time through precipitation processes (following Eirund et al., 2019b).A background CCN concentration of 100 cm −3 at every grid point and every height with a mean diameter of 0.05 m was chosen based on aerosol measurements at Jungfraujoch (JFJ) in the Swiss Alps during the simulated period in August 2017 (Figure S1).This background concentration corresponds approximately to the annual median value observed at the JFJ at an annual median cloud peak supersaturation value of 0.35% (Jurányi et al., 2011;Hammer et al., 2014) In contrast to the CTRL simulation, we performed one simulation with the one-moment cloud microphysics scheme (1-M) by Seifert and Beheng (2001) in order to represent the operational set-up.The scheme includes prognostic equations for the mass mixing ratios of QI, QC, QR and QS.The rain formation process can be modified by changing theconstant cloud droplet number concentration, which influences the cloud droplet autoconversion rate.In our model set-up this constant cloud droplet number concentration has been set to 100 cm −3 , to match the set-up of the CTRL simulation.Note that hail is by design not included in the one-moment scheme, which in addition to the number of moments depicts another important difference between the 1-M and CTRL simulations.Contrasting the CTRL and the 1-M simulations provides a first assessment of the importance of hydrometeor number and size information in forecasting precipitation during the observed Saharan dust outbreak.

Model perturbation simulations
To study the effects of aerosol perturbations on precipitation, a set of simulations with fully prognostic aerosols was performed.CCN and INPs were released for the first 14 hr of the simulations in the southeastern corner of the model domain, vertically limited to the lower 5 km of the model atmosphere to match the observed Saharan dust plume.
Based on JFJ measurements (Figures S1 and S2) the CCN perturbation mode was assumed to be larger than the background mode, with a count mean diameter of 0.12 m.Both aerosol perturbations are fully prognostic, meaning that aerosols are advected throughout the domain, depleted by cloud droplet or ice crystal formation and precipitation, and are released back into the atmosphere through evaporation or sublimation.CCN perturbations were set to 500 cm −3 (CCN_perturb) and to 5,000 cm −3 (CCN_highperturb), INP perturbations were increased by a factor of 10 (additional 100 L −1 ; Table 1 and Figure S2).According to measurements taken during Saharan dust events close to the source (Boose et al., 2016) as well as in the Swiss Alps (Lacher et al., 2018a), INP concentrations can increase up to several hundreds per litre.Thus, a perturbation of 100 L −1 can be considered realistic for the Alpine environment as simulated in our case.

SYNOPTIC EVOLUTION
The synoptic evolution during the heavy precipitation event from 31 August to 02 September 2017 over Switzerland is schematically outlined in Figure 1 and can be divided into three distinct stages: • 31 August 2017, 0000-1100 UTC: prefrontal precipitation and cold front passage in northern Switzerland.
• 31 August, 1100 UTC-01 September 2017, 1100 UTC: heavy rainfall along a band stretching from southern Switzerland northeastward across the Swiss Alps.
Most of the precipitation originated from strong convective cells which initiated on the southern side of the Alps and were transported downstream to northeastern Switzerland.At this stage, increased aerosol concentrations potentially influenced cloud processes (also Section 3.3).
• 01 September, 1200 UTC-02 September 2017, 1200 UTC: the presence of a cut-off low with northerly flow leading to precipitation along the northeastern Alps.
In southern Switzerland, fair weather with northerly f'´ohn winds prevailed.This last phase is not of interest for this study, and thus is not described further.Note: The aerosol perturbations were applied in the southeastern corner of the domain and lasted for 14 hr only (Figure S2).

Stage 1: Cold front passage in the north, convection initiation in the south
period dominated by high pressure over Central Europe.Ahead of the trough, strong southwesterly flow impinged the Alps (Figure S2).NOAA HYSPLIT backward trajectories (Stein et al., 2015) computed with the READY web application (Rolph et al., 2017) and based on the 0.5 • Global Data Assimilation System dataset (not shown) suggest that these air masses partly originated from the North African desert areas and were loaded with dust.A cold front passed over northern Switzerland between 0600 and 1200 UTC.It decelerated thereafter, remaining stationary along the northwestern slopes of the Swiss Alps for the rest of the day.
On the southern side of the Alps, very warm and potentially unstable air was present ahead of the cold front.A strong upper-level jet extended from the western Alpine arc northeastward across Germany towards Scandinavia.Its right entrance area approached the area of this study, thereby promoting large-scale ascent and supporting the favourable synoptic setting for convective activity during the event.

