Approximating Dispersive Mechanisms Using Debye Model

Approximating Dispersive Mechanisms Using the
Debye Model with Distributions of Dielectric

Parameters
Nathan Gibson
Assistant Professor
Department of Mathematics
gibsonn@math.oregonstate.edu
In Collaboration with:
Prof. H. T. Banks, CRSC
Dr. W. P. Winfree, NASA Langley
Karen Barrese Smith, OSU

Neel Chugh, Tufts
Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 1 / 45

1 Background
Maxwells Equations
The One Dimensional Problem
Dielectric Parameters of Interest
2 Cole-Cole and Debye Models

Cole-Cole and Debye Models
Distributions
3 Inverse Problems
Frequency-domain Inverse Problem
Time-domain Inverse Problem

Background
1 Background
Maxwells Equations

Distributions
3 Inverse Problems

Background Maxwells Equations
Maxwells Equations
D
+J=H (Ampere)
t
B
= E (Faraday)
t
D= (Poisson)
B=0 (Gauss)
E= Electric field vector D= Electric displacement

H= Magnetic field vector B= Magnetic flux density
= Electric charge density J= Current density
We impose homogeneous initial conditions and boundary conditions.

Background Maxwells Equations
Constitutive Laws
Maxwells equations are completed by constitutive laws that describe the

response of the medium to the electromagnetic field.
D = E + P
B = H + M
J = E + Js
P= Polarization = Electric permittivity

M= Magnetization = Magnetic permeability
Js = Source Current = Electric Conductivity

Background The One Dimensional Problem
Maxwells Equations in One Space Dimension
Assume that the electric field is polarized to oscillate only in the y

direction, propagates in x direction, and everything is uniform in z
direction.
y
Equations involving Ey and Hz .
Ey Hz dP Ey
= Ey
t x dt .........
........
..........................
.......... ..
..
.........
........
..........................
.......... ..
.
Hz Ey x
........ ... ........ ...
....... ....... ..
..
..
........
. .
... ..
..
........
. ...
=
.. ..

.. ....... ... .......
.... ..
..
..
. .
. ..
..
..
.
.. ........ ... ........
........ ........
t x
..... ......... ..... .........
...... ..................... ...... .....................
............. .............
Hz
If = 0 and P = 0, then E = Ey satisfies the 1D wave equation with

c = 1/
2 E (x, t) 2
2 E (x, t)
= c
t 2 x 2
Background Dielectric Parameters of Interest
Constitutive Relations
Recall
D = E + P
where P is the dielectric polarization.
We can generally define P in terms of a convolution
Z t
P(t, x) = g E(t, x) = g (t s, x; )E(s, x)ds,
0
where g is a general dielectric response function (DRF), and is

some parameter set.

DRF Examples
Debye model
g (t, x) = 0 (s )/ e t/
(or P + P = 0 (s )E)
Lorentz model
g (t, x) = 0 p2 /0 e t/2 sin(0 t)

+ 1 P + 02 P = 0 p2 E)
(or P


Frequency Domain
Converting to frequency domain via Fourier transforms
= ()E
D
Debye model
s
() = + +
1 + i i 0
Cole-Cole model
s
() = + 1
+
1 + (i ) i 0

1 Background
Maxwells Equations

Distributions
3 Inverse Problems

Cole-Cole and Debye Models Cole-Cole and Debye Models
Multi-pole models
In general there are multiple mechanisms at various scales that account for
polarization. To attempt to account for several of these over a range of
frequencies, researchers tend to use multi-pole models:
Multi-pole Debye model:
n
X m
()D = + +
1 + i m i 0
m=1
Multi-pole Cole-Cole model:

n
X m
()CC = + (1 )
+
1 + (i m ) m i 0
m=1

Cole-Cole and Debye Models Dry skin data
True Data
3
10 Debye Model
ColeCole Model
2
10
1
10 2 4 6 8 10
10 10 10 10 10
f (Hz)
Figure: Real part of (), , or the permittivity.

