J.M.

Burgerscentrum
Research School for Fluid Mechanics

Computational Fluid Dynamics II

Practical exercises
Dr.ir. F.J. Vermolen and Prof.dr.ir. C. Vuik

2015

Delft University of Technology

Faculty of Electrical Engineering, Mathematics and Computer
Science
Delft Institute of Applied Mathematics
 Practical exercises CFD II 2

Introduction
In this note we treat some exercises that are meant as illustration for the
theory treated in the course CFD 2. The exercises can be seen as two
separate parts.
Part 1 (exercises 1-13) is meant to illustrate the use of finite elements in the
flows. Besides that, implicitly some linear algebra aspects will be treated.
Part 1 is executed with the aid of the finite element package SEPRAN. Part
1 is planned for Monday and Wednesday till Friday.

part 2 (exercises 14-15) deals only with linear algebra aspects. This part is
planned for Tuesday and must be executed with the aid of MATLAB.

The exercises are planned for several days. Of course one is free to finish the
exercises of the previous day. The exercises are ordered in increasing degree
of difficulty and it is not necessary to make all of them. Exercises marked
with a star (*) are only meant for those who have time left. In general they
are more difficult than the other ones.

The following extra manuals are meant for the exercises:

A short introduction of the login procedure and some information con-

cerning the linux system will be given at the first lab afternoon. Where
we refer to editor and use edit, the editor of the linux system is
meant.
You may use for example kwrite or you may open your favorite editor
by pressing the left icon on the control panel and choose utilities
and later on editors.

The SEPRAN manuals can all be found at the following Internet ad-
dress:

http://ta.twi.tudelft.nl/nw/users/vuik/burgers/burg.html

press Practical work to find the manuals.

Alternatively they can also be found on the local system in the following
way:

sepman intro shows the introduction

 Practical exercises CFD II 3

sepman um shows the users manual

sepman sp shows the manual standard problems
sepman exams shows the examples manual

Schedule
Monday: Exercises 1, 2 and 3
Tuesday: Exercises 14 and 15
Wednesday: Exercises 4 and 5
Thursday: Exercises 6 and 7
Friday: Exercises 8, 9 and 16
The exercises 10-13 are meant for those that have time left. These exercises
require much more effort and have more an investigating character.

All these exercises have been preprogrammed. In your directory you will find
2 subdirectories: sepran and matlab. In sepran you will find the sepran files,
in Matlab the Matlab files. It is wise to create a new directory, for example
workdir in which you run the sepran exercises and update the files. Make a
copy of the files corresponding to a specific exercise from the sepran directory
to this new directory. In this way the original files are kept undisturbed and
may be reused. For example in case of Exercise 1:
mkdir workdir
cp sepran/exercise1c.msh workdir
cp sepran/exercise1d.prb workdir
cd workdir
After that you may run SEPRAN.

In case of Exercises 14 and 15 you should proceed as follows. After login in

go to the main window and give the following commands:
cd matlab
matlab
After that the exercises 14 and 15 can be made. Matlab may be left by the
command quit. Logging out is done by the command exit in the main
window.
 Practical exercises CFD II 4

Exercise 1 heat conduction in a plate with a

hole
This exercise is meant to solve a simple temperature problem in order to get
acquainted with SEPRAN as well as the notions mesh generation, computa-
tional part and postprocessing. Besides that, the notion of boundary element
is introduced.
In this exercise we consider a square plate with a hole like sketched in Fig-
ure 1. In this plate we want to solve a stationary heat conduction problem,
h

b0

h/2 h
0

0 b/2 b

Figure 1: Square plate with hole

which is described by the two-dimensional Poisson equation

divkT = . (1)

T(x,y) denotes the unknown temperature , k the heat conduction coefficient

and the production of heat per unit of volume and time.
In order to solve this problem uniquely, it is necessary to prescribe boundary
conditions on the whole boundary. Types of boundary conditions that can
be applied are for example:

Temperature T prescribed at the wall.

Heat flux prescribed at the wall.

qn = 0: insulated wall
 Practical exercises CFD II 5

qn = q0 : prescribed heat flux

with qn normal component of the heat flux q = kT .

Heat transfer relation of Newton

qn = (Tw T )
with the heat transfer coefficient,
Tw , T : the temperature on the wall and far away from the wall re-
spectively.

We assume that in the case of the plate in Figure 1, the left and right-hand
wall are kept at a constant temperature T = 0. The lower wall is kept at
temperature T = 1. The heat production is assumed to be constant.
On the boundary of the hole we assume that qn = 0 (insulated) or that
qn = q0 (constant heat flux).
In first instance we assume that the upper wall is insulated (qn =0) and in
second instance we use the heat transfer relation T + kT n = 0.

Exercises:

1a Verify that the temperature with the prescribed boundary conditions

is symmetric with respect to the line x = b/2.
What is the correct boundary condition on this line motivated by the
symmetry of the solution?

