The CLOUDLAB project aims to improve the understanding of ice formation and growth through glaciogenic cloud seeding and entails three field campaigns: (1) January–March 2022, (2) December 2022–February 2023, and (3) December 2023–February 2024. During the first two campaigns, 55 successful seeding experiments were conducted at various temperatures, under various wind conditions, and using various seeding particle emission rates. detail the observational setup, including the analysis of four seeding experiments from the first two campaigns. show the possible seeding patterns conducted in the field experiments. Their Fig. 7 displays how in-cloud experiments consist of several seeding legs per experiment, during which the UAV is flying back and forth while the seeding flare is burning for approximately 6 min. The field experiments discussed in this study were performed using four respective 400 m legs. Here, we focus on five seeding experiments that were conducted on 2 consecutive days: three experiments on 25 January 2023 (S25-2, S25-2.5, and S25-3), which are also discussed in , and two experiments on 26 January 2023 (S26-2.5a and S26-2.5b). We closely followed the seeding distances from the field site, the seeding patterns, and the seeding start time (Table ). The seeding simulations are named by combining their experiment date and their seeding distance, e.g., the seeding experiment on 25 January 2023 at 2.5 km distance is called S25-2.5. For 26 January 2023, where both experiments are identical in their setup, we introduced an additional identifier in the form of “a” and “b”, i.e., S26-2.5a and S26-2.5b. These two identical setups serve as a test for the validity of the signal that we observe in the field experiments as well as in the model simulations.

Both January days were characterized by low stratus clouds and low temperatures, which are the targeted conditions for conducting seeding experiments in this project. Here, we only briefly discuss the relevant instrumentation; a full description of the experimental approach and instrumentation can be found in .

For model validation, we compared the model simulations to the atmospheric profiles measured by radiosondes (Sparv S1H3, Windsond) and to the radar reflectivity observed by a vertically pointing cloud radar (FMCW-94-DP, Radiometer Physics GmbH) (Sect. ). The seeding simulations (Sect. ) are compared to elevation scans conducted by a cloud radar (Mira-35, METEK) and in situ measurements obtained with a tethered-balloon system (TBS) (Sect.  and ). The TBS is equipped with a holographic imager for microscopic objects (HOLIMO) to observe cloud characteristics, such as cloud droplet and ice crystal number concentrations and size distributions. The HOLIMO measures particle diameters in the range between 6 µm and 2 mm . Given the nonspherical shape of ice crystals, a mean area equivalent radius is calculated for a more direct comparison to the model data in Fig. . The differentiation between cloud droplets and ice crystals is based on the particle's shape for radii larger than 25 µm. Anything below that is identified as cloud droplets due to the resolution limit of HOLIMO. After applying a neural network to classify the particles into cloud droplets and ice crystals, a manual labeling of all ice crystals was conducted to minimize misclassifications. Overall the following uncertainties apply: cloud droplet number concentration (±5 %), ice crystal number concentrations for particles larger than 100 µm in diameter (5 %–10 %), and ice crystal number concentrations for particles smaller than 100 µm in diameter (15 %) (for further information, see , and ). The HOLIMO data were analyzed at a 5 Hz frequency during the seeding experiment, whereas the background cloud was analyzed at 1 Hz. Subsequently, the HOLIMO data were averaged over five data points over time, resulting in 1 s averages for the seeding signal and 5 s for the background. The time periods before and after seeding are used to observe the unperturbed background characteristics of the cloud. We further assume that the highly localized measurements by HOLIMO serve as a representative sample of the cloud characteristics inside the seeding plume and of the background due to persistent low stratus clouds in a stable boundary layer.

