Professional Documents
Culture Documents
July 2016
Introduction*
This chapter charts the most direct navigable path from first principles to the mathematical models used to
provide the description of the velocity field in the mixing tank. The aim is to provide physical insight, and
to show how we arrive at the final model used for the flow field. Throughout, the goal will be to answer the
question:
"How can we provide a detailed heterogeneous description of the fluid velocity field in order to make
predictions about the behaviour of the particles therein?"
The following discussions assume the reader has an understanding of Cartesian tensors and Einstein
summation notation.
Lagrangian*Description*–*“riding*along*with*the*fluid*parcel…”*
Firstly, we define the continuum concept of a Lagrangian ‘fluid parcel’1. This is a very small volume
of fluid that moves according to the local velocity field, carrying its properties (such as momentum
and density) with it. Because it moves with the fluid, the relative velocity between the parcel and the
surrounding fluid is zero. Consequently, there is no entry or exit of fluid particles – rather, the fluid
parcel stretches and deforms according to the flow field.
Eulerian*Description*–*“staying*put,*watching*fluid*pass*through…”*
Instead of following the fluid parcel as it moves through the flow field, we can sit at a fixed point in
space and record how the fluid properties at that location change over time. To this end, we introduce
a new concept – that of the Eulerian 'control volume'. This is a very small volume at a fixed location
in space2, through which fluid flows.
1
Often called the ‘material volume’.
2
Its position in space is fixed only for an inertial reference frame. For a moving reference frame, its position relative to the reference frame
will be fixed.
Conveniently, at the moment a fluid parcel passes through a laboratory point, the Lagrangian
description (of that fluid parcel) and the Eulerian description (at that point) are instantaneously the
same. This will prove useful, as we shall now see.
Let us imagine we are riding along with the fluid parcel, and we are interested in some scalar quantity
– say, temperature – of the parcel. The change in temperature experienced by the fluid parcel due to
tiny little changes in time (!!) and space (!!, !!, !") is:
We’re interested in recording the changes in temperature of the fluid parcel as it moves through the
fluid (giving a Lagrangian description), therefore, !! = !! !!!, !! = !! !!! and !! = !! !!!.
Consequently:
!" !" !" !"
+ !! + !! + !
!" !" !" !" !
!" !"
+ !! ⋅
!" !!!
If we reduced these tiny little changes in space down to infinitesimally small changes, we would end
up describing how the temperature changes with time at a fixed point in space – in other words, we’d
end up with an Eulerian description! This concept can be extended to any other quantity besides
temperature, including vectors and tensors. In general:
! ! !
!!!! = + !! ⋅
!" !" !!!
Eq. 1
! !
= +!⋅!
!" !"
Material Derivative
So, the rate of change of a quantity with time of a fluid parcel is given by:
• Partial derivative of Lagrangian field
• Material derivative of Eulerian field
The equations that follow are mostly in non-conservative form (for the blue cube, the moving fluid
parcel) in terms of the material derivative. Some are written in conservative form (for the green cube,
the fixed control volume), in terms of surface and volume integrals. The bridge between the two forms
is the Reynolds Transport Theorem. In Chapter 4, we will see that in order to solve the system of
equations developed in this chapter, we must use the conservative form.
!" !!!
+! =0
!" !!!
Eq. 2
!"
+ !!! ⋅ ! = 0
!"
Continuity Equation
Physically, this states that the change in density of the fluid parcel with time depends on the spatial
divergence of the velocity field it traverses3. In this work, we consider only incompressible fluid in
which case:
!"
=0
!"
! ⋅ ! = 0.
3
For those unfamiliar with Gauss’ Theorem, we can make intuitive sense of the equation: think that if fluid particles are diverging away
from a point, then the density must be decreasing (converging would increase the density). See Appendix A on velocity field.
! Eq. 3
!!!! ! = !!"#$!!"#$%& + !"#$%&'!!"#$!"
!" !
Momentum Equation
Body*Forces*
!!
!"#$!!"#$% = !"#$%&#&%'(#)!!"#$% = − !!!!
!!!
Since the gravitational field (Ψ) is a scalar potential field, its gradient is a conservative force field
!!
(conservative body force). Note that close to the Earth, = !!~!9.81!!/! ! .
!!!
Surface*Forces**
where !! is the outward unit normal vector to the surface ‘j’. Using Gauss’ divergence theorem, we
can re-write this as:
! !
! !!! + !!!" ! !!" + !!!"
!! !(!!" + !!!" )!!! = !!" + !!!" !!! !!" = !!" = !!
!!! !!!
!! !!
! 0 0
!!!!!!!!!!!!!" = 0 ! 0
0 0 !
Figure 4
[Left] The stress due to strain and the stress due to pressure in matrix form.
[Middle] The surface stresses acting on a cubic fluid parcel.
[Right] Illustration that 'i' denotes the direction the force acts, while 'j' denotes the face on which it acts.
• The strain can be decomposed into an isotropic (!!" ) and anisotropic component (!!" )4:
1 !!!
!!" = ! = ! ⋅ ! = 0!
3 !!! !"
!
1 !!! !!!
!!" = + − !!" !
2 !!! !!!
!
Thus, due to incompressibility (! ⋅ ! = 0), we are left only with the anisotropic strain – and
its consequent anisotropic stress, which acts tangentially. This is called the shear stress (!!" ).5
We require a constitutive law6 to relate the strain rate experienced by the fluid to the shear
stress applied. For a Newtonian fluid, we can use Newton’s law of viscosity:
!!! !!!
!!" = !!! + !
!!! !!! Eq. 4
!!!" = !!!!(2!!!" )
Shear Stress
• In an incompressible fluid, pressure waves travel infinitely fast [11] – consequently, the
pressure at every point on the surface of fluid parcel is the same, and acts normal to the
surface. The pressure then, is a normal, isotropic stress.
!!!!" !!!"
!"#$$%"#!!"#$% = (!!) ! !ℎ!"#!!"#$% = !!
!!! !!!
NavierDStokes*Equation*
Therefore, the momentum equation (Eq. 3) for an incompressible Newtonian fluid now becomes:
! !! !! ! !!!
!!!! !! = ! − !!!! − (!!) + (!")!!!!!
!" !!! !!! !!! !!! !
Note that the pressure force becomes a (conservative) body force, which allows is to define the
modified pressure as:
! = !" + !
4
See Appendix A (Velocity Field).
5
For second order tensors such as this, isotropic ≣!spherical, anisotropic ≣!deviatoric
6
A constitutive law is simply a relation between two physical quantities that is specific for a given material. Typically this is the kinematic
response of the material to an external kinetic stimulus. [10]
! !" ! !!!
! !! = ! − + !!!!!
!" !!! !!! !!!
Eq. 5
!
! ! = ! −!! + !!!!!! ! !
!"
Navier-Stokes Equation
The number of particles in a fluid parcel can change due to particles being transported in or out, by the
addition from some ‘source’ or removal by some 'sink':
!" !"
= + !!"#$%!&'$% = !"##$%"&' + !"#$%&"'( + !"#$%& − !"#$
!" !"
Where it is assumed there is no electric field for migration. The diffusion, can be described using
Fick’s Law of Diffusion:
!" Eq. 6
+ !! ⋅ !! = !! ! ! + !"#$%& − !"#$
!"
Figure 5 – The "number of particles" as a passive scalar in a fluid parcel. It is worth noting that the 'local change' is that experienced by a
control volume. The differential forms of the non-conservative and conservative equations are the same.
Figure 6 – Viscous stress acting tangentially on the surface of a cubic fluid parcel of volume δV and surface area δS.
In a fluid flow, this conversion is done by the viscous stress (!!" ). Thus, let us consider the rate of
working of this viscous stress on the fluid parcel. We showed earlier that the viscous stress is:
!!" = !!!!(2!!!" )
where !! is the unit normal vector to the surface ‘j’. Using Gauss’ Theorem, we can re-write this as:
! !
! !
(!! !!!" )!!! !!! = (!! !!!" )!!! !!" = (!! !!!" ) !" = !! (! ! )
!!! !!! ! !"
!! !!
Therefore, the rate of working by viscous forces per unit volume is:
! !!!" !!!
Ẇ= (!! !!" ) = !!! + !!"
!!! !!! !!!
