Introduction
Improving the simulation of tropical convective clouds in
convection-permitting simulations is an important yet challenging endeavour.
Forecasting centres are beginning to use operational numerical weather
prediction models with horizontal grid spacing of order of 1 km and while these
models have been shown to improve the diurnal cycle of convection and the
distribution of rain rates (e.g. Clark et al., 2007; Weusthoff et al., 2010),
there are numerous deficiencies at these resolutions that impact the
accuracy of the forecasts and the confidence in using these models to help
guide parameterisation development for coarser-resolution models and develop
retrieval algorithms for remotely sensed cloud properties (e.g. Del Genio and
Wu, 2010; Shige et al., 2009). One salient aspect of forecasting tropical
meteorology is the high ice water contents that are responsible for numerous
aircraft safety incidents as discussed by Fridlind et al. (2015). These
incidents tend to occur in fully glaciated conditions in the vicinity of deep
convection where high ice water contents can cause engine power loss (e.g.
Lawson et al., 1998; Mason et al., 2006; Strapp et al., 2015). In recognition
of this, an international field campaign called the High Ice Water Content
(HIWC) study was conducted out of Darwin in the beginning of 2014 and
provided a high-quality database of ice cloud measurements associated with
deep tropical convective systems. These observations are a valuable resource
for evaluating convection-permitting model simulations and cloud
microphysical parameterisations. In this work cloud properties are evaluated
from an operational model with the focus on the model's ability to simulate
high ice water contents generated from the outflow of deep convection and to
understand what modelled processes control the phase composition of the
simulated tropical convective clouds.
Many previous convection-permitting modelling studies of tropical convection have
documented common biases amongst models including excessive reflectivities
above the freezing level, lack of stratiform cloud and precipitation and too
much frozen condensate (e.g. Blossey et al., 2007; Lang et al., 2011;
Fridlind et al., 2012; Varble et al., 2014a, b). Lang et al. (2011) modified
a single-moment microphysics scheme to reduce the biases in simulated radar
reflectivities and ice sizes in convective systems and found better success
in a weakly organised continental convective case compared to a stronger
oceanic MCS (mesoscale convective system). The reason could be due to dynamical errors in the model that
had a greater influence on the microphysical characteristics in the
simulations of stronger convection. Varble et al. (2014a) compared cloud-resolving
and limited-area model simulations with the extensive database of
observations from the Tropical Warm Pool – International Cloud Experiment (TWP-ICE). They
found excessive vertical velocities even at 100 m horizontal grid spacings
and suggested that the overly intense updraughts are a product of
interactions
between the convective dynamics and microphysics. These strong updraughts
transport condensate and moisture to the upper levels, which contributes to the
larger amount of frozen condensate seen in simulations, and the reduced
detrainment at lower levels could play a role in the lack of generation of
significant stratiform cloud and precipitation. This has been seen by Tao et
al. (1993), who showed the importance of the horizontal transport of
hydrometeors from the convective to the stratiform regions for the generation
of stratiform rainfall. An increase in stratiform rain was also shown by
Ferrier et al. (1996) to occur when the rearward transport of condensate was
promoted through more upshear tilted updraughts. Morrison et al. (2009)
compared squall line simulations using a single- and double-moment
microphysics scheme and determined that the greater stratiform precipitation
region produced from the double-moment scheme was due to both a reduced rain
evaporation rate in the stratiform region and an increased evaporation rate
in the convective region. This had the effect of reducing the intensity of
the convection and increasing the mid-level horizontal flux of positively
buoyant air from the convective to the stratiform regions. In the operational
model used in this study, the microphysics scheme is a single-moment bulk
scheme. Model intercomparison studies have shown that double-moment
microphysics schemes do not necessarily perform better than single-moment
schemes and, in fact, provided that the intercept parameters are not fixed and
are able to vary, these more simple schemes can match or even outperform the
more complex double-moment schemes in their representation of cloud and
rainfall properties (e.g. VanWeverberg et al., 2013; Varble et al., 2014b).
The aims of this study are twofold: firstly to test different configurations
of the dynamics, turbulence and microphysical formulations in the model to
determine those that best represent tropical convective cloud systems and to
understand the sensitivities in the modelled cloud and dynamical properties
to these changes; secondly to determine what processes control the phase
composition and ice water content in the model. As mentioned previously,
observations of HIWC (defined here as > 2 g m-3 at 1 km
resolution) typically occur in glaciated conditions. However, as will be
shown, the model is unable to replicate this and instead produces mixed-phase
clouds under the same temperature regimes. For this reason we examine which
processes control the modelled phase composition in order to understand how
the model produces HIWC. This understanding will aid in improving the
representation of these clouds in the model and produce a better forecasting
capability. The following section describes the model and observations used
in this work. Section 3 compares the simulations with the available
observations including a time series comparison with the satellite data,
comparison of the simulated radar reflectivity characteristics with those
from the Darwin radar and an investigation into the controls on phase
composition in the model and how the IWC (ice water content) and ice particle sizes compare with
the in situ observations. This is followed by a summary of the results in
Sect. 4.
Description of the model and observations
The Met Office Unified Model (UM) version 8.5 is used to create a series of
one-way nested simulations. The global model configuration GA6 (Walters et
al., 2015) is the driving model, which uses the Even Newer Dynamics for
General atmospheric modelling of the environment (ENDGame) dynamical core
(Wood et al., 2014). The global model has a resolution of N512
(∼ 25 km) with 70 vertical levels and is run with a 10 min time step.
The convection scheme is based on Gregory and Rowntree (1990) and uses a
vertical velocity-dependent convective available potential energy (CAPE)
closure. The Prognostic Cloud Prognostic Condensate (PC2) scheme of Wilson et
al. (2008) is used with the microphysics scheme described by Wilson and
Ballard (1999) but with numerous modifications including prognostic
rain, cloud droplet settling and the Abel and
Boutle (2012) rain drop size distribution. The boundary layer scheme used is
based on Lock et al. (2000) and the radiative fluxes are determined by the
Edwards and Slingo (1996) scheme. The global model is initialised at
00:00 UTC using the Australian Community Climate and Earth System Simulator
(ACCESS; Puri et al., 2013) operational analysis for the case study date of
18 February 2014.
The first nested simulation within the global model is a 4 km grid length
simulation. These simulations are run with a 100 s time step and are forced
at the boundaries every 30 min. At this resolution the Smith (1990)
diagnostic cloud scheme is used where the critical relative humidity is 0.8
above 800 m and increases to 0.96 at the lowest model level. The cloud
microphysical parameterisations are the same as the global model except
that prognostic graupel is included and the generic ice particle size
distribution (PSD) scheme of Field et al. (2007) is used. The convection
scheme at this resolution has a modified CAPE closure that scales with grid
box area, which allows for more of the convective activity to be modelled
explicitly. The other difference from the global model is the diffusion.
While there is no horizontal diffusion in the global model, in the 4 km
model this is modelled by a Smagorinsky (1963)-type scheme and the vertical
diffusion coefficients are determined using a scheme that blends those from
the boundary layer scheme and the Smagorinsky scheme (Boutle et al., 2014).
The older dynamics scheme (named New Dynamics; Davies et al., 2005) is used
in the control model configuration, as that dynamical core was the one being
used in the high-resolution operational model forecasts for this version of
the model. However, the effects of the dynamics are also tested by using
ENDGame in a sensitivity experiment.
A suite of 1 km simulations are nested in the 4 km simulation that
investigates the effects of the dynamics, turbulence and microphysical
parameterisations on the simulations of tropical convective clouds. There are
80 vertical levels and the model is run with a time step of 30 s. The domain
is 500 × 500 km2 centred on the location of the Darwin radar
(12.25∘ S, 131.04∘ E) as shown in Fig. 1 and the convection
is modelled explicitly. Given that the focus of this work is primarily on the
cloud microphysics, a description of the scheme used in the model is
provided, with the details of the other parameterisations available in the
previously cited references. The microphysics scheme is described by Wilson
and Ballard (1999) but with numerous modifications. The single-moment scheme
carries water in four variables: vapour, liquid, ice and rain, with an
additional graupel variable in the 1 and 4 km simulations. The 4 km and
control version of the 1 km model use the generic ice particle size
distribution of Field et al. (2007), where the aggregates and crystals are
represented by a single prognostic aggregate variable. This parameterisation
is based on the idea of relating moments of the size distribution to the
second moment, which is directly proportional to the ice water content when
mass is equal to the square of the particle size. In using this
parameterisation there is no need to specify an intercept parameter for the
PSD and instead the microphysical transfer rates are derived from the moment
estimation parameterisation that is a function of ice water content and
temperature. The mass–diameter relationships take the form of a power law.
