COMPUTATIONAL FLUID

DYNAMICS
Instructor: Dr. Luca Biancofiore
Email: [email protected]
Office: At Felix

Lesson V: TURBULENCE
Outline

Introduction
Reynolds Averaged Navier-Stokes (RANS) equations
Turbulent channel flow
Correlation, eddies and mixing length model
Where can we find turbulence? - 1

Nearly all macroscopic flows encountered in the natural

world and in engineering practice are turbulent.
Where can we find turbulence:
1 Winds and currents in the atmosphere and ocean;
2 Flows through residential, commercial, and municipal
water/airdelivery systems;
3 Flows past transportation devices (cars, aircraft, ships,
etc.);
4 Flows through turbines, engines and for power generation
and conversion reactors
Where can we find turbulence? - 2

Turbulence is an enigmatic state of fluid flow that may be

simultaneously beneficial and problematic.
In air-breathing combustion systems, it is exploited for
mixing reactants but, within the same devices, it also leads
to noise, and efficiency losses.
Within the earth’s ocean and atmosphere, turbulence sets
the mass, momentum, and heat transfer rates involved in
pollutant dispersion and climate regulation.
What is turbulence?

Turbulence involves fluctuations that are unpredictable in

detail, and it has not been conquered by deterministic or
statistical analysis.
However, useful predictions about it are still possible and
these may arise from physical intuition, dimensional
arguments, direct numerical simulations, or empirical
models and computational schemes.
In spite of our everyday experience with it, turbulence is
not easy to define precisely and there is a tendency to
confuse turbulence with randomness.
A turbulent fluid velocity field conserves mass, momentum,
and energy while a purely random time-dependent vector
field need not.
Turbulence: definition

Usually a flow is said to be turbulent if it features the

following symptoms (not completely perfect but that’s it)
1 The flow is unsteady (fluctuations)
2 The flow contains vortical motions (i.e. turbulent eddies)
3 The flow contains a wide range of length scales (constantly
evolving & self-sustained)
4 The flow enhance mixing (large diffusivity)

A simple (but imperfect) definition of turbulence as:

A dissipative flow state characterized by nonlinear
fluctuating three-dimensional vorticity
But be careful in atmosphere turbulence can be 2D (or
geostrophic) due to the restriction of the vertical motion by
stratification and Coriolis acceleration.
Turbulence characteristics: fluctuations

1 Fluctuations
Turbulent flows contain fluctuations in the dependent-field
quantities (velocity, pressure, temperature, etc.) even when
the flow’s boundary conditions are steady.
Turbulent fluctuations appear to be irregular, chaotic, and
unpredictable.
Turbulence characteristics: non-linearity

2 Nonlinearity:
The momentum conservation equation contains the
nonlinear advective-acceleration term, and even in ideal
flows this nonlinearity causes pressure to depend on the
square of the velocity.
Turbulence represents an even further assertion of this
nonlinearity, and occurs when the relevant nonlinearity
parameter (i.e. Re) exceeds a critical value.
The enhanced nonlinearity of turbulence is evident in vortex
stretching (i.e a key mechanism that produces
three-dimensional fluctuations).
Turbulence’s nonlinearity causes the time and length scales
of the flow’s initial and boundary conditions to be smeared
into fluctuations having continuous spectra involving a
range of frequencies and wave numbers.
Turbulence characteristics: vorticity

3 Vorticity:
Turbulence is characterized by fluctuating vorticity.
A cross-section view of a turbulent flow typically appears as
a diverse collection of streaks, strain regions, and swirls of
various sizes that deform, coalesce, divide, and spin.
Identifiable structures in a turbulent flow, particularly those
that spin, are called eddies.
Turbulence always involves a range of eddy sizes and the
3
size range increases with Re 4 .
The characteristic size of the largest eddies is the width of
the turbulent region (i.e. 𝛿 in a boundary layer).
Such layer-spanning eddies commonly contain most of the
fluctuation energy in a turbulent flow and may be orders of
magnitude larger than the smallest eddies.
Turbulence characteristics: dissipation

4 Dissipation:
On average, the vortex stretching mechanism transfers
fluctuation energy and vorticity to smaller and smaller
eddies via nonlinear interactions
This occurs until velocity gradients become so large that
fluctuation energy is converted into heat (i.e., dissipated) by
the action of viscosity and the motions of the smallest
eddies.
Persistent turbulence therefore requires a continuous
supply of energy from an imposed velocity or pressure
difference to make up for this energy loss.
Turbulence characteristics: diffusivity

5 Diffusivity:
Within a turbulent flow, the prevalence of fluctuations and
vortical overturning motions increases the mixing and
diffusion of chemical species, momentum, and heat
They are orders of magnitude faster than molecular
transport in equivalent laminar flows!
Implications for engineers?

