L2.

3 Introduction to turbulent flow

Recommended reading: Fay ch. 7

At the completion of this lecture you should be able to…

• Describe the ‘energy-cascade’ phenomenon of turbulent flow.

• List three common CFD approaches (DNS, LES, RANS) and how they differ.
• Explain why it is useful to average the Navier-Stokes equations and what the
consequences of doing this are.

2.3.1. Characteristics of turbulent flow

In section 1 of this module we introduced the Navier-Stokes equations. These equations were
derived by considering a small material element of fluid and applying Newton’s second law.

Following the derivation of these equations, we investigated a number of simple flows which
satisfied the Navier-Stokes equations (e.g. steady Poiseuille and Couette flow). These examples
showed how viscosity in a fluid acts to slow a fluid down in the vicinity of a solid wall, so that
the flow satisfies the no-slip condition. Poiseuille and Couette flow are examples of laminar
flow. In a laminar flow, the fluid particles travel along smooth paths in a predictable manner.
However, there is another type of viscous flow which also obeys the Navier-Stokes equations
but in which the particles move in a chaotic manner. These flows are known as being turbulent.

There is no exact solution of the Navier-Stokes equations for turbulent flows. However, we
will show later in this lecture that by making certain assumptions turbulent flow can be
modelled reasonably accurately.

The transition to turbulence is dependent on the viscosity of the fluid. In laminar flows, the
viscosity of the fluid is large enough so that instabilities in the flow (which could, for example,
be introduced by a rough surface) are damped by viscosity. Let us consider the flow of a viscous
fluid in a circular pipe. In such a flow we can define a Reynolds number, Re, based on the pipe
diameter, 𝐷𝐷, the mean velocity in the pipe, 𝑈𝑈, and the kinematic viscosity, 𝜈𝜈.

𝑈𝑈𝑈𝑈
Re =
𝜈𝜈
The Reynolds number is a measure of the relative strength of inertia ‘forces’ (mass ×
acceleration) to viscous forces within the flow. At low Reynolds numbers viscous forces damp
out instabilities which occur and the flow remains laminar. If the Reynolds number is increased,
then at Re ~ 2300, the flow begins to become unsteady. As Reynolds number is increased
further, bursts of turbulent flow occur until at some point the flow becomes fully turbulent.
This transition to turbulent flow over a range of Reynolds numbers is typical of turbulent flow
in pipes and boundary layers. Note that the transition Reynolds number for a flat plate aligned
with the flow direction is

𝑈𝑈∞ 𝛿𝛿
Retr = � � ~ 3900
𝜈𝜈 𝑡𝑡𝑡𝑡
where 𝑈𝑈∞ is the velocity well away from the flat-plate and 𝛿𝛿 is the boundary layer thickness.

The images below show flow in a transparent pipe at different Reynolds numbers. Dye is
injected into the flow at a point. At low flow speeds (Reynolds numbers) the flow is laminar
and the streakline formed by the dye is straight and steady. As flow velocity, and Reynolds
number increases, the streakline begins to oscillate, indicating that transition is occurring. As
the Reynolds number increases further the flow eventually becomes fully turbulent.

Figure 2.3.A. Dye visualisation of laminar flow in a pipe (top), flow transition (middle),
turbulent flow (bottom). Image courtesy of Terry Johnson, Berkeley.

2.3.2. Averaging

In a turbulent flow, the velocity field fluctuates randomly in time and is highly disordered in
space. However, these velocity fluctuations are generally relatively small in comparison to the
average velocity in the flow itself. 1

𝑢𝑢1′ (𝐱𝐱, 𝑡𝑡)

𝑢𝑢1 (𝐱𝐱, 𝑡𝑡)

𝑢𝑢�1 (𝐱𝐱, 𝑡𝑡)

𝑡𝑡

1
P. A. Davidson, Turbulence: an introduction for scientists and engineers, Oxford University Press, 2004
 Figure 2.3.B. Velocity in a turbulent flow plotted as a function of time (𝐱𝐱 is constant).

