Simulating Turbulent Flows: Issues and Challenges

Shyam Sunder
Assistant Professor
Mechanical Engineering

BITS Pilani
November 19, 2023
Shyam Sunder Assistant Professor MechanicalSimulating
Engineering
Turbulent Flows: Issues and Challenges 1 / 41
 Table of contents
1 Characteristics of turbulent flows

2 Turbulent Kinetic Energy and its Dissipation

3 Descriptors of turbulent flows

4 Reynolds-Averaged Models

5 Issues with Reynolds averaging concept

6 Methodology in Large Eddy Simulation

7 Spatial filtering of equations

8 Filtered unsteady Navier-Stokes equations

9 Sub Grid Scale stresses

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 2 / 41
 Characteristics of turbulent flows

A turbulent flow is always unsteady and the flow field variables

including velocity and pressure fluctuate with time continuously and
randomly
This is due to a great many eddies of different sizes, life spans,
orientations, and energy levels in the flow
The flow disturbances are random in nature and their growth is a
complex nonlinear process, the resultant eddies also have a random
aspect
It is difficult to predict their exact characteristics like size, life span,
orientation, and energy level, etc.
The instantaneous flow field of a turbulent flow is always 3-D and
unsteady, as we cannot force the disturbances or eddies to stay on a
specific plane, although the time-averaged flow field can be 2-D, and
statistically steady

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 3 / 41
 Characteristics of turbulent flows
The eddies interact with the mean flow field as well as with each other
Such eddy motions have a diffusive effect to the mean flow field in
the sense that they can smooth out differences in the mean flow field
The differences within the mean flow field (“mean shear” as it is
called) in fact can stretch the eddies and make them spin faster. That
is, kinetic energy can transfer from the mean flow to the eddies
This transfer is most effective for the largest eddies. The same energy
transfer also occurs between eddies
Since larger eddies usually are also more energetic, energy typically
flows from the large eddies to the smaller ones, and then from the
smaller ones to the even smaller ones. This is the so-called energy
cascade in turbulent flows
Does this cascade process continue without limit? Or is there a
certain minimum eddy scale at the end of this chain of energy
transfer?
Shyam Sunder Assistant Professor MechanicalSimulating
Engineering
Turbulent Flows: Issues and Challenges 4 / 41
 Characteristics of turbulent flows
In fact there exists such a smallest eddy scale. The key to understand
this fact is the momentum equation
The diffusion term in this equation becomes more and more
significant as the Reynolds number becomes less and less
At the smallest scale the eddy velocity is also the lowest, therefore the
Reynolds number, being proportional to length scale and velocity
scale, is very small (of order 1) at the smallest eddy scale
And the viscous diffusion can dissipate the kinetic energy of the
smallest eddies instantly (into heat) and no smaller eddies can survive
One should understand that although energy cascade is a net kinetic
energy transfer from large to small scales, it does not mean there is
no opposite transfer from small to large scales
In fact, energy transfer in both directions is present in real turbulent
flows but the forward-scatter typically dominates

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 5 / 41
 Characteristics of turbulent flows
There are some significant differences between the large and the small
eddies
The size of the largest eddies in a turbulent flow usually is comparable
to the characteristic length scale of the flow field; while the size of the
smallest eddies is independent of such scales
The large eddies usually have some preferred directions, which are the
directions along which they can be most effectively stretched by the
mean flow
The smaller eddies gradually lose such direction preferences and
become more and more isotropic as the eddy size decreases
The largest eddies obviously have the most significant contribution to
the flow field fluctuations i.e. the turbulent kinetic energy
The smallest eddies on the other hand contribute the most to the
viscous dissipation of such fluctuations

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 6 / 41
 Turbulent Kinetic Energy and its Dissipation

Based on these considerations, we might estimate the length, velocity,

and time scales of the eddies at the two ends of the eddy size
spectrum
Define the (specific) turbulent kinetic energy as
 
k = 12 u ′2 + v ′2 + w ′2 . Units: m2 /s 2
Here the prime stands for fluctuations: ui′ = ui − u i and the overbar
denotes either time-average or ensemble-average
The time-averaged flow field is obtained by averaging the flow field
over a long enough time period while the ensemble-averaged flow field
is the mean of the flow fields realized in a large number of repetitions
of the same experiment