Stage 2: Convection in the south and downstream transport across the Alps
During 31 August, surface heating further decreased the stability in the lower atmosphere south of the Alps.Convective cells were triggered repeatedly in southern Switzerland and adjacent northern Italy, some with isolated and supercellular characteristics, others with a more linear shape aligned with the flow (not shown).Rainfall accumulations with peaks of over 200 mm within 24 hr occurred in these areas during this stage of the event (Figure 2).
The southwesterly flow advected the convective cells towards and over the Alpine ridge, thereby dissipating and interacting with counter-flowing cold air at lower levels north of the Alps.A secondary maximum of precipitation occurred in northeastern Switzerland and adjacent western Austria, with local peaks of 100 mm rainfall within 24 hr.

Evidence of Saharan dust
Both model simulations and lidar observations suggest that aerosol concentrations were increased ahead of the cold front during this flooding event.A band of increased aerosol optical depth (Figure 3a) in the midlatitudes indicated that dust was advected by the southwesterly flow (Figure 3c) from the Saharan desert across the Mediterranean Sea towards the Alps.This ensemble model forecast is supported by the aerosol backscattering measured by the CALIOP lidar on board the CALIPSO satellite.The lidar data indicate that the Saharan dust plume had a thickness of approximately 5 km (Figure 3b).The Saharan dust event was also detected with the optical in situ measurements on the JFJ using the inverted wavelength dependence of the single-scattering albedo (Coen et al., 2004).The detected plume lasted at the JFJ until midnight of 31 August (Figure S1).

Simulated background conditions
To compare the background conditions as simulated in CTRL with observations, total surface precipitation accumulated over 24 hr over Switzerland in the CTRL simulation is shown in Figure 4a.Elevated precipitation rates (>50 mm) are simulated in eastern Switzerland as well as southern Germany, in agreement with the observations shown in Figure 2.However, the overall precipitation amount is underestimated in the CTRL simulation, and the regions of precipitation maxima are shifted in time (Figure 4b).The precipitation maxima in southern Switzerland are weaker in the CTRL simulation than observed and shifted eastward.Regionally, the accumulated precipitation exceeds 100 mm north of the Alpine ridge, but over a much smaller spatial extent than in the observations.The temporal evolution of the heavy rainfall event north of the Alpine ridge (as indicated by the black rectangular box outlined in Figures 2 and 4a) can be subdivided into three peak events.The first peak of rainfall between 1200 and 2100 UTC was initiated by pre-frontal precipitation development in the south and is represented in both the observations and the CTRL simulation (Figure 4b).The precipitation magnitude during the first peak is comparable in the CTRL simulation and the observations, indicating that the large-scale conditions are reasonably well simulated by the model.However, the model shifts the precipitation maximum to 1600 UTC, 3 hr earlier than in the observations (Figure 4b).
Figure 5 shows simulated total cloud liquid (QLIQ = QC + QR) and solid (QICE = QS + QI + QG + QH) mass mixing ratios during the three periods averaged over the flow-aligned cross-section outlined in Figure 4a.Deep convection was initiated during Peak 1 (Figure 5a), with cloud tops reaching 13 km in altitude.High QLIQ and QICE (up to 0.5 and 0.9 g⋅kg −1 , respectively) led to elevated precipitation rates especially along the southern Alpine ridge during this period (Figure 5d).During the second and third peak, a combination of moisture transport towards southern Germany and the increased moisture loss during the heavy precipitation in Peak 1 decreased QLIQ and QICE (Figure 5b, c).The thinner cloud and the spatially more homogeneously spread surface precipitation indicate the transition to a more stratiform character of precipitating clouds (Figure 5e, f).Small-scale convection at the northern Alpine ridge led to a small precipitation increase during Peak 3, but simulated surface precipitation remained well below the simulated precipitation north of the Alpine ridge during Peak 1, in contrast to the observations.
To investigate the effect of the microphysics parametrization on precipitation, we show differences between the 1-M and the CTRL simulations in terms of total surface precipitation, the cloud development, and accumulated surface precipitation along the flow-oriented cross-section in Figure 6.Throughout all three periods, simulated surface precipitation is lower in 1-M than in CTRL in southeast Switzerland (Figures 6a-c).The reduced precipitation rates in 1-M are related to a substantially lower QLIQ within the cloud and lower QICE at altitudes below 7 km, especially during Peak 1 (Figure 6d).Here, the reduction in precipitation formation efficiency in 1-M versus CTRL is two-fold.Firstly, the lower QICE below 7 km reduces the riming efficiency of cloud droplets onto ice crystals at lower altitudes (Field and Heymsfield, 2015).The reduced riming rates also have a negative bearing on secondary ice formation induced by the Hallett-Mossop process at temperatures between −3 and −8 • C (Hallett and Mossop, 1974).Secondly, in the 1-M microphysics scheme merely the hydrometeor mass is prognostically determined.The neglected size dependence delays the conversion of cloud water and ice to rain and snow, respectively, and thus leads to less efficient precipitation formation in the 1-M simulation on the windward side of the Alpine ridge during Peak 1.The more available QICE at higher altitudes in 1-M is transported towards southern Germany during Peaks 2 and 3, which leads to higher QICE in 1-M as compared to CTRL, as shown in Figure 6e, f.The sedimenting ice crystals on the leeward side of the Alpine ridge melt and form liquid drops, and thereby increase the QLIQ at lower altitudes.In addition, the lack of hail formation in 1-M as compared to CTRL also reduces QICE in 1-M (not shown), which in turn contributes to reduced riming processes.These results suggest that the microphysical parametrization, especially the size dependence, strongly impacts the cloud properties and thus also the regional surface precipitation rates.For summertime deep convective clouds such as are simulated in our case, we find especially the conversion of cloud-to precipitation-size particles to be important for ice crystal sedimentation and enhanced precipitation formation.Our simulated reduction in precipitation in 1-M is in contrast to findings from Weger et al. (2018) and Igel et al. (2015) who simulated overestimated precipitation rates in a one-moment scheme as compared to a two-moment scheme and observations.However, none of these studies simulated flow over orographic terrain, which highly impacts the dynamics and related cloud microphysical conversion rates in our case.

