Assessment of the WRF Model as a Guidance Tool Into Cloud Seeding Operations in the United Arab Emirates

With the projected expansion of arid/semi‐arid regions in a warming world, precipitation enhancement activities such as cloud seeding will become increasingly popular and relied upon. Due to the inherent costs, a successful planning is crucial, which involves accurate model predictions. In this study, the usefulness of the Weather Research and Forecasting (WRF) model forecasts for guidance into seeding operations in the United Arab Emirates, where seeding activities have been conducted for more than two decades, is assessed. The WRF predictions are compared with ground‐based, satellite‐derived and radar reflectivity data, and in‐situ observations onboard the airplanes used to perform the seeding operations. WRF is found to have higher skill in simulating the observed cloud top pressure/temperature than the cloud fraction, with the model vertical velocity predictions also more skillful than those of the radar reflectivity. A stronger Arabian Heat Low (AHL) in the model leads to drier conditions which, together with a surface cold bias, limits the spatial extent and vertical depth of the simulated convective clouds. Development of convective rolls in the boundary layer is reported in both observations and simulations and their interaction with cold pools from convective clouds promote the development of secondary convection. Sensitivity to the choice of the Planetary Boundary Layer (PBL) scheme is also noticed, with the Yonsei University PBL scheme giving the best performance. When considering the two factors needed for a successful seeding operation that is, the presence of an updraft and clouds, the model‐predicted seeding regions largely match the areas where precipitation was observed. As the proposed WRF set up can be used operationally, the model forecasts will bring added value to the seeding activities in the country.

The paper is structured as follows: Section 2 gives an overview of the meteorological conditions on 23 August 2019 in the Arabian Peninsula using reanalysis data, while a summary of the WRF model setup and the observational datasets and verification diagnostics employed in this study is presented in Section 3. In Section 4, the evaluation of the WRF forecasts is performed. On that day, hygroscopic cloud seeding was conducted by the UAE's National Center of Meteorology (NCM) on the northeastern side of the country, where clouds developed. Together with ground-based, satellite-derived and radar reflectivity measurements, the observations by the pilots who performed the seeding, in particular with respect to the presence and strength of convective updrafts, are used to gauge the WRF predictions' capability of being used as guidance for cloud seeding operations in the UAE. The main findings of the study are outlined in Section 5.

Description of the Event
Figures 1a-1b show the large-scale atmospheric conditions on 23 August 2019 at 06, 12, and 18 UTC from ERA-5 reanalysis data (Hersbach et al., 2020), available on a 0.25° × 0.25° grid and on an hourly basis. The subtropical high on this day stretched from northeastern Africa to northern Saudi Arabia, Iraq and western Iran, with a mid-level trough ahead of it extending all the way to the UAE, Oman and southeastern Saudi Arabia. The presence of a mid-level trough preconditions the environment for the occurrence of convection as it leads to a conditionally unstable stratification ( Fonseca, Francis, Nelli, and Thota (2021). (b) The top panels give the 2-m temperature (shading; (K) and 10-m wind (arrows; m s −1 ); the middle panels give the 2-m water vapor mixing ratio (shading; g kg −1 ) and 10-m wind (arrows; m s −1 ); the bottom panels give the sea-level pressure (shading; hPa) and 850 hPa wind (arrows; m s −1 ) for the domain 20-30°N and 49-61°E. All fields are given at 06, 12 and 18 UTC and are extracted from ERA-5 data.
FONSECA ET AL. 10.102910. /2022EA002269 4 of 29 et al., 2021. Also favoring the triggering of convection is the ITD: the associated low-level wind convergence is over the Al Hajar mountains and northern parts of Oman. The AHL is the dominant feature in the sea-level pressure maps, and is in general agreement at 06 UTC with that estimated from the low-level atmospheric thickness (LLAT; 700-925 hPa geopotential height thickness) given by the stippled region. As noted by Fonseca, Francis, Nelli, and Thota (2021), while the sea-level pressure and LLAT-defined thermal low are in general agreement at night, during daytime the former metric gives a better representation of the AHL. This is because the low-level temperature field is more impacted by the heterogeneous surface properties, in particular by the surface albedo pattern. The southeasterly winds associated with the cyclonic circulation of the AHL extend into central parts of the Arabian Gulf, with the northwesterly (Shamal) winds restricted to the northern side. This is the typical synoptic pattern in the region in the months July and August (Al Senafi & Anis, 2015).
At more local scales, Figure 1b, the southwest monsoon flow turns northwards and then northwestwards around Oman reaching the eastern UAE coastline and the Strait of Hormuz. As the more moist tropical air encounters the Al Hajar mountains, it is forced to rise and the associated moisture condenses, leading to the formation of low-level clouds which persisted throughout the local morning hours, Figure 2a. This moist air mass, with mixing ratios in excess of 20 g kg −1 , penetrates inland and is later reinforced by the afternoon sea-breeze circulation. It converges with the cyclonic circulation associated with the AHL, the sea-breeze from the Arabian Gulf, and the ITD which remained over the mountains throughout the day (Figure 1a). The low-level wind convergence and mid-level trough therefore set the scene for the afternoon convection around the mountains, Figures 2a and 3a-3c, a similar set up to that discussed in Francis, Chaboureau, et al. (2021) for 05 September 2017 convective event. The interaction of the sea-breeze from the Arabian Gulf with the southeasterly winds from the Arabian Sea around 18 UTC also led to the development of clouds in parts of southern UAE and adjacent Saudi Arabia, Figures 4a-4c. As expected, these clouds are shallower than the afternoon convective clouds that formed around the Al Hajar mountains (cf. Figures 3b-3c with 4b-4c). It is interesting to note that the AHL's circulation, reinforced by the monsoon winds blowing from the Arabian Sea, leads to more than a doubling of the 2-m water vapor mixing ratio over inland areas in western UAE and southeastern Saudi Arabia in the evening and nighttime hours, rising from about 12 g kg −1 at 12 UTC to more than 24 g kg −1 at 18 UTC. However, and for the same reason, in the deep inland desert the air dries out considerably at night with mixing ratios dropping below 4 g kg −1 , more than six times smaller than those over coastal areas. The land-sea temperature gradients of more than 10°C and the referred moisture gradients seen on this day are comparable to those observed on 05 September 2017 (Francis, Chaboureau, et al., 2021).

