Transcript otws2 7790
Shape Optimization for
Elliptic Eigenvalue Problems
Chiu-Yen Kao
Collaborators:
He Lin, Yuan Lou, Stanley Osher,
Fadil Santosa, Eli Yabolnovich, Eiji Yanagida
IPAM, Numerics and Dynamics for Optimal Transport
April 17, 2008
Motivation
• Control the resonance frequencies of devices:
– Maximize or minimize certain frequencies
– Maximize the gap between adjacent frequencies
– Maximize the ratio between real part and imaginary part of
eigenvalues (quality factor optimization)
– Minimize the principle eigenvalue
• Application:
–
–
–
–
Vibration system control
Photonic crystal design
Optical Resonator
Population biology
2
Shape Optimization for
Elliptic Eigenvalue Problem
• Goal: Minimize a certain design objective
min F ( )
such that
is the eigenvalue of
L(u, ( x), ) 0 , 1 ( x) 2
subjects to boundary condition on
elliptic differential operator.
D
and L is the
3
Motivation I : Shape of the drum
Consider an open bounded D R n , a positive : D R ,
and ( , u ) satisfies the elliptic eigenvalue problem
u ( x) ( x)u ( x)
u ( x) 0
x D,
x D.
The eigenvalues are
0 1 ( D, ) 2 ( D, ) 3 ( D, ) ...
Q1: Shape problem: D n ( D, ), 1, D A
Q2: Composition problem: n ( D, ), 1 2 , D dx M
4
Theoretical Results
Q1: Shape problem:
1.
2.
Rayleigh (1877) conjectured, and Faber (1923) and Krahn (1925) proved,
that if you fix the area of a drum, the lowest eigenvalue is minimized
uniquely by the disk.
Payne, Pólya, and Weinberger conjecture (1955): The disk maximizes the
ratio of 2 to 1 has been proved by Ashbaugh and Benguria (1992)
Q2: composition problem:
1.
2.
Krein (1955) provided one dimensional optimal density distribution for
maximal and minimal n .
Cox and McLaughling (1993): minimal n for higher dimensions.
( 1 )( 2 ) 0 a.e. in D
5
Numerical Approach
Q2: composition problem:
n ( D, ), 1 2 , dx M
D
( 1 )( 2 ) 0 a.e. in D
Let 2 1 D \ , | D | const
F (k ) or max F (k )
Find min
shape optimization problem:
6
Eigenvalue Separation for Drums
• Shape Optimization of a drum head with a fixed domain D
u ( x)u x D
{
u0
x D
• Let be a domain inside D
1 for x
, and ( x) {
2 for x
• Solve the optimization:
•
max (k k 1 ) max F () for a given fixed k
• Subject to the constraint:
|| || const G () 0
7
Shape Mapping
• The set ( I )
is defined by
{x ( x) | x }
• The vector field ( x ) is the displacement of .
8
Shape Derivative
• Framework of Murat-Simon:
• Let be a reference domain. Consider its variations
• ( I ) with W 1, ( R N ; R N )
• Definition: the shape derivative of F () at is the
Frechet differential of F (( I )) at 0 .
F (( I h )) F ()
d s F ()( ) lim
h 0
h
9
Shape Derivatives of
Eigenvalues and Area
Compute shape derivatives
1. F () (k k 1 )
d s F ()( )
k ( 2 1 ) 2
k 1 ( 2 1 ) 2
uk nds
uk 1 nds
2
2
( x)uk dx
( x)uk 1dx
D
D
2.G () 1dx const
d s G ()( ) ndx
10
Lagrange Multiplier Method
• max F () subject to G () 0
use Lagrange multiplier method:
L ( ) F ( ) G ( )
the necessary condition for a minimizer is
d S L()( ) d S F ()( ) d S G()( ) 0
together with the constraint
G () 0
allows us, in principle, to find and .
11
Gradient Ascent
• Shape derivative
L' ()( ) ( f (uk , uk 1 , k , k 1 , ) g ) ( x) nds
Gradient descent algorithm for the shape
new ( I ( f (uk , uk 1 , k , k 1 , ) g )n old ) old
The normal advection velocity of the shape is f g .
We solve the level set equation:
t ( f g ) | | 0
Because of G () 0, we need to have ( f g )ds 0 .
