Sensitivity of precipitation formation to secondary ice production in winter orographic mixed-phase clouds

The discrepancy between the observed concentration of ice nucleating particles (INPs) and the ice crystal number concentration (ICNC) remains unresolved and limits our understanding of ice formation and hence precipitation amount, location and intensity. Enhanced ice formation through secondary ice production (SIP) could be accounting for this discrepancy. Here, we present the results from a sensitivity model study in the Eastern Swiss Alps with additional simulated in-cloud SIP on precipitation formation and consequently on surface precipitation. The SIP processes considered include rime splintering, 5 droplet shattering during freezing and breakup through ice-ice collisions. We simulated the passage of a cold front at Gotschnagrat, a peak at 2281 m above sea level (a.s.l.), on 7 March 2019 with COSMO, at a 1 km horizontal resolution, as part of the RACLETS field campaign in the Davos region in Switzerland. The largest simulated difference in the ICNC at the surface originated from the breakup simulations. Indeed, breakup caused a 1 to 3 order of magnitude increase in the ICNC compared to SIP from rime splintering or without SIP processes in the control simulations. The ICNCs from the collisional breakup 10 simulations at Gotschnagrat were in better agreement with the ICNCs measured on a gondola near the surface. However, these simulations were not able to reproduce the ice crystal habits near the surface. Enhanced ICNCs from collisional breakup reduced localized regions of higher precipitation and thereby improving the model performance in terms of surface precipitation over the domain.

this study, we simulate the cold front passage between 9:30 and 14:45 UTC on March 7, 2019. Hourly initial and boundary conditions analysis data at a horizontal resolution of 7×7 km, supplied by MeteoSwiss, were used to force COSMO.
Simulations were conducted with the SIP processes, where several SIP processes were active at the same time and a control simulation (CNTL) where none of the SIP processes was active. For each of these simulations, 5 ensemble simulations are conducted by perturbing the initial temperature conditions at each grid point through the model domain with unbiased Gaussian 130 noise at a zero mean and a standard deviation of 0.01 K (Selz and Craig, 2015;Keil et al., 2019).

Cloud microphysics scheme
We use a two-moment cloud microphysics scheme within COSMO with six hydrometeor categories, including hail, graupel, snow, ice crystals, raindrops and cloud droplets (Seifert and Beheng, 2006). INPs available for immersion freezing is prognostic throughout the simulations and are implemented following Possner et al. (2017) and Eirund et al. (2019b). The immersion  (DeMott et al., 2010) and, therefore, we also used the retrieved aerosol concentration from the upward-pointing LIDAR that was situated at Davos Wolfgang. The LIDAR retrieval gave a full vertical profile of the atmosphere. At temperatures between 258 and 243 K, the aerosol concentration larger than 0.5 µm was between 1.8 and 2.5 cm −3 for which we then accordingly chose 2 cm −3 as input for the D15 parameterization. At a temperature of 243 K, the estimated INP concentration was 23 L −1 . The freezing of raindrops occur heterogeneously and independent of INPs. This is because the raindrop number 145 concentration is at least 10 3 smaller than the CNDC and has an insignificant contribution to the primary production of ice crystals upon freezing (Figs. S4 and S5i and j). It is included in our simulations because without this parameterization droplet shattering can not occur. In the spectral partitioning of freezing rain, only the frozen raindrops that are partitioned as ice crystals (Blahak, 2008) can cause multiplication through droplet shattering as discussed in the following section.

Secondary ice processes parameterizations 150
Besides the primary ice formation pathways through homogeneous and heterogeneous nucleation of cloud droplets and raindrops, rime splintering is the only process included in the standard version of COSMO that can enhance the ICNC otherwise.
