turbulence, rain drops and the l number density lagang/eprints/eprintlovejoy/... · 11/29/07 1...
TRANSCRIPT
11/29/07 1
Turbulence, rain drops and the l1/2 number density law
S. Lovejoy1, D. Schertzer2
1Department of Physics, McGill University, 3600 University st., Montreal, Que., Canada
2CEREVE, Ecole Nationale des Ponts et Chaussées, 6-8, avenue Blaise Pascal, Cité Descartes, 77455
MARNE-LA-VALLEE Cedex, France
Abstract:
We treat rain drops as low Stokes number particles in a highly turbulent flow and we derive
equations for the number and mass densities (n, ρ) and their variance fluxes (ψ, χ). By showing
that χ, ψ are only dissipated at small scales, we conclude that ρ should follow Corrsin-Obukhov k-5/3
spectra and that n should follow a k-2 spectrum corresponding to fluctuations Δρ, Δn at scale l
scaling as l1/3 and l1/2 respectively. We directly verify these predictions as well as many of the
necessary assumptions using a unique data set of three dimensional drop positions and masses from
the HYDROP experiment. While the Corrsin-Obukhov law has never been observed in rain before,
it’s discovery is perhaps not surprising; in contrast the Δn ≈ l1/2 number density law is quite new.
The key difference between the Δρ, Δn laws is that the time scale for the transfer of χ is determined
by the strongly scale dependent turbulent velocity whereas the time scale for ψ is determined by the
weakly scale dependent drop coalescence speed. The combination of number and mass density
laws can be used to develop stochastic compound multifractal Poisson processes which are useful
new tools for studying and modeling rain.
We discuss the implications of this for the rain rate statistics including a simplified model
which can explain the observed rain rate spectra.
1. Introduction
Rain is a highly turbulent process yet there is wide gap between turbulence and precipitation
research. On the one hand turbulence is increasingly viewed as a highly intermittent, highly
1 Corresponding author: Department of Physics, McGill University, 3600 University st., Montreal, QC, Canada, H3A
2T8,
e-mail: [email protected]
11/29/07 2
heterogeneous process with turbulent energy, passive scalar variance and other fluxes concentrated
into a hierarchy of increasingly sparse fractal sets; over wide ranges the fields are multifractal (see
e.g. [Anselmet, 2001] for a recent review). Furthermore, advances in high powered lidars have
produced turbulent atmospheric data sets of unparalleled space-time resolutions. Analysis of such
data from aerosols have shown that if classical Corrsin-Obukhov theory of passive scalar turbulent
advection is given multifractal extensions to account for intermittency and scaling anisotropic
extensions to account for atmospheric stratification and scaling wave-like space-time extensions to
account for wave behaviour, that these rejuvenated classical theories account remarkably well for
the statistics of passive pollutants ([Lilley, et al., 2004], [Lovejoy, et al., 2007a], [Lilley, et al.,
2007], [Radkevitch, et al., 2007]). In contrast, applications of turbulence theory either to
interpreting radar echoes, or to disdrometer experiments almost invariably assume that the
turbulence is uniform so that rain has either homogeneous Poisson rain statistics (e.g. [Marshall and
Hitschfeld, 1953], [Wallace, 1953]), or that it is only weakly heterogeneous (e.g. [Uijlenhoet, et al.,
1999], [Jameson and Kostinski, 1999]).
Combining turbulence theory with rain drop physics involves several related difficulties. First,
rain is particulate. This is usually handled by considering volumes large enough so that many
particles are present and spatial averages can be taken. However since rain is strongly coupled to
the (multifractal) wind field even these supposedly continuous average fields turn out to be
discontinuous ([Lovejoy, et al., 2003]). This means that due to the systematic, strong dependence
on the scale/resolution over which the rate is estimated, the classical treatment of the rain rate R(x,t)
as a mathematical space-time field (without explicit reference to its scale/resolution) is not valid.
Finally rain does not trivially fit into the classical turbulence framework of passive scalars: rain
simultaneously modifies the wind field while moving with speeds different than that of the ambient
air. However the treatment of particles in flows has seen great progress especially with the
pioneering work of [Maxey and Riley, 1983], [Maxey, 1987] relating particle and flow velocities.
Unfortunately with the main exception of [Falkovitch and Pumir, 2004], [Falkovitch, et al., 2006]
these particle/fluid dynamics advances have not been applied to cloud or rain physics. While the
latter authors consider the case with Stokes numbers (St) ≈1 (applicable to cloud drops), in this
paper we develop the theory for rain particles with St <<1 and put the results in the context of
turbulence theory including multifractal cascades.
[Falkovitch and Pumir, 2004], [Falkovitch, et al., 2006] have admirably attacked the full
drop/turbulence interaction at its most fundamental level, but progress has been difficult.
Fortunately, at a more phenomenological level, things have been easier. [Schertzer and Lovejoy,
11/29/07 3
1987] proposed that even if rain is not a passive scalar that it nevertheless has an associated scale-
by-scale conserved turbulent flux. This proposal was made in analogy with the energy and passive
scalar variance fluxes based on scale invariance symmetries but it lacked a more explicit, detailed
justification. It nevertheless leads to a coupled turbulence/rain cascade model which predicts
multifractal rain statistics over wide ranges of scale. Today, this cascade picture has obtained wide
empirical support, particularly with the help of recent large scale analyses of global satellite radar
reflectivities which quantitatively show that from 20,000 down to 4.3 km, that the reflectivities (and
hence presumably the rain rate) are very nearly scale by scale conserved fluxes [Lovejoy, et al.,
2007b]. These global satellite observations show that cascade behaviour is followed to within
±4.6% over nearly four orders of magnitude. Other related predictions - based essentially on the
use of scaling symmetry arguments - are that rain should have anisotropic (especially stratified)
multifractal statistics (see e.g. [Lovejoy, et al., 1987] and that rain should belong to three -parameter
universality classes ([Schertzer and Lovejoy, 1987], [Schertzer and Lovejoy, 1997]), see the review
[Lovejoy and Schertzer, 1995]). However these empirical studies have been at scales larger than
drop scales and outstanding problems include the characterization of low and zero rain rate events
and the identification of the conserved (cascaded) flux itself (see however [Lovejoy, et al., 2007b]).
In other words, up until now the connection with turbulence has remained implicit rather than
explicit.
In order to bridge the drop physics - turbulence gap and to further pursue phenomenological
approaches, data spanning the drop and turbulence scales are needed. Starting in the 1980’s,
various attempts to obtain such data have been made; this includes experiments with chemically
coated blotting paper ([Lovejoy and Schertzer, 1990]) and lidars, ([Lovejoy and Schertzer, 1991]).
The most satisfactory of these was the HYDROP experiment ([Desaulniers-Soucy., et al., 2001])
which involved stereophotography of rain drops in ≈10 m3 volumes typically capturing the position
and size of 5,000 - 20,000 drops (nominal rain rates were between 2 and 10mm/hr; see fig. 1 for an
example, and table 1 and [Lilley, et al., 2006] for information about HYDROP). Analyses to date
([Lovejoy, et al., 2003], [Lilley, et al., 2006]) have shown that at scales larger than a characteristic
scale determined by both the turbulence intensity and the drop size distribution (but typically
around 20-30 cm see e.g. fig. 2), that the liquid water content (LWC; i.e. the mass density), number
density and other statistics behave in a scaling manner as predicted by cascade theories. While
these results suggest that rain is strongly coupled to the turbulent wind field at scales larger than 20-
30 cm (potentially explaining the multifractal properties of rain observed at much larger scales), the
11/29/07 4
analyses did not find an explicit connection with standard turbulence theory. See also [Marshak, et
al., 2005] for similar results for cloud drops.
