Sie sind auf Seite 1von 4

IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS, VOL.

10, 2011

143

Comparative Study of FDTD-Adopted Numerical


Algorithms for Kerr Nonlinearities
Ivan S. Maksymov, Andrey A. Sukhorukov, Andrei V. Lavrinenko, and Yuri S. Kivshar

AbstractAccurate finite-difference time-domain (FDTD) modeling of optical pulse propagation in nonlinear media usually implies the use of auxiliary differential equation (ADE) techniques.
The updating of electric field in full-vectorial 3-D ADE FDTD modeling of the optical Kerr effect and two-photon absorption in optical media is proceeded conventionally through the iterative solution of nonlinear algebraic equations. Here, we study three approaches for the field update including simple noniterative explicit
schemes. By comparing them to the analytical results for optical
pulse propagation in long nonlinear media (nonlinear phase incursion for the pump wave of about radians), we demonstrate convincingly that simple noniterative FDTD updating schemes, which
are commonly believed to be inaccurate and unstable, produce accurate results and drastically speed up the computation as compared to ADE approaches. Such schemes can significantly reduce
the CPU time for nonlinear computations, especially in 3-D models.
Index TermsFinite-difference time domain (FDTD), four-wave
mixing (FWM), nonlinearity, optical Kerr effect.

I. INTRODUCTION
HE CAPABILITY of a finite-difference time-domain
(FDTD) method to solve problems of electromagnetic
wave interactions with materials having frequency-dependent nonlinear optical properties [1][3] keeps this method
among the most universal and powerful numerical tools for
optics, electrodynamics, antennas, and waveguides theory.
An essential part of every modern FDTD simulator is the
implementation of auxiliary differential equation (ADE) FDTD
techniques that drastically broaden the use of the FDTD method
and its calculation accuracy. Recently, Dissanayake et al. [4]
proposed an extended 3-D FDTD scheme for modeling optical
phenomena in silicon waveguides, which takes into account
material anisotropy of the third-order electronic and Raman
susceptibilities. Such FDTD features make it possible to

Manuscript received December 01, 2010; revised January 15, 2011; accepted
February 04, 2011. Date of publication February 14, 2011; date of current version March 14, 2011. This work was supported by the Australian Research
Council. The work of A. V. Lavrinenko was supported by the Danish Research
Council for Technology and Production Sciences under the THz COW Project
and COST Action MP0702.
I. S. Maksymov, A. A. Sukhorukov, and Y. S. Kivshar are with the Nonlinear Physics Centre and Centre for Ultrahigh Bandwidth Devices for Optical
Systems (CUDOS), Research School of Physics and Engineering, Australian
National University, Canberra, ACT 0200, Australia (e-mail: mis124@physics.
anu.edu.au).
A. V. Lavrinenko is with the Department of Photonics Engineering,
Technical University of Denmark, 2800 Kongens Lyngby, Denmark (e-mail:
alav@fotonik.dtu.dk).
Color versions of one or more of the figures in this letter are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LAWP.2011.2114319