The rime splintering process has been parameterized, implemented and tested in numerical weather models (Blyth and Latham, 1997;Ovtchinnikov and Kogan, 2000;Phillips et al., 2006;Milbrandt and Morrison, 2015;Phillips et al., 2017). In COSMO, rime splintering occurs exclusively after collisions between supercooled cloud droplets of diameter greater than 25 µm or rain-155 drops with ice crystals, snow, graupel or hail, all larger than 100 µm (e.g Phillips et al., 2006) at temperatures between -3 and -8 • C (Hallett and Mossop, 1974). The predominant theory is that within this temperature range the supercooled droplets that rime on large ice particle freeze resulting in a buildup of internal pressure whereby the pressure is relieved when the frozen shell cracks and produces secondary ice particles (Hallett and Mossop, 1974). At temperatures colder than -8 • C the ice shell of the frozen droplet is too strong to break (Griggs and Choularton, 1983) and at warmer temperatures than -3 • C the supercooled 160 droplet spread over the ice particle not causing any SIP (Dong and Hallett, 1989). However, SIP has also been observed at temperatures where the Hallett-Mossop droplet size and temperature requirements were not satisfied.
Droplet shattering produces maximum splinters at around -15 • C when large droplets freeze and shatter if the internal pressure build-up is high enough to eject fragments (e.g. Kolomeychuk et al., 1975;Leisner et al., 2014;Wildeman et al., 2017;Lauber et al., 2018;Keinert et al., 2020). So far, other than the temperature at which maximum droplet shattering occurs, 165 there is no temperature constraint on this process which can be specifically important at temperatures close to -15 • C (Korolev et al., 2020;Lauber et al., 2020). The pressure build-up occurs mainly due to the unique characteristic of liquid water having a higher density than ice and thus expanding when it freezes. Larger droplets are more likely to shatter and likely produce more ice splinters (Kolomeychuk et al., 1975;Lauber et al., 2018). However, as of yet the number of splinters that are produced during droplet shattering could not be quantified. A more rigorous formulation for the fragment number remains a challenge 170 due to the lack of measurement in laboratory studies. Lauber et al. (2018) showed the highest fragment rates occur at 258 K, however, this was only for droplet sizes of 83 and 310 µm. Droplet shattering is parameterized as the product of a fixed fragment number, a temperature-dependent shattering probability given by a normal distribution in temperature and the existing droplet freezing tendency used by Seifert and Beheng (2006). The normal distribution is centered at 258 K with a standard deviation of 5 K and a maximum probability of 10 % as illustrated in Fig. 2a).
The collisional breakup of ice particles, in ice-ice collisions, was introduced and studied in laboratory experiments by Vardiman (1978) and Takahashi et al. (1995) and found to be most effective at -15 • C. Takahashi et al. (1995) enfored the collision of large, 1.8 cm in diameter, heavily rimed ice particles with one another and generated secondary ice particles of up to 10 3 per collision. Yano and Phillips (2010b) and Yano et al. (2016) have demonstrated the generation of massive enhancement of the ICNC by SIP due to ice-ice collision in a dynamical system-type model. Recently, Phillips et al. (2017) 180 developed a more physically robust theoretical parameterization, that was also applied in numerical simulations that consider energy-conservation. In our case, following Sullivan et al. (2018a), we take a more simplified approach. The collisional breakup of ice particles in the laboratory work of Takahashi et al. (1995) resulted in a temperature-dependent parameterization of the fragment number: where α is the scale factor, F BR is the fragments generated during each collision, T is the temperature in Kelvin and γ BR is the decay rate of fragment number at warmer temperatures. In our breakup simulations, F BR and γ BR are the experimental factors that are adapted for evaluating the collisional breakup parameterization. The fragment number ℵ BR is multiplied by the 190 collisional tendency ∂N j /∂t of the colliding hydrometeor pairs to calculate the number of ice crystals generated per time step ∂N ice /∂t in Eq. (2). As shown in Fig. 2b), no collisional breakup occurs for temperatures below 252 K. Takahashi