Consider the 𝑥𝑥1 component of velocity 𝑢𝑢1 (𝐱𝐱, 𝑡𝑡) in a turbulent flow at point 𝐱𝐱 and time 𝑡𝑡 .
𝑢𝑢1 (𝐱𝐱, 𝑡𝑡) can be considered to be the sum of a time-averaged component, 𝑢𝑢�1 (𝐱𝐱) and a time-
varying component, 𝑢𝑢1′ (𝐱𝐱, 𝑡𝑡), which is usually much smaller than 𝑢𝑢�1 (𝐱𝐱) (note that here the
overbar denotes that the quantity is time-averaged – not dimensionless). Thus we have

𝑢𝑢1 (𝐱𝐱, 𝑡𝑡) = 𝑢𝑢�1 (𝐱𝐱) + 𝑢𝑢1′ (𝐱𝐱, 𝑡𝑡)

where
𝑇𝑇
1
𝑢𝑢�1 (𝐱𝐱) = � 𝑢𝑢1 (𝐱𝐱, 𝑡𝑡)𝑑𝑑𝑑𝑑.
𝑇𝑇
0

and 𝑇𝑇 is a sufficiently long time period to ensure that the average converges (i.e. does not
change if 𝑇𝑇 is increased further). Note that the time averaged value of the fluctuating
component of velocity is zero.
𝑇𝑇
1
���
𝑢𝑢1′ (𝐱𝐱) = � 𝑢𝑢1 (𝐱𝐱, 𝑡𝑡) − 𝑢𝑢�1 (𝐱𝐱)𝑑𝑑𝑑𝑑 = 𝑢𝑢�1 (𝐱𝐱) − 𝑢𝑢�1 (𝐱𝐱) = 0
𝑇𝑇
0

A useful way of quantifying the ‘level’ of turbulence in a flow is by calculating the time-
average of the kinetic energy per unit mass associated with the time-averaged and fluctuating
components. The total kinetic energy per unit mass is given by

1 1
𝑢𝑢𝑖𝑖 (𝐱𝐱, 𝑡𝑡)𝑢𝑢𝑖𝑖 (𝐱𝐱, 𝑡𝑡) = �𝑢𝑢�𝑖𝑖 (𝐱𝐱) + 𝑢𝑢𝑖𝑖′ (𝐱𝐱, 𝑡𝑡)��𝑢𝑢�𝑖𝑖 (𝐱𝐱) + 𝑢𝑢𝑖𝑖′ (𝐱𝐱, 𝑡𝑡)�
2 2
or

1 1
𝑢𝑢𝑖𝑖 𝑢𝑢𝑖𝑖 = (𝑢𝑢�𝑖𝑖 𝑢𝑢�𝑖𝑖 + 2𝑢𝑢�𝑖𝑖 𝑢𝑢𝑖𝑖′ + 𝑢𝑢𝑖𝑖′ 𝑢𝑢𝑖𝑖′ )
2 2
Time-averaging this expression yields the following equation for the time-averaged kinetic
energy per unit mass in a flow

1 1 1 ′ ′
𝑢𝑢𝚤𝚤 𝑢𝑢𝚤𝚤 = 𝑢𝑢�𝑖𝑖 𝑢𝑢�𝑖𝑖 + ������
����� 𝑢𝑢 𝑢𝑢 .
2 2 2 𝚤𝚤 𝚤𝚤
Thus, the time-averaged kinetic energy in the flow is equal to the kinetic energy associated
with the steady flow + the kinetic energy associated with the turbulent component of flow. In
general the turbulent kinetic energy is only a few percent of the steady flow kinetic energy.
However, the random velocity field produces viscous stresses which are much larger than those
which would exist if the flow were laminar.
4.1.3. Turbulence scales and modelling approaches

When dye is injected into a pipe containing turbulent flow of a liquid, the streakline formed by
the dye distorts into a very complex pattern. The turbulent flow contains eddies – or vortices
of different sizes. These are regions of swirling flow which, for a short time, convect through
the fluid before breaking up into smaller eddies. Thus, the velocity field produced by a turbulent
flow can be thought of as a superposition of a large number of eddies of various sizes. The
largest size of an eddy in a flow is limited only by the transverse dimension of the flow field.
For example, in pipe flow the maximum eddy size is limited by the diameter of the pipe, while
in boundary layer flow the maximum eddy size is limited by the boundary layer thickness.

