Sei sulla pagina 1di 8

A LAGRE EDDY SIMULATION (LES) CODE FOR

TURBULENT FLOWS:
VALIDATION AGAINST A TURBULENT FLOW AROUND
A SINGLE SURFACE-MOUNTED CUBICLE OBSTACLE
EXPERIMENT.
Timothy Ganesan
Masters of Science in Mechanical Engineering
Universiti Technologi PETRONAS
E-mail: tim.ganesan@gmail.com

The mass conservation laws state that mass is either


created or destroyed during the flow (or in the process)
and thus, the net differential of the mass towards time is
equal to zero.

AbstractThe unique and smooth analytical solution of


the Navier-Stokes equations thus far does not exist.
However, for various flow regimes, the numerical solution
can be obtained. The study of obtaining the solution of the
Navier-Stokes
equation
numerically
is
called
Computational Fluid Dynamics (CFD). Therefore, the
object of this research is to develop an in-house computer
code in Visual C++ using the Finite Volume Method to
compute numerically the solutions of the Navier-Stokes
equation under turbulent conditions. The computer code
consists of a steady three-dimensional turbulent solver
using the Large Eddy Simulation (LES) turbulent model.
The velocity flow profile obtained from the computer code
is then compared with experimental results. The
experimental results consist of velocity profiles for a
turbulent flow around a single surface-mounted cubicle
obstacle.

The mass flow rate across the faces of the fluid element
is the product of density, area and the velocity
component normal to the face. As a frame of reference
let the inflow into the control volume is a positive term
and the outflow is vice versa. Thus, the summation will
be in the form of a differential which will result as the
following:

( u ) ( v ) ( w)

0
t
x
y
z

where u, v
and w are the velocity components in the x, y and z
direction respectively.

Index Termscomputational fluid dynamics, in-house


computer code, large eddy simulation, and finite volume
method

I.

(1)

Conservation of Momentum
In the momentum conservation principles, the concept of
Newtons Second law of motion plays an important role
in the derivations involved. Forces acting on fluid
particles are referred into two categories that are the
surface forces and the body forces. The conservation
principles are required in order to form a partial
differential equation (conservation equation). The
solution to this partial differential equation would
describe the behavior of the momentum in the flow.

INTRODUCTION

In the study of CFD, the numerical solution of the


Navier-Stokes equations describes the approximate
behavior of the fluid flow. The governing laws are
required to be modified using the Finite Volume Method
to prepare the equations to be in computational form. In
using the Finite Volume Method, the numerical scheme
applied is the first order UPWIND[1] scheme. The
general solution method used is the Semi-Implicit
Method for Pressure Linked Equations Revised
(SIMPLER) [1]. Since, it is required to solve the linear
algebraic equations numerically; a matrix solver
component (in three-dimensions) was also written using
the Tri-Diagonal Matrix method.

The final form of the momentum equation will be as the


following:

u
div uv div gradu Su
t

Mass Conservation
1

(2)

whereby the u t is the unsteady term,


div uv is the convective term, div gradu is
the diffusive term and Su is the source term. The in
the diffusion term is the diffusion coefficient which is a
function of the viscosity.
II.

of equations with the direct method, the allocated


memory is generally large and the rate of convergence is
oscillatory. Thus, in CFD applications the iterative
method is highly favored. The number of iterations using
the iterative solver will be generally large, however it
will eventually lead to convergence that is to say
convergence of the solution of the linear equations is
somewhat guaranteed.

NUMERICAL SCHEME

When the continuity equation is solved numerically, a


series of linear equations for each cell are obtained. The
SIMPLER (Semi-Implicit Pressure Linked Equations
Revised) algorithm is a guess-correct method. According
to the SIMPLER algorithm, first some value of pressure
is guessed. Then the numerical solution of the pressure
correction equation is required to compute the real
pressure and velocity terms (since it is coupled). Then
the real pressure terms will be used to correct the
guessed pressure terms iteratively until the real pressure
terms converge. Prior to that, the upwind numerical
scheme is used to compute the coefficients which are the
function of the convective and the diffusive terms. These
coefficients are used to solve the linear equations that
arise in the pressure correction continuity equation which
is as the following:

Guessed initial value

Buffer function
Back substitute and reiterate
Compute the source
term

a I , J pI , J a I 1, J pI 1, J a I 1, J pI 1, J a I , J 1 pI , J 1

Compute the Thomas


algorithm coefficients

a I , J 1 pI , J 1 bI, J
(3)
If error

where, a I , J is the coefficients of the linear equation,


pI , J is the pressure correction and bI , J is the source
term .

