I. Overview - Computer Engineering Group

Download Report

Transcript I. Overview - Computer Engineering Group

Finite Difference Method


For conductor exterior, solve Laplacian
equation
k
l
In 2D:
7/16/2015
i
m
j
ELEN 689
1
Uniform Grid
i, j+1
i–1, j
i+1, j
i, j
i, j–1
7/16/2015
ELEN 689
2
Basic Properties

Consider two conductors
–Q

+Q
S
Let v=f(Q). From Gauss’ law
if we double the amount of charge, E will also double
Really?
since the equation is linear
 Therefore, v and Q are linearly related, or Q=Cv
7/16/2015
ELEN 689
3
Multiple Conductors

Consider conductors 1, 2, …, n
Q1
Qn
Q2

Apply the above argument for every pair of
conductors i and j
7/16/2015
ELEN 689
4
Capacitance Matrix
7/16/2015
ELEN 689
5
BEM Review


Partition conductor surfaces into panels
Build coefficient matrix P, where
and G is Green’s function, such as


Solve linear system Pq=v
Add charges to get capacitance
7/16/2015
ELEN 689
6
Make It Faster


Discretization: O(n)
Compute P: O(n2) O(n)


Since P is size nn, P can not be
constructed explicitly
Solve Pq=v: O(n3) O(n)

7/16/2015
Iterative methods
ELEN 689
7
Fast Multipole Methods


N-body problem: Given n particles in 3D
space, compute all forces between the
particles
Fast multipole algorithms



7/16/2015
Appel 85
Rokhlin 86, Greengard & Rokhlin 87
O(n) time
ELEN 689
8
Basic Idea of Multipole



A cluster of charges at distance can
be approximated by a single charge
Reduce operations from k2 to k
Form all clusters recursively in O(n)
time — hard part!
potential
k charges
7/16/2015
at k points
ELEN 689
9
Solve Ax=b Iteratively

Approximate Ax–b=0:
Initial solution x
Compute Ax
If Ax–b > t/b,
modify x


Bottleneck: Matrix-vector product Ax
A is not used elsewhere
7/16/2015
ELEN 689
10
Example: Jacobi Method
7/16/2015
ELEN 689
11
Example: Jacobi Method

Transformation


Iterations




Ax = b  Dx=Dx–Ax+b  x = (I–D–1A)x+D–1b
x(i+1)= (I–D–1A) x(i)+D–1b
x(0) = 0, x(1) = D–1b, x(2) = (I–D–1A) x(1)+D–1b, …
If diagonal dominate, then the method
converges
Better iterative methods exist that converge
under weaker conditions
7/16/2015
ELEN 689
12
Fast Algorithm HiCap

Conductor surface refinement:


Approximate P and store it in a hierarchical
data structure of size O(n)


Adaptively partition conductor surfaces into small
panels according to a user supplied threshold 
The data structure permits O(n) time matrixvector product Px for any n-vector x
Solve linear system Pq=v using iterative
methods
7/16/2015
ELEN 689
13
Adaptive Panel Partition

If interaction between Ai Aj > , refine
Ai and Aj. Otherwise, record Pij in P.
A
C
C
B
E
F G
H
7/16/2015
1
2
I
M N
L
J
J
3
ELEN 689
4
5
14
Representation of Matrix P

P is stored as links in a hierarchical data
structure
A
B
D
F
7/16/2015
H
C
I
E
G
J
K
M
ELEN 689
L
N
15
Example

If area/dist  1, refine the panel
A
2
C
4
H
1/7
1/5
B
C
1/5
I
J
1/3
B
1
4
I
J
7/16/2015
ELEN 689
16
Example (cont’d)

If area/dist  1, refine the panel
A
2
4
1
H
1/7
1/5
C
B
C
1/5
I
J
E
F G
M N
L
4
J
7/16/2015
D
F
E
G
K
M
ELEN 689
L
N
17
A
Full 8x8
B
D
matrix P:
B
H
K
E
D
E
A
C
K
I
L
H
J
M
N
I
C
1/4.6
1/4.6
1/5.5
J
L
A
Implicitly
stored P:
B
B
D
H
K
E
D
E
A
C
K
I
L
H
J
I
C
1/5
J
L
Properties of P

P positive, symmetric, positive definite



If fully expanded, P is size nn
P can be approximated by O(n) block
entries, where n is the number of panels


Positive definite: xPx > 0 for all x
This is because each panel interacts with
constant number of other panels
The block entries allow O(n) time matrixvector product Px for any x
7/16/2015
ELEN 689
20
Mat-Vec Pq, Step 1

Compute charge for all panels
A
B
D
F
7/16/2015
H
C
I
E
J
K
G
M
ELEN 689
L
N
21
Mat-Vec Pq, Step 2

Compute potential for all panels
A
B
D
F
7/16/2015
H
C
I
E
G
J
K
M
ELEN 689
L
N
22
Mat-Vec Pq, Step 3

Distribute potential to leaf panels
A
B
D
F
7/16/2015
H
C
I
E
G
J
K
M
ELEN 689
L
N
23
Solving Linear Systems



Use fast iterative methods GMRES
Each iteration requires a matrix-vector
product Pq that can be completed in
O(n) time
Solution obtained in 10-20 iterations,
regardless of n
7/16/2015
ELEN 689
24
Error and Time Complexity


Error of approximation is controlled
by 
Time complexity is O(n) because step
takes O(n) time
7/16/2015
ELEN 689
25
Multi-layer Dielectric

Kernel independent
methods


Multi-layer Green’s function
Kernel dependent methods



7/16/2015
=8.0
m3
=4.0
Discretize dielectric-dielectric m2 =3.9 m2 m2
interface
=4.1
Introduce interface variables
m1
and modify linear system
Expensive
ELEN 689
26
Other Dielectric Problems



Conformal dielectric
Voids
Air gap
m3
m2
m2
m1
7/16/2015
ELEN 689
27
Comparison of Methods

FastCap: O(n)


Random Walk


Kernel independent
Singular Value Decomposition: O(nlogn)


Kernel independent, QuickCap
Pre-corrected FFT: O(nlogn)


Kernel dependent (1/r)
Kernel independent, Assura RCX
HiCap: O(n)

7/16/2015
Kernel independent
ELEN 689
28