Professional Documents
Culture Documents
Instructor
Dr. Raymond Rumpf
(915) 747‐6958
rcrumpf@utep.edu
EE 5337
Computational Electromagnetics (CEM)
Lecture #26
Introduction to
Variational Methods
These notes may contain copyrighted material obtained under fair use rules. Distribution of these materials is strictly prohibited
Lecture 26 Slide 1
Outline
• Overview
• Galerkin Method
• Example
• Finite Element Method
• Method of Moments
• Other Worthy Methods
– Boundary element method
– Spectral domain method
Lecture 26 Slide 2
1
7/21/2017
Overview
Lecture 26 Slide 3
Classification of Variational Methods
Finite Element Method Boundary Element Method
• Utilizes a volume mesh • Utilizes a conformal surface mesh
• Matrices are sparse • Matrices are full
• Requires boundary conditions • Good for devices described efficiently
by their surfaces
• Does not require boundary conditions
Method of Moments Spectral Domain Method
• Typically makes a PEC approximation • BEM/MoM in Fourier‐Space
• Utilizes a conformal surface mesh • Matrices are full
• Can be 1D for thin wire structures • Excellent for periodic structures
• Matrices are full • Does not require boundary conditions
• Good for devices described efficiently
by their surfaces
• Does not require boundary conditions
Lecture 26 Slide 4
2
7/21/2017
Discretization
Variational Method
Governing Equation Matrix Equation
Weighted Residual
Method
Both the variational method and the method of weighted residuals can be used
to write a governing equation in matrix form.
Both approaches yield exactly the same matrices.
The Galerkin method is the most popular special case of weighted residual
methods.
Lecture 26 Slide 5
Galerkin Method
Boris Grigoryevich Galerkin
1871 ‐ 1945
Lecture 26 Slide 6
3
7/21/2017
Linear Equations
Consider the following linear homogeneous equation.
L f x g x L Linear operation
f x unknown solution
g x known driving function
The linear operator L[] has the following properties
L f1 x f 2 x L f1 x L f 2 x
L af x aL f x
L1 L2 f x L2 L1 f x
Lecture 26 Slide 7
Inner Product
An inner product is a scalar quantity that provides a measure of
similarity between two functions.
f x , g x inner product between f x and g x
f,g 0 f and g are orthogonal It is common to scale functions such that
f , g small number f and g are very different f , f * 1
f , g big number f and g are very similar
An appropriate inner product must satisfy
f , g g, f
f g, h f , h g, h
0 f 0
f, f*
0 f 0
We will use the following inner product: f z , g z f z g z dz
z
Lecture 26 Slide 8
4
7/21/2017
Expansion Into a Set of Basis Functions
Let f(x) be expanded into a set of basis functions.
f x unknown function
f x an vn x an coefficient of n th basis function
n
vn x n th basis function
Subdomain Basis Functions Entire Domain Basis Functions
We choose the basis functions with two considerations: (1) ease of calculations, and
(2) minimize how many are needed in the expansion to accurately portray the field.
Lecture 26 Slide 9
Linear Equation in Terms of Basis Functions
First we substitute the expansion into the original linear equation.
f x an vn x
n
L f x g x
Using the properties of linear operations, we get
L f x g x We try to choose vn(x) so that
L[vn(x)] is easy to calculate and
efficiently represents f(x).
L an vn x g x
n
L an vn x g x a L v x g x
n
n n
n
Lecture 26 Slide 10
5
7/21/2017
Method of Weighted Residuals (1 of 2)
Similar to how we choose a set of basis functions, we choose another
set of weighting functions.
wn x
We start with our linear inhomogeneous equation and calculate the
inner product with wm(x) of both sides. Here we “test” both sides
with wm(x).
a L v x g x
n
n n
wm x , a L v x
n
n n wm x , g x
an
n wm x , L vn x wm x , g x
Lecture 26 Slide 11
Method of Weighted Residuals (2 of 2)
This equation can be written in matrix form as
a n
n wm x , L vn x wm x , g x zmn an g m
w1 , L v1 w1 , L v2
w , L v w2 , L v2
zmn 2 1
wM , L vN
a1 w1 , g
a
w2 , g
n 1
a m
g
aN wM , g
Lecture 26 Slide 12
6
7/21/2017
Galerkin Method
The Galerkin method is the method of weighted residuals, but the
weighting functions are made to be the same as the basis functions.
wm x vm x
The matrix equation becomes
an
n vm x , L vn x vm x , g x zmn an g m
v1 , L v1 v1 , L v2 a1 v1 , g
a
v , L v v2 , L v2 v ,g
zmn 2 1 an 1 g m 2
vM , L v N
aN vM , g
Lecture 26 Slide 13
Summary of the Galerkin Method
The Galerkin method can be used to find the solution to any linear inhomogeneous
L f g
equation.
