Turbulent Diffusion
A Venkatram, University of California – Riverside, Riverside, CA, USA
S Du, California Air Resources Board, Sacramento, CA, USA
Ó 2015 Elsevier Ltd. All rights reserved.
This article is reproduced from the previous edition, volume 6, pp 2455–2466, Ó 2003, Elsevier Ltd.
Glossary
1 AU Astronomical Unit ¼ the mean Earth–Sun separation.
Dobson Unit Unit for total column ozone
(1 DU ¼ 2.69  1016 molecule cm2). A column amount
of 300 Du, which is a typical global average that
corresponds to a 3-mm layer of pure ozone at standard
temperature and pressure.
RAF Radiative amplification factor.
Nomenclature
~ Instantaneous concentration.
C t Travel time from source to receptor.
C Mean concentration.
c Concentration fluctuation defined as C ~  C.
Cpeak Peak concentration.
y
C Crosswind-integrated concentration.
D Molecular diffusivity of species in air.
hs Source height.
H0 Surface heat flux.
K Eddy diffusivity.
L Monin–Obukhov length.
l Length scale characterizing turbulent fluctuations.
N Brunt–Vaisala frequency.
Q Source strength for continuous emissions.
Qm Total mass of a puff.
RLv Lagrangian autocorrelation for crosswind velocity.
s Puff dimension (or spread).
Tav Averaging interval.
Tc Timescale of concentration fluctuations.
TLv Lagrangian timescale for crosswind velocity.
TLw Lagrangian timescale for vertical velocity.
Introduction
Turbulent diffusion or, more correctly, turbulent dispersion
addresses the problem of estimating the concentration field in
a turbulent flow. Our ignorance of the details of the turbulent
velocity field translates into uncertainty in estimating the
details of the concentration field. Thus, as in other problems
related to turbulence, the goal of turbulent dispersion is limited
to understanding the ensemble-averaged statistics of the
concentration field and the associated fluxes that govern the
field.
Encyclopedia of Atmospheric Sciences 2nd Edition, Volume 6 http://dx.doi.org/10.1016/B978-0-12-382225-3.00441-2
SZA Solar zenith angle: the angle between the local vertical
and the Sun (SZA ¼ 90  solar elevation).
SPF Sun protection factor (for sunscreens).
TOMS Total Ozone Mapping Spectrometer (a satellite-
borne ozone sensor).
~ Instantaneous wind speed.
U
U Average wind speed.
u Wind speed fluctuation defined as U ~  U.
u* Surface friction velocity.
ws Gravitational settling velocity of a heavy particle.
d Deviation of time-averaged mean about the ensemble
mean.
3 Rate of dissipation of turbulent kinetic energy.
r Density of air.
sc Standard deviation of concentration fluctuations.
sv Standard deviation of crosswind velocity component.
sw Standard deviation of vertical velocity component.
sy Lateral dimension (or spread) for Gaussian plume
model.
sz Vertical dimension (or spread) for Gaussian plume
model.
s Surface shear stress.
sLeff Effective Lagrangian timescale of a heavy particle.
sparticle Relaxation timescale of a heavy particle.
Most of our understanding of turbulent dispersion is couched
in terms of semiempirical models that have been developed by
fitting tentative theories to observations. For example, models for
dispersion in the surface boundary layer are based on tracer
experiments conducted in Prairie Grass, Nebraska, USA, in the
1950s. The development of models for dispersion in the
convective boundary layer (CBL) has been guided by laboratory
experiments conducted by Willis and Deardorff in the 1970s.
Thus, to a large extent, our understanding of turbulent dispersion
consists of a patchwork of semiempirical models, each of which
describes a limited set of observations. The primary goal of this
277
278 Turbulence and Mixing j Turbulent Diffusion
article is to provide the reader an understanding of these semi-
empirical models. It is only over the past 10 years that attempts
have been made to develop methods that can be applied to a large
class of problems. These methods will be discussed in the last part
of this article.
This article assumes that the reader is familiar with the
fundamentals of turbulence in the atmospheric boundary layer.
To keep this article to a manageable size, we will not deal with
several important topics including the effects of turbulence on
chemical reactions. We begin with the statement of the problem
that the techniques of turbulent dispersion attempt to solve.
The Problem of Turbulent Dispersion
The evolution of the concentration field of a species is governed
by the mass conservation equation, eqn [1], where the
‘squiggle’ overbar refers to the instantaneous field, and D is the
molecular diffusivity of the chemical species in the fluid.
~
vC v
~ ~ ~
v2 C
þ Ui C ¼ D [1]
vt vxi vxi vxi
Because the instantaneous velocity field is unknown, we
have to limit ourselves to estimating the ensemble-averaged
mean and the associated statistics of the concentration field.
To define an ensemble, we express the instantaneous
velocity field as the sum of a known velocity field, Ui, and the
unknown deviation, ui, as in eqn [2].
~ i ¼ Ui þ u i
U [2]
Then, an ensemble is defined as the infinite set of possible
concentrations corresponding to a given Ui. Note that the
definition of an ensemble is arbitrary because it depends on
what we know about the velocity field. Each member of this
ensemble corresponds to the component of the unknown
velocity field, ui; we assume that we know something about the
statistics of this unknown field. The concentration associated
with each member of the ensemble can be written as eqn [3],
where C is the average obtained over all possible concentration
fields in the ensemble.
