You are on page 1of 14

Optical and Quantum Electronics (2006) 38:843–856 © Springer 2007

DOI 10.1007/s11082-006-9058-0

An FDTD method for the simulation of dispersive


metallic structures

w. h . p . p e r n i c e1,∗ , f. p . p ay n e1 a n d d . f. g . g a l l a g h e r2
1 Department of Engineering Science, University of Oxford, Parks Road, Oxford, OX1 3PJ, UK
2 Photon Design, 34 Leopold Street, Oxford, OX4 1TW, UK
(∗ author for correspondence: E-mail: wolfram.pernice@eng.ox.ac.uk)

Abstract. In this paper, we present a formulation of the finite-difference time-domain method for the
simulation of metallic structures. The frequency dependent dielectric function of metals is approximated
by a combined Drude–Lorentzian multi-pole expansion and fitting errors of only a few percent are
obtained. An auxiliary differential equation technique is used to extend the standard FDTD algorithm
with the dispersive material equations. The algorithm is validated by calculating reflection and trans-
mission coefficients for thin metal layers, elliptical nano-particles and by simulating a surface plasmon
resonance device. Excellent agreement between the FDTD simulations and exact theoretical results are
obtained.

Key words: FDTD method, dispersive media, surface plasmon resonance, metals

1. Introduction

A very successful class of algorithms for the time-domain integration


of the differential form of Maxwell’s equations is based on the finite-
difference scheme proposed by Yee (1966) called the finite-difference time-
domain (FDTD) method. Based on second-order central differences, the
FDTD method implements the space derivatives of the curl operators
via finite differences in regular interleaved (dual) Cartesian space meshes
for the electromagnetic fields in order to directly solve the time-depen-
dent EM field in a volumetric region for a given time interval. The
FDTD method is one of the most popular computational techniques
due to its ease of implementation and flexibility. However the applica-
tion of the original FDTD method is limited in a sense, because it
applies only to isotropic non-dispersive materials. In many cases, it is
desirable to include dispersive media, for example for the simulation of
metal structures. Metals show strong dispersive behavior in the optical
and near infrared wavelength range. In order to get accurate simula-
tion results for structures containing metallic layers or nano-particles it is
therefore paramount to take the material dispersion into account. In the
past, predominantly two material models have been incorporated into the
844 W.H.P. PERNICE ET AL.

FDTD scheme, the Lorentzian and Debye models, Fan and Liu (2000),
Young and Nelson (2001). By choosing appropriate parameters a num-
ber of materials can be modeled over a wide frequency range. In this
paper, we show that metals can be efficiently modeled in the frequency
domain by fitting the complex dielectric susceptibility to just a set of
Lorentzian functions.
The incorporation of material dispersion into the FDTD method is often
achieved by using either the recursive convolution (RC) approach (Luebbers
et al. 1990; Luebbers and Hunsberger 1992; Vial et al. 2005) or an auxil-
iary differential equation (ADE), Taflove (1995) and Korner and Fichtner
(1997). The storage requirements for additional variables and the accuracy
of the scheme depends on the discretization of the time derivatives and the
representation of the convolution integrals for the electric flux density.
In this paper we adopt the ADE scheme. We show how the complex
dielectric functions of metals can be modeled accurately by fitting a multi-
pole Lorentzian model to the experimental data. We show that our fit-
ting routine yields highly accurate results which is demonstrated by fitting
Silver, Nickel, Tungsten, Aluminium and Platinum to experimental data.
The FDTD equations are formulated to include the material dispersion. We
derive an FDTD scheme that only uses polarization and electric fields lead-
ing to a memory efficient algorithm that uses less additional variables than
alternative schemes (Korner and Fichtner 1997). In the case of a dielectric
function that can be modeled as a Drude material we are able to reduce
the number of additional variables needed further, which leads to an algo-
rithm that we believe to use the minimal amount of additional variables
possible. The algorithm is validated by performing several numerical simu-
lations. First we computed the reflection and transmission coefficients from
thin metal layers at normal incidence in vacuum and in an extended ver-
sion of the test the transmission through a Fabry-Perot etalon with metallic
mirrors. Secondly, we determined the reflection spectrum for a surface plas-
mon resonance device and finally we consider the transmission through an
infinite chain of metallic nano-particles.