Step 1 – Expand the unknown function into a set of basis functions.
f an vn L an vn g a L v g
n n
n n n
Step 2 – Test both sides of the equation with the basis functions using an inner
product
vm , an L vn vm , g a n vm , L vn vm , g
n n
Step 3 – Form a matrix equation
zmn an g m
a1 v1 , g
v1 , L v1 v1 , L v2 a
an 1 v ,g
zmn v2 , L v1 v2 , L v2 g m 2
aN vM , g
Lecture 26 Slide 14
7
7/21/2017
Example
Lecture 26 Slide 15
Governing Equation and Its Solution
Governing Equation
We will apply the Galerkin method to solve the following differential
equation.
d2 f
2
1 4 x2 f 0 f 1 0 0 x 1
dx
Simple Analytical Solution
This is a simple boundary value problem with the following solution.
5x x2 x4
f x
6 2 3
We wish to solve this using the Galerkin method.
Lecture 26 Slide 16
8
7/21/2017
Choose Basis Functions
Basis Functions
For this problem, it will be convenient to choose as basis functions:
vn x x n 1
Expansion of the Function into the Basis
We expand our function into this set of basis functions as
N
f x an x x n 1
n 1
Lecture 26 Slide 17
Choose Testing Functions
Testing Functions
The Galerkin method uses the same function for basis and testing.
wn vn x x n 1
Lecture 26 Slide 18
9
7/21/2017
Form of Final Solution
Recall that we are converting a linear equation into a matrix equation
according to
L f g zmn an g m
v1 , L v1 v1 , L v2
zmn v2 , L v1 v2 , L v2
a1 v1 , g
a
v ,g
an 1 g m 2
aN vM , g
vm , L vn
To do this, we need to evaluate and .
vm , g
Lecture 26 Slide 19
First Inner Product
vm , L vn vm L vn dx
mn
zmn vm , L vn
x
m n 1
1 2
x x m 1
dxd 2 x x n1 dx
0
1
x x m 1 n n 1 x n 1 dx
0
1
n n 1 x n x m n dx
0
1
x n 1 x m n 1
n n 1
n 1 m n 1 0
1 1
n n 1
n 1 m n 1
m n 1 n 1
n n 1
n 1 m n 1 n 1 m n 1
m
n n 1
n 1 m n 1
mn
m n 1
Lecture 26 Slide 20
10
7/21/2017
Second Inner Product
vm , g vm gdx
x
m 3m 8
1
x x m1 1 4 x 2 dx
g m vm , g
0
2 m 2 m 4
1
x 4 x3 x m1 4 x m 3 dx
0
1
x2 x m 2 4 x m 4
x4
2 m 2 m4 0
1 1 4
1
2 m2 m4
3 1 4
2 m2 m4
3 m 2 m 4 2 m 4 8 m 2
2 m 2 m 4 2 m 2 m 4 2 m 2 m 4
3 m 2 m 4 2 m 4 8 m 2
2 m 2 m 4
3 m 2 6m 8 10m 24
2 m 2 m 4
3m2 8m
2 m 2 m 4
m 3m 8
2 m 2 m 4
Lecture 26 Slide 21
Try N=1
For N=1, our matrix equation isn’t even a matrix equation.
mn
z11 a1 g1 z11a1 g1 zmn
m n 1
Applying our equations for zmn and gm, we get m 3m 8
gm
1 1 1 2 m 2 m 4
z11
1 1 1 3
1 3 1 8 11
g1
2 1 2 1 4 30
The coefficient is
g1 11 30 11
a1
z11 13 10
Finally, the solution for N=1 is
11x 11x 2
f x a1 x x 2 Not correct. Need larger N.
10 10
Lecture 26 Slide 22
11
7/21/2017
Try N=2
For N=2, our matrix equation is zmn
mn
z11 z12 a1 g1 m n 1
z m 3m 8
21 z22 a2 g 2 gm
2 m 2 m 4
Applying our equations for zmn and gm, we get
1 1 1 1 2 1 2 1 1 22 4
z11 z z z22
1 1 1 3 12 1 2 1 2 21 2 1 1 2 2 2 1 5
1 3 1 8 11 2 3 2 8 7
g1 g2
2 1 2 1 4 30 2 2 2 2 4 12
The coefficients are
1 1 1
a1 z11 z12 g1 13 30
11
101
a z 72
2
2 21 z22 g 2 12 4
5 12 3
Finally, the solution for N=2 is
23 x x 2 2 x3
f x a1 x x 2 a2 x x3 Still not correct. Need larger N.
