High-resolution simulations of convective cold pools over the northwestern Sahara

Transcription

1 Click Here for Full Article JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 114,, doi: /2008jd011271, 2009 High-resolution simulations of convective cold pools over the northwestern Sahara P. Knippertz, 1 J. Trentmann, 1,2 and A. Seifert 2 Received 9 October 2008; revised 22 January 2009; accepted 11 February 2009; published 24 April [1] Cooling by evaporation of convective precipitation in the deep and dry subcloud layer over desert regions can generate intense downdrafts and long-lived and extensive atmospheric density currents. The strong gusts at their leading edges can cause so-called haboob dust storms. Despite their importance for the dust cycle, the ability of state-of-theart numerical weather prediction models to realistically simulate the associated convective cold pools has been investigated very little to date. During the first field campaign of the Saharan Mineral Dust Experiment in southern Morocco in May/June 2006, several density currents were observed. They were triggered by deep moist convection over the Atlas Mountains during the afternoon and propagated into the foothills in the course of the evening. Here we present numerical simulations of three of these density currents using the nonhydrostatic Consortium for Small-Scale Modelling model with 2.8-km horizontal grid spacing, which allows an explicit treatment of deep convection. The model is capable of simulating the timely initiation of convective cells over the Atlas Mountains and the subsequent formation of long-lived, extensive cold pools with a realistic threedimensional structure. Deviations from available surface and satellite observations are closely related to model deficiencies in simulating precipitating convection over the Algerian Sahara. Sensitivity studies with modified microphysics reveal a large influence of raindrop size distributions on evaporation and surface rainfall but a rather moderate influence on the cold pool evolution. Decreasing the length scale for turbulent vertical mixing in the boundary layer leads to more widespread but weaker precipitation, more evaporation, and a faster and more extended cold pool. Citation: Knippertz, P., J. Trentmann, and A. Seifert (2009), High-resolution simulations of convective cold pools over the northwestern Sahara, J. Geophys. Res., 114,, doi: /2008jd Introduction [2] The evaporation and melting of convectively generated hydrometeors sedimenting into the subsaturated subcloud layer can lead to deep, intense downdrafts and the formation of large cold pools at the surface that spread laterally as density currents [Simpson, 1997]. This process can be enhanced through the evaporation of cloud water due to lateral entrainment of subsaturated ambient air [Knupp and Cotton, 1985]. Conditions favorable to produce strong downdrafts penetrating to the surface include a very deep, dry-adiabatic mixed layer, high rain water mixing ratios at cloud base, and small raindrop sizes [Kamburova and Ludlam, 1966; Srivastava, 1985; Proctor, 1988, 1989; Takemi and Satomura, 2000]. The former condition is of major importance, since more stable environmental lapse rates allow descending parcels to attain neutral buoyancy through compressional heating. Downdraft intensity can be enhanced through precipitation load, in particular in wet 1 Institute for Atmospheric Physics, Johannes Gutenberg University Mainz, Mainz, Germany. 2 German Weather Service, Offenbach, Germany. Copyright 2009 by the American Geophysical Union /09/2008JD011271$09.00 microbursts often related to short-lived, small-scale, heavily precipitating thunderstorms in weakly sheared environments [Fujita, 1985; Goodman et al., 1988], and then depends upon the width, type, and duration of precipitation [Proctor, 1989]. Typical vertical velocities in convective downdrafts are 5 10 m s 1, but up to 20 m s 1 have been observed [Knupp and Cotton, 1985]. [3] The high momentum and density of the cold-pool air causes a strong, divergent wind field pattern at the ground, usually with a marked convergent gust front. In the case of dry-adiabatic deep subcloud layers even shallow high-based clouds producing small quantities of precipitation can generate strong downdrafts and surface gusts [Knupp and Cotton, 1985]. The evolution and depth of the cold pool is sensitive to the ambient vertical wind shear at lower levels [Weisman and Rotunno, 2004] and other factors such as convective momentum transports. For a given density difference, a balance between the cold pool and the shear circulations results in deep lifting at the leading edge of the density current [Rotunno et al., 1988; Xu et al., 1996; Xue, 2000], which can trigger the formation of shallow arc clouds, or, if the atmosphere is conditionally unstable, of new convective cells [Knupp and Cotton, 1985]. This way cold pools can become a major ingredient in the organiza- 1of16

2 tion of moist convection from the meso-g to the meso-b scale and even meso-a scale [Takemi, 1999; Weisman and Rotunno, 2004]. [4] Over deserts, the deep, well-mixed, hot, and dry boundary layer allows for a substantial amount of latent cooling and the formation of extraordinarily long-lived and extensive cold pools [Takemi, 1999; Knippertz et al., 2007; Miller et al., 2008]. The associated density currents propagate at speeds on the order of 10 m s 1 over hundreds of kilometers, sometimes far away from the parent convective clouds [Freeman, 1952; Miller et al., 2008]. The passage of the leading edge is usually associated with abrupt changes in temperature, dew point, pressure, visibility, and with strong, gusty winds and high turbulence. The latter can lift fine material from the ground leading to fast-moving, dramatic, billowing walls of dust that for this reason attract more attention than similar systems that move over vegetated surfaces. These so-called haboobs have been documented for the Sahel [Sutton, 1925; Farquharson, 1937; Freeman, 1952; Lawson, 1971; Williams et al., 2008], the northwestern Sahara [Knippertz et al., 2007], semiarid parts of the USA [Idso et al., 1972; Chen and Fryrear, 2002], northwest China [Takemi, 1999], and the Arabian Peninsula [Membery, 1985; Miller et al., 2008]. Typical temperature drops are on the order of 7 K, while relative humidity typically increases by 15% [Idso et al., 1972; Miller et al., 2008]. A typical depth of the dust layer found in the literature is 1 2 km [Sutton, 1925; Freeman, 1952], but vertically pointing Doppler radar measurements by Williams et al. [2008] indicate substantial dust loadings well above 2 km also. Most