Lecture Notes in

Engineering
Edited by C. A. Brebbia and S. A. Orszag

70

E.K Bruch

The Boundary Element Method

for Groundwater Flow

,---------, Springer-Verlag
Berlin Heidelberg New York
London Paris Tokyo
Hong Kong Barcelona Budapest
Series Editors
C. A. Brebbia . S. A. Orszag

Consulting Editors
J. Argyris . K.-J. Bathe· A. S. Cakmak' J. Connor' R. McCrory
C. S. Desai, K. -Po Holz . F. A. Leckie· G. Pinder· A. R. S. Pont
J. H. Seinfeld . P. Silvester· P. Spanos' W. Wunderlich· S. Yip

Author
Dr. Erwin Karl Bruch
Max Strasse 5
B-4721 Neu-Moresnet
Belgien

ISBN-13: 978-3-540-54407-4 e-ISBN-13:978-3-642-84577-2

001: 10.1007/978-3-642-84577-2

This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation,
broadcasting, reproduction on microfilms or in other ways, and storage in data banks. Duplication
of this publication or parts thereof is only permitted under the provisions of the German Copyright
Law of September 9, 1965, in its current version, and a copyright fee must always be paid.
Violations fall under the prosecution act of the German Copyright Law.
© Springer-Verlag Berlin, Heidelberg 1991

The use of registered names, trademarks, etc. in this publication does not imply, even in the
absence of a specific statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general use.
Typesetting: Camera ready by author

61/3020-543210 Printed on acid-free paper.

 Preface

During the last few years rapid progress has been made on the solution of partial
differential equations, which are now routinely solved using computers. Their
implementation is based numerically in three methods which in chronological order are
the finite difference method, the finite element method and the boundary element
method. In comparison to the two other methods, the later presents some important
advantages which have motivated its rapid development in the last decade or so.
In this book the application of the boundary element method to the solution of the
Laplace equation is examined. This equation is of fundamental importance in
engineering and science as it describes different types of phenomena, including the
groundwater flow applications highlighted in this book. Special subjects such as
numerical integration, subdivision of the domain into regions and other computational
aspects are discussed in detail in the first chapters. To demonstrate the accuracy and
efficiency of the boundary element method, results obtained when solving the Laplace
equation have been compared against known analytical solutions. Other chapters deal
with problems such as steady and unsteady flow in addition to infiltration problems.
The applications demonstrate that the boundary element method provides a powerful
solution technique which can be effectively applied to solve this type of problem.
 Table of Contents

Chapter 1 Porous Media Flows. 4

1.1 Introduction . 4
1.2 Application of Continuum Mechanics to Porous Media 4
1.3 Equation Describing the Porous Media Flow 5
1.4 The Boundary Conditions 8
Chapter 2 The Laplace Equation 13

2.1 Introduction 13
2.2 The Mathematical Model. 13
2.3 Mathematical Aspects of the Laplace Equation . 14
2.4 Solution of the Laplace Equation 16
Chapter 3 Introduction to the Boundary Element Method . 17
3.1 Introduction 17
3.2 Governing Equation . . 17
3.3 The Basic Equation of the Boundary Element Method . 18
3.4 The Weighting Function. 21
3.5 Analysis of the Integrals . 22
3.6 Discretization of the Problem 25
3.7 Special Problems . 26
3.8 Conclusion 27
Chapter 4 Numerical Integration 28
4.1 Introduction 28
4.2 Regular Integration 28
4.3 Singular Integration of Geij 31
4.4 Singular Integration ofHeij 38
4.5 Conclusions. . . . 45
Chapter 5 Numerical Aspects of the Boundary Element Method . 46
5.1 Introduction 46
5.2 Comparison of the Integration Methods. 46
5.3 Study of the Discretization . . . . . . 48
5.4 Sensitivity of the Results to the Variations of Geometry 50
5.5 Conclusions 51
 3

Chapter 6 Division of the Domain into Sub-regions 52

6.1 Introduction •••• 52

6.2 Discretization of the Integrals 52
6.3 The Linear System • . 54
6.4 Earlier Solutions . •••• 54
6.5 Analysis of the Numerical Problem • 55
6.6 The Additional Equation . 60
6.7 SUlIllllIlIY of the Theory . 61
6.8 Applications. . 62
6.9 Special Interfaces. 66
6.10 Conclusions. . 70
Chapter 7 Steady State Flow Through Porous Media • 71

7.1 Introduction • . . • • • 71
7.2 Iterative Detennination of the Free Surface Position 71
7.3 Applications • • 75
7.4 Conclusions 84

Chapter 8 Unsteady Flow Through Porous Media. 85

8.1 Introduction.. . 85
8.2 Iterative Detennination of the Free Surface Position 85
8.3 Applications . . . 91
8.4 Choice of the Discretization . · 102
8.5 Conclusions. . . • · 104
Chapter 9 Infiltration Problems . · 106
9.1 Introduction .106
9.2 Previous Solutions .106
9.3 Iterative Detennination of the Free Surface Position · 107
9.4 Computational Examples. · 109
9.5 Conclusions • · 113
Chapter 10 Conclusions · 114
Chapter 11 References 116
Acknowledgments. . 120
 CHAPTER 1
POROUS MEDIA FLOWS

1.1. INTRODUCTION

The study of groundwater and porous media flow has become important during the last years in
fields as different as civil engineering, aquifer managing and petroleum engineering.
The fundamental equations describing these physical phenomena are well known [5], [6] and
before the computers allowed them to be solved numerically, many authors developed analytical
solutions for special cases [57], [66]. Nowadays, analytical solutions are very important
because they allow to check the numerical methods. Different numerical methods have been
developed to solve numerically porous media flow problems. The finite difference method
(FDM), the finite element method (FEM) and the boundary element method (BEM) have
successfully been applied to solve the equations. The application of these three methods to the
computation of groundwater flow has shown that the BEM is the most suitable, since it allows
the most efficient handling of free surfaces, infinite flow domains and singularities. Besides,
the BEM gives more accurate results than the two other methods because the weighting
functions used by this method are analytical solutions of the governing equations. In this book,
the application of the BEM to the different types of groundwater flows, especially to those
described by the Laplace equation, will be discussed.

1.2. APPLICATION OF CONTINUUM MECHANICS TO POROUS MEDIA

In this study, only saturated porous media will be considered and the capillary pressure will be
neglected. In this chapter, the partial differential equations describing the porous media flow
are described. These equations have different parameters depending on the characteristics of the
porous media. Whereas in a partial differential equation, these parameters refer to the
characteristics of the porous media at one point, a certain volume around the point is needed to
determine them. Indeed, the porous media is made up of grains and voids and its intrinsic
properties are defined as mean values on a certain volume. Thus, a porous media is
fundamentally discontinuous because the grains are considered to be impervious in comparison
with the voids where the fluid flows.
The continuum concept is a very powerful way to analyse a large range of physical phenomena.
In order to introduce the properties of the porous medium into the mathematical model, one has
to choose between a deterministic [5], [6] or a stochastic approach [51], [31]. In the
framework of the present study, the first approach only is applied which is the one generally
applied to practical problems.
Thus, the variations of the parameters of the fluid and the porous medium are described in a
representative elementary volume (REV) around each porous media point and this define the
properties there by mean values on this REV. To apply this method, the REV must be small
enough so that its dimensions are negligible with regard to those of the domain under
consideration and subsequently, the REV is considered as a material point in space. On the
other hand, the REV must be large enough so that the different parameters may be defined by
averages in the representative volume.
 5

1.3. EQUATIONS DESCRIBING '!HE POROUS MEDIA FLOW

1.3.1. Introduction
In this paragraph, the continuity equation and the momentum equation that are the two basic
equations describing the porous media flow will be established. After that, these two equations
will be combined in order to obtain the Laplace equation that will give the mathematical model
of the porous media flow. Throughout this paragraph, some assumptions will be made, which
will impose limitations to the types of problems represented by the model. This topic has
already been expounded in detail in many books [57], [66], [5], [37], [6], and consequently.
only the main points of the theory will be discussed here.

1.3.2. The Continuity Equation

For a fluid flowing through a saturated porous media, the continuity equation is given by the
equation below. This formula expresses that in the REV and during a given time interval. the
quantity of stored fluid is equal to the variation of the volume of the voids, ie.

a(np) .
- - = '-div (p q) (1.1)
at
with: n: porosity of the porous medium;
p: density of the fluid;
q: velocity vector.
Notice that the above velocity is not the real velocity of the fluid in the porous medium. To
obtain the real velocity, one has to consider that the fluid occupies only a part n (porosity) of the
REV and that the fluid follows a very tortuous route between the grains. In the relation (1.1).
the variations of n and p take into account the compressibility of the porous medium and of the
fluid due to the variation of the flow conditions and the boundary conditions. In order to
simplify (1.1). one applies the two following assumptions:
- the density of the fluid depends only on the pressure p. ie.

p = pep) (1.2)

- the total stress (0) resulting from the water pressure (p) and the overburden loads remains
constant at each point. Thus, if 0' is the effective stre!;s supported by the porous medium
matrix, one obtains that the total stress (0) does not vary with the flow conditions, ie.

o=o'+p (1.3)
do' =-dp (1.4)

The introduction of these two assumptions into equation (1.1) gives:

ap .
p [ a (1 - n) + J3n] dt = - div (pq) (1.5)

with: a: compressibility coefficient of the solid matrix;

13: compressibility coefficient of the fluid.
In this book, only the flow of water through a porous media will be considered. As the
compressibility of water is very small (13 = 4,829.10- 10 m2/N), the density of water is nearly
 6

constant for a large range of pressures. Furthermore, for the major part of the applications that
will be considered, the influence of the compressibility of the porous medium on the flow
conditions is negligible. Thus, the left hand side terms of (1.5) are generally negligible with
regard to the right hand side terms of the same equation and hence, the following continuity
equation is obtained.

div q =0 (1.6)

1.3.3. The Darcy law

Like other types of fluid flow, porous media flows are in principle described by the Navier-
Stokes equations. However in the present case, these equations cannot be applied directly
because the boundary conditions that they require (the geometry of the grains) are unknown. fu
this case, the momentum equations can be obtained by expressing an equilibrium equation
between the different forces acting on the fluid.
These forces are:
- the internal pressure of the fluid;
- the gravity;
- the inertia forces;
- the viscous resistance force that the porous media oppose to the motion of the fluid.
The equilibrium of these forces gives-the following equation.

dq
dt + q grad q = - g grad u + F (1.7)

where: q, velocity vector,

g, gravity;
u = pip + Y(p), velocity potential;
F, viscous resistance force per unit mass of the porous media;
Yep), height of the point P, measured from a reference level.
The type of flow (laminar or turbulent) depends on the Reynolds number, which is given in the
present case by the relation.

Re=qd/V (1.8)

with: q: velocity of the fluid (scalar);

d: representative length of the porous media matrix that is generally taken equal to the
mean diameter of the grains;
v: the kinematic viscosity of the fluid (for water, v = 10-6rn2/s).

In practice, the Reynolds number in these cases is generally lower than 1 and so, the flow is
laminar [5]. If so, one obtains:
- that the inertia forces in equation (1.7) are negligible [5] with regard to the right hand
side terms of (1.7);
- that the viscous resistance force F is proportional to the flow velocity, with K the
hydraulic conductivity tensor.
Consequently, the velocity vector can be expressed as:

q =KF/g (1.9)

Thus, in case of laminar flow conditions, the momentum equation (1.7) gives the classical
Darcy law, ie.

q =K grad (u) (1.10)

 7

1.3.4. Hydraulic Conductivity and Penneability

In equation (1.10), the components of the hydraulic conductivity tensor K are not only
dependent on the characteristics of the porous medium, but also on those of. the fluid as given
by:
K=(glv)k (1.11)

The components of the permeability tensor k are only dependent on the characteristics of the
porous media. The values of the components of this tensor depend on the associated reference
system. For each tensor exists a principal reference system for which the non diagonal terms of
the tensor are equal to zero, ie.

_- °Ky0)
K-(~ (1.12)

The principal reference system will be used in what follows to define the geometry of the
problem. If the components Kx and Ky of (1.12) are equal, the porous media is isotropic. If,
Kx and Ky are different, the porous media is anisotropic. The components of the tensor K are
determined by experiment. The table 1.1 gives some characteristic values

Porous media log K (ern/s)

gravel . 2 to - 1
sand -lto-3
silt and loess - 3 to -7
clay -7to-11
non cracked rock -8 to-tO
Table 1.1

Several authors have developed formulae relating the penneability k to various properties of the
porous medium matrix, such as the mean diameter of the grains [6].

1.3.5. The Laplace equation

The previous two sections were devoted to the basic equations of the porous media flow, the
continuity equation (1.6) and the Darcy law (1.10).

div q =0 (1.6)
q =-K grad (u) (1.10)

After the elimination of q between these two relations, one obtains the following equations
where the reference system (x,y) corresponds to the principal axes of the tensor K.

(1.13)

In the case of homogeneous porous media, the last two terms of equation (1.13) are equal to
zero and one obtains the well known Laplace equation (1.14) or (1.15), ie.

a2u a2u
KX2+~2 =0 (1.14)
ax ay
 8

K(.1u) = 0 (1.15)

In order to solve the Laplace equation on a given domain, one has to impose conditions on the
boundaries of the domain as will be discussed in section 1.4.

1.3.6. The assumptions

Throughout the different steps taken to obtain the Laplace equation, the following assumptions
have been made:
- the porous media is saturated;
- the fluid is incompressible;
- the inertia forces are negligible;
- the flow is laminar and so, the viscous resistance force is proportional to the flow
velocity;
- the porous media is homogeneous.
In practice, one needs to verify that the flow respects these assumptions. If one of them is not
satisfied, the Laplace equation will not describe the physical phenomenon. In particular, the
Laplace equation is not valid close to the boundaries where the REV is partially located outside
the flow domain. However, this case has generally no significant influence on the flow inside
the domain.
Another case is when the solution of the Laplace equation gives very high, or even infinite
velocities at certain points. In this case, the Laplace equation is not valid near these points
because the flow does not remain laminar and hence, the resistance force is no longer
proportional to the flow velocity. In this case, equation (1.9) must be replaced by a more
elaborated relation [5] . If such problems are present only in some boundary regions, they will
have no significant influence on the flow inside the domain.

1.3.7. Concluding remarks

In this section 3, the governing equation for groundwater flow problems described by the
Laplace equation have been established. It is important always to bear in mind the different
assumptions made to obtain this equation.
To solve the Laplace equation, one needs to impose conditions along the boundary of the
problem. The Laplace equation is independent of time and so, for unsteady flows, the boundary
conditions must take into account the transient character of the flow. This point will be
considered in the next paragraph.

1.4. THE BOUNDARY CONDITIONS

1.4.1. Introduction

To solve the Laplace equation (1.14) on a domain Q, one has to impose conditions concerning
the potential u or its derivative along an outward vector q on the different parts of the boundary
r. Usually, one of the three following conditions on each part of the boundary is imposed:
-) u =Uo ' the Dirichlet condition
-) q =qo ' the Neumann condition;
-) yu + 'JIq + t = 0 , the Cauchy condition with y,'JI, and t some parameters.

1.4.2. The steady state case

The steady porous media flow is described by the Laplace equation (1.14). To solve this
equation in the domain Q, boundary conditions given by (1.18) need to be imposed on the
different parts of the r boundary.
 9

Figure (1.1) shows the classical case of water seeping through a domain with:
AB : the upstream face of the domain;
BC : the free surface;
CD : the seepage surface;
DE : the downstream face of the domain;
EA : the base of the domain.
y

h x
Figure 1.1

To determine the boundary conditions, one can uses the definition of the potential (see (1.7»
and the Darcy law (1.1 0), with p the pressure of the fluid and Yep) the height of the point P
under consideration, measured from a reference level (see figure 1.1). Thus,

u = RIp + Y(p) (1.16)

q = -Ii grad (u) (1.17)

The application of equations (1.16) and (1.17) along the different parts of the boundary r gives
the following boundary conditions.

AB:u=H
BC : q =0 and u =Y
CD:u=Y (1.18)
DE :u=h
EA:q=O

One has two boundary conditions along the free surface Be because its position is unknown. In
the figure 1.1, system (X,Y) is the general reference system used to define the geometry of the
domain and (x,y) is the reference system corresponding to the principal directions of the
permeability tensor .K.

1.4.3. The unsteady case

In the case of unconfmed unsteady groundwater flows, the influence of the compressibility of
either the fluid or the porous medium is negligible with regard to the variation of the position of
the free surface and the flow remains generally laminar. Thus, unconfined unsteady
groundwater flows are described by the Laplace equation that is independent of time and, the
boundary conditions consider the transient character of the flow.

AB u = H (t) DE u= h (t)
CD u=Y M q=O

BC
~ = ec~:~~~ P) (:: + N (n») et u=Y
(1.19)

with : pand y : the two angles showed in the figure 1.2;

 10

£ : effective porosity of the porous media;

N(o) : normal recharge along the free surface.

Figure 1.2 shows the case of unsteady flow though a classical domain.
Y~
\

\1 \
y
~O'
I
'h\\
I
I 1\\ \\
I
,
I \ \
\ \
\
\
\

H(t)

I Y-B

h(t)
x
r
Figure 1.2

Notice that one can use the definition of the potential given by equations (1.16) and (1.17) to
determine the conditions along the different parts of the boundary. The boundary conditions are
obtained by generalizing those of the steady state case equations (1.18) with the exception of the
condition along the free surface BC, ie.
The case of the boundary condition along the free surface will be studied in detail in the
following section [19].

1.4.4. The kinematic condition along the free surface in the unsteady case

Consider a free surface with a point I (figure 1.2) the position of which varies in time along an
oblique line forming an angle "( with the vertical Y, and consequently parallel to y' (figure 1.3).
Along the free surface, one has a recharge N forming an angle (X with the vertical Y. The
components of N are positive along the increasing x' and the decreasing y'. As showed in
figure 1.3, the angles (X and "( are positive and the angle f3 is negative. The angle f3 is the slope
of the free surface in the general reference system (X,Y) and the vectors 0 and t are the normal
and the tangential vector to the free surface showed in figure 1.3. The reference system (x,y)
corresponding to the principal directions of the permeability tensor K is not shown in figure
1.3. This figure also gives the following relations between the two reference systems (X,Y)
and (x',y'), ie.

x = cos "( x' - sin "( y'

Y = sin "(x' + cos "(y'
x' = cos "( X + sin "( Y (1.20)
y' = - sin "( X + cos "( Y
heX, t) = sin "( x' + cos "( h'(x',t)
 11

y'

h(X,t)
y
x'

-=~~~ __________ __ __ __ __________

~ ~ ~ ~ - L_ _ _ X

Figure 1.3 ~
In order to develop a kinematic boundary condition along the free surface, one needs to defme a
function F constant in time at each free surfac y point The total derivative of this function F is
equal to zero [5], ie.

dF aF aFax' aF ay'
dt = (i+dx' Tt + dfTt = 0 (1.21)

In the present case, the function F is given by the following relationship:

F(x', y', t) = h' (x',t) - y' = 0 (1.22)

The introduction of (1.22) into (1.21) gives (1.23).

ah' ah' ax' ay'

Tt+dx'Tt-Tt=O (1.23)

In equation (1.23), the partial derivatives of x' and y' with regard to time give the velocities of
the free surface point along these axis, with e the effective porosity of the medium, ie.

and
(1.24)

The components of the velocity vector for the case of an anisotropic domain in function of the
velocity potential and in the (X,Y) system are given by the following relations.

au au
Vx = - Kxx ax - KXY dY + Nx
au au (1.25)
Vy = - KXY dX - Kyy iN + Ny
 12

Taking into account the relations between the two reference systems (1.20), the introduction of
(1.24) and (1.25) into (1.22) gives the following equation.

ah' 1 ah'[ au aU]

dt - -; dx' (Kxx cOlf'( + Kxy siny) dX + (Kxy cOlf'( + Kyy siny) iN

1[ (Kxx siny - Kxy cosy) dX

--
~ + (Kxy siny - Kyy cosy) iN
~]
(1.26)
e
1 ah' 1 ah'
+ - Nx ("CT cosy + siny) + - Ny ("CT siny - cosy) = 0
e ax e ax

This relationship is non linear because its second term contains the product of two partial
derivatives. In order to linearize it, one can write (see figure 1.3):

ah'
dx' = - tg (y - ~) (1.27)

The second and the third terms of equation (1.26) contain the derivative of the potential along
the axis X and Y. But, as the unknowns of the problem are the potential u and its derivative
au/an, one needs to express X and Y in terms of derivatives along the normal vector n and the
tangential vector t (1.28).

au [ au aU]
dX = A. (Kxy cos~ + Kyy sin~) do + (Kxy sin~ - Kyy cos~) dt

au [ au aU]
eN = - A. (Kxy cos~ + Kxy sin~) do + (Kxx sin~ - Kxy cos~) dt (1.28)

1
with: A. = - - : - - - - -
2
<Kxy) - Kxx Kyy

Introduction of equations (1.27) and (1.28) into (1.26) gives after several manipulations the
following equation.

au cosy ( au )
- do - cos ~ Ny + sin ~ Nx
dt - e cos (y- ~) (1.29)

If the vector N (n) is defined as the normal recharge along the free surface, the relation (1.29)
gives the following equation.

au - cosy ( au )
- do + N(n)
dt - e cos (y-~) (1.30)

This equation represents the kinematic boundary condition along the free surface as it varies in
time and space.
Notice that a fully anisotropic porous media has been considered and that condition (1.30) is
applicable to anisotropic as well as isotropic cases.
 CHAPTER 2
THE LAPLACE EQUATION

2.1. INTRODUCfION

The subject of this book is the study of steady and unsteady porous media flow.
In order to study a physical problem, one can describe it by a mathematical model. In the
present case, this process leads to the Laplace equation (see previous chapter) which is
one of the fundamental equations of engineering analysis.
Different methods have been presented to solve the Laplace equation using computers.
After the fmite difference method and the finite element method, the boundary element
method has been developed by combining the methodology of the finite element method
with the theory of boundary integral equations.
The Laplace equation allows to describe many physical problems and it will be interesting
first to study some special points concerning the application of the boundary element
method to this equation. These special points concern especially numerical integration
and the division of the domain into sub-n;:gions. After that, the boundary element method
will be applied to solve different types of porous media flow problems which are
described by the Laplace equation.

2.2. 1HE MATIIEMATICAL MODEL

In order to develop the mathematical model which describes a physical problem, one
needs to make some assumptions which reduce the field of application of the model. One
generally obtains an equation which contains variables describing the problem and some
parameters function of the characteristics and the properties of the domain under
consideration. These assumptions have been made to obtain the mathematical model and
one has to verify that the problem does not violate any of those assumptions.
In many cases, the Laplace equation (2.1) yields the following mathematical model.

K(~u) =0 (2.1)

In this case, the potential u is the variable describing the problem and the components of
the tensor K introduce the characteristics and the properties of the medium into the model.
Equation (2.1) describes many problems like porous media flows, thermal conduction
and electrostatics.
 14

2.3. MATHEMATICAL ASPEcrs OF THE LAPLACE EQUATION

2.3.1. Introduction
In its most general form, the Laplace equation is given by the following relationship:

(2.2)

It is always possible to transform equation (2.2) by an appropriate change of variables

into the following relationship.

a2u a2u
K - +~- = 0 (2.3)
x ax2 ay2
In this chapter, some mathematical aspects of the Laplace equation (2.3) will be
considered. The equation can also be written as given below.

K(t.,U) =0 (2.4)

where t. is the Laplace operator and K the tensor function of the properties of the domain.

2.3.2. Boundary conditions

To solve the Laplace equation on a given domain n, one needs to impose conditions
along its boundary i.

...

r,o
y £n

r1

Figure 2.1

One can impose one of the three following conditions at any point on the boundary i:
1) u =uo' the Dirichlet condition;
2) q = qo, the Neumann condition;
3) yu + 'IjIq + 1: =0, the Cauchy condition with the parameters y, '" and 1: depending on
some characteristics of the domain or on the geometry of the boundary in the vicinity
of the point where this condition is imposed.
Usually, one needs one condition on any point on the boundary. An exception to this rule
is given by boundaries the positions of which are an unknown of the problem. Along
 15

such boundaries, one needs two boundary conditions in order to detennine their positions
by an iterative procedure. In case of groundwater flows, the positions of the free surface
is initially unknown and consequently, two conditions are required there.

2.3.3. The anisotropic case

The most general form of the Laplace equation is given by the following relationship:

(2.2)

The coefficients multiplying the partial derivatives of the potential u are the components of
the tensor K (2.5). They depend on the orientation of the axis of the reference system.

(2.5)

Generally, the Laplace equation is written with regard to the principal directions of the
tensor K. In this case, the components of the tensor K which are not located on its
principal diagonal are equal to zero, ie.

Kx 0,)
K= (OK y (2.6)

The corresponding Laplace equation is given by the following equation:

(2.7)

The coefficients of the two tensors (2.5) and (2.6) are linked by the following two
relationships:

(2.8)

Transformation of equation (2.2) into equation (2.7) corresponds to a rotation of the axes
of the reference system around the origin. The angle 9 of this rotation is given by the
relation:
Tan 29 = 2Kxy I (Kxx-Kyy) (2.9)

Transformation from the reference system (X,Y) to the reference system (x,y) is given by
the relations:
x = Xcos9+Ysin9 Y = - X sin 9 + Y cos 9 (2.10)
 16

A second change of variables allows to transfonn the anisotropic problem (2.8) into an
isotropic one, ie.

and (2.11)

(2.12)

In many practical applications, the domain n will be divided into several sub-regions
having different anisotropic properties. In this case, the first change of variables (2.10)
gives a different reference system to each sub-region corresponding to the principal
direction of their tensors K. As these changes of variables correspond to rotations of the
reference systems, they do not distort the geometry of the domain. On the other hand, the
second change of variables (2.11) correspond to a lengthening or a shortening along the
axes of the reference system and consequently, it distorts the geometry of the domain.
One only needs to consider the fITst change of variables in practice (equation (2.10».