>

1%
Perform the Forward
Elimination Procedure

The upwind differencing scheme is suitable for a wide


range of Peclet numbers greater than two where the
central differencing scheme fails to provide results that
are stable. To determine the computations at each cell,
the upwind or donor cell differencing scheme that
factors in the flow direction is used. In the upwind
differencing scheme, the flow can be taken to be from
west to east, north to south and up to down or vice versa
in the same order. In the current numerical scheme, the
flow from west to east that is in the positive direction is
utilized.

Perform
the
Back
Substitution Procedure and
obtain

Compute
unknown

Then, the remaining unknowns are that of the scalar


property which can be computed by solving the linear
algebra equations.

Solve

the

If error < 1%
Figure 1. Matrix Solver Algorithm

III.

THE TRI-DIAGONAL MATRIX ALGORITHM

where * is the guessed scalar term, is the


corrected scalar term and is the real scalar term. As
can be seen in Figure 1, the linear equations arising as a
solution to the scalar continuity equation is solved
numerically using the Thomas algorithm. The seminal
mathematical operations involved in this algorithm are
the Forward Elimination procedure and the Back-

Numerical solutions of matrices generally fall into two


techniques that are the direct method and the iterative
method. The Thomas algorithm is a derivative of the
Gauss-Seidel method. In CFD applications, the sets of
linear equations conventionally generated are large, that
is to say that the number of the vertical and the
horizontal entries are myriad. Hence, to solve such a set
2

nb

*
vnb

Substitution procedure. Thus, the matrix solver algorithm


is used to compute the pressure correction terms and
other scalar terms.

Navier-Stokes equation and the filter function (Gaussian


filter, Fourier filter or the Tophat filter) producing the
spatial-averaged Navier-Stokes equation. Hence, the
convolution averages-off the small eddies (which effects
are estimated using semi-empirical models called subgrid scale (SGS) models) whereby computation is only

For the three-dimensional case some modifications to the


algorithm are necessary. In three-dimensions the sweep
direction is maintained (that is for this case the direction
of the flow field which is west to east). The difference in
the three-dimensional case and the two-dimensional case
is instead of line-by-line computation, a phase-by-phase
computation is used.

START
IV.

SOLUTION METHOD

Initial guess

The SIMPLER (Semi-Implicit Pressure Linked


Equations Revised) algorithm [3] is a guess-and-correct
method for the calculation of pressure and pressure
which was explained previously.

Step 1: Solve Discretised


Equations

The discretised momentum equations for two dimensions


are as the following:
ai , J u

*
i, J

anbu

*
nb

(p

*
I 1, J

v * ,u *

Step 2: Solve Pressure Correction

p p* , v v* ,

p I , J ) Ai , J bi , J

Equations

u u*

Step 3: Correct pressure and


velocities

(4)

aI , j v* I , j anb vnb* ( pI*,J 1 p I , J ) AI , j bI , j


*

p , v, u , *

(5)

Step 3: Solve all other descretised


transport equations

*
whereby the a nb u nb
are the neighbor nodes can
be computed by using the upwind numerical scheme, the
pressure, p at this initial stage are guessed, the A term
is the surface area of the grid at the given points
(referring to the index value) and the b is the source
term. Where the p , v and u values are the correction
values and the p * , v * and u * are the guessed values
and the p, v and u are the real values. Figure 2 gives
overview of the simplified SIMPLER algorithm.

NO

convergence?
YES

STOP!

Figure 2. SIMPLER algorithm

V.