The first term simply represents the rate of working of the viscous stress on the fluid parcel, acting to
change the kinetic energy of the fluid parcel:
!!!" Eq. 7
!!
!!! !
Total Working of Viscous Stress
The second term represents the rate that kinetic energy is converted to heat as a result of viscous
dissipation. This rate of viscous dissipation can be re-written as:
Therefore, we can finally define the rate of viscous dissipation (per unit mass) to be:
Thus we have 4 time-dependent equations, along with the boundary condition (U = 0 at stationary
solid surfaces) – but we have 5 unknowns (!, !, !! , !! , !! ). We almost have a closed system.
Relating*Pressure*to*Velocity*
Given that the velocity field is solenoidal (! ⋅ ! = 0), we can relate it to the pressure field by taking
the divergence of the Navier-Stokes Equation (Eq. 5). Doing so obtains:
!
! ⋅ ! ⋅ !!! = ! −! !
!
The Biot-Savart Law[12] may be used to invert this equation, giving us:
! ! ⋅ (! ⋅ !!) ′
! ! = ! !!!′
4! |! − !! |
Thus, by writing the pressure in terms of the velocity, we have reduced the unknowns to
(!, !! , !! , !! ).
“Out*of*the*woods…”*
We now have a closed system of coupled partial differential equations and associated boundary
conditions! In principle, this system could be solved using numerical methods, and we could give the
description of the velocity field that we require to predict the aggregation and breakage behaviour of
the particles! It appears as though we're out of the woods.
To understand why things may not be so straightforward, let’s consider an important characteristic of
the Navier-Stokes Equation:
!!! !!! !" !!!
! + !! ⋅ =!− + !!!!
!" !!! !!! !!! !!!!
The convective term is clearly non-linear (quadratic). This means that while our system may be
closed, it is non-linear dynamic – the hallmark of mathematical chaos.
As we shall see, this takes us very much out of the woods and into the jungle.
7
The Navier-Stokes Equation (Eq. 5) is often referred to as the ‘Navier-Stokes Equations’, since it can be separated into three velocity
components (in Cartesian co-ordinates, this would be the velocity in the x, y and z directions – !! , !! , !! ).
! ! ! !"#$%!&'!!"#$%& Eq. 9
!" = =
! !"#$%&#!!"#$%&
Reynolds Number Equation
The operating conditions in our mixing tank are such that 2000 < Re < 8000 – indicating that
turbulence is present.
What consequences will this 'turbulence' have for the quest of providing the description of the velocity
field in the mixing tank? To answer this question, we might first want to consider where this
'prediction' of turbulence comes from, and the meaning of the phrase 'inertial to viscous forces'.
2.2.1 – Origins
Consider the fate of the smoke from the burning
cigarette (Figure 7). As the smoke emerges, it
accelerates directly upwards in a controlled manner8. It
then begins to 'wobble', before suddenly busting into
seemingly chaotic and random behaviour. We call this
chaotic behaviour of fluids Turbulence.
Disturbances*
Even in the most rigorously controlled experiments,
there exists small disturbances and perturbations that
lie out with experimental control.
In the lower part of the flume, the smoke has not yet
gathered much velocity – consequently, the Reynolds
number is low:
!!! !"#$%!&'!!"#$%&
!" = ! ! = !!
! !"#$%&#!!"#$%&
Figure 7 – Transition to turbulence in a cigarette plume
In other words, the viscous forces dominate the inertial forces. These comparatively large viscous
forces act to dampen the small, unavoidable disturbances – through viscous dissipation, as seen in the
previous chapter. Consequently, they cannot grow and manifest into appreciable flow instabilities. We
say that the flow is 'laminar'.
However, as the smoke accelerates upwards, the inertial forces begin to dominate the viscous forces.
The small, unavoidable disturbances can no longer be sufficiently damped, and they reveal themselves
in the 'wobbling' of the plume. These un-damped disturbances interact with one another, producing
greater disturbances, which interact further still – this growth and interaction leads to the significant
disturbances which we see in the upper plume, and which we call turbulence.
8
The smoke is warmer and less dense than the surrounding air, and rises due to buoyancy effects.
The mixing tank is subject to continual disturbances of varying amplitudes and frequencies, and the
Reynolds number indicates that the inertial forces do indeed outweigh the viscous forces – hence why
we predicted the presence of turbulence in the tank.
2.2.2 – Characteristics
To explore the characteristics of turbulence,
we will look at a classic experimental example
– flow across a cylinder. (Figure 8)
Sensitivity*to*Disturbances*
The result is velocity profiles like those in
Figure 9. Clearly, the velocity profile is
markedly different for each experimental run.
• Turbulent flows display an acute sensitivity to finite perturbations in initial conditions, boundary
conditions and material properties – meaning a precise deterministic prediction of their evolution
is impossible.9
Mean*&*Turbulent*Velocities*
To observe another important characteristic, let's
introduce the simplest of statistical quantities –
the mean (Figure 10). The blue line represents
the (time-averaged) mean velocity profile of
both experimental runs. It's clear that the mean
is steady with time and reproducible –
predictable, one could say.
Figure 10 – The mean velocity profile is reproducible in nominally
identical runs of the experiment.*
9
This acute sensitivity to initial conditions is popularly known as the 'butterfly effect', the mathematical study of which is known as 'chaos
theory'. It is due to the non-linear term identified at the end of the previous chapter – we will further discuss this non-linearity later.
––––––––––––––––––––––––––––––––––––––––––
Suppose we were able to visualise these mean
and turbulent velocity components – what would
they look like?
Absolute*vs.*Statistical*approach*
So what approach should we take for describing the velocity profile in the mixing tank?
• Absolute:
We could attempt to provide a description of the total velocity field by numerically solving the
Navier-Stokes equation. This would, as we know, only give us one particular realisation of the
velocity field, which would look entirely different from one realisation to the next.
• Statistical:
Otherwise we could search for a solution based on the statistical properties (such as the mean
velocity), which we know are reproducible and predictable.
Consider an analogous dilemma – suppose we were interested in the behaviour of a gaseous system.
Would we be more concerned with the motion of every single gas molecule, or the statistical
properties of the system, such as pressure?
Before deciding, let us discuss three further fundamental characteristics of turbulence, which may
make the decision rather easy:
• Vorticity
• Scales
• The Energy Cascade.
Figure 13 –– (a) Illustrative example of turbulence as a tussling tangle of vortex tubes and blobs; (b) Warping of a vortex tube by the fluctuating
velocity field; (c) Stretching of a vortex tube by the mean velocity strain
Turbulence can visualised as a tussling tangle of vortex tubes, sheets and blobs, continually being
teased and stretched out of shape by the local velocity field (Figure 13). This seething mess of
vorticity exhibits chaotic behaviour on a wide range of time and length scales. To endow some kind of
physical meaning to this rather hand-wavy definition, we have to introduce some dynamics.
Vorticity*Dynamics*
Let us consider a small blob of fluid, which we will define as being instantaneously spherical10 – in
effect, a spherical fluid parcel.
The angular momentum (!) of this blob is related to its moment of inertia (!) and vorticity (!) by:
1
!!!! !=
2
Now the angular momentum of this blob can only change due to surface forces (viscous forces) given
that the only body forces of interest – the 'modified pressure' !"/(!!! ) – is a conservative body force.
We have already stated that the pressure is the same at every point on the surface of the fluid parcel,
and acts normal to the surface. Therefore, given that the blob is spherical, pressure cannot change its
angular momentum – only the viscous forces can:
!" 1 ! !!!
!!!!!!!!!!!! !! = !! ! = (!"#$%&#!!"#$%&)
!" 2 !"
!! !"
! = −!! + 2!×! !"!"#$!!!"#$%&
!" !"
Therefore, we can intensify the vorticity of fluid elements by stretching them (decreasing !), as in
Figure 13 (c) above. It will become clear, much later, that this vortex stretching transfers kinetic
energy from the mean flow and into the turbulence.
10
Note: we are free to define a spherical fluid parcel anywhere in the flow at any instant of time.
The*Energy*Cascade*
The above poem actually represents one of the most useful and notable advances in the study of
turbulence – but what exactly did Richardson mean?
The idea is that the largest eddies (at the so-called integral scale) are created by instabilities in the
mean flow, which arise due to insufficient viscous damping as we discussed earlier. These eddies have
a size comparable to the characteristic length of the flow (like the diameter of the mixing tank).