Cole-Cole and Debye Models Dry skin data
1
10
0
10
1
10
2
10
3
10 True Data
Debye Model
ColeCole Model
4
10 2 4 6 8 10
10 10 10 10 10
f (Hz)
Figure: Imaginary part of (), , or the conductivity.

Cole-Cole and Debye Models Distributions
Distributions of Parameters
To account for the possible effect of multiple parameter sets , consider

Z
h(t, x; F ) = g (t, x; )dF (),
N
where N is some admissible set and F P(N ).

Then the polarization becomes:
Z t
P(t, x) = h(t s, x)E(s, x)ds.
0
Motivation: match data even better than multi-pole Cole-Cole, and more
efficient to simulate.

Inverse Problems
1 Background
Maxwells Equations

Distributions
3 Inverse Problems

Inverse Problems
Two Approaches
We will consider the problem of determining the distribution of dielectric

parameters which describe a material by using the following as data:
Complex permittivity (frequency-domain)
Electric field (time-domain)

Inverse Problems Frequency-domain Inverse Problem
Inverse Problem for F
}j we seek to determine a probability measure F , such

Given data {
that
F = min J (F ),
F P(N )
where, for example,
j ] 2 .
X
J (F ) = [(j ; F )
j
As () is complex, we define e = [R((j )), R((j )i j 0 )] and

minimize the 2 -norm of the relative error between e(F ) and e.
Given a trial distribution Fk we compute (j ; Fk ) and test J (Fk ),
then update Fk+1 as necessary.

Monte Carlo Simulations
To compute (; Fk ) we perform N Monte Carlo (MC) simulations.

Each MC simulation consists of drawing trial values of one or more of
the following according to the definition of the distribution F :
, , ,
We then compute

() = + +
1 + (i ) i 0
The term (; F ) is simply computed as the sample mean of the

(; F ) ,
N
1 X
()DD = () .
N
=1

Convergence of MC
We need to select N (the number of MC simulations in the computation

of ()DD ) sufficiently large so as to reduce variability.
0.01
10
0.02
10
0.03
10

0.04
10
N = 10,000
10
0.05 N = 100,000
N = 1,000,000
9.03 9.04 9.05 9.06 9.07 9.08 9.09
10 10 10 10 10 10 10
f (Hz)
Multi-pole Example
Consider
n
X m
() = + +
1 + (i m ) i 0
m=1
For each pole m, we randomly sample each m and m where
m U [(1 am )m , (1 + bm )m ] ,
and
m U [(1 cm )m , (1 + dm )m ]
for some given reference values of m and m .
Thus, F is determined by am , bm , cm and dm , i.e., they are the values
of interest in our inverse problem.

Dry Skin Problem
We use complex permittivity measurements from [GLG96] describing

dry skin as data.
We use the estimates from [GLG96] for , , m and m as our
reference values.
The constraints on the distribution parameters were
a1 [0, 1] b1 [0, 1]
a2 [.5, 1.5] b2 [1, 2]
c1 [0, 1] d1 [0, 1]
c2 [0, 1] d2 [0, 1]
The results from DIRECT (global constrained optimization) were
a1 = 0.1337 b1 = 0.6646
a2 = 1.0000 b2 = 1.7840
c1 = 0.4630 d1 = 0.5000
c2 = 0.5988 d2 = 0.4630
J = 12.1945
0.04
Initial
Minimizer
0.035
0.03
0.025
0.02
F
0.015
0.01
0.005
0
0 500 1000 1500 2000 2500 3000 3500

Figure: Uniform distributions for values in multi-pole Debye model for dry
skin.

12
Initial
Minimizer
10
8
log(F)
0
11 10.5 10 9.5 9 8.5 8 7.5
log()
Figure: Uniform distributions for values in multi-pole Debye model for dry skin.

