Numerical Methods Book PDF

Boundary Element Method for Impedance and Optical Tomography
Jan Sikora
30
8 222 217 104 103 79 216 221 100
218 223 219 231 225
13 113 224 14 12
116 115 118 114 232 18 119 117 19 17
25
20
15
10
0 0 5 10 15 20 25 30
Warsaw 2007
Contents
1 Twodimensional potential problems
19
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 1.2 Laplace equation . . . . . . . . . . . . . . . . . . . . . . . . . 21 1.2.1 Division of the boundary into constant boundary elements . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 Numerical integration of the kernels . . . . . . . . . . . 25 Division of the boundary into linear boundary elements 30 Numerical integration of the kernels . . . . . . . . . . . 33 Division of the boundary into quadratic boundary elements . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 Numerical integration of the kernels . . . . . . . . . . . 41 Numerical integration of the c(r) coecient . . . . . . 47 Internal points calculation . . . . . . . . . . . . . . . . 48 Symbolic calculation . . . . . . . . . . . . . . . . . . . 51
1.2.2 1.2.3 1.2.4 1.2.5
1.2.6 1.2.7 1.2.8 1.2.9
1.3 Diusion equation . . . . . . . . . . . . . . . . . . . . . . . . . 53 3
4 1.3.1 1.3.2 1.3.3
CONTENTS Treatment of singularity . . . . . . . . . . . . . . . . . 55 Internal points calculation . . . . . . . . . . . . . . . . 56 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 57
1.4 Diusion equation in frequency domain . . . . . . . . . . . . . 57 1.4.1 1.4.2 Treatment of singularity . . . . . . . . . . . . . . . . . 61 Internal points calculation . . . . . . . . . . . . . . . . 61
1.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 1.5.1 1.5.2 1.5.3 Cartesian coordinate system . . . . . . . . . . . . . . . 62 Polar coordinate system . . . . . . . . . . . . . . . . . 64 Distributed source for a diusion model for light transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 Point source located on the boundary for a diusion model for light transport . . . . . . . . . . . . . . . . 70 Point source located inside the region for a diusion model for light transport . . . . . . . . . . . . . . . . 76 Comparison FEM and BEM results of calculation . . . 80 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . 81
1.5.4
1.5.5
1.5.6 1.5.7
1.6 Anisotropic medium . . . . . . . . . . . . . . . . . . . . . . . 81 1.6.1 1.6.2 1.6.3 Anisotropy model . . . . . . . . . . . . . . . . . . . . . 82 Treatment of singularity . . . . . . . . . . . . . . . . . 84 Comparison FEM and BEM results . . . . . . . . . . . 86
CONTENTS
1.7 Galerkin formulation of boundary integral equations . . . . . . 88 1.7.1 1.7.2 Analytical integrations of coincident integrands . . . . 91 Numerical integrations of coincident integrands . . . . 93
2 Three-dimensional potential problems
97
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 2.2 Singular and nearly singular integrals . . . . . . . . . . . . . . 99 2.3 Governing equations . . . . . . . . . . . . . . . . . . . . . . . 101 2.4 Zeroorder interpolation functions . . . . . . . . . . . . . . . . 103 2.4.1 2.4.2 2.4.3 Jacobian . . . . . . . . . . . . . . . . . . . . . . . . . . 104 Integration of nonsingular integrals over the triangle . 106 Integration of singular integrals . . . . . . . . . . . . . 107 . . . . . . . . . . . . . . . 108
2.5 Firstorder interpolation functions
2.6 Secondorder interpolation functions . . . . . . . . . . . . . . 112 2.6.1 2.6.2 2.6.3 2.6.4 Triangular boundary elements . . . . . . . . . . . . . . 112 Numerical integration of singular integrals . . . . . . . 114 Quadrilateral boundary elements . . . . . . . . . . . . 120 Integration of nonsingular integrals over the square . . . . . . . . . . . . . . . . . . . . . . 121 Integration of singular integrals over the square . . . . 122
2.6.5
2.7 Treatment of Boundary Conditions . . . . . . . . . . . . . . . 130
6 2.7.1 2.7.2 2.7.3 2.7.4
CONTENTS Dirichlet boundary conditions . . . . . . . . . . . . . . 130 Neumann boundary conditions . . . . . . . . . . . . . . 130 Robin boundary conditions . . . . . . . . . . . . . . . . 131 Mixed boundary conditions . . . . . . . . . . . . . . . 131
2.8 Nonhomogeneity . . . . . . . . . . . . . . . . . . . . . . . . . 131 2.9 Index mismatched diusive/diusive interfaces . . . . . . . . . 133 2.9.1 2.9.2 Approximate interface conditions . . . . . . . . . . . . 135 Complete interface conditions . . . . . . . . . . . . . . 140
2.10 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . 142 2.10.1 Cube . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 2.10.2 Two concentric spheres . . . . . . . . . . . . . . . . . . 143 2.10.3 The proximity eect . . . . . . . . . . . . . . . . . . . 157 2.10.4 Results for spatially nonhomogeneous region in 2D space . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
3 Diusion model for light transport in the frequency domain
161
3.1 Governing equations . . . . . . . . . . . . . . . . . . . . . . . 162 3.1.1 3.1.2 3.1.3 Two dimensional space . . . . . . . . . . . . . . . . . . 163 Three dimensional space . . . . . . . . . . . . . . . . . 163 The Boundary Element Method . . . . . . . . . . . . . 163
CONTENTS
3.2 Numerical Implementation . . . . . . . . . . . . . . . . . . . . 164 3.2.1 3.2.2 Jacobian . . . . . . . . . . . . . . . . . . . . . . . . . . 166 Matrix Assembly . . . . . . . . . . . . . . . . . . . . . 166
3.3 Numerical integration of singular integrals in 3D . . . . . . . . 168 3.3.1 3.3.2 3.3.3 Mapping formula for triangular constant element . . . 168 Isoparametric triangular quadratic element . . . . . . . 171 Isoparametric quadrilateral quadratic element . . . . . 175
3.4 Results for 3D . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 3.4.1 3.4.2 3.4.3 Validation of numerical results . . . . . . . . . . . . . . 179 Measures of the accuracy . . . . . . . . . . . . . . . . . 179 Quadratic meshing . . . . . . . . . . . . . . . . . . . . 181
3.5 Multilayered model of the neonatal head . . . . . . . . . . . . 182
4 Light propagation in diusive media with nonscattering regions 187 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 4.2 Governing equations for nonscattering sphere embedded in a diusive spherical region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 4.2.1 4.2.2 The Boundary Element Method . . . . . . . . . . . . . 189 Matrix form of integral equations . . . . . . . . . . . . 193
4.3 The Form Factor . . . . . . . . . . . . . . . . . . . . . . . . . 194
8 4.3.1 4.3.2
CONTENTS The Form Factor calculated analytically . . . . . . . . 194 The Form Factor calculated numerically . . . . . . . . 195
4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 4.4.1 4.4.2 The steady state . . . . . . . . . . . . . . . . . . . . . 196 The frequency domain solution 100MHz . . . . . . . 198
4.5 Nonscattering gap between two diusive regions of a spherical shape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 4.5.1 4.5.2 4.5.3 4.5.4 4.5.5 4.5.6 Form Factor calculated analytically . . . . . . . . . . . 199 Visibility function calculated analytically . . . . . . . . 200 Visibility function calculated numerically . . . . . . . . 202 Point in triangle test . . . . . . . . . . . . . . . . . . . 204 Integral equations . . . . . . . . . . . . . . . . . . . . . 206 Matrix form of integral equations . . . . . . . . . . . . 207
4.6 Results for the void gap . . . . . . . . . . . . . . . . . . . . . 208 4.6.1 4.6.2 The steady state . . . . . . . . . . . . . . . . . . . . . 208 The frequency domain solution 100MHz . . . . . . . 208
4.7 Consistency checks for BEM solutions . . . . . . . . . . . . . . 209 4.7.1 4.7.2 Diusion equation case . . . . . . . . . . . . . . . . . . 210 RadiosityDiusion equation case . . . . . . . . . . . . 212 . . . . . . . . . . . . . . . . . . . . 213
4.8 Nuuttis 2D test example
CONTENTS 4.8.1 4.8.2 4.8.3 4.8.4
9 Analytical solution for diusive boundary conditions . 214 Analytical solution for P1 boundary conditions . . . . 221 Numerical results . . . . . . . . . . . . . . . . . . . . . 223 Multilayered neonatal head model with the CSF layer . 224
4.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
5 BEM formulation for thin layers
229
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229 5.2 Standard BEM formulation . . . . . . . . . . . . . . . . . . . 230 5.3 Modication for closely spaced surfaces . . . . . . . . . . . . . 232 5.4 Integration of singular integrals . . . . . . . . . . . . . . . . . 235 5.5 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
6 Wavelet based techniques for CPU time reduction
239
6.1 Discrete wavelet transform . . . . . . . . . . . . . . . . . . . . 242 6.2 Time acceleration . . . . . . . . . . . . . . . . . . . . . . . . . 244 6.3 DWT with permutations . . . . . . . . . . . . . . . . . . . . . 245 6.4 DWT when the size of the coecient matrix is not a number equal to 2n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
10 7 FEMBEM coupling
CONTENTS 251
7.1 2D space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252 7.1.1 Incorporating the BE equations to the FE ones . . . . 253
7.2 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . 257 7.2.1 7.2.2 7.2.3 Simple benchmark problem . . . . . . . . . . . . . . . 257 Two squares one immersed in the other . . . . . . . . . 262 Concentric circles . . . . . . . . . . . . . . . . . . . . . 268
7.3 2D void comparison . . . . . . . . . . . . . . . . . . . . . . . . 274 7.3.1 7.3.2 7.3.3 Boundary conditions at diusive/nondiusive interfaces275 P1 boundary conditions . . . . . . . . . . . . . . . . . 276 Diusive boundary conditions . . . . . . . . . . . . . . 277
8 Miscellaneous
279
8.1 Introduction to FEM mixed formulation . . . . . . . . . . . . 279 8.2 Mixed formulation for Laplace equation . . . . . . . . . . . . . 279 8.2.1 8.2.2 Discretization of mixed forms . . . . . . . . . . . . . . 280 Robin boundary conditions . . . . . . . . . . . . . . . . 281
8.3 Mixed formulation for Diusion equation . . . . . . . . . . . . 284 8.3.1 Three nodes triangle with three degrees of freedom P1 triangle . . . . . . . . . . . . . . . . . . . . . . . . 285 Three nodes triangle with the middle sides nodes . . . 286
8.3.2
CONTENTS 8.3.3
11 Three nodes triangle with a bubble function at the barycenter P1 + triangle . . . . . . . . . . . . . . . . . 287
12
CONTENTS
List of symbols
represents any potential function (electric potential [V] in Electrostatics, temperature [0 C] in heat conduction problems or photon density in Diusive Optical Tomography. T temperature [0 C] or [0 K].
r) q normal derivative of ( ( ). n
G the Green function. r position vector. n unit outward vector normal to the boundary. x, y, z cartesian coordinate system. r, , z or , , z cylindrical coordinate system. domain under consideration. or S surface of the domain . or local coordinate system. Ni or Mi shape functions of the i th node. J ( ) Jacobian of transformation. [A] and [B] matrices containing integrals of the Green function normal derivative and Green function respectively. 13
14 R distance between load and observation points.
CONTENTS
wi weight function (for example in the GaussLegendre quadrature). L length of the boundary element. c(r) coecient of integral equation. k wave number [cm1 ]. D diusion coecient [cm]. K0 modied Bessel function of the second kind and zero order. K1 modied Bessel function of the second kind and rst order. L[] Laplace transform. i= 1.
H vector of magnetic eld intensity [A/m]. J vector of current density [A/m2 ]. the angular frequency [ s1 ]. f frequency [1/s]. material conductivity [1/( m)]. magnetic permeability [H/m]. In modied Bessel function of the rst kind and the n th order. A coecient depending on refractive index. c speed of light [mm/ps]. a or absorbing coecient [mm1 ]. s scattering coecient [mm1 ]. s reduced scattering coecient [mm1 ].
CONTENTS Ld diusion length [mm]. Dirac delta function. ni or i refracting index of the i th diusive region. Cn refracting index mismatch. G radiocity kernel. V Boolean visibility function. g Form Factor [1/m2 ]. Uco visibility cuto angle. s directional vector. W onedimensional wavelet transform. threshold (truncation level). F functional. coecient equal to 2 or to 3 depending on the space dimensionality.
15
16
CONTENTS
Introduction
There is a lot of excellent books devoted the Boundary Element Method (BEM). That is why I would not dare to write one more. The main goal of this book is only to show the problems of modelling the forward problems of Diusive Optical Tomography (DOT) or Electrical Impedance Tomography (EIT) when the Boundary Element is used in 2D or 3D space. The modern Inverse Problems are solved by the iterative process in which we have to solve many times the forward problems for which the design parameters are improved due to the Sensitivity Analysis (SA). The scope of this book is limited to the forward problems and methods presentation of their eective solution in 2D and 3D space with the aid of the Boundary Element Method. The special attention was paid to the basic of the BEM because majority of the monographes devoted this method, do not pay enough attention to the problems of singular integrals integrations. This problem is already solved by many famous authors ( maybe with the exception of the 3D Galerkin approach) but is spread in the literature, in most cases, not easily accessible. That is why it seems to be reasonable to present this problem in full details in this work. A lot of attention was devoted to the problem of modelling of the infant head for the DOT. The BEM seems to be more elastic tool than the FEM. This
17
18 work tries to prove this opinion.
CONTENTS
However, some disadvantages of BEM exist, for example anisotropy of neonatal brain, which force us to develop hybrid method. Such problems are presented in the last chapters of this book. Taking above remarks into account someone may say that this book is not only addressed to the community dealing with DOT, but also to the engineers, last years students or PhD students dealing with medical or industrial tomography. This book might be helpful to the students of Technological Universities, dealing with numerical simulation used in dierent areas of technology. As an Author I am conscious that not all problems of the BEM were exhaust in this book. Many of them were merely announced. The subject is so fascinated that further research will allow to remove those disadvantages. For all critical remarks send by e-mail (the actual address may be nd in the internet) I will be extremely grateful and I will consider them carefully. Jan Sikora
Chapter 1 Twodimensional potential problems
1.1
Introduction
Starting to deal with Boundary Element (BE) method someone would compare it to its main rival, the Finite Element (FE) method. The BEM advantages and disadvantages either in 2D or 3D space can be summarized as follows. Advantages of the BE Method 1. Less data preparation time. This is a direct result of the surfaceonly modelling. Some of the authors say that it is reduction of dimensionality by one [14]. Furthermore, subsequent changes in meshes are made easier. This advantage is particularly important in problems where remeshing is required, such as inverse problems, for example optimal shape design. 2. High resolution regarding the gradient of the state function. Solution is exact and fully continuous inside the domain. This makes the BE me19
20
CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS thod very suitable for modelling problems of rapidly changing gradients of the state function. 3. Both state function and its normal derivative racy.
n
of the same accu-
4. Less computer time and storage. For the same level of accuracy, the BE method uses less number of nodes and elements (but contrary to the FEM fully populated matrix of coecients). However, there are problems where either FEM/BEM can give rise to the smaller system and quickest solutionit depends partly on the volume to surface ratio. 5. Less unwanted information. Since internal points in BE solution are optional, the user can focus on a particular interior region rather than the whole domain. 6. Open boundary problems more easy to handle. For problems involving innite or semiinnite domains, BEM is to be favoured. Disadvantages of the BE Method 1. Unfamiliar mathematics. Integrals are more dicult to evaluate and some contain integrands which become singular. The singular integrals have a signicant eect on the accuracy of the solution, so these integrals need to be evaluated accurately. However many FE numerical procedures are directly applicable to BE solutions. 2. A fundamental solution must be found (or at least an approximate one) before BEM can be applied. 3. The interior has to be discretized in many problems (Poisson equation or nonlinear material problems). That means that the method is loosing its main advantage of the reduction dimensionality. 4. Poor behavior for thin regions (for example the void problems). 5. Fully populated nonsymmetric, not positively dened matrix and not as good conditioned as in FEM.
1.2. LAPLACE EQUATION
21
So, which method is better? To answer this question we have to know what kind of problem we would like to solve. Certainly the BEM is very suitable and more accurate for the linear 2D and 3D problems with rapidly changing variables such as for example light transport ones.
1.2
Laplace equation
The dierential equation representing Laplaces equation in two dimensions can be written as follows: 2 (r) = 2 (r) 2 (r) + x2 y 2 (1.1)
where (r) represents any potential function (electric potential in electrostatics, or temperature in heat conduction problems) and x and y are the Cartesian coordinates. Consider now the physical domain of the problem.
n
enclosed solution domain
y surface 0 x z
Figure 1.1: Region in local coordinate system
Figure 1.1 shows an arbitrary domain where the solution is sought. In this solution domain, assume that there is an interior point r (usually called the load point) of coordinates X and Y , and consider any point on the boundary r (usually called the eld point) with coordinates x and y . The use
22
CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS
of capital letters for the coordinates indicates a xed point, whereas small letters indicate a variable point.
a)
111111111 000000000 000000000 111111111 000000000 111111111 000000000 111111111 domain 000000000 111111111 000000000 111111111 000000000 111111111
11111111111111111 00000000000000000 00000000000000000 11111111111111111 00000000000000000 11111111111111111 b) 00000000000000000 11111111111111111 domain 00000000000000000 11111111111111111 00000000000000000 11111111111111111 00000000000000000 11111111111111111 00000000000000000 11111111111111111 00000000000000000 11111111111111111 hole n 00000000000000000 11111111111111111 00000000000000000 11111111111111111 00000000000000000 11111111111111111 00000000000000000 11111111111111111 00000000000000000 11111111111111111 00000000000000000 11111111111111111 00000000000000000 11111111111111111 00000000000000000 11111111111111111
Figure 1.2: a) Anticlockwise numbering direction for a closed domain b) Clockwise numbering direction for an opened domain
1.2.1
Division of the boundary into constant boundary elements
The boundary of the region under consideration will be divided into M straight line segments (boundary elements) as it is shown in Fig. 1.3. The nodes where the unknown values are considered, are taken to be in the middle of each boundary element. The values of or q = are assumed to be conn stant on each element and equal to the value at the mid-node of the element. Let consider the following integral equation [3, 14]: c(r)(r) +
G(|r r |) (r )d(r ) = n
(r ) d(r ) (1.2) G(|r r |) n
where r usually called the load point and the point on the boundary, r is called eld point. The coecient c(r) will be described later in section 1.2.7 and G is the fundamental solution of Laplace equation. Now the Boundary Integral Equation (BIE), equivalent to Laplace equation, (see Eq.1.1) can be written in terms of local coordinate (see Fig. 1.4), instead of the boundary curve .
1111111111111111111111 0000000000000000000000 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 q 0000000000000000000000 1111111111111111111111 0000000000000000 1111111111111111 4 0000000000000000000000 1111111111111111111111 0000000000000000 1111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000 1111111111111111 q3 0000000000000000000000 1111111111111111111111 0000000000000000 00000000000000000000001111111111111111 1111111111111111111111 0000000000000000 1111111111111111 0000000000000000000000 1111111111111111111111 i=4 0000000000000000 1111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000 1111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000 1111111111111111 4 j=3 0000000000000000000000 1111111111111111111111 0000000000000000 1111111111111111 3 0000000000000000000000 1111111111111111111111 0000000000000000 1111111111111111 j=4 0000000000000000000000 1111111111111111111111 0000000000000000 1111111111111111 0000000000000000000000 1111111111111111111111 i=3 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 i=5
j=5 domain j=2
23
i=0
j=0
j=1
i=2
i=1
Figure 1.3: Boundary discretization for a constant elements with unknowns and q = over some part of the boundary n
11111111111111111111111 00000000000000000000000 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 q0 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 k=1 00000000000000000000000 11111111111111111111111 N0=1 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 =+1 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 =0 00000000000000000000000 11111111111111111111111 0 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 j=0 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 00000000000000000000000 11111111111111111111111 k=0
=1
Figure 1.4: Constant element (the shape function N0 for this element is equal to one), in local coordinates system
24
It can be easily veried that the coordinates of a point on element e with the intrinsic coordinate are given: x( ) = x i + xj xj x i + 2 2 (1.3) y ( ) = yi + yj yj yi + 2 2
Substituting the global numbering of the nodes by the local system and rearranging the Eq.(1.3) we will get: 1 1 x( ) = (1 )x0 + (1 + )x1 2 2 (1.4) 1 1 y ( ) = (1 )y0 + (1 + )y1 2 2 From now on we will work in the local numbering system and use the global numbering system to extract coordinates. Now Eq.(1.4) can be written in matrix notation. x = [x, y ] =
T k=1 k=0
Nk ( )xk
(1.5)
where Nk ( ) is basis interpolation function at node k th often called the shape function, x is a vector containing coordinates of a point on particular element and xk is a vector of coordinates of the k th node of that element. Note that vector x can have up to three components (x, y, z ). For the two node constant element just derived the shape functions are: 1 N = [Nk=0 ( ), Nk=1 ( )]T = (1 + k ) 2 where: is the local intrinsic coordinate (consult Fig. 1.4), and for: k=0 k = 1 (1.7) k=1 k = +1 (1.6)
The shape function is equal to one at its own node and equal to zero at the other nodes (please consult the Fig. 1.7 and Fig. 1.11).
25
1.2.2
Numerical integration of the kernels
Now we turn our attention to the numerical integration of the kernels. It is important to note that the choice of the local coordinate system from -1 to +1 was not arbitrary, because it happens to be the same as the limits used in the Gaussian quadrature technique. The boundary curve is now divided into elements j and the numerical integration performed over each element using the local intrinsic coordinate rather than the boundary segment j . f (x, y )d =
j 1
+1
f (x( ), y ( ))J ( )d
(1.8)
where: f means any function. The Jacobian1 of transformation is as follows: d J ( ) = = d ( dx( ) d )2 + ( dy ( ) d )2 (1.9)
In the boundary element method it will be necessary to work out the direction normal to a line or surface element. The best way to determine these directions is use of vector algebra [15]. Therefore the components of unit outward normal are given by [14]: [ ] 1 dy ( ) nx = J ( ) d where: dx( ) dN0 ( ) dN1 ( ) = x0 + x1 d d d (1.11) dy ( ) dN0 ( ) dN1 ( ) = y0 + y1 d d d
The Jacobian is shorthand for either the Jacobian matrix or its determinant, the Jacobian determinant. In most cases Jacobian means determinant.
1
[ ] dx( ) 1 ny = J ( ) d
(1.10)
26
and the dierentials of the shape functions are easily determined as follows: ( ) dN0 ( ) d 1 1 = (1 ) = d d 2 2 (1.12) ( ) dN1 ( ) d 1 1 = (1 + ) = + d d 2 2 Using results of Eq.(1.12) from Eq.(1.11) we will get: dx( ) x1 x0 = d 2 (1.13) dy ( ) y1 y0 = d 2 Introducing above relations into Eq.(1.9) nally we have: d J ( ) = = d ( x1 x0 2 )2 + ( y1 y0 2 )2 1 = L 2
(1.14)
where L denotes the length of the boundary element. Taking into account above relations the expressions for components of the unit outward normal vector in case of the constant element become very simple and constant over the element. nx = y1 y0 L ny = x1 x0 L (1.15)
Now the BIE of Eq.(1.2) can be written in terms of local coordinate instead of the boundary curve , as follows: c(r)i (r) +
M 1 j =0 M 1 j =0
j (r ) j (r ) n
+1 1
G(|r r |) J ( )d = n G(|r r |)J ( )d (1.16)
+1 1
where M is the total number of elements.
27
The integral functions (containing the kernels) can be lumped together in the functions Ai,j and Bi,j as follows: c(r)i (r) +
M 1 j =0
j (r )Ai,j (r, r ) =
M 1 j =0
j (r ) Bi,j (r, r ) n
(1.17)
To form a set of linear algebraic equations, we take each node in turn as a load point r and perform the integrations indicated in Eq.(1.16). This will result in the following matrices: [ ] [A][] = [B ] (1.18) n where the matrices [A] and [B ] contain the integrals of the kernels normal rr |) derivative G(|n and the kernels G(|r r |) respectively, i.e. the functions Ai,j and Bi,j of Eq.(1.17). For 2D system, the fundamental solution for Laplace equation is equal: G(|r r |) = 1 1 1 1 ln = ln 2 |r r | 2 (x x )2 + (y y )2 (1.19)
The entries of the matrices are: Ai,j (r, r ) =

+1 1 +1 1
G(|r r |) J ( )d n (1.20) G(|r r |)J ( )d
Bi,j (r, r ) =
where r depends on index i and r depends on index j . If R denotes the distance between point r and point r then: R = |r r | = (x x )2 + (y y )2 (1.21)
The normal derivative of Greens function in case of Laplace equation is: [ ] G(|r r |) G R G R x R y = = + = n R n R x n y n 1 [(x x)nx + (y y )ny ] (1.22) = 2R2
28 where:
R x x ; = x R x = nx ; n
R y y = y R (1.23) y = ny n
where nx and ny are dened by Eq.(1.10).
Integration of nonsingular integrals In case of nonsingular integrals, when point r and point r do not belong to the same boundary element, the standard GaussLegendre quadrature can be easily applicable to integrals of the general form:
+1
f ( )d =
1
g 1 i=0
wi f (i )
(1.24)
where g is the total number of Gaussian integration points, and i is the Gaussian coordinate with an associated weight function wi . The most frequently used values are listed in the Table 1.1 [18].
Integration of singular integrals The integrals can be calculated using Gauss quadrature rules for all elements except the one corresponding to the node under consideration. For this particular case the Ai,i coecients are equal zero and the Bi,i integrals can be calculated analytically. The distance R (see Eq.(1.21)) between point r and point r ( ) depends on local coordinate system in a following way: R( ) = |r r ( )| then Bi,i (R( )) =
+1 1
(1.25)
L G(R( ))J ( )d = 2
+1
1 1 ln d 2 R( )
(1.26)
1.2. LAPLACE EQUATION Table 1.1: Onedimensional Gaussian quadrature

n 4 6 i 1 2 1 2 3 1 2 3 4 5 1 2 3 4 5 6 1 2 3 4 5 6 7 8 Abscissa, i 0.33998104358485626480 0.86113631159405257522 0.23861918608319690863 0.66120938646626451366 0.93246951420315202781 0.14887433898163121089 0.43339539412924719080 0.67940956829902440623 0.86506336668898451073 0.97390652851717172008 0.12523340851146891547 0.36783149899818019375 0.58731795428661744730 0.76990267419430468704 0.90411725637047485668 0.98156063424671925069 0.09501250983763744019 0.28160355077925891323 0.45801677765722738634 0.61787624440264374845 0.75540440835500303390 0.86563120238783174388 0.94457502307323257608 0.98940093499164993260 Weight, wi 0.65214515486254614263 0.34785484513745385737 0.46791393457269104739 0.36076157304813860757 0.17132449237917034504 0.29552422471475287017 0.26926671930999635509 0.21908636251598204400 0.14945134915058059315 0.06667134430868813759 0.24914704581340278500 0.23349253653835480876 0.20316742672306592175 0.16007832854334622633 0.10693932599531843096 0.04717533638651182719 0.18945061045506849629 0.18260341504492358887 0.16915651939500253819 0.14959598881657673208 0.12462897125553387205 0.09515851168249278481 0.06225352393864789286 0.02715245941175409485
29
10
12
16
where in case of constant element J ( ) is dened by Eq.(1.14) and L denotes the length of the boundary element. This integral could easily be calculated analytically [23]. Hence Eq.(1.26) becomes, 1L Bi,i (R( )) = 2 ( ln 1
L 2
) +1 (1.27)
For more complex cases a special logarithmically weighted numerical inte-
30

q4 i=4 q3 4 j=4 i=5 y j=5 r i=0 j=0 x 0 i=1 r() j=3 domain
g=3
i=3
numerical integration points
rr
g=2 g=1 g=0
j=2
j=1
i=2
Figure 1.5: Numerical integration for constant boundary elements
gration formula can be used [21, 43, 45, 50, 54]:

0 1
f ( ) ln
1 d = wi f (i ) i=0
gl1
(1.28)
where: gl is the total number of logarithmic Gaussian integration points used and i is the Gaussian coordinate with an associated weight function wi (see Table 1.2). Note that the limits of integration are from 0 to 1 instead of the -1 to +1 range used in the nonsingular integrals of BIE.
1.2.3
Division of the boundary into linear boundary elements
Although, the constant element is a primitive one, in fact is a very convenient one. Particularly there is no problem with the corners of the region where the special treatment must be introduced to calculate the normal derivative of the state function, when the higher order interpolation is applied. Unlike
1.2. LAPLACE EQUATION Table 1.2: Onedimensional logarithmic Gaussian quadrature

n 4 i 1 2 3 4 1 2 3 4 5 6 7 8 i 0.04144848019938322080 0.24527491432060225194 0.55616545356027583718 0.84898239453298517465 0.01332024416089246501 0.07975042901389493841 0.19787102932618805379 0.35415399435190941967 0.52945857523491727771 0.70181452993909996384 0.84937932044110667605 0.95332645005635978877 wi 0.38346406814513512485 0.38687531777476262734 0.19043512695014241536 0.03922548712995983245 0.16441660472800288683 0.23752561002330602050 0.22684198443191912637 0.17575407900607024499 0.11292403024675905186 0.05787221071778207239 0.02097907374213297804 0.00368640710402761901
31
the second order isoparametric quadratic element, the linear element has no signicant advantages over the constant element. Geometry of the region is interpolated in the same way. Only better interpolation of the unknowns is oered by this element, but the user has to pay for that with the corner troubles. Consider the region divided by the linear elements as it is shown in Fig. 1.6. At each node of the particular boundary element indicated by upper index (j ) two unknowns and q (j ) = are present. The same transformation n (as for constant element) is used to transfer the nodes from global to local coordinate system (see Fig. 1.7).
x( ) =
1 k=0
Nk ( )xk = N0 ( )x0 + N1 ( )x1 (1.29)
y ( ) =
1 k=0
Nk ( )yk = N0 ( )y0 + N1 ( )y1
32

q4 1111111111111111111111111111111111 0000000000000000000000000000000000 000000000000000000000 111111111111111111111 0000000000000000000000000000000000 1111111111111111111111111111111111 000000000000000000000 111111111111111111111 i=4 000000000000000000000 111111111111111111111 0000000000000000000000000000000000 1111111111111111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 0000000000000000000000000000000000 1111111111111111111111111111111111 000000000000000000000 111111111111111111111 (j=4) 000000000000000000000 111111111111111111111 0000000000000000000000000000000000 1111111111111111111111111111111111 (j=3) 000000000000000000000 q111111111111111111111 000000000000000000000 111111111111111111111 q3 0000000000000000000000000000000000 1111111111111111111111111111111111 5 000000000000000000000 111111111111111111111 (j=4) (j=3) 000000000000000000000 111111111111111111111 0000000000000000000000000000000000 1111111111111111111111111111111111 000000000000000000000 111111111111111111111 4= 1111111111111111111111111111111111 000000000000000000000 111111111111111111111 0000000000000000000000000000000000 k=0= k=1 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 0000000000000000000000000000000000 1111111111111111111111111111111111 000000000000000000000 111111111111111111111 j=3 000000000000000000000 111111111111111111111 0000000000000000000000000000000000 1111111111111111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 0000000000000000000000000000000000 (j=3) 1111111111111111111111111111111111 j=4 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 0000000000000000000000000000000000 1111111111111111111111111111111111 000000000000000000000 111111111111111111111 = 000000000000000000000 111111111111111111111 q4 5= k=1
111111111111111111111 000000000000000000000 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 (j=4) 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111
(j=4)
(j=3)
i=3
k=0
i=5
j=5
domain
j=2
i=0
j=0
i=2 j=1
i=1
Figure 1.6: Boundary discretization for a linear elements
N() N0()
N1() 1
(j)
00000000000000000000000000 111111111111111111111111111 00000000000000000000000000 11111111111111111111111111 =+1 11111111111111111111111111 00000000000000000000000000 00000000000000000000000000 11111111111111111111111111 00000000000000000000000000 11111111111111111111111111 00000000000000000000000000 11111111111111111111111111 j 00000000000000000000000000 11111111111111111111111111 =0 00000000000000000000000000 11111111111111111111111111 (j) 00000000000000000000000000 11111111111111111111111111 0 00000000000000000000000000 11111111111111111111111111 00000000000000000000000000 11111111111111111111111111 00000000000000000000000000 11111111111111111111111111 00000000000000000000000000 11111111111111111111111111
k=0 =1
k=1
Figure 1.7: Local coordinates of linear element number j
1.2. LAPLACE EQUATION For the linear element the shape functions are: 1 N( ) = [Nk=0 ( ), Nk=1 ( )]T = (1 + k ) 2 where: k=0 k = 1
33
(1.30)
(1.31) k=1 k = +1
The shape function dened by Eq.(1.30) are linear functions such as: Nk ( ) = 1 at the node k = 0 and Nk ( ) = 0 at the node k = 1, as it is shown in Fig. (1.7). Using the isoparametric linear elements the same basis interpolation functions are used for the solution variables representation: ( ) = and: ( ) k 0 1 Nk ( ) = = N0 ( ) + N1 ( ) n n n n k=0
1 1 k=0
Nk ( )k = N0 ( )0 + N1 ( )1
(1.32)
(1.33)
1.2.4
In case of linear boundary elements the numerical integration of the kernels can be done in a similar way as for constant elements. The Jacobian of transformation and the components of unit outward normal are calculated according to Eq.(1.14) and Eq.(1.10) respectively. The boundary curve is divided into boundary elements (segments) j . Now the BIE of Eq.(1.2) will take the form: M 1 G(|r r |) (j ) (r )d(r ) = c(r)i (r) + n j =0 j M 1 (j ) (r ) = G(|r r |) d(r ) (1.34) n j =0 j
34
The numerical integration is performed over each boundary element j using the local intrinsic coordinate , as follows: c(r)i (r) +
M 1 +1 j =0 1 +1 1
M 1 j =0
G(|r r |) (j ) k (r )Nk ( )J ( )d = n k=0

1
G(|r r |)
1 (j ) (r ) k k=0
Nk ( )J ( )d
(1.35)
where M is the total number of quadratic elements. The nodal values are constant (do not depend on local coordinate ), so we can nally write Eq.(1.35) in the following form: c(r)i (r) +
M 1 1 j =0 k=0 (j ) k (r ) (j )
+1 1
G(|r r |) Nk ( )J ( )d = n G(|r r |)Nk ( )J ( )d (1.36)
M 1 1 j =0 k=0
k (r ) n
+1 1
If we denote the terms containing the integrals of the kernels normal derirr |) vative G(|n and the kernels G(|r r |) as a and b respectively, we will get: +1 G(|r r |) (j ) ai,k (r, r ) = Nk ( )J ( )d (1.37) n 1
(j ) bi,k (r, r )
+1 1
G(|r r |)Nk ( )J ( )d
(j ) (j )
(1.38)
For the smooth boundary the integral functions ai,k and bi,k can be lumped together in the global functions Ai,j and Bi,j (see Eq.(1.40)) as follows: c(r)i (r) +
M 1 1 j =0 k=0
ai,k (r, r )(r )k =

(j ) (j )
M 1 1 j =0 k=0
bi,j (r, r )
(j )
(r )k n
(j )
(1.39)
where r depends on index i and r depends on index j . To form a set of linear algebraic equations, we take each node in turn as a load point r and perform the integrations indicated in Eq.(1.36). This will
1.2. LAPLACE EQUATION result in the following matrices: [A][] = [B ] n [ ]
35
(1.40)
where the matrices [A] and [B ] are of the same size (only for the smooth boundary). The normal derivatives of Greens functions are calculated in the same way as for constant element (Eq.(1.22)). In case of the concave or convex corners a special treatment is needed.
The points r and r located in dierent elements Dealing rst with the second kernel G(|r r |) we have to consider two cases. The rst case is when the load and eld points are in dierent elements. Then integrals are not singular. And a bit more dicult problem when those two points are in the same element. In this case the singularity may occur. Let denote distance between two points as: R = |r r | = (x x)2 + (y y )2 (1.41)
then to calculate the integrals we have to consider the following two cases: The point r is the rst node (k = 0) of the linear element: R2 = [x ( ) x0 ] + [y ( ) y0 ] =
2 2
= [N0 ( )x0 + N1 ( )x1 x0 ]2 + [N0 ( )y0 + N1 ( )y1 y0 ]2 (1.42) where: N0 ( ) and N1 ( ) are expressed by Eq.(1.52). So: [ R
2
= [ +
1 1 (1 )x0 + (1 + )x1 x0 2 2 1 1 (1 )y0 + (1 + )y1 y0 2 2
]2 + ]2 (1.43)
36
The point r is a second node (k = 1) of the linear element: ]2 [ 1 1 2 R = (1 )x0 + (1 + )x1 x1 + 2 2 [ ]2 1 1 + (1 )y0 + (1 + )y1 y1 (1.44) 2 2 The points r and r are in the same element but r = r In this case, kernels are singular but the shape function Nk ( ) in the vicinity of r is of the order r. Therefore, the product of the kernels and the shape function is not singular, and the integrals can be evaluated using the standard Gaussian quadrature. So far, all the odiagonal coecients of the matrices [A] and [B ] have been calculated.
The points r and r are in the same linear element but r = r so R0 In this case, the standard Gaussian quadrature cannot be used because of 1 the singularity of the kernels. Dealing with the kernel G(r, r ) = 21 ln R 1 rst, it is clear that as r coincides with r , the singularity is of the form ln as 0. Fortunately, this form of integral can be calculated by using the special logarithmic Gaussian quadrature scheme [21,43,45,50,54,85,86] given bellow: 1 gl1 1 f ( ) ln d = wi f (i ) (1.45) 0 i=0 where: gl is the total number of logarithmic Gaussian integration points used and i is the Gaussian coordinate with an associated weight function wi (see Table 1.2). Note that the limits of integration are now from 0 to 1 instead of the -1 to +1 range used in the nonsingular integrals. A simple linear transformation can be used to transform the integral variable from to as follows:

a) k=1
(0)
37
b) k=1 =+1 (1) =0
=+1 (0) =1
=0 k=0 =1 (0) =0 k=0 (1) =1 (1) =1
=0
Figure 1.8: Coordinates transformation. a) r is in the rst node of the element, b) in the second node
1. if the position vector r is the rst node (k=0) of the element: (k=0) = 0.5(1 + ) 2. if the position vector r is the second node (k=1) of the element: (k=1) = 0.5(1 ) (1.47) (1.46)
For the rst node the Eq.(1.43) can be rearranged in the following way: [ ] ( )2 R2 = [0.5(1 + )]2 (x1 x0 )2 + (y1 y0 )2 = (0) L2 where L is a length of linear boundary element. For the second node the Eq.(1.44) can be rearranged in the following way: [ ] ( )2 R2 = [0.5(1 )]2 (x0 x1 )2 + (y0 y1 )2 = (1) L2 (1.49) (1.48)
Therefore, a general expression can be written for the logarithmic term as follows: G(|r r |) = 1 1 1 1 1 1 1 ln = ln (k) = ln (k) ln L (1.50) 2 |r r | 2 L 2 2
38
where: k indicates the node number and (k) transforms the integration limits from (1 to +1) to (0 to +1) for k = 0 and from (+1 to 1) to (0 to +1) for k = 1. Examining closely the Green function (Eq.(1.50)) as r approaches r , this function can be divided into two distinct parts: logarithmic part and a nonlogarithmic one (see Eq.(1.50)). Dealing now with the rst kernel (normal derivative of the Green function), it can be shown that it contains terms of order (1k) as (k) 0. Therefore, we can no longer use the Gaussian quadrature technique, even if a very large number of Gaussian points are used. Furthermore, we also need to explicitly calculate the parameter c(r) because its contribution is added to the diagonal terms of the [A] matrix. However, because all nondiagonal coecients of the [A] matrix can be calculated, there is a way to overcome this problem, which will be discussed later.
1.2.5
Division of the boundary into quadratic boundary elements
Particularly in optical or in impedance tomography we have to deal with very complicated shapes of the regions. That is why the main attention will be devoted to the second order interpolation over boundary elements in 2D and 3D space. Let consider the region presented in Fig. 1.9. We can dene a new coordinate system that is local to the element using an intrinsic variable , with its origin at the midpoint of the element and values of -1 and +1 at the end nodes as it is shown in Fig. 1.10. x( ) =
2 k=0
Nk ( )xk = N0 ( )x0 + N1 ( )x1 + N2 ( )x2 (1.51)
y ( ) =
2 k=0
Nk ( )yk = N0 ( )y0 + N1 ( )y1 + N2 ( )y2
where Nk ( ) are quadratic functions such as: Nk ( ) = 1 at its own node for

k=2 0000000000000000 1111111111111111 1111111111111111111111111111111 0000000000000000000000000000000 0000000000000000 1111111111111111 0000000000000000000000000000000 1111111111111111111111111111111 (j=3) 0000000000000000 1111111111111111 0000000000000000000000000000000 q 7 = q k=1 0000000000000000 1111111111111111 q 1111111111111111111111111111111 = q(j=4) 0000000000000000000000000000000 1111111111111111111111111111111 i=8 0000000000000000 1111111111111111 10 k=2 0000000000000000000000000000000 1111111111111111111111111111111 0000000000000000 1111111111111111 0000000000000000000000000000000 1111111111111111111111111111111 0000000000000000 1111111111111111 i=9 0000000000000000000000000000000 1111111111111111111111111111111 i=7 0000000000000000 1111111111111111 (j=3) (j=4) 0000000000000000000000000000000 1111111111111111111111111111111 = 0000000000000000 1111111111111111 = k=2 j=3 k=0 0000000000000000000000000000000 1111111111111111111111111111111 8 (j=3) (j=2) 0000000000000000 1111111111111111 0000000000000000000000000000000 1111111111111111111111111111111 q = q k=0 = q k=2 0000000000000000 1111111111111111 j=4 0000000000000000000000000000000 1111111111111111111111111111111 6 0000000000000000 1111111111111111 0000000000000000000000000000000 1111111111111111111111111111111 0000000000000000 1111111111111111 i=6 0000000000000000000000000000000 1111111111111111111111111111111 (j=3) (j=2) 0000000000000000000000000000000 1111111111111111111111111111111 6= k=0 = k=2 0000000000000000000000000000000 1111111111111111111111111111111 i=10 (j=4) 8
39
q = q k=0= q(j=3)
(j=4)
q9= q(j=4) k=1
10= k=2
i=11
j=5
domain
j=2
i=5 i=4
j=0 i=0 i=1 i=2
j=1 i=3
Figure 1.9: Boundary discretization for a quadratic boundary elements

k=2 =+1 endpoint k=1 =0 midpoint
k=0 =1 endpoint
Figure 1.10: An isoparametric quadratic element
example the node k = 0 and Nk ( ) = 0 at the other two nodes k = 1 and k = 2, resulting in the following: N0 ( ) = (1 ) = 0.5 ( 1) 2 N1 ( ) = (1 + )(1 ) = 1 2 N2 ( ) = + (1 + ) = 0.5 ( + 1) 2
(1.52)
40