30 10 3
Lecture 26 Slide 23
Try N=3
For N=3, are matrix equation is zmn
mn
m n 1
z11 z12 z13 a1 g1
z m 3m 8
gm
21 z22 z23 a2 g 2 2 m 2 m 4
z31 z32 z33 a3 g 3
Applying our equations for zmn and gm, we get
1
13 12 53 a1 11 30 a1 13 1
2 3
5 11
30 12
1 4 1 a 7 1
2 5 2 12 a2 2
4
5 1 7
12 0
53 1 79 a3 70
51
a3 53 1 97 70
51
3
1
Finally, the solution for N=3 is
5x x2 x4
f x a1 x x 2 a2 x x3 a3 x x 4 Exact solution!
6 2 3
Lecture 26 Slide 24
12
7/21/2017
Try N=4
For N=4, we get a larger matrix equation, but it also converges to the
exact solution.
In fact, the method converges to the exact solution for all N3.
N
5x x2 x4
f x an x x n 1 Exact solution
n 1 6 2 3
Lecture 26 Slide 25
Finite Element Method
Lecture 26 Slide 26
13
7/21/2017
What is a Finite Element?
Like the finite‐difference method, the finite element method (FEM)
discretizes the problem space. In FEM, it is divided into small regions
called finite elements.
Most common
Node finite
element
Lecture 26 Slide 27
Meshing
Meshing describes the “intelligent” process of subdividing space into
non‐overlapping finite elements. Usually this is done to conform
perfectly to the shapes of devices.
2
1
• FEM mesh conforms very well to curved geometries.
• Mesh can be locally refined to resolve small features or abruptly varying fields.
• Adaptive meshing – Refines the mesh based on a prior solution.
Lecture 26 Slide 28
14
7/21/2017
Adaptive Meshing
A model is performed using a coarse mesh. The mesh is then refined
where the field is varying rapidly and the model is run again. This
continues until convergence.
http://www.cradle‐cfd.com/product/tetrav7/tetra_pre.html
Lecture 26 Slide 29
FEM Vs. Finite‐Difference Grids
The finite element mesh offers minimal discretization error.
discretization error
Lecture 26 Slide 30
15
7/21/2017
Node Elements Vs. Edge Elements
Node Elements
• Field is known and stored at the element nodes.
• Elsewhere, field is interpolated.
• Suffers from spurious solutions due to lack of
enforcing divergence conditions.
• Matrices better conditioned.
Edge Elements
• Vector field is known at the edges of the elements.
• Elsewhere, field is interpolated.
• Solves the spurious solution problem because the
divergence conditions are enforced through
continuity of the tangential field components.
• Matrices have poorer conditioning.
Lecture 26 Slide 31
Shape Functions
The field within the elements is
expanded into a set of basis
functions that are called
“shape” functions.
3
E x, y ui N i x, y
i 1
For linear triangular elements,
the shape functions are
N i x, y ai x bi y ci
Lecture 26 Slide 32
16
7/21/2017
Element Matrix K
The element matrix is N×N where N is the number of nodes in an
element. It is derived to enforce Maxwell’s equations within an
element.
1
E 2 E 0
u3
e
u2
e
u1 e
Element e K11 e K12 e K13 e u1 e b1 e
e e e
e e
K 21 K 22 K 23 u2 b2
e
Nodes are normally labeled
K 31 K 32 e K 33 u3 e b3 e
e
going counter‐clockwise.
Element Matrix K e
Lecture 26 Slide 33
Global Matrix
The global matrix is the overall matrix wave equation.
1 E 2 E b K E b
Global matrix
? ? ? ? ? ? ? ? ?
? ? ? ? ? ? ? ? ?
? ? ? ? ? ? ? ? ?
? ? ? ? ? ? ? ? ?
K
b
? ? ? ? ? ? ? ? ?
? ? ? ? ? ? ? ? ?
? ? ? ? ? ? ? ? ?
? ? ? ? ? ? ? ? ?
Lecture 26 Slide 34
17
7/21/2017
Assembling the Global Matrix (1 of 5)
The global matrix is assembled by adding the elements of the
element matrices to the corresponding elements of the global matrix.