et al. (1995) forced the collision between heavily rimed ice particles at a velocity of 4 m s −1 . However, when the collision speed from the experiment is used to calculate the size of the involved graupel particles, a 4 m s −1 fall speed corresponds to graupel particles in the range of 2.5 mm in diameter according to Lohmann et al. (2016b). Considering the size-mass and fall-mass relations 195 that are used in COSMO, Blahak (2008) showed that for graupel and hail falling at 4 m s −1 the effective diameters were 4 mm and 1.4 mm respectively. Also, in the simulations conducted here, graupel sizes would very rarely exceed 3 mm in diameter (not shown here). Therefore, having smaller graupel particles than what was used in Takahashi et al. (1995), we expect ℵ BR to be less and thus we introduce α to prevent extreme overestimations in ℵ BR (Table 1). Another consideration to take into account is that COSMO treats snowflakes as unrimed particles and as soon as riming occurs on snowflakes, the snow mixing 200 ratio is converted to the graupel mixing ratio causing especially large graupel mixing ratio (Otkin et al., 2006). Since graupel is the only contributor to SIP through collisional breakup, increased graupel mixing ratios could lead to excessive SIP. Further sensitivity studies were conducted with γ BR of 2.5 instead of 5 as described in the parameterization used by Sullivan et al. (2018a). When γ BR is 2.5, ℵ BR will be reduced at warmer temperatures (Fig. 2b).
In the standard version of COSMO, the ICNC after each model time step is limited to 500 L −1 for each level. However, 205 measurements showed that the ICNC within MPCs produced higher ICNC of up to 1014 L −1 (Lohmann et al., 2016a). Korolev Table 1. Sensitivity settings for the collisional breakup parameterization. α is the scale factor, FBR the fragments generated and γBR the decay rate of fragment number at warmer temperatures. In bold is γBR = 5 as used in by Sullivan et al. (2018a).  small faceted ice crystals. Therefore, we increased the ICNC limit to 2000 L −1 in our model setup.
Using droplet shattering as the only active SIP process in our case study yields very similar results to the CNTL simulations with the exception that the SIP rate between 4 and 5 km was 0.01 L −1 s −1 . Therefore, it was not included in the rest of our 210 analysis (Fig. S1). Also, the initial analysis of the simulations with collisional breakup (for γ BR = 2.5) showed that the SIP rate is between 1× 10 −3 and 7 L −1 s −1 below 5 km yielding an ICNC between 0.1 and 300 L −1 at the surface ( Fig. S2a, g). Contrasted against these simulations are the collisional breakup simulations (for γ BR = 5) that showed an increased SIP rate between 100 to 1000 L −1 s −1 yielding ICNC of 2000 L −1 at the surface at Gotschnagrat (Fig. S3a, g). These ICNCs were strongly influenced by the ICNCs extending from the surface to 5 km. We chose the BR2.8_T settings because the ICNC was represented best near 215 the surface at Gotschnagrat and that the ICNC hard limit did not impact the simulations and therefore collisional breakup and its impacts on the MPC can be understood better. Further, we used the BR28 simulation as a comparison to understand what the effect would be by reducing the SIP at warmer temperatures than 263.5 K while at the same time increasing the SIP at colder temperatures than 263.5 K compared to the BR2.8_T settings (Fig. 2b). Higher ice particles number concentrations at colder temperatures can increase the competition for available cloud liquid water and glaciate the clouds at a faster rate slowing down 220 precipitation formation. Due to their smaller size as a result of the aggressive collisional breakup, these ice particles should have slower sedimentation velocities. We do expect that precipitation formation will be slower in the BR28 simulations and that there will be a leeward shift in surface precipitation assuming that the lower part of the MPC is supersaturated with respect to water.

225
3.1 Modeling ICNC and ice crystals growth rates at Gotschnagrat

Modeling ICNC
During each ∼2 min ascent, the HoloGondel platform on the cable car recorded ice crystal concentrations averaged over three altitudes (1808-1961, 1961-2113 and 2113-2266 m). We first show the ICNC inferred from HoloGondel measurements on each of its three ascents (at 11:54, 12:54 and 13:26 UTC) and as the outcome of COSMO simulations at Gotschnagrat (Fig. 1a).