LARGE EDDY SIMULATION (LES)

done on the large eddies. Thus, the LES model gives a


spatial-filtered Navier-Stokes equations such as the
following:

The CFD computer code uses the Large Eddy Simulation


(LES). Large eddies depend and extract energy from the
bulk flow, their behavior is more anisotropic than small
eddies (which are isotropic at high Reynolds number
flows) [6]. Therefore, the flow characteristics are very
dependent on the behavior of these eddies at high
Reynolds numbers. Models such as the Reynolds
Averaged Navier-Stokes (RANS) equations also describe
the solution that comprises of the behavior of eddies in a
time-averaged fashion. However, the LES model
provides a means to compute a certain range of eddy
scales (large eddies) through a statistical procedure
called filtering (spatialaveraging). The filtering
procedure is done by performing a convolution on the

ij
ui
p
div ui u j i div gradui
t
xi
x j

(6)

ij u i u j u i u j
(7)

where u i are the filtered velocities, p is the filtered

strain, S ij of the resolved flow, which can be


formulated as the following:

pressure, the term ij is the stress tensor that results


from
the statistical procedure (analogous to the
Reynolds stresses in the RANS equations). These
stresses are considered to arise from the interactions
between the unresolved small scale eddies or better
known as the sub-grid scale (SGS) stresses. The SGS
stresses are decomposed (Leonard Decomposition [7])
into three groups of stresses that are the Leonard stresses
( Lij ), the Cross stresses ( Cij ) and the LES Reynolds
stresses ( Rij ) which can be formulated as the

S ij 0.5(

Rij SGS S ij

(8)

Lij ( ui u j ui u j )

The stresses Cij , Lij and Rij can be used together in


the finite volume method [4] and thus resulting in one
SGS turbulence model which can be formulated as the
following:

(9)

Rij uiu j

(11)

1
ij 2 SGS S ij ii ij
3

The Leonard stresses are caused by the effects at the


resolved scales. Besides, they also came into being due
to the second filtering operation that generates a
difference to a filtered flow variable. The Cross stresses
arise due to the property interactions between the SGS
eddies and the resolved flow. The LES Reynolds stresses
arise due to the convective momentum transport which is
a result of the interactions among SGS eddies per se and
are modeled with a SGS turbulence model. The SGS
turbulence model used in this computer code is the
Smagorinsky-Lilly SGS model.

1
Rii ij
3

where SGS is the artificial or the sub-grid scale


viscosity which acts as the constant of proportionality
and has the units Pa.s. The second term which contains
the Kronecker delta function which acts to make sure
that the sum of the modeled normal stresses are
equivalent to the total kinetic energy of the SGS eddies.

ij ( ui u j ui u j ) ( ui u j uiu j ) uiu j

(10)

(12)

(13)

following:

Cij ( ui u j uiu j )

u j
ui

)
x j
xi

(14)
This model builds on the mixing length model. The size
of the SGS eddies are determined by the filter choice as
well as the filter cut-off width which is used during the
averaging operation. The SGS viscosity can be obtained
by the following semi-empirical formulation:
SGS C SGS 2

2 S ij S ij

where is the filter cut-off width and C SGS is an


empirical constant which is usually specified in a range,
0.19 C SGS 0.24 [5].

18

160

80

480

(15)

xFlow direction
z
1000
1600
*dimensions are in ,milimeters

LABEL

Figure 3. Geometry of the test section.

This models main axiom is that the LES Reynolds


stresses are taken to be proportional to the local rate of

*dimensions are in millimeters

VI.

RESULTS AND DISCUSSION

TABLE 2: Properties initialized in the numerical simulation

Experimental Setup
The geometry of the test section in the experiments
performed by Martinuzzi and Tropea [8], [9] is as Figure
3. the fluid type in this experiment is water (with density
1.0 10 3 kg / m 3
=
and
viscosity
=
1.0 10 3 kg / m.s ). The cubicle obstacle is placed

in the symmetrical line with reference to the z-direction


of the test section at a distance of 0.8m (five times the
channel height of the test section). Besides that, the
width (z-direction) to channel height (y-direction) ratio
must be greater than 6 (which in this study is set to 6.25).
As can be seen in Figure 3, the flow is of the x-direction.
The Reynolds number which is based on the channel
height must lie within the range of 8 10 4 to

Properties

Values

Number of iterations
for velocity and
pressure

500

Convergence criterion
for velocity correction

50%

Relaxation factor for


velocity

0.9 m/s

Relaxation factor for


pressure

0.9 m/s

Convergence criterion
for pressure correction

1.2 10 4 . In this study the Reynolds number of


1.2 10 4 is used and based on the channel height, the
fluid velocity at the inlet would be 0.467 m / s . The
flowsconsideredinthisexperimentwereinvestigatedby
means of static pressure measurements [8], laser light
sheet, and oilfilm and crystal violet visualization
techniques.

