Characteristics of a Hailstorm over the Andean La Paz Valley

The iconic hailstorm and flash flood episode of 19 February 2002 over La Paz city is numerically investigated in this article. Large scale atmospheric circulation is dynamically downscaled in order to take into account the complex orography forcing and local features. Satellite observations suggests late morning shallow convection over the Altiplano that becomes deep convection in the early afternoon around complex orography. The control simulation captures well the cloud evolution and suggest a two-stage precipitation mechanism. First, early convection occurred around 1200 LST and originated from 5 thermodynamic instability combined with lake breeze and orographic lifting. Rainfall discharge then generated cold pools. During the second stage, cold pools around complex orography were propagated by lake breeze and encountered the La Paz Valley breeze, triggering the deep convection near La Paz city around 1400 LST. We assess the importance of local features through numerical experiments, which include modification of orography, suppression of surface heat fluxes, changes of surface lake temperature and removal of the lake. We show the importance of orographic configuration as triggering mechanism for 10 convection initiation and for mesoscale circulation, the role of lake temperature for frontal breeze and propagation of cold pools, and of surface heat fluxes for atmospheric instability. This study highlights the complex interaction between lakes, surface heating and orography that favour deep convection and hailstorm formation, which is especially relevant around the Titicaca lake region.

this iconic hailstorm can be found on-line. This natural disaster (that resulted in 69 casualties and still keeps in people's minds) exposed the city structural vulnerability, providing valuable but at the same time hard lessons to the local crisis managers.
Even if the impacts of hail are well known over the Altiplano region, there is no systematic official hail observation system; neither ground observations nor remotely by radar. Furthermore, no data from insurance companies on hail are available. Most of the knowledge is taken from local people's perception of hail frequency and intensity. The lack of formal physical process 5 knowledge about local thunderstorms formation over this region is evident as we take as example the explanation given by the SENAMHI about the plausible mechanisms for this particular cell formation.
The SENAMHI argued that a stream of moist air came from the south-east through the La Paz valley towards La Paz city.
Once the stream reached the end of the valley, it was lifted by thermal and orographic effects. Upon reaching a given height of about 4000 meters above sea level, moisture began to condensate. These factors, coupled with atmospheric instability, 10 generated a super-cell over the city (Soruco, 2012). This explanation might sound trivial for a super-cell formation but a formal study about the atmospheric characteristics of this precipitation and flash-flood event has never been done.
The goal of this paper is to better understand the atmospheric processes leading to this hail and flash flood episode over the west side of the Andes-Amazon interface. For this purpose, here we will conduct several high resolution numerical simulations.
In section 2 we present the datasets and methods used, section 3 shows the main results including the large and mesoscale 15 conditions observed during the event and results from numerical experiments. We discuss the main findings in section 4 and we summarize the main conclusions in section 5.

Data and Methods
In order to study the atmospheric characteristics of this hailstorm, we combine satellite and stations observations with numerical studies. We focus on the time period preceding the cell formation on 19 February 2002 at 1430 LST. The investigation area 20 includes the Andes-Amazon interface connected by several valleys and the Altiplano region near Titicaca lake.

Meteorological conditions from Reanalysis
For assessing large scale meteorological features, we use the ECMWF global reanalysis ERA-interim (Dee et al., 2011). It has a temporal resolution of 6 hours and a spatial resolution of around 0.75 • × 0.75 • lat-lon with 60 vertical levels. We take the 25 geopotential fields at 200 hPa, and specific humidity and winds at 500 hPa from the 19 February 2002 at 1400 LST file. The ERA-interim dataset is also used as initial and boundary conditions for the following numerical simulations. The simulation horizontal domains are shown in Fig. 1a. 2 Nat. Hazards Earth Syst. Sci. Discuss., https://doi.org/10.5194/nhess-2019-27 Manuscript under review for journal Nat. Hazards Earth Syst. Sci. Discussion started: 5 February 2019 c Author(s) 2019. CC BY 4.0 License.