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 7 / 41
 Turbulent Kinetic Energy and its Dissipation

Let the time rate of the turbulent kinetic energy dissipation per unit
mass of fluid is ϵ with units W /kg or m2 /s 3
Since the largest eddies contribute the most to kinetic energy k, their
velocity scale uL should be of the order of k 0.5
Their time scale tL should be of the order of k/ϵ and their length
scale lL should be of the order of uL tL ∼ k 1.5 /ϵ
The smallest eddies dissipate energy at the rate of by means of
viscosity. It is reasonable to relate scales of the smallest eddies with
and the kinematic viscosity whose SI unit is m2 /s
By requiring the unit of ν a ϵb to be the same as the unit of the desired
scales, we can find the length scale of the smallest eddies, lν , which is
of the order of ν 0.75 /ϵ0.25 ; their velocity scale uη is of the order of
ν 0.25 ϵ0.25 and their time scale tη is of the order of ν 0.5 /ϵ0.5 . Such
scales are called Kolmogorov scales.

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 8 / 41
 Length and Time Scales in Turbulent Flows

If we define a Reynolds number based on the largest eddy scales

ReL = uLνlL , we can find that
lL uL tL
lη ∼ ReL0.75 uη ∼ ReL0.25 tη ∼ ReL0.5
Since Reynolds number is typically very high in a turbulent flow, the
scales of the smallest eddies are usually way smaller than the largest
scales
For example, if ReL = 105 , and the largest eddy size is lL ∼ 1 m, the
size of the smallest eddies is only lη ∼ 10−4 m, which is roughly the
same as the thickness of a piece of paper
Also you can see if the time scale of the largest eddy is tL ∼ 1 s, the
time scale of the smallest eddies is only tη ∼ 10−3 s at the same
Reynolds number

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 9 / 41
 Direct Numerical Simulation
DNS
One way to simulate turbulent flows is simply solving the
Navier–Stokes equations with very fine mesh and very small time step
sizes so that even the eddy motion at the Kolmogorov scale is
calculated
In this way no turbulent models are needed since every flow detail is
resolved and nothing is left for modeling. This method is called direct
numerical simulation (DNS)
DNS is very accurate. In fact DNS data are usually used to calibrate
turbulent flow measuring techniques!
DNS however is very expensive. The reason is because DNS must use
a computational domain big enough to contain the largest eddies, yet
with a grid spacing small enough to capture the smallest eddies
The number of mesh cells in each spatial direction should be of the
order of llηL ∼ ReL0.75 and the total number of mesh cells should be
N ∼ (ReL0.75 )3 = ReL2.25 since turbulence is intrinsically 3D
Shyam Sunder Assistant Professor MechanicalSimulating
Engineering
Turbulent Flows: Issues and Challenges 10 / 41
 Direct Numerical Simulation
DNS
Moreover, the total flow time a DNS simulates should be longer than
the time scale of the largest eddies, but the time step size ∆t should
be less than the time scale of the smallest eddies so that behavior of
all eddies can be discerned
In fact usually an even more stringent criterion is applied to ∆t in
DNS i.e. ∆t has to be proportional to the time required for a
small-scale eddy to pass a fixed point when advected by the
l
large-scale eddy i.e. ∆t ∼ uηL
The total number of time steps needed for such a simulation is
tL
therefore of the order of ∆t ∼ llηL ∼ ReL0.75
Each time step involves solving equations whose number is of the
order of ReL2.25
The total number of equations is thus of the order of ReL3
For example, if we use DNS to simulate a flow with a Reynolds
number ReL = 105 , and if 100 floating point operations have to be
executed to solve one equation, the total computing load is of the
order of 1017 flops EngineeringTurbulent Flows: Issues and Challenges
Shyam Sunder Assistant Professor MechanicalSimulating 11 / 41
 Direct Numerical Simulation
DNS

For this reason DNS is almost exclusively done on super computers

and is restricted to relatively simple flows e.g. the fully developed
turbulent channel flow and duct flow at relatively low Reynolds
numbers
For most turbulent flows of practical engineering interest, which are
usually much more complicated than such simple flows and with
much higher Reynolds numbers, DNS is still not a realistic option at
least for now
Instead we will have to ignore certain, even most of the eddy scales
and only calculate the flow field of the rest scales so that the
simulation can be accomplished in reasonable time frame with
affordable resources
The effects of those neglected scales on the resolved scales then have
to be modeled