Figure 15 – The length scale of successively smaller eddies effectively lowers the Reynolds number, until eventually, at the smallest
scales, viscous dissipation becomes significant and mechanical energy is converted to heat.
However, these large eddies are also subject to inertial instabilities and the distorting effects of other
eddies, and so are teased out into many smaller eddies. Crucially, this process is essentially inviscid –
this is because !" = !!/! is large, and so the viscous forces acting on the large eddies is negligible.
So the offspring essentially steal kinetic energy from the original eddy.
These smaller eddies suffer a similar fate; evolving into even smaller eddies. Hence, there is a
continual, inviscid cascade of energy from the large scales to the smaller scales.
Things change, however, when we get down to the very smallest scale (the so-called Kolmogorov
Microscale). Here, !" = !!/! is much smaller (being based on the smaller eddy size) – consequently,
the viscosity is no longer negligible, and finally the energy starts to be dissipated. Viscosity, therefore,
acts to mop up whatever energy happens to cascade down from above, and has little effect at the larger
scales.
Table 1 – Length scales, time scales, turnover time and Reynolds number of the smallest and largest eddies in turbulent flows
Let's start with the largest eddies – those responsible for passing energy down the cascade. We know
that their kinetic energy (per unit mass) !! !~!! ! , and so the rate at which energy passes down the
cascade is:
!! !!
!!~ =
(!/!) !
–––––––––––––––––––––––––––––––––––––––––––
In a steady state situation (such as our mixing tank after some time), the rate at which energy passes
down the cascade must exactly match the rate at which energy is dissipated. We know from our earlier
discussion that the viscous dissipation rate (per unit mass) is:
! = 2!!!!!" !!!" = 2!!!!!" !!!" + 2!!!!′!" !!′!"
! = ! + !′
We also know that viscous dissipation is predominantly at the smallest eddies, and so the viscous
dissipation is simply:
!′ = 2!!!!′!" !!′!"
Where !′!" is the turbulent rate of strain associated with the smallest eddies, given of course by
!′!" !~!!/!. The viscosity is, as always, given by υ.
–––––––––––––––––––––––––––––––––––––––––––
Therefore, the energy passing down the cascade from the largest eddies balances with the energy
being dissipated by the smallest eddies:
!!!!! = !′
!! ! !
!!~!!!!
! !
!
!!~! !/!
!!!
So then, we find that the larger the Reynolds Figure 16 – The smallest scales of turbulent motion are smaller for
number, the smaller the smallest eddies will be! higher Reynolds number flows
*
It is well documented[15] that in mixing tanks, the turbulent component of the velocity is in the order
of 5-10% of the total velocity. Therefore, we can expect the Reynolds number of the largest scales,
based on this turbulent velocity, to be 200 < !!! < 700.
This means that the Kolmogorov Microscale of our mixing tank can be expected to be:
Therefore, the size of the eddies in the mixing tank range from as small as 0.5mm to as large as 80mm.
But what does this mean for providing the description of the velocity field, and how has it helped us
decide on which approach to take?
So*–*Absolute*or*Statistical?*
Let's make the case for numerically solving the total velocity field – that is, providing as accurate a
specification of the initial, boundary and material conditions as possible, and then integrating the
Navier-Stokes equation forward in time.
This would require us to numerically compute flow structures ranging from 80mm to 0.5mm, over a
range of time scales. If we did not resolve the flow structures of the Kolmogorov Microscales, then we
would essentially be spatially averaging, and in doing so would miss out on some of the physics. If we
did not resolve the largest scales, we would be doing the same – worse still, we would be spatially
averaging the scales containing the most energy! Resolving the full range of scales in this manner is
called direct numerical simulation (DNS). It is prohibitively computationally expensive for all but the
simplest of geometries, and for all but the lowest of Reynolds numbers. Therefore, such a calculation
for the mixing tank simply isn't practical on a desktop computer.[16]
We indicated earlier that a solution based on the total velocity field might not be so useful, given its
irreproducibility. Now we have shown that it isn't computationally practical either. Thus, we had
better begin searching for a statistical solution!
We start with the unknown quantities from the closed system we presented at the end of the previous
chapter: (!, !! , !! , !! ). We can decompose these unknowns (along with the passive scalar, let's say
number of particles '!') into their mean and fluctuating components:
!!!! = !! + !! ′
Eq. 10
! = ! + !′ ! = ! + !′
! = ! + !′
Reynolds Decomposition
With the above definitions to hand, and the time-averaging rules given in Appendix B, we may now
set about re-defining the closed system of equations in time-averaged form – commonly referred to as
Reynolds-averaging.
ReynoldsDaveraged*Continuity*Equation*
We begin with Reynolds-averaging the continuity equation:
!"
+ !!! ⋅ (! + !) = 0
!"
Therefore:
!⋅ ! +!⋅ ! =0
!⋅!=0
Thus, the turbulent velocity field is also solenoidal. So the mean and turbulent velocities satisfy the
same form of mass conservation as the total velocity field does – which we would have expected,
since the continuity equation is linear.
ReynoldsDaveraged*NavierDStokes*(RANS)*Equation*
We now repeat this procedure for the Navier-Stokes equation, using Reynolds decomposition to re-
express the equation, and then Reynolds-averaging:
! !! + !! ! !! + !! ! !+! ! !! + !!
! + !! + !! ⋅ =− + !!!!
!" !!! !!! !!! !!!!
!!! !!!
!!" = !!!! + !
!!! !!! Eq. 11
And hence:
! ! !!! !!!
!!" = !!!! +
!!! !!! !!! !!!
!!! !!! !! !
! + !! ⋅ =− + ! − !!!! !!
!" !!! !!! !!! !"
Eq. 12
! !
! (!! ) = −!! + ! !!" + !!"
!!
Where we have defined a new form of the material derivative – the mean material derivative – as:
! ! ! Eq. 13
= + !! ⋅
!! !" !!!
Thus, it seems as though these new unknowns may be interpreted as a surface force – a 'stress', in
other words. This new term is called the Reynolds Stress:
It's immediate consequence for our goal of providing the velocity field description in the mixing tank
is obvious – it adds a further six unknowns. Given that we have only four new equations – one from
the conservation of mass (Reynolds-averaged continuity equation) and three from the conservation of
momentum (RANS) – the system is no longer closed.
This raises two immediate questions, which we will attempt to answer in the forthcoming sections:
(i) – What is the physical meaning of this so-called Reynolds "Stress"
(ii) – How can we provide additional equations for the Reynolds Stress to close the system?
!!
= !! ! ! + !"#$%& − !"#$
!"
Introducing the Reynolds averaged quantities and re-arranging, just like we did previously, we obtain:
! Eq. 14
! ! = !!! ! !! !! − !! ∇ ∙ !! !! + !"#$%& − !"#$
!!
Where the final term on the right (!! !! ) is the scalar flux. This arose in a similar way to they Reynolds
Stress, with one important difference – the turbulent velocity in the above equation is uncorrelated,
meaning it is not multiplied by another turbulent velocity in the way that the Reynolds Stress is. This,
of course, is because the passive scalar equation is linear. Nonetheless, it is an additional unknown.
This equation will prove to be important in how we relate the velocity field to the number of particles
in the mixing tank – which is, after all, a scalar quantity. Since the volume fraction of the particles is
very small compared to the fluid itself, we assume that the solid particles in dispersion have no effect
on the flow field (meaning the ‘number of particles’ is a passive scalar). Therefore we are left to
consider only the effect of the flow on the particles, which of course is the primary consideration of
this work.
The unknown scalar flux term must be closed. We will attend to this matter later.
For now though, we return to consider the two proposed questions about the Reynolds Stress.
(i)*What*is*the*physical*meaning*of*this*soDcalled*Reynolds*"Stress"?*
Let's consider the Navier-Stokes equation applied to a control volume, which we can obtain from the
non-conservative form by application of Reynolds Transport Theorem[17]:
! ! ! !
!
!!!! !!" = − !!!! ! !!!! !!" + (!!" !! )!!" − (! !! )!!"
!"
!! !" !" !"
Conservative*form*of*RANS*
If we now use the Reynolds decomposition and time-average (just like we did previously), then:
! ! ! ! !
!