1b Derive the weak formulation for the heat conduction problem. If you
have a choice, use the most general of the boundary conditions includ-
ing the symmetry boundary conditions.
Why is it not necessary to use the symmetry boundary condition ex-
plicitly in the finite element method?

1c In order to solve the temperature equation, half the region is subdivided

into elements , where the symmetry axis is part of the boundary. Sub-
divide the region in triangular elements with the aid of the SEPRAN
grid generator SEPMESH. Use coarse with unit = 0.05 and as sub-grid
generator GENERAL We choose the following data:
h=1
b=1
h0 = b0 = 0.4
 Practical exercises CFD II 6

In order to run this exercise we apply the next commands.

mkdir workdir
(workdir is your own work directory,
you may choose the name freely)
cp sepran/exercise1c.msh workdir
cd workdir
sepmesh exercise1c.msh (Make the mesh)
sepview sepplot.001 (View the mesh)

You can inspect the mesh file with the editor:

kwrite exercise1c.msh

sepmesh creates a file meshoutput, which will be used by the subsequent

programs.
Make a sketch of the boundary, using the input file and the plot shown
by sepview. Indicate the user points and the curves in this sketch
(reverse engineering).

1d Compute the temperature and the heat flux vector q by SEPCOMP

using the following data:
k=1
=1
The hole and the upper walls are insulated.

The procedure is as follows:

cp ../sepran/exercise1d.prb .
sepcomp exercise1d.prb (Solve problem)

You can inspect the input file with the editor:

kwrite exercise1d.prb

Program sepcomp solves the problem prints the temperature and plot
the isotherms and make a vector plot of q.
You can inspect the plots by sepview:
 Practical exercises CFD II 7

sepview sepplot.001 (View the plots)

Interpret the results. Are the boundary conditions satisfied?

1e Perform the same computations, with a coarseness of 0.025.
1f Instead of an insulated upper wall we use the mixed boundary condition
T + kT n = 0, with = 500.
Repeat the computations.

In this case we need to define more boundaries than in the first case.
Therefore two new files are needed:

exercise1f.msh
exercise1f.prb

First run sepmesh and then sepcomp with these files and view the re-
sults.

sepmesh exercise1f.msh
sepcomp exercise1f.prb
sepview sepplot.001

Explain the boundary condition at the upper wall. Is this boundary

condition satisfied?
1g Repeat the computations of Exercise 1f, now with T=0 on the right
wall and a heat flux qn = 1 through the hole. Try some combinations
of physical data yourself. Also try some other configurations.

Also in this case the mesh is adapted since more boundaries are needed
to define the boundary conditions. These new files are:

exercise1g.msh
exercise1g.prb

Explain the results.

After finishing the computations all files can be removed from your work
directory,
 Practical exercises CFD II 8

Exercise 2 Memory usage and computational

time for direct methods
The goal of this exercise is the study the memory usage and the computa-
tional time of a direct method as function of the number of elements and the
numbering used.

For that reason we solve the Poisson equation on the rectangle (0,5) (0,1).

T = 1. (2)

As boundary conditions we choose T = 0 on the whole boundary.

Exercises:

2a Generate a mesh by sepmesh with 250 elements in horizontal direc-

tion and 50 elements in vertical direction. Use the submesh generator
SQUARE. Use bi-linear rectangular elements.

The input file for sepmesh is

exercise2.msh

2b Solve the Poisson equation with a direct method and the matrix stored
as symmetrical matrix. Inspect the number of entries of the matrix and
the computation time required for building the equations and solving
the linear system of equations. In order to measure the memory usage
and the computational time we put two extra lines in the input file:

set output on
set time on

The adapted input file is called exercise2b.prb

These lines generate a large quantity of extra output. The number of

entries in the large matrix is given by the line:

the large matrix contains .... real entries.

 Practical exercises CFD II 9

The computational time is given (accumulated) in the lines:

time is ... seconds in subroutine ....

In order to measure the computational time in the matrix building we

subtract the computational time before the call of build from the one
after the call of build. It concerns the lines

time is ... seconds in subroutine buildbf

The line

time is ... seconds in subroutine solvelbf

gives the time after the linear solver.

Carry out the same exercise, but now with the matrix stored as non-
symmetrical matrix.

2c In exercise 2b we implicitly used the fact that SEPRAN renumbers the

nodes. To investigate the effect of the renumbering, we switch of the
renumbering. This is done by putting the lines 1

start
renumber not
end

in the input file for sepcomp before the part

problem
.
.
end

The adapted input file is called

exercise2c.prb
1
if you put a # before renumber not, the command is ignored
 Practical exercises CFD II 10

Repeat exercise 2b.

Try to estimate the memory usage if a band method and a horizontal
is chosen. Do the same for a vertical numbering. Compare these with
the printed values. Note that SEPRAN uses a profile method.