3 Data
10 Debye (27.79)
ColeCole (10.4)
Model A (13.60)
Model B (12.19)

2
10
2 4 6 8 10
10 10 10 10 10
f (Hz)
Figure: Real part of (), , or the permittivity. Model A refers to the Debye
model with distributions only on . Model B refers to the Debye model with
distributions on both and . Note: U156 = 18.0443,
2 (4) : = {.05, .01, .001} = = {9.49, 13.28, 18.47}
0
10
Model A relative cost
Model B relative cost
1
10
Relative Cost
2
10
3
10
4
10
5
10 2 4 6 8 10 12
10 10 10 10 10 10
f (Hz)
Figure: The relative costs between Model A and the true data and between
Model B and the true data. Model A refers to the Debye model with distributions
only on . Model B refers to the Debye model with distributions on both and
.

Comments on Optimization
Levenberg-Marquardt failed to find a local minimum.

In addition, programs such as fminsearch and fmincon
(fminsearch subject to a set of constraints) were also tried.
This difficulty was mentioned in [GLG96].
The randomness of the inverse problem implies that it is ill-posed;
gradient-based algorithms will often choose a non-descent direction.
Methods for implementation of such local minimum searches is an
area which should be explored further.

To compare the time-domain response of each model of dry skin (Debye,

Cole-Cole, and Distributed Debey), we simulate a broad-band pulse
through the materials.
5
x 10
5
Free space
Electric Field
Dielectric
0
5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time (s) 9
x 10
12
x 10
4
Spectral Amplitude
0
0 1 2 3 4 5 6 7 8 9 10
f (Hz) 9
x 10
Figure: The top plot shows the value of the electric field at a fixed point in space
as time varies. The bottom shows the FFT of the two signals.
5
x 10
2
Debye Model
1.5 Matchup to ColeCole
Matchup to Data
1
0.5
0
E
0.5
1.5
2.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t x 10
9
Figure: Forward simulations with different distributions of dielectric parameters.

Inverse Problems Time-domain Inverse Problem
Inverse Problem for F
Given data {E }j we seek to determine a probability measure F , such

that
F = min J (F ),
F P(N )
where, for example,

X 2
J (F ) = E (0, tj ; F ) Ej .
j
Given a trial distribution Fk we compute (j ; Fk ) and test J (Fk ),

then update Fk+1 as necessary.
Need a (fast) (numerical) method for computing E (x, t; F ).

Stability of Inverse Problem
Continuity of F (E , E ) = continuity of F J (F )
Compactness of N = compactness of P(N ) with respect to the

Prohorov metric
Therefore, a minimum of J (F ) over P(N ) exists

1D Example
x
1
0
0
1
0
1 111111
000000
000000
111111
0
1
0
1 000000
111111
0
1 000000
111111
000000
111111
0
1
0
1 000000
111111
0
1 000000
111111
000000
111111
0
1
0
1 000000
111111
0
1 000000
111111
000000
111111
0
1
0
1 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
E(t,z) 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111 z
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
Js (t, z) = 0 (z) sin(t)I[0,tf ] (t)

Numerical Discretization
E H dP
= E
t x dt
H E
=
t Z x
Z t
P(t, x) = g (t s, x; )E (s, x)ds dF ().
N 0
Second order FEM in space

piecewise linear splines
Second order FD in time
Crank-Nicholson (P)
Central differences (E )
en pn en+1 pn+1
Use quadrature (trapezoidal) for distribution

Discrete Distribution Example
Mixture of two Debye materials with 1 and 2

Total polarization a weighted average
P = 1 P1 (1 ) + 2 P2 (2 )
Corresponds to the discrete probability distribution
dF ( ) = [1 (1 ) + 2 (2 )] d

Discrete Distribution Inverse Problem
Assume the proportions 1 and 2 = 1 1 are known.

