2284 JOURNAL OF THE ATMOSPHERIC SCIENCES VOLUME 55

A Flux Method for the Numerical Solution of the Stochastic Collection Equation
ANDREAS BOTT
Institut für Physik der Atmosphäre, Johannes Gutenberg-Universität Mainz, Mainz, Germany
(Manuscript received 29 January 1997, in final form 20 October 1997)

ABSTRACT
A new mass conservative flux method is presented for the numerical solution of the stochastic collection
equation. The method consists of a two-step procedure. In the first step the mass distribution of drops with mass
x9 that have been newly formed in a collision process is entirely added to grid box k of the numerical grid mesh
with x k # x9 # x k11 . In the second step a certain fraction of the water mass in grid box k is transported to k 1
1. This transport is done by means of an advection procedure.
Different numerical test runs are presented in which the proposed method is compared with the Berry–Reinhardt
scheme. These tests show a very good agreement between the two approaches. In various sensitivity studies it
is demonstrated that the flux method remains numerically stable for different choices of the grid mesh and the
integration time step. Since a time step of 10 s may be used without significant loss of accuracy, the flux method
is numerically very efficient in comparison to the Berry–Reinhardt scheme.

1. Introduction One of the most popular schemes for the numerical

solution of the SCE has been developed by Berry and
Numerical modeling of clouds requires the simulation
Reinhardt (1974). In this approach the SCE is solved at
of various processes such as the formation of cloud
discrete points of the drop spectrum. If necessary, high-
droplets, their diffusional growth, and the evolution of
order Lagrangian polynomials are used to interpolate
a raindrop spectrum. One of the most important mech-
the drop spectrum at intermediate points. Seeßelberg et
anisms for the formation of raindrops is the collision
al. (1996) presented a stochastic simulation algorithm
and subsequent coalescence of smaller cloud droplets.
Analytically the collision/coalescence of two drops is for the solution to the SCE. The main purpose of their
treated as a stochastic process and is described by the method is to produce benchmark calculations for situ-
stochastic collection equation (SCE). Many different ations where no analytical solutions of the SCE are
methods have been developed for the numerical solution available. By comparing their stochastic simulation al-
of the SCE. The approaches of Kovetz and Olund (1969) gorithm with the Berry–Reinhardt scheme, Seeßelberg
and Bleck (1970) are known to be numerically very et al. found a very good agreement between both meth-
efficient. However, due to their strong numerical dif- ods, thus demonstrating the high accuracy of the latter
fusivity both methods produce unsatisfactory results as scheme.
expressed by an artificial broadening of the raindrop As constructed, the Berry–Reinhardt scheme is not
spectrum. Soong (1974) extended Bleck’s algorithm by mass conservative. In addition, it is computationally rel-
representing in each grid box the mass distributions by atively time-consuming. These are two serious disad-
exponential functions. This yielded a clear improvement vantages when the scheme is implemented into a dy-
of the numerical results. The method of Tzivion et al. namic cloud model with detailed microphysics. Fur-
(1987) is based on a transformation of the SCE into a thermore, since the scheme solves the SCE in an integral
set of equations for the moments of the drop spectrum. way, one gets no information on the collision partners
Solving this equation set for the first two moments of being involved in a particular collision–coalescence pro-
the drop spectrum (the liquid water mass and the total cess. However, such information is necessary if, for in-
number concentration), Tzivion and coworkers also dis- stance, one is interested in the redistribution of aerosol
tinctly improved the results in comparison to Bleck’s mass or other chemical substances that are contained
method. within the colliding droplets. In order to avoid these
shortcomings, in the present paper a new flux method
is proposed for the numerical solution of the SCE. The
new scheme is exactly mass conservative and is com-
Corresponding author address: Dr. Andreas Bott, Institut für Phy-
sik der Atmosphäre, Johannes Gutenberg-Universität Mainz, D-55099
putationally very efficient. A detailed description of the
Mainz, Germany. flux method will be given in the following section. Sec-
E-mail: [email protected] tion 3 presents some numerical results for different test

q 1998 American Meteorological Society

Unauthenticated | Downloaded 03/30/22 11:16 AM UTC

1 JULY 1998 BOTT 2285

cases. The quality of the results will be checked by x i with drops of mass x j yields a change in the mass
comparing them with corresponding results obtained distributions g i , g j ,
with the Berry–Reinhardt scheme. Section 4 deals with
K (i, j)
the numerical stability of the flux method while section g i (i, j) 5 g i 2 g i g jDyDt (5)
5 summarizes the findings of the paper. xj
K ( j, i)
2. Model description g j ( j, i) 5 g j 2 g j g iDyDt. (6)
xi
The stochastic collection equation is given by (Prup- Here, g i , g j are the mass distribution functions at grid
pacher and Klett 1997) points i, j before the collision process while g i (i, j) and
]n(x, t)
]t
5 E x1

x0
n(x c , t)K(x c , x9)n(x9, t) dx9
g j ( j, i) represent the new mass distributions after the
collision. Also, K(i, j) is an average value of the col-
lection kernel that has been obtained by two-dimen-