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 12 / 41
 Descriptors of turbulent flows
Reynolds averaging decomposes a flow property into an average part
and a time varying fluctuating part with zero mean value:
ϕ(t) = Φ + ϕ′ (t)
Time average or Mean
Z ∆t Z ∆t
1 1
Φ= ϕ(t)dt ⇒ ϕ′ = ϕ′ (t)dt = 0
∆t 0 ∆t 0

In the above equation ∆t has to larger than the time scale associated
with the slowest variations of property ϕ. This happens due to the
influence of largest eddies
This definition of ‘average’ is only adequate for steady mean flows
For time dependent flows, one has to take ‘Ensemble average’
Ensemble average is the average of the instantaneous values of a
property over a large number of repeated identical experiments
Shyam Sunder Assistant Professor MechanicalSimulating
Engineering
Turbulent Flows: Issues and Challenges 13 / 41
 Descriptors of turbulent flows
The descriptors used to indicate the spread of the fluctuations ϕ′
about the mean value ΦR are
∆t
▶ Variance (ϕ′ )2 = 1
∆t 0
(ϕ′ )2 dt
q h R ∆t i1/2
▶ Root Mean Square Value ϕrms = (ϕ′ )2 = 1
∆t 0
(ϕ′ )2 dt
The RMS values of velocity components are particularly important
▶ These can be easily measured and express the average magnitude of
velocity fluctuations
▶ Variances of velocity fluctualtion u ′2 , v ′2 , w ′2 will appear as part of
momentum fluxes induced by turbulent eddies
▶ These will appear as additional normal stresses experienced by fluid
elements in turbulent flows
▶ The total kinetic energy per unit mass k at any location in a turbulent
flow is given by
⋆ k = 21 (u ′2 + v ′2 + w ′2 )
▶ The turbulence intensity I is the average RMS velocity divided by a
reference mean flow velocity Uref and is linked to the turbulence kinetic
energy k √as follows
2k
3
⋆ I = Uref

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 14 / 41
 Descriptors of turbulent flows
We also have to consider moments of different fluctuating variables

Important details of the structure of the fluctuations are contained in

moments constructed from pairs of different variables

Let ϕ = Φ + ϕ′ and ψ = Ψ + ψ ′ be two different physical properties

of the flow with ϕ′ = 0 and ψ ′ = 0
1
R ∆t ′ ′
Their second moment is defined as ϕ′ ψ ′ = ∆t 0 ϕ ψ dt

If velocity fluctuations in different directions were independent

random fluctuations, then the values of the second moments of the
velocity components u ′ v ′ , v ′ w ′ , u ′ w ′ would be equal to zero

Turbulence is associated with the appearance of vortical flow

structures and the induced velocity components are chaotic, but not
independent, so in turn the second moments are non-zero

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 15 / 41
 Descriptors of turbulent flows
Additional information relating to the distribution of the fluctuations
can be obtained from higher-order moments
1
R ∆t ′ 3
▶ Skewness (ϕ′ )3 = ∆tR 0 (ϕ ) dt
1 ∆t ′ 4
▶ Kurtosis (ϕ′ )4 = ∆t 0 (ϕ ) dt

More detailed information relating to the structure of the fluctuations

can be obtained by studying the relationship between the fluctuations
at different times and spatial locations
▶ Autocorrelation in time
Rϕ′ ϕ′ (t ∗ ) = ϕ′ (t) ϕ′ (t + t ∗ )
R ∆t ′
1
= ∆t 0
ϕ (t) ϕ′ (t + t ∗ )dt
▶ Autocorrelation in space
Rϕ′ ϕ′ (⃗x ∗ ) = ϕ′ (⃗x ) ϕ′ (⃗x + ⃗x ∗ )
R ∆t ′
1
= ∆t 0
ϕ (⃗x ) ϕ′ (⃗x + ⃗x ∗ )dt

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 16 / 41
 Reynolds-Averaged Models
Reynolds Averaging