~ ¼ Cþc
C [3]
Substituting eqn [3] into eqn [1] and averaging over the
ensemble yield eqn [4], where the overbar refers to the
ensemble average, the subscripted index refers to coordinate
direction, and repeated indices imply summation. This equa-
tion can be solved only by modeling the turbulent flux term,
ui c, using known properties of the velocity field. Before we take
up models for this term, let us consider the problem in
turbulent dispersion that has to a large extent motivated the
development of the field.
vC v   v2 C
þ Ui C þ ui c ¼ D [4]
vt vxi vxi vxi
The Point Source in the Atmospheric Boundary Layer
The classical problem of turbulent dispersion in the atmo-
spheric boundary is that of a continuous source emitting
material at some height above the ground (see Figure 1). If the
ground is taken to be the reference height, with the x-axis of
the coordinate system aligned along the wind direction at the
source, empirical evidence indicates that the time-averaged
(typically 1 h) concentration field can be described in terms
of the Gaussian distribution by eqn [5], where y is the cross-
wind coordinate, Q is the source strength (mass/time), hs is the
height of the source above ground, U is the time-averaged wind
speed at source height, and sy and sz are the plume spreads
corresponding to the Gaussian distribution.
" #
Q ðz  hs Þ2 y2
Cðx; y; zÞ ¼ exp   2 [5]
2psy sz U 2s2z 2sy
The effect of the ground on concentrations is accounted for
by making sure that there is no flux of material through the
plane at z ¼ 0. The mathematical trick to achieve this is to
place an ‘image’ source at a distance z ¼ hs; the upward flux
from this image source essentially cancels out the downward
flux from the real source. Then, the concentration can be
described by eqn [6].
" #
Q y2
Cðx; y; zÞ ¼ exp  2
2psy sz U 2sy
8 " # " #9 [6]
< ðz  hs Þ2 ðz þ hs Þ2 =
 exp  þ exp 
: 2s2z 2s2z ;
In the real atmosphere, dispersion in the upward direction
is limited by the height of the atmospheric boundary layer. This
limitation of vertical mixing is incorporated into the Gaussian
formulation by reflecting off the top of the mixed layer. When
the material is ‘reflected’ from both the ground as well as the
top of the mixed layer, it is necessary to account for the infinite
set of ‘reflections’ from the two surfaces. This can be readily
accounted for in the Gaussian formulation.
The point source solution is the kernel for the integral that is
used to estimate dispersion from a variety of sources, including
line and volume sources. Most dispersion models that apply to
spatial scales of tens of kilometers are based on eqns [5] and
[6]. The plume spreads or s values are empirically derived from
observations. It turns out that we continue to use plume-spread
formulations first recommended by Pasquill in the 1960s.
These formulations, largely based on the Prairie Grass experi-
ment, relate plume spread to surface meteorological measure-
ments such as wind speed and cloud cover. The advances in
micrometeorology during the 1970s provided the incentive to
develop dispersion models that relied on theoretical under-
standing of dispersion.
The preceding expression for the concentration field is
essentially an empirical description of observations. Equation
[5] is a formal solution to the point source problem only when
the turbulence field is homogeneous, and the velocity distri-
bution is Gaussian. Under these circumstances, the concentra-
tion field can be analyzed using a statistical approach, first
proposed by Taylor.
Statistical Analysis of Dispersion
Consider a source located at ‘s’ emitting particles continuously
into a turbulent flow. If the mean flow and turbulence are

Plume
hs
z x and y directions
Figure 1 Turbulent dispersion from a point source.
steady, the ensemble-averaged concentration at ‘r,’ C(r) can be
shown to be
ZN
CðrÞ ¼ Q pðrjs; tÞdt
0
where Q is the mass emission rate of particles, and p(rjs,t)dV is
the probability that a particle released at ‘s’ will be found in
a volume dV surrounding ‘r’ after a travel time t from release.
Then, the problem of calculating the concentration reduces to
estimating the probability density function, p(rjs,t), of particle
positions as a function of travel time from the source. A good
approximation for this function is the Gaussian distribution.
Placing the origin of our coordinate system at the source, we
can express the distribution as
1 is sufficient to interpret the timescale as roughly the time over
pðrjs; tÞ ¼ 3=2
ð2pÞ sx sy sz
!!
1 ðxr  xÞ2 ðyr  yÞ2 ðzr  zÞ2
 exp  þ þ
2 s2x s2y s2z
where sx, sy, and sz are the standard deviations of particle
positions about their mean positions x, y, and z after a travel
time, t, from release, and xr, yr, and zr are the receptor coordi-
nates. These statistics are derived by averaging over an infinite
number of particles for a flow with fixed mean and turbulent
characteristics; the statistics are functions of travel time, t.
The probability distribution function, eqn [8], represents an
ensemble average over all possible particle positions for a fixed
travel time from the source; the travel time is the difference
between the arbitrary time at which the particle is released and
time at which the particle is observed at a location. In principle,
it can be constructed experimentally by releasing particles
serially from a source, and recording the coordinates of these
particles at specified travel times. Thus, eqn [8] does not
Turbulence and Mixing j Turbulent Diffusion 279
z y
Concentration
distribution
is assumed to
be Gaussian in
describe the distribution of particle positions within a ‘puff, ’
which usually describes an entity at an instant of time.