230
The holographic images yield size and habit information of the ice particles. From this information, the precipitation forming processes can be inferred ( Fig. 3a, b). Note that the classification process for irregular ice particles is difficult to assign accurately as it can include aggregates of ice particles, blowing snow, ice particles sedimenting through different growth regimes or collisional breakup of ice particles. Therefore, the irregularly shaped particles can't be assigned to a specific category (e.g. SIP). The total ICNCs from the observations were averaged over altitude for each ascent and were 16±5, 19±4 and 9±3 L −1 at 235 11:54, 12:54 and 13:26 UTC respectively. The simulated ICNC from the CNTL and rime splintering simulations was less than 0.1 L −1 below 2.15 km for both 12:00 and 13:00 UTC (Figs. 4 and 5a) and were at least 2 orders of magnitude smaller than the observations at Gotschnagrat (Fig. 3). Between 2.2 and 3.2 km, the SIP rate from rime splintering reached 5× 10 −3 L −1 s −1 (Figs. 4 and 5g). Rime splintering originated exclusively from raindrops, with diameters between 150 and 250 µm at a concentration of 0.1 L −1 that rimed onto ice particles (Fig. S4e, o). A diameter of 25 µm required for rime splintering to be active 240 was not met by the cloud droplets, which only reached diameters of 20 µm ( Fig. 4 and 5f). The CNTL and rime splintering simulations reached cloud droplet number concentrations (CDNCs) of 100 cm −3 and droplet diameters between 10 and 22 µm, aiding in the primary ice production. The low ICNC by the CNTL and rime splintering simulations emphasized the need to explore SIP with collisional breakup. The ICNC in the BR28 simulation was between 1 and 2 L −1 at 12:00 and 13:00 UTC (Figs. 4 and 5a), within an order of magnitude of the HoloGondel observations. The ICNC from the BR2.8_T was higher 245 between 10 and 12 L −1 which compared better to the HoloGondel observations, albeit with a high uncertainty below 3 km at 12:00 UTC. The SIP rate of collisional breakup was 2 orders of magnitude larger than rime splintering for 265 K≤T≤270 K.
The ICNC from collisional breakup was significantly larger than the CNTL simulation above 3 km. This process resulted in lower liquid water and rain mass mixing ratios preventing primary ice production because of the absence of liquid water (Figs. 4 and 5a, b, c, g).

Modeling ice crystal growth rates
The enhanced SIP through collisional breakup had a significant impact on mostly glaciating the cloud above 2.5 km and can be seen in the nearly non-existent primary ice production rates due to the lack of cloud droplets (Fig. 4b, c). This also meant that the growth of ice crystals was mostly through vapor deposition that, in turn, suggests that the ice crystals were mostly pristine above 2.5 km (Fig. 4d, h). Below 2.5 km a very shallow liquid layer, that was subsaturated with respect to water was 255 present, causing ice particles to grow through the WBF process and/or riming at 12:00 and 13:00 UTC (Figs. 4,5a,b,d,f) and S4d, e). The shallow mixed-phase cloud layer was of interest, because close to the surface the HoloGondel observations showed that rimed particles were dominant, making up 47 -72 % of the total ICNC, indicating that the cloud (at least close to the surface), from 11:54 to 13:26 UTC, was in a mixed-phase state while passing over Gotschnagrat (Fig. 3b). Interestingly, observed irregularly shaped ice crystals were also present that could have been artifacts from collisional breakup of ice or snow. 260 However, it is also possible that the irregular ice crystals could have either been from ice crystals that fell through different growth regimes or they were from blowing snow. This would mean that only a subsection of the irregular ice crystals could have been from collisional breakup. Considering the lowest 700 m in the model we calculated the growth fraction of each of the growth mechanisms (riming, deposition and aggregation) of ice crystals (e.g. riming(%) = riming/(riming + deposition + aggregation) from Table 2). This was done to compare the ice crystal classification to the HoloGondel observations. This 265 was not meant to be a direct comparison because the ice crystals in the model could be rimed while also growing further by deposition and vice versa that could lead to double counting of processes. However, following this approach the BR2.8_T simulation showed that 2.6, 80.5 and 16.9 % of the growth was by riming, deposition and aggregation respectively at 12:00 UTC compared to the BR28 simulation showed growth fractions of 0.2, 80.3 and 19.5 %. At 13:00and 13:30 UTC the collisional breakup simulations had growth by riming fractions of below 5.4 % suggesting that most of the ice crystals would have been 270 either pristine or aggregated whereas the ice crystals observations were predominantly rimed in the observations. Evidently, the cloud liquid water that acts as rimers is underestimated in the collisional breakup simulations. These comparisons could not be carried over to the CNTL and rime splintering simulations due to the underestimated ICNC. The ice crystals that formed through primary ice production above 3 km sedimented through the overestimated liquid layer (Fig. 6b)