of the suspended dust is contained within the cold pool [Lawson, 1971] where near-surface wind gusts of 15 m s 1 and even 25 m s 1 are commonly observed [Sutton, 1925; Freeman, 1952; Miller et al., 2008]. Idealized modeling work by Takemi [2005] has shown that turbulent mixing can entrain dust into the system-relative front-to-rear flow at midlevels and even into the cloud updraft [Takemi, 2005, Figure 7]. How much dust remains in the atmosphere after the decay of a convective cold pool is still a matter of debate [Williams, 2008]. Naturally, cold pools that quickly separate from the parent convection can more efficiently mobilize dust as they avoid scavenging by precipitation. In this sense the optimal condition for the generation of a haboob is enough moisture to produce rain, but not quite enough to allow it to reach the ground before it is evaporated. This leads to a distinct annual cycle of haboob occurrence that is closely, but not exactly, tied to the climatology of convective activity [Farquharson, 1937; Freeman, 1952; Miller et al., 2008]. The amounts of dust being mobilized by a system with a given intensity also depend on annual changes in vegetation and soil moisture. [5] Despite their importance for the dust cycle and their potential threat to aviation safety, most research on haboobs to date has been restricted to observational analyses of single cases or multiyear statistics at single observation sites, while numerical modeling studies are rare. Miller et al. [2008] developed a simplified haboob deflation model for dust budget analysis, while Takemi [2005] used a threedimensional cloud resolving model with a dust scheme in an idealized setup. Reinfried et al. [2009] employed the COSMO (Consortium for Small-Scale Modelling) model, the operational nonhydrostatic, limited area model of the German Weather Service (Deutscher Wetterdienst, DWD; see section 2.1), with an additional dust module to simulate a density current that occurred during the first field campaign of the Saharan Mineral Dust Experiment (SAMUM) (see Heintzenberg [2009] for an overview) in southern Morocco in May/June Such systems are usually triggered by deep moist convection over the High Atlas Mountains during the afternoon and then spread into the southern foothills in the course of the evening and night [Knippertz et al., 2007]. During SAMUM their occurrence was closely tied to the presence of upper tropospheric troughs, which help to destabilize the atmosphere and to advect hydrometeors toward the down-shear Saharan side of the mountain range. The observed systems had lifetimes of up to 10 h and leading edge extensions of several hundred kilometers as identified from infrared (IR) satellite imagery. Reinfried et al. [2009] were able to satisfactorily reproduce the most important characteristics of the cold pool when using a horizontal grid spacing of 2.8 km that allows an explicit treatment of deep convection. The existence of a realistic cold pool in the simulations is closely linked to the quality of the precipitation forecast, which in turn is highly sensitive to the treatment of convection. [6] Here, the work of Reinfried et al. [2009] is extended by conducting convection permitting COSMO simulations (without dust module) of three SAMUM density currents. The goal of this paper is (1) to further corroborate the ability of the model to generate realistic cold pools with respect to their temporal evolution and three-dimensional structure, and (2) to test the sensitivity of the cold pool evolution to changes in the microphysics and turbulence schemes of the model. Previous idealized studies suggest an influence of raindrop size on downdraft intensity and low-level cooling [e.g., Srivastava, 1985], but to our knowledge this has never been tested with a state-of-the-art numerical model in a realistic setup. Changes to the treatment of boundary layer turbulence can be expected to affect both the precipitation generation and the density current propagation, for example, through the modified entrainment of ambient air [Proctor, 1988]. In our view, convective cold pools in the northwestern Sahara are an ideal test bed for such a study as the causative precipitation is strongly related to forcings by upper level troughs and the Atlas Mountains, and is therefore more realistically represented in numerical models than convection resulting from air mass instabilities alone. In addition the deep dry boundary layer in the Sahara favors the formation of large, long-lived cold pools, whose macroscopic characteristics are well determined. The remainder of the paper is structured as follows: In section 2 the COSMO model and the observational data used for model evaluation are briefly introduced. Section 3 discusses reference simulations for three cases including a comparison to observations. The results from four sensitivity experiments are presented in section 4. Section 5 contains a summary and conclusions. 2. Model and Data 2.1. COSMO Model [7] All simulations for this study were conducted with the COSMO model version 4.0 [Steppeler et al., 2003; Schättler et al., 2008], the nonhydrostatic limited-area 2of16

3 Figure 1. Model domain with orography. The 1000-m isohypse is highlighted. Meteorological stations and geographical terms used in the text are indicated. weather prediction model used for operational weather forecasts in Germany, Switzerland, Italy, Greece, Romania, Poland and Russia. The results shown in this paper are from simulations with 2.8-km horizontal grid spacing nested into COSMO runs with 7-km horizontal grid spacing, which are in turn nested into operational analyses of the global model GME (this model has been named GME as it replaced the operational global model (GM) and the regional model (EM) for central Europe in 1999) of the DWD [Majewski et al., 2002]. The configuration of the COSMO model used here applies an efficient split-explicit Runge-Kutta solver [Wicker and Skamarock, 2002], a Lin-type one-moment cloud microphysics scheme that predicts cloud water, rain water, cloud ice, snow, and graupel [Lin et al., 1983; Reinhardt and Seifert, 2006], and a boundary layer scheme using a prognostic turbulent kinetic energy (TKE) equation based on a Mellor-Yamada level 2.5 turbulence closure [Mellor and Yamada, 1974; Raschendorfer, 2001]. In the high-resolution simulation with 2.8-km grid spacing deep moist convection is explicitly treated, but shallow convection with cloud depth below 300 hpa is parameterized using a simple mass-flux formulation based on the Tiedtke [1983] scheme [Doms and Förstner, 2004]. The simulations are initialized at 0000 UTC on 3 days during the