!!!! !" = − !!!! ! !!!!! !!" + −!!! !! !! !!"! + !!" !! !!" − !!!! !"
!"
!! !" !" !" !"
This is the RANS equation in integral form. The terms in this equation can be understood as meaning:
• The rate of change of (mean) momentum in the cube (of volume δV).
• Minus the rate at which (mean) momentum flows out through the surface (of area δS), carried
by the mean velocity
• Minus the rate at which (mean) momentum flows out through the surface (of area δS), carried
the turbulent velocity
• The net (mean) viscous force (of area δS).
• The net (mean) modified pressure force acting on the surface (of area δS).
So the Reynolds Stress isn't actually a stress at all – it's a fraud! To better understand its true physical
meaning, let's consider the total momentum flux out of the control volume. We'll consider only the !-
direction for simplicity.
The momentum fluxes through the side of the cube which contribute to a rate of change of momentum
in the !-direction are:
(ii)*How*can*we*provide*the*further*Equations*we*require*to*close*the*system?*
To answer the second question, let's first consider how we provided the necessary equation for the
viscous stress.
The viscous stress can be thought of as mixing of momentum at the molecular scale. We then want to
determine the relationship this viscous stress has to the flow – in other words, a relation between a
kinetic property (viscous stress) and a kinematic property (strain) – in other words a 'constitutive'
relationship! Since the scale of the viscous stress is very small, and the scale of the flow is much
larger, we can easily find an intermediary scale with which to average the required constitutive
property (viscosity) over – in other words, the continuum hypothesis can be considered valid. There
are several reasons why this wouldn't work for the Reynolds stress.
Firstly, and most critically, it's not a real stress, it's a flux of momentum (a kinematic property, not a
kinetic property). Therefore, we're trying to relate one kinematic property of the flow to another
kinematic property of the flow – this is not the job of 'constitutive' relationships!
Secondly, the Reynolds Stress is the mixing of momentum at the scale of the flow. The mean strain
(by definition) is also on the scale of the flow. So even if we were to consider the Reynolds Stress as a
real kinetic property, we cannot find an intermediary scale with which to average some kind of
fabricated, ad hoc constitutive property over.
Both these points, of course, are a manifestation of the fact that the turbulence is a property of the
flow, not of the fluid!
So how can we provide the additional equations we require? One possibility would be to provide a
governing transport equation for the Reynolds Stress. Let's see if we can concoct such an equation.
! !! ! !
! (!! ) = − + ! + !!" !!!!!!!!!!!!!!!!!!!!!!!!!!(!"#$!!"#$%&'()
!! !!! !!! !"
Step*2*–*Transport*Equation*for*the*turbulent*velocity* !! !! *&* !! !! *
Multiplying the first equation by !! and averaging, we obtain:
!!! !!! !!! !!! !! ! !! ! !"
!! !! + !! ! !! + !! !! + !! !! = −!!! ! + !! !!!!!!!!!!!!!, ! = 1,2,3
!" !!! !!! !!! !!! !!!
Step*4*–*ReDarrangement*
The above equation can be understood in a more physically intuitive form by simply re-arranging the
terms:
! !!! !!! ! ! ! ! !
! !! !! = !!
!" + !!
!" + !!! !! !! − ! !! − ! !! + 2!! ! !′!"
!! !!! !!! !!! !!! !!!
!"#
!"#$%&'(#) !"#$%&'"( !"#$%&'"(
!! !!! !!!
+! ! ! !!! !! − 2!"
!!! !!! !!!
!"#$%&'"( !"#"$%&%"'(11
!"##"$%&"'(
The physical meaning of these terms will become important shortly, in a discussion on kinetic energy.
For now we note that:
• Production – the production of Reynolds Stress by the mean velocity gradients.
• Dissipation – the mean viscous dissipation of Reynolds Stress due to fluctuating velocity.
• Transport – all these terms represent the mean spatial redistribution of the Reynolds Stress
by the turbulent velocity, the pressure fluctuations and the viscous stresses respectively.
• PSC – the 'pressure – strain-rate correlation' gives the correlation between
the fluctuating pressure (!′) and the fluctuating strain rate (!′!" ).
11
Note that this term only represents the viscous dissipation of homogeneous flow, which is not the case for sheared flows such as the mixing
tank. However, at the Kolmogorov microscales, the.[18]
We will see later that this 'production' is involves a transfer of kinetic energy from the mean flow to
the turbulence.
However, it should be immediately obvious what the problem is from our perspective – we have
generated a considerable number of new unknowns. Not all of these new unknowns are independent,
because some can be derived from others[19] – however, we've certainly opened a can of worms!
Let's focus on one of these new unknowns in particular – the triple turbulent velocity correlation,
!! !! !! – and how we could go about closing this term.
Let's recall how we ended up in this slump. In the "Equations of fluid motion" chapter, we developed a
perfectly closed system capable, in principle, of providing the Eulerian description of the velocity field
we required. However, we detected the presence of turbulence in the mixing tank, and learned that the
velocity would behave chaotically, and over a wide range of scales. This meant that solving the system
numerically would be too computationally expensive.
Conversely, we saw that the statistical properties do not behave chaotically, and were in fact
predictable. To that end we developed a statistical system capable of making such predictions – only
to find that the non-linear term had come back to bite us, giving birth to the Reynolds Stress for which
we can provide no closed equation. This is what we meant when we said the non-linearity would take
us "out of the woods and into the jungle".
The closure problem has haunted all of the brilliant minds to tackle the subject, from its earliest
pioneer, Osborne Reynolds. Though many great explorers have mapped out important sections of this
mysterious jungle – Prandtl, Richardson, Taylor, Kolmogorov, Corrsin – we are still no closer to
understanding whether a grand theory of turbulence even exists.
When we first predicted the presence of turbulence, we asked: what consequences will this 'turbulence'
have for our quest of providing an Eulerian description of the velocity field in the mixing tank?
We should be warned, however, that this alternate path – that of proposing turbulence closure models
– is riddled with questionable hypotheses, which are often applicable to narrow classes of problems
and heavily reliant on empirical data.
Notwithstanding, we proceed to consider one such category of models – the eddy viscosity models.
Boussinesq*Hypothesis*
To this end, we return to the notion that the Reynolds Stress is actually a kinetic property – that is, a
real stress. Recall that this is what the RANS equation actually suggested in the first place:
! !! ! !
! (!! ) = − + ! + !!"
!! !!! !!! !"
In this case, as far as the mean flow is concerned, the mixing of momentum caused by the Reynolds
Stress is analogous to the molecular mixing of momentum by the viscous stress. It might then occur to
provide a similar constitutive quantity – a viscosity – to quantitatively relate this stress to the
kinematic property of interest – this time the mean strain.
This ad hoc viscosity is called the eddy viscosity (!! ), and its proposed effect is to increase simply the
effective viscosity:
!!"" = ! + !!
! !!! !!! 1
!!" = −!!!! !! = !!!!! + − !! !!! !! !!!"
!!! !!! 3
! 1
!!!!!!" ! + !!! !!! !! !!!" ! = !! !!!! ! 2!!!" Eq. 15
3
!"#$"%&'(& !"#$!!"#$
Boussinesq Hypothesis
!
Since the normal stresses of !!" must sum to −!!!! !! , the term on the right is required.
!
!!" = −!!!!! !!
Therefore, the above diagram can be re-interpreted as portions of fluid being tossed between layers by
the turbulent velocity. This analogy between the mixing of momentum due to thermal agitation, and
mixing of momentum due to the turbulent velocity led Prandtl to propose an analogous macroscopic
kinetic theory for turbulent flows:
Eq. 16
!!!!!!!!!!!! = ! !! !!
Prandtl Hypothesis
Where !! is the so-called 'mixing length', while !! is an appropriate measure of the magnitude of the
turbulent velocity |!|. The physical intuition is rather clear:
• The portion of fluid being jostled around will conserve its properties for a characteristic length,
!! , before mixing with the surrounding fluid.
• The greater !! , the more energetic the turbulence, hence the greater the tossing around of fluid
portions, hence the greater mixing of momentum, hence the larger the 'viscosity' !! .
!
Thus, we started out with the question of "what is !!" ?"
We then used the eddy viscosity hypothesis to reformulate this question to "what is !! ?"
Now we have used Prandtl's mixing length theory to finally re-formulate it as "what are !! & !! ?"