1 km simulation domain with the radar location denoted by the red
triangle and the 150 km range of the radar shown by the red circle. The
aircraft flight track is shown by the blue line with the domain used in the
aircraft comparison given by the blue circle.
![]()
Parameters used to define the mass–diameter relationships
(Eq. 1) and particle size
distributions (Eq. 2), where fT is given by Eq. (3).
Parameter
Units
Rain
Aggregates
Crystals
Graupel
a
kg m-b
523.56
2.3 × 10-2
2.3 × 10-2
261.8
b
3.0
2.0
2.0
3.0
N0
m-4
0.22λ2.2
2 × 106 fT
40 × 106 fT
5 × 1025λ-4
μ
0
0
0
2.5
mD=aDb
The particle size distributions are
generalised gamma functions.
ND=N0Dμe-λD,
where N0 is the intercept parameter, μ is the shape parameter and
λ is the slope parameter. The coefficients for each hydrometeor
species are given in Table 1, where the aggregate and crystal PSD
coefficients are for the simulations that use an explicit PSD and not the
generic ice PSD parameterisation. The explicit ice size distributions have a
temperature-dependent intercept parameter that decreases with warming
temperatures, representing larger particles and the effect of aggregation
(Houze et al., 1979), where in Table 1
fT=exp-maxTc,-45∘C8.18∘C,
following Cox (1988) with Tc the temperature in degrees Celsius.
Fall speeds are parameterised from power laws with the coefficients for
crystals and aggregates from Mitchell (1996), graupel from Ferrier (1994) and
rain from Abel and Shipway (2007).
Ice can be formed by homogeneous and heterogeneous nucleation processes. At
-40 ∘C and below, homogeneous nucleation instantaneously converts
all liquid water (both cloud water and rain) to ice. Heterogeneous nucleation
requires cloud water to be present at temperatures at or below
-10 ∘C. The process is dependent on relative humidity and the mass
of the number of active nuclei produced from the temperature-dependent
function from Fletcher (1962). Once ice has been formed it can grow by vapour
deposition, riming, collection and aggregation. The autoconversion of snow to
graupel occurs when snow growth is dominated by riming, with the additional
conditions that the snow mass threshold is exceeded and the temperature is
below -4 ∘C. Once graupel has formed it grows by riming and
collection. The ice hydrometeors experience sublimation, evaporation and
melting. There are a number of graupel transfer terms that have not been
included in the model as their rates are significantly smaller than the
dominant processes (Wilkinson, 2013). The graupel terms not included are
deposition and sublimation, wet mode growth, collection of ice crystals and
heterogeneous freezing of rain by ice nuclei.
The control model (denoted by nd) in the set of 1km simulations uses the New
Dynamics and the sensitivity to dynamical formulation is investigated by
testing the ENDGame dynamical core in the simulation denoted eg. Modelling
the vertical turbulent mixing using the 3-D Smagorinsky scheme rather than
the blended scheme used in the control simulation is labelled 3d. The other
experiments test aspects of the microphysical parameterisations:
nopsd – rather than use the generic ice PSD as in the control
experiment, explicit PSDs are used for ice where the single ice prognostic is
diagnostically split as a function of the temperature difference from cloud
top into two categories to represent the smaller more numerous ice crystals
and larger aggregates (Wilkinson, 2013);
qcf2 – as for nopsd but the crystals and aggregates are represented as two
separate prognostic variables;
qcf2hm – as for qcf2 but with the inclusion of an ice splintering
parameterisation that increases the deposition rate in the Hallett and
Mossop (1974) temperature zone of -3 to -8 ∘C. This
parameterisation represents the increase in the ice particle number
concentration due to ice splinter production during riming and is dependent
on the supercooled liquid water content, and as such the riming rate, as well
as the temperature that allows for increased deposition at temperatures
colder than -8 ∘C due to the vertical transport of ice splinters
(Cardwell et al., 2002);
qcf2ndrop500 – as for qcf2 but with an increase in the cloud droplet number
concentration from 100 to 500 cm-3;
qcf2sr2graupel – as for qcf2 but with the restriction that snow–rain
collisions do not produce graupel;
qcf2noqgr – as for qcf2 but without the inclusion of graupel;
qcf2rainfreeze – as for qcf2 but with the inclusion of a heterogeneous rain-freezing parameterisation based on the stochastic parameterisation of
Bigg (1953) following Wisner et al. (1972). This process represents the
heterogeneous freezing of rain by heterogeneous nucleation by ice nuclei;
qcf2raindsd – as for qcf2 but with the Marshall and Palmer (1948) rain drop
size distribution.
The Darwin C-band polarimetric (CPOL) radar (Keenan et al., 1998) collects a
3-D volume of observations out to a range of 150 km. The radar observations
have been interpolated onto the model 1 km grid, and the analysis of radar
reflectivities is for the area encompassed by the radius < 150 km from
the radar (see Fig. 1). The precipitation rates derived from the radar
reflectivity have uncertainties of 25 % at rain rates greater than
10 mm h-1 and 100 % for the lowest rain rates (Fridlind et al.,
2012). The satellite observations of outgoing longwave radiation (OLR) and
ice water path (IWP) were derived from the geostationary satellite MTSAT-1R
following Minnis and Smith (1998) and Minnis et al. (2008, 2011).
Observations from the French Falcon 20 aircraft include the IWC measurement
made with the isokinetic evaporator probe IKP-2 (Davison et al., 2009), and
the ice particle size distribution reconstructed from images of individual
particles from the 2-D stereo (Lawson et al., 2006) and precipitation imaging
probes (Baumgardner et al., 2001). The particle probes were fitted with
anti-shattering tips and the processing of the size observations accounted
for any possible remaining ice shattering by consideration of the
interarrival times and the ratio between the particle surface and lengths
(Leroy et al., 2015). Since the IKP-2 measures the total water content,
liquid water and water vapour contributions should be subtracted to obtain
IWC. Unfortunately, the hot-wire liquid water content (LWC) sensor on the
aircraft was unable to measure LWC below about 10 % of the IWC in
mixed-phase conditions, and LWC levels exceeding this value were very rare.
Fortunately the Goodrich Ice Detector could be used to detect the presence of
liquid water. Two such regions in two very short flight segments for this
case, research flight 23, were identified at -10 ∘C, and these
regions have been excluded from the analysis. The minimum detectable IWC of
the IKP-2 is determined by the noise level of the water vapour measurements
of the IKP-2 and background probes. This resulting noise level of the
subtraction of the background humidity from the IKP-2 humidity is a function
of temperature: it is about 0.1 g m-3 at -10 ∘C, dropping
rapidly to about 0.005 g m-3 at -50 ∘C. Since most data
were taken at temperatures colder than about -25 ∘C, a minimum IWC
of 0.05 g m-3 was chosen as the threshold to include in our analysis.
Two sources of vertical velocity are used from the Falcon 20. Position,
orientation and speed of the aircraft are measured by a GPS-coupled inertial
navigation system. The 3-D air motion vector relative to the aircraft is
measured by Rosemount 1221 differential pressures transducer connected to a
Rosemount 858 flow angle sensor mounted at the tip of the boom, ahead of the
aircraft and by a pitot tube which is part of the standard equipment of the
aircraft. Wind in local geographical coordinates is computed as the sum of
the air speed vector relative to the aircraft and the aircraft velocity
vector relative to the ground. Both computations use classical formulas in
the airborne measurement field described in Bange et al. (2013). The other
vertical air velocity measurement used is retrieved from the multibeam cloud
radar observations using the 3-D wind retrieval technique described in Protat
and Zawadzki (1999), and we use the technique described in Protat and
Williams (2011) to separate terminal fall speed and vertical air velocity.