perform numerical investigations beyond the restrictions of


the single-mode or slowly varying envelope approximations.
ADE FDTD schemes for materials with multiple-pole linear
Lorentz, instantaneous Kerr, two-photon absorption-induced,
and Raman nonlinear polarizations require finding the solution
of a system of nonlinear coupled equations at every time step.
In the case of anisotropic media, the coupled update equations
for the electric field components involve three auxiliary variables [4] that, in turn, are nonlinear functions of all three electric
field components. Additionally, the denominators of two out of
three auxiliary variables contain electric field terms that can, in
principle, diverge if corresponding field values approach zero.
Therefore, general 3-D FDTD modeling of nonlinear waves
in nonlinear waveguides might encounter numerical problems
with the proposed iterative algorithm.
To the best of our knowledge, there exist at least two alternative FDTD approaches accounting for the instantaneous
Kerr effect [1], [5][8], which allows one to avoid any iterative scheme in the electric field updating. The updated electric
in the first approach [1], [6][8] relies on the elecfield
from the previous time step ; it also implies the
tric field
at the curknowledge of the electric displacement vector
rent
th time step. Interestingly, the validity of this approach is confirmed indirectly by the fact that sufficient accuracy of the ADE iterative scheme may be obtained after just
chosen as the initial
one iteration [9] with the electric field
guess. The second noniterative approach analytically solves a
values for given
cubic equation [5], [10] to determine the
and
. Both noniterative approaches were supposedly
proclaimed to be inaccurate and unstable and are nearly abandoned in recent publications [4].
In this letter, we perform a comprehensive comparison of
both approaches against the iterative one and the approximate
analytical solution to clarify their constraints and limits in applications. The findings support our hypothesis that the noniterative schemes are stable and reasonably accurate. They can
be recommended for a broad range of problems especially addressing the designs with complex 3-D topology, where iterative solutions of coupled nonlinear equations can be unstable
and resource-demanding.

II. MAXWELLS EQUATIONS


We start with Maxwells equations describing the propagation of light in nonlinear media [11]:
and
. We aim to
compare FDTD algorithms for Kerr nonlinearities, so we consider only the nonlinear effects associated with the third-order

1536-1225/$26.00 2011 IEEE

144

IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS, VOL. 10, 2011

susceptibility tensor
. The induced nonlinear polarization is
related to the electric field as

Newtons method is used to determine zeros of the objective


function (6) with the initial guess chosen to be equal to the elecknown from the previous time step.
tric field
B. Noniterative Updating Scheme

(1)
which is valid in the electric-dipole approximation assuming
that the medium response is local [11]. For nonresonant,
incoherent (intensity-dependent) nonlinear effects, the
third-order nonlinear polarization can be described by the
BornOppenheimer approximation [12] as
(2)

where
is the nonlinear response function normalized in a
manner similar to the delta function [11]. The upper limit of integration in (2) extends only up to because the response function
must be zero for
to ensure causality. According to [11], the nonlinear response function
, where represents the relative strength of the
is a Dirac delta function
Kerr and Raman polarizations and
that models the instantaneous Kerr nonresonant virtual transiinto (2) gives the polarization and
tions. The substitution of
)
the polarization current for the Kerr nonlinearity (for
(3)
(4)
Hereafter, Equation (4) is referred to as auxiliary.

Now, we formulate the first noniterative electric-field updating scheme [1], [6][8] (referred to as Method 2), which
will be compared against the iterative updating scheme. We
note that the numerical integration in (2) up to the previous
naturally ensures the fulfillment of the causality
time step
principle. It makes it possible to rewrite (3) as
(7)
By substituting (7) into the constitutive equation
, we obtain the following noniterative
electric-field updating equation for
:
(8)

C. Noniterative Updating Scheme With Cubic Equation


The second noniterative scheme, referred to as Method 3
[5], [10], analytically solves a cubic equation derived directly from the constitutive equation
as

(9)
where
. Once is found, the electric field is
.
calculated as
In Methods 2 and 3, we have to keep the displacement field
, which is updated through Amperes Law.

III. DISCRETIZATION
We consider a 1-D model that describes the evolution of the
and
. A part of the nonlinear FDTD
field components
method in updating the -fields comprises the following explicit equation, which is the same in all three algorithms:
(5)
where the amplitudes of the pump
and signal
waves
at the th time step are defined by initial conditions. In the fol. Herelowing, we compare different methods for updating
after, we formulate three
updating schemes corresponding
to the compared methods.
A. ADE Updating Scheme
In the first algorithm considered in our letter, referred to
as Method 1, we obtain a finite-difference expression for the
ADE (4) as
and determine the electric field
by iterations
from the Ampere Law [2] as

(6)