2.4. SOLUTION OF TIlE LAPLACE,EQUATION

If conditions like those discussed above are imposed along the different boundaries of the
domain, it will be possible to solve the Laplace equation. Before the computers were
used for the numerical solution of the Laplace equation, many analytical solutions were
developed. These analytical solutions were obtained with the help of methods like
separation of variables or confonnal transfonnation. Unfortunately, these methods are
only applicable if the geometry of the domain and the boundary conditions are simple.
The development of computational methods now allow to solve numerically the Laplace
equation. To this end, different methods have been developed in last few decades. In
chronological order, these methods are: - the finite difference method (FDM);
- the finite element method (FEM);
- the boundary element method (BEM).
In comparison with the FDM and the FEM, the BEM presents important advantages for
groundwater flow applications as will be demonstrated in the following chapters.
 CHAPTER 3
INTRODUCTION TO THE BOUNDARY ELEMENT METHOD

3.1. INTRODUCTION

The boundary element method was developed at the University of Southampton by

combining the methodology of the finite element method with the boundary integral
method. The first international conference devoted to the boundary element method took
place in 1978 at Southampton [7]. Since that time, many books have been published ([8],
[9], [10], ...) and the numerous contributions to the annual conferences like BEM and
BE1ECH show the rapid development of the new method for all the engineering fields.
In this chapter, the weighed residual method will be used to develop the boundary element
method for the case of anisotropic Laplace problems. The weighed residual method is the
most general technique, because it can also be applied to develop the finite difference
method and the finite element method for instance.
The application of the boundary element method to Laplace problems is presented in some
well known references like the four books mentioned above. Because of this, only some
special points will be described in this chapter, especially those relating to the anisotropic
characteristics of the medium.
Finally, the application of the boundary element method to Laplace problems leads to
special numerical integration problems and problems connected with the subdivision of
the domain into sub-regions. These points will be investigated in chapters 4 and 6.

3.2. GOVERNING EQUATION

Let Q be a domain in which the following anisotropic Laplace equation applies, ie.

(3.1)

The values of the components of the tensor K depend on the characteristics of the domain
Kx and Ky. They are the principal values associated to the corresponding principal axis x
and y. In order to determine the value of the potential u on the domain Q, one has to
impose Dirichlet (along r 1) or Neumann (along r 2) conditions on the boundaries of the
domain:
 18

u = u along f, and q = q along f2 (3.2)

The bars on the terms in (3.2) indicate that those are the imposed values of the potential
(u) or the normal velocity (q) along the boundary.
y

x
Figure 3.1

In case of a Cauchy boundary condition ('YU + 'l'q + 't =0), one of the two unknowns will
be expressed in terms of the other and afterwards introduced in the same way as a
Dirichlet or a Neumann boundary condition into the linear system.

3.3. TIffi BASIC EQUATION OF TIffi BOUNDARY ELEMENT METHOD

3.3.1. The weighed residual method

When one applies a numerical method to solve the Laplace equation (3.1), the numerical
result will generally give an error EI' This error must be as small as possible in order to
obtain accurate results. Thus, the approximation of the potential field u in the Laplace
equation gives the following result:

K (~u) , :;to
= E (3.3)

In order to obtain an accurate solution, the error EI must be as small as possible. EI is also
called the residual and in order to minimize it, one can use a weighting function FI to
weight the residual on the domain. Thus, the weighed residual method applied to relation
(3.3) leads to :

(3.4)

The accuracy of the solution depends on the choice of the weighting function Fl'
The application of a numerical method to solve the Laplace equation does not allow to
satisfy it rigorously on the domain Q. Consequently, the solution will be only approached
 19

along the boundary where the boundary conditions will produce errors Ez and ~ :

(3.5)

Thus, one proceeds with the boundary conditions by using two weighting functions F2
and F3, which are generally different from Fl. So, one obtains the following
relationships for the two parts of the boundary r.

(3.6.a)

f ',F,dr, f
r2
=
r2
(3.6.b.)

As each of these three weighting residual relations (3.4) and (3.6) must be equal to zero,
one can group them together in a single relation, ie.

(3.7)

As each term of (3.7) must be separately equal to zero, one can choose their signs to
allow some simplifications.
Before proceeding with this developments, one needs to make a choice concerning the
weighting functions F l , F2 and F3. In order to obtain the boundary element method, one
can define,

(3.8)

The special choice concerning F2 will be justified showing that it produces a useful
relationship which leads to the cancellation of certain terms. One can also consider the
units.ofthe two terms of the right part of (3.7) to justify the special choice for F2.
After the introduction of (3.8) into (3.7), one obtains the following weighed residual
relation:

(3.9)
 20
-
dU dU
with: q = dii and q = dii

In order to continue the transformation of (3.9), one now requires the Green's theorem to
transfer the Laplace operator from the potential u to the weighting function w.

3.3.2. The Green's theorem

The Green's theorem is given by the following relationship:

I !l
W K (Au) dQ = I !l
uK (Aw) dQ + I~ I
r
w dr -
r
(3.10)

In the case of an anisotropic domain, the tensor K is given by equation (2.6) and it is
possible to show that the operator d/dn is as follows,

(3.11)

The coefficients nx and ny are the components of the unit outward normal vector to the
boundary.
In order to transfer the Laplace operator from the unknown potential u to the weighting
function w, one can use the Green's theorem (3.10) in equation (3.9).

3.3.3. The basic relation of the boundary element method

After the introduction of (3.10) into (3.9), one obtains after several steps the following
equation:

I!l
uK(~w)dQ=
I I
r
udildI"-
dw
r
Wdiidr
au (3.12)

with : u = u on r and u = u on r
2 1
(3.13)
q = q on r and q = q on r
1 2

The first term of (3.12) contains an integral in the domain Q, whereas its second term
contains two integrals on the boundary r of the domain. A judicious choice of the
weighting function w will allow to eliminate the integral on the domain Q.
 21

3.4. 1HE WEIGHTING FUNCTION

3.4.1. Choice of the weiihtini function

In order to eliminate the remaining integral in the domain in (3.12), one can choose as
weighting function w the fundamental solution of the Laplace equation with Sj, the Dirac
function or distribution. Thus, the weighting function and its derivate along the outward
normal vector can be written as,

(3.14)

~
2
-1 x y
w =21tjKx"Ky In Kx +"Ky (3.15)

(3.16)

Taking into account the properties of the Dirac distribution, the introduction of (3.14) into
(3.12) gives:

(3.17)

The two remaining integrals of (3.17) apply only on the boundary of the domain. This
relationship is the starting point to develop the boundary element method.

3.4.2. Localization of the collocation point

The relation (3.14) is the continuity equation corresponding to a unit source located at the
point i. This point is called the collocation point and its position can not be chosen
arbitrarily. Indeed, the weighting function (3.15) and its derivative (3.16) are infinite at
the collocation point i. Considering this property, one has three possibilities regarding the
localization of the collocation point:
- firstly, one can locate the collocation point i on the boundary r. In this case, the
integrands of (3.17) are infinite at i and one has to be careful when carrying out the
numerical integrations.
- secondly, one can locate the collocation point i outside the domain Q. In this caSe, the
integrands of (3.17) do not tend towards infinity. It is however difficult to define the
optimal position for i. The choice of the position of i will influence the accuracy of the
results [63], [60].
- finally, the collocation point can be located inside the domain, in which case the
integrals here have finite values. In this case, the potential at the collocation point
introduces an additional unknown into the system.
 22
In the present case the fIrst approach (which is the classical one) will be used. One will
see later on that the detailed study of the numerical integrations will allow to obtain very
accurate results. One also has to analyse the behaviour of the integrals of (3.17) in the
vicinity of the collocation point i. This point will be investigated in the following section.

3.5. ANALYSIS OF THE INTEGRALS

3.5.1. Introduction

In this section, the behaviour of the integrals of (3.17), in the vicinity of the collocation
point i will be studied. Indeed, x and y are equal to zero at i and consequently, the
weighting function w (3.15) and its derivative (3.16) are infInite there. Consequently,
one needs to verify if the two integrals of (3.17) are integrable and if the singularities at
this point could have any special contribution to equation (3.17).

3.5.2. The fIrst integral

Taking into account the relation (3.15) for the weighting function w, the second integral
in (3.17) is given by:

11 =
-1 I doau (3.18)
2njKxKy n
r
Y

Yj

Yi-r---------r.-r~~-----------------x~

Figure 3.2

The integrand of (3.18) tends towards infInite at the collocation point i, but as In r is
integrable on each interval [0 IL], with IL different from infinite, II is integrable. Thus,
the numerical integration of (3.18) do not pose problems, except in the vicinity of i where
one has to use a numerical integration appropriate to the behaviour of the integrand there.
In the case of this fIrst integral (3.18), the singularity does not give a specifIc
contribution. To show this, one can consider the collocation point as isolated from the
rest of the boundary with the help of a circular notch r e the radius e of which will tend
towards zero (fIgure 3.2).
 23
Thus, II gives the following relationship:

(3.19)

The first integral of (3.19) tends towards 11. To analyse the second one, one can change
from cartesian coordinates to polar ones. Thus, one obtains the following relationships:

x =£cosO y = £ sinO (3.20)

The introduction of (3.20) into the second integral of (3.19) gives:

lim (3.21)
£~O

This last relation shows that the singularity at the collocation point does not result in any
specific contribution to the integral 11.

3.5.3. The second integral

Taking into account the relation (3.16) for the derivative of the weighting function, the
second integral is given by :

(3.22)

The analysis of the integral (3.22) shows that its integrand is infinite at the collocation
point i, but does not tend towards infinite at this point. Consequently, the integrand is
continuous at the collocation point i. It is possible to show that the value of the derivative
of the weighting function at the collocation point is given by :

an x any
aw ny ar - nx ar
-1
lim (3.23)
x~O.y~O
do

Due to the special behaviour of the integrand of 12 in the vicinity of the collocation point i,
the singularity will give a non zero contribution to the integral. In order to determine this
contribution in the anisotropic case [13], [19], one can use the same technique as for the
first integral (see figure 3.2). In the present case, 12 gives:
 24

(3.24)

The first integral of (3.24) tends towards 12. If one introduces now (3.24) into the
relationship (3.17), one obtains the relation (3.25) because u tends towards Uj when E
goes to zero.

au aw
CjU j =
J
r
w- dr-
an J r
u do dr (3.25)

with : Cj =1+
lim
E-tO
I £
aw
- dr
an

FOIIDula (3.26) shows that the value of Cj depends on the following (see figure 3.2)
(3.26)

- the angle (8b - 8a );

- the characteristics Kx and Ky of the medium;
- the position of the angle a in the principal reference system (x, y) associated with the
characteristics of the medium.
The integration of (3.26) gives:

cj = 1-~[ 2n
Arctg (
J[K:t
R;
g 8}f
e
(3.27)
a

The primitive of the integral of the equation (3.26) given by (3.27) presents a
discontinuity at 8 =± n/2 which one needs to take into account when computing (3.27).
In the case of isotropic characteristics (K = Kx = Ky), (3.27) gives:

Cj = 1 - ro I 2n (3.28)

where ro = 8b - 8a. Notice that in the case of a smooth boundary (ro = n), equation (3.28)
gives'cj = 0,5.
As have been seen, the integrand of (3.22) does not tend towards infinite at the
collocation point i, but is continuous. Nevertheless, in order to integrate numerically
(3.22) in the vicinity of i, special integration methods will be required.
 25

3.6. DISCRETIZATION OF THE PROBLEM

To discretize in equation (3.25) the potential u and its nonnal derivative q on the boundary
r, one can use the classical isoparametric shape functions [8], [9], [10] for two, three and
four node elements. These elements correspond respectively to linear, quadratic and
cubic shape functions. The use of the shape functions entails a change of variables.
Thus, each element is transfonned from the general reference system (X,Y) to its own
non-dimensional reference system. This is done by the relation (3.29) with j varying
from 1 to Nn(e), the number of nodes of the element being e. As in the classical
isoparametric finite element, the different shape functions Nj are multiplied by the
corresponding node values 2j of the coordinates or the unknowns.

(3.29)

The figure 3.3 shows the change of variables in the case of a three node element

y
(2) (3)

(l)~ -1'
(l) (3)
1
.. ~

Figure 3.3

To discretize the whole boundary, Ne elements like those shown by the figure 3.3 will be
used. Thus, one obtains :
Ne
r = L re (3.30)
e=!

The introduction of the discretization of the boundary and the unknowns into the
boundary integral equation (3.25) gives the following discretized boundary element
equation:

Ne (Nn(e) e) Ne (Nn(e) e )
L L q.G..
J IJ
- L L u·J R ..IJ (3.31)
e=! j=! e=! j=!

W1·th : H IJ..e = i
re
dW
"'C"N.dr
an J e (3.32), and G~.D = ~(
re
w N. ctr
J e
(3.33)

Equation (3.31) contains one unknown at each node of the discretization. For this
reason, one locates successively the collocation point i at each node of the discretization
and obtains a linear system representing the boundary value problem.
 26

3.7. SPECIAL PROBLEMS

3.7.1. Determination of the coefficients oeij and Heij

In the discretized boundary element equation (3.31), the coefficients oeij and Hei} given
by the relations (3.32) and (3.33), need to be determined. These numencal integrations
will be considered in the next chapter. Indeed, their integration requires special
re
techniques when the collocation point i is located on the element under consideration.

3.7.2. Determination of the ci coefficient

As shown by the relation (3.26), the determination of the ci coefficient is more difficult
when the domain has anisotropic characteristics. In order to get round this problem, one
can use equation (3.31) to determine the ci coefficient. As shown by the relations (3.26),
(3.32) and (3.33), the ci coefficient, oeij and Heij depends only on the geometry of the
domain and its characteristics and not on the boundary conditions. Consequently, if one
considers that the potential u has a constant value on the whole domain, equation (3.31)
gives:
Ne (Nn(e) )
c·1 = - ~ ~ H~.
1J (3.34)
e=l j=l

3.7.3. The division of the domain into sub-regions

The method so far described in this chapter requires that the coefficients Kx and Ky are
constant on the whole domain. But in many practical cases, the domain needs to be
divided into several parts, each of them having different characteristics. In these cases,
one has to divide the domain into several sub-regions having different characteristics.
The sub-regions will be separated by interfaces along which one has to impose the
continuity of the potential u and the normal velocity q. The intersection points of several
interfaces or of an interface and a boundary lead to special numerical problems. The
division of the domain into sub-regions and these special problems will be considered in
chapter 6.

3.7.4. The junction point of two boundaries

At the junction point of two boundaries r A and r B (figure 3.4), there is uncertainty
regarding the continuity of the potential u and its derivative along the two outward normal
vectors corresponding to the two boundaries at this point.
In order to solve this problem, different methods have been suggested :
- to place a small curvature at the corner at the point P to obtain there one single
normal vector [39];
- the use of two physically separated nodes in the vicinity of P or the use of
discontinuous elements [64], [63];
- the use of an additional equation to describe the potential field in the vicinity of P
[2], [43], [55];
- the double node technique [70].
The double node technique consists in the localization of two nodes at the point P,
belonging respectively to the two boundaries r A and r B. This method avoids the
 27

ambiguity between the two boundary conditions at this point and is the one used here.
However, this technique can lead to problems at the junction point of several interfaces.
These problems will be analysed in chapter 6 which studies the division of the domain
into sub-regions .

...n p

Figure 3.4 Figure 3.5

3.7.5. Detennination of the unknowns at interior points

Mter the solution of the boundary value problem, the potential and the nonnal velocity at
every boundary point are known and equation (3.17) allows to determine the potential at
any interior point. If one locates the collocation point i at an interior point, the only
remaining unknown in (3.17) is the potential at this point. Finally, in order to obtain the
gradient at this point, one can differentiate the relation (3.17) [8].

3.8. CONCLUSIONS

In this chapter, the boundary element method for the case of a Laplace problem on an
anisotropic domain has been presented.
The chapter has shown that the numerical integrations poses some problems due to the
fact that the weighting function wand its nonnal derivative are infmite at the collocation
point i. The study of this problem is the subject of the next chapter.
 CHAPTER 4

NUMERICAL INTEGRATION

4.1. IN1RODUCITON

After the introduction to the boundary element method in the previous chapter, the
numerical integration of the integrals (3.32) and (3.33) will be considered here, ie.

Gij = i
re
wN.J dre ( 4.1.)

H.·eIJ .: i
re
dW N· d r
"-
oR J e (4.2.)

As mentioned in the previous chapter, one has to study very meticulously the numerical
integration of (4.1) and (4.2), since their integrands become infinite if the collocation
point i is located on the element r e under consideration (singular integration). If the
collocation point is not located on that element, the numerical integration of (4.1) and
(4.2) does not pose any special problem and in this case, one always use the Gauss
integration method (regular integration).

4.2. REGULAR INTEGRATION

4.2.1. Introduction
In the previous chapter, the geometry of the boundary, the potential u and the normal
velocity have been discretized with the help of the isoparametric shape functions Nj.
These shape functions are functions of a new variable~. Therefore, the integrals (4.1)
and (4.2) require a change of variables from the reference system (X,Y) to the new
variable~. Thus, one obtains the following relationships with J(~), the jacobian of the
change of variables.
 29

Hij = f ;:
-1
(SJ N;C<) J <<> d< (4.3)

f
1

Gij = w <<> N;C<> J © d< (4.4)

-1

One needs to consider the different terms before computing these two integrals.

4.2.2. The shape functions

The shape functions Nj (~) depend only on the new variable ~ and are continuous and
derivable on the interval [-1 1]. Thus, the computation of the shape functions and their
derivatives presents no problem.

4.2.3. The jacobian

The jacobian of the change of variables is given by the following formula:

with: x = Xj Nj @
Y = Yj Nj @ (4.6)
~r
~x
n
j = 1,.... , Nn (e) X
Nn (e) : number of nodes of
the element re
Figure 4.1

After the introduction of the equation (4.6) into (4.5), one obtains the following
relationship with Nj(~) the derivatives of the shape functions with regard to ~.

(4.7)

Like the computation of the shape functions, those of the jacobian do not present any
special problem.

4.2.4. The wei&htin& function

The weighting function w and its derivative along the outward normal vector n are given
by equations (3.15) and (3.16). Unlike the shape functions and the jacobian, these two
relationships also depend on the characteristics of the domain. As these functions are
 30

infinite at the collocation point i, one needs to be sure that the integration points are not
located at a node. Finally, to compute the derivative of the weighting function, one has to
calculate first the components of the unitary outward nonnal vector.

4.2.5. The outward nonnal vector

Figure (4.2-A) shows the position of the unitary outward normal vector and its two
components along the axis X and Y. Figure (42-B) illustrates the differential variation of
the boundary in the vicinity of the point P. The result of this is that nx and ny are given
by the following relationships, with J the jacobian of the change of variables and x and y
given by (4.6).
y

r
%jdY
dx

x (B)
(A)
Figure 4.2

. dy 1 ay
nx = -slha = -- =-- -
dr J a~
(4.8)
1 ax dx
n =cosa=-=--
Y dr J a~

Thus, the computation of nx and ny does not present any special problem.

4.2.6. The numerical intewtion

After the computation of the different terms which make up the integrands of (4.1) and
(4.2), one can consider now their numerical integration by the Gauss integration method.

f f(x) dx =
k=l
f f(xk )· Wk,G (4.9)
-1

The application of this method to the integrals (4.1) and (4.2) gives:

e P aw
Hij = :E - (~kl. Nj(~ . J(~kl' W k G (4.10)
k=1 an '
P
G~. =k :E=1 w(~J. NJ'(~J . J(~J . W k,G (4.11)
1J
 31

with : ~k the integration points;

:
W k,G : the associated weights;
p : the number of integration points per element
The integration points and their associated weights are given in table 4.1 [1], [10].
w,
n=2
0.5773502691 89626 1.00000 00000 00000
n=3
0.00000 00000 00000 0.88888 88888 88888
0.774596669241483 0.555555555555555
n=4
0.33998 1043584856 0.652145154862546
0.86113 63]]5 94053 0.3478548451 37454
n=5
0.00000 00000 00000 0.56888 88888 88889
0.53846 93101 05683 0.4786286704 99366
0.90617 98459 38664 0.2369268850 56189
n=6
0.2386191860 83197 0.46791 3934572691
0.6612093864 66265 0.36076 1573048139
0.932469514203152 0.1713244923 79170
n=7
0.00000 00000 00000 0.4179591836 73469
0.40584 51513 77397' 0.381830050505119
0.741531185599394 0.279705391489277
0.949107912342759 0.1294849661 68870
w,
n=8
0.183434642495650 0.362683783378362
0.5255324099 16329 0.313706645877887
0.79666 64774 13627 0.22238 10344 53374
0.96028 98564 97536 0.10122 8536290376
n=9
0.00000 00000 00000 0.33023 93550 01260
0.3242534234 03809 0.31234 70770 40003
0.613371432700590 0.26061 06964 02935
0.83603 1107326636 0.18064 81606 94857
0.968160239507626 0.081274388361574
n= 10
0.148874338981631 0.295524224714753
0.4333953941 29247 0.269266719309996
0.67940 95682 99024 0.21908 6362515982
0.86506 33666 88985 0.14945 13491 50581
0.97390 6528517172 0.066671344308688
n=12
0.1252334085 11469 0.2491470458 13403
0.36783 1498998180 0.2334925365 38355
0.58731 79542 86617 0.203167426723066
0.76990 26741 94305 0.16007 83285 43346
0.90411 7256370475 0.106939325995318
0.98 J56 06342 46719 0.047175336386512

Table 4.1
4.2.7. Conclusions

In the above section, the Gauss integration method has been applied to compute the
integrals (4.1) and (4.2) giving the components of the linear system. To use this
integration method, the collocation point i must be located on an element other than the
one under consideration. If this condition is not verified, one needs to apply singular
numerical integration methods to compute these two integrals. These methods will be
developed in the following sections. The developments presented up to now regarding
the shape functions, the jacobian and the components of the outward normal vector will
also be applied for the study of the singular integrals.

4.3. SINGULAR INTEGRATION OFGeij

4.3.1. Introduction

This paragraph deals with the integration of the equation below when the collocation point
i is located on the element under consideration,

Gij "I re
wNj dr, (4.1)
 32
The singular integration of (4.1) is based on the splitting of its integrand [62], [65], [19].
Thus, one will obtain five integrals which will be integrated with similar accuracy as that
provided by the Gauss method in the case of regular integrals.

4.3.2. Singular integration

After the introduction of the weighting function w given by (3.15) into (4.1), this integral
yields:

Gij = k
21t K K
x Y
fIN/~). J(~).
-I
In r @ d~ (4.12)

with:r= (4.13)

If the collOCation point i is located at the abscissa ~ = ~i' the integral (4.12) gives:

(4.14)

with : (4.15)

IB = f
l

-1
Nj(~) . J @ .In
(I ~ -~.I )
2 I d~ (4.16)

The integral IA (4.15) is not singular at the collocation point Indeed, the limit yields:

(4.17)

This limit is only equal to zero if the two squares are together equal to zero. As shown by
equation (4.7), the jacobian is also equal to zero in this case. Thus, the integrand of
(4.15) does not tend towards infinity at the collocation point and consequently, this
integral can be integrated by the standard Gauss integration method
To analyse the second integral IB (4.16), it is more convenient because of the absolute
value to divide the integration interval into two parts. Thus, one obtains,
 33

(4.18)

with: f ~ (~.T-~ )
-1
In NjJd~ (4.19)

(4.20)

To simplify the relations, one can defme the following change of variable :

~. -~ = ~__
-~.
IBI : A
1
= _1_
2
IB2 : A
2 2
1 (4.21)

This change of variables also affect the shape functions Nj and the jacobians J of the
integrals (4.19) and (4.20) as they are functions of~. The introduction of these two new
variables into (4.19) and (4.20) yields the following relationships:

(4.22)

J d /.. (4.23)
2

These two integrals require another change of variables to return the integration interval
from 0 to 1. This second change is given by :

2\
11 = - - (4.24)
4 1- ~

The introduction of these two new variables into (4.22) and (4.23) yields the following
two relationships:
 34

IBl=(l+~)InC:1;j) iN 1

o
j Jd1'\2-(1+1;) i 0
1

In(:
2
)N j Jd1'\2 (4.25)

IB2 = (1 - 1;j) In C; 1;j) iN

0 0
1

j J d1'\4 - (1 - 1;;) i 1

In (:
4
)N
j J d1'\4 (4.26)

The integrands of the first integrals of (4.25) and (4.26) are never singular, since they
only contain the product of the shape functions by the jacobian. Thus, these two integrals
will be computed by the standard Gauss integration method. However, to carry out the
integrations, they require a third change of variables to bring back their integration interval
to the standard -1 to + 1. The second integrals of (4.25) and (4.26) instead present
singularities at (Tj = 0). To compute these integrals, the Berthod-Zaborowisky method
considered in section 4.3.3 will be applied. The third change of variables is given by:

(4.27)

The introduction of these two new variables into the first integrals of (4.25) and (4.26)
yields respectively:

(4.28)

(4.29)

By grouping together the relations (4.14), (4.18), (4.28) and (4.29), one obtains the
relationship (4.30).

Ofj =-
21t
FA lfl
1
K K
In -I2r @
~--I
~ -~j
j
N .(~) J@ d~
J
x y .1

I
1+~ 1+~jf
+ -2- In -2- NjCll l ) J(11 I ) d1l 1
·1
I

- (1 + ~i) fin.!. "\

N·(11 )J (11 ) dll
J 2 2 2
(4.30)
o
 35

Finally, taking into account the relationships (4.21), (4.24) and (4.27), one fmds :

~.1 - 2~- ~.1 - 2~ +1 ~.1 -~

11 1 = (4.31)
~.-
1
1

In conclusion, to compute equation (4.30), the Gauss integration method can be applied
to evaluate its first, second and founh integral and the Berthod-Zaborowisky method to
evaluate its third and fifth terms.