Comparisons near flight altitude with the aircraft in situ vertical velocity
measurements show that the vertical velocity retrieval is accurate to within
0.3 m s-1. All observations are averaged to the model 1 km grid and
exclude observations when the aircraft roll angle exceeds 5∘.
Comparison of the simulations with observations
On 18 February 2014 the monsoon trough was stalled near the base of the Top
End with active conditions continuing about the northern coast. There was a
deep moisture layer and low-level convergence that produced a mesoscale
convective system. At 14:30 UTC, satellite imagery shows the convection
around Darwin was somewhat isolated in nature, with a convective cell
developing close to the radar (Fig. 2). This convection developed into a
larger organised oceanic mesoscale convective system by 18:00 UTC with deep
convective cells producing cloud top temperatures of -80 ∘C. A
widespread region of anvil cloud produced from the outflow of deep convection
was seen to develop from 18:00 UTC and persist for over 8 h. The HIWC
research flight penetrated convective cores in a region north-east of the
radar at 22:00–24:00 UTC (Fig. 1) with peak ice water content up to
5 g m-3 at 1 s resolution. There was almost no supercooled water
detected during the flight, even at -10 ∘C, and graupel was
intermittently observed. The absence of supercooled water coupled with the
occasional presence of graupel is due to the system being sampled at the
mature-decaying stage, where the supercooled water had been consumed in the
production of graupel. Most of the time the particle images were of dense ice
aggregates at flight level, except within some convective cores where graupel
was observed, as also indicated by strong W-band attenuation.
Top row: time series of enhanced infrared satellite imagery over the
Darwin region on 18 February 2014: (a) 14:30, (b) 17:30,
(c) 20:30 and (d) 23:30 UTC. The temperatures range from
230 K in blue through to 190 K in white and purple. Middle row: time series of
observed outgoing longwave radiation centred on the Darwin radar, where the
pixel-level satellite data have been interpolated onto the 1 km model grid.
Last row: as above, but for the modelled outgoing longwave radiation from the
control experiment, labelled nd.
![]()
Time series of domain mean: (a) precipitation (mm h-1)
and (b) ice water path (g m-3). The observations are from the
CPOL radar in (a) and the satellite retrieval in (b); note
that the observed IWP is only plotted from 22:30 to 23:30. The time period
spans 12:00–24:00 UTC on 18 February 2014. (c) 2.5 km observed
radar reflectivity averaged over 17:00–18:00 UTC. (d) As in
(c) except for the modelled reflectivity from the control simulation
(nd), (e) as in (c) except for 23:00–24:00 UTC,
(f) as in (d) except for 23:00–24:00 UTC.
![]()
Comparison of the modelled outgoing longwave radiation (OLR) with the
satellite observations in Fig. 2 show that in general, the control simulation
represents the life cycle of the MCS fairly well. The location of the mostly
oceanic convective cells look reasonable; however, the modelled MCS is larger
and composed of more numerous and deeper convective clouds than what was
observed in the pixel-level satellite OLR data and seen in the low-level
radar reflectivity fields shown in Fig. 3. The model also produces more
convection over the Tiwi Islands than what was observed at 17:30 UTC. As the
MCS transitions from a developing-mature system through to a mature-decaying
system, the observed reduction of deep convective cells with time is
simulated, although the OLR remains significantly underestimated. During the
research flight at 23:30 UTC, the modelled MCS shows cloud positioned in a
similar location to that observed with respect to the MCS structure; however,
the modelled cloud is shifted somewhat to the north-east (Fig. 2h, l).
The mean precipitation rates and ice water path (IWP) (Fig. 3) calculated for
the radar domain shown in Fig. 1 demonstrate that a larger IWP implies a
larger surface rainfall rate as seen in previous tropical studies (e.g. Liu
and Curry, 1999). The radar-derived precipitation shows that the simulations
overestimate the domain mean rainfall rate during the development stages of
the MCS and produce the peak in precipitation about 2 h earlier than is
observed. The model precipitation maximum occurs when the simulated
convection is strongest, as measured by the largest domain mean vertical
velocity at 500 hPa and the maximum vertical velocities. The observed domain
mean rainfall maximum corresponds to the time when the domain mean cloud top
height is highest (not shown) and, together with the infrared satellite
imagery (Fig. 2), suggests that the generation of significant anvil cloud
occurs before the domain mean precipitation maximum, rather than when the
convection is strongest as is the case in the simulations. Note that the
simulated domain mean precipitation rate at both the earlier and later times
is outside of the uncertainty range of the radar-derived rainfall rate
(Fridlind et al., 2012).
The underestimate in modelled surface rainfall for the later times when the
MCS has matured is not due to an underestimate in the domain mean upper
tropospheric cloud cover, as both the model and satellite observations show
mostly overcast conditions, but rather the underestimate in condensate
reaching below the freezing level (Fig. 3f). The observed IWP is only valid
for the daytime from about 22:30 UTC or 08:00 local time and, while the
simulations with the generic PSD parameterisation compare well with the
satellite-derived value, the comparison of VISST (Visible Infrared Solar-Infrared Split Window Technique)-derived IWP with CloudSat in
tropical regions was shown by Waliser et al. (2009) to be underestimated by
25 %, likely due to the maximum retrieved optical depth being limited to
128. Together with the CloudSat uncertainties (30 bias and 80 % root mean
square error; Heymsfield et al., 2008), this suggests that the modelled
domain mean IWP may be underestimated from 22:30 to 23:30 UTC. Other studies
have documented the lack of stratiform rainfall in convective-scale
simulations and some attributed the error to excessive evaporation in
single-moment microphysics schemes that use a constant intercept parameter in
the rain DSD (drop size distribution) (Morrison et al., 2009). That is not the case in this work and
rather the cause is likely due to overly strong convection (Figs. 2 and 3d)
that detrains too high and does not produce enough condensate in the lower
stratiform regions as has been shown by Ferrier et al. (1996), Tao et
al. (1993) and Morrison et al. (2009).
The greater IWP in the simulations that use the generic ice PSD
parameterisation is associated with larger relative humidity in the upper
troposphere (Fig. 4a: nd, eg, 3d). In a study comparing different
microphysics schemes, VanWeverberg et al. (2013) found the same result and
associated the increased moisture with the sublimation of ice particles due
to the scheme with the slowest ice fall speeds producing the greatest
condensate and moisture. That is not the case for this current study where
the larger IWP and relative humidity is produced by the microphysics
configuration that produces larger mean mass-weighted particle sizes
(Fig. 4c) but similar ice fall speeds above about 12 km and faster speeds below
this height. Figure 4b shows the fall speeds for the ice crystals and
aggregates/snow particles. All simulations use the same formulation for snow
and, even though the generic PSD only represents a single hydrometeor
category,
there are two fall speeds used to enable a representation of both fast and
slow sedimenting particles based on size. The method when using the generic
PSD is described by Furtardo et al. (2014) where, for narrow size
distributions and small mean sizes, the fall speed used is that shown for the
ice crystals in Fig. 4b, and for broader size distributions and larger mean
sizes the snow fall speed is used (the crossover is around
600 µm). Looking at the mean mass-weighted ice diameters in Fig. 4c
and d shows larger sizes for the simulations that use the generic PSD; however, the slower ice crystal fall speed used in these cases produces a
similar mean fall speed to the simulations that use two ice prognostics.
(a) Simulated relative humidity is for the area encompassed
by the 150 km radius centred on the Darwin radar on 18 February 2014 from
23:00 to 24:00 UTC. (b) Ice fall speeds (m s-1) as a function
of diameter (µm) for the snow category and the ice crystals used in
the simulations with the explicit and generic PSD, see text for details.
(c) Mean mass-weighted snow diameter (µm) as a function of
temperature (∘C) where the observations are from the aircraft and
have been averaged to be representative of a 1 km2 grid cell.