To make progress, we need expressions for sx , sy , and sz.
While the mean particle positions are determined by the mean
[7]
flow, the standard deviations depend on the characteristics of
the turbulent flow. Taylor derived expressions for the variance
of particle positions as a function of travel time from a fixed
release point in a steady flow in which turbulent statistics do
not depend on location. His expressions for the asymptotic
behavior of plume spread are
sy ¼ sv t for t  TLv
[9]
sy ¼ sv ð2tTLv Þ1=2 for t[TLv
where TLv is the Lagrangian timescale, which can formally defined
in terms of the statistics of the turbulent flow. For our purposes, it
which a particle retains its initial velocity. For small travel times,
a particle’s velocity remains essentially unchanged from its value
at the release point, and the particle trajectory is a straight line.
This explains the result that, for small travel times, the spread of
particles is proportional to the travel time from the source (eqn
[8]
[9]). On the other hand, when the travel time is large compared
to the Lagrangian timescale, the plume spread is proportional to
the product of the ‘average’ step size, sv TLv , and the square root of
the number of steps, t/TLv, taken by the particle.
In order to obtain an expression for the concentration, we
still have to integrate eqn [7] after inserting eqn [8] with
appropriate expressions for plume spread. Let us first consider
an idealized flow that is used to model dispersion in the
atmospheric boundary layer. In this flow, the mean wind U is
along the x-axis, and the turbulence is homogeneous and
stationary. These assumptions lead to
x ¼ Ut; y ¼ 0; and z ¼ 0 [10]
If we make the assumption that along-wind dispersion, sx,
is small compared to transport by the mean wind, the
280 Turbulence and Mixing j Turbulent Diffusion

exponential term in eqn [8], associated with downwind conservation equation. Let us examine this approach in some
dispersion, becomes a Dirac delta function in the limit of sx detail because it yields useful results for dispersion in the
going to zero. This allows us to integrate eqn [7] for arbitrary sy surface boundary layer.
and sz to obtain
!!
Q 1 y 2 z2
CðrÞ ¼ exp  þ [11] Solving the Species Conservation Equation
2psy sz U 2 s2y s2z
The species conservation equation (eqn [4]) can be rewritten as
where the plume spreads are evaluated at t ¼ xr/U. This equa-
eqn [14].
tion is identical to the empirical expression presented earlier.
We can obtain an expression for the concentration even when vC v v   v2 C
þ ðUi CÞ ¼ ui c þ D [14]
downwind dispersion is not small if we can express plume vt vxi vxi vxi vxi
spreads in terms of the asymptotic limits of eqn [9].
One way of modeling the turbulent flux term is to postulate
Observations of plume spread from elevated releases are
the concept of eddy diffusivity. It is based on an analogy with
often summarized in the form
molecular diffusion, in which the flux of material in any
sv t direction is proportional to the gradient of the concentration.
sy ¼ [12]
ð1 þ t=2TLv Þ1=2 For example, the turbulent flux of species is according to eqn
[15], where Ki is the so-called eddy diffusivity.
to ensure consistency with theory of eqn [9]. For ground-level
releases, there is no simple way of relating travel time to vC
turbulent flux ¼ ui c ¼ Ki [15]
distance because the velocity varies rapidly with height near the vxi
ground. Under these circumstances, we need to use methods
where the bar over i negates the summation convention. This
that are discussed next.
relationship cannot be justified rigorously for turbulent trans-
port. However, it has heuristic value, and is useful for devel-
Dispersion in an Inhomogeneous Boundary Layer oping semiempirical models of turbulent transport. The use of
the eddy diffusivity in the mass conservation equation is often
The theory presented thus far applies to a boundary layer in referred to as K-theory.
which the mean and turbulent properties are constant in space With eqn [15], eqn [14] can be rewritten in the form of eqn
and time. To apply it to a real boundary layer in which the [16],
properties are highly inhomogeneous, we can use one of two  
vC v v vC
approaches. The first is to average the turbulence and mean þ ðUi CÞ ¼ Ki [16]
vt vxi vxi vxi
properties over the region of interest, and use the average
properties in the (homogeneous) formulations discussed where we have neglected the molecular diffusion term in
earlier. This is the most straightforward approach, except that comparison to turbulent diffusion. While molecular diffusion
the averaging procedure is necessarily arbitrary. The validity of can often be ignored in calculating the ensemble mean, it plays
the method needs to be established by comparing the results a major role in determining the statistics of concentration
obtained from the formulations with observations or theory fluctuations, as we will see later.
that accounts for inhomogeneity more explicitly. In general, One way of checking whether the use of the eddy diffusivity
empirical knowledge derived from observations plays a major is plausible is to see whether eqn [16] yields solutions that are
role in the development of practical models of dispersion. As in compatible with observations. We will apply eqn [16] to the
most turbulence research, theory can suggest plausible forms point source problem. If we assume that transport along the
for a dispersion model, but the model almost always contains mean wind dominates over the corresponding diffusion term,
parameters that have to be estimated from observations. and turbulent properties are homogeneous, eqn [16] can be
Even if we could treat the boundary layer as vertically reduced to the form given by eqn [17].
homogeneous, the presence of boundaries, such as the ground vC v2 C v2 C
and the top of the mixed layer, makes it difficult to estimate the U ¼ Kz 2 þ Ky 2 [17]
vx vz vy
Lagrangian timescale, TLv, from a priori considerations. Thus,
the timescale is often treated as an empirical parameter that is It turns out that eqn [17] yields the empirical Gaussian
derived by fitting eqn [12] to observations. Alternatively, we solution of eqn [5] if we ensure that the eddy diffusivity is
can postulate an expression for TLv in terms of a known length related to the plume spread according to eqn [18].
scale l as shown in eqn [13].
