Silicon

DOI 10.1007/s12633-014-9267-2

ORIGINAL PAPER

A Genetic Algorithm Based Approach for the Extraction

of Optical Parameters
Md. Ghulam Saber Rakibul Hasan Sagor
Ashik Ahmed

Received: 4 February 2014 / Accepted: 26 November 2014

Springer Science+Business Media Dordrecht 2015

Abstract We present the Debye model parameters of four

materials which are extensively used in the field of electronics and photonics. The parameters have been extracted using
a Genetic Algorithm (GA) based technique. We have determined the complex relative permittivity using the extracted
modeling parameters and compared with the experimentally
obtained ones. A very good agreement has been found in
each case which validates our extracted parameters. The
associated root-mean-square (RMS) deviations have been
found to be 0.3894, 0.026, 0.8163 and 0.4370 for graphene,
graphene oxide, aluminum zinc oxide (AZO) and gallium
zinc oxide (GZO) respectively.
Keywords Genetic algorithm Debye model Optical
properties Material optics

1 Introduction
Materials interaction with electromagnetic (EM) waves
is described by polarization which can be both electrical
and magnetic. However, in the case of naturally occurring materials, the magnetic polarization can be neglected
for frequencies higher than several hundred terahertz. The
complex relative permittivity or dielectric function is used
to describe the electrical polarization of a material. Many
parametric models have been developed to determine the
dielectric properties of the materials over a broad frequency

M. G. Saber () R. H. Sagor A. Ahmed

Department of Electrical and Electronic Engineering,
Islamic University of Technology (IUT),
Board Bazar, Gazipur, 1704, Bangladesh
e-mail: gsaber@iut-dhaka.edu

range. However, the modeling parameters need to be optimized in order to obtain the best fit to the experimentally
measured data.
Simulation of EM problems involving arbitrary geometries and inhomogeneous materials are often performed
using computational electromagnetics methods such as the
finite-difference time-domain (FDTD) method [1] which
requires the parametric models in order to integrate the
dispersive properties of different materials. Since FDTD
solves the EM fields as a function of time, the parametric
model we choose needs to be expressed in the time domain.
One such model is the Debye model which can be applied
conveniently in both time and frequency domain.
Optimization techniques are required to fit such material
models. Rakic et al. [2] used the acceptance-probabilitycontrolled simulated annealing algorithm in order to solve
the parameter estimation problem. Pernice et al. [3] utilized
an adaptive nonlinear least square algorithm to fit the data.
Recently some researchers have focused on non-traditional
optimization methods like particle swarm optimization
(PSO), ant colony optimization and genetic algorithm (GA).
Kelley et al. [4] used a hybrid particle swarm-least squares
optimization approach to fit the Debye model function.
Clegg et al. [5] fitted the tissue dielectric properties for the
Debye model using GA. We utilized the GA to minimize the
difference between the experimentally obtained values and
the Debye model values. The GA has been chosen because
of its robustness in finding solutions from a large search
space and its ability to overcome the problem of getting
trapped in local minima.
The modeling parameters for several materials using different parametric models have been reported. Deinega et al.
[6] proposed a simple effective model in which the dielectric polarization depends both on the electric field and its
first time derivative and they extracted the parameters for

Silicon

silicon in the spectral range of 3001000 nm. Sagor et

al. [7] reported the extracted parameters for five metals
using the modified Debye model. Rakic et al. [2] presented
the modeling parameters for eleven materials using both
the Lorentz-Drude model and the Brendel-Bormann model.
Saber et al. [8] determined the parameters of Si0.6 Ge0.4
using both modified Debye and Lorentz models.
In this paper, we report the extracted parameters of
graphene, graphene oxide, AZO and GZO for the Debye
model. The modeling parameters have been optimized using
the GA. In order to validate our obtained results, we have
compared them with the experimental ones. We have determined the root-mean-square (RMS) deviation for each case
and an excellent agreement has been observed. The obtained
RMS deviations are 0.3894, 0.026, 0.8163 and 0.4370
respectively for graphene, graphene oxide, AZO and GZO.
The frequency range for which we have determined the
modeling parameters focuses on the plasmonic applications.
A lot of research work is going on in the field of plasmonics due to its ability to overcome the diffraction limit [9].
To the best of our knowledge, this is the first time one has
extracted the modeling parameters for these materials using
the Debye model. It is expected that the extracted parameters will be useful for efficiently defining materials in the
simulation model which will give birth to new types of
plasmonic devices.