IV. RESULTS
To perform the testing, we compare the performance of the
three nonlinear FDTD algorithms applied to the analysis of
parametric four-wave mixing (FWM) in an isotropic optical
Kerr medium using the numerical conditions of [13]. FWM
is a nonlinear process arising from the third-order optical
nonlinearity [11]. If a strong pump wave at the frequency
and a weak signal wave at the frequency
propagate together in a nonlinear medium, two additional waves
are generated due to a refractive index modulation with the
and
. FWM transfers
frequencies
energy from a strong pump wave to the waves upshifted and
downshifted in frequency from the pump frequency [11].
We use an analytical asymptotic approximate formula that
expresses the FWM signal conversion efficiency as a function
of the pump electric field amplitude
after the propagation
distance as [13]
(10)
where the linear refractive index of the entire medium is chosen
to be
, the nonlinear susceptibility is
m V , and is the speed of light.

MAKSYMOV et al.: COMPARATIVE STUDY OF FDTD-ADOPTED NUMERICAL ALGORITHMS FOR KERR NONLINEARITIES

145

Fig. 1. Power frequency spectrum for the FDTD electric field detected at the
:
V/m. Unlike [13], a Gaussian filter is applied to
output for E
facilitate comparison of the methods below 140 dB. (left inset) Enlarged plot
at 189 THz. (right inset) Power spectrum at 189.2 THz as a function of the
spatial discretization step x.

Fig. 2. Signal-conversion efficiency in the nonlinear Kerr medium for the prop; ; and m as a function of the pump electric
agation distances L
field amplitude E . (inset) Nonlinear refractive index change as a function of
V/m.
E . Notice hardly attainable in practice n values for E >

Fig. 1 shows typical power frequency spectra obtained with


m.
Methods 13 for the nonlinear medium length
The presented data are detected at the output. The spectra are
V/m
calculated for the pump wave amplitude
and normalized such that the peak value of the pump field power
at 192 THz is 0 dB. According to the initial conditions, the signal
wave (194.8 THz) is taken at 20 dB lower in intensity than the
pump wave. A brief overview of the spectrum shows that there is
a negligible discrepancy (absolute error of 0.08% at 189.2 THz)
between the results produced by Method 2 and Method 3. Since
the almost undistinguishable behavior of both methods has been
in the range
observed for the pump wave amplitudes
to
V/m, we will not further consider Method 3,
leaving for the analysis and comparison Methods 1 and 2 only.
The discrepancy between the results obtained with Method 1
and Method 2 is also very small; it becomes visible only in the
enlarged plot (the left inset in Fig. 1). Even below 140 dB,
both methods produce very close results. The converted signals observed at 189.2 THz are in a good agreement with the
approximate FWM theory [13] [compare analytically predicted
29.7 dB with 29.8 dB (Method 1) and 29.6 dB (Method 2)
relative to the signal wave].
To continue with the nonlinear analysis, we test our numerical schemes for convergence. The right inset in Fig. 1 shows the
convergence of Method 1 (solid line) and Method 2 (dashed line)
. A remarkable
as a function of the spatial discretization step
result is that both methods converge for
, where
is the pump wavelength. Nevertheless, both methods diverge
because of the round-off errors in the computaat smaller
tion [14]. Unlike its opponent, Method 2 diverges substantially
faster. This difference is attributed to higher numerical noise
levels produced by Method 2 with regard to Method 1. If it were
desired to improve the performance of Method 2 at such small
, this could be done by decreasing the Courant number [1]
and optimizing the parameters (such as conductivity profile
and number of layers) of perfectly matched layers (PMLs)
absorbing boundary conditions [15]. However, the use of step
is usually more than enough for the most exigent
FDTD simulations [1].

We further test both FDTD schemes for energy conservation.