Fig. 1: Second scene from experiment f295. For clarity only the 10% largest drops are shown,
only the relative sizes and positions of the drops are correct, the colours code the size of the drops. The boundaries are defined by the flash lamps used for lighting the drops and by the depth of field of the photographs.
Perhaps the key new result of this paper concerns the fluctuations in the particle number density
Δn which we find - both theoretically and empirically - follows a new Δn ≈ l1/2 law with prefactors
depending on turbulent fluxes. This explicit link between the particle number density fluctuations
and turbulence is expected to hold under fairly general circumstances and provides the basis for
constructing compound multifractal –Poisson processes. While the traditional approach to drop
modeling (see e.g. [Srivastava and Passarelli, 1980], [Khain and Pinsky, 1995]) is to hypothesize
specific parametric forms for the drop size distribution and then to assume spatial homogeneity in
the horizontal and smooth variations in the vertical, our approach on the contrary assumes extreme
turbulent - induced variability governed by turbulent cascade processes and allows the drop size
distributions to be defined only implicitly with their essential variations being due to turbulence.
11/29/07 5
This paper is structured as follows. In section 2 we carefully discuss the relation between drop
and wind speeds showing that rain drops typically have low Stokes numbers so that Maxey’s
relations can be used and we directly estimate the mean relaxation times, lengths and speeds from
the HYDROP experiment. In section 3 we derive the basic equations for number density, mass
density, the corresponding fluxes density (including rainrate), and the coalescence speed. In section
4 we derive the l1/3, l1/2 laws for mass and number densities and verify the results empirically on the
HYDROP data, in section 5 we conclude.
2. The connection between drops and turbulence
2.1 Drop Relaxation times, distances, velocities:
Raindrops are inertial particles moving in turbulent flows under the influence of gravity. For an
individual drop moving with velocity v in a wind field u, Newton’s second law implies:
dv
dt= g +
g
VR!v
!vVR
"
#$%
&'
(d )1
; !v = u ) v (1)
where we have taken the downward acceleration of gravity (g) as positive, and assumed a power
law drag with exponent ηd. VR(M) is the “relaxation velocity” of a drop mass M. In still air, after
an infinite time, VR is the theoretical terminal drop velocity (we use capitals for single drop
properties, lower case for various averages). The corresponding relaxation times TR(M) and
distances LR(M) are given by:
TR=VR
g
LR=VR
2
g
(2)
These are the typical time and length scales over which the drops adjust their velocities to the
ambient air flow. According to classical dimensional analysis for drag on rigid bodies, the low
Reynold’s number regime is independent of fluid density while the high Reynold’s number regime
is independent of fluid viscosity. This implies that in the first regime, the drag is linear (ηd = 1;
“Stokes flow”), while in the second, it is quadratic (ηd = 2) so that the drag coefficient (CD) is
constant. For these regimes we have:
TR=L2!
w
"v
; LR=gL
4!w
2
"v
2; V
R=gL
2!w
"v
; Re << 1 (3)a
11/29/07 6
TR=
!DL"
w
g"a
#
$%&
'(
1/2
; LR= !
D
"w
"a
#
$%&
'(L; V
R=
!DgL"
w
"a
#
$%&
'(
1/2
; Re >> 1 (3)b
where L is the drop diameter, ρw, ρa are the densities of water and air !D= " / 3C
D( ) , CD is the
standard drag coefficient and !v is the dynamic viscosity. In the Stokes flow, we have ignored
numerical factors of order unity. Since at 20oC, pressure = 1 atmosphere, !w/ !
a" 1.29X103 , for
drops with diameters 0.2 < L < 4 mm (the range detected by the HYDROP experiment discussed
below), we expect LR to be typically of the order 20 cm – 4 m.
According to theory and experiment reported in [Le Clair, et al., 1972], for spheres, the high
Re limit is obtained roughly for Re>102 where CD≈1 (although it decreases by another factor of 2-3
as Re increases to ≈ 5000). According to semi-empirical results presented in [Pruppacher, 1997],
Re ≈ 1 is attained for drops with diameter (L) ≈ 0.1mm, whereas Re = 102 is attained for drops
diameter 0.5mm. On this basis, Pruppacher proposes breaking the fall speed dynamics into three
regimes with (roughly) L <0.01 mm, 0.01<L<1 mm, L>1 mm. The small L “Stokes” regime has
ηd = 1, the second regime is intermediate and the third (high Re) regime has roughly ηd = 2 up to a
“saturation” at L ≈ 4mm. Although for rain, this regime is more complicated than for rigid spheres
due to both drop flattening and due to internal drop flow dynamics, according to data and numerics
reviewed in [Pruppacher, 1997], the basic predictions of the ηd = 2 drag law (VR! L
1/2 ,
LR! "
w/ "
a( )L ) are reasonably well respected over the range 0.5 – 4 mm. For example, using
Pruppacher’s data on LR at 700 mb and a drag coefficient of CD = 0.8, the predictions of eq. 3 are
verified to within ±25% over this range.
2.2 Empirical estimates of mean relaxation lengths using drop stereophotography:
According to the above analysis, we expect the transition from rain drop behaviour dominated by
turbulence to behaviour dominated by inertia to occur at LR; the latter being roughly a linear
function of drop size and of the order of 1 m. Before discussing other aspects of turbulence theory,
we give the first direct estimates of the average LR in rain using the HYDROP data.
Fig. 2 shows the 3D isotropic (angle averaged) spectrum of the 19 stereophotographic drop
reconstructions averaged over each of the five storms. Each storm had 3 - 7 “scenes” (from
matched stereographic triplets) with 5,000 – 20,000 drops each taken over a 15 - 45 minute period
(see table 1). The low wavenumber reference line indicates the theoretical Corrsin-Obukhov k-5/3
angle integrated spectrum. Since Gaussian white noise has constant spectral energy density, the
11/29/07 7
integrated spectrum varies as k2 and is indicated by the high wavenumber reference line. It can be
seen that the transition between passive scalar behaviour - where the drops are highly influenced by
the turbulence and the high wavenumber regime, where the drops are totally chaotic (white noise)
occurs at scales of roughly 40 – 75 cm.
More details on the HYDROP experiment can be found in [Desaulniers-Soucy., et al., 2001] and
for these data sets see [Lilley, et al., 2006]. The main differences between the part of the data
analyzed in the latter and below is that in the latter, a very conservative choice of sampling volume
was used. By using only data from the region best lit and most sharply in focus, >90% of the drops
with diameter >0.2mm were identified. In this study, we sought to get a somewhat wider a range of
scales with as many drops as possible by exploiting the fact that fourier techniques are very
insensitive (except at the lowest wavenumbers) to slow falloffs in the sensitivity (and hence drop
concentrations) near the edges of the scene (this is a kind of empirically produced spectral
“windowing” close to the numerical filters used to reduce leakage in spectral estimates). We
therefore took all the data in a roughly rectangular region about 4.4X4.4X9.2 m in size (geometric
mean = 5.6 m), and then analyzed only the spectrum for wavenumbers k ≥ 2 (i.e. spatial scales ≤ 2.3
m). This somewhat larger scene size with respect to the previous HYDROP analyses yielded an
increase in the range of scales by a factor of about 2.8 and about 2-3 times more drops.