What do you want first?

1 The good news
2 The bad news
Good news for engineering
The turbulence is embedded in the Navier-Stokes equations (no
further equation is required!)
Continuity equation + momentum equations + boundary/initial
conditions + high Re → TURBULENCE!

My colleagues at Imperial, Touber & Leschziner performed

channel-flow simulation on the UK’s National Supercomputer
(HECToR), Re = 2 × 104 .
Some numbers. Mesh-size 0.6 billion nodes. Runtime 10 years
on today’s 8-core CPUs. Actual cost $150k
Output. 40TB of research data on turbulence. A better
understanding of near-wall turbulence. 3 top-journal papers...
but no engineering design!
Bad news for engineering

Solving the system described in the previous slide

precisely is time-consuming and costly
The actual time-resolved solution is extremely sensitive to
both the boundary and the initial condition.
The engineer’s approach.
Do I care about the precise time history of the flow?
If not, can I make useful predictions about the statistical
properties of the flow instead? (e.g. time-averaged speed &
pressure distribution)
Car-design example

A car designer may ask: what is the average drag force on

the car at a given speed?
The answer will translate into fuel consumption and
eventually how well the car sells.
A back of the envelope calculation gives that Re > 106 . So
the flow is turbulent.
How to predict the average drag on the car if the flow
is turbulent?

Option 1. Run wind-tunnel experiments.

But it is extremely costly. Wind-tunnel tests can cost

$25M/day
Only to test the final-computer-based design
How to predict the average drag on the car if the flow
is turbulent?

Option 2. Solve the Navier-Stokes equations.

Attractive but the required computing resources are not

widely available.
A practical approach is to solve the time-averaged
Navier-Stokes (RANS) equations instead.
Reynolds’ decomposition - 1

Let us consider a signal (say from a probe in a flow) measuring

some quantity (say the velocity component u) at a given point in
space as a function of time t

We then choose a time interval [t0 , t0 + T ] and define the

time-averaged velocity ū (time-mean velocity)

∫︁ t0 +T
1
ū = u(t)dt
T t0
Reynolds’ decomposition - 2

At any given time it is therefore possible to define the difference

between the instantaneuous velocity and the time-mean velocity.
That difference is referred to as the velocity fluctuation
(u ′ = u − ū)

Therefore at any given time inside the chosen time interval, the
instantaneous velocity is nothing but the sum of the time-mean
velocity with the velocity fluctuation:
∀t ∈ [t0 , t0 + T ] u(t) = ū + u ′ (t)
The above decomposition of the velocity into a time-mean
component and a fluctuating component is known as Reynolds’
decomposition.
Some properties of the time-averaging operator
(overbar)

The time-averaging is denoted by an over-bar. We will use

this notation extensively.
In practice we want to apply this operator to sums,
products and differentials. It is therefore convenient to build
a library of rules relating to this operator.
Let us start with three rules.
1 Projection. The average of the average is the average
¯ = ū
ū
2 Linearity. The average of the sum is the sum of the
averages. If u1 and u2 are two time signals and 𝜆 is a real
constant u1 + 𝜆u2 = ū1 + 𝜆ū2
3 Commutativity. The average of the derivative is the
𝜕 ū
derivative of the average: 𝜕u
𝜕x = 𝜕x
Time-averaging of the continuity equation

Starting from the continuity equation for an incompressible

flow
𝜕u 𝜕v 𝜕w
+ + =0
𝜕x 𝜕y 𝜕z
We can arrive to...
𝜕 ū 𝜕 v̄ 𝜕 w̄
+ + =0
𝜕x 𝜕y 𝜕z
Time-averaging of the continuity equation