2d In order to find a relation between the computational time and memory

usage as function of the number of points we subdivide the region again
with SEPMESH, however, now with 20 x 20, 40 x 40, 80 x 80 and 160
x 160 elements. Repeat exercise 2b for these subdivisions.
 Practical exercises CFD II 11

Exercise 3* Writing a user element

Remark: this exercise is meant for those students that want to write an ele-
ment subroutine themselves. It is not part of the standard exercises.

The goal of this exercise is practicing the writing of your own standard el-
ement. For that reason we reuse exercise 1, however, in this case we use
element type 1 for the inner region and element types 2 and 3 respectively
for the boundary elements. Due to this choice it is necessary to add your
own subroutine ELEM to program SEPCOMP.

Exercises:

3a Derive the Galerkin equations corresponding to the weak formulation

of 1b. Compute the element matrix and element vector using linear
triangles and the also the linear boundary element. Assume that the
triangle may have an arbitrary shape and also that boundary elements
may be positioned anywhere in the domain.

3b Program subroutine ELEM using the results of 3a. Use the users man-
ual Section 4.2. Repeat the exercises 1c, 1d and 1f using your own
element.
 Practical exercises CFD II 12

Exercise 4 Application of iterative methods

The goal of this exercise is to solve a simple three-dimensional problem and
to compare the effect of direct solvers with the effect of iterative solvers.

For that reason we compute the temperature in a straight pipe. The lower
wall and the upper wall of the pipe consist of parallel circles with radius
1. The outer wall is perpendicular to these circles and has height 5. The
temperature in the pipe satisfies the Laplace equation

divkT = 0. (3)

The pipe is insulated on the upper wall and on the under wall we have a
constant temperature T0 . On the outer wall we assume the heat transfer
relation: qn = (Tw T )
To solve this problem, we can use the axi-symmetry or we can solve the
problem as a three-dimensional one.
First we carry out the axi-symmetrical approach.

Exercises:

4a Choose k = 1 and = 0.5. Solve the heat conduction problem with 20

elements in the r direction and 100 elements in the z direction. Record
the number of nodes in the mesh and the number of elements of the
matrix as well as the computation time required to build the matrix
and to solve the system of equations .
The corresponding files are:

exercise4a.msh
exercise4a.prb

4b Repeat the exercise with 40 100 and 80 200 elements.

4c Now repeat the exercises 4a and 4b with an iterative solver (CG) and
an accuracy of 103 . Compare the results.
The corresponding file is:

exercise4c.prb
 Practical exercises CFD II 13

4d We will now repeat the exercise with a real three dimensional mesh.
Although this is senseless for this exercise, in practice many problems
are purely three-dimensional.
Subdivide the lower surface into triangular elements by the submesh
generator GENERAL. Use 20 elements along the circle.
The upper face must be constructed by translation of the lower face.
The side wall must be generated by the generator PIPESURFACE. Use
25 elements in the z-direction.
To generate the 3D grid the volume generator PIPE is used.
Carry out the same exercises as in 4a with a direct solver.
The corresponding files are:

exercise4d.msh
exercise4d.prb

4e Repeat 4d with an iterative solver.

The corresponding file is:

exercise4e.prb

4f We refine the mesh by taking 40 elements along the circle instead of

20. repeat exercises 4a and 4e.

Since this exercise may take a lot of computation time it is advised to

do it only if there is time left.
 Practical exercises CFD II 14

Exercise 5 heat transfer and temperature in a

Poisseuille flow
The goal of this exercise is to study the influence of outstream boundary con-
ditions on the convection-diffusion equation as well as to study the influence
of the streamline Petrov Galerkin upwinding.

Consider the flow in a pipe as sketched in Figure 2. We suppose that the flow

T1

0
T0
uz z

ur r

Figure 2: Pipe with Poisseuille flow

is a stationary Poisseuille flow. The heat dissipation is neglected and there

are no other heat sources. The temperature equation is considered station-
ary. Changes of the velocity due to temperature differences are neglected.

We consider the flow in a straight, cylindrical tube with radius R. For z = 0

the walls have a temperature T0 and for z > 0 a temperature T1 . We consider
the region 0 z l and assume that the liquid on z = l has a homogeneous
temperature. Since flow and temperature are axi-symmetrical it is sufficient
to solve the problem in a cross-section with axi-symmetrical coordinates.
Define the inlet length lin as the minimal length for which temperature var-
ious at most 1% through the cross section. For z lin , it can be shown that

lin = 0.6 PRe ,

Umax R
Pe = a
,
 Practical exercises CFD II 15

with P e the Peclet number , Umax the maximum velocity and a the so-called
temperature adjustment coefficient.

The convection-diffusion equation for the temperature can be written in di-

mensionless form as:
1
T + u T = 0. (4)
Pe
In our example we use the next data:

T0 = 0

T1 = 1

ur = 0, uz = 1 - r 2

R = 1 P e = 700

On the lower boundary we have T = T0 , on the side wall T = T1 . On the

symmetry axis we have of course the symmetry boundary condition Tn
= 0.
From these data it follows that the inlet length is equal to 420.
We shall now compute the temperature for various choices of the outflow
boundary and various boundary conditions on the outflow boundary.