4.3.3. The Berthod-Zaborowisky integration method

The Berthod-Zaborowisky integration method is similar to the Gauss integration method

and is defined as follows,

(4.32)

This integration method has the same prpperties as in the Gauss, and equation (4.32) has
p integration points which compute the integral exactly if the degree of f(x) is equal or
lower than (2p - 1).
The integration points and their associated weights are given in table 4.2 [IOJ.
n w, n /;, w,

2 "
0.11200880 0.71853931 8 0.13320243 ( -1) 0.16441660
0.60227691 0.28146068 0.79750427 ( -1) 0.23752560
3 0.63890792 ( - 1) 0.51340455 0.19787102 0.22684198
0.36899706 0.39198004 0.35415398 0.17575408
0.76688030 0.94615406 (-1) 0.52945857 0.11292402
4 0.41448480 (-1) 0.38346406 0.70181452 0.57872212 (-1)
0.24527491 0.38687532 0.84937932 0.20979074 ( -1)
0.55616545 0.19043513 0.95332645 0.36864071 (-2)
0.84898239 0.39225487 ( -1) 9 0.10869338 (-1) 0.14006846
5 0.29134472 (-1) 0.29789346 0.64983682 ( - 1) 0.20977224
0.17397721 0.34977622 0.16222943 0.21142716
0.41170251 0.23448829 0.29374996 0.17715622
0.67731417 0.98930460 ( - 1) 0.44663195 0.12779920
0.89477136 0.18911552 (-1) 0.60548172 0.78478879 (-I)
6 0.216344005 (-1) 0.23876366 0.75411017 0.39022490 ( - I)
0.12958339 0.30828657 0.87726585 0.13867290 (-1)
0.31402045 0.24531742 0.96225056 0.24080402 ( - 2)
0.53865721 0.14200875 10 0.90425944 ( - 2) 0.12095474
0.75691533 0.55454622 ( -1) 0.53971054 (-1) 0.18636310
0.92266884 0.10168958 (-1) 0.13531134 0.19566066
7 0.16719355 (-1) 0.19616938 0.24705169 0.17357723
0.10018568 0.27030264 0.38021171 0.13569597
0.24629424 0.23968187 0.52379159 0.93647084 (-1)
0.43346349 0.16577577 0.66577472 0.55787938 ( -1)
0.63235098 0.88943226 ( - 1) 0.79419019 0.27159893 (-1)
0.81111862 0.33194304 (-1) 0.89816102 0.95151992 (-2)
0.94084816 0.59327869 ( - 2) 0.96884798 0.16381586 (-2)

Note: Numbers are to be multiplied by the power of 10 in parentheses.

Table 4.2
 36
4.3.4 Comparison of the different integration methods

4.3.4.l.Introduction
In this section, the results given by different integration methods will be compared with
those given by Gauss. The fIrst example is simple to make an analytical computation
possible.

Gij =
21t
fix flN}~).J@. 1nr@d~
KK
x Y -1
(4.12)

The collocation point i will be located successively at each of the Nn(e) nodes of the
element and at each node, the integral (4.12) will be computed. The Nn(e) shape
functions Nj(~) will produce a square matrix with Nn(e) lines and columns, each of its
terms corresponding to an integral of the type given by equation (4.12).

4.3.4.2 First case

The fIrst example [19] is shown in fIgure 4.3, considering that the domain is isotropic (K
= Kx = Ky).

(1) (2) (3)

~.~----~.~----~.---- ~
-1 fe 0 1

Figure 4.3
In this case, equation (4.12) gives:

-1
Gij = (4.33)
21tK

with: i andj = 1,2 and 3

~l =-1, ~2=O and~3 = 1
Nl = ~(~ - 1) / 2, N2 = (1 - ~)( 1 +~) and N3 = ~(~ + 1) / 2

The analytical computation of (4.33) gives the following results :

with : gl = 17 + 3 In (04) g2 = 4 .(5 + 3 In (K/4»

g3 = -1 + 3 In (K/4) g4=2+3InK (4.34)
gs = 4.(8 + 3 In K)
 37

Figure 4.4 compares the analytical solution given by (4.34) with the singular integration
as resulting from (4.30) and the Gauss integration of equation (4.33) using 2, 4, 6 and 10
integration points. The figure 4.4 shows that the singular integration method gives the
same results as the analytical formulas (4.34). Indeed, the numerical integration methods
used here compute the integrals of the functions f(x) of the power (2p - 1) with P
integration points exactly (see relations 4.9 and 4.32). It is possible to show that in this
case all the functions f(x) are of power 2 and consequently, the singular integration
method gives effectively the analytical solution (apart from the rounding errors).

- analytical
singular integration
- + - Gauss integration
NIP: Number of integration
points

I
I
-1,0 I
I

;l-
I
-1,5

Figure 4.4

If the results given by the Gauss integration method are compared with those of the
analytical solution, one notices a slow convergence, especially for the terms
corresponding to the principal diagonal (gl and gS)' The use of the singular integration
method given by (4.30) is therefore essential to obtain satisfactory results.

4.3.4.3. Second case

The second example [19] is illustrated in figure 4.5. In this case, one also considers a
domain with isotropic properties (Kx = Ky = 1).

3 ( 4)

2 (3)

~O"",-=~-'--_ _", X
(1) 1

Figure 4.5
 38

In the present case, the analytical integration of (4.12) is not possible and the figure 4.6
only compares the results given by the singular integration methods with those given by
the Gauss integration method. Figure 4.6 shows that the results given by the Gauss
integration method converge slowly towards those given by the singular integration
method. Furthermore,this figure demonstrates that curvilinear elements require a great
number of integration points in order to obtain accurate results.

1, °
-.- singular integration
0,8 - + - Gauss integration
NIP: Number of integration
0,6 points
a:i=j=1
0,4 b: i =j =2
c: i =2 and j =3
0,2
-.--+. - --:.p.----
,,-- 1
-'-_-- ______ -!<. NIP
I ,,._.oor "

-0,2 /
.....
~
".,.---.-. _.-.:::::./c
-0,4
Figure 4.6

4.3.4.4. Conclusion

In this section, a special numerical integration method to compute Geij when the
collocation point i is located on the element under consideration has been developed.
Several changes of variables have allowed to split up the initial integral (4.12) into five
other integrals (4.30) for which well adapted numerical integration methods are available.
Two examples have shown that this method gives good results.
Finally, it has been demonstrated that curvilinear elements require a great number of
integration points in order to obtain accurate results.

4.4. SINGULAR IN1EGRATION OF H eij

4.4.1. Introduction

In this section, the numerical integration of Heij when the collocation point i is located on
the element under consideration will be studied. The integral under consideration is:

H e.. =
IJ
I ow :;-N.dr
un J e
(4.2)

re
 39
Before studying the singular integration of equation (4.2), one has to analyse its
integrand. As already mentioned at the previous chapter, the integrand of (4.2) does not
tend towards infinity at the collocation point. The Cj coefficient takes the contribution of
the singularity into account and the integrand is continuous at the collocation point. In
spite of this, one needs to develop a special integration method to obtain results as
accurate as possible for the case of curvilinear elements.

4.4.2. Behaviour of the integrand at the collocation point

As already pointed out in the previous chapter, the normal derivative of the weighting
function, and consequently the integrand of (4.2) tends towards a finite limit (L) at the
collocation point.

-1
L=---- Inn (4.35)
21t jKXKy x -+ O,y -+ 0

One can now replace x and y above by (4.6) and nx and ny by (4.10). Thus, one
introduces a change of variables from the general reference system (x,y) to the non
dimensional abscissa ~ and equation (4.35) becomes:

(4.36)

To evaluate this limit, one needs to apply twice the l'Hospital theorem which gives:

<mx <my
n --n-
aw -1 y ar x ar
lim dn = 2 2 (4.37)
x -+ o,y -+ 0 n 41t ~ (ny) (nx )
x y +
~-x;-
This last relation proves that the integrand of (4.2) tends towards a finite limit at the
collocation point (because nx2 + nl= 1) and consequently, is continuous there.

4.4.3. The strai~ht line element

In the previous point, it has been shown that the integrand of Heij never tends towards
infinity. To analyse its behaviour in the case of a straight line element, one considers the
numerator (4.38) of the derivative of the weighting function.

Num =nx . x + Dy • Y (4.38)

 40

y y
-I-
n

-+______~~~~~-----------------x
-+______~--------------------+x
Figure 4.7

If the collocation point i is located on the element under consideration or on the line on
which this element is located, one obtains for the different terms of (4.38) (see figure
4.7).
x=rcosa y = rsina nx = - sin a ny=cosa (4.39)

The introduction of (4.39) into (4.38) shows that the numerator is equal to zero. The
result of this is that the integral (4.2) is equal to zero when the collocation point i is
located on a straight line element .

4.4.4. The singularintemtion

Even if the integrand of (4.2) does not tend towards infinity at the collocation point, the
use of a special integration method [36] when the collocation point is located on the
element under consideration improves the solution with respect to the classical Gauss
integration method. To develop this special integration method, one needs to take into
account the position of the element in the reference system (x,y) [11]. That is why the
case of "horizontal elements" (element with a horizontal axis, element A on the figure 4.8)
and "vertical elements" (elements with a vertical axis, element B on the figure 4.8) will be
considered here.

-~--(~,
-+__________________________ X

Figure 4.8

In the case of an element having an oblique axis (element C on the figure 4.8), one can
use the formulas obtained for either horizontal or vertical elements. To study the special
integration method, one can introduce the weighting function w given by equation (3.15)
into (4.2). Thus, the following formula is obtained:
 41

e
Hij =
-1 f1

N/~).J @
nx x + Ilyy
2 2 d~ (4.40)
2x /KxK
--y -1
~+K
K '!..-
x y

The components of the unit outward normal vector are given by the relations (4.10) which
can also be written as follows:

nx = - y'/1@ (4.41)

Mter the introduction of (4.41) into (4.40), one obtains:

1
e
Hij =
-1 f x'y - xy'
N}~) 2 2 d~ (4.42)
2x /Kx~ ~ + '!..-
-1 K K
x y

From here on, one needs to distinguish between "horizontal elements" (EH) and "vertical
elements" (EV). Indeed, these two cases require different changes of variables in formula
(4.42). These changes of variables are given by the following relationships.

EH v@ =- ~Y
~ x
(4.43)

EV : t(~)=
~' --
Kx Y
(4.44)

After the introduction of (4.43) or (4.44) into (4.42), one obtains the relation (4.45) with
f(~) equal to v(~) or t(~) according to the position of the element in the reference system
(x,y).

Hij = -
-1
2x
f f (~= + 1)
1
Nj @ - - df
1+r
(4.45)
f(~= -1)

One can now take into consideration the following property of the Arctg function, ie.

df d (Arctg f@)
-= d~ (4.46)
1 +r d~
The introduction of (4.46) into (4.45) gives:

H~J. ~ f
=
2x
I N .@
J
d (Arctg f@)
----d~
d~
(4.47)

-1
 42

After the integration by parts of (4.47), one obtains:

To simplify this last relation, one can take into account a fundamental property of the
shape function as given in table 4.3, ie.

N j(f,; =-1) N j (f,;=+I)

j=1 1 0
1 <j < Nn (e) 0 0
j = Nn(e) 0 1

Table 4.3

The introduction of the results given in table 4.3 into equation, (4.48) produces the
following relationships:

1
- (Arctgf(~=-I) +1.) if j = 1
2x J

e 1
Hij= - I. if l<j<Nn(e)
2x J
1
(Arctg f (~= +1) - Ij ) if j = Nn (e)
2x
with : (4.49)
1

I = J N; ©
j Amgf© d;
. 1

f (;) = - j K,..!...
~ x
fm bOOmntal demon" Ell

f@ = j K,
Kx
::.
y
fm vertkal elomen" EV

If the numerator and the denominator of f(~) tend together towards zero for a given value
of ~, one can apply l'Hospital theorem to detennine the value of f(~) at this point.
Since the integral Ij does not present any singularity, it will be evaluated by the standard
Gauss integration method.
For "oblique elements" one can use one or the other relation to determine f(~). Indeed, it
is possible to show that in this case, the two relationships lead to the same results if one
applies the well known property of the function Arctg, ie.
withw
 43

Arctg (a) + Arctg ( ~) = ~ sign (a) (4.50)

Finally, a detailed analysis of the relation (4.49) demonstrates that the distinction between
horizontal and vertical elements is imposed by the discontinuity of the function Arctg
between plus and minus infinity [11].

4.4.5. Comparison test of the integration methods

At this point one can compare the singular integration method and the Gauss method with
an analytical solution for the case of a curvilinear element [19].
In this case, the collocation point i will be located successively at each of the Nn(e) nodes
of the element and at each node, one will compute the following integral with regard to the
Nn(e) shape functions.

e
Hij =
J1 N j (~) . J (~)
dw
do d~ (4.51)

-1

Thus, one obtains a square matrix with Nn(e) lines and columns and each of its terms
corresponds to an integral such as (4.51).
The figure 4.9 shows this example with an isotropic domain (Kx =Ky = 1).

1 (2)

.,..r,~-....i.a---X

Figure 4.9

After the computation of (4.51) for i and j varying from 1 to 4, one obtains the following
analytical solution.

h2 h3
h6 h7
h7 h6
h3 h2
 44

where: hi =- 3 a 2 + 2 (a2 - 2) . A 13 + 4 a B
h2 =15 aL ( 3 a 2 - 4) • A - 16 a B
h3 =- 2Ia2 + 2( 3 a 2 - 2) . A + 20 a B
h.i =9 a L (11 a - 4) • A 13 + 2 a (a2 - 4) B
hs = 5 a 2 + 2( a 2 - 2) C 13 - 4 aD
h(; = - 9 a 2 + (a2 + 4) C + 2 a (a2 + 4) D (4.52)
h7 = 3 a L 2 ( a 2 + 2) C - 4 aD
hs = a 2 + ( a 2 + 4) C 13
A =In (1 + 9 ( 2/4)
B = Arctg ( 3 a I 2)
C = In «(1 + ( ( 2 /4) I ( 1 + ( 2 »
D = Arctg (a) + Arctg ( a 12)

It is possible to show that all the components of the manix Wl tend towards zero when a
tends towards zero by applying three times the l'Hospital theorem. Figure 4.10 compares
the logarithm of the relative error in t~e Gauss and the singular integration methods for
two terms of the matrix Ife when a is equal to 1.
The analysis of the results given by this figure shows that the accuracy given by 2
integration points is always insufficient For a higher number of integration points (NIP),
the accuracy improves and the singular integration method gives bener results than the
Gauss integration method.

4 6 8
NIP e : relative error
NIP: number of integration
-1
points
a: i= 1 and j=3
-2 , b: i=j=2
,'+ - : singular integration
-3
\
\
.'' -+ - : Gauss integration
\\
-4

-5.
.~, \., \ lb
, +,
.~
T,
,
-6 \~,/a

Figure 4.10
The comparison of the integration methods for a three node element [13] leads to similar
conclusion.
 45

4.4.6. Conclusion

In this section, it has been shown that the integrand of Heij (4.2) tends to a finite limit at
the collocation point and that this integral is equal to zero for straight line elements when
the collocation point is located on the element or on its prolongation.
The main purpose of this section has been the development of a special numerical
integration method to compute formula (4.2) when the collocation point is located on the
element under consideration. The development of this method has shown that the
discontinuity of the Arctg function between plus and minus infmity leads to different
changes of variables according to the position of the element in the reference system.
Nevertheless, one can obtain the same relationships (4.49) in those two cases.

4.5. CONCLUSIONS
In this chapter, special numerical integration methods has been developed to compute
formula (4.1) and (4.2) when the collocation point i is located on the element under
consideration. By comparing the results given by the present method with those obtained
by Gauss integration and the analytical solution, it has been demonstrated that the present
method provides better results than Gauss. The use of singular integration method is very
important since they allow for the cOlnputation of the terms located near and on the
principal diagonal of the matrices. Indeed, these terms have a great influence on the
accuracy of the boundary element soluti9n.
In order to avoid the singular integrations, many authors use constant elements for which
all the integrals are computed analytically or linear elements with a higher number of
integration points. However, the computation of practical cases requires higher order
elements like quadratic or cubic ones, to obtain accurate results. These higher order
elements allow a more accurate representation of the potential field and of the geometry of
the boundaries. These higher order elements require the use of these singular integration
methods.
In the previous chapter, the boundary element method to solve potential problems has
been developed. In the present chapter, the numerical computation of the integrals giving
the components of the linear system has been studied. In the next one, the results of these
two chapters will be applied to solve some simple examples for which analytical solutions
are available. These examples will provide some interesting conclusions concerning the
numerical aspects of the boundary element method.
 CHAPTER 5
NUMERICAL ASPECTS OF THE BOUNDARY ELEMENT METHOD

5.1. INTRODUCTION

In order to check the accuracy of the boundary element method, the method will be
applied to several examples for which analytical solutions are available. In spite of the
importance of these comparisons when developing new software, only few examples
comparing numerical results with analytical solutions have been presented in the literature
[7], [41], [64], [12], [75], [55], [15]. The aim of the present chapter is the study of
some numerical aspects of the boundary element method, like the comparison between the
singular and the Gauss integration method, the choice of the number of integration points
and the discretization of the boundaries.

5.2. COMPARISON OF INTEGRATION METIIODS

In this paragraph, it will be shown that the use of singular integration methods is essential
to obtain accurate results. In order to bring this to the fore, the simple problem of heat
propagation shown in figure 5.1 will be studied first In this case, the domain is a square
(d = 6 cm) and the potential is equal to the temperature (T in 0c) and the coefficients Kx
and JS. are the thermal diffusivity (Kx = Ky = 1 cm/h).

y=6 1.:
4
~ q=O
7 8
9

3 10
T=300 T=O
2 11

1,5 14 13 12
--~~~-<~~~~¢-----x
1 16 q=O x=6

Figure 5.1
 47

Each of the four boundaries of the domain is discretized with three linear two node
elements. At the corner, the double node is considered. Thus, the discretization of the
whole boundary requires 16 nodes and 12 elements.
The analytical solution describing this problem is given by:

T=300- SOX (5.1)

The present problem has been studied by several authors [7], [41], [64] who have
obtained results of similar accuracy. To compute the integrals giving the components of
the linear systems, those authors have used the Gauss integration method with 4
integration points per element. Table 5.1 compares the analytical solution (5.1) with the
results obtained by [7] and[64].

I'b:le Unknown Numerical Analytical

Solution Solution

9 q -50,13 -50,00

10 q -50,10 -50,00

13 T 2,07 0,00

14 T 99,97 100,00
15 T 200,0 200,00

16 T 298.1 300,00

Table 5.1

Figure 5.2 shows the logarithm of the relative error obtained for the potential at node 15
and the normal velocity at node 10 (see figure 5.1) in the following cases:
1) Gauss integration method used for all elements;
2) singular integration method applied if the collocation point is located on the
element under consideration and Gauss integration method in all other cases.
log(e)
- . - singular integration
-2 - + - Gauss integration
e: relative error
-4 NIP: number of integration
points
-6 a: node 15
b: node 10
-8
-10
-+__~__~__~__~~~__ NIP
2 4 6 8 10
Figure 5.2

Table 5.1 and figure 5.2 show that the singular integration method always gives more
accurate results than the regular integration method. Figure 5.2 in particular demonstrates
that singular integration with 2 integration points provides an accuracy of the same order
as the regular integration with 10 integration points. The figure also shows that the
 48

accuracy improves rapidly with the number of integration points if one uses the singular
integration method. This is due to the fact that the singular integration provides the terms
located near and on the principal diagonal of the system matrix. As these terms are in
absolute value the largest terms of the system, they have a great influence on the accuracy
of the solution.

5.3. STUDYOFlHEDISCRETIZATION

In order to bring to the fore the influence of the choice of element type and the number of
nodes on the results, a porous media flow problem has been studied. The figure 5.3
shows the geometry of the domain and the boundary conditions.
y

Figure 5.3
Assuming that the domain is isotropic (K = Kx =Ky = 1), one can obtain the analytical
solution as given below.
41nr
u= In5 +1 (5.2)

with: r = j X2 + y2

From this analytical solution one can deduce the pressure P on the faces BC and AD, as
well as the discharge Q through the whole domain, ie.

PBe =4 ( 3- In\) PAD 8

= (3-~)
In 5 Q =
21t K
In5
(5.3)

The four following figures are function of the number of nodes of the discretization (NN)
and the element type used and show the variation of the following terms:
-) the maximum relative error in the normal velocity along the boundary CD (figure
5.4);
-) the relative error in the discharge along the boundary CD (figure 5.5);
-) the maximum relative error in the potential along the boundary AD (figure 5.6);
-) the relative error in the pressure along the boundary AD (figure 5.7).
 49

In these figures, the curves denoted by (2), (3) and (4) correspond respectively to 2, 3
and 4 node elements.
These four figures lead to some interesting conclusions regarding the discretization.
Firstly, the accuracy of the solution improves appreciably when one uses higher order
elements. Indeed, as the potential varies non linearly (see equation (5.2» and as a part of
the boundaries are curved, higher order elements allow a more reliable discretization of
the potential field and the geometry of the boundaries. Figures 5.4 and 5.6 show that the
error in the normal velocities is appreciably more important than the error in the potential.
The figures also show that the accuracy of the results does not improve indefinitely if we
refine the discretization. Indeed, when the nodes get close, the equations of two
adjoining nodes become very similar. Thus the conditioning number of the linear system
increases and consequently, the accuracy of the solution decreases.

E(%)
E(%)

Figure 5.4 Figure 5.5

E(%)
E(%) 4

',(4)
'ir"-::~._ 0

...... :.::::'--.~.-. -. _(31-0_

-T~15~-2~O~-2~5~~30~~35~~4~0----NN ~--~~~~~~==~=-NN
25 30 35 40

Figure 5.6 Figure 5.7

 50

Finally, a detailed analyses of the results shows that the maximum errors are always
located at corners A, B, C and D (see figure 5.3)

5.4. SENSITIVITY OF THE RESULTS TO THE VARIATIONS OF GEOMElRY

When studying porous media flow, one has to determine the position of free surface by
on iterative procedure. Indeed, the position of the free surfaces is one of the problem
unknowns. Thus at each iteration, one corrects the position of the free surface nodes and
consequently, it is important to investigate the sensitivity of the potential and the normal
velocities to variations of the domain geometry. To study this, one can consider the
problem shown in figure 5.8 for which an analytical solution is available as given by the
following equations.

u=Y K= 1 qAD =-1 qBc=+1 (5.4)

y y
u =3 c _!c!..
3 ----------
B B
d

q 0 q =0

.1
1
d
D
X X
A u =0 3 A

Figure 5.8 Figure 5.9

d Case A CaseB CaseC

u q u q u q
0 0 0 0 0 0 0
0,05 0,0287 0,1609 0,0291 0,1882 0,0231 0,1765
0,10 0,0564 0,3215 0,0569 0,3781 0,0450 0,3514
0,20 0,1089 0,6351 0,1089 0,7513 0,0853 0,6857
0,30 0,1582 0,9276 0,1566 1,0986 0,1219 0,9853

Table 5.2
Maximum errors

To study the sensitivity of the results to variations of the geometry, one can give to the
four faces of the square different eccentricities d (see figure 5.9). The three following
cases have been considered here: -) discretization A: 16 two node elements;
-) discretization B: 8 three node elements;
-) discretization C: 4 four node elements.
 51

Table 5.2 shows the maximum errors in the potential (u) and the normal velocities (q) for
different eccentricities d.
Table 5.2 demonstrates clearly that the normal velocities are 5 to 10 times more sensitive
to boundary geometry variations than the potentials. The result of this is that along the
free surfaces, one should choose the potential as unknown, and not the velocities, to
ensure the better convergence and stability of the iterative procedure.

5.5. CONCLUSIONS

This chapter has demonstrated the following:

- that the use of singular integration method is of primary importance in order to obtain
accurate results;
- that higher order elements improve the accuracy of the results ;
- that the normal velocities are 5 to 10 times more sensitive than the potentials to
boundary geometry variations.
All these conclusions will be useful in the applications studied in the following chapters.
 CHAPTER 6

DIVISION OF THE DOMAIN INTO SUB-REGIONS

6.1. INTRODUCTION

The division of the domain into sub-regions is a very useful technique particularly in two
cases. First, if the domain is divided in several parts having different properties, and
second, if the domain has a complicated geometry. In this second case, it will give more
accurate results if one subdivides the domain into several sub-regions rather than
considers the whole domain without divisions. The sub-regions are separated by
interfaces along which both the potential (u) and the normal velocity (q) are unknown.
The conditions along these interfaces are given by the continuity conditions, ie.

and (6.1)

The division of the domain into several sub-regions leads (see figure 6.1) to special
numerical problems at the intersection points of several interfaces (point M) or at the
intersection of an interface and a boundary (points N, P and Q). Similar problems occur
at the junction point of two boundari~s.
Q

Figure 6.1

During the past years, several methods have been suggested to overcome the numerical
problems associated with these intersection points. All these methods are based on the
fact that these problems mainly affect the values of the normal velocities (q) along the
interfaces or boundaries near the intersection points. The methods modify in various ways
the description of the potential field near these points to improve the solution.
Before analysing these methods, the discretization of the interfaces and the assembling of
the whole linear system will be explained.

6.2. DISCRETIZATION OF THE INTERFACES

When the domain is divided in several sub-regions, each of them will have its own
reference system (x, y) corresponding to the principal directions associated with the
 53

characteristics of the tensor K of the sub-region under consideration. Hence, the interface
points have different coordinates in the two sub-regions to which they belong and the
interfaces have to be discretized twice.
Since the interfaces are discretized twice, to each interface point correspond two nodes
with their coordinates in the two adjoining sub-regions (see figure 6.2). Finally, the sub-
regions will be coupled by the continuity conditions (6.1).
At the junction point of two boundaries, the correct discretization of the geometry of the
boundary, the potential (u) and the normal velocity (q) imposes the use of the double node
technique [70]. At the junction point of several interfaces (see figure 6.3 and table 6.1) or
of an interface and a boundary (see figure 6.4 and table 6.2), the application of the double
node technique in each sub-region leads to a multiple node technique.