11/29/07 8
Fig. 2: This shows the 3D isotropic (angle integrated) spectrum of the 19 stereophotographic drop
reconstructions, for ρ, the particle mass density. Each of the five storms had 3 - 7 “scenes” (from matched stereographic triplets) with ≈ 5,000 – 40,000 drops (see table 2) each taken over a 15 - 30 minute period (orange = f207, yellow = f295, green = f229, blue green = f142, cyan = f145). The data were taken from regions with roughly 4.4 m X 4.4 m X 9.2 m in extent (slight changes in the geometry were made between storms). The region was broken into 1283 cells (3.4 cm X 3.4 cm X 7.2 cm); we use the approximation that the extreme low wavenumber (log10k=0) corresponds to the geometric mean, i.e. 5.6 m, the minima correspond to about 40 - 70 cm; see table 2). The single lowest wavenumbers (k = 1) are not shown since the largest scales are nonuniform due to poor lighting and focus on the edges. The reference lines have slopes -5/3, +2, the theoretical values for the Corrsin-Obukhov (l1/3) law and white noise respectively.
11/29/07 9
Storm 207 295 229 142 145 Number of Triplets/scenes
7 3 3 3 3
Wind speed at 300m (in m/s)
27.5 10 17.5 2.5 22.5
Nominal Rain rate (mm/hr)
6-10 1.4-2.2 2-4 2-4 1.4-2.2
Lmin (m) 0.49±0.08 0.53±0.09 0.44±0.06 0.75±0.11 0.59±0.07
L (mm) 1.29±0.12 1.24±0.06 1.01±0.02 0.98±0.09 1.36±0.07 Number of drops
19300±13100 21400±1400 33000±8200 34800±9800 13000±2500
mean LWC at 70cm scale (g/m3)
0.50±0.63 0.44±0.42 0.39±0.57 2.40±1.35 0.44±0.58
maximum LWC at 70cm scale (g/m3)
3.89 1.65 2.48 5.93 2.16
Wind at 300 m height (m/s)
27.5 10 17.5 2.5 22.5
Mean vR,n
at 70cm resolution in m/s
1.26±0.12 1.32±0.03 1.23±0.06 1.10±0.04 1.17±0.09
Mean vR,! at
70cm resolution in m/s
1.34±0.13 1.40±0.04 1.33±0.03 1.55±0.004 1.27±0.09
Table 1: Comparison of scale of the spectral minimum (Lmin) estimated from the drop mass spectra with the corresponding mean drop diameters; this is an estimate of the mean relaxation length. Fig. 3a shows the individual values, fig. 3b shows the generally small effect of weighting the spectrum to smaller or to larger drops. The spread is the storm-to-storm variability based on 3 triplets per storm (the exception being storm 207 for which there were 7 triplets). As expected from theory, the relaxation time is not dependent of the wind/turbulence level which varied considerably (the data at 300 m altitude are from a UHF windsonde 250 m south of the experiment). Mean (of the 19 triplets) diameter = 1.19±.17 mm (i.e. ±14%; spread is triplet to triplet variation, not standard deviation for each triplet), mean relaxation distance =0.56±0.13 m (i.e. ±23%), mean number of drops 23200 ± 11800 (i.e. ±51%). These numbers are larger than those in [Lilley, et al., 2006] since all the reconstituted drops were used, not only those in the more reliable central region. The LWC statistics is for the well-lit central region only, averaged at 70cm scale (roughly the relaxation scale). To obtain the ratio of LWC to air density, divide by ρa = 1290 g/m3 for air.
11/29/07 10
Fig. 3a: The mean drop diameter L versus the ratio of the mean (mass density based, η = 1) relaxation length lmin (estimated by the spectral minimum), to L for the 19 triplets. The overall mean ratio l
min/ L =
2.8±0.8X102. This result shows that if lmin is used to estimate the mean relaxation length: lR ≈ lmin then the
theoretical formula (eq. 3b) lR!
"3C
D
#
$%&
'()w
)a
#
$%&
'(L is valid to within ±28% for the 19 triplets with CD=2.90.
Fig. 3b: For each storm, this shows the mean distance (in m) corresponding to the minimum of the
spectrum of ρη as a function of the order of moment η (from top to bottom on the right hand side: 142, 145, 295, 207, 229). Since higher η values weight the spectra to the larger drops so that there is a slow increase of
11/29/07 11
Lmin with increasing η. The main exception is the 142 case which is the most dominated by small drops and shows a near doubling in Lmin when comparing the number density (η=0) with the mass density (η = 1).
2.3 Relations between Drop and wind velocities:
In regimes of “weak turbulence” where Du
Dt<< g , and power law drag (ηd >0) we may follow
[Maxey, 1987], [Maxey and Riley, 1983], [Falkovitch and Pumir, 2004], [Falkovitch, et al., 2006]
(who considered ηd = 1), and obtain an expansion for the drop velocity v in terms of the wind
velocity u:
v ! u + TR g "Du
Dt
#$%
&'(+O TR
2D2u
Dt2
#$%
&'(
(4)
although he only investigated the η d = 1 case in detail [Maxey, 1987] pointed out the insensitivity of
this result to the exact form of the drag law (up to first order, it is independent of ηd; note that we
use the notation D / Dt = ! / !t + u "# for the Lagrangian derivative of the wind). We now consider
the limits of validity of this approximation in a turbulent flow when it is applied to eddies at the
mean relaxation scale lR. We should note that this equation implies that even if the wind field is
incompressible, that the rain velocity field is compressible:
! " v =VR
g! " u "!u( ) =
VR
g
# 2
2$ s2
%&'
()*
(5)
where ω is the vorticity, and s is the strain rate [Maxey, 1987].
We now check that eq. 4 is indeed justified for raindrops. According to fig. 2, for l>lR, the
spectrum is nearly that of the theoretical Corrsin-Obukhov passive scalar form so that the standard
turbulence estimate of derivatives at scale l can be used:
Du
Dtl
n
! "1/2#e
1/2$n! %u
l#e
$n; #
e= l
2 /3"$1/3 (6)
where τe is the eddy turn over time (i.e. eddy lifetime) for eddies of size l. Applying this in eq. 4
we obtain:
v ! u + VR
!z " St#u
l( ) +O St2( )#ul ; St =
TR
$e,R
(7)
11/29/07 12
where St is the Stokes number evaluated at the eddy turn over time: !e,R
= LR
2 /3"#1/3 . We see that as
long as St<1, the series converges. To see how strong a constraint this is, we can use the measured
mean relaxation distances from section 2.2 and eqs. 2, 6 to obtain:
St =!R
" e=
LR
g
#$%
&'(
1/2
l)2 /3*1/3 (8)
Note that due to the weak LR dependence above and (roughly linear) dependence of LR on drop
diameter L (c.f. eq. 3b) that the use of a single average/effective lR for all drops may be a reasonable
approximation.
The criterion St<1 also ensures that the eddy scale turbulent accelerations are much less than the
acceleration of gravity; this can be seen by using:
1
g
Du
Dt
!"#
$%&lR
' g(1)1/2* R(1/2
= St( )3/2 (9)
In section 3.3 we discuss direct evidence from drop sondes that the vertical wind accelerations at
scales 5 - 10 m rarely exceed 2% of g. The critical Stokes number St = 1 is thus obtained for eddies
of critical size lc:
lc =LR
g
!"#
$%&
3/4
'1/2 (10)
Using the overall mean value lR ≈ 0.56 m, (TR = 0.24 s, vR = 2.34 m/s; these are the means of
VR(M), ΤR(M)) and the atmospheric mean ε (which is often estimated at 10-4m2s-3), we find lc≈1mm.