Exercises:

5a Choose the outflow boundary on z = 500 and choose T = 1 as boundary

condition on z = 500. Subdivide the region into 4 elements in r direction
and 10 elements in z direction. Use the submesh generator quadrilateral
and rectangular elements. This solution will be the reference for the
other parts. Make a contour plot (or a colored contour plot) of the
temperature .
The corresponding files are:
exercise5a.msh
exercise5a.prb
5b Choose the outflow boundary on z = 250 and use the same bound-
ary conditions. Repeat exercise 5a. The temperature will show non-
physical oscillations. Explain these oscillations.
The corresponding file is:
 Practical exercises CFD II 16

exercise5b.msh

Is it possible to remove the oscillations by refining the mesh?

Investigate this in practice.

5c In order to suppress the oscillations, we apply the SUPG method in-

stead of the standard Galerkin method. Now repeat exercise 5b.
Compare the results.
The corresponding file is:

exercise5c.prb

Repeat the exercise for finer grids.

T
5d Repeat part 5b with SGA, using the outflow boundary condition n
= 0.
Explain the results.
The corresponding file is:

exercise5d.prb
 Practical exercises CFD II 17

Exercise 6 The rotating cone problem

In this exercise we shall solve the rotating cone problem from the lecture
notes, however, with a coarser grid.
!
y
Consider the region in Figure 3. We consider a rotating velocity field u = .
x

0.5

-0.5
-0.5 0 0.5

Figure 3: Definition of region for the rotating cone problem

The velocity field starts on the line B, rotates counter clockwise and ends on
the line B. We convect a cosine shaped concentration by the velocity field.
The concentration c satisfies the convection-diffusion equation

c + u c = 0. (5)

On the outer boundary we choose c = 0 as boundary condition. On the

inflow boundary B we choose c = cos(2(y + 41 )) as boundary condition.
We suppose that there is a gap of thickness 0 between inflow boundary and
outflow boundary. In other words we choose two different curves for in-
flow boundary and outflow boundary. On the outflow boundary we choose
c
n
= 0 as boundary condition. We shall solve this problem both with SGA
and SUPG.

Exercises:
 Practical exercises CFD II 18

6a Make a mesh consisting of triangles by the submesh generator QUADRI-

LATERAL. Use 20 elements along each outer side.
The corresponding file is:

exercise6a.msh

6b Choose = 106 and solve the convection-diffusion equation with

SGA. Use a direct solver for the linear system. Make a contour plot
and a 3D plot of the concentration.
The corresponding file is:

exercise6b.prb

6c Repeat 6b, however, now with SUPG.

The corresponding file is:

exercise6c.prb

6d Repeat 6b and 6c, but use a mesh consisting of quadrilaterals

The corresponding file is:

exercise6d.msh
 Practical exercises CFD II 19

Exercise 7 The use of iterative methods for the

convection-diffusion equation
In this exercise we use the convection-diffusion equation as described in the
SEPRAN manual Examples, Section 3.1.
Consider the example of the curved pipe with given concentration on the
inflow boundary as described in the manual Examples Section 3.1.2. Using
this example we shall investigate the behavior of iterative methods.

First we start with the Poisson equation, in which the discretization matrix
is symmetric and positive definite. This is done by using poisson as type in
the problem input block.

7a Solve the system with a direct method

matrix_structure: symmetric # matrix is symmetric

Record the memory usage and the computation time used. Plot the
computed solution by sepview. Repeat this by adding the following
input block before end_of_sepran_input

solve
positive_definite
end

to the file.
The corresponding files are:

exercise7.msh
exercise7a.prb

7b Now we shall perform some experiments with the CG method. Use

matrix_structure: storage_scheme = compact, symmetric

and
 Practical exercises CFD II 20

solve
iteration=method = cg, print_level = 1, preconditioner = ilu
end

Read the memory usage, the number of iterations and the computa-
tion time required. Choose the following preconditioners: diagonal,
eisenstat, ilu and mod eisenstat. If time is left vary the number of dis-
cretization points.
The corresponding file is:

exercise7b.prb

7c We now consider the effect of renumbering on the iterative method.