Satellite Information
We collect satellite image data from the NOAA Geostationary Operational Environmental Satellite (GOES-8) Imager in the visible channel (0.55-0.75 µm). The spatial resolution is 1 km and Full Earth-Disk images are produced 8 times per day. We take the images at 745, 1115, 1345 and 1745 LST, in order to investigate early shallow convection stages and the hailstorm evolution. We use the Tropical Rainfall Measuring Mission TRMM 3B42 version 7 satellite product (Huffman and Bolvin, 5 2013) to complement the GOES images. The TRMM Multisatellite Precipitation Analysis (TMPA) are available on a 3-h temporal resolution at 0.25 • × 0.25 • lat-lon spatial resolution. Despite known uncertainties of precipitation estimates over complex terrain (Rasmussen et al., 2013), they provide area-wise estimates with a fair temporal resolution. This information is particularly useful in remote regions with low weather stations density like the Bolivian Andes. We take the finest simulation domain region (D4 in Fig. 1a is equivalent to all maps shown in Fig. 2) for this purpose.

Weather Stations Network
The nature of this event demands a high spatio-temporal resolution precipitation dataset to reveal important characteristics of the hailstorm. The city of La Paz possesses a relatively high spatial (around 1 km distance) but low temporal (24 hours) resolution network of rain gauges; the network is maintained by SENAMHI. We use the available data over the region covered in Fig. 3a. Many of the stations are located on the hill slopes in the city.

15
Some data quality issues that affects large fractions of station datasets are well known in this region. Nevertheless recent efforts (Hunziker et al., 2017(Hunziker et al., , 2018 have addressed these problems and have produced flags that evaluate the quality of the observations. We use the original dataset taking into accounts the quality flags.

20
The Weather Research and Forecasting Model WRF-ARW version 3.9.1.1 (Skamarock et al., 2008) is used to investigate the physical mechanisms of convection. WRF is a non-hydrostatic next-generation mesoscale numerical weather prediction system designed for both atmospheric research and operational forecasting needs.
The study of a hailstorm atmospheric characteristics requires high resolution simulations. We define four one-way nested domains over the Bolivian central Andes D1, D2, D3 and D4 of 54, 18, 6 and 2 km of grid size, respectively (Fig. 1a).

25
This configuration allows an explicit treatment of deep convective processes in the finest domain. All domains uses Mercator projection and 60 vertical levels with a top level at 50 hPa. The most relevant model configuration can be found in Table 1 and is detailed as follows; we use the Thompson microphysics scheme (Thompson et al., 2008), the YSU planetary boundary layer scheme (Hong et al., 2006) for turbulent fluxes, the Noah land surface model (Ek et al., 2003), and the Rapid Radiative Transfer Model (Mlawer et al., 1997) for long and short-wave radiation. The Kein-Fritsch scheme (Kain, 2004) is used for We couple in addition the one-dimensional hail growth model WRF-HAILCAST (Adams-Selin and Ziegler, 2016) integrated into the WRF-ARW, in order to explore its capabilities for hail production. This approach was used with success in a recent study in the alpine region (Trefalt et al., 2018) and here we test it in the very high tropical Andes. This main configuration 5 (control run, hereafter CTRL) is used to investigate the physical processes leading to the hailstorm formation and it remain fixed for all set of simulations (sensitivity experiments).
The sensitivity experiments are all initialized at the same time as the control run (1400 LST on 17 February 2002) and they are useful to asses the individual importance of the main features in the region: lake, orography and surface heating. The main differences with the control run and the goals of each experiment are summarized in Table 2. 10 We asses the role of orography by modifying the terrain. The Smoothed Terrain Experiment (SMTR) has the valley partially filled up but enlarged, it reduces also (but not drastically) the mountain peaks above the lake. The Reduced Terrain Simulation (RDTR) explores the Altiplano circulation under reduced mountain heights above the lake level following the function, where, H n is the new orography, the real orography is H r and the lake level is H L .