1 ds2i
Ki ¼ U [18]
TLv ¼ al=sv [13] 2 dxi
The parameter a has to be obtained by fitting estimates of If we take plume spread to follow the behavior described in
plume spread to observations. In unstable conditions, l, usually eqn [9], we see from eqn [18] that the eddy diffusivity is
scales with the depth of the boundary layer, while in stable proportional to the travel time, x/U, from the source, for travel
conditions, the relevant length scale is taken to be sw =N, where times less than the governing Lagrangian timescale. What this
N is the Brunt–Vaisala frequency. means is that the eddy diffusivities corresponding to two
The second approach to accounting for inhomogeneity in different sources displaced along the wind will have different
the boundary layer is based on the solution of the species values at the same location. It is only at large travel times that

eddy diffusivities become independent of travel time (eqn
[19]), where TLv and TLw are the Lagrangian timescales for the
horizontal and vertical velocity fluctuations, respectively.
Ky ¼ s2v TLv and Kz ¼ s2w TLw [19]
The eddy diffusivity, Kz, can be related to turbulent flow
properties by appealing to ‘mixing length’ theory, which
suggests the relationship [20], where sw is the standard devia-
tion of the vertical velocity fluctuations, and lz is the ‘length
scale’ of turbulence for vertical transport, defined by eqn [21],
which is consistent with eqn [13].
Kz ¼ sw lz [20]
lz ¼ sw TLw [21]
We are now in a position to make some additional state-
ments on the applicability of the eddy diffusivity concept. We
saw earlier that eqn [19] is valid when the travel time is much
larger than the Lagrangian timescale, expressed by the rela-
tionship [22].
x
t ¼ [TLw [22]
U
If we combine this condition with the expression for plume
spread, eqn [9], and use eqn [21], we obtain eqn [23].
sz [lz [23]
Thus, the eddy diffusivity concept is most applicable when
the scale of concentration variation, sz, is much larger than the
scale of the eddies responsible for plume spreading.
Equation [20] is useful because we can guess at the
appropriate form of lz, and then see whether the consequences
of our assumption agree with observations. Over the years, we
have developed enough experience with different types of
flows to prescribe useful forms for the mixing length (or eddy
diffusivity) for these flows. Our initial guesses for K are usually
based on measurement of fluxes and the associated gradients
for a limited set of situations. This K is then extrapolated to
situations different from those used to derive it. For example,
we can derive a K for heat flux, and find out whether it works
for pollutant transport. It is this type of semiempirical argu-
ments that form the basis of practical calculations for turbu-
lent flows.
Equation [15] represents only one possible approach to
expressing the turbulent flux. In principle, we can write
conservation equations for the turbulent fluxes, but these
equations contain ‘third-order’ terms that are essentially the
fluxes of the second-order terms. These third-order terms have
to be parameterized using some sort of flux-gradient approxi-
mation. At this point, there is no compelling evidence to
suggest that these approaches yield much better results than the
closure of eqn [15]. One way of improving upon a simple
prescription of the eddy diffusivity is to formulate semi-
empirical conservation equations for the components of eddy
diffusivities: the turbulent velocity and the length scale in eqn
[20]. In practice, the turbulent velocity is related to the turbu-
lent kinetic energy, k, and the length scale is related to the
turbulent dissipation rate, 3 . While the k  3 approach is
popular in modeling turbulent flows, it has found limited
application in modeling dispersion.
Turbulence and Mixing j Turbulent Diffusion 281
The eddy diffusivity formulation is almost exclusively used
in comprehensive Eulerian air quality models that include
details of atmospheric process such as gas- and aqueous-phase
chemistry. The main reason is that the species conservation
equation, formulated in terms of the eddy diffusivity, is
a convenient framework to incorporate a number of processes,
including nonlinear chemistry. The resulting mass conservation
can be solved using numerical methods. Results from compre-
hensive air quality models indicate that modeling dispersion
with the eddy diffusivity model has some practical value, even
though the underlying theoretical justification is weak.
We saw earlier that the eddy diffusivity concept is likely to be
most applicable when the turbulent length scales are smaller than
or comparable to the concentration space scales. Thus, we might
expect it to apply to dispersion in the surface boundary layer,
where plume spread in the vertical spread is comparable to the
length scale of the eddies responsible for vertical transport. It
turns out that K-theory provides useful results for dispersion in
the surface boundary layer even though it is characterized by steep
gradients in temperature and velocity. However, the gradients of
fluxes and turbulence levels are negligible. In the surface
boundary layer, semiempirical theories, referred to as Monin–
Obukhov similarity, provide useful relationships between
velocity and temperature gradients and the corresponding heat
and momentum fluxes. These relationships are cast in terms of
length and velocity scales, which are the surface friction velocity
u* and the Monin–Obukhov length, L, defined by eqns [24],
rffiffiffi
s
u ¼
r
[24]
T0 u3 rCp
L ¼ 
g kH0
where s is the surface shear stress, r is the air density, Cp is the
specific heat of air, T0 is the surface temperature, k is the von
Karman constant, g is the acceleration due to gravity, and H0 is
the surface heat flux. These relationships can be used to derive
eddy diffusivities for heat and momentum.