2. Modeling the dielectric function of metals

When metallic structures are used for FDTD calculations, which cover large
wavelength intervals in the optical and near-infrared regime, the dispersive
properties of such materials can no longer be ignored. Many metals exhibit
a complex dielectric permittivity ε(ω) with negative real part in the optical
range. Because negative values of ε can not be directly included in the stan-
dard set of Maxwell’s equations, special techniques are necessary to allow
the FDTD calculation to proceed.
AN FDTD METHOD FOR METALS 845
In this article, we use the auxiliary differential approach. For this
approach we fit the dielectric function to be considered in the frequency
domain with a function that obeys Kramers–Kronig relations in order to
ensure causality of the FDTD simulation. In the frequency domain we
model the displacement fields in Maxwell’s equations by using the formal-
ism of a macroscopic polarization as follows


D(ω) 
= ε(ω)E(ω) 
= ε0 (ε∞ + χ(ω))E(ω) 
= ε0 (ε∞ E(ω) + P (ω)). (1)

Here we use the frequency dependent susceptibility function χ (ω) and


the static permittivity ε∞ , which is set equal to unity in the subsequent
numerical simulations. The susceptibility function needs to be fitted to
experimental data with model functions that maintain causality of the
FDTD algorithm. We find that a set of Lorentzian functions are a good
choice because they can be used efficiently in the FDTD algorithm and also
describe the susceptibility function well over a large frequency range. We
therefore model the frequency dependent susceptibility as a material with
N poles as follows
N
 Ap
χ (ω) = . (2)
ω + 2iωp − ω2
2
p=1 p

In the above equation, Ap denotes the pole strength, ωp the resonant fre-
quency and p the damping constant, respectively.
A special case of the Lorentzian model is the Drude model, which
describes an oscillator with zero restoring force and is a suitable model to
describe the high frequency behavior of metals. The susceptibility function
for the Drude model is given by

ωP2
χ (ω) = . (3)
ω(iν − ω)
Here, ν denotes the collision rate and ωP the plasma frequency, respectively.
Combining both models allows us to represent the susceptibility function of
metals in the optical and near infrared wavelength range with good accu-
racy and the full material model is then given by

 N
ωP2 Ap
χ (ω) = + . (4)
ω(iν − ω) ωp + 2iωp − ω2
2
p=1

In order to fit the experimental data to our model, we use the adap-


tive non-linear least squares fitting routine NL2SOL, which is described in
Dennis et al. (1981). The algorithm requires two input functions, one that
846 W.H.P. PERNICE ET AL.

Table 1. Fitted parameters for metals in the wavelength range from 300 to 1000 nm.
Aluminium Nickel Platinum Tungsten Silver

Plasma frequency ωp – 9902.1 16096 9183 13349


[THz]
Collision rate ν [THz] – 403.64 1122.1 51.739 49.371
Amplitude A1 [THz2 ] 3.9155 × 108 6.4846 × 109 3.4822 × 108 4.0873 × 107 3.0916 × 1011
Damping constant 1 44.906 96204 7.3396 × 107 812.64 1.6224 × 106
[THz]
Resonance frequency ω1 537.28 21419 1.9579 × 108 2892.4 3.1437 × 104
[THz]
Amplitude A2 [THz2 ] 3.7422 × 107 2.2844 × 107 8.4141 × 108 1.6915 × 109 –
Damping constant 2 1565.9 243.66 1.1948 × 104 7933.7 –
[THz]
Resonance frequency ω2 4246.5 886.13 9489.5 8689.4 –
[THz]
Amplitude A3 [THz2 ] 4.1708 × 107 1.5450 × 107 – 1.2108 × 106 –
Damping constant 3 761.23 748.99135 – 101.34 –
[THz]
Resonance frequency ω3 2959.1 2545.5447 1378.3 –
[THz]
Amplitude A4 [THz2 ] 5.9605 × 107 1.5852 × 107 – 16179 × 107 –
Damping constant 4 343.34 691.85 – 376.96 –
[THz]
Resonance frequency ω4 2368.8 2016.1 – 1569.7 –
[THz]