SAMUM field campaign, on which density currents were observed (i.e., 07 and 31 May, and 03 June 2006), and then run for 30 h. If not noted otherwise, the employed time step was 30 s. The model has 50 unevenly spaced vertical levels up to a height of 22 km and the horizontal domain comprises grid points spanning the area from the Moroccan Atlantic coast to the Algerian Sahara Atlas in the west east and from the southern tip of Spain to northern Mauritania in the north south direction (Figure 1). Close to the center of the domain are the High Atlas Mountains with a maximum elevation of 3509 m in the model (4167 m in reality). To evaluate the COSMO-simulated clouds with satellite observations, model-derived pseudo satellite images are used [Keil et al., 2006] Observational Data [8] For the evaluation of model-simulated near-surface temperature, dew point, pressure, wind, and precipitation, observations from automatic weather stations (Campbell Scientific Inc., Logan, Utah) operated by the German research initiative IMPETUS (An Integrated Approach to the Efficient Management of Scarce Water Resources in West Africa) [Speth and Diekkrüger, 2006] are used. We will concentrate on the station El Miyit (EMY hereafter) that is located in the Drâa Valley in southern Morocco (30.36 N, 5.63 W, 792 m; see Figure 1). More details on station locations and instrumentation are given by Knippertz et al. [2007]. In addition, standard 3-h SYNOP reports from the Moroccan stations Ouarzazate (WMO 60265, N, 6.90 W, 1140 m) and Errachidia (60210, N, 4.40 W, 1042 m), and from the Algerian stations Bechar (60571, N, 2.23 W, 816 m), Beni Abbes (60602, N, 2.17 W, 505 m), Timimoun (60607, N, 0.28 E, 317 m), and Adrar (60620, N, 0.28 E, 283 m), also indicated in Figure 1, are used. Model-simulated brightness temperatures in the IR channel (10.8 mm) are compared to corresponding values from Meteosat Second Generation. For the evaluation of modelsimulated precipitation, the Tropical Rainfall Measuring Mission (TRMM) and Other Rainfall Estimate (3B42 V6) in horizontal resolution as described by Huffman et al. [2007] are used. These data are 3-h accumulated, combined TRMM precipitation radar-microwave-ir estimates (with gauge adjustment) and were downloaded from operated by the National Aeronautics and Space Administration (NASA). 3. Reference Simulations 3.1. Cold Pool on June 2006 [9] The first case to be studied is the rather extended density current that formed on 03 June 2006 [Knippertz et al., 2007, section 5.2], which was also investigated by Reinfried et al. [2009] Temporal Evolution [10] Figure 2 shows the evolution of the cold pool in the course of the afternoon and evening of 03 June as simulated by the COSMO model with 2.8 km horizontal grid spacing. At 1600 UTC the model simulates four precipitating convective cells slightly to the southeast (i.e., downwind with respect to the westerly to northwesterly flow at upper levels) of the highest peaks of the Atlas Mountains (Figure 2a). Surface precipitation accumulated over the previous hour clearly exceeds 5 mm at some grid points. The 10-m wind vectors in Figure 2a indicate a convergence between the mainly southwesterly flow from the Sahara and the northwesterlies from the Atlantic Ocean over the mountain crest that has helped to trigger the convection. Each of the four storms is associated with wind speeds exceeding 10 m s 1, usually at the Saharan, i.e., the down-shear side of the strongest precipitation (red lines in Figure 2a). This value is above the mobilization threshold for typical dust sources in the region [e.g., Chomette et al., 1999]. The grid point 3of16

4 Figure 2. Temporal evolution of the cold pool on June 2006 as simulated by the COSMO model. Shown is the precipitation accumulated over the previous hour in millimeters (shading according to scale), 10-m wind vectors together with 10 m s 1 isotachs (red lines) at (a) 1600 UTC 03 June, (b) 1900 UTC 03 June, (c) 2200 UTC 03 June, and (d) 0000 UTC 04 June. The black dots mark grid points where the vertical velocity exceeds 3 m s 1 below 700 hpa. The thin black lines are political borders, and the thick black lines mark the 1000-m isohypse of the COSMO model orography. The green line with the two crosses in Figure 2c indicates the location of the cross section and the vertical profiles, respectively, shown in Figure 3. maximum at this time is 16.5 m s 1. These strong winds are associated with convective downdrafts of up to 4.9 m s 1 in agreement with typical values in the literature [Knupp and Cotton, 1985]. The black dots in Figure 2a mark grid points with ascent of more than 3 m s 1 below 700 hpa. Over the cloudless Sahara, where daytime heating is very intense, they indicate the turbulent upward motions related to deep dry boundary layer convection. At the southern and eastern edges of the convective cells over the Atlas they indicate that the cold pools begin to spread laterally and lift ambient air. This criterion will be used in the following to identify the dynamically active part of the leading edges of density currents (see discussion at the end of this subsection). At 1600 UTC the leading edges are mostly oriented perpendicular to the low-level flow as has been observed for many other cases [e.g., Weisman and Rotunno, 2004]. Maximum uplift is 10.2 m s 1 at this time, which is in the upper part of the range of vertical motions observed by Williams et al. [2008] for the Sahel. [11] Three hours later, at 1900 UTC, the precipitation zone has shifted further downwind and also down-slope to the Moroccan-Algerian border (Figure 2b). The decreasing rainfall amounts at the ground might be an indication of increasing evaporation. The precipitation regions and cold pools of the three western cells have merged, while the larger and more intense cell to the northeast is still somewhat isolated. The strong low-level ascent at the leading edges to the south, which still reaches 7.6 m s 1 at this time, has supported the regeneration of convection. Winds behind the leading edges largely exceed 10 m s 1 with a maximum 4of16