2 Material Models
The Debye model describes dielectrics in terms of electric
dipoles. The dipoles respond to an applied electric field and
follow the behavior of the field applied with some relaxation
time. The polarization of the material depends on the oscillating frequency. If the frequency is fast, polarization will
be low and vice-versa.
The frequency dependent complex permittivity function
for the single pole-pair Debye model [10] is given by the
following equation
r () = +

s
(1 + j )

(1)

where, is the infinite frequency relative permittivity, s

is the zero frequency relative permittivity or the static permittivity, j is the imaginary unit and is the relaxation
time.
From (1), it can be observed that the single pole-pair
Debye model can be described by three parameters which
are , s and . However, in the case of metal another term,
, related to conductivity is added and (1) becomes,
r () = +

j
(1 + j )
o

(2)

In this case, we have four variables; however, a relationship

can be obtained among them by comparing with the Drude
model equation as,

= o ( s )

(3)

Thus, there are three parameters that need to be extracted

for fitting the materials and the other can be obtained
from equation (3).

3 Optimization Method
The GA is a very robust and powerful non-traditional optimization technique when the problem has a very large
parameter search space [11]. The general procedure of
the GA requires a randomly picked initial population of
probable solutions and for our case we choose an initial population of 500. Roulette selection (also known as stochastic
sampling) is then used to select the parents from this initial
population. In roulette selection, the chance of an individual getting selected is proportional to its fitness. The fitness
function for the roulette selection is given in equation (4).
After selecting the parents, crossover is done in order to
produce offspring. The final step is to perform mutation on
the new offspring and this process is continued until the
stopping criteria are met.
In this study we have utilized the optimization toolbox of
MATLAB in order to optimize the modeling parameters
of the materials. We have used a double vector population type and constraint dependent creation function. The
proportional scaling function has been chosen for fitness
scaling. For mutation, an adaptive feasible function has been
used and for crossover, we have used the heuristic function. We have run the simulation for 150 generations and
found it to be sufficient for the convergence of the minimization function. For other parameters we have used the
default values (Fig. 1).
Inside the GA, we have allowed , s and to vary for
the convergence of the minimization function. In the case
of the search space was very large in comparison to the
search space of and s . We have used logarithmic values
of in the GA in order to avoid this problem. The fitness
function that we employed:
ff =

(rexp ropt )2 + (imexp imopt )2

(4)

where, rexp is the real part of the experimentally obtained

permittivity function, ropt is the real part of the permittivity
function obtained using optimized parameters, imexp is the
imaginary part of the experimentally obtained permittivity
function and imopt is the imaginary part of the permittivity
function obtained using optimized parameters.

Silicon

be used to model the optical behavior of graphene in the visible range. The complex refractive index that they reported
is in the visible range of electromagnetic spectrum:
C1

(5)
3
The approximation that the refractive index n is constant
over the visible spectral range is reasonable because there
is no sharp resonance within this range. We have used this
method of obtaining n and k and determined the complex
relative permittivity using the obtained n and k values within
the visible and near-infrared spectral range. Then we have
fitted the Debye function for the obtained complex relative
permittivity using the GA.
n = 3.0;

k=

4.2 Graphene Oxide

Fig. 1 Flow chart of the GA

The boundary conditions that need to be maintained

for the Debye parameters to be integrated in the FDTD

algorithm are > 1, s < and o ( s ) .

4 Results and Discussion

4.1 Graphene
Graphene has been the subject of extensive research since
the first experimental demonstration by Novoselov et al.
[12]. It has unique band structure and high mobility
[13] which have attracted the attention of the researchers.
Recently Bruna et al. [14] reported that a dielectric function
can be simply derived by imposing the consistence with various experiments. The optical constants they obtained can

Chemical exfoliation technique yields a large number of

single layers of graphene. However, this alternative method
of producing single layer graphene also produces heavily oxidized layers of graphene oxide. Since the optical
properties of a material are sensitive to its oxide as well,
they are useful in estimating the optical behavior of that
particular material. Jung et al. [15] reported the dispersion functions for the refractive index, n and the extinction coefficient, k of single and multiple-layer graphene
oxide measured using the imaging spectroscopic ellipsometry technique. They measured the dispersion parameters of
a 100 nm thick graphene oxide stack and determined the
dispersion parameters of individual graphene oxide sheets
using an effective medium approximation. In order to optimize the parameters for fitting the Debye function using
the GA for graphene oxide, we have used this data as the
reference.
4.3 AZO and GZO
AZO and GZO are high performance conductive zinc oxides
which have potential applications in transparent electronics and photonics. From the reported optical characteristics
of these conductive films, it has been found that at optical communication wavelength AZO and GZO have lower