According to theoretical considerations in [Wang and Maxey, 1993] and experimental results in
[Fessler, et al., 1994], the Stokes number which has the maximum effect in causing preferential
concentration of particles occurs at St = 1, we thus conclude that at the drop collision scales (≈1mm
here) that the drops are quite decoupled from the wind field, a fact that we use here.
In the model developed below, we use eq. 4 for drop velocities only down to LR, this will be
justified only if St << 1 at that scale. Applying eq. 10 to eddies at the relaxation length, (setting
LR=lc) we obtain:
St =!R
" e,R
= #1/3g$2 /3
!R
$1/3= #
1/3g$1/2LR$1/6 (11)
Hence in terms of the measured mean lR:
St < 1! " < g3/2lR1/2 (12)
Using the above value for the mean relaxation length (inferred from the spectra) lR≈0.56m, we
obtain a limit ε < 25 m2 s-3, which is a very large value for atmospheric turbulence. For example,
11/29/07 13
using standard estimates for multifractal intermittency parameters (e.g. [Schmitt, et al., 1996]), and
taking the scale ratio 107 (planetary scale/ relaxation scale), we find that the probability of
exceeding this “weak turbulence” limit is of the order 10-9). Alternatively, taking the estimate 10-4
m2 s-3 for the mean ε, we find that the eddy turn over time at the mean relaxation length scale is ≈ 15
s so that St ≈ 1.6X10-2.
3. The basic equations:
3.1 The drop mass distribution function N :
Having established that at the relaxation scales, the Stokes number is typically small we can
systematically use eq. 4 to relate the turbulent wind field to the drop velocities. To do this, we first
introduce N = N M,x,t( ) which is the drop mass distribution function i.e. the number of drops with
mass between M and M+dM per unit volume of space at location x, time t. Ignoring condensation,
the usual coalescence (Smolukowski) equation (used for example in cloud and rain modeling
[Srivastava and Passarelli, 1980]) can be written:
!N
!t= N " N (13)
The right hand side term is the coalescence operator in compact notation (see [Lovejoy, et al.,
2004]), N1! N2 :
N1! N2 =
1
2! M " #M ,M,x,t( )N
1M " #M ,x,t( )N
2#M ,x,t( )d #M
0
M
$
-N1
M,x,t( ) ! #M ,M,x,t( )N2
#M ,x,t( )d #M
0
%
$
(14)
If we now assume that the drop-drop collision mechanism is space-time independent then the full
coalescence kernel H can be been factored H !M ,M,x,t( ) = ! x,t( )H !M ,M( ) so that, space-time
variations in the coalescence rate are accounted for by ϕ which is the coalescence speed; we return
to this in section 3.4. The budget equation for N , is thus:
!N
!t= "# $ Nv( ) + ! x,t( ) N H N (15)
The validity of this turbulence-drop coalescence equation can be verified by integrating over a
volume; the left hand side is the change in the volume of particles between M and M+dM, the first
right hand side term is the flux of such particles across the bounding surface, the second is the
change in the number within the volume due to coalescence. In this equation, v(M,x,t) is the
11/29/07 14
velocity of a particle mass M. While this equation was invoked in [Falkovitch and Pumir, 2004],
only the static v=0 case was studied. Eq. 15 is also used in conventional drop size distribution
modeling but typically with the additional assumption of horizontal homogeneity (only smooth
vertical variability is considered).
3.2 The mass and number densities and fluxes:
We now consider the first two moments of the turbulence-drop coalescence eq. (15); the number
density (n) and drop mass density (ρ), as well as their fluxes: the mass flux r and the number flux
! :
n x,t( ) = N M , x,t( )dM
0
!
" ; # x,t( ) = N M , x,t( )MdM0
!
" (16)
! = v M , x,t( )N M , x,t( )dM! (17)
r = Mv M , x,t( )N M , x,t( )dM! (18)
Note that the vertical component of r is the usual rain rate; R=rz. (see section 3.4).
With r, ! we can obtain the budget equations for the particle number and mass densities, by
respectively integrating eq. 15 with respect to M and by multiplying eq. 15 by M and then by
integrating to obtain:
!n
!t= "# $! +% N H N (19)
!"
!t= #$ % r (20)
where we have used the fact that the coalescence operator conserves drop mass:
N H NM = 0 (21)
but not number density n. We see that for the ρ equation (20), coalescence is not directly relevant
whereas it is relevant for n.
We now introduce the average relaxation velocities vR,n
, vR,! weighted by number density
and mass density respectively:
vR,n
=1
nVRM( )N M( )dM! (22)
vR,! =
1
!VRM( )N M( )MdM" (23)
11/29/07 15
We obtain:
! = n!n; !
n= u + v
R,n
"z "
1
g
Du
Dt
#$%
&'(
(24)
r = !"!; "! = u + vR,!!z #
1
g
Du
Dt
$%&
'()
(25)
so that αn, αρ are the effective velocities of the drops with respect to the wind.
Fig. 4 This shows scatterplots for the 70 cm resolution estimates of v
R,! versus vR,n
for the five storms (yellow =207, green =295, blue =229, purple =142, red = 145). The points clustered along the bisectrix have a single drop in the averaging volumes.
It is of interest to consider the spatial variability of vR,! , v
R,n and also how they are related
to each other. Fig. 4 shows the scatter plots obtained from the HYDROP data by using the high Re
theoretical relaxation formula (eq. 3b) with the empirically determined mean drag coefficient CD =
2.90 (fig. 3a). Table 1 shows the storm by storm mean as well as the spread between scenes within
the same storm, averaging over a 70 cm resolution, i.e. at about the relaxation scale. From the
table, we can see that the mean over each storm is fairly stable, whereas from fig. 4 we see that the
11/29/07 16
variability is very large. Although there is clearly no one to one relation between vR,! and v
R,n the
following relation is roughly satisfied: vR,n
= avR,!
b with a = 1.17±0.09, b = 0.70±0.10 (the standard
deviations being the storm to to storm spread in values). This shows that vR,! and v
R,n cannot be
taken to be equal. To get an idea of the spatial variability (important below), we also calculated the
mean horizontal spectra ofvR,! , v
R,n (fig. 5a) although the range of scales is small, the figure shows
that there is evidence that the latter has a roughly Corrsin-Obukhov spectrum at low wavenumbers.
While this result follows if ρ has a Corrsin-Obukhov spectrum (since the vR’s are powers of ρ; see
section 4.4, fig. 8c), note that it is contrary to the usual assumption that vR
is horizontally
homogeneous (being determined by a spatially homogeneous drop size distribution).
To investigate the variability in the vertical, we determined the spectra of the vertical
gradients of vR,! , v
R,n (using finite differences estimates of !v
R,n/ !z , !v
R," / !z ), these are needed
in section 3.3) and are shown in fig. 6a, b. We see that in both cases the spectrum shows a slight
tendency to rise at the higher wavenumbers; if the spectrum is E(kz) ≈ kz-β, then roughly β ≈ -0.3.
Recall that that whenever β<1, that the variance is dominated by the small scales/large
wavenumbers (an “ultra violet catastrophe”). Since the spectrum of !vR/ !z is kz
2 times the
spectrum of vR
, this is consistent with a near Kolmogorov value β ≈ 1.7 in the vertical. Finally, we
find that the spectrum of !vR
2/ !z which has nearly the same β as !v
R,n/ !z (the small differences
are due to intermittency corrections).
Fig. 5a: Horizontal Spectra of the number weighted mean relaxation velocities vR,n with highest wavenumber = (70 cm)-1 showing a roughly k-5/3 (Kolmogorov) spectrum at low wavenumbers (see black reference line), flattening out near the mean relaxation scale at the right. Yellow=207,
Fig. 5a: Same but for the density weighted mean relaxation velocities vR,ρ.