Numerical Model
The WRF model version 4.2.1 is used in this study. Details of the model configuration are given in Table 1. As seen in Figure 5a, WRF is run in a three-nest configuration with an 18 km outermost grid comprising the entire Arabian Peninsula and surrounding region, a 6 km grid extending over the Arabian Gulf and neighboring regions, and a 2 km grid centered over the UAE. The simulation lasts from 22 to 24 August 2019, with the first 24 hr regarded as spin-up and the output discarded, and is forced with the 0.25° × 0.25° Global Forecast System (GFS) forecast data initialized on 22 August 2019 at 00 UTC. A 24-hr spin-up time has been used in similar studies such as in Francis, Chaboureau, et al. (2021). In the vertical 60 levels are considered, with the first model level at ∼27 m above the surface and a total of 33 levels between 1 and 10 km. This increased vertical resolution is needed to properly represent the convective clouds and associated updrafts (Gu et al., 2011). With this set up, the 2-day WRF run takes roughly 3 hr to complete with 104 Central Processing Units (CPUs) or two computational nodes at the Khalifa University High Performance Computing server. This means that the proposed model configuration can be easily extended to real-time operational purposes, also because the GFS forecast data are freely available in near real-time. In fact, the short latency time of the simulations allows for more than one run to be performed, with the resulting ensemble giving further confidence into the WRF predictions.
The physics schemes selected reflect the findings of Schwitalla et al. (2020) and Francis, Chaboureau, et al. (2021). Given the crucial role of the PBL scheme on the model-predicted convection and precipitation in the UAE, as highlighted by Taraphdar et al. (2021)    Asymmetric Convective Model 2 (ACM2; Pleim, 2007aPleim, , 2007b. PBL schemes differ in the way the turbulent fluxes, which arise when the Reynolds decomposition is applied to the primitive equations, are parameterized (Cohen et al., 2015(Cohen et al., , 2017. As a complete statistical description of turbulence requires an infinite number of equations, a "closure approximation" is needed. Some schemes retain the prognostic equations for the mean variables and approximate the second-order terms, "first-order closure schemes", while others solve the equations for the mean and variance and approximate the third-order terms, "second-order closure schemes". It is also possible for a PBL scheme to be first-order with respect to some variables (e.g., temperature and moisture), and second-order with respect to others (e.g., momentum variables), what is called a "one-and-a-half-order closure" scheme. For the PBL schemes considered here, the YSU and ACM2 schemes are first-order closure schemes while the MYNN is a one-and-a-half-order closure scheme. Besides differing in the highest order of the turbulent term considered, PBL schemes also represent the vertical gradient in the turbulent term in different ways. In "local" schemes, only vertical levels adjacent to a given model grid-point can affect the variables at that location. They assume that the size of the turbulent eddies is smaller than the vertical grid spacing used in the model, and therefore favor more

Interior nudging
Grid (or analysis) nudging (Stauffer & Seaman, 1990;Stauffer et al., 1991) employed in 18 and 6 km grids above ∼800 hPa on a time-scale of 1 hr toward GFS data. Fields nudged include the horizontal wind components, water vapor mixing ratio and potential temperature perturbation. This nudging configuration follows the findings of Fonseca, Francis, Weston, et al. (2021) Parameterization scheme Option selected