15
We study he role of land surface fluxes for convection development by turning off the energy fluxes between land and atmosphere (NOHEAT experiment). Furthermore, we investigate the lake effects by adding to the lake surface temperature 3 • C (experiment LK+3). The last experiment consists of removing the lake and replacing it by the surrounding land surface type (experiment NOLAKE)

20
We assess the presence of the main ingredients for a hailstorm to occur (moisture, instability and lifting) through a series of atmospheric variables and diagnostics derived from the model output.
The low level moisture transport vectors were calculated following, 25 with g = 9.81 ms −2 the gravity constant, q as the specific humidity in gkg −1 , u and v as the horizontal wind in ms −1 and dp in hPa as the pressure thickness. It is calculated from the surface SF C up to 200 hPa.
The instability is assessed by the calculation of the convective available potential energy (CAPE). We also asses the surface heat fluxes contribution to buoyancy and explore the lifting mechanisms using vertical speed as a proxy.

Results
During the hailstorm and flash flood event's precedent weeks, several heavy precipitation episodes were registered over La Paz city. The continuous water contribution from rainfall kept the soil saturated and limited the absorption capabilities, favouring 5 surface runoff (Hardy, 2009). In the following we present the atmospheric situation of the event itself.

Synoptic Situation
On 19 February 2002 at 1400 LST, the well known anticyclone at 200 hPa (also called Bolivian High) was located over the north-east part of Bolivia (Fig. 1b). The intensity and position of the Bolivian High usually drives the large scale moisture 10 transport towards the Altiplano. A rather northern location allows the establishment of a westerly wind circulation which suppresses moist air transport from the Amazon towards the Altiplano (Garreaud et al., 2003).
The atmospheric conditions at 500 hPa confirms the presence of westerly winds over La Paz city in part due to the northward displacement of the Bolivian High and also because of the presence of a strong anticyclone over the Pacific Ocean (not shown).
We find a considerable amount of water vapour over the Bolivian Altiplano due to the continuous precipitation episodes 15 registered during precedent weeks.
This synoptic configuration is favourable for isolating the Altiplano from Amazonian influences, allowing the development of mesoscale features in the presence of a sufficient amount of humidity. It is still unknown how frequently this synoptic configuration occurs during thunderstorms and hail events, since a formal circulation classification is still lacking over the Central Andes. Early morning (Fig. 2a) was characterized by cloudiness over Titicaca Lake, the Amazon region and the eastern cordillera,

25
TRMM is able to capture convective rainfall over the Amazon and near the cordillera; the presence of low level water vapour is not well captured in this band but it's corroborated with infra-red image at 12 µm (not shown). The late morning image ( Titicaca and surrounding La Paz city. The northern cell was located southwards from the cordillera and it corresponds very well to the hailstorm location, while the southern cell is located around complex orography inside the Altiplano (called hereafter serranias). At this point the infra-red images are almost the same as the visible channel (not shown). Finally the convective cloud development arrives to its term during late afternoon (Fig. 2d). The Altiplano was then almost completely covered by clouds and TRMM shows important rainfall localized all over the cloud cover. We conclude that the development of the cell leading to the hailstorm event is mainly due to mesoscale features starting from shallow convection around the Altiplano and later deep convection around complex orography. The cordillera acts as a barrier not allowing much influence from the Amazon, consistently with the synoptic situation. Thus WRF is able to simulate the event with its most important features.

Rainfall estimations from the SENAMHI network
The SENAMHI stations network provides mostly 24h cumulated precipitation measurements from rain gauges. Hourly observations are rare and often incomplete. Nevertheless, they provide an idea of the intensity and the spatial distribution of the rainfall during that particular day. The most important rainfall quantity was therefore registered around La Paz next to the mountain slopes (with measured values of around 50 mm). Some places registered no precipitation, which is a sign of hetero-20 geneous precipitation spatial pattern (Fig. 3a). Consistently with satellite observations, station observations confirms that an important quantity of rainfall fell down close to complex orography and lake.

Physical processes for cell development from Control Run
The analysis of the large scale characteristics and the few observations available provides insufficient information about the three basic ingredients for a thunderstorm: moisture, instability and lifting. After a confirmation that WRF is able to reproduce 25 the event's main features, we complement the analysis with the output of the CTRL experiment.