5 of 17.5 m s 1 to the south of the eastern cell. Owing to the decreased insolation shortly before sunset the boundary layer convection over the Sahara has ceased as indicated by the absence of black dots in Figure 2b. Another 3 h later, by 2200 UTC, the precipitation zone has continued shifting southeastward into Algeria (Figure 2c). The initial cells have merged, but there are still indications of convective regeneration farther to the west with uplift reaching maximum values of 6.6 m s 1 at this time. Precipitation intensity has generally decreased. The leading edge of the cold pool is now clearly separated from the parent precipitation and forms an almost straight line that stretches some 500 km from 29 N, 6 W to 30 N, 1 W. Right behind the leading edge is an extended area with 10-m winds exceeding 10 m s 1, indicating potential for substantial dust mobilization. A maximum of 18.2 m s 1 is reached at this time. The wind vectors indicate a continuation of the cold-pool spreading to the northeast and west of the leading edge identified with the vertical motion criterion. Finally at 0000 UTC 04 June the precipitation has almost ceased and the cold pool begins to weaken (Figure 2d). While the active leading edge with strong horizontal and vertical velocities stretches for about 600 km across the Sahara, the area with indications of a density current in the wind field reaches much farther west illustrating the impressive distances a cold pool can spread in this desert environment. Maximum 10-m wind speeds have decreased to 15.6 m s 1, the maximum ascent at the leading edge is now 5.4 m s 1. [12] Between 1900 UTC 03 June and 0000 UTC 04 June (Figures 2b and 2d) the leading edge has traveled between 220 km in its western part and 270 km farther to the east, where the source of evaporative cooling is. This corresponds to propagation velocities between 12 and 15 m s 1. The fact that the maximum low-level winds behind the leading edge are considerable faster than the leading edge movement itself is consistent with observations of haboobs in other places [Lawson, 1971] and more theoretical arguments on density current behavior (see discussion by Smith and Reeder [1988]). It is remarkable that the maximum horizontal wind speed behind the leading edge, its propagation velocity, and the maximum uplift above the leading edge vary rather little over this period of 5 h indicating a stable dynamical evolution. After midnight, however, the leading edge becomes more and more diffuse, and is not detectable at 0300 UTC anymore, giving a total lifetime of the cold pool of 10 h. Potential reasons for the decay are the following: Radiative cooling of the surrounding desert surface at night stabilizes the atmosphere (see Figure 3b) and impedes the lifting of environmental air above the approaching density current and therefore its propagation [Proctor, 1989]. It also reduces the density difference between the cold pool and the ambient air. The reduced lifting and the penetration of the density current into a very dry environment suppress the regeneration of moist convection at the leading edge, which cuts off the source of evaporative cooling as indicated by the decaying precipitation. Without this source, vertical mixing quickly reduces the density contrast between the cold pool and the less dense environmental air Structure [13] Figure 3 shows the three-dimensional structure of the density current at 2200 UTC 03 June 2006, when the system is fully developed (see Figure 2c). The cold pool can be identified by the elevated 306-K isentropic surface to the southeast of the Atlas, which is almost vertical at the southern and eastern flanks reflecting the very strong baroclinicity along the leading edge (Figure 3a). Farther to the northeast (i.e., to the right-hand side of the plot) the 306-K isentrope gently slopes upward marking a weak baroclinic zone. The isentropic surfaces are color-coded with specific humidity showing that particularly the western part of the density current is substantially moister than the cool air to the northeast. The isolated nature of the cold pool and its high moisture content are clear signs of a local generation as opposed to an advection into this region. While the isentropic surface indicating the baroclinic zone in the northeast is relatively smooth, the cold pool has a more jagged structure, which points to effects of updrafts and downdrafts in the precipitation/evaporation zone, and high turbulence in the regions of strong horizontal winds (see also Figure 3b). Turbulent mixing in the wake of the head of a density current is a typical element found in laboratory and idealized numerical experiments [Simpson, 1997]. Details of this process, however, are not well resolved with 2.8-km horizontal grid spacing. The vertical cross section in Figure 3b shows that a vertical orientation is found for all isentropes between 306 and 312 K, with the latter spanning a vertical range from near the surface to about 680 hpa at the leading edge. The height minimum of the isentropes some 15 km from the leading edge might be an indication of descent behind the head of the density current [cf. Williams et al., 2008, Figure 3]. Vertical profiles of temperature and dew point on either side of the leading edge show that the regions with substantial cooling reach from the surface to about 800 hpa, while the moistening affects levels up to 750 hpa (Figure 3c). A height of 1.5 km of the main bulk of the cold pool agrees well with observations found in the literature [Sutton, 1925; Freeman, 1952], while weaker effects of the density current can be seen as high as 4 km as also observed by Williams et al. [2008]. [14] The air mass that the density current is intruding into (left side of Figure 3b and black profile in Figure 3c) is characterized by a deep well-mixed layer up to 600 hpa with an almost constant potential temperature q of about K, which is on the order of maximum daytime 2-m temperatures in this region. Winds in this layer are from easterly directions and the water vapor mixing ratio is almost constant. As the sun set almost 3 h before the time shown in Figure 3, radiative cooling has created a shallow stable layer at the surface with potential temperature being 2 3 K cooler than the mixed-layer air aloft, which is likely to contribute to the decay of the cold pool in the course of the night (see discussion at the end of section 3.1.1). The top of the boundary layer near 600 hpa is marked by an increase in q by 4 K within 50 hpa with no apparent differences between the region of the cold pool and the undisturbed surroundings (Figure 3b). The neutral stratification of the ambient air up to a level well above the height of the cold pool is an important prerequisite for its longevity [Takemi and Satomura, 2000]. The free troposphere differs only little between the regions ahead of and behind the leading edge with winds coming predominantly from westerly directions (Figure 3c). The high relative humidity and almost moist-adiabatic lapse rate above the 600-hPa level 5of16