Figure 6.2

Ncxle Interface Sub-region

1 A-C A
2 A-B A
2 A-B B
4 B-C B
5 B-C C
6 A-C C

Figure 6.3 Table 6.1

Ncxle Situation Sub-region

1 Boundary A
2 Interface A
3 Interface B
4 Boundary B

Figure 6. 4 Table 6.2

 54

In order to detennine the reason for the numerical problems, the linear system formed by
the equations resulting from the collocation at the different nodes at the intersection points
will be analysed [15], [17], [19], [21]. Later, the same method will be applied to study
the numerical problems at the junction point of two boundaries [22].

6.3. THE LINEAR SYSTEM

In the chapter 3, it has been shown that the discretization of the boundary value problem
leads to a linear system of equation, ie.

( HeI).. + 8I).. c·1 ) u'J = GeI).. q.J

with: Hejj given by (2.32)

Gejj given by (2.33) (6.2)
Cjgiven by (2.34)
8ij =0 if i *j and Bij = 1 if i =j
This linear system is obtained by applying collocation at each node of the discretization.
Thus, one fmds the following linear system of equations.

Hu = Gq (6.3)

This system is reorganized with the unknowns grouped on the left hand side and the
boundary conditions on the right hand side.
In the present case, one obtains a linear system (6.3) for each sub-region. However,
these linear systems can not be solved since at each interface node, two conditions co-
exist. That is why one needs to group these linear systems to obtain one only system
describing the whole problem. To couple the different sub-regions, one introduces for
each interface node the continuity conditions (6.1) into the system.
In practice, one does not compute the linear system of the different sub-regions, but
computes directly the linear system of the whole problem. This is done by discretizing
twice the interfaces and by numbering continuously all the nodes of the discretization.
Thus, if NA and NB are numbers of the nodes A and B (see figure 6.1), the introduction
of the continuity conditions (6.1) into the linear system will lead to two unknowns at this
point. In practice, the unknown number NA will be the potential at this point and the
unknown NB the normal velocity to the interface at this point. Thus, the number of
unknowns of the problem will be equal to the number of nodes of the discretization.
After these transformations, one obtains a linear system which solution should provide
the values of the unknowns. However, in practice, some numerical problems distort the
accuracy of the solution. The purpose of this chapter is to determine the origin of these
numerical problems and to develop a method to overcome them. Since these numerical
problems are well known, different methods have been suggested during the past years to
overcome them. The following section gives a review of these methods and their
limitations.

6.4. EARLIER SOLUTIONS

During the past years, several methods have been suggested to overcome the numerical
problems present at the intersection points. These methods are based on the fact that these
problems mainly affect the values of the normal velocities (q) along the interfaces or
boundaries near the intersection points. The methods modify in various ways the
description of the potential field near those points to improve the solution.
 55

For sub-regions with isotropic characteristics, one of the methods uses the analytical
solution of the Laplace equation expressed in polar coordinates around the intersection
point of several interfaces. The analytical solution allows to develop special shape
functions [44] describing the potential field near the intersection points. Thus, the
elements along each interface or boundary that are in contact with the intersection point are
described by these special shape functions. The analytical solution allows to deduce a
continuity equation around the intersection point [30]. Furthermore, the analytical
solution shows that theoretically, the velocity (q) is either infinite or zero at the
intersection point. Thus, a single velocity vector may be imposed at this point and the
normal velocities to the different interfaces are obtained by projections. The method
requires a very fine mesh in the vicinity of the intersection point [72], [29], [30].
Furthermore, it gives accurate results only if the different sub-regions have similar
characteristics [12].
Another method to overcome these numerical problems removes the collocation points
from the intersection point by using special discontinuous shape functions. When these
discontinuous elements [64] are applied, an important point is the determination of the
optimal distance between the intersection point and the points where the new unknowns
are located [53]. Studies show that beside the location of the collocation points, the
numerical integrations have great influence on the accuracy of the solution. Although the
use of discontinuous elements improves the solution, it leads to an increase in the number
of nodes, and hence of the unknowns, especially in three-dimensional problems.
Another method can be used to overcome these numerical problems when all the sub-
regions have isotropic or the same anisotropic characteristics. When this method is
applied, at each node of the discretization, the fundamental equation of the boundary
element method (3.25) is not only writteft with regard to the SUb-region to which the node
under consideration belongs, but also with regard to all the other sub-regions adjacent to
the one under study. Then, it is possible to show that all the integrals along the interfaces
relative to the normal velocity (q) are equal to zero [76]. However, the application of this
method to sub-regions which have different characteristics leads to a poorly conditioned
linear system that affect the accuracy of the solution [19]. Indeed, the different authors
who have applied this method in such cases have obtained different results for the same
problem [76], [61], [19].
The methods briefly described above solve the numerical problems present at the
intersection points only in some particular cases. These methods also require a fine mesh
near the intersection points, and so they generally lead to an increase in the number of
unknowns to ensure the improvement of the solution.

6.5. ANALYSIS OF THE NUMERICAL PROBLEMS

6.5.1. Introduction

Nooe InteIface Sub-region

1 A-C A

2 A-B A

2 A-B B

4 B-C B

5 B-C C

6 A-C C

Figure 6.5 Table 6.3

 56
Although the methods presented in the previous section provide some forms of solution to
the problem. none of them explain the origin of the numerical problem. The purpose of
this section is to determine the actual reason for the numerical problem at the junction
points and to develop an appropriate solution. Thus. the three following cases will be
analysed: -) the junction point of several interfaces;
-) the junction point of an interface and a boundary;
-) the junction point of two boundaries.

6.5.2. Junction point of several interfaces

Notice that there are two unknowns along each interface. the potential (u) and the normal
velocity (q). and by using the double node technique. the intersection of N boundaries
leads to 2N equations at this point. belonging two by two to each N sUb-regions (Figure
6.4 and table 6.3).
To study the numerical problems arising at the junction point of several interfaces (the
case of three sub-regions will be studied explicitly). the system of the six equations
relating to six nodes will be discussed.
The necessary stages are:
a) grouping the unknowns on the left hand side of the equal sign and the boundary
conditions on the other side;
b) application of the continuity conditions (6.1);
c) assembling unknowns other thljn those at the junction point in the vector X. the
boundary conditions in the vector C, and the corresponding terms of the matrices G
and H in vectors F and S (the Einstein summation convention is used below).
These operations produce the foJlowhlg system of equations:

As all the nodes are located at the same point and they belong two by two to the same sub-
regions, some relations link the components of the matrices G and H together.
These relations are given by:

Hk,k = H k+1,k+l H k,k+l H k+1,k (a)

Gk,k G k+1,k Gk+1,k+l G k,k+l (b)
= (6.5)
Hk,p

Gk,p
H k+1,p

Gk+1,p
} => { Fk,jXj

Sk,Fj
F k + 1,jXj

Sk+l,Fj
(c)

with:k = 1.3 and 5

p = 7 •...• NN
NN : number of nodes of the discretization

After the introduction of the relations (6.5) into the linear system (6.4). and substracting
term by term equations (2) and (I) of the linear system, (4) and (3). and (6) and (5). one
obtains the following linear system.
 57

(H21 - Hn) Ul/6 + (H22 - H 12) U2/3 =0

{ (H43 - H 33 ) U 2/3 + (H44 - H 34 ) U4/S =0 (6.6)
(H66 - Hs6) UI / 6 + (H6S - ass) u4/ S =0

The combination of the first relation (6.5) with the linear system (6.6) yields :

ul/6 = u2/3 = 114/S = u (6.7)

Thus, one obtains the logical result that all the potentials at the junction point are identical.
If one introduces the relation (6.7) into the linear system (6.4), one immediately notices
that the equations belonging two by two to the same sub-regions are identical and so; the
linear system becomes theoretically singular. In practice however, the rounding errors
and the approximations inherent to the numerical integrations lead to a system that is not
singular, but poorly conditioned. This is the origin of the numerical problems often
noticed in the past.
As the equations are identical two by two, one needs to eliminate one of them in each sub-
region at the intersection point. Thus, one obtains the following system.

(Hll + H12)u - G 12q2/3 - G n ql/6 + FAliXi SA1jq =

{ (H33 + H )u +G q2/3 - G q3/4 + FB 3iXj = SB3jCj
34 33 34 (6.8)
<Hss + HS~u + GSS q3/4 + G S6Ql/6 + pCSiXj = SCSjCj

The linear system (6.8) is composed of three equations, but contains four unknowns at
the intersection point: the potential at this point and the normal velocities along each
interface. So, one needs to add an additional equation to the linear system be able to solve
it. The additional equation will be discussed in section 6.6.
The case of three sub-regions concurring at one point has been considered here, although
it is not difficult to extent this analysis to any other number N of sub-regions (N ~ 2), the
interfaces of which intersect at one point. Independently of the number N of sub-regions,
one arrives to similar conclusions as presented above.

6.5.3. Junction point at intersection of an interface and a boundmy

In this section, the case of an intersection point between an interface and a boundary will
be studied by applying the same method as at the previous section. The three following
cases can occur (see figure 6.6 and table 6.4):

Node Situation Sub-region

1 Boundary A
2 Interface A
3 Interface B

4 Boundary B

Figure 6.6 Table 6.4

 58

-)casel: PMandPN: u=up (Dirichlet conditions)

-) case 2: PM: q = qpl and PN: q = qp4 (Neumann conditions)
-) case 3: PM: u = Up and PN: q=qp4

In the present case, one obtains four unknowns at the junction point. To study the
numerical problem arising at the junction point of an interface and a boundary, one locates
a collocation point i at each of these four nodes. Thus, one finds the following system
that is similar to (6.4), ie.

Huul - Gnql + H12u2/3 - G12q2/3 + P\iXi =S\Fj

H21u1 - G21 qI + H22u2/3 - G22q2/3 + pA2i~ = SA2jCj (6.9)
H33 u2/3 + G33q2/3 + H34u4 - G34~ + pB3iXi =SB3Fj
H43 u2/3 + G43 q2/3 + H44u4 - G44~ + pB4iXi =SB4Fj

As before, all the nodes are located at the same point P and belong two by two to the same
sub-regions A and B. It follows from this that the components of the matrices H and G
of these four equations are linked by the relations (6.5), with k equal to 1 and 3 and p =
5, ...... ,NN. After the introduction of these relations into the linear system (6.9) and
subtracting terms by terms the equati9ns located two by two in the same sub-regions, one
obtains again the identity of the potentials at the intersection point P, ie.

(6.10)

If one introduces the relation (6.10) into,the linear system (6.9), one immediately notices
that the equations belonging two by two to the same SUb-regions are identical and so, the
linear system becomes singular again. Thus, at the intersection point, one has to eliminate
one equation in each two sub-regions. After that, one obtains the following system, ie.

(Hll + H12)u - Gllql - G 12q2!3 + F\iXi = S\Fj

{ (6.11)
(H33 + H34)u + G 33q2!3 - G34~ + FB3iXi = SB3jq
To continue this analysis, one has to consider the different combinations of boundary
conditions near the intersection point P. For the three cases mentioned above, the linear
system (6.11) yields:

-) PM and PN : u = up

{
- G U ql - G 12q2/3 + F;liXi = -(Hll + H 12)t;, + SA1Fj
(6.12)
G 33q2!3 - G34~ + F 3iX i = -(H33 + H 34)up + SB3Fi

-) PM: q = qpl and PN: q = qp4

(Hll + H 12)u - G 12q 2!3 + FAliXi = Gllqpl + S\Fj

{ (6.13)
(H33 + H34)u + G33 q 2!3 + FB 3iXi = G 34qp4 + SB 3Fj

-) PM: u = up and PN: q = qp4

- Gllql - G 12q2!3 + FAliXi = -(Hll + H 12)t;, + SAI·C.

{ B J J (6.14)
G33q2!3 + F 3i~ = -(H33 + H 34)up + G 34 qp4 + SB3Fj

In the cases 2 and 3, the whole system remains regular. On the other hand, in the first
case, the system (6.12) is composed of two equations at the intersection point, but they
contain three unknowns there. So, the first case requires an additional equation to restore
 59

the system to its regular form. The choice of the additional equation will be considered in
the next section.

6.5.4. Junction point of two boundaries

In this section, the case of an intersection point of two boundaries will be studied by
applying the same method as in the two previous sections. The following three cases can
occur (see figure 6.7).

Figure 6.7

-) case 1: PM and PN : u = up (Dirichlet conditions)

-) case 2: PM: q = qpl and PN: q = qp2 (Neumann conditions)
-)case3: PM: u = up and PN: q =qp2

In the present case, one obtains two unknowns at the junction point. To study the
numerical problems at this junction point, one can locate the collocation point i
successively at the two nodes. Thus, one obtains a linear system which is similar to (6.4)
and (6.9), ie.

{
HU ul - GIl ql + H12 u2 - G 12 qz + FH Xl = Slj q
(6.15)
H2l ul - G2l ql + H22 u2 - G22 qz + F2i Xi = S2j Cj

The linear system (6.15) was obtained after grouping the unknowns other than those
corresponding to the junction point in the vector X, the boundary conditions in the vector
C and the corresponding terms of the matrices G and H in the vectors F and S. Finally,
the indices 1 and 2 refer to the corresponding node numbers (see figure 6.7).
To analyse the numerical problems at this junction point, one uses the same method as at
the intersection point of several interfaces or of an interface and a boundary, in the case
where the domain is divided into several sub-regions. As the two nodes are located at the
same point, the components of (6.15) are linked by the following relationships.

HU =H22 G11 = 021 012 = G22

(6.16)

with: i = 3, ......, NN
j = 3, ...... , NN
NN: number of nodes of the discretization.

After the5ntroduction of relations (6.16) into the linear system (6.15) and subtracting term
by term the two equations, one obtains the following identity of the potentials at the
junction point P, ie.
(6.17)

If one introduces the identity (6.17) into the linear system (6.15) and if one takes
relationship (6.16) into account, one immediately notices that the two equations become
identical and consequently, that the whole system becomes singular. However in
practice, the rounding errors and the approximations inherent in the numerical integrations
lead to a linear system that is not singular, but poorly conditioned. This is the origin of
the numerical problems often noticed in the past. Thus, one needs to suppress one of
these equations and one obtains the following equation at the junction point:
 60

(6.18)

To continue this analysis, one needs to consider the different combinations of boundary
conditions near the junction point.

a) Dirichlet COnditions
If Dirichlet conditions are imposed along the two boundaries PM and PN, the equation
(6.18) gives:
(6.19)

Here, the only remaining equation at the junction point contains two unknowns at the
point P, the two normal velocities ql and q2' Hence, one needs to add an additional
equation to the system in order to restore it to regular. The choice of the additional
equation will be discussed in the following section.

b) Neumann COnditions
If Neumann conditions are imposed along the two boundaries PM and PN, the equation
(6.18) gives:

(6.20)

In the present case, the only remaining equation at the junction point contains only one
unknown at the point P, the potential,u at this point, and so the system does not require an
additional equation.

c) Mixed Boundmy COnditions

If mixed boundary conditions are imposed along the two boundaries PM and PN, the
equation (6.18) gives:

(6.21)

In this case, the only remaining equation at the junction point contains only one unknown
at the point P, ie the normal velocity along the boundary PN, and so, the system also does
not require an additional equation.

6.6. TIffiADDmONALEQUATION

As has been seen at the previous paragraph, certain cases require an additional equation to
the linear system to restore its regular character. The choice of this additional equation is
important, because it has great influence on the accuracy of the solution. For this reason,
it will be proposed to use the basic relation of the boundary element method (3.25) as
additional equation. Indeed, for a collocation point i located outside the domain or outside
the sub-region under consideration, the ci coefficient is equal to zero (see point 3.5.3).
Thus, one obtains an equation that contains no additional unknowns and this equation can
be added to the linear system to restore it to regular. This equation is:

(6.22)
 61

Figure 6.8

In the present case, the boundary r c is the boundary of a sub-region of the domain. One
has to locate the additional collocation point outside the considered sub-region. For the
case in figure 6.8, if the points M and N require additional equations and if one wants to
write them with respect to the sub-region A, the collocation points can be at I or J for the
point M and at K or L for the point N. One will use the equation (6.22) as an additional
equation at the junction point of several interfaces or at an interface and a boundary.
At the junction point of two boundaries, one could use the same equation (6.22) as
additional equation and this method has been used successfully for two dimensional
problems [55]. However, in the case of three dimensional problems, the results given by
this method were not very accurate [35]. In the present case, the boundary conditions and
the geometry of the boundaries near the junction point impose some relations between the
two unknown normal velocities at this point Thus, the additional equation will vary from
problem to problem and depends on these relations.

6.7. SUMMARY OF THE THEORY

In the previous section, the numerical problems occurring at the junction point of several
interfaces, at the junction point of an interface and a boundary and at the junction point of
two boundaries have been studied. The study has shown that these numerical problems
occur because the linear system becomes theoretically singular at these points. Hence, to
overcome these problems, the following procedure has been developed in this book:

a. Intersection point between interfaces:

+ The identity of the potentials at the intersection point is explicitly imposed.
+ At the intersection point, one equation is suppressed in each sub-region.
+ The additional equation is introduced into the linear system.
b. Intersection point between an interface and a boundary:
+ The identity of the potentials at the intersection point is explicitly imposed
+ At the intersection point, one equation is suppressed in each of the two sub- region.
+ If a Dirichlet condition is imposed along the boundaries of the two sub-regions at
the intersection point, the additional equation is introduced into the linear system.
c. Intersection point of two boundaries:
+ The identity of the potentials is explicitly imposed at the two nodes.
+ One of the two equations is suppressed at the junction point.
+ If Dirichlet conditions are imposed along the two boundaries at the junction point,
an additional equation is introduced into the linear system to restore it to regular.

This method has been applied to many examples a few of which are presented in the
following section. From these examples and other numerical tests, the following
conclusions about the method have been deduced.
Concerning the optimal position of the collocation points, it is concluded that if they are
located too close to the boundary of the sub-region under consideration, the errors
become important due to the inaccuracies of the Gauss integration method and if they are
 62
moved too far away from the point that requires the additional equation, the errors also
increase because the additional equation does not describe properly the potential field near
the point. However, even if the optimum position for the collocation points is not
chosen, the resulting errors only affect the solution near the point requiring the additional
equation.
For problems with several points requiring additional equations, one can conclude that a
maximum of them have to be written with regard to a sub-region that is next to a
maximum of other sub-regions of the problem under consideration.
Finally, in comparison to other methods, the ones presented here has several advantages.
For instance, at the intersection point of N interfaces, the discontinuous elements lead to
2N equations. On the other hand, the present method leads to only N + 1 equations at
this point. This point is especially important for three-dimensional problems for which
the new method is easily generalizable.

6.8. APPLICATIONS

6.8.1. Introduction

In this paragraph, the method developed above will be applied to two examples. The first
example will compare the new method with an analytical solution. After that, another
example will be considered to show not only the effectiveness of this method, but also to
deduce some rules concerning the optimum position of the collocation point for the
additional equation.

6.8.2. Example with an analytical solution

First, the method shall be applied to an example with an analytical solution [15], [17].
The problem is described by figure 6.9 and the corresponding analytical solution is given
by the following relations with Kx = Ky = 0.1 cm/s, ie.

u=Y qx = O. qy = 0.1 (6.23)

Figure 6.10 shows the division of the square in the three sub-regions and the
discretization of the boundaries and the interfaces with 30 quadratic elements. Thus, the
number of nodes of the discretization is equal to 73 and the application of the above
described method at the different intersection points leads to a linear system with 62
degrees of freedom. It is easy to see that one obtains a reduction of 15% in comparison to
the same discretization with discontinuous elements.

y y y
u=lD
0
R3
p B xp p
q=O 6U=0 q=O SA
Yp
R1 R2 Yc
C X Xc
u=O x s 10 c x
Figure 6.9 Figure 6.10 Figure 6.11

In accordance with the conclusions of the previous part of this chapter, one has to write
two additional equations, one for the point P and another for the point C (figure 6.10).
These two equations are written with respect to the sub-regions Rl and R2 and their
collocation points are located respectively in the sub-regions R2 and Rl, where they are
referenced by the coordinate systems (xp, yp) and (xc, Yc) of figure 6.11.
 63

The figures 6.12, 6.13 and 6.14 give the maximum (M) and the mean (m) error (e, the
errors multiplied by 109) in the computed values of the potential (u) and the normal
velocities (q) with respect to the analytical solution (6.23), and according to the position
(Xi and Yi) of the two points where the additional equations are written. The maximum
error occurs at points P and C where one has to write the additional equations.
The three figures show that one obtains a good accuracy for a large range of variation of
the positions of the collocation points of the additional equations. These results
demonstrates that the cancellation of certain equations, which are theoretically identical
after setting the identity of the potentials at these points, results in a good conditioning of
the linear system and that the selected additional equation are well suited to the linear
system and contribute to this excellent accuracy.
The analysis of the results in these three figures show not only that the values of the
normal velocities are distinctly more sensitive than those of the potentials to the variations
of the position of the collocation points corresponding to the additional equations, but also
that the relative errors in the normal velocities are about a hundred times greater than those
in the potentials.
70 70 70

l/
60 60 60
so so
40 40 40
30 30 30
20 20
10 10
4
0 0 0
1,0 1,5 2,0 2,5 3,0 1,0 l,S 2,0 2,5 3,0 1,0 1,5 2,0 2.5

Figure 6.12 Figure 6.13 Figure 6.14

(Yi =0.625) (Yi =1.250) (Yi =1.875)

where: 1: Maximum error (multiplied by 1(9) in the velocity

2: Mean error (multiplied by 1(9) in the velocity
3 : Maximum error (multiplied by 1(9) in the potential
4 : Mean error (multiplied by 109) in the potential
The previous remarks regarding the optimum position of the collocation points are also
confinned by these figures,

6.8.3. Symmetrical problem

Here the first example [15]. [17] for sub-regions having different characteristics, like
those given at the table 6.5 where the angle corresponds to ex of the figure 6.17 will be
discussed.

Sub-region kxx kyy kxy kx ky angle

R1 0.2 0.4 0.1732 0.1 0.5 60'
R2 0.2 0.4 -0.1732 0.1 0.5 120'
R3 0.3 0.3 0 0.3 0.3 O·

Properties of the different sub-regions

Table 6.5
 64

In the present case, one uses for the discretization of the interfaces and the boundaries on
which Dirichlet conditions are imposed 4 node elements and along the boundaries on
which Neumann conditions are imposed 3 node elements. Thus, the boundaries and the
interfaces were discretized with the help of 86 nodes and the application of the present
method leads to a linear system with 75 degrees of freedom.

y jM y y
u=10 u=10
I
I R3
R3
q= 0 i q= o
q=O q=O
RI R2 Y
Rl
X
u=O X
1M u=O
!
Figure 6.15 Figure 6.16 Figure 6.17

As shown in figure 6.15 and table 6.5, this problem is symmetrical (geometry, boundary
conditions and properties of the sub-regions) with regard to the axis MM' (figure 6.15).
The result of this is that the potentials and the normal velocities also will be distributed
symmetrically with respect to the axis MM', and that along this axis, the normal velocity
will be equal to zero. It follows from this that one can also compute this problem
considering only half the domain, with a Neumann boundary condition (q = 0) along the
axis MM'. Figure 6.16 shows half the domain with its boundary conditions and its two
sub-regions the characteristics of which are given in table 6.5. In accordance with the
conclusions of the theoretical part of this chapter, the computation of the half domain
problem does not require any additional equations, contrary to the whole domain where
they are essential at the points P and C (figure 6.17).
In the present case, one obtains similar conclusions as for the case where all the sub-
regions have the same characteristics. Furthermore, the comparison of the half-domain
and the whole domain results confirm that bad location for the collocation points only
affect the solution near the points P and C. In the present case, figure 6.18 shows that the
errors become rapidly negligible at the other nodes. Indeed, the collocation points of the
additional equations (Xi = 0,750 I Yi = 2,083) are located too far away from the
intersection points requiring the additional equations and too close to the sub-regions with
respect to which the additional equations are written.

V(c:m!s)

D p
1.0 -- -- half domain
-Xi.0.7S Y, .2.083

A
c
x

Errors in the velocities (Xi = 0.750 I Yi = 2.083)

Figure 6.18
 65

Figure 6.19 shows the equipotentials and the streamlines for half the domain [15]. This
figure clearly demonstrates that for anisotropic case, the velocities are not perpendicular to
the equipotential lines.

9 8 7

u.l0

9 8 7

Equipotentials and Streamlines.

F~gure 6.19.

Finally, to conclude this example, figure 6.20 gives the maximum absolute error in the
velocities according to the position of the points where the additional equations are
written.

error
-.-.... -- y. 0.~167
- - - Y. 0.8333
0.6 - - Y. 1.250
.: , _.-. y. 1.667
!: 1 .: ------ Y.2.083 :
I:
.,
I
I :: J:
0.5 111 '
I" :! '
j!: :; "
0.4 i! :h !I ; i
·If.'/1I i:,. ~"r\
#' • , "
I
0.3 V: rUr\,
u, :'
[: I : ,; i
1: I :j' 1\ :
.: I ,~ \ I
\: I ,': \ '
0.2 I I • \ "
\ I: , ; \ "
[·-..ri·-·}.. \ \ '
:: I:
I: II \ ~," , .,'"
0.1 : \.\ ' .....)/...... .
".... .,.'" '" ,<..:.-::::......
",:-~.-, /'
" .. ..
2. 3. 4. s. X

Maximal errors in the velocities.

Figure 6.20.
 66

For the first example, one know the analytical solution of the potential field which was a
simple solution. For the second example, one obtains a very complicated potential field
with highly varying velocities (see figures 6.18 and 6.19 corresponding to the half
domain problem). One can use similar discretizations in these two cases and that is why
the accuracy is lower in the second one. But, it is interesting to notice that if important
errors appear, they remain concentrated at the points needing the additional equations and
do not affect the values at the other nodes appreciably (figure 6.18).
Figure 6.20 allows to conclude, regarding the optimization of the position of the
collocation point where the additional equation is written, that:
- if this point is located too close to the boundary of the concerned sub-region, the
errors become important due to the unavoidable inaccuracies of the numerical integrations.
Moreover, the bad conditioning of the linear system causes important and strong
variations of these errors.
- if this point moves too far away from the point that requires this additional equation,
the errors increase also because the compatibility equation does not describe properly the
potential field in the vicinity of the point under consideration.
Finally, figure 6.20 shows that after the strong variations of the errors for small values of
X, the errors have a parabolical behaviour for greater values of X.