For the CG method without preconditioning the number of iterations
must be independent of the ordering. Check this by using the next
orderings:

start
renumber not
end

After that the Cuthill profile and Sloan profile renumbering:

start
renumber cuthill profile
end

or

start
renumber sloan profile
end

In case of a preconditioned CG method the ordering will have some

influence.
In this case the mesh generator has already renumbered the nodes in an
almost optimal way, so the effect of further renumbering is very small.
Choose the mod eisenstat preconditioning and use all previous reorder-
ing strategies.
 Practical exercises CFD II 21

7d Next we solve the original convection diffusion. Use central differences

and compute the solution with a direct method. Plot the solution
with sepview. Next solve the problem iteratively with the ilu precon-
ditioning and Bi-CGSTAB, CGS, IDR, and GMRES. Choose for the
iteration_method choose bicgstab, cgs, idr, and gmres.
The corresponding files are:

exercise7d.prb

In the case of a direct method we do not define the matrix storage,

but use the default one, which implies a profile storage scheme for an
asymmetric matrix.
In case of the iterative methods we must use a compact storage scheme:

matrix_structure: storage_scheme = compact

and add an extra block:

solve
iteration_method = bicgstab, print_level = 1, preconditioner = ilu
end

7e Repeat these exercises with upwind discretization. Compare the so-

lution and the computational time required with that of the previous
part.
The corresponding file is:

exercise7e.prb

The difference between central and upwind scheme appears to be very

small. However, make the diffusion parameter 10 times smaller
(0.0005 instead of 0.005) and compare the results of both schemes.
Also try a decrease with a factor 100.

7f From now on we use upwind discretization. Solve the system with the
Bi-CGSTAB method without preconditioning. Record also the begin
and end residual. Repeat this with eisenstat, ilu and mod eisenstat
preconditionings. Conclusion?
 Practical exercises CFD II 22

7g Use ilu as preconditioning and Bi-CGSTAB, CGS, IDR, and GMRES

as iterative methods. Compare the results.

7h If time is left repeat the exercises with various diffusion parameters.

 Practical exercises CFD II 23

Exercise 8 Development of a flow in a straight

pipe
The goal of this exercise is to give an introduction in the solution of the
stationary Navier-Stokes equations by the finite element method.

Consider the flow between two flat plates. At the inflow we prescribe a con-
stant velocity. After some time the flow will have been converted to a fully
developed flow . In this exercise we shall simulate this numerically. Consider
the configuration as sketched in Figure 4. Due to symmetry it is sufficient
y

h/2

0 x
l

Figure 4: Region of definition for exercise 8

to consider a half channel only. The liquid satisfies the Navier-Stokes equa-
tions. At the inflow we prescribe a constant velocity field parallel to the flat
plates. On the fixed walls we have a no-slip condition. At the outflow we
assume that the flow is parallel to the plates. Furthermore we suppose that
the pressure is equal to 0.

Exercises:

8a Make a mesh consisting of quadratic triangles by the submesh generator

QUADRILATERAL. Use 8 elements in x direction and 3 in y direction.
Choose the height h equal to 1 and the length l equal to 10.
The corresponding file is:

exercise8.msh
 Practical exercises CFD II 24

8b Solve the Stokes equations by the grid created in 8a. Use as inflow
velocity u = 1, a density = 1 and a viscosity = 0.01. What is
the corresponding Reynolds number based on the distance h? Choose
= 106 as penalty parameter, i.e. = 106 .
Make a vector plot of the velocity and a streamline plot.
The corresponding files are:

exercise8b.prb

8c Investigate the dependence by comparing the solution for a number

of values of . Use values between 10 and 1012 .

8d Repeat exercise 8b, using the Navier-Stokes equations instead of Stokes.

Note that in case we have to solve a non-linear problem. The iteration
process is explicitly written in the structure block.
The corresponding file is:

exercise8d.prb

Why is the difference between both solutions so small?

 Practical exercises CFD II 25

Exercise 9 Flow over a backward facing step

In this exercise we shall compute the flow over a backward facing step, as
described in the lecture notes. The effect of the outflow boundary conditions
is studied. Consider the backward facing step as sketched in figure 5. Define

H
l

Figure 5: Region of definition backward facing step

the Reynolds number by Re = Umax Hh

. At inflow we assume a fully devel-
oped velocity profile. At outflow you may choose between several types of
outflow boundary conditions. At the walls there is a noslip condition.

Exercises:

9a Create a mesh by the submesh generator QUADRILATERAL consist-

ing of quadratic triangles. Subdivide for that reason the region into
two rectangles and apply QUADRILATERAL to each of the quadri-
laterals. Increase the length of the elements in the direction of the
outflow boundary. Do not use too many elements since otherwise the
computational time increases too much. It is a good practice to start
with a coarse grid and refine later if necessary.
Use the next data:

H=1
 Practical exercises CFD II 26

h = 0.5

l=6

L = 38
The corresponding file is:

exercise9.msh

9b Solve the Stokes equations for Re = 50. Check if the solution satisfies
the boundary conditions.
The corresponding files are:

exercise9b.prb

9c Solve the Navier-Stokes equations for Re = 50. Make a stream line plot
and a vector plot of the velocity. You can also make a contour plot of
the pressure.
The corresponding file is:

exercise9c.prb

9d Solve the Navier-Stokes equations for Re = 150. If the iteration process

does not converge try to construct a method to reach convergence.