Table 1 Optimized Debye model parameters for the materials

Parameters

Graphene

Graphene Oxide

Aluminum zinc oxide

Gallium zinc oxide

s
(sec)
(S/m)
Range of wavelength (nm)
RMS Deviation

8.00
77.00
4.118e-15
N/A
3001200
0.3894

5.0771
3.327
4.7143e-17
N/A
370790
0.026

1.0011
40.5
2.863e-15
1.2834e5
13002100
0.8163

2.1011
51.5
2.863e-15
1.6576 e5
14002200
0.4370

Silicon

Table 2 Covariance of the optimized parameters for different materials

Covariance of the complex permittivity

Covariance of the real part
Covariance of the imaginary part

Graphene

Graphene oxide

Aluminum zinc oxide

Gallium zinc oxide

8.8065 2.0881i
0.7303
8.0762

0.0112 0.00021i
0.0012
0.0099

1.6284 1.3392i
1.8407
0.2124

2.0921 2.4363i
2.6437
0.5515

Curve fitting for graphene

Relative permittivity

12

Curve fitting for Graphene Oxide

5

Re[Ref.]
Imag[Ref.]
Re[Simulated]
Imag[Simulated]

Re[Ref.]
Imag[Ref.]
Re[Simulated]
Imag[Simulated]

4.5
4
Relative Permittivity

14

10

3.5
3
2.5
2
1.5
1

0.5
2
300

400

500

600

700

800

900

1000

1100

0
350

1200

400

450

600

(i)

(ii)

650

700

750

800

Curve fitting for Gallium Zinc Oxide

6

Re[Ref.]
Imag[Ref.]
Re[Simulated]
Imag[Simulated]

Relative Permittivity

Relative Permittivity

550

Wavelength (nm)

Curve fitting for Aluminum Zinc Oxide

4

500

Wavelength (nm)

-2

-4

Re[Ref.]
Imag[Ref.]
Re[Simulated]
Imag[Simulated]

2
0
-2
-4

-6

-6
-8

-8
1400

1500

1600

1700

1800

1900

1500

2000

Wavelength (nm)

1600

1700

1800

1900

2000

2100

Wavelength (nm)

(iv)

(iii)
Fig. 2 Comparison of relative permittivity between our results and
experimental values for (i) Graphene, (ii) Graphene Oxide, (iii) AZO
and (iv) GZO, obtained using the single pole-pair Debye model. The
Fig. 3 Plot of number of
iterations vs time elapsed

Time Elapsed (seconds)

1500

red color indicates the imaginary part and the black color indicates the
real part of the complex relative permittivity

Six-pole Lorentz-Drude
Single Pole-pair Debye

1000

500

1000

2000

3000

4000
5000
6000
Number of Iterations

7000

8000

9000

10000

Silicon

loss than silver [1618]. This characteristic makes AZO and

GZO potential candidates in the field of nanophotonics. We
have obtained the experimental values of the real and imaginary part of the complex relative permittivity for AZO and
GZO from West et al. [19]. These materials offer lower
losses at telecommunication wavelength which is of particular interest in the field of optical communication. The
extracted parameters using the GA for the Debye function
are presented in Table 1. We have determined the RMS
deviation of the complex permittivity in order to find out
the difference between the experimental results and our
extracted parameters. A maximum RMS deviation of 0.8163
has been observed in the case of AZO and for other materials it is less than that. The graphical representation of
our obtained results and experimental results are given in
Fig. 2. The solid lines indicate the results obtained using our
extracted parameters and the dotted lines indicate the results
obtained from the references. It can be easily understood
from the figures that our obtained results agree well with
the results provided in the references for the range of the
wavelengths we have used. The RMS deviation for nickel
using the parameters reported by Rakic et al. [2] has been
determined as 4.38 in the wavelength range 6001100nm
and for palladium it is 3.1251 in the wavelength range of
300700nm while we have obtained a maximum RMS deviation of 0.8163 for AZO which indicates robustness of our
approach in determining the modeling parameters. In the
case of graphene and AZO, the simulated real part of the
complex relative permittivity exhibits slight deviation from
the experimental ones. This is the best we could obtain using
the single pole-pair model. It can be improved by increasing the number of poles, however, that will result in more
computational time.
Covariance is another method that estimates the error
of a fitted parameter. We have determined the covariance
of the fitted parameters which are presented in Table 2.
The covariance of the experimental and optimized complex
permittivity as well as the real and imaginary part of the
experimental and optimized complex permittivity has been
determined.
In our study, we have used the single pole-pair Debye
model to fit the dielectric functions of the materials. This
provides us with an advantage of less computation time
over multiple-pole models. If we would have used multiplepole models, then the computation time would have been
much higher. In order to make an estimate of how much
longer it would take to simulate using multiple-pole models, we have developed a one-dimensional FDTD model
and simulated optical signal propagation through graphene
and silver. For defining graphene in the simulation model,
we have used our extracted parameters for the single polepair Debye model and for silver, we have used the six-pole
Lorentz-Drude (LD) model parameters reported by Rakic