2 E `

x0
n(x, t)K(x, x9)n(x9, t) dx9, (1)
sional linear interpolation:
1
K (i, j) 5 [K(i 2 1, j) 1 K(i, j 2 1) 1 4K(i, j)
8
where n(x, t) is the drop number distribution function
at time t and K(x c , x9) is the collection kernel describing 1 K(i 1 1, j) 1 K(i, j 1 1)]. (7)
the rate at which a drop of mass x c 5 x 2 x9 is collected Note that the collection kernel is symmetrical in i and
by a drop of mass x9 thus forming a drop of mass x. j so that K( j, i) 5 K(i, j) and g j ( j, i) 5 g j (i, j). Due
Here, x 0 is the mass of the smallest drop being involved to the collision of drops in grid box i with drops in grid
in the collection process and x1 5 x/2. Following Berry box j new drops with mass x9(i, j) 5 x i 1 x j are pro-
(1967), the mass distribution function g( y, t) will now duced:
be introduced by
x9(i, j)
1 g9(i, j) 5 g K (i, j)g j DyDt. (8)
g(y, t)dy 5 xn(x, t)dx, n(x, t) 5 g(y, t), (2) xi xj i
3x 2
Now the well-known problem arises that usually x9(i,
where y 5 lnr and r is the radius of drops with mass j) differs from the discretized mass points x k as defined
x. Substituting (2) into (1) yields the SCE for the mass in (4). Instead, one has
distribution function

E
x k # x9(i, j) # x k11 . (9)
]g(y, t) y1
x2
5 g(y c , t)K(y c , y9)g(y9, t) dy9 Thus the mass density g9(i, j) has to be split up in grid
]t x c2 x9
y0
boxes k and k 1 1. In the present model this partitioning
2 E `

y0
g(y, t)
K(y, y9)
x9
g(y9, t) dy9. (3)
is done in the following two-step procedure. First g9(i,
j) is entirely added to grid box k, yielding

The first integral on the right hand side of (1) and (3) g9k (i, j) 5 g k 1 g9(i, j). (10)
describes the gain rate of drops of mass x by collision In the second step a certain fraction of the new mass,
and coalescence of two smaller drops, while the second g9k (i, j), is transported into grid box k 1 1. This transport
integral denotes the loss of drops with mass x due to is calculated by means of an advection process through
collection by other drops. the boundary k 1 1/2 between grid box k and k 1 1.
For the numerical solution of (3) a logarithmically The procedure is schematically illustrated in Fig. 1.
equidistant mass grid is introduced, that is, Here, the dashed lines indicate the initial mass distri-
x k11 5 ax k , k 5 1, . . . , m, (4) butions in grid boxes i, j, k, and k 1 1. The full lines
denote the mass distributions after the collision process.
yielding the y grid mesh equally spaced, Dy k 5 Dy 5 The stippled area in grid box k represents the mass that
lna/3. Note m is the total number of grid points. In the will be transported into grid box k 1 1 while the dark
present model a may be arbitrarily chosen with a . 1. shaded areas are the final mass increase in grid boxes
Many authors use a 5 21/2 , yielding a doubling of the k and k 1 1.
drop mass after two grid cells. This value has also been Formally the advection step may be written as
utilized in the numerical calculations presented in the
following section, thus facilitating the comparison of g k (i, j) 5 g9(i,
k j) 2 f k11/2 (i, j)
the results with already existing models. However, sen- g k11 (i, j) 5 g k11 1 f k11/2 (i, j), (11)
sitivity studies with different a-values will also be pre-
sented. whereby f k11/2 (i, j)Dy/Dt is the mass flux through the
In discretized form the collision of drops with mass boundary k 1 1/2.

Unauthenticated | Downloaded 03/30/22 11:16 AM UTC

2286 JOURNAL OF THE ATMOSPHERIC SCIENCES VOLUME 55

FIG. 1. Schematic illustration of the flux method. See text for details.

