123

23. TURBULENCE, PART II 2. THE MEAN MOMENTUM EQUATION

In the last chapter we talked about the basic nature of This analysis can be repeated for the other basic equa-
turbulence: what it’s like and how it’s described statis- tions; I’m not going to write down all the steps, how-
tically. In this chapter I carry on with more “classical ever. The momentum equation starts as
turbulence”, namely the effect of turbulent stresses on
the mean-field flow; and say a little about “modern tur- ∂
(V + v)+(V + v) · ∇(V + v) =
bulence”, including how 2D turbulence is different, and ∂t
1 (23.5)
turbulence on small scales (which includes vorticity & − ∇(P + p) + ν∇2 (V + v)
intermittency). ρ

A. Mean Field Equations The mean of this equation becomes – written out in
Cartesian
In many systems, we can treat the large-scale flow as ∂Vi ∂Vi 1 ∂P
steady, or at least slowly varying, with the turbulence as +Vj =−
∂t ∂xj ρ ∂xi
a rapidly fluctuating additive term. That is: let V be the ! (23.6)
mean velocity, and v the turbulent one, so that the net ∂ 2 Vi ∂ 2 Vi ∂
velocity field is V+v. Treat the pressure field similarly, +ν + − hvi vj i
∂x2i ∂x2j ∂xj
P + p; ditto for temperature. (As long as we stay in
the incompressible limit – which is where most turbu- Thus: the turbulent field contributes a net reaction
lence analysis stays – there are no density fluctuations, back on the mean field. This last term is called the
right?) If these sums are put into the basic dynam- Reynolds stress. Even though hvi i = 0 for each com-
ical equations, we can (borrowing terminology from ponent i, the mean of their product does not necessar-
MHD turbulence), isolate the dynamical equations for ily vanish: hvi vj i =
6 0. It turns out that this is so for
the “mean field” quantities, and find how the turbulence anisotropic turbulence – and real turbulence is com-
affects the mean flow. We will find, for instance, that monly anisotropic.1
the mean-flow momentum equation contains what are Note that the Reynolds stress arisews from averag-
called Reynolds stresses: non-zero terms involving sec- ing the nonlinear convective term in the full Navier-
ond moments of the fluctuating velocity field. Thus, Stokes equation; it is not truly a stress term, more like
the mean flow and turbulence are intimately connected, the mean momentum flux due to the turbulent fluctua-
with the one affecting the other. tions. Nonetheless, it is often talked about in terms of a
1. THE CONTINUITY EQUATION
turbulent viscosity. Because this extra force term in the
basic momentum equation is the gradient of a tensor, it
For instance, take the incompressible continuity equa- is generally combined with the viscous stress tensor, as
tion:  
∂Vi ∂Vj
∇ · (V + v) = ∇ · V + ∇ · v = 0 (23.1) σij = −P δij + ρν + − ρhvi vj i (23.7)
∂xj ∂xi
Now, take means: that means ensemble averages, or We can go further, and write the turbulent contribution
time averages for the case of stationary flows. This to the viscous stress tensor as
gives  
dVi dVj
∇ · hVi + ∇ · hvi = 0 (23.2) σij,turb = −ρhvi vj i = ρνt + (23.8)
∂xj ∂xi
But now: first, we have hVi = V; that is the “mean This last step is an assumption which turns out to be
flow” velocity. And, we assume the turbulent fluctua- quite well justified. That is, the Reynolds stress is lin-
tions have zero mean: hvi = 0. Thus, (23.2) becomes early proportional to the mean flow shears. But then,
∇·V =0 (23.3)
1
(recovering incompressibility of our mean state). And, This is the case because, microscopically, viscosity mixes the rel-
subtracting (23.2) from (23.1) gives ative phases of each vi component; they do not stay in phase, and
thus the mean of their product is non-zero. Alternatively: think
∇·v =0 (23.4) about how Brownian motion, on the microscopic level, trans-
ports momentum components laterally.. this gives us the usual
showing that our turbulent field is also incompressible macro stress tensor. Here, the turbulent fluctuations provide a
(by itself). mean transport of momentum...
 124