Topographic features and rain propagation
We therefore explore the model's rainfall propagation over the red lines in (Fig. 3a) in order to explore the chronology of the precipitation. The Hovmoeller diagram along line L1-L2 (Fig. 3b)  cordillera and south serrania; forming a short duration rain-band (around one hour) over the Altiplano after propagation (Fig.   3c).
A closer look to the maximum radar reflectivity (in dBZ) spatio-temporal evolution in the model reveals late morning convection in places where lake and/or valley breeze encounter complex orography (Fig. 4a). Later on, the lake breeze becomes more intense and pushes the rain spots towards the east (Fig. 4b-c); at 1300 LST deep convection is already present and even 5 hailstones of around 5 mm are simulated at the centre of the two formed cells. The cells finally encounters each other and form a rain-band with isolated hailstorms at the convergence zone between the lake and valley breezes (Fig. 4d-e). Finally the lake breeze becomes weaker and the rain-band dissipates putting an end to this event (Fig. 4f).

Low level wind circulation and moisture transport
While the surface specific humidity over the Altiplano follows the lake breeze, the La Paz valley water vapour comes from 10 the Amazon avoiding the cordillera obstacle ( Fig. 5a-c). The calculated integrated vapour transport (IVT from surface up to 200 hPa) confirms the partial suppression of moisture transport from the Amazon over the cordillera (Fig. 5a-c). Our results suggest the La Paz valley and Titicaca lake as main humidity sources, with the moisture transport from the lake increasing slightly across time.
Mesoscale conditions a 1100 LST show a rather weak moisture transport following the lake breeze towards the cordillera 15 and both serranias (Fig. 5a). The lake breeze front is accompanied by strong winds at 500 hPa towards the cordillera and a bit weaker towards the serranias; a strong convergence next to the lake and serrania south also appears (Fig. 5d). Early afternoon (1230 LST) is characterized by stronger IVT from the lake with a small change of direction (Fig. 5b). We observe at the same time an intensification of previous convergence zones around complex orography; with a propagation of the convergence areas from the previous zones towards each other (Fig. 5e). Around 1400 LST, the lake breeze is dominant (Fig. 5c) and displaces 20 the well formed convergence line towards the edge of the valley (Fig. 5f).

Instability
At 1100 LST, instability (indicated by relatively high CAPE values of more than 200 J kg −1 ) develops around the mountain slopes (both cordillera and serrania) and over the Amazon, as shown in Fig. 6a. At the same time, sensible heat is released in non-cloudy areas (similar spatial distribution as Fig. 4a). During early afternoon at 1230 LST, surface heating intensifies 25 the land sensible energy flux over the Altiplano and La Paz valley, accompanied by an expansion and intensification of CAPE previously identified (Fig. 6b). The highest CAPE values (excluding the Amazon) are located at the north slope of the La Paz valley. Finally at 1400 LST (Fig. 6c) the CAPE intensifies over the valley and sensible heat flux from the surface is released anywhere except the cordillera, serranias and in the proximity if the rain-band.
The city heat-island behaviour is well captured by the model, providing a permanent heat source that contributes to warming and still close to saturation (CAPE of 970 J kg −1 in Fig. 6e)). At 1400 LST, the atmosphere is saturated until 400 hPa with important wind shear favouring hail and graupel formation (Fig. 6f).

Lifting and propagation mechanisms
Previous sections reveal convection initiation around humid and unstable zones near complex orography, resulting in the formation of two important cells (also detected by satellite). Both cells then propagates towards each other forming a strong 5 convection band. The location of this band overlaps the lake-valley breeze convergence zone. The evolution of the intensity of vertical velocity at 4000 meters above ground level (magl) and wind shear from surface to 6000 magl (Fig. 7a-c) gives an idea about the severity of afternoon convection and resulting storm. The responsible lifting mechanisms identified until now are orography and lake-valley breeze convergence. Therefore, we focus on these areas and explore the vertical structure between lake and valley (line A-B in Fig. 7a) and between cordillera and serrania (line C-D in Fig. 7a).

10
The lake-valley cross section confirms the early light convergence over the Altiplano and surface heating ( Fig. 7d-f). Nevertheless, convergence is not enough to explain deep convection. The later appearance of a cold pool (Fig. 7f) over the convergence zone (which coincides with the city's location), combined with orographic forcing, may have triggered the deep convection. Low level moisture is then rapidly lifted and encounters a supercooled atmospheric environment, highlighted by the low freezing level of around 1000 magl (also in Fig. 7f).