Ne K 8 8 global matrix
K K e K e 4 4 element matrix
e 1
The global matrix is initialized as
0 0 0 0 0 0 0 0 Element Element
0 0 0 0 0 0 0 0
1 2
0 0 0 0 0 0 0 0
Element
0 0 0 0 0 0 0 0
K global node 3
0 0 0 0 0 0 0 0 numbers
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
Lecture 26 Slide 35
Assembling the Global Matrix (2 of 5)
Add Element 1
1. We match the local nodes to the global nodes.
Local Node Global Node
Number Number
1 1 3
4
2 8
Element Element
3 3
1 2
4 2 1 2
Element
Kij e K mn
global 3
im
jn We use this chart to map
local matrix elements to
global matrix elements.
Lecture 26 Slide 36
18
7/21/2017
Assembling the Global Matrix (3 of 5)
Local Global
Add Element 1 1 1
2. We add [K(1)] to the global matrix. 2 8
3 3
4 2
3
4
Element Element
1 2
1 2
Element
3
Lecture 26 Slide 37
Assembling the Global Matrix (4 of 5)
Local Global
Add Element 2 1 8
1. Match local to global node numbers. 2 5
3 4
2. Add [K(2)] to the global matrix. 4 3
3
Element Element
1 2
1 2
Element
3
Lecture 26 Slide 38
19
7/21/2017
Assembling the Global Matrix (5 of 5)
Local Global
Add Element 3 1 7
1. Match local to global node numbers. 2 6
3 5
2. Add [K(3)] to the global matrix. 4 8
Element Element
1 2
4 3
Element
3
2
Lecture 26 Slide 39
Don’t Forget about [b]
We follow the same procedure for [b] and get
Element Element
1 2
Element
3
Lecture 26 Slide 40
20
7/21/2017
Overall Solution
We solve this problem as
K E b
Element Element
1 2
E K b
1
Element
3
Lecture 26 Slide 41
Domain Decomposition Methods
The grid is split into a number of smaller grids. The overall solution is
found by solving the smaller subdomains and iterating their solutions.
Lecture 26 Slide 42
21
7/21/2017
Method of Moments
Lecture 26 Slide 43
Two Common Forms
The method of moments is most commonly applied to one of two
special cases:
Thin Wire Devices Thin Metal Sheets
Lecture 26 Slide 44
22
7/21/2017
Thin Wire Devices
Thin wire devices are most commonly modeled using 1D integral
equations solved along the length of the wire.
Pocklington’s Integral Equation
Pocklington’s integral equation is the most famous, but converges
slower due to the derivative operator operating directly on a
singularity in the kernel. It is the simplest to solve.
j L2
2 e jkr
Ezinc
L 2
I z z k 2 2
z 4 r
dz
Hallen’s Integral Equation
Hallen’s integral equations is a modified version of Pocklington’s
integral equal which fixes the singularity. It is more complicated to
solve, but converges faster.
j 2 2 L2
e jkr
Ezinc k 2
z L 2
I z z
4 r
dz
Lecture 26 Slide 45
Thin Metallic Sheets
Metal sheets are most commonly modeled using 2D integral
equations solved along the length of the wire.
Triangular meshes are typically used where the Rao‐Wilton‐Glisson
(RWG) edge element is most famous.
S. M. Rao, D. R. Wilton, A. W. Glisson, “Electromagnetic scattering by surfaces of arbitrary shape,” IEEE Trans. Antennas and Propagation,
Vol. 30, No. 3, pp. 409‐418, 1982.
Lecture 26 Slide 46
23
7/21/2017
Fast Multipole Method (FMM)
A drawback of the method of moments is that it leads to full
matrices where each element is coupled to every other element.
This leads to prohibitively large matrices for very large structures.
The fast multiple method is a way of treating far away elements as
a single element, thereby significantly reducing the number of
elements in the final matrix.
Lecture 26 Slide 47
Other Worthy Methods
Lecture 26 Slide 48
24
7/21/2017
Boundary Element Method
The wave equation describes the field through volumetric space. The solution
requires the field throughout the volume to be resolved. This leads to a large
number of elements. Matrices are sparse. Boundary conditions are needed.
Maxwell’s equations can be manipulated into the form of a surface integral. For
many structures, this formulation requires many fewer elements and provides
faster solution. Matrices are full. Boundary conditions are needed.
2
1
FEM Mesh BEM Mesh
183 nodes 24 nodes
321 elements 23 elements
sparse matrix full matrix
Lecture 26 Slide 49
Spectral Domain Method
The spectral domain method is essentially the method of moments
in Fourier‐space. Many variations of this method exist.
Array Antennas and
Transmission Line Analysis
Frequency Selective Surfaces
1
2
3
Lecture 26 Slide 50
25