calculates the residual error for the current fitting status and another one
that obtains the Jacobian matrix for the function model to be fitted. An
alternative fitting routine is described by Vial et al. for the dielectric func-
tion of gold using a Drude-Lorentzian model with one Lorentzian pole.
Using NL2SOL for the fit yields results with fitting errors of only a few
percent. The fitted parameters for the following examples are given in
Table 1.
In Fig. 1, we show the obtained result for platinum.
It can be seen from above figure that the model is able to describe both
the real and the imaginary part of the susceptibility function with high
accuracy. For the above model a Drude–Lorentzian material function with
three poles was used.
In Fig. 2, we show results for fitting the susceptibility function of silver.
We show two fits, one obtained for a pure Drude model and a second
one using an additional Lorentzian pole. Choosing the Drude model, the
resulting error is in the range of 10% whereas adding an additional pole
reduces the error to 5%.
Because every additional pole leads to a higher computational load in
the FDTD calculation, we are interested in using the smallest number of
AN FDTD METHOD FOR METALS 847

Fig. 1. Fitting result for the susceptibility function of platinum in the optical and near infrared wave-
length spectrum.

Fig. 2. Fitting result for silver. Shown are the fits for a Drude model and a combined Drude–Lorentz
model.

poles necessary. We find that one pole is needed for every resonance peak
in the frequency spectrum, in order to get the shape of the susceptibility
function correct. However, when the magnitude of the pole is also impor-
tant, additional terms in the expansion are required. This is illustrated in
Fig. 3 for aluminium.
Using a simple Drude model, we obtain fitting errors of the order
of 10%. Adding another Lorentzian pole resolves the resonance peak at
800 nm, but does not give an accurate fit for the imaginary part. Adding a
second Lorentzian pole yields errors that are now of the order of 1%. We
therefore conclude that roughly two poles are required to model both the
magnitude and the resonance frequency of a peak in the frequency spec-
trum. In the following section we describe how to incorporate the fitted
functions into the FDTD equations.
848 W.H.P. PERNICE ET AL.

Fig. 3. Dependence of the fitting result on the number of poles used. The fit was performed for
aluminium.

3. The FDTD scheme

In order to include the Lorentzian poles in the FDTD equations, we choose


the auxiliary differential equation approach. Each pole in the frequency
domain will lead to an additional differential equation in the time domain
with each pole treated as a separate polarization contribution. Transform-
ing the susceptibility function into the time domain yields

∂t2 Pp + 2p ∂t Pp + ωp2 Pp = Ap E.


 (5)

The susceptibility function for a Drude material can be rewritten as a


Debye model together with an effective static conducitivty as follows

ωP2 ωP2 σ
χ(ω) = =− + . (6)
ω(iν − ω) ν(iω + ν) iε0 ω

In Equation (6), we have defined an effective static conductivity σ = ε0 ωP2 /ν.


The static conductivity can be added to Maxwell’s equations as a current
term without the need to allocate memory for previous polarization fields
and results in a formulation for a Drude material that is more memory effi-
cient.
Transforming the first term on the right hand side of Equation (6) into
the time domain yields the FDTD scheme for a Drude pole as

ωP2 
∂t PD + ν PD = − E. (7)
ν
AN FDTD METHOD FOR METALS 849
Then together with the standard set of Maxwell’s equations we obtain the
coupled differential equations
  
ε∞ ∂t E + ∂t PD + N ∂ P
p=1 t p
 = 1
ε0
∇ × 
H − σ 
E
(8)

∂t H = − 1 ∇ × E.
µ0

The above equations are discretized in the following fashion for a Drude-
Lorentzian material with one Drude and N Lorentzian poles
  t
H n+ 2 − H n− 2 = − ∇ × E n ,
1 1