1
N0
0.5
0 1 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
N1
0.5
0 1 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
N2
0.5
0 1 0.8 0.6 0.4 0.2
0.2
0.4
0.6
0.8
Figure 1.11: Basis interpolation functions distribution
Using the isoparametric elements the same basis functions are used for the solution variables:
2 k=0
( ) =
Nk ( )k = N0 ( )0 + N1 ( )1 + N2 ( )2
(1.53)
and: k 0 1 2 ( ) = Nk ( ) = N0 ( ) + N1 ( ) + N2 ( ) n n n n n k=0
2
(1.54)
The basis functions distribution over the boundary element are presented in Fig. 1.11.

q k=0= q q9= q(j+1) k=2 111111111111111111111111111111 000000000000000000000000000000 0000000000000000 1111111111111111 k=1 000000000000000000000000000000 111111111111111111111111111111 0000000000000000 1111111111111111 000000000000000000000000000000 111111111111111111111111111111 0000000000000000 1111111111111111 000000000000000000000000000000 111111111111111111111111111111 0000000000000000 1111111111111111 q(j) q = q(j+1) 000000000000000000000000000000 k=1 0000000000000000 1111111111111111 10 111111111111111111111111111111 k=2 000000000000000000000000000000 111111111111111111111111111111 0000000000000000 1111111111111111 000000000000000000000000000000 111111111111111111111111111111 k=2 0000000000000000 01111111111111111 000000000000000000000000000000 111111111111111111111111111111 0000000000000000 1111111111111111 (j+1) k=1 000000000000000000000000000000 111111111111111111111111111111 0000000000000000 = (j) k=0 1111111111111111 000000000000000000000000000000 111111111111111111111111111111 k=2 0000000000000000 1111111111111111 1 000000000000000000000000000000 111111111111111111111111111111 0000000000000000 1111111111111111 j q(j) 000000000000000000000000000000 111111111111111111111111111111 k=0 0000000000000000 1111111111111111 000000000000000000000000000000 111111111111111111111111111111 j+1 k=0 0000000000000000 1111111111111111 000000000000000000000000000000 111111111111111111111111111111 (j1) (j) 000000000000000000000000000000 111111111111111111111111111111 k=2 = k=0 000000000000000000000000000000 111111111111111111111111111111 (j+1) (j+2) 2 q k=2 9 k=1 0000000000000 1111111111111 111111111111111111111111111111 000000000000000000000000000000 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 q(j) 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 q = q(j+1) k=1 10 k=2 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 0 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 k=2 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 (j+1) 0000000000000 000000000000000000000000000000 111111111111111111111111111111 k=1 k=0 =1111111111111 (j) 1 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 k=2 q(j) k=0 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 j 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 j+1 k=0 0000000000000 1111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 (j1) 000000000000000000000000000000 111111111111111111111111111111 2 k=2 = (j) k=0 k=2 = (j+2) k=0
(j+1) (j+1) (j)
41
k=2 = k=0
q =q(j+1)
q(j+1) k=0
(j)
Figure 1.12: Variation of q = along the boundary line interpolated by qun adratic boundary elements without (upper) and with (lower) jumps between elements
1.2.6
In case of quadratic boundary elements the numerical integration of the kernels could be done in a similar way as for constant and linear elements. The Jacobian of transformation and the components of unit outward normal are calculated according to Eq.(1.14) and Eq.(1.10) respectively. The components of unit outward normal (see Eq.(1.10)) are functions of local coordinate : [ ] 1 dy ( ) nx ( ) = J ( ) d [ ] 1 dx( ) ny ( ) = J ( ) d
(1.55)
42 where:
dN0 ( ) dN1 ( ) dN2 ( ) dx( ) = x0 + x1 + x2 d d d d (1.56) dy ( ) dN0 ( ) dN1 ( ) dN2 ( ) = y0 + y1 + y2 d d d d and the dierentials of the shape functions are easily determined as follows: ( ) dN0 ( ) d 1 (1 ) = = d d 2 2 dN1 ( ) d = ((1 + )(1 )) = 2 (1.57) d d ( ) dN2 ( ) d 1 = + (1 + ) = + d d 2 2 The numerical integration is performed over each boundary element j using the local intrinsic coordinate , as follows: 2 M 1 +1 G(|r r |) (j ) k (r )Nk ( ) c(r)i (r) + J ( )d = n j =0 1 k=0 2 M 1 +1 (j ) k (r ) = G(|r r |)Nk ( )J ( )d (1.58) n 1 j =0 k=0 where M is the total number of quadratic elements. The nodal values are constant, so we can nally write Eq.(1.58) in the following form: +1 M 1 2 G(|r r |) (j ) c(r)i (r) + k (r ) Nk ( )J ( )d = n 1 j =0 k=0 M 1 2 (j ) k (r ) +1 G(|r r |)Nk ( )J ( )d (1.59) = n 1 j =0 k=0 If we denote the terms containing the integrals of the kernels normal derirr |) and the kernels G(|r r |) as a and b respectively, we will vative G(|n get: +1 G(|r r |) (j ) Nk ( )J ( )d (1.60) ai,k (r, r ) = n 1
1.2. LAPLACE EQUATION +1 (j ) bi,k (r, r ) = G(|r r |)Nk ( )J ( )d

1 (j ) (j )
43 (1.61)
For the smooth boundary the integral functions ai,k and bi,k can be lumped together in the global functions Ai,j and Bi,j (see Eq.(1.63)) as follows: c(r)i (r) +
M 1 2 j =0 k=0 (j ) (j ) ai,k (r, r )(r )k
M 1 2 j =0 k=0
(j ) (j ) (r )k bi,j (r, r )
(1.62)
where r depends on index i and r depends on index j . To form a set of linear algebraic equations, we take each node in turn as a load point r and perform the integrations indicated in Eq.(1.60) and in Eq.(1.61). This will result in the following matrices: [ ] [A][] = [B ] (1.63) n where the matrices [A] and [B ] are of the same size (only for the smooth boundary). The normal derivatives of Green functions are calculated in the same way as for constant and linear elements ( see Eq.(1.22)). In case of the concave or convex corners, a special treatment is needed.
The points r and r located in dierent elements Dealing rst with the second kernel G(|r r |) we have to consider two cases. The rst case when load and eld points are in dierent elements. Then integrals are not singular. And a bit more dicult problem when those two points are in the same element. In this case the singularity may occur. If distance between two points we will denote as: R = |r r | = (x x)2 + (y y )2 (1.64)
then to calculate the integrals we have to consider the following three cases:
44
The point r is the rst node (k = 0) of the element: R2 = [x ( ) x0 ] + [y ( ) y0 ] =

2 2
= [N0 ( )x0 + N1 ( )x1 + N2 ( )x2 x0 ]2 + + [N0 ( )y0 + N1 ( )y1 + N2 ( )y2 y0 ]

2
(1.65)
where: N0 ( ), N1 ( ) and N2 ( ) are expressed by Eq.(1.52). So: R

2
[ ]2 = (1 )x0 + (1 + )(1 )x1 + (1 + )x2 x0 + 2 2 [ ]2 (1.66) + (1 )y0 + (1 + )(1 )y1 + (1 + )y2 y0 2 2
The point r is second node (k = 1) of the element: R

2
[ ]2 = (1 )x0 + (1 + )(1 )x1 + (1 + )x2 x1 + 2 2 [ ]2 + (1 )y0 + (1 + )(1 )y1 + (1 + )y2 y1 (1.67) 2 2
The point r is third node (k = 2) of the element: R

2
[ ]2 = (1 )x0 + (1 + )(1 )x1 + (1 + )x2 x2 + 2 2 [ ]2 + (1 )y0 + (1 + )(1 )y1 + (1 + )y2 y2 (1.68) 2 2
The points r and r are in the same element but r = r In this case, kernels are singular but the shape function Nk ( ) in the vicinity of r is of the order r. Therefore, the product of the kernels and the shape function is not singular, and the integrals can be evaluated using the standard Gaussian quadrature. So far, all the odiagonal coecients of the matrices [A] and [B ] have been calculated.
1.2. LAPLACE EQUATION The points r and r are in the same element but r = r , so R 0
45
In this case, the standard Gaussian quadrature cannot be used because of 1 the singularity of the kernels. Dealing with the kernel G(r, r ) = 21 ln R rst, it is clear that as r coincides with r , the singularity is of the form 1 as 0. Fortunately, this form of integral can be calculated by using ln the special logarithmic Gaussian quadrature scheme [21, 43, 45, 50, 54] given bellow: 1 gl1 1 wi f (i ) (1.69) f ( ) ln d = 0 i=0 where: gl is the total number of logarithmic Gaussian integration points used and i is the Gaussian coordinate with an associated weight function wi (see Table 1.2). Note that the limits of integration are now from 0 to 1 instead of the 1 to +1 range used in the nonsingular integrals.
k=2 =+1 =1 =0 =1 k=0 =0

(0) (0)
(0)
=+1
(1)
(1) =+1
=1 =0 =0
=0
(2)
k=1
=1
(1)
=0 = 0 =1
(1)
(1)
= 1
(1)
(2)
=1
(2)
Figure 1.13: Coordinates transformation. a) r is in the rst node of the element, b) second node and c) third node
A simple linear transformation can be used to transform the integral variable from to as follows: 1. if the position vector r is the rst node (k = 0) of the element: (k=0) = 0.5(1 + ) (1.70)
46
CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS 2. if the position vector r is the second node (k = 1) of the element - the element is divided into two subelements: (k=1) = for 1 < < 0 and (k=1) = for 0 < < 1 (1.71) 3. if the position vector r is the third node (k = 2) of the element: (k=2) = 0.5(1 ) (1.72)
For the rst node the Eq.(1.66) can be rearranged in the following way: { R2 = [0.5(1 + )]2 [( 2)x0 + 2(1 )x1 + x2 ]2 + (1.73) [ } ( ) ( ) ( )2 ] 2 2 (0) (0) + [( 2)y0 + 2(1 )y1 + y2 ]2 = (0) fx ( ) + fy ( ) For the second node the Eq.(1.67) can be rearranged in the following way: { R2 = 2 [0.5( 1)x0 x1 + 0.5( + 1)x2 ]2 + (1.74) [( } ( ) ) ( )2 ] 2 2 (1) (1) + [0.5( 1)y0 y1 + 0.5( + 1)y2 ]2 = (1) fx ( ) + fy ( ) For the third node the Eq.(1.69) can be rearranged in the following way: { R2 = [0.5(1 )]2 [x0 + 2( + 1)x1 2( + 2)x2 ]2 + (1.75) [ } ( ) ( ) ( )2 ] 2 2 (2) 2 (2) (2) + [y0 + 2( + 1)y1 2( + 2)y2 ] = fx ( ) + fy ( ) Therefore, a general expression can be written for the logarithmic term as follows: G(|r r |) = 1 1 1 ln = ln 2 |r r | 2 1 )2 ( )2 = ( k ) ( k ) (k) fx ( ) + fy ( ) 1 1 1 [( (k) )2 ( (k) )2 ] 1 (1.76) ln (k) ln fx ( ) + fy ( ) = 2 2 2 (
where: k indicate, the node number and (k) transforms the integration limits from (1 to +1) to (0 to +1) and changes its value according to the position of r on the element. As for linear boundary element the second kernel as r approaches r , can be divided into two distinct parts: logarithmic part and a nonlogarithmic one (see Eq.(1.76)). Dealing now with the rst kernel, it can
47
1 be shown that it contains terms of order as 0. Therefore, we can no longer use the Gaussian quadrature technique, even if a very large number of Gaussian points are used. Furthermore, we also need to explicitly calculate the parameter c(r) because its contribution is added to the diagonal terms of the [A] matrix. However, because all nondiagonal coecients of the [A] matrix can be calculated, there is a way to overcome this problem. Next section is devoted to this particular problem.
1.2.7
Numerical integration of the c(r) coecient
We have two sets of points: points r where the unit sources are applied and points r where we have to satisfy boundary conditions. The problem is that some integrals in Eq.(1.59) only exist in the sense of a limiting value as r approaches r . This is explained in Fig. 1.14 for twodimensional potential problems [15]. We dene a region of exclusion around point r, with radius and we integrate around it. The integrals in Eq.(1.59) can now be split up into integrals over S S that are, the part of curve without the exclusion zone and into integrals over s that is, the circular boundary. As goes to zero it does not matter if we integrate over s or S as shown in Fig. 1.14. The lefthand side coecient c(r) of the Eq.(1.59) for 2D problems could be written as: c(r) = 1 2 (1.77)
where (see Fig. 1.15) is dened as the angle subtended at r by s in the limit as tends to zero. For 3D problems the same procedure could be applied (as it is sketched in Fig. 1.16) but for general corners expression similar to Eq.(1.77) is more complicated [15]. The function c(r) does not have to be calculated explicitly, and can be obtained indirectly by utilizing some physical considerations. It is based on the fact that the Boundary Integral Equations (BIE) matrices must apply to any physical problem with a unique solution. Any physical problem can be chosen as long as the solution does not depend on the geometry. The simplest issue to consider is the case where the potential is constant throughout the solution domain because that
48
d r y r
SS
Figure 1.14: For twodimensional potential problem semicircles around the boundary point r
leads to zero potential gradients everywhere. Therefore the righthand side of Eq.(1.59) becomes zero. Since the values are the same, the sum of all coecients in any row of the lefthand side matrices must be zero. The diagonal term can be determined as the sum of the nondiagonal ones as follows: Ai,j =
N
j =0
Ai,j
for
i = 0, 1, 2, . . . , N
(1.78)
j =i
where the i and j are the row and column counters respectively, while the N is the total nodes number.
1.2.8
Internal points calculation
Very close to the problem of singularities mentioned in the previous section is the problem of the so called quasisingularity, which occurs in the Boundary Element Method (BEM) analysis. Namely, when calculating the internal points placed near the boundary, one faces the small arguments of the kernels.
49
r y r n
SS
Figure 1.15: Boundary point r is located in the corner
000 111 d d 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 n

s
000 111
1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 SS 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111 0000000000000000000000000000000000000000 1111111111111111111111111111111111111111
Figure 1.16: For the threedimensional potential problem hemisphere around the boundary point
Due to this reason the numerical integration may be not precise enough as we can observe in Fig. 1.17. To avoid this situation some of the authors [49] propose to use the classical technique consisting of adding and subtracting a part of the integrand, which produces a quasisingularity. The others [14] suggest that instead of
50
Figure 1.17: Internal points calculation for 16 (left) and 32 (right) boundary elements discretization division the kernel into singular and nonsingular parts, simply multiplying 1 and dividing it by ln as follows:
0 1
f ( )d =
0
1 f ( ) ln 1 ln
d =
0
1 wi h(i ) h( ) ln d = i=0
gl1
(1.79)
The righthand side of the above equation is exactly as required by the logarithmic Gaussian quadrature. This latter approach is much simpler to program and is only slightly less accurate than the former approach of dividing the kernel into singular and nonsingular parts. However, such a problem could be avoided when the number of elements increases as it is presented in Fig. 1.17. For the internal points the function c(r) is equal one, so the state values could be calculated as follows: G(|r r |) (r ) (r) = (r )d(r ) + G(|r r |) d(r ) (1.80) n n It is worth to mention that for internal points usually more Gaussian integration points are demanded to achieve satisfactory precision. In this particular case, for the boundary variables the four points Gaussian quadrature integration rule was applied when for the internal point variables the ten points Gaussian quadrature integration rule was adopted. That makes integration a very timeconsuming process, so some authors recommend to avoid calculation variables in the internal points and plan division of the structure for some substructures in such a way that the interesting us internal points could be treated as the points belonging to the interface between substructures.
51
The gradient components at point r in the x and y directions are computed by taking the derivative of Eq.(1.80) in x and y direction respectively: (r) = x
(r ) d G ( | r r | ) n x (1.81) ( ) (r ) (r) G(|r r |) = (r )d + G(|r r |) d y n n y y
G(|r r |) n
(r )d +
Where the derivatives of the kernel x ( G(|r r |) n )
G n
are well known and are given as: ( G(r, r ) n ) = 1 y y (1.82) 2 R 2
1 x x = and 2 2 R y
The derivatives of the kernel G are a little bit more complicated: [ ] G(|r r |) 1 x x y y = nx + ny = x 2 x R2 R2 [ ( )] nx 2(x x) x x 1 y y 2 + nx + ny (1.83) = 2 x R R2 R2 R2 [ ] 1 x x G(|r r |) y y = nx + ny = y 2 y R2 R2 [ ( )] 1 ny 2(y y ) x x y y = 2 + n x + ny (1.84) 2 y R R2 R2 R2
1.2.9
Symbolic calculation
Symbolic calculation of matrix entries became very popular in the Finite Element Method (FEM). It speeds up the calculation and improves precision as well. However, for the BEM it is much more dicult due to the fact that we have much more complicated integrals. But even though it is still possible for some of those integrals. Symbolic integration example for non singular integrals in case of Laplaces equation approximated by second order boundary elements is presented below.
52
a_aux[0]=(yp*(0.07957747154594767*cordx[0] - 0.1061032953945969*cordx[1] +0.026525823848649224*cordx[2]))/ (xp*xp - 2.*xp*xq + xq*xq + yp*yp - 2.*yp*yq + yq*yq) + (0.07957747154594767*yq*cordx[0] - 0.1061032953945969*yq*cordx[1] + 0.026525823848649224*yq*cordx[2] + 0.07957747154594767*xp*cordy[0] - 0.07957747154594767*xq*cordy[0] - 0.1061032953945969*xp*cordy[1] + 0.1061032953945969*xq*cordy[1] +0.026525823848649224*xp*cordy[2] - 0.026525823848649224*xq*cordy[2])/ (-xp*xp + 2.*xp*xq - xq*xq - yp*yp + 2.*yp*yq - yq*yq);
a_aux[1]=(0.1061032953945969*(yp*(cordx[0] - cordx[2]) + yq*(-cordx[0] + cordx[2]) - (xp - xq)*(cordy[0] - cordy[2])))/ (xp*xp - 2.*xp*xq + xq*xq + yp*yp - 2.*yp*yq + yq*yq);
a_aux[2]= (0.026525823848649224*yp*cordx[0] - 0.026525823848649224*yq*cordx[0] - 0.1061032953945969*yp*cordx[1] + 0.1061032953945969*yq*cordx[1] + 0.07957747154594767*yp*cordx[2] - 0.07957747154594767*yq*cordx[2] - 0.026525823848649224*xp*cordy[0] + 0.026525823848649224*xq*cordy[0] + 0.1061032953945969*xp*cordy[1] - 0.1061032953945969*xq*cordy[1] - 0.07957747154594767*xp*cordy[2] + 0.07957747154594767*xq*cordy[2])/ (-xp*xp + 2.*xp*xq - xq*xq - yp*yp + 2.*yp*yq - yq*yq);
where: xp, yp are equivalent to x, y and xq , yq are equivalent to x , y respectively, cordx and cordy denotes coordinates of the nodes as it is shown for example in Fig.1.10, and a_aux stands for the coecients ai,k - consult Eq.(1.60). As we can see the above expressions are so complicated that such approach cannot be recommended on the stage of software prototyping. So far it is dicult to judge how much the symbolic integration will help to speed up calculation and improve the precision.
1.3. DIFFUSION EQUATION
53
1.3
Diusion equation
For the homogeneous regions and a steady state the diusion equation could be written as follows: D2 (r) + (r) = 0 or: 2 (r) (r) = 0 D (1.86) (1.85)
where k 2 = D is so called wave number and for the steady state, in some cases (neutron transport problem) is a real number [42].
For optical tomography, usually diusion coecient D = 0.03 cm and = 0.1 cm1 , so k = D = 1.8257 cm1 . 2 (r) k 2 (r) = 0 (1.87)
The boundary integral equation corresponding to Eq.(1.87) can be written in the form [21] G(|r r |) (r ) c(r)(r) + D (r )d(r ) = D G(|r r |) d(r ) n n (1.88) For 2D system, a modied Bessel function of the second kind and zero order, can be chosen as a fundamental solution [21] G(|r r |) = where R = |r r | = 1 K0 (k R) 2D (1.89)
(x x )2 + (y y )2
rr |) To calculate the rst kernel G(|n , the function G(|r r |) is dierentiated with respect to the unit outward normal at the point r : ) ( G(|r r |) 1 = K0 (k R) = n n 2D (1.90) ) ( ) ( R k R 1 = K0 (k R) = K1 ( k R ) R 2D n 2D n
54
where: K1 is the modied Bessel function of the second kind and rst order [2]. The derivative of the distance R with respect to the unit outward normal n = 1x + 1y at the point r is as follows: R R x R y R R = + = nx + ny n x n y n x y where: R x x = x R and R y y = y R (1.92) (1.91)
Therefore, the kernel from Eq.(1.90) can be expressed: ( )( ) G(|r r |) k x x y y = K1 ( k R ) nx + ny n 2D R R
(1.93)
So the equation (1.88) becomes: )( ) ( x x y y k c(r)(r) + D K1 (k R) nx + ny (r )d = 2D R R (r ) 1 K0 (k R) d (1.94) = D n 2D The boundary of the solution domain is divided into a number of connected elements j . Over each element, the variation of the geometry and variables must be described. Let assume, that as in case of Laplace equation the isoparametric quadratic element will be used. The BIE can be solved numerically by dividing the boundary into elements using appropriate shape functions in local coordinate system, the same as in Laplace equation section. Than Eq.(1.94) will take the form: c(r)i (r) + M 1 2 k 2
j =0 l=0 M 1 2 j =0 l=0
(j ) l (r ) (j )
+1
( K1 (k R)
) y y x x nx + ny Nl ( )J ( )d R R (1.95)
1 = 2
l (r ) n
+1
K0 (k R)Nl ( )J ( )d
where: M is the total number of elements. Now k denotes the wave number (see Eq.(1.87)) and l represents the local number of the nodes of particular boundary element (j ).
1.3. DIFFUSION EQUATION
55
1.3.1
Treatment of singularity
For the small arguments x n , the modied Bessel function becomes, asymptotically simple power (see Eq.(1.97)) of their arguments [2]: for n = 0: 1 K0 (x) = ln(x) = + ln x and for n > 0: (n 1)! ( x )n Kn (x) = 2 2 So for the rst order we will get:
80
BesselK(0,x) ln(x) 4 3.5 3 2.5
(1.96)
(1.97)
4.5
70 60 50 40
BesselK(1,x) 1./x
30
1.5 1 0.5 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
20 10 0 0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Figure 1.18: Comparison between the modied Bessel function of the second kind zero order (left) and the rst order (right) and the function 1/x for the arguments less than 1 (1 1)! ( x )1 1 K1 (x) = (1.98) = 2 2 x Taking above into account, the Green functions for the small arguments could be written as follows: 1 1 1 K0 (k R) ( ln(k R)) = (ln(k ) + ln(R)) = G(|r r |) = = 2D 2D 2D 1 1 = ln(k ) + ln 2D (l) (l) ( l ) 2 2 fx ( ) + fy ( ) ] [ ) 1 1 1 ( (l ) 2 2 (l ) = ln(k ) ln fx ( ) + fy ( ) + ln (l) 2D 2 =
(1.99)
56
where k is the wave number, superscript (l) denotes the number of the node in particular boundary element, is local coordinate system (see the Fig. 1.13) and fx and fy are the auxiliary functions for the rst time introduced in Eq.(1.73). The rst two terms of Eq.(1.99) are nonlogarithmic and the last one is the logarithmic type. That is why this kernel needs a special treatment (see the chapter 1.2.2 which is devoted to the Laplace equation). ( ) G(|r r |) k R k 1 x x x y y y = = K1 ( k R ) + = n 2D n 2D k R R n R n [ ] 1 x y = (x x ) + (y y ) (1.100) 2DR2 n n
1.3.2
After solving the system of linear equations, all the values of potential and potential gradient are determined on the boundary. So, it is straightforward matter to determine the potentials and gradient components in x and y direction, at any interior point. Because the interior point never lies on the boundary, there is no possibility of r and r coinciding and the kernels are no longer singular. )( ) ( k x x y y (r) = K1 (k R) nx + ny (r )d + 2 R R (r ) 1 K0 (k R) d (1.101) + n 2 (r) = x +
)( )] [( x x y y k nx + ny (r )d + K1 (k R) 2 R R (1.102)
1 K0 (k R) (r ) d 2 x n
In the similar way the y component of the gradient can be calculated. It is worth to mention that all dierentiations are made in the point r not in the point r .
1.4. DIFFUSION EQUATION IN FREQUENCY DOMAIN
57
1.3.3
Example
As an example let consider the steady state for the diusion equation (for example Eq.(1.85)) in a square region of dimension a for Dirichlet boundary conditions: (x = 0) = 0 and (x = a) = 10 and Neumann boundary = = 0. = conditions: n y y
y =0 y =a
So in reality, this problem is 1D, but formally it can be solved in 2D space. Comparison of the results with the analytical solution for two dierent discre for dierent tizations is presented in Fig. 1.19. But the relative error of n zero order boundary elements density discretization is presented in Fig. 1.20. We can observe an exponential error decreasing with respect to the number of elements. Also it is very interesting to compare the results of two numerical method like FEM and BEM with the analytical solution. One can easily notice that for FEM the error achieves enormously big values, when for BEM the error remains constant within the region and is equal to about 2 % (see Fig. 1.21). This is an immanent and very positive feature of the Boundary Element Method.
1.4
Diusion equation in frequency domain
The time dependent diusion equation for potential problems can be written as follows: D2 (r, t) + (r, t) =0 t (1.103)
where r represents space dimensions and t represents time. Material properties assumed constant in space and time were described in previous section. Initial values of (r, t) must be prescribed at time t = t0 or zero, as follows: (r, t) |t=t0 = (r, t0 ) = const.
58
10 10 10 10 10 10 10 10 10
1 0

10
1
Analytic BEM
10 10 10 10 10 10 10 10
Analytic BEM
9 x
10
9 x
10
Figure 1.19: Solution for 16 boundary elements (left) and for 128 boundary elements (right)
50
40 [ %]
30
20
10
16 32
64
96
128 number of elements
256
Figure 1.20: For the zero order boundary elements the relative error for versus number of elements
100
16 14
80
12 10
60 error [%]
error [%]
8 6 4 2 0 2
FEM BEM
40
20
20 0
x [mm]
10
4 0
x [mm]
10
Figure 1.21: Error comparison for FEM (right) and BEM (left) in case of the example presented in subsection 1.3.3
59
The boundary conditions are usually in the form of a prescribed state function on part of the boundary 1 and prescribed state (potential) gradient function on another part of the boundary 2 as follows: (r, t) = h(r, t) on 2 (1.104) t where the functions g (r, t) and h(r, t) are known as functions in time. (r, t) = g (r, t) on 1 ; The Laplace transform is used to remove the time dependency of the integral equations. To obtain the physical quantities, an inverse Laplace transform is applied. The Laplace transform can be dened as follows [14]: L [(r, t)] = (r, t)est dt = L (r, s) (1.105)
0
where s is called the transform parameter. Therefore, using the properties of Laplace transforms, the dierential equation can be written as follows: 1 2 L (r, t) s L (r, t) = 0 (r, t0 ) (1.106) D D where k 2 = s is so called wave number. D Note that the boundary conditions in the form of the functions g and h (Eq. (1.104)) also have to be expressed as Laplace transforms. It can be veried that the fundamental solution for this dierential equation is given by the following expression (see [14]): ( ) 1 s GL (|r r | , s) = K0 R (1.107) 2D D where K0 is the modied Bessel function of the second kind of zero order. Using the same procedure as that outlined in case of Laplace equation for twodimensional potential problems, we can use Greens second identity to arrive at a boundary integral equation as follows: GL (|r r | , s) L (r , s)d = c(r)L (r, s) + D n (1.108) L (r , s) = D GL (|r r | , s) d + GL (|ri r | , s)0 (r , t0 )d, n
60
where r and r , ri is a position vector of i th point source.

rr |,s) It is apparent that the potential kernels GL (|r r | , s) and GL (|n may be written more explicitly in the form: 1 GL (|r r | , s) = K0 (k R, s) (1.109) 2D ( ) GL (|r r | , s) 1 R k = K0 (k R, s) = K1 (k R, s) (1.110) n n 2D 2D n The derivative of the distance R = |r r | = (x x)2 + (y y )2 with respect to the unit outward normal n is calculated as for the steady state (see Eq.(1.91), Eq.(1.92) and Eq.(1.93)).
Therefore, the normal derivative of the Greens function can be rewritten in more explicit form as follows: [ ]( ) GL (r, s) k x x y y = K1 (k R, s) nx + ny (1.111) n 2D R R So the equation (1.108), under assumption that initial conditions of are equal to zero, can take the following form: ]( ) [ x x k y y c(r)(r, s) + K1 (k R, s) nx + ny (r , s)d = 2 R R (1.112) 1 (r , s) K0 (k R, s) d = n 2 The BIE of Eq.(1.108) can be solved numerically by dividing the boundary into elements and the domain into cells using appropriate shape functions. The numerical implementation is very similar to that of the steadystate problems described in Laplace equation section, since the form of singularity is of the same order. +1 M 1 2 G(|r r | , s) (j ) Nl ( )J ( )d = c(r)i (r, s) + l (r , s) n 1 j =0 l=0
M 1 2 j =0 l=0 (j ) l (r , s)
(1.113)
+1
1 G(|r r | , s)Nl ( )J ( )d 2
61
where, for simplicity we have skipped subscript L understanding that notation G(|r r | , s) means the Laplaces transform of fundamental solution, M is the total number of elements and the index l expresses the local numbers of the nodes of boundary elements, in this case the isoparametric quadratic boundary elements. If one makes the assumption that is varying harmonically in time then Eq. (1.106) becomes: 1 (r, ) = 0 (r, t0 ) D D
2 (r, ) i
(1.114)
where i =
1, = 2f is an angular frequency, k 2 = and k = i . D D
In this case the wave number is complex, so we have a modied Bessel function of complex arguments.
1.4.1
In case of the solution in the frequency domain the singular integrals are treated in the same way as in case of the steady state (see for example section 1.2.6).
1.4.2
For the frequency domain formulation, the internal point values like or [ ]T , can be calculated in the similar way as it was described in section x y 1.3.2 (see Eqs. (1.101) and (1.102)).
62
1.5
Examples
To show the exibility and powerfulness of the BEM let consider two examples described by the Helmholtz equation. The rst one is for the Cartesian coordinate system, while the second one is for the polar coordinate one.
1.5.1
Cartesian coordinate system
To start with, let consider the Helmholtz equation in Cartesian coordinate system of the form [49]. 2 H = i H (1.115)
where i = 1 and [ s1 ] is the angular frequency, [1/( m)] is a material conductivity and [H/m] is a magnetic permeability. Let consider the following example: the onedimensional bounded conducting plate of thickness 2 b is placed in the sinusoidal magnetic eld (see Fig. 1.22).
z
0 1 0 1 11111111111 00000000000 0 1 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 H z 1z 0 1 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 b b 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 J x 1x 11111111111 0 1 00000000000 0 1 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 x 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 0 1 00000000000 11111111111 0 1 00000000000 11111111111
Figure 1.22: Conducting plate of nite thickness
Both magnetic eld strength and eddy currents have one component only, in z and x direction, respectively. Solving Eq.(1.115) the analytical solution for
1.5. EXAMPLES eddy current component in x direction takes the form: Jx (y ) = where = dHz sinh(y) = Hz b dy cosh( ) 2
63
(1.116)
i and b is a thickness of the plate.
Numerical solution, in the frequency domain of Eq.(1.115) for two dierent boundary discretization, is presented in the Fig. 1.23 and Fig. 1.24. The solid line is the analytical solution, but the diamonds represents the boundary element solution. The relative error for the modulus and the angle
32 boundary elements 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 0.05 0 0.05
100 150 0.05 0 0.05 0 50 100 50 32 boundary elements 150 analitical solution BEM solution
analitical solution BEM solution
Figure 1.23: Modulus (left) and phase shift (right) of the function Hz changes across the plate thickness for 32 boundary elements discretization
128 boundary elements 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 0.05 0 0.05 100 150 0.05 0 50 100 50 analitical solution BEM solution 150
128 boundary elements analitical solution BEM solution
0.05
Figure 1.24: Modulus (left) and phase shift (right) of the function Hz changes across the plate thickness for 128 boundary elements discretization
64
shift for the second order boundary element are presented in the following Fig. 1.25.
55 50 45 40 35 30 25 20 15 10 5 0 16 32 64 96 128 256 55 50 45 40 35 30 25 20 15 10 5 0 16 32 64 96 128 256
Figure 1.25: Maximal relative error for modulus (left) and maximal relative error for the phase shifting (right) in function of number of boundary elements
As we can see from the Fig. 1.25 the phase shift is much more sensitive than the magnitude and demands better discretization in order to achieve the same relative error. This example shows that the BE method is very ecient because even for the coarse discretization the relative error for both the magnitude and the phase shift is really small (see for example discretization for 96 boundary elements for which the maximal relative error is less than 5%).
1.5.2
Polar coordinate system
Now, let consider more complicated example described by the 2D Helmholtz equation on a disc radius r = b [m], which we express in polar coordinates [7]: 2 (r, ) + k 2 (r, ) = 0 where k is a wave number. For this equation it exists an analytical solution of the form: (r, ) =
n=0
(1.117)
cn In (k r)ein
(1.118)
1.5. EXAMPLES
65
where In is the modied Bessel function of the rst kind and the n th order. The state function and its normal derivative on the boundary are then n given by: (r, )|r=b =
n=0
cn In (k b)ein (1.119)
(r, ) n
=
r =b
n=0
cn k In (k b)ein
Assume that the Robin boundary condition applies, such that a known input ux is specied leading to a boundary condition = D ( ) on the right hand side of: (b, ) + 2A D (b, ) = (b, ), n b (1.120)
where: D is the diusion coecient and coecient A depends on refractive index (see chapter 2.9.1). Than we have:
n=0 ( ) cn In (k b) + 2A Dk In (k b) ein = n ein n=0
(1.121)
which leads to cn = n . In (k b) + 2A Dk In (k b) (1.122)
Now we have the analytic solution expressed as an innite series of modied Bessel functions.
1.5.3
Distributed source for a diusion model for light transport
Using this simple example (simple in the sense that this example posses analytic solution, see Eq.(1.119)), we consider three cases for the distributed
66
source on the boundary, when in Eq.(1.121), we will take n = 0, n = 3 and n = 9 terms of the solution. For n = 0 the Robin boundary condition will take the form: ( ) 0 = c0 I0 (k b) + 2A Dk I0 (k b) .
(1.123)
It means that incoming ux is constant regarding the magnitude and the phase shift along the whole boundary. That is why in this case only the distribution of the and the along the radius of the disc are presented n (see Fig. 1.26). For the boundary conditions dened by the Eq.(1.123), even coarse discretization (48 elements of the second order) produces nice results an relative error less than 5% (see Fig. 1.27). When the number of boundary elements will be increased to 192 then the relative error will drop below 1%. For n = 3 the Robin boundary condition will take the following form: ( ) 3 = c3 I3 (k b) + 2A Dk I3 (k b) ,
(1.124)
where I3 is the modied Bessel function of the rst kind and the third order. The distribution of the and
n
along the boundary is presented in Fig. 1.28.
For this kind of boundary conditions we need more boundary elements in order to keep the relative error on the same level as in case of n = 0. For n = 9 the Robin boundary condition becomes: ( ) 9 = c9 I9 (k b) + 2ADk I9 (k b) .
(1.125)
where I9 is the modied Bessel function of the rst kind and the ninth order. The distribution of the state function and its normal derivative along the perimeter of the region is presented in the Fig. 1.30.
1.5. EXAMPLES
67
Figure 1.26: Analytical solution (solid lines) and numerical one for 48 boundary elements and n = 0 versus the radius of the disc
4.5 4 3.5 relative error [%] relative error [%] 3 2.5 2 1.5 1 0.5 0 48 96 192 number of elements 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 48 96 192 number of elements
Figure 1.27: The modulus of the (left) and the phase shift of the (right) as a function of the boundary elements number
68
Figure 1.28: Analytical solution for n = 3 along the perimeter of the disc in radians
6 5 relative error [%] 4 3 2 1 0 48 96 192 384 768 number of the second order boundary elements
Figure 1.29: The modulus of the as a function of discretization density
The numerical solution for 96 second order boundary elements is shown in the next gure (see Fig. 1.31). In this case the oscillation of the external
1.5. EXAMPLES
69
Figure 1.30: Analytical solution for n = 9 along the perimeter of the disc in radians
values of and are clearly visible. The relative error distribution with n respect to the number of boundary elements is presented in the Fig. 1.32.
Concluding this relatively simple example when we have to deal with a distributed source, which is modelled by the Robin boundary conditions, the precision of the solution strongly depends on kind of boundary conditions imposed. In case of the example considered, increasing the number of boundary elements by 2, the relative error is dropping by factor 2 as well (consult Fig. 1.29 and Fig. 1.32).
70
Figure 1.31: (upper left) and phase shift (bottom left) distribution along the boundary; (upper right) and phase shift of (bottom right) along n n the boundary
9 8 7 relative error [%] 6 5 4 3 2 1 0 48 96 192 384 768 number of the second order boundary elements
relative error [%] 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 48 96 192 384 768 number of the second order boundary elements
Figure 1.32: Relative error of modulus with respect to the number of boundary elements (left) and modulus of oscillation with respect to the number of boundary elements
1.5.4
Point source located on the boundary for a diusion model for light transport
So far, distributed sources placed on the boundary were considered. Now let present some results when the source is located in one or two elements
1.5. EXAMPLES
71
(constant element) as it is shown in Fig. 1.33 or in one point (rst, second or higher order elements) on the boundary. Such cases are much more dif30 25 20 15 10 5 0 5 10 15 20 25 30 30 25 20 15 10 5 0 5 10 15 20 25 30 The source 30 25 20 15 10 5 0 5 10 15 20 25 30 30 25 20 15 10 5 0 5 10 15 20 25 30 The source
Figure 1.33: Discretization of the boundary region by 48 (left) and 192 (right) elements cult because they demand a very careful discretization around the source point. This is one of the possibilities of the point sources treatment in numerical approach. Interested readers will nd more informations in the work of S. Arridge for example [6, 8].
Constant element Let consider diusion equation with Dirichlet boundary conditions: = 0 on the boundary of the region with exception of one or two elements (midpoint of the selected element) where the point source is located. We will investigate inuence of discretization in the solution. The region is a typical benchmark object for optical tomography, namely the circle of radius a = 25 mm. Basic optical data: c = 0.21 mm/ps, a = 0.025 mm1 , s = 2.0 mm1 , f = 200 MHz. 48 boundary elements For such coarse discretization (regular one) we can observe a very big oscillation around the point source, particularly for the phase shift.
72
10 10 10 modulus of / n 10 10 10 10 10 10 10 10
0 1 2 3 4 5 6 7 8

180
150 phase shift of / n
120
90
60
30
9 10
60
120
180 240 angle [deg.]
300
360
0 0
60
120
180 angle [deg.]
240
300
360
Figure 1.34: Modulus (left) and the phase shift (right) of the output ux on the boundary in case of 48 elements 192 boundary elements
In order to remove those oscillations it is necessary to employ ne discretization (192 boundary elements). They are four times shorter. As we can see in the Fig. 1.35 the oscillations almost despaired.
10 10 10 modulus of / n 10 10 10 10 10 10 10 10
0 1 2 3 4 5 6 7 8
180
150 phase shift of / n
120
90
60
30
9 10
60
120
300
360
0 0
60
120
180 angle [deg.]
240
300
360
Figure 1.35: Modulus (left) and the phase shift (right) of the output ux on the boundary in case of 192 elements
Isoparametric quadratic element In order to compare the previous results with more sophisticated approximation let apply the second order BEM to discretize diusion equation for light transport problem. To make the comparison possible, the same examples were considered.
1.5. EXAMPLES 48 boundary elements

10 10 10 10
0 1 2 3 4 5 6 7 8 9
73
360 330 300 phase shift of / n

0 60 120 180 240 angle [deg.] 300 360
270 240 210 180 150 120 90 60 30 0 0 60 120 180 angle [deg.] 240 300 360
modulus of / n
10 10 10 10 10 10 10
10
Figure 1.36: Modulus of the output ux on the boundary (left) and the n phase shift of the output ux n on the boundary in case of 48 elements
192 boundary elements

10 10 10 modulus of / n 10 10 10 10 10 10 10 10
0 1 2 3 4 5 6 7 8 9
360 330 300 phase shift of / n 270 240 210 180 150 120 90 60 30
10
60
120
300
360
0 0
60
120
180 angle [deg.]
240
300
360
Figure 1.37: Modulus of the output ux on the boundary (left) and the n phase shift of the output ux n on the boundary in case of 192 elements
As we can see for coarse discretization the solution is highly oscillating near the point source, even more than for the zero order approximation (Fig. 1.36). It is due to the basis function distribution for the end points of boundary element (see Fig. 1.11). However if discretization becomes dense the oscillations fade away, as it was in case of the constant elements. Interesting results could be achieved if nonuniform boundary element division was applied. Let discretize the boundary by the same number of
74
elements, for example 48, but in the way presented in Fig. 1.38. The right semicircle is discretized by 32 elements of the length equal to 2.454 mm < Ld when the left semicircle is discretized by 16 elements of the length equal to 4.909 mm > Ld . Where Ld denotes the diusion length1 [69], which in the case considered, is equal to: D 1 = = 2.582 [mm] (1.126) Ld = a 3a s where a is absorbing coecient and s stands for reduced scattering coecient.
30 25 20 15 10 5 0 5 10 15 20 25 30 30 25 20 15 10 5 0 5 10 15 20 25 30 The inner circle The point source
Figure 1.38: Nonuniform boundary element discretization 48 elements
The phase shift oscillations around the point source are much smaller due to the fact that better discretization in this part of boundary have been used. We have kept the same number of elements, so in the far region the boundary elements become bigger, the oscillations occurred for the phase
The diusion length Ld , as opposed to the absorption length la = 1/a , is the distance at which the average intensity decreases by a factor of e.
1
1.5. EXAMPLES
75
shift distribution, as we can see in Fig 1.39. Situation is even more serious if we consider the phase shift of the along the circle located inside the region 1 = 0.5 mm from the boundary (see Fig. 1.40). at a distance equal to s To eliminate such oscillations we have to increase the number of boundary elements. In case of 96 elements oscillations decreased dramatically as we can see in Fig. 1.41.
10 10 10 modulus of / n 10 10 10 10 10 10 10 10
0 1 2 3 4 5 6 7 8 9
360 330 300 phase shift of / n 270 240 210 180 150 120 90 60 30
10
60
120
300
360
0 0
60
120
180 angle [deg.]
240
300
360
on the boundary (left) and the Figure 1.39: Modulus of the output ux n phase shift of the output ux n on the boundary in case of 48 elements
10 10 10 10 modulus of 10 10 10 10 10 10 10
0 1 2 3 4 5 6 7 8
360 330 300 270 phase shift of 240 210 180 150 120 90 60 30
9 10
60
120
300
360
0 0
60
120
180 angle [deg.]
240
300
360
Figure 1.40: Modulus of the (left) and phase shift of the (right) along the circle located at a distance 0.5 mm from the surface
Those results for 96 nonuniform discretization are almost as good as the results achieved for the uniform discretization by 192 boundary elements. The conclusion is quite obvious. Near the point source, discretization have to
76
10 10 10 modulus of / n 10 10 10 10 10 10 10 10
0 1 2 3 4 5 6 7 8 9

360 330 300 phase shift of / n 270 240 210 180 150 120 90 60 30
10
60
120
300
360
0 0
60
120
180 angle [deg.]
240
300
360
Figure 1.41: Modulus of the output ux on the boundary (left) and the n on the boundary in case of 96 elements phase shift of the output ux n
be as dense as possible. However, the length of the boundary elements never should exceed a diusion length. Otherwise, we can expect nonphysical oscillations in the solution (see for example Fig. 1.41).
1.5.5
Point source located inside the region for a diusion model for light transport
Let consider the following circular region with internal source as indicated in the following picture: Inside the region the governing equation is as follows: 2 (r, ) k 2 (r, ) = q a where k = i cD is a so called wave number, q = D for 2D space or D = 3( 1 for 3D space. + )
s a
(1.127)
qs D
and D =
1 2(s +a )
On the boundary the Robin boundary conditions are imposed: (r, ) + 2Dn (r, ) = 0 r (1.128)
For the boundary element method we are dealing with a couple of unknowns r, ) and ( , so it is convenient to present the boundary conditions in the n
1.5. EXAMPLES
30 25 20 15 10 5 0 5 10 15 20 25 30 30 25 20 15 10 5 0 5 10 15 20 25 30 The point source
77
Figure 1.42: Internal point source and boundary element discretization by 48 isoparametric quadratic elements
following form: (r, ) 1 = (r, ) n 2D r
(1.129)
Note that the boundary conditions also have to be expressed as Laplaces transforms. It can be veried that the fundamental solution for this dierential equation is given by the following expression (see [14]): G(|r r | , ) = 1 K0 (k |r r | , ) 2 (1.130)
where K0 is the modied Bessel function of the second kind of zero order. Using Greens second identity to arrive at a boundary integral equation as
78 follows:
CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS c(r)(r, ) +
G(|r r | , ) (r , )d = n (r , ) d n
G(|r r | , )
(1.131) G(|ris r | , ) q d
where r and r , ris . In optical tomography, the concentrated sources are frequently used and fortunately they are very simple to handle in boundary element method. They are a special case for which the function q at the internal point ris becomes: q = Qis is (1.132)
where Qis is the magnitude of the source and is (r ris ) is a Dirac delta function which integral is equal to 1 at the point ris and zero elsewhere. Assuming that, a number of these functions Qis exists and is equal to p, and can be written as follows: G(|r r | , ) c(r)(r, ) + (r , )d = n p1 (r , ) = G(|r r | , ) d G(|ris r | , ) Qis (1.133) n is=0 where G(|ris r | , ) is the value of the fundamental solution at the point ris , p is the total number of point sources inside the domain. It is apparent that kernels in Eq.(1.133) may be written more explicitly in the form: ) ( G(|r r | , ) 1 = K0 (k |r r | , ) = n n 2 k R = K1 (k |r r | , ) (1.134) 2 n where K1 is the modied Bessel function of the second kind of order one.
1.5. EXAMPLES
79
The derivative of the radius R with respect to the unit outward normal n at the point r is calculated as for the steady state (see Eq.(1.91) Eq.(1.92)). Therefore, the Eq.(1.134) can be rewritten in more explicit form as follows: G(|r r | , ) k = K1 (k |r r | , ) n 2 ( x x y y nx + ny R R ) (1.135)
So the equation (1.133) could take the following form: c(r)(r, ) + ( ) k x x y y K1 (k |r r | , ) nx + ny (r , )d = 2 R R p1 (r , ) 1 K0 (k |r r | , ) d G(|ris r | , ) Qis = n 2 is=0
(1.136)
The BIE can be solved numerically by dividing the boundary into elements j using appropriate shape functions in local coordinate system: c(ri )(ri , ) + M 1 2 (j ) l (r j , ) +
j =0 l=0 yj yi
+1 1
k K1 (k |ri r j | , ) 2
xj xi nx |j + Rij
ny |j Nl ( )J ( )d = (1.137) Rij M 1 2 (j ) l (r j , ) +1 1 = K0 (k |ri r j | , ) Nl ( )J ( )d + n 2 1 j =0 l=0

p1 M 1 j =0 is=0
G(|ris r j | , ) Qis
where: M is the total number of elements and l (before was k but now it coincides with the wave number) expresses local numbers of the nodes of boundary elements, in this case the isoparametric quadratic boundary element.
80
Treatment of singularity In case of the frequency domain, the singularity is treated in the same way as in the steady state (see section 1.3.2).