Define the following least squares optimization problem:
X 2
min J = min E (tj , 0; (1 , 2 )) Ej ,

(1 ,2 ) (1 ,2 )
j
where Ej is synthetic data generated using (1 , 2 ) in our simulation

routine.

Discrete Distribution J using 106 Hz
f=1e6,1=.5
3
log(J)
7.2
7.4 7.2
7.6 7.4
7.6
7.8
7.8
log(tau2) 8 8
log(tau1)
The solid line above the surface represents the curve of constant
:= 1 1 + (1 1 )2 . Note:
.15 < 1.
Inverse Problem Results 106 Hz
1 2
Initial 3.95000e-8 1.26400e-8 2.60700e-8
LM 3.19001e-8 1.55032e-8 2.37016e-8
Final 3.16039e-8 1.55744e-8 2.37016e-8
Exact 3.16000e-8 1.58000e-8 2.37000e-8
Levenberg-Marquardt converges to curve of constant

Traversing curve results in accurate final estimates

Discrete Distribution J using 1011 Hz
f=1e11,1=.5
4
log(J)
7.2
7.4
7.6 7.1
7.2
7.3
7.4
7.8 7.5
7.6
log(tau2) 7.7
7.8
8 7.9
8 log(tau1)
The solid line above the surface represents the curve of constant
:= 1 = 1 + 2 . Note:
15000 > 1.
c c1 c2
Inverse Problem Results 1011 Hz
1 2

Initial 3.95000e-8 1.26400e-8 0.174167
LM 4.08413e-8 1.41942e-8 0.158333
Final 3.16038e-8 1.57991e-8 0.158333
Exact 3.16000e-8 1.58000e-8 0.158333


Log-Normal Distribution of
Gaussian distribution of log( ) with mean and with standard deviation :
(log )2

1 1 1
dF ( ; , ) = exp d,
2 2 ln 10 2 2
Corresponding inverse problem:

X 2
min E (tj , 0; (, )) Ej .

q=(,)
j

Estimated density of as log normal

12
Converged estimate (+) Initial estimate (*)
and true estimate (o)
10
6
f
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
7
x 10
Shown are the initial density function, the minimizing density function and
the true density function (the latter two being practically identical).
Bi-gaussian Distribution of log
Bi-gaussian distribution with means 1 and 2 and with standard deviations

1 and 2 :
dF ( ) = 1 d F ( ; 1 , 1 ) + (1 1 )d F ( ; 2 , 2 ),
where
(log )2

1 1 1
d F ( ; , ) = exp d,
2 2 ln 10 2 2
Corresponding inverse problem:
X 2
min |E (tj , 0; q)| |Ej | .

q=(1 ,1 ,2 ,2 )
j

Bi-gaussian Results with 106 Hz
case 1 1 2 2
Initial 1.58001e-7 0.036606 3.16002e-9 0.0571969 8.1201e-8
1 ,2 4.27129e-8 0.036606 4.24844e-9 0.0571969 2.36499e-8
Final 3.09079e-8 0.0136811 1.63897e-8 0.0663628 2.37978e-8
Exact 3.16000e-8 0.0457575 1.58000e-8 0.0457575 2.37957e-8

Note: for this continuous distribution,
Z
= dF ( ).
T

Bi-gaussian Results with 1011 Hz
case 1 1 2 2

Initial 1.58001e-7 0.036606 3.16002e-9 0.0571969 0.538786
1 ,2 1.58001e-7 0.036606 1.12595e-8 0.0571969 0.158863
Final 3.23914e-8 0.0366059 1.56020e-8 0.0571968 0.158863
Exact 3.16000e-8 0.0457575 1.58000e-8 0.0457575 0.158863

Note: for this continuous distribution,
1
Z

= dF ( ).
T c

Comments on Time-domain Inverse Problems
We have shown well-posedness of the problem for determining

distributions of dielectric parameters
Our estimation methods worked well for discrete distributions
Our estimation methods worked well for the continuous uniform
distribution and gaussian distributions
We are currently only able to determine the means in the bi-gaussian
distributions, the data is relatively insensitive to the standard
deviations

Homogenization
A good fit when (or ) is constant suggests using a single , even

for the bi-gaussian case
This modeling approach concludes that the effective parameter

should be if < 1, else 1/c
We have also considered a traditional homogenization method based
on periodic unfolding (See [BBC+ 06] for details)
This approach allows us to use information about the periodic
structure, i.e., hexagonal cells.