The simplest approach to calculate f k11/2 (i, j) is an

upstream formula, yielding E 1/2

1/22ck
gk,l (z) dz 5 O (s 1a1)2
l

s50
k,s
s11
[1 2 (1 2 2ck ) s11 ],

f k11/2 (i, j) 5 c kg9k (i, j)w(i, j). (12) (17)

Here, c k may be interpreted as a Courant number and yielding the mass flux through the boundary k 1 1/2 as

O (s 1a1)2
is calculated as a function of the position of x9(i, j) l

f k11/2 (i, j) 5 w(i, j) k,s

between x k and x k11 : s11
s50
x9(i, j) 2 x k
ck 5 . (13) 3 [1 2 (1 2 2c k ) s11 ]. (18)
x k11 2 x k
In (12) the weighting function w(i, j) has been intro- The coefficients a k,s are obtained by interpolation of
duced because, in contrast to the normal advection pro- the mass distribution function as in Bott (1989b). For
cess, for c k 5 1, that is, for x9(i, j) 5 x k11 , the advective l 5 0 one obtains again UFM with a k,0 5 g9k (i, j). The
flux through the boundary k 1 1/2 is given by g9(i, j) linear flux method (LFM) results from l 5 1,
and not by g9k (i, j). Consequently, we have a k,0 5 g9k (i, j), a k,1 5 g k11 2 g9k (i, j), (19)
g9(i, j) and the parabolic flux method (PFM) with l 5 2 is given
w(i, j) 5 . (14)
g9(i,
k j) by
Substitution of (14) into (12) shows that in the up- 1
stream scheme f k11/2 (i, j) is simply given by a k,0 5 2 [g 2 26g9(i, j) 1 g k21 ]
24 k11 k

f k11/2 (i, j) 5 c k g9(i, j), (15) 1

a k,1 5 (g k11 2 g k21 )
resulting in the same partitioning of g9(i, j) as in the 2
method of Kovetz and Olund (1969), except that in their
1
model the SCE is solved for the number distribution a k,2 5 [g k11 2 2g9(i,
k j) 1 g k21 ]. (20)
n(x, t) rather than for the mass distribution. The model 2
version with f k11/2 (i, j) calculated according to (15) will Positive definiteness of the advection scheme is
henceforth be called upstream flux method (UFM). achieved by applying the flux limiter
In the following section presenting some model re-
sults it will be seen that UFM yields rather broad drop 0 # f k11/2 (i, j) # g9(i, j). (21)
spectra. This is explained by the large numerical dif- In order to treat all collisions of drops during the time
fusion produced by the upstream advection scheme. In step Dt the following iterative procedure is applied. Let
order to reduce the numerical diffusivity of the flux us denote the grid box of the smallest and the largest
method, f k11/2 (i, j) will now be calculated according to drops being involved in the collision process by i 5 i 0
the higher-order advection schemes of Bott (1989a,b). and i 5 i1 , respectively. In the first step the collision of
These advection schemes are positive definite thus the smallest drops with drops of grid box j 5 i 0 1 1 is
avoiding the occurrence of negative mass concentra- calculated yielding new values g i 0 (i 0 , i 0 1 1), g i 011 (i 0 ,
tions. In grid box k the constant value g9k (i, j) is replaced i 0 1 1), g k (i 0 , i 0 1 1), and g k11 (i 0 , i 0 1 1) according
by a polynomial of order l, to Eqs. (5), (6), and (11). In the next step the collision

Oa of the remaining drops in i 5 i 0 having now the new

l

g k,l (z) 5 k,s zs, (16) mass distribution function g i 0 (i 0 , i 0 1 1) with drops in
grid box j 5 i 0 1 2 is calculated. This is continued until
s50

with z 5 ( y 2 y k )/Dy and 21/2 # z # 1/2. Now the all collisions of the drops in grid box i 5 i 0 with drops
term c kg9k (i, j) in (12) is replaced by the integral relation of grid boxes j 5 i 0 1 1, i 0 1 2, . . . , i1 are accounted