6.8.4. Conclusion

The purpose of the two examples presented in this section was to show that the new
method provides an effective solution to the numerical problems at the junction point of
several interfaces and at the junction p'oint of an interface and a boundary. Similar
problems occur at the junction points of two boundaries where one can use the double
node technique. The two examples considered here demonstrate such properties and the
possibility of obtaining accurate results as well. In the following chapters, other
examples containing such points will be considered and it will be shown that this method
provides accurate results in those cases.
Beside the two examples presented here and those in the following chapters, the new
method has been applied to several other problems and good results have been obtained
which allow to conclude that the simple technique developed in this chapter is an effective
solution to the numerical problems caused by the use of the double and multiple node
technique at junction points. Furthermore, the method does not have any limitation like
those presented in paragraph 4 of this chapter and reduces the number of unknowns of the
problem. This last conclusion is especially important for three dimensional problems for
which this method can be generalized. Indeed, in this case, this method applies to the
edges of the boundaries, and consequently to an important part of the nodes of the
discretization.

6.9. SPECIAL INTERFACES

6.9.1. Introduction

Up to now, the interfaces separating two sub-regions have been considered as ideal
surfaces along which one can impose the continuity of the potential and the normal
velocity. However, in practice, these interfaces have often specific properties that differ
appreciably from those of the two adjoining sub-regions. So, the characteristics of these
interfaces have to be introduced into the discretization of the problem.
In practice, these interfaces are very thin layers. Thus, their discretization by finite
elements or as additional sub-regions will lead to a bad conditioned system. Indeed, their
thin shape.leads to problems when performing the numerical integrations. To introduce
these thin layers into the software, one can use special continuity conditions along the
interfaces taking into account the properties of the layers.
 67
6.9.2. Discretization of thin layers

When the thickness of a thin layer is negligible with respect to the size of the whole
domain, one can discretize it like a classical interface by neglecting its thickness. In this
case, one needs to impose special conditions along this new interface to take the
properties of this layer into account (see figure 6.21), ie.

(6.24)

The coefficient A. depends on the properties (hydraulic conductivity, thickness, ..•) of the
layer and its value has to be determined experimentally. If the coefficient A. is equal to
zero, one retrieves again the classical interface.

6.9.3. Intersection point of several interfaces

As there &re two unknowns along each interface, the potential and the normal velocity, and
as the multiple node technique is being used, the intersection of N interfaces leads to 2N
equations at this point, belonging two by two to each N sub-regions and to the N
interfaces. Figure 6.21 and table 6.6 show this for 3 interfaces concurring at one point.

Node Interlik:e Sub-region

1 A-C A
2 A-B A
2 A-B B
4 B-C B
5 B-C C
6 A-C C

Figure 6.21 Table 6.6

In previous sections of this chapter, the numerical problems at the junction point of one
classical interface (A. =0) or of an interface and a boundary have been studied. In these
cases, it has been concluded that the linear system becomes singular and a method has
been developed to overcome this singularity.
In the present case (A. ¢ 0), the same method has been applied to analyse the behaviour of
the linear system at these junction points and it has been concluded that, the system does
not become singular. Notice that one always uses multiple nodes at the junction points but
that they do not require any special treatment in the present case.

6.9.4. The linear system

The relations (6.24) show that three unknowns exist along each interface : the two
potentials (uA and uB) and the normal velocity to the interface (qAB)' To solve the linear
system, one has to eliminate one of these unknowns by expressing it as a function of the
two other ones.
In the present case, one can choose to keep the two potentials as unknowns. Indeed, it
has been shown [19] that the normal velocities are much more sensitive than the potentials
to numerical inaccuracies and discretization effects. Thus, after the solving of the linear
system, the values of the normal velocities along the interfaces are computed with the help
of the first equation (6.24).
 68

6.9.5. Example with an analytical solution

Figure 6.22 shows the geometry of the domain that is divided into two parts by an
interface [23]. The properties of the two sub-regions and of the interface are given by:

AAB = 10 (6.25)

Considering the geometry of the domain and the characteristics of the two sub-regions
and of the interface, one obtains the following analytical solutions for the potential field u
and the velocities q.

UB = 5+ Y/2
qy = 0.5 (6.26)

16 three node elements with a total of 40 nodes were used for the discretization of this
problem. The application of the new method for double and multiple nodes reduces the
number of unknowns of the problem from 40 to 32.

y
10 u=10

B
q=O
5
q=O A
X
u=O

Figure 6.22

The analysis of the results and their comparison with the analytical solutions (6.26) lead
to the following conclusions:
maximum absolute error in the potential: 5,44.10-5
maximum absolute error in the normal velocity: 4,89.10-7
maximum relative error in the potential: 8, 46.10-5
maximum relative error in the normal velocity: 9,78.10-5

The analysis of the results also shows that these maximum errors do not occur along the
interface, but on the boundaries of the domain. Thus, in the case where an analytical
solution is available, one corroborates that the boundary element method is well adapted
to solve such interface problems.

6.9.6. Second Example

Figure 6.23 shows the geometry of a domain divided into three sub-regions [23]. The
characteristics of the three sub-regions and the three interfaces are given by:

AAC =100 ABC =50

KA = 0,1 KC = 1,0 (6.27)

For the discretization of this problem, 29 three node elements with a total of 71 nodes
have been used. The application of the method for double and multiple nodes reduces the
number of unknowns of the problem from 71 to 58. Figure 6.24 shows the potential
field and figure 6.25 the normal velocities to the boundaries and the interfaces of the three
sub-regions.
 69

Y Y
u=10 10
9,9
q=O 9,8
9,7
9,6
9,5

X
1
X
u=O 5 6 10 0 0
Figure 6.23 Figure 6.24

0.0026

0, 131

Figure 6.25
Although in the present case, no other solution is available for comparison with the
presented results, the discharges through the different boundaries and interfaces are
coherent with the geometry of the different sub-regions, their characteristics and those of
the interfaces.

6.9.7. Conclusions
Classical potential problems requiring the division of their domain into sub-regions
presents special numerical problems at the junction point of several interfaces. These
numerical problems occur because the whole linear system becomes singular in those
cases. To avoid this singularity, a new method has been developed which has the
additional property to reduce the number of unknowns of the problem.
In this section, the analysis has been extended to the case where the interfaces have
specific properties that differ appreciably from those of the sub-regions that they separate.
To investigate this new problem, one can use the same technique as had been developed
for sub-regions separated by ideal interfaces (A = 0). Thus, one can conclude that the
linear system does not become singular and that the solution of these problems does not
present any special numerical problems.
Contrary to the ideal interfaces, one has in the present case three unknowns along the
interface. To solve the linear system, one can express one of them as a function of the
other two. Thus, two unknowns are considered at each interface node and the value of
the third unknown is given by the fIrst relation (6.25) after solving the system. Based on
previous experience, one can keep the two potentials as unknowns to solve the linear
system and deduce the normal velocity from their values.
 70

The generalized interrace conditions (6.25) have been incorporated into the software and
applied to solve two different problems. The first is a very simple case for which an
analytical solution is available and the results were highly accurate. The second example,
which is more complicated, shows that the method also provides good results in that case.

6.10. CONCLUSIONS

In this chapter, it has been shown that the application of the double node technique and
the multiple node technique can lead to singular systems in the three following cases:
- the intersection point of several intexfaces ;
- the intersection point of an interrace and a boundary ;
- the intersection point of two boundaries.
After the identification of the origins of these numerical problems, a new method has been
developed to restore the conditioning of the system. This method is divided into four
steps:
- use the double or multiple node technique at every intersection point;
- impose in the linear system the identity of the potentials at the intersection points;
- suppress in each sub-region one of the two equations at the intersection points (because
they have become identical) ;
- add an addition'al equation to the'linear system in the cases where the number of
unknowns is higher than the number of the equations at the junction point.
The two examples presented in this chapter show that the new method provides good
results and one can conclude from them some recommendations concerning the optimum
position of the additional collocation point. It has also been shown that the errors due to a
bad choice of the position of the additional collocation point only affect the solution at the
point requiring the additional equation.
These conclusions concern interraces that do not have specific properties. In this case,
the linear system does not become singular because the junction point will have different
potentials in the sub-regions meeting there.
This method has been developed for two dimensional problems, but can be easily
generalized to three dimensional problems. In those cases, it will lead to an important
reduction in the number of unknowns of the problem. Finally, this method is also
applicable to problems described by equations other than Laplace's.
 CHAPTER 7

STEADY STATE FLOW THROUGH POROUS MEDIA

7.1. INTRODUCTION

In this chapter, the boundary element method is applied to solve steady porous media
flows described by the Laplace equation. In the present case, the potential u = Y + pIp has
to satisfy the Laplace equation. The potential contains two terms of which the fIrst
represents the influence of the gravity and the second the effect of the water pressure. In
some special cases, the influence of the pressure is negligible. This case will be analysed
in a later chapter.
This chapter will be devoted to the study of groundwater flows where the pressure has a
non negligible contribution to the potential, except along the possible free surface. To
study theses types of porous media flows, one needs to distinguish between confined and
non- confined flows. The confIned flows do not present free surfaces and are of the same
type as the cases analysed in the chapters dealing with numerical problems and the
division of domains into sub-regions. The examples presented there illustrate that the
boundary element method applies favorably to this type of problems.
To stress the application of the boundary element method to solving steady state porous
media flows with free surfaces (non-confined flow), the method will be applied to several
examples for which analytical solutions are available. These examples also show that the
boundary element method provides accurate results for the potential as well as the normal
velocities to the boundaries.

7.2. ITERATIVE DETERMINATION OF THE FREE SURFACEPOSmON

7.2.1. Boundary conditions

To solve the Laplace equation on a given domain Q, one has to impose conditions along
its boundary. The boundary conditions have been discussed in the fIrst chapter and for
free surface problem, one obtains the following relationships.

AB: u =H DE:u = h
BC: q = 0 and u =Y EA:q = 0 (7.1)
CD: u= Y

To start the numerical solution of this problem, one needs to discretize the boundaries of
the problems. This discretization applies also to the free surface the actual position of
which is one of the problem unknowns. Thus, one starts by choosing an initial position
of the free surface that is compatible with the other boundary conditions and which
 72

describe the expected flow. Now, since the actual position of the free surface is one of
the problem unknowns, two conditions exist along this boundary. One of these two
conditions (in practice the first one, q= 0) will be used to solve the linear system and the
other condition (u=Y) will be used to update the free surface position.
y

h x

Figure 7.1

7.2.2. Discretization of the free surface

If one uses for the discretization of the free surface, for instance, two noded elements, the
discretized free surface will not be flat at the junction point of two elements (see figure
7.2).

Figure 7.2

In this case, one has to use double nodes and to apply the conclusions of chapter 6. As
Neumann condition (q=O) is applied along the free surface to solve the linear system, one
obtains one equation at each of the junctions of two elements.
This equation is given by the following relationship:

(6.20)

In the present case, the boundary condition used to solve the linear system says that the
normal velocities along the free surface are equal to zero and that so, it is not necessary to
use the double node technique at the junction of two elements. In this case, a unique node
there leads to the same linear system as the double node technique.
In the next section, the required iterative procedure and convergence criteria will be
discussed.

7.2.3. Iterative procedure and convergence criteria

The iterative determination of the free surface position can be started from an initial
position (Xo, Yo) of the free surface nodes and the iterative procedure described below
can be applied.
- As already mentioned before, one can use the first boundary condition along the free
surface (q=O) to solve the linear system. Then, at each free surface node the potential
(Uj) can be obtained (j is the number of the iteration under consideration).
 73

- Thus, if after the previous iteration, the height at the node under consideration was Yj-l.
the condition (u =Y) along the free surface gives the new elevation for the node, ie.

(7.2)

Notice that equation (7.2) contains a weighting coefficient 0). It has been shown, for
underground water flows that this iterative procedure is always stable and converges if
0) is equal to 0,5. Hence, equation (7.2) can be written as:

(7.3)

After the correction of the free surface position, one has to check whether the
convergence has been reached. To do this, one can use a convergence criterion based
on the absolute or on the relative error, with eo as the imposed accuracy, ie.

absolute error: Maxj I Yj - Yj-1 I ::; Eo (7.4)

relative error: Maxj I [Yr Yj-1]/Yj I::;Eo (7.5)

For the applications in this chapter, the relative criteria (7.5) will be used. Other
convergence criteria have been proposed [61] but (7.5) is the most commonly used.
- If the convergence criteria is satisfied at all free surface node, convergence is reached
and the actual position of the free surface is obtained, with a precision depending on eo.
If the convergence criteria is not satisfied at all free surface node, the iteration procedure
continues with the updated free surface position.
At this point, the correction of the level of the free surface nodes (along the axis Y) during
the iterative procedure has been considered. In practise, the geometry of the problems can
also impose corrections of the position of the free surface nodes as will be seen in the next
section.

7.2.4. Correction of the position of the free surface nodes

The relationships (7.2) or (7.3) allow to update the height of the free surface nodes after
each iteration and to displace them vertically to find the new free surface position. But, as
figure 7.3.a shows, one can be confronted with problems near the junction points (B and
C) of the free surface with other boundaries.

y
I
I
I
:.---- --1
.......... I
" -rI
1,
I " "-
Y.J- 1 I ~ ...

A E

Figure 7.3.a Figure 7.3.b

 74

Indeed, if the free surface nodes are moved vertically, one takes the risk that the free
surface moves out of the domain or loses contact with the adjoining boundary. To avoid
these problems, one needs to update the position of each free surface node I along a line
IP, with P the intersection point of the two lines defining the adjoining boundaries AB
and CD of the free surface. Figure 7.3.b shows that when doing so, a vertical
displacement of the free surface also leads to a horizontal displacement of its nodes.
Practice has shown that this horizontal displacement does not affect the convergence and
the stability of the iterative procedure. In the special case of two parallel adjoining
boundaries AB and CD, the positions of the free surface nodes will be updated parallel to
them.
The conditions imposed along the adjoining boundaries of the free surface can also
require special treatments as will be seen in the next section.

7.2.5. Conditions alon& the adjoinin& boundaries

At this point, the influence on the iterative procedure of the conditions along the adjoining
boundaries to the free surface will be discussed.

----r-~--~--L---~------·x
A J I H

Figure 7.4

The figure 7.4 shows a dam divided into three sub-regions. The upstream part of the dam
(ABCJ), the core of the dam (CDEIJ) and the downstream part of the dam (EFGHI). In
practice, the hydraulic conductivity of the core of the dam is always lower than those of
the two other parts, to reduce as far as possible the discharge through a dam. Indeed, the
water seeping through a dam compromises its stability. The consequence of the low
hydraulic conductivity of the core is an inner seepage surface DE along the interface
separating the core and the downstream face of the dam. In each of the three sub-regions,
one fmds free surfaces and their behaviour near their junction points with other
boundaries or with interfaces needs to be analysed.
As only steady state flow conditions are considered in this chapter, the level of the point B
remains constant and will not be affected by the iteration of the free surface position.
Thus, the free surface does not require special treatment near such points.
On the other hand, the level of the points D,E and F that are intersection points of the free
surfaces (CD and EF) and the seepage surfaces (DE and FG) depend on the actual
position of the free surfaces. Thus, the positions of these points have to vary from one
iteration to another. However, along the seepage surfaces, the boundary condition (u= Y)
exists and hence, solving the linear system does not give the new potential Uj like at other
free surface nodes. Thus the levels of these points will not be corrected by the procedure
described in the previous section.
If one does not consider this special case, the correction of the free surface position (see
figures 7.4 and 7.5) will give a free surface MND. To avoid this problem, the new
position of the intersection point of the free surface and the seepage surface will be given
by the intersection point of the free surface with the line defined by the points M and N.
Thus, one obtains the new free surface position MND' (see figure 7.5).
 75

j
j+l

Figure 7.5

Finally, the point D is a free surface point in the two sub-regions to which it belongs.
Hence, one needs only to ensure that this point keeps the same level in the two sub-
regions.
Besides some special points of the free surfaces, inner seepage surfaces like DE (see
figure 7.4) also require a special treatment Since such inner seepage surfaces are because
one sub-region (the kernel) has a lower hydraulic conductivity than another (the
downstream part of the dam), the discharge QDE leaving the kernel creates an unsaturated
flow along the seepage surface in the downstream part of the dam. To reintroduce this
flow into the domain near the point E,one can use a vertical triangular recharge along the
frrst tree surface element downstream the point E (figure 7.6), with~' =QDE.

Figure 7.6

7.3. APPLICATIONS

7.3.1. Introduction

In the previous two chapters 5 and 6 concerning the numerical aspects of the boundary
element method and the division of the domain into sub-regions, several examples of
groundwater flow have been presented. These problems did not have a free surface and
consequently, they correspond to confined flow. Several other examples of confined
groundwater flows were also presented in previous publications [16] , [18] , [21] .
In what follows, two applications, one of them with an analytical solution, will be
described.

7.3.2. Example with an analytical solution

7.3.2.1. Introduction
 76

q/K A q/K
E
1,6 0,8 •l.o L 1
2,4 3,2
I
.. !

Figure 7.7

For phreatic groundwater flows, Muskat [57] gives an analytical solution for a rectangular
dam with isotropic hydraulic conductivity. Figure 7.7 shows the geometry of the dam and
the analytical solution for :
1- the normal velocity along the upstream face AB
2- the free surface position Be
3- the tangential velocity along the bottom AE
4- the pressure distribution along the bottom AE
5- the normal velocity along the seepage surface CD
6- the normal velocity along the downstream face DE
For the velocities, the dimensionalees variable qlK is used, with K the hydraulic
conductivity. In the present case, the following values for the hydraulic conductivity and
the geometrical dimensions of the dam are used:

H= 0,322 m L= 0,162 m h = 0,084 m K=O,l m/h

The boundary conditions of this problem are given by equation (7.1).

7.3.2.2. Discharge throu&h the dam

The total discharge through the dam is given by the following equation:

Q = K ( H2 - h2 ) /2 L (7.6)

This relation was obtained by applying the Dupuit assumption that the equipotential lines
are vertical. Thus, this assumption leads to a free surface given by the curve (4) of figure
7.7: Although the position of this free surface is different from its actual position (curve
(2) of the figure 7.7), this assumption leads to an analytical solution [66] of the discharge,
Q = 0,02982 m 3 / h.

7.3.2.3. Discretization and comers

To solve this problem by the boundary element method, different discretizations with a
number of nodes varying from 20 to 111 will be considered.
 77

Figure 7.6 shows an example which presents five points where the conclusions of chapter
6 concerning the junction points of two boundaries have to be taken into consideration.
These conclusions depend on the different conditions along the two concurring
boundaries. In the present case, these conditions are given by the following relationships:

AB: u=0,322 DE: u =0,084

BC : u = Y and q =
CD:u=Y
° AE:q=O (7.7)

Thus, one has :

- Along the bottom of the darn (AE), a Neumann condition is imposed and along the
upstream (AB) and the downstream (DE) face of the dam, Dirichlet conditions are
imposed. Thus, the points A and E do not require additional equations.
- Along the free surface, one has two boundary conditions of which the Neumann
condition is used to solve the linear system. Hence, the points B and C also do not
require additional equations.
- Finally, one needs to consider the point D [19] .
At this point, one has different Dirichlet boundary conditions along the seepage surface
CD and the downstream surface DE, and hence, the normal velocity to these boundaries is
infinite at D. Thus, one needs to introduce as an additional equation into the linear system
the identity of the two normal velocities at the point D, ie.

q =qD (CD) = qo (DE) (7.8)

It follows from this that equation (6.19) yields in the present case the following
relationship: '
(7.9)

Thus, there is now only one equation at the point D and one remaining unknown, and the
system is regular.
This example will show that the remaining equation (7.9) leads to accurate results.
Another solution to this problem was suggested by Liggett [47] who has developed an
analytical solution of the potential field near this point. He then proceeds to carry out an
analytical integration of the boundary integrals near this singular point.
Thus, the application of the concepts developed in chapter 6 to the points where the
double node technique is used reduce by five the number of equations of the linear
system.

7.3.2.4. Results of the computation

Case NN NDL NI T(sec) Y·(m) ~(m /sec) e (%)

A 2U 15 9 4,35 0,2133 0,02979 3,55
B 25 20 9 5,34 0,2132 0,02981 3,50
C 31 26 11 8,44 0,2111 0,02982 2,48
D 41 36 13 14,62 0,2085 0,02982 1,21
E 52 47 14 32,44 0,2072 0,02982 0,58
F 61 56 16 34,08 0,2065 0,02982 0,24
G 71 66 17 47,31 0,2059 0,02982 0,05
H 91 86 18 77,39 0,2057 0,02982 0,15
I 111 106 20 132,87 0,2054 0,02982 0,29

Table 7.1
 78

To solve this problem, nine different discretizations have been considered as shown in
table 7.1:
In this table, the following abbreviations are used:

NN : the number of nodes of the discretization;

NDL : the number of degrees of freedom of the linear system (NDL = NN - 5) ;
NI : the number of iterations to reach the convergence with CO =0.5 and £0 =0.00 1 ;
T (sec) : the CPU time in seconds used to solve the problem on an IBM 4341
computer;
YC (m) : the hight of the intersection point C of the free surface and the seepage
surface;
Q (m3/h) : the discharge through the darn ;
e (%) : the relative error of the level of the point C with the corresponding analytical
solution (yC =0,206 m) .

- - : analytical solution
----: numerical solution

1,6 0,8
L

- : analytical solution
• : numerical solution

H (5)