Using the eddy diffusivity for heat in the mass conservation
equation has provided concentration estimates that compare
well with observations made in field experiments conducted in
Prairie Grass, Nebraska, in the 1950s. We note that data from
this experiment, conducted with relatively primitive equip-
ment, are still the most complete for the analysis of surface
layer dispersion.
The solutions of the mass conservation equations, using the
eddy diffusivity, have a number of useful asymptotic forms for
y
the crosswind-integrated concentration, C , as shown by eqn
y
[25], where x ¼ x=jLj C ¼ C u jLj=Q, with Q representing the
source strength. These asymptotic forms are useful because they
can be patched together to obtain analytic expressions that
span the entire range of stability. Unknown parameters in these
expressions have been obtained by fitting them to observations
from Prairie Grass.
C  x1 for neutral conditions
 x2=3 for stable conditions [25]
 x2 for unstable conditions
These expressions for crosswind-integrated concentrations
can be converted to yield centerline concentrations through
282 Turbulence and Mixing j Turbulent Diffusion
eqn [26], where sy is the crosswind spread, and the crosswind
distribution is taken to be Gaussian.
y
! released from origin at time 0 by eqn [30], where Qm is the total
C y2
Cðx; y; 0Þ ¼ pffiffiffiffiffiffiffiffiffiffiffi exp  2 [26]
2psy 2sy
Equation [25] can be used to derive expressions for the
vertical plume spread. These expressions depend on distance
from the release location because travel time has little meaning
near the ground. In fact, most dispersion models used in
practical applications express the concentration in terms of
a Gaussian distribution, where the plume spreads are empir-
ically derived functions of source–receptor distance and
micrometeorology.
Puff Dispersion
In the preceding sections, we have discussed dispersion of
plumes, which refers to a continuous release from a source.
Often we are interested in concentrations associated with puffs
of material that are released almost instantaneously, as in an
explosion. The concentration of material in the plume is
determined by the spread of the material about the center of
mass of the moving puff. The analysis of such puffs is more
complicated than that of plumes because the spread depends
on both space and time correlations between particles in the
puff; as Taylor’s analysis indicates, plume dispersion can
consider the motion of particles to be independent of each
other. These space–time correlations depend on the properties
of the turbulent eddies that contribute to puff spread at any
instant of time. The length scale of the relevant eddies is
roughly proportional to the size of the puff. Eddies smaller
than this length scale disperse the material within the puff,
while eddies larger than the puff transport the puff as a whole.
When the puff size is of the order of the Kolmogorov
microscale, the puff spreads by molecular diffusion, which
implies that the puff spread, s, is proportional to the square
root of time. When the puff size is comparable to eddies in
the inertial subrange, dimensional considerations suggest that
the rate of puff spread is represented by eqn [27], where 3 is the
dissipation rate of turbulent kinetic energy.
ds
 ðεsÞ1=3 [27]
dt
Integration yields expression [28].
s  ε1=2 t 3=2 [28]
This rapid growth phase ends when the puff size is
comparable to the largest eddy of dimension L. Then, the puff
spread can be written as [29], where sv refers to the standard
deviation of the turbulent velocity fluctuations in the direction
of the spread.
s  ðsv LÞ1=2 t 1=2 [29]
These three regimes of puff growth can, in principle, be
patched together to provide a continuous description of puff
spread. But this is rarely done in practice, and one usually
resorts to empirical descriptions of puff spread. The actual
concentration in a puff is usually estimated with a Gaussian
distribution about the puff center of mass.
The simplest puff model is the Gaussian puff model that
relates concentration at receptor (x, y, z) at time t due to a puff
mass of the puff and s is the puff spread corresponding to the
Gaussian distribution.
" #
Qm ðx  UtÞ2 þ y2 þ z2
Cðx; y; z; tÞ ¼ exp  [30]
ð2pÞ3=2 s3 2s2
In practical applications, a continuous release in a wind
field that varies in space and time can be modeled through
a series of puffs, each of which is allowed to follow a different
trajectory. The concentration at a receptor at any given time is
calculated by summing the contributions from these puffs. One
advantage of this approach is that it can deal with situations
when the mean wind is calm.
Dispersion of Heavy Particles
The previous discussions implicitly assumed that the material
being dispersed by turbulence has the same density as air. This
assumption is clearly not valid for aerosol particles, whose
densities are typically over 1 g cm3. Two effects influence
dispersion of such particles. One is related to the finite time
required by the particle to respond to turbulent velocity fluc-
tuations. The other is the so-called trajectory crossing effect
related to particle trajectories being different from fluid parcel
trajectories because of gravitational settling. Let us consider
each of these effects.
The particle inertia effect is related to the difference between
the fluid velocity, uf, and the particle velocity, up, whose
difference is proportional to the reaction time to turbulent
velocity fluctuations. This can be expressed symbolically by eqn
[31], where the relaxation timescale of the particle is given by
eqn [32].