Unauthenticated | Downloaded 03/30/22 11:16 AM UTC

1 JULY 1998 BOTT 2287

for. In the next step the collisions of drops in grid box

1 2
L x
i 5 i 0 1 1 with all larger drops j 5 i 0 1 2, i 0 1 3, n(x, t 5 0) 5 exp 2 , (25)
. . . , i1 are treated in the same way. This procedure is x 2
x
repeated for all drops i 5 i 0 1 2, . . . , until in the last where L is the total cloud water content and x is the
step the collision of drops i 5 i1 2 1 with drops j 5 i1 mean drop mass. Here x is related to the mean drop
has been calculated. If for a particular kernel (such as radius r by x 5 (4/3)prr 3 where r is the density of
the Golovin kernel to be presented in the next section) water.
the collision of equal-sized drops is also possible, in Three different kernels will be applied in the calcu-
each grid box the procedure starts with the collision of lations. In the first test the Golovin kernel K(i, j) 5 b(x i
drops i with drops j 5 i. Due to this iteration approach, 1 x j ) with b 5 1500 cm 3 g21 s21 will be used. For this
after each collision the drop distribution is updated be- kernel the analytic solution to the SCE has been pre-
fore the next collision process is calculated. Analytically sented by Golovin (1963) and will, therefore, be com-
this is accomplished by replacing in (5)–(11) after the pared with the numerical results of the flux method. It
first collision process g i , g j , g k , and g k11 with g i (i, j 2 should, however, be noted that for the Golovin kernel
1), g j (i, j 2 1), g k (i, j 2 1), and g k11 (i, j 2 1), respec- BRM yields almost the same results as the analytical
tively. solution. Note that L and r have been chosen to L 5 1
From (5) and (6) it is easily seen that g i (i, j) or g j ( j, g m23 and r 5 10 mm.
i) may become negative so that the positive definiteness Figure 2a shows the mass distributions g(r) of the
of the flux method would be violated. In order to avoid three model versions UFM, LFM, and PFM in com-
this situation the numerical time step has to be limited parison to the analytical solution as obtained after 30
in the following way: and 60 min. It is clearly seen that UFM produces dis-
tinctly too large drops, yielding a shift of the mass dis-
xj tributions toward larger radii. In contrast to this, for
Dt # (22)
g j (i, j 2 1)DyK (i, j) LFM and PFM the location of the curves agrees much
better with the analytical solution. However, in both
and model versions the maxima of the distributions are
xi somewhat too high. In order to compare the numerical
Dt # if j ± k. (23) diffusivity of the three model versions, Fig. 2b depicts
g i (i, j 2 1)DyK (i, j) the same results (after 30 min) now, however, with a
The inequality (22) guarantees that after each collision logarithmic scale of the ordinate. Obviously the nu-
process g i (i, j) $ 0, while restriction (23) does not apply merical diffusion of the flux method is decreasing with
if j 5 k. The reason for this is that for j 5 k the mass increasing order of the advection scheme. While for
subtracted from g j ( j, i) in (6) will be added again in LFM the concentration of the largest drops is still some-
(10). Note that according to (8), what higher than in the analytical solution, in PFM the
production of big drops is slightly retarded. This is ex-
K ( j, i) pressed by the steep gradient of the dash-dotted curve
g9(i, j) . g j (i, j 2 1) g i (i, j 2 1)DyDt (24) at large radii.
xi
The numerical results presented in the following are
since x9(i, j) . x j . Thus, for j 5 k after the collision obtained by utilizing for the solution of the SCE the
process one obtains g j ( j, i) $ 0 without applying the hydrodynamic kernel (Pruppacher and Klett 1997)
limitation (23) to the kernel.
K(i, j) 5 p (r i 1 r j ) 2E|w(r i ) 2 w(r j )|. (26)

3. Model results Here, E is the collision efficiency taken from Long

(1974) while w is the terminal velocity of the falling
In this section some numerical results of the flux drops, which is obtained following Beard (1976). Fig-
method will be presented and compared with the Berry– ures 3a and 3b depict for the three model versions and
Reinhardt model (henceforth abbreviated as BRM). In for BRM the distributions after 30 and 60 min (only
all model runs of BRM an integration time step, Dt 5 Fig. 3a). As was to be expected, in UFM the evolution
1 s, has been used while the grid mesh has been deter- of the raindrop spectrum is again too fast. After 30 min
mined according to (4) with a 5 21/2 . In the flux method the maximum of the large drops is already distinctly
Dt 5 10 s and a 5 21/2 have been chosen. From nu- higher than in BRM while the concentration of the small
merous model runs these values turned out to be a rea- cloud droplets is too low. The curves of LFM agree
sonable compromise between numerical accuracy and very well with BRM whereby, similar to the Golovin
efficiency of the scheme. The sensitivity of the flux kernel, after 60 min the maximum of the distribution is
method regarding different choices of Dt and a will be somewhat too high. The results of PFM are surprisingly
discussed in section 4. The initial cloud droplet distri- bad. In contrast to the other model runs, after 30 min
bution is given by a gamma function of the form the second maximum of the distribution has not yet

Unauthenticated | Downloaded 03/30/22 11:16 AM UTC

2288 JOURNAL OF THE ATMOSPHERIC SCIENCES VOLUME 55

FIG. 2. (a) Solution to the stochastic collection equation for the Golovin kernel, L 5 1 g m23 , and r 5 10 mm with different versions
of the flux method in comparison to the analytic solution (full curve). Short dashed curve: UFM; long dashed curve: LFM; dash-dotted
curve PFM. Linear scale of the ordinate. Curves are shown after 30 and 60 min. (b) Same as (a) except with a logarithmic scale of the
ordinate. Curves are shown after 30 min.