The rationale for Reynolds averaging is that we are not interested in

the part of flow solution that can be described as “turbulent
fluctuations”

Instead, it is the mean (velocity, pressure, lift, drag) that is of interest

Looking at turbulent flow, it may be steady in the mean in spite of

turbulent fluctuations
If this is so, and we manage to derive the equations for the mean
properties directly, we may reduce the cost by orders of magnitude:
▶ It is no longer necessary to perform transient simulation and assemble
the averages: we are solving for average properties directly
▶ Spatial resolution requirement is no longer governed by the
Kolmogorov micro-scale! We can tackle high Reynolds numbers and
determine the resolution based on required engineering accuracy

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 17 / 41
 Reynolds-Averaged Models
Reynolds Averaged Navier-Stokes Equations

Repeating from above: decompose u and p into a mean and

fluctuating component:
▶ u = ū + u′
▶ p = p̄ + p ′

Substitute the above into original equations. Eliminate all terms

containing products of mean and fluctuating values, we get:

▶ ∇ · ū = 0

∂ū
▶
∂t + ∇ · ūū = − ρ1 ∇p̄ + ν∇2 ū − ρ1 ∇ · τ R

▶ Here τ R = −ρu′ u′ is the Reynolds stress tensor

Different Reynolds averaging based turbulence models provide
expressions for τ R

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 18 / 41
 Reynolds-Averaged Models
Reynolds Averaged Navier-Stokes Equations: Modelling Paradigms

We saw that one new term: the Reynolds stress tensor τ R comes into
picture
τ R = −ρu′ u′
which is a second rank symmetric tensor. We have seen something
similar when we wrote the constitutive equation for viscous stresses
In order to close the system, we need to describe the unknown value
τ R as a function of the solution. Two ways of doing this are:
▶ Write an algebraic function, resulting in eddy viscosity models:
τ R = f (ū, p̄)
▶ Add more differential equations, i.e. a transport equation for τ R ,
producing Reynolds Transport Models. As we keep introducing new
equations, the above problem will recur. The first option will need to
be used at some level
Both options are in use today, but first massively out-weights the
second
Shyam Sunder Assistant Professor MechanicalSimulating
Engineering
Turbulent Flows: Issues and Challenges 19 / 41
 Eddy Viscosity Models
Dimensional Analysis
Looking at τ R , the starting point is to find an appropriate symmetric
second rank tensor. Remember that the terms acts as diffusion of
momentum, appears in the equation under divergence and appears to
act as diffusion
Based on this, the second rank tensor is the symmetric velocity
gradient S:
τ R = f (S̄)
where
S̄ = 12 ∇ū + ∇ūT
 

Under divergence, this will produce a ∇ · ∇ū kind of term, which

makes physical sense and is numerically well behaved
Using dimensional analysis, it turns out that we need a pre-factor with
dimensions of viscosity: for laminar flows, this will be [m2 /s] and
because of its equivalence with laminar viscosity we may call it
turbulent viscosity νt
Shyam Sunder Assistant Professor MechanicalSimulating
Engineering
Turbulent Flows: Issues and Challenges 20 / 41
 Eddy Viscosity Models
Dimensional Analysis

The problem reduces to finding νt as a function of the solution.

Looking at dimensions, we need a length and time-scale, either
postulated or calculated. On second thought, it makes more sense to
use velocity scale U and length-scale ∆
We can think of the velocity scale as the size of u′ and length-scale as
the size of energy-containing vortices. Thus:
τ R = νt 21 ∇ū + ∇ūT
 

and
νt = AU∆
where A is a dimensionless constant allowing us to tune the model to
the actual physical behaviour

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 21 / 41
 Eddy Viscosity Models
Velocity and Length Scale

Velocity scale is relatively easy: it represents the strength of turbulent

fluctuations. Thus, U ≡ |u′ |. Additionally, it is easy to derive the
equation for turbulence kinetic energy k:
k = 32 u′ 2
directly from the momentum equation in the following form:
∂k T 2−ϵ
1

∂t + ∇ · (ūk) = ∇ · (νeff ∇k) + νt 2 ∇ū + ∇ū
Here ϵ is turbulent dissipation which contains the length scale:
3/2
ϵ = Cϵ k ∆
where Cϵ is a model constant and ∆ a turbulent length scale that
needs to be specified
The above equation is based on the dimensional arguments and with
the assumption that the dissipation is a function of flow properties
and independent of fluid properties