11/29/07 17
green=295, blue=229, purple=142, red= 145.
Fig. 6a: Spectra of the vertical gradients of the number weighted mean relaxation velocities !v
R,n/ !z , showing their slowly increasing
character up to the mean relaxation scale (the maximum at the far right). Yellow=207, green=295, blue=229, purple=142, red= 145.
Fig. 6a: Same but for the density weighted mean relaxation velocities !v
R," / !z .
3.3 The mass and number fluxes:
To put the mass flux vector into a more useful form, we can appeal to the dynamical fluid
equations:
!a
Du
Dt" g = "#p +$#2
u + Fr
# %u = 0
(26)
Where we have used the incompressible continuity equation for u adequate for our purposes, (see
[Pruppacher, 1997]) and Fr is the reaction force of the rain on the wind. The reaction force per
volume of the rain on the air is:
Fr= ! g "
dv
dt
#$%
&'() ! g "
Du
Dt
#$%
&'(+ !
vR,!
g
D2u
Dt2
(27)
plus higher order corrections. We therefore obtain:
g !Du
Dt=
1
" + "a
#p !$#2u ! "
vR,"
g
D2u
Dt2
%&'
()*
(28)
11/29/07 18
We can now note that the mean relaxation scale lR is typically much greater than the turbulence
dissipation scale ldiss so that at the scale lR, we can neglect the viscous term. In addition, since
St<<1 at lR, we can neglect the D2u
Dt2
term. We therefore obtain the approximation:
g !Du
Dt"
1
# + #a
$p (29)
and hence using this in eqs. 24, 25 we obtain:
!n" u +
vR,n
g # + #a( )$p (30)
!"# u +
vR,"
g " + "a( )$p (31)
Finally if we make the hydrostatic approximation ( !p = g " + "
a( )!z ), we see that this reduces to
a simple form:
!
n" u + v
R,n
!z (32)
!
"# u + v
R,"
!z (33)
i.e. the effective velocities are equal to the wind velocities plus the appropriate mean relaxation
speeds. To judge the accuracy of the hydrostatic approximation, we can consider the dimensionless
difference 1! 1
g"a
#p
#z which in non precipitating air at scales above the viscous scale, should be
equal to the dimensionless vertical acceleration: !1
g
Dw
Dt. Although we did not have measurements
during HYDROP, this difference has been evaluated empirically using state of the art drop sondes
with roughly 5 - 10 m resolutions in the vertical during the “NOAA Winter Storms 04” experiment
(north Pacific ocean). It was found that the typical mean value at the surface was about 0.02 with
fluctuations of the order ±0.005 so that, even at 10 m scales we may use the hydrostatic
approximation to reasonable accuracy. In addition, at the higher wavenumbers, the vertical
spectrum of 1! 1
g"a
#p
#z$1
g
Dw
Dt (fig. 7) increases nearly linearly with wavenumber showing its strong
dependence on the small scales. This suggests that the (notoriously difficult to measure) vertical
velocity (the integral of the acceleration) has a spectrum k2 times smaller, i.e. roughly k-1 (see also
the discussion on the vertical velocity in [Lovejoy, et al., 2007a] where similar conclusions are
reached).
11/29/07 19
Fig. 7: This shows the quadratically detrended vertical acceleration compensated by multiplying by k-1,
bottom 4km, 10 m resolution. The flat spectrum corresponding to scales smaller than about 400m corresponds to Ea(k) ≈ k+1 behaviour. The data are average spectra from 16 drop sondes during the Pacific 2004 experiment over the north Pacific Ocean.
Using the hydrostatic approximation we therefore obtain the following equations for the
number and mass densities:
!n!t
+ vR,n
!z + u( ) "#$ = %
n
2
! vR,n( )
2
!z+$v
R,n
g
1
2& 2 % s2'
()*+,%- N H N (34)
!"!t
+ vR,"!z + u( ) #$" = %
"2
! vR,"( )
2
!z+"v
R,"
g
1
2& 2 % s2'
()*+,
(35)
where we note the additional coalescence term in the equation for n.
3.4 The rain rate:
In a similar manner, we can obtain the z component of the mass flux r which is the rain rate
R:
11/29/07 20
R = r( )z= ! w + v
R,!( ) (36)
where for the vertical wind we have used the notation w = u( )z. Although the full ramifications of
this equation for rain will be developed elsewhere, it should be noted that this approximation will
break down in regions where the particle number density n (strongly correlated with ρ) is so small
that there is low probability of finding a drop (see the discussion of the necessary compound
cascade/Poisson model in section 4.3, 4.4 and eq. 61). In this way zero rain rate regions simply
correspond to regions with very small n.
To understand how this may affect the rain statistics, consider the following simple model for
this. Since n and ρ are highly correlated, we may set ρ to zero whenever it is below some threshold
ρt. If for the moment we ignore the vR,ρ term the model is:
R = !tw; !
t=! ! > t
0 ! < t (37)
We now take ρ to have Corrsin-Obukhov statistics (Δρ ≈ lH with H = 1/3 corresponding to a
spectrum k-β ; ignoring intermittency corrections, β = 1+2H= 5/3 as found in the HYDROP
experiment), and w to have H ≈ 0 statistics (β ≈ 1; as inferred indirectly from the drop sonde data,
fig. 7). We seek the spectrum of R. First consider the effect of the threshold. At large scales it
imposes a fractal support which flattens the spectrum; at high wavenumbers it remains smoother
than w so that the product ρtw will have high wavenmber statistics dominated by the vertical wind ≈
k-1, while at low wavenumbers, it will be much shallower (depending somewhat on the threshold
and the value of H). In addition if we confine our rain analyses to regions with no zeroes, then the
break disappears and we expect the roughly k-1 statistics. If we now consider the (neglected) !vR,!
in eq. 36, we see that that it will have the much smoother, nearly Corrsin-Obukhov (β ≈ 5/3)
statistics; its contribution will to the spectrum will thus be dominated by the vertical wind term.
These conclusions have been substantiated by numerical simulations and will be the subject of a
future paper.
This simple model may well be sufficient to explain observations of rain from high
(temporal) resolution raingauges: several studies have indeed found corresponding ω-0.5 spectra at
long times but ω-1 spectra at short times (see e.g [Fraedrich and Larnder, 1993] with transitions
occurring at scales of 2-3 hours (see [de Montera, et al., 2007]).
11/29/07 21
3.5 Energy flux, Number and mass variance fluxes:
In fig. 1, we noted that at scales l>lR, ρ accurately obeyed the Corrsin-Obhukov law for passive
scalars: Eρ(k)≈k-5/3. The reason is that the variance flux ! = "#$2
#t is “conserved” by the nonlinear
terms and can therefore only be dissipated at small scales. To see this, multiply eq. 24 by -2ρ; to
obtain:
! = "#$2
#t= 2$% & r = % & $2'$( ) + $2% &' $ (38)
Integrating over a volume and using the divergence theorem, we see that only the !2" #$!
represents a dissipation of the variance flux. We therefore obtain:
!diss " #2$ %& # " #2'vR,#'z
+vR,#
g( 2/ 2 ) s2( )
*+,
-./
(39)
(with the hydrostatic approximation). These terms are only important at small scales. This is true
of the vorticity/shear term since Eω(k) = Es(k) = k2Eu(k) (the standard result in isotropic turbulence
which is a consequence of the fact that ω, s are derivatives of the velocity). In Kolmogorov
turbulence therefore both ω2 and s2 have roughly k1/3 spectrum. Furthermore, we have seen form
fig. 6 that the spectrum of !vR!z
is also dominated by the small scales and also has a near k1/3
spectrum so that we may expect both dissipation terms to be important. Unlike the variance flux
dissipation in passive scalar advection which is due purely to molecular diffusion (and hence
depends only on the Laplacian of the velocity (the curl of the vorticity), here the dissipation
depends on the field itself (ρ2) but also on the local relaxation velocity, its vertical derivative, the
vorticity and shear (the latter actually reduces the dissipation).