We employed the ICON-v2.6.5 numerical weather prediction model in a one-way nested mode (Fig. ). The outermost nest has a 1 km horizontal resolution, 80 vertical levels (up to 22 km), a model time step of 10 s, and is based on the operational grid of the Swiss National Meteorological Service (MeteoSwiss) . The second nest with a 260 m horizontal resolution and a model time step of 2.5 s is located over the Swiss Plateau, centered around the field site. The final nest has a 130 m horizontal resolution with a 1 s model time step and a domain size of 40 × 30 km2 (Fig. ). Both inner nests also have 80 vertical levels and a vertical grid spacing ranging from approximately 20 m (at cloud base) to 80 m (at cloud top). The initial and boundary conditions for the outermost nest are based on hourly Consortium for Small-scale Modeling (COSMO) analysis data generated by MeteoSwiss, and the simulations were initialized at 00:00 UTC and performed for 23 h. For each simulation, we conducted a reference (no seeding) and a seeding simulation to quantify the impact of cloud seeding. The innermost nest with the highest resolution is the focus of this analysis. All three nests employ the ecRad scheme , the “3D Smagorinsky diffusion” turbulence scheme , and the two-moment microphysics scheme. The cloud condensation nuclei concentration was set to 1000 particles cm-3 following for rural and continental areas (Melpitz, Germany; CESAR tower, the Netherlands; and Vavihill, Sweden) during wintertime and is uniformly distributed in the domain. The frequency of model output was set to 5 min, after also testing a 1 min output frequency (which showed similar results as in the 5 min output). Moreover, calculating the expected arrival time of the seeding plume at the field site (seeding start and growth time; see Table ) shows that the expected arrival and a full 5 min model output time step are very close (within ±1 min).

In the following, we provide a short description of the two-moment microphysics scheme used in the model. For more details, please refer to . The scheme tracks mass and number mixing ratios for six hydrometeors – cloud droplets, ice crystals, snowflakes, graupel, hail, and rain drops – by assuming a Gamma distribution for the underlying size distributions. All relevant cloud processes are parameterized, such as cloud droplet activation, ice crystal nucleation, growth of ice crystals by water vapor deposition (i.e., the WBF process), riming, melting, and sublimation of ice crystals. The maximum diameter and terminal fall velocity are parameterized following power laws with constant coefficients . As we are investigating the ice processes within mixed-phase clouds, we provide additional information relevant to ice particles. The ice crystal shape is set to be spherical, which is a simplification in the scheme given the large variety of shapes . In this study, we do not change the shape of the ice crystals, as we want to investigate the ice crystal growth rate in the default configuration of the model. During the conducted seeding experiments, we mostly measured needles or columns. When we compare the ice crystal sizes in Fig. , we investigate the mean equivalent radius of ice crystals.

Simulations S26-2.5a and S26-2.5b are identical with respect to their setup and serve as a proof of concept for the seeding approach in the model. Simulations S25-2, S25-2.5, and S25-3 highlight the impact of the seeding distance on the ice crystal growth time. These three simulations have been conducted at a warmer temperature than S26-2.5a and S26-2.5b. We use this difference in environmental conditions to evaluate the temperature dependence of ice nucleation induced by the seeding particles.

At subzero temperatures secondary ice production can occur, which is also parameterized in the model. For secondary ice production to occur in the model, rimed graupel particles are needed; however, their concentrations are close to zero in the model. Hence, we can exclude the effect of secondary ice production in our analysis. During the field experiments, we also expect a low secondary ice production rate, given that only a few larger cloud droplets (with radii > 20 µm) are present. Riming on the ice crystals was also only visible in two out of the five experiments (S26-2.5a/b). In addition, if splinters occurred, they probably did not grow large enough to be detected, given the short growth time during the experiments (see Table ).

The selection of the presented seeding simulations was constrained by how accurately the model reproduced the observed environmental conditions. Unfortunately, the model overestimated the temperatures for 25 January 2023 (Fig. a), while the temperatures on 26 January 2023 were simulated adequately (Fig. b). For this reason, we decided to utilize the simulation of 26 January 2023 for all seeding experiments conducted on 25 and 26 January 2023 (see also Sect. ), given the presence of persistent low stratus clouds with northeasterly to easterly winds on both days. While the experiments on 26 January 2023 (S26-2.5a and S26-2.5b) were simulated at the same seeding height as the field experiments (Fig. ), the experiments on 25 January 2023 (S25-2, S25-2.5, and S25-3) were simulated at a lower height such that the model temperature matches the seeding temperature from the field (-5.5 °C; Table ).