we can see from this that the turbulent viscosity coef- the constancy of τ we get τ = ydPw /dx + τw and
ficient νt ∼ vt L, if L is the local gradient scale. This τw is another constant, the mean viscous shear stress
connects back to our dimensional argument, above. at the wall y = 0. We find its value by noting that
the flow must be symmetric about the midplane; thus
3. EXAMPLE : 2D CHANNEL FLOW τw = −DdPw /dx. Using this in the definition of τ , we
I follow Mathieu & Scott here. Consider a channel get
in the x-direction, with y the transverse coordinate.  y dVx
(Once again, compare this to the laminar flows in chap- τw 1 − + ρhvx vy i = ν (23.11)
D dy
ter 2). Assume the mean flow independent of z, and
that Vz = 0. Note, this does not imply that the flow This is almost the answer for the mean flow .... as we’re
is fully 2D, indeed the turbulence will be 3D. Rather, turbulent, we know Re ≫ 0, so that the RHS is small
we’re assusming that it’s 2D in the mean, and all sta- except very close to the walls. Thus, away from the
tistical properties are unchanged under reflection about boundary layer, −ρhvx vy i ≃ τw (1 − y/D), that is lin-
the z = 0 plane. As before, put the overall mean flow in ear behavior. And, if this balance holds, we expect Vx
x direction, and let it be driven by some dP/dx. Also to be independent of y; and “measurements find Vx is
as before, all quantites except the pressure depend only approximately constant in this region”. Near the wall,
on y; we extend this here to include Re stresses. we expect steep gradients in Vx , so that viscosity mat-
ters...and a more complicated behavior of the Reynolds
111111111111111111111111111111111111111111
000000000000000000000000000000000000000000
stress.
y

dp/dx 4. THE MEAN ENERGY EQUATION

x

Vx Finally, the mean-field energy equation is worth noting.

IF we multiply the mean-field momentum equation by
111111111111111111111111111111111111111111
000000000000000000000000000000000000000000 Vi and sum on i, we get
Figure 23.1. Geometry and solution of turbulent channel
flow. The flow is symmetric about the midplane, y = 0; the 1 dVi2 1 ∂Vi2 1 ∂ 1 ∂Vi
walls are at y = ±D. Following Mathieu & Scott Figure + Vj = (Vi τij ) − τij
2 ∂t 2 ∂t ρ ∂xj ρ ∂xj
4.3.
(23.12)
Expanding the τij term back out, defining
Now: the assumed symmetry in the z direction
means only the offdiagonal term hvx vy i in the Re stress 
1 ∂Vi ∂Vj

will be nonzero; and it depends only on y. Thus, the Eij = + (23.13)
mean force equation has two nonzero components: 2 ∂xj ∂xi

dhvy2 i and noting that δij ∂Vi /∂xj = ∂Vi /∂xi = 0, we have a
∂P
+ρ =0 useful form:
∂y dy
(23.9)    
d2 Vx ∂P dhvx vy i D 1 2 ∂ P Vj
ν 2 = +ρ V = − + 2νVi Eij − hvi vj iVi
dy ∂x dy Dt 2 i ∂xj ρ
∂Vi (23.14)
From the first of these, we get P (x, y) = Pw (x) − − 2νEij Eij + hvi vj i
∂xj
ρhvy2 i: where Pw (x) is the pressure at the wall (where
vy = 0 by no-slip); and the second term is the diagonal Here, the first 3 terms (inside the brackets) are advective
contribution of the Re stress to the net pressure. Putting energy transport; the term with ν is viscous dissipation;
this into the second equation in (23.9), we get and the last term is energy lost (in a decelerating flow)
  to drive the turbulence.
dPw dτ d ∂Vx
= = ν − ρhvx vy i (23.10)
dx dy dy ∂y 5. WHAT ABOUT THE TURBULENT TERMS ?

The total mean stress is defined as τ = −ρhvx vy i + In these notes I’ve said nothing about the dynamics of
νdVx /dy. But now: the left hand side of (23.10) is the turbulent terms (p, v, etc). That was on purpose
only a function of x, while the right hand side is only .... refer back to the scale separation we used in de-
a function of y; thus, each side is a constant. The left riving the mean-field equations (23.3), (23.6), (23.14).
hand side is just the driving pressure gradient. From We could equally well have separated out the D/Dt
 125