Modeling precipitation
To analyze the impact of ICNC on precipitation, we compared the cloud radar precipitation rate and the LWP from the microwave radiometer with that of the simulated precipitation rate and LWP. The observed precipitation started at 8:30 UTC and continued until 14:00 UTC reaching maximum precipitation rates over 15 min intervals of 3.75 mm h −1 at 12:00 UTC as 280 the cold front passed over Davos Wolfgang. All the simulations were unable to time the onset of the precipitation and underestimated the precipitation before 10:45 UTC (Fig. 6a). After 10:45 UTC the collisional breakup simulations overestimated the precipitation as compared to observations. The CNTL simulation also outperformed the rime splintering simulation which represented the precipitation rate more accurately. This underlines the difficulty that models have in simulating mountainous weather in general (Rotach and Zardi, 2007;Panosetti et al., 2018). The collisional breakup simulations, from 11:30 to 285 13:30 UTC, mostly underestimated the microwave radiometer liquid water path which resulted in the MPC consisting of less than 10 % liquid water (Fig. 6b, c).
As MPCs approach glaciation, precipitation formation through the WBF process slows down when the updraft velocity is not high enough to stabilize the MPC (Korolev and Mazin, 2003). In the collisional breakup simulations, in which the ICNCs were between 10 1 and 10 3 L −1 , the updraft velocities were on the order of -0.2 to 0.6 m s −1 at altitudes between 1.7 and 4.3 km. above the surface from 12:00 to 13:00 UTC. Most of this layer, however, was subsaturated with respect to water and the ice particle growth was via the WBF process aided in precipitation formation.
In Fig. 7 we consider the precipitation over the larger domain (e.g. the red box in Fig. 1). The average wind direction between 12:00 and 14:00 UTC, and between 2 to 4 km above the surface, came from the south-west (indicated by the red 295 arrow on Fig. 7a). In general, the spatial precipitation patterns of all the simulations over the domain have a moderate to strong relationship, correlations above 0.68 with high confidence, with CombiPrecip (Table. 3). Important here is that the Pearson correlation, in general, is sensitive to outliers in a skewed distribution which can lead to being a less desirable statistic. This was the case with surface precipitation over a domain and, therefore, we also calculated the interquartile ranges with the 25th and 75th percentiles to compare the observations with the simulations. CombiPrecip had a narrower precipitation distribution 300 indicated by the interquartile range between the 25th and 75th percentiles of 1.09 mm h −1 than all the simulations meaning that the precipitation had less variability. This can be seen in the domain as all the simulations had regions of localized high precipitation rates. This resulted in a higher variability with an interquartile range between 1.46 and 1.79 mm h −1 of which 75 % of the precipitation were between 1.7 and 2.14 mm h −1 (Fig. 8). Higher variability, e.g. in the CNTL and rime splintering the SIP processes were excluded, causing the higher variability, can most likely be attributed to the dynamics, rather than the cloud microphysics scheme. Craig and Dörnbrack (2008) demonstrated that a model resolution of ∼1 km is approximately equal to the characteristic turbulence scales of convective structures. Therefore, using a 1D turbulence scheme, which we used and is generally used in cloud-resolving models, is not optimal and in the "grey zone". Also, horizontally homogeneous conditions are assumed in most turbulence schemes and have been validated over flat terrain (e.g. Mellor and Yamada, 1982;310 Rotach and Zardi, 2007) which is not the case for our study. Earlier and higher precipitation rates can occur if mountains are high enough to force the flow into an elevated mixed layer leading to a faster transition of deep convection (Panosetti et al., 2018).