Basically the calculation of and at the internal points can be calculated n in the same way as before, with one exception, namely the point sources placed inside the domain at points is should be added as in Eq.(1.137).
1.5.6
Comparison FEM and BEM results of calculation
The main goal is to compare the BE formulation with the FE one with respect to precision of approximation and number of unknowns involved. The source term has the value Q0 , and the Robin boundary conditions are dened: (r, ) + 2Dn (r, ) = 0, where D is the diusion coecient.
10
0
40 BEM BEM 35 30 FEM FEM
10
10 amplitude
phase shift [deg.] 60 120 180 240 angle [deg.] 300 360
25 20 15 10
10
10
10
5 0 0 0 60 120 180 240 angle [deg.] 300 360
10
Figure 1.43: Modulus of distribution along the boundary (left) and the phase shift distribution along the boundary (in radians)
As we can see, the excellent agreement was achieved regarding the surface values as well as the internal one along the radius of the circular region. But it is worth to mention that FE needs more than 7000 unknowns and BEM
1.6. ANISOTROPIC MEDIUM

10
0
81
40 BEM FEM 30
BEM 10
1
FEM
10 amplitude
10
phase shift [deg.] 5 0 5 distance [mm] 10 15 20 25
20
10
10 10
5
10 25 20 15 10
0 25 20 15 10
5 0 5 distance [mm]
10
15
20
25
Figure 1.44: Modulus of (left) and phase shift (right) distribution along the radius of the domain only 200. In the Fig. 1.43 and Fig. 1.44 the solid lines represent the FEM solution, but the dotted lines the BEM ones.
1.5.7
Conclusion
Interesting results were obtained in case of zero order boundary element implementation. The numerical experiments proved that even for the zero order BEM, we were able to achieve a stable and precise results. In the same time the second order elements were have a tendency to oscillations All those numerical experiments proved that boundary element method could be very useful for modelling the forward problem in optical tomography. The boundary element method is precise and, what is more, it is very economic in respect to the number of elements. This problem becomes critical for three dimensional space and particularly for the inverse problem solution.
1.6
Anisotropic medium
Let consider the same homogeneous circular region as before (see Fig. 1.42), but this time the ability of the BEM dealing with the anisotropic media will be shown.
82
1.6.1
Anisotropy model
In two dimensions, we write the diusion tensor D as [40] D(r) = R(r)diag(1 (r), 2 (r))R(r)T where R(r) is a 2 2 orthogonal matrix, and i (r) = 1 2 [a (r) + (1 bi (r))s (r)] , i = 1, 2. (1.139) (1.138)
Representation (1.138) can also be interpreted as an eigenvalue decomposition of the diusion tensor. The eigenvalues i , i = 1, 2, present the strength of the anisotropy, and the direction of anisotropy is conned in the matrix R. Let satisfy the diusion equation in the frequency domain: (D(r) k 2 )(r, ) = 0 (1.140)
is the wave number and where k = a i c [ ] [ ] b d a d 1 D det D D= , D = det , det D = |D| = ab d2 .(1.141) d a d b det D det D On the boundary the Robin boundary conditions are imposed: (r, ) + 2n D(r, ) = 0 Introduce the new coordinate system: r = D1/2 r, the Eq. (1.140) will take the form:
2 2 r, ) = 0 ( r k )(
(1.142)
r = D1/2 r,
(1.143)
(1.144)
Fundamental solution for above dierential equation is given by the following expression [40]: G( r , ) = ( ) 1 K0 k (rT D1 r)1/2 1 / 2 2 |D| (1.145)
83
where K0 is the modied Bessel function of the second kind of zero order, k wave number and |D|1/2 is the square root of determinant of matrix D. Using Greens second identity a boundary integral equation is dened as follows c(r)(r, ) + (n DG(|r r | , )) (r , )d = = G(|r r | , ) (n D(r , )) d + (1.146) G(|ris r | , ) q d
where the potential kernels G and n DG are dened as follows G= where k = [ ] 1 1 K0 k (r r )T D(r r ) = K0 (k t) 1 / 2 2 |D| 2 |D|1/2 and t is an auxiliary function dened as: (1.147)
1 |D|1/2
[ ]1/2 t = b(x x )2 + a(y y )2 2d(x x )(y y ) . The kernel on the left hand of integral equation demands a little more attention. G G n DG = (a + d) + (b + d) = x y [ ] 1 (a + d) K0 (k t) + (b + d) K0 (k t) = 2 |D|1/2 x y [ ] 1 t t (a + d) K0 (k t) + (b + d) K0 (k t) = 2 |D|1/2 t x t y [ 1 b(x x) d(y y ) ( a + d ) + 2 |D|1/2 t ] a(x x) d(y y ) (b + d) K0 (k t) (1.148) t t
= = = +
After some mathematics we will come to a relatively simple form: n D G = k (x x) + (y y ) K1 (k t) 2 t (1.149)
where K1 is the modied Bessel function of the second kind of order one.
84
So the equation (1.146) under assumption that initial conditions are equal zero, can take the following form: ] [ k (x x) + (y y ) c(r)(r, ) + K1 (k t) (r , )d = 2 t p 1 1 1 (r , ) K ( k t ) K (k tis ) Qis (1.150) = d 0 1 / 2 1/2 0 2 | D | n 2 | D | is=0 where tis = [b(xis x )2 + a(yis y )2 2d(xis x )(yis y )]
1/2
The BIE can be solved numerically in the same way as for the isotropic case.
1.6.2
The singularity is treated in a similar way as for the isotropic material. First, let consider the kernel G for quadratic interpolation function as it was for the Laplace equation. G= 1 1 1 K0 (k t) ln(k t) = (ln k ln t) (1.151) = 1 / 2 1 / 2 2 |D| 2 |D| 2 |D|1/2
As before, rst dealing with the situation when r is the rst node of the element containing r , the distance t(r, r ) can be written as follows t2 = b(x( ) x0 )2 + a(y ( ) y0 )2 2d(x( ) x0 )(y ( ) y0 ) Substituting the quadratic shape functions Eq. (1.52), we obtain t
2
(1.152)
= +
]2 1 (1 + ) [( 2)x0 + 2(1 )x1 + x2 ]2 + b 2 [ ]2 1 a (1 + ) [( 2)y0 + 2(1 )y1 + y2 ]2 + 2 [ ]2 1 2d (1 + ) [( 2)x0 + 2(1 )x1 + x2 ] 2 [( 2)y0 + 2(1 )y1 + y2 ] ] [ ( ( (0) )2 ) 2 (0) (0) (0) ( ) ( )fy ( ) 2d fx ( ) + a fy 2 b fx [
(1.153)
1.6. ANISOTROPIC MEDIUM where = 0.5(1 + ).
85
Repeating this procedure for the situation when r is the second node of the element containing r , we obtain ]2 [ 1 2 2 1 t = b ( 1)x0 x1 + ( + 1)x2 + 2 2 [ ]2 1 2 1 + a ( 1)y0 y1 + ( + 1)y2 + 2 2 [ ] 1 2 1 2d ( 1)x0 x1 + ( + 1)x2 (1.154) 2 2 [ ] 1 1 ( 1)y0 y1 + ( + 1)y2 2 2 [ ( ] ) ( )2 2 (1) (1) (1) (1) = 2 b fx ( ) + a fy ( ) 2d fx ( )fy ( ) where = and the third node [ ]2 1 2 t = b (1 ) [x0 + 2(1 + )x1 + ( + 2)x2 ]2 + 2 ]2 [ 1 (1 ) [y0 + 2(1 + )y1 + ( + 2)y2 ]2 + + a 2 [ ]2 1 2d (1 ) [x0 + 2(1 + )x1 + ( + 2)x2 ] 2 [y0 + 2(1 + )y1 + ( + 2)y2 ] = [ ( ] ) ( (2) )2 2 (2) (2) (2) = 2 b fx ( ) + a fy ( ) 2d fx ( )fy ( ) where = 0.5(1 ) (see Fig. 1.13). Therefore a general expression can be written for the logarithmic term of Eq. (1.151) as follows 1 (ln t) = 2 |D|1/2 ( ) ( )2 ( )2 1 (0) (0) (0) (0) = ln b fx ( ) + a fy ( ) 2dfx ( )fy ( ) = 2 |D|1/2 1 1 ln + 1 / 2 2 |D| ] [ ( ( (0) )2 )2 1 (0) (0) (0) ( ) ( ) f 2 df ( ) + a f ( ) ln b f y x y x 4 |D|1/2 =
(1.155)
(1.156)
86
and analogically for the second and the third nodes.
1.6.3
Comparison FEM and BEM results
Two sets of data dening the anisotropy of the medium were considered [40]. The rst set is: a = 0.008333, b = 0.025, d = 0.0. If we assume that
10 10 10 10 modulus of 10 10 10 10 10 10
2
50
FEM BEM
45 40 35 phase shift of 30 25 20 15 10 5 FEM BEM
60
120
180 angle [deg.]
240
300
360
0 0
60
120
180 angle [deg.]
240
300
360
Figure 1.45: Modulus (left) and phase shift (right) distribution along the boundary
10 10 10 10 modulus of 10 10 10 10 10 10
2
40
FEM BEM
FEM BEM 30 phase shift of
20
10
1.5
0.5
0 radius
0.5
1.5
0 2
1.5
0.5
0 radius
0.5
1.5
Figure 1.46: Modulus (left) and phase shift (right) distribution along the radius of the domain the FEM results are a correct one, than the relative error of BEM results against the FEM results can be calculated. Discrepancies between them for dierent BEM discretizations are presented in the following gures (please consult Fig. 1.45Fig. 1.51). We can notice that relative error is below 10 % and decreasing when the number of boundary elements increases. The main

35 [%] 30 25 20 15 10 5 55 50 45 40 35 30 25 20 15 10 5 0 1 2 3 4 5 6 [rd] 0 1 2 3 4 5 6 [rd] [%]
87
Figure 1.47: Relative error distribution for modulus (left) and phase shift (right) along the boundary discretized by 256 nodes
[%] 25 [%] 20
20
15
15 10 10 5 5 0 1 2 3 4 5 6 [rd] 0 1 2 3 4 5 6[rd]
Figure 1.48: Relative error distribution for modulus (left) and phase shift (right) along the boundary discretized by 512 nodes
20 25 [%] 20 14 15 12 10 10 8 6 5 4 2 0 1 2 3 4 5 6 [rd] 0 1 2 3 4 5 6[rd] 18 16 [%]
Figure 1.49: Relative error distribution for modulus (left) and phase shift (right) along the boundary discretized by 768 nodes dierences between results are caused by dierent mathematical models of the point sources.
88
CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS b = 0.008333, d = 0.0.

FEM BEM
The second data set is dened by: a = 0.025,

10 10 10 modulus of 10 10 10 10 10
2
70
FEM BEM
60 50 phase shift of 40 30 20 10 0 0
10
12
60
120
180 angle [deg.]
240
300
360
60
120
180 angle [deg.]
240
300
360
Figure 1.50: Modulus (left) and phase shift (right) distribution along the boundary
10 10 10 modulus of 10 10 10 10 10
10 FEM BEM 10 10 modulus of 10 10 10 10 10
FEM BEM
10
10
12
12
1.5
0.5
0 radius
0.5
1.5
1.5
0.5
0 radius
0.5
1.5
Figure 1.51: Modulus (left) and phase shift (right) distribution along the radius of the domain
1.7
Galerkin formulation of boundary integral equations
In this section the concept of the Galerkin approach to Boundary Element Analysis (BEA) will be described, however with restriction to the constant elements and 2D space. Arise the justied question why? So far it is not proved that Galerkin approach is more ecient then standard BEM,
1.7. GALERKIN FORMULATION OF BIE
89
especially for 3D space, due to a very complicated procedure concerning with the singular integrals. The collocation approach to BEM, discussed in the previous chapters, involved the satisfaction of boundary integral equations at a discrete number of source points r. The alternative Galerkin BEM (GBEM) approach involves the satisfaction of the governing boundary integral equations in an integral or weighted residual sense, leading to double surface integrations. The GBEM can generate symmetric sets of algebraic equations, a property of considerable benet in many contexts. The collocation process described in previous chapters does not treat the source and sample (observation) points symmetrically. In order to illustrate this situation let consider the situation presented in Fig. 1.52. In the Fig. 1.52a) and Fig. 1.52b), the roles of the source and sample (observation) nodes have been reversed.
a)
i
b)
i
r y j r() 0 x
0
r ()
y j r x
Figure 1.52: a) Four element object with Source and Observation points reversed
It is quite obvious that the values of the integrals computed in these two
90
cases, will not be the same. Bi,j (r, r ( )) =

+1
G(|r r ( )|)J ( )d (1.157)
Bj,i (r( ), r ) =
+1
G(|r r( )|)J ( )d
One important dierence in these two quantities is the value of Jacobian determinant J . For the constant elements, J is constant and proportional to the element length. Thus, the values of J ( ) and J ( ) will generally be dierent. Now, we came up to the understanding that an approach with the potential to produce symmetry in the BEM coecient matrices involves two integrations. The conventional governing equations (see for example Eq.(1.2)), can be multiplied by a weight (or test) function and integrated with respect to r around for a second time [1]. In order to treat r( ) and r ( ) symmetrically, we can choose the weight to be the same as the boundary element interpolation function associated with a particular node. This is called a Galerkin approach because the same functions are used both to approximate the response and to test it. For example, the quantities associated with the situation shown in Fig. 1.52 become: +1 +1 Bi,j (r, r ( )) = Ni ( )Nj ( )G(|r r ( )|)J ( )J ( )dd
1 1
Bj,i (r( ), r ) =
+1
(1.158)
+1
Nj ( )Ni ( )G(|r r( )|)J ( )J ( )dd
where N is the weight function2 . It should be obvious that Bi,j (r, r ( )) = Bj,i (r( ), r ). When more sophisticated elements are employed, the analogous and still symmetric quantities are calculated. However, the coecients of the matrix A (see for example
2
For the constant boundary element the weight (test) function is equal to one.
91
Eq.(1.20), the matrix that has the contributions associated with integrals involving the rst derivative of the Green function) are asymmetric. That is why, for the mixed boundaryvalue problem, lack of symmetry is associated with the Galerkin BEM. The apparent loss of symmetry associated with the Galerkin BEA approach to the mixed boundaryvalue problem can be remedied, but this is not the subject of this chapter. Computer implementation of the Galerkin BEM requires essentially the same capabilities present in the ordinary BEM. In the Galerkin approach, the integration over the two lines (in 2D) is mathematically equivalent to the integration over a surface. Thus, the surface integral numerical integration techniques described in chapter 2 for 3D BEM can be employed here. When we integrate both sides of Eq.(1.2) and multiply then by the weight function we will get: ) G(|r r |) c(r)(r)d(r) + (r )d(r ) d(r) = n (1.159) ) ( (r ) = G(|r r |) d(r ) d(r) n (
1.7.1
Analytical integrations of coincident integrands
Nonsingular integrals we can integrate in a similar way as for conventional BEM (see chapter 1.2.2, Eq.(1.26)) using twice the GaussLegendre integration rules. However, integration of the singular integrals can be somehow dicult. As it was mentioned earlier, the 3D integration over the surface rules could be applied in this case. But fortunately, for the constant elements singular integrals can be calculated analytically (compare Eq.(1.27)).
92
As a benchmark example, the Dirichlet problem for the Poisson equation 2 (r) = f (r), (r) = = 0, where r = [x1 , x2 ]. is solved in a quadratic twodimensional domain = [0, 1] [0, 1]. At the boundaries the potential function is kept to zero. The interior is subjected to stationary source function f . The boundary is divided into eight elements, two for each side (see Fig. 1.53). Taking into account that the r ; (1.160) r ,
x2 1
=0
=0
=0
x1 =0 1
Figure 1.53: Quadratic domain discretized by eight boundary elements
potential function is vanishing at the boundaries, the entries of matrix Ai,i became zero and the nonzero entries of the matrix Bi,i can be calculated analytically: 1 Bi,i = 2
0 L
ln
(r r )2 dr dr =
L2 (3 + 2 ln L) 4
(1.161)
where L denotes the length of the boundary element.
93
The constant element is interesting because, having results of analytical integration, we are able to estimate an error of numerical integration, which is the only one possible solution in case of more sophisticated boundary elements.
1.7.2
Numerical integrations of coincident integrands
In case of Galerkin approach to boundary element method, we have to deal with double integrals. Formally, it is equivalent to integration over the surface of the square region. So, the singularity will occur along the diagonal of the square (see Fig. 1.54). In order to get correct results, seems to be reasonable to use the Gauss Legendre integration rule which is applying dierent number of integration points in each direction (see Eq.(1.162)). This method can be called direct integration method. (m1 ) +1 +1 n1 I= f (1 j , 2 i )wj wi (1.162) f (1 , 2 )d1 d2 =
1 1 i=0 j =0
where n number of integration points in 1 direction, m number of integration points in 2 direction. The complete list of the weight coecients and Gaussian integration points can be found in [20]. In the table 1.3 results of numerical experiment are presented. The results are compared with the regularization method (Fig. 1.54) and using the analytical results the relative error is calculated. Inspecting the table 1.3 we can see that for the last case the error drops beneath the level of 0.5%. This satisfactory result was achieved using huge number of integration points. Such number of integration points is not justied. That is why new, more eective method with better precision and less expensive regarding the time of execution is needed. The idea was derived from the surface integration methods of 3D BEM (see the next chapter).
94
CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS Table 1.3: Numerical integration of singular double integrals relative error of relative error of direct integration regularization method method [%] [%] 5.56 2.81 0.88 0.46
3 T 1 T2 JT 1 0 1 JT 2 0 t 2
case number 1 2 3 4
number of integration points 20 ( 4 in x and 5 in y ) 90 ( 9 in x and 10 in y ) 992 (31 in x and 32 in y) 3660 (60 in x and 61 in y)
0.90 0.26 0.03 0.008
1 0
Figure 1.54: Regularization method of integration of double singular integrals
Among many dierent methods [4, 21, 35], the regularization method was selected as a very simple, eective and producing excellent results as we can see in the table 1.3. For the GBEM the regularization method rely on subdividing the square
95
region into two subtriangles and next those subtriangles are mapped into two squares (see Eq.(1.163)) as it is shown in Fig. 1.54. 1 = t 1 + t 2 2 = 1t 1 = 1 + 2t 2 = 1 (1.163)
Next we can use the standard GaussLegendre rule for numerical integration (see Eq.(1.162)). The Jacobian of transformation is the same in both cases 1 and is equal to JT = JT 1 = JT 2 = 1+ . 2
96
Chapter 2 Three-dimensional potential problems
2.1
Introduction
Like Finite Element Method (FEM), the Boundary Element Method (BEM) provides a general numerical tool for the solution of complex engineering problems. In the last decades, the range of its applications has remarkably been enlarged. This chapter is dealing with the application of BEM to Electrical Impedance or Optical Tomography, a new to my best knowledge eld of application. Nevertheless, the BEM still demands an explicit expression of a fundamental solution, which is only known in simple cases. Therefore, in Optical Tomography BEM is restricted to a diusion approximation of transport equation. But hopefully this restriction can be lifted in future, as just recently an alternative BEMformulation, based on the Fourier transform, has been released [30]. The new formulation only needs the Fourier transform of the fundamental solution, which can be constructed, in contrast to the fundamental solution itself, for all linear and homogeneous dierential operators 97
98
CHAPTER 2. 3D POTENTIAL PROBLEMS
by a simple matrix inversion of the transformed dierential operator. Hence, the realm of applications of BEM can be extended remarkably. In recent applications in Optical or Impedance tomography, researchers attention is focused on threedimensional problems. They are a lot more dicult than those dened in the twodimensional space. Mainly due to the geometry which demands a sophisticated discretization with enormously big number of unknowns. Such problems are named Large Scale Problems . BEM is characterized by the boundaryonly property of the algorithm. This property reduces the number of unknowns in BEM as compared to those in methods of the domain type such as Finite Dierence Method (FDM) or Finite Element Method (FEM). However, the reduced number of unknowns does not necessarily lead to improved eciency, because BEM generally produces a full asymmetric matrix of coecients, while the matrices for FDM or FEM are usually sparse and very often symmetric. Because of this drawback, BEM has so far been considered to be less ecient than these domain type competitors in large scale problems. However, the situation is changing with the recent breakthrough introduced by the so called fastBEMs based on techniques such as multiple methods [24], panel clustering, the use of wavelet bases, etc. These fast BEMs can compute potential functions at the all collocation points with O(N ) O(N (log N )m ) (m 0) operations in problems with N unknowns. This is a dramatic improvement over the conventional BEMs which have O(N 2 ) number of operations. Development of the fast BEM is certain to further enhance the status of BEM as a solver of large scale problems.
2.2. SINGULAR AND NEARLY SINGULAR INTEGRALS
99
2.2
Singular and nearly singular integrals
In three dimensional boundary element analysis, computation of integrals is an important aspect since it governs the accuracy of the analysis and also because it usually takes the substantiable part of the CPU time. The integrals which determine the inuence matrices, the internal eld and its gradients contain nearly singular kernels of order 1/R ( = 1, 2, 3, 4, ...) where R is the distance between the source point and the integration point on the boundary element [38]. For planar elements, analytical integration may be possible [60]. However, it is becoming increasingly important, in practical boundary element codes, to use curved elements, such as the isoparametric elements, to model general curved surfaces [77]. Since analytical integration is not possible for general isoparametric curved elements, one has to rely on numerical integration. When the distance between the source point and the element over which the integration is performed is suciently large, compared to the element size, the standard GaussLegendre quadrature formula works eciently. However, when the source is actually on the element, the kernel becomes singular and the straight forward application of the GaussLegendre quadrature formula breaks down. These integrals will be called singular integrals. Singular integrals occur when calculating the diagonals of the coecient matrix. When the source is not on the element, but very close to the element, although the kernel is regular in the mathematical sense, the value of the kernel changes rapidly in the neighborhood of the source point. In such case the standard GaussLegendre quadrature formula is not practical, since it would require a huge number of integration points to achieve the required accuracy. These integrals will be called nearly singular integrals. Nearly singular integrals occur in practice when calculating inuence matrices for thin struc-
100
tures, where distances between dierent elements can be very small compared to the element size. Such situation is very common for scull or CSF layer modelling in Impedance or Optical Tomography. They also occur when calculating the eld or its derivatives at the internal point, very close to the boundary element (see for example Fig. 1.17). Numerous research works have already been published on this subject, for example [4, 27, 32, 38, 44, 47, 62], and they may be classied as in Table 2.1. Let briey review the methods Table 2.1: Methods of integration for singular and nearly singular integrals in three dimensional Boundary Element Method
No. I II 1 2 3 4 4a 4b 5 Weighted Gauss Singularity subtraction and Taylor expansion Variable transformation Coordinate transformation Triangle to quadrilateral Polar coordinates Finite part integrals Singular integrals No. Nearly singular integrals Analytical (for planar elements only) Numerical 1 2 2a 2b 3 3a Element subdivision Variable transformation Double exponential transformation Cubic transformation Coordinate transformation Polar coordinates and modication
which using the kernel 1/R as the weight function for generating the Gauss integration points. The singularity subtraction with Taylor expansion method [4] expands the singular kernel by the local parametric coordinates. The main terms containing the singularity are subtracted and integrated analytically and the remaining well behaved terms are integrated by Gaussian quadrature. Then there are the coordinate transformation methods. The rst type is the method of transforming a triangular region into a quadrilateral region, so that the node corresponding to the singularity is expanded to an edge of the quadrilateral, so that the singularity is weakened. This method is used in the following chapters (see for example 2.6.2). The second type is the method
2.3. GOVERNING EQUATIONS
101
of using polar coordinates (r, ) around the source point in the parameter space [21]. This introduces a Jacobian which cancels the singularity 1/R type. For higher singularities of order 1/R2 , which appears in elastostatic (out of the scope of this book), the method for calculating nite part integrals may be used. Nearly singular integrals turn out to be more dicult and expensive to calculate compared to singular integrals. They are becoming more and more important in practical boundary element codes, since the ability and eciency to calculate nearly singular integrals governs the codes versatility in treating objects containing thin structures (scull or CSF layer of the human head). The reader interested in this particular problem may consult [38] where a new quadrature scheme for the accurate and ecient evaluation of these nearly singular integrals is presented.
2.3
Governing equations
Let consider the Poissons equation in threedimensional space: 2 (r) = b (2.1)
where the stands for the arbitrary potential function as temperature or electric potential. On the surface of the volume the Robin boundary conditions are imposed: (r) = mR (r) + nR n where mR and nR are known Robin boundary condition coecients. The fundamental solution for 3D space is: 1 4R where R = |r r | is a distance between r and r , given by R = |r r | = (x x )2 + (y y )2 + (z z )2 G(|r r |) = (2.3) (2.2)
(2.4)
102
As in the twodimensional formulation, we can use Greens theorem in its second form to derive an integral equation applicable on the surface: G(|r r |) (r ) (r )d(r ) = G(|r r |) d(r ) + (r) + n n b G(|r r |)d(r ) (2.5) +
To make this equation a truly boundaryonly equation, we move the interior load point r to the boundary which results in the following equation: G(|r r |) (r ) c(r)(r) + (r )d(r ) = d(r ) + G(|r r |) n n + b G(|r r |)d(r ) (2.6)
The function c(r) can be calculated by surrounding the boundary point r by a small sphere of radius and taking each term of Eq.(2.6) in the limit as 0. However, as shown in previous chapters for the twodimensional problems, the term c(r) does not need to be calculated explicitly, and can be obtained indirectly by utilizing some simple physical considerations [14]. To calculate the kernel of the rst integral in Eq.(2.6), the Greens function is dierentiated with respect to the unit normal at the point r , as follows: ( ) [ ( ) ( ) ( )] G(|r r |) G R G R x R y R z = = + + (2.7) n R n R x n y n z n where the derivatives of the coordinates x, y and z with respect to the unit outward normal n in point r are the components of the outward normal as follows: nx = and x x R = x R y y R = y R z z R = z R (2.9) x n ny = y n nz = z n (2.8)
2.4. ZEROORDER INTERPOLATION FUNCTIONS Therefore, the rst kernel can be written as follows: G(|r r |) 1 = [(x x)nx + (y y )ny + (z z )nz ] 3 n 4R
103
(2.10)
To solve the threedimensional problem numerically, the surface has to be discretized into elements. The zero linear and quadratic order elements were used. The numerical implementation of threedimensional problems follows the similar procedure to that of twodimensional problems described before.
2.4
Zeroorder interpolation functions
Let consider the zeroorder interpolation functions for at triangular element. We can dene a new local coordinate system using intrinsic variables 1 and 2 with its origin at the node number zero as it is shown in the Fig. 2.1.
1111111111111111111111 0000000000000000000000 3 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0 0 0000000000000000000000 1111111111111111111111 n 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0
0 1 x z
2
2 2
(0,1) 2 2 =1 1
y 1
0 (0,0)
1 (1,0)
Figure 2.1: Global coordinates of the at triangle (left) and local one (right)
So the global coordinates may be expressed as a function of the local coordinates in the usual form:
104
x(1 , 2 ) =
2 k=0 2
Nk (1 , 2 )xk = N0 (1 , 2 )x0 + N1 (1 , 2 )x1 + N2 (1 , 2 )x2 ,
y (1 , 2 ) =
Nk (1 , 2 )yk = N0 (1 , 2 )y0 + N1 (1 , 2 )y1 + N2 (1 , 2 ) y2 , (2.11)
z (1 , 2 ) =
k=0 2 k=0
Nk (1 , 2 )zk = N0 (1 , 2 )z0 + N1 (1 , 2 )z1 + N2 (1 , 2 ) z2 ,
where the basis interpolation functions are: N0 (1 , 2 ) = 1 1 2 , N1 (1 , 2 ) = 1 , N2 (1 , 2 ) = 2 . The rst derivatives of the standard interpolation functions with respect to the 1 and 2 are given by: N0 (1 , 2 ) = 1, 1 N1 (1 , 2 ) = 1, 1 N2 (1 , 2 ) = 0, 1 N0 (1 , 2 ) = 1, 2 N1 (1 , 2 ) = 0, 2 N2 (1 , 2 ) = 1. 2 (2.12)
(2.13)
The unknown values of or q = as it was in 2D case are assumed to n be constant on each triangle and are xed in the geometrical center of the triangular element.
2.4.1
Jacobian
To study boundary elements which are twodimensional structures placed in the 3D space, rst we need to dene the way in which we can pass from the xyz global Cartesian system to the 1 , 2 , 3 system dened over the element, where 1 , 2 are oblique coordinates and 3 is in the direction of the normal
2.4. ZEROORDER INTERPOLATION FUNCTIONS
105
(see Fig. 2.1). The transformation for a given function u is related through the following: , (2.14)
u 1 u 2 u 3
x 1 x 2 x 3
y 1 y 2 y 3
z 1 z 2 z 3
u x u y u z
where the square matrix is the Jacobian matrix (or Jacoby matrix). Transformation of this type allows us to describe dierentials of surface in the Cartesian system in terms of the curvilinear coordinates. A dierential of area will be given by:
d = nd1 d2 =
r r d1 d2 = nx 2 + ny 2 + nz 2 d1 d2 , 1 2
(2.15)
where
nx =
y z y z , 1 2 2 1 x z z x ny = , 1 2 2 1 x y x y nz = . 1 2 2 1
(2.16)
This mapping introduces the Jacobian1 J proportional to the magnitude of the area of the mapped boundary element [14].
Jacobian is shorthand for determinant of Jacoby matrix.
106
2.4.2
Integration of nonsingular integrals over the triangle
Now the boundary integral equation Eq.(2.6) after discretization will take the form: M 1 G(|r r |) c(r)(r) + (r )dj (r ) = n j =0 =
M 1 j =0 j j
(r ) G(|r r |) dj (r ) + n
b G(|r r |)d(r ) (2.17)
where j is the j th boundary triangle. And next can be written in terms of local coordinates 1 , 2 : c(r)i (r) +
M 1 j =0 M 1 j =0
j (r )
1 =0 2 =0
1 =1 2 =1 1
G(|r r |) Nj Jj (1 , 2 )d1 d2 = n
j (r ) n
1 =1 2 =1 1
G(|r r |)Nj Jj (1 , 2 )d1 d2 +
(2.18)
1 =0
2 =0
b G(|r r |)d(r )
where Nj = 1 for the zero order triangle. The numerical integration rules discussed in this section, perform well when the distance from the source point is relatively far from the element being integrated. Twodimensional integrals over the triangle can be evaluated by application of the equation similar to Eq.(1.24) for onedimensional integrals.
0 1
11
f (1 , 2 )d1 d2 =
g 1 i=0
wi f (1 i , 2 i )
(2.19)
where g is the total number of Gaussian integration points. The abscissas and associated weights are given in Table 2.2 [21, 89].
2.4. ZEROORDER INTERPOLATION FUNCTIONS Table 2.2: Gauss points and weights over the triangle
n 4 i 1 2 3 4 1 2 3 4 5 6 7 1 2 3 4 5 6 7 8 9 10 11 12 1 1./3. 0.6 0.2 0.2 0.3333333333333333 0.0597158717897698 0.4701420641051151 0.4701420641051151 0.7974269853530873 0.1012865073234563 0.1012865073234563 0.06308901449150223 0.87382197101699554 0.06308901449150223 0.24928674517091042 0.50142650965817920 0.24928674517091042 0.31035245103378440 0.63650249912139870 0.05314504984481695 0.63650249912139870 0.05314504984481695 0.31035245103378440 2 1./3. 0.2 0.6 0.2 0.3333333333333333 0.4701420641051151 0.0597158717897698 0.4701420641051151 0.1012865073234563 0.7974269853530873 0.1012865073234563 0.87382197101699554 0.06308901449150223 0.06308901449150223 0.50142650965817920 0.24928674517091042 0.24928674517091042 0.63650249912139870 0.31035245103378440 0.63650249912139870 0.05314504984481695 0.31035245103378440 0.05314504984481695 wi -9./32. 25./96. 25./96. 25./96. 0.1125000000000000 0.0661970763942531 0.0661970763942531 0.0661970763942531 0.0629695902724136 0.0629695902724136 0.0629695902724136 0.025422453185103408 0.025422453185103408 0.025422453185103408 0.058393137863189683 0.058393137863189683 0.058393137863189683 0.041425537809186787 0.041425537809186787 0.041425537809186787 0.041425537809186787 0.041425537809186787 0.041425537809186787
107
12
2.4.3
Integration of singular integrals
The singular integrals in zeroorder triangular element over a at area can be evaluated analytically. The symbols used in the following closed forms are shown in Fig. 2.2. [ tan [(0 + 1 )/2] 2 1 1 + dS = ln 3 r12 tan(1 /2) r ] 1 tan [(1 + 2 )/2] tan [(2 + 0 )/2] 1 + ln + ln r20 tan(2 /2) r01 tan(0 /2)
(2.20)
108 and
CHAPTER 2. 3D POTENTIAL PROBLEMS ( ) 1 dS = 0 r
(2.21)
where is an area of the triangular boundary element in 3D space.