Land surface model (LSM)
Noah LSM with MultiParameterization options (Niu et al., 2011;Yang et al., 2011) configured as in Weston et al. (2018) Sea surface temperature (SST) 6-hourly GFS SSTs + simple skin temperature scheme (Zeng & Beljaars, 2005)   local-scale mixing. In "nonlocal" PBL schemes, multiple vertical levels are considered, as they are built on the premise that larger-scale eddies can transport the atmospheric fluid over some distance in the vertical before it is mixed by smaller-scale turbulent eddies. Some PBL schemes comprise features of "local" and "nonlocal" closure and are deemed as "hybrid" schemes. The YSU is a "nonlocal" PBL scheme, the MYNN is a "local" PBL scheme, and the ACM2 is a "hybrid" PBL scheme, the latter featuring local and nonlocal upward mixing and local downward mixing, with the non-local transport turned off for stable or neutral flows. Each PBL scheme has its own strengths and weaknesses: for example, "local" schemes are known to have problems with localized stability maxima, whereas "nonlocal" schemes can lead to an overly deep, warm and dry convective boundary layer due to excessive vertical mixing. A full discussion of the PBL schemes is provided in Cohen et al. (2015Cohen et al. ( , 2017. While a cumulus scheme is switched on in the 18 and 6 km grids, no cumulus scheme is employed in the 2 km nest. However, this nest's spatial resolution may still be too coarse to explicitly represent shallow clouds, and hence a shallow cumulus scheme should be used. Following Schwitalla et al. (2020), the Global and Regional Integrated Model System (GRIMs; Hong et al., 2013) shallow convection scheme is activated in the YSU and ACM2 simulations. In the MYNN simulation, and instead of switching on a shallow cumulus scheme, the mass-flux scheme embedded in the MYNN PBL scheme, which is designed to simulate the non-convective component of the subgrid clouds (Olson et al., 2019), is employed.
The sea surface skin temperature (SSKT) scheme described in Zeng and Beljaars (2005) is considered, which adds a diurnal variation to the SSKT which itself is controlled by the 6-hourly sea surface temperatures (SSTs) from the GFS input data. An updated soil texture and landuse land cover over the UAE described in Temimi et al. (2020), which has been found to improve the model's surface and near-surface predictions over the country, is employed in all grids. In addition, and following Wootten et al. (2016) and Fonseca, Francis, Weston,  2021), interior grid (analysis) nudging is employed in the two outermost nests, a set-up which is found to improve the model's innermost grid forecasts. Following the latter study, the potential temperature perturbation, horizontal wind components and water vapor mixing ratio are relaxed toward the forcing data above ∼800 hPa and on a time-scale of 1 hr.

Observational Datasets
A total of six observational datasets are employed in this study, with the main features summarized in Table 2. The observed cloud cover and properties are estimated from the measurements collected by the Spinning Enhanced Visible and Infrared Imager (SEVIRI) instrument, on board the Meteosat Second Generation spacecraft, and the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument, on board the National Aeronautic and Space Administration's Terra and Aqua satellites. Radar reflectivity snapshots are available every 6 min over the country, and 15-min weather measurements at the location of the 48 stations given in Figure 5b, are used to better understand the evolution of the convective event and its surface and near-surface signatures, as well as for model evaluation purposes. The model-predicted precipitable water is also assessed against that estimated from satellites as given by the Morphed Integrated Microwave Imagery at the Cooperative Institute for Meteorological Satellite Studies -Total Precipitable Water version 2 (MIMIC-TPW2) dataset. In addition, in-situ observations of vertical wind speed at the altitude where seeding was conducted are available for comparison with the WRF forecasts. These unique observations were obtained from the airplanes used to conduct the seeding on that day at the eight sites highlighted with a star in Figure 5c.

Verification Diagnostics
The set of verification diagnostics proposed by Koh et al. (2012) is used here to assess the WRF performance. They comprise the bias, normalised bias (μ), correlation (ρ), variance similarity (η), and normalised error variance The radiances are converted to reflectance following the documentation available on EUMETSAT's website: http://eumetrain.org/IntGuide/PowerPoints/ Channels/conversion.ppt Weather station data provided by UAE's National Center of Meteorology (NCM) 15-min air temperature, relative humidity, sea-level pressure, horizontal wind direction and speed and surface downward shortwave radiation flux measurements at standard heights. The data is available at the location of 46 weather stations given in Figure 5b Observations by the pilots who conducted cloud seeding operations on the day of the event provided by the NCM In situ measurement of the updraft speed at the seeding altitude at the eight locations given in Figure 5c Radar reflectivity maps over the UAE provided by the NCM 6-min radar reflectivity (dBZ)   (α), and are defined in the equations below. Here, gives the discrepancy between the WRF forecast and the observations , while and are, respectively, the mean and standard deviation of the variable .
The model bias, > , is the mean discrepancy between the WRF predictions and the observations. The normalized bias is given by the bias divided by the standard deviation of the discrepancy, , and is used to check whether the bias can be regarded as significant. As detailed in Koh et al. (2012), if | | < 0.5 , the contribution of the bias to the Root-Mean-Square-Error is less than roughly 10%, and hence the model bias can be considered as not significant. The correlation ( ) and the variance similarity ( ) give an indication of the phase and amplitude agreement between the observed and modeled data, respectively, with the referred two sources of error accounted for in the normalised error variance ( ) diagnostic. The optimal scores are = 1 , = 1 and = 0 , with a model forecast regarded as practically useful if 1 as in this case it is more skillful than a random forecast based on the climatological mean for which = 0 and hence = 1 . Some of the strengths of the , and diagnostics are that they are non-dimensional, symmetric with respect to the observations and forecasts, and applicable to both scalar and vector variables, making them ideal to be used in this work.