Similarly, for the dissipative part of the number variance flux:
!diss
= "#n2
#t= n
2$ %&n+ 2n' N H N (40)
so we have:
! diss " n2
#vR,n#z
+vR,n
g$ 2/ 2 % s2( )
&'(
)*++ 2n, N H N (41)
(again using the hydrostatic approximation). The dissipation of number density flux thus has a first
term of the same form as the mass density flux, but an addition coalescence term which is actually
cubic in n and which does not depend on spatial gradients. We therefore expect it to contribute to
11/29/07 22
dissipation of ψ over a wide range of scales with intensity strongly dependent on the local drop
concentration. The dissipation mechanism is thus quite different than for ρ.
3.6 Coalescence:
We now consider in more detail the coalescence term:
N ! M, "M ,x,t( ) N = ! x,t( ) N H M, "M( ) N (42)
where we have factored the general kernel ! M, "M ,x,t( ) into a purely mass dependent
H M, !M( ) and a
time-space varying speed term ϕ. The H M, !M( ) kernel describes the basic collision mechanism,
whereas the ϕ determines the speed. Fairly generally, we may write:
! M, "M ,x,t( ) = E M, "M( ) L + "L( )
2#v M, "M ,x,t( ) (43)
where E is the collision efficiency (typically taken as a power law of the mass ratios, c.f. [Beard
and Ochs III, 1995]), and Δv is the difference in velocities of the colliding drops. Following the
empirical spectrum which indicates a rapid transition from drops following the flow at scales larger
than LR to drops with apparently white noise statistics at smaller scales, it is natural to consider the
simplified model that the particles follow the flow down to LR (roughly independent of particle
mass) and then that they are independent at smaller scales. When two particles collide, we may
thus assume that since they decoupled from the flow a distance LR away they have followed roughly
linear trajectories. More precisely, denoting the corresponding positions by x, x’ and masses M, M’.
This velocity difference is therefore:
!v = v M , x( ) " v #M , x#( ) = u x( ) " u x#( ) " TRM( )
Du x( )
Dt" T
R#M( )Du x#( )Dt
$
%
&&&
'
(
))); x " x# = L
R
(44)
where we evaluate the velocity differences at scale LR. If we take M>M’, so that TRM( )>T
R!M( )
we see that the second term is bounded by:
TRM( )
Du x( )
Dt! T
R"M( )Du x"( )Dt
< TRM( )#a; #a =
Du x( )
Dt!
Du x"( )Dt
(45)
11/29/07 23
However; taking !u = u x( ) " u x#( ) = $lR
1/3LR
1/3 and using the estimate (see eq. 6) !a = !u / "e,R
, we
see that the second term is of the order TR/ !
e,R= St times the first. However since the Stokes
number St is <<1 at LR, this can be neglected with respect to the turbulent velocity gradient at the
relaxation scale.
Using the average relaxation scale lR, we therefore obtain the estimate:
!v = !u = "lR
1/3lR
1/3 (46)
which is independent of the particle mass. Using this in the interaction kernel (eq. 43), we obtain:
! x,t( ) N H N = " x,t( )lR
1/3lR
1/3N #
3
4#
!
"####
$
%&&&&
1/3
E M, 'M( ) M1/3 + 'M 1/3( )2N (47)
or:
H M, !M( ) =3!2
4
"
#
$$$$
%
&
'''''
1/3
E M, !M( ) M1/3 + !M 1/3( )2
" x,t( ) = # x,t( )lR
1/3lR
1/3
(48)
i.e. the coalescence speed ϕ is equal to the turbulent velocity gradient at the mean relaxation scale
(the subscript on the integral denotes spatial averaging at the relaxation scale).
Although ε(x,t) and hence the speed ϕ(x,t) is highly variable, it defines characteristic velocities
for the coalescence process: !lR
1/3lR
1/3 . We note that this approximation implies that the interdrop
collision rate is determined by the turbulence and not by the mass dependent relaxation velocities
(as is usually assumed).
4. The scaling laws:
4.1 The Corrsin-Obukhov law:
We have seen in the previous section that except at small dissipation scales, the mass density
variance flux (χ), number density variance flux (ψ) and energy flux (ε) are conserved:
! = !"u
2
"t; " = !
"#2
"t; $ = !
"n2
"t (49)
The standard Kolmogorov and Corrsin-Obukhov laws are obtained by assuming that there is a
quasi constant injection of the passive scalar variance flux χ and energy flux ε fluxes from large
scales and that this is thus transferred without significant loss to the small scales where it is
11/29/07 24
dissipated. Between an outer injection scale and the dissipation scale, there is no characteristic
scale, hence at any intermediate scale l one expects:
!l!"u
l
2
"e,l
; #l!"$
l
2
"e,l
(50)
where Δul is the typical shear across an l sized eddy, and Δρl is the corresponding typical gradient of
passive scalar; the τe,l’s are the corresponding transfer times (the “eddy turnover time”). For both ε,
χ the time scale is determined by the turbulent velocity at the length scale l:
!l
=l
!ul
(51)
This is the time scale for Δρl as well as Δul since neither process is affected by coalescence. This
leads to:
!l!"u
l
3
l; "
l!"#
l
2"u
l
l (52)
Solving for Δul, Δρl, we obtain the classical Kolmogorov and Corrsin Obukhov laws:
!u
l= !
l
1/3l1/3 (53)
!!
l= "
l
1/2#l
"1/6l1/3 (54)
Finally, we might add that by invoking a third property of the equations – that they are “local” in
fourier space, i.e. interactions are strongest between neighbouring scales, we obtain the standard
cascade phenomenology, the basis of cascade models and multifractal intermittency.
4.2 The new number density scaling law
We have noted (section 3.6) a key difference between the mass density and the number density
equations: due to the conservation of mass in coalescence processes, the mass density equation was
independent of coalescence processes while on the contrary the number density was dependent on
the latter.
Whereas the speed of variance flux transfer for ε, χ is the strongly scale dependent Δul the
corresponding speed for the number density variance flux ψ is determined by the weakly scale
dependent speed of the coalescence processes at the scale l:
!
l= "
lR
1/3lR
1/3( )l
! "l
1/3 "#
1/3lR
1/3; # = l / lR
(55)
11/29/07 25
the subscript on the bracketed term indicates spatial averaging at the scale l. We have used the
multiplicative property of the ε cascade so that it factors into a product of a cascade from the
external scale down to scale l with another smaller scale cascade over the range l to lR , i.e. over the
ratio λ. We then use the approximation that at scales below l, that spatial averaging may be
replaced by ensemble averaging and that lR is roughly constant. For such cascades, we have
!"
q = "K q( ) where K(q) is the moment scaling exponent characterizing the statistics of the cascade,
hence !"