H. T. Banks, V. A. Bokil, D. Cioranescu, N. L. Gibson, G. Griso, and

B. Miara.
Homogenization of periodically varying coefficients in electromagnetic
materials.
28(2):191221, 2006.
HT Banks and NL Gibson.
Electromagnetic inverse problems involving distributions of dielectric
mechanisms and parameters.
QUARTERLY OF APPLIED MATHEMATICS, 64(4):749, 2006.
S. Gabriel, RW Lau, and C. Gabriel.
The dielectric properties of biological tissues: III. Parametric models
for the dielectric spectrum of tissues (results available online at
http://niremf.ifac.cnr.it/docs/DIELECTRIC/home.html).
Phys. Med. Biol, 41(11):22712293, 1996.

Approximating Dispersive Mechanisms Using Debye Model

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Approximating Dispersive Mechanisms Using Debye Model

Diunggah oleh

Hak Cipta:

Format Tersedia

Approximating Dispersive Mechanisms Using the

Debye Model with Distributions of Dielectric

Karen Barrese Smith, OSU

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 1 / 45

2 Cole-Cole and Debye Models

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 2 / 45

2 Cole-Cole and Debye Models

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 3 / 45

E= Electric field vector D= Electric displacement

We impose homogeneous initial conditions and boundary conditions.

Maxwells equations are completed by constitutive laws that describe the

P= Polarization = Electric permittivity

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 5 / 45

Maxwells Equations in One Space Dimension

Assume that the electric field is polarized to oscillate only in the y

If = 0 and P = 0, then E = Ey satisfies the 1D wave equation with

where g is a general dielectric response function (DRF), and is

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 7 / 45

g (t, x) = 0 p2 /0 e t/2 sin(0 t)

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 8 / 45

Converting to frequency domain via Fourier transforms

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 9 / 45

2 Cole-Cole and Debye Models

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 10 / 45

Multi-pole Cole-Cole model:

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 11 / 45

Figure: Real part of (), , or the permittivity.

Figure: Imaginary part of (), , or the conductivity.

To account for the possible effect of multiple parameter sets , consider

where N is some admissible set and F P(N ).

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 14 / 45

2 Cole-Cole and Debye Models

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 15 / 45

We will consider the problem of determining the distribution of dielectric

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 16 / 45

Inverse Problem for F

}j we seek to determine a probability measure F , such

where, for example,

As () is complex, we define e = [R((j )), R((j )i j 0 )] and

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 17 / 45

Monte Carlo Simulations

To compute (; Fk ) we perform N Monte Carlo (MC) simulations.

The term (; F ) is simply computed as the sample mean of the

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 18 / 45

We need to select N (the number of MC simulations in the computation

For each pole m, we randomly sample each m and m where

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 20 / 45

Dry Skin Problem

We use complex permittivity measurements from [GLG96] describing

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 22 / 45

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 23 / 45

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 25 / 45

Levenberg-Marquardt failed to find a local minimum.

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 26 / 45

To compare the time-domain response of each model of dry skin (Debye,

Figure: Forward simulations with different distributions of dielectric parameters.

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 28 / 45

Inverse Problem for F

Given data {E }j we seek to determine a probability measure F , such

where, for example,

Given a trial distribution Fk we compute (j ; Fk ) and test J (Fk ),

Nathan Gibson (OSU-Math) Approximating Dispersive Mechanisms Oct 2008 29 / 45

Stability of Inverse Problem