2 (x2 ,y2 ,z 2) 2 r 2 1 0 (x g,y g,z g) 2 r 01 r1 1 1 (x ,y ,z )
1 1 1
r 20
r 12
r0 0 (x0 ,y0 ,z 0) 0
Figure 2.2: Notations for geometric information of triangular element
2.5
Firstorder interpolation functions
Let consider the rstorder interpolation functions for a at linear triangle. The geometry of the boundary discretization is the same as in case of the zeroorder triangle, so the same basis interpolation functions (see Eq.(2.12)) can be used for the state functions and its normal derivative representation:
(1 , 2 ) =
2 k=0
Nk (1 , 2 )k = N0 (1 , 2 )0 + N1 (1 , 2 )1 + N2 (1 , 2 )2 (2.22)
(1 , 2 ) n
2 k=0
Nk (1 , 2 )
k 0 1 2 = N0 (1 , 2 ) + N1 (1 , 2 ) + N2 (1 , 2 ) n n n n
In case of linear triangular boundary elements the numerical integration of the kernels may be done in a similar way as for constant elements. The
2.5. FIRSTORDER INTERPOLATION FUNCTIONS

z 2 2 n 2 0 0 n 0 0 1 x y 1 11 n 2
109
Figure 2.3: Global and local coordinates of the three nodes triangle
Jacobian of transformation and the components of unit outward normal are calculated according to Eq.(2.15) and Eq.(2.16) respectively. The boundary curve is divided into linear boundary elements j . The numerical integration is performed over each linear boundary element j using the local intrinsic coordinate 1 and 2 , as follows:
c(r)i (r) +
M 1 1 j =0 0
1 1
G(|r r |) (j ) k (r )Nk (1 , 2 )J (1 , 2 ) d1 d2 = n
2 k=0
M 1 1 j =0 0
(2.23)
1 1
G(|r r |)
2 (j ) (r ) k k=0
Nk (1 , 2 )J (1 , 2 ) d1 d2
where M is the total number of linear boundary elements. The nodal values are constant, so we can nally rewrite Eq.(2.23) in the following form:
c(r)i (r) +
M 1 2 j =0 k=0
k (r )
(j )
11
G(|r r |) Nk (1 , 2 )J (1 , 2 ) d1 d2 = n (2.24)
M 1 2 j =0 k=0
(j ) k (r )
11
G(|r r |)Nk (1 , 2 )J (1 , 2 ) d1 d2
110
2 n

2 n 2
n 0 r z y x 2
2
0 1 z y r 1 z
0 r y J1 2
1
J0 2
0
J2
J rT0
2 2
J rT 1 2
0
JrT2 2
1
1 1
Figure 2.4: Linear boundary element mapping procedure
Node 0 Based on Eq.(2.36) the transformation from global to local coordinate system is: x = N0 (1 , 2 )x0 + N1 (1 , 2 )x1 + N2 (1 , 2 )x2 = = (1 1 2 )x0 + 1 x1 + 2 x2 = x0 + 1 (x1 x0 ) + 2 (x2 x0 ) y = N0 (1 , 2 )y0 + N1 (1 , 2 )y1 + N2 (1 , 2 )y2 = z = N0 (1 , 2 )z0 + N1 (1 , 2 )z1 + N2 (1 , 2 )z2 = = (1 1 2 )z0 + 1 z1 + 2 z2 = z0 + 1 (z1 z0 ) + 2 (z2 z0 ) (2.25) = (1 1 2 )y0 + 1 y1 + 2 y2 = y0 + 1 (y1 y0 ) + 2 (y2 y0 )
2.5. FIRSTORDER INTERPOLATION FUNCTIONS
111
Jacobian J0 (1 , 2 ), will be associated with the change of variables from global coordinates to the local one and is given by: J0 = Node 1 For singularity located at node 1: x = N0 (1 , 2 )x1 + N1 (1 , 2 )x2 + N2 (1 , 2 )x0 = = (1 1 2 )x1 + 1 x2 + 2 x0 = x1 + 1 (x2 x1 ) + 2 (x0 x1 ) y = N0 (1 , 2 )y1 + N1 (1 , 2 )y2 + N2 (1 , 2 )y0 = z = N0 (1 , 2 )z1 + N1 (1 , 2 )z2 + N2 (1 , 2 )z0 = = (1 1 2 )z1 + 1 z2 + 2 z0 = z1 + 1 (z2 z1 ) + 2 (z0 z1 ) A Jacobian J1 (1 , 2 ) is given by: J1 = Node 2 For singularity located at node 2: x = N0 (1 , 2 )x2 + N1 (1 , 2 )x0 + N2 (1 , 2 )x1 = = (1 1 2 )x2 + 1 x0 + 2 x1 = x2 + 1 (x0 x2 ) + 2 (x1 x2 ) y = N0 (1 , 2 )y2 + N1 (1 , 2 )y0 + N2 (1 , 2 )y1 = z = N0 (1 , 2 )z2 + N1 (1 , 2 )z0 + N2 (1 , 2 )z1 = = (1 1 2 )z2 + 1 z0 + 2 z1 = z2 + 1 (z0 z2 ) + 2 (z1 z2 ) A Jacobian J2 (1 , 2 ) is given by: J2 =
x 1 y 1 x 2 y 2 x 1 y 1 x 2 y 2 x 1 y 1 x 2 y 2
= (x1 x0 )(y2 y0 ) (x2 x0 )(y1 y0 )
(2.26)
(2.27)
= (1 1 2 )y1 + 1 y2 + 2 y0 = y1 + 1 (y2 y1 ) + 2 (y0 y1 )
= (x2 x1 )(y0 y1 ) (x0 x1 )(y2 y1 )
(2.28)
(2.29)
= (1 1 2 )y2 + 1 y0 + 2 y1 = y2 + 1 (y0 y2 ) + 2 (y1 y2 )
= (x0 x2 )(y1 y2 ) (x1 x2 )(y0 y2 )
(2.30)
112
The transformation from a standard triangle to a standard square is: 1 = 1 + 1 (1 + 1 )(1 + 2 ) , 2 4 2 = (1 + 1 )(1 + 2 ) . 4 (2.31)
The Jacobian Jr (1 , 2 ), will be associated with the change of variables which is given by: J r T0 = J r T1 = J r T2 = J r =
1 1 2 1 1 2 2 2
1 + 1 . 8
(2.32)
2.6
Secondorder interpolation functions
As with many numerical methods, the accuracy of the solution obtained with the BEM, increases with the number of discretization points and/or surface elements at the cost of increased computation time and memory requirements. In this section following [32], it will be shown how the accuracy of the BEM could be improved if secondorder interpolation functions for the potential and variation over the surface elements and for the shape of n these surface elements are introduced.
2.6.1
Triangular boundary elements
Let focus our attention on six nodes triangular isoparametric element presented in Fig. 2.5. The shape functions are given by the following formulas: N0 (1 , 2 ) = 3 (1 23 ), N2 (1 , 2 ) = 1 (1 21 ), N4 (1 , 2 ) = 2 (1 22 ), where 3 = 1 1 2 . The rst derivatives of the standard interpolation functions with respect to N1 (1 , 2 ) = 41 3 , N3 (1 , 2 ) = 41 2 , N5 (1 , 2 ) = 42 3 , (2.33)
2.6. SECONDORDER INTERPOLATION FUNCTIONS

z
113
4 1 2 3
r 0
y
1
Figure 2.5: Global and local coordinates of the six nodes triangle
the 1 and 2 are given by:
N0 (1 , 2 ) 1 N1 (1 , 2 ) 1 N2 (1 , 2 ) 1 N3 (1 , 2 ) 1 N4 (1 , 2 ) 1 N5 (1 , 2 ) 1
= 1 43 , = 4(3 1 ), = 1 + 41 , = 42 , = 0, = 42 ,
N0 (1 , 2 ) 2 N1 (1 , 2 ) 2 N2 (1 , 2 ) 2 N3 (1 , 2 ) 2 N4 (1 , 2 ) 2 N5 (1 , 2 ) 2
= 1 43 , = 41 , = 0, = 41 , = 1 + 42 , = 4(3 2 ). (2.34)
In contrast with the threenode triangular element, the quadratic variation in 1 , 2 of the functions N0 .... N5 allows the creation of curved surfaces with curved edges.
114
2.6.2
Numerical integration of singular integrals
The standard integration rules gives an exact results except when the load point is near to or coincident with one of the nodes of the element. The singularity occurs in this case, so that special treatment of the integration is required. Two of the most eective methods will be described. These are the regularization method and the subtraction/expansion method. As it was for 2D case the exact integration is not possible for the quadratic boundary elements. In the following gures the mapping procedure, when the singular point (surrounded by a circle) is presented in the nodes 0, 1 or 2. At rst, three dimensional Cartesian space is mapped into a local twodimensional Cartesian space by interpolation functions.
5 i=0 5 i=0 5 i=0
x =
Ni (1 , 2 )xi ,
y =
Ni (1 , 2 )yi ,
(2.35)
z =
Ni (1 , 2 )zi .
If the singularity is in the vertex nodes then the element in local coordinates system 1 , 2 is mapped into a standard square (see Fig. 2.6), but when singularity is in one of the midside nodes than triangle is divided into two subtriangles and then such triangles are mapped into a standard squares (subelement coordinates)(see Fig. 2.6). The transformation from local elements to subelements coordinates is given by the following expressions:
2 i=0 2 i=0
1 =
Mi (1 , 2 )1 i ,
2 =
Mi (1 , 2 )2 i ,
(2.36)
where 1 and 2 are the local coordinate of the subelement and the shape
2.6. SECONDORDER INTERPOLATION FUNCTIONS functions are given by: (1 + 1 )(1 2 ) , 4 (1 + 1 )(1 + 2 ) M1 (1 , 2 ) = , 4 1 1 M2 (1 , 2 ) = . 2 M0 (1 , 2 ) =
115
(2.37)
To integrate numerically an arbitrary function over the standard square, the

4 5 3 1 0
2
4 5 3 1 2
4 5 3 1 0 2
0
2
4 2 5 2 0 1 1 3 1
4 1 5
0 2
1
11111111 00000000 1 00000000 11111111 00000000 11111111 0 3 00000000 11111111 1 00000000 11111111 2 00000000 11111111 T1 T 00000000 11111111 2 2 0 00000000 11111111 2 1 00000000 11111111 1 2 2 2 0 1 T 1
4 1 5 1 2 1 2 2 2 0 1
1
2 4 1
0 2
1
1
0 4
2
1
0 4
Figure 2.6: The singular point (circled): in the node 0 (left), in the node 1 (middle) and in the node 2 (right) conventional Gaussian quadrature scheme can be easily carried out there [15]. The regularization method transforms triangular domain into a square, in that way introduces a further Jacobian which cancels out the singularity of the integrand. Node 0
116
Based on Eq.(2.36) the transformation is: 1 = (1 + 1 )(1 2 ) , 4 2 = (1 + 1 )(1 + 2 ) . 4 (2.38)
A regularizing Jacobian Jr (1 , 2 ), will be associated with the change of variables which is given by: Jr = Node 1 For subtriangle T1 : 1 = 2 J r T1 For subtriangle T2 : 1 = 2 J r T2 Node 2 1 = 1 1 , 2 2 = (1 + 1 )(1 2 ) . 4 (2.42) 2 (1 + 1 )2 , 4 (1 + 1 )(1 + 2 ) , = 4 1 + 1 = . 16 1 1 , 4 (1 + 1 )(1 2 ) = , 4 1 + 1 = . 16
1 1 2 1 1 2 2 2
1 + 1 . 8
(2.39)
(2.40)
(2.41)
A regularizing Jacobian Jr (1 , 2 ) is given by: Jr = 1 + 1 . 8 (2.43)
The mapping procedure when the singular point is in the nodes 3, 4 or 5. Node 3

4 5 3 1 0
2
117
4 5
4 5 3 1
1
2
2
2

2
1111 0000 0000 1111 0 0000 1111 3 0000 T2 1 1111 0000 1111 2 5 11111111 00000000 0000 1111 2 00000000 11111111 2 1 0000 1111 1 0 00000000 11111111 0000 1111 T1 1 00000000 11111111 0000 1111 1
4 0 2 2 2 2 1 T 3 2 1
4 2 5 0
1
3 2
1 1 1 2 2 1 2
1
11111111 00000000 00000000 11111111 1 2 11111111 00000000 00000000 11111111 1 00000000 11111111 3 2 T2 5 11111111 00000000 2 00000000 11111111 2 00000000 11111111 0 T1 1 0 00000000 11111111 1
4 0 2 1 2 2 1 T 1
0 0
4 1
0 0
1
0 0
Figure 2.7: The singular point (circled): in the node 3 (left), in the node 4 (middle) and in the node 5 (right) For subtriangle T1 : 1 = 2 J r T1 For subtriangle T2 : 1 = 2 J r T2 Node 4 1 = (1 + 1 )(1 + 2 ) , 4 2 = 1 1 . 2 1 1 , 4 2 (1 + 1 )2 = , 4 1 + 1 = . 16 2 + (1 + 1 )2 , 4 1 1 = , 4 1 + 1 = . 16
(2.44)
(2.45)
118
A regularizing Jacobian Jr (1 , 2 ) as in case of node 0 and node 2 is given by: Jr = (1 + 1 )/8. Node 5 For subtriangle T1 : 1 = 2 J r T1 For subtriangle T2 : 1 = 2 J r T2 (1 + 1 )(1 2 ) , 4 2 + (1 + 1 )2 = , 4 1 + 1 = . 16 (1 + 1 )(1 + 2 ) , 4 1 1 = , 4 1 + 1 = . 16 (2.46)
(2.47)
(2.48)
In the Fig. 2.8 it is illustrated how splitting and mapping procedure concentrates the Gaussian points around the singularity node.
119
Figure 2.8: The Gaussian quadrature points used on the square mapped back to the at triangle space (1 , 2 ). The concentration of points around the singularity node 1 is clearly visible
120
2.6.3
Quadrilateral boundary elements
Let consider eight nodes quadrilateral isoparametric element presented in Fig. 2.9.
z
6 7
5 3 2
r 0 y
1
Figure 2.9: Local coordinates of the quadrilateral boundary element
The shape functions are given by the following formulas: N0 (1 , 2 ) = (1 1 )(1 2 )(1 + 1 + 2 )/4,
2 N1 (1 , 2 ) = (1 1 )(1 2 )/2,
N2 (1 , 2 ) = (1 + 1 )(1 2 )(1 1 + 2 )/4,

2 )/2, N3 (1 , 2 ) = (1 + 1 )(1 2
N4 (1 , 2 ) = (1 + 1 )(1 + 2 )(1 1 2 )/4,

2 N5 (1 , 2 ) = (1 1 )(1 + 2 )/2,
(2.49)
N6 (1 , 2 ) = (1 1 )(1 + 2 )(1 + 1 2 )/4,

2 N7 (1 , 2 ) = (1 1 )(1 2 )/2.
The rst derivatives of the standard interpolation functions with respect to
2.6. SECONDORDER INTERPOLATION FUNCTIONS the 1 and 2 are given by: N0 (1 , 2 ) = (1 2 )(21 + 2 )/4, 1 N2 (1 , 2 ) = (1 2 )(21 2 )/4, 1 N4 (1 , 2 ) = (1 + 2 )(21 + 2 )/4, 1 N6 (1 , 2 ) = (1 + 2 )(21 2 )/4, 1
121
N1 (1 , 2 ) = 1 (1 2 ), 1 N3 (1 , 2 ) 2 = (1 2 )/2, 1 (2.50) N5 (1 , 2 ) = 1 (1 + 2 ), 1 N7 (1 , 2 ) 2 = (1 2 )/2, 1
N0 (1 , 2 ) = (1 1 )(22 + 1 )/4, 2 N2 (1 , 2 ) = (1 + 1 )(22 1 )/4, 2 N4 (1 , 2 ) = (1 + 1 )(22 + 1 )/4, 2 N6 (1 , 2 ) = (1 1 )(22 1 )/4, 2
N1 (1 , 2 ) 2 = (1 1 )/2, 2 N3 (1 , 2 ) = 2 (1 + 1 ), 2 (2.51) N5 (1 , 2 ) 2 = (1 1 )/2, 2 N7 (1 , 2 ) = 2 (1 1 ). 2
2.6.4
Integration of nonsingular integrals over the square
Threedimensional boundary element analysis involves the integration of certain functions over the surface of boundary elements. For quadratic elements we can characterize the generic integration task by: I=
1 1 +1
+1
f (1 , 2 )d1 d2 =
( n n
i=1 j =1
) f (1 j , 2 i )wj wi (2.52)
Values of the Gaussian integration points and weights coecients are listed in the table 2.3.
122
CHAPTER 2. 3D POTENTIAL PROBLEMS Table 2.3: Gauss points and weights

n 3 i 1 2 3 1 2 3 4 1 2 3 4 5 6 1 -0.77459666924148337704 0.0 +0.77459666924148337704 -0.86113631159495257522 -0.33998104358485626480 +0.33998104358485626480 +0.86113631159495257522 -0.93246951420315202781 -0.66120938646626451366 -0.23861918608319690863 +0.23861918608319690863 +0.66120938646626451366 +0.93246951420315202781 2 -0.77459666924148337704 0.0 +0.77459666924148337704 -0.86113631159495257522 -0.33998104358485626480 +0.33998104358485626480 +0.86113631159495257522 -0.93246951420315202781 -0.66120938646626451366 -0.23861918608319690863 +0.23861918608319690863 +0.66120938646626451366 +0.93246951420315202781 wi 0.55555555555555555556 0.88888888888888888889 0.55555555555555555556 0.34785484513745385737 0.65214515486254614263 0.65214515486254614263 0.34785484513745385737 0.17132449237917034504 0.36076157304813860757 0.46791393457269104739 0.46791393457269104739 0.36076157304813860757 0.17132449237917034504
2.6.5
Integration of singular integrals over the square
The standard integration rules (see table 2.3) give an exact result except when the load point is near to or coincident with one of the nodes of the element. The singularity occurs in this case, so a special treatment of the integration is required. Two of the most eective methods will be described. These are the regularization method and the subtraction/expansion method [82]. As it was for 2D case, the exact integration is not possible for the quadratic boundary elements. In the following gures the mapping procedure, when the singular point (surrounded by a circle) is in the nodes 0, 1 or 2, is presented. At rst, threedimensional Cartesian space is mapped into a local twodimensional Cartesian space by interpolation functions. x =
7 i=0 7 i=0 7 i=0
Ni (1 , 2 )xi ,
y =
Ni (1 , 2 )yi ,
(2.53)
z =
Ni (1 , 2 )zi .
123
Then the standard square element in local coordinates system 1 , 2 is divided into two or three triangular subelements, depending on the location of the singular point, as it is shown in Fig. 2.10. Lastly, those subtriangles
6 n 0 1 r z y x
5 4 111111111111 000000000000 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 2 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 7 111111111111 3 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 1 000000000000 111111111111 000000000000 111111111111 6
2 2
5 7 4 3 0 2 z J r y x 1
6 r 2 r 1 2
r 7 2
5 4 3 0 z J y r x 7 r 1 1
5 n r 2 3 4
r 1
J rT2
JrT1
2
2 4 5 111111 111111 000000 000000 000000 111111 000000 111111 000000 111111 000000 111111 3 000000 111111 000000 111111 111111 000000 000000 111111 000000 111111 000000 111111 000000 111111 000000 111111 000000 000000 111111 7 111111 3 000000 111111 000000 111111 000000 111111 000000 111111 000000 111111 000000 111111 1 2 000000 111111 000000 000000111111 111111 000000 111111
6
1 1 2 2
2 5 111111111111 000000000000 000000000000 111111111111 4 000000000000 111111111111 1 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 1 000000000000 111111111111 000000000000 7 111111111111 3 000000000000 111111111111 000000000000 111111111111 2 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111
6
2 2
JrT1
4 1 1 2 2 0 1 1 6 2
JrT3
6 1 1 4
JrT2
2 4 1 2 2
JrT2
2 0 1 2 6 2 6
JrT1
6 1 0 4
Figure 2.10: The singular point in the node 0 (left), the singular point in the node 1 (middle) and the singular point in the node 2 (right)
are mapped into the standard square again, so the conventional Gaussian quadrature scheme can be easily carried out there. The regularization method using transformation of a triangular domain into a square and in that way introduces a further Jacobian which cancels out the singularity of the integrand. Node 0
124 For triangle T1 : 1 = 1 , For triangle T2 : 1 =
2 =
1 + 1 + (1 + 1 )2 . 2
(2.54)
1 + 1 (1 + 1 )2 , 2
2 = 1 .
(2.55)
A regularizing Jacobian JrT1 = JrT2 = Jr (1 , 2 ), is associated with the change of variables and is given by Jr = Node 1 For triangle T1 : 1 = 1 + 1 , 2 2 = 1 + 1 (1 + 1 )2 . 2 (2.57)
1 1 2 1 1 2 2 2
1 + 1 . 2
(2.56)
A regularizing Jacobian JrT3 (1 , 2 ), is expressed by: J r T1 = For triangle T2 : 1 = 1 + 1 , 2 2 = 1 + 1 + (1 + 1 )2 . 2 (2.59) 1 + 1 . 4 (2.58)
A regularizing Jacobian JrT3 (1 , 2 ), is expressed by: J r T2 = For triangle T3 : 1 = (1 + 1 )2 , 2 2 = 1 . (2.61) 1 + 1 . 4 (2.60)
A regularizing Jacobian JrT3 (1 , 2 ), is expressed by: J r T3 = 1 + 1 . 2 (2.62)
2.6. SECONDORDER INTERPOLATION FUNCTIONS Node 2 For triangle T1 : 1 = For triangle T2 : 1 = 1 , 2 = 1 + 1 (1 + 1 )2 . 2 1 1 (1 + 1 )2 , 2 2 = 1 .
125
(2.63)
(2.64)
As it was in case of the node 0 the regularizing Jacobian Jr (1 , 2 ) for triangle T1 and for the triangle T2 is the same and is given by: Jr = 1 + 1 . 2 (2.65)
The mapping procedure when the singular point is placed in the nodes 3, 4 or 5 are presented below. Node 3 For triangle T1 : 1 = 1 1 + (1 + 1 )2 , 2 1 + 1 . 4 2 = 1 + 1 , 2 (2.66)
J r T1 = For triangle T2 : 1 =
(2.67)
1 1 (1 + 1 )2 , 2 1 + 1 . 4
2 =
1 + 1 , 2
(2.68)
J r T2 = For triangle T3 : 1 = 1 ,
(2.69)
2 =
(1 + 1 )2 . 2
(2.70)
126
6 7 r 1 1 2 z y x
5 4 111111111111 000000000000 000000000000 111111111111 000000000000 111111111111 2 000000000000 111111111111 3 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 7 111111111111 3 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 1 000000000000 111111111111 6
2 2 1 2

5 n 7 3 r 4 2 0 z J y x 2 2 4 6 5 1111111111111 0000000000000 0000000000000 1111111111111 0000000000000 1111111111111 2 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 3 7 1 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 1 0000000000000 1111111111111 0000000000000 1111111111111
2 2
5 r 2 3
r 1 4
n n 7 6 r 2 3 5 r 1 4
1 r 2 J z
0 y r
1 2 J x
1 J rT 2
J rT3
1 0 1 3 6 2
J rT1
2 1 3 0
J rT 2
6
1
1111111 0000000 000000 111111 4 0000000 1111111 000000 111111 2 1 0000000 1111111 000000 111111 0000000 1111111 000000 111111 0000000 1111111 000000 111111 1 0000000 1111111 000000 111111 0000000 1111111 000000 111111 7 3 0000000 1111111 000000 111111 0000000 1111111 000000 3 111111 0000000 1111111 000000 111111 0000000 1111111 000000 111111 0000000111111 1111111 000000
2 2 2 1
0
1
J rT1
2
J rT 1
2
1
0
1
J rT 3
2 1 5 0
JrT 2
4
1
4 6
5 6
Figure 2.11: Singularity in the node 3 (left), in the node 4 (middle) and in the node 5 (right) A regularizing Jacobian JrT3 (1 , 2 ), is associated with the change of variables and is given by: 1 + 1 J r T3 = . (2.71) 2 Node 4 For triangle T1 : 1 = For triangle T2 : 1 1 (1 + 1 )2 . (2.73) 2 A regularizing Jacobian Jr (1 , 2 ) for both subtriangles is given by: 1 + 1 Jr = . (2.74) 2 1 = 1 , 2 = 1 1 + (1 + 1 )2 , 2 2 = 1 . (2.72)
2.6. SECONDORDER INTERPOLATION FUNCTIONS Node 5 For triangle T1 : 1 = 1 + 1 , 2 2 = 1 1 (1 + 1 )2 , 2
127
(2.75)
J r T1 = For triangle T2 : 1 = 1 + 1 , 2 2 =
1 + 1 . 4
(2.76)
1 1 + (1 + 1 )2 , 2 1 + 1 . 4
(2.77)
J r T2 = For triangle T3 : 1 =
(2.78)
(1 + 1 )2 , 2
2 = 1 .
(2.79)
A regularizing Jacobian JrT3 (1 , 2 ), is given by: J r T3 = Node 6 For triangle T1 : 1 = For triangle T2 : 1 = 1 , 2 = 1 1 + (1 + 1 )2 . 2 (2.82) 1 + 1 + (1 + 1 )2 , 2 2 = 1 . (2.81) 1 + 1 . 2 (2.80)
A regularizing Jacobian Jr (1 , 2 ), is associated with the change of variables and is given by: Jr = 1 + 1 . 2 (2.83)
128
n r 2 7

r 1 6
5 4
n 7
r 2 6 r 1
5 4 3
0 r z y
3 1 2 z J x
6 5 2 4 1111111111111 0000000000000 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 1 0000000000000 1111111111111 1 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 3 7 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 2 0000000000000 1111111111111 0000000000000 1111111111111
2 2
0 y
1 2 J x
J rT 2
5 111111111111 000000000000 000000000000 111111111111 4 1 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 1 000000000000 111111111111 000000000000 7 111111111111 3 000000000000 111111111111 3 000000000000 111111111111 000000000000 111111111111 2 000000000000 111111111111 000000000000 111111111111 6
2 2 1 2
2
1
J rT1
2
J rT 1
J rT 2
2 6
1
J rT 3
4
1
2 1 7 0
4
1
6 0
7 4
Figure 2.12: Singularity in the node 6 and in the node 7
Node 7 For triangle T1 : 1 = 1 + 1 + (1 + 1 )2 , 2 J r T1 = For triangle T2 : 1 = 1 + 1 (1 + 1 )2 , 2 J r T2 = 1 + 1 . 4 2 = 1 + 1 , 2 (2.86) 1 + 1 . 4 2 = 1 + 1 , 2 (2.84)
(2.85)
(2.87)

6 5 4
6 5 4
129
Figure 2.13: The Gaussian points concentration around the singular node 0
6 5 4
T 1 1 2
Figure 2.14: The Gaussian points concentration around the singular node 1
For triangle T3 : (1 + 1 )2 . 2
1 = 1 ,
2 =
(2.88)
A regularizing Jacobian JrT3 (1 , 2 ), is given by: J r T3 = 1 + 1 . 2 (2.89)
In the g. 2.13 and in the g. 2.14 it is illustrated how splitting and mapping procedure concentrates the Gaussian points around the singularity nodes in rst two cases.
130
2.7
Treatment of Boundary Conditions
When the boundary conditions are imposed on the whole boundary, the number of unknowns is reduced from 2N to N . We can distinguish the following cases:
1. Dirichlet boundary conditions, 2. Neumann boundary conditions, 3. Robin boundary conditions, 4. Mixed boundary conditions.
All of those cases will be described in the following sections.
2.7.1
Dirichlet boundary conditions
For Dirichlet boundary conditions vector = D is specied and the vector is unknown. Then the Eq.(2.2) which in matrix form is A = B +q n n can be rearranged to the following form: B = AD q = CD n (2.90)
2.7.2
Neumann boundary conditions

n = ( ) is known and vector is unknown: n N
In this case the vector
A = (B
) + q = CN n N
(2.91)
2.8. NONHOMOGENEITY
131
2.7.3
Robin boundary conditions
Robin boundary conditions (see Eq.(2.2)) express linear dependency of both vectors and . So the procedure is as follows: dene as a function of n n and then introduce to Eq.(2.90): (A + 1 B) = q = CR 2D (2.92)
2.7.4
Mixed boundary conditions
When on some part of the boundary surface the Dirichlet boundary conditions are imposed and on the rest of the surface the Robin boundary conditions are imposed, then the rearranging process is more complicated. Let assume that rst m1 rows of vector are the Dirichlet boundary conditions and the rest m m1 are of Robin type. [ ][ ] [ ] [ ][ ] A11 A12 D B11 B12 q 1 n = + (2.93) 1 A21 A22 2D B21 B22 q2 After some mathematics nal system of equations has the following form: [ ] [ ] 1 (A12 + 2D B12 ) B11 A + q 11 D 1 = (2.94) 1 A + q 21 D 2 (A22 + 2D B22 ) B21 n
2.8
Nonhomogeneity
Basically the BEM method is designed for the solution in the homogeneous areas. However very often we have to nd out the solution inside the regions which are spatially homogeneous. Then each region is considered separately and the solution is sewed (see Eq.(2.95)) on an interface or on the interfaces if we have more than two subregions. Such approach is general and describes 2D and 3D problems. Let introduce the superscripts (i) denoting the ith region.
132
The photon density and the current of photons along the nodes on the interface must therefore satisfy the following conditions: i D(i1) n
(i1) i (i1)
= i
(i)
(2.95) = D(i) n
(i) i
The system of equations describing the solution in the structure consisting of n subregions is: G(1) (|r r |, ) (1) (1) c(r) (r) + (r )d(r ) = n =

(r ) G (|r r |, ) d(r ) Qs G(1) (|rs r|, ) n s=1

(1) (1) ns
. . . c(r)
(n)
(2.96)
(r) +
G(n) (|r r |, ) (n) (r )d(r ) = n G(n) (|r r |, ) (n) (r ) d(r ) n
The equations (2.96) may be presented in a matrix form: (n1) A21 0 (1) B11 (1) B21 . . = . (n1) B21 0 A11 (1) A21 . . .
(1)
... ... ...
A12 (1) A22 . . .

(n1)
(1)
. . . A22 (n) . . . A11 ... ... .. . B12 (1) B22 . . .

(1)
. . . B22 (n) . . . B11
(n1)
= (n1) n (n) n (1) (1) q1 Jn 1 (1) (1) Jn 2 q2 . . + . . . . (n1) 0 Jn n (n) 0 Jn n
1 (1) 2 . . .
(1)
(2.97)
2.9. INDEX MISMATCHED DIFFUSIVE/DIFFUSIVE
133
Including the Robin boundary conditions Eq.(2.92) to the Eq.(2.97) we can (1) reduce the number of unknowns by the number of the values Jn 1 . =
(1)
A11 A12 (1) (1) A21 A22 . . . . . . 0 0 0 0 B11 (1) B21 . . . 0 0

(1)
(1)
(1)
... ... .. .
0 0 . . .
(n1)
0 0 . . . 0 (n) A11 0 0 . . . 0 (n) B11
. . . A22 ... 0 ... ... ...
(n1) n (n) n
(1) 1 (1) Jn 2
1 (1) 2 . . .
(1)
= (2.98)
(1) B12 (1) B22
. . . 0 0
0 0 . . .
(n1)
. . . B22 ... 0
. . . (n1) Jn n (n) Jn n
q1 (1) q2 (1) . . . 0 0
where B11 and B21 are the matrices modied by Robin boundary conditions. Both sides of the Eq.(2.98) have the unknown vectors. In order to solve such a system we have to reorganize it. The unknown values should be transferred to the lefthand side and nally we come to the blockbounded form: (1) (1) (1) (1) A11 B11 A12 B12 . . . 0 0 1 q1 (1) (1) (1) (1) A21 B21 A22 B22 . . . 0 0 2 q2 (1) (2) 0 0 +B11 . . . 0 0 Jn 2 0 = (2.99) . . . . . . .. . . . . . . . . . . . . . . . . (n) 0 0 0 . . . A11 0 n 0 (n1) (n) Jn n 0 0 0 0 . . . 0 +B11
(1)
2.9
Index mismatched diusive/diusive interfaces
In a previous section we have considered the common case when the refractive indices are equal to each other and are constant throughout both media. Ho-
134
wever, in practice, existing refractive index mismatch causes some problems and the boundary conditions Eq.(2.95) can not be imposed [29, 69]. For refracting index mismatched (n0 = n1 ) along diusive/diusive interfaces the following boundary conditions have to be imposed: ( )2 n1 i + = Cn Jn i n0 (2.100) ) ( i + i Jn = D0 = D1 n i n i where the total ux at the boundary must be continuous i.e., J(r) n = J+ = 1 J = Jn 1 and where Cn =
0 1 2 R1 R0 J J 0 R1
+
i
D0 n1 D1
0
+
Jn+ Jn
Figure 2.15: Interface between diusive regions with dierent refractive indices
0 0 In case when refracting index n0 = n1 one gets R1 = 1 and R1 = 0.5 J and Cn = 0 obtaining again conditions dened by Eq.(2.95). As it is stated
135
in [69] the most common range of refractive indices in biological media is 1.3 < n < 1.5 and therefore maximum expected value for n1 > n0 is Cn 5. In such cases it is suggested to implement approximate interface conditions ( )2 n1 which are valid as long as we can consider n0 i + >> Cn Jn and this condition holds when the refractive indices are within the range mentioned above.
2.9.1
Approximate interface conditions
Let start with a simplied version of the boundary/interface conditions (see Eq.(2.100)). Then the approximate interface conditions will be expressed as follows: ( )2 n1 i i + n0 (2.101) ( ) i + i D0 = D1 n i n i Let focus our attention on the spherical region with internal interface as shown in Fig. 2.16. The system of equations describing the solution in the structure consisting of two subregions is: 1. An equation for subregion 0 bounded by surface 0 and surface + 1 is: G(0) (|r r |, ) (0) (0) (r )d(r ) = c(r) (r) + n
(2.102) G(0) (|r r |, ) (0) (r ) d(r ) n

ns s=1
Qs G(0) (|rs r|, )
where (0) consists of unknowns on the boundary 0 0 and on the (see Fig. 2.16), interface + 1 + 1
136

0
0
+ 1
+
1 1
Jn Jn
1
Jn
+ 1
R1 R0
Figure 2.16: Crosssection of spherical region with the interface.
for surface 0 Robin boundary conditions 0 mJn0 = 0 (2.103)
where m = 2A [69, 75]. The coecient A can be calculated as follows: A= 2/(1 R0 ) 1 + | cos c |3 1 | cos c |2 (2.104)
where c = arcsin(1/n) and R0 = (n 1)2 /(n + 1)2 . For the value of refractive index n = 1.333 we get A = 2.3645. 2. The second equation for subregion 1 bounded by surface 1 is: (r) + c(r) 1
G(1) (|r r |, ) (r )d(r ) = 1 n (2.105) G(1) (|r r |, ) (r ) 1 n
d(r )
and for the surface 1 we impose the approximate interface boundary
2.9. INDEX MISMATCHED DIFFUSIVE/DIFFUSIVE conditions (see Eq.(2.101)). ( 1 n1 n0 )2 1 +
137
(2.106) Jn
1
= Jn
The above two equations we may rewrite in the following matrix form: 0 0 (0) 0 0 A33 0 0 0 (0) (0) B11 B12 0 B(0) B(0) 0 22 = 21 (0) 0 0 B33 0 0 0
(0) A11 (0) A21 (0) A12 (0) A22
0 0 + 0 1 0 0 (1) A44 1 Jn 0 0 Jn + 0 1 0 Jn 0 (1) B44 Jn 1
= q1 (0) q2 (0) 0 0 (2.107)
Reducing the number of unknowns using approximate interface conditions (0) and replacing by Jn we get: n c(r) (r) + 1 = D0
(0) (0)
G(0) (|r r |, ) (0) (r )d(r ) = n

ns s=1
G (|r r |, )Jn 0 d(r )
Qs G(0) (|rs r|, ), 0 mJn0 = 0,
( c(r)
n1 n0
)2 (r) + + 1
G(1) (|r r |, ) n
n1 n0
)2 (r )d(r ) = + 1 (2.108)
1 = D1
) ( d(r ). G(1) (|r r |, ) Jn + 1
138
Now, eliminating the unknowns Jn0 the system of equations in a matrix form is as follows: (0) (0) A11 A12 0 0 (0) A21 A(0) + 0 22 1 = ( )2 (1) n1 + 0 0 A44 n0 1 (0) (0) 0 0 q1 (0) B 11 B12 (0) (0) = B (0) (2.109) B22 0 Jn + + q2 21 1 (1) 0 0 B44 Jn + 0 1 where B 11 = B11
(0) (0)
B33
(0)
)1
A33 .
(0)
Now, we have only three group, of unknowns: 0 , + and Jn + . Transfer1 1 ring unknowns to the left hand side of the system we will nally get: (0) (0) (0) A11 A12 B12 0 q1 (0) (0) (0) (0) (0) A21 A B22 (2.110) = q2 + ( )22 1 2 (1) (1) n1 Jn + 0 A44 B44 0 n0 1 Let consider two concentric spheres which crosssection is shown in Fig. 2.16. Their dimensions are: R0 = 25 mm and R0 = 22.5 mm. The amplitude and phase distribution on the perimeter of the region are presented in the Fig. 2.17.

10
1
139
1.333/1.6 1.333/1.333 1.333/1.0
complex intensity 1.333/1.6 1.333/1.333 1.333/1.
x 10
complex intensity
10
3.6308
10
3.4674
10
10
10
10
3.3113 3.1623 160
60
120
180 angle [deg.]
240
300
360
170
180 angle [deg.]
190
200
phase angle [deg.] 80 70
complex intensity 0.025 1.333/1.6 1.333/1.333 1.333/1.0
0.02
60 phase shift [deg.] 50 40 30 20 10 0 0 1.333/1.6 1.333/1.333 1.333/1.
0.015 0.01 0.005 0 0
60
120
180 angle [deg.]
240
300
360
10 angle [deg.]
15
20
Figure 2.17: Left column of the guresapproximate interface conditions ( ) n1 results for refractive index mismatch case n0 = 1.33/1.6 = 0.83, refractive ( ) 1 index match case n = 1.33/1.33 = 1.0 and refractive index mismatch n0 ( ) 1 case n = 1.33/1. = 1.33; the right column of the guresin linear scale n0 presents enlarged amplitude distribution to show that graphs on top left are close each to other especially in logarithmic scale but are not the same
( ) 1 For approximate interface conditions, when the coecient n increases n0 the phase shift decreases signicantly (see Fig. 2.17) but amplitude remains almost unchanged for all cases. That seems to be not justied from physical point of view. That is why we will implement the complete boundary conditions Eq.(2.100).
140
2.9.2
Complete interface conditions
If we consider complete or saltus [69] interface boundary conditions then 1 will read: ( )2 n1 1 = 1 + Cn Jn + (2.111) 1 n0 Now, introducing Eq.(2.111) into Eq.(2.105) we will get:
c(r) (r) + 1 = D0
(0) (0)
G(0) (|r r |, ) (0) (r )d(r ) = n

ns s=1
G (|r r |, )Jn 0 d(r )
Qs G(0) (|rs r|, ),
0 + mJn0 = 0, ( c(r) n1 n0 )2 + (r) + 1
G(1) (|r r |, ) n
n1 n0
)2 + (r )d(r ) = 1
= c(r)Cn Jn + + 1
G(1) (|r r |, ) Cn Jn + d(r ) + 1 n 1 + G(1) (|r r |, )Jn + d(r ). (2.112) 1 D1
In a matrix form, the system of integral equations (0) (0) (0) A11 A12 B12 (0) 0 (0) (0) A21 + B22 A 1 ( )22 2 (1) (1) n1 (1) Jn + 0 A44 B 44 B44 n0 1
Eq.(2.112) is: q1 (0) (0) = q2 0
(2.113)
where the new term of the system is equal: G(1) (|r r |, ) (1) + d(r ) B 44 = c(r)Cn Jn + Cn Jn + 1 1 n
141
Numerically, solving the integral equations Eq.(2.112) requires discretization of the boundary curve or the boundary surface. In case of curved boundary, the limit of validity of saltus condition expressed in Eq.(2.100), should be studied, since this expression was originally derived for a locally plane interface. In the work [69] it was proved that the error committed by using k the expression Rj ,J for a plane interface is never higher than 0.5%. The Author of [69] for practical applications has used values of the refractive index n1 between 1.0 and 3.0. Considering our example we can conclude that we have big enough curvature radius and small enough boundary elements (the segment is not bigger than 10o ), to achieve satisfactory results. Some
10 10 10 10 10 10 10 10 10
1
complex intensity 1.333/1.6 1.333/1.333 1.333/1.
10 10 10 10 10 10 10 10 10
complex intensity 1.33/1.6 1.33/1.33 1.33/1.0
60
120
180 angle [deg.]
240
300
360
60
120
180 angle [deg.]
240
300
360
phase angle [deg.] 80 70 60 phase shift [deg.] 50 40 30 20 10 0 0 phase shift [deg.] 80 70 60 50 40 30 20 10 0 0
phase angle [deg.] 1.33/1.6 1.33/1.33 1.33/1.0
1.333/1.6 1.333/1.333 1.333/1. 60 120 180 angle [deg.] 240 300 360
60
120
180 angle [deg.]
240
300
360
Figure 2.18: BEM with saltus interface conditions results for refractive index mismatch (Cn = 0.2)left column and FMC resultsright column typical values of Cn are shown in [69], where we see that in cases in which n1 < n0 , we obtain values of Cn < 2.5. The most relevant case corresponds to n0 = 1.33, since it is the most common value in biological media.
142
For calculation we have assumed that Cn = 0.2. The results are very close to the results presented in Fig. 2.17. In order to emphasis the dierences for dierent refractive indices all cases were put in one gure. For comparison, the results achieved with the aid of the Frequency Monte Carlo method were also presented (Fig. 2.18 top and bottom right). As we and can see in the Fig. 2.18 the BEM and the FMC results for the cases 11..33 6 1.33 are really close for the amplitude and the phase shift. Unfortunately the 1.33 case 11..33 demand further validation as the results seem to be not correct. 0
2.10
Numerical examples
All examples presented in this section are selected in such a way to make sure that boundary element method will provide a reliable result in impedance or optical tomography. For Electrical Impedance Tomography (EIT) and its industrial applications we have to deal with regions possessing sharp edges and corners. One of such examples is presented in Fig. 2.19 and is concern with monitoring of humidity inside the brick walls. For details consult [17]. For Diuse Optical Tomography (DOT) mostly we have to deal with medical applications where the regions possess rather smooth surfaces without sharp edges and corners. That is why we have also studied close placed surfaces of spheres as well as inhomogeneous regions with refractive index match and mismatch in case of DOT. Whenever that was possible, results were compared with analytical solution [6] or with the other numerical methods like the Finite Element [7,8] or MonteCarlo [9]. The main attention is focused on 3D cases but more interesting 2D examples were also considered (see the previous sections).
2.10. NUMERICAL EXAMPLES
143
2.10.1
Cube
Let consider the Laplaces equation in a cube region (edge length = 1.0) which was tessellated using eight subdivisions along each edge resulting in 768 triangles with 1538 vertices (see the top left pictures in Fig. 2.20). It is worth to notice that objects of such shapes must be often considered in cases of industrial tomography. The cube seems to be a trivial shape but its edges and vertices denitely would cause numerical troubles, due to the lack of continuity of the normal derivatives (see Fig. 2.19right).
e13 e14 e15 e16 e17 e18 e19 e20 e21 e22 e23 e24
e1 e2 e3 e4 e5 e6 e7 e8 e9 e10 e11 e12
Figure 2.19: An example of industrial application of EIT (left) and the cube edges and vertex with discontinuities in particular nodes (right) n
2.10.2
Two concentric spheres
Two concentric spheres can be recognized as a simplied numerical model of the CSF layer of the babys head. That is why so much attention will be devoted to this particular region conguration. We will consider the inuence of the gap length on the relative error of the solution, starting with the spherical object of standard dimensions of 25 mm simulating the babys head. This investigation we will start rst with Laplace equation for dierent discretization and dierent sizes of the internal embedded spheres.
144
10 9 8 7 6 5 4 3 2 1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
2 1.5 1 0.5 0 0.5 1 1.5 2 0
2 1.5 1 0.5 0 0.5 1 1.5 2 0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
100
200
300
400
500
600
700
800
Figure 2.20: The cube dense discretization, analytical and numerical solution (solid and dotted line respectively)top row and relative error for inside the region and for in case all nodes on the surfacebottom row n The errors of the solution on the boundary as well as inside the structure will be controlled. Next, the more dicult problem of diusion equation will be considered. Now, let start with the following structure:
Internal radius equal to 10 mm, external equal to 25 mm Region under consideration was discretised by 384 isoparametric 6 nodes triangular boundary elements. That gives 772 nodes in total. As we can observe in Fig. 2.22, solution when approaching boundary becomes less precise.
145
It is an immanent feature of BEM. In order to improve the behavior of the solution, the special treatment was introduced in this case [76]. When radius had become less than the biggest sidesize of the boundary element, than the singular integration procedure was adopted (integrals became almost singular). That will improve results, but not totally. According to [32], the boundary elements subdivision would help in such cases. The relative error distributions for the surface values are presented in the Fig. 2.23. Next, the same region was discretised by 2*768 isoparametric 6 nodes triangular boundary elements. That gives 3074 nodes in total. As we can see, the discretization is four times denser (see Fig. 2.24). This time we can observe much better results regarding internal and surface points as well (Fig. 2.25 and Fig. 2.26).
146
25 20 15
25 20 15
10 5 0 5 10 15 20 25 20
25 10 5 0 5 10 15 20 20
10 0 10 20 20 0 10 10
20
20
10
10
20
20
10
10
Figure 2.21: The outer and inner shell meshes (left) and external surface discretization (right)
10 9
1.5 2
8 7 Potential 6 5 4 3 2 1 0 10 12.5 15 17.5 Radius 20 22.5 25

1.5 2 10 Relative error 1 0.5 0 0.5 1
12.5
15
17.5 Radius
20
22.5
25
Figure 2.22: Analytical and numerical solution (solid and dotted line respectively) and relative error distribution for potential function inside the region
147
1 0.8 0.6 0.4 relative error 0.2 0 0.2 0.4 0.6 0.8 1 0 200 400 600 800 number of elements*6 1000 1200 relative error
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 0 200 400 600 800 number of elements*6 1000 1200
Figure 2.23: Relative error distribution for the d on the external sphere dn d (right) and the error distribution for the dn on the internal sphere (left)
25 20 15 10 5 0 5 10 15 20 25 20 10 0 10 20 20 0 10 10 20
25 20 15 10 5 0 5 10 15 20 25 20 10 0 10 20 20 0 10 10 20
Figure 2.24: The outer and inner (left) shell meshes and external surface discretization
148
11 10
2 1.5
9 8
Relative error 1 0.5 0 0.5 1 1.5
7 Potential 6 5 4 3 2 1 0 10 12.5 15 17.5 Radius 20 22.5 25
2 10
12.5
15
17.5 Radius
20
22.5
25
Figure 2.25: Analytical and numerical solution (solid and dotted line respectively) on the left and the relative error distribution for potential function inside the region
1 0.8 0.6 0.4 Relative error 0.2 0 0.2 0.4 0.6 0.8 1 0 200 400 600 800 Number of elements*6 1000 1200
Relative error 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 0 200 400 600 800 Number of elements*6 1000 1200
Figure 2.26: Relative error distribution for the d on the external sphere dn d (left) and relative error distribution for the dn on the internal sphere (right)
2.10. NUMERICAL EXAMPLES Internal radius equal to 20 mm, external equal to 25 mm
149
The region with internal radius equal to 20 mm and external equal to 25 mm presented in Fig. 2.27 was discretised by 2*768 isoparametric 6 nodes triangular boundary elements, that gives 3074 nodes.
25
20
25 20 15
10 15
10 5 0 5 10 15 20 25
15 10 0 5
20 10 0 10 20 20 0 10 10
25 25 20 15 10 5 0 5 10 15 20 25
20
20
Figure 2.27: The outer and inner shell meshes (left) and external surface discretization
The internal solution for this conguration is presented in the Fig. 2.28 (left) and its relative error (right).
10 9 8
1 2 1.5
7 6 5 4 3
1 0.5 0 0.5
2 1 0 20 20.5 21 21.5 22 22.5 23 23.5 24 24.5 25