CCN perturbation experiments
Air masses containing high aerosol loadings can potentially provide additional CCN, which, once activated, form cloud droplets.The additional latent heat release during condensation may intensify convection (e.g., Fan et al., 2018), which could have contributed to the high precipitation event on the leeward side of the Alpine ridge.To test the hypothesis of convective invigoration for our simulated case, we show differences in total surface precipitation, QLIQ and QICE, as well as the accumulated precipitation along the flow-oriented cross-sections for the two CCN perturbation scenarios CCN_perturb and CCN_highperturb (with perturbations of 500 and 5,000 CCN cm −3 in the first 14 hr, respectively) with regard to the CTRL simulation in Figures 7 and 8. Increased levels of CCN increase QLIQ in the upper cloud layers and decrease QLIQ in the lower cloud layers (Figures 7d-f and  8d-f).This is caused by a shift in QLIQ to smaller hydrometeors, resulting in an increase in cloud liquid water in the upper layers and a decrease in rain water in the lower layers (not shown), which in turn is caused by a reduction in collision-coalescence efficiency in the perturbed simulations (as also shown by Khain et al., 2009).As a result, increased CCN concentrations decrease the total surface precipitation in the south of the domain relative to CTRL (Figures 7a-c and 8a-c).These differences in precipitation are characterised by southwest-northeast oriented bands, which can be attributed to a slight shift of the initial convection in the simulations.
In-cloud adjustments to increased CCN concentrations reflect the spatial pattern of CCN perturbations in the F I G U R E 7 As Figure 6, but for CCN_perturb minus CTRL F I G U R E 8 As Figure 6, but for CCN_highperturb minus CTRL domain.South of the Alpine ridge where the CCN perturbations were applied, the high number of available CCN (Figure S3), as well as the suppressed precipitation, substantially increase QLIQ at higher altitudes (3-9 km) by up to 0.12 g⋅kg −1 in the CCN_perturb simulation, and decrease QLIQ in the lower 3 km during Peak 1 (Figures 7d   and 8d).The increase in QLIQ is even more pronounced in the CCN_highperturb simulation, with a higher concentration south of the Alpine ridge of up to 4.4 g⋅kg −1 during the second peak (Figure 8e).Following the main flow (Figure 3c), the anomalies propagate to the leeward side of the Alps.As during Peaks 2 and 3 most of the perturbation CCN are transported up to the tropopause via convection (Figure S3), the differences in QLIQ decrease (Figures 7e, f and 8e, f).Even though CCN perturbations imply only a direct effect of liquid cloud properties, QICE is also impacted by elevated CCN concentrations.On the one hand, more available CCN favour the formation of cloud droplets, which may lead to more latent heat release throughout the cloud and hence a decrease in cloud ice mainly between 4 and 9 km (Figures 7d-f and  8d-f).On the other hand, the increased cloud liquid water content increases the cloud long-wave emissivity (Garrett and Zhao, 2006), which increases cloud-top cooling and promotes ice formation at the lower temperatures aloft (Figures 7d-f, 8d-f and Possner et al., 2017;Eirund et al., 2019b;Ramelli et al., 2021).Also, excess QLIQ in CCN_perturb and CCN_highperturb leads to more homogeneous nucleation of ice crystals at temperatures below −38 • C in the perturbed simulations.
The lower QICE and QR (not shown) at altitudes below 9 km and the smaller cloud droplets with a reduced sedimentation efficiency on the southern side of the Alpine ridge cause a reduction in accumulated surface precipitation over the cross-section of up to 4 and 8 mm during Peak 1 for a perturbation of 500 CCN cm −3 (Figure 7g) and 5,000 CCN cm −3 (Figure 8g), respectively.However, on the northern side of the Alpine ridge, precipitation is slightly increased throughout the three peaks.In both CCN perturbation simulations, the increased CCN concentrations cause more numerous, but smaller, cloud droplets which delay the precipitation process.This decreased loss of cloud water on the southern side of the Alpine ridge leads to a build-up in QLIQ on the northern side of the Alpine ridge which in turn increases precipitation formation according to the spillover effect (Figures 7g-i, 8g-i and Jiang, 2003;Muhlbauer and Lohmann, 2008).Increased riming rates in the perturbed simulations may further contribute to the spillover effect and the increase in surface precipitation north of the Alpine ridge (Figure S4a).In CCN-perturbed environments, sedimenting ice crystals can feed on the more available QLIQ, increasing the riming rate and contributing to surface precipitation.Enhanced riming in response to increased aerosol concentrations and correspondingly higher QLIQ is in agreement with Lohmann (2004), but in contrast to Rosenfeld (2005) and Heikenfeld et al. (2019).
Overall, the simulated changes in precipitation over the flow-aligned cross-section indicate a substantially lower total impact on precipitation resulting from CCN perturbations as compared to a change in microphysics between the 1-M and the CTRL simulations (Figure 6).
Thermodynamical cloud adjustments through convective invigoration are shown in Figure 9 in terms of differences in temperature due to latent heating as well as updraught velocity (w > 0) for the three analyzed peaks between the CCN_highperturb and CTRL simulations.Indeed, the enhanced cloud droplet activation leads to a local warming through latent heat release up to 4 K⋅hr −1 .This warming can suppress in-cloud freezing as shown in Figures 7 and 8.In addition, the local heating causes an increase of the temporally averaged updraught velocity by up to 2.2 m⋅s −1 in CCN_highperturb relative to CTRL, which promotes higher supersaturation and deeper clouds, in which cloud particles become larger at higher levels.This then allows for growth by collisions and recycling of hydrometeors that are caught again in the updraughts by recirculation.However, this increase in updraught velocity is spatially confined and the average change in updraught velocity remains small (around 2%).