, where energy
We define a confidence factor
and
are
fluxes
calculated from instantaneous magnitudes of
and
detected at the output and input, respectively [13].
(98.5%) is found for
V/m with Method 1
V/m, we obtain 97.2% and 97%.
(Method 2). For
Since the losses increase as
increases, they are attributed to
a PML numerical artifact often observed in numerical models at
high power levels [15].
Fig. 2 compares the signal-conversion efficiencies obtained
by both methods as a function of the pump wave amplitude for
different nonlinear medium lengths (the number of time steps
increases up to more than 10 as for larger ). Notice that for
V/m and
m, the nonlinear phase
incursion for the pump wave is radians. All the FDTD results are in a good agreement with the predictions of the analytical approximate formula (10) within the region of validity
of the asymptotic approximation. It enables us to demonstrate
clearly that both methods produce comparable results even in
the regime when the conversion efficiency saturates and the approximate analytical formula (10) is no longer valid. To the best
of our knowledge, neither of the methods has been tested under
this condition previously.
A characteristic feature of Method 1 is the determination of
the updated electric field from (6) by iterations. According to
textbooks [14], the iteration process is stopped when an accuis reached. If the pump wave ampliracy of
tude is low, then such accuracy is attained in just two to three
iterations, but this number is typically larger for stronger pump
waves. To characterize the speed of both methods, we evaluate a run-time for one update of the fields in one grid point:
, where
and
denote, respectively, the total number of nodes in the FDTD grid
and the total number of time steps. Fig. 3 shows the FOM as a
function of the pump electric field amplitude. The computation
time required by Method 1 (solid line) gradually increases as the
V/m is due to
pump power is increased. A jump at
an increase in the number of iterations in electric field updating

= 5 1 2 10

= 10 100

400
1

110

146

IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS, VOL. 10, 2011

V. CONCLUSION
We have demonstrated convincingly that the optical Kerr
effect can be accurately modeled by a completely explicit
numerical scheme using the previous electric field value for
updating nonlinear FDTD equations. Being implemented with
other ADE FDTD techniques, this approach is expected to drastically accelerate computations, especially in multidimensional
models. On the other hand, we have confirmed higher accuracy
and stability of the standard iterative ADE FDTD scheme
formulated using polarization currents. The latter must be used
if the accuracy is of primary importance. We anticipate that
similar conclusions apply to the case of two-photon absorption
FDTD modeling [16].
Fig. 3. FDTD computation time for one spatial and time grid point as a function
of the pump electric field amplitude E for Method 1 (solid line) and Method 2
(dashed line). Insets: Absolute error of (a) Method 1 and (b) Method 2 for L
m as a function of E . Absolute error of (c) Method 1 and (d) Method 2
for L
m as a function of E .

10

= 100

subroutine. In contrast, the time required by Method 2 (dashed


.
line) is much lower and almost independent of
We turn our attention to the accuracy of the compared algorithms. We calculate the absolute error as the deviation of
computed conversion efficiencies from approximate analytical
values obtained with (10). Since (10) fails to predict conversion
m, the latter case is excluded from
efficiencies for
further discussions. We also do not pay much attention to the
V/m because such amplitudes lead to
region of
unfeasible nonlinear refractive index changes (see the inset in
Fig. 2) and the conversion efficiency saturation.
The inset in Fig. 3 shows the absolute error versus pump
power graphs of Method 1 (left column) and Method 2 (right
m (top row)
column) for the nonlinear medium lengths
m (bottom row). The absolute error of Method 1
and
averages about 0.3 dB for
m and
m; that
m, but attains 1 dB
of Method 2 is of about 0.3 dB for
m.
for
It is also instructive to examine the relative error of the
methods with respect to the analytical solution. On average, at
189.2 THz (see Fig. 1), the relative error of Method 1 is about
0.6%; that of Method 2 is between 0.6% and 2%. However,
as the main panel of Fig. 3 shows, the computation time of
Method 1 increases tenfold with regard to that of Method 2.
We recall that these results have been obtained with a 1-D
model. Computation costs of the noniterative Method 2 for 2-D
and 3-D models can be estimated keeping the time per one
FDTD grid point the same as in 1-D and scaling only the number
in 2-D and
in 3-D. We predict a conof grid points as
siderably larger, nonlinear increase in computation time for 2-D
and 3-D ADE FDTD algorithms attributed to additional efforts
in solving systems of nonlinear coupled equations.