uf  up sparticle
 [31]
uf sturbulence
ws
sparticle  [32]
g
Here, ws is the gravitational settling velocity, which is
a function of the size and density of the particle and the
viscosity of the fluid. The turbulence timescale can be expressed
as [33], where l is a measure of eddy size and sw is the associ-
ated velocity fluctuation.
l
sturbulence  [33]
sw
Thus, inertia effects can be neglected if the ratio of these two
scales is small (see eqn [34]).
ws sw
1 [34]
gl
If we take small to mean 0.1, and consider 100 mm particles
with settling velocities of the order of 1 m s1, and take the
turbulent velocity to be 1 m s1, the inertia effect is not likely to
be important for length scales greater than 1 m. However, it
could play a role in the dispersion of large particles under very
stable conditions, close to the ground.

l ,
TLw ~ particle ~
w
If settling
ws
velocity ws >> w,
l
w required to bring material from the release point to the ground-
Figure 2 Timescales that affect dispersion of heavy particles.
The trajectory crossing effect can be examined by
considering the extreme case when the time taken for
a falling particle to traverse an eddy is much smaller than
the Lagrangian timescale for vertical dispersion. Then, the
effective Lagrangian timescale becomes this traversal time
because it corresponds roughly to the decorrelation time.
The effective Lagrangian timescale and the associated eddy
diffusivity for the heavy particle can now be represented by
eqn [35].
l s2 l
sLeff ¼ and Keff ¼ w
ws ws
Figure 2 shows the timescales that govern dispersion of
heavy particles. In practice, the effects of particle settling are
not important in determining plume spread because settling
velocities for most particles are generally much smaller than
turbulent velocities. However, even particle velocities of the
order of a few centimeters per second lead to mean downward
motion of the plume, and hence increase concentrations at
ground level. This effect can be described approximately by
‘tilting’ the plume toward the ground, represented by eqn
[36].
ws x
hs ðwith particlesÞ ¼ hs 
U increases in computing power, direct numerical simulation
Alternatively, the entire concentration profile can be moved
‘into’ the ground by a distance wsx/U after computing the
concentration corresponding to passive dispersion. It is simple
to account for mean motion of particles in the eddy diffusivity
formulation through the advection term, ws(vC/vz).
Other Approaches to Modeling Dispersion
The eddy diffusivity approach does not generally apply to
sources far removed from the ground. For example, it is
difficult to justify its application to the convective atmo-
spheric layer, where the turbulent length scales are large
compared to the spatial scales of the concentration field. The
pdf approach can provide useful results under these
circumstances. Studies show that dispersion in the CBL is
strongly influenced by the relative longevity of convective
downdrafts and updrafts. The majority of particles released
in downdrafts travel continuously downward until they
Turbulence and Mixing j Turbulent Diffusion 283
reach the vicinity of the ground. Numerical experiments
ws indicate that to a useful degree of approximation, the
g velocity of these particles can be taken to be constant at the
value at the release point.
This assumption allows one to express the crosswind-
y
l integrated ground-level concentration, C , in terms of the pdf
~
Leff
ws of the vertical velocity, p(wjhs), at the height of release, hs, by
eqn [37], where the vertical velocity corresponds to that
level receptor at x, given by [38].
y 2Q
C ¼ Pðwjhs Þ [37]
x
w ¼ Uhs =x [38]
The simple formulation, which can be readily modified to
account for the presence of the mixed layer, provides an
excellent description of laboratory observations of dispersion
in the CBL. Note that eqn [37] reproduces the empirical
Gaussian distribution if we make the reasonable assumption
that the pdf of vertical velocity fluctuations is normal, giving
eqn [39], where sz is expressed by eqn [40].
rffiffiffiffi
y 2 Q h2
C ¼ exp  s2 [39]
p Usz 2sz
sz ¼ sw x=U [40]
[35] The Gaussian formulation is not reliable in the CBL
because the pdf is positively skewed. The associated negative
mode of the pdf leads to the descent of the plume centerline
when the release is elevated, and leads to concentrations that
are about 30% higher than that predicted with the Gaussian
formulation.
In principle, if we could simulate all the scales of turbulent
motion, there would be no need for models of turbulent
dispersion. We could use the Navier–Stokes equations to
generate an ensemble of flows, obtain the corresponding
concentration fields from the species conservation equation,
and average over them to obtain the ensemble average as well
as the statistics of concentration fluctuations. With the rapid
[36]
(DNS) is becoming a reality, and we have been able to obtain
useful information for low Reynolds number flows. However,
it will be some time in the future before we will be able to use
the technique for routine applications.
The large eddy simulation (LES) technique avoids the
computational demands of DNS by only simulating the
energy-containing eddies. The effects of the unresolved scales of
motion are modeled using a variety of parameterizations. It is
believed that the most important features of flow are insensi-
tive to these parameterizations, because the subgrid scales
contain a small fraction of the total energy. This assumption
has been vindicated by LES of CBLs that was pioneered by
Deardorff in the 1970s. Velocity fields generated by LESs
compare well with observations, and continue to provide
information that is difficult to obtain in the field. Lamb used
the velocity fields generated by Deardorff to simulate disper-
sion in the CBL. The simulation consisted of releasing a large
number of particles and tracing their motion using the LES
velocity field. Then, the crosswind concentration averaged over
284 Turbulence and Mixing j Turbulent Diffusion
Δz
Q In homogeneous turbulence, we can show that a and b are
U given by eqns [44],
y Fraction of particles
U Δz C = that pass through Δz
Q C0 ε
Figure 3 Calculation of concentrations using Lagrangian stochastic
simulation of particle trajectories.
a vertical distance Dz is given by eqn [41], where f is the fraction
of the particles released that pass through Dz.
y Q
C ¼ f [41]
U Dz
Figure 3 justifies this equation.