(d) As for (c) except for the mean mass-weighted ice
crystal diameter (µm).
![]()
(a) Vertical profile of convective updraught
(> 1 m s-1) mean horizontal mass divergence
(10-4 kg s-1 m-3) at 18:00 UTC. (b) Scatter plot
of θe against Δθd at 14 km for two
simulations that change the turbulent mixing (3d), and add an additional ice
prognostic variable and have smaller ice sizes (qcf2). (c) Histogram
of Δθd at 14 km. (d) As in (b) except
for 6 km and comparing the control (nd) and the 3d simulations and
(e) as in (c) except for 6 km. See text for details.
![]()
The higher RH in the simulations using the generic ice PSD could be due to
the larger, faster falling particles in the levels below 12 km, removing more
of the LWC via riming, which would allow for greater supersaturation. To be
able to conclude this with certainty would require additional experiments
that isolate individual processes, something that is beyond the scope of this
study; however, the subsequent results to be presented support this possible
line of thinking. More riming would release more latent heat, which, along
with the larger ice particles being more effectively offloaded, could lead
to the generation of stronger updraughts with less entrainment and higher RH in
the upper troposphere. This is illustrated in the convective updraught
(> 1 m s-1) horizontal mass divergence profiles shown in Fig. 5a.
As discussed by Yuter and Houze Jr. (1995), the presence of decelerating
updraughts and accelerating downdraughts can be largely explained by entrainment.
Entrainment reduces the buoyancy of updraughts, slowing and eventually stopping
the air parcel, which is where divergence is expected. In contrast,
entrainment into downdraughts enhances evaporative cooling, increasing the
downward mass transport and convergence. Note that above 16 km the vertical
velocities show oscillatory motions consistent with gravity waves; therefore, above this height the mass divergence appears to be driven by
these waves.
Figure 5a shows that horizontal mass divergence in the mixed-phase regions of
the convective updraughts is the most sensitive to the turbulence formulation
in the model, with the simulation with greater turbulent mixing (3d) showing
greater mass divergence, indicative of greater entrainment, in the range of
5–8 km. This contrasts with the upper ice-only regions of the convective
updraughts that show that the largest control on horizontal mass divergence is
the ice sizes. The simulations with smaller sized particles have more mass
divergence above 12 km, indicating more entrainment and a larger reduction
in the buoyancy in the upper levels of convective updraughts than the
simulations with larger sized ice particles. This is confirmed by examining
the convective updraught buoyancy properties at 14 km shown in Fig. 5b and c.
The buoyancy, Δθd, is calculated from the difference in
the density potential temperature (which includes condensate) from the slab
mean for the convective updraughts with vertical velocity > 1 m s-1.
Comparing the equivalent potential temperature as a function of Δθd at 14 km (Fig. 5b) between simulations with larger (3d) and
smaller (qcf2) ice sizes shows that for the positively buoyant updraughts, the
simulation with smaller ice sizes has fewer occurrences of high
θe. This gives support to the argument derived from the
convective updraught horizontal mass divergence that entrainment is larger in
the upper ice-only convective updraughts when the ice sizes are smaller,
although we do note that some of this difference could be due to differences
in freezing. To analyse this in more detail, the histogram of convective
updraught buoyancy (Fig. 5c) shows a greater number of occurrences of more
positively buoyant clouds at 14 km for the simulations that have larger
sized ice particles, supporting the argument that less horizontal mass
divergence represents less entrainment with more positively buoyant updraughts
that penetrate higher (as confirmed by examining the cloud top height
distributions; not shown). Similarly, comparing θe as a
function of Δθd at 6 km between the control simulation
(nd) and the one that increases turbulent mixing (3d) shows that the case
with greater mixing has significantly more occurrences of low
θe, consistent with greater entrainment. Note that the
increased number of occurrences of positively buoyant convective updraughts at
6 km in the 3d simulation is due to the increased cloudiness at these levels
as shown in Fig. 6 and discussed in the next section.
Radar reflectivity characteristics
The model hydrometeor fields have been converted into radar reflectivities by
assuming Rayleigh scattering, with no consideration of the effects of
attenuation or attempt to model the radar bright band. Due to the long
wavelength of the CPOL radar (5.3 cm), modelled reflectivity is calculated
following Hogan et al. (2006) where the reflectivity is considered
proportional to mass squared
Z=R∫0∞MD2NDdD,
where R=1018K20.936πρ2, ρ is the particle density and the mass M and particle
size distribution N(D) are defined by Eqs. (1) and (2). For cloud liquid
water the reflectivity is calculated from the constant number concentration
of 100 cm-3 in the simulations with the size distribution ND=PD2exp-λD, where P=N/2λ3 following
McBeath et al. (2014). The dielectric factor K2 is set to
0.93 for water and 0.174 for ice. The particle densities used in the
calculation of R are 1000 kg m-3 for rain, 917 kg m-3 for
aggregates and crystals and 500 kg m-3 for graupel. For the
simulations that use the generic ice PSD parameterisation, the aggregate
reflectivity is proportional to the fourth moment of the PSD, which is
calculated from the Field et al. (2007) moment estimation parameterisation.
Statistical radar coverage analysis
To examine the temporal evolution of the mesoscale convective system and
evaluate the modelled MCS life cycle and the simulated reflectivities, a
statistical coverage product has been produced following May and Lane (2009).
The data used to construct the statistical product are reflectivity fields
from CPOL and the simulations every 30 min for 12 h from 12:00 to 24:00 UTC.
At each height the fraction of the total area within the radar domain covered
by reflectivity thresholds is calculated, with the thresholds chosen as 10,
20, 30 and 40 dBZ.
The observed statistical radar coverage product shown in Fig. 6 illustrates
the development of the MCS. At 12:00 UTC the radar domain has a low
fractional area coverage of up to 0.15 for the 10 dBZ threshold, showing
that at 12:00 UTC there were radar-detectable hydrometeors covering
5–15 % of the radar sampling area between the lowest detectable altitude
of 1.5 and 8 km. Highest reflectivity echo tops of 11 km are seen in the
> 10 dBZ fractional coverage at 17:30 UTC, which coincides with the time
that the very cold cloud tops associated with deep convective cells were seen
in the satellite imagery (Fig. 2). The maximum coverage of the domain by
hydrometeors with reflectivities > 10 dBZ is 85 %, seen at
21:00–22:00 UTC, which is when the large anvil cloud shield appears a few
hours after the deepest convection occurs. The observed areas of reflectivity
> 10 dBZ are fairly uniform with height from 2 to 6 km, demonstrating
little variability of the reflectivity echo coverage from the low levels to a
couple of kilometres above the freezing level. Fractional areas larger than
0.05 with reflectivities > 20 dBZ are mostly confined to below 6 km,
with the maximum fraction of 0.65 occurring at 21:00 UTC at 4 km. The
> 30 dBZ area is not greater than 10 % until 16:00 UTC and is
maximum between 20:30 and 22:00 UTC at 4 km with a value of 0.35. There is no
fractional area of the domain > 0.05 that contains observed reflectivities
greater than 40 dBZ.
The observed (top panels), simulated by the control model (middle
panels) and with a change to the turbulent mixing (lower panel)
fraction of radar-detected area covered by reflectivities greater
than (a, e, i) 10, (b, f, j) 20, (c, g, k) 30 and
(d, h, l) 40 dBZ for 12:00–24:00 UTC on 18 February 2014.
![]()
While the statistical radar coverage product produced for the control
simulation, nd, does show a transition to widespread stratiform cloud
regions, as shown by the peak < 10 dBZ coverage at 21:00 UTC and
predicts the timing of the deepest clouds generally well (Fig. 6), there are
clear deficiencies in the simulated evolution of the MCS. There are much
larger high dBZ fractional areas, deeper clouds occur too early in the
simulation and there is a strong vertical gradient in the area coverage with
height. The less uniform vertical area coverage shows that the simulated
clouds have more variability in reflectivity with height compared to the
observations. In coarse-resolution models, a common model error is too little
detrainment at the freezing level (e.g. Franklin et al., 2013); however, in
this convection-permitting simulation the change in hydrometeor area with
height is mainly due to too little stratiform cloud and rain area, which
explains the reduction in area below the melting level and the
convective–stratiform-modelled ratio being skewed towards more convection
than is observed (discussed in Sect. 3.1.3).