Most eddies are much smaller than the maximum eddy size and thus a fully developed turbulent
flow will consist of a broad range of eddy sizes. The large eddies within the flow are inherently
unstable and after a short period of time ‘break-up’ into smaller eddies. These smaller eddies
in turn break-up into even smaller eddies and the process continues so that smaller and smaller
eddies are produced. This process in known as an energy cascade, because energy from the
larger eddies is passed on - or cascades down to the smaller eddies which form from the larger
eddies, as shown in figure C below.

Figure 2.3.C. An illustration of the concept of energy cascade in a turbulent flow. Large eddies
break up into small eddies which in turn break up into even smaller eddies. The kinetic energy
in the large eddies gets passed to the smaller eddies.

Let’s consider a flow where Reynolds number Re0 = 𝑢𝑢0 𝑥𝑥0 /𝜈𝜈 is large. In this flow, 𝑢𝑢0 is a
typical value of the magnitude of turbulent velocity associated with the large-scale eddies and
𝑥𝑥0 is the typical size of the large-scale eddies. Because the Reynolds number based on the
size/velocity of the turbulent eddies is large, the inertial forces within the eddies is sufficient
to overcome any viscous forces within the fluid. As the eddies break-up and become smaller,
viscous stresses become more and more important. When the eddies become small enough so
that the Reynolds number is of the order 1 then viscous stresses within the fluid dissipate the
energy contained within the smallest vortices.
Thus, turbulence can be thought of in the following way. Large-scale instabilities (eddies) will
form provided that the viscosity is low enough or the mean flow velocity is high enough. These
eddies are naturally unstable and break-up into smaller eddies which in turn break-up into even
smaller eddies. This process continues until the eddies become small enough for viscous forces
to become important and energy is dissipated. The energy dissipation which occurs in the
smallest eddies means that a turbulent flow needs a continuous supply of energy to maintain it.

The distribution of turbulent kinetic energy with eddy size is often plotted as a turbulent energy
spectrum (see figure D below). The plot below shows the distribution of turbulent kinetic
energy versus turbulence wavenumber ‘k’ which is inversely proportional to eddy size.

log(energy) Energy
cascade

Energy
generation
Energy
dissipation

log(k)

Figure 2.3.D. turbulent energy spectrum.

There are three eddy size ranges

• Large scale eddies (low wavenumber) where turbulent kinetic energy is ‘created’
as instabilities form. This process is dependent on the characteristic lengthscale and
are only weakly dependent on the characteristic kinematic viscosity of the fluid.
• Intermediate ‘inertial’ scales (medium wavenumbers) where energy is simply
transferred from large to small eddies (or small to large wavenumbers)
• Small scale eddies where turbulent kinetic energy is destroyed by viscous
dissipation. This process is dependent on the characteristic kinematic viscosity of
the fluid (and may be assumed to be independent of the characteristic lengthscale).

Because these small eddies are essentially independent of the particulars of the flow, they are
more easily modelled mathematically than the large eddies which are very dependent on the
particulars of the flow.
The typical lifetime of a large eddy, 𝜏𝜏0 , turns out to be related to the typical length and velocity
of the eddy by the following relationship
𝑥𝑥0
𝜏𝜏0 ~ .
𝑢𝑢0

The rate of turbulent kinetic energy production by the large eddies must be equal to the rate of
turbulent energy dissipation by the smallest eddies, 𝜖𝜖. The kinetic energy (per unit mass) of the
largest eddies is thus approximately equal to 𝑢𝑢02 (to the nearest order of magnitude). The rate
of energy production and therefore dissipation (per unit mass) can be calculated to the nearest
order of magnitude by the following expression.

𝑢𝑢02 𝑢𝑢03
𝜖𝜖~ ~ .
𝜏𝜏0 𝑥𝑥0

Large-scale turbulent motion is roughly independent of viscosity (because Re0 ≫ 1). The
small-scale turbulent eddies are however very dependent on viscosity (because Re0 ~1). These
small-scale motions are also statistically independent of the relatively slow large-scale
turbulent fluctuations. Using dimensional arguments, a famous Russian physicist named
Andrey Kolmogorov showed that the length, 𝜂𝜂, velocity, 𝑢𝑢𝜂𝜂 and time-scales 𝜏𝜏𝜂𝜂 of the smallest
‘energy dissipating’ eddies are only dependent on the rate of energy dissipation per unit mass
𝜖𝜖 and kinematic viscosity 𝜈𝜈 and are given by the following expressions.
1/4
𝜈𝜈 3 𝜈𝜈 1/2
𝜂𝜂 = � � , 𝜏𝜏𝜂𝜂 = � � , 𝑢𝑢𝜂𝜂 = (𝜖𝜖𝜖𝜖)1/4
𝜖𝜖 𝜖𝜖