1/3 = "K 1/3( )
= l / lR
( )K 1/3( )
; empirically in atmospheric turbulence K(1/3) ≈ -0.07 so that
this “intermittency correction” is small (see e.g. [Schmitt, et al., 1992], [Schmitt, et al., 1996]). For
the coalescence variance flux transfer ψ we therefore obtain:
!l!"n
l
2
"l,R
; "l,R
=l
#l
(56)
so that the ψ cascade is determined by the weakly scale dependent coalescence processes not by the
strongly scale dependent turbulent processes. We thus obtain:
!l="n
l
2
#l
="n
l
2
l$l
1/3l
lR
%
&'(
)*
K 1/3( )
lR
1/3 (57)
Hence for the scaling of the fluctuations in number density at scale l:
!nl"#
l
1/2$l
%1/6lR
%1/6l1/2 (58)
where we have ignored the small intermittency correction with exponent –K(1/3)/2 (which
empirically is of the order 0.03). This l1/2 law (eq. 58) is the key result of this section, in fourier
space (ignoring multifractal intermittency corrections), this implies En(k)≈ k-2 whereas for ρ, we
have the classical Corrsin-Obukhov result: Eρ(k)≈ k-5/3.
4.3 Coupled χ, ψ cascades and the l1/2 law:
If we combine the l1/3 law from Δρ, and the l1/2 law for Δn we obtain simple relation:
!!l
="
l
#l
"
#
$$$$
%
&
'''''
1/2
lR
l
"
#
$$$$
%
&
'''''
1/6
!nl (59)
which has no explicit dependence on the energy flux ε and shows that mass density fluctuations
tend to damp out with respect to the number density fluctuations at larger and larger scales.
11/29/07 26
Eq. 59 raises the question of the statistical couplings between the ε, χ, ψ cascades. While the link
between ε and χ, ψ is not obvious, χ, ψ are indirectly linked through the number size distribution N,
so that they cannot be statistically independent of each other. Indeed, the ratio χ/ψ has dimensions
of mass2 and it seems clear that at least at large enough scales the two should be related by the
(ensemble, i.e. climatological) drop mass variance M
2 :
!
Lext
! "L
ext
M2 (60)
where Lext is the external scale of the cascades. A “strongly coupled” model would take
!
l! "
lM
2 , i.e. it would assume this relation to hold at all scales l, but this is likely to be too
strong an assumption to be realistic.
In order to exploit these statistical relations in order to make stochastic processes with the
corresponding statistics, we may use standard multifractal simulation techniques to simulate ε, χ, ψ
and from them (by fractional integration) to obtain the extra l1/3, l1/2 scalings, the, u, ρ, n fields.
However the n field determines the probability per unit volume of finding a drop; in other words it
should control a compound Poisson-multifractal process. In a future paper, we show how to make
such processes which produces stochastic realizations respecting all the above the statistics, and
which determines implicitly the drop size distributions. For the moment, we note that the
particulate nature of rain in fact imposes a length scale equal to the mean interdrop distance:
linter
= n!1/3 (61)
(this is the actual number density, not its fluctuation Δn). Therefore, wherever linter>lR , lR must be
replaced by linter. This will occur at low rain rates and will modify the low rain rate statistics. This
behaviour in fact provides a natural “cutoff” mechanism for the transition from rain to no rain; in
effect for lower and lower rain rates, the inner scale of the cascade becomes larger and larger. We
can study this using compound cascade Poisson processes.
4.4 The empirical number density law:
In order to empirically test the l1/2 law for the number density, we replaced each drop mass by the
indicator function and produced the spectra shown in fig. 8a.
11/29/07 27
Fig. 8a: Same as previous but for n, the particle number density (calculated using an indicator function
on a 1283 grid. The reference lines have slopes -2, +2, the theoretical values for the l1/2 law and white noise respectively.
We see that the convergence to the low wavenumber theoretical k-2 behaviour (straight line)
occurs at slightly smaller scales than for the k-5/3 behaviour of ρ since n is less variable (smoother)
than ρ. In fig. 8b we show the ratio of the ensemble spectra (all 19 triplets) for Eρ(k), En(k); this
shows that the number density field really is smoother (by about k1/3) than the corresponding
spectrum for the mass density. Because we have taken the ratio of the spectra, the y axis is “blown
up” with respect to the previous; this is quite a sensitive indicator that the basic theory is roughly
correct. Finally, in fig. 8c, we show the exponents of spectra of ρη; η = 0, 1 are the number and
mass density spectra whereas η = 1/6, 7/6 are important for vRn, vR,ρ respectively.
11/29/07 28
Fig. 8b: This shows the ratio of the ensemble spectra Eρ(k)/En(k); for each of the 5 storms, and the overall
ensemble (purple), with theoretical reference line slope -1/3. This is a sensitive test of the prediction of eqs. 54, 58.
11/29/07 29
Fig. 8 c: These are mean β values for the different η’s with the bars indicating storm to storm differences.
The fits were for wavenumbers < (70cm)-1 (the turbulent regime). Note that for η = 0 (number density) and η =1 (mass density), the theoretical predictions (2, 5/3; ignoring intermittency corrections) are well verified since β = 2.02±0.11, 1.74±0.12 respectively.
5. Conclusions:
Cloud and rain physics have been divorced from turbulence theory for too long. While attempts
to unify them at the most fundamental level are very difficult, phenomenological turbulent cascade
models provide an attractive alternative. This is because they have proved to be not only highly
accurate but even indispensable for understanding and modelling both rain and turbulence; they
thus provide the natural framework for explicitly marrying the two. In this paper, we benefited
from two key advances: the general and elegant relation between particle and wind velocities
([Maxey, 1987]) and a new and unique source of data spanning the drop/turbulence scales: the
HYDROP experiment. The latter provides the positions and masses of rain drops in a roughly 10
m3 volume; with Maxey’s relations, they enabled us to convincingly formulate dynamical equations
for both liquid water mass and particle number densities and to empirically verify many of the
assumptions. With the exception of a mean vertical velocity, the liquid water mass equations were
11/29/07 30
identical to those of standard passive scalar advection and therefore predicted that the rain should
follow the Corrsin-Obukhov (l1/3 , k-5/3 law) of passive scalar advection. However, the l1/2 (i.e. k-2
spectrum) number density law was quite new and arises from the fact that while coalescence
processes conserve total mass, they do not conserve total particle numbers. This allows the (weakly
scale dependent) coalescence speed to determine the rate of scale by scale transfer of number
variance flux whereas the speed of the corresponding mass density variance flux depends on the
strongly scale dependent turbulent velocity. In both cases, the HYDROP data was able to verify
that both the l1/3 and l1/2 laws were obeyed with reasonable accuracy.
Although in this paper we only touch on it, an important consequence concerns the rain rate field
R which is the vertical component of the liquid water mass flux vector. Within the hydrostatic
approximation (which is of acceptable accuracy in this context), we showed that R is the product of
the mass density and the sum of the vertical velocity and mean relaxation speed. However, when
the number density (n) field (which is correlated with the ρ field) is sufficiently close to zero, it will
imply zero rain rates since the probability of finding a drop becomes very low (a full study of this
effect must be done using compound Poisson-multifractal processes discussed below). In this way,
a natural rain / no-rain mechanism emerges and we argue that it can quantitatively explain
observations of rain rate spectra.
The conventional methods of modelling the evolution of raindrops (e.g. [Srivastava and
Passarelli, 1980], [Khain and Pinsky, 1995], [Khairoutdinav and Kogan, 1999]) give turbulence at
most a minor (highly “parameterized”) role: the atmosphere is considered homogeneous and the
spatial variability of the drop size distribution arises primarily due to complex drop interactions.