Then f ds
. s
gd
12
max (
2
1 )
13
max (
3
2 )
14
max (
3
2 )
15
max (
5
4 )
16
Photonic Crystals
17
Motivation II : Photonic Crystal
• Shape Optimization of a photonic crystal with a fixed domain
1
2
i i En 2 En , En ( x a ) En
c
1
2
i i Hn 2 Hn , Hn ( x a ) Hn
c
1 for x S
Let S be a domain inside , and
( x)
We begin with a shape with
2 for x S
Ej ( 0 , ) 02 / c 2 Ej 1 ( 0 , )
E
E
Hj ( 0 , ) 02 / c 2 Hj 1 ( 0 , )
H
H
K
and we want to maximize
J1 ( ) inf EjE 1 ( , ) sup EjE ( , ) J 2 ( ) inf HjH 1 ( , ) sup HjH ( , )
J 3 ( ) inf inf HjH 1 ( , ),inf EjE 1 ( , ) sup sup HjH ( , ),sup EjE ( , )
18
Maximize
Dielectric: 1 : 11.4
Initial Gap: none Final Gap: 0.1415
1 2
TM
Band Gap
19
Maximize
1 2
TE
Band Gap
20
Motivation III: Quality Factor Optimization
• Design the material to have lower loss of energy
– Mechanics Systems
Ex: damped mass spring
– Electrical Systems
Ex: RLC circuit, quartz crystal
– Optical Systems
Ex: photonic crystal
Pictures:
http://en.wikipedia.org/wiki/Quartz_clock
http://minty.stanford.edu/PBG/
21
The Mass Spring System (1)
The displacement satisfies:
For small damping
(1)
, the solution is
The total energy is
where the period
and
.
22
The Mass Spring System (2)
The quality factor is defined as :
23
1-D Schrödinger’s Equation
• Finite potential well:
24
Optical Resonator
• One-Dimensional Case
• Higher-Dimensional Case
•Goal: minimize the quality factor
Re( )
sup
2 Im( )
25
1-D Forward Eigenvalue Solver (1)
Solve
by finite element method. Apply the test function
By incorporating the boundary condition
26
1-D Forward Eigenvalue Solver (2)
Thus
The equation can be written as
It is a nonlinear Eigenvalue Problem !!
27
1-D Forward Eigenvalue Solver (3)
The equation can be written as
28
2-D Forward Eigenvalue Solver (1)
• Boundary Integral Method
It is a nonlinear Eigenvalue Problem !!
29
2-D Forward Eigenvalue Solver (2)
30
2-D Forward Eigenvalue Solver (3)
• Nonlinear Eigenvalue Problem: (Newton’s method) inverse iteration
31
Gradient Flow
In terms of a single mode damped oscillator, the quality factor is proportional
to the ratio between the real part and the imaginary part of the eigenvalue.
Q
real ( )
2imag ( )
Our goal here is to maximize the quality factor subject to the wave equation
we discussed previously which can be written in the general eigenvalue problem
p( )VR 0
Suppose there is a small perturbation s.t.
We keep only the first order term
Premultiplying by the corresponding eigenvector leads to
Thus
32
Algorithm
33
1D Numerical Results (1)
34
1D Numerical Results (2)
35
2D Numerical Results (1)
36
2D Numerical Results (2)
37
IV: Eigenvalue with Indefinite Weight
• Consider the elliptic eigenvalue problem with indefinite
weight
u m( x)u x D
{
u / n 0
x D
Let be a domain inside D , and
Solve the optimization:
m1 0 for x
m( x ) {
m2 0 for x
min 1
Subject to the constraint:
mdx M 0
D
38
Diffusive Logistic Equation
t [m( x) ]
0
n
( x,0) 0, ( x,0) 0
where
represents the density of a species and
1. If 1 (m) ,
2. If
( x, t ) 0
uniformly as
1 (m), ( x, t ) * ( x)
xD
x D
xD
is the growth rate.
t
uniformly as
t
The effect of dispersal and spatial heterogeneity in population dynamics.
39
Eigenvalue and Eigenfunction
40
Minimizers in 1D
• Theorem: When D ( a, b) is an interval, then there are exactly
two global minimizers of (m) . For the logistic model, this
means that a single favorable region at one of the two ends of
the whole habitat provides the best opportunity for the species
to survive.
41
Minimizers in High Dimensions
• In general domain: open question
m 0 for x (c,1) (0, b)
• Suppose D (0,1) (0, b). m( x) { 1
m2 0 for x (0, c) (0, b)
If b / c , then is not minimal. In particular, the strip
at the end with much longer edge can’t be the optimal favorable
region.
42
Square domains
43
Square domains
44
Rectangular domains
45
Ellipse domain
46
More general domains
47
The End
Thank you
48