We can use these expressions to determine the length, time and velocity ratios between the
small-scale and large-scale eddies.
𝜂𝜂 −3/4 𝜏𝜏𝜂𝜂 −1/2 𝑢𝑢𝜂𝜂 −1/4
= Re0 , = Re0 , = Re0
𝐿𝐿0 𝜏𝜏0 𝑈𝑈0

Thus, in a turbulent flow, the large-scale eddies will be several orders of magnitude larger than
the smallest eddies. This clearly shown in table A below.

Re0 (𝜂𝜂/𝐿𝐿0 )−1 (𝜏𝜏𝜂𝜂 /𝜏𝜏0 )−1 (𝑢𝑢𝜂𝜂 /𝑈𝑈0 )−1

104 1×103 1×102 1×101
6 4 3
10 3×10 1×10 3×101
108 1×106 1×104 1×102
Table 2.3.A. Small-to-large scale ratios for different Reynolds number flows

This has important implications for an engineer who is trying to model a turbulent flow using
a computer simulation. In a Computational Fluid Dynamics (CFD) simulation, the region
containing the flow is discretized, or ‘divided up’ into a (usually) large number of smaller cells
or elements (see figure E below).
Figure 2.3.E. CFD simulation showing mesh and pressure magnitude on a F1 car (courtesy of
the ANSYS FLUENT website).

Within each of these elements, flow variables such as velocity, pressure etc. are specified at an
initial point and time. A number of different methods can then be used to determine the value
of these flow variables in these elements a short time later (by solving the governing equations
using a numerical method such as finite differences). In order to successfully model all the
structures within the flow, the length of the elements and the size of the time step must be
somewhat smaller than the scale of the smallest turbulent eddy which is present. A CFD
simulation which attempts to resolve all turbulence scales within a flow by solving the
appropriate governing equations is referred to as a Direct Numerical Simulation (DNS).

As an example, let’s say we are going to try and predict the following flow system shown in
figure F below which has large-scale turbulent velocity, length and time-scales 𝑈𝑈0 , 𝐿𝐿0 and 𝜏𝜏0 .

Figure 2.3.F. Turbulent flow which we will try and predict using a CFD simulation.

There are ~10 large structures in our domain. The method which we are going to use to solve
the governing equations requires ~8 elements to resolve one eddy in 3D. Thus, in order to
resolve the large-scale turbulent structures in this flow we need ~80 elements. Now, what if we
 −3/4
want to resolve the small-scale structures? We can use the expression 𝜂𝜂/𝐿𝐿0 = 𝑅𝑅𝑒𝑒0 to
determine the number of elements which are required. As we can see in table B below, as
Reynolds number increases, the number of elements required to correctly resolve all scales of
turbulence present in our flow increases dramatically.

Re0 (𝜂𝜂/𝐿𝐿0 )−1 Number of elements

4
10 1×103 8×1010
106 3×104 2×1015
8 6
10 1×10 8×1019
Table 2.3.B. Number of elements required to resolve small-scale turbulence based on the
Reynolds number of the large scale turbulence Re0

Remember, a CFD simulation needs to be correctly resolved in time as well as space. Let’s
assume that our CFD method needs 10 time steps to resolve a single eddy lifetime. We’ll also
assume that 𝐿𝐿0 ~ 0.01m, 𝜈𝜈 = 10−5 m2.s-1 and 𝑈𝑈0 = 10−4 Re0 . Now, 𝜏𝜏0 ~𝐿𝐿0 /𝑈𝑈0 , so 𝜏𝜏0 =
102 Re−1
0 . The step-size required to resolve the small scale turbulence is shown in table C below.
Once again, as the Reynolds number increases, the computational effort required to correctly
simulate all scales of turbulence dramatically increases.