Contoured frequency with altitude diagrams of radar reflectivity for
the region within 150 km of the radar for the times 23:00–24:00 UTC.
(a) Observations, (b) control simulation,
(c) ENDGame dynamical core simulation, (d) no use of the
generic ice PSD parameterisation, (e) additional ice prognostic and
(f) inclusion of heterogeneous rain-freezing parameterisation. See
text for details on different simulations.
![]()
A clear difference between the observations and the simulation is the
> 20 dBZ reflectivity areas above the freezing level. The observations
show some hydrometeors present 1–2 km above the freezing level that have
reflectivities > 20 dBZ, but no areas that meet the minimum threshold of
5 % that have reflectivities > 30 or 40 dBZ. The simulation on the
other hand shows > 20 dBZ fractional areas > 0.6 indicative of
larger ice particles in the model than in the observations, which will be
explored in detail later. The simulated reflectivity area > 30 dBZ above
5 km is due to the presence of both ice and rain, and the > 40 dBZ areas
are almost exclusively due to rain. The simulated rain above the freezing
level that is not observed suggests that the model has faster updraughts than
observed, which loft large rain particles upwards, and/or the heterogeneous
freezing of rain that is not represented in the model is an important process
in tropical convection and/or other errors exist in the representation of the
rain DSD. This result is what motivated the experiment with the addition of a
heterogeneous rain-freezing parameterisation, as observations in oceanic
convection have shown that most drops freeze between about -6 and
-18 ∘C (Stith et al., 2002, 2004; Heymsfield et al., 2009).
All simulations show the same main errors in the statistical radar coverage
as the control case, nd. The simulation that uses a differing turbulent
mixing formulation (3d) produces the closest representation of the observed
fractional areas for the dBZ thresholds of 10 and 20 dBZ in the larger areas
below the melting level (Fig. 6i, j). This can likely be attributed to
greater horizontal mass divergence between 5 and 8 km at the earlier
convective times (Fig. 5), indicative of increased entrainment and mixing of
environmental air in this simulation, which acts to increase the amount of
IWC (Figs. 3 and 13) and the area of precipitation.
Contoured frequency by altitude diagrams
The CPOL contoured frequency by altitude diagram (CFAD) using the
observations from 23:00 to 24:00 UTC every 30 min exhibits a fairly narrow
distribution at the heights above the freezing level, with the altitude range
of 12–13 km having little variability, reflecting the dominance of small
ice particles growing primarily by deposition in the uppermost cloud levels
(Fig. 7a). Below 10 km the distribution shows increasing reflectivity with
decreasing height as particles grow rapidly through aggregation, with
reflectivities centred on the modal value of 10 dBZ. At altitudes below the
melting level the distribution widens and the reflectivities extend from
5 to 35 dBZ with the largest occurrences around 30 dBZ. The lack of a
predominant bright band in the observations is likely due to the data being
collected from volumetric scans; however, there are slightly higher
reflectivities seen at 4 km indicating a bright band.
The simulations all show the common errors of clouds within these
reflectivity regions extending too high, reflectivities that are too large
between 4 and 6 km, greater reflectivity range below 4 km and disjointed
profiles due to separate hydrometeor categories. The simulations show more of
a convective-type profile with broader distributions above the freezing level
compared to the observations. The more numerous high-reflectivity outliers in
the simulations indicate a larger number of deep convective cells and/or a
smaller proportion of convective–stratiform area.
The simulation with the different dynamical core, ENDGame (eg) shown in
Fig. 7c, shows higher clouds and a broader range of reflectivities at
14–16 km. This latter result suggests the presence of large particles being
lofted into the upper cloud levels by intense convective cores, as can be
seen by the 40 dBZ reflectivities at 17 km. The observations do show some
sign of this lofting occurring at 11–12 km; however, the reflectivities are
constrained to be < 20 dBZ. This feature can also be seen in the cases
that include the ice splintering process, the limited graupel case and the
increased droplet number concentration case (not shown). The simulations that use the
generic ice PSD parameterisation (Fig. 7b and c; nd and eg) overestimate the
occurrence of low reflectivities above 10 km and have a modal reflectivity
at 6–8 km that is too low compared to the observations. Using explicit ice
PSDs produces a closer match to the observed reflectivity distribution above
10 km, although the simulated clouds still have greater vertical extent, and
the modal value of the reflectivities at 6–8 km with the explicit PSDs is
approximately 5 dBZ too large (nopsd, qcf2). The inclusion of a
heterogeneous rain-freezing parameterisation reduces the number of
occurrences of reflectivities > 20 dBZ between 5 and 10 km and reduces
the cloud top heights (qcf2rainfreeze). Both of these results agree better
with the observations suggesting that this process may be important in
tropical convective cloud systems. However, given the errors in the dynamics
and microphysics in the model for this case, further study is required to
better understand the effects of this process. Even in the simulation without
graupel, the reflectivities are overestimated at the melting level (not shown)
and this is due to the ice aggregate PSD.
Focussing on the 2.5 km reflectivity distribution shown in Fig. 8a allows an
evaluation of the rain properties from the simulations, in particular the
rain DSD. All simulations except for one use the Abel and Boutle (2012) rain
DSD, with the remaining simulation testing the sensitivity of rain drop sizes
by using the Marshall and Palmer (1948) DSD. The Abel and Boutle rain DSD
represents the observed rain reflectivity distribution fairly well; however,
the observed peak of 30 dBZ is underestimated and there are too many
occurrences in the tails of the distribution. The contribution from the
convective updraughts is demonstrated by the largest occurrences in the high-reflectivity tail coming from the simulation with the different dynamical
core (eg). It is this ENDGame simulation that produces the strongest updraughts
(Fig. 11) and is the least representative of the observed rain reflectivity
distribution for the reflectivities > 40 dBZ. The simulation using the
Marshall–Palmer DSD (qcf2raindsd) peaks at too low a reflectivity at around
10 dBZ and produces too many small rain drops with low reflectivities.
Radar reflectivity probability density functions for two heights,
(a) 2.5 and (b) 6 km.
![]()
At 6 km the observations again show a bimodal reflectivity distribution,
with the largest peak centred on approximately 16 dBZ (Fig. 8b). The
simulations show a more complicated distribution at this height with multiple
modes due to the presence of multiple hydrometeor species. The simulations
that use the generic ice PSD parameterisation peak at -1 dBZ (nd, eg, 3d).
When this parameterisation is not used and the explicit ice size distribution
is used, the peak is too high at 24 dBZ (nopsd). When an additional ice
prognostic is added this peak is reduced and compares better to the
observations at 18 dBZ (all qcf2 simulations); however, the tail of the
distribution in these cases is too long with too many occurrences at high
reflectivities. While the tail of the distribution for the generic ice PSD
cases is also too long compared to the observed reflectivity distribution,
these cases represent the graupel reflectivities better than the cases that
use the explicit PSD even though all cases use the same graupel PSD. The
better graupel representation with the generic ice PSD coupled with the
significantly larger occurrence of weak reflectivities around 0 dBZ is
similar to the result found by Lang et al. (2011). They modified microphysics
parameterisations to reduce the occurrence of excessive large reflectivities
and found that this resulted in too many low reflectivities due to a shift in
the reflectivity distribution, as is the case here when comparing the
generic and explicit ice PSD cases.