evolved while after 60 min the raindrops are clearly too a way that the mass transport is calculated only if g9k (i,
small. j) exceeds a critical value gmin . Thus, in LFM the flux
Obviously, the strong antidiffusive character of the limitation (21) is modified in the following way:
second-order advection scheme inhibits the evolution of
the raindrop spectrum. This behavior is explained by
the use of a second-order polynomial for the calculation
of the mass flux in Eq. (18) together with the positive
f k11/2 (i, j) 5
# g9(i, j) if g9(i,
50
k

if g9(i,
k
j) . gmin
j) # gmin .
(27)

definite flux limitation (21). In contrast to polynomials In order to determine an optimum value for gmin , the
of order l 5 0, 1 higher-order polynomials with l $ 2 model run presented in Fig. 3 has been repeated with
may produce negative values for the mass flux. This the modified linear flux method (MFM), which utilizes
happens particularly in regions with steep gradients of the flux limitation (27) instead of (21). In different mod-
the mass distribution, that is, if g9k (i, j) k g k11 , and el runs the values of gmin were varied between 1029 and
when at the same time the Courant number c k is small. 1026 g m23 . The corresponding results are summarized
Both conditions are met at the right tail of the drop in Fig. 4. The dashed curve shows the distribution of
distribution. In the model calculations presented in this BRM after 30 min. Curves 1 and 6 denote the results
paper on the average more than 50% of the collisions as obtained with LFM and PFM. Curves 2–5 are the
yielded c k , 0.01. distributions of MFM with gmin 5 1029 , 1028 , 1027 , and
If in a particular situation f k11/2 (i, j) , 0, the flux 1026 g m23 , respectively. It is clearly seen that with
limitation (21) replaces the negative value by zero, thus increasing value of gmin the numerical diffusivity of the
inhibiting the transport of water mass from grid box k advection scheme is decreasing. For gmin $ 1027 g m23
into grid box k 1 1. Therefore, the mass transport (curves 4 and 5) the flux limitation is again too restric-
through the boundary k 1 1/2 plays an important role tive, thus retarding the evolution of the raindrop spec-
for the numerical diffusivity of the flux method. The trum. These findings suggest that gmin 5 1028 g m23
best results are obtained with LFM. However, from Figs. seems to be an appropriate choice for the flux limiter
2b and 3b it is seen that LFM still suffers from some in (27).
numerical diffusion, thus producing a broader drop In order to test the performance of the two linear flux
spectrum than BRM. From the results of PFM one may methods LFM and MFM, numerous sensitivity studies
conclude that the numerical diffusivity of LFM could with different choices of L and r have been carried out.
further be reduced if in some situations the mass trans- Two examples for these tests are presented in Figs. 5a,b,
port from grid box k to k 1 1 is omitted, which is and 6a,b, and depicting the results for L 5 2 g m23 , L
equivalent to setting f k11/2 (i, j) 5 0. Since the numerical 5 15 mm and L 5 0.5 g m23 , r 5 10 mm. For the
diffusion is most evident for small values of g it seems collection kernel in (26) the same collision efficiencies
appropriate to modify in LFM the flux limitation in such are used as described in Seeßelberg et al. (1996). For

Unauthenticated | Downloaded 03/30/22 11:16 AM UTC

1 JULY 1998 BOTT 2289

FIG. 3. (a) Solution to the stochastic collection equation for the hydrodynamic kernel, L 5 1 g m23 , and r 5 10 mm with different
versions of the flux method in comparison to BRM (full curve). Short dashed curve: UFM; long dashed curve: LFM; dash-dotted curve
PFM. Linear scale of the ordinate. Curves are shown after 30 and 60 min. (b) Same as (a) except with a logarithmic scale of the ordinate.
Curves are shown after 30 min.