Kinetic*Energy*of*the*Mean*Flow*
To determine the kinetic energy of the mean flow, we must firstly recall the transport equation for the
mean flow itself – the RANS equation:
! !! !!!" !!!!! !!
! (!! ) = − + −
!! !!! !!! !!!
We can therefore find an expression for the kinetic energy (per unit volume) of the mean flow simply
by multiplying this equation by !! and halving as follows:
!
! 1 !! !!!" !!!!
! !!! !!! ! = −!! ! + !! ! + !! !
!! 2 !!! !!! !!!
Re-arranging, we obtain:
! 1 ! ! ! ! Eq. 17
! !! = −!!!!!" !!! ! + ! !!" !!! ! + ! !!" !!! − !!" !!!" − 2!"!!!" !!!"
!! 2 !!!
Kinetic Energy (per unit volume) of the Mean Flow
We can identify the first term on the right (the divergence term) as being the rate of working of the
mean pressure, the mean viscous stress and the Reynolds stress on the boundary. If we consider the
mixing tank as a whole, which is of course a closed domain, then this term will integrate to zero. The
last term on the right should be easily recognisable as the rate of mean kinetic energy dissipation rate
(!!). For the mean flow, this term will be very small – this is because at the scale of the mean flow,
the Reynolds number is high12. We will find out the meaning of the unknown term shortly.
!"#$%&"'!!ℎ!"#$ !"!"#$%&'!!" !
= − ! !!" !!!" ! − !! !!!!!"##"$%&"'(!!"!!"#$!!! !!!(!!) !!
!"!!"#$!!! !"!!"#$!!!
!"!!"#$"%"&'$(!!"##"$%&"'(!!"!!"#$%&'(!!"#$%
12
Note, however, that the viscous dissipation isn't negligible at the boundaries (i.e. near the walls), where the no slip condition (! = 0)
means that the Reynolds number (!" = !"/!) is small. Consequently, the viscous forces are important near the walls, and the mean viscous
dissipation also becomes important.
! !!! !!! ! ! ! ! !
! !! !! = !!
!" + !!
!" + !!!! !! !! − ! !! − ! !! + 2!! ! !′!"
!! !!! !!! !!! !!! !!!
!"#
!"#$%&'(#) !"#$%&'"( !"#$%&'"(
!! !!! !!!
+! ! ! !!! !! − 2!"
!!! !!! !!!
!"#$%&'"( !"#"$%&%"'(
!"##"$%&"'(
! !!! !!! ! ! ! ! !
! !! !! = !!
!" + !!
!" + !!!! !! !! − ! !! − ! !! + 2!! ! !′!!
!! !!! !!! !!! !!! !!!
!"#
!"#$%&'(#) !"#$%&'"( !"#$%&'"(
!! !!! !!!
+! ! ! !!! !! − 2!"
!!! !!! !!!
!"#$%&'"( !"#"$%&%"'(13
!"##"$%&"'(
! 1 !
! 1 ! Eq. 18
! !! = −! ! !!" !!! ! + ! !! !!′!" + !!!! !! !! + !!" !!!" − 2!"!!′!" !!′!" 14
!! 2 !!! 2
Kinetic Energy (per unit volume) of the Turbulence
When we presented the Reynolds Stress transport equation, we gave some physical meaning to the
terms involved – and so these meanings can be understood in the new context of kinetic energy. The
first term on the right is the spatial redistribution of kinetic energy by pressure fluctuations, by mean
viscous stress and by the turbulent velocity itself. The last term on the right is the
mean turbulent kinetic energy dissipation rate (!!′). Unlike the mean rate of strain, this term is not
negligible, because viscous forces are important at these small scales (see Section 2.2.4). Once more
we have an unknown term in our equation.
!"#$%&"'!!ℎ!"#$ !"#$%&'"(!!"
= + ! !!!" !!!" ! − !! !"##"$%&"'(!!"!!"#$"%&'!!!! !!! !!′ ! !!
!"!!"#$"%&'!!!! !"!!"#$"%&'!!!!
!"##"$%&"'(!!"!!"#$"%"&"'!!"#$%!!!"!!"#$%&!"#
Transfer*of*Energy*from*the*Mean*Flow*to*the*Turbulence*
If we compare the equation for the kinetic energy of the mean flow with that of the turbulence:
!"#$%&"'!!ℎ!"#$ !!!"#$%&'"(!!"!! !
= − ! !!" !!!" ! − !! !!!!!"##"$%&"'(!!"!!"#$!!! !!!(!!) !!
!"!!"#$!!! !"!!"!"!!!
!"!!"#$"%"&'$(!!"##"$%&"'(!!"!!"#$%!!"!!"#$!!"#$
!"#$%&"'!!ℎ!"#$ !"#$%&'"(!!" !
= + ! !!" !!!" ! − !! !"##"$%&"'(!!"!!"#$"%&'!!!! !!! !!′ ! !!
!"!!"#$"%&'!!!! !"!!"#$"%&'!!!!
!"##"$%&"'(!!"!!"#$"%&'!!!"#$%!!!"!!"#$%&!"#
The meaning of the unknown term in both equations should now be obvious – it clearly represents the
transfer of kinetic energy out of the mean flow and into the turbulence!
13
As noted previously, when written in this form, the dissipation term is only valid for homogeneous flows.
14
However, written in this form, it represents the true viscous dissipation.
(1) If we take the physical interpretation of the Reynolds Stress as a real stress, then we can arrive at
an understanding by analogy with the viscous stress. The viscous stress acts to resist the deformation
of the fluid parcel caused by the instantaneous rate of strain. The energy expended in doing so ends up
as internal energy. By analogy, the Reynolds Stress resists the deformation of the fluid caused by the
mean rate of strain. The energy expended in doing so ends up in the turbulence (in the turbulent
velocity). Thus, the Reynolds Stress removes kinetic energy from the mean flow and transfers it to the
turbulence – but how can we physically interpret this transfer? Yes, it comes from the work the
Reynolds Stress does against the mean rate of strain – and so rain does come from the sky!
(2) To paint a simplified mental picture, let us return to the idea
of a tangled tussle of vorticity. Consider the fate of such a tangled
blob of vorticity in a simple sheared flow (Figure 21). The mean rate
of strain in this simple example acts to stretch the blob of vorticity –
thereby decreasing its moment of inertia. Through the conservation
of angular momentum (!), this leads to an increase in the angular
velocity – and so increases the kinetic energy of the turbulence: Figure 21 – The stretching of a portion of
vorticity by the mean strain
1
!= !!!(!!×!!)
2
This, of course, is the vortex stretching term we identified earlier, and we now realise its importance.
The*Energy*Cascade*Revisited*
With an insight into how kinetic energy is transferred, it would be useful to connect this to what we
already know about what happens to energy in turbulent flow – thus we recall the energy cascade.
We can now define the rate of generation of kinetic energy (per unit mass) in the energy cascade
(from the mean flow into eddies at the integral scale) as:
!
!!" !!!"
!= = ! −!!! !! !!!"
!
We have previously defined the rate of mean turbulent kinetic energy dissipation rate (per unit mass)
at the Kolmogorov Microscales to be:
Eq. 19
!′ = 2!!!′!" !!′!"
However, molecules have a fundamentally different nature of physical interaction. They are discrete,
and their collisions are intermittent. Turbulent eddies, on the other hand, are distributed and in a state
of constant interaction.
In addition, for molecular mixing of momentum, the molecular mean-free path is significantly smaller
than the characteristic length of the mean flow – in turbulent flows of course, the integral scale is
comparable to the characteristic scale of the mean flow.
––––––––––––––––––––––––––––––––
The eddy hypothesis extends the above notion by proposing that the relation between the Reynolds
!
Stress (!!" ) and the mean rate of strain (!!" ) can be defined in a similar constitutive manner as the
!
viscous stress. In effect, the two symmetric, anisotropic components !!" !&!!!" each have five
independent components, and the eddy viscosity hypothesis assumes these can be related using the
scalar coefficient, the eddy-viscosity (!! ).
!
By using this scalar quantity (!! ) to define a linear relationship between (!!" ) and (!!" ), we are
essentially assuming that the flow is isotropic, such that:
! ! !
!! = !! = !!