Performance of the WRF Model
The potential use of the WRF predictions for guidance into cloud seeding operations in the UAE. In the following subsections the model performance is evaluated against the ERA-5 reanalysis data in Section 4.1, NCM's weather station data in Section 4.2, satellite-derived and radar estimates in Section 4.3, and the in-situ observations collected during the seeding operations in Section 4.4. Figure 6 shows the large-scale fields from WRF's 18 km (outermost) and 6 km (intermediate) nest, and can be directly compared with the ERA-5 plots given in Figure 1. The model-predicted 500 hPa geopotential height and 200 hPa horizontal winds are in good agreement with those from ERA-5. As grid nudging toward the forcing dataset is employed in the two outermost nests above ∼800 hPa (Table 1), this indicates that the large-scales in the GFS forecast data are reliable at least in the first 48 hr of the run. At lower-levels, the AHL is slightly displaced southwards in the model. This bias is reported in , and can be explained by the colder surface environment: as seen in the 2-m temperature plots (Figures 1 and 6b), WRF has a cold bias at the surface which is more pronounced in the morning to mid-afternoon hours when it can exceed 6 K. This is a well-known feature of the WRF predictions in arid regions and has been reported in the Arabian Peninsula (e.g., Temimi et al., 2020) and in the Sahara Desert (e.g., Fekih & Mohamed, 2019). The cold bias has been attributed to deficiencies in the land surface model and forcing data, and an incorrect representation of surface/soil properties and concentration of greenhouse gases and dust Temimi et al., 2020). Despite the colder temperatures, the AHL is stronger in the model compared to the reanalysis dataset. This is because the cold bias in WRF is limited to lower altitudes: as seen in a cross-section at 52.5°E at 12 UTC, Figure S1 in Supporting Information S1, above about 925 hPa the model is actually warmer than ERA-5, with a more vigorous vertical circulation. As a result of a stronger AHL, the atmosphere is drier than in ERA-5 in eastern UAE, as the enhanced southeasterly winds FONSECA ET AL.

10.1029/2022EA002269
13 of 29  Figures 1 and 6b). The drier environment in the model is also seen when the WRF-predicted precipitable water is compared with that estimated from satellite data ( Figure S2 in Supporting Information S1). This inhibited the development of clouds in the afternoon in the region (Figure 2 at 10 and 12 UTC). The moist air from the Arabian Sea that flows around Oman in WRF weakens before reaching the eastern UAE coastline. Together with the offshore winds from the desert, this led to a much reduced cloud cover when compared to observations in the morning hours (Figure 2 at 05 UTC). On the other hand, on the western side of the UAE and neighboring Saudi Arabia and Qatar, the AHL winds reinforce the daytime sea-breeze circulation, resulting in a deeper inland moisture penetration compared to ERA-5.