To examine to what extent the generic ice PSD parameterisation is
misrepresenting the observed reflectivities or how much the erroneous cloud
dynamics are responsible for errors in the modelled reflectivities, the PSD
moments derived from the generic PSD parameterisation using the observed IWC
and temperature are shown in Fig. 9. In calculating the predicted moments, the
observed mass–diameter relation was used, m=4.97×10-3D2.05, and
the observed moments are calculated only for particle sizes
> 100 µm in diameter and for IWC > 10-3 g m-3 to
be consistent with the data used to derive the Field et al. (2007)
parameterisation. The fourth moment is equivalent to radar reflectivity when
mass is proportional to the square of the particle diameter, and it can be
seen in Fig. 9a that the parameterised reflectivity results in
an overestimate of the larger reflectivities. The generic ice PSD
parameterisation underestimates the zeroth and first moments and has a good
representation of the third moment. The underestimate of the number
concentration (Fig. 9d) is consistent with the overestimation of particle
sizes and reflectivities. The observations in this case may be in a different
type of cloud environment from the data used to construct the Field
parameterisation, as suggested by the observed number concentration being
below the lower range shown in Field et al. (2007).
Moments (fourth, third, first and zeroth) of the observed particle size
distribution by the aircraft (for particles with diameters
> 100 µm) and predicted using the PSD parameterisation with the
observed ice water content (> 10-3 g m-3), temperature and
mass–diameter relationship.
![]()
Maximum reflectivity profiles and vertical velocities
In agreement with many previous studies (e.g. Blossey et al., 2007; Varble et
al., 2011), the model overestimates the reflectivity above the freezing level
as can be seen in the profiles of maximum reflectivity shown in Fig. 10, as
well as overestimating the rain reflectivities below 5 km. From the set of
simulations it can be seen that graupel is not the sole cause of the
significantly higher reflectivities, as the simulation without graupel also
displays this bias. The largest difference between simulated and observed
maximum reflectivity at 23:00–24:00 UTC occurs above 7 km and
increases with height for many of the simulations, with the difference
between the simulation with the different dynamical core (eg) and the
observations at 10 km equal to 40 dBZ. The observations show a decrease in
the maximum reflectivity with height from approximately 2 km, whereas the
simulations tend to show a more constant profile. The observed reduction in
height may be due to large raindrops falling out of strong updraughts or due to
raindrops falling through weak updraughts and growing due to the accretion of
cloud droplets. The likely overestimate in updraught strength in the
simulations (shown next) will advect the raindrops upwards, allowing these
particles to be collected by the existing ice, generating larger ice
particles and maximum reflectivities above the freezing level, as well as
acting as a source of latent heating to further fuel convective updraughts. The
simulation that decreases the maximum reflectivity with height the most is
the simulation with differing subgrid turbulent mixing (3d; Fig. 10b), which
suggests weaker updraughts. The addition of a rain heterogeneous freezing
parameterisation (qcf2rainfreeze) follows the different turbulence simulation
(3d) in reducing the maximum reflectivity from the freezing level up to
8 km, reflecting the reduction in rain and a better representation of the
reflectivities.
Profiles of maximum radar reflectivity for the times
(a) 17:00–18:00 UTC and (b) 23:00–24:00 UTC.
![]()
At 17:00–18:00 UTC, when the greatest amount of deep convection occurs in
all of the simulations and the coldest satellite-derived cloud top
temperatures are observed, the CPOL maximum reflectivity profile has a more
constant profile with a slower reduction of reflectivity with height compared to the later less convective times (Fig. 10). The observed 40 dBZ
contour reaches 8 km in agreement with the results of Zipser et al. (2006),
who showed that radar echoes of this strength rarely occur above 10 km. The
profile of maximum reflectivity from the simulation that uses the new
dynamical core (eg) shows essentially the same profile at these strong
convective times as for the later times when the MCS has matured, unlike the
observations and the majority of the simulations, suggesting that there is
less variability in maximum updraught when using ENDGame. There is little
spread in the maximum reflectivity profile across the simulations at
17:00–18:00 UTC, with strong updraughts > 20 m s-1 in all
simulations (not shown) that allow large particles to be advected into the
upper troposphere. There is a clear difference between the two simulations that
limit or exclude graupel (qcf2noqgr, qcf2sr2graupel), demonstrating that at
the time of strongest convection, the vertical advection of graupel is
responsible for the largest error in the maximum reflectivities in the upper
troposphere.
(a) Maximum vertical velocity observed by the aircraft and
derived from RASTA (Radar SysTem Airborne) for the times 23:00–24:00 UTC.
Solid lines use the highest-resolution observations, dashed lines use the observations averaged to the 1 km resolution. Modelled in-cloud
vertical velocity statistics (m s-1) over the radar domain for the
times 23:00–24:00 UTC: (b) maximum, (c) updraught mean,
(d) mean, (e) updraught 90th percentile and
(f) updraught 99th percentile.
![]()
Comparing the control case with the cases that use a different dynamical core
and different turbulent mixing parameterisation (nd, eg, 3d) shows that the
reduction in maximum reflectivity with height at 23:00–24:00 UTC is well
correlated with the reduction in maximum vertical velocity shown in Fig. 11b.
These cases all use the generic ice PSD and the differences are likely due to
the different entrainment and water loading that affects the cloud buoyancy
and the strength of the updraughts that advect large particles into the upper
troposphere. The ENDGame simulation (eg) produces significantly larger
maximum updraughts and has less accumulated ice water (see Fig. 13). Conversely
there is greater accumulated IWC for the simulation with the different
turbulent mixing parameterisation (3d) compared to the control case (nd),
supporting the argument that water loading differences likely contribute to
the differences in maximum vertical velocities and maximum reflectivities.
Ice water content (g m-3) as a function of vertical velocity
(m s-1) for four temperature regimes: (a) -5 to 0,
(b) -10 to -5, (c) -20 to -10 and
(d) -30 to -20 ∘C. Panels (e, f) show liquid water
content (g m-3) as a function of vertical velocity for the two coldest
regimes: (e) -20 to -10 and (f) -30 to
-20 ∘C.
![]()
Comparing the differences in maximum vertical velocity across the simulations
for the times 23:00–24:00 UTC shows that the largest sensitivity tends to
come from the choice of dynamics and turbulence (eg, 3d). The reduction in
updraught strength at these times with the 3-D Smagorinsky turbulence scheme is
also achieved with the inclusion of a heterogeneous freezing rain
parameterisation (qcf2rainfreeze). Both of these cases tend to have larger
ice water contents in strong updraughts (see Fig. 12), which will reduce buoyancy
through the effect of water loading. While there is different sampling
between the aircraft observations and the simulations, the aircraft
observations of maximum updraught strength shown in Fig. 11 are smaller than
the ENDGame simulation (eg) by as much as 20 m s-1. In this simulation
it seems as though the stronger and deeper updraughts are able to generate
enough latent heating that this effect on buoyancy is larger than that of
entrainment and water loading compared to the other cases. The in-cloud
mean vertical velocity for this simulation is also larger than the other
cases from 4 to 8 km, as well as the 99th percentile of upward vertical motion
(Fig. 11). The shape of the mean updraught velocity is similar for the ENDGame
case and the simulation without graupel (qcf2noqgr), both showing greater
mean updraught strength from 3 to 7 km. These two simulations produce the
largest domain mean rain rate (Fig. 3a) at these times and show that
dynamical changes to the cloud system can be achieved through changes to the
model's dynamical core and the cloud microphysics.
For the aircraft analysis region (150 km radius from the mean
aircraft track), the total accumulated water contents (kg kg-1) over
the domain from 23:00 to 24:00 UTC. (a) Cloud liquid water,
(b) rain water, (c) total ice, (d) ice
aggregates/snow, (e) ice crystals and (f) graupel.
![]()
While the maximum updraughts produced by the simulations at these times are
within the range of observed maximum tropical updraughts from other field
campaigns at Darwin (e.g. < 25 m s-1 in TWP-ICE; Varble et al.,
2014a), the maximum updraughts produced throughout the MCS life cycle are much
larger and in excess of 50 m s-1 for the ENDGame simulation (eg) at
17:00–18:00 UTC. These values are well outside the range of maximum
vertical velocities presented for oceanic convection by Heymsfield et
al. (2010) and agree with other studies showing excessive tropical vertical
velocities simulated by convection-permitting models. Hanley et al. (2014)
demonstrated that the UM with a grid length of 1.5 km simulated convective
cells that were too intense and were initiated too early, as was also shown
by Varble et al. (2014a), suggesting that convection is underresolved at
grid lengths of order of 1 km. Improved initiation time was shown by Hanley et
al. (2014) to occur when the grid length was reduced to 500 and 200 m.