µ0
  N 
 
ε∞ E n+1 − E n + PDn+1 − PDn + Ppn+1 − Ppn
p=1
t  σ   n+1  n 
∇ × H n+ 2 −
1
= E +E ,
ε0 2
P n+1 + PDn ω2  
PDn+1 − PDn + νt D = − P t E n+1 + E n ,
2 2ν
   
Ppn+1 − 2Ppn + Ppn−1 + p t Ppn+1 − Ppn−1 + ωp2 (t)2 Ppn

= Ap (t)2 E n , p = 1, 2, · · · (9)

The spatial derivatives in the curl operator are evaluated in standard Yee-
fashion, as described for example by Taflove (1995). These equations can be
easily solved to yield explicit expressions for the new polarization, electric
and magnetic fields. Our scheme requires storage for only the electric and
polarization fields, thus eliminating the need to keep displacement fields in
memory. For a Lorentzian pole we therefore require only two additional
variables, P n and P n−1 , as opposed to three additional variables, D n , P n
and P n−1 , as in the scheme presented by Korner et al. Furthermore, we
note that transforming the Lorentzian formalism of a Drude material to the
extended Debye formalism allows us to simulate a number of nobel met-
als with minimum memory usage. Our scheme requires only one additional
variable, P n , in comparison to three additional variables, P n , P n−1 and D n ,
for the FDTD scheme presented by Korner et al. Because memory is the
limiting factor for FDTD simulations, we are able to simulate larger struc-
tures than alternative FDTD algorithms. For the model of silver described
below, our scheme will require three additional variables as opposed to five
for the FDTD scheme of Korner. Reducing the number of additional vari-
ables makes our FDTD scheme superior in both storage and speed, because
the run-time of FDTD simulations is largely affected by memory through-
put through the CPU.
850 W.H.P. PERNICE ET AL.

4. Numerical simulations

In order to validate our FDTD scheme we performed a number of numer-


ical simulations described in detail in the following sections.

4.1. Calculation of reflection coefficients for thin metal layers

The reflection coefficient for thin metal layers is obtained by performing


two separate FDTD calculations. In the first computation we determine
the incident fields by performing the FDTD simulation without the dis-
persive material present. The electromagnetic fields are sampled closely on
both sides of the proposed position of the metal sheet. The second com-
putation is performed with the metal sheet included. Subtracting this result
from the first yields the reflected fields. The reflection coefficient can then
be computed as the ratio of the discrete Fourier transforms of the incident
and reflected fields. A plane wave with gaussian temporal profile centered
around 0.5 microns is launched from the left hand side of the computa-
tional domain. We choose a grid spacing of 5 nm and propagate the fields
for 100 fs, corresponding to 10400 time steps.
In a first simulation we investigated the reflection at normal incidence
from a 15 nm thick aluminium film in vacuum. The results of this simula-
tion are presented in Fig. 4.
The metal was modeled as a Lorentzian material with four poles. The fit-
ting parameters for the complex susceptibility are given in Table 1. The rms
fitting error was 0.8% for the real part and 1.2% for the imaginary part and
therefore agrees very closely with the experimental data. The results from

Fig. 4. The reflection and transmission spectrum calculated for a 15 nm thick aluminium film. The
aluminium was modeled as a Lorentzian material with 4 resonances.
AN FDTD METHOD FOR METALS 851

68 60

Transmission coefficient in %
Reflection coefficient in %

64 55
60
50
reflection, matrix method
56 reflection, FDTD
transmission, matrix method 45
52 transmission, FDTD
48 40

44 35
40
30
200 300 400 500 600 700 800 900 1000
Wavelength in nm

Fig. 5. The reflection and transmission spectrum calculated for a 50 nm thick platinum film. The metal
was modeled as a Drude–Lorentzian model with two Lorentzian resonances.

the FDTD calculation can be compared to analytical results, which can be


calculated using the matrix method described by Tubb et al. (1997). The
agreement between both methods is excellent.
As a second example, we calculated the reflection and transmission at
normal incidence through a 50 nm platinum layer in vacuum. The results
are shown in Fig. 5.
The platinum was modeled as a Drude–Lorentzian material with one
Drude and two Lorentzian poles. The agreement between the exact matrix
method and the FDTD method is again excellent.