9e Solve the Navier-Stokes equations for Re = 150. Choose L = 12 and try

both the two different types of outflow boundary conditions mentioned
in the lecture notes. What is your conclusion?
If time left, repeat these exercises with a finer mesh.
 Practical exercises CFD II 27

Exercise 10 Flow through a bend

In this exercise we consider the flow through a bend. We restrict ourselves
to a two-dimensional Cartesian configuration. The region of definition is
sketched in Figure 6. Compute and plot velocity and pressure in this bend

Figure 6: Region of definition bend problem

for various Reynolds numbers, where the Reynolds number is based on the
width of the channel. For Re = 500 the flow must be laminar and stationary.
The corresponding files are:

exercise10.msh
exercise10.prb

Repeat the exercise for smaller mesh sizes.

 Practical exercises CFD II 28

Exercise 11 Natural convection in a square cav-

ity
In this exercise we consider the flow in a square cavity. The left and right
walls are kept at a constant temperature. The left-hand wall gets the tem-
perature T=1, the right-hand side wall: T=0. The lower wall and upper wall
are assumed to be insulated. Compute and plot velocity, temperature and
pressure in this cavity for various Rayleigh numbers in the range 103 to 106 .
Try both the coupled and the decoupled approach.
The corresponding files are:

exercise11.msh
exercise11a.prb

exercise11b.prb

Repeat the exercise for smaller mesh sizes.

 Practical exercises CFD II 29

Exercise 12* Heat exchanger

This exercise is meant for those that have time left.
In this exercise we consider the temperature distribution in a heat exchanger
as sketched in Figure 7. It concerns two concentric pipes. Only the cross

Figure 7: heat exchanger

section for r > 0 has been drawn. In the inner pipe the liquid is warm in the
outer pipe it is cold. The outer temperature is assumed to be constant.

In first instance we assume that in both pipes we have a parabolic velocity

profile and that the temperature distribution satisfies the convection-diffusion
equation.

Formulate the differential equations and the boundary conditions. Devise

a method to model the heat transfer between inner and outer pipe, where
the temperature is different in both pipes. Use a number of values of the
parameters.
Consider also the case that the liquid in the outer pipe has the opposite
direction of that in the inner pipe.
Finally you can investigate if the velocity in the pipes is influenced by the
temperature . For that purpose you have to solve the Boussinesq equations.
Keep in mind that SEPRAN can only be used to solve laminar flows.
There are no example files available. Take an existing file and adapt it.
 Practical exercises CFD II 30

Exercise 13* Flow around a cylinder

This exercise is meant for those that have time left.
An application of a time-dependent flow, is the flow around a cylinder. it
is known that for Re > 50 we can expect so-called von Karmann eddies.
The Reynolds number is defined by the diameter of the cylinder. These von
Karmann eddies can only arise if the flow is activated somewhere, for example
by disturbing the symmetry in the velocity field.

Try to generate a von Karmann eddy by solving the time-dependent Navier-

Stokes equations.
See the manual SEPRAN examples for some help.
 Practical exercises CFD II 31

Exercise 14
Experiments with the Gaussian elimination method
In this exercise we consider various properties of the Gaussian elimination
method. The function [a,f] = poisson(n1,n2,u1,u2,disc) is used to
construct a matrix, which originates from a discretization of the following
equation:
2
c 2c c c
( x 2 + y 2
) + u1 x + u2 y = g where x, y (0, 1) (0, 1),

c(x, y) = 0 for x {0, 1} or y {0, 1}.

The number of point in the x- and y-direction is denoted by n1, n2 respec-

tively. The value disc = central gives a central discretization, whereas
disc = upwind leads to an upwind discretization. The resulting matrix is
a and the right-hand-side vector is f . To get more information of the Matlab
programs one can use the help-function. Try for instance help poisson.

14a First the fill-in is investigated for small matrices. Therefore the com-
mand [l,u,p] = gauss(a); can be used. This function computes the
LU-decomposition of a with pivoting such that pa = lu. Furthermore
the number of nonzero elements and the bandwidth in l and u are given.
Finally the nonzero structures of the matrices are shown.

Execute the following commands in Matlab:

[a,f] = poisson(3,10,0,0,central);

[l,u,p] = gauss(a);

In the subroutine gauss the Matlab subroutine lu is called. Repeat this

exercise with [a,f] = poisson(10,3,0,0,central); and compare
both examples.

Related theory: part II, Section 1.2, and 1.7

Exercise aims:

The matrices L and U are less sparse than A.

 Practical exercises CFD II 32

Numbering of the unknowns is important. Numbering such that

n1 n2 is better than n1> n2.

14b In this exercise we take u1 6= 0, or u2 6= 0 and use central differences.

It appears that for some choices partial pivoting is used.

Take the following choices:

[a,f] = poisson(5,5,0,0,central);
[l,u,p] = gauss(a);

[a,f] = poisson(5,5,100,0,central);
[l,u,p] = gauss(a);

[a,f] = poisson(5,5,0,100,central);
[l,u,p] = gauss(a);

Try also some other values. Warning: for large matrices the CPU-time
can increase considerably (ctrl c kills a job).