Lamb’s simulations provided valuable insight into disper-
sion in the CBL, including the observation that the locus of the
maximum concentration descended toward the ground. This
behavior is related to the negative mode of the pdf of the
vertical velocity fluctuations.
A technique that is not as computationally demanding as
DNS or LES is called Lagrangian stochastic simulation (LSS). It
is attractive because it only uses the statistics of the velocity field
such as velocity variance, and dissipation rate of turbulent
kinetic energy. Because this technique is being used to examine
routine dispersion problems, we discuss it in some detail in the
next section.
Lagrangian Stochastic Models
In Lagrangian stochastic modeling, turbulent dispersion is
modeled by tracing the motion of a large number of fluid
particles that are tagged at the source; these particles are treated
mathematically as points. The evolution of the velocity of each
particle depends on turbulence properties at the current loca-
tion of the particle. To illustrate the basic ideas, let us trace
a particle that is only affected by vertical velocity fluctuations.
Then, the vertical velocity of a parcel at time t þ Dt is related to
the velocity at time t through eqn [42], where w0 is a random
velocity drawn from the distribution of vertical velocity
fluctuations.
wðt þ DtÞ ¼ awðtÞ þ bw0 [42]
Thus, the future velocity of the particle consists of
a deterministic component that depends on its current
velocity, and a random component that depends on the
turbulence at the location of the particle. We can show that
eqn [42] is equivalent to the Langevin equation in which
a particle is subject to a linear viscous force and a random
pressure force.
The particle position is traced through eqns [43].
Dz ¼ wðt þ DtÞDt
[43]
Dx ¼ U Dt
   
Dt 2 Dt
a ¼ 1 and b ¼ [44]
TLw TLw
where the Lagrangian timescale, TLw, can be related to the
turbulent dissipation rate according to eqn [45], where C0 is
a constant.
2 s2w
TLw ¼ [45]
While this simple model has produced useful results, it is
not applicable to boundary layers with gradients in turbulence
properties. New developments in LSS are best described by
recasting eqn [42] as eqn [46], where a is given by eqn [47], and
the dz is a normal random variable with zero mean and vari-
ance dt.
dw ¼ aw dt þ ðC0 εÞ1=2 dz [46]
C0 ε
a ¼ [47]
2s2w
During the 1980s, several ad hoc formulations for the term
‘a’ in eqn [46] were proposed to account for flow complexities
such as inhomogeneity and non-Gaussianity of the turbulent
velocities. This unsatisfactory situation was remedied by
Thomson in 1987 when he proposed a systematic method for
constructing formulations for ‘a’ by invoking the constraint
that the model for this term should preserve a well-mixed
concentration field. This is equivalent to insisting that the
Lagrangian pdf for fluid particles marked at the source should
become identical to that of unmarked fluid particles at large
travel times. The evolution of the pdf is governed by the
Fokker–Planck equation corresponding to eqn [46]. Using the
solution of this equation, Thomson was able to derive
formulations for the term ‘a’ that accounted for inhomoge-
neity and non-Gaussianity of the turbulent velocities. For
example, the modified form of eqn [46] can be represented by
eqn [48].
 
C0 ε 1 w2 vs2w
dw ¼  2 w dt þ 1þ 2 dt þ ðC0 εÞ1=2 dz [48]
2sw 2 sw vz
It turns out that ‘a’ can be expressed uniquely only for one-
dimensional turbulence or isotropic turbulence; for two- and
three-dimensional turbulent flows, ‘a’ is not a unique function
of the velocity field. This nonuniqueness problem can be se-
rious because for a given turbulent flow, substantially different
concentration fields can be obtained from two models, both of
which satisfy the well-mixed condition.
One way of alleviating this problem is to drive Lagrangian
stochastic models with velocity fields obtained from large eddy
simulation because large eddies that contribute to inhomoge-
neity and anisotropy of the velocity field are treated explicitly in
both LES and LSS. The subgrid-scale eddies can be considered

locally homogeneous and locally isotropic. Because only the
particle motion associated with subgrid energy is modeled with
LSS, errors associated with the nonuniqueness of the LSS model
can be minimized.
Concentration Fluctuations and Model Uncertainty
Any given observed concentration will deviate from the
ensemble average. This deviation is caused by the intrinsic
variability of the concentration field called concentration fluc-
tuations. In principle, the statistics of these concentration
fluctuations can provide insight into the expected deviation of
the model-predicted ensemble mean from a corresponding
observed concentration.