The localized regions of invigorated precipitation rates were suppressed by including the SIP processes. Because collisional breakup is a mechanical process it doesn't contribute directly to the latent heat budget and, therefore, shouldn't invigorate the 315 updraft velocities. However, larger number concentrations of ice particles can cause increased depositional growth and thereby change the buoyancy structure of the cloud. Stronger updrafts could then loft the smaller ice particles to higher altitudes reducing their sedimentation velocities towards the surface. Evidence for the effect of the strong SIP rate on ice particle size can be seen in Figs. S4 and S5k, l and m. The BR28 simulation, which was able to produce the largest SIP rates, resembled the the BR28 simulations matched the observed one better but was shifted towards the south which negatively influenced the correlation. In summary, we have shown that COSMO benefits from the inclusion of collisional breakup processes in simulating 355 ICNC and precipitation.

Discussion
Our study suggests that including SIP through collisional breakup can enhance the in-situ ICNC and consequently surface precipitation. The collisional breakup simulations led to ICNCs at Gotschnagrat of one to two orders magnitude larger than in the rime splintering simulations. More precisely, the numbers were between 1 and 2 L −1 and 10 and 12 L −1 for BR28 and 360  BR2.8_T, respectively, compared to the rime splintering simulations of less than 0.1 L −1 . The BR28 simulation most closely represented observations, albeit leading to ICNC an order of magnitude smaller than observed at Gotschnagrat. However, blowing snow cannot be excluded as a contributor to the measured ICNC (e.g. Farrington et al., 2016;Beck et al., 2018) even though we were not able to quantify this process in our analysis. As a consequence, the current discrepancy in ICNC between the BR28 simulation and the observations could be overestimated.
365 Surprisingly, our SIP rate through collisional breakup was between 10 4 to 10 6 times larger than what Sullivan et al. (2018a) reported in their cold front rainband study. We hypothesize that this large difference is due to the availability of cloud liquid  water that rimes onto ice crystals and snowflakes. This process converts ice and, especially, snow quickly to large quantities of graupel which causes higher SIP rates through collisional breakup. In the Sullivan et al. (2018a) case, graupel was noticeably low, which limited graupel-ice/snow interactions. The low graupel concentrations meant that the SIP from collisional breakup 370 simulation was at least 10 3 times smaller than that of the rime splintering simulation which was at around 10 to 100 L −1 .
The excessive SIP rate through collisional breakup in our simulations, in combination with a larger graupel mixing ratio (Fig.   11a), is likely due to ideal conditions for collisional breakup (Fig. 9 Sullivan et al., 2018b). Indeed, our simulations have a cloud base temperature of around 273 K and averaged updraft velocities through our cross-section of up to 0.6 m s −1 . The large SIP rates, in our case, counteracted the stronger precipitation in localized regions. This decreased precipitation is in direct 375 contrast to Sullivan et al. (2018a) showing that when using secondary ice parameterizations, regions of invigorated precipitation became even wetter. A possible explanation for the decreased precipitation is that the smaller ice particles in our collisional breakup simulations are lofted to higher regions within the cloud which is already glaciated in winter orographic MPC. These ice particles take a longer time to sediment to the surface and alter the location and intensity of the surface precipitation.