terms involving the turbulent fields. Remember, now, scales: one author described it as the buildup of coher-
that flow equations are nonlinear (the v · ∇v term, for ent vortices, in which nonlinear distortions nearly van-
instance). When one isolates the dynamical equations ish, and which continue to grow by vortex coalesence.
for the fluctuations, one finds higher-order terms are in- One can use arguments similar to those in §22.3 to
volved. For instance, the equation for Dhvi i/Dt in- predict the turbulent spectrum. Let ko be the driving
volves ∂hvi vk i/∂xk ; the equation for Dhvi vj i/Dt in- wavenumber, and assume it is well between the sys-
volves ∂hvi vj vk i/∂xk ; and so on. Some additional as- tem scale ks , and the dissipation scale kd . Energy cas-
sumption is always needed to close a system like this; cades to lower wavenumbers; the arguments of §22.3
and pursuing that would take us too far afield. still apply, so that a power spectrum W (k) ∝ k−5/3
should obtain for kmin < k < ko , where kmin cor-
B. Two-dimensional Turbulence
responds to some largest scale reached by the reverse
. . . differs significantly from turbulence in three- cascade. Above ko , the enstrophy cascade rules. De-
dimensions. The most striking difference is that tur- scribing enstrophy and turnover time at scale l as ωl ∼
bulent power in 2D cascades both forward (to higher vl2 /l2 , τl ∼ l/vl , we get the enstrophy power to be
wavenumber) and backwards (to lower wavenumber). ǫl ∼ Ωl /τl ∼ vl3 /l3 . But this last must be indepen-
This is still a topic of active discussion in the literature. dent of l (repeating the Kolmogorov argument); from
I mostly follow Biskamp’s discussion in these notes. this we need vl ∝ l, so that kW (k) ∝ vk2 ∝ k−2 . This
should describe the cascade up to the dissipation range.
Recall 3D turbulence: we saw that turbulent power
Thus, for 2D fluid turbulence driven at ko , we expect
undergoes only a forward cascade. We also recall that
three-dimensional fluid flow has two significant invari-
ants:
1
Z
W (k) ∝ k−5/3 ; kmin < k < ko
E= v 2 dV, energy (23.15) (23.19)
2
W (k) ∝ k −3
; ko < k < kd
and
1
Z
HV = v · ∇ × vdV, velocity helicity (23.16)
2 This is supported by observations. As in 3D turbulence,
(Invariance of the first should be obvious; invariance of the high-k part of the cascade can be steady-state. En-
the second may be proved in the homework.2 ) Both en- ergy input to the system at ko will cascade forward,
ergy and velocity helicity are conserved in the mode- reaching the dissipation scale kdiss where it goes into
mode interactions which set up the cascade. Such hear. On the contrary, however, the low-k part of the
small-scale interactions drive both E and HV forward, cascade cannot be stable. There is no low-k dissipation
to smaller wavenumbers. The second invariant in 2D to balance the energy input. One would expect kmin to
turbulence, however, is different. The two invariants in decrease with time, until the system size ks is reached;
2D are at and after this point energy will continue to accumu-
late at the largest scales allowed in the system.
1
Z
E= v 2 dV, energy (23.17) What sets the cascade direction? I have not found
2 any clear answer in the literature. On small scales,
and the nature of energy, enstrophy and helicity transfer in
mode-mode interactions (say, k1 +k2 → k3 ) must have
1
Z
Ω= ω 2 dV, enstrophy (23.18) a preference for forward or reverse transfer. The details
2 of these processes seem not to be obvious, however.
Biskamp reports on work in the literature addressing
(The enstrophy is a fancy name that turbulence types
the (thermodynamic) equilibrium distribution of E(k),
like to use for the mean square vorticity.) In 2D tur-
HV (k), and Ω(k). It appears that the equilibrium dis-
bulence, wave-wave interactions drive Ω forward, but
tributions of both E(k) and HV (k) are both weighted
drives E to smaller wavenumbers – in an inverse cas-
towards high k’s in 3D; while in 2D the distribution of
cade. This cascade drives power to larger and larger
Ω(k) is weighted to high k, but that of E(k) is weighted
to low k. The inference, then, is that the cascade di-
2
The methods are similar to the proof that magnetic helicity is rection is set by the tendency of the system to move
invariant, back in chapter 15. towards a statistical equibrium state.
 126