1.5 2 20
21
22
23
24
25
Figure 2.28: Analytical and numerical solution (solid and dotted line respectively) on the left and relative error distribution for potential function inside the region
150
The relative error distributions for the surface values of photon intensity and current photons are presented in the Fig. 2.29. Both values are calculated with a high precision, and error does not exceed 0.8%.
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 0 100 200 300 400 500 600 700 800
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
on the external sphere Figure 2.29: Relative error distribution for the d dx d (left) and relative error distribution for the dx on the internal sphere
Zeroorder element; Internal radius equal to 24 mm, external equal to 25 mm This case is particularly absorbing as we are interested in a thin layers which can model such subregions as cerebrospinal uid or a scull of a human head. The following numerical experiments show that without a special treatment of this problem it is very dicult to achieve precise results. Region under consideration, for which the gap between surfaces was only 1 mm, was discretised by 2*2500 3 nodes triangular constant boundary elements. That gives 5000 unknowns. The results are presented in Fig. 2.30 and Fig. 2.31. To see the inuence of discretization on the results in case of at constant elements, region under consideration was discretised by 2*3456 elements. That gives 6912 unknowns. Results are shown in Fig. 2.32 and Fig. 2.33. Comparing results of calculations, it is worth to highlight that increasing

10 9
600 700
151
8
500
7 6 5 4 3
100 400 300 200
2
0
1 0 24 24.1 24.2 24.3 24.4 24.5 24.6 24.7 24.8 24.9 25

100 24 24.1 24.2 24.3 24.4 24.5 24.6 24.7 24.8 24.9 25
Figure 2.30: Analytical and numerical solution (solid and dotted line respectively) on the left and relative error distribution for potential function inside the region (right)coarse discretization
5 15 10 0 5 0 5 10 10 15 15 20 20 0 25 0
500
1000
1500
2000
2500
3000
500
1000
1500
2000
2500
3000
on the exterior surface Figure 2.31: Relative error distribution for the d dx d (left) and relative error distribution for the dx on the interior surfacecoarse discretization
discretization by 34% produces only slightly better precision, as it is shown in Fig. 2.31 and Fig. 2.33. The separate problem is the calculation of internal quantities. The standard approach without special techniques, which are beyond interest of this book, does not provide a positive results as it is presented in the Fig. 2.30 and Fig. 2.32. Particularly big errors are visible near the surfaces dening the region under consideration. The situation is even worse when the surfaces become really close each to other (for example 1 mm).
152
10 9 8 7 6 5 4 3 2 1 0 24 24.1 24.2 24.3 24.4 24.5 24.6 24.7 24.8

100 80 60 40 20 0 20 40 60 80 100 24 24.1 24.2 24.3 24.4 24.5 24.6 24.7 24.8 24.9 25
24.9
25
Figure 2.32: Analytical and numerical solution (solid and dotted line respectively) on the left and relative error distribution for potential function inside the regionne discretization
0
16 14
12
4
10
6
8 6 4
10
2
12
0
14 0 500 1000 1500 2000 2500 3000 3500
2 0
500
1000
1500
2000
2500
3000
3500
on the exterior surface Figure 2.33: Relative error distribution for the d dx (left) and relative error distribution for the d on the interior surfacene dx discretization
Secondorder element; Internal radius equal to 24 mm, external equal to 25 mm To see the dierence between zero order and second order interpolation the same region was discretised by 2*768 isoparametric 6 nodes triangular boundary elements. That gives 3074 nodes (see Fig. 2.34) much less than in case of the at triangle. Solution for the internal function and its relative error is shown in Fig. 2.35. Comparing results we can see that the potential distribution however insuciently precise is much better than for the zero order interpolation functions,
153
presented in the previous section in Fig. 2.32 and Fig. 2.35. That gives the hope for a satisfactory result if discretization is properly dense. The relative error distributions for the surface values are presented in the Fig. 2.36. Looking at the results for the at triangle and at the results for isoparametric 6 nodes triangular boundary elements surprisingly high relative error occurs for a certain nodes (see Fig. 2.36).
Internal radius equal to 24 mm, external equal to 25 mm in case of discretisation by quadrilateral elements It is interesting how the precision of thin layers calculation depends on the kind of the boundary elements. Region under consideration was discretised by 2*384 isoparametric 8 nodes quadrilateral isoparametric boundary elements. That gives 2308 nodes. Discretisation of the region is presented in Fig. 2.37. The internal eld solution and the relative error are shown in the Fig. 2.38. The relative error distributions for the surface values are presented in the Fig. 2.39. The results are still unsatisfactory, so in order to investigate the inuence of the discretization on the precision of the solution, this region was discretised by 2*1536 isoparametric 8 nodes quadrilateral boundary elements (see Fig. 2.40). That gives 9220 nodes in total. The relative error distributions for the surface values are presented in the gures 2.41 and 2.42. This time the relative error for the surface quantities dropped even below 0.3%.
154

25
20
25 20 15
10 15
10 5 0 5 10 15 20 25
15 10 0 5
20 10 0 10 20 20 0 10 10
25 25 20 15 10 5 0 5 10 15 20 25
20
20
Figure 2.34: The outer and inner shell meshes (left) and external surface discretization by 6 nodes isoparametric triangles
10 9 8 7 6 5 4 3 2 1 0 24 24.1 24.2 24.3 24.4 24.5 24.6 24.7 24.8 24.9 25
10 8 6 4 2 0 2 4 6 8 10 24 24.1 24.2 24.3 24.4 24.5 24.6 24.7 24.8 24.9 25
Figure 2.35: Analytical and numerical solution (solid and dotted line respectively) on the left and relative error distribution for potential function inside the region (right)
100
40
80
20
60
40
20
20
40
60
20
80
40 0
500
1000 1500 2000 2500 3000 3500 4000 4500 5000
100 0
500
1000 1500 2000 2500 3000 3500 4000 4500 5000
Figure 2.36: Relative error distribution for the d on the external sphere dx d (left) and relative error distribution for the dx on the internal sphere (right)

25
155
20
25 20 15
10 15
10 5 0 5 10 15 20 25
15 10 0 5
20 10 0 10 20 20 0 10 10
25 25 20 15 10 5 0 5 10 15 20 25
20
20
Figure 2.37: The outer and inner shell meshes (left) and external surface for coarse discretization
10 9 8 7 6 5 4 3 2 1 0 24 24.1 24.2 24.3 24.4 24.5 24.6 24.7 24.8 24.9 25
10 8 6 4 2 0 2 4 6 8 10 24 24.1 24.2 24.3 24.4 24.5 24.6 24.7 24.8 24.9 25
Figure 2.38: Analytical and numerical solution (solid and dotted line respectively) (left) and relative error distribution for potential function inside the region
5 4 3 2 1 0 1 2 3 4 5 0 500 1000 1500 2000 2500 3000 3500 5 4 3 2 1 0 1 2 3 4 5 0 500 1000 1500 2000 2500 3000 3500
Figure 2.39: Relative error distribution for the d on the external sphere dx (left) and relative error distribution for the d on the internal sphere dx
156

25
20
25 20 15
10 15
10 5 0 5 10 15 20 25
15 10 0 5
20 10 0 10 20 20 0 10 10
25 25 20 15 10 5 0 5 10 15 20 25
20
20
Figure 2.40: The outer and inner shell meshes (left) and external surface for dense discretization
10 9 8 7 6 5 4 3 2 1 0 24 24.1 24.2 24.3 24.4 24.5 24.6 24.7 24.8 24.9 25
50 40 30 20 10 0 10 20 30 40 50 24 24.1 24.2 24.3 24.4 24.5 24.6 24.7 24.8 24.9 25
Figure 2.41: Analytical and numerical solution (solid and dotted line respectively) on the left and relative error distribution for potential function inside the region on the right
0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5 0 2000 4000 6000 8000 10000 12000 14000 0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5 0 2000 4000 6000 8000 10000 12000 14000
Figure 2.42: Relative error distribution for the d on the external sphere dx d (left) and relative error distribution for the dx on the internal sphere (right)
157
2.10.3
The proximity eect
Because the purpose is to create the 3D BEM model of the babys head, the behavior of the method in case of diusion equation formulated for the frequency domain is very important when the surfaces become close to each other. In the previous chapters the thin layers structures were investigated in the case of Laplace equation. Now we are interested in more complicated case, due to the fact that all quantities become complex. Some of the integrals for the thin layers became nearly singular. We need to know the error which may be introduced by that kind of geometry. For this purposes we have created the structures with 1 mm, 2 mm and 3 mm distance between the surfaces in the homogeneous region, for which we know the analytical solution [6, 9]. In all three cases discretization remains the same. Comparison of the results are shown in Fig. 2.43 and Fig. 2.44.
40 30
14 12
4 10
amplitude erorr
20 10 0
8 6 4 amplitude erorr
3 amplitude erorr
Phase shift erorr 10
2 0 phase shift erorr
20 30 0
phase shift erorr
2
60 120 180 240 300 360
4 0
60
120
180
240
300
360
1 0
60
120
180
240
300
360
Figure 2.43: Thin layers solution, on the left for 1 mm gap, in the middle for 2 mm gap and on the right for 3 mm gap
Concluding this numerical experiment, we can say that magnitude is much more sensitive on the distance between the surfaces (gap) then the phase shift. The bigger gap the smaller relative error.
158
10 10 10 10 10 10 10 10 10
1
60
120
180
240
300
360
Figure 2.44: The amplitude distribution for 1mm gap (one can hardly recognize the dierences with the exact solution
2.10.4
Results for spatially nonhomogeneous region in 2D space
In this section possibility of solving forward problem, for optical tomography using boundary element approach, will be investigated. The signicance of changes introduced by the blob inside the circular region as well as the inuence of the refractive index mismatch on the quality of the signal achieved on the boundary. Comparison with the FEM solution will be made. Let consider the simple 2D example having the material coecients as indicated in Fig. 2.45. As we can see in the gures 2.46, 2.47 and 2.48 the BEM results matched very well to the FEM ones. The next question is how big dierences in signal (perturbation in amplitude or phase shift distribution along the perimeter of the region) are caused by taking into account refractive index mismatch on the interface between two subregions. Calculations were carried out for several dierent sets of a and s coecients, but nally the parameters remain the same as in Fig. 2.45. Results are not very sensitive on them.

25 20 15 10 5 0 5 10 15 20 25 25 20 15 10 5 0 5 10 15 20 25 a=0.01, s=1. point source a=0.05, s=2. * a=0.01, s=1.
159
Figure 2.45: Region under consideration
In all cases the refractive index for the background medium was n0 = 1.333 and refractive index for the blob, which could simulate the breast tumor ( 1.0 n1 3.0 [69]) was set to 1.6 and then to 3.0. For the rst case the refractive index does not introduce a signicant changes to the solution (see Fig. 2.47), when in the second case the signal increased signicantly (see Fig. 2.48). To conclude this numerical experiment, we can say that if implementing refractive index mismatch is easy (BEM), it is worth to do it, but if its not (FEM for example) we can skip this problem as the quantity of the signal is much more sensitive on geometry, not on the value of the refractive index itself.
160
10 10 10 10 10 10 10 10
0

70
FEMblobin FEMblobout BEMblobin
60 50 40 30 20 10 0 0 FEMblobin FEMblobout BEMblobin 60 120 180 240 300 360
60
120
180
240
300
360
Figure 2.46: Amplitude and the phase shift solutiondotted line for homogeneous region; solid line for the nonhomogeneous region
10 10 10 10 10 10 10 10
0
60
1.333/1.6 1.333/1.333
1.333/1.6 1.333/1.333 50
phase shift [deg.]
40
30
20
10
60
120
180 angle [deg.]
240
300
360
0 0
60
120
180 angle [deg.]
240
300
360
Figure index match case ( ) 2.47: Comparison of the solution for the ( refractive ) n1 n1 = 1.333/1.333 = 1.0 with mismatch case n0 = 1.333/1.6 = 0.83 n0
10 10 10 10 10 10 10 10
0
60
1.333/3.0 1.333/1.333
1.333/3.0 1.333/1.333 50
phase shift [deg.]
40
30
20
10
60
120
180 angle [deg.]
240
300
360
0 0
60
120
180 angle [deg.]
240
300
360
Figure index match case ( ) 2.48: Comparison of the solution for the ( refractive ) n1 n1 = 1.333/1.333 = 1.0 with mismatch case n0 = 1.333/3.0 = 0.44 n0
Chapter 3 Diusion model for light transport in the frequency domain

In recent years Optical Tomography (OT) has emerged as a highly active and viable eld of research, due to advances both in measurement technology and in theoretical and practical understanding of the nature of the image of reconstruction problem. For recent reviews see [6, 9, 19, 37, 39, 57, 83]. An increasingly active topic within this eld is the development of an ecient and accurate method for calculating the intensity of light internal to, and on the boundary of, an object under experimental investigation, sometimes referred to as the Forward Problem. Existing methods are either deterministic, based on the solutions to governing equations, or stochastic based on simulations of the individual scattering and absorption events undertaken by each photon. The former includes analytical expressions based on Green functions [7], and numerical methods based on Finite Dierence method (FDM) or Finite Element Methods (FEM) [10, 56, 61, 74, 75]. However, a generally applicable model of the forward problem in threedimensional space is still not a fully solved problem. In this chapter we introduce another standard technique for solving Partial Dierential Equations (PDE) in general geometries: the Boundary Element
161
162
CHAPTER 3. DIFFUSION MODEL FOR LIGHT TRANSPORT
Method (BEM), which has received substantial attention in numerical modelling of elds [27, 28, 32, 42, 53, 60, 81]. The advantages and disadvantages of BEM are well known [14,15,21,35] and they will not be repeated here, but instead we will concentrate on some specic features useful in OT. Recently BEM has been used in Diusing-wave spectroscopy for determining the correlation function for dierent boundary conditions and source properties in a cone-plate geometry [81], but has received very little attention in OT, although it is closely related to the integral method of Ripoll [69], which was based on the extinction theorem method. In computational mathematics this method was introduced by Kantorovich and Krylov [48], and since then has been frequently refereed to as the Method of Moments (MOM) [36, 58, 79].
3.1
Governing equations
Let consider a domain with boundary . Light transport in scattering tissue is commonly described by the diusion approximation to the transport equation [7], a second order elliptic partial dierential equation: ( D a + i/c)(r, ) = q0 (r, ) r \ (3.1)
Where stands for photon density, diusion coecient D = 3(a1 , a + s) is an absorbing and s is reduced scattering coecient, the speed of light c(r) = c0 / (r), where (r) is the refractive index and c0 is the speed of light in a vacuum, and q0 is a source of light with modulation frequency . Boundary conditions are usually taken as Robin type [75]: (r, ) = 0, r n where coecient depends on refractive index. (r, ) + 2D (3.2)
In the case where scattering and absorption are homogeneous, Eq.(3.1) re a duces to a Helmholtz equation with complex wave number k = i cD : D q0 (r, ) r\ D with the same boundary condition as those expressed by Eq.(3.2). 2 (r, ) k 2 (r, ) = (3.3)
3.1. GOVERNING EQUATIONS
163
3.1.1
Two dimensional space
For the diusion equation the fundamental solution (i.e. the Green function), for 2D space is [7]: 1 G(|r r |, ) = K0 (k |r r |) (3.4) 2 where K0 is the modied Bessel function of the second kind of zero order and |r r | represents the distance between the source point and the eld point. The normal derivative of the Green function in a direction n can be written [7, 69]: ( ) k rx ry n G = K1 (k |r r |) nx + ny (3.5) 2 |r r | |r r |
3.1.2
Three dimensional space
For the diusion equation the fundamental solution is [7]: 1 G(|r r |, ) = ek|rr | 4 |r r |
(3.6)
The normal derivative of the Green function in a direction n can be written: ( ) 1 k n G = n ek|rr | (3.7) 2 4 |r r | 4 |r r | where =
rr . |rr |
3.1.3
The Boundary Element Method
The Boundary Element Method (BEM) proceeds by applying Green theorem in its second form [14] to derive an integral equation applicable on the surface: G(|r r |, ) c(r)(r) + (r )d(r ) = n (r ) d(r ) q0 G(|r r |, )d (3.8) = G(|r r |, ) n

164
Since it is assumed that the observation point r lies on the boundary, a singularity will occur in the boundary integral if this point was included. Instead, the integral is taken only outside which represents an articial hemisphere with innitesimal radius centered at the observation point and the integral over the singularity is accounted by the extra function c(r), which can be calculated by taking each term of Eq.(3.8) in the limit as 0 [15]. However, as shown in [14, 15], the term c(r) does not need to be calculated explicitly, and can be obtained indirectly by utilizing some simple physical considerations. In particular, we have c(r) = 1/2 when the observation point lies on a smooth surface, which is the case considered in this work. In OT, concentrated sources are frequently used and are very simple to handle in BEM. They are a special case for which the function q0 at the internal point rs becomes: q0 = Qs (rs ) (3.9)
where Qs is the magnitude of the source and (rs ) is a Dirac delta function whose integral is equal to 1 over any volume containing the singularity point rs and zero elsewhere. Assuming that a number, ns of these functions exists one can write: G(|r r |, ) c(r)(r) + (r )d(r ) = n =
ns 1 (r ) d(r ) Qs G(|rs r|, ) G(|r r |, ) n s=0
(3.10)
3.2
Numerical Implementation
Some of the most important problems of numerical implementation for three dimensional problems are described in detail in [14] or [16]. In this section are given the basic derivation of a discrete implementation of Eq.(3.8). Two dierent element types will be considered with particular attention to the treatment of singular integrals.
3.2. NUMERICAL IMPLEMENTATION
165
To solve the threedimensional problem numerically, the surface is discretised into M elements each containing 1, 3 or 6 nodes. The total number of nodes is N . Our attention is focused on two types of elements: planar constant triangular elements (1 node) and curvilinear elements (6 nodes for triangular and 8 nodes for quadrilateral) which provide a better t to the modelled surfaces. For the second type of elements, both an isoparametric quadratic triangular element and an isoparametric quadratic quadrilateral element will be developed. In contrast to the FEM approach in which the derivative of the eld is expressed through derivatives of the local basis functions, in the BEM approach both the eld and its normal derivative are represented in the nite space spanned by these functions Nk . In the case of the isoparametric quadratic triangular element we will have: (r) =
6 k=0
k Nk (r) ;
(r) = n k=0
6
k n
) Nk (r) (3.11)
r) Note that this representation enforces at least C0 continuity on ( if the n boundary is assumed smooth as we do here. Extensions are possible to allow the representation of explicit jumps in the eld, but this is outside the scope of the present chapter.
Now Eq.(3.10) for the at constant triangular elements is represented discretely as: ci i +
M 1 ( j =0 M 1 j =0 j =i
j
j
G(|ri r|, ) Nj (r)dr = n

ns 1 s=0
)
j j
G(|ri r|, )Nj (r)dr
Qs G(|ri rs |, )
(3.12)
where the basis interpolation functions Nj are equal to 1 in this case. Above equation in matrix form is [14, 21]: A = B +q n (3.13)
166
where A and B are N N fully populated asymmetric matrices and and are the vectors of dimension N . For the actual construction of matrices n A and B, it is convenient to split the support of each basis function Nj into pieces on each element j and evaluate these by mapping to a local coordinate system.
3.2.1
Jacobian
To study boundary elements which are twodimensional structures placed in the 3D space, rst we need to dene the way in which we can pass from the xyz global Cartesian system to the 1 , 2 , 3 system dened over the element, where 1 , 2 are oblique coordinates and 3 is in the direction of the normal. The transformation for a given function u is related through the Eq.(2.15) from the chapter 2.4.1. Transformation of this type allows us to describe dierentials of surface in Cartesian system in terms of the curvilinear coordinates. A dierential of area will be given by the Eq.(2.16).
3.2.2
Matrix Assembly
Matrix assembly is carried out by performing integration in local coordinates on each element and then adding terms into the global system matrix for nodes which are shared by more than one element. This can be summarized in algorithmic form as shown in Algorithm 1 For elements considered in this section E is equal to 1, or 6 depending on the element and on the interpolation which was used. e denotes the area of boundary element. The diagonal terms in coecients matrix A, contrary to the Laplace equations
3.2. NUMERICAL IMPLEMENTATION
167
Algorithm 1 Construction of the BEM system matrics. for all nodes i = 1 . . . N do for all elements e = 1 . . . M do for all nodes j = 1 . . . E do n = global(j ) if (n = i) then ai,n = 1/2 (e) else ai,n + = ai,j (e) bi,n + = bi,j end for end for end for where ] [ G(|ri (1 , 2 ) rj (1 , 2 )|, ) (e) (e) (e) ai,j = Nj (1 , 2 )J (1 , 2 ) d1 d2 n e ] [ (e) (e) (e) bi,j = G(|ri (1 , 2 ) rj (1 , 2 )|, ) Nj (1 , 2 )J (1 , 2 ) d1 d2
e
case, can be determined as the sum of the nondiagonal coecients as follows:
Ai,j =
N 1 j =0 j =i
ai,j
i = 0, 1, 2, . . . , N 1
(3.14)
The other terms are handled by numerical integration rules such as Gaussian quadrature. When the source point is not close to one of the nodes of the element the quadrature method gives an accurate result. When it is close to or coincident with one of the nodes of the element, a singularity occurs in this case. Then a special treatment of the integration is required, which was already discussed in the previous sections.
168
3.3
Numerical integration of singular integrals in 3D
There are two main approaches for singular integration: integration by regularization and by subtraction and series expansion. Subtraction and series expansion method of calculating the singular integrals is described by Hall in the monograph [35] and for optics by Ripoll in [69]. The regularization method is described in [15], and is a convenient way to handle higher order elements such as the isoparametric types considered here. However, a detailed comparison of the two methods is outside the scope of this book. In the following sections the mapping procedures, which are essential for the regularizing method, will be presented. Singular points are denoted by circle surrounding a particular node.
3.3.1
Mapping formula for triangular constant element
In the constant element approach, the basis functions are taken to be piecewise constant on elements and the representing value of the eld is obtained at the central point of the element ri , which is where the singularity occurs, as shown in Fig. 3.1. For the Laplace equation analytical calculation of singular integrals was presented in chapter 2.4.3. However, in more complicated cases than the Laplace equation, it is necessary to know how to deal with numerical integration of singular integrals over the at triangular elements. In order to calculate the singular integrals, the parent element is divided into three subtriangles. Those are mapped to the local coordinate system in such a way that the singular point is always in the origin. Jacobian of this transformation is described in details in section 2.4.1. In the nal step the subtriangles are mapped to squares. The sequence of mappings is shown in Fig.3.1 and briey described below.
3.3. SINGULAR INTEGRALS IN 3D
T2
169
2 2 1 n
T1 T0 T2 T1
2
T2 T1 T0
n C
n C
0 1 z y r
0 1 r 2 z y 0 J
0
C 1 1 2 0 r z y x x J
2
T0
0 2 1
x 2 2
J 1 2 0
T0
T1
T2
1 C
2 C
J rT0
2
J rT 1 2 2 2 C
1
JrT2 0 1 C 1
1 C
Figure 3.1: Three nodes triangular element subdivision into a subtriangles centered at point C
The triangular element from the x, y, z global coordinates system is mapped to the local coordinates system 1 , 2 by:
170
For triangle T0 x = N0 (1 , 2 )xC + N1 (1 , 2 )x1 + N2 (1 , 2 )x2 = = (1 1 2 )xC + 1 x1 + 2 x2 = xC + 1 (x1 xC ) + 2 (x2 xC ), y = N0 (1 , 2 )yC + N1 (1 , 2 )y1 + N2 (1 , 2 )y2 = z = N0 (1 , 2 )zC + N1 (1 , 2 )z1 + N2 (1 , 2 )z2 = = (1 1 2 )zC + 1 z1 + 2 z2 = zC + 1 (z1 zC ) + 2 (z2 zC ). For triangle T1 x = N0 (1 , 2 )x1 + N1 (1 , 2 )x2 + N2 (1 , 2 )xC = (1 1 2 )x1 + 1 x2 + 2 xC = x1 + 1 (x2 x1 ) + 2 (xC x1 ), y = N0 (1 , 2 )y1 + N1 (1 , 2 )y2 + N2 (1 , 2 )yC z = N0 (1 , 2 )z1 + N1 (1 , 2 )z2 + N2 (1 , 2 )zC = (1 1 2 )z1 + 1 z2 + 2 zC = z1 + 1 (z2 z1 ) + 2 (zC z1 ). For triangle T2 x = N0 (1 , 2 )x2 + N1 (1 , 2 )xC + N2 (1 , 2 )x1 = (1 1 2 )x2 + 1 xC + 2 x1 = x2 + 1 (xC x2 ) + 2 (x1 x2 ), y = N0 (1 , 2 )y2 + N1 (1 , 2 )yC + N2 (1 , 2 )y1 z = N0 (1 , 2 )z2 + N1 (1 , 2 )zC + N2 (1 , 2 )z1 = (1 1 2 )z2 + 1 zC + 2 z1 = z2 + 1 (zC z2 ) + 2 (z1 z2 ). where Nk (1 , 2 ) are the shape functions: N0 (1 , 2 ) = 1 1 2 , N1 (1 , 2 ) = 1 , N2 (1 , 2 ) = 2 . The transformation from a standard triangle to a standard square is: 1 = 1 + 1 (1 + 1 )(1 + 2 ) , 2 4 2 = (1 + 1 )(1 + 2 ) . 4 (3.19) (3.18) (3.17) = (1 1 2 )y2 + 1 yC + 2 y1 = y2 + 1 (yC y2 ) + 2 (y1 y2 ), (3.16) = (1 1 2 )y1 + 1 y2 + 2 yC = y1 + 1 (y2 y1 ) + 2 (yC y1 ), (3.15) = (1 1 2 )yC + 1 y1 + 2 y2 = yC + 1 (y1 yC ) + 2 (y2 yC ),
171
The Jacobian Jr (1 , 2 ), will be associated with the change of variables which is given by: J r T0 = J r T1 = J r T2 = J r =
1 1 2 1 1 2 2 2
1 + 1 . 8
(3.20)
The regularization method introduces a transformation of a triangular domain into a square in such a way that coordinate 1 is always equal to 1 in the singular point. Than the regularizing Jacobian becomes Jr (Ti ) (1 , 2 ) = 0. In this way the integral becomes zero, precisely where the singularity occurs.
3.3.2
Isoparametric triangular quadratic element
For the isoparametric triangular quadratic element the singular integration procedure was already presented in chapter 2.6.2 for Laplace equation. In case of diusion equation the situation is more complicated, because in order to achieve the same level of accuracy, better integration procedure of singular integrands is needed. That is why a new mapping procedure will be shown. This procedure is similar to the previous one only for the rst two nodes. The rest nodes can be treated in a similar way. To gain sucient precision of calculations, each element containing a singularity is divided not into two subelements (see Fig. 2.6) but into four subtriangles T0 , T1 , T2 and T3 . Those subelements containing the singular point, for example T0 in Fig. 3.2 and subtriangles T0 , T1 and T2 in the same gure are mapped again into a square (coordinates 1 , 2 ). Sub triangles which do not contain the singular point are numerically integrated using for example the seven point Gaussian quadrature rule for triangles (see Table 2.2) [21, 88, 89]. The second transformation introduces Jacobian Jr or JrT0 , JrT1 , JrT2 (see Fig. 3.2), according to the singularity position.
172

4
2
4 5 T T 0 1 2 2 T 1 T 3 3
r 2 n
5 T 0 T2
T T
3 3 1
0
0 y
r 1 1
2 J
1
z y
0 2
x
0
2
11111111 00000000 4 11111111 00000000 00000000 11111111 1 T 00000000 11111111 2 3 00000000 11111111 5 3 T 00000000 11111111 1 2 00000000 11111111 T T 2 00 00000000 11111111 1 00000000 11111111
0 1 2 J r 2 5 1
x J
2
1111 0000 2 T 0000 1111 3 0000 1111 3 1 0000 5 1111 T 4
1 0
0 2 2 2 T 0 T1 1 2 1 2 J 2 5 1 1 T 2 3 1 1 T 1 2 rT 2 J rT 1 2 2
1
rT 0
0 0 1
1
1 T
2 0
3 1
0 5
Figure 3.2: Isoparametric triangle subdivision in case when the singular point is located in the node 0 (left) and in the node 1 (right)
Fig. 3.3 illustrates how the splitting and mapping procedures concentrate the Gaussian quadrature points around the singularity node. The mathematical expressions are given below. The isoparametric triangular element from the x, y, z global coordinates sys-

0
173
4
5 T T 0 1
2 T 1 2
Figure 3.3: The Gaussian point concentration around the singularity located at the node 1 tem is mapped to the local coordinates system 1 , 2 by: x=
5 k=0 5
Nk (1 , 2 )xk , Nk (1 , 2 )yk , Nk (1 , 2 )zk , (3.21)
y=
z=
k=0 5 k=0
where Nk (1 , 2 ) are the shape functions: N0 (1 , 2 ) = 3 (1 23 ), N2 (1 , 2 ) = 1 (1 21 ), N4 (1 , 2 ) = 2 (1 22 ), and 3 = 1 1 2 . First let consider the singularity at node 0 and than at node 1: Node 0 (Fig. 3.2) The transformation from a subtriangle T1 to a standard square is: 1 = (1 + 1 )(1 2 ) , 8 2 = (1 + 1 )(1 + 2 ) . 8 (3.23) N1 (1 , 2 ) = 41 3 , N3 (1 , 2 ) = 41 2 , N5 (1 , 2 ) = 42 3 , (3.22)
174
The Jacobian Jr (1 , 2 ), will be associated with the change of variables which is given by: Jr =
1 1 2 1 1 2 2 2
1 + 1 . 32
(3.24)
Node 1 (Fig. 3.2) For subtriangle T1 : 1 = 1 1 , 4 2 = (1 + 1 )(1 2 ) , 8 (3.25) JrT1 = (1 + 1 )/32. For subtriangle T2 : 1 = 2 (1 + 1 )(3 2 ) 1 1 + , 8 4 (1 + 1 )(1 + 2 ) = , 8 JrT2 = (1 + 1 )/32. For subtriangle T3 : 1 = 2 (1 + 1 )(1 2 ) 1 1 + , 8 4 1 + 1 = , 4 JrT3 = (1 + 1 )/32. The similar procedure can be used for the next nodes 2, 3, 4 and 5, using the transformation from a local coordinates system of the particular element 1 , 2 to subelement coordinates 1 , 2 given by: 1 =
2 i=0
(3.26)
(3.27)
Mi (1 , 2 )1 i ,
2 =
2 i=0
Mi (1 , 2 )2 i ,
(3.28)
175
where i is the local number of subtriangle node (see italic numbers inside the subelements in Fig. 3.2 or Fig. 3.2) and Mi are the new shape functions: M0 = M1 M2 (1 + 1 )(1 2 ) , 4 (1 + 1 )(1 + 2 ) = , 4 1 1 = . 2
(3.29)
3.3.3
Isoparametric quadrilateral quadratic element
The quadrilateral element from the x, y, z global coordinates is mapped to the local coordinates system 1 , 2 : x=
7 k=0 7
Nk (1 , 2 )xk , Nk (1 , 2 )yk , Nk (1 , 2 )zk , (3.30)
y=
z=
k=0 7 k=0
where Nk (1 , 2 ) are the shape functions: 1 N0 (1 , 2 ) = (1 1 )(1 2 )(1 + 1 + 2 ), 4 1 2 N1 (1 , 2 ) = + (1 1 )(1 2 ), 2 1 N2 (1 , 2 ) = (1 + 1 )(1 2 )(1 1 + 2 ), 4 1 2 ), N3 (1 , 2 ) = + (1 + 1 )(1 2 2 1 N4 (1 , 2 ) = (1 + 1 )(1 + 2 )(1 1 2 ), 4 1 2 )(1 + 2 ), N5 (1 , 2 ) = + (1 1 2 1 N6 (1 , 2 ) = (1 1 )(1 + 2 )(1 + 1 2 ), 4 1 2 ). N7 (1 , 2 ) = + (1 1 )(1 2 2
(3.31)
176
Let consider in details the case when the singular point is in the node 0 and next in the node 1. Node 0 (Fig. 2.10(a)) For triangle T1 : 1 = 1 , For triangle T2 : 1 = 1 + 1 (1 + 1 )2 , 2 2 = 1 . (3.33) 2 = 1 + 1 + (1 + 1 )2 . 2 (3.32)
For both subtriangles the Jacobian Jr (1 , 2 ), will be the same: Jr =

1 1 2 1 1 2 2 2
1 + 1 . 2
(3.34)
The subelements local coordinates are selected in the same way as it was for the triangular boundary elements. Node 1 (Fig. 2.10(b)) For triangle T1 : 1 = For triangle T2 : 1 = + For triangle T3 : 1 = (1 + 1 )2 , 2 2 = 1 . (3.37) 1 + 1 , 2 2 = 1 + 1 + (1 + 1 )2 . 2 (3.36) 1 + 1 , 2 2 = 1 + 1 (1 + 1 )2 . 2 (3.35)
For all subtriangles the Jacobians JrTi (1 , 2 ) i = 1, 2, 3, remain the same as in Eq.(3.34) and are associated with the mapping from the subtriangular to the square regions once again.
3.4. RESULTS FOR 3D
177
3.4
Results for 3D
To test the inuence of the discretization in the BEM on the accuracy of the method, we will use the homogeneous spherical region presented in Fig. 3.4. To discretize the surface, three dierent boundary elements were used: at triangle, isoparametric triangular and isoparametric quadrilateral elements. In a gray color scale the logarithm of the solution on the surface of the sphere is visualized. As a result of the solution of equation (3.1) with Robin boundary condition (3.2), the amplitude and the phase shift of the photon density on the circumference of the sphere have been plotted, together with an analytical solution taken from [7]. Calculations have been performed for the cases of a point source in a sphere 50 mm in diameter. The source was located in a distance rd = 1/s mm from the surface, and a refractive index of 1.0 was assumed, so the speed of light was c = 0.3 mm ps1 . Values for a of 0.025 mm1 and for s of 2.0 mm1 were used, which are representative for brain tissue [7]. The modulation frequency was equal to 200 MHz. In each case the equation (2.92) was solved with = 1. In order to achieve the solution of this problem, the GMRES [73] numerical solver was used. To discretize the surface, two dierent boundary elements were used: at triangular elements and isoparametric quadratic triangular ones. An important factor in mesh construction is to make the size of a distance between the nodes of elements no bigger than the diusion length (consult Eq.(1.126)): 1 D (3.38) Ld = = a k =0 The diusion length is a distance at which the photon density decreases by a factor of e, which derives directly from (3.6) (when = 0). Since this quantity is very small especially for nearinfrared (NIR) light in the wavelength range 650 to 1200 nm, the diusion equation is dicult to solve numerically.
178
For the case of at triangular elements the sphere surface was discretized by 3072 boundary elements, which gave 1538 nodes (see Fig. 3.4). For the case of isoparametric quadratic triangular elements the sphere surface had been discretized by 768 triangular boundary elements, which gave rise to the same 1538 nodes (see Fig. 3.6). Thus each quadratic element corresponds to four planar constant elements, but the interpolation in the former is closer to the sphere boundary than in the latter. Note that the number of unknowns is equal to the number of elements in the constant element case and equal to the number of nodes in the isoparametric element case. Thus the size of matrix is larger in the former case.
20 10 0 10 20 20 10 0 10 20 20 10 0 10 20
Figure 3.4: The sphere surface discretized by triangular elements
Amplitude and a phase shift distribution of against angle for the equatorial plane of the sphere for both cases are presented in Fig. 3.5, and Fig. 3.7. Conformity with the analytical solution is very good, so it is hard to distinguish both curves. In addition, the logarithm of the amplitude of the solution is visualized on the surface of the sphere in gures 3.4 and 3.6. The timings for a solution on a Pentium IV1.0 GHz is collected in table 3.1. If we compare the numerical results for both kinds of triangular elements and their performance, there is no doubt that isoparametric triangular boundary element is superior, what denitely was expected.
3.4. RESULTS FOR 3D

10 10 10 Amplitude 10 10 10 10 10
0 2 4 6 8 10 12 14
179
90 80 Phase shift (degrees) 70 60 50 40 30 20 10 0 0
60
120 180 240 Angle(degrees)
300
360
60
120 180 240 Angle (degrees)
300
360
Figure 3.5: Comparison of the photon density results for a at triangle discretization () with analytical solution (solid line) (a) Amplitude (b) Phase shift as a function of angle Table 3.1: Time performance for dierent kind of elements
type of element triangle const. triangle izopar. time (sec) 22.7 4.8 no of unknowns 3072 1538
3.4.1
Validation of numerical results
To estimate the precision of the numerical method we need some measure of the accuracy.
3.4.2
Measures of the accuracy
In [32,55] several measures of the accuracy of the calculated potential function distributions have been dened. They all compare the numerical solution against the exact one. The most widely used measure is the relative dierence measure (RDM), which is dened as follows: RDM =
S
( exact )2 dS 2 dS S exact
(3.39)
180
20 10 0 10 20 20 10 0 10 20 20 10 0 10 20
Figure 3.6: The sphere surface discretized by quadratic triangular elements
10 10 10 Amplitude 10 10 10 10 10
0 2
90 80 Phase shift (degrees) 70 60 50 40 30 20 10 0 0
4 6 8 10 12 14
60
300
360
60
300
360
Figure 3.7: Comparison of the photon density results for a isoparametric triangular quadratic discretization () with analytical solution (solid line) (a) Log Amplitude (b) Phase shift as a function of angle However, such a measure gives us the global error estimation, when in Optical Tomography we are mainly interested in local error distribution. The relative error distribution for the amplitude and the phase shift is presented in the Fig. 3.8. For the at triangular approximation we can observe large oscillations in the vicinity of the point source (see Fig. 3.8b), even though the discretization
3.4. RESULTS FOR 3D

100 80 60 40 20 0 20 40 60 80 100 0
100 80 60 40 20 0 20 40 60 80 100 0 Phase shift relative error (%)
181
Amplitude relative error (%)
60
300
360
60
300
360
Figure 3.8: Distribution of the relative error for constant triangle (dashed line) and isoparametric triangle (solid line) : (a) for the log amplitude (b) for the phase shift was quite dense, the isoparametric case gives a more accurate result.
3.4.3
Quadratic meshing
Similar results were achieved using quadrilateral boundary element described in the previous section. This time, the surface of the sphere was discretized by 1536 elements with 4610 nodes (Fig. 3.9) providing better results, because the discretization is twice as dense as for triangular mesh. This is particularly true for the phase shift, where for triangular discretization we can observe small oscillations (Fig. 3.7), disappearing for more dense discretization. Amplitude and a phase shift distribution of against the angle for all three considered cases are presented in the Fig. 3.5, Fig. 3.7 and Fig. 3.10. The analytical solution is a solid line but numerical is a line marked with triangles or diamonds for triangular or quadrilateral discretization, respectively. Coincidence with the analytical solution is very good, so it is hard to distinguish both curves. In order to validate the numerical results it were compared with the analytical solution [7], however it is unknown how precise the method is. To estimate
182
20 10 0 10 20 20 10 0 10 20 20 10 0 10 20
Figure 3.9: The sphere surface discretized by quadratic quadrilateral elements
the precision of the numerical method we need some measure of the accuracy. The relative error distribution for the amplitude and the phase shift are presented in the Fig. 3.11.
3.5
Multilayered model of the neonatal head
Finally, the nonstructural meshes provided by David Holder coworker were used [78, 86]. Next results for a four layer head model will be presented. The surfaces modeled were the outer skin, the skull, the CSF layer and the brain. Those surfaces were generated from an MRI scan following tissue segmentation. The complexity of these meshes is that a volume mesh (like this) used for a nite element method (FEM) would be very dicult to construct. The optical parameters and mesh sizes are given in table 3.2
3.5. MULTILAYERED MODEL OF THE NEONATAL HEAD

10 10 10 Amplitude 10 10 10 10 10
0 2
183
90 80 Phase shift (degrees) 70 60 50 40 30 20 10

0 60 120 180 240 Angle (degrees) 300 360
4 6 8 10 12 14
0 0
60
300
360
Figure 3.10: Comparison of the photon density results for a quadrilateral discretization (diamond) with analytical solution (solid line) (a) Amplitude (b) Phase shift as a function of angle
10 0 Phase shift relative error (%) 60 120 180 240 Angle (degrees) 300 360 Amplitude relative error (%) 30 20 10 0
10 20 30 40 50 60 0
10 20 30 40 0 60 120 180 240 Angle (degrees) 300 360
Figure 3.11: For quadrilateral discretization (a) Amplitude relative error (b) Phase shift relative error In this chapter the 3D BEM numerical model for the forward problem formulated for Optical Tomography has been presented. The standard optical tomography benchmark has been solved for the diusion equation in order to highlight the main advantages of BEM which could be useful in OT. The results were compared with those presented in [7], showing good conformity, even though the BEM mesh was relatively coarse. As the main advantages of BEM from the optical tomography applications point of view, the following may be mentioned: an easy and precise way to represent a source point (see Eq.(3.10)),
184
Table 3.2: Optical parameters and mesh discretisations for the four layer head model
a outer shell middle shell inner shell 0.0149 0.01 0.0178 s 0.8 1 1.25
nodes 2849 3294 2098
elements 1402 1646 1048
60
4 30 25 8 20 10 15 30 10 5 0 25 20 15 10 5 0 0 10 20 20 12 40 50 6
14
20
16 30 10 18
Figure 3.12: Four layers model with the blob inside
more easy to generate a good quality 3D boundary mesh than 3D volume one, adaptive mesh renement as easy as for 2D FEM problems, BEM guarantees high and a constant precision of the solution inside the whole region under consideration for both quantities: and the gradient of .
3.5. MULTILAYERED MODEL OF THE NEONATAL HEAD

4 4
185
6 30 8 25 20 15 10 5 14 0 25 20 15 10 5 0 0 10 18 20 12 10 5 0 25 16 20 15 10 5 0 0 4 10 20 25 10 20 15
30 8
10
12
14
30
30
16
18
30 8 25 20 15 12 10 5 0 25 20 15 10 5 0 0 10 18 20 10
14
30
16
Figure 3.13: One point source solution on the surfaces of the baby head
60
8 222 217 104 103 79 216 221 100 218 223 219 231 225 13 113 224 14 12
30
50 30 20 10 0 0 5 10 15 20 20 25 5 30 0
0 0 5 10
100 216 7 104 217 225 8 103 219 221 222 12 218 224 9 223 232 13 231 118 17 113 116 117 14 115 18 114 119 19 25
116 115 118 114 232 18 119 117 19 17
25
40
20
30
15
20
10
15 10
10
15
20
25
30
Figure 3.14: One point source solution on the surfaces of the baby head
186
60 40 20 0 20 40 60 50 0 50 40 20 0 20 40
60 40 20 0 20 40 60 50 0 0 50 20 40 40 20
60 40 20 0 20 40 60 50 0 50 40 20 0 20 40
60 40 20 0 20 40 60 50 0 0 50 20 40 40 20
60 40 20 0 20 40 60 50 0 50 40 20 0 20 40
60 40 20 0 20 40 60 50 0 0 50 20 40 40 20
60 40 20 0 20 40 60 50 0 50 40 20 0 20 40
60 40 20 0 20 40 60 50 0 0 50 20 40 40 20
Figure 3.15: Four layer baby head model without a void gap. Left amplitude, right phase
Chapter 4 Light propagation in diusive media with nonscattering regions
4.1
Introduction
The main purpose of this chapter is to develop an ecient method for mathematical modelling the babys head with clear layers. Previously such a model was created using almost exclusively the FEM, see for example [8, 1113, 41, 6771]. In those 2D or 3D models the region under consideration was treated as spatially homogeneous, in order to avoid too many nite elements. So, why do not apply the BEM, which in this particular case provide almost the same ability, regarding discretization, and has advantages over FEM, as mentioned before? Fig. 4.1 illustrate two types of void and the hollow case and the thin layer will be discussed in this chapter. These have been chosen as they are analogous to anatomical features one might expect to see in brain, the hollow regions to the ventricles and the thin layer being analogous to the SubArachnoid Space (SAS), as illustrated in Fig. 4.2. 187
188
CHAPTER 4. MEDIA WITH NONSCATTERING REGIONS