small droplets these data are taken from Davis (1972) tive cloud. Due to the high liquid water content and the
and Jonas (1972), whereas for larger drops the dataset large mean radius of the initial cloud droplet distribution
of Hall (1980) has been utilized. The first case (Figs. the collision process is very efficient, yielding a fast
5a,b) may be interpreted as being typical for a convec- evolution of the raindrop spectrum. The lower liquid
water content and the smaller mean radius of the second
model run (Figs. 6a,b) are more characteristic for strat-
iform clouds. Here the evolution of the drizzle spectrum
occurs more slowly than in the convective cloud. Both
test runs reveal the following characteristic behavior of
LFM and MFM. 1) At all times the numerical results
agree fairly well with BRM. 2) The largest deviations
between the flux method and BRM are observed at the
time where the second maximum of the mass distri-
bution evolves. This seems to be a critical stage in the
development of the raindrop spectrum. 3) With increas-
ing time, the mass distributions converge toward a qua-
si-steady state. During this time, the agreement between
LFM, MFM, and BRM is continuously improving. 4)
MFM and LFM produce very similar curves except at
the right tail of the drop spectra where the numerical
diffusion of MFM is clearly lower than in LFM.
The modified flux limitation (27) has also been ap-
plied to UFM and PFM. Similar to LFM, the results
were improved for 1028 # gmin # 1026 . However, the
unmodified linear flux method is still better than the
modified UFM and PFM versions. Thus it is concluded
FIG. 4. Solution to the stochastic collection equation for the hy- that the success of the flux method is mainly based on
drodynamic kernel, L 5 1 g m23 , and r 5 10 mm with different the particular choice of the polynomial fitting technique
versions of the flux method in comparison to BRM (short dashed than on the use of the flux limiter gmin .
curve). Curve 1: LFM; curve 2: MFM with gmin 5 1029 ; curve 3:
MFM with gmin 5 1028 ; curve 4: MFM with gmin 5 1027 ; curve 5:
In principle it is possible to fit the mass distribution
MFM with gmin 5 1026 ; curve 6: PFM. Curves are shown after 30 in each grid box with other functions than the poly-
min. nomials of (17). For instance, one could also take an

Unauthenticated | Downloaded 03/30/22 11:16 AM UTC

2290 JOURNAL OF THE ATMOSPHERIC SCIENCES VOLUME 55

FIG. 5. (a) Solution to the stochastic collection equation for L 5 2 g m23 and r 5 15 mm with different versions of the flux method in
comparison to BRM (full curve). Short dashed curve: LFM; long dashed curve: MFM. Curves are shown after 10 and 20 min. (b) Same as
(a) except with a logarithmic scale of the ordinate.

exponential function that has already been applied by 4. Numerical stability and convergence
Soong (1974). This treatment has the advantage that no
undershooting occurs in the distributions like this is the This section deals with the investigation of the sen-
case in PFM. In fact, the flux method has also been sitivity of the flux method regarding a variation of the
performed by fitting the mass distributions as iterative collision procedure, of the integration timestep,
and the numerical grid mesh. Since in the iterative pro-
g i 5 a i exp(b i z), (28) cedure as described in the previous section the drop
where a i and b i have been determined according to spectrum is updated after each collision process, one
Soong. However, the model results obtained with the may argue that a variation of the collision sequence may
exponential fit were not better than those of LFM. alter the numerical results. In principle four different
Therefore their presentation has been omitted in this collision sequences are possible by changing the direc-
paper. tions of the running indices i and j of the colliding drops:

FIG. 6. (a) Same as Fig. 5a except L 5 0.5 g m 23

and r 5 10 mm. Curves are shown after 40 and 60 min. (b) Same as (a) except with a
logarithmic scale of the ordinate.

Unauthenticated | Downloaded 03/30/22 11:16 AM UTC

1 JULY 1998 BOTT 2291

FIG. 7. Solution to the stochastic collection equation for the hy- FIG. 8. Solution to the stochastic collection equation for the hy-
drodynamic kernel, L 5 1 g m23 , and r 5 10 mm with BRM. Com- drodynamic kernel, L 5 1 g m23 , and r 5 10 mm with MFM. Com-
parison of different time steps. Full curve: Dt 5 0.1 s; short dashed parison of different time steps. Full curve: Dt 5 0.1 s; short dashed
curve: Dt 5 2 s; long dashed curve: Dt 5 5 s; dash-dotted curve: Dt curve: Dt 5 10 s; long dashed curve: Dt 5 30 s; dash-dotted curve:
5 10 s. Curves are shown after 30 and 60 min. Dt 5 60 s. Curves are shown after 30 and 60 min.

1) i 5 i 0 , . . . , i1 and j 5 i, . . . , i1 ; 2) i 5 i 0 , . . . , i1 is concluded that BRM should not be used with too