Our approach is in many ways the opposite: down to a small relaxation scale (below which a kind
of white noise “drop chaos” exists), the drops are shown to be strongly coupled with the highly
variable (multifractal) wind field. The coalescence speed is then determined by the turbulent
velocity gradient at this decoupling scale. By considering the zeroth and first moments of the drop
size distribution (n, ρ), and showing how they are related to scale by scale conserved turbulent
fluxes, we obtain strong (but implicit) constraints on the drop size distribution which we expect to
be highly variable simply because it is controlled by turbulent cascades. This idea can be taken
further: if we are given a turbulent number density field, it can be used to control a compound
Poisson process which determines the locations of the drops. Once the positions are determined,
the corresponding ρ field can then be used to attribute masses to the particles. In this way we
obtain a three dimensional model of particle positions and masses which implicitly determines the
highly variable drop size distribution in the vicinity of each point and which respects the turbulence
11/29/07 31
laws and is compatible with the microphysics (see [Lovejoy and Schertzer, 2006] for a preliminary
model). Such models can be used to solve longstanding observers problems in radar meteorology,
satellite rain algorithms, rain amount estimation and cloud physics.
6. Acknowledgements:
This research was carried for purely scientific purposes, it was unfunded.
7. References
Anselmet, F., R. A. Antonia, L. Danaila (2001), Turbulent flows and intermittency in laboratory experiments, Plan. and Space Sci., 49, 1177-1191.
Beard, K. V., and H. T. Ochs III (1995), Collisions between small precipitation drops. Part II: Formulas for coalescence, temprorary coalescence, and satellites, J. Atmos. Sci., 52, 3977-3996.
de Montera, L., et al. (2007), Multifractal properties of rain at high resolution derived from Earth-to- satellite links fading, paper presented at Ninth Inter. Precipitation Conf., U. Paris Est, Paris.
Desaulniers-Soucy., N., et al. (2001), The continuum limit in rain and the HYDROP experiment, Atmos. Resear., 59-60, 163-197.
Falkovitch, G., and A. Pumir (2004), Intermittent distribution of heavy particles in a turbulent flow, Phys. Fluids, 16, L47-L50.
Falkovitch, G., et al. (2006), Rain initiation time in turbuloent warm clouds, J. Applied Meteorology and Climatology, 45, 591-599.
Fessler, J. R., et al. (1994), Preferential concentration of heavy particles in turbulent channel flow, Phys. Fluids, 6, 3742-3749.
Fraedrich, K., and C. Larnder (1993), Scaling regimes of composite rainfall time series, Tellus, 45 A, 289-298.
Jameson, A. R., and A. B. Kostinski (1999), Fluctuation properties of precipitation: Part V: Distribution of rain rates – theory and observations in clustered rain, J. Atmos. Sci., 56, 3920–3932.
Khain, A. P., and M. B. Pinsky (1995), Drop Inertia and its contribution to trublent coalescence in coonvective clouds. Part I: Drop fall in the flow with random horizontal velocity, J. Atmos. Sci., 52, 196-206.
Khairoutdinav, M., and Y. Kogan (1999), A new cloud physics paramterization in a Large-Eddy Simulation model of marine stratocumulus, Mon. Wea. Rev., 128, 229-243.
Le Clair, B. P., et al. (1972), J. Atmos. Sci, 27, 308. Lilley, M., et al. (2007), Scaling turbulent atmospheric stratification, Part II: empirical study of the
the stratification of the intermittency, Quart. J. Roy. Meteor. Soc., in press. Lilley, M., et al. (2006), Multifractal large number of drops limit in Rain, J. Hydrol.,
doi:10.1016/j.jhydrol.2005.11.063, Lilley, M., et al. (2004), 23/9 dimensional anisotropic scaling of passive admixtures using lidar
aerosol data, Phys. Rev. E, DOI 10.1103/70, 036307-036301-036307. Lovejoy, S., et al. (2007a), Scaling turbulent atmospheric stratification, Part I: turbulence and
waves, Quart. J. Roy. Meteor. Soc., (submitted). Lovejoy, S., et al. (2004), Bubble distributions and dynamics: the expansion-coalescence equation,
J. Geophys. Res. (Solid Earth), 109, B, doi:10.1029/2003JB002823., 11203.
11/29/07 32
Lovejoy, S., et al. (2003), Large particle number limit in Rain, Phys. Rev. E, 68, 025301. Lovejoy, S., and D. Schertzer (1990), Fractals, rain drops and resolution dependence of rain
measurements, Journal of Applied Meteorology, 29, 1167-1170. Lovejoy, S., and D. Schertzer (1991), Multifractal analysis techniques and the rain and clouds fields
from 10-3 to 106m, in Non-linear variability in geophysics: Scaling and Fractals, edited by D. Schertzer and S. Lovejoy, pp. 111-144, Kluwer.
Lovejoy, S., and D. Schertzer (1995), Multifractals and Rain, in New Uncertainty Concepts in Hydrology and Water Ressources, edited by Z. W. Kunzewicz, pp. 62-103, Cambridge University Press.
Lovejoy, S., and D. Schertzer (2006), Stereophotography of rain drops and compound Poisson-multifractal cascade processes, paper presented at Cloud Physics conf., American Meteor. Soc., Madison Wisc., July 2006.
Lovejoy, S., et al. (2007b), The remarkable wide range spatial scaling of TRMM precipitation, submitted to Atmos. Res. August., J. Atmos. Research, submitted 8/7.
Lovejoy, S., et al. (1987), Functional Box-Counting and Multiple Dimensions in rain, Science, 235, 1036-1038.
Marshak, A., et al. (2005), Small-Scale Drop-Size Variability: Empirical Models for Drop-Size-Dependent Clustering in Clouds, J. Atmos. Sci., 62, 551-558.
Marshall, J. S., and W. Hitschfeld (1953), Interpretation of The fluctuating Echo from Random Distributed Scatterers; Part I., Canadian Journal of Physics, 31, 962-994.
Maxey, M. (1987), J. Fluid Mech., 174, 441. Maxey, M., and J. Riley (1983), Phys. Fluids, 26, 883. Pruppacher, H. R. (1997), Microphysics of Clouds and Precipitation, 954 pp., Kluwer, Dordrecht,
Holland. Radkevitch, A., et al. (2007), Scaling turbulent atmospheric stratification, Part III: empIrical study
of Space-time stratification of passive scalars using lidar data, Quart. J. Roy. Meteor. Soc., in press.
Schertzer, D., and S. Lovejoy (1987), Physical modeling and Analysis of Rain and Clouds by Anisotropic Scaling of Multiplicative Processes, Journal of Geophysical Research, 92, 9693-9714.
Schertzer, D., and S. Lovejoy (1997), Universal Multifractals do Exist! J. Appl. Meteor., 36, 1296-1303.
Schmitt, F., et al. (1992), Empirical Determination of Universal Multifractal Exponents in Turbulent Velocity Fields, Physical Review Letter, 68, 305-308.
Schmitt, F., et al. (1996), Universal multifractal structure of atmospheric temperature and velocity fields, Europhysics Lett., 34, 195-200.
Srivastava, R. C., and R. E. Passarelli (1980), Analytical solutions to simple models of condensation and coalescence, J. Atmos. Sci., 37, 612–621.
Uijlenhoet, R., et al. (1999), Towards a stochastic model of rainfall for radar hydrology: testing the Poisson homogeneity hypothesis, Phys. Chem. Earth (B), 24, 747–755. Wallace, P. R. (1953), Interpretation of The fluctuating Echo from Random Distributed Scatterers;
Part II, Canadian Journal of Physics, 31, 994-1009. Wang, L., and M. Maxey (1993), Settling velcoity and concentration distribution of heavy particles
in homogeneous isotropic turbuloence, J. Fluid Mech., 256, 27.