To understand the effect of concentration fluctuations,
consider the following model of the concentration time series
in which the concentration is a peak value, Cpeak, or zero. Then,
we can show that the variance of the instantaneous concen-
tration about the ensemble-averaged mean is given by eqn [49].
s2c Cpeak
¼ 1 [49]
C2 C ratio, which determines the concentration variance, is simply
Because the peak-to-mean concentration ratio can be as
large as 10 or even 100, especially close to an elevated source,
the standard deviation of the concentration fluctuations can be
several times the mean. If we are interested in predicting a time-
averaged concentration, we have to estimate the timescale, Tc
that governs the concentration fluctuations. This then allows us
to estimate the number of independent concentration events
that we are likely to observe during the averaging interval Tav, as
Tav/Tc. Then, the deviation, d, of the time-averaged mean about
the ensemble mean is given by eqn [50].
1=2 neous plume, which is inversely proportional to s2, where s
d Tc sc
 [50]
C Tav C averaged spread of the plume, the peak-to-mean ratio, and the
Qm
V0
Qm
C0 ~
V0
Figure 4 Factors that affect concentration fluctuations. Cpeak  C0 in the absence of molecular diffusion, and s2c =C2  Cpeak =C  V=V0 .
Turbulence and Mixing j Turbulent Diffusion 285
Tc roughly corresponds to the time taken for the instanta-
neous plume to pass a fixed observer. If we take the instanta-
neous plume size to be 100 m and the wind speed to be
5 m s1, the timescale, Tc, is 20 s. Assuming that we are inter-
ested in an averaging time of 1 h, d works to be about 25% of
the mean for a peak-to-mean ratio of 10. This exercise identifies
the variables that might govern the deviation between model
estimates and observations. Actual comparisons between
model estimates and observations indicate that the error is
much larger because of model formulation and input errors. A
model is generally considered adequate if its estimates are
consistently within a factor of two of the observations.
We can use a simple model to discuss the role of molecular
diffusion in determining concentration fluctuations. Assume
that we release Qm mass units in an initial volume of V0, so that
the initial concentration is Qm/V0. As the released material is
stretched over a larger volume, V, in space, the volume marked
by the material remains unchanged if the fluid is incompress-
ible and molecular diffusion is negligible (see Figure 4). The
mean concentration, corresponding to material spread over V,
is proportional to 1/V, while the peak concentration remains
unchanged from the initial value. Then, the peak-to-mean
V/V0. In the presence of molecular diffusion, the material is no
longer confined to its initial volume, and the peak concentra-
tion has to decrease with time. Thus, molecular diffusion
decreases concentration fluctuations relative to the mean value.
The meandering plume model, used to estimate the vari-
ance of concentration fluctuations, is an extension of the
previous concept. Here, the time-averaged plume is assumed to
be composed of instantaneous plumes whose dimensions are
determined by relative dispersion. Then, the peak concentra-
tion is determined by the concentration within the instanta-
refers to the spread by relative dispersion. Then, if s is the time-
Volume of marked material
is still V0 but material is
spread over V in space
V
Qm
C~ Peak concentration
V is decreased
by molecular
diffusion through
strands of marked
material
286 Turbulence and Mixing j Turbulent Diffusion
normalized concentration variance is of the order of s2 =s2 as
long as we can neglect molecular diffusion within the instan-
taneous plume. Molecular diffusion will eventually smear the
concentration over the instantaneous plume and thus decrease
the normalized concentration variance.
The statistics of concentration fluctuations can be used in
estimating the air quality impacts of species for which small
time exposures are important. These statistics can be used to
estimate the probability that a certain threshold is exceeded.
Thus, there is great interest in formulating models for
concentration fluctuations. Some of the more recent models
have combined a version of LSS with LES velocity fields to
model relative dispersion and thus concentration fluctuations.
In the future, we are likely to see more use of DNS in under-
standing turbulent dispersion. This does not mean that our
ability to predict concentrations will improve substantially,
because the nature of turbulence places practical limits on
predicting individual realizations of concentrations.
See also: Aerosols: Observations and Measurements; Aerosol
Physics and Chemistry. Boundary Layer (Atmospheric) and Air
Pollution: Modeling and Parameterization; Overview. Aviation
Meteorology: Clear Air Turbulence. Numerical Models:
Parameterization of Physical Processes: Turbulence and
Mixing. Turbulence and Mixing: Overview; Turbulence, Two
Dimensional.
Further Reading
Arya, S.P., 1999. Air Pollution Meteorology and Dispersion. Oxford University Press,
New York.
Csanady, G.T., 1973. Turbulent Diffusion in the Environment. Reidel, Dordrecht,
Holland.
Nieuwstadt, F.T.M., van Dop, H. (Eds.), 1982. Atmospheric Turbulence and Air
Pollution Modeling. Reidel, Dordrecht, Holland.
Pasquill, F., Smith, F.B., 1983. Atmospheric Diffusion, third ed. Ellis Horwood Limited,
John Wiley & Sons, New York.
Rodean, H.C., 1996. Stochastic Lagrangian Models of Turbulent Diffusion. American
Meteorological Society, Boston, MA.
Seinfeld, J.H., Pandis, S., 1999. Atmospheric Chemistry and Physics of the Atmo-
sphere. Wiley-Interscience, New York.
Taylor, G.I., 1921. Diffusion by continuous movements. Proceedings of the London
Mathematical Society 20, 196–211.
Tennekes, H., Lumley, J., 1972. A First Course in Turbulence. MIT Press,
Cambridge, MA.
Venkatram, A., Wyngaard, J. (Eds.), 1988. Lectures on Air Pollution Modeling.
American Meteorological Society, Boston, MA.