and j 5 i1 , . . . , i; 3) i 5 i1 , . . . , i 0 and j 5 i, . . . , i1 ; large a time step.
and (4) i 5 i1 , . . . , i 0 and j 5 i1 , . . . , i. Sensitivity Figure 8 shows the model results for MFM with time
studies with LFM and MFM have been carried out in steps Dt 5 0.1, 10, 30, and 60 s. Again the curves are
which the four combinations have been compared. The graphically identical if Dt is varied between 0.1 and 1
studies showed that the numerical results are practically s. As can be clearly seen from the figure, MFM produces
not affected by a particular choice of the collision se- very similar results even for large time steps exceeding
quence (not shown). 30 s. For these time steps, after 30 min the evolution
Of particular interest is the model sensitivity regard- of the raindrop spectrum is somewhat accelerated. How-
ing the integration time step Dt since in a dynamic cloud ever, the most interesting feature of this sensitivity study
model the numerical treatment of the coalescence pro- is the fact that for Dt 5 10 s the mass distributions are
cess is computationally relatively time consuming. From still very similar to those of Dt 5 0.1 s (compare the
BRM it is known that the model is not mass conservative short-dashed curve with the solid curve). Numerous oth-
whereby the error is increasing with increasing time er test cases (not shown) revealed the same stability
step. Figure 7 shows the results of BRM that have been behavior of the flux method, thus justifying the use of
obtained with Dt 5 0.1, 2, 5, and 10 s. The initial Dt 5 10 s for the numerical calculations. Finally, it is
conditions are the same as in the second model run (see also noteworthy that in all calculations presented in this
Fig. 3). For Dt 5 1 s the results are graphically identical paper and also in all other model runs that have been
to those of Dt 5 0.1 s. With increasing time step, the performed so far it was never necessary to apply re-
method produces less satisfactory results. After 30 min striction (22) or (23) on the time step. This is explained
artificial ripples are observed in the raindrop spectrum. by the fact that large values of the collection kernel
Sometimes this behavior is observed even with Dt 5 1 occur only if small droplets are collected by large drops.
s; see, for example, Fig. 6a. A reason for this is probably In this case the number concentration of the large drops
the use of high-order polynomials in regions with very is usually very small so that g9j (i, j 2 1) in the denom-
steep gradients of the mass distribution. Furthermore, inator of (22) is small. Furthermore, for large values of
for large timesteps the evolution of the raindrop spec- K(i, j) in most cases we have j 5 k so that (23) does
trum is clearly retarded. After 60 min the mass loss is not apply. If the colliding drops have similar size, that
less than 1% for Dt 5 1 s and around 4% for Dt 5 10 is, j ± k, then the terminal velocities of the falling drops
s. It should, however, be noted that in another model are similar so that the collection kernel is small; see Eq.
run where the liquid water content was increased to L (26).
5 2 g m23 , the use of Dt 5 10 s resulted in a mass loss The sensitivity of the flux method to a variation of
of 40% during the first 30 min. From these findings it the numerical grid mesh was also investigated. Figure

Unauthenticated | Downloaded 03/30/22 11:16 AM UTC

2292 JOURNAL OF THE ATMOSPHERIC SCIENCES VOLUME 55

FIG. 9. (a) Solution to the stochastic collection equation for the hydrodynamic kernel, L 5 1 g m23 , and r 5 10 mm with MFM.
Comparison of different grid meshes. Full curve: BRM; short dashed curve: a 5 21/2 ; long dashed curve: a 5 21/4 ; dash-dotted curve: a
5 21/8 . Curves are shown after 30 and 60 min. (b) Same as Fig. 9a except for PFM.

9a shows the corresponding results of MFM as obtained 5. Conclusions

with the hydrodynamic kernel, L 5 1 g m23 and r 5
10 mm. Three model runs are presented with a 5 21/n In the present paper a new mass conservative method
and n 5 2, 4, 8. It is seen that a refinement of the grid was introduced for the numerical solution of the sto-
mesh decelerates the evolution of the raindrop spectrum. chastic collection equation SCE. The method consists
With increasing value of n the curves converge toward of a two-step procedure. In the first step the mass dis-
the solution of BRM (solid line). Similar to the previous tribution of drops with mass x9 that have been newly
sensitivity studies the differences between the three formed in a collision process is added to grid box k with
curves decrease with increasing time. Figure 9b shows x k # x9 # x k11 . By solving an advection equation, in
the corresponding results as obtained with PFM. Con- the second step a certain fraction of the cloud water
sidering the fact that the results are rather poor for n 5 mass is transported from k to k 1 1.
2 it is surprising how fast the distributions improve with For the solution of the advection step three different
increasing n. The reason for this behavior is that the use versions of the positive definite advection scheme of
of a fine grid mesh reduces the number of undershoot- Bott (1989b) were utilized. The upstream solution UFM
ings of the second-order polynomials yielding f k11/2 (i, was numerically too diffusive while the parabolic flux
j) , 0 in (18) so that the nonlinear flux limitation (21) method PFM clearly retarded the evolution of the rain-
is performed less often than this is the case for n 5 2. drop spectrum because of its strong antidiffusive char-
Due to the possibility of using relatively large inte- acter. Among the three model versions the linear flux
gration time steps, the flux method is computationally method LFM produced the best results. In an attempt
very efficient. By comparing for a 5 21/2 the CPU time to decrease the numerical diffusivity of LFM, the pos-
consumption of BRM as obtained with Dt 5 1 s with itive definite flux limitation was modified yielding the
the linear flux method the following speed up factors e modified linear flux method MFM.
were obtained: In different numerical test studies the results of LFM
and MFM were compared with the Berry–Reinhardt
e 5 3.9 for Dt 5 1 s method (BRM). The agreement between both methods
e 5 33.8 for Dt 5 10 s. (29) and BRM was generally very good. Numerical sensi-
tivity studies were presented showing that the flux meth-
Larger time steps increase the value of e almost linearly. od behaves numerically stable for different choices of
Since in LFM a time step of 10 s may be used without the numerical grid as well as different values of the
losing numerical accuracy, the CPU time saving is con- integration time step. For 0.1 # Dt # 10 s the curves
siderable. The numerical efficiency, the accuracy, and were very similar thus allowing a relatively large time
the fact that LFM is absolutely mass conserving are step for the numerical solution of the SCE. Utilizing in
certainly the most important advantages of the scheme. LFM or MFM Dt 5 10 s yielded a decrease of the CPU