Related theory: part II, Section 1.2, 1.5, and 1.7

Exercise aims:

Pivoting destroys the non-zero structure of the matrix.

The amount of fill-in is comparable to that without pivoting.

14c Do the same exercises as before for upwind differences. Compare the
results.

Related theory: part II, Section 1.2, 1.5, and 1.7

Exercise aims:

The choice of discretization influences the Gaussian elimination

algorithm.
No pivoting is necessary if AT is diagonal dominant.

14d In this exercise matrices are considered where the amount of fill-in can
be very large. The matrices are not related to the Poisson equation.
The matrices used here are known as arrowhead matrices. This type
of matrix occurs in Domain Decomposition methods. Call uparrow
 Practical exercises CFD II 33

and substitute 25 for the dimension. Execute thereafter [l,u,p] =

gauss(a);. Repeat the exercise for the downarrow matrix and compare
the results.

Related theory: part II, Section 1.2, and 1.8

Exercise aims:

Pivoting (or renumbering) can be used to minimize fill-in.

14e In the following part we investigate the solution of the discretized Pois-
son equation. Execute [a,f] = poisson(10,10,0,0,central);.
The LU-decomposition of the matrix a can be obtained by:

[l,u,p] = lu(a);

Compute with this LU-decomposition the solution of the linear system

ax = f . The solution of ly = f can be done by:

y = l\f;

To check if you have computed the correct solution compute:

norm(a*x-f)

The resulting solution x can be visualized using the call:

plotsol(10,10,x)

Related theory: part II, Section 1.2, and 1.3

Exercise aims:

Compute the solution of a linear system when the LU-decomposition

is available.
Visualization of the solution of the Poisson equation.

14f Repeat the exercise for

[a,f] = poisson(10,10,200,0,central);
 Practical exercises CFD II 34

Compute again norm(a*x-f). Its value should be less than 1010 . Fi-
nally the same problem can be solved using upwind differences. Com-
pare both solutions.

Related theory: part II, Section 1.2, and 1.5

Exercise aims:

Compute the solution of a linear system when the LU-decomposition

with pivoting is available.
A comparison of the approximations using central and upwind
differences.

14g Finally some experiments are done with iterative improvement. First
the matrix and right-hand side are formed by [a,f] = poisson(5,5,0,0,central);
Compute the LU-decomposition with

[l,u,p] = lu(a);

It is not easy to work with single precision numbers in Matlab. To

simulate iterative improvement we disturb the matrix l as follows:

l = random(l,10^(-5));

This function changes the lower triangular matrix l with random num-
bers of order less than 105 . Compute norm(a-l*u) to check that a 6=
lu (the matrix u can also be changed). Write a function iterna(a,l,u,p,f)
which uses the disturbed matrix l (and possible the disturbed matrix
u). First construct a termination criterion and make the function such
that it stops after 100 iterations.

Related theory: part II, Section 1.2, and 1.5

Exercise aims:

Get experience with the iterative improvement algorithm.

 Practical exercises CFD II 35

Exercise 15
Experiments with iterative methods
In this exercise we consider various iterative methods to solve linear systems.
We start with systems where the coefficient matrix is symmetric and positive
definite. These systems are solved by basic iterative methods and the Con-
jugate Gradient method. Finally we investigate if these methods can also be
applied to non-symmetric systems.

15a Construct the matrix a and the right-hand-side vector f using the call
[a,f] = poisson(10,11,0,0,central); Compute the solution of
the linear system with the Gauss-Jacobi method:

x = jacobi(a,f,10^-10);

Write down the number of iterations for comparison with other meth-
ods.

Related theory: part II, Section 2.2

Exercise aims:

Gauss-Jacobi is a slowly converging method.

A linear converging iteration method.

15b First try to modify the program jacobi.m to obtain the Gauss-Seidel
method. To check your implementation you also use:

x = seidel(a,f,10^-10);

Compare the results with those obtained in exercise 15a.

Related theory: part II, Section 2.2

Exercise aims:

Gauss-Seidel converges two times as fast as Gauss-Jacobi.

15c In this exercise we use the SOR method. The estimate of the optimal
is not straightforward. Therefore we do some experiments with various
choices of . Type
 Practical exercises CFD II 36

x = sor(a,f,10^-10,omega);

where omega is a real number between 0 and 2. Try to approximate

the optimal value. Using the call

x = sor(a,f,10^-10,0);

an approximation of the optimal is calculated from theory. Compare

the results.

Related theory: part II, Section 2.2

Exercise aims:

SOR converges much faster than the previous methods.

The convergence behavior can be non-linear.
There are values of which leads to somewhat less iterations than
the optimal calculated from theory.

15d In this part we consider the Conjugate Gradient method.

x = cg(a,f,10^-10);

Compare the results with the results obtained with the basic iterative
methods.