Starting from the continuity equation for an incompressible

flow
𝜕u 𝜕v 𝜕w
+ + =0
𝜕x 𝜕y 𝜕z
We can arrive to...
𝜕 ū 𝜕 v̄ 𝜕 w̄
+ + =0
𝜕x 𝜕y 𝜕z
Time-averaging of the momentum equation

Starting from the x-component of the momentum equation,

assuming the fluid to be Newtonian with constant density 𝜌0 and
constant viscosity 𝜇0 , you will show that if we time-average
1 𝜕p 𝜇0 𝜕 2 u 𝜕2u 𝜕2u
(︂ )︂
𝜕u 𝜕u 𝜕v 𝜕w
+u +u +u =− + + +
𝜕t 𝜕x 𝜕y 𝜕z 𝜌0 𝜕x 𝜌0 𝜕x 2 𝜕y 2 𝜕z 2
we obtain
𝜕 2 ū 𝜕 2 ū 𝜕 2 ū
(︂ )︂
𝜕 ū 𝜕 v̄ 𝜕 w̄ 1 𝜕 p̄ 𝜇0
ū + ū + ū =− + + + +
𝜕x 𝜕y 𝜕z 𝜌0 𝜕x 𝜌0 𝜕x 2 𝜕y 2 𝜕z 2
𝜕u ′ u ′ 𝜕u ′ v ′ 𝜕u ′ w ′
− − −
𝜕x 𝜕y 𝜕z
The first line looks very familiar (as if we simply added an
overbar to the variables in the original momentum equation
assuming steady flow).
However there are some other extra forcing terms on the second
line...
The RANS equations

Eventually time-averaging the Navier-Stokes equations gives the

Reynolds-Averaged Navier-Stokes (RANS) equations.
𝜕 ū 𝜕 v̄ 𝜕 w̄
+ + =0
𝜕x 𝜕y 𝜕z

𝜕 2 ū 𝜕 2 ū 𝜕 2 ū
(︂ )︂
𝜕 ū 𝜕 v̄ 𝜕 w̄ 1 𝜕 p̄ 𝜕u ′ u ′ 𝜕u ′ v ′ 𝜕u ′ w ′
ū + ū + ū =− + 𝜈0 + + − − −
𝜕x 𝜕y 𝜕z 𝜌0 𝜕x 𝜕x 2 𝜕y 2 𝜕z 2 𝜕x 𝜕y 𝜕z
(︂ 2
𝜕 2 v̄ 𝜕 2 v̄
)︂
𝜕 v̄ 𝜕 v̄ 𝜕 v̄ 1 𝜕 p̄ 𝜕 v̄ 𝜕v ′ u ′ 𝜕v ′ v ′ 𝜕v ′ w ′
ū + v̄ + w̄ =− + 𝜈0 2
+ 2
+ 2
− − −
𝜕x 𝜕y 𝜕z 𝜌0 𝜕y 𝜕x 𝜕y 𝜕z 𝜕x 𝜕y 𝜕z
(︂ 2 2 2
)︂ ′ ′ ′ ′
𝜕 w̄ 𝜕 w̄ 𝜕 w̄ 1 𝜕 p̄ 𝜕 w̄ 𝜕 w̄ 𝜕 w̄ 𝜕w u 𝜕w v 𝜕w ′ w ′
ū + v̄ + v̄ =− + 𝜈0 2
+ 2
+ 2
− − −
𝜕x 𝜕y 𝜕z 𝜌0 𝜕y 𝜕x 𝜕y 𝜕z 𝜕x 𝜕y 𝜕z

1st term: team-mean convective acceleration. 2nd term:

time-mean pressure gradient. 3rd term: time-mean viscous
diffusion, 4th term: some extra forcing terms?
The closure issue - 1

The extra forcing terms add 6 unknowns to the systems since

u ′ v ′ = v ′ u ′ , u ′ w ′ = w ′ u ′ and w ′ v ′ = v ′ w ′
The systems is not closed and so it is not solvable. We have four
equations for 4 + 6 = 10 unknowns.
If we want to continue with the idea of solving the time-mean
flow field, we need to model those 6 terms through a relations
with the time-mean velocity and pressure
Modelling the extra terms is one of the most salient issue in fluid
mechanics
So far, engineering are only provided with a catalogue of
models. Each model is specific to a flow configuration and its
use in a slightly different context (a common situation in design)
brings its share of uncertainties
Finding a universal model to close the system is the holy grail of
turbulence research
The closure issue - 2