Unauthenticated | Downloaded 03/30/22 11:16 AM UTC

1 JULY 1998 BOTT 2293

time consumption by a factor of about 34 compared to by collection: Part I. Double distributions. J. Atmos. Sci., 31,
BRM. 1814–1824.
Bleck, R., 1970: A fast, approximative method for integrating the
The exact mass conservation as well as the numerical stochastic coalescence equation. J. Geophys. Res., 75, 5165–
stability and efficiency of the flux method are the most 5171.
important advantages when the model is compared with Bott, A., 1989a: A positive definite advection scheme obtained by
the Berry–Reinhardt scheme. Therefore, LFM or MFM nonlinear renormalization of the advective fluxes. Mon. Wea.
Rev., 117, 1006–1015.
are very attractive alternatives to BRM for solving the , 1989b: Reply. Mon. Wea. Rev., 117, 2633–2636.
SCE in dynamic cloud models dealing with explicit mi- Davis, M. H., 1972: Collisions of small droplets: Gas kinetic effects.
crophysics. J. Atmos. Sci., 29, 911–915.
Golovin, A. M., 1963: The solution of the coagulation equation for
cloud droplets in a rising air current. Izv. Akad. Nauk. SSSR. Ser.
Acknowledgments. The author wishes to thank Dr. T. Geofiz., 5, 783–791.
Trautmann for many helpful discussions in this research Hall, W. D., 1980: A detailed microphysical model within a two-
and for providing the Berry–Reinhardt computer code. dimensional dynamic framework: Model description and prelim-
The constructive criticism of an anonymous reviewer inary results. J. Atmos. Sci., 37, 2486–2507.
Jonas, P. R., 1972: The collision efficiency of small drops. Quart. J.
who discovered an error in the first model version are Roy. Meteor. Soc., 98, 681–683.
also acknowledged. A FORTRAN code of the flux meth- Kovetz, A., and B. Olund, 1969: The effect of coalescence and con-
od is available upon request via e-mail from the author densation on rain formation in a cloud of finite vertical extent.
at [email protected]. This research is fund- J. Atmos. Sci., 26, 1060–1065.
Long, A., 1974: Solutions to the droplet collection equation for poly-
ed by the Deutsche Forschungsgemeinschaft, SFB 233, nomial kernels. J. Atmos. Sci., 31, 1040–1052.
Chemie und Dynamik der Hydrometeore. Pruppacher, H. R., and J. D. Klett, 1997: Microphysics of Clouds and
Precipitation. Kluwer Academic, 954 pp.
Seeßelberg, M., T. Trautmann, and M. Thorn, 1996: Stochastic sim-
REFERENCES ulations as a bench-mark for mathematical methods solving the
coalescence equation. Atmos. Res., 40, 33–48.
Beard, K. V., 1976: Terminal velocity and shape of cloud and pre- Soong, S.-T., 1974: Numerical simulation of warm rain development
cipitation drops aloft. J. Atmos. Sci., 33, 851–864. in an axisymmetric cloud model. J. Atmos. Sci., 31, 1262–1285.
Berry, E. X., 1967: Cloud droplet growth by collection. J. Atmos. Tzivion, S., G. Feingold, and Z. Levin, 1987: An efficient numerical
Sci., 24, 688–701. solution to the stochastic collection equation. J. Atmos. Sci., 44,
, and R. L. Reinhardt, 1974: An analysis of cloud droplet growth 3139–3149.

Unauthenticated | Downloaded 03/30/22 11:16 AM UTC