However, the intensity of the convective cells was not necessarily improved,
with the results being case-dependent. Varble et al. (2014a) showed that in
the tropics the intensity of the updraughts remained overestimated even at the
100 m grid length. Both of these studies suggest that there are missing
processes in the model and/or the interactions between convective dynamics
and microphysics are incorrectly represented.
Most of the simulations show a double peak in vertical velocities with maxima
at 3 km and in the upper troposphere at about 13 km. The upper-level
updraught peak has been observed (e.g. May and Rajopadhyaya, 1999) and is
argued to be due to the deep column of convectively available potential
energy in the tropics, coupled with latent heat released by freezing
condensate and the unloading of hydrometeors, both of which increase parcel
buoyancy. A bimodal peak has been observed but tends to be correlated with
the freezing level rather than a couple of kilometres lower as in the
simulations. The apparent lack of observational support for the low-level
peak is likely due to the inability of many observations to distinguish
between non-precipitating cloud and clear air, and dual profiler measurements
during TWP-ICE do show some evidence of a low-level peak (Collis et al.,
2013).
Phase composition and comparison with in situ observations
Due to the small sample size of observations from the single research flight
on 18 February 2014, the observations from 18 of the Darwin HIWC flights have
been used to allow for a more robust comparison of the model to the
observations (Figs. 11, 12 and 14). The majority of the flight time for these
cases was in clouds with temperatures < -10 ∘C and vertical
motions within the range of -2 to 2 m s-1. Therefore, when comparing
the model to the aircraft observations, the focus is on this subset of cloud
conditions as there are limited observational samples outside of these
ranges.
Mean mass-weighted ice particle size (µm) as a function of
ice water content (g m-3) for four temperature regimes:
(a) -5 to 0, (b) -10 to -5, (c) -20 to
-1 and (d) -30 to -20 ∘C.
![]()
In the simulations, the relationship of IWC to vertical velocity changes with
the temperature regime, as shown in Fig. 12. For the warmest range of 0 to
-5 ∘C the IWC reduces as the strength of the updraught increases
from 1 m s-1. For the two intermediate temperature regimes, -5 to
-10 and -10 to -20 ∘C, the IWC is fairly constant, with
vertical velocities greater than 2 m s-1 and with the colder regime
consisting of 1 g m-3 more ice for a given vertical velocity. For the
coldest regime analysed the IWC increases as the vertical velocity increases.
For the warmest temperature regime the decline of IWC with updraught speed is
offset by the strong increase in LWC, with the fraction of condensate that is
supercooled cloud water reaching 0.8 at 15 m s-1 (not shown). In this
temperature regime there is no new ice being formed as heterogeneous freezing
in the model does not occur until the temperature cools to -10 ∘C.
Any ice in this regime has formed above and has been recirculated into these
updraughts and as the vertical velocity increases the saturation-specific
humidity increases faster than the supercooled water can be removed by
deposition and riming, resulting in the large LWC. The circulation of ice from
high levels to those below was suggested by Black and Hallett (1999) to be a
factor in the observed rapid glaciation of clouds in hurricanes. The no-graupel and limited-graupel cases (qcf2noqgr, qcf2sr2graupel) do not show the
same decline in IWC in the warmest temperature regime. For these cases the
fraction of condensate that is supercooled water is lower so there is less
competition for the available water vapour, which results in greater
depositional ice growth. In these simulations the greater proportion of ice
mass with slower fall speeds leads to greater in-cloud residence times,
producing larger accumulated IWC than the other cases with two ice
prognostics (see Fig. 13). This shows that when graupel is included in the
simulations and allowed to grow unrestricted, the removal of LWC by ice
processes is less efficient in this temperature regime. The other simulation
with different behaviour and larger IWC in this warmest regime is the case
that includes rain heterogeneous freezing (qcf2rainfreeze). In this
simulation there is an additional source of ice and this results in greater
IWC in strong updraughts due to the rain that is advected upwards, freezing
rather than remaining as liquid water as in the other simulations. The impact
of this on the cloud liquid water is to increase the cloud water content in
strong updraughts as shown in Fig. 12. This is due to the reduction in the
riming of cloud water by graupel compared to the accretion of cloud water
by rain.
The large IWC in the downdraught regions of the warmer temperature regime is
where graupel is expected, which is often located behind and below the
convective updraughts (Barnes and Houze, 2014), where the suggestion is that
these larger particles help to generate downdraughts through mass loading
(Franklin et al., 2005; Jung et al., 2012). This argument is supported by
analysis of the downdraught IWC that shows that the majority of the ice in the
downdraughts is graupel. For example in the control simulation, 82 % of the
ice mass is graupel for the warmest regime downdraught of 5 m s-1. The
simulated increase in IWC with increasing downdraught speed is observed, with
many of the simulations representing the downdraught IWC quantitatively well.
The colder regime of -10 to -5 ∘C shows IWC invariable to
vertical velocity. These colder temperatures will produce a greater
difference in saturated vapour pressure and saturated vapour pressure over
ice, therefore resulting in larger depositional growth rates via the
Bergeron–Findeisen process than the warmest temperature regime.
Compared to the warmer temperature regimes, the temperature regime of -20
to -10 ∘C shows a small increase in IWC with vertical velocity
(Fig. 12c) due to the effects of heterogeneous freezing (which occurs at
temperatures < -10 ∘C) on increasing the mass of ice and
further increases in the vapour pressure. In agreement with the observations,
the simulations increase the IWC from 0 to 5 m s-1, with the mean
modelled IWC increasing from 0.2 to about 3 g m-3, which is in good
agreement with the observed IWC. For the coldest temperature regime, the
modelled relationship of IWC to vertical velocity is represented well for
updraught strengths < 9 m s-1; however, the modelled IWC tends to be a
bit larger, particularly for the simulations that have larger sized snow
particles. The spread in IWC across the simulations is typically not
statistically significant, particularly for the stronger updraughts; however,
the differences can be attributed to the effects that the changes have on
producing and removing LWC, with different dynamics, turbulence and
microphysics all displaying sensitivities to the amount and distribution of
IWC within tropical clouds.
Across the temperature regimes the simulations show an increase in cloud LWC
with updraught strength (Fig. 12e, f), with the LWC reducing as the temperature
cools along with the fraction of condensate that is supercooled liquid water.
The strongest updraughts are associated with convective cores that will have
minimal entrainment and consequently high supersaturations. The simulations
that use the generic ice PSD (nd, eg, 3d) tend to have lower liquid water
contents for a given vertical velocity, likely due to the increased accretion
and riming growth due to the larger ice particle sizes compared to the
explicit PSD (Figs. 4 and 14). Increasing the cloud droplet number
concentration in the model (qcf2ndrop500) only directly impacts the
microphysical process of autoconversion between cloud droplets and rain, and
reduces the precipitation efficiency. For this case the reduced
autoconversion rate does not make a significant difference to the surface
rainfall, since the ice processes dominate the rainfall production (see
Fig. 3). However, the less efficient transfer of cloud water mass to rain
does change the cloud structure with more LWC and a larger amount and
fraction of condensate being supercooled water for the temperatures between
-10 and -30 ∘C (Fig. 12). As cloud water is the only liquid
water source used in the model for deposition growth via the
Bergeron–Findeisen mechanism and can freeze heterogeneously, this
implies potentially greater growth rates for ice.
The other simulation that produces more cloud water for updraughts
> 5 m s-1 in the coldest temperature regime is the simulation that
includes ice splintering or the Hallett–Mossop (H–M) process (qcf2hm; Fig. 12f).