At the present time there are three approaches to the

closure problem.
1 The first, known as RANS closure modeling involves using
as closure equations developed from dimensional analysis,
intuition, symmetry requirements, and experimental results.
2 The second, known as direct numerical simulations (DNS)
involves numerically solving the time-dependent equations
of motion and then Reynolds averaging the computational
output to determine mean-flow quantities.
3 The third, known as large-eddy simulation (LES), combines
elements of the other two and involves some modelling and
some numerical simulation of large-scale turbulent
fluctuations.
Reynolds’ stresses

There are six Reynolds’ stresses: u ′ v ′ , u ′ w ′ , v ′ w ′ , u ′ u ′ ,

v ′v ′, w ′w ′.

Three Reynolds’ normal stresses: u ′ u ′ , v ′ v ′ and w ′ w ′

Three Reynolds’ shear stresses: u ′ v ′ , u ′ w ′ and v ′ w ′
Case of a fully-developed turbulent channel flow - 1

We consider a turbulent channel flow

We measure the instantaneous streamwise velocity profile
Case of a fully-developed turbulent channel flow - 2

After a sufficient amount of measurements we can

compute the time-mean profile ū
Case of a fully-developed turbulent channel flow - 3

Compare with the laminar channel flow solution with

identical mass flow rate
Recall that the wall shear stress is 𝜏w = 𝜇 𝜕u
𝜕y |y=0
Friction drag on a commercial airlines

The turbulence in the boundary layers developing on the

aircraft generates 30% of the total drag!
Huge potential for fuel savings if only we could maintain a
laminar flow on the wing.
Fully-devoloped turbulence in x - 1

We say that the turbulence is fully-developed in the

streamwise direction if both the time-mean velocity & the
Reynolds stresses are independent of the streamwise
position
Fully-devoloped turbulence in x - 2

Mathematically, fully-developed turbulent flow in the

streamwise direction x translates into:
𝜕 ū 𝜕 v̄ 𝜕 w̄
= = =0
𝜕x 𝜕x 𝜕x
𝜕u ′ u ′ 𝜕u ′ v ′ 𝜕u ′ w ′
= = =0
𝜕x 𝜕x 𝜕x
𝜕v ′ v ′ 𝜕v ′ w ′ 𝜕w ′ w ′
= = =0
𝜕x 𝜕x 𝜕x
Homogeneous turbulence in z

What about the spanwise direction?

At any given time the velocity field strongly depends on the
spanwise position.
We say that the turbulence is homogeneous in the
spanwise direction if both the time-mean velocity & the
Reynolds stresses are independent of the spanwise
position.
𝜕 ū 𝜕 v̄ 𝜕 w̄
= = =0
𝜕z 𝜕z 𝜕z
𝜕w ′ u ′ 𝜕w ′ v ′ 𝜕w ′ w ′
= = =0
𝜕z 𝜕z 𝜕z
𝜕u ′ u ′ 𝜕u ′ v ′ 𝜕v ′ v ′
= = =0
𝜕x 𝜕z 𝜕z
Governing equations for the fully-developed turbulent
channel flow

Let us write the RANS equations

for a fully-devoloped turbulent
channel flow
We make the following assumptions
1 The fluid is Newtonian with constant density and constant
dynamic viscosity
2 There is no applied body force

This allows us to use the RANS equations previously

obtained.
In addition we assume that
1 The turbulence is fully-developed in x.
2 The turbulence is homogeneous in z
3 There is no mean flow in z (w̄ = 0)
Final form of the governing equations - 1

Therefore, the most reduced form for the governing

equation is
d ū (︁ y )︁
𝜇0 − 𝜌0 u ′ v ′ = 𝜏¯w 1 −
dy h
LHS: laminar shear stress (1st term) + Turbulent shear
stress (2nd term). RHS: Total shear stress.