To simulate glaciogenic cloud seeding in ICON, we implemented a deterministic freezing parameterization specifically for the seeding particles (AgI) used in the field. hypothesized that the seeding particles are highly hygroscopic and postulated that these are either taken up by existing cloud droplets or undergo cloud droplet activation before freezing. Furthermore, their clear-sky seeding experiments showed that the seeding particles have a mean particle diameter of between 100 and 400 nm. The freezing ability of AgI particles for sizes between 20 and 400 nm was measured by in a laboratory setup (their Fig. 1), and the aforementioned work showed a strong size and temperature dependence for the frozen fraction (FF). However, for sizes larger than 40 nm, the freezing curves are fairly similar; thus, we decided to follow the 400 nm measurements. We obtained the following relationship: 1FF=-b1+exp⁡-k(T-T0)+b, with the parameters b=0.97, k=0.88, and T0=263.95 K, where T is the temperature (in kelvin). Figure a shows the laboratory measurements (400 nm) of , indicating an increase in the FF with decreasing temperature, highlighting the strong temperature dependence of the ice activity of AgI particles. Within our model setup, we introduced the Marcolli parameterization in the two-moment microphysics scheme before the dust freezing parameterization takes place . For the analysis, we assume immediate ice nucleation, given the high ice nucleation activity of AgI below -5 °C. Further, we limit the ice-nucleating activity by the availability of cloud droplets. Hence, the number of ice crystals cannot exceed the number of available cloud droplets.

In the model, the seeding particles were introduced along a 400 m leg to mimic the seeding pattern conducted in the field experiments (see Sect. ). We used the coordinates of the seeding legs to define the injection area for the seeding particles, as shown in Fig. b. Each conducted simulation received the same seeding particle emission rate: 106 seeding particles m-3 s-1 were released in three model grid boxes for 6 min, corresponding to the burning time of one seeding flare. This seeding particle emission rate is based on a series of sensitivity simulations for seeding experiment S26-2.5a, where we injected different concentrations of seeding particles into the model and compared the simulated ice crystal number concentrations to the observations. The seeding particle emission rate and, thus, the seeding setup were constrained by the ice crystal number concentrations observed by HOLIMO and chosen in such a way that they match seeding simulation S26-2.5a (Sect. ).

We applied a simple method to extract the seeding signal from the background. We took the difference in ice crystal number concentrations between a seeding simulation and a reference simulation (no seeding) to remove the background and isolate the seeding plume. The seeding plume was then defined by a threshold ice crystal number concentration of 0.001 particles cm-3. We used the identified seeding plume as a mask to extract further quantities in the seeding simulation but also in the difference between the seeding and reference simulation, such as cloud droplet number concentrations, temperature, and updraft changes caused by the seeding perturbation. The analysis of the seeding plume and related quantities is based on this approach to quantify the cloud-seeding impact. For the comparison to in situ measurements, we considered the model output time step closest to the expected arrival of the seeding signal at the field site using the calculated growth time in Table . For simulations S25-2, S25-2.5, and S25-3 the growth times differ from those reported in , as we use the time between the ignition of the seeding flare and the first increase in radar reflectivity in a vertically pointing radar at the field site (i.e., arrival of seeding plume) in this work. used the wind profiler for the wind measurement and estimated the advection time from this. However, the lowest level of measurements in the wind profiler was consistently higher than the actual seeding altitude, which leads to an overestimation of the wind speed and, thus, an underestimation of the ice crystal growth time.

To validate the model, we compared the simulated to the observed conditions on 25 and 26 January 2023. Figure  shows the vertical profiles of temperature and relative humidity measured by a radiosonde (solid lines) launched from the field site and predicted by the model (dashed and dotted lines). Both days were characterized by low-level clouds with subzero temperatures, with 26 January showing a deeper cloud and colder temperatures. In both cases, the model did not reproduce the sharp inversions at cloud top; moreover, for 25 January 2023, it did not simulate cold enough temperatures for the seeding particles to be effective. Therefore, we used the 26 January 2023 simulation as a surrogate for all seeding simulations (see also Sect. ), as it better represents the observations on 25 January 2023 (Fig. a). The morning of 26 January 2023 was characterized by a strong temperature inversion at around 1.6 km a.m.s.l. (above mean sea level), limiting the cloud top to that height. The model predicted a lower and weaker inversion compared with the observations, with a corresponding lower cloud top. This discrepancy does not pose a problem, as the observed and simulated temperatures at the seeding height (see temperature in Table ) are in good agreement, which is the strongest constraint for our seeding simulations. Regarding the wind speed, the model performs reasonably well with slight underestimations for four of the cases (0.4–0.7 ms-1; Table ). However, the model performance is worse with respect to the wind direction (±13°; Table ). The discrepancy in wind speed and direction does not impact the ability of the model to simulate seeding experiments.