4.2. Calculation of transmission through a Fabry–Perot etalon

In an extended version of the previous experiment, we simulated the trans-


mission through a Fabry–Perot resonator. The structure under test was
composed of an air cavity of 1.05 micron length surrounded by metallic
mirrors. The cavity is embedded in quartz glass with a refractive index of
1.477. As a metal we chose tungsten, which can be fitted by our algorithm
to high accuracy using a combined Drude–Lorentz material with 5 poles.
The fitted parameters are given in Table 1. The cavity was excited with a
plane wave with Gaussian temporal profile centered around 1.1 microns.
A uniform mesh of 5 nm was used to discretize the computational domain.
Because light trapped in the cavity takes a long time to decay, we ran the
simulation for 2000 fs, corresponding to 209800 time steps.
The FDTD results were compared to computations carried out using
the commercial optical modeling software FIMMWAVE, which calculates
the transmission by eigenmode expansion (Gallagher and Felici, 2003). We
852 W.H.P. PERNICE ET AL.

Fig. 6. Transmission through a metallic Fabry–Perot etalon with varying mirror width. The FDTD
results are shown by solid and dashed lines, the points were calculated using FIMMWAVE.

investigated different thicknesses of the metallic mirrors ranging from 10 nm


to 40 nm. The results are presented in Fig. 6.
For all structures we observe excellent agreement with the results from
the Eigenmode expansion method.

4.3. Calculation of reflection spectrum for a SPR device

In this simulation we consider a device that shows surface plasmon


resonance (SPR) phenomena. SPR devices have been studied in depth and
analytical solutions are available Tubb et al. (1997). This example was cho-
sen because the determination of the resonance frequency is highly sensitive
to the calculation parameters and therefore provides a serious test for our
FDTD algorithm. We investigate an SPR device in the Kretschmann con-
figuration shown in Fig. 7.
Three materials are present, the incident medium with a refractive index
of 1.9, the substrate with a refractive index of 1.132 and a metallic thin
film of 50 nm thickness. In this example we used silver and fitted the com-
plex dielectric function using a Drude model. Using a plasma frequency
of 12120.749 THz and a collision rate of 116.10586 THz yields a good fit
over a wide wavelength spectrum. In order to detect the resonant frequency
of the metal film we could either scan the incident angle θ of the field
or equivalently the wavelength of the exciting source. Sweeping the angle
would require us to run a separate FDTD calculation for each angle. We
therefore chose the wavelength scan, because we are simulating in the time-
domain and therefore only need to run one calculation in order to obtain
the reflection coefficient at each wavelength.
AN FDTD METHOD FOR METALS 853
x

substrate
metal film
z
incident
inciden θ
medium

Fig. 7. Kretschmann geometry for the SPR device used in the FDTD simulations.

The simulation was performed with a grid spacing of 40 nm in both


the x and z directions with a finer mesh size of 5 nm in the metallic
areas. The total simulation time was chosen as 283.3 fs with a time step
of 0.009533 fs resulting in a total of 25000 time-steps. This simulation time
allows the fields to decay sufficiently after propagating through the device.
A p-polarized wave is injected from the incident medium with a narrow
Gaussian temporal profile with a central wavelength of 400 nm. The inci-
dent angle θ was set to 43.5 degrees. The incident and reflected fields are
recorded in order to allow us to calculate the reflection coefficient for the
structure. In Fig. 8 we show the calculated fields after 10000 time-steps.

Fig. 8. Electric field computed after 10000 time-steps during the FDTD calculation. The surface plas-
mon can be seen propagating along the top of the metal film.
854 W.H.P. PERNICE ET AL.