Quite remarkably the sum of the

laminar and turbulent shear stresses
(i.e. the total shear stress) is a linear
function of y
Notice that away from the walls,
most of the shear stress is due to
the turbulence
Final form of the governing equations - 2

d ū (︁ y )︁
𝜇0 − 𝜌0 u ′ v ′ = 𝜏¯w 1 −
dy h

Ultimately we want to solve for the velocity ū

Unfortunately we do not know how the turbulent shear
stress varies with y.
This make the integration impossible
We will therefore have to model the turbulent shear stress.
Integration

We integrate from the bottom wall to some y value:

∫︁ y ∫︁ y
𝜏¯w (︁ 𝜂 )︁ 𝜌0 ′ ′
ū = 1− + u v d𝜂
0 𝜇0 h 0 𝜇0

The first term represent the laminar part and the second one,
the turbulent part

The Reynolds’ shear stress

significantly modifies the
laminar velocity profile
If we knew its dependence
on y , we would be able to
compute the second
integral and predict the
time-mean velocity. Can
we model it?
Notion of correlation

Mathematically the Reynolds’ shear stress (u ′ v ′ ) is a

measure of the correlation between streamwise and
wall-normal velocity fluctuations.
A non-zero correlation between two variables (say A and
B) indicates that A and B depend on one another.
For example, the price of a product is correlated to its
demand
Correlated/uncorrelated events - 1

Suppose we ask 50 people the following questions: on a

scale from 1 to 5
1 Do you consider yourself wealthy? (A) [1=very poor, 5=very
rich]
2 Do you consider yourself happy? (B) [1=very unhappy,
5=very happy]
We perform the survey in two different countries and put
the two questions together
Correlated/uncorrelated events - 2

If all data fit on the line B = A then A and B are perfectly

correlated.
If all data fit on the line B = −A then A and B are perfectly
anti-correlated.

If given N values for A and B the

correlation coefficient is defined by:
∑︀N
− Ā)(Bi − B̄)
i=1 (Ai
R = √︂[︁ ]︁ [︁∑︀ ]︁
∑︀ N 2 N 2
i=1 (Ai − Ā) i=1 (Bi − B̄)

1
∑︀N 1
∑︀N
where Ā = N i=1 Ai and B̄ = N i=1 Bi
Perfectly correlated R = 1, perfectly anti-correlated R = −1,
R = 0 un-correlated data
If all Ai and Bi values are random, the correlation coefficient R
must be equal to zero.
Warning: correlation does not imply causation
Correlation between stream-wise and wall-normal
velocity fluctuations - 1

Then for any point in time, we plot v ′

against u ′ to see whether they
correlate
We find that u ′ and v ′ are
anti-correlated
Correlation between stream-wise and wall-normal
velocity fluctuations - 2

Then for any point in time, we plot v ′

against u ′ to see whether they
correlate
We find that u ′ and v ′ are not
correlated
Link with the Reynolds’ shear stress?

If we do the limit of an infinitely large sample of the correlation

coefficient
∑︀N
i=1 (ui − ū)(vi − v̄ ) u′v ′
lim R = lim √︂[︁ = √︁
N→∞ N→∞
]︁ [︁ ]︁
(u ′ u ′ )(v ′ v ′ )
∑︀N 2
∑︀N 2
i=1 (ui − ū) i=1 (vi − v̄ )

The Reynolds’ shear stress is the correlation coefficient without

normalisation by the standard deviations.
About the sign of u ′ v ′ : a conceptual picture

Both cases give a negative correlation. On average one

expects the correlation to be negative in the lower part of
the channel.
About the sign of u ′ v ′ : in practice - 1

White vectors represent the velocity fluctuations. The contours

represent positive u ′ (light grey) and v ′ (dark grey) regions.
About the sign of u ′ v ′ : in practice - 2

White vectors represent the velocity fluctuations. The contours

represent positive u ′ (light grey) and v ′ (dark grey) regions.
Notion of eddies

A very useful way of looking at the velocity fluctuations is to

imagine that they emerge from the combination of tornadoes of
different sizes and strengths...
Recall: one descriptive symptom of turbulence is vertical motion
in all directions, of all sizes
We refer them as turbulent eddies
Conceptual picture of turbulence