In addition, we compared the observed and predicted cloud cover at the field site by taking the radar reflectivity of a vertically pointing radar as a proxy for cloud cover and the computed cloud cover from the prognostic cloud water mass in the model (Fig. ). We see that the model predicts a long-lasting low cloud that reaches slightly lower cloud top heights than observed by the radar, with the seeding simulations still being fully inside the cloud. The lower cloud top can also be seen in the comparison of relative humidity in the radiosonde profiles.

In the following, we discuss the seeding impact on cloud droplet and ice crystal number concentrations. We first show the evolution of the seeding plume in terms of time and location and then compare it to radar scans (Sect. ). We next compare the observed and predicted ice crystal number concentrations for all simulations (Sect. ). This is followed by the investigation of the WBF process in the model by utilizing the in situ measurements (Sect. ). We conclude with the impact of the seeding particle emission rate on ice crystal number concentration and cloud droplet reductions (Sect. ).

The in situ observations taken during the seeding experiments allow us to investigate the WBF process in greater detail. Figure a shows the measured cloud droplet number concentrations, while Fig. b presents the ice crystal number concentrations during seeding experiment S26-2.5a, including measurements of the undisturbed background cloud prior to and after seeding. The analysis for the other experiments is shown in Appendix  (Fig. ), and an overview of all experiments is given in Table . The background cloud had an ice crystal number concentration of 0 particles cm-3 and a median cloud droplet number concentration of 320 particles cm-3 (Fig. ). During the seeding experiment, the ice crystal number concentrations increase up to 1 particle cm-3, whereas the cloud droplet number concentrations simultaneously decrease by up to 300 particles cm-3.

To quantify the reduction in cloud droplets throughout the seeding experiment, we first defined the cloud droplet number concentration of the background state by calculating the median over a 20 min period of the background. Second, we considered all times in our analysis during which the observed ice crystal number concentration was larger than 0.001 particles cm-3. The time periods considered in the analysis are marked by the vertical brown lines in Fig. a and b, while the relative reductions in the cloud droplet number concentrations with regard to the median concentration are shown as a frequency distribution in Fig. c–e for different model output time steps. The observations indicate that, between the start of seeding and the measurements (i.e., 8 min), the liquid phase was entirely depleted during some time periods (i.e., cloud droplet reductions of 100 %), emphasizing the high efficiency of the WBF process observed in the field. Thus, the freshly nucleated ice crystals are highly efficient with respect to consuming water vapor from the evaporating cloud droplets, leading to increased ice crystal growth and, eventually, isolated glaciated patches inside the cloud. These strong reductions in cloud droplets can partly originate from riming, where the cloud droplets immediately freeze onto the ice crystals. Only for experiments S26-2.5a and S26-2.5b was riming observed with HOLIMO. Here, we solely focus on the WBF process, but we are aware that riming can also lead to additional reductions in the liquid phase.

Next, we compare the observations to the simulated changes in the model by taking the model output time step that is closest to the time the seeding signal is expected to arrive at the field site. The changes in the model are computed from the difference between the reference and seeding simulation. In addition, we constrained the seeding plume by the available liquid water content inside the cloud: we only considered grid cells in the analysis with an ice crystal number concentration larger than 0.001 particles cm-3 and liquid water content larger than 0.1 gm-3. This way, we only include grid cells in which the WBF process actually could take place. Figure c shows both the observations and the model response, and we can see that the model is not able to reproduce such strong reductions in cloud droplet number concentrations, indicating that the ice crystal growth in the model is underestimated. This discrepancy may originate from the computation of the ventilation coefficient, which determines the speeding up of the diffusional growth due to turbulent motions. This hypothesis will be investigated in future studies. When looking at later model output time steps, we see that the model predicts a comparable reduction in cloud droplet number concentrations to that observed (Fig. d–e). This further points to the fact that the WBF process in the model in its current form is not efficient enough, as more time is needed to reach similar cloud droplet reductions.