INP perturbation experiment
Following Section 4.2, we performed one additional simulation with elevated INP perturbations in order to represent the Saharan dust outbreak.Similar to the effect of CCN perturbations (Figures 7a-c and 8a-c), differences in precipitation in response to elevated INP concentrations are characterised by stripe-like bands, most pronounced south of the Alpine ridge (Figure 10a-c).However, simulated changes in surface precipitation remain below 5 mm, with locally increased as well as decreased precipitation amounts.During Peaks 1 and 2 (when the domain is continuously supplied with new INPs from the south), the accumulated surface precipitation is mainly increased south of the Alpine ridge.The positive precipitation changes in the south thus agree with previous findings of INPs promoting windward orographic precipitation as compared to pristine conditions (Ault et al., 2011;Creamean et al., 2013;Fan et al., 2017).Within cloud, increased INP concentrations generally increase QICE below and above the −38 • C threshold.During Peak 1, the more available INPs increase the ice crystal number concentration through immersion freezing and mass through subsequent growth by the WBF, which leads to a decreased QLIQ in the upper cloud layers (4-8 km), consistent with previous studies (e.g., Fan et al., 2017;Eirund et al., 2019b) In total, the effect of elevated INP concentrations on surface precipitation is comparable to the effect of a 500 CCN cm −3 perturbation in CCN_perturb (Figure 7).While INP perturbations locally increase surface precipitation at and south of the Alpine ridge, the response to a 100 INP L −1 perturbation also does not explain the high precipitation event in northern Switzerland.