1%

Comparison of Numerical and Experimental Results

Initial Conditions for the Numerical Simulation

The velocity field obtained from the numerical


simulation is then compared with the velocity field
obtained from the experiment. The u-velocity is the
velocity in the x-direction relative to the y- direction.

The boundary conditions and the geometry of the


experimental setup is initialized in the numerical
simulation. The data for the initial conditions are as the
following:

Four data sets of the u-velocity field are obtained


throughout the test section. The first data set is taken at
the flow development region five obstacle heights before
the obstacle.
Two data sets are taken on the obstacle, that is directly
above the obstacle and finally one data set is taken after
the obstacle at the eddy dissipation region (which is two
and a half obstacle heights from the obstacle coherent
with the experimental data sets). The first data set (line 1
in Figure 3) is the comparison of the u-velocities
obtained at five obstacle heights before the obstacle, and
one obstacle height towards the z-direction from the
centre of the obstacle or with reference coordinate
system (x = 400mm, z = 580mm, in Figure 3).

TABLE 1: Properties initialized in the numerical simulation

Properties

Values

Meshsize(symmetric)

0.02m

Inlet u-velocity

0.467 m/s

Inlet v-velocity

0 m/s

Inlet w-velocity

0 m/s

Inlet pressure

101.11 kPa

Fluid density

1000kg/m3

Fluid viscosity

0.001kg/m.s

Filter cut-off width

0.025m

SGS Constant, C SGS

0.2

Table 1 provides the values of the SGS constant and the


average strain rate, and the filter cut-off width. The
numerical parameters which are also required to be
initialized are as the following:

Figure 5. Normalized u-velocity versus normalized distance y/H at line


1.

Figure 8. Error (%) obtained from the comparison of the numerical


results relative to the experimental results of the u-velocity profile
versus normalized distance y/H at line 2.
Figure 6. Error (%) obtained from the comparison of the numerical
results relative to the experimental results of the u-velocity profile
versus normalized distance y/H at line 1.

It can be observed that the normalized u-velocity profile


in the second (Figure 7) and third (Figure 9) data set
obtained by the numerical simulation is very rigid as
compared to the experimental data (which is oscillates)
especially at certain ranges (approximately at 0.8 to
1.4m). This is because at those positions that is along the
top side of the obstacle is where the flow separation
occurs. In this regions, due to recirculation, eddies are
generated. The effect of these eddies causes the flow to

It can be observed in Figure 6 that the maximum velocity


obtained numerically is higher than the experimental
values. The flow profile obtained from the numerical
simulation is oscillatory at some points as compared to
the experimental results. There error range through out
the channel height can be observed as also oscillatory.
The possible reason for this sort of behavior is that the
convergence criterion set for the velocity is too low (that
is 50 %). Therefore the velocity fails to stabilize due to
reaching convergence at a very early stage of the
iteration. To increase the stability of the values of the
velocity the convergence criterion should be more
stringent so that more iteration can be performed. The
average error in the first data set is 15.6%.
The second (line 2 in Figure 6) and the third (line 3 in
Figure 6) data set is the comparison of the normalized uvelocities obtained along and through the obstacle at x =
840mm and z =480mm and 820mm and z=480mm.

Figure 9. Normalized u-velocity versus normalized distance y/H at line


3.

Figure 7. Normalized u-velocity versus normalized distance y/H at line


2.

Figure 10. Error (%) obtained from the comparison of the numerical

results relative to the experimental results of the u-velocity profile


versus normalized distance y/H at line 3.