Fig. 9. Reflection spectrum obtained using the FDTD method. Also shown is the exact reflectivity com-
puted using the matrix method.

The excited surface plasmon is clearly visible propagating on the top of


the silver film resulting in a frequency dip in the reflection spectrum. The
calculated spectrum is shown in Fig. 9.
The FDTD results are compared to a reflection spectrum obtained using
the exact matrix method described by Tubb et al. (1997). The agreement
between both methods is very good. In particular, the resonant frequency
of the surface plasmon was accurately predicted by our FDTD method.

4.4. Calculation of transmission through an infinite chain


of metallic nano-particles

As a last numerical example we investigated the transmission coefficient


for a chain of elliptical Nickel nano-particles of varying size. The chain is
embedded in a non-dispersive medium of refractive index 1.0. The simula-
tion geometry is shown schematically in Fig. 10.
We chose the short axis of the particles to be 100 nm and varied the long
axis from 100 to 300 nm. The infinite chain was modeled in the FDTD by
applying perfectly matched layers (PMLS) at the top and bottom boundary
and periodic boundary conditions on the left and right side. The computa-
tional domain size was chosen to be 2000 × 580 nm. Nickel was modeled as
a Drude–Lorentzian material with five poles yielding very close agreement
with the experimental data.
In order to reduce stair-casing errors from approximating the elliptic
particle, we used a fine grid resolution of 2 nm. A plane wave with Gauss-
ian temporal profile centered around 0.5 microns was launched from the left
AN FDTD METHOD FOR METALS 855

Fig. 10. Computational geometry for the simulation of an infinite chain of nano-particles.

Fig. 11. Transmission spectra calculated for varying values of the long axis of a Nickel nano-particle.
Solid lines are results from the FDTD, markers are FIMMWAVE calculations.

hand side. Fields were monitored every time step before and after the parti-
cle. The fields were propagated for a total of 150 fs, corresponding to 39300
time steps. The results are presented in Fig. 11.
By comparing it to the solution obtained with FIMMWAVE we observe
again very close agreement between the two different calculation methods.

5. Summary

We have presented a formulation of the finite-difference time-domain


method that is capable of accurately simulating metals. The FDTD algo-
rithm is based on the auxiliary differential equation approach. This approach
is coupled with a fitting routine that fits the complex dielectric function of
metals with a set of Lorentzian and Drude poles, yielding errors of only a
856 W.H.P. PERNICE ET AL.

few percent. We have shown that the proposed algorithm can be used to
accurately model the reflections from thin metal layers, elliptical particles,
and also to analyze surface plasmon resonance devices.

Acknowledgements

Wolfram Pernice would like to thank Photon Design Ltd. for the support
of this research.

References

Dennis, J.E. Jr., D.M. Gay and R.E. Welsch. ACM Trans. Math. Softw. 7(3) 348, 1981.
Fan, G.X. and Q.H. Liu. IEEE T. Antenn. Propag. 48(5) 637, 2000.
Gallagher, D.F.G. and T.P. Felici. Proc. SPIE 4987 69, 2003.
Korner, T.O. and W. Fichtner. Opt. Lett. 22 1586, 1997.
Luebbers, R., F. Hunsberger, K. Kunz, R. Standler and M. Schneider. IEEE Trans. Electromagn.
Compat. 32 222, 1990.
Luebbers, R. and F. Hunsberger. IEEE T. Antenn. Propag. 40 1297, 1992.
Taflove, A. Computational Electrodynamics: The Finite-Difference Time-Domain Method. Artech
House, Boston, MA, 1995.
Tubb, A.J.C., F.P. Payne, R.B. Millington and C. Lowe. Sensors Actuators B 41 71, 1997.
Vial, A., A. Grimault, D. Macials, D. Barchiesi and M.L. de la Chapelle. Phys. Rev. B 71 085416,
2005.
Yee, K.S. IEEE T. Antenn. Propag. 14(3) 302, 1966.
Young, J.L. and R.O. Nelson. IEEE Antenn. Propag. Magazine 43(1) 61, 2001.

You might also like