et al. [2]. We have run the simulation for different number

of iterations and recorded the corresponding time elapsed.
The simulation has been done using a first generation Intel
CoreTM i5 processor based computer and the simulation
model has been developed in MATLAB version R2012a
(7.14.0.739). The obtained data are presented graphically in
Fig. 3.
From the figure, it is clearly visible that the computation time required for simulation using the six-pole LD
model is much higher than the single pole-pair Debye case
and the difference between the two increases drastically
as the number of iterations are increased. When the number of iterations is 10000, the six-pole LD model requires
about 3.12 times higher computation time than the single
pole-pair Debye model. Therefore, it is evident that our
extracted parameters exhibit negligible deviations and are
computationally efficient.
5 Conclusion
In summary, we have extracted the Debye model parameters for four materials using the GA. The accuracy of
the obtained parameters has been determined by comparing them with the results provided in the literature. Our
extracted parameters for the single pole-pair Debye model
also have an advantage of computational efficiency over
multiple-pole models and it enables us to run simulations
in the time domain. We believe that these material modeling parameters will help researchers in simulating new
nano-structures and devising new plasmonic applications.

References
1. Yee K (1966) Numerical solution of initial boundary value problems involving Maxwells equations in isotropic media. IEEE
Trans Antennas Propag 14(3):302307
2. Rakic AD, Djurisic AB, Elazar JM, Majewski ML (1998) Optical properties of metallic films for vertical-cavity optoelectronic
devices. Appl Opt 37(22):52715283
3. Pernice WHP, Payne FP, Gallagher DFG (2007) A general framework for the finite-difference time-domain simulation of real
metals. IEEE Trans Antennas Propag 55(3):916923
4. Kelley DF, Destan TJ, Luebbers RJ (2007) Debye function expansions of complex permittivity using a hybrid particle swarm-least
squares optimization approach. IEEE Trans Antennas Propag
55(7):19992005
5. Clegg J, Robinson M (2012) A genetic algorithm for optimizing
multi-pole Debye models of tissue dielectric properties. Phys Med
Biol 57(19):6227
6. Deinega A, John S (2012) Effective optical response of silicon
to sunlight in the finite-difference time-domain method. Opt Lett
37(1):112114
7. Sagor RH, Saber MG, Al-Amin MT, Al Noor A (2013) An
optimization method for parameter extraction of metals using
modified Debye model. SpringerPlus 2(1):15

Silicon
8. Saber MG, Sagor RH (2013) Extraction of optimized parameters for Si0. 6Ge0. 4 material and SPP mode propagation through
Si0. 6Ge0. 4/Ag/Si0. 6Ge0. 4 waveguide. Optoelectronics Lett
9(6):454457
9. Gramotnev DK, Bozhevolnyi SI (2010) Plasmonics beyond the
diffraction limit. Nat Photonics 4(2):8391
10. Kunz KS, Luebbers RJ (1993) The finite difference time domain
method for electromagnetics. CRC
11. Holland JH (1975). U Michigan Press
12. Novoselov KS, Geim AK, Morozov S, Jiang D, Zhang Y,
Dubonos S, Grigorieva I, Firsov A (2004) Electric field effect in
atomically thin carbon films. science 306(5696):666669
13. Geim AK, Novoselov KS (2007) The rise of graphene. Nat Mater
6(3):183191
14. Bruna M, Borini S (2009) Optical constants of graphene layers in the visible range. Appl Phys Lett 94(3):031901-031901031903

15. Jung I, Vaupel M, Pelton M, Piner R, Dikin DA, Stankovich S,

An J, Ruoff RS (2008) Characterization of thermally reduced
graphene oxide by imaging ellipsometry. J Phys Chem C
112(23):84998506
16. Agura H, Suzuki A, Matsushita T, Aoki T, Okuda M (2003) Low
resistivity transparent conducting Al-doped ZnO films prepared
by pulsed laser deposition. Thin Solid Films 445(2):263267
17. Hiramatsu M, Imaeda K, Horio N, Nawata M (1998) Transparent conducting ZnO thin films prepared by XeCl excimer laser
ablation. J Vac Sci Technol A 16(2):669673
18. Kim H, Pique A, Horwitz J, Murata H, Kafafi Z, Gilmore C,
Chrisey D (2000) Effect of aluminum doping on zinc oxide thin
films grown by pulsed laser deposition for organic light-emitting
devices. Thin Solid Films 377:798802
19. West PR, Ishii S, Naik GV, Emani NK, Shalaev VM, Boltasseva A
(2010) Searching for better plasmonic materials. Laser Photonics
Rev 4(6):795808