6 Figure 3. Structure of the cold pool at 2200 UTC 03 June 2006 as simulated by the COSMO model. (a) Three-dimensional view of the area 27 N 35 N, 7.5 W 2.5 E looking from SE toward NW. Displayed are the model orography, the 306-K isentropic surface colored with specific humidity according to the scale, and regions with vertical velocity exceeding 4 m s 1 are indicated by white isosurfaces. (b) Vertical cross section along the green line in Figure 2c showing potential temperature (shading according to scale), horizontal wind speed (black contours every 5 m s 1 ), and the 3 m s 1 vertical velocity isotach in white. (c) Vertical profiles of temperature, dew point, and wind depicted in the form of a skew-t log-p diagram. The black (red) lines and arrows show a location ahead of (behind) the leading edge as indicated by the crosses in Figure 2c. indicate the cloud layer. This profile clearly fulfills the dry/ wet condition (i.e., dew point temperature depression greater (smaller) than 8 C at 700 hpa (500 hpa)) favorable for supporting dry microbursts as defined by Miller et al. [2008]. The wind profile shows that the cold pool is mainly spreading in the direction of the moderate westerly shear between the subcloud and the cloud layer. As discussed above there is a distinct maximum of near-surface horizontal winds behind the leading edge (Figure 2). Figures 3b and 3c show that the abrupt jump in wind speed and the change in direction from easterly to northerly winds prevails up to about 750 hpa with a maximum of more than 20 m s 1 at around 900 hpa. This leads to massive convergence in this layer and to strong ascent of more than 3 m s 1 along an extended stretch of the leading edge and throughout most of the mixed layer (Figures 3a and 3b). Maximum values of vertical wind reach 7 m s 1 at around 740 hpa in close agreement with modeling results by Takemi [2005] and observations by Williams et al. [2008]. This marked lowlevel ascent region is so unique that the authors decided to 6of16

7 use it as a criterion to identify the active part of the leading edge in all horizontal distributions shown in this paper (e.g., Figure 2). [15] The vertical profiles shown in Figure 3 allow an estimate of the density difference across the leading edge, which is closely related to the propagation velocity c (and also varies across the cold pool). In a simple pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi dry, frictionless two-fluid system c is proportional to ghdq=q 0, where g is acceleration of gravity, H is the height of the current, and q is potential temperature [e.g., Smith and Reeder, 1988]. In a moist system with vertically varying density contrasts it is more appropriate to use the integrated buoyancy to calculate c in the following form [Weisman and Rotunno, 2004]: Z H c 2 ¼ 2 g Dq þ 0:61ðq q 0 q 0 Þ dz; 0 where q is the mixing ratio of water vapor. Subscript zero stands for undisturbed background values. The effects of precipitation load can be assumed to be small in this dry environment and has been neglected. If the two profiles in Figure 3c are used as representatives of the disturbed and undisturbed environments, a theoretical propagation speed of 23.4 m s 1 is obtained, which is largely dominated by the temperature effect. This value is substantially larger than the 15 m s 1 determined from the propagation of the leading edge analyzed in Figure 2 (see section 3.1.1). Possible reasons for this discrepancy are friction and three-dimensional effects, both neglected in equation (1) Comparison to Observations [16] The foregoing discussion reveals that the COSMO model is able to simulate an extended, long-lived cool pool with a realistic three-dimensional structure. In the following these results will be compared with available satellite and ground-based observations. In IR satellite imagery, the leading edge is subjectively identified by tracing the boundary between the cold pool and the ambient air in cloud free regions or, even better, by tracing arc clouds forming owing to ascent at the leading edge similar to the method used by Knippertz et al. [2007] (e.g., red line in Figure 4d). [17] During the initial stage at 1600 UTC 03 June 2006, deep convection in the model is restricted to the region close to the High Atlas and there are only a few cloud streaks over the adjacent Sahara (Figure 4a). The corresponding Meteosat image equally shows deep convective cells over the mountains, but also over the Algerian Sahara (Figure 4b), which is consistent with ground observations. Surface stations Bechar and Errachidia (for station locations, see orange-filled circles in Figure 4) report thunderstorms and, respectively, 2 and 3 mm of rain between 1000 and 1800 UTC, and even the arid station Beni Abbes reports traces of precipitation before 1800 UTC. This difference in precipitation and therefore in the production of evaporatively cooled air has a pronounced impact on the density current extent, orientation, and propagation as we will see below. The reason why the model struggles to produce the convective cells over Algeria is unclear. Among the possible causes are too dry initial conditions in this extremely observation-sparse region, but midday dew point observations from synoptic stations and from the model both indicate a lifting condensation level around 600 hpa. It ð1þ appears possible that the model struggles to realistically simulate the evolution of the highly complex and very deep summertime desert boundary layer with its high aerosol content. It is also conceivable that ascent in the weak baroclinic zone described above (see Figure 3a) is underrepresented in the simulation with the COSMO model. [18] In the mature stage of the density current at 2200 UTC clouds and precipitation have shifted southeastward (Figures 4c and 4d). The COSMO model simulates an extensive region of mostly convectively generated cirrus clouds stretching in a west east direction across the Algerian Sahara. Single convective cells are hard to make out in this pseudo satellite image, but precipitation rates at the ground of more than 1 mm h 1 (blue lines in Figure 4c) indicate active convection, predominantly in the central part of this cloud band. The leading edge of the cold pool is located below the southern and western flanks of the clouds. In contrast, the corresponding Meteosat image shows a much smaller cloud band with several single cold-cloud-top cells, notably a large system around 30.5 N, 1 W (Figure 4d). The tendency of the COSMO model to produce too widespread cirrus in this situation has already been noted by Reinfried et al. [2009]. The observed leading edge is more extended than in the model and spans from the Moroccan- Algerian border to near the Saharan station Adrar. The cold pool is clearly separated from the cloud band