In mixing tank operations, particularly in the impeller discharge[20], the turbulence is known to have
an anisotropic nature – meaning that the assumption of such correlations is generally not valid.
––––––––––––––––––––––––––––––––
A further consequence of the eddy viscosity hypothesis is that is assumes the Reynolds Stress is
determined by the local mean strain rate (!!" ), and not by the history of the strain.
However, we know that the Reynolds Stress comes from the turbulence itself (it is a correlation of
!
turbulent velocites). Therefore, we can expect !!" to be increased in accordance with the turbulence.
We also learned that the turbulence can be increased by the stretching of vortex blobs and tubes.
With this in mind, consider again the simple sheared stretching of vorticity we presented earlier
(Figure 21). It is clear from this graphic that the vorticity will depend on the history of the mean rate of
strain – and so too, by extension, will the Reynolds Stress.
If the turbulent fluctuations occurred on a timescale which was much smaller than the timescale of the
meal flow, then we could make the case that the local conditions would have reached a relative
statistical equilibrium. However, the Reynolds Stress stems predominantly from the large eddies,
whose timescale is comparable to that of the mean vorticity.
Therefore, we cannot simply consider that at any instant, the flow will reach a statistical equilibrium
rapidly enough for us to consider only the instantaneous rate of strain.
––––––––––––––––––––––––––––––––
The take-forward message from the above limitations is clear: any turbulence closure model based on
the eddy-viscosity hypothesis must be viewed in light of these limitations, and regarded with caution.
With this in mind, we will consider one such model which has become the standard working
turbulence closure model used in engineering: the k-ε model.
Before doing so, we will briefly shift focus to consider a passive scalar quantity again.
The scalar flux is the transport (or flux) of the scalar quantity (!) resulting from the turbulent velocity
fluctuations – the gradient-diffusion hypothesis states that this flux is down the mean scalar gradient
(−∇!!). Therefore:
!! !! = −!!! !∇!!
Where !! is defined as the turbulent diffusivity. Thus, we can re-define our mean passive scalar
equation:
!
! ! = !!! ! !! !! − !! ∇ ∙ !! !! + !"#$%& − !"#$
!!
!
! ! = !!! ! !! !! − !! ∇ ∙ !! !∇!! + !"#$%& − !"#$
!!
!
! ! = ! ! + !! !! ! !! ! + !"#$%& − !"#!
!!
It worth noting that the form of the equation has returned to that of the original passive scalar
equation, with the exception that it now contains Reynolds averaged quantities. This is mathematically
analogous to Fick's law of diffusion.
Figure 22 – The Reynolds-averaged "number of particles" in a fluid parcel after the Gradient-Diffusion Hypothesis
The similarity to the Eddy-viscosity hypothesis is clear. We started with three unknown terms in this
case (one for each component of turbulent velocity) and ended up with a single unknown scalar
quantity. Thus we now have to close both the turbulent viscosity and the turbulent diffusivity.
However, it has been shown [16][14] that for most flows, the turbulent Schmidt number for a passive
scalar is approximately unity:
!!
!"! = !~!1
!!
! !! Eq. 20
!!=! ! + !! ! !! ! + !"#$%& − !"#$
!! !"!
Reynolds-averaged Passive Scalar Equation (Gradient-Diffusion Hypothesis)
Thus, we can proceed to search for closure of the turbulent viscosity alone. Firstly, however let's give
a brief critique of the gradient diffusion hypothesis.
The gradient diffusion hypothesis assumes that the three components of the scalar flux (!! !! ) can be
related to the mean scalar gradient vector (! ! !!) through a scalar coefficient (!! ). This assumes these
vectors are aligned. However, it has been shown that for homogeneous turbulent shear flow, these
quantities are considerably misaligned.
! 1
!!" = !!!!! !!!" − !! !!! !! !!!"
3
Then, using the Prandtl Hypothesis, we had re-formulated it once more to be a question of
determining the mixing length (!! ) and a suitable measure of the magnitude of the
turbulent velocity (!! ):
!!!!!!!!!!!! = ! !! !!
––––––––––––––––––––––––––––––––
Given the assumption that the Reynolds Stress is uniquely determined by the local conditions, it
would seem reasonable that the eddy-viscosity !! would be determined by the characteristics of the
large eddies at that the local position in question. We are also working on the assumption that there is
an absence of severe anisotropy. Therefore, it is natural to define:
!! !~! !
!! = !
!
1 !
!! !~! !! ! = ! !/!
2 ! !
Where ! is the kinetic energy (per unit mass) of the turbulence, and ! is the intregral scale associated
with the largest eddies in the turbulent flow. Hence we can re-express the turbulent viscosity as:
!! !~!! !/! !!
––––––––––––––––––––––––––––––––
We also noted at the end of our discussion on the energy transfer to turbulence that:
!! ! ! ! !/!
!!~!!!~! !~!!′!!!!!!!!!!!!!!!!!!!!!!"#!!"!!!!!!!!!!!!!!!!!!!!!!!!!!~! =
! !′ !′
Where !′ is the mean turbulent kinetic energy dissipation rate (per unit mass) at the Kolmogorov
Microscales. Therefore, re-expressing the eddy viscosity once more we obtain:
! !/!
!! !~
!′
!! Eq. 21
!! ! = !!
!′
k-ε Hypothesis
Where !! = 0.09 is a constant[21]. This constant has been determined by numerous iterations of data
interpolation and fitting for a wide range of turbulent flows.[14]
––––––––––––––––––––––––––––––––
Thus, our search for closure takes on yet another form, this time for the determination of the turbulent
kinetic energy (!) and the viscous dissipation rate (!′). As we will see, the !– ! model provides semi-
empirical transport equations for these terms.
Stuart McCready (mccready.stuart@gmail.com) 37
2.6.1 – The 'k' equation
Recall the equation for the kinetic energy of the turbulence:
! 1 ! 1 !
!!!! !! = −! ! !!" !!! ! + ! !! !!′!" + !!!! !! !! + !!" !!!" − 2!"!!′!" !!′!"
!! 2 !!! 2
!
! = !!! ⋅ !!! + !!!!!!! − !!!!′!!!
!! !"#$%&'"( !"#"$%&'(# !"##"$%&"'(
––––––––––––––––––––––––––––––––
If we consider the newly defined transport term, we find that it contains three unknowns:
1
![!! ] = −! ! !!" !!! ! + !!!!!! !!′!" + !!!! !! !!
2
We know already that viscous dissipation is concentrated in the smallest scales of turbulence – hence
we can expect this term to be very small above. The key modelling step is that the k-ε model assumes
that the pressure fluctuations caused by the turbulent velocity fluctuations act to spatially redistribute
the turbulent kinetic energy in a diffusive manner, from regions of high intensity to low intensity. The
same applies to the triple turbulent velocity correlations. Specifically:
!"#$%&'"(! = ! = !! !!
where !! is an unknown turbulent diffusivity. This diffusivity is typically taken to be equal to the
eddy-viscosity, since we have seen that the turbulent Schmidt number (!"! ) for most flows is
approximately unity[14].
!!
!"! = !~!1!!!!!!!!!!!!"#!!ℎ!"!#$"!:!!!!!!!!!!! = !! !!
!!
––––––––––––––––––––––––––––––––
Thus, the turbulent kinetic energy can be expressed as a simple transport equation, with an associated
sink and source term:
!
! !!"
! = ! ⋅ !! !! + !! − 2!!!′!" !!′!"
!! ! !"
!
! = ! ⋅ !! !! + ! − !′
!!
This equation is then typically re-written in the following more general form, which allows for
adjustment of the diffusivity by specifying the turbulent Schmidt number (!"! ) [22]:
! !! Eq. 22
! =!⋅ !+ !!! + ! − !′
!! !"!
Therefore, the 'k' equation amounts to an exact equation, where the transport term is modelled as a
simple gradient diffusion. The other terms in the equation (the mean material derivative of 'k', the
generation and the viscous dissipation) are all in closed form. Thus the 'k' equation has been closed.
!
! !! !!!′ !′ Eq. 23
!′ = ! ⋅ (! + )!!!′ ! + ! !! − ! !!
!! !"! ! !
'ε' Transport Equation
Where !! , !! , !! are additional empirical constants, which normally take the following values:
!! = 1.3!;!!!!!!!!! = 1.44!;!!!!!!! = 1.92
Thus, with the inclusion of these empirical constants, the 'ε' equation has been closed.