DISCUSSION AND CONCLUSIONS
This study provides insights into aerosol-cloudprecipitation interactions in an Alpine environment based on a case-study in the Alps.In summer 2017, heavy Shown is a temporal snapshot at 2 hr, shortly before the heavy precipitation peak north of the Alps (Figure 4b).The grey shading represents regions above 1,500 m height in order to mask cold pools over steep terrain precipitation associated with the passage of a cold front led to local flooding and damage in the northeastern parts of Switzerland.The severe flooding event was underestimated by one warning level in the local weather forecast.To follow up on our hypothesis that the constraints on the cloud microphysical parametrization in the COSMO-1 forecast model (including only a one-moment scheme and no aerosol perturbations) might have been responsible for the lack of forecasting skill in predicting the heavy precipitation event, we performed a series of sensitivity simulations including a one-and two-moment cloud microphysics scheme (Seifert and Beheng, 2001;2006, respectively), as well as different aerosol perturbations over Switzerland.
The model simulations show that increasing either CCN or INPs has a profound effect on the convective dynamics south of the Alps, where new cells develop, grow, precipitate, and repeatedly trigger secondary cells as cold pools spread and interfere with the orography (Figure 11).Differences in cold pool strength reach up to 5 K (Figure 11b-e), with less intense cold pools being present in the 1-M, CCN_perturb and CCN_highperturb simulations as compared to the CTRL simulation.These weaker cold pools induce a positive feedback, by triggering less convection and lower precipitation rates south of the Alps (Figures 6a,7a,8a,10a).The meso-gamma scale variability introduced by CCN or INPs nevertheless does not significantly change the large-scale evolution of the flow in this case.However, from the perspective of operational weather forecasting, where ensemble model systems play an increasingly important role in short-term forecasting of convection, our results suggest that aerosols could be used for model perturbations both in the assimilation as well as in the forecast step.
In addition to the aerosol forcing, other possible causes for the extreme precipitation event observed in northeastern Switzerland and the lack of forecasting skill could arise from the large-scale synoptic situation or to be linked to model deficits.In terms of shortcomings regarding the model set-up, previous work identified model resolution (Bryan et al., 2003;Possner et al., 2016;Panosetti et al., 2019) and orography (Lohmann et al., 2016;Henneberg et al., 2017) being crucial when simulating atmospheric boundary-layer processes and MPCs in complex terrain.Orographic lifting determines the amount of condensate in low-and mid-level clouds, which can result in surface precipitation.With the chosen horizontal and vertical resolution of our set-up, as well as the operational COSMO-1 set-up, the highest elevation in the model orography is merely 88% of the actual terrain (Section 2).As a result, the simulated air mass ascent over the southern Alpine ridge might be too weak, resulting in an underestimation of condensate and precipitation downstream of the Alpine ridge (comparable to results from Henneberg et al., 2017;Henneberg, 2017).On the sub-synoptic scale, various other processes, which are notoriously under-represented in NWP models due to their smaller scale and complex nature, could have contributed to the strong precipitation event.For instance, rearward of the front, low-level winds on the northern side of the Alps turned from southwest to north, favouring shallow orographic lift and counteracting downslope winds originating from the southerly upper-level flow.This change in wind direction may have enhanced precipitation by the seeder-feeder mechanism (Bergeron, 1965), where hydrometeors from an upper cloud layer ("seeder") sediment through low-level orographic clouds ("feeder") and can increase precipitation by collision-coalescence or accretion.
Apart from its indirect effect on cloud properties, (Saharan) dust is also known to impact the atmospheric radiative balance firstly through its direct interaction with solar and thermal radiative fluxes (Slingo et al., 2006) and secondly through its deposition on snow surfaces (Di Mauro et al., 2015).This altered radiative balance can in turn lead to changes in atmospheric stability (Lohmann and Feichter, 2005) and induce feedbacks with clouds (Zelinka et al., 2014).In addition, local dust radiative heating interacts with atmospheric dynamics by affecting cold pools (Gläser et al., 2012), low-level frontogenesis (Chen et al., 1994), or frontal propagation velocities and strength (Tummon et al., 2010;Deetz et al., 2018).However, the prognostic aerosol perturbations such as implemented in our scheme do not interact directly with radiation, thus we are missing any direct radiation-induced feedbacks in our study.Such feedbacks could likely add to the aerosol-cloud-precipitation interaction-induced signatures in cold pool dynamics south of the Alps (Figure 11) and affect the intensity and timing of the cold front.Investigating prognostic aerosols during Saharan dust outbreaks and their feedbacks with atmospheric radiation and dynamics in addition to aerosol-cloud-precipitation interactions remains subject to future work.
To conclude, the main findings of our study can be summarised as followed: • The choice of microphysics scheme has the largest impact on surface precipitation.In summer convective clouds such as simulated in our case, the conversion rates from cloud droplets to raindrops and ice crystals to snow crystals are reduced in the 1-M simulation.As a result, QLIQ in the 1-M simulation is substantially lower than in the CRTL simulation throughout the cloud.Similarly, QICE is decreased below 6 km on the south of the Alpine ridge early on in the 1-M simulation.The lower conversion rates in 1-M strongly delay the initial precipitation formation, leading to a drastic reduction in surface precipitation over most of the flow-oriented cross-sections covering Switzerland during Peak 1.However, the larger amount of QICE at higher altitudes is transported over the Alpine ridge, where growing and sedimenting ice crystals lead to increased surface precipitation during Peaks 2 and 3.
• Increased CCN concentrations lead to increased surface precipitation downstream in northeastern Switzerland, but their effect is too small to explain the observed heavy precipitation event.Elevated CCN concentrations substantially increase QLIQ, and decrease QICE between the melting and the homogeneous freezing levels south of the Alpine ridge during Peak 1.The build-up of QLIQ is carried towards the north of the Alpine ridge where surface precipitation is increased according to the spillover effect (Jiang, 2003;Muhlbauer and Lohmann, 2008).In addition, the position of deep convective cells shift in the CCN perturbation simulations as compared to CTRL, which may also cause locally large differences in surface precipitation.
• INP perturbations lead to locally increased windward surface precipitation, but do not induce elevated precipitation rates downstream of the mean flow, north of the Alpine ridge.Increased INP concentrations cause a reduction in QLIQ, as more of the QLIQ is converted to QICE through heterogenous nucleation and the WBF process throughout the cloud during Peak 1.As the INP perturbation gets advected through the domain during Peaks 2 and 3, QICE decreases with time, which results in a decrease rather than an increase in surface precipitation during Peak 3.
Our study highlights the competing impact of synoptic forcing and microphysics on cloud properties and precipitation formation.We show that microphysical perturbations impact smaller-scale convective events downstream, but any effect on the meso-gamma scale is masked by the strong large-scale forcing.A way forward to investigate the possibility of aerosol-induced precipitation changes in such dynamically forced environments could be the analysis of ensemble runs, perturbed by aerosols.In addition, simulations at higher horizontal and vertical resolutions would improve the representation of orography and thus better resolve updraughts and moisture transport.Thus, we encourage future work to continue investigating the competing effects of microphysics and dynamics in NWP and cloud-resolving models.