Looking at the accumulated ice crystal mass between the simulations that do
and do not include an ice splintering parameterisation (Fig. 13, qcf2 and
qcf2hm) shows that, while there tends to be less crystal mass at most heights
when the H–M process is included, there are crystals present in updraughts up
to 15 m s-1, whereas in the qcf2 case there are no crystals present in
updraughts > 4 m s-1 (not shown). Similarly for the aggregates there
is ice spread across a wider range of updraughts when the H–M process is
included, particularly for the colder temperatures, resulting in a larger
accumulated amount of snow and total ice (Fig. 13). The generation of a
larger quantity of ice crystal mass in the H–M zone allows for a larger
amount to be transported to the upper cloud levels by the convective updraughts
where the crystals then grow through deposition, riming and aggregation,
producing a larger mass of snow.
The observed mean mass-weighted ice diameter from research flight 23 shown in
Fig. 14 increases with warmer temperatures and shows a strong dependence on
IWC, with the characteristic size decreasing with increasing IWC, reflecting
the dominance of smaller particles for higher IWC. This contrasts with the
lack of dependence of mean ice particle size on IWC which has been observed in
earlier flights over Darwin and Cayenne in 2010–2012 (Fridlind et al., 2015)
but agrees with more recent findings by Leroy et al. (2015). By analysing
research flights 12, 13 and 16, they showed that regions of high IWC over
Darwin could be generated in various environments, with the most common
result being high concentrations of small crystals, but sometimes smaller
concentrations of larger crystals. Figure 14 shows that when using all of the
Darwin research flights there is little variability in mean diameter for the
temperature range between -10 and -20 ∘C, as there were also
flights that showed an increase in mean diameter with increasing IWC (Leroy
et al., 2015). These findings show similar results to those documented by
Gayet et al. (2012), with high concentrations of ice crystals occurring in
regions of ice water content > 1 g m-3 sustained for at least
100 s at Darwin (Leroy et al., 2015) and > 0.3 g m-3 in the overshooting convection in the midlatitudes in western Europe (Gayet et al.,
2012). Gayet et al. (2012) proposed that the high concentration of ice
crystals, which appeared as chain-like aggregates of frozen drops, could be
generated by strong updraughts lofting supercooled droplets that freeze
homogeneously. However, using updraught parcel model simulations, Ackerman et
al. (2015) showed that this process produced a smaller median mass area
equivalent diameter than is observed. They proposed a number of other
possible microphysical pathways to explain the observations, including the
Hallett–Mossop process and a large source of heterogeneous ice nuclei coupled
with the shattering of water droplets when they freeze.
The modelled mean snow diameter increases with increasing temperature,
reflecting the process of aggregation; however, the modelled snow PSD also
increases the mean diameter with increasing IWC, with the rate of increase
being similar in both the generic ice PSD and the explicit specified gamma
size distribution. Note that both of the modelled PSDs generally lie within
1 standard deviation of the observations. The mean diameter from the
generic ice PSD tends to agree reasonably well with the observed size for
IWC < 0.5 g m-3; however, the sizes are significantly
overestimated for IWC > 0.5 g m-3. Given that the number
concentration is dependent on the size of the particles for a given IWC, this
implies that the generic ice PSD simulates larger concentrations of larger
particles than the observations. This reflects the data that were used to
develop the generic ice PSD coming largely from stratiform clouds with
smaller IWC and larger ice particles. The explicit gamma PSD shows the
opposite behaviour, underestimating the mean ice diameter for
IWC < 0.5 g m-3 and matching the observed size for higher IWC. To
more accurately represent the snow sizes in the model for this case
a double-moment microphysics scheme is required to be able to better capture the observed
variability of the PSD or the use of a wider data set that includes high IWC
observations to generate a more applicable generic ice PSD parameterisation
for modelling tropical convective cloud systems.
Conclusions
A set of 1 km horizontal grid length simulations has been analysed to
evaluate the ability of the UM to simulate tropical convective cloud systems
and to investigate the impacts of different dynamical, turbulent and
microphysical representations on the cloud properties, including the phase
composition. The case study is 18 February 2014, where active monsoon
conditions produced a mesoscale convective system in the Darwin area.
Analysing 12 h of observed and simulated radar reflectivity has shown that
the simulations capture the intensification and decay of convective strength
associated with the life cycle of the MCS. However, convection occurs too
early in the simulations, the radar detectable cloud top heights are
overestimated, as are the maximum reflectivities and areas above the freezing
level with reflectivities greater than 30 dBZ. The observed maximum domain
averaged precipitation rate coincides with the generation of significant
anvil cloud, whereas the simulations generate the highest mean precipitation
rate a few hours too early at the times of deepest convection. Observations
of maximum vertical velocity suggest that the new dynamical core simulation
(eg) overestimates the strength of convection at the mature-decaying stage of the
MCS. In this case the stronger updraughts contribute to the excessive
reflectivities above the freezing level, but this was apparent in all of the
simulations albeit to a lesser degree, suggesting that both the updraught
dynamics and the particle sizes are responsible for this error.
The simulated reflectivity CFADs show more of a convective-type profile
compared to the observations, with broader distributions and a greater
occurrence of high-reflectivity outliers. This suggests a larger number of
convective cells in the simulations, as was apparent in the plan views of OLR
and 2.5 km radar reflectivity, which has been seen in tropical
convective-scale model intercomparison studies (e.g. Varble et al., 2014a).
The simulation with the differing turbulence parameterisation showed the best
agreement with the observed maximum reflectivity at the later times of
23:00–24:00 UTC. The change to the 3-D Smagorinsky scheme induces greater
mixing, resulting in a reduction of the maximum vertical velocities and
reflectivities during the mature decaying MCS stages. This same reduction in
the vertical velocity and reflectivity up to 8 km was also found with a
change to the microphysics formulation with the addition of a rain
heterogeneous freezing parameterisation. At 17:00–18:00 UTC, the time of
deepest convection, all simulations showed a similar error in maximum
reflectivity regardless of dynamics or turbulence formulation due to the
larger and less variable maximum updraughts across all of the simulations at
these times.
The largest sensitivities in the maximum updraught velocities are generally
produced by changes to the dynamical and turbulence formulations in the
model. However, the spread across the simulations for the mean and
percentiles of updraught velocity show the greatest sensitivity coming from
changes to the microphysical parameters and processes. Changing the
microphysics affects the dynamics by altering the vertical distribution of
latent heating. The horizontal mass divergence was shown to be most
sensitive to the turbulence parameterisation in the mixed-phase regions of
the updraughts, where greater mixing generated larger mass divergence,
indicative of greater entrainment at these heights. The upper ice-only
regions of the convective updraughts showed that the control on updraught
buoyancy was the size of the ice particles. Simulations with smaller
particles have fewer occurrences of positively buoyancy convective updraughts,
reflecting the importance of the microphysical processes on the convective
dynamics.
Analysing the relationship between phase composition and vertical velocity
for four different temperature regimes showed that the phase composition in the
modelled convective updraughts is controlled by the following:
the size of the ice particles, with larger particles growing more
efficiently through riming, producing larger IWC;
the efficiency of the warm rain process, with greater cloud water
contents being available to support larger ice growth rates;
exclusion or limitation of graupel growth, with more mass contained in
slower falling snow particles resulting in an increase of in-cloud residence
times and more efficient removal of LWC.
The evaluation of a tropical mesoscale convective system in this study has
documented a number of model shortcomings and developments that improve the
model performance.
Excessive areas with high reflectivities improve with reduced ice sizes,
inclusion of a heterogeneous freezing rain parameterisation, an additional
ice prognostic variable and increased turbulent mixing through the use of the
3-D Smagorinsky turbulence scheme.
Too much rain above the freezing level is reduced with the inclusion of a
heterogeneous rain-freezing parameterisation.
Too little stratiform cloud and rain area (Fig. 6, Sect. 3.1.1) is
increased with increased turbulent mixing.
While the listed model changes do improve aspects of the simulations, none of
these produce a simulation that closely matches all of the observations. This
study has shown the need to include a better representation of the observed
size distribution, which could be achieved through the use of a double-moment
microphysics scheme. Being able to predict both the number concentration and
mass would allow the model to better represent the observed variability of
the PSD, which would impact the model's representation of the ice water
contents and reflectivities, as well as the convective dynamics through the
effects of latent heating and water loading on buoyancy.