111111111111111111111111111 000000000000000000000000000 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 1 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 2 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 3 3 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 2 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000 111111111111111 000000000000000000000000000 111111111111111111111111111 1 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111
111111111111111111111111111 000000000000000000000000000 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 1 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 2 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 2 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 1 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111 000000000000000000000000000 111111111111111111111111111
Figure 4.1: A diagram illustrating two types of void inclusion a hollow region (left) and a thin layer region (right)
Figure 4.2: A diagram illustrating the biological analogies of the void types
4.2
Governing equations for nonscattering sphere embedded in a diusive spherical region
Let consider a spherical domain with boundary . For the diusion equation the fundamental solution (i.e. the Green function), for 3D space is [7]: G(|r r |, ) = 1 ek|rr | 4 |r r | (4.1)
4.2. NONSCATTERING SPHERE

y
111111111111111111111111111111 000000000000000000000000000000 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 Diffuse 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 1 000000000000000000000000000000 111111111111111111111111111111 r 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 1 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 2 Non 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 r Scattering 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 2 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 0 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 2 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 1 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111
189
Figure 4.3: The crosssection of the region under consideration with a non scattering subregion
where |r r | represents the distance between the source point and the eld (always outside of the region point, and its normal derivative in a direction n under consideration) can be written [7, 69]: ( n G = n )
1 k 2 4 |r r | 4 |r r |
ek|rr |
(4.2)
where =
rr . |rr |
4.2.1
The Boundary Element Method
The Boundary Element Method (BEM) proceeds by applying Green theorem in its second form [14] to derive an integral equation applicable on the surface
190
CHAPTER 4. MEDIA WITH NONSCATTERING REGIONS =

= 1 2 : c(r)(r) +
G(|r r |) (r )d(r ) = n Jn ( r ) G(|r r |) d(r ) D

(4.3) q0 G(|r r |)d
r 1 2 , r 1 2
(r ) r) where: JnD = ( and c(r) = 1/2 when the observation point lies on n a smooth surface, which is the case considered in this book. The Green function, photon density function and photons current function Jn are depended on , but in order to simplify the equation, this symbol was skipped.
It is worth to notice that the vector Jn has an opposite direction to the outward normal unit vector n and to the vector . It means that it is n directed inside the region. Boundary conditions on the boundary 1 are usually referred to the zero ux or partial ux boundary conditions [69], and by means of Fick law, J(r) = D can be rewritten as: 1 (r) = 1 (r) Jn 1 = 0, r 1 (4.4) n where coecient = (2 RJ )/RU depends on a refractive index mismatch. We have assumed that RJ = 1 and RU = 1/2 so = 2 in case of a refractive index matched [69]. The variables i and Jn i indicates to which part of boundary they belong. 1 (r) + D On the other part of the boundary = 2 the nonlocal boundary condition between the diusive and nonscattering media is imposed [69, 71]: 2 (r) = Jn 2 (r) + ] [ RJ 1 2 (r ) + Jn (r ) G (|r r |)d2 (r ) r 2 (4.5) + RU 2
2
where G is the radiocity kernel representing diuse-diuse propagation of light in free space which will be explained in details later.
4.2. NONSCATTERING SPHERE
191
We can rewrite previous equation in a standard BEM way as follows: 1 1 2 (r) G (|r r |)2 (r )d2 (r ) = 2 2 2 1 = Jn 2 (r) + G (|r r |)2Jn 2 (r )d2 (r ) r 2 (4.6) 2
2
where: 1 1 G (|r r |) = (|r r |)V (r r ) cos() 2 2 In the above equation V (r r ) is a Boolean visibility function: { 1 if r and r are mutually visible, V (r r ) = 0 otherwise. (4.7)
(4.8)
In the considered case (see Fig. 4.3) V (r r ) is equal to unity so can be skipped, and is expressed as follows: e[a0 +i(n0 /c)]|rr | (|r r |) = cos( ) 2 |r r |
(4.9)
where: cos() = n (r r ) |r r | (4.10) cos( ) = n (r r) |r r |
where n is the unit outward normal vector as indicated in the Fig. 4.4. The component section 4.3).
cos() cos( ) |rr |2
is often reered to as the Form Factor (see
For concentrated sources for which the function q0 at the internal point rs becomes: q0 = Qs (rs ) (4.11)
192

n
r y
rr x n
0 r
Figure 4.4: Angles illustration for 2D space
where Qs is the magnitude of the point source and (rs ) is a Dirac delta function whose integral is equal to 1 over any volume containing the singularity point rs and zero elsewhere. Assuming that a number ns of these functions exists, the Eq.(4.3) , Eq.(4.4) and Eq.(4.6) are making the system of equations describing our problem: 1 (r) + 2 i =
G(|r r |) i (r )d(r ) = n
n s 1 Jn i (r ) d(r ) G(|r rs |)Qs D s=0
G(|r r |)
(4.12)
r 1 2 , r 1 2 , rs 1
1 (r) = Jn 1 (r) r 1 2 1
(4.13)
4.2. NONSCATTERING SPHERE and 1 1 2 (r) G (|r r |)2 (r )d2 (r ) = 2 2 2 1 = Jn 2 (r) + G (|r r |)2Jn 2 (r )d2 (r ) r 2 2
2
193
(4.14)
where i (r) = (r)|i , Jn i (r) = Jn (r)|i and i = 1, 2. If we assume that sphere 1 and sphere 2 have the same number of N unknowns then the total number of unknowns is equal to 4N .
4.2.2
Matrix form of integral equations
Equations Eq.(4.12) to Eq.(4.14) can be presented in a matrix form: A11 A12 A21 A22 A31 0 0 A42 [ ] = B11 B12 B21 B22 I 0 0 B42 [ ] + q1 q2 0 0 (4.15)
1 2
Jn 1 Jn 2
By including the Robin boundary conditions (Eq.(4.13)) to Eq.(4.12), we can reduce the number of unknowns to the values 1 (r), 2 (r) and Jn 2 (r) only.
] ] q1 B11 A31 B12 [ A11 A12 [ 1 1 + q2 = B21 A31 B22 A21 A22 Jn 2 2 0 0 B42 0 A42
(4.16)
If the rst column of the righthand matrix is transferred to the lefthand side, the unknown vector Jn 2 would be included to the left hand side 1 and 2 unknowns vectors. Finally we come to the form:
194
A11 B11 A31 A12 B12 1 q1 A21 B21 A31 A22 B22 2 = q2 0 A42 B42 Jn 2 0
(4.17)
4.3
The Form Factor
For the Form Factor enclosed in Eq.(4.7) we will now develop analytic expressions and numerical approach for the case presented in Fig. 4.3. 1 e[a0 +i(n0 /c)]|rr | cos() cos( ) G (|r r |) = = 2 2 |r r |2 e[a0 +i(n0 /c)]|rr | = g (u, v ) 2
(4.18)
where g (u, v ) is the standard Form Factor expression; u and v are the standard parametric representation for a sphere (u < 0, > and v < 0, 2 >) [66]:
g (u, v ) =
cos() cos( ) . |r r |2
(4.19)
4.3.1
The Form Factor calculated analytically
The Form Factor (4.19) for a sphere of radius r2 (see Fig. 4.3 and consult Fig. 4.4) can be easily calculated analytically: g (u, v ) = 1 cos() cos( ) = . 2 2 |r r | 4r2 (4.20)
The above relation signicantly simplies calculation due to the fact that integrals in Eq.(4.14) are not singular any more.
4.3. THE FORM FACTOR
195
4.3.2
The Form Factor calculated numerically
In case of arbitrary shaped surfaces there is a need to calculate the form factor numerically. That means (see Eq.(4.19)) that not only both cosine functions have to be calculated numerically (Eq.(4.10), but also a visibility function. cos() = n (r r ) nx |r r |x + ny |r r |y + nz |r r |z = |r r | |r r | (4.21) cos( ) = n (r r) = |r r |
nx |r
r|x +
ny |r r|y |r r|
nz |r
r|z
For an isotropic second order triangle the normal vector components in BEM can be expressed in a following way y z z y 1 2 1 2 z x x z ny = 1 2 1 2 x y y x nz = 1 2 1 2
nx =
(4.22)
(4.23) where x, y, z are the coordinates of the points r or r and i dene the local curvilinear coordinate system. Regarding the numerical calculation of the visibility function we have already tried two basic approaches. The rst one when the visibility function (basically the continuous function) was calculated in each Gaussian numerical integration point and the second one where was calculated only in the nodes of the boundary elements. Both approaches producing almost identical results. However, the rst approach seems to be more ecient numerically, so only this approach will be exploited in the following chapters.
196

4 5 3 n 1 0 2 outer surface rr r 0 r z y x 4 inner surface 5 n 3 1 2
Figure 4.5: A diagram illustrating the way of the Form Factor numerical calculations (doted lines denote the node to Gaussian node visibility matrix calculation)
4.4
Results
BEM results for the clear sphere embedded in a diusive spherical region were compared with the MC method [68, 70]. Results for the steady state and for the frequency of 100MHz are presented below.
4.4.1
The steady state
In the begining let consider the most simple case: namely, the steady state, but in the following subsections, for the frequency domain case, both the amplitude and the phase data will be presented.
4.4. RESULTS
197
In the Fig. 4.6 there are presented results of photon density amplitude distribution along the circumference on the outer and inner spheres calculated with the aid of BEM. In order to see the inuence of the nonscattering sub-
10 10 10 10 10 10 10 10
inner surface r[2]
outer surface r[1] FEM r[1]
diffusive r[1]
8
60
120
180
240
300
360
Figure 4.6: Amplitude distribution for a region with a nonscattering sub region
region the thin solid line represents the case when this subregion is absent. It means that only diusive region was considered. The great inuence of nonscattering subregion is clearly visible. Additionally, BEM results are compared with the FEM one borrowed from [66]. The idea here had been compared to the BEM and the FEM models with each other. As the FEM model is much more matured [8, 67], it has been using as a benchmark. We can see some discrepancy between BEM and FEM results (see Fig. 4.6). One possibility is the fact that the BEM type boundary conditions are slightly dierent from the FEM model (consult [69,71] and [8]). One more possibility is that in fact the BEM model (or rather its implementation) is not working correctly. It needs more attention so the chapters 4.7, 4.8 and 4.9 will be devoted to this problem.
198
4.4.2
The frequency domain solution 100MHz
10 10 10 10 10 10 10 10
40 35 30 25 diffusive r[1] 100MHz
inner surface r[2] 100MHz

20
outer surface r[1] 100MHz FEM r[1] 0MHZ
15 10
outer surface r[1] 100MHz
inner surface r[2] 100MHz 5
diffusive r[1] 100MHz

8
60
120
180
240
300
360
0 0
60
120
180
240
300
360
Figure 4.7: Amplitude and a phase shift distribution for a region with a nonscattering subregion
4.5
Nonscattering gap between two diusive regions of a spherical shape
Let study the threedimensional conguration depicted in Fig. 4.8, where the outer sphere of xed radius r1 = 25 mm, lled with a diusive medium of parameters s = 1 mm1 and a1 = 0.01 mm1 with a nonscattering gap of outer radius r2 = 20 mm and inner radius r3 = 17 mm between two diusive regions. The absorption coecient for the nonscattering region is a 0 = 0.005 mm1 . In this case the visibility factor V (r r ) is either unity if both points r and r can be joined by a straight line without intersecting an interface (i.e., when they are visible to each other) or zero when such a straight line does not exist. For the case of concentric spheres it may be calculated analytically.
4.5. NONSCATTERING GAP

y
111111111111111111111111111111 000000000000000000000000000000 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 1 Diffuse 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 r 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 1 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 Non Scattering 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 2 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 3 Diffuse r 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 2 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 0 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 r 000000000000000000000000000000 111111111111111111111111111111 3 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 3 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 2 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 1 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111
199
Figure 4.8: The crosssection of the region with a nonscattering gap
4.5.1
Form Factor calculated analytically
We now attempt to derive the analytical Form Factor for more interesting case then the one presented in the previous section, namely that of two concentric spheres, outer sphere radius r2 and inner sphere radius r3 [68]. Considering any point on the outer sphere we have two cases:
1. the visible point lies on the inner sphere,
2. the visible point lies on the outer sphere.
In case when the visible point is laying on the inner surface the analytical expression for the Form Factor becames (see Fig. 4.9): (r2 r3 cos U )(r3 r2 cos U ) cos() cos( ) = , 2 2 2 |r r | (r2 + r3 2r2 r3 cos U )2
g (u, v ) =
(4.24)
200

y r U 0 r 2 r 3 2 3 x r
Figure 4.9: Analytical calculation of the Form Factor for the case number 1
where r2 r3 cos U cos = 2 , 2 r2 + r3 2r2 r3 cos U (4.25) , cos = 2 2 r2 + r3 2r2 r3 cos U and cos U = r r . |r||r | (4.26) r3 r2 cos U
But for the second case, when the visible point is on the outer surface the expression remains the same as in Eq.(4.20). The visibility matrix image is shown in Fig. 4.10.
4.5.2
Visibility function calculated analytically
If we are dealing with a regular surfaces as spheres, then the visibility factor (ref. Eq.(4.7)) may be calculated analytically. Clearly, for a hollow sphere

700 700
201
600
600
500
500
400
400
300
300
200
200
100
100
100
200
300
400
500
600
700
100
200
300
400
500
600
700
Figure 4.10: Visibility matrix image calculated analytically (left) and a dierence between the matrix calculated analytically and calculated numerically (right)
the visibility function is equal to 1 because all points (nodes of boundary elements) can see all other points. But for the gap between a concentric spheres (see Fig. 4.8) it is not that simple. We have to consider three cases: 1. both points r and r are on the surface 2 (see Fig. 4.11left) are visible if vis < 2Uco . Angle Uco is the visibility cuto angle for this case, 2. one of the points is on the surface 2 and the other on the surface 3 (see Fig. 4.11right) are visible if vis < Uco , 3. both points r and r are on the surface 3 are not visible. It is now possible to calculate the visibility cutof ( Uco |2 and Uco |3 ) for concentric spheres geometry. From geometry (see Fig. 4.8) we have that r3 Uco |3 = 1 Uco |2 = arccos( r ), giving us a simple expression for the point to 2 2 point visibility for all points on the domain boundary in this case. The angle between two arbitrarily chosen points r and r is equal vis = arccos r r . |r||r | (4.27)
202

y Non Scattering 2 Diffuse 3 0 vis vis 3 2 r 2 y Non Scattering Diffuse 3 0 vis r 3 2 3 r
r 2 r 3
r 2 r
Figure 4.11: The visibility angle between two points on the 2 surface (left) and between two points on 1 and 2 surface (right)
4.5.3
Visibility function calculated numerically
Visibility function calculated numerically leads to the following basic problem. We have to nd the intersection point (if it exists) between a line segment dened by two points r and r and a planar three vertex facet. Each boundary element is subdivided by four at facets in order to simplify the procedure (see Fig. 4.12). At the beginning we will adopt the brutal force
4 5 3 1 0 2
000000000000 111111111111 4 111111111111 000000000000 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 5 000000000000 111111111111 000000000000 111111111111 00000000 11111111 000000000000 111111111111 000000000 111111111 00000000 11111111 000000000000 111111111111 000000000 111111111 00000000 11111111 000000000000 111111111111 000000000 111111111 00000000 11111111 000000000000 111111111111 000000000 111111111 00000000 11111111 000000000000 111111111111 000000000 111111111 00000000 11111111 000000000000 111111111111 000000000 111111111 00000000 11111111 000000000 111111111 000000000 111111111 00000000 11111111 000000000 111111111 000000000 111111111 00000000 3 11111111 000000000 111111111 000000000 111111111 00000000 11111111 000000000 111111111 000000000 111111111 00000000 11111111 000000000 111111111 000000000 111111111 00000000 11111111 000000000 111111111 000000000 111111111 00000000 11111111 000000000 111111111 000000000 111111111 00000000 11111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 1 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 000000000 111111111 0
Figure 4.12: The boundary element subdivided by four at subtriangles
method which rely on checking if every subtriangle has an intersection point with the line segment dened by two points mentioned before.
4.5. NONSCATTERING GAP The solution involves the following steps: 1. Check that the line and the plane are not parallel.
203
2. Find the intersection of the line, on which the given line segment lies, with the plane containing the facet. 3. Check that the intersection point lies along the line segment. 4. Check that the intersection point lies within the subtriangle. The intersection point ri is found by substituting the equation for the line ri = r u(r r ) by the general form of equation for the plane Ax + By + Cz + D = 0. Note that the values of A, B , C are the components of the normal to the plane which can be found by taking the cross product of any two normalized edge vectors and then D is found by substituting one vertex into the equation for the plane. A = z1 (y3 y2 ) + z2 (y1 y3 ) + z3 (y2 y1 ), B = z1 (x2 x3 ) + z2 (x3 x1 ) + z3 (x1 x2 ), (4.28) C = x1 (y2 y3 ) + x2 (y3 y1 ) + x3 (y1 y2 ), D = x3 A y3 B z3 C, where three vertices r1 , r2 and r3 determine the plane. This gives an expression for u from which the point of intersection can be found. u= Ax + By + Cz + D . A(x x ) + B (y y ) + C (z z ) (4.29)
If the denominator above is equal to 0 then the line is parallel to the plane and they do not intersect. For the intersection point to lie on the line segment, u must be between 0 and 1. Those conditions allow us an early return.
204
And last but not least, we need to check whether or not the intersection point lies within the triangle. If the point of intersection is inside of the boundary element than the point r and point r are not visible.
4.5.4
Point in triangle test
Boundary elements mesh usually is quite dense so the crucial point of a test is its eciency. We will constrain ourself to a triangular elements only due to the fact that they are the most popular and every quadrilateral boundary element can be split into two triangular elements. A common way to check whether a point is in triangle, is to nd the vectors connecting the point to each of the triangles three vertices and sum the angles between those vectors. If sum of the angles is 2 then the point is inside the triangle, otherwise it is not. It works, but it is very slow. We need a faster and much easier method [84]. Set of the points, let them call A, B and C forms the triangle and lines AB , BC and CA each split space in half and one of those halves is entirely outside the triangle. That is what we will take advantage of. For a point to be inside the triangle ABC it must be below AB and left of BC and right of AC . If any one of these conditions fails we can return early. In order to test whether the point is on the correct side of the line we have to take the cross product of AB and AP (see Fig. 4.13), we will get a vector pointing out of the plane. On the other hand, if we take the cross product of AB and AP , we will get a vector pointing into the plane. The question is what direction the cross product should point in? Because the triangle can be oriented in any way in the 3D space, we need some reference

B
000000000000000000000000 111111111111111111111111 111111111111111111111111 000000000000000000000000 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 P 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 P 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 A 111111111111111111111111 000000000000000000000000 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111 000000000000000000000000 111111111111111111111111
205
Figure 4.13: Illustration of the point in triangle test
point a point that we know is on a certain side of the line. In our case is just the third point C . So, any point P where AB cross AP does not point in the same direction as AB cross AC is not inside the triangle. If the cross products do point in the same direction, than we need to test P with the other lines as well. If the point was on the same side of AB as C and is on the same side of BC as A and on the same side of CA as B , then it is in the triangle. We can implement this algorithm in the following way: Algorithm 2 Test Point in Triangle. function SameSide(p1,p2, a,b) cp1 = CrossProduct(ab, ap1) cp2 = CrossProduct(ab, ap2) if DotProduct(cp1, cp2) >= 0 then return true else return false function PointTriangle(p, a,b,c) if SameSide(p,a, b,c) && SameSide(p,b, a,c) && SameSied(p,c, a,b) then return true else return false
This algorithm is simple, eective and has no square roots and the other
206
functions like arc cosines etc.
4.5.5
Integral equations
In case of three surfaces immersed one in the other, as it is presented in Fig. 4.8, the system of equations will be more complicated:
1 (r) + 2 i =
G(|r r |) i (r )d(r ) = n
n s 1 Jn i (r ) d(r ) G(|r rs |)Qs G(|r r |) D s=0
(4.30)
r 1 2 , r 1 2 , rs 1
1 (r) = Jn 1 (r) r 1 2 1 1 1 V (r) G (|r r |)V (r )dV (r ) = 2 2 V 1 = Jn V (r) + G (|r r |)2Jn V (r )dV (r ) 2
V
(4.31)
r r V = 2 3 and 1 G(|r r |) 2 (r) + 2 (r )d3 (r ) = 2 n 3 Jn ( r ) d3 (r ) r r 3 . = G(|r r |) 2 D

3
(4.32)
(4.33)
207
4.5.6
Matrix form of integral equations
Equations Eq.(4.30) to the Eq.(4.33) can be presented in a matrix form: A11 A21 A31 0 0 0 A12 0 A22 0 0 0 A42 A43 A52 A53 0 A63 B11 B21 I 0 0 0 B12 0 B22 0 0 0 B42 B43 B52 B53 0 B63 q1 q2 0 0 0 0 (4.34)
1 = 2 3
Jn 1 Jn + 2 Jn 3
By including the Robin boundary conditions from Eq.(4.31) to the Eq.(4.30) we can reduce the number of unknowns to the values 1 (r), 2 (r) , 3 (r), Jn 2 (r) and Jn 3 (r) only. A11 A21 0 0 0 A12 0 A22 0 A42 A43 A52 A53 0 A63 1 = 2 3 B11 A31 B21 A31 0 0 0 B12 0 B22 0 B42 B43 B52 B53 0 B63 1 Jn + 2 Jn 3 q1 q2 0 0 0 (4.35) If the rst column of the righthand matrix will be transferred to the left hand side, then we have to rearrange the system of equations in such a way that unknown values should be on the lefthand side and the source terms on the righthand side. A11 B11 A31 A21 B21 A31 0 0 0 A12 A22 A42 A52 0 B12 0 0 B22 0 0 B42 A43 B43 B52 A53 B53 0 A63 B63 1 2 Jn 2 3 Jn 3 q1 q2 0 0 0
(4.36)
208
4.6
4.6.1
Results for the void gap

The steady state
In the Fig. 4.14 there are outcomes of comparison between the MC and BEM results. In case of 5 mm gap the visibility cuto angle is equal to 41.4o and for 3 mm gap the visibility cuto angle is equal to 31.7o .
10 10 10 10 10 10 10 10 10
10 10 10 10
2 3
3 4
10 10 10 10
60
120
180
240
300
360
60
120
180
240
300
360
Figure 4.14: Amplitude distribution for a region with a nonscattering 5mm gap (left) and 3mm gap (right), comparison between MC (thick line) and BEM (dotted line)
4.6.2
The frequency domain solution 100MHz
In order to validate the steady state BEM solution, the two other solutions of well known methods like Monte Carlo (MC) and FEM are presented in the same gure (see 4.14). Inspecting all the gures of this section we can conclude that discrepancies between dierent methods still need more attention and investigation.
6 7
n 7
5 r 2
r 1 4
n n 7 6 r 2 3 5 r 1 4
4.7. 0
z y
r 4 r 2 1 3 CONSISTENCY CHECKS 1 2 r z J y x 2 0 1 r
FOR BEM SOLUTIONS 3

2 J
30 25 20
6 5
209
40
0 y r
1 2 J x
z 35
x
5 4 111111111111 000000000000 000000000000 111111111111 000000000000 111111111111 2 000000000000 111111111111 3 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 7 111111111111 3 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 000000000000 111111111111 1 000000000000 111111111111 6
2 2 1 2
1 J rT 2
2
J rT1AmplitudeJ and 1 Figure a J 4.15: rT3 1 rT

2 2 2 2 1 1
rT phaseJ rT shift distribution for a region with a J rT J rT 0 0 nonscattering 5mm 6 gap 100MHz, the 2thick line 0is a steady state4 MC, just 2 2 to see between the curves; the5 solid line region only 4 4 5 is a diusive 5 3 a dierences 3
2 1 1 2 2 2 3 2 1 1 1 1 1 1
2 4 6 5 1111111111111 0000000000000 0000000000000 1111111111111 0000000000000 1111111111111 2 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 3 7 1 0000000000000 1111111111111 0000000000000 1111111111111 0000000000000 1111111111111 1 0000000000000 1111111111111 0000000000000 1111111111111
2 2
15 10 5 0 0
1
1111111111111 0000000 000000 4 0000000 1111111 000000 111111 2 1 0000000 1111111 000000 0000000111111 1111111 000000 111111 0000000 1111111 000000 0000000111111 1111111 000000 1 111111 0000000 1111111 000000 111111 7 3 0000000 1111111 000000 111111 0000000 1111111 000000 3 111111 0000000 1111111 000000 111111 60 120 180 240 0000000 1111111 000000 111111 0000000111111 1111111 000000
2 2 2 1 1
300
360
10 10 10 10 10 10 10
40 35 30 25 Diffusive BEM FEM MCF
20
5
15
6
MCF
& FEM
10 5 0 0
BEM DIFFUSIVE 120 180
60
240
300
360
60
120
180
240
300
360
Figure 4.16: Amplitude and a phase shift distribution for a region with a nonscattering 3mm gap 100MHz, the solid line is a diusive region only
4.7
Consistency checks for BEM solutions
Additional validation of the BEMvoid software except for comparison with FMC method and FEM, in case of spherical regions, would be a consistency check described below.
210
10 10 10 10 10 10 10 10
1

40 35 30 FEM 25
DIFFUSIVE
20
5
MCF
BEM
BEM
6
15 10 5
FEM MCF DIFFUSIVE 60 120 180 240 300 360
0 0
60
120
180
240
300
360
10 10 10 10 10 10 10 10
1
200 150 100 50 0
50
6
100 150 200 0
60
120
180
240
300
360
60
120
180
240
300
360
4.7.1
Diusion equation case
Consider the Diusion Partial Dierential Equation (PDE) in a frequency domain. There are two commonly used possibilities
1. Inhomogeneous PDE, homogeneous boundary conditions: i (r; ) D(r)(r; ) + a (r)(r; ) = Q0 (r; ) c (m; ) 2Jn (m; ) = 0 m r (4.37)
(4.38)
4.7. CONSISTENCY CHECKS FOR BEM SOLUTIONS For convenience of BEM formulation, Eq.(4.37) becomes: i a (r) (r; ) + (r; ) (r; ) = cD(r) D(r) Q0 (r; ) = r D(r)
211
(4.39)
2. Homogeneous PDE, inhomogeneous boundary conditions: i (r; ) D(r)(r; ) + a (r)(r; ) = 0 c (m; ) 2Jn (m; ) = (m; ) In both cases, we assume Fick law: J = D, and so dene: Jn (m) = D(m) (m; ) (4.43) (4.42) r (4.40) m (4.41)
In a BEM solution we impose the boundary condition (4.38) or (4.41) only on the outer boundary 1 and obtain two variables on each internal boundary. We can derive some simple consistency checks on the solution. Consider the case = 0, a = 0; = 1 (DC, lossless, refractive index matched). In this case, we have: Jn (m)d 1 (m) = D(r)(r)d(r) (4.44)
1
In case 1, the right hand side of (4.44) is 1, so we have: 1 (m)d 1 (m) = 1 Jn (m)d 1 (m) = 2 1 1 In case 2, the right hand side of (4.44) is 0, so we have: 1 Jn (m)d 1 (m) = ((m) (m))d 1 (m) = 0 2 1 1
(4.45)
(4.46)
212
on every other interface k we have: Jn (m)d k (m) = 0

k
(4.47)
The check sums are easy to do in BEM. Since we have: Jni i (m) Jn (m)d k (m) =
k i
(4.48)
where i is the shape function at i th node.
4.7.2
RadiosityDiusion equation case
In the case where there is a void we use the RadiosityDiusion Model. In this case, inside the void, we replace the diusion equation by the transport equation in the absence of scattering ( ) i s (r, s; ) + a (r) + (r, s; ) = 0 (4.49) c where s means direction of power ux ow at a point r. We also have to note that photon current is equal J( r; ) = s (r, s; )dS (r)
S N 1
(4.50)
where N is the dimension of the space. Thus we have (DC, lossless case) Jn (m)d V oid (m) = (s (r, s)) dV oid (r) = V oid V oid = s (r, s)dV oid (r) = 0 (4.51)
V oid
Physically, this condition is just the same as Eq.(4.47): all energy entering a domain, must exit the domain (because there are no sink and losses). Two cases are considered. First, for the nonscattering sphere embedded in a diusive spherical region (see Fig. 4.3) and the second one for a non scattering gap between two diusive regions of a spherical shape (see Fig. 4.8). Results of the consistency check are presented in Tab. 4.1 and in Tab. 4.2 respectively.
4.8. NUUTTIS 2D TEST EXAMPLE
213
outcoming flux
20 15 4 10 5 0 5 10 2 15 20 20 10 0 10 10 20 20 1.5 1 3.5 3 2.5 4.5
incoming flux
10 0
20
0.5 0
Figure 4.19: Incoming and outcoming current distribution
Table 4.1: Consistency test for dierent discretizations

surfaces outer sphere inner sphere Number of elements 192 768 0.9698885 0.0212911 0.991551 0.00525638
4.8
Nuuttis 2D test example
In order to check how both methods approximate the nonlocal boundary conditions (4.32), let consider the simple 2D example having the analytical solution [41]. The structure of the region is similar to the cross section of the 3D domain under consideration (Fig. 4.8). However, to simplify the calculations the whole region was normalized to the following values: r1 = 1.0, r2 = 0.8 and r3 = 0.5. In the strongly scattering regions the following values were assumed: a = 0.5 and s = 50 and in nonscattering region a = 0.25. This kind of coecient values is reecting a neonatal brain
214
CHAPTER 4. MEDIA WITH NONSCATTERING REGIONS Table 4.2: Consistency test for dierent discretizations
surfaces outer sphere inner sphere Number of elements 192 768 0.994072 -0.0292366 0.992358 0.00444954
of diameter 100 mm [41]. For the sake of simplicity the unit input ux on the outermost boundary was imposed and a steady state was considered.
4.8.1
Analytical solution for diusive boundary conditions
Let consider 2D region shown in Fig. 4.20. Our plan is to rst solve the
y
111111111111111111111111111111 000000000000000000000000000000 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 1 Diffuse 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 r 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 1 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 Non Scattering 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 2 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 3 Diffuse r 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 2 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 0 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 r 000000000000000000000000000000 111111111111111111111111111111 3 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 3 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 2 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 1 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111 000000000000000000000000000000 111111111111111111111111111111
Figure 4.20: 2D region with a nonscattering gap
problem in the polar coordinates thus, we x the origin to be the center of the circular region and write diusion equation in polar coordinates: ( ) ( ) 1 1 2 D a = 0. (4.52) r + 2 r r r r 2
215
If we assume that incoming ux is constant then the boundary conditions on the outermost boundary 1 are equal: ( ) D = 1 (4.53) + 2 n 1 where D =
1 . 2(a +s )
Next, we do the standard trick of separating variables. Namely, we write (r, ) = R(r)() and deduce easily that the functions R and must satisfy the pair of equations: () = C() ( ) r2 R (r) + rR (r) 2 r2 + C R(r) = 0
a where C is some real constant and 2 = . By using the fact that function D must be periodic, we straight away obtain:
() = ein and, hence, C = n2 . Since neither the input ux nor the geometry of depends on the variable , it seems reasonable to assume that the only case that we are interested in is n = 0, meaning that the function is assumed to be a constant. Thus, for the function R we get the dierential equation: r2 R (r) + rR (r) 2 r2 R(r) = 0 for which the complete solution is well known: (r, ) = cI Io (r) + cK Ko (r) (4.54)
where Io and Ko mean the Bessel functions of the rst and second kind and zero order, respectively. For a diusive region 1 :
K 1 (r, ) = cI 1 Io (r ) + c1 Ko (r )
(4.55) 1 (r, ) K = cI 1 I1 (r ) c1 K1 (r ) n
216
where r2 r r1 . Due to the singularity of the Ko function when r tends to zero, solution for a diusive region 3 becomes: 3 (r, ) = cI 3 Io (r ) (4.56) 3 (r, ) = cI 3 I1 (r ) n where 0 r r3 . As we can see the solution does not depend on the angle .
K I We have three unknown coecients cI 1 , c1 and c3 , which could be calculated from boundary conditions on the 1 , 2 and 3 as follows:
1. for the boundary 1 : 1 (r) =2 n K I K 2cI 1 Io (r1 ) + 2c1 Ko (r1 ) + D (c1 I1 (r1 ) c1 K1 (r1 )) = 2 (4.57) 21 (r) + D
K cI 1 (2Io (r1 ) + DI1 (r1 )) + c1 (2Ko (r1 ) DK1 (r1 )) = 2
2. for the boundary 2 r 2 and r 2 3 : 2 (r) 2 (r) = D n 2 (r) = D ( ) 1 +

i =2 3
i (r )G (|r r |)di (4.58)
2 (r) 1 + 2 (r )G (|r r |)d2 + n 2 1 + 3 (r )G (|r r |)d3

3
where in our case = 1.
4.8. NUUTTIS 2D TEST EXAMPLE 3. for the boundary 3 r 3 and r 3 2 : 3 (r) 1 3 (r) = D + n
i =2 3
217
i (r )G (|r r |)di (4.59)
3 (r) = D
3 (r) 1 + n
2 (r )G (|r r |)d2
where the operator G (|r r |) is equal zero when the vector r 3 .
Points on the most inner circle are not visible for each other. Implementing Eq.(4.55) and Eq.(4.56) into Eq.(4.58) and Eq.(4.59) respectively, we will get K two missing equations which allow us to calculate unknown coecients cI 1 , c1 and cI 3 . But rstly we have to be more precise about the operator G (|r r |) in those two cases described by Eq.(4.58) and Eq.(4.59). Let consider the case where r 3 and r 2 (see Fig. 4.21)
n 0 r 2 r 3 2 3 U r Uo r
rr=(rr) x
Figure 4.21: Visible part of boundary 2 when r 3
218
CHAPTER 4. MEDIA WITH NONSCATTERING REGIONS cos Uo = r3 r2
cos = n cos because:
(r r) (|r r|) (r r ) = n (|r r |)
(4.60)
2 2 1/2 (|r r|) = (|r r |) = (r2 2r2 r3 cos U + r3 ) r2 cos U r3 cos = 2 2 1/2 (r2 2r2 r3 cos U + r3 ) r2 r3 cos U cos = 2 2 1/2 (r2 2r2 r3 cos U + r3 )
(4.61)
g1
arccos( r3 )
G (|r r |)d2 = 2r2

0
(r )
cos cos 1 a |rr | e dU = |r r | 2 (4.62)
arccos( r3 )
2r2
2 2 2 2 1/2 cos U 1 a (r2 cos U r2 r3 (1 + cos2 U ) + r3 r2 2r2 r3 cos U +r3 ) e dU 2 2 3 / 2 2 (r2 2r2 r3 cos U + r3 )
Let consider the case where r 2 and r 2 (see Fig. 4.22):
U (|r r|) = (|r r |) = 2r2 sin( ) = r2 (2(1 cos U ))1/2 2 |r r| cos = 2r2 |r r | cos = 2r2
2 2 arccos( r3 ) 2 arccos( r3 )
(4.63)
g3
2r2 G (|r r |)d2 = 2 2

0
r 2
(r )
cos cos 1 a |rr | e dU = |r r | 2
1 2 2
(1 cos U )1/2 ea r2 (2(1cos U ))
1/2
dU
(4.64)
219
n Uo 0 r 2 r 3 2 3 Uo r U r Uo
rr=(rr) x
And nally, let consider the case where r 2 and r 3 (see Fig. 4.23):
2 2 1/2 (|r r|) = (|r r |) = (r2 2r2 r3 cos U + r3 ) r2 r3 cos U cos = 2 2 1/2 (r2 2r2 r3 cos U + r3 ) r2 cos U r3 cos = 2 2 1/2 (r2 2r2 r3 cos U + r3 )
(4.65)
g2
3 arccos( r3 )
arccos( r3 )
G (|r r |)d3 = 2r3

0
(r )
cos cos 1 a |rr | e dU = |r r | 2
(4.66)
2r3
2 2 2 2 1/2 r2 cos U r2 r3 (1 + cos2 U ) + r3 cos U 1 a (r2 2r2 r3 cos U +r3 ) e dU 2 2 3 / 2 2 (r2 2r2 r3 cos U + r3 )
g2 =
r3 g1 r2
(4.67)
Now we can return to the boundary conditions on 2 and 3 (Eq.(4.58) and
220
Uo 0 r 2 r 3 2 3 Uo
U
r Uo r
rr=(rr)
Eq.(4.59)).
1 2 (r) 2 (r) = D + (r ) G (|r r |)d2 + n 2 2 1 + 3 (r ) G (|r r |)d3

3
2 (r) + 2 (r )g3 + 3 (r )g2 2 (r) = D n

2
(4.68)
3 (r) 1 3 (r) = D + (r ) n 2 3 (r) = D
G (|r r |)d2
3 (r) + 2 (r )g1 (4.69) n By substituting the solution for the Eq.(4.68) and Eq.(4.69) we have got:
K cI 1 [Io (r2 ) DI1 (r2 )] + c1 [Ko (r2 ) + DK1 (r2 )] = [ ] K I = g3 c I 1 Io (r2 ) + c1 Ko (r2 ) + g2 c3 Io (r3 ) K I I cI 3 Io (r3 ) = Dc3 I1 (r3 ) + g1 c1 Io (r2 ) + g1 c1 Ko (r2 )
(4.70)
4.8. NUUTTIS 2D TEST EXAMPLE where 2 =

a . D
221
By adding the Eq.(4.57) and doing some maths we will get a system of linear K I equations for cI 1 , c1 and c3 unknowns.
K cI 1 (2Io (r1 ) + DI1 (r1 ) + c1 (2Ko (r1 ) DK1 (r1 )) = 2 K cI 1 [(g3 1)Io (r2 ) + DI1 (r2 )] + c1 [(g3 1)Ko (r2 ) DK1 (r2 )] +
+cI 3 g2 Io (r3 ) = 0
K I cI 1 g1 Io (r2 ) c1 g1 Ko (r2 ) + c3 [Io (r3 ) + DI1 (r3 )] = 0
4.8.2
Analytical solution for P1 boundary conditions
P1 boundary conditions suggested by Rippol in [69], are dierent from diusive conditions used in FEM introduced by Arridge in [68]. Now, by using the 2D benchmark presented in Fig. 4.20, we are interested in what dierences are caused by using Diusive or P1 boundary conditions. In order to check this, once again we have to return to the boundary and interface conditions. We have three unknown coecients which can be calculated by boundary conditions on the 1 , 2 and 3 in a similar way as in previous paragraph: 1. for boundary 1 on this boundary conditions are unchanged:
K cI 1 (2Io (r1 ) + DI1 (r1 )) + c1 (2Ko (r1 ) DK1 (r1 )) = 2
(4.71)
2. for boundary 2 r 2 and r 2 3 : ( ) 2 (r) 2 (r) = D + n ( )] [ 1 RJ 2 (r) + 2 (r ) D G (|r r |)d2 + RU n 2 ] [ RJ 3 (r) 1 3 (r ) D G (|r r |)d3 (4.72) + RU n
3
222
3. for boundary 3 r 3 and r 2 :
3 (r) + (4.73) 3 (r) = D ] [ n 1 RJ 2 (r ) + 2 (r ) D G (|r r |)d2 RU n

2
By implementing Eq.(4.55) and Eq.(4.56) into Eq.(4.72) and Eq.(4.73) we will get missing two equations which allow us to calculate unknown coecients K I cI 1 , c1 and c3 . Operator G (|r r |) will be the same as for diusive boundary conditions.
[ ] 2 (r) 1 RJ 2 (r ) 2 (r) = D + 2 (r ) + D G (|r r |)d2 + n RU n 2 [ ] 1 RJ 3 (r ) G (|r r |)d3 + 3 (r ) D RU n

3
( 2 (r) = D
2 (r) n
(4.74) [ ] RJ 2 (r ) + 2 (r ) + D g3 + RU n [ ] RJ 3 (r ) + 3 (r ) D g2 RU n
[ ] 3 (r) 1 RJ 2 (r ) + 3 (r) = D 2 (r ) + D G (|r r |)d2 n RU n

2
[ ] 3 (r) RJ 2 (r ) 3 (r) = D + 2 (r ) + D g1 n RU n
(4.75)
223
By substituting solution for the Eq.(4.74) and Eq.(4.75) we have got: [ ( )] RJ I c1 Io (r2 ) + DI1 (r2 ) + g3 Io (r2 ) + D I1 (r2 ) + RU [ ( )] RJ K +c1 Ko (r2 ) DK1 (r2 ) + g3 Ko (r2 ) D K1 (r2 ) + RU ] [ RJ I I1 (r3 ) = 0 +c3 g2 Io (r3 ) D RU (4.76) [ I I K cI 3 Io (r3 ) = Dc3 I1 (r3 ) + g1 c1 Io (r2 ) + c1 Ko (r2 )+ )] RJ ( I +D c1 I1 (r2 ) cK 1 K1 (r2 ) RU By adding the Eq.(4.71) and doing some maths we will get a system of linear K I equations for cI 1 , c1 and c3 unknowns.
K cI 1 (2Io (r1 ) + DI1 (r1 ) + c1 (2Ko (r1 ) DK1 (r1 )) = 2
cI 1
) ] [ ( [ RJ K I1 (r2 ) + c1 (g3 1)Ko (r2 ) + (g3 1)Io (r2 ) + D + g3 RU ( ) ] ] [ RJ RJ I D + g3 K1 (r2 ) + c3 g2 Io (r3 ) D I1 (r3 ) = 0 RU RU (4.77) ] ] [ [ RJ RJ cI I1 (r2 ) + cK K1 (r2 ) + 1 g1 Io (r2 ) D 1 g1 Ko (r2 ) + D RU RU +c I 3 [Io (r3 ) + DI1 (r3 )] = 0
4.8.3
Numerical results
Comparison of the solution for the internal eld distribution along the radius of the domain presented in Fig. 4.20, is shown in Fig. 4.24. As a reference to the analytic, FEM and BEM solutions with a clear layer, the solution for the homogeneous diusive region is shown in the same gure. Between analytic and the FEM solution with the void (nonscattering/clear region) the conformity is almost ideal, slightly worse for the BEM solution
224
3