1.0

0.5 h

~~~~--~~--~----~----~~,~~~~~~~~~~~----~
K 1.6 12 0.8 0.4 f--_",-L_ _ -j 0.4 0.8 1.2 1.6 2.0 2.4 2.8 3.0 K

Figure 7.9
 79

The figures 7.8 and 7.9 compare the analytical solution with numerical results obtained
for the cases A and E. The numbers used on these figures correspond to those of figure
7.7.

7.3.2.5. Analysis of the results

Table 7.1 and figures 7.8 and 7.9 allow to draw the following conclusions:
- In the case A, one can obtain a linear system with only 15 unknowns. The
discretization corresponds to the minimal number of nodes to describe properly the
variation of the unknowns along the boundaries. Here, the errors along the
boundaries are not negligible. However, if one takes the number of nodes of the
discretization into account, the solution is good.
- In the case E [18], one obtains good agreement between the analytical solution of
Muskat [57] and the numerical solution. Table 7.1 also shows good agreement
between the computed discharge and using the analytical solution (7.6).
- The free surface requires a finer discretization than the other boundaries of the
domain to obtain a similar accuracy.
- An appropriate discretization provides good results for the potentials as well as for the
normal velocities. As shown earlier, the accuracy of the potential is always higher
than this of the normal velocities.
- Table 7.1 shows that the number of iterations to reach convergence increases
noticeably with the number of nodes of the discretization.
- Table 7.1 also shows that the accuracy of the solution does not improve continuously
with the number of nodes in the discretization. The accuracy even slightly decreases if
the discretization becomes very refined.
- The discharge through the domain does not show significant variations with the change
of discretization.
- The junction point D of the seepage surface and the downstream surface is well
represented by the method presented here.
- The choice of different types of elements (2, 3 and 4 nodes) to discretize the free
surface has no significant influence on the solution.

7.3.2.6. Conclusion

The application of the boundary element method to the present free surface problem for
which an analytical solution is available demonstrates that the method is well adapted to
solve this type of problem.

7.3.3. Zoned dam

7.3.3.1. Introduction

In the second example, the flow through a zoned dam will be studied. In the first
application, the dam will be considered to be built on an anisotropic foundation [17] as
shown in figure 7.10. Afterwards, the dam will be built on an impervious foundation [12]
as presented in figure 7.11.
Table 7.2 gives the hydraulic conductivity of the different sub-regions the numbers of
which are defined in the figure 7.10.
Table 7.2 shows that the foundation presents two regions with different anisotropic
characteristics, that the upper and the lower part of the dam have the same characteristics
and that the hydraulic conductivity of the core of the dam is a hundred times lower than
those of the two other parts of the dam.
 80

Figure 7.10

Figure 7.11

Sub-region Kxx Kxy Kyy

1 0,19 0,0375 0,06
2 0,1 0 0,1
3 0,001 0 0,001
4 0,1 0,01
5 0,1 ° 0,1
°
Table 7.2

7.3.3.2. Analysis of the first case

The purpose of the first case is to show that the method can deal with sub-regions in the
case of a complex example. For the discretization of the boundaries and the interfaces,
quadratic and cubic elements with a total of 154 nodes have been used. The application of
the method developed in the previous chapter leads to a reduction of the number of
 81

unknowns of the problem. In the present case, a linear system of 130 degrees of freedom
is obtained.
The application of the relationships concerning the different junction points between
interfaces and interfaces with boundaries, produces additional equations for the six points
A,B,C,D,E, and F of figure 7.10. The points corresponding to the additional equations
for these six points are also shown in this figure and they are noted A', B', C', D', E'
and F'. Table 7.3 gives the coordinates of all these points and the number of the sub-
regions of the additional equations.
In the present case, the hydraulic conductivity of the core of the dam is about hundred
times lower than the hydraulic conductivity of its upstream and downstream parts. Hence,
an inner seepage surface appears along the interface separating the core and the
downstream part. To find the actual position of the free surface, one can use the iterative
procedure and the convergence criterion described by (7.2) and (7.5) :

m= 0.5
In the present case, one obtains the convergence after 15 iterations. The results for this
first case will be compared with those obtained for the second case.

Intersection Point Source Point Sub-region

Point X Y Point X Y Number
A -9 0 A' -12 1 2
-2
B
C -1
°
-3
'8'
C'
-5
-4
1
-6
3
3
D 1 -3 D' 4 -6 3
E >2 0 E' 5 1 3
F 9
° F' 12 1 5

Table 7.3

7.3.3.3. Analysis of the second case

This case is shown in figure 7.11 where the hydraulic conductivity of the kernel (k2) is
lower than that of the upstream and the downstream (k l ) part of the dam. As in the first
case, the width of the dam (b) is considered to be equal to 1 meter and the discharge
through the dam (Q) is given by the parameter ~ =Q / (bk:z). The iterative procedure and
convergence criterion are as before. In the present case, the number of iterations to reach
convergence varies between 9 and 12 for different values of a = k2/ k 1. Figure 7.12
shows the position of the free surface for 4 values of a [12]. The figure also shows that
the inner seepage surface appears for values of a lower than 0,5.
This problem was also studied by Liu and Liggett [49] for the case a = 1 and the present
results are in agreement with theirs. The figure 7.13 shows the variation of the coefficient
~ and the heights of the points M,N,P and Q with respect to 10g(1/a). This figure
confrrms that the inner seepage surface appears for values of a lower than 0,5. If a is
greater than 0,5, the points Nand P have the same height and the free surface crosses
continuously the interface. For values of log(l/a) higher than 1,2, the seepage surface at
the downstream face of the dam vanishes and the height of the point Q becomes equal to
the downstream water level. For higher values of 10g(1/a), the discharge continues to
 82
decrease and also the height of the point P. If 10g(1/a) becomes higher than 3, the
influence of the upstream and the downstream part of the dam on the flow in its core is
negligible.
+-+:a =1
,,-v: a = 0.5
0-0: a = 0.1
-: a =0.05

10

Figure 7.12
Z(m)
14
a
12

10
H
8 N

4
Q
2

log(a)
2 3 4
Figure 7.13

In the following section, the results obtained in these two cases are compared for the case
a = 0,01.

7.3.3.4. Comparison of the results

The results in the two cases obtained for a=O,Ol will be compared in this section. First,
the positions of the different free surfaces are compared. The table 7.4 shows the heights
of the points M, N and P in the two cases.
Table 7.4 shows that the same heights are obtained for these two cases for the points M
and N. The table also shows that in the first case, the level of the point P is slightly higher
than in the second case. Table 7.5 demonstrates that along the interface EF, a flow from
the sub-region 4 to sub-region 5 involving a slight rising of the free surface in the
downstream part of the dam exists. The table 7.5 gives the discharges through the
different boundaries and interfaces in the two cases. Along the boundaries, the discharges
are positive in the inside direction and along the interfaces, they are positive from the first
to the second sub-region.
 83

Discharges (m3/h)
Surface Case 1 Case 2
F- I 0,1164 -
1-2 -0,0606 0
1 -3 0,0034 0
Level of the points(m)
Point
1-4 0,1735 -
Case I Case 2 F-2 0,0734 0,0133
M 9,94 9,97 0,0133
2-3 0,0128
N 8,43 8,40 F- 3 -0,0041 -0,0061
P 3,63 3,34
3-4 0,0048 0
3-5 0,0073 0,0072
F-4 -0,1236 -
4-5 0,0548 0
F-5 -0,0621 -0,0133

Table 7.4 Table 7.5

In the first case, the total discharge through the dam is equal to 0.1895 m 3 /h and in the
second case is 0.0133 m 3 /h. Thus, 91,5'% of the discharge seeps through the foundation
of the dam. In practice, the permeability of the foundation under the core will be reduced
by injections or other techniques to diminish as much as possible the discharge. These
discharges are not desirable as they not only reduce the available water volume but also
compromise the stability of the dam.

7.3.3.5. Conclusion

Although in the present case, no other solution is available for comparison, the discharges
through the different boundaries and interfaces are coherent with regard to the geometry
of the sub-regions and their characteristics.
In case of the impervious foundation, any junction point at an interface and a boundary or
of two boundaries requires a special treatment, whereas in the first case, six points of the
discretization require additional equations. As already shown in chapter 6, the
inaccuracies due to a non-optimal location of the additional points only affect the solution
near the intersection points requiring the additional equations.
This example confirms also the conclusions of chapter 6: one has to write a maximum of
these additional equations with regard to a sub-region that is next to a maximum of other
sub-regions of the domain. In the present case, the additional equations for the points B,
C ,D and E (see figure 7.10) have been written with regard to the core of the dam. The
core is the unique sub-region which is next to all its other sub-regions.The two other
additional equations, for the points A and F, have been written respectively with regard to
the upstream and the downstream part of the dam. These two sub-regions have "long"
common interfaces with the core of the dam. The fact the all the additional equations have
been written with regard to sub-regions the geometry of which changes after each iteration
does not affect the stability, the convergence of the iterative procedure, or the accuracy of
the results. Several tests were carried out where the additional equations where written
with regard to one or other sub-region meeting at the point under consideration. In these
cases, however, the results were not satisfactory.
 84

7.4. CONCLUSIONS

In this chapter, the boundary element method has been applied to two steady state porous
media flows. In the fIrst example, results are compared with an analytical solution and
good agreement has been obtained.
In the second example, two cases were considered for which the sub-regions have
different characteristics and other types of interfaces passing through the free surfaces.
This example allows also to test the method developed at chapter 6 to overcome the
numerical problems at the junctions of boundaries and interfaces.
The two examples presented in this chapter show that the boundary element method is
well adapted to solve stationary free surface porous media flows. Interesting results were
obtained for the applications under study. Notice that the boundary element method does
not require any remeshing after each iteration because this is automatically done by the
updating of the free surfaces.
 CHAPTERS

UNSTEADY FLOW THROUGH POROUS MEDIA

8.1. INTRODUCTION

In this chapter, the boundary element method will be applied to solve unsteady porous
media flows described by the Laplace equation. In the present case, as in the steady state
case, the potential u = Y + P / p has to satisfy the Laplace equation. The potential
contains two terms of which the first represents the influence of the gravity and the
second the influence of the water pressure. Thus, the potential is independent of time and
consequently, only the boundary conditions take the transient character of the flow into
account
In this chapter, unconfined groundwater flows, which require the iterative determination
of the free surface position in space and time will be studied. For confined unsteady
groundwater flows, the compressibility of the fluid and the porous media are not
negligible and the combination of (1.5) and (LlO) gives [6]:

K (~u) = - p [a (l-n)/n + P] ou/ot (8.1)

The above is called the diffusion equation. This equation is one of the fundamental
relationships of mathematical physics and describes problems as different as transient
head conductions, wave problems, diffusion problems, confined transient groundwater
flows and many others [48], (3], [74].
In the case of steady state flows, analytical solutions are available to test the results
provided by the numerical computations. For unsteady flows, only a few simplified
analytical solutions are known. Thus, in order to test the software, one needs to analyse
the convergence of unsteady flows to steady state flows and to compare results with those
provided by other authors.

8.2. ITERATIVE DETERMINATION OF THE FREE SURFACE POSmON

8.2.1.Boundaty conditions

To solve the Laplace equation on a given domain n, one needs to impose conditions
along the boundary and since the Laplace equation is independent of time, these
conditions have to take the transient character of the flow into account. The boundary
conditions have been discussed in the first chapter and for transient free surface flows, the
following conditions were obtained, ie.
 86

AB: u = H(t) DE: u = h(t)

CD: u=y EA: q=O (8.2)

au - cos 1 (au )
Be: = +N (n) and u=Y
dt E cos (y-P) do

One needs to discretize all the boundaries for a given problem to start its numerical
solution. This concerns also the initial position of the free surface which is given as data
of the problem. To determine the free surface behaviour in time, one needs two
conditions along this boundary. The fIrst of these two conditions will be used to solve
the linear system and the second condition (u = Y) will be applied to update the free
surface position at each time step.

Y~
\

\1
~ 0'
y

II h'\
I
I I'\ \\
I , \
I \ '
I, \
\
\

H(t) r/, - -}s---

,~C ........
I Y-B "
h(t)
_ _~~~~~_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _~~_ _- L_ _ _ _ _ _~X

Figure 8.1

8.2.2. Transformation of the first boundruy condition along the free surface

The fIrst condition along the free surface is given by the following relationship:

au - cos 1 (aU ) (8.2)

= + N (D)
dt E cos (y-P) do

The angles 1 and (y-P) are shown in fIgure 8.1 and N(D) is the normal recharge to the
free surfaces. Figure 8.1 shows the positive directions of the angles Pand 1, and of the
normal recharge vector.
The two unknowns along the free surface are the potential (u) and the normal velocity
(au/aD). The relation (8.2) contains also the derivative of the potential with regard to the
 87

time t. To introduce this relation as boundary condition into the linear system, one can
express the derivative of the potential with regard to the time in the form of finite
differences between two instants t and t+6t (with 6t the time steps) [47], [67], [18].
Thus, equation (8.2), with q = au/an, gives:

HAt t HAt HAt HAt t t t

Uj -u -1 ecos Ij . (qj +N (0» + (l-e) cosy . (q + N (n»
=- (8.3)
6t e e' cos (Y-~)j
HAt A t
+ (l-e') cos (')'-1-')

In equation (8.3), the coefficients e and e' are weighting coefficients that give the free
surface between the two instants t and t+6t The index j relates to the iterations at the time
step under consideration. The relation (8.2) was obtained after a linearization and the
transformation of the derivative with regard to time into a finite difference form introduces
a second linearization. Consequently, one needs to choose time steps (6t) that are small
enough to ensure the convergence and the stability of the iterative procedure. After
applying these small time steps, one can bring about the following simplifications in
equation (8.3) :
-) As the position and the shape of the free surface only changes slightly between two
time steps, the angle (y-~) also varies slightly and consequently, one can assume the
coefficient e' equal to zero

e' = 0 (8.4)

-) Like in the steady state case, the free surface nodes will move along lines between two
time steps. This line will be the line 0'1 in the case of figure 8.1. Thus, the angle y
remains constant at each free surface node, ie.

(8.5)

-) As the coefficient e' is taken equal to zero, one can assume that the angle (y-~) is equal
to its value at the last iteration. Thus, one obtains:

t
(y-~) = (y-~) ifj =1
t+At (8.6)
(y-~) = (y- ~)j-l ifj > 1

-) The recharge along the free surface is one of the data of the problem and consequently,
does not vary as the iterative procedure proceeds to determine the free surface position
at the time (t+6t). Thus, one can write:

Nt,t+At =e Nl+At (n) + (l-e) Nt (n) (8.7)

 88

The introduction of (8.4), (8.5), (8.6) and (8.7) into the equation (8.3) gives:

- cos y [9q~+ll.t + (1-9) qt + Nt.t+ll.t] (8.8)

E cos (y-P) J

8.2.3. Transformation of the linear system

8.2.3.1. Introduction

Contrary to the case of other boundaries, both the potential (u) and the normal velocity (q)
are unknown along the free surface. Thus, at each free surface node, one has two
unknowns and hence, an additional equation is needed at each of these nodes to obtain a
regular system. These additional equations are provided by the relation (8.8). Thus, if
the number of nodes of the discretization is equal to NN and if the number of nodes of the
free surface is equal to NFS' the linear system will consist of NN+NFS equations. Notice
that relation (8.8) contains only two unknowns at each time step (t+.1.t) : the potential and
the normal velocity. Thus, along the free surface, it is possible to express one of these
two unknowns in function of the other, and afterwards, to eliminate one of them and
consequently, to reduce the number of unknowns to NN. One has the choice of the
elimination of either the potential, or the normal velocity, at the next time step. This will
be discussed in what follows.

8.2.3.2. Elimination of the potential

This method has been developed by Liggett [47] and several other authors [27], [56],
[45] have applied it to solve various unsteady free surface flow problems. In this case,
the solution of the linear system gives the normal velocity (qt+.1.t) at the free surface
nodes, but one needs the potential at the free surface nodes in order to update their
positions. The potential at these nodes is given by the following relation, deduced from
(8.8), ie.

t+ll.t _ t .1.t cos y [9 l+ll.t (I 9) t Nt.t+ll. t]

uj - U - q. + - q + (8.9)
E cos (y-P) J

After the computation of the potentials, the same iterative procedure (6.2) and the same
convergence criterion (6.4) as in the steady state case are used in order to update the free
surface position and to check the convergence at the given time step (t+.1.t).
Unfortunately, this method gives rise to some numerical problems and requires the use of
smoothing procedures along the free surface [27], or to use a very small time step [56], in
order to guarantee the stability and convergence of the iterative procedure.
The sensitivity of the results to geometry variations for the case of a very simple problem
has been studied in chapter 5 and it has been concluded that the normal velocities are 5 to
10 times more sensitive to boundary geometry variations than the potentials. Hence, in
the present case, one can compute first the normal velocities from which one can deduce
by the way of the relation (8.9) the potential. Thus, the errors that affect the normal
velocities will have repercussions on the potential, which may be more inaccurate.
 89

Further to this conclusion and the numerical experience of the author, a new method
based on the elimination of the normal velocity at the free surface nodes has been
developed. This method is presented at the following section.

8.2.3.3. Elimination of the normal velocity

Contrary to the first method, the elimination of the normal velocity from the system
provides directly the values of the potential along the free surface [14], [16], as in the
steady state case. Thus, after solving the linear system, one can apply directly the
relations (7.2) and (7.4) in order to update the free surface position and to check the
convergence at any considered time step. If the convergence criterion (7.4) is not
verified, one can iterate again. On the other hand, if convergence is reached, one has
obtained the free surface position at that time step and the relation (8.10), deduced from
(8.8), gives the nonnal velocities at the free surface nodes, ie.

l+t.l e cos ()'-~) I t+t.l I-a I N1,l+llt

q. = - - - ( u -u. ) - - q - - - (8.10)
J a
t1t cos y J a a
Seve,al applications of this method [14]} [18], [19], [23] have been presented and they
show that no smoothing procedure is required to ensure convergence and stability.
n1

Figure 8.2

8.2.4. Discretization of the free surface

As in the steady state case, the discretization of a smooth free surface will present an angle
at the junction point of two elements (see figure 8.2).
Hence, one has to use there double nodes and to apply the conclusions of chapter 6. The
elimination of the normal velocity is carried out by introducing relationship (8.10) as
Neumann boundary condition at each free surface node. Thus, one equation is obtained
at each junction point of two elements. This equation is given by the foIIowing
relationship.

(6.20)

The unknowns u, qpl and qp2 are all at time (t+t1t). Thus, the relation (8.10) gives the
following expressions for PI and P2, ie.

e cos (y- ~ )
1 I I -a I
- - - - ( u -u)
t1t a cos 'Y
-a- qP1 a
(8.11)
 90

£ cos (y- ~2) 1- 9

t t
qp2 = At9cosy
(u - u) - -
. 9
CIp2
e
(8.12)

If one uses for the discretization of the free surface elements that are small enough so that
the discretized free surface is nearly smooth, the following simplifications can be
inttoduced:

q = (qpl+qp2) /2
Nt.t+t.t = (N1+t.l + N+t.l) / 2 (8.13)
pi p2

Thus, introduction of (8.11), (8.12) and (8.13) into (6.20) yields the following fonnula:

[ Hn + Hl2 + £
&e~y
• ( cos (y-
I
13) On + cos (y- 132) °
12 ) ] u + Fli
.
Xi

(8.14)

with the following additional simplification,

(8.15)

Combination of (8.11) and (8.12) yields the following equation.

£ cos (y- ~) 1 1 - 9 1 Nt.l+t.l

q = (u - u) - - q - -- (8.16)
At9cosy 9 9

Thus as in the steady state case, one can use a unique node at the junction point of two
free surface elements if the discretization of the free surface is refined enough to obtain a
nearly smooth surface.

8.2.5 The coefficient 9

The method developed above uses the weighting factor 9 between time steps t and t+At
The choice of this weighting factor is of prime importance on the convergence and
stability of the solution. The influence of this factor on the solution and the determination
of its optimal value has been studied by several authors [46], [48]. Following these
studies, one can use the relation (8.17) in order to determine the optimal value of 9 [14],
ie.
 91

=-(~+
/,. ~l
e-1
(8.17)

1t jKxKy ~t
with: /,.= -
e ~x

e : the effective porosity

~ : the horizontal distance between two free surface nodes

8.2.6. Conclusion

In the above section, the two methods that allow to introduce the finite difference
boundary condition along the free surface, into the linear system have been studied. This
relation was obtained after two linearizations (see relation (1.27) and the fmite difference
in time) and hence, the use of small time steps is essential. The choice of small time has
also allowed some simplifications (see point 8.2.2.). The conclusions of the chapter
concerning the numerical aspects of the boundary element method and the experience of
the authors who have applied the elimination of the potential, lead one to assert that it is
better to apply the elimination of the normal velocity. The use of a finite difference in time
introduces the weighting factor 6. Ony can use the relation (8.17) to determine this
coefficient in order to ensure the convergence and the stability of the iterative procedure.
In the next section, some applications of the method developed above will be presented.

8.3. APPLICATIONS

8.3.1. Introduction

In this section, the boundary element method will be applied to solve several unsteady
free surface porous media flows.
In the unsteady case, only a few analytical solutions are available in the literature and they
have been obtained after some linearizations of the real problem and thus, they are not
rigorously accurate. Contrary to the steady state case, one can only compare results with
other numerical and experimental results. The examples presented in this paragraph have
been chosen to cover a range of cases as large as possible.

8.3.2. Unsteady flow through a rectangular dam

The figure 8.3 shows a rectangular dam [14], [18] having initially (t<O) the same water
level (24 m) upstream and downstream. At the time t=O, the downstream water level is
supposed to have an instantaneous draw down to h=4 m. It follows that this produces a
transient water flow inside the dam and a movement of the free surface towards a new
steady state position corresponding to the new downstream water level. For the
discretization of this problem, one can use linear and quadratic elements with a total of 50
nodes and 45 degrees of freedom. The boundaries were discretized as follows:
 92

AB : 5 quadratic elements BC: 16 linear elements

CD : 4 quadratic elements DE : 2 quadratic elements
AE: 7 linear elements
The characteristics of the porous medium and the values of the different computation
parameters are given hereafter :

K=O,Q1 cm/s £=0,4 ~t= 3204 s

Eo =0,001 (j) =0,5 a =0,69
H=24m h=4m L=16m

The figure 8.3 shows the geometry of the dam and the position of the free surface after
several time steps (N = t / ~t). The free surface converges to a new stationary position
imposed by the new downstream water level. The convergence was reached after 60 time
steps. The table 8.1 gives the number of iterations per time step. As shown in figure
8.3, the free surface decreases rapidly during the first time steps and so, the number of
iterations is important during this period. Afterwards, the number of iterations to achieve
convergence to the desired accuracy decreases rapidly and after the 35 th time step, the
relative correction at each time step is directly lower than the imposed value of Eo.
tYlm'
-r~Bi-==-_ _~'----t_-N.O

Time steps number of iterations

1 7
2-7 8
8 - 12 7
c
13 - 16 6
H 17 - 20 5
21- 24 4
25 - 29 3
30- 34 2
o N>O 35 - 60 1

A x~
L

Figure 8.3 Table 8.1

The figure 8.4 shows discharge entering and leaving the dam through its upstream (AB)
and its downstream (CD + DE) face. Their values converge rapidly towards Q = 6,075
m3/h which corresponds to the analytical solution of the steady state case given by (6.6).

Computations Level of point C (m)

Our computation (unsteady) 13.03
Our computation (steady) 12.98
Polubarinova -Kochina [65] 12.95
Finnemore and Perry, FDM [32] 12.85
Cheng and Li, FEM [27] 12.20
Todson, FDM [72] 11.60
Table 8.2
 93

3r---~ __________
2 DE

N
to 20 30 40 50 60

Figure 8.4

This example has been studied by several other authors who have applied the finite
difference method (FDM) [32], [72] and tl}e finite element method (FEM) [27]. The table
8.2 gives the final position of the intersection point C of the free surface and the seepage
surrace obtained by these authors and by the present study in the steady and in the
unsteady case. The analytical solution of. Polubarinova-Kochina [65] corresponds to the
steady state case to which the free surface ought to converge. The table 8.2 shows that
the present solutions are in a good agreement with this analytical solution.
The analysis of the results shows a good convergence to the new steady state case and at
each time step, a good convergence to the corresponding free surface position. These
results have been obtained without the use of any smoothing procedure.

8.3.3. Rechar~e prob1em

Groundwater flows with vertical recharges, due to infiltrations along their free surface
often occur in practice and they involve a raising of the free surface. Such a problem is
shown in figure 8.5 where a vertical recharge N is imposed along a part of the free
surface [18], [19].

l ,N-i
n.311 - - -- -- r -- - --- -- --~--+
t,:; : X
;:;:;:;;:: :::::::::,,::,,:;~:,,:::::;:>::::;.;:;:y:;:::":::::;:;:::;::::~::::":::::~~;';::::::':::::;:":::;:::::;:;::::: .......... $:::::<::::::::::;:;::;:::j::;:::;:::::::;::::::;';:::; •
23.8 243.8
Figure 8.5

Linear and quadratic elements have been used for the discretization of this problem with a
total of 34 nodes and 30 degrees of freedom. The vertical recharge has been taken equal
 94

to 0,056 cm3/cm2s. The characteristics of the porous medium and the values of the
different parameters are given hereafter.
K=0,42cm/s £=1,0 ~t=I5s
£0 = 0,005 CJ) = 0,5 e = 0,78
This problem was studied by several authors in the past. An analytical solution has been
developed by Hantush [38] who assumed that the water level at each point is equal to the
averaged head at this point. This assumption leads to a linearized differential equation in
terms of this head. Marino [50] used a Hele-Shaw model in order to obtain experimental
results for this problem. Neuman et al. [59] applied the finite element method to this
problem and Liu and Liggett [49] the boundary element method.
The figure 8.6 compares the free surface positions after 1, 5 and 9 minutes obtained by
these authors with our results.
The results presented in figure 8.6 allow to conclude:
- that during the first time steps, a good agreement between the experimental and the
numerical results is obtained. Afterwards, some problems with the Hele-Show model
[38] lead to a less satisfactory agreement.
- that the results show a good agreement between the finite element results and the present
study.