Related theory: part II, Section 3.2 and 3.3

Exercise aims:

CG converges very fast.

It is not necessary to estimate an optimal parameter.
The convergence behavior is super linear.

15e In the theory used to analyze CG three quantities are used: kx xi k2 ,

kx xi kA , and kb Axi k2 . These quantities are plotted by the call

x = cganal(a,f,10^-10);
 Practical exercises CFD II 37

Compute the condition of a by the Matlab command cond(a) and

estimate the number of required CG iterations. Compare this with
your experiments.

Related theory: part II, Section 3.2 and 3.3

Exercise aims:

All quantities decrease.

The number of required estimations obtained from the theory is
an upper bound.

15f There are matrices where the theory for the CG method no longer
holds, due to rounding errors. One of them is a discretization of the
bending beam equation:
d4 c
= f.
dx4
The matrix can be constructed by beam; Take for the dimension 40
(thereafter also try 80 and 160). Solve the system by

x = cganal(a,f,10^-10);

and observe the results.

Related theory: part II, Section 3.3

Exercise aims:

kx xi kA forms a monotone decreasing sequence.

kb Axi k2 increases at some iterations.
CG has not been converged when the number of iterations is equal
to the dimension of a.

15g In the final exercises we investigate if the methods given in Chapter 2

and 3 can be used when the matrices are non-symmetric. Construct the
matrix as follows: [a,f] = poisson(5,5,0.1,0,central); and
apply Gauss-Jacobi, Gauss-Seidel, SOR, and CG.

Exercise aims:

All methods converge although the theory for CG is not valid.

 Practical exercises CFD II 38

15h Construct the matrix as follows: [a,f] = poisson(5,5,1,0,central);

and apply Gauss-Jacobi, Gauss-Seidel, SOR, and CG.

Exercise aims:

The basic iterative methods converge but CG does no longer con-

verge.

15i Construct the matrix as follows: [a,f] = poisson(5,5,100,0,central);

and apply Gauss-Jacobi, Gauss-Seidel, SOR, and CG.

Exercise aims:

The iterative methods are divergent.

15j Construct the matrix as follows: [a,f] = poisson(5,5,100,0,upwind);

and apply Gauss-Jacobi, Gauss-Seidel, SOR, and CG.

Exercise aims:

The basic iterative methods converge.

The choice of the discretization influences the convergence behav-
ior of an iterative method.
 Practical exercises CFD II 39

Exercise 16 Pressure in the earths surface

In this exercise, we are investigating the convergence behavior of the ICCG
process for problems with layers with large contrasts in the coefficients. For
that reason we simplify the equation considerably and assume that we have
to solve the stationary linearized 2D diffusion equation, in a layered region:

div(p) = 0 , (6)

with p the excess pressure and the permeability. At the earths surface the
excess pressure is prescribed. When the pressure field is required in some
reservoir it is not practical to calculate the pressure in every position of the
earths crust. Therefore the domain of interest is restricted artificially. We
assume that the lowest layer is bounded by an impermeable layer, so there
is no flux through this boundary. The artificial vertical boundaries are taken
at a sealing fault, or far away from the reservoir. Again a zero flux condition
is a reasonable assumption at these boundaries.For our model problem we
assume that in sandstone is equal to 1 and in shale is equal to 107 .
Furthermore the Dirichlet boundary condition at the earths surface is set
equal to 1. The solution of equation (6) with these boundary conditions is of
course p = 1, but if we start with p = 0 or a random vector, our linear solver
will not notice the difference with a real problem. Numerical experiments
show that the choice of one of these start vectors has only marginal effects
on the convergence behavior. An advantage of this problem is that the exact
error can easily be calculated. We consider this problem on a rectangular
domain with 7 straight layers. In each layer 10 elements in the horizontal
and 5 elements in the vertical direction are used.

Exercises:

Before starting this exercise copy the files exercise16* from the sepran direc-
tory to your working directory.

16a We start by solving this problem without preconditioning and stop the
CG iterations if the residual has been reduced by a factor accuracy =
106 . Give the following commands:

sepmesh exercise16.msh
 Practical exercises CFD II 40

sepcomp exercise16a.prb
sepview sepplot.001

Note that the residual is small but the solution is not close to the exact
solution p = 1. Repeat the computation with accuracy = 107 and
accuracy = 108 . Inspect the solution and measure the number of
iterations.

16b We now include IC preconditioning. Therefor you can use the file

exercise16b.prb

Again inspect the solution and compare the number of iterations.

16c Make a table with the number of iterations where the k shale in exer-
cise16b.prb is chosen equal to 102, 104 , and 107 .

16d Repeat exercise c with deflation. The deflation choices are specified in

exercise16d.prb

16e Finally, use the file exercise16d.prb and chose k shale equal to 107 and
vary accuracy: 102 , 104 , 106 , and 108 . Compare the number of
iterations and the quality of the solution.