Local-Scale: Weather Station Data
Besides a comparison with reanalysis data, the model's 15-min forecasts of the 2 km grid are assessed against the observed measurements at the location of 46 weather stations given in Figure 5b. Figure 7a shows the WRF and observed data at the location of the eight stations that comprise the bulk of the rainy region on this day. Similar results are obtained at the location of the other stations (not shown), with the biases having a slightly higher magnitude for the higher elevation sites in line with the findings of Branch et al. (2021) and Temimi et al. (2020). The verification diagnostics are presented in Table 3.
The cold bias in the model is mostly a daytime feature, with a peak value of about 8.5 K at 13 Local Time (LT), but with magnitudes generally within ±2 K between 16 LT and 07 LT. It is interesting to note that, while here WRF captures generally well the nighttime temperature but has a pronounced daytime cold bias, in  the model showed a large nighttime cold bias, exceeding 10 K at some stations, but reproduced well the observed daytime temperatures. The reason for this discrepancy is likely to be the use of a climatological aerosol distribution in the cloud microphysics scheme and representation of the aerosol direct and indirect effects in the model in the runs discussed in this paper (Table 2). In fact, the air temperature biases in Figure 7a resemble those obtained by Fonseca, Francis, Weston, et al. (2021) which employed a similar WRF configuration in their simulations. Aerosols are known to scatter and absorb the Sun's shortwave radiation and trap the longwave radiation emitted by the Earth's surface (Francis, Chaboureau, et al., 2021). As a result, they lead to cooler daytime and warmer nighttime temperatures when compared to clearer conditions, in particular in aerosol-rich regions such as the UAE which is part of the Arabian Desert (Bou Karam Francis et al., 2017;Nelli et al., 2021). The daytime cold bias is largely insensitive to the choice of the PBL scheme, at least for the three considered in this study, and may act to delay the onset of convection in the model. The clearer skies in WRF explain why the predicted temperatures in the evening hours are up to 2 K higher than those observed. As expected, the scores for the three runs, given in Table 3, have comparable magnitudes. Besides the bias, the correlation is also poor, varying between 0.44 and 0.56, which may be attributed to the slower increase in temperature during the day in WRF likely a result of the increased aerosol loading. However, 0.6 for all three simulations and hence the WRF forecasts can be considered skillful.
The observed wind speed is generally overestimated by the model, with biases up to 5 m s −1 in the early morning hours. As opposed to the air temperature, for this variable there are clear differences between the three simulations, with the WRF-YSU run largely giving the best scores (Table 3). In fact, the daily-mean wind speed bias in this simulation is roughly half of that obtained in the other two runs. This arises mostly from a superior representation of the observed nighttime winds, which are weaker in WRF-YSU because of a more stable nighttime PBL, as evidenced by the cooler nighttime temperatures. One reason for this is a shift in the wind direction for some of the stations, as highlighted by the larger error bars for this run which represent one standard deviation from the mean, and resulting cooler air advection. The stronger daytime winds may also contribute to the cold bias as they promote enhanced vertical mixing, as noted by Fonseca, Francis, Weston, et al. (2021). The overestimation of the wind speed in WRF may arise from the deeper AHL and associated enhanced circulation in the model (cf. Figure 6 with Figure 1), which also advects the drier desert air into the region. The predicted water vapor mixing ratio has a negative bias in all simulations with a daily-mean magnitude between 2.5 g kg −1 and 3.3 g kg −1 ( Table 3). The drier conditions in WRF are also seen in comparison with satellite-derived precipitable water estimates ( Figure S2 in Supporting Information S1). Deficiencies in the simulation of the wind's subgridscale variability and an incorrect representation of the surface drag are also likely to play a role in its overestimation by WRF (Nelli et al., 2020b).  In line with previous research (e.g., Fonseca et al., 2020;Wehbe et al., 2019), WRF underpredicts the observed cloud cover, with the modeled downward shortwave radiation flux exceeding that observed by more than 400 W m −2 and with a daily-mean bias of ∼60 W m −2 ( Table 3). The reduced cloud cover in the model is also consistent with the drier environment. The observed sea-level pressure time-series shows a minimum at about 01-02 and 15-16 LT and a maximum at 09-10 and 21-22 LT which are due to the semi-diurnal atmospheric tide and its interaction with the slope flows from the nearby Al Hajar mountains (Nelli et al., 2020a). In WRF, the phase of the diurnal variability is well captured even though the extrema occur roughly 1-2 hr later than in observations. A comparable phase lag has been reported by other authors (e.g, Santoro et al., 2011), and may be explained by an incorrect simulation of the local-scale circulations. The mismatch in magnitude is attributed to an incorrect representation of the topography by the model, as some of the stations are over high terrain. This is the    (Francis, Chaboureau, et al., 2021), as the Al Hajar mountains are highly heterogeneous.
The observed sea-level pressure data shows a jump of about 0.8 hPa in 30 min at 16 LT, at the time when the wind suddenly changes direction from easterly to westerly and its speed increases by up to 2.6 m s −1 , with the air temperature dropping from 312 K at 15:30 LT to 309.5 K at 16:30 LT. These are the signature of the passage of a MCS (Francis, Chaboureau, et al., 2021), which is seen in the satellite (Figure 2a) and radar reflectivity (Figure 8a) images. In particular, in the YSU simulation, WRF captures the MCS with an overall sea-level pressure increase of ∼0.4 hPa, a temperature drop of ∼2 K, and a wind speed rise of about 3.2 m s −1 . However, the MCS in the model has a reduced spatial extent (Figure 8), consistent with the drier and cooler conditions. An analysis of Figure 7a and Table 3 reveals that, for the evaluation against surface measurements, the WRF-YSU run generally outperforms the other two simulations. In fact, it is the only run where the model predicts precipitation at the location of the eight stations considered.