In Table , we show the median, mean, and maximum ice crystal number concentrations (absolute values). For cloud droplets, we report the reduction in the median, mean, and maximum cloud droplet number concentration relative to the undisturbed background (relative values) to account for the lower median cloud droplet number concentration of approximately 100 particles cm-3 in the model compared with the observations (Fig. ). The maximum ice crystal number concentrations are in good agreement (within ±0.3 particles cm-3) with observations in four out of five simulations. Only the S25-2 simulation strongly underestimates the maximum ice crystal number concentration by 1 particle cm-3 (see Sect.  and ), whereas the simulated median ice crystal number concentrations match the observations well. This is not the case for the other four simulations, where the median concentration is underestimated by an order of magnitude. When we also consider the mean values, we see that the model generally only has a few grid cells with high ice crystal number concentrations, whereas a lot of grid cells have very low ice crystal number concentrations. Regarding the changes in cloud droplets, the model fails to reproduce the maximum cloud droplet reductions, where four out of five simulations show almost no reduction. Only in simulation S26-2.5a is a stronger reduction in cloud droplet number concentration notable. However, the median and mean cloud droplet reductions are strongly underestimated for all simulations.

As a next step, we aim to identify when suitable conditions for the WBF process exist inside the seeding plume by applying a set of theoretical equations : 2w′<w<w*,wherew′=es,i-es,wes,wNwr‾wχandw*=es,w-es,ies,iNir‾iη. Here, w is the vertical velocity; w* and w′ are the computed vertical velocity thresholds; es,i and ew,i are the saturation vapor pressures with respect to ice and water, respectively; Nw and Ni are the number concentrations for cloud droplets and ice crystals, respectively; and r‾w and r‾i are the mean radius for cloud droplets and ice crystals, respectively. η and χ are terms dependent on ambient temperature and pressure used to calculate the thermodynamic equilibrium. For w>w*, both cloud droplets and ice crystals grow, whereas they shrink for w<w′.

Conditions favorable for the WBF process are therefore constrained by the vertical velocity and the integral ice crystal and cloud droplet radius. For positive vertical velocities, i.e., updrafts, the ice crystal number concentrations and mean radii define the WBF regime, whereas for negative vertical velocities, i.e., downdrafts, cloud droplet number concentrations and mean radii define the WBF regime. Hence, for 0<w<w*, the ice phase acts as a sink for the supersaturation generated by the updrafts. For 0>w>w′, the liquid phase is constraining the WBF process, as the downdrafts reduce the supersaturation. As long as cloud droplets are present, they serve as a source for the ice crystals to grow. We call these two conditions updraft WBF (WBF↑) and downdraft WBF (WBF↓). To quantify the occurrence of the WBF process, we determined which of these two regimes prevails for each model grid box inside the seeding plume (Fig. a–g). For most of the time steps shown, roughly 50 % or more of the plume grid boxes are within the WBF regime. Both cloud droplets and ice crystals can grow inside the plume about 40 % of the time (w>w*). The simultaneous evaporation and sublimation (w<w′) of the hydrometeors occurs least often (< 2 % of the time) within the seeding plume. Under WBF conditions, the WBF↓ is dominant, which is further supported by the vertical velocities in Fig. h–n, which indicate the presence of downdrafts inside the seeding plume. We note here that we cannot distinguish between the microphysical (latent heat release) and dynamical (topography and wind field) influence on ice crystal growth and evaporation of cloud droplets. discussed that some updraft invigoration may occur due to latent heat release upon ice nucleation; however, this is still under debate.

Based on the prognostic mass and number mixing ratio for the ice and cloud droplet phases, we calculated the ice crystal and cloud droplet radii assuming a spherical shape for all particles (Fig. o–ab). We can see that the mean ice crystal radius increases over time, whereas the mean cloud droplet radius slightly decreases, indicating that the WBF process in fact takes place. Also, the mean radius of cloud droplets in the reference simulation is consistently larger than in the seeding simulation (see Fig.  in Appendix ). Additionally, we show the mean radius for ice crystals and cloud droplets measured by HOLIMO (see Sect. ) during the seeding event (see t2). While the cloud droplet radius distributions match well (Fig. w), the model underestimates the mean ice crystal radius by almost 30 µm (Fig. p). Also, the observations show a broader range of ice crystal radii than was simulated. This clearly indicates that the WBF process in the model is too slow to reproduce the observed growth rates in the field and/or the turbulent motions are too weak. Finally, we diagnosed the rate of water vapor deposition onto the ice crystals during the model simulation (Fig. ac). Initially, we simulated the largest vapor deposition growth rates (up to 1.5 × 10-4 g m-3 s-1), with ice crystals growing the fastest, which can also be seen in the evolution of the ice crystal radius over time (Fig. o–ab). At later model output time steps, the rate decreases to almost 0 gm-3s-1, when ice crystal no longer grow due to vapor deposition.