The precise production/destruction/motion of eddies

occupies a large number of researchers.
Historical note: Leonardo da Vinci (5 centuries ago)

“The small eddies are almost numberless, and large things

are rotated only by large eddies and not by small ones, and
small things are turned by both small eddies and large.”
Formalising the idea of diving and rising lumps - 1

1 Suppose that a fluid particle is thrown from y0 + ℓ to y0 retaining

its linear momentum. On average it arrives at y0 with velocity
ū(y0 + ℓ).
2 Suppose that a fluid particle is thrown from y0 − ℓ to y0 retaining
its linear momentum. On average it arrives at y0 with velocity
ū(y0 − ℓ).
Assuming that the time-mean velocity is sufficiently well-behaved (i.e.
its n-th derivative exists and is continuous), we proceed by expanding
it in Taylor series about y0 ± ℓ:
1 2 d 2 ū ⃒⃒ 1 3 d 3 ū ⃒⃒
⃒ ⃒ ⃒
d ū ⃒⃒
ū(y0 ± ℓ) = ū(y0 ) ± ℓ + ℓ ± ℓ
dy ⃒
y0 2 dy 2 ⃒
y0 6 dy 3 ⃒
y0
Formalising the idea of diving and rising lumps - 2

If ℓ << h, we can make the following approximation

⃒
d ū ⃒⃒
ū(y0 ± ℓ) ≈ ū(y0 ) ± ℓ
dy ⃒y0

Therefore, the relative velocity difference of the lump of fluid

coming from top/bottom is

u ′ = ū(y0 ± ℓ) − ū(y0 )

This implies that lumps of fluid coming from the top (case 1) and
bottom (case 2) each induce a velocity fluctuation which scales
with the time-mean velocity gradient (at any value of y ):
d ū
u ′ ≈ ±ℓ
dy

Note that + is for case 1 and − for case 2.

Formalising the idea of diving and rising lumps - 3

The diving and rising motions of lumps is related to the action of

turbulent eddies. If we assume the eddy to be axi-symmetric in z, we
expect u ′ and v ′ to scale similarly:
Typical scales: du ′ ∼ u ′ , dx ∼ 𝛿, dv ′ ∼ v ′ ,
dy ∼ 𝛿.
Order of magntifude analysis: from the 2D
continuity eq. we can arrive to
(︂ )︂ (︂ )︂
′ ′ 2 d ū d ū
|u v | ∼ ℓ
dy dy
Formalising the idea of diving and rising lumps - 4

Recall the discussion on the sign of the Reynolds shear stress

d ū
1 If dy is positive, the Reynolds’ shear stress is negative.
d ū
2 If is negative, the Reynolds’ shear stress is positive.
dy
⃒ ⃒
This means that −u v ∼ ℓ ⃒ ddyū ⃒ ddyū
′ ′ 2 ⃒ ⃒
Mixing length

If we want to change to the equal sign, we need to include the

multiplication by a constant
⃒ ⃒
2 ⃒ d ū ⃒ d ū
⃒ ⃒
′ ′
−u v = Cℓ ⃒ ⃒
dy dy

We do√not know what C is, and simply redefine ℓ to account for it

ℓm = Cℓ which is often called mixing length
Finally, we find that the Reynolds’ shear stress can be modelled
as ⃒ ⃒
⃒ d ū ⃒ d ū
−u ′ v ′ = ℓ2m ⃒⃒ ⃒⃒
dy dy
If we know how to model the mixing length we can solve for the
time-mean velocity:
d ū (︁ y )︁
𝜇0 − 𝜌0 u ′ v ′ = 𝜏¯w 1 −
dy h
Interpretation of the mixing length - 1

Kinetic theory
Molecules undergoing random motion Turbulent mixing
(due to thermal vibration) Lumps of fluid thrown around and
jostled (due to turbulence)

On average, each molecule travels a

distance ℓ before colliding with another Following the kinetic-theory success to
molecule: this is knows as the mean predict the molecular viscosity, Prandtl
free-path length proposed a direct analogy for turbulent
mixing
Each molecule will have a
root-mean-squared velocity V Viewed as lumps of fluid exchanging
(following many collisions) momentum over a distance ℓm at the
root-mean-squared velocity VT ,
Kinetic theory links such fundamental
effectively creating a turbulent
motions to the kinematic viscosity of
viscosity 𝜈t : 𝜈t = ℓm VT
the fluid: 𝜈0 = 13 ℓV
Interpretation of the mixing length - 2