Cloud Properties: Satellite-Derived and Radar Observations
In addition to the evaluation against surface meteorological observations, it is also of interest to assess how the cloud properties are simulated by the model. In particular, a correct representation of the clouds and their characteristics such as top pressure/temperature is crucial if the WRF forecasts are to be used for guidance into cloud seeding activities. Figures 3 and 4 show the cloud fraction and cloud top pressure/temperature as estimated from two MODIS' overpasses at ∼10 UTC and ∼22 UTC, respectively, and the corresponding WRF plots. At a given grid-box, the WRF cloud fraction is taken as the maximum cloud fraction in the column, while the cloud top pressure and temperature are the pressure and temperature of the highest model layer where clouds are present, respectively.
The overpass at ∼10 UTC (14 LT) captures some of the daytime convective clouds. The cloud top pressure and temperature of the clouds that developed over the Zagros mountains in Iran between 56° and 59°E are reasonably well simulated by the model, extending up to ∼500 hPa and having a top temperature of about 260 K. However, the contrast between the shallower/warmer northern half and the deeper/cooler southern half is not fully captured by WRF. A possible reason is the drier environment seen in comparison with both ERA-5 (cf. Figures 1 and 6) and satellite ( Figure S2 in Supporting Information S1) data. Overall these clouds are better represented in the WRF-YSU and WRF-ACM2 simulations, and are too widespread in the WRF-MYNN run. In fact, the MYNN PBL scheme has a tendency to overpredict shallow clouds at all times of the day in particular over the water bodies, as also seen in comparison with the SEVIRI images in Figure 2. This is consistent with the fact that the local mixing in the MYNN scheme favors lower Lifting Condensation Level (and hence cloud base) heights, as noted for example, by Milovac et al. (2016). Also likely due to the drier conditions, the model underpredicts the observed cloud cover over eastern UAE, with the WRF clouds being shallower (top around 400 hPa vs. 200 hPa) and warmer (top temperature of 250 vs. 220 K) as well as smaller in size when compared to the MODIS' estimates. The reduced convective cloud cover in the WRF-ACM2 simulation may arise from too deep PBLs in this run, a drawback of the scheme noted by Cohen et al. (2015). This is further supported by the drier conditions in the region seen in comparison with surface (Table 3) and satellite ( Figure S2 in Supporting Information S1) data. Both the SEVIRI images in Figure 2 and the MODIS snapshot in Figure 3 reveal that the clouds in WRF are shifted upstream, closer to the Al Hajar mountains, compared to observations. This has also been noted by Schwitalla et al. (2020) who attributed it to an incorrect simulation of the location of the low-level convergence as a result of the still coarse spatial resolution.
The mid-level clouds over parts of Saudi Arabia and Oman, remnants of the afternoon convective clouds, in the 22 UTC (02 LT) overpass, Figure 4, are largely missed by WRF. The model only predicts scattered clouds deeper into the Rub' Al Khali desert, even though their cloud top pressure and temperature are generally comparable to those estimated from satellite observations. This is not surprising as the daytime convection is weaker and less extensive in the model, as seen in Figures 2 and 3

10.1029/2022EA002269
19 of 29 the PBL scheme for the simulation of an extratropical cyclone in the North Atlantic, in particular with respect to the low-level clouds. Figure 8 shows two-dimensional radar reflectivity maps from 10 UTC (14 LT) to 13 UTC (17 LT) during the bulk of the convective event. Consistent with the satellite images, Figure 2, convective clouds developed along the Al Hajar mountains in the early to mid-afternoon hours, with an intense convective cell over northeastern UAE just to the east of Dubai around 12 UTC. WRF partially simulates this MCS but, likely because of the drier environment ( Figure S2 in Supporting Information S1), it has a smaller spatial extent. Nevertheless, in both the model forecasts and observations the maximum reflectivity values are ∼60 dBZ, an indication of heavy precipitation (Liu & Fu, 2001). A comparison of the three WRF configurations reveals that the WRF-YSU simulation gives the best agreement with the observations, with a better developed MCS at 12 UTC that persists well into 13 UTC. In the WRF-ACM2 run, and likely because of the deeper PBLs and resulting drier conditions (Table 3 and Figure  S2 in Supporting Information S1), convection is almost exclusively limited to Oman.
In order to assess how the vertical extent of the observed clouds is captured by the model, Figure 9 shows histograms of radar reflectivity at three sets of heights (1-5 km, 5-10 km and 10-15 km) for 11-13 UTC and for the domain 54.5-57°E and 24-26°N which comprises where the majority of the convection took place. For a given reflectivity bin, the number of counts were divided by the total number of grid-boxes so as to allow for a direct comparison with the observed data. The reduced spatial extent of the convective cells in WRF seen in the radar reflectivity maps in Figure 8 is also present in Figure 9, as is the better performance of the YSU scheme. The observed radar reflectivity values are almost exclusively below 60 dBZ, with higher occurrences in the mid-troposphere (5-10 km). The lower reflectivities above 10 km are expected as the presence of the stratosphere and associated higher static stability acts as a lid on cloud development, and as the radar reflectivity of an ice particle is smaller than that of a liquid drop (Huo et al., 2021). On the other hand, the higher cloud bases, which are expected for arid regions given that the drier environment promotes the evaporation of raindrops (Liu & Zipser, 2013), likely explain why the reflectivities are lower in the 1-5 km range compared to 5-10 km. The trend in radar reflectivity seen in the observations is present in the WRF-MYNN and WRF-YSU runs, while in the WRF-ACM2 simulation the reflectivity decreases with height with almost no clouds above 10 km. The results in Figure 9 suggest that the updrafts in the model are likely weaker than those observed owing to the smaller vertical extent of the modeled convective clouds, which is also seen in vertical cross-sections of radar reflectivity (not shown). As noted before, this likely arises from the cold surface temperatures (Figure 7a and Table 3) and drier environment ( Figure S2 in Supporting Information S1 and Table 3) in the model.