15
The cross section between complex orography shows important surface heating next to mountain slopes during early afternoon (Fig. 7i). Important convection is then initiated over the serrania and propagates towards the cordillera over the lake-valley convergence zone (Fig. 7h-i). Figure 7i reveals later appearances of cold pools over the mountains slopes which, combined with breeze convergence, may explain the multiple zones of deep convection. This cross section shows that the cold pool seen in Fig. 7f is part of the cold pool that propagated from the cordillera.

20
These results suggest that the two orographic cells produced rain that cooled down the surface and originated cold pools. The propagation of these cold pools, combined with instability, convergence and orographic forcing, then triggered deep convection resulting in a rain-band that included isolated hailstorms.

Sensitivity studies
After we described the hailstorm dynamics from the control run, we assess the principal elements participating to the storm 25 formation and propagation.

Orography influence
The smoothed terrain experiment (SMTR) provides a larger La Paz Valley extent, allowing a better organization and expansion of valley breeze. Since the mountains summits are still high, the orographic convection is still present. Nevertheless, the stronger valley breeze enhances convergence with lake front breeze and the cold pool lifting is no longer necessary for deep convection (Fig. 8a-c). The resulting rain region is less organized and the band expands heterogeneously with hail originated by valley breeze and orographic interactions (Fig. 10b) In the case of the reduced terrain (RDTR experiment), we purposely reduced the mountain heights and observed a drastic reduction of deep convection regions (Fig. 8d-f). Rainfall still exists but it is weaker (Fig. 10c), highlighting the importance of high mountains for convective initiation. Nevertheless, hail can still be produced by the model.

Soil Fluxes
In the experiment with the surface heat fluxes suppressed (NOHEAT), convection is still present without thermodynamic instability; but cells are very isolated. The precipitation regions (Fig. 10d) corresponds to complex orography, showing that breeze and orographic lifting are enough for producing rainfall.

10
A warmer lake (experiment LK+3) reduces the temperature gradient with the land surface and weakens the lake breeze. Cross section over A-B line shows that the La Paz valley wind becomes predominant and vertical motion is less intense (Fig. 9a-c).
This shifts the rainband towards the lake and suppresses hail formation (Fig. 10e).
When the lake is removed (NOLAKE experiment) the surface where the lake was located becomes a shallow valley and the wind circulation is similar to a lake breeze. Nevertheless, the main mechanism for convection is surface heating and orographic 15 lifting ( Fig. 9a-c). In this experiment, convection occurs later than experiments containing the lake and it concentrates around the serranias, leaving the cordillera storm free with isolated hailstorms (Fig. 10f).