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 22 / 41
 Eddy Viscosity Models
Zero and One-Equation Models

Zero equation model: assume local equilibrium above: k = ϵ, with

no transport. The problem reduces to the specification of
length-scale. Example: Smagorinsky model

νt = (CS ∆)2 |S̄|

where CS is the Smagorinsky “constant”. The model is actually in

active use but not in this manner. It is used in LES based models

One equation model: solve the k equation and use an algebraic

equation for the length scale. Example: length-scale for airfoil
simulations can be determined from the distance to the wall

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 23 / 41
 Eddy Viscosity Models
Two-Equation Model

Two-equation models are the work-horse of engineering simulations

today. Using the k equation from above, the system is closed by
forming an equation for turbulent dissipation ϵ and modelling its
generation and destruction terms
Other choices also exist. For example, the Wilcox model uses eddy
turnover time ω as the second variable, claiming better behaviour
near the wall and easier modelling
Two-equation models are popular because these accounts for
transport of both the velocity and length-scale and can be tuned to
return several canonical results

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 24 / 41
 Eddy Viscosity Models
Standard k − ϵ Model

This is the most popular 2-equation model, now on its way out. There
exists a number of minor variants, but the basic idea is the same
Turbulence kinetic energy equation
∂k
∂t + ∇ · (ūk) − ∇ · (νeff ∇k) = G − ϵ
where
1
2
G = νt 2 ∇ū + ∇ūT
Dissipation of turbulence kinetic energy equation
∂ϵ 2
∂t + ∇ · (ūϵ) − ∇ · (νeff ∇ϵ) = C1 G kϵ − C2 ϵk

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 25 / 41
 Eddy Viscosity Models
Standard k − ϵ Model
Once we know k and ϵ, things can be calculated
Turbulent viscosity:
2
νt = Cµ kϵ
Reynolds stress τ R :
τ R = νt 21 ∇ū + ∇ūT
 

Model constants are tuned to canonical flows. Which?

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 26 / 41
 k − ϵ model
Backward Facing Step

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 27 / 41
 Issues with Reynolds averaging concept
In spite of a huge effort to develop RANS based turbulence models, a
general-purpose model suitable for a wide range of practical
applications has so far proved to be elusive
This is to a large extent attributable to differences in the behaviour of
large and small eddies
The smaller eddies are nearly isotropic and have a universal behaviour
(for turbulent flows at sufficiently high Reynolds numbers at least)
On the other hand, the larger eddies, which interact with and extract
energy from the mean flow, are more anisotropic and their behaviour
is dictated by the geometry of the problem domain, the boundary
conditions and body forces
When Reynolds-averaged equations are used the collective behaviour
of all eddies must be described by a single turbulence model, but the
problem dependence of the largest eddies complicates the search for
widely applicable models

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 28 / 41
 Methodology adopted in LES
A different approach to the computation of turbulent flows accepts
that the larger eddies need to be computed for each problem with a
time-dependent simulation
The universal behaviour of the smaller eddies, on the other hand,
should hopefully be easier to capture with a compact model. This is
the essence of the large eddy simulation (LES) approach to the
numerical treatment of turbulence
Instead of time-averaging, LES uses a spatial filtering operation to
separate the larger and smaller eddies. The method starts with the
selection of a filtering function and a certain cutoff width with the
aim of resolving in an unsteady flow computation all those eddies
with a length scale greater than the cutoff width
In the next step the spatial filtering operation is performed on the
time-dependent flow equations. During spatial filtering information
relating to the smaller, filtered-out turbulent eddies is destroyed

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 29 / 41
 Methodology in LES

This interaction between the larger, resolved eddies and the smaller
unresolved ones, gives rise to sub-grid-scale stresses (SGS) stresses
Their effect on the resolved flow must be described by means of an
SGS model
If the finite volume method is used the time-dependent, space-filtered
flow equations are solved on a grid of control volumes along with the
SGS model of the unresolved stresses
This yields the mean flow and all turbulent eddies at scales larger
than the cutoff width