The normalized velocity profile in Figure 11 on the


fourth data set describes the flow after the obstacle that
during eddy dissipation. It can be seen that the
experimental results show some fluctuations but it
vanishes nearing the wall regions which is similar as the
numerical simulation results. The only difference is that
the numerical simulation results maintain its rigid
behavior and fails to capture the finer properties of the
flow. The average error in the fourth data set is 13.54%.

oscillate. There are several factors that can cause the


numerical simulation to not capture these flow behaviors.
Firstly, the numerical simulation has a very coarse mesh
as compared to the experimental data. Since the filter
cut-off width has to be larger than the mesh size, hence
the simulation is unable to capture the finer flow
behavior. The method to encounter this difficulty is by
defining a finer mesh size as well as a smaller filter cutoff width.
Another factor that can contribute to this deviation in
results, is the numerical scheme used in this simulation.
The numerical scheme used in this simulation is the first
order UPWIND. This numerical scheme is the
extrapolation technique used during the process of
descritising the flow equations where the Taylor
expansion is truncated at the first order and hence limits
accurate computation. Thus, the current numerical
scheme would need to be upgraded to the second order
UPWIND scheme to increase the accuracy of the flow
computations. The average errors in the second and third
data sets are 8.783% and 8.73% respectively.
The fourth data set (line 4 in Figure 6) is the comparison
of the normalized u-velocities obtained after the obstacle
at x/H = 2.5, z/H =1.00 (x = 1000mm and z = 580mm).

Figure 12. The standard deviation (%) obtained from the comparison of
the numerical results relative to the experimental results of the uvelocity profile versus normalized distance x/H for all four data sets .

In Figure 13 although the average error seems to be very


low at second and third data sets, the range of errors
from the mean is higher in those data sets. The standard
deviation for the second and third data sets is higher than
the first and second data sets. This shows that the second
and third data sets contain the highest range of error
fluctuation from the mean error. This emphasizes the fact
that there various flow oscillations that could not be
captured by the numerical simulation.
VII.

CONCLUSION

The CFD computer simulation research project is at the


development stage. Although the three-dimensional
turbulent solver using the LES turbulence model source
codes are complete, further advancement and validation
with empirical results with different conditions are
necessary.
Further development of the numerical simulation is
necessary, for instance the upgrading of the numerical
scheme, the refining of the meshing, testing the flow
properties with different filter cut-off width as well as
testing it with a different SGS models such as the
dynamic SGS models.

Figure 11. Normalized u-velocity versus normalized distance y/H at


line 4.

VIII.

REFERENCES

[1] H.K. Versteeg, W. Malalasekera, An Introduction to


Computational Fluid Dynamics, Longman Scientific and
Technical (1995), 114-117.

Figure 12. Error (%) obtained from the comparison of the numerical
results relative to the experimental results of the u-velocity profile
versus normalized distance y/H at line 4.

[2] J.H. Ferziger, M.Peric, Computational Methods for Fluid


Dynamics, Springer Publications (2002) 10-80,
[3] S.V. Pantankar, Numerical Heat Transfer and Fluid
Flow, Hemisphere Publishing Corporation, Taylor and
Francis Group, New York, (1980).
[4] R.Peyret, and E.Krause, Advanced Turbulent Flow
Computation, CISM Courses and Lectures No.395,
International Centre for Mechanical Sciences, SpringerVerlag, Vienna, (2000).
[5] J.W.Deardoff, A Numerical Study of Three-dimensional
Turbulent Channel Flow at
Large
Reynolds
Numbers,J.FluidMech., Vol 41, (1970), 453-480.
[6] J.Smagorinsky, General Circulation Experiments with the
Primitive Equations I. TheBasic Experiment, ,Mon.
Weather Rev. Vol 91, No.3, (1963), 99-164.
[7] A.Leonard, Energy
Cascade
in
Large
Eddy
Simulations of Turbulent Fluid Flows, Adv in
Geophysics. A18, (1974), 237-248.
[8] R.Martinuzzi & C.Tropea, The flow around surfacemounted, prismatic obstacles placed in a fully developed
channel flow, Journal of Fluid Engineering, Vol. 115,
(1993), 85.
[9]

R.Martinuzzi. & M.Psud'homme,


Higher-order
correlations for the turbulent flow around a surfacemounted cube placed in a channel, TSF9, Kyoto, Japan,
(1993), 211-1.

Potrebbero piacerti anche