- that the analytical solution provides satisfactory results, despite the simplifications.

Y(rn)
- - - analytical [37]
25 - - present results [18]
BIEM [48]
* FEM [58]
• experimental [49]

20

15

____ _______
11.3
--~~~L- ~ ~

10 X(rnl
10 20 30 40 50 60 70 80 90 100
Figure 8.6

The table 8.3 shows the number of iterations per time step.

number ot Iterations umestep

4 1-2
5 3-5
6 6 -14
7 15 - 36
Table 8.3
 95

It was noticed that after 36 time steps, the free surface did not reach its steady state
position and consequently, the number of iterations per time step did not decrease. Like
before, this example did not require any smoothing procedure to ensure stability.

8.3.4. Dam with a sinusoidal varying upstream water level

The present example considers a case where the free surface goes up and down
simultaneously in different parts of domain. Consider a rectangular dam and along its
upstream face, a sinusoidal varying water level H(t) in time. Figure 8.7 shows the
geometry of the domain and the relation (8.18) defines the variation in time of the
upstream water level.

H(t) = 5 + sin (0,002618 t) (8.18)

One applies linear and quadratic elements with a total of 42 nodes and 38 degrees of
freedom to discretize the boundary. The characteristics of the porous medium and the
values of the different parameters are given as,

K=0,3m/h dt = 1 hour CI) =0,5

e= 0,76 £0 =0,001 £=0,5

y ,

"" 8
C
HIt) 5
A 0 X
11l

Figure 8.7

The figure 8.8 shows the free surface position at five different moments (N =t I dt).
y

c X
N=3 10

Figure 8.8

Figure 8.8 shows that the free surface moves up and down with regard to the imposed
variations of the water level along the upstream face of the dam. The figure also shows
that for the 30 to the 40 time step, the free surface rises again in the upstream part of the
 96

dam, but continues to go down in the downstream part. This comple:, behaviour of 0e
free surfaces is easy to represent using boundary element method as It does not requtre
any special treatment. Convergence at each time step was reached after 5 iterations at the
most
H(m)

Time step Y (X - 3)
0 5
10 5,505
20 5,354
30 4,635
40 4,693
50 5,468
60 5,347
70 4,633
80 4,692
90 5,468
100 5,347

Table 8.4 Figure 8.9

The figure 8.9 shows the variation 6f the upstream water level (H) and the discharges
entering (positive, Qup) the domain along the upstream face and leaving (negative, Qjown)
the domain along its downstream fac y in function of the time step N.
Figure 8.9 shows a lack of phase of approximately 4 hours in the discharges with respect
to the upstream water level. Indeed, the maximum discharge enters the dam through its
upstream face before the upstream water level is maximum and the discharge leaving the
dam through its downstream face continues to increase at the same time the upstream
water level begins to decrease.
The table 8.4 shows the free surface water level inside the dam, 3 meters from its
upstream face. Notice that like the upstream water level H(t), the variation of the free
surface inside the dam becomes rapidly periodical. The number of iterations per time step
varies between 1 and 5. Like the two previous examples, this one requires no smoothing
procedure along the free surface in order to ensure stability. A similar problem has been
considered with a 25 meter long dam [18] and similar results have been obtained.

8.3.5. Interior points of the domain

As explained in chapter 3, one can compute the potential and the velocity at every interior
point of the domain after solving the boundary value problem. Then, the potentials are
given by the relation (3.17) and the two components of the velocity vector are given
respectively by the differentiation of (3.17) with regard to x and y.
The figure 8.10 shows a trapezoidal dam [14] along the upstream face on which a linear
water level going down from the level HI to Hz during 400 minutes has been imposed.
This example will be used to show what happens at some interior points during the
unsteady flow occurring inside the dam.
The potential and the velocity will be computed at the nine points shown in figure 8.lD.
The same example was studied previously by Kellogg [42] who obtained experimental
results and by Chang [27] who applied the boundary element method. Following the
original method due to Liggett [47], Chang uses the elimination of the potential along the
free surface in order to solve the linear system and like the other authors who have applied
this method, he has to apply a smoothing procedure along the free surface in order to
ensure the convergence and the stability of the iterative procedure. Chang has avoided
these numerical problems by averaging the velocity vectors spacewise and timewise along
 97

the free surface. For our computations [14], the approach of eliminating the normal
velocity along the free surface has been used and this procedure did not require any
smoothing.

H~

Figure 8.10

The characteristics of the porous media, the dimensions of the dam (i : the slope of the
two faces of the dam) and the values of-the different computation parameters are given
hereafter.

K = 0,024 m/min H 1 =15m 'H2=5,4m' B=28m i = 51,34°

.1t = 10 min £ = 0,5 co = 0,5 £0 = 0,005 Sop =0,75
The value of the hydraulic conductivity corresponds to the experimental value of Kellogg
[42]. The hydraulic conductivity is a parameter that cannot be very accurately determined.
Moreover, it can significatively change from point to point, even in laboratory conditions.
In practice, the value of K is never constant on a domain. Thus, the assumed value of K
is the equivalent hydraulic conductivity that describes as best as possible the behaviour of
the real problem. The figure 8.11 shows the free surface position with regard to the
number of time steps (N = tI.1t). After the 40 th time step, the upstream water level
remains constant and the flow conditions inside the dam converge to a new steady state.

N=O
N=10

N=20

N=30

N>40

Figure 8.11
 98

°
The free surface positions at t = and t = 00 correspond to steady state conditions and
their respective positions have to be determined by considering the corresponding steady
state flows with their specific boundary conditions.
The behaviour of the potential at some interior points of the domain will now be analysed.
Figure 8.12.a shows the localization of several interior points and figure 8.12.b the
variation of the potential in function of time at those points.
u(m)

-.-...~==== 1
~
~--6

~__~~__~__~~~t
o 40 80 120
Figure 8.12.a Figure 8.12.b

Only points 1,3,5 and 6 are located inside the flow domain at all times. The other points
leave the flow domain sooner or later due to the movement of the free surface. Figure
8.12.b shows that the potential decreases continuously at the interiors points, except at
point 1. Indeed, this point is located close to the upstream boundary of the dam and
consequently, the potential at this point is mainly influenced by the boundary conditions
imposed there, and not by the variation of the free surface position.
The hydraulic conductivity determined by Kellogg [42] was equal to 0,024 m/min. The
tables 8.5.a to 8.5.c compare the experimental results of Kellogg [42] and the numerical
results of Chang [27] with our results for four different values of the hydraulic
conductivity. Figure 8.12.a shows the position of the three points considered here.

lit Kellogg Chang k=0.0 24 k=0.032 k=0.04O k=0.048

0 14.7 14.7 14.7 14.7 14.7 14.7

10 12.5 10.9 12.4 12.4 12.4 12.4
20 9.9 8.4 10.1 10.1 10.1 10.1
30 7.& 6.6 7.9 7.8 7.& 7.&
40 5.6 4.8 5.7 5.6 5.6 5.6

Table 8.5.a - Point 1

lit Kellogg Chang k=0.024 k=0.032 k=0.040 k=0.048

0 12.9 12.4 12.9 12.9 12.9 12.9

10 10.6 10.4 11.3 11.2 11.2 11.2
20 9.& 9.0 9.8 9.7 9.6 9.5
30 8.3 7.7 8.5 8.2 &.0 7.9
40 6.7 6.6 7.4 7.0 6.7 6.5

Table 8.S.b - Point 2

 99

At Kellogg Chang k=0.021j. k=O.OJ2 k=O.OIj.Q k=0.01j.8

0 9.1 9.0 9.2 9.2 9.2 9.2
10 7.9 8.3 8.6 8.6 8.6 8.5
20 7.4 7.7 8.0 7.9 7.8 7.7
30 6.6 7.0 7.4 7.2 6.9 6.7
Ij.Q 5.8 6.2 6.8 6.3 6.0 5.7

Table 8.5.c - Point 7

Table 8.S.a shows that close to the upstream face of the dam, the potential is almost
independent of the hydraulic conductivity, but depends mainly on the boundary
conditions imposed there. At the two other points (2 and 7), the potential depends mainly
on the free surface level, which depends on the hydraulic conductivity. Tables 8.5
suggest that the actual hydraulic conductivity K may be slightly higher than the assumed
value. Thus, the study of the evolution of the potential at some interior points can be used
to calibrate the value of K. After the analysis of the potential values, the variation of the
velocity at interior points will be studied. Figure 8.13.a shows the localization of the
points under consideration, the figure 8.13.c the variation of the velocity vector at the
point· 1 and the figure 8.13. b the variation of the velocity vector at the other interior
points. .

Figure 8.13.a

-x

0.5 cm/~in
y
Figure 8.13.b
 100

40

o
Figure 8.13.c

Figures 8.13.b and 8.l3.c show the behaviour of the velocity vectors in function of time.
Their origin is always the points P and their heads move on the corresponding curves.
The dots on the different curves correspond to segments of ten time steps.
The figures 8.13.b and 8.13.c allow to draw the following conclusions regarding the
flow through this dam :
- the flow velocity increases progressively from the upstream to the downstream part of
the dam ;
- like the potentials, the evolution of the velocities of the points located close to the
upstream face present a discontinuity at the fortieth time step;
- as showed by the figure 8.11, the free surface decreases not as fast as the imposed
upstream free surface level. Consequently, close to the upstream face of the dam, the
velocity changes twice its direction;
- at the point I, the velocity vector remains nearly parallel to a direction imposed by the
boundary conditions along the upstream face and the bottom of the dam. At the other
interior points, the direction of the velocity vectors varies appreciably in function of
time.

0,15

Figure 8.14
 101

The figure 8.14 shows the discharges through the different boundaries of the dam. This
figure confirms that during 60 time steps, water leaves the dam through both its upstream
and its downstream face.
Whereas the discharge leaving the dam (positive) through its downstream face varies
continuously, the discharge through the upstream face (positive when it enters into the
dam) as well as along the free surface present strongly discontinuities at the fortieth time
step. The discharge "through" the free surface corresponds to the integration of the
normal velocity along this boundary. This discharge tends towards zero whereas the two
other discharges, like for N = 0, tend towards a new common value corresponding to a
new steady state case (N = 00).
The different results presented here show that the flow tends very slowly to a new steady
state case. The analysis of the potentials and the velocities at some interior points show
that variable boundary conditions can lead to a flow appreciably more variable inside the
domain.

8.3.6. Conclusions

In this section, the boundary element method has been applied to four cases of unsteady
free surface flows. The first example studies a flow through a rectangular dam. For the
given boundary conditions, the free surface converges to a new position corresponding to
a new steady state case. The free surface position after convergence has been favorably
compared with several other numerical solutions and an analytical solution.
The second example considers a recharge problem. Like for the first example, the results
have been compared against other numerical solutions, as well as with experimental
results and an analytical solution that was obtained after a further linearization of the
problem.
The third example considers an unsteady flow inside a dam along the upstream face of
which a sinusoidal varying water level was imposed. In this case, the behaviour of the
free surface becomes complex because it goes up and down simultaneously in different
parts of the domain. However, the free surface variations tend towards a periodical one,
like the variation of the upstream boundary condition.
The last example shows the results obtained at the interior points of the domain. For the
dam under consideration, some interesting conclusions have been obtained concerning the
flow conditions inside the dam in the transient phase. The analysis of the behaviour of
the velocity vectors inside the dam has showed that in the upper part of the dam, the flow
changes twice its direction.
The classical method that uses the elimination of the potential along the free surface needs
the use of smoothing procedures along the free surface and the discretization into small
time steps to ensure the convergence and the stability of the iterative procedure. The
method developed here, instead uses the elimination of the normal velocity along the free
surface and does not require any of those two procedures to ensure stability and
convergence.
The four examples presented here have shown that the boundary element method is very
well adapted to solve unsteady free surface problems because it does not require any
remeshing of the discretization, like other numerical methods. Indeed, the new
discretization of the boundaries is set up simultaneously with the correction of the free
surface position at the time of each iteration.
The four examples have demonstrated that the boundary element method provides good
results, even for boundary conditions varying strongly in time.
 102

8.4. CHOICE OF THE DISCRETIZATION

8.4.1. Introduction

When a numerical method is used to solve a physical problem, an important point is the
choice of the discretization. For an unsteady flow through a porous media, one has to
e
choose the discretization of the boundaries, the time step l!.t and the coefficient (see
point 8.2.5). If the finite element method is used to solve such problems, the
discretization has to be chosen judiciously to ensure the convergence and the stability of
the iterative procedure. Different authors [52],[32],[4],[34] have studied these questions
and they have suggested convergence and stability criterion.
To show the influence of the discretization on the results, the boundary element method
has been applied to a problem for which an analytical solution is available in the steady
state case. Consider the rectangular dam that has already been studied in the chapter
dealing with steady state flow conditions (see point 7.3.2). At the time t > 0, assume that
upstream and downstream water level corresponding to those of the analytical solution of
Muskat [57] in the steady state case.

8.4.2. Convergence to the steady case

As in fIrst example, the convergence of the free surface to the steady case and particularly
the position of the resurgence point C when the convergence is reached has been studied.
Numerical tests have shown that the fInal position of the resurgence point is only a
e
function of the discretization and so, is independent of the time step l!.t. The coefficient
was determined according to the relation (8.17) :
The fIgure 8.15 presents the fInal level of the resurgence point in function of the number
of nodes of the discretization (NDE), for both steady and unsteady flows.
yc{m)

steady flow
0,215 unsteady flow

0,210
\
\

0,205
".,'.'.,,--------.---0- --0--
"
-. ----.. -
-+~~~~-T~~ __ ~~~ __ ____ NDE
~

50 100 150
Figure 8.15

The comparison with the analytical solution giving an elevation of 0,206 m for the
resurgence point C, shows that the relative error is less than 1% when the number of
nodes of the discretization is higher than 45.
No stability or convergence problems occur during these computations. Furthermore, the
elimination of the normal velocity along the free surface does not need any smoothing
procedure along the free surface.
 103

Even if the convergence of the free surface to the same steady state position is
independent of the time step for a fixed number of nodes, the time evolution of the free
surface, the time necessary to reach the convergence and the total number of iterations
vary in function of the chosen time step. The figure 8.16 presents the level of the
resurgence point C obtained after 30 hours, for different discretizations and different time
steps.
The figure 8.16 allows to conclude that the free surface moves more slowly when using
longer time steps. This conclusion is confirmed by the figure 8.17 that shows that the
time necessary to reach the convergence increases with the time step length. On the other
hand, the number of iterations necessary to reach the convergence decreases with the time
step.

Y(m)

0,225

1 ~t = 1 hour
2 ~t = 2 hours
0,220 3 ~t = 3 hours
4 ~t = 4 hours
5 ~t = 5 hours

0,215
-t-~-~-~~-~-~-~-~-NDE
20 40 60 80

Figure 8.16

lit (h)

100 125 150

6
--i--150
.'
I

90 , ;',' / / I " , " 140

'~',/
_,/ /,' -' ,120
3
, " .,.~
, " ,
~~' " ,'100
2 ~ ... / , .. " ......
•• ' •• ' 80
175 -----
200 225
NOE
20 40 60 80

Figure 8.17

- : number of hours to reach convergence

----: number of iterations to reach convergence
 104

The figure 8.17 shows that for a one hour time step, the time necessary to reach
convergence is nearly independent of the number of nodes of the discretization. On the
other hand, it can vary considerably for longer time steps.

8.4.3. Influence of the e coefficient

The e coefficient is always determined by the relation (8.18) and the numerical tests allow
to conclude that the iterative procedure becomes rapidly unsteady for e < 0,5.
e
For > 0,5, one obtains significant differences in the position of the free surface during
the first time steps when e is varying. When e is smaller than the optimal value
determined by the relation (8.18), the free surface moves more rapidly. But for a given
discretization and a given time step, the free surface always converges to the same steady
position, which is reached after a number of time steps that increases a little with the
considered time step.
In practice, one always determines the e coefficient by the relation (8.18).

8.4.4. Conclusions

Thanks to the different tests presented in this paragraph, one can formulate the following
recommendations for the choice of the different factors that influence the problem:

- Concernin~ the discretization:

One has to choose a discretization that gives satisfying results for the corresponding
steady flow. If such a case is not available, one can take a density of nodes similar to
that used in other cases and giving satisfactory results. In all cases, it is recommended
to use a discretization with a higher density of nodes (about two times more) along the
free and the downstream surfaces than along the other boundaries of the domain.

- Concerning the e coefficient:

The e value is determined with regard to the relation (8.18).

- Concerning the time step llt :

Use a time step that gives a e value about 0,7 to 0,8.

8.5. CONCLUSIONS

In this chapter, the boundary element method has been applied to solve transient free
surface groundwater flows described by the Laplace equation. Since this equation is
independent of time, the boundary conditions have to take the transient character of the
flow into account. The condition that describes the evolution of the free surface in time
has been established in chapter one. To introduce it into the linear system, it has been
expressed in a finite difference form. Unlike the previous authors who have applied the
elimination of the potential along the free surface, the present work has introduced the
elimination of the normal velocity which does not require any smoothing procedure to
ensure convergence and stability.
 105

The examples presented in this chapter deal with a wide range of unsteady flows and gave
in all these cases good results.
Finally, the influence of the discretization in space and time has been studied. This study
has allowed to formulate some recommendations concerning the discretization of the
boundary, the time step ~t and the coefficient 6.
 CHAPTER 9

INFILTRATION PROBLEMS

9.1. INTRODUCTION

In this chapter, infiltration problems of water from canals or rivers through porous media
soils to underlying groundwater will be studied. The study of these infiltrations is very
important in areas where the agriculture is dependent on irrigation. In fact, infiltration and
evaporation produce water losses that have to be taken into account to develop a water
management project
In chapter 7, steady state problems has been presented where at each point, the
contribution of the pressure (p) to the potential (u = Y + pip) is not negligible (except for
points close to the free surface). On the other hand, in case of infiltration problems, the
contribution of the gravity (Y) to the potential is important with respect to the pressure. In
consequence, the free surfaces are nearly vertical (see figure 9.1) and their determination
is not easy.
y

__ ~~ ____________-r__ ~X

Figure 9.1

During the last years, several authors have applied the finite difference method and the
fmite element method to solve such problems. The application of these methods imposes
the use of special formulations of this problem in order to ensure the convergence and the
stability of the iterative determination of the free surface position. The application of the
boundary element method to these cases only requires a suited iterative procedure to
determine the actual position of the free surface as will be seen in what follows.

9.2. PREVIOUS SOLUTIONS

The figure 9.1 shows that the infiltration from canals to underlying groundwater is
characterized by nearly vertical free surfaces. In addition to that, the position of the two
 107

free surfaces influence each other. As a result, the iterative procedure to detennine the
free surface position must be chosen judiciously to ensure convergence and stability.
During the last years, different authors have applied the finite difference method [40],
[26], [54] and the finite element method [58], [26], [69], [40] to infIltration problems
and as these problems are very sensitive to numerical instabilities, they have had to
develop more sophisticated techniques than those used to compute classical free surface
flows.
Among these special techniques, the following should be mentioned:
- auxiliary functions depending on the velocity potential (u) and the stream function
[40];
-a generalized velocity potential and stream function [26], [69], [54] ;
- special smoothing procedures along the free surfaces [58] ;
- a modified functional [34].
In this chapter, the boundary elements method is applied to infiltration problems. As will
be shown, the boundary elements method does not require any particular numerical
treatment to ensure stability and convergence of the iterative procedure.

9.3. ITERATIVE DETERMINATION OF THE FREE SURFACE POSmON

9.3.1. BoundaI)' Conditions

In order to solve the Laplace equation on a domain n, one has to impose conditions along
its boundary r.
y y

r n

A
L......._ _ _~D
h E
h D
X X

Figure 9.2.a Figure 9.2.b

In the case of the figure 9.2.a, the following conditions are obtained:

BC: u=H AD:u=h (9.1)

ABandCD: u=Y and q=O

If the problem has an axis of symmetry (figure 9.2.b), one obtains the following
conditions:

FC: u=H ED: u=h EF: q=O (9.2)

CD: u=Y and q=O

The potential imposed along the lower boundary (AD or DE) corresponds to the level of a
drainage layer where the fluid seeps nearly horizontally or to the free surface of the
groundwater. This simplifying hypothesis does not affect significantly the discharge and
the free surface position near the points A and D. The two conditions along the free
surfaces will be used here in order to develop a new iterative procedure to determine its
steady state position.
 108

9.3.2. Iterative Procedure and Convergence Criteria

As the positions of the free surfaces are problem unknowns, one needs the following two
conditions along these boundaries in order to determine their positions with the help of an
iterative procedure, ie.

q = 0 and u=y (9.3)

The first of these two conditions (q = 0) is used to solve the linear system and the second
one (u = Y) to update the free surface position by an iterative procedure. At the time of
the iteration j, the solution of the linear system gives the potential Uj at the free surface
nodes. Then, the second condition in (9.3) gives the relationship require to update the
free surface nodes positions, where the sufftx j is the iteration number, ie.

y.J = cou·J + (1 - co) y.J-1 (9.4)

This procedure is used until convergence is reached for a given relative precision Eo'
using the following criterion.

(9.5)

In the case of classical groundwater fl<~w problems, the coefficient co is taken equal to 0,5
[16]. In the present case, the fluid pressure becomes rapidly negligible and consequently,
the potential at the free surface nodes becomes rapidly equal to their corresponding
vertical coordinate. For this reason, one can choose in the present case lower values for
the coefftcient co.
y
,

Figure 9.3

The relation (9.4) leads naturally to a vertical correction of the positions of the free surface
nodes. But, since the free surfaces are also nearly vertical, the following procedure (see
figures 9.3 and 9.4) has been developed to correct horizontally the free surface node
positions [19], [20], ie.
a) To start the iterative procedure, one imposes a vertical free surface COo (see figure
9.3) as a first solution of the linear system.
 109

b) After the flI'St solution of the linear system, one only updates the position of the flI'St
free surface node (Mo) downstream the point C. The position of the node Mo is
updated along the line PM o (see figure 9.3) in accordance with the relation (9.4).
Thus, the node Mo moves to M'o' Then, one moves the point M'o to MI which will
have the same height as Mo (see figure 9.4). Thus, after the flI'St iteration, the node
Mo moves to MI'
c) The other free surface nodes (downstream the node MO> are moved horizontally to the
same abscissa as the point MI' Thus, the node Do moves to D} (see figure 9.3). In
this way, one carries out a horizontal correction of the free surface nodes.
d) After the second solution of the linear system, one applies the same method as
described in b to the two first free surface nodes downstream of the point C (the
points M} and N I , figure 9.4). Thus, the nodes MI and NI move respectively to M2
and N2 (see figure 9.4).
e) The other free surface nodes (downstream of node N 1) are moved horizontally to the
same abscissa as the node N 2. Thus, DI moves to D2'
f) The process described in the two previous paragraphs d and e is applied as follows.
After each iteration, the procedure described in d is applied to an additional free
surface node, until the position of the first node upstream the point D is updated by
this method. At each iteration, the positions of the remaining free surface nodes are
updated in accordance with paragraph e.
g) If the free surface comprises N nodes (including those at C and D), one will apply
from de (N-l)th iteration the procedure described in d to all the free surface nodes.
Finally, the intersection point (D) of the free surface and the lower boundary will be
determined by the intersection of two straight lines defined respectively by the lower
boundary and the two last free surface nodes above D.
One can use the relationship (9.5) to check the convergence of the iterative procedure for a
given approximation eo. For a free surface node R (see figure 9.4), one will have Yj - I =
Y(Rj _1) and Yj = Y(Rj_I)'

9.4. COMPUTATIONAL EXAMPLES

9.4.1. Introduction

In this section, three computational examples of steady state infiltration flows will be
presented. In order to check the method, results have been compared with three cases
having analytical solutions.

9.4.2. Symmetrical Problem

The first example is shown in figure 9.5 which compares the corresponding analytical
solution (continuous line) [25] with experimental results (crosses) [25] and the solution
obtained by the boundary element method (points) [19], [20]. The experimental results
take into account the capillary fringe that does not contribute significantly to the discharge,
but the extent of which can be appreciable in certain cases.
As the problem is symmetrical, only the half domain with a Neumann boundary condition
(q=O) along the symmetry axis needs to be considered. For the discretization of this
problem, 14 quadratic three node elements with a total of 32 nodes and 28 degrees of
freedom have been used. The reduction in the number of unknowns from 32 to 28 is
obtained by a special method initially developed for problems requiring the division of the
domain into sub-regions. For the iterative determination of the free surface position, the
following values have been chosen for the coefficients of equations (9.4) and (9.5).
eo = 0.0005 co = 0.25
 110

To start the iterative procedure, one needs to impose a vertical free surface at the abscissa
of the point C. The figure 9.6 shows for each iteration (N) the value of the abscissa of
the points D and the evolution of the convergence defined by the coefficient 8 as given by
(9.6), with £ defined as in (9.5), ie.

(9.6)

The slight increase of the coefficient 8 after the twelfth iteration corresponds to the
transition from the iterative procedure (d), (e) and (f) to the procedure (g).
Along the free surface, the pressure is theoretically equal to the atmospheric pressure that
has been taken equal to zero. When convergence was reached for the given approximation
£0' the computed pressure along the free surface varied between 0.063 m and -0.097 m.
At the interior points of the domain, one obtains variations of the same magnitude. The