Discussion
The iconic heavy hailstorm and flash flood on 19 February 2002 over La Paz city was an exceptional natural disaster that still remains in people's memory. There were several factors that increased the city damage: the hail location, the water management 20 system and the fast cell development.
Satellite images shows early shallow convection over the Altiplano during late morning that became deep convection over complex orography in the early afternoon; forming notably two cells, one over the cordillera and another over the south serrania. The two cells then propagated towards each other, producing strong precipitation and isolated hailstorms. Synoptic conditions were favourable to the mesoscale circulation development at near surface levels, produced mainly by lake breeze 25 and orographic thermal circulation.
Using WRF simulations, we present new insights on the local atmospheric conditions leading to the formation of deep convection and later hailstorm and flash flood episode over La Paz. We focus on the dynamical mechanisms before and during the event under the described synoptic conditions; notably moisture sources, atmospheric instability and lifting mechanisms.
The results from CTRL experiment are consistent with satellite and stations observations, synoptic circulation and historical evidence. The simulations output are able to reproduce the spatial pattern of cloud cover and precipitation. They are therefore able to reveal plausible key processes.
On 19 February 2002, surface wind over the altiplano was guided by thermal lake, mountain and valley breeze effects.
The significant precipitation falling on previous days over the region saturated the soil and provided a local moisture source (Hardy, 2011). On top of these already moist conditions, even more water vapour entered the region from the valley and the 5 lake following the thermo-topographic circulation.
The first convective cells were formed during late morning over convergence zones that were located over complex orography. They were triggered by a mix of low level wind convergence, surface heating and orographic forcing. Later on, the atmospheric unstable zones grew in surface and strength with a bigger amplitude over the valley. The surface heating increased the lake-land temperature gradient and exacerbated the lake breeze; keeping the northern cell over the cordillera and pushing 10 it towards La Paz. The southern cell at the same time also grew but stayed stationary.
However, the most remarkable results from WRF simulations concerns the secondary trigger mechanism. It is revealed that the first convective cells formed cold pools that propagated following the mesoscale circulation. This propagation allowed both cells to join each other resulting in a precipitation band. This auto-propagation mechanism has been observed by previous works over the Alps (Trefalt et al., 2018;Kunz et al., 2018) or in idealized situations (Schlemmer and Hohenegger, 2014), but 15 not yet over the central Altiplano.
The presence of sufficient wind shear extends and supports the organization of convective storms in terms of multicells, supercells or mesoscale convective systems. Wind shear is often measured between surface and 6 km above surface and for hail formation they can be moderate (Trefalt et al. (2018) found values of around 10 m s −1 ), or unusually high (Kunz et al. (2018) observed values of more than 40 m s −1 . In our case we find a rather moderate 0-6 km wind shear (around 10 m s −1 ), 20 but it becomes strong if we take a 0-9 km basis (around 20 m s −1 ), as shown in Fig. 7f.
The sensitivity experiments highlights the importance of orography for low level wind circulation and for triggering deep convection. A reduced mountain altitude suppresses deep convection while a smoothed but still high orography modifies the upslope and valley wind circulation. They also reveal the crucial influence of surface energy fluxes for atmospheric instability.
The lake breeze effect turns out to be determinant for storm propagation. The sensitivity experiments shows that a higher lake 25 surface temperature results in a smaller temperature gradient between land and lake, which weakens the lake breeze and shifts the convergence zones eastwards. Similar numerical studies found a comparable behaviour over the Tibetan plateau (Gerken et al., 2014). Gerken et al. (2014) found that the lake breeze strength controls the location of convection over a mountain.
While we get similar results, we also observe a competition between lake and valley breezes over the convergence line. A lake suppression eliminates the deep convection over the cordillera and highlights the lake breeze role in the northern cell formation.

30
While hailstorms are perceived to increase in frequency over the Altiplano (Boillat and Berkes, 2013;Yasukawa, 2011); hail hotspots have not been formally identified yet. Nevertheless, several studies have studied the formation of severe thunderstorms in the southern Andes that can be initiated by gravity waves (de la Torre et al., 2015(de la Torre et al., , 2011 or the passage of a cold front (Teitelbaum and D'Andrea, 2015). The common ground with our study is the importance of orography and a supplementary lifting mechanism. We find two lifting mechanisms in chronological order: early convection is initiated by a mix of breeze and orography, and later deep convection is originated by a combination of cold pool and orographic lifting at the end of valley.

Summary and Conclusions
This study analyses the main atmospheric characteristics of the historical hailstorm of 19 February 2002 over La Paz City. A first assessment from scarce station observations, satellite information and reanalysis suggests that this severe event was in fact 5 part of a mesoscale convective system. The large scale moisture transport towards the Altiplano was in part blocked by the high Cordillera at the interface Andes-Amazon due to the position of the Bolivian high. This synoptic condition allowed the formation of mesoscale thermal circulation.
The control run (previously evaluated with TRMM dataset) offers a better insight of the mesoscale atmospheric dynamics preceding the hailstorm. A first convection episode is generated by thermal instability and triggered by lake breeze and oro- The analysis conducted in this study highlights the complex interaction between large scale circulation, orography and local features in the formation of hailstorms over the tropical Altiplano. A semi-comprehensive scheme of participating mechanisms can be found in Fig. 11.
Our results are consistent at the same time with the few low resolution observations, the historical context and the known 25 atmospheric dynamics; showing that it is possible to complement the observations with numerical techniques in order to better understand local processes resulting in severe weather over the tropical Andes' complex orography.
The combination of large scale analysis and high resolution numerical techniques used in this study offer new elements for forecasting purposes. Nevertheless, we stress that the proposed mechanisms of this hailstorm formation should be confirmed by high resolution observations and further numerical investigations of similar high-impact events.