and is propagating into southerly directions in contrast to the southeastward propagation in the model. This analysis is supported by station observations. Timimoun is reached by the cold pool between 1800 and 2100 UTC indicated by a dew-point increase of 8 K, a temperature drop of 7 K, a pressure increase of 3.1 hpa, and a drop in visibility from 10 km to 1.5 km, accompanied by observations of dust mobilization near the station. At 2100 UTC sustained 10-m winds reach 15 m s 1 and light rain is observed. Adrar reports a similar evolution but without precipitation between 2100 UTC 03 June and 0000 UTC 04 June [19] The differences between the model simulation and observations become very apparent in 24-h accumulated precipitation amounts for 03 June. Close to the Atlas, model and TRMM precipitation estimates agree well and there is some confidence for the latter through observations from Errachidia, Bechar, Beni Abbes, and Timimoun (Figure 5). Precipitation records at the five high-mountain stations of the IMPETUS climate network (located near 31.5 N, 6.4 W and between 1900 and 3850 m [see Knippertz et al., 2007]) of mm during the period from UTC are also in good agreement with the satellite data. Over the eastern part of the domain, however, discrepancies become quite large. TRMM precipitation estimates exceed 10 mm in this region, but unfortunately there is no ground truth to confirm such amounts (Figure 5b). This rather intense precipitation over the desert consists an additional source of evaporationally cooled air that is not present in the model and explains the larger, southward shifted, southward propagating cold pool in the observations (compare black dots in Figure 5a with black line in Figure 5b for 0000 UTC 04 June). The observed propagation speed, subjectively estimated from IR imagery between 1900 UTC 03 June and 0000 UTC 04 June, is about 11 m s 1 and thus considerably slower than in the model (see section 3.1.1). This points to a smaller density difference between cold pools and environ- 7of16

8 Figure 4. Comparison of brightness temperatures in K (left) simulated by the COSMO model with (right) corresponding Meteosat data for (a and b) 1600 UTC and (c and d) 2200 UTC on 03 June Red lines in Figures 4a and 4c mark the 10 m s 1 isotach of the 10-m wind as in Figure 2, and blue lines indicate instantaneous precipitation rates of 1 mm h 1. The red line in Figure 4d is the subjectively analyzed leading edge of the cold pool as in work by Knippertz et al. [2007]. Political borders, the 1000-m model orography isohypses, and synoptic stations (orange solid circles) are also indicated. ment, and thus overall weaker evaporative cooling. The larger extent is therefore mainly the consequence of a larger cold-air source. This comparison demonstrates that a correct initiation of convection is an important prerequisite for the correct simulation of the associated cold pools. [20] Unfortunately available observations are not sufficient to document the three-dimensional structure of the density current. However, observations from the surface station EMY in southern Morocco with 15-min temporal resolution can be used to compare observed and modeled characteristics of the leading edge. This station was passed by the cold pool in the late afternoon of 03 June. With respect to 2-m temperature both data sets show an abrupt drop on the order of 4 6 K in half an hour with the model leading by about 45 min (Figure 6a). The model satisfactorily reproduces the temperature evolution before this time with a slight overestimation of the diurnal cycle. The dew point time series also agree well showing a jump of about 6 K concomitant with the temperature drop (Figure 6b). The model output indicates a first (unverified) moisture jump around 0200 UTC, leading to a moist bias of about 2 K over most of the day. A jump and associated bias are also found for wind speed (Figure 6c). Closer inspection of horizontal winds, clouds, and precipitation suggest that some very small convective cells are triggered during the model spinup that cause these spurious signals (not shown). In the model, the wind signal associated with the passage of the afternoon cold pool consists of a relatively short high-wind 8of16

9 Figure 5. Comparison of precipitation accumulated over the 24-h period 0000 UTC 03 June to 0000 UTC 04 June 2006 (a) simulated by the COSMO model with (b) TRMM estimates and selected station observations.the cross marks the station Adrar, where no precipitation was observed. The black and red lines in Figures 5a mark regions of, respectively, strong vertical and horizontal winds at 0000 UTC 04 June 2006 as in Figure 2. The black line in Figure 5b is the corresponding, subjectively identified leading edge as in Figure 4d. phase with a peak velocity of 15 m s 1, while observations show more sustained winds of more than 10 m s 1 behind the leading edge. This difference can be expected to reduce dust emissions in the model owing to the relatively short time, during which the mobilization threshold is exceeded. Both data sets show an abrupt change from southwesterly to easterly winds and a pressure increase by almost 2 hpa at the leading edge (Figures 6c and 6d). The more abrupt pressure rise in the model is consistent with the differences in wind speed evolution. The strong decrease in pressure during the afternoon is most likely caused to a substantial degree by a combination of atmospheric tides and the strong daytime heating of the boundary layer over the desert. [21] The previous discussion reveals that the COSMO model is able to simulate realistically most of the general characteristics of the density current such as size, propagation speed, initiation time, and air mass characteristics. Problems with the triggering of convection over the desert, however, lead to differences in the precipitation distribution that affect the location and propagation direction of the system Other Simulated Cases [22] In order to test the representativeness of these results, reference simulations for two other cases documented by Knippertz et al. [2007] were conducted. These are the horizontally rather extended density current on 07 May 2006 and the more localized event on 31 May For 07 May the model quite realistically reproduces several convective cells that form over the High Atlas in the early afternoon and have propagated into the Algerian Sahara by 2100 UTC, similar to the case discussed in section 3.1. At this time the vertical-velocity identification criterion (black lines in Figure 7a) shows separate leading edges of four cells (Figure 7a), while the satellite image indicates only two already