F
I G U R E 1 500 hPa geopotential height contours in dm and surface fronts at 0000 UTC on (a) 31 August and (b) 01 September 2017, and the 1,500 m height contour of the operational COSMO-1 model orography (brown shading).In (a), the grey dashed box denotes the model domain used for the simulations in this study, geographical terms used are given, and the asterisk marks the location of the JFJ high-altitude measuring site . INPs available for immersion freezing are prognostic throughout the simulations and are implemented following Possner et al. (2017) and Eirund et al. (2019b).The scheme parametrizes immersion freezing following the DeMott et al. (2015) temperature dependence and captures the depletion and replenishment of INPs.The initial concentration of INPs was chosen according to measurements at JFJ during summer by Lacher et al. (2018b) and set to 10 INP L −1 .The INPs are released in 15 separate temperature bins each of width 0.5 • C. The highest and lowest temperature bins are at −11 • C and −18.5 • C, respectively.
On 31 August 2017, an upper-level trough approached the Alps from western Europe and terminated a very warm TA B L E 1 Overview of the performed simulations Observed accumulated precipitation over 24 hr from 1200 UTC on 31 August 2017, obtained from the Meteoswiss CombiPrecip product.The solid box represents the averaging region shown in Figure 4b.The dashed parallelogram indicates the area used for the flow-oriented cross-sections shown below