Molecular viscosity Turbulent viscosity

The molecular viscosity induces a The turbulent viscosity induces a
friction force per unit area in the friction force per unit area in the
presence of a velocity gradient: presence of a velocity gradient
d ū turb d ū
𝜏¯xy = 𝜌0 𝜈0 𝜏¯xy = 𝜌0 𝜈 t
dy dy
But the friction due to the turbulence is nothing but the Reynolds’ stress:
turb
𝜏¯xy = −𝜌0 u ′ v ′
Recalling our ⃒ previous
⃒ estimate for the u ′ − v ′ correlation
2 d ū d ū
−u ′ v ′ = ℓm ⃒ dy ⃒ dy :
⃒ ⃒
⃒ ⃒
⃒ d ū ⃒
𝜈t = ℓ2m ⃒⃒ ⃒
dy ⃒
The turbulent viscosity 𝜇t = 𝜌𝜈t is associated with the turbulent eddies
0
and we often call it the eddy viscosity (𝜈t is the kinematic turbulent
viscosity).
The mixing length represents the characteristic size of the eddies
A key interpretation of turbulence - 1

So, the fully developed turbulent channel flow equation

may be written:
(︂ )︂
𝜈t d ū (︁ y )︁
𝜇0 1 + = 𝜏¯w 1 −
𝜈0 dy h
⃒ ⃒
where 𝜈t = ℓ2m ⃒ ddyū ⃒
⃒ ⃒

The mixing length represents the mixing induced by

turbulent eddies, that is their characteristic sizes.
From this equation, turbulence is found to bump up the
molecular viscosity.
Turbulence is therefore seen as a diffusive process (in the
time-averaged sense)
However such a diffusion acts on larger scales (it is not
acting at molecular level)
A key interpretation of turbulence - 2

The idea of a turbulence viscosity is not restricted to the

Reynolds’ shear stress and was extended to all 6
components (Boussinesq hypothesis)
(︂ )︂ 3
𝜕 ūi 𝜕 ūj 𝜌0 ∑︁ ′ ′
−𝜌0 ui′ uj′ = 𝜌0 𝜈t + + 𝛿ij uk uk
𝜕xj 𝜕xi 3
k=1

In a similar way to what we have done for the stress tensor

(︂ )︂ 3
𝜕ui 𝜕uj ∑︁ 𝜕uk
𝜏ij = 𝜇 + + 𝜆𝛿ij (i, j) ∈ {1, 2, 3}
𝜕xj 𝜕xi 𝜕xk
k =1

This is an approach used to compute complex time-mean

3D flows (e.g. on an aircraft, car, etc.), provided that there
exists a good model for the eddy viscosity
A simplistic model for the mixing length

In the vicinity of the wall, the eddy size is restricted by the

distance between the wall and the centre of the eddy.
Further away from the wall, the eddy size may reach a
maximum size which scales with the channel half-height
Mixing-length and eddy viscosity from actual data

The variation of the mixing length with the wall distance is

seen to have distinctive behaviours.
The linear variation, although not well supported by the
data, turns out to give good results
A piecewise modelling approach

If we model the turbulent viscosity 𝜈t in

d ū (︁ y )︁ d ū
𝜇0 − 𝜌0 u ′ v ′ = 𝜏¯w 1 − with − u ′ v ′ = 𝜈t
dy h dy
with a piece-wise model (left figure) we can integrate it to
obtain the velocity (right figure).
Normalised Reynolds’ shear stress and time-mean
velocity
The velocity is made of 3 distinct layers: the viscous sub-layer,
the log-law and the wake region.
It is possible to demonstrate the behaviour of the velocity profile
in these layers
The region between the viscous sublayer and the log-law region
is called the buffer layer (where the production of turbulence is
maximal)
The end

Lesson VI is over. It’s time to check what is happening it

the fluid is compressible... but please remember to read
the historical notes about turbulence (very interesting!)