merged leading edges marked by arc clouds (Figure 7c). As in section 3.1, the COSMO model simulates a too extended cirrus shield that covers the leading edge at this time (Figure 7b). The propagation velocity of the leading edge between 1800 UTC 07 May and 0000 UTC 08 May is on the order of 14 m s 1, again with a slower propagation toward the west. Maximum grid point winds associated with the density current reach somewhat higher values of 16 m s 1 at the time shown in Figure 7a. By midnight the cold pool reaches an west east extension of more than 500 km (not shown). The southward displacement of the leading edge in the satellite data by approximately 50 km is mainly due to an earlier initiation of convection by about an hour than by a faster propagation of the cold pool. Overall the size, location, and propagation of the observed system are quite realistically reproduced. In agreement with this, the model-simulated precipitation Figure 6. Time series between 0000 and 2400 UTC 03 June 2006 of (a) temperature, (b) dew point, (c) wind speed (lines) and direction (diamonds), and (d) pressure observed at the IMPETUS station EMY (red lines and circles), and simulated by the COSMO model and interpolated to the station location (blue lines and circles). The slightly higher model orography at this point leads to the systematic differences in pressure in Figure 6d, which is measured with a rather coarse resolution of 1 hpa at EMY. Temporal resolution is 15 min for both time series. The time of passage of the leading edge between 1700 and 1900 UTC is indicated by orange shading. 9of16

10 Figure 6 10 of 16

11 Figure 7. As in Figure 2 but for the cold pools (a) at 2100 UTC 07 May 2006 and (d) at 2000 UTC 31 May The corresponding brightness temperature images (b and e) simulated by the COSMO model and (c and f) from Meteosat data are shown as in Figure of 16

12 grid point winds are also weaker, reaching 15.4 m s 1 at the time shown in Figure 7d. [24] This analysis corroborates the general ability of the COSMO model with 2.8-km horizontal grid spacing to generate cold pools with a realistic temporal evolution, structure, air mass characteristics, and propagation speed. The cold pool evolution reacts sensitively to temporal and spatial details of the precipitation distribution in the model as will be shown in section 4. The authors believe that one of the reasons for this overall satisfactory model performance is the robust orographic trigger provided by the Atlas Mountains when compared to the sometimes very problematic initiation of convection over flatter terrain [Montmerle et al., 2006]. Figure 8. Raindrop size distributions for different values of the dimensionless shape parameter m. D is drop diameter in millimeters, and n(d) is the corresponding number density in m 3 mm 1. matches the TRMM estimates well over the Sahara (not shown). Maximum values are reached over Algeria close to the southeastern corner of Morocco with storm totals exceeding 20 mm in the model, TRMM estimates, and surface observations. [23] Despite its smaller horizontal extent, the density current on 31 May is also reasonably well reproduced, although with some time delay. In satellite imagery deep convection over the Atlas is visible as early as 1200 UTC, while the model initiates the first precipitating cells about 2 h later (not shown). By 2000 UTC the leading edge has passed EMY [see Knippertz et al., 2007, Figure 7] and spreads into Algeria as indicated by brightness temperatures below (above) 295 K within the cold pool (in the surroundings) (Figure 7f). Owing to the delayed triggering of convection, the associated cirrus shields are less extended in the model than in the observations (Figure 7e) and the leading edge lags behind by about 100 km (Figure 7d). The propagation direction to the south and a propagation speed on the order of 10 m s 1, however, are in good agreement, and the accumulated model-generated precipitation satisfactorily matches TRMM estimates (not shown). In contrast to the other two cases, there are hardly any grid points with low-level ascent above 3 m s 1 along the leading edge (note the absence of black dots in Figure 7d). The authors suspect that this is due to the orientation of the leading edge almost parallel to the fairly strong westerly shear on this day, which impedes an optimal organization of the system [Weisman and Rotunno, 2004]. Weaker ascent is consistent with precipitation being more confined to the mountains and a slower propagation than in the other two cases. Maximum 4. Sensitivity Studies [25] In this section the sensitivity of the cold pool evolution over the northwestern Sahara to the treatment of microphysical and turbulent processes in the COSMO model will be tested, again using the case of 03 June described in section 3.1. The following subsection contains information on the physical meaning of the parameters to be varied in the sensitivity experiments, while the second subsection discusses the results Rationale [26] Several studies have shown that convective downdrafts are fostered by the existence of relatively small raindrops that can be more effectively evaporated in the subcloud layer [Kamburova and Ludlam, 1966; Srivastava, 1985; Proctor, 1989]. In the bulk microphysics scheme used in the COSMO model, evaporation of rain is highly parameterized and drop size distributions are represented by a gamma distribution of the form nd ð Þ ¼ N 0 D m expð ldþ; ð2þ where n(d) is the number density, D is the drop diameter, N 0 the intercept parameter, l the slope, and m the shape parameter. In one-moment microphysical schemes, N 0 and m are set constant, and l is a unique function of the rain water content (RWC). With some additional assumptions, the evaporation rate can then be calculated from the RWC. [27] In the following simulations the impact of the choice of m on the evolution of the density current is evaluated. In the standard version of the COSMO model m = 0.5 is used, which is to some extent motivated by the study of Schlesinger et al. [1988], who found a value of about 0.4 to be optimal on the basis of radar observations of evaporating rain. Here, results from additional model simulations using values of zero and unity are presented. For these runs, the intercept parameter, N 0, is determined as a function of m using equation (27) of Ulbrich [1983], N 0 ¼ expð3:2mþ: ð3þ [28] The corresponding raindrop size distributions are shown in Figure 8. The standard value of 0.5 used for the reference simulations gives a maximum near 0.4 mm. The distribution with m = 0 describes an exponential decrease of drop numbers with diameter and corresponds to the 12 of 16