! 1
Boussinesq Hypothesis !!" = !!!!! !!!" − !! !!! !! !!!"
3
!!
k-ε Hypothesis !! ! = !!
!′
! !!
'k' transport equation ! =!⋅ !+ !!! + ! − !′
!! !"!
!
! !! !!!′ !′
'ε' transport equation !′ = ! ⋅ (! + )!!!′ ! + ! !! − ! !!
!! !"! ! !
where:
!
!!"
!= !!!" !!!!!!!!!!!"#!!!!!!!!!!!′ = 2!!!′!" !!′!"
!
––––––––––––––––––––––––––––––––––––––––
It seems like the model relies on a totally fabricated equation along with some arbitrary coefficients,
designed to reproduce a few laboratory results. Indeed, the k-ε is essentially a highly sophisticated
interpolation between data sets! However, it has proven to be largely effective for a wide variety of
engineering applications, with the exception of flows containing large pressure gradients[14] (not
present in the mixing tank).
We stated already that while the 'k' equation had an identifiable physical origin – being based on the
derived equation for the kinetic energy of the turbulence – the 'ε' equation was simply plucked from
thin air. Is there any inherent rationale behind the equation?
15
Note that both the constants for the turbulent Schmidt numbers are approximately unity, as stated previously.
!!′
!!~
!
Suppose now that there is a homogeneous shear flow in one-dimension:
!!! !
! = !! ! ! ê! !!!!!!!!!!!!! = !!"
!"
Where ê! is the unit normal vector in the !-direction. We have seen already that these large scale
eddies can be distorted by the mean flow, and that this process increases the vorticity of these eddies.
Therefore, it might be concluded that after some time, the vorticity of the large scale eddies would
eventually settle down to a value in the order of the mean rate of strain:
!!"
!!~!!!" !!!!!!!!!!!!!!"!!!!!!!!!!!!!!! =
!
Where ! is a constant. Therefore, we might also conclude that if at any instant, the vorticity of the
large eddies is different from this value, then it would tend towards this value on a timescale of
!!
!!" such that we could define the overall behaviour in a heuristic equation:
!
!! !!"
= !! ! − ! ! !!!!!… … … … … … … … … … … !!!". (!)
!" !
!
!" ! !!′ ! !!′ !′
= ! − ! !′!!!!!!!!!!!!!!!!"#!!!!!!!!!!!!!! = !! ! − ! !!
!" !" ! !
Where:
!
! !! ! ! !!"
! = !! ! !!" =
!′
In the case of the homogeneous, one-dimensional flow, these !– ! equations can be re-arranged:
!! !
!
= (!! − 1)!! !!" − (!! − 1)!! ! !… … … … … … … … !!!". (!)
!"
!!
! ! = !! − 1 !!!!!!!!!!!!!!!!!!!!"#!!!!!!!!!!!!!!!!!!!!!!! = !! !! − 1
!!
Therefore, in conclusion, if !! ,!!! > 1 then the particular form of the '!' equation means that in one-
dimensional homogeneous shear flow, the vorticity has reasonable behaviour. Therefore, one
interpretation of the '!' equation is that it is a transport equation for the vorticity of the large eddies,
which causes their vorticity to tend towards that of the mean flow.
Where ! is the unit normal vector to the surface of the fluid parcel. The terms of the equation can be
understood as:
It is useful to note that there are a large number of variants of Reynolds transport theorem present in
the literature. This is because the formula is extremely general, and can be applied in a variety of
different contexts. As such, different literature will inevitably have equations that often look different
than the above equation in both appearance and complexity.
• Spatial and temporal changes are described by finite steps, rather than infinitesimal steps
• The closed system of equations can be written in algebraic form
There are several approaches to the spatial discretisation, including finite element (FEM) method,
finite difference method (FDM) and finite volume method (FVM). In these methods, each flow
variable is defined at every discrete point in the entire spatial domain. The general method of spatial
discretisation is different for each of the numerical methods – our focus will be on the finite volume
method, since the CFD software used to compute the solution (ANSYS Fluent) is based on this
method.
The principle of the finite volume method is local conservation – and this is the key reason why it is
successful in computational fluid dynamics applications. As we have seen in the previous chapter,
fluid mechanics problems are governed by local conservation principles – the continuity equation is
based on the conservation of mass, while the Navier-Stokes is based on the conservation of
momentum. In this way, using Gauss' theorem, these laws of conservation can be recast for each cell
as linear algebraic equations – and these equations can be solved iteratively.
The mesh must be fine enough to capture the important details of the
flow and provide an appropriate resolution – however, a finer mesh also
means more cells, which means greater computational cost.
The mesh used for the mixing tank is shown in Appendix C. Figure 24 – Polyhedra possible from
3D meshing operation
To illustrate how these conservation equations can be discretised, we will consider the transport of
some passive scalar property in two dimensions. The conservative form of the mean passive scalar
equation, for some passive scalar '!' is given by:
!!" ! ! !!
+ !!!! !! ! = ! !! ! + !!"#$%&
!" !!! !!! !!!
Discretised*Equations*
The process discretisation begins with integrating the conservative form of the mean scalar transport
equation over the control volume (cell) of interest, and using Gauss' theorem to convert the volume
integrals to surface integrals. Doing so, we obtain the integral form of the conservative transport
equation:
!! !!
! !! ! !! ! !!! ! − ! !! ! !! ! !!! !! = !!! !! − ! !! ! !! + !! !"#$%&
!" ! !" !
Where the final term is the volume integral of the source term. The problem is, of course, that the
equation requires values at the faces, while the values themselves are stored in the computational
nodes in the cell centres. Therefore, interpolation – or a 'differencing scheme' – is required.
• Conservativeness – The flux out of a cell should be equal to the flux into the corresponding
neighbouring cell. We noted already that the finite volume method is based on the principle of
conservation, and so this criterion is automatically satisfied.
• Boundedness – The face value should be neither larger nor smaller than the cell values used for its
computation. In other words, the face value must lie within the cell values used to calculate it. This
ensures that no spurious minima/maxima are present within the domain.
• Accuracy – The differencing schemes assume shape for the variable of interest (Φ). This is
approximated by a Taylor series expansion:
!′ !! !′′ !! ! ! !!
! !! = ! !! + (!! − !! ) + (!! − !! )! !!…!! (!! − !! )!
1! 2! !!
Upwind*Differencing*
For !" >> 1, convection dominates, as would be expected in the mixing tank – meaning an upwind
differencing scheme is used. This means that the values required for the faces can be assumed to be
dependent only on the values at the computational nodes further upstream (upwind).
!! = !!
!! − !! !! − !!
= !!!!!!!ℎ!"#!:
!! − !! !! − !!
The major disadvantage of the second-order upwind differencing scheme is that it is unbounded. This
can lead to convergence and stability issues.
Numerical*Diffusion*
Numerical diffusion is an inherent error in CFD calculations, and presents itself as an increase in the
diffusion coefficient. Therefore, in the mean passive scalar transport equation, the diffusion coefficient
would appear to be higher than it actually is. Similarly, in the RANS equation, the viscosity would
appear to be higher than it actually is. The relative effect of numerical diffusion is greater when
diffusion is small.
Numerical diffusion is more problematic in course meshes. Therefore, it is important that a suitably
fine mesh is used, particularly for the regions where accuracy is required, such as those where a high
velocity gradient would be expected. Accordingly, as the mesh in Appendix C shows, the regions near
the impeller and the wall have a finer mesh structure.
In addition, numerical diffusion is more pronounced when the flow is misaligned with the mesh.
Clearly, in the turbulent mixing tank, this misalignment will be significant, and the chaotic flow field
will rarely be aligned with the mesh. To minimise numerical diffusion, a higher order discretisation
scheme can be used. Hence, in the calculation of the final solution, the second-order upwind scheme is
used. This is highlighted in Figure 28.
• Segregated – This approach involves solving one variable at a time, for the entire spatial
domain. For example, !! is solved across the entire domain, followed by !! , then !! and so
on and so forth. The 'iteration' is complete only when all of the variables have been solved in
this way.
• Coupled – This approach solves all of the equations simultaneously, but for one cell at a time.
The 'iteration' is complete only when this process has been repeated for all of the cells in the
entire spatial domain.