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 30 / 41
 Spatial filtering of equations
Filters are familiar separation devices in electronics and process
applications that are designed to split an input into a desirable,
retained part and an undesirable, rejected part
The details of the design of a filter – in particular its functional form
and the cutoff width ∆ determine precisely what is retained and
rejected
In LES we define a spatial filtering operation by means of a filter
function G (x, x′ , ∆) as follows:
Z ∞ Z ∞ Z ∞
ϕ(x, t) = G (x, x′ , ∆)ϕ(x′ , t)dx1′ dx2′ dx3′
−∞ −∞ −∞

where ϕ(x, t) = filtered function

ϕ(x, t) =original (unfiltered) function
∆ = filter cutoff width

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 31 / 41
 Spatial filtering of equations

Note that overline now indicates spatial filtering, not time-averaging

as was the case with RANS equations
Notice that filtering is an integration, just like time-averaging in the
development of the RANS equations, only in the LES the integration
is not carried out in time but in three-dimensional space
It should be noted that filtering is a linear operation
The common filtering function in three-dimensional LES are as
follows:
Top hat or Box filter
(
1
if |x − x′ | ≤ ∆/2
G (x, x′ , ∆) = ∆3
0 if |x − x′ | > ∆/2

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 32 / 41
 Spatial filtering of equations

Gaussian filter
 
 γ 3/2 |x−x′ |2
−γ
G (x, x′ , ∆) = e ∆2

π∆2
where γ = 6
Spectral cutoff filter

3
Y sin[(xi − xi′ )/∆]
G (x, x′ , ∆) =
(xi − xi′ )
i=1

The top-hat filter is used in finite volume implementations of LES.

The Gaussian and spectral cutoff filters are preferred in the research
literature

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 33 / 41
 Spatial filtering of equations

The Gaussian filter was introduced for LES in finite difference context
by the Stanford group, which, over a period of more than three
decades, has been the centre of research on LES and has established
a rigorous basis for the technique as a turbulence modelling tool
Spectral methods (i.e. Fourier series to describe the flow variables)
are also used in turbulence research, and the spectral filter gives a
sharp cutoff in the energy spectrum at a wavelength of ∆/π
This is attractive from the point of view of separation of the large and
small eddy scales, but the spectral method cannot be used for general
purpose CFD

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 34 / 41
 Spatial filtering of equations

The cutoff width is intended as an indicative measure of the size of

eddies that are retained in the computations and the eddies that are
rejected
In principle, we can choose the cutoff width ∆ to be any size, but in
CFD computations with the finite volume method it is pointless to
select a cutoff width that is smaller than the grid size
In this type of computation only a single nodal value of each flow
variable is retained on each grid cell, so all finer detail is lost anyway
The most common selection is to take the cutoff width to be of the
same order as the grid size
In three-dimensional computations with grid cells of different length
∆x, width ∆y and height ∆z, the cutoff width is often taken to be
the cube root of the grid cell volume ∆ = (∆x ∆y ∆z)1/3

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 35 / 41
 Filtered unsteady Navier-Stokes equations

The unsteady Navier-Stokes equations for a fluid with constant

viscosity are
∂ρ
∂t + ∇ · ρ⃗v =0
▶
∂ρu
▶
∂t + ∇ · ρu⃗ v = − ∂p
∂x + µ∇ · ∇u + ρgx
∂ρv ∂p
▶
∂t + ∇ · ρv ⃗
v = − ∂y + µ∇ · ∇v + ρgy
▶ ∂ρw + ∇ · ρw ⃗ v = − ∂p
∂t ∂z + µ∇ · ∇w + ρgz
Let us use the same filtering function G (x, x′ ) throughout the
computational domain, i.e., G is independent of position x
If we use such a uniform filter function we can, by exploiting the
linearity of the filtering operation, swap the order of the filtering and
differentiation with respect to time, as well as the order of filtering
and differentiation with respect to space coordinates
Similar commutative property was used when the time-averaged
RANS flow equations were derived

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 36 / 41
 Filtered unsteady Navier-Stokes equations