Cloud-Scale Updrafts: In-Situ Airborne Observations
A more direct evaluation of the model's skill can be achieved by comparing the WRF forecasts with the observations made during the seeding operations. A total of eight seeding flights took place on this day, most just east of Dubai (Figure 5c), with the observed vertical wind speeds at the seeding altitude given in Table 4. At sites #5-6, and at the seeding altitude, moderate to strong updrafts were reported with speeds of about 5 m s −1 , whereas at sites #1, 4 and 8 the speeds were roughly half as strong. Qualitatively, WRF predicts an updraft at the seeding altitude at three to five of the eight sites, yielding a skill score of ∼40-60%. It is interesting to note that in the configuration that gives the highest score, WRF-ACM2, the model does not predict clouds at any of the seeding sites, while the WRF-YSU run features the largest number of non-zero radar reflectivity values (two) but gives the lowest score with respect to the updrafts. This suggests that the WRF predictions of vertical velocity may be more useful for guidance into cloud seeding operations than those of radar reflectivity. The updraft speeds in the model are about one to two orders of magnitude smaller than those observed, which can be attributed to a mismatch in the location of the convective cells. In fact, the updraft speeds in the simulated cells shown in Figure 8 are of 5-6 m s −1 and therefore comparable to those observed. This stresses the need to put more effort into getting the dynamics right so as to better capture the location and timing of the observed cells. The observed radar reflectivity is typically in the range 40-50 dBZ at the seeding elevation, with generally higher reflectivity values when the updrafts are more intense. In fact, the highest radar reflectivity of about 54 dBZ co-occurs with the highest measured updraft speed of 4.7 m s −1 , while the lowest reflectivity of 39.5 dBZ co-occurs with the lowest updraft speed of 2.9 m s −1 . This is in line with the findings of Heymsfield et al. (2010), and suggests the potential use of near real-time observations of radar reflectivity for guidance into seeding operations by targeting areas associated with maximum reflectivity values for seeding.
FONSECA ET AL.
10.1029/2022EA002269 20 of 29 The WRF vertical velocity and radar reflectivity profiles at each of the sites are given in Figure 10. For the WRF-MYNN simulation, and in particular for sites #1, 3, 5-7, the vertical velocity profiles generally exhibit ascent at lower levels and descent above, resembling those observed in shallow convective events (e.g., Haertel & Kiladis, 2004;Uma & Rao, 2009). This is consistent with the fact that cloud seeding is normally performed when clouds are still developing. The profiles for events #2 and 4 show rather small vertical velocities from the surface up to 8 km, whereas that for event #8 is indicative of stratiform clouds with ascent around 4-5 km and descent below 4 km likely due to evaporative cooling. A mixture of shallow cumulus and stratiform clouds are also present in the WRF-YSU run, with rainfall falling at sites #2 and 5 for which descending motions with speeds up to 15 m s −1 are forecasted by the model. As opposed to the WRF-MYNN and WRF-YSU simulations, and in line with the cloud cover maps in Figure 2 and the radar snapshots in Figure 8, the vertical wind speed at the seeding sites in the WRF-ACM2 run is rather small, generally within ±5 m s −1 , in line with the absence of convection in the region.
In order to gain further insight into the model clouds, Figures 11a-11d show two-dimensional maps of the vertical wind speed for the WRF-YSU run at the closest model level to the seeding height and seeding time for events #2, 4, 5 and 7. Two main features can be seen: (a) convective rolls that extend inland as the sea-breeze penetrates deeper into the country in the mid-afternoon hours, and (b) convective updrafts and downdrafts and cold pools resulting from rainfall reaching the surface and spreading over time. In order to check that the vertical velocity structures in (a) are associated with convective rolls, Figure 11e gives a vertical cross-section at Sweihan, a station highlighted with a cross in Figures 11a-11d. A comparison with the convective roll signature in the plotted fields summarized in Dailey and Fovell (1999) and Chen et al. (2015) confirm the occurrence of these clouds, with alternating patterns of ascending/descending motion (accompanied by the associated horizontal wind convergence/divergence at lower-levels and the opposite at upper-levels) and with the clouds oriented parallel to the prevailing wind direction. Convective rolls form in the presence of strong surface heating and low-level wind shear induced by the sea-breeze, and as noted by Schwitalla et al. (2020), they play a role in convection initiation over land. In particular, the sea-breeze advects the moist marine-air inland, and the convective rolls act to transport it vertically making the atmosphere more suitable for the occurrence of convection. An inspection of the near-surface measurements at Sweihan and Saih Al Salem, Figure 7b, reveals that these features are also present in the observations, as seen by the shifts in horizontal wind direction/speed and the variations in the downward shortwave radiation flux (Saih Al Salem is closer to the coast than Sweihan and hence the rolls reach this site first). The cross-section at Al Dhaid, Figure 11f, shows the occurrence of deep convection followed by stratiform precipitation, with the model predicting roughly 6 mm of rainfall at this location. Further analysis of the WRF output reveals that the collision of the spreading cold pools with the sea-breeze leads to development of new clouds (e.g., Bou Karam et al., 2008), as seen in the cross-sections at 56°E presented in Figure S3 in Supporting Information S1. The interaction of these two processes is a common cloud formation mechanism, as highlighted by Grant and van den Heever (2014).
A better way to assess the suitability of the model predictions for guidance into cloud seeding operations is to plot the regions which, according to WRF, provide suitable conditions for seeding. This is done in Figure 12, where a given grid-point is deemed seedable according to the model if, as noted in the Introduction, an updraft is present (in the 900-500 hPa layer) and WRF forecasts clouds in the column. A comparison with the observed radar reflectivity reveals the superior performance of WRF-YSU, with WRF-MYNN predicting extensive favorable seeding regions due to the overprediction of the observed cloud cover, as noted in Figures 2 and 3, while WRF-ACM2 shows reduced potential seeding areas due to the weaker convective activity, Figures 8 and 9.