C. Small scales and intermittency time/space scales. This turns out not to be so: on
small enough scales (high enough frequencies, think of
Much of the action these days seems to be “what’s hap- a high-pass filter), you find intermittent “bursts” of en-
pening on small scales”, which means close to the Kol- ergy. Connect this to statistics: a self-similar random
mogorov dissipation scale. This is potentially a big signal, which is the same on different scales, is a white
topic; I’m storing only a brief overview here. noise signal, and has Gaussian statistics. When you go
intermittent, you have a greater chance of getting large
1. INTERMITTENCY bursts, and less chance of getting low-amplitude signal
.. so the probability distribution function (PDF) flat-
The term intermittency is used in two different ways — tens. Several authors show PDF plots with tails much
which drove me crazy when I was trying to learn this flatter than Gaussian. This becomes conspicuous only
field. on scales comparable to, or smaller than, the dissipa-
Older: macroscopic intermittency. Turbulence tion scale. Thus, it is characteristic of the dissipation
observed on large scales (comparable to the system range, and does not imply breakdown of the entire K41
size, well above the dissipation range) is intermittent: analysis.
only a fraction of the volume is filled with turbulent
spots or eddies at any instant. Sit at one point in a tur-
3. THE ROLE OF VORTEX FILAMENTS
bulent region .. you will find periods of high frequency
fluctuations, and also quiet periods, as the turbulent
patches evolve into/out of your region. This seems to One more commenet on the topic of statistics: the sta-
be how the term is used in the older literature. tistical approach we introduced in §22.C was based on
Newer: microscopic intermittency. Turbulence the power spectrum of two-point correlations. It threw
shows quite interesting behavior when observed on away phase information about the Fourier components
scales comparable to the dissipation scale. On these of the flow, as well as any information about higher-
small scales, you find intermittent random “bursts” of order correlations. In particular, that means that the
energy (think of measuring the velocity as a function of power-spectrum approach can’t say anything about lo-
time); this which translates into non-Gaussian statistics, cally anisotropic structures, such as small-scale vortex
which many authors use as a definition of intermittency. ropes. But we are learning, more and more, that vor-
Physically, careful observations and numerical simula- ticity is important on small scales. So, we can’t just
tion reveal the existence of a tangle of intense, slender rely on analytic methods, let alone scaling arguments.
vortex ropes on these small scales. This is the way the We need to turn to numerical simulations – which have
term is used in the current literature. made great strides in recent years, as computers get
faster and sophisticated numerical codes are developed
2. TURBULENCE ON SMALL SCALES to take advantage of them. We can also turn to careful
experiments, designed to probe small scales.
The current discussion on small-scale structure is quite The picture emerging from simulations and exper-
interesting. Problems with the Kolmogorov picture iment is as follows. Essentially all simulations show
(called K41) on small scales seem to have been real- persistent and extended tubes, sheets and blobs of
ized early on. It was noted that velocity derivatives did small-scale vorticity. The filaments are tubes with ap-
vary with the Reynolds number (which is inconsistent proximately circular cross section, and diameter on the
with K41 assumptions). Batchelor and Townsend in- order of the dissipation scale. Their length is some-
terpreted this as “a tendency to form isolated regions where between the Taylor scale and the driving (outer)
of concentrated vorticity”. Moffatt reports a comment scale. The internal structure and dynamics of these little
from Landau to Kolmogorov, that in local regions of vortices is still unclear, it seems – attempts to use ana-
higher ε, the cascade will proceed more vigorously. lytic models (such as the Hill’s vortex we saw in home-
Thus an intermittent distribution of ε(x, t) is expected; work, or a linear vortex model called Burgers) have not
and this should affect the 5/3 exponent slightly, but been all that successful. There is definitely a sense in
should have stronger effect on higher-order statistical the literature that these dissipation-scale vortes tubes
quantities (such as the velocity derivative). play an important role in the overall structure and dy-
This can also be interpreted statistically. A cen- namics of turbulence ... but, again, just how that works
tral assumption of the Kolmogorov theory is the self- is far from clear. So this field is still evolving. Mean-
similarity of the random (velocity) field at all scales. while, I like the quote from Moffatt, Kida and Ohkitani
Therefore, you should find the same statistics on all (1994):
 127

Just as sinews serve to connect a muscle with a

bone or other structure, so the concentrated vor-
tices of turbulence serve to connect large eddies
of much weaker vorticity; and just as sinews can
take the stress and strain of muscular effort, so the
concentrated vortices can accomodate the stress
associated with the low pressure in their cores and
the stress imposed by relative motion of the eddies
into which they must merge at their ends.

References
The mean-field material is “traditional”, and can be
found in various books which treat turbulence math-
ematically. Tennekes & Lumley, or Hinze, are good
sources.
The newer turbulence material, especially on inter-
mittency, I’ve taken from “here and there”. The An-
nual Review of Fluid Mech has several useful reviews
(including the Moffat homage-to-Batchelor article from
2003).