If we apply the filtering operation to the Navier-Stokes equations, we

get
∂ρ
∂t + ∇ · ρ⃗v =0
▶

▶ ∂ρu + ∇ · ρu⃗ v = − ∂p
∂t ∂x + µ∇ · ∇u + ρgx
∂ρv ∂p
▶
∂t + ∇ · ρv ⃗
v = − ∂y + µ∇ · ∇v + ρgy
▶ ∂ρw + ∇ · ρw ⃗ v = − ∂p
∂t ∂z + µ∇ · ∇w + ρgz
The issue now is that we do not have functional relationships for
terms like u⃗v , v ⃗v , w ⃗v
In order to proceed further, we use the following trick
∇ · ρϕ⃗v = ∇ · ϕ ⃗v + (∇ · ρϕ⃗v − ∇ · ϕ ⃗v )
| {z }
Modelled with Sub Grid Scale model
where ϕ is any of the velocity components u, v , w

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 37 / 41
 Filtered unsteady Navier-Stokes equations
We get after the modification
∂ρu
▶
∂t + ∇ · ρu⃗v = − ∂p
∂x + µ∇ · ∇u − (∇ · ρu⃗
v − ∇ · ρu⃗v ) + ρgx
∂ρv ∂p
▶
∂t + ∇ · ρv ⃗v = − ∂y + µ∇ · ∇v − (∇ · ρv ⃗v − ∇ · ρv ⃗v ) + ρgy
∂ρw
▶
∂t + ∇ · ρw ⃗v = − ∂p
∂z + µ∇ · ∇w − (∇ · ρw v − ∇ · ρw v ) + ρgz
⃗ ⃗

The underlined terms on the RHS are caused by the filtering

operation just like the Reynolds stresses in the RANS equations

These can be considered as a divergence of a set of stresses τijS called

the sub grid scale stresses

The name ‘stress’ stems from the way in which it is treated rather
than its physical nature

It is in fact the large scale momentum flux caused by the action of

the small or unresolved scales. The name ‘subgrid scale’ is also
somewhat of a misnomer
Shyam Sunder Assistant Professor MechanicalSimulating
Engineering
Turbulent Flows: Issues and Challenges 38 / 41
 Filtered unsteady Navier-Stokes equations

We finally have the tensor form τ S = ρ⃗v ⃗v − ρ⃗v ⃗v

In order to proceed further, the velocity variable is written as a
filtered part plus the fluctuations: ⃗v = ⃗v + ⃗v ′
Putting back in the sub grid scale stress expression we get
τ S = ρ(⃗v + ⃗v ′ )(⃗v + ⃗v ′ ) − ρ⃗v ⃗v
= ρ(⃗v ⃗v + ⃗v ⃗v ′ + ⃗v ′ ⃗v + ⃗v ′ ⃗v ′ ) − ρ⃗v ⃗v
Noting that the filtering operation is a linear one, we get
τ S = ρ⃗v ⃗v + ρ⃗v ⃗v ′ + ρ⃗v ′ ⃗v + ρ⃗v ′ ⃗v ′ − ρ⃗v ⃗v
Rearranging, we get:
τ S = ρ⃗v ⃗v − ρ⃗v ⃗v + ρ⃗v ⃗v ′ + ρ⃗v ′ ⃗v + ρ⃗v ′ ⃗v ′
| {z } | {z } | {z }
Leonard stresses Cross stresses LES Reynolds stresses

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 39 / 41
 Origin of the different sub grid scale stresses

Remember that we have divided the sub grid scale stresses into
Leonard stresses
▶ L = ρ⃗v ⃗v − ρ⃗v ⃗v
▶ The Leonard stresses are due to things occurring at resolved scale
▶ A second filtering operation makes a change to the already filtered
variable, i.e, ϕ ̸= ϕ
Cross stress terms
▶ C = ρ⃗v ⃗v ′ + ρ⃗v ′ ⃗v
▶ The cross stresses are due to interaction between the small eddies and
the resolved flow
LES Reynolds stresses
▶ R = ρ⃗v ′ ⃗v ′
▶ These are caused by the convective momentum transfer solely due to
the interaction of small eddies
▶ These are modelled by a Sub Grid Scale turbulence model

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 40 / 41
 Thank You

Shyam Sunder Assistant Professor MechanicalSimulating

Engineering
Turbulent Flows: Issues and Challenges 41 / 41