Discussion and Conclusions
In this work, the usefulness of the Weather Research and Forecasting (WRF) forecasts to support cloud seeding operations in the United Arab Emirates (UAE) is assessed for a typical summertime convective event in the eastern UAE. As boundary layer dynamics play a crucial role in convection initiation, three PBL schemes are considered: the "local" Mellor-Yamada Nakanishi Niino (MYNN), the "nonlocal" Yonsei University (YSU), and the "hybrid" Asymmetric Convective Model 2 (ACM2   . Further information regarding the radar reflectivity and seeding data is given in Table 2. FONSECA ET AL.

10.1029/2022EA002269
23 of 29 (GFS) data, which is freely available online in near real-time, and with the proposed configuration the model forecasts can be made available in time to help guide seeding activities.
The convection around the Al Hajar mountains on 23 August 2019 was triggered by the low-level convergence of the Arabian Heat Low (AHL), topography-driven and daytime sea-breeze circulations along the Intertropical Discontinuity (ITD) region, and was further supported by the presence of a mid-level trough. The development of convective rolls in the boundary layer, due to the extreme heat during the day, was reported in the observations and in the model simulations. The interaction between the convective rolls and the cold pools from deep convection promoted the development of new convective clouds.
An evaluation of the model predictions against the in-situ observations in the precipitating region indicates that the WRF-YSU run generally gives the best scores, and generates the most realistic simulation of the observed convection. A comparison of the modeled cloud cover and properties with those estimated from satellite observations revealed that the MYNN scheme has a tendency to overpredict shallow clouds in particular over water bodies, with all configurations failing to simulate the deep convective clouds that developed over eastern UAE in the afternoon. While the WRF-MYNN simulation forecasts excessive clouds, in the WRF-ACM2 run, and likely as a result of deeper (and hence drier) boundary layers, the model predicts a clearer atmosphere when compared to the other two simulations.
When the radar reflectivity forecasted by WRF is evaluated against that observed, the WRF-YSU configuration again performs the best, both in terms of the horizontal and vertical distribution. The reduced spatial extent of the convection in WRF is also seen in the radar reflectivity. In both WRF and observations the reflectivity values are higher in the 5-10\km depth compared to the 10-15 km, where ice prevails and the presence of the stratosphere limits cloud development, and the 1-5 km, likely due to the higher cloud bases common in arid regions.
A more direct assessment of the suitability of the WRF forecasts as guidance for seeding operations is achieved by comparing the model forecasts with the measurements taken by the pilots who performed the seeding. The sign of the vertical velocity predicted by the model at the location of the eight seeding sites agrees with that observed roughly 40%-60% of the time, and after taking into account the mismatch in the location of the convective cells, the observed and modeled speeds are also comparable. The correlation between the radar reflectivity (here the 40-50 dBZ surface) and vertical velocity observed by Heymsfield et al. (2010) is also present for this event, meaning that it can be used as a proxy for updraft speeds and, if available in near real-time, to support the seeding activities. Further analysis of the WRF output indicated the occurrence of convective rolls, whose signature is also seen in the in-situ data, that propagate inland and help to precondition the environment for the later occurrence of convection.
The most important conditions for a successful seeding operation are the presence of updrafts and cloud cover. The combined model predictions of these two variables agree best with the observed convective regions in the WRF-YSU simulation, being too widespread in WRF-MYNN due to the excessive cloud cover, and more spatially restricted in WRF-ACM2 that underpredicts the observed convective activity.
We conclude that the WRF model with proposed experimental set up can be used to guide cloud seeding operations in the country in near-real-time by predicting potential rainy regions based on the identification of convective updrafts and suitable cloud properties. An extension of this work would be to consider additional case studies, so as to assess the robustness of the findings reported here, as well as to employ data assimilation in an attempt to correct the surface cold bias in WRF.

Conflict of Interest
The authors declare no conflicts of interest relevant to this study.