3 no void P1 BC
2.5
2.5
1.5
1.5
1 Analytical 0.5 FEM 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 BEM
0.5
0 0
0.2
0.4
0.6
0.8
Figure 4.24: Internal eld calculated analytically dashed line ( [41], Fig. 4.4) and its comparison with the FEM dotted line and BEM solid line solutions (left) and on the right picture there is an analytical solution for P1 boundary conditions used in BEM code
in a range of radius equal to 0.8 1.0. But within a range of radius equal to 0.0 0.5 it is just the opposite. BEM shows an excellent conformity with the analytical solution when the FEM solution is below them. This is due to the way in which the nonlocal boundary conditions are treated by both methods (for details see [8, 6971]). The BEM is using the same nonlocal boundary conditions as the analytical approach. Such experiment does conrm that BEM software is constructed correctly giving the same results as the analytical solution, but the main question, which results better reect reality, still remains.
4.8.4
Multilayered neonatal head model with the CSF layer
The multiregion strategy described in previous sections was employed. The target is to create the multilayered baby head numerical model taking into account the CSF layer. To achieve this goal we have started with the three layer spherical geometry rst than with the three layer structural meshes generated on a babys head. To achieve the structural meshes, parametric

3
3
225
2.5
2.5
P1 BC
1.5
1.5 Diffusive
P1 BC
0.5 Diffusive BC 0 0 0.2 0.4 0.6 0.8 1
0.5
0 0
10
15
20
25
Figure 4.25: Internal eld calculated analytically for P1 nonlocal boundary conditions: r1 = 25 mm, r2 = 22.5 mm, r3 = 19.5 mm, a = 0.01 mm1 and s = 1.0 mm1 (left) and comparison with the diusive boundary conditions (right)
surfaces were generated with the aid of the algorithm based on spherical harmonics described in [85, 87]. To get a reference point, rst the threelayered spherical structure was considered (see Fig. 4.26) and then the babys head discretized by the structural mesh without a clear layer (Fig. 4.27 left) and with the clear layer (Fig. 4.27 right). As we can see from Fig. 4.27, there is a big dierence between both solutions. At the moment, the comparison is only done quantitatively as FEM or MC solutions do not exist for the babys head so far.
226
2 4 25 20 15 10 5 0 5 10 15 20 25 20 10 0 10 20 20 2 4 25 20 15 10 5 0 5 10 15 20 25 20 10 0 10 20 20 2 4 25 20 15 10 5 0 5 10 15 20 25 20 10 0 10 20 20 0 10 10 20 20 22 24 16 18 12 14 10 6 8 0 10 10 20 20 22 24 16 18 12 14 10 6 8 0 10 10 20 20 22 24 16 18 12 14 10 6 8
Figure 4.26: Spheres for the 3mm gap
227
30 25
30
4
25
6 20 15 10 5 0 12 30 20 10 0 20 15 10 5 10
20 15
8
10 5 0 30 20
10
12
14 0 16
2
10 0 20 15 10 5 0
14
16
25
25
30 25
30 25
6 20 15 10 5 0 30 20 10 0 20 15 10 5 10
6 20 15 10 5 0 10
12
12 30 20 10
14 0 16
14 0 16
2
25
25
20
15
10
30 25
30 25
6 20 15 10 5 0 12 30 20 10 0 20 15 10 5 10
20 15 10 5 0 30 20 10
10
12
14 0 16
14 0 16
25
25
20
15
10
Figure 4.27: Babys head model without a void gap (left) and with a 3mm void gap between the second and the third surface (right)
228
4.9
Conclusion
The application of the boundary element method to the regions containing the nonscattering inclusions was presented in this chapter. Such an approach seems to be more natural than the nite element one, due to the fact that clear regions introduce to the discrete form of FEM a nonsymmetric and fully populated matrix, which cost of loosing the most desiring features of FEM like sparse and bounded coecient matrix. This is not an disadvantage of the BEM and what is more, the nonlocal boundary conditions could be treated in the same way as the multiregion BEM but with a dierent function as a fundamental solution (see eq.(4.32)). All this, as well as the more simple procedure of discretisation of the surfaces against those ones for the volume discretisation, the Author is convinced to the conclusion that the BEM might be an ecient and exible tool for neonatal babys head modelling in Optical Tomography.
Chapter 5 BEM formulation for thin layers

Thin layers like a skull in Electrical Impedance Tomography or a CSF layer (void) in Optical Tomography are causing many problems due to geometrical representation. In order to discretize such subregions, huge amount of elements (nodes) should be involved to achieve a satisfactory results. It is dicult to aord such a ne discretization especially in the Inverse Problem solution, were the calculation time is the main issue. In this chapter, following ideas of mechanical and thermal engineering [60,65], new approach for such thin layered subregion discretization will be presented in Boundary Element formulation.
5.1
Introduction
To start with this problem let focus our attention on the Laplace equation and a constant interpolation function on at triangular boundary element. If such approach is useful for the Laplace equation, it would be quite straightforward to expand it to a diusion equation as well. 229
230
CHAPTER 5. BEM FORMULATION FOR THIN LAYERS
5.2
Standard BEM formulation
Consider a threedimensional region closed by external surface S + and internal surface S as shown in Fig. 5.1. The unitnormal vector at any point on S + is opposite to the outward unitnormal on S . Let a suciently smooth function satisfy Laplace equation in the volume .
2 (r) = 0 r
(5.1)
On both surfaces S + and S Dirichlet boundary conditions were imposed.

y n+
+ Q Q
n Midsurface
p
0 z Negative side S
+ Positive side S
Figure 5.1: Schematic diagram of two spherical surfaces close to each other
For the Laplace equation the fundamental solution (i.e. the Green function), for 3D space is [14, 15]: G(p, Q) = 1 4R (5.2)
where the vector r(p, Q) represents the distance between the point p and the
5.2. STANDARD BEM FORMULATION point Q, given by: R(p, Q) = |r(p, Q)| = (Xp xQ )2 + (Yp yQ )2 + (Zp zQ )2
231
(5.3)
where: p , Q S + S . We can use Green theorem in its second form [14] to derive an integral equation: (Q) (p) + K1 (p, Q)(Q)d(Q) = K2 (p, Q) d(Q) (5.4) n

where the potential kernels K1 and K2 are dened as follows: K1 (p, Q) = G(p, Q) n (5.5)
K2 (p, Q) = G(p, Q)
(5.6)
Since measurements of the function are made at the boundary, we consider point P S + S , which results in: (Q) c(P )(P ) + K1 (P, Q)(Q)d(Q) = K2 (P, Q) d(Q) (5.7) n

The extra function c(P ), arises due to the singularities on the boundary. It can be calculated by surrounding boundary point P by a small sphere of radius and taking each term of Eq.(5.7) in the limit as 0 [15]. However, as shown in [14, 15], the term c(P ) does not have to be calculated explicitly, and can be obtained indirectly by utilizing some simple physical considerations. To calculate the rst kernel K1 (P, Q), the function G(P, Q) is dierentiated with respect to the unit normal at the point Q, as follows: ( ) [ ( ) ( ) ( )] G(p, Q) G R G R x R y R z = = + + (5.8) n R n R x n y n z n
232
where the derivatives of the coordinates x, y and z with respect to the unit outward normal n in point Q, are the components of the outward normal: nx = x n ny = y n nz = z n (5.9)
and R(p, Q) x Q Xp = , x R(p, Q) R(p, Q) yQ Yp = , y R(p, Q) R(p, Q) z Q Zp = . z R(p, Q) Therefore, the rst kernel can be written: K1 (p, Q) = 1 [(xQ Xp )nx + (yQ Yp )ny + (zQ Zp )nz ] 4R3 (5.11)
(5.10)
To solve the threedimensional problem numerically, the surface was discretized into piecewise constant elements.
5.3
Modication for closely spaced surfaces
Starting with the interior formula (c(p) = 1 in Eq.(5.4)), we shall rst shrink the surfaces S + and S into (see Fig. 5.1), then let p approach . The order of the limiting operations is important: 1. limit S + , limit S , 2. limit p , limit p . After limiting operation 1, the kernels in Eq.(5.7) are related to: K1 (P, Q+ ) = K1 (P, Q ) K2 (P, Q+ ) = K2 (P, Q ) (5.12)
(+) ()
5.3. MODIFICATION FOR CLOSELY SPACED SURFACES
233
where Q+ and Q are images of one another on S + and S , respectively. The negative sign leading K1 in Eq.(5.12) is caused by the opposite signs of n on Q+ and Q . We shall adopt the sign convention that n(Q+ ) = n(Q ) and n(Q ) = n(Q ). By applying Eq.(5.12) to Eq.(5.7), we obtain a modied interior formula: (Q) (p) + K1 (p, Q)(Q)d(Q) = K2 (p, Q) d(Q) (5.13) n

where (Q) n = (Q+ ) (Q ) + n n
(Q) = (Q+ ) (Q ). Now we perform limiting operation 2. In the limit p , Eq.(5.13) becomes: [ ] (P + ) c(P + ) (P + ) (P ) + (5.14) (Q) + K1 (P, Q)(Q)d(Q) = K2 (P, Q) d(Q) n
() (+)
while in the limit P : [ ] (P ) c(P ) (P + ) (P ) + (5.15) (Q) + K1 (P, Q)(Q)d(Q) = K2 (P, Q) d(Q) n

At the rst glance, the two equations above appear to be distinct. However, after noting that c(P + ) + c(P ) = 1, one sees that Eq.(5.14) and Eq.(5.15) are indeed identical. In the Eqs.(5.13), (5.14) and (5.15) we have removed the diculties associated with the integration in Eq.(5.7) over closely spaced surfaces. Unfortunately, these new formulas are not very useful, because they cannot be combined with the usual set of boundary conditions to form wellposed boundary value problem. For example, suppose we specify on and wish
234
Eq.(5.14) (or its twin, Eq.(5.15)) to yield on the surface . Since apn n + (Q ) (Q ) pears in the integrals over only in the form = n + n , these n on equations determine only previously mentioned linear combination of n . The solution is not unique because Eqs.(5.14) and (5.15) are redundant. We can complete the set with the normal derivative of the boundary formula. First, take the gradient of the interior formula (Eq.(5.13)), with respect to p:
P (p) +

P K1 (p, Q)(Q)d(Q) = P K2 (p, Q)
(Q) n
d(Q)
(5.16)
Next, let p in Eq.(5.16) and perform the dot product with the unit normal to at P to obtain: [ ] (P + ) c(P + ) (P + ) + (P ) + (5.17) (Q) + K1 (P, Q)(Q)d(Q) = K2 (P, Q) d(Q) n

(+)
where is the outward unit normal at P , K1 = 2 K1 /n, K2 = K2 / and = / . With Eq.(5.14),(5.15) and (5.16), we can now eciently model the volume between the two closely spaced surfaces. Surface may be meshed using standard elements with the understanding that two degrees of freedom are associated with every node on (it means that every node contains four variables: (P + ), (P ), (P + )/n and (P )/n as it is shown in Fig. 5.2. Compared to the standard procedures, the numerical quadrature is a bit more complicated due to the singular integrand in Eq.(5.16). One of several procedures may be used to reduce the order of the singularity [35].
5.4. INTEGRATION OF SINGULAR INTEGRALS

+ + , n
235
, n
Figure 5.2: Variables associated with a boundary element
5.4
Integration of singular integrals
Let consider the singular integral
K1 d(Q) over a at triangular area
(), which includes the point P . As usually, we shall transform to polar coordinates (, ), centered on P , to lower the degree of singularity. We shall also hold Q, a distance d of the surface S until the integral is evaluated, then take the limit d 0, i.e.: ] ] [ [ 2 3( r)(n r) (n ) 3d 1 K1 d(Q) = d = 3 d = R5 R3 R5 R

= = 1 2
] 3d 1 d d = (2 + d2 )5/2 (2 + d2 )3/2 0 [ ] () 2d2 2 d (5.18) (2 + d2 )3/2 (2 + d2 )1/2 0

2
() [
where r = P Q, and () describes the extent of in polar coordinates. The crucial step is to note that the lower limit of Eq.(5.18) is zero, independent of d, and to drop this limit before letting d 0. in the limit d 0.
[ K1 d(Q) =
] 1 d ()
This integral can be evaluated analytically in a simple manner. Let consider
236
the at triangle in a local coordinate system.

(0,1,0) 2
2
1 =1 2
0 (0,0,1)
1 (1,0,0) 1
Figure 5.3: Local coordinates of the triangle
The shape functions are dened: N0 (1 , 2 ) = 1 1 2 , N1 (1 , 2 ) = 1 , N2 (1 , 2 ) = 2 . The rst derivatives of the standard interpolation functions with respect to the 1 and 2 are given by: N0 (1 , 2 ) = 1, 1 N1 (1 , 2 ) = 1, 1 N2 (1 , 2 ) = 0, 1 N0 (1 , 2 ) = 1, 2 N1 (1 , 2 ) = 0, 2 N2 (1 , 2 ) = 1. 2 (5.19)
(5.20)
The singular integrals in zeroorder triangular element over a at area can be evaluated analytically (consult Eq.(2.20)).
5.5. FUTURE WORK
237
5.5
Future work
The CSF layers cause many problems for numerical modelling. That is why extension of the idea presented in this chapter seems to be very promising for Diuse Optical Tomography. The thin layers with varying thickness d
d 2
d1
Figure 5.4: Thin layered region with varying distance d
(see Fig. 5.4) could approximate eectively the CSF subregion.
5.6
Conclusion
In this chapter the 3D BEM numerical model for the particular forward problem has been presented. As the main advantages of presented approach to modelling closely spaced surfaces from the point of view of Impedance Tomography or Optical Tomography applications the following ones can be mentioned: an easy way to represent a scull or a void region in a numerical model of the humans head, less number of nodes needed to construct a reliable mesh,
238
more precise results, easier calculations of visibility function in Optical Tomography.
Chapter 6 Wavelet based techniques for CPU time reduction

One of the biggest advantages of the BEM is less data preparation time, comparing to the FEM as a direct result of the surfaceonly modelling. Some of the authors say that it is a reduction of dimensionality by one. It is well known that the application of the boundary element method for the numerical solution of boundaryvalue problems produce nonsymmetric, dense matrices, which are computationally expensive to solve using direct methods such as Gauss elimination, requiring an eort of order 0 (N 3 ) for systems with N unknowns. To reduce the computational eort involved in the solution, several papers have been published on iterative methods such as conjugate gradients or GMRES and their respective preconditioners [63]. Despite some relative success, the time necessary to achieve accurate solutions, even with welltuned iterative solvers, is still of order 0(N 2 ). There is a great eort, at present, on the development of very fast solvers for the BEM, of order 0(N logN ), based on the idea of reducing the redudancy by using panel clustering [34] or fast multipole expansions [72]. Although very ecient, these techniques require completely new numerical formulations and related computer programs. Another alternative exploited in this chapter is
239
240
CHAPTER 6. WAVELET BASED TECHNIQUES
Wavelet series expansion of and d /dn
Boundary data
Traditional matrix assembler
New quadrature algorithms
Panel Clustering, Fast Multipole
Dense Matrix
Sparse Matrix
Wavelet compressor
Direct (Gauss) or Iterative (GMRES) solver
Iterative Solver
Final solution
Figure 6.1: Schematic view of commonly used solver techniques in boundary element methods
to apply wavelet transforms. The main advantage of this technique is that it can be directly used as a black box in existing BEM programs. The application of wavelet compression techniques for the solution of BEM systems of equations is still in its infancy and only very simple problems of potential theory have been dealt so far [5, 31]. Robust wavelet compression algorithms are necessary if the technique is to be applied for the solution of practical engineering problems. The two basic ways of implementing wavelet techniques into the boundary element method are schematically shown in Fig. 6.1, and include direct matrix compression, and wavelet series expansions of the state variables and their normal derivatives.
241 The wavelet series expansion of the variables avoids the calculation of the complete matrix, substantially reducing the total computation time to the order 0(N logN ). The drawback is that it requires a completely new numerical formulation, since element matrices are now calculated in the wavelet space, and completely new quadrature techniques to accurately represent them. Therefore, standard boundary element codes (dashed line path in Fig. 6.1) cannot be adopted to this technique. The former form, which is the subject of this chapter, can be applied as a black box to existing boundary element codes with minimal changes. It is illustrated in Fig. 6.1 by the thick path leading to shadowed boxes. The main reason is that wavelet compression techniques are only applied to the system matrix before the solution of the nal system starts. Basically, this approach takes dense matrices and produces corresponding sparse ones. Therefore, all previous theory and numerical algorithms for the BEM can still be used, since the wavelet transform is directly applied to the matrices resulting in the original discretization. In order to show how the solution of dense matrices is important, let consider the time consumption versus number of unknowns. To solve the system of equations produced by the BEM, the Generalized Minimum Residuals (GMRES) method has been used. This method, rstly proposed by Saad and Schultz [73], nowadays is almost exclusively used to solve large and non symmetric linear system of algebraic equations. Comparison of the CPU time solution of BEM equations for dierent numbers of unknowns is presented in Table 6.1. For better comparison the solution of
Table 6.1: Solution time for dierent number of unknowns

GMRES tolerance 1e-6 1e-9 Number of unknowns 1024 2048 4096 8192 0.0625 0.1875 0.3125 1.000 8.375 15.500 552.6 5477.5
242
the system for 2048 unknowns and tolerance equal to 1e-9 was assumed to be one. All computations were done on the Intel Pentium 4 1.7 GHz PC with 256 MB RAM. We can see how fast the solution time for the fully populated matrices is increasing with the number of unknowns. It is very dicult to accept such behavior for the forward solver, especially when we plan to create the multilayered babys head and more than 15000 equation is the minimal number which we will need. So this problem might be qualied as a Large Scale one. One of the methods recently developed [1, 5, 26, 31] to deal eciently with LS problems is a Discrete Wavelet Transform (DWT).
6.1
Discrete wavelet transform
The discrete wavelet transform is a fast, linear operation that operates on a data vector whose length is an integer power of two (which is a disadvantage but as we can see later it can be overcome quite easily), transforming it into a numerically dierent vector of the same length [64]. It causes some loss of information, but preserves important features of the problem. One of the most promising wavelet applications is linear algebra. The basic idea is to think of an integral operator, that in case of BEM is a large dense matrix, as a digital image. Suppose that the operator compresses well under a two dimensional wavelet transform, ie. the large fraction of its wavelet coecients is so small as to be negligible. Then system of equations generated by BEM becomes a sparse system in the wavelet basis. So, to solve the system of equations: Ax = b (6.1)
First we transform the operator A and the righthand side b in the following way: WAWT , A Wb b (6.2)
6.1. DISCRETE WAVELET TRANSFORM
243
where W represents the onedimensional wavelet transform, then we solve: A x=b and nally transform to the answer by the inverse wavelet transform: x = WT x (6.4) (6.3)
coecients When transformation is completed we can neglect all the matrix A with magnitudes smaller than assumed threshold (truncation level) :
i,j =1,2...N
aij : |aij | < aij = 0
(6.5)
where N is the matrix dimension. is [64]: The number of negligible elements in matrix A 10N log10 1 (6.6)
is a ngerlike sparse matrix and is presented in Finally, the matrix A Fig. 6.2. Matrix elements with magnitude greater than are shown in black color. The standard DWT with Daubechies lters of 4th , 12th and 20th order has been tested. For all lters results are similar, so only the 12th order lter is shown in the Fig. 6.2.
Figure 6.2: The BEM matrix after DWT with Daubechies 12 order lter and with threshold 1e-6 (left) 1e-5 (centre) and 1e-4 (right)
244
The transformation matrix for the simplest member of Daubechies lters called DAU B 4 [64], has the following form: c0 c1 c2 c3 c3 c2 c1 c0 c c c c 0 1 2 3 c3 c2 c1 c0 . . . . . .. (6.7) . . c0 c1 c2 c3 c c c c 3 2 1 0 c0 c1 c2 c3 c1 c0 c3 c2 where c0 = (1 + 3)/4 2, c2 = (3 3)/4 2, c1 = (3 + 3)/4 2, c3 = (1 3)/4 2.
For higher order of Daubechies lters the values of coecients can not be expressed in a close form (except of an order 6) and theirs numerical values may be found in [64].
6.2
Time acceleration
By applying DWT some acceleration of the time solution of the BEM system of equation has been achieved (see Table 6.2). The acceleration has been achieved only for the system of 4096 and more equations, which is the case we are the most interested in (babys head model). In case of 8192 elements, the acceleration can be even equal to 5477.5/174.56 = 31.4 times depending on GMRES tolerance. Of course, we have to pay for such acceleration by the increase of relative error comparing to the solution of the original BEM system. Taking into account those numerical experiments we can conclude that for the Optical Tomography it is worth to implement DDAUB12 lter with the
6.3. DWT WITH PERMUTATIONS Table 6.2: Solution time for a sparse system after DWT
GMRES tolerance 1e-6 treshold 1e-4 1e-5 1e-6 1e-4 1e-5 1e-6 Number of unknowns 1024 2048 4096 8192 0.4375 0.4375 0.4375 0.2500 0.2500 0.2500 0.9375 0.9375 1.0000 0.9375 0.9375 1.0000 15.00 15.40 16.30 15.75 14.50 15.19 123.62 161.06 163.62 168.88 181.00 174.56
245
1e-9
threshold 1e6 or even 1e5 . The latest threshold reduces more than 80% of the entries of a dense BEM matrices, increasing only the relative error by about 2%.
6.3
DWT with permutations
To obtain bandlike pattern matrix, instead of ngerlike one, we have used new nonstandard DWT with permutations based on Haar and Daubechies lters [26]. In this case a new onelevel DWT matrix will take the form: c0 O c 1 O c 2 O c3 O . . . . . . O I O O O O O O ... ... c3 O c2 O c1 O c0 O . . . . . . O O ... ... O O O I O O . . . . . . .. .. . . . . . . (6.8) . . . . . . ... ... . . c2 O c 3 O . . . . . . c 0 O c 1 O O O O O ... ... O I O O c1 O c0 O . . . . . . c3 O c2 O O O O O ... ... O O O I Here I is an identity matrix of dimension 2L1 1 (where L is the wavelet level) and the O s are block zero matrices. For L = 1, both I and O are of
246
dimension zero [31], so the wavelet lter is the same as in standard DWT. Resulting matrices for dierent wavelet lters are shown in Fig. 6.3.
Figure 6.3: The BEM matrix after DWT with permutations, treshold 1.0e-5, level 3 DAUB4 lter (left) DAUB12 lter (centre) and DAUB20 (right)
For the DWT with permutation and for Haar or Daubechie lters, the time consumption is assembled in table 6.3. As one can see the DWT with perTable 6.3: Solution time after DWT with permutations
GMRES tolerance 1e-6 DWT lter haar daub4 daub12 daub20 haar daub4 daub12 daub20 Number of unknowns 1024 2048 4096 8192 0.0625 0.0625 0.0625 0.1250 0.0625 0.0625 0.0625 0.1250 0.2500 0.1875 0.3125 0.6250 0.3125 0.2500 0.4375 0.6250 5.2500 5.0625 5.0000 8.1875 6.0625 5.9375 4.9375 7.8125 208.50 211.44 195.63 198.19 206.13 206.56 196.44 201.00
1e-9
mutation is very useful for the system of about 4000 or more equations. The CPU times are shorter than for the standard DWT or for the systems with dense matrices. But for largescale problems (see the 8192 equations case)
6.4. DWTFOR NOT 2N
247
the CPU time is still much shorter comparing to a dense system, but slightly greater than for the standard DWT.
6.4
DWT when the size of the coecient matrix is not a number equal to 2n
The boundary Element Method application to Diusive Optical Tomography is very attractive thanks to the fact that is much easier to generate good quality patient devoted surface mesh (BEM) than the volume one (FEM). One of the consequences is patients dependant matrix coecient dimension (see Eq.(6.1)). In most cases dimension is not equal to the number 2n . Therefore we have to apply generalized algorithms [52] or we can increase the number of equations (dimension of coecients matrix) in order to complete the number of equations to the number 2n . In case when the number of equations is not a number equal to 2n there is a simple method which can use a standard DWT algorithm. It is enough to add such a number of auxiliary equations that the total number of equations becomes equal to 2n . ak,k xk = bk (6.9)
where subindex k is changing from an arbitrary number of equations till the number of equations equals to 2n . The diagonal coecients of matrix A (see Eq.(6.1)) must be selected in such a way that the condition number of matrix A will not increase signicantly [64]. The main advantage of such approach is its simplicity. However, the disadvantage is that we have to solve much bigger problem. In such cases the advantage of DWT might be problematic. Dierent but typical for DOT structure of coecient matrix is a block bounded one. Such kind of matrices arise when multilayered structures are mo-
248
1

1 1
32
32
32
64
64
64
96
96
96
128
32
64 nz = 7189
96
128
128
32
64 nz = 4692
96
128
128 1
32
64 nz = 2774
96
128
Figure 6.4: The BEM matrix after DWT, treshold 1.0e-5, (left) treshold 1.0e-4 (centre) and treshold 1.0e-3 (right)
1 1 1
128 160
128 160
128 160
256 1
128 160 nz = 15915
256
256 1
128 160 nz = 10837
256
256 1
128 160 nz = 5724
256
Figure 6.5: The BEM matrix after DWT, treshold 1.0e-5, (left) treshold 1.0e-4 (centre) and treshold 1.0e-3 (right)
Figure 6.6: The BEM block matrix (left) after DWT treshold 1.0e-5 (centre) and treshold 1.0e-3 (right)
delled with the aid of BEM. Then it is very dicult to achieve the number of
6.5. CONCLUSION
249
unknowns equal to 2n . DWT, as it was in previous cases describing the homogeneous regions, transforms of blockstructured matrix to the sparse one with a number of nonzero coecient depending on the threshold applied (see Fig. 6.6).
600 without DWT with DWT 500
400 solution time
300
200
100
0 4500
5000
5500
6000 6500 7000 number of equations
7500
8000
8500
Figure 6.7: Solution time versus number of unknowns without and with wavelet tansformation
6.5
Conclusion
Using the standard DWT or DWT with permutations the solution time reduction has been tested in this chapter. In all cases the GMRES algorithm has been used to solve the resulting sparse system of equations. The wavelet transformation seems to be a very ecient tool especially for largescale problems, which are the most interesting from optical tomography point of view. However, after numerical experiments carried out in this chapter, we can say that the most useful, regarding the time reduction and the low level error preservation, is the standard DWT with Daubechies 12 lter. Due to the large solution errors, DWT with permutations for Haar and order 4 of Daubechies wavelet lters are rather useless.
250
Chapter 7 FEMBEM coupling

There is a lot of discussion about the advantages and disadvantages of the BE method when compared to the FE one. Clearly, there are certain applications where one technique is more suitable than the other. But for optical tomography problems combining both techniques in the same computer program would be the most ecient way of modelling the light point sources or the void regions for example. In order to take advantage of FEM and BEM, their coupling has been investigated extensively in several engineering elds, such as geomechanics [15, 16], solid mechanics [88], fracture mechanics [3] and electromagnetics [22, 25, 33, 51, 80] There are several dierent methods of coupling BEM and FEM [22, 51, 59]. The methods discussed in this chapter are limited to the direct coupling of the Boundary Element with the Finite Element matrices.
251
252
CHAPTER 7. FEMBEM COUPLING
7.1
2D space
This problem is closely related to the multiregion problem of the BE method such as presented in Fig. 7.1. The multiregion analysis has to full
2 n 2 n i 1 1 1 n n n i n 2 2 n
Figure 7.1: The multiregion boundary element analysis
continuity conditions along the interface line i between 1 and 2 regions. This results in the following two relationships: i n
(1) i (1)
= i =
(2)
(7.1) n
(2) i
Let the subregion 1 be discretised by the Finite Elements and the 2 by the Boundary Elements. Along the common interface, two conditions must be satised Eq.(7.1): continuity (the rst one) and equilibrium (the second one), as it was in case of multiregion BEM. Continuity of the state function can be maintained by using the same order of basis functions in both FE and BE formulations. Thus, if a three nodes isoparametric quadratic boundary element is used, an equivalent nite element
7.1. 2D SPACE
253
such as for example eight nodes quadrilateral quadratic element or six nodes isoparametric triangle has to be used for the nite element approximation. The essentiality of the problem lies in the fact that the interpolation for the derivatives of the potential for the FEM lies one order lower than the order of the potential itself, whereas for the BEM formulation developed here, the interpolation functions have the same order not only for the potential but also for its normal derivatives. Such unequal interpolation of the normal derivatives on the interface implants an error to the resulting system of equations. This problem will be discussed later in this chapter. Because along the interface, the continuity and equilibrium conditions have to be fullled, that is why for the FE approach we have additional unknown ux, which is expressed by Neumann boundary conditions.
7.1.1
Incorporating the BE equations to the FE ones
Such approach seems to be more useful for Optical Tomography, due to already existing FE code or due to the necessity of Inverse Problem solution when the environmental parameters have to be reconstructed. The boundaryvalue problem for the FE subregion is dened by the second order dierential equation [46]: x ( D x ) y ( D y ) + k2 = 0 (7.2)
in conjunction with the boundary conditions: | = 0 (7.3) n =

i
254
The equivalent variational problem for the boundaryvalue problem dened above is given by: F () = 0 with the boundary conditions (7.3), where 1 F () = 2 D
(F E )
(7.4)
[(
)2 +
)2 +k
2 2
] d +
di
i
(7.5)
To discretize the functional (7.5), the FE subregion is divided into M elements and interface boundary i (see Fig. 7.1) is broken into Mi segments. Usually, M is much larger than Mi . Within each area element, the potential function (electric potential or photon density) is expressed as: (x, y ) =
ne i=0 eT e eT e Nie (x, y )e i = N = N
(7.6)
where ne number of nodes within the nite element. On each segment of the interface line, the eld is expressed as: (x, y ) =
i 2 i=0 i i i i Nii (x, y )i i = N = N T T
(7.7)
Assuming that the interface boundary i is a smooth contour, the normal derivative of the boundary eld, which is denoted as , is well dened at each interface node and therefore can also be expressed as: (x, y ) =
i 2 i=0 i Nii (x, y )i = Ni i = i Ni T T
(7.8)
This is a weak point of this approach because the and its normal derivative are approximated by the same shape functions. Results of such approach will be demonstrated later. Substituting Eq.(7.67.8) into Eq.(7.5), we obtain:
Mi 1 M 1 1 eT e e T F = A + i Bi i 2 e=0 i=0
(7.9)
7.1. 2D SPACE where A =

e e
255 [
] Ne Ne T Ne Ne T 2 e eT D + dxdy k0 N N dxdy x x y y
e
(7.10)
and B =
i
Ni Ni di
(7.11)
Providing that the element length of the interface is small, the Jacobian of transformation to local coordinate system may be assumed constant and taken out of the integral sign in Eq.(7.11) without causing signicant errors. Therefore, by substituting the explicit expressions for the shape functions, it is easy to perform the indicated integrations analytically. So, the entries of matrix Bi , in case of the quadrilateral three nodes isoparametric elements of interface in local coordinate system, are dened by: 2 1 4 i i i i i i N1 N1 N1 N2 N1 N3 15 15 15 i i 2 16 2 i i i i J ( ) = Bi = N2 J ( ) (7.12) N1 N2 N2 N2 N3 15 15 15 1 2 4 i i i i i i 15 15 N3 N1 N3 N2 N3 N3 15 Then, performing the assembly Eq.(7.9) can be written as: 1 F = T A(F E ) + T B(F E ) 2 (7.13)
where A(F E ) is an N N square matrix, B(F E ) is an N Mi rectangular matrix, is a column vector representing the nodal values of eld intensity and is a column vector representing the nodal values of on the Mi n segments of the interface. Dierentiating F with respect to each nodal value of the and equating the resulting expression to zero yields a system of linear equations: A(F E ) (FE) + B(F E ) (F E ) =0 n (7.14)
Matrix form of the BE subregion can be written as: A(BE ) (BE ) B(BE ) (BE ) =q n (7.15)
256
Now, the system of boundary equations can be incorporated to the system of nite elements. (F E ) (F E ) (F E ) A11 B12 0 0 0 A(F E ) B(F E ) (F E ) 0 0 21 n 0 22 = (7.16) (BE ) (BE ) 0 0 A11 B12 (BE ) q1 (BE ) (BE ) (BE ) 0 0 A21 B22 q2 n Note that the superscripts (BE) and (FE) indicate to which subregion a particular matrix is prescribed. The simplest approach to solving Eq.(7.15) and Eq.(7.14) is to solve them simultaneously, that is to solve an (M + 3Mi + 2M2 ) (M + 3Mi + 2M2 ) matrix system, where M2 is the number of unknowns on the boundary 2 in Fig. 7.1. The resulting matrices have to be rearranged to accommodate the continuity and equilibrium conditions (Eq.(7.1)) as well as the prescribed boundary conditions. As a result we will get the matrix with the following structure (see the Fig. 7.2).
11111 00000 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 FEMBEMl 000000000000000000000 111111111111111111111 BEM 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111
111111111111111111111 000000000000000000000 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 FEM 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111
Figure 7.2: The FEMBEM matrix structure (left) and an exemplary image for the discretization shown in Fig. 7.4 The matrix is asymmetrical with much bigger bandwidth with two additional groups of nonzero elements caused by the interface between FE and BE subdomains. Rearrangement of resulting matrix is very similar to that required in multiregion BEM problems. This method has the advantage of
FEMBEMu
257
not requiring the inversion of the B(BE ) matrix. However, choosing which approach to use depends on whether the FE or BE is the dominant and whether it is necessary to make the BE matrix symmetric to implement it in an existing FE code.
7.2
Numerical examples
Let consider several examples starting from the simplest one and then slowly move to the standard optical tomography benchmark. Our attention we will focus on the local errors which can occur on the interface FE/BE. By the numerical experiments we would like to prove that hybrid modelling is able to provide results with a high precision. Such a feature is particularly important in case of DOT.
7.2.1
Simple benchmark problem
As a benchmark problem let consider the steady state for a diusion equation in a square region (see Fig. 7.4): 2 (x) a (x) = 2 (x) k 2 (x) = 0 D (7.17)
where k is a so called wave number and for the steady state k in some cases (for example the neutron transport problem) is a real number. The following boundary conditions are imposed: Dirichlet boundary conditions: (x)|x=0 = 0 and Neumann boundary conditions: (x) n =
x=0
(x)|x=a = 10
(7.18)
(x) y
=
x=0
(x) y
=0
x=a
(7.19)
258
This problem, due to the geometry and boundary conditions can be reduced to onedimensional problem. Formally, this problem will be solved in 2D space and results will be compared with the analytical solution which are: ( ) 1 D k(ax) 1 k(ax) D kx 1 kx (x) = (a) e J (a) e + (0) e + J (0) e D 2 2k 2 2k where for Optical Tomography, usually diusion coecient D = 0.03 cm and a = 0.1 cm1 so k = a /D 1.8257 cm1 , and J (0) = (x) kD = (cosh(k )(0) (a)) n sinh(k ) (7.20) J (a) = kD sinh(k )(0) + cosh(k )kD (cosh(k )(0) (a)) sinh(k )
First, only the solution for FEM and BEM, and then the comparison of the results with the analytical solution for two dierent discretizations are presented. In order to achieve the following results with the aid of FEM, region was discretised by 256 elements with 833 nodes (bandwidth equal 53) and with the aid of BEM. In case of BEM region was discretised by 128 elements and 256 nodes. Both discretization are presented in Fig. 7.3. For FEM, solution is characterized by linearly changing precision of the solution along the x axis. Using the BEM, the relative error of the solution inside the region is constant (immanent feature of BEM) and the level of the maximal relative error is almost the same in both cases as we can see in Fig. 7.3. Now we know of what level of the error we can expect. So we may use the hybrid method for the solution of this kind of benchmark and what is more we can see how the relative error due to the hybrid method application will change. Part of the region was discretized by BEM and the rest of it by FEM. Coarse and dense discretization was applied (see Fig. 7.4). One thing is obvious after inspecting Fig. 7.4, that relative error increased signicantly only due to the Hybrid Method application. The second thing
259
is the error increment on the common line for FEM and BEM. This line we will call the interface.
260
10
10
0 0
10 10 10 10 10 10 10 10 10
1
0 2 4
FEM Analytic solution FEM
0
10
4
BEM
10
10
Analytic solution BEM
10
amplitude
10
10
5 x
10
0
10 8
10
0.2 0 0.2 0.4 relative error [%] 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 0 1 2 3 4 5 x 6 7 8 9 10 coarse mesh dense mesh
6 4 relative error [%] 2 0 2 4 6 8 10 0 1 2 3 4 5 x 6 7 8 9 10
Figure 7.3: Discretization and solution for the FEM (left) and BEM (right)
261
10
10
0 0
10 10 10 10 10 10 10 10 10
1
0 2 4 6 8 10
10 BEM 10 10 10 FEM interface 10 10 10 10 10 1 2 3 4 5 6 7 8 8.5 9 10
1
10
FEMBEM coupling Analytic solution FEMBEM
BEMFEM coupling Analytic solution FEMBEM
BEM
FEM
4
interface
0
0 2 4 6 relative error [%] 8 10 12
0
14 12
8 8.5 9
10
interface
10 8 relative error [%] 6 4 2 0 2 4 FEM BEM interface
BEM
FEM 14 16 18 20 0 1 2 3 4 5 x 6 7 8 8.5 9 10
6 0
5 x
8 8.5 9
10
Figure 7.4: Discretization and solution for the coarse mesh (left) and dense (right) for the hybrid solution
262
7.2.2
Two squares one immersed in the other
Now let consider hybrid method application in case of a slightly more advanced example when the FE subregion is immersed in the BE region as it is shown in Fig. 7.5. The boundary conditions remain as in previous example (see Eq.(7.18) and Eq.(7.19)). Solutions will be controlled along two basic lines. One line AA is placed on the external boundary discretized by BEM and the second one BB is going through the subregion discretized by FEM. Thanks to those lines we can control the inuence of geometry on the solutions relative error. But rst, some remarks about a corner problem which arise in the Boundary Element applications. This simple example has two kind of corners: convex and concave ones. For the convex corners the function c(r) according to the 2 Eq.(1.77) becomes c(r) = 1 2 = 1 / = 3/4 and for the concave corners 2 3/2 the function c(r) = 1 2 = 1 2 = 1/4. The literature of the subject of the FEM/BEM coupling is really vast and it is enough to indicate the following positions [3,15,16,22,33,51,59] among the many others. The main goal of this section is to focus readers attention on the problems of geometry inuence on the model precision which is absolutely essential for DOT. Concluding this problem we can say that: Dimension of the BEM subregion has a great inuence on the solution precision (what is intuitively expected) but the scale of this problem must be numerically investigated (compare the results presented in Fig. 7.6 and in Fig. 7.7 where the error increases from 5.3% up to 9% on the interface). Interesting is that in the remain part of the region the error remains almost unchanged. Results are not sensitive when discretization is dense or coarse. This is a very good message from the point of view of numerical model creation
7.2. NUMERICAL EXAMPLES with the aid of hybrid method.
263
Now, let consider the same numerical examples, but the Finite Element sub region will be discretized by isoparametric quadratic triangular elements (see Fig. 7.8). This numerical experiments proved that kind of nite elements are not a very important regarding the precision of the solution.
264
10 10 8 8 6 6 B 4 4 2 2 0 A 0 2 4 6 8 A 0 10 0
10 10
2 1
BEMFEM coupling Analytic solution BEM
10
BEMFEM coupling Analytic solution FEM
10
10
10 interface FEM interface 10 10 BEM BEM 10 10
10
interface
FEM
interface
10
6 5
BEM
BEM
0
5 4 3 relative error [%]
10
10
0
8 6 4
10
interface
interface
relative error [%]
2 1
2 0 2 4 6 interface 8 BEM 10 0 1 FEM interface
BEM 0 1 2 BEM 3 0 1 FEM 2 3 4 5 x 6 7 8 9 10
BEM 2 3 4 5 x 6 7 8 9 10
Figure 7.5: Discretization of the domain under consideration and solution along the line AA for (left) and along the line BB (right) for coarse FEM mesh
265
10
10
B
4
4
2 A 0 2 4 6 8 A
10
10
1
10
FEMBEM coupling Analytic solution FEMBEM
BEMFEM coupling Analytic solution FEM
10
10 10
10
interface
4
FEM interface
10 10 10
interface
3
FEM interface
10
BEM 10
6
BEM 10 10
5
BEM
BEM
0
5 4
10
0
6
10
3 relative error [%] 2 1 0 1 2 BEM 1 2 3 4 FEM interface
interface
relative error [%] 2 interface 0
BEM
interface FEM
4 BEM BEM 6 0
3 0
5 x
10
5 x
10
Figure 7.6: Discretization of the domain under consideration and solution along the line AA (left) and along the line BB (right) for dense FEM mesh
266
10
6 5 4
relative error [%]
10 9 8 7 6 5 4 3 2 1 0 1 interface 2 3 4 5 0 0.5 1 2 BEM
interface
FEM 3 4 5 x 6 7 8
9 9.5 10
10
Figure 7.7: Discretisation of the domain under consideration and solution along the line BB for a dense FEM mesh in case of a small gap between BEM and FEM subregions
BEM
267
10
10
6 5 4 B B
6 5 4 B B
0 0
8 6
0 2 4 6 8 10
6
10
4 4 interface relative error [%] relative error [%] 2 0 2 BEM 4 6 8 10 0 FEM BEM 2 interface 0
interface FEM
4 interface 1 2 3 4 5 x 6 7 8 9 10 BEM BEM 6 0 1 2 3 4 5 x 6 7 8 9 10
Figure 7.8: Coarse (left) and dense (right) discretization, solution and relative error along the line BB
268
7.2.3
Concentric circles
Let again consider hybrid method application in case of subregions discretised by FEM immersed in BEM regions. Contrary to the previous section no corners will be considered. Such situation might be analogous to the cross section of a humans head when the inuence of the CSF layer is necessary to consider. Our attention will be focused on the interface between BEM and FEM subregions and a precision of the solution on the interface. The geometry of the region (the gap dimension) and the kind of discretisation inuence on the solution precision will be investigated, as it is a primary problem for the DOT numerical models. This time let consider point source located on the boundary for the steady state of the diusion model for light transport as it was already considered in the section 1.5.4. This section was devoted to the dierent ways of modelling the point sources and inuence of boundary discretization on the results. Let consider more dicult problem of the source placed on the boundary of the region and modelled by boundary conditions of the second kind (see for example Fig. 1.33). However, this time we will be interested not only with the solution on the boundary, but also inside of the region, which will be divided into two subregions: one discretized with the aid of the BEM but the second one with the FEM as it is presented in Fig. 7.9. Achieved results for the hybrid FEMBEM approximation was compared with the BEMonly solution which was treated as the "exact" one. Two cases were considered: coarse and dense (only FEM subregion) discretization. Discretization of the most external part of BEM subregion boundary was xed as it was constrained by the solution precision due to the Neumann boundary conditions modelling the point source). As we can see in the second row in the Fig. 7.9 in both cases we have got excellent conformity with the "exact" solution. But looking inside the region and particularly on the interface, the situation is not so simple. The relative error level reaches about 10%. Unfortunately, coupling procedure itself generates the error on the interface.
269
To illustrate this problem let consider two FEM subregions presented in Fig. 7.10. Using continuity and equilibrium conditions (Eq.(7.1)) both sub regions are combined in one region shown in Fig. 7.11. As we can see, in this gure such approach generates the error on the interface equal even to 8% but outside the interface the internal error dropped to zero. In order to reduce interface error, more attention should be devoted to Eq.(7.11). This example seems to be simple, but it is not, due to the dynamic range of the solution which is really wide. In the case presented in Fig. 7.9 solution drop to the values smaller than the maximal one of about six orders. That demand special attention when numerical model is constructed. Lots of numerical experiments (dierent kind of point sources and dierent kind of elements) justify the following remarks about the hybrid FEMBEM approach to the DOT forward problem: Dense discretization improves precision of the solution but not in such extend as we may expect. This is rather good message as we do not need to apply particularly dense discretization inside the region under consideration (consult Fig. 7.9) in some diusive problems. Unfortunately the "void" problem complicates numerical model and causes that discretization has much more inuence on the precision of the solution (see Fig. 7.12).
270
25 20 15 10 5 A 0 5 10 15 20 25 25 20 15 10 5
10 10 10 10 10 10 10 10 10
0