Re0 𝜏𝜏0 𝜏𝜏𝜂𝜂 /𝜏𝜏0 𝜏𝜏𝜂𝜂 Time step size

4 -2 -2 -4
10 1×10 1×10 1×10 1×10-5
106 3×10-4 1×10-3 3×10-7 3×10-8
8 -6 -4 -10
10 1×10 1×10 1×10 1×10-11
Table 2.3.C. Time-step size required to resolve small-scale turbulence based on the Reynolds
number of the large scale turbulence Re0

As you can by now see, a full Direct Numerical Simulation (DNS) of a turbulent flow is very
challenging. In fact, it turns out that for almost all turbulent flows of practical interest a DNS
is impossible. Figure G below, shows a 2D DNS of a jet at very low Reynolds number which
took 105 cpu hours to simulate.

Figure 2.3.G. DNS of a jet at low Reynolds number (Hu, Morfey and Sandham AIAA 2003-
3166).

Fortunately, as engineers we have a way around this. Instead of trying to directly simulate a
turbulent flow using a DNS, we can time-average the governing equations and use these to
predict what happens to time-averaged quantities.
2.3.4. RANS

The time-averaged Navier-Stokes equations are usually referred to as the Reynolds Averaged
Navier-Stokes (RANS) equations 2. In this section we will derive the RANS equations.

Starting with the incompressible, constant viscosity form of the conservation of mass and
Navier-Stokes equations

𝜕𝜕𝑢𝑢𝑗𝑗
=0
𝜕𝜕𝑥𝑥𝑗𝑗

𝜕𝜕𝑢𝑢𝑖𝑖 𝜕𝜕𝑢𝑢𝑖𝑖 1 𝜕𝜕𝜕𝜕 𝜕𝜕 2 𝑢𝑢𝑖𝑖

+ 𝑢𝑢𝑗𝑗 =− + 𝜈𝜈 + 𝑔𝑔𝑖𝑖
𝜕𝜕𝜕𝜕 𝜕𝜕𝑥𝑥𝑗𝑗 𝜌𝜌 𝜕𝜕𝑥𝑥𝑖𝑖 𝜕𝜕𝑥𝑥𝑗𝑗 𝜕𝜕𝑥𝑥𝑗𝑗

We will now take the time-average of both of these equations, where the time average of a
function f(t) is defined by the equation below.
𝑇𝑇
1
𝑓𝑓 ̅ = � 𝑓𝑓(𝑡𝑡)𝑑𝑑𝑑𝑑
𝑇𝑇
0

As before, the time-average will be denoted by an overbar. The time-averaged equations are
thus

�����
𝜕𝜕𝑢𝑢𝚥𝚥
= 0,
𝜕𝜕𝑥𝑥𝚥𝚥

��������
𝜕𝜕𝑢𝑢𝚤𝚤 �����������
𝜕𝜕𝑢𝑢𝚤𝚤 1 �������
𝜕𝜕𝜕𝜕 ������������
𝜕𝜕 2 𝑢𝑢𝚤𝚤
� � + �𝑢𝑢𝚥𝚥 � = − � � + 𝜈𝜈 � � + 𝑔𝑔𝑖𝑖
𝜕𝜕𝜕𝜕 𝜕𝜕𝑥𝑥𝚥𝚥 𝜌𝜌 𝜕𝜕𝑥𝑥𝚤𝚤 𝜕𝜕𝑥𝑥𝚥𝚥 𝜕𝜕𝑥𝑥𝚥𝚥

Note that density, viscosity and gravity are assumed to be constants so they can be excluded
from the averaging process.

Now, let’s split the flow velocity into a time-averaged component and a fluctuating component
i.e. 𝑢𝑢𝑖𝑖 (𝐱𝐱, 𝑡𝑡) = 𝑢𝑢�𝑖𝑖 (𝐱𝐱) + 𝑢𝑢𝑖𝑖′ (𝐱𝐱, 𝑡𝑡).