F
I G U R E 3 (a) Aerosol optical depth (AOD) forecast for 0000 UTC on 01 September 2017 from the multimodel median produced by the WMO SDS-WAS programme.The black line indicates the path of the CALIPSO satellite between approximately 0120 and 0140 UTC.(b) Vertical feature mask of the CALIOP lidar mounted on the CALIPSO satellite.The depicted sector corresponds to the path shown in (a).(c) Water vapour satellite image (6.2 m) showing the 300 hPa geopotential height in dm (white contours) and wind (coloured barbs) U R E 4 (a) CTRL simulation of accumulated precipitation over 24 hr from 1200 UTC on 31 August 2017.(b) Averaged surface precipitation north of the Alps over the box indicated by the solid lines in (a) from the observations (red) and the CTRL simulation (blue).The vertical black lines in (b) separate the three distinct precipitation peaks and averaging periods over the flow-oriented cross-section (dashed lines in (a)) shown in Figures 5-10

F
I G U R E 5 (a-c) Simulated liquid (QLIQ = QC + QR; colours) and solid (QICE = QS + QI + QG + QH; contours) hydrometeor mass mixing ratios in the CTRL simulation averaged over the flow-oriented cross-section indicated in Figure 4a, for the three peak periods.The red (blue) lines denote the 0 • C (−38 • C) elevations.The red shading represents the topography along the cross-section.(d-f) show the total accumulated surface precipitation during the respective peak periods along the flow-oriented cross-section

F
I G U R E 6 (a-c) Differences in total accumulated surface precipitation (mm/hr) between the 1-M simulation and CTRL from the start of the simulation (1100 UTC) until the end of periods (a) Peak 1, (b) Peak 2, and Peak 3. (d-f) differences in QLIQ (colours) and QICE (contours) hydrometeor mixing ratios between the 1-M and the CTRL simulation in the flow-oriented cross-section indicated in (a) during the respective peak periods.The red (blue) lines denote the 0 • C (−38 • C) elevations.(g-i) show differences in accumulated surface precipitation during the respective peak periods along the cross-section Differences of (a-c) temperature change due to latent heating and (d-f) updraught velocity between the CCN_highperturb and the CTRL simulations over the flow-aligned cross-section F I G U R E 10 As Figure 6, but for INP_perturb minus CTRL Thus, convective invigoration cannot sufficiently explain the larger-scale heavy precipitation event as seen in the observations.
. Above 9 km, increased convection increases homogeneously nucleated ice crystals and hence QICE.These anomalies propagate northward during Peak 2, although the response in the ice phase decreases.During Peak 3, less QICE is present in INP_perturb compared to the CTRL simulation.The precipitation increase in INP_perturb during Peaks 1 and 2 leads to a loss of INP, which is reinforced by the northward advection of INPs and the discontinued supply of perturbation INPs after 14 hr simulation time.

F
I G U R E 11 Differences in cold pool strength shown as virtual potential temperature differences over the lowest ten model levels (corresponding approximately to the lowest 250±10 m) and 10 m horizontal wind over the box outlined in (a) for the (b) 1-M and CTRL, (c) CCN_perturb and CTRL, (d) CCN_highperturb and CTRL, and (e) INP_perturb and CTRL simulations.