If the middle to the upper part of the cloud were not glaciated, we would have indeed, similar to Sullivan et al. (2018a), 380 expected localized increase in surface precipitation due to faster ice particle growth. The SIP rates in our simulations compared better with Phillips et al. (2017) even though they simulated a convective storm with updrafts exceeding 5 m s −1 which is not comparable to the meteorological situation we simulated. However, in their case, the collisional breakup parameterization was more physically robust by including the kinetic energy of two particles, the fragility coefficients, humidity-and temperaturedependent collision types into account. As stated in 2.2.3, we set an ICNC threshold of 2000 L-1 in our model. This threshold strongly restricts the ICNC. Therefore, the full effect of using higher coefficients for F BR could not be realized. We ran simulations with no ICNC threshold and observed evidence for violations against mass conservation in the model. This ICNC threshold is necessary, particularly in the collisional breakup simulations.
Another aspect that could enhance collisional breakup is the conversion from snow to graupel. For collisional breakup to 390 occur, graupel formation is necessary. Graupel formation can only happen when either cloud droplets or raindrops rime onto ice crystals or snow. Because snow in the model is described as pristine (unrimed), a tuning parameter is used to rapidly convert snow to graupel when raindrops rime onto snow (Seifert and Beheng, 2006). Alongside this tuning parameter, Seifert and Beheng (2006) set a threshold to convert ice crystals and snow to graupel only if they are larger than 500 µm in diameter to suppress the early formation of very small, 200 to 400 µm, graupel. As soon as graupel forms within the collisional breakup 395 regime (temperatures warmer than 252 K), collisional breakup occurs. In the standard model version, the threshold has been set to 200 µm (e.g. Seifert and Beheng (2006)) which encourages earlier graupel formation. Using this threshold in conjunction with collisional breakup could be the a reason why we saw such large SIP rates. It is worthwhile to consider the size conversion threshold when secondary ice processes are used in cloud microphysics schemes.    Takahashi et al. (1995) and used here with different configurations of F BR and γ BR to account for the hydrometeor size scaling.
To conclude, our main findings can be summarised as follows: -Droplet shattering did not show significant differences to the CNTL simulation. This is mainly due to the low raindrop number concentration present during the cold front passage over Gotschnagrat. Aside from using a shattering probability 410 of 10 %, which is similar to Sullivan et al. (2018a), freezing raindrops in COSMO are spectrally partitioned into ice ICNC at the surface. Secondarily formed ice crystals through rime splintering was at most 0.9 L −1 at 3 km and decrease significantly below 3 km. The fact that small concentrations, less than 0.5 mg m −3 , of raindrops between 100 and 150 µm were available in the rime splintering region and that Gotschnagrat was not in the rime splintering temperature regime on 7 March limited the rime splintering process. At Davos Wolfgang, all the simulations could not adequately represent the radar precipitation.

420
-The BR28 and BR2.8_T simulations showed enhanced ICNC at the surface compared to the CNTL simulations and represented the observed ICNC at Gotschnagrat very well. The enhanced SIP production did impact the cloud liquid water, reducing the LWP significantly through the cloud layer and underestimating the LWC at Davos Wolfgang. A 700 m shallow layer of cloud liquid water near the surface was maintained causing cloud droplets and raindrops to rime onto the ice crystals corresponding to rime fractions of less than 5.4 %). However, this could not explain the HoloGondel 425 observations showing that over 50 % of the ice crystals were rimed.
-Including collisional breakup showed to be beneficial for simulating precipitation over the domain. The simulations presenting the stronger SIP, BR28, showed the most improvement in timing and amount of surface precipitation during 7 March. Regions of invigorated precipitations in the CNTL and rime splintering simulations generally were more suppressed when the secondary ice parameterizations were used bringing the precipitations rates closer to the observations.