In the questions below you are asked to show that the time-averaged conservation of mass and
Navier Stokes equations reduce to

𝜕𝜕𝑢𝑢�𝑗𝑗
= 0,
𝜕𝜕𝑥𝑥𝑗𝑗

𝜕𝜕𝑢𝑢�𝑖𝑖 1 𝜕𝜕𝑝𝑝̅ 𝜕𝜕 2 𝑢𝑢�𝚤𝚤 ������

𝜕𝜕𝑢𝑢 ′ ′
𝚤𝚤 𝑢𝑢𝚥𝚥
𝑢𝑢�𝑗𝑗 =− + 𝜈𝜈 − + 𝑔𝑔𝑖𝑖 .
𝜕𝜕𝑥𝑥𝑗𝑗 𝜌𝜌 𝜕𝜕𝑥𝑥𝑖𝑖 𝜕𝜕𝑥𝑥𝑗𝑗 𝜕𝜕𝑥𝑥𝑗𝑗 𝜕𝜕𝑥𝑥𝑗𝑗

2
The famous physicist Osborne Reynolds was the first to propose this time-averaging process and thus it is
named after him.
Note that these equations are dependent on the constants 𝜈𝜈, 𝜌𝜌 and 𝑔𝑔𝑖𝑖 and the variables 𝑢𝑢�𝑖𝑖 , 𝑝𝑝̅
������
and 𝑢𝑢 ′ ′ ������
′ ′
𝚤𝚤 𝑢𝑢𝚥𝚥 . The quantity 𝜌𝜌𝑢𝑢𝚤𝚤 𝑢𝑢𝚥𝚥 is known as the Reynolds’ stress which is a new parameter which
must be modelled. The modelling of this Reynolds’ stress tensor is a subject of much research,
however, these models are beyond the scope of this course.

An engineer can solve the RANS equations to determine the time-averaged flow which he/she
is interested in. The Reynolds’ stress contains information about the time-averaged turbulence.
Most commercial CFD software packages contain a RANS solver to allow these types of
calculation to be performed. The advantages of using a RANS method are:

• The RANS equations are simple and therefore quick to solve.

• The method generally gives reasonable estimates for mean flow quantities (𝑢𝑢�𝑖𝑖 , 𝑝𝑝̅ and
������
𝑢𝑢 ′ ′
𝚤𝚤 𝑢𝑢𝚥𝚥 ).

Disadvantages are

• Loss of information – time histories cannot be generated (e.g. turbulent vortex

shedding is not captured).
• A model for the Reynolds’ stress must be selected from a range of candidates
• Loss of accuracy.

2.3.5. LES

We have introduced two alternative CFD methods. In a DNS, we try and solve the exact
governing equations. This approach is very costly and is only possible for very simple, low
Reynolds number flows. Alternatively, we can use a RANS approach. This has the advantage
of being quick, but the disadvantage of a loss of accuracy and information. Fortunately, there
are alternative approaches which lie mid-way between these extremes. The most common of
these alternative approaches is referred to as Large Eddy Simulation (LES).

Because small scale turbulence is relatively homogenous and isotropic it can be accurately
approximated using a simple model which does not require the CFD method to resolve the
details of this small-scale turbulence. Large-scale turbulence is generally not homogenous or
isotropic and is generally difficult/impossible to model accurately. This has led to the
development of a CFD method called Large Eddy Simulation (LES) in which the large-scale
turbulence is directly simulated and the small-scale turbulence is modelled. The size of the
turbulent eddies which are directly simulated in a LES is determined by the size of the
elements/time-step. Simulations which use smaller element sizes/time steps are thus able to
directly simulate smaller scale turbulence. Turbulence scales which are too small to directly
simulate are ‘filtered’ out of the simulation and their effect on the flow is modelled using an
appropriate turbulence model. LES requires greater computational resources than RANS
methods but is far cheaper than DNS methods.
Questions

2.3.1. Show that the time-averaged conservation of mass and Navier Stokes equations reduce
to

𝜕𝜕𝑢𝑢�𝑗𝑗
= 0,
𝜕𝜕𝑥𝑥𝑗𝑗

𝜕𝜕𝑢𝑢�𝑖𝑖 1 𝜕𝜕𝑝𝑝̅ 𝜕𝜕 2 𝑢𝑢�𝚤𝚤 ������

𝜕𝜕𝑢𝑢 ′ ′
𝚤𝚤 𝑢𝑢𝚥𝚥
𝑢𝑢�𝑗𝑗 =− + 𝜈𝜈 − + 𝑔𝑔𝑖𝑖 .
𝜕𝜕𝑥𝑥𝑗𝑗 𝜌𝜌 𝜕𝜕𝑥𝑥𝑖𝑖 𝜕𝜕𝑥𝑥𝑗𝑗 𝜕𝜕𝑥𝑥𝑗𝑗 𝜕𝜕𝑥𝑥𝑗𝑗