13 Marshall-Palmer distribution as used in most bulk microphysical schemes [Lin et al., 1983]. Compared to m = 0.5 it has a reduced (increased) number of large (small) drops. For such a modification the evaporation in the subcloud layer is expected to increase, which should decrease precipitation at the ground and intensify the cold pool owing to the enhanced density differences to the environment. The opposite can be expected for a change to m = 1, which shifts the maximum of the distribution to about 0.8 mm (Figure 8). In the sensitivity experiments the changes to m were only applied in the parameterization of evaporation and the sedimentation velocity of rain, but not in other parts of the microphysics package. [29] In addition, the more sophisticated two-moment microphysics scheme of Seifert and Beheng [2006] (SB hereafter) is applied in one simulation. The SB two-moment scheme has recently been extended to include a separate hail category (U. Blahak, Towards a better representation of high density ice particles in a state-of-the-art two-moment bulk microphysical scheme, paper presented at 15th International Conference on Clouds and Precipitation, International Association of Meteorology and Atmospheric Sciences, Cancun, Mexico, 2008), and a new parameterization of evaporation of raindrops with a diagnostic relation for m [Seifert, 2008]. For the experiment with the SB twomoment scheme the time step had to be reduced from 30 to 10 s. In order to rule out an influence of the change in time step on the results, an additional simulation with the standard configuration but a time step of 10 s was run, and only negligible differences were found (not shown). It should be stressed that the usage of the SB two-moment scheme affects all microphysical processes and not just the parameterization of evaporation as in the experiments with the modified m parameter. [30] In another simulation the asymptotic turbulence length scale l 1 was modified. The TKE scheme of the COSMO model uses the classic Blackadar-Deardorff formulation of the turbulent mixing length l turb ¼ kzl 1 kz þ l 1 ; with height z and the von Karman constant k = 0.4. The standard value is l 1 = 200 m. In one simulation a decreased l 1 of 60 m is used, which reduces vertical mixing. This can affect the density current in two ways. First, reduced mixing of environmental air into the cold pool will slow its demise. Second, reduced mixing in the planetary boundary layer will sustain larger vertical temperature gradients during daytime. The associated destabilization of the lower layers can help to initiate convection and increase evaporation, which would also favor the formation of cold pools. The reduction of l 1 to 60 m was found to improve forecasts of deep convection over Germany (A. Seifert et al., The challenge of convective-scale quantitative precipitation forecasting, paper presented at 15th International Conference on Clouds and Precipitation, International Association of Meteorology and Atmospheric Sciences, Cancun, Mexico, Cancun, Mexico, 2008) and therefore this modification has recently been made operational at the DWD. ð4þ 4.2. Results [31] Figure 9 shows accumulated precipitation, the 10 m s 1 isotach and leading edge identification for the four sensitivity experiments described in section 4.1. The leading edge from the corresponding figure for the reference simulation, Figure 5a, is indicated in green. The drop size distribution used for the evaporation computation has a substantial impact on the amount of precipitation reaching the ground (Figures 9a and 9b). While the overall structure of the precipitation field remains almost unchanged, amounts over the Algerian desert are increased by several millimeters over a large area with m = 1, and reduced to traces with m = 0. The impact on the propagation of the density current, however, is moderate, resulting in differences in leading edge position of up to 30 km as compared to the reference simulation. As expected, the area of winds above 10 m s 1 is extended (reduced) with maximum winds of 16 m s 1 (14.4 m s 1 )form =0(m = 1) demonstrating the influence of the microphysics on the dynamics. This has some implications for dust mobilization, which, being a threshold problem with an approximately cubic relation between wind and vertical dust flux, reacts rather sensitively to changes in peak winds [Cakmur et al., 2004]. [32] The use of the SB two-moment scheme (Figure 9c) produces a precipitation pattern similar to the reference simulation (Figure 5a) with somewhat higher amounts over the Sahara (up to 10 mm at some grid points). The leading edge position agrees with the one-moment simulation within less than 10 km but extends more than 200 km farther to the north according to the vertical motion criterion (compare green and black dots in Figure 9c). Surprisingly the nearsurface winds behind the leading edge are weaker than in all other simulations with a maximum of only 12.4 m s 1 at 0000 UTC 04 June (15.6 m s 1 in the reference simulation), pointing to a weaker density current. The explanation for this unexpected result lies in a different temporal evolution as compared to the reference simulation. The convection initiation is concomitant, but the simulation with the SB two-moment scheme subsequently generates more precipitation and a more intense density current during the afternoon and early evening. At 1900 UTC the run with the SB two-moment scheme generates maximum wind speeds of 22.7 m s 1 as opposed to 17.5 m s 1 in the reference simulation (not shown). At 2000 UTC the eastern part of the leading edge lies about 42 km ahead of the one in the reference simulation (not shown). According to equation (1), the faster propagation in the sensitivity experiment should be related to a larger integrated density difference between the cold pool and the surroundings and thus to a stronger evaporative cooling associated with a more localized (in space and time) evolution of precipitation. Between this time until midnight, precipitation rates in the simulation with the two-moment scheme decrease significantly and the leading edge slows down. The described differences have many potential reasons that are difficult to unravel from a fully nonlinear model simulation. They include the usage of hail in addition to low-density graupel, the different treatment of particle sedimentation (no gravitational sorting in the one-moment scheme), and last but not least, the different parameterizations of evaporation of raindrops. [33] One possible explanation of the comparably small sensitivity of propagation velocity to changes in evaporation 13 of 16