REFERENCES
[1] R. W. Ziolkowski, The incorporation of microscopic material models
into the FDTD approach for ultrafast optical pulse simulations, IEEE
Trans. Antennas Propag., vol. 45, no. 3, pp. 375391, Mar. 1997.
[2] J. H. Greene and A. Taflove, General vector auxiliary differential
equation finite-difference time-domain method for nonlinear optics,
Opt. Exp., vol. 14, no. 18, pp. 83058310, Sept. 2006.
[3] M. Fujii, M. Tahara, I. Sakagami, W. Freude, and P. Russer, Highorder FDTD and auxiliary differential equation formulation of optical
pulse propagation in 2-D Kerr and Raman nonlinear dispersive media,
IEEE J. Quantum Electron., vol. 40, no. 2, pp. 175182, Feb. 2004.
[4] C. M. Dissanayake, M. Premaratne, I. D. Rukhlenko, and G. P.
Agrawal, FDTD modeling of anisotropic nonlinear optical phenomena in silicon waveguides, Opt. Exp., vol. 18, no. 20, pp.
2142721448, Sep. 2010.
[5] P. Tran, Photonic-band-structure calculation of material possessing
Kerr nonlinearity, Phys. Rev. B, Condens. Matter, vol. 52, no. 15, pp.
1067310676, Oct. 1995.
[6] E. P. Kosmidou and T. D. Tsiboukis, An FDTD analysis of photonic
crystal waveguides comprising third-order nonlinear materials, Opt.
Quantum Electron., vol. 35, no. 10, pp. 931946, Aug. 2003.
[7] I. S. Maksymov, L. F. Marsal, and J. Pallares, Finite-difference timedomain analysis of band structures in one-dimensional Kerr-nonlinear
photonic crystals, Opt. Commun., vol. 239, no. 13, pp. 213222, Sep.
2004.
[8] D. Pinto, S. S. A. Obayya, B. M. A. Rahman, and K. T. V. Grattan,
FDTD analysis of nonlinear bragg grating based optical devices, Opt.
Quantum Electron., vol. 38, no. 13, pp. 12171235, Jan. 2006.
[9] P. M. Goorjian, A. Taflove, R. M. Joseph, and S. C. Hagness, Computational modeling of femtosecond optical solitons from Maxwells
equations, IEEE J. Quantum Electron., vol. 28, no. 10, pp. 24162422,
Oct. 1992.
[10] I. S. Maksymov, L. F. Marsal, and J. Pallares, An FDTD analysis of
nonlinear photonic crystal waveguides, Opt. Quantum Electron., vol.
38, no. 13, pp. 149160, Jan. 2006.
[11] G. P. Agrawal, Nonlinear Fiber Optics, 3rd ed. San Diego, CA: Academic, 2001.
[12] R. W. Hellwarth, Third-order optical susceptibilities of liquid and
solids, J. Prog. Quantum Electron., vol. 5, no. 1-A, pp. 168, 1977.
[13] M. Fujii, C. Koos, C. Poulton, I. Sakagami, J. Leuthold, and W.
Freude, A simple and rigorous verification technique for nonlinear
FDTD algorithms by optical parametric four-wave mixing, Microw.
Opt. Technol. Lett., vol. 48, no. 1, pp. 8891, Jan. 2006.
[14] R. L. Burden and J. D. Faires, Numerical Analysis, 5th ed. Boston,
MA: PWS, 1993.
[15] A. P. Zhao, Application of the material-independent PML absorbers
to the FDTD analysis of electromagnetic waves in nonlinear media,
Microw. Opt. Technol. Lett., vol. 44, no. 3, pp. 165168, Feb. 1998.
[16] I. S. Maksymov, L. F. Marsal, and J. Pallares, Modeling of two-photon
absorption in nonlinear photonic crystal all-optical switch, Opt.
Commun., vol. 269, no. 1, pp. 137141, Jan. 2007.

Das könnte Ihnen auch gefallen