25
BE FE
20 15 10 5
BE FE
A
0 5
10 15 20
10
15
20
25
25 25 20 15 10 5
10
0
10
15
20
25
BEMFEM coupling FEMBEM FEMBEMinterface BEM BEM
BEMFEM coupling FEMBEM FEMBEMinterface BEM BEM
10 10 10 10 10 10 10 10
amplitude
amplitude 60 120 180 angle [deg.] 240 300 360
0
0
0
0
60
120
180 angle [deg.]
240
300
360
BEMFEM coupling FEMBEM BEM Interface FEM BEM Interface

amplitude
BEMFEM coupling FEMBEM BEM
10 10 10 10
10 10 10 10
Interface
amplitude
10 BEM 10 10 10
5
10 BEM 5 10 10 10
6
Interface
FEM
BEM
10 25
20
15
10
5 0 5 distance [mm] BEMFEM coupling
10
15
20
25
10 25
20
15
10
5 0 5 distance [mm] BEMFEM coupling
10
15
20
25
0 Interface 10
0 Interface 10 Interface
error [%]
error [%]
20 FEM 30 BEM 40 BEM Interface
20 FEM 30
40 BEM BEM
50 25
20
15
10
5 0 5 distance [m]
10
15
20
25
50 25
20
15
10
5 0 5 distance [mm]
10
15
20
25
Figure 7.9: Hybrid coarse discretization (top, left) and dense (top, right), in the second row there are the solutions on the perimeter of the region and along the interface, in the third row there are internal solutions and in the last row there are the relative errors for the internal solutions for Neumann boundary conditions case
271
25 20 15 10 5 0 5 10 15 20 25 25 20 15 10 5 0 5 10 15 20 25 n
25 20 15 10 5 0 n
5 ext 10 15 20 25 25 20 15 10 5 0 5 10 15
int 25
20
Figure 7.10: Two FE subregions discretization
272
25 20 15 10 5 0 5 10 15 20 25 25 20 15 10 5
10 10 10 10 10 10 10 10
0

25 20 15 10 5 0 5 10 15 20 0 5 10 15 20 25 25 25 20 15 10 5
10 FEMFEMinterface FEMFEM FEMinterface FEM 10 10 10 10 10 10 10
0
10
15
20
25
FEMFEMinterface FEMFEM FEMinterface FEM
0
0
45
90
135
180
225
270
315
360
0
0
45
90
135
180
225
270
315
360
FEMFEM coupling
FEMFEM coupling
10 10 10 10 10 10 10
10 10 FEMext FEMext 10 10 10 interface 10 10
FEMint
interface
3
FEMint interface FEMext FEM FEMFEM
interface
FEM FEMFEM 20 15 10 5 0 5 10 15 20 25
10 25
10 25
FEMext 20
15
10
10
15
20
25
FEMFEM coupling 1 interface 1 2 3 4 5 6 7 8 9 25 20 15 10 5 0 5 10 15 20 interface FEMext 7 6 FEMint 5 4 3 2 FEMext 1 0 1 25 interface 20 15 10 FEMext 0 9 8
FEMFEM coupling
FEMint
interface 5 0 5 10 15 20 25
25
Figure 7.11: Coarse discretization (left) and dense (right) of a multi nite element domain
FEMext
273
25 20 15 10 5 A 0 5 10 15 20 25 25 20 15 10 5
10
0
25 20 15 10 5 A A 0 5 10 15 20 0 5 10 15 20 25 25 25 20 15 10 5
10 FEMvoidFEM FEMFEMinterface
0
10
15
20
25
FEMvoidFEM
FEMvoidFEM FEMvoidFEM FEMFEMinterface
10
10
10
10
10
10
10
10
10
10
10
0
0
60
120
180
240
300
360
10
0
0
60
120
180
240
300
360
FEMvoidFEM FEMvoidFEM BEM
FEMvoidFEM FEMvoidFEM BEM
10
10
10
10
10
10
10
10
10
10
10
10
10 25
20
15
10
10
15
20
25
10 25
20
15
10
10
15
20
25
Figure 7.12: Solutions for the coarse discretization (left) and dense (right) of a nite element domain with a nonscattering gap
274
7.3
2D void comparison
The light propagation in scattering media, in presence of nonscattering regions (void problem), were considered in numerous papers [8,1113,41,6771]. So far those problems were mainly approximated by FEM (see for example [8, 67, 68] and in case of 2D space they were solved by Integral Equations approach described in [69, 71]. Chapter 4 is devoted to such problems but in 3D space, described by Integral Equations and approximated by BEM, the following work is accurate [69]. J. Rippol expressed the boundary conditions by Eq.(4.5) (see chapter 4) on the interface between diusive and nondiusive media named as a complete boundary conditions, whereas in the FEM model the approximate boundary conditions are employed in his work [69]. We will rather refer to those boundary conditions as to the P1 boundary conditions and diusive boundary conditions, respectively. To nd out the reason of dierences between BEM and FEM results (see Fig. 4.16), both kind of boundary conditions will be applied to FEM code in 2D space and results compared, hopefully allowing to come to some conclusions.
7.3. 2D VOID COMPARISON
275
7.3.1
Boundary conditions at diusive/nondiusive interfaces
Let summarize all boundary conditions used in the literature. BEM with P1 BC 1 2 (r) = Jn 2 (r) + ] [ RJ Jn (r ) G (|r r |)d2 2 (r ) + RU 2
2
(7.21)
BEM with diusive BC
1 2 (r) = Jn +
r) where Jn = D ( . n
2 (r )G (|r r |)d2
(7.22)
FEM with P1 BC (r) 2 (r) = D n 1 + [

2
RJ (r ) 2 (r ) D RU n
] G (|r r |)d2
2
(7.23) FEM with diusive BC
(r) 2 (r) = D n
1 +
2 (r )G (|r r |)d2
(7.24)
All of presented above boundary conditions will be used in FEM and BEM code and the results will be compared in the following sections.
276
7.3.2
P1 boundary conditions
The expressions for the boundary conditions in two dimensions are quite useful in validation process. (r) = Jn 2 (r) + ] [ 2 RJ 1 2 (r ) + Jn 2 (r ) G 2D (|r r |)d2 r 2 + RU
2
(7.25)
where: = and G 2D (|r r |) = cos() cos( ) [a0 +i(n0 /c)|rr |] V (r r ) e 2 |r r | (7.27)

1 2
2 RJ RU
(7.26)
For a refractive index match case following [69] we can assume, that RU = and RJ = 1. Then a coecient = 2.
After including number under integral sign and rearranging the Eq.(7.25) according to BEM standards we will get: 1 2 (r) 2 (r ) G 2D (|r r |)d2 = 2 1 = 2Jn 2 (r) + 2 Jn 2 (r ) G 2D (|r r |)d2 (7.28)
2
and Eq.(7.27) will be: 1 2D cos() cos( ) e[a0 +i(n0 /c)|rr |] G (|r r |) = V (r r ) |r r | 2
(7.29)
Using the idea of coupling FEM and BEM above, Eq.(7.28) and Eq.(7.29) were implemented into FEM code. Comparison of that solution with BEM voidBEM one is presented in the Fig.7.13. The maximum values and minimum ones t very well particularly for a dense discretization but still there are some discrepancies regarding the shape of both results.
7.3. 2D VOID COMPARISON

25 20 15 10 5 0 5 10 15 20 25 25 20 15 10 5
10
0
277
25 20 15 10 5 0 5 10 15 20
10
15
20
25
25 25 20 15 10 5
10
0
10
15
20
25
FEM/BEMvoid comparison BEMvoidBEM FEMvoidFEM
10
10
10
10
10
10
60 120 180 angle [deg.] 240 300 360
10
10
10
10
10
10
60
120
180 angle [deg.]
240
300
360
Figure 7.13: Comparison of the FEMvoidFEM solution with the BEM voidBEM solution in case of P1 nonlocal boundary conditions
7.3.3
Diusive boundary conditions
In order to check if any deviation between both methods is caused by the dierence in the boundary conditions, we apply the diusive (approximate according to [69]) boundary conditions as follows: 1 2 (r) = Jn + 2 (r )G 2D (|r r |)d2 = 2 1 1 = Jn + 2 (r ) G 2D (|r r |)d2 (7.30) 2
2 1 2D G (|r r |) is dened by Eq.(7.29). where
278
25 20 15 10 5 0 5 10 15 20 25 25 20 15 10 5
10 10 10 10 10 10 10 10 10
0
25 20 15 10 5 0 5 10 15 20 0 5 10 15 20 25 25 25 20 15 10 5
10 BEMvoidBEM FEMvoidFEM 10 10 10 10 10 10 10 10
0
10
15
20
25
FEM/BEMvoid comparison
60
120
180
240
300
360
60
120
180
240
300
360
Figure 7.14: Comparison of the FEMvoidFEM solution with the BEM voidBEM solution in case of diusive nonlocal boundary conditions
Chapter 8 Miscellaneous
8.1
Introduction to FEM mixed formulation
The mixed formulation for Finite Element Method is well known in the literature [89], but for optical tomography it is not. This approach will be implemented in this chapter, presenting some advantages in comparison to classical FEM, like a better precision or easier merging with Boundary Element Method. However, there are some diculties for commonly used optical parameters of the diusion equation. Those diculties and some of the methods how to avoid them will be discussed in the following sections.
8.2
Mixed formulation for Laplace equation
Let consider the following simple example: q = k, q = [ qx , q y ] , (8.1)
where k is a parameter describing material properties of the region under consideration. 279
280
CHAPTER 8. MISCELLANEOUS
The conservation equation can be written as: q= qx qy + =Q x y (8.2)
If the above equations are satised in and the boundary conditions: = on or qn = q n on q (8.3)
are obeyed then the problem is solved. Elimination of vector q is possible and simple substitution of Eq.(8.1) into Eq.(8.2) leads to: (k) + Q = 0 in (8.4)
with appropriate boundary conditions expressed in terms of or its gradient. If the discretized solutions start from this point, no further elimination of variables is possible, the formulation is irreducible [89]. On the other hand, if we are to start the discretization from Eq.(8.1) to Eq.(8.2) the formulation would be mixed. Stationarity of the variational principle given below is equivalent to the differential equations (8.1) and (8.2) together with the boundary conditions (8.3). 1 T 1 = q k q d ( q Q) d + ( qn q n ) d. (8.5) 2 q
8.2.1
Discretization of mixed forms
We assume that each of the unknowns is approximated in the usual manner by appropriate shape functions and corresponding unknown parameters. Thus: = Nq q q =q and = N = (8.6)
where q and stand for nodal (see Fig. 8.3a) or more generally element parameters (see Fig. 8.6) that have to be determined.
8.2. MIXED FORMULATION FOR LAPLACE EQUATION
281
are satised by the choice Assuming that the boundary conditions for = of the expansion, for Eq.(8.1) the weighted statement of the problem is: ) ( T d = 0 + Wq k1 q (8.7)
and for Eq.(8.2) the natural boundary conditions1 : T T Q) d W ( q W ( qn q n ) d = 0.

q
(8.8)
Selection of Wq = Nq and W = N (8.9)
will yield symmetric equations of the form [ ][ ] [ ] q A C f1 = CT 0 f2 where A=
(8.10)
1 NT q k Nq
d and C=
NT q Nq d
(8.11)
and f1 = 0 and f2 =
NT Q d
+
q
NT d. q
(8.12)
8.2.2
Robin boundary conditions
The boundary conditions of the third kind to be considered are given by: (k) n q = 0. The photon current is equal2 : J = k,
1 2
(8.13)
(8.14)
Natural boundary conditions means homogeneous Neumann boundary conditions. Please compare the Eq.(4.3) in section 4.2.1.
282
so the boundary conditions of the third kind can be rewritten: J n = + q. (8.15)
For particular node of the grid Eq.(8.15) can be written in the following form: Jx nx + Jy ny = + q. (8.16)
As a result the can be expressed as a linear combination of two other unknowns: = 1 1 (Jx nx + Jy ny ) q. (8.17)
The components of the unit outward normal (of unit length) are given by: [ ] [ ] 1 dy ( ) 1 dx( ) nx = , ny = . (8.18) J ( ) d J ( ) d The dierentials of the coordinates x( ) and y ( ) with respect to are given by: dx( ) dN1 ( ) dN2 ( ) = x1 + x2 , d d d (8.19) dy ( ) dN1 ( ) dN2 ( ) = y1 + y2 , d d d where 1 N1 ( ) = (1 ), 2 1 N2 ( ) = (1 + ). 2 (8.20)
Then Eq.(8.19) can be expressed as follows: 1 1 dx( ) = x1 + x2 , d 2 2 dy ( ) 1 1 = y1 + y2 . d 2 2
Because of the transformation of the variable from the boundary curve to the intrinsic coordinate , we must calculate the Jacobian of transformation: ( )2 ( )2 ( )2 ( )2 d dy ( ) y2 y1 dx( ) x2 x1 J ( ) = = + = + (8.21) d d d 2 2
8.2. MIXED FORMULATION FOR LAPLACE EQUATION Finally, we have got: y2 y1 nx = , (x2 x1 )2 + (y2 y1 )2 x2 x1 (x2 x1 )2 + (y2 y1 )2
283
ny =
. (8.22)
Those are the components of unit outward normal vector to the boundary line, but we need them in particular node. Due to the linear approximation left and right values are dierent. The coarse discretization the bigger dierences between them. That is why we have taken the average values to the computation. In order to check the algorithm the following example was considered: the outer surface of the region shown in Fig. 8.1 is immersed in a uid of temperature Tf and the inner surface temperature is constant and equal to T1 .
1.5
1.5
0.5
0.5
0 0 0.5 1 1.5 2
0 0 0.5 1 1.5 2
Figure 8.1: FE coarse mesh (left) and ne mesh (right) for the cylinder problem The numerical values used are: R1 = 1.0, R2 = 2.0, T1 = 10.0 and Tf = 6.0 with conductivity of k = 3.5 and a heat transfer coecient, h = 1.2. Newtons law states that the heat ux at the interface of the uid/solid surface is given by [14]: h h T = T + Tf . (8.23) n k k The analytical solution of the problem is: T = A ln r + B (8.24)
284
where for this particular boundary conditions A = 1.8592 and B = 10. As we have taken only one quarter of the cylindrical region into account, the natural boundary conditions ( T = 0) are imposed on the boundary y = 0 n and x = 0. The Fig. 8.2 shows the computed FE results compared to the equivalent analytical solution, where the coincidence depends on discretization, and is better for the ne mesh.
10
9.5
8.5
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
Figure 8.2: Results of calculation - solid line for analytical solution; triangle down at each data point for mixFE solution (ne mesh) and x-mark for the coarse solution
8.3
Mixed formulation for Diusion equation
Let consider a three nodes triangle with three degrees of freedom for a mixed formulation. As it is shown in [89] this type of triangle passes the patch test. We would like to know if such a triangle would be suitable for optical tomography.
8.3. MIXED FORMULATION FOR DIFFUSION EQUATION
285
8.3.1
Three nodes triangle with three degrees of freedom P1 triangle
At the beginning let consider three nodes triangle with three degrees of freedom in vertex nodes as it is shown in Fig. 8.3a. As a test problem the a)
y Jy 2 2
2
b)
10
Jx
Jy 0
Jy 0 1 Jx 0 1
Jx
x
1
0 0
10
Figure 8.3: a) The P1 triangle and b) region discretization by 16x16 layers square region was considered (see Fig. 8.3b), where the steady state diusion equation is valid: ( where k 2 =
. D
) 2 k 2 = 0
(8.25)
The following boundary conditions were applied: on the boundaries dened by the equations: y = 0 and y = 10, natural boundary conditions; on the boundary dened by the equations: x = 0 and x = 10, the Dirichlet boundary conditions: = 0 and = 10 respectively. The region discretized by the P1 triangle is shown in Fig. 8.3b. Such a nite element discretization produces good results for Laplace equation and diusion equation when a diusion coecient is not less than
286
0.1. At this value some oscillations in the solution occur (see Fig. 8.4a and Fig. 8.4b), which we are not able to remove by increasing the number of elements and nodes as it is shown in Fig. 8.5. a)
10
1
b)
10
1
10
10
10
10
10
10
10
10
10
10
10
10
Figure 8.4: a) Good behaviour of mixed formulation for diusion equation D=0.15 and b) Oscillation start to occur in the solution when D<0.1
The oscillation in the solution has a structural nature and in order to remove them we are forced to design the next more sophisticated nite element.
8.3.2
Three nodes triangle with the middle sides nodes
As a second nite element for the mixed formulation, the following nite element was considered as it is shown in Fig. 8.6. This time the variable and the variables Jx and Jy were separated geometrically. The region under consideration is shown in the Fig. 8.7, and as in the previous case, the same kind of boundary conditions were imposed. Points mark the vertex nodes when the middle side nodes with Jx and Jy variables are marked by squares. The results of calculation are presented in Fig. 8.8a and in Fig. 8.8b. As we can see, there is no oscillation in the solution in spite of the very low value of diusion parameter (D = 0.03).
8.3. MIXED FORMULATION FOR DIFFUSION EQUATION a)

10 10 10 10 10 10 10 10 10 10
1 0
287
b)
10 10 10 10 mixed 10 10 analytical 10 10 10 2 4 6 8 10 10
1 0
mixed
analytical
10
c)
10 10 10 10 10 10 10 10 10 10
1 0
d)
10 10 10 10 10 mixed 10 10 10 analytical 10 2 4 6 8 10 10
1 0
mixed
analytical
8 8
10
Figure 8.5: Solution for : a) 64 b) 128 c) 256 and d) 512 layers in x direction The solution is smooth without any osscillations but some discrepancies exist for both and Jx variables, which are as big as 40% (see Fig. 8.8). Even much more dense discretization is not able to improve the results. So, another nite element is needed which can remove disadvantages of both already considered triangles.
8.3.3
Three nodes triangle with a bubble function at the barycenter P1 + triangle
The P1 triangular element (Fig. 8.3a) is the simplest one and works perfectly well for the problems described by the Laplace equation. However for
288
y 2 2 Jy
0
J y2 Jx
0 2
Jx
Jy 0 0 0
1 Jx 1 1 x
Figure 8.6: Three nodes triangle with the middle side nodes for a mixed formulation
1.5 1 0.5 0 0 1 2 3 4 5 6 7 8 9 10
Figure 8.7: Region under consideration
the diusion equation and particularly for the data used in Diuse Optical Tomography problems, the oscillations occurred (see Fig. 8.4b). Such oscillations poses a structural nature and even very dense discretization is not able to remove them, as it was proved by numerical experiment presented in Fig. 8.5. Therefore, for Optical Tomography we need a new triangle, which can provide stable results for a diusion equation approximating Optical Tomography problems. Let consider P1 + triangle which diers from P1 triangle by the adjunction of a bubble function 3 (see Fig. 8.9). The bubble function taking the value of 1 at the barycenter and zero on the
8.3. MIXED FORMULATION FOR DIFFUSION EQUATION a)

10
2
289
b)
10
0
10
10
10
10
10
1
10
10
2
10
10
3
10
10
10
10
10
10
10
10
10
10
10
10
Figure 8.8: a) Solution of and b) solution of Jx for the nite element from Fig. 8.6 boundary of the element.
y Jy 2 2
2
Jx
3 Jy 0
0
Jy 0 1 Jx 0 1
Jx
x
1
Figure 8.9: Three nodes triangle with a bubble function at the barycenter P1 + triangle
= N0 0 + N1 1 + N2 2 + N0 N1 N2 3 where 3 is the bubble function.
(8.26)
290
The basis functions in triangular coordinates [89] are dened: N0 = L0 N1 = L1 N2 = L2 N3 = L0 L1 L2 (8.27)
0 x+c0 y 1 x+c1 y 2 x+c2 y where L0 = a0 +b2 , L1 = a1 +b2 , L2 = a2 +b2 , and the coecients: b0 = (y1 y2 ), c0 = (x2 x1 ), b1 = (y2 y0 ), c1 = (x0 x2 ), b2 = (y0 y1 ), c2 = (x1 x0 ), = 0.5(x1 y2 + x0 y1 + x2 y0 y0 x1 y2 x0 y1 x2 ).
The same region with the same boundary conditions was discretized by the P1 + triangle and results are compared with P1 triangle as we can see in Fig. 8.10. The results are smooth without any trace of the oscillations and, what is more, are very precise (relative error is less than 3%) even for a very coarse discretization. a)
10 10 10 10 10 10 10 10 10 10
1 0
b)
3 2.5 2
P1
1.5 1 0.5
P1+
0 0.5 0
10
10
Figure 8.10: a) Comparison of the solutions for P1 and P1 + triangles discretization. b) Relative error distribution of the solution for coarse (64 layers) discretization by P1 + triangles
Conclusions and future perspective

Some problems of numerical simulation with the aid of BEM for Diusive Optical Tomography and Industrial Tomography were presented in this book. Both kinds of tomographies are so fascinating that in the future I will try to continue this work together with my PhD students. Keeping in my mind that two, among the many others medical or industrial applications are particularly important. The rst one is monitoring brain haemorrhage of the new born babes (this is the only one harmless technology presently known), the second is the breast cancer (DOT can serve as supporting technology). That is why the future works will concentrate on the following subjects:
1. Incorporate special boundary elements (gap elements) in order to create more eectively clear CSF layers or innite ones to make the breast model more precise (the boundary conditions between the breast and the chest wall are unknown). 2. Further development of mixed formulation in FEM to make hybridization of BEM with FEM easier. 3. Development of methods allowing solution of "Large scale" problems. The multilayer neonatal head model belong to the "Large scale" problem. In order to consider the inverse problem the forward problem have to be solved more eectively than it is done nowadays. I strongly 291
292
CHAPTER 8. MISCELLANEOUS believe that the following directions of research will lead to the solution of the "Large scale" problem. (a) Application of Discrete Wavelet Transform in order to accelerate the forward problem solution. (b) Research on Multipole approach to the BEM. (c) Using method of Region decomposition, particularly that multi layered head model seems to be ideal for such approach
All those methods are already well known, but it is not obvious which one will be the best for DOT or industrial tomography applications. As we can see the plan of the future work is very ambitious and shows how much work have to be done in order to state that BEM is able to provide eectively precise and reliable results leading to the image building algorithms.
Bibliography
[1] Kazuhisa Abe, Kazuhiro Koro, and Kayoko Itami. An hhierarchical Galerkin BEM using Haar wavelets. Engineering Analysis with Boundary Elements, 25:581591, 2001. [2] M. Abramowitz and I. A. Stegun. Handbook of mathematical functions with formulas, graphs and mathematical tables. John Wiley, New York, 1973. [3] M. H. Aliabadi. The Boundary Element Method; Volume 2; Applications in Solids and Structures. John Wiley & Sons, LTD, 2002. [4] M. H. Aliabadi and W.S. Hall. The regularizing transformation integration method for boundary element kernels. Comparison with series expansion and weighted Gaussian integration methods. Engineering Analysis with Boundary Elements, 6(2):6670, 1989. [5] K. Amaratunga. A waveletbased approach for compressing kernel data in largescale simulations of 3D integral problems. Computing in Science and Engineering, pages 3445, July/August 2000. [6] S. R. Arridge. Optical tomography in medical imaging. Inverse Problems, 15(2):R41R93, 1999. [7] S. R. Arridge, M. Cope, and D. T. Delpy. Theoretical basis for the determination of optical pathlengths in tissue: Temporal and frequency analysis. Physics in Medicine and Biology, 37:15311560, 1992. 293
294
BIBLIOGRAPHY
[8] S. R. Arridge, H. Dehghani, M. Schweiger, and E. Okada. The nite element model for the propagation of light in scaterring media: A direct method for domains with nonscattering regions. Medical Physics, 27(1):252264, 2000. [9] S. R. Arridge and J. C. Hebden. Optical imaging in medicine: II. Modelling and reconstruction. Physics in Medicine and Biology, 42:841853, 1997. [10] S. R. Arridge, M. Schweiger, M. Hiraoka, and D. T. Delpy. A nite element approach for modeling photon transport in tissue. Medical Physics, 20(2):299309, 1993. [11] G. Bal. Particle transport through scattering regions with clear layers and inclusions. J. Comp. Phys., 180:659685, 2002. http://www.columbia.edu/gb2030/pubs.html. [12] G. Bal. Transport through diusive and nondiusive regions, embeded objects, and clear layers. SIAM J. Appl. Math, 62:16771697, 2002. http://www.columbia.edu/gb2030/pubs.html. [13] G. Bal and K. Ren. Generalized diusion model in optical tomography with clear layers. http://www.columbia.edu/gb2030/PAPERS/GDMLocal.ps. [14] A.A. Becker. The Boundary Element Method in Engineering. A complete course. McGrawHill Book Company, 1992. [15] G. Beer. Programming the Boundary Element Method. An Introduction for Engineers. John Wiley & Sons, 2001. [16] G. Beer and J.O. Watson. Introduction to Finite and Boundary Element Methods for Engineers. John Wiley & Sons, 1992. [17] P. Berowski, J. Sikora, S. Wjtowicz, and S. F. Filipowicz. Monitoring of dehumidication process of the wall using 3D EIT system, 2005. Proc. of 15th Conference on the Computation of Electromagnetic Fields COMPUMAG05, Shenyang, China, paper PD112 CSY0114.
BIBLIOGRAPHY [18] P. Bettess. Innite Elements. Penshaw Press, 1992.
295
[19] D. A. Boas, D. H. Brooks, E. L. Miller, C. A. DiMarzio, M. Kilmer, R. J. Gaudette, and Q. Zhang. Imaging the body with diuse optical tomography. IEEE Sig. Proc. Magazine, 18(6):5775, 2001. [20] C. Bond. A new integration method providing the accuracy of gausslegandre with error estimation capability. http://www.crbond.com/papers/gbint.pdf. [21] M. Bonnet. Boundary Integral Equation Methods for Solid and Fluids. John Wiley & Sons, 1995. [22] C.P. Bradley, G.M. Harris, and A.J. Pullan. The computational performance of a highorder coupled FEM/BEM procedure in elektropotential problems. IEEE Transactions on Biomedical Engineering, 48(11):1238 1250, November 2001. [23] C.A. Brebbia and J. Dominguez. Boundary Elements. An Introductory Course. McGraw-Hill Book Company, 1989. [24] A. Buchau, C.J. Huber, W. Rieger, and W.M. Rucker. Fast BEM computations with the Adaptive Multilevel Fast Multipole Method. IEEE Transactions on Magnetics, 36(4):680684, 2000. [25] M.V.K. Chari and S.J. Salon. Numerical Methods in Electromagnetism. Academic Press, 2000. [26] Ke Chen. Discrete wavelet transforms accelerated sparse preconditioners for dense boundary element systems. Electronic Transactions on Numerical Analysis, 8:138153, 1999. [27] J.C. de Munck. A linear discretization of the volume conductor boundary integral equation using analitically integrated elements. IEEE Transactions on Biomedical Engineering, 39:986990, 1992. [28] J.C. de Munck, T.J.C. Faes, and R.M. Heethaar. The boundary element method in the forward and inverse problem of electrical impedance tomo-
296
BIBLIOGRAPHY graphy. IEEE Transactions on Biomedical Engineering, 47(6):792800, 2000.
[29] H. Dehghani, B. Brooksby, K. Viswanath, B.W. Pogue, and K. D. Paulsen. The eects of internal refractive index variation in nearinfrared optical tomography: a nite element modelling approach. Physics in Medicine and Biology, 48:27132727, 2003. [30] F.M.E. Duddeck. Fourier BEM. SpringerVerlag, 2002. Lecture Notes in Applied Mechanics, Volume 5. [31] J. M. Ford. Waveletbased preconditioning of dense linear systems. PhD Thesis, University of Liverpool, 2001. [32] J.H.M. Frijns, S. L. de Snoo, and R. Schoonhoven. Improving the accuracy of the boundary element method by the use of secondorder interpolation functions. IEEE Transactions on Biomedical Engineering, 47(10):13361346, 2000. [33] I. Guven and E. Madenci. Transient heat conducting analysis in a piecewise homogeneous domain by a coupled boundary and nite element method. Int. Journ. for Numerical Meth. in Engineering, 56:351380, 2003. [34] W. Hackbusch and Z.P. Nowak. On the fast matrix multiplication in the boundary element method by panel clustering. Numerische Mathematik, 54:463491, 1989. [35] W.S. Hall. The Boundary Element Method. Kluwer Academic Publishers, 1994. [36] R.F. Harrington. Field computation by Moment Methods. FL: Krieger, 1968. Malabar. [37] D. Hawysz and E. M. Sevick-Muraca. Developments towards diagnostic breast cancer imaging using nearinfrared optical measurements and ourescent contrast agents. Neoplasia, 2(5):388417, 2000.
BIBLIOGRAPHY
297
[38] K. Hayami. A Projection Transformation Method for Nearly Singular Surface Boundary Element Integrals. SpringerVerlag, 1992. [39] J. C. Hebden, S. R. Arridge, and D. T. Delpy. Optical imaging in medicine: I. Experimental techniques. Physics in Medicine and Biology, 42:825840, 1997. [40] J. Heino, S. R. Arridge, J. Sikora, and E. Somersalo. Anisotropic eects in highly scattering media. Physical Review, E:03190818, 2003. [41] N. Hyvnen. Analysis of Optical Tomography with NonScattering Regions. Msc Thesis, Helsinki University of Technology, Institute of Mathematics, 2000. [42] M. Itagaki and C.A. Brebia. Boundary element method applied to neutron diusion problems. In Proc. 10th Int. Conf. BEM, Southampton Univ. SpringerVerlag, 1988. [43] P. Jaboski. Metoda elementw brzegowych w analizie pola elektromagnetycznego. Wydawnictwo Politechniki Czstochowskiej, 2003. in Polish. [44] S.A. Jankins and J.R. Bowler. Numerical evaluation of singular matrix elements in three dimensions. IEEE Transactions on Magnetics, 27:44384444, November 1991. [45] A. Jaworski. Metoda Elementw Brzegowych: Zagadnienia potencjalne. Ocyna Wydawnicza Politechniki Warszawskiej, 2000. in Polish. [46] Jianming Jin. The Finite Element Method in Electromagnetics. John Wiley & Sons, 1993. [47] J.H. Kane, A. Gupta, and S. Saigal. Reusable intrinsic sample point (RISP) algorithm for the ecient numerical integration of three dimensional curved boundary elements. Int. Journ. for Numerical Meth. in Engineering, 28:16611676, 1989.
298
BIBLIOGRAPHY
[48] L.V. Kantorovich and V.I. Krylov. Approximate Methods of Higher Analysis. John Wiley & Sons, 1964. New York, Translated from Russian by C.D. Benster. [49] A. Krawczyk and J.A. Tegopoulos. Numerical modelling of eddy currents. Clarendon Press, 1993. Oxford. [50] E. Kurgan. Analiza pola magnetostatycznego w rodowisku niejednorodnym metod elementw brzegowych. Uczelniane Wydawnictwa NaukowoDydaktyczne, 1999. Rozprawy Monograe 81, in Polish. [51] S. Kurz and S. Russenschuck. Accurate calculation of magnetic elds in the end regions of superconducting accelerator magnets using the BEMFEM coupling method, 1999. Proceedings of the 1999 Particle Accelerator Conference, New York. [52] Y. Li, Y. Yan, and G. Song. Fast wavelet transform based on binary partition techniques. Journal of Xidian University (Natural Science), 33(1):107110, 2006. [53] Tao Lu and D. O. Yevick. Boundary element analysis of dielectric waveguides. Journal of the Optical Society of America, 19(6):11971206, June 2002. [54] E. Majchrzak. Metoda Elementw Brzegowych w Przepywie Ciepa. Wydawnictwo Politechniki Czstochowskiej, 2001. Seria Monograe Nr. 76, in Polish. [55] J.W.H. Meijs, O.W. Weiner, M.J. Peters, and A. van Oosterom. On the numerical accuracy of the boundary element method. IEEE Transactions on Biomedical Engineering, 36:10381049, 1989. [56] R. Model, M. Orlt, M. Walzel, and R. Hnlich. Reconstruction algorithm for nearinfrared imaging in turbid media by means of timedomain data. Journal of the Optical Society of America A: Optics Image Science and Vision, 14(1):313324, 1997.
BIBLIOGRAPHY
299
[57] G. Muller, B. Chance, R. Alfano, S. Arridge, J. Beuthan, E. Gratton, M. Kaschke, B. Masters, S. Svanberg, and P. van der Zee, editors. The forward and inverse problems in timeresolved infrared imaging, 1993. Proc. SPIE, Medical Optical Tomography: Functional Imaging and Monitoring. [58] M.M. Ney. Methods of moments as applied to electromagnetics problems. IEEE Transactions on Macrowave Theory and Technique, 33(10):972980, 1985. [59] O.K. Panagouli and P.D. Panagiotopulos. The FEM and BEM for fractal boundaries and interfaces. Applications to unilateral problems. Computers and Structures, 64(14):329339, 1997. [60] Seong Jin Park and Tai Hun Kwon. Sensitivity analysis formulation for 3D conduction heat transfer with complex geometries using a boundary element method. Int. Journ. for Numerical Meth. in Engineering, 39:28372862, 1996. [61] K. D. Paulsen and H. Jiang. Spatiallyvarying optical property reconstruction using a nite element diusion equation approximation. Medical Physics, 22(6):691701, 1995. [62] H.L.G. Pina, J.L.M. Fernandes, and C.A. Brebia. Some numerical integration formulae over triangles and squares with a 1/R singularity. Appl. Math. Modelling, 5:209211, 1981. [63] K.G. Prasad, J.H. Kane, D.E. Keyes, and C. Balakrishna. Preconditioned Krylov solvers for BEA. Int. Journ. for Numerical Meth. in Engineering, 37:16511672, June 1994. [64] W.H. Press, S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery. Numerical Recipes in C. The art of scientic computing. Cambridge University Press, second edition, 1992. [65] Youssef F. Rashed. Boundary Element Formulations for Thick Plates. WIT Press, 2000.
300
BIBLIOGRAPHY
[66] J.D. Riley. Modelling of light propagation in mixed diusing and non diusing domains. PhD Thesis, University College London, Department of Computer Science, 2003. [67] J.D. Riley, S. R. Arridge, Y. Chrysanthou, H. Dehghani, E.M.C. Hillman, and M. Schweiger. Radiosity diusion model in 3D. In Stefan Andersson-Engels and Michael F. Kaschke, editors, Photon Migration, Optical Coherence Tomography, and Microscopy, pages 153164, 18 21 June 2001. Proceedings of SPIE. [68] J.D. Riley, H. Dehghani, M. Schweiger, S. R. Arridge, J. Ripoll, and M. Nieto-Vesperinas. 3D optical tomography in the presence of void regions. OPTICS EXPRESS, 7(13), 2000. [69] J. Ripoll. Light diusion in turbid media with biomedical application. PhD Thesis, University of Madrid, 2000. [70] J. Ripoll, M. Nieto-Vesperinas, and S. R. Arridge. Eect of roughness in nondiusive regions within diusive media. Journal of the Optical Society of America A: Optics Image Science and Vision, 18(4):940947, 2001. [71] J. Ripoll, M. Nieto-Vesperinas, S. R. Arridge, and H. Dehghani. Boundary conditions for light propagation in diusive media with nonscattering regions. Journal of the Optical Society of America A: Optics Image Science and Vision, 17(9):16711681, 2000. [72] V. Rokhlin. Rapid solution of integral equations of classical potential theory. J. Comp. Physics, (60):187207, 1983. [73] Y. Saad and M. H. Schultz. GMRES: A generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM J. Sci. Statist. Comput., 7(3):856869, April 1986. [74] M. Schweiger and S. R. Arridge. The nite element model for the propagation of light in scattering media: Frequency domain case. Medical Physics, 24(6):895902, 1997.
BIBLIOGRAPHY
301
[75] M. Schweiger, S. R. Arridge, M. Hiraoka, and D. T. Delpy. The nite element model for the propagation of light in scattering media: Boundary and source conditions. Medical Physics, 22(11):17791792, 1995. [76] Yin sheing Li and Tatsuo Obata. Some improvements of accuracy and eciency in three dimensional direct boundary element method. Int. Journ. for Numerical Meth. in Engineering, pages 14511464, 1992. [77] J. Sikora, J. Riley, S. R. Arridge, A. D. Zacharopoulos, and J. Ripoll. Analysis of light propagation in diusive media with nonscattering regions using 3D BEM. In Proceedings of XIIth International Symposium on Theoretical Electrical Engineering ISTET03, pages 511514, 2003. Warsaw, Poland, July 69. [78] J. Sikora, A. D. Zacharopoulos, A. Douiri, M. Schweiger, L. Horesh, S. R. Arridge, and J. Ripoll. Diuse Photon Propagation in Multilaiered Geometries. Physics in Medicine and Biology, 51(3):497516, 2006. [79] B.J. Strait. Applications of the Method of Moments to Electromagnetics. FL: SCEEE, 1980. St. Cloud. [80] M. Trlep, L. Skerget, B. Krea, and B. Hribernik. Hybrid nite elementboundary element method for nonlinear electromagnetic problems. IEEE Transactions on Magnetics, 31(3):13801383, May 1995. [81] Loic Vanel, P.A. Lemieux, and D.J. Durian. Diusing wave spectroscopy for arbitrary geometries: Numerical analysis by boundary element method. Applied Optics, 40(24):13361346, 2001. [82] L. C. Wrobel. The Boundary Element Method; Volume 1; Applications in ThermoFluids and Acoustics. John Wiley & Sons, LTD, 2002. [83] A. Yodh and B. Chance. Spectroscopy and imaging with diusing light. Phys. Today, pages 3840, 1995. [84] Point in triangle test. http:// www.blackpawn.com/ texts/ pointinpoly/ default.html.
302
BIBLIOGRAPHY
[85] A. Zacharopoulos, S.R. Arridge, O. Dorn, V. Kolehmainen, and J. Sikora. 3D Shape Reconstruction in Optical Tomography Using Spherical Harmonics and BEM. Journal of Electromagnetic Waves and Applications, (13):18271836, 2006. [86] A. Zacharopoulos, S.R. Arridge, O. Dorn, V. Kolehmainen, and J. Sikora. Three-dimensional reconstruction of shape and picewise constant region values for optical tomography using spherical harmonic parametrization and a boundary element method BEM. Inverse Problems, pages 124, 2006. [87] A.D. Zacharopoulos and S. R. Arridge. Surface mesh creation for 3D objects with sphere topology by homeomorphism on a unit sphere. In Proceedings of IVth International Workshop: Computational Problems of Electrical Engineering, pages 211215, 2002. Zakopane, Poland, September 25. [88] O.C. Zienkiewicz, D.W. Kelly, and P. Bettess. The coupling of the nite elements method and boundary solution procedures. Int. Journ. for Numerical Meth. in Engineering, pages 355375, 1977. [89] O.C. Zienkiewicz and R.L. Taylor. The Finite Element Method. McGraw-Hill, 1993. 4th edn., New York.

Numerical Methods Book PDF

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Numerical Methods Book PDF

Enviado por

Direitos autorais:

Formatos disponíveis

Boundary Element Method for Impedance and Optical Tomography

8 222 217 104 103 79 216 221 100

218 223 219 231 225

116 115 118 114 232 18 119 117 19 17

1 Twodimensional potential problems

1.2.2 1.2.3 1.2.4 1.2.5

1.2.6 1.2.7 1.2.8 1.2.9

1.3 Diusion equation . . . . . . . . . . . . . . . . . . . . . . . . . 53 3

4 1.3.1 1.3.2 1.3.3

CONTENTS Treatment of singularity . . . . . . . . . . . . . . . . . 55 Internal points calculation . . . . . . . . . . . . . . . . 56 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 57

2 Three-dimensional potential problems

2.5 Firstorder interpolation functions

2.7 Treatment of Boundary Conditions . . . . . . . . . . . . . . . 130

6 2.7.1 2.7.2 2.7.3 2.7.4

3 Diusion model for light transport in the frequency domain

3.5 Multilayered model of the neonatal head . . . . . . . . . . . . 182

4.3 The Form Factor . . . . . . . . . . . . . . . . . . . . . . . . . 194

4.8 Nuuttis 2D test example

CONTENTS 4.8.1 4.8.2 4.8.3 4.8.4

4.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228

5 BEM formulation for thin layers

6 Wavelet based techniques for CPU time reduction

14 R distance between load and observation points.

18 work tries to prove this opinion.

Chapter 1 Twodimensional potential problems

of the same accu-

1.2. LAPLACE EQUATION

enclosed solution domain

Figure 1.1: Region in local coordinate system

CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS

Division of the boundary into constant boundary elements

(r ) d(r ) (1.2) G(|r r |) n

1.2. LAPLACE EQUATION

j=5 domain j=2

CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS

1.2. LAPLACE EQUATION

Numerical integration of the kernels

CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS

G(|r r |) J ( )d = n G(|r r |)J ( )d (1.16)

where M is the total number of elements.

1.2. LAPLACE EQUATION

The entries of the matrices are: Ai,j (r, r ) =

G(|r r |) J ( )d n (1.20) G(|r r |)J ( )d

CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS

where nx and ny are dened by Eq.(1.10).

1.2. LAPLACE EQUATION Table 1.1: Onedimensional Gaussian quadrature

For more complex cases a special logarithmically weighted numerical inte-

CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS

g=2 g=1 g=0

Figure 1.5: Numerical integration for constant boundary elements

gration formula can be used [21, 43, 45, 50, 54]:

Division of the boundary into linear boundary elements

1.2. LAPLACE EQUATION Table 1.2: Onedimensional logarithmic Gaussian quadrature

Nk ( )xk = N0 ( )x0 + N1 ( )x1 (1.29)

Nk ( )yk = N0 ( )y0 + N1 ( )y1

CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS

Figure 1.6: Boundary discretization for a linear elements

Figure 1.7: Local coordinates of linear element number j

Numerical integration of the kernels

CHAPTER 1. TWODIMENSIONAL POTENTIAL PROBLEMS

G(|r r |) (j ) k (r )Nk ( )J ( )d = n k=0

G(|r r |) Nk ( )J ( )d = n G(|r r |)Nk ( )J ( )d (1.36)

ai,k (r, r )(r )k =