table 9.1 gives the pressure at some interior points located at the abscissa of the node C.
The table confirms that the fluid pressure becomes rapidly equal to the atmospheric one
and consequently, that the gravity forces mainly influence infiltration flows. Finally, the
discharge through the domain is equal to Q=8.01K (m 3Jh per meter canal). This discharge
is rather important and in practice, measures have to be taken to reduce this discharge.
Xo(m)
9
y
8 2,5
46 7

42 6 2,0
40
+
+ 5

1,5 N
+
+ o 5 10 15 20
30
Figure 9.6
+
+
+ Y(m) p(m)

20 + 5 0,0618
+
10 0,0355
+ 15 -0,0917

10 + 20 -0,1382
+ 25 -0,0879
30 -0,0795
+
+ 35 0,1003
E X
10 40 0,7847

Figure 9.5 Table 9.1

 111

9.4.3. Nonsymmetrical Problem

This second example analyses the infiltration from a nonsymmetrical canal. In the present
case, one has to compute the positions of two free surfaces which will influence each
other.
This second example is shown in figure 9.7 that compares the corresponding analytical
solution (continuous lines) [68] with a solution obtained by the finite element method
(dashed lines) [69] and the present solution obtained by the boundary elements method
(points) [19], [20].
16 quadratic three node elements with a total of 37 nodes and 32 degrees of freedom were
used for the discretization of this problem. For the iterative determination of the free
surface positions, the following values have been chosen for the coefficients of
relationships (9.4) and (9.5).

EO = 0.0005 COAB = 0.25 COnE = 0.10

To start the iterative procedure, one needs to impose two vertical free surfaces. The
figure 9.8 shows for each iteration (N) the values of the abscissas of the points A and E.
The figure 9.8 gives also the evolution of convergence as defined by the coefficient o.
The nearly constant values of 0 around the fifteenth iteration correspond to the transition
from the iterative procedure (d), (e) and.(f) to the procedure described by (g). Figure 9.8
shows the fast convergence of the iterative procedure and that the reciprocal influences of
the two free surfaces do not damage the convergence. Finally, table 9.2 gives the
discharge (q = QIK, with Q the total discharge and K the hydraulic conductivity) per
meter of canal.
Table 9.2 confrrrns that the discharge from the canal is important. Hence, the bottom and
the banks of the canals are often made partially or fully impervious by an appropriate
covering. The third example will deal with such a canal.

Y(m~
XA XE IS
20 19 :-w... 3,00
\ ',IS
80 \ \ 2,75
180
\ \.
.r\',;,
15 X~ 2,50
160 \ \', XE

.i .\ '\~,.-..-- 2,25
75
140
10 ....-..,
! \ \, XA 6'.\
/X E
2,00
120 70
E X{m) 6 I ·\'"..-t 1,75
---/ \
N
A 20 40 60 80 0 5 10 15 20

Figure 9.7 Figure 9.8

Method q(m)
Analytical [68] 67,60
BEM [20] 67,39
FEM [69] 67,05

Table 9.2
 112

9.4.4. Canal with a panlally impervious bottom

This third example studies the infiltration fann a canal with a partially impervious bottom.
In the present case, the canal is symmetrical and so, one only needs to consider the half
domain with an impervious boundary along the axis of symmetry AB. The flat bottom
BC of the canal is pervious and the sloping bottom CD is impervious. This example is
shown in figure 9.9 that compares the corresponding analytical solution (continuous line)
[71] with a solution obtained by the finite difference method (dashed line) [54] and the
present solution by the boundary element method (points) [19], [20]. Figure 9.9 also
shows the variation of the pressure along the axis of symmetry and confirms that the
pressure becomes rapidly very low and thus, the flow nearly vertical.
13 quadratic three node elements with a total of 31 nodes and 26 degrees of freedom have
been used for the discretization of this problem. For the iteration of the free surface
position, the following values have been chosen for the coefficients of relationships (9.4)
and (9.5).
EO = 0.0005 co = 0.1
To start the iterative procedure, one has to impose a vertical free surface (DE) at the
abscissa 0.6 m. The figure 9.10 shows for each iteration (N) the abscissas of the points
D and E and the evolution of the convergence according to the coefficient 0 defined by
(9.6).
Just like the two previous examples, the nearly constant values of the coefficient 0
correspond to the transition from the first to the second iterative procedure. Finally, the
table 9.3 gives the discharge q per meter of canal.
This third example confirms that the results obtained by the present procedure are in good
agreement with the analytical solution.

Y(m) 1,3 X{m) 0

13,96 \ ........----., 2,75
1,2 'Z //"- XE .-....-

~~6
' .
\:
15° 1,1
1,0 ></ ,50

i. XE/ \ ,25
\ 12 \:
\;
0,9
!. "._-
\ 11
I ;/-,\
'\.
'\.
- " , <5
\.
,00

,75
D.,.-.-..- ....-- ~\._
X "
p(m) 10 iA E X N
1,0 0,5 I 1,0 2,0 5 10

Figure 9.9 Figure 9.10

Method q(m)
Analytical [71] 1,25
FDM [54] 1,25
BEM [20] 1,22

Table 9.3
 113

9.5. CONCLUSIONS

InfIltration flows from canals to underlying groundwater are characterized by nearly

vertical free surfaces. This peculiarity requires a very judiciously choice of the iterative
procedure used to detennine the free surlace position.
Previous authors who have applied the finite difference method and the finite element
methods to such problems have had to develop very sophisticated techniques in order to
ensure convergence and stability of the iterative procedure. The boundary element method
on the other hand is simple to apply and only requires an appropriate technique to carry
out the correction of the free surface nodes after each iteration.
The three computational examples shown in this chapter confirm that the present
technique leads to good agreement with the available analytical solutions.
 CHAPTER 10

CONCLUSIONS

The subject of the studies presented in this book is the application of the boundary
element method to two dimensional potential problems described by the Laplace equation,
with special reference to the study of free surface groundwater problems. These studies
lead to the following conclusions:
1. The application of the boundary element method to the solution of problems described
by the Laplace equation requires special care of the resulting numerical integrations.
The use of special numerical integration methods is essential when the collocation point
is located on the element under consideration. For linear elements, four integration
points per element are sufficient. For elements with a strong curvature, one has to use
higher order elements with a maximum of integration points. Although the integrals
relative to the potential are not singular when the collocation point is located on the
element under consideration, it is suitable for accuracy to use an integration method
well adapted to the special variations of the corresponding integrands. As has been
shown, this method depends on the position of the element in the reference system.
Finally, the results provided by these special integration methods have been compared
with analytical solutions in order to show the high accuracy they give, especially with
respect to the standard Gauss integration method.
2. The different applications presented in this book show that the use of higher order
elements (elements with 3 or 4 nodes) furthers the accuracy of the solution. Indeed,
these elements not only allow a more accurate discretization of the geometry, but also
describe more accurately the potential field.

3. The examples presented here show that using the boundary element method, one
obtains accurate results for the potential as well as for the normal velocities to the
boundaries. This is due to the fact that the weighting function w of the boundary
element method is a fundamental solution of the Laplace equation. However, it has
been shown that the accuracy of the potential is always higher than those of the normal
velocity.

4. When the geometry of the domain or its properties require its division into sub-regions,
these sub-regions will be separated by interfaces. Then, at the junction point of several
interfaces and at the junction point of an interface with a boundary, special numerical
problems occur. This study shows that these numerical problems are due to the fact
that the resulting linear system becomes singular at these points. The method
developed in chapter 6 involving extra equations provides good results. The examples
considered in that chapter allow to formulate some recommendations concerning the
localization of the collocation point of the additional equation. In particular, one can
conclude that the errors due to a non optimal choice of the position of the additional
 115

collocation point only affect the results in the vicinity of the point requiring the
additional equation. Funhennore, one can conclude that these additional equations
must be written with regard to sub-regions next to a maximum of other sub-regions of
the domain. Finally, this method is easily generalizable to three dimensional problems
and to other problems requiring the division of the problem into several sub-regions.

5. The case of sub-regions separated by interfaces having specific properties has also been
studied. In these cases, any numerical problem occurs at the junction points of several
interfaces or of an interface with a boundary. The examples presented demonstrate the
applicability of the presented method to these cases.
6.When analytical solutions were available, the results provided by the boundary element
method have been compared against them. These comparisons have shown that, the
boundary element method results are accurate for the potential as well as for the
velocities.
7. The use of mathematical models to describe physical problems often leads to infinite
velocities at some boundary points. They are due to the fact that the underlying
simplifications and linearizations lead to a singular potential field at these points. Such
infmite velocities have obviously no physical meaning, but their presence give rise
numerical problems. If the finite element method is used to solve such problems, the
discretization of the domain must be refined in the vicinity of such points, not only to
obtain accurate results there, but also to obtain accurate results on the domain. Several
examples considered in this study present such points and one can conclude that the
application of the boundary element method to such problems does not require a special
discretization near those points and that the possible errors only affect the accuracy of
the results there.

8. Whenthe boundary element method is applied to solve free surface flows, one has to
choose between the elimination of the potential or of the nonnal velocity along the
moving free surface. In the past, authors who have applied the boundary element
method to solve such problems have proposed the elimination of the potential along the
free surface. This produces stability and convergence problems requiring some
smoothing procedures along the free surface. In chapter 5, it has been shown that the
nonnal velocities are more sensitive to boundary geometry variations than the
potentials. In accordance with this important conclusion, the elimination of the normal
velocities at the free surface nodes has been used in this book. The application of this
method has always given satisfactory results without the use of any smoothing
procedure along the moving free surface.

9.Systematic tests have been carried out in the unsteady case in order to fonnulate some
recommendations about the discretization of the boundaries, the time step and the value
of the ecoefficient

10. infIltration flows are mainly governed by the gravity force. That is why they are
very sensitive as shown by earlier applications of the finite difference method and the
finite element method to such problems. The application of these methods to such
problems imposes the development of special numerical procedures. To apply the
boundary element method to such problems, it is sufficient to develop a simple special
iterative procedure for the correction of the position of the free surface nodes, which
takes the special features of this type of flows into account. As shown by the
examples. good results have been obtained.

Finally, it is important to point out that the methods developed in this book are not only
applicable to steady and unsteady porous media flows. but also to many other problems
described by the Laplace equation.
 CHAPTER 11

REFERENCES

[1] ABRAMOWITZ M. and STEGUN I., Handbook of Mathematical Functions,

Dover Publications Inc., (1965).
[2] ALARCON A., MARTIN A. and PARIS F., Improved Boundary Elements in
Torsion Problems, Proc. Int. Conf. on Recent Advances in B.E.M.,
Southampton University, pp. 149-165, (1978).
[3] BANERJEE P. K., BUTTERFIELD R. and TOMLIN G.R., Boundary Element
Methods for two-dimensional problems of transient ground water flow,
International Journal for Numerical and Analytical Methods in Geomechanics,
5(1), pp. 15-31, (1981).
[4] BARENDS F.B.J., Non-linear Groundwater Flow Simulation, 3th International
Conference on Finite Elements in Water Resources, pp. 2.56 - 2.65, (1980).
[5] BEAR J., Dynamics of Fluids in Porous Media, American Elsevier Environmental
Science Series, New-York, (1972).
[6] BEAR J., Hydraulics of Groundwater, Mc Graw-Hill Series in Water Resources
and Environmental Engineering; London, (1979).
[7] BREBBIA C.A. (Ed.), Recent Advances in Boundary Element Methods,
Proceedings of the fIrst Int. Conf. on BEM, Southampton University, Pentech
Press, London, (1978). '
[8] BREBBIA C.A. and WALKERS S., Boundary Element Techniques in
Engineering, Newnes-Butterworths, London, (1980).
[9] BREBBIA C.A., TELLES J.C.P. and WROBEL L.C., Boundary Element
Techniques, Springer Verlag, (1984).
[10] BREBBIA C.A. and DOMINGUEZ J., Boundary Elements - An Introductory
Course, Computational Mechanics Publications and Mc Graw-Hill Book
Company, (1989).
[11] BRUCH E., Resolution par elements frontieres des ecoulements permanents en
milieux poreux It surface libre eventuellement indeterminee, Travail de fIn
d'etudes, Universite de Liege, (1985).
[12] BRUCH E., GRILLI S. and LEJEUNE A., Computation of the Fluid Flow in
Zoned Anisotropic Porous Media and Determination of the Free Surface Seepage,
Proceedings of the 8th Boundary Element Conference, Tokyo, pp. 889-903,
(1986).
[13] BRUCH E. and GRILLI S., Computation of the Transient Flow in Zoned
Anisotropic Porous Media by the Boundary Element Method, Notes on
Numerical Fluid Mechanics, Vol. 21, pp. 43-51, (1987).
[14] BRUCH E. and GRILLI S., Computation of the Transient Seepage Problem in
Anisotropic Porous Media by the Boundary Element Method, Proceedings of the
9th Boundary Element Conference, Stuttgart, Vol. 3, pp. 329-341, (1987).
[15] BRUCH E., An Effective Solution of the Singularities at Multi-Domain Points for
Laplace Problems, Proceedings of the European Boundary Element Meeting,
Bruxelles, pp. 2.65-2.91, (1988).
 117

[16] BRUCH E., Application of the BEM to the Steady and Unsteady Two-
Dimensional Phreatic Groundwater Flow, Proceedings of the 10th Boundary
Element Conference, Southampton, Vol. 2, pp. 115-128, (1988).
[17] BRUCH E. and LEJEUNE A., An Effective Solution of the Numerical Problems
at Multi-domain Points for Anisotropic Laplace Problems, Proceedings of the 11 th
Boundary Element Conference, Cambridge, Vol. 2, pp. 2-14, (1989).
[18] BRUCH E. and LEJEUNE A., Applications of the Boundary Element Method to
Groundwater Flows, HYDROSOFl', Vol. 3, Nr. 2, pp. 73-87, (1990).
[19] BRUCH E., Contribution ala resolution de l'equation de Laplace par la methode
des elements frontieres avec applications aux ecoulements dans les milieux
poreux, these de doctorat, Universite de Liege, Belgique, (19989).
[20] BRUCH E., LEJEUNE A. and MERGEAI B., Application of the BEM to free
surface seepage from canals to anisotropic soils, Proceedings of the International
Boundary Element Symposium, Nice, pp. 249-259, (1990).
[21] BRUCH E. and LEJEUNE A., Division of Domain in several sub-regions for
anisotropic Laplace problems using BEM, BETECH 90, Delaware, Vol. 1, pp.
195-208, (1990).
[22] BRUCH E. and LEJEUNE A., Treatment of Comer Node Problems in Laplace
Fields by BEM, BETECH 90, Delaware, Vol. 1, pp. 131-141, (1990).
[23] BRUCH E., LEJEUNE A. and MERGEAI B., Study of unsteady flows in
porous media governed by Laplace equation, BETECH 90, Delaware, Vol. 1, pp.
185-194, (1990).
[24] BRUCH E., LEJEUNE A. and hlERGEAI B., Computer Simulation of the use
of Geotextiles in Hydraulic Works, Proceedings of the 12th Boundary Element
Conference, Sapporo, Vol. 2, pp. 503-512, (1990).
[25] BRUCH J.C., Jr and STREET R.L., Free Surface Flow in Porous Media,
Journal of the Irrigation and Drainage Division, ASCE, Vol. 93, n03, pp. 125-
145, (1967).
[26] BRUCH J.C., Jr, FREDERICK C.S. and SLOSS J.M., Seepage from a
Trapezoidal and a Rectangular Channel using Variational Inequalities, Journal of
Hydrology, Vol. 63, pp. 247-260, (1978).
[27] CHANG C.S., Application of Boundary Element Method in Transient Flow
Problems with Moving Boundaries, Proceedings of the 8th Boundary Element
Conference, Tokyo, Vol. 2, pp. 849-858, (1986).
[28] CHENG R.T. and LI C., On the Solution of Transient Free Surface Flow
Problems by the Finite Element Method, Journal of Hydrology, Vol. 20, pp. 49-
63, (1973).
[29] DEFOURNY M., Original Treatment of Singularities in a B.E. Program applied
to Electrostatics, Proceedings of the 8th Boundary Element Conference, Tokyo,
pp. 259-267, (1986).
[30] DEFOURNY M., Singular Point Theory in Laplace Field, Proceedings of the
10th Boundary Element Conference, Southampton, Vol.2, pp. 165-180, (1988).
[31] DE MARSILY G., Quantitative hydrogeology groundwater hydrology for
engineers, Academic Press, New-York, (1986).
[32] DESAI C.S., Finite Element Methods for Flow in Porous Media, Finite Elements
in Fluids, Vol. 1, Viscous Flow and Hydrodynamics, John Wiley and Sons,
Chapter 8, pp. 157-182, (1975).
[33] FINNEMORE E.J. and PERRY B., Seepage Through an Earth Dam Computed
by the Relaxation Technique, Water Resources Res., pp. 1059-1067, (1968).
[34] GlODA G. and GENTILE C., A Nonlinear Programming Analysis of
Unconfined Steady-state Seepage, International Journal for Numerical and
Analytical Methods in Geomechanics,Vol. 11, pp. 283-305, (1987).
[35] GRA Y L.J., Numerical Experiments with a Boundary Element Technique for
Comers, Proceedings of the 11 th Boundary Element Conference, Cambridge,
Vol. 1, pp. 243-251, (1989).
 118

[36] GR1LLI S., Etude experimentale et numerique du comportement hydrodynamique

de grandes portes flottantes automotrices pour ecluses et barrieres marees-
tempetes, These de doctoral, Universite de Liege, (1985).
[37] HALEK V. and SVEC J., Groundwater Hydraulics, Developments in Water
Science, Vol. 7, Elsevier Scientific Publishing Company, (1979).
[38] HANTUSH M.S., Growth of Ground-Water Ridge in Response to Deep
Percolation, Symposium on Transient Ground Water Hydraulics, Fort Collins,
Colorado, (1963).
[39] JASWON M. and SYMM G.T., Integral Equation Methods in Potential Theory
and Elastostatics, Academic Press, London, (1977).
[40] JEPPSON R.W., Seepage from Ditches - Solution by Finite Differences, Journal
of Hydraulics Division, ASCE, Vol. 94, nOl, pp. 259-283, (1968).
[41] KATZ C., Some Improvements in 2D Boundary Elements Using Integration by
Pans, Proceedings of the 4th Boundary Element Conference, Southampton, pp.
72-85, (1982).
[42] KELLOGG F.H., Investigation of Drainage Rates Affecting Stability of Earth
Dams, American Society of Civil Engineers, Transactions, Vol. 113, pp. 1261-
1309, (1941).
[43] KELMANSON M.A., An Integral Equation Method for the Solution of Singular
Slow Flow Problems, Journal of Comput. Physics, Vol. 51, pp. 139-158,
(1983).
[44] LAFE O.E., MONTES J.S., CHENG A.H.D., LIGGETI J.A. and LIU P.L.-
F., Singularities in Darcy Flow Through Porous Media, Journal of Hydr. Div. ,
ASCE 106 (HYG), pp. 977-997, (1980).
[45] LENNON G.P., Unconfined Groundwater Flow, Topics in Boundary Element
Research, Vol. 4, Chapter 8, pp~ 155-172, (1987).
[46] LIGGETI J.A. and CUNGE J.A., Numerical Methods of Solution of the
Unsteady Flow Equation, Unsteady Flow in Open Channels, Chapter 4, Fort
Collins, Colorado, Water Resources Publications, (1975).
[47] LIGGETI J.A., Location of Free Surface in Porous Media, Journal of the
Hydraulics Division, ASCE, Vol. 103, N°4, pp. 353-365, (1977).
[48] LIGGETI J.A. and LIU P.L.-F., The Boundary Integral Equation Method for
Porous Media Flow, George AlIen and Unwin, London, (1983).
[49] LIU P .L.-F. and LIGGETI J.A., Boundary Solutions to two Problems in Porous
Media, Journal of Hydraulics Division, ASCE, Vol. 105, N° 3, pp. 171-183,
(1979).
[50] MARINO M.A., Hele-Shaw Model Study of the Growth and Decay of
Groundwater Ridges, Journal of Geophysical Research, Vol. 72, pp. 1195-1205,
(1967).
[51] MA THERON J., Elements pour une theorie des milieux poreux, Masson & Cie,
Paris, (1967).
[52] MC CORQUODALE J. A., Variational Approach to Non-Darcy Flow, Journal of
Hydraulics Division, ASCE, Vol. 96, W11, (1970).
[53] MIGEOT J.L., Some Experimental Results with Non-Conforming Elements,
Proceedings of the 2th Boundary Element Technology Conference, Computational
Mechanics Publication, pp. 807-819, (1986).
[54] MIRNATEGHI A. and BRUCH J.C., Jr., Seepage from Canals having Variable
Shape and Partial Lining, Journal of Hydrology, Vol. 64, pp. 239-265, (1983).
[55] MITRA A.K. and INGBER M.S., Resolving of Difficulties in the BIEM caused
by Geometric Corners and Discontinuous Boundary Conditions, Proceedings of
the 9th Boundary Element Conference, Stuttgart, Vol. 1, pp. 519-532, (1987).
[56] MIZUMURA K., Time Dependent Ground Flow Analysis, Topics in Boundary
Element Research, Vol. 4, Chapter 7, pp. 145-154, (1987).
[57] MUSKAT M., The Flow of Homogeneous Fluids Through Porous Media, J.W.
Edwards Inc., Ann Arbor, Michigan, (1946).
 119

[58] NEUMAN S.P. and WITIffiRSPOON P.A., Finite Element Method of Analysing
steady State Seepage With Free Surface, Water Resources Research, Vol. 6, N°
3, pp. 889-897, (1970).
[59] NEUMAN S.P., NARASIMHAN T.N. and WITHERSPOON P.A., Application
of mixed explicit-implicit Finite Element Method to Nonlinear Diffusion-type
Problems, International Conference on Finite Elements in Water Resources,
Princeton University, (1976).
[60] OUAZAR D. and BENMANSOUR N., Location of Free Surface in Porous
Media by the Regular Boundary Element Method, Proceedings of the 2nd
Boundary Element Technology Conference, MIT, (1986).
[61] OUAZAR D., NAMLI A., SAIDI A. and BREBBIA C.A., An Efficient
Boundary Element Formulation for Zoned Porous Media Problems with Free
Surfaces, Proceedings of the 9th Boundary Element Conference, Stuttgart,
(1987).
[62] PARIS F., MARTIN A. and ALARCON E., Potential Theory, Progress in
Boundary Element Methods, Vol. 1, Pentech Press, London, pp. 45-83, (1981).
[63] PATTERSON C. and SHEIKH M.A., Regular Boundary Integral Equations for
Fluid Flow, Int. J. Num. Meth. in Fluids, Vol. 2, pp. 239-251, (1982).
[64] PATTERSON C. and SHEIKH M.A., Interelement Continuity in the Boundary
Element Method, Chapter 6, Topics in Boundary Element Research, Vol. 1, pp.
123-141, (1985).
[651 PINA H.L.G., FERNANDES J.L.M. and BREBBIA C.A., The Effect of Mesh
Refinement in the Boundary Element Solution of Laplace's Equation With
Singularities, Proceedings of the 3th Boundary Element Conference, Irvine, pp.
443-456, (1981). '
[66] POLUBARINOVA-KOCHINA P.Y., Theory of Ground Water Movement,
Translated from the Russian by J.M.R. de Wiest, Princeton University Press,
(1962).
[67] RADOJKOVIC M.M., Application of Boundary Element Method in Seepage
Problems, New Developments in Boundary Element Methods, CML
Publications, pp. 151-161, (1980).
[68] REDDY A.S. and BASU U., Seepage from Trapezoidal Canal in Anisotropic
Soil, Journal of the Irrigation and Drainage Division, ASCE, Vol. 102, N° 3, pp.
349-361, (1976).
[69] REMAR J., BRUCH J.C., Jr and SLOSS J.M., Free Surface Seepage from
Nonsymmetric Channels, International Journal for Numerical Methods in
Engineering, Vol. 18, pp. 1547-1563, (1982).
[70] RICCARDELLA P., An Implementation of the Boundary Integral Techniques for
Plane Problems in Elasticity and Elasto-Plasticity, Ph. D. Thesis, Carnegie-
Mellon University, Pittsburgh, (1973).
[71] SUBRAMANJA K., MADRA V M.R. and MISHRA G.D., Studies of Seepage
from Canals with Partial Lining, Journal of Hydraulics Division, ASCE, Vol. 99,
N° 12, pp. 2333-2351, (1973).
[72] TAKUMA T., KOUNA T. and MATSUDE H., Field Behaviour near Singular
Points in Composite Dielectric Arrangements, IEEE Trans. on Electr. Insul., Vol.
13, N° 6, pp. 426-435, (1978).
[73] TODSON M., On the Solution of Transient Free Surface Flow Problems in
Porous Media by Finite-difference Methods, Journal of Hydrology, Vol. 12, pp.
177-210, (1971).
[74] WROBEL L.C. and BREBBIA C.A., Time-dependent potential problems,
Progress in Boundary Element Methods, Vol. 1, pp. 192-212, (1985).
[75] WROBEL L.C., Potential Problems, Topics in Boundary Element Research,
Chapter 8, Vol. 3, pp. 235-264, (1987).
[76] ZHENG J.D., Location of Free Surface for Zoned Problems - A New Economical
Boundary Element Method, Proceedings of the 5th Boundary Element
Conference, Hiroshima, (1983).
 Acknowledgements

This book presents part of the research on boundary element methods carried out
during the last years under the leadership of Prof. A. LEJEUNE at University of Liege
in Belgium.
The author is indebted to Dr. C. A. BREBBIA for his useful suggestions and editing of
the manuscript.
The tables 4.1 and 4.2 of this book come from the book "Boundary Elements - An
Introductory Course" of C.A. BREBBIA and J. DOMINGUEZ and are used with their
permission.
 Lecture Notes in Engineering
Edited by C.A. Brebbia and S.A. Orszag

Vol. 40: R. Borghi, S. N. B. Murhty (Eds.) Vol. 49: J. P. Boyd

Turbulent Reactive Flows Chebyshev & Fourier Spectral Methods
VIII, 950 pages. 19a9 XVI, 79a pages. 19a9

Vol. 41: W. J. Lick Vol. 50: L. Chibani

Difference Equations Optimum Design of Structures
from Differential Equations VIII, 154 pages. 19a9
X, 2a2 pages. 19a9
Vol. 51: G. Karami
Vol. 42: H. A. Eschenauer, G. Thierauf (Eds.) A Boundary Element Method for
Discretization Methods Two-Dimensional Contact Problems
and Structural Optimization - VII, 243 pages. 19a9
Procedures and Applications
Proceedings of a GAMM-Seminar Vol. 52: Y. S. Jiang
October 5-7, 19aa, Siegen, FRG Slope Analysis Using
XV, 360 pages. 19a9 Boundary Elements
IV, 176 pages. 19a9
Vol. 43: C. C. Chao, S. A. Orszag, W. Shyy (Eds.)
Recent Advances in Computational Vol. 53: A. S. Jovanovic,
Fluid Dynamics K. F. Kussmaul, A. C. Lucia,
Proceedings of the US/ROC (Taiwan) Joint P. P. Bonissone (Eds.)
Workshop in Recent Advances in Expert Systems in Structural
Computational Fluid Dynamics Safety Assessment
V, 529 pages. 19a9 X, 493 pages. 19a9

Vol. 44: R. S. Edgar Vol. 54: T. J. Mueller (Ed.)

Field Analysis and Low Reynolds Number
Potential Theory Aerodynamics
XII, 696 pages. 19a9 V, 446 pages. 19a9

Vol. 45: M. Gad-el-Hak (Ed.) Vol. 55: K. Kitagawa

Advances in Fluid Mechanics Boundary Element Analysis
Measurements of Viscous Flow
VII, 606 pages. 19a9 VII, 136 pages. 1990

Vol. 46: M. Gad-el-Hak (Ed.) Vol. 56: A. A. Aldama

Frontiers in Experimental Filtering Techniques for
Fluid Mechanics Turbulent Flow Simulation
VI, 532 pages. 19a9 VIII, 397 pages. 1990

Vol. 47: H. W. Bergmann (Ed.) Vol. 57: M. G. Donley, P. D. Spanos

Optimization: Methods and Applications, Dynamic Analysis of Non-Linear
Possibilities and Limitations Structures by the Method of
Proceedings of an International Seminar Statistical Quadratization
Organized by Deutsche Forschungsanstalt fUr VII, la6 pages. 1990
Luft- und Raumfahrt (DLR), Bonn, June 19a9
IV, 155 pages. 19a9 Vol. 5a: S. Naomis, P. C. M. Lau
Computational Tensor Analysis
Vol. 4a: P. Thoft-Christensen (Ed.) of Shell Structures
Reliability and Optimization XII, 304 pages. 1990
of Structural Systems 'aa
Proceedings of the 2nd IFIP WG 7.5 Conference
London, UK, September 26-2a, 19aa
VII, 434 pages. 19a9

For information about Vols. 1-39 please contact your bookseller or Springer-Verlag.
 Lecture Notes in Engineering
Edited by C.A. Brebbia and S.A. Orszag

Vol. 59: K. P. Herrmann, Z. S. Olesiak (Eds.) Vol. 68: T. G. B. DeFigueiredo

Thermal Effects in Fracture A New Boundary Element Formulation
of Multiphase Materials in Engineering
Proceedings of the Euromech Colloquium 255 IX, 198 pages. 1991
October 31 - November 2, 1989, Paderborn, FRG
VII, 247 pages. 1990 Vol. 69: In preparation

Vol. 60: U. Schumann (Ed.) Vol. 70: E. K. Bruch

Air Traffic and the Environment - The Boundary Element Method
Background, Tendencies and Potential for Groundwater Flow
Global Atmospheric Effects II, 120 pages. 1991
Proceedings of a DLR International Colloquiul\1
Bonn, Germany, November 15/16, 1990
VI, 170 pages. 1990

Vol. 61: A. Der Kiureghian,

P. Thoft-Christensen (Eds.)
Reliability and Optimization
of Structural Systems '90
Proceedings of the 3rd IFIP WG 7.5 Conference
Berkeley, Califomia, USA, March 26-28,1990
VIII, 405 pages. 1991

Vol. 62: Z. Zhao

Shape Design Sensitivity Analysis and Optimization
Using the Boundary Element Method
VIII, 192 pages. 1991

Vol. 63: H. A. Eschenauer, C. Mattheck,

N. Olhoff (Eds.)
Engineering Optimization in Design Processes
Proceedings of the International Conference
Karlsruhe Nuclear Research Center, Germany
September 3-4,1990
XIV, 355 pages. 1991

Vol. 64: A. Elzein

Plate Stability by Boundary
Element Method
VII, 205 pages. 1991

Vol. 65: J, Holnicki-Szulc

Virtual Distortion Method
VIII, 176 pages. 1991

Vol. 66: M. A. Jabri

An Artificial Intelligence Approach
to Integrated Circuit Floorplanning
XIII, 149 pages. 1991

Vol. 67: S. Takahashi

Elastic Contact AnalYSis
by Boundary Elements
VI, 217 pages. 1991