Since we are assuming incompressible flow, and the mixing tank operation is turbulent, the segregated
approach is preferred [29].
PressureDVelocity*Coupling*
Since there is no explicit equation for the pressure, other approaches have been devised to extract it –
these are so-called pressure-velocity coupling methods. The particular method used in these
simulations was the semi-implicit method for pressure-linked equations (SIMPLE).
In the SIMPLE approach, a guess is made for the value of pressure so that the RANS equation can be
solved. This 'guessed' value is of course the last updated value from the previous iteration. Using this
pressure, new velocities are calculated, which will not satisfy the Reynolds averaged continuity
equation. Therefore, the velocities are corrected accordingly, and used to determine a new value for
the pressure. Thereafter, the other variables are calculated, and if convergence is not met, a new
iteration begins.
Residuals*
The solution of the algebraic equations involves an iterative process – therefore, at any step in this
process, the calculation is based in the inexact values calculated from the previous iteration. These
inexact values originate from the initial 'guessed' value, and in a convergent solution, are refined
through repeated iterative calculations.
The total residual for a given equation is the sum of the residuals of all the cells in the entire domain.
In the final calculations (using the sliding mesh approach, for both flow field simulations and also for
the simulations involving the solution to the PBE), the following convergence criteria were used:
Equations. .Absolute.Criteria.
Velocities#(x,y,z)# 0.00001#
k# 0.00001#
ε# 0.00001#
It is important to note that this type of convergence criteria only indicates when overall balances are
achieved. It requires engineering discretion to verify that convergence has indeed been reached.
Since we are interested in providing a description of the steady state velocity field in the mixing tank,
it is wise to consider other convergence criteria in addition to the absolute residuals above. To this
end, the impeller moment, average velocity magnitude and average tangential magnitude were
monitored. These monitors, along with the residuals, are discussed in more detail in the results section.
UnderDrelaxation*
We have seen already that the iterative solution procedure involves using information calculated from
the previous iteration. Therefore, some small change in the value (∆!) is added to the value from the
previous iteration (!! ) to yield the new value (!!!! ):
!!!! = !! + ∆!
However, in iterative processes involving the solution of a set of couples equations, it is common to
introduce a 'relaxation factor' (!), such that:
!!!! = !! + !∆!
This is called 'under-relaxation', and helps to reduce unstable oscillations in the flow solution, but
naturally increase the time required for convergence. The values of the relaxation factors typically
range from 0 to 1. In the case of the mixing tank, the standard Fluent relaxation factors were used:
.. UnderMrelaxation.Factors.
Pressure# 0.3#
Density# 1#
Body#Forces# 1#
Momentum# 0.7#
Turbulent#Kinetic#Energy# 0.8#
Turbulent#Dissipation#Rate# 0.8#
Turbulent#Viscosity# 1#
In order to solve a temporal solution, the time derivative must be discretised. For the example of the
mean passive scalar transport equation:
!!" ! ! !!
!!! !!! + !!!! !! ! = ! !! ! + !!"#$%&
!" !!! !!! !!!
!"#$
!"#$%&'$%"
The first order discretisation of this time derivative for incompressible flow is:
! !!! − ! !
= !(!)
∆!
Where ! ! is the value of the variable at time '!', while ! !!! is the value at time '! + ∆!'. The spatially
discretised part of the equation is denoted by !(!), and has been described previously. This spatially
discretised part can be evaluated at time '!' or at time '! + ∆!' – this is referred to as an explicit and
implicit approach respectively. For the mixing tank operation, the preference is to use an implicit
approach, and as such:
! !!! − ! !
= ! !!!!
∆!
The underlying assumption is that the newly calculated value for the variable (!) is constant
throughout the entire time step – that is, from time '!' to time '! + ∆!'.
To do so, an implicit or explicit approach may be used. In the implicit approach requires experimental
data of the time-averaged velocity in the fluid region immediately adjacent to the impellers. The
explicit method involves modelling the modelling the impeller geometry directly, and specifying either
a time-averaged or steady state approach to meshing and solving.
We consider only explicit approaches, using the physical geometry of the impellers and shaft.
Since mixing tank operation involves the rotation of the impellers in a stationary tank in an otherwise
stationary geometry, a stationary mesh specified at time '!' of the operation will no longer conform to
the new geometric configuration at time '! + ∆!' – at this point the impeller will be at a different
position relative to the remaining geometry of the tank.
This requires us to split our mesh into multiple fluid zones – one for the region surrounding the
moving impeller and shaft (hereafter called the "impeller zone"), and another for the remaining
stationary tank and impellers (hereafter called the "stationary zone"). This naturally requires
information to be shared between zones at the zone interface. There are two approaches to this
problem – the multiple reference frame approach and the sliding mesh approach.19 Both of these
approaches will be discussed, before the final solution strategy is described.
It is important to realise that this approach cannot account for the relative motion of the impeller zone
with respect to the stationary zone – this is because the mesh remains fixed. Therefore, the solution
calculated using this approach is valid only for one particular impeller-baffle orientation. In other
words, we are freezing the motion of the impeller at some point, and calculating the instantaneous
flow field at that orientation.
Consequently, if the interaction between the impeller and the baffles is strong, then this solution
method would only be valid for the particular impeller orientation used. Considering the geometry of
the mixing tank in question, the impeller-baffle interaction can be expected to be significant. The flow
at the interface of the impeller zone and stationary zone is therefore unlikely to be uniform.
As such, this approach is not used for the final solution required to determine the aggregation and
break-up behaviour of the particles. We will see later, however, that this approach can still prove very
useful in the overall solution strategy.
19
There is s third approach, known as 'mixing plane model', which was not used and it therefore not discussed.
Unlike the multiple reference frame approach, in which the solution is valid only for the particular
orientation of the impeller relative to the baffles, the sliding mesh model allows a time-accurate
solution in which the equations are solved for successive impeller-baffle orientations as the baffle
rotates.
The time-iteration for a sliding mesh model requires further comment. The 'sliding' of the mesh
interfaces past one another is not continuous. Rather, the set of conservation equations is solved
iteratively for a particular impeller-baffle orientation until convergence is achieved. The impeller and
surrounding mesh then rotate – the rotational distance dictated by the designated 'time step' size (e.g
0.01s) and the rotational rate of the impeller (e.g. 100rpm) – and the equations are once again solved
iteratively until convergence. During each of the quasi-steady calculations, information is exchanged
at the interface between the impeller zone and the stationary zone.
Recall that the goal is to determine the aggregation and break-up behaviour of particles in the mixing
tank after at steady state conditions. Therefore, we are not particularly interested in the start-up
behaviour of the mixing tank (from when the impeller is initially at rest in the fluid). As such,
performing a time-accurate solution (using the sliding mesh approach) would require many
computational 'time-steps' before the mixing tank arrived at steady state.
Therefore, the multiple reference frame approach was used to determine the instantaneous solution at a
particular impeller-baffle configuration. The converged solution was then simply used as initial
conditions for the sliding mesh approach, thereby reducing the overall computational load (and time).
Operation at 50rpm, 100rpm and 150 rpm impeller speeds were modelled.
This velocity field (in other words, the spatial gradient of velocity in all three dimensions) is described
by the second-order tensor:
As with any second-order tensor, this can be decomposed into symmetric and skew components. To
highlight the physical meaning of these components, we will consider an example flow in two
dimensions only (the third component of velocity, which would point out of the screen, is assumed to
be zero everywhere). Note that the same ideas apply to the mean velocity (!! ).
This gives the rate of shear of the fluid parcel, This gives the rate of expansion (divergence) of the
without any associated change in volume. fluid parcel, which produces a change in volume.
For incompressible fluids, ! ⋅ ! = 0, and hence the isotropic component is zero: !!" = 0. For this
reason, we often refer to the symmetric, anisotropic component simply as the rate-of-strain.
!!! !!!
!! 1 !! 1 !!
= !" = ! !" =
!" ∆! !" ∆! !"
! !
!!! !!!
!" 1 !" 1
= !" = ! !" = 0
!" ∆! !" ∆!
! !
!=!
!+!=!+!
!⋅! =!⋅!
!" !!
=
!" !"
! !" = ! !"
!∙! ≠!⋅!
Here, the interface between the impeller zone and the stationary zone is clearly visible.