6th European Conference on Antennas and Propagation (EUCAP)

Moving-Window Propagation Model Based on an

Unconditionally Stable FDTD Method
Cláudio Garcia Batista∗ and Cássio Gonçalves do Rego†
∗ GraduateProgram in Electrical Engineering, † Department of Electronic Engineering
Federal University of Minas Gerais - Av. Antonio Carlos 6627, 31270-901, Belo Horizonte, MG, Brazil
e-mails: {claudio, cassio}@cpdee.ufmg.br

Abstract—This work introduces a finite-difference time-domain of the FDTD computation domain (meaning that is not a
(FDTD) propagation model based on a moving window algorithm. function of the dielectric properties of the material). Thus, the
The FDTD is evaluated by an unconditionally stable (US) method total simulation time can be drastically reduced by the use of
combined with a material independent (MI) perfectly matched
layer (PML) formulation. Thus, the time step used in simulation bigger time steps and the formulation can be efficiently applied
is no longer restricted by the Courant-Friedrich-Levy (CFL) to any real propagation scenarios (mixed paths, presence of
stability condition and the formulation can be efficiently applied obstacles, atmosphere effects, etc.) without any modification.
to any real propagation scenario without any modification. The In order to validate the propagation method, we first test
US-IPML formulation is tested through an analytic problem. The the US-IPML formulation through the problem of a perfectly
proposed propagation model is applied to two idealized terrain
profiles involving HF and VHF signals. conducting cylinder illuminated by a 2D current line source.
After that, we study two idealized mixed terrains (Gaussian
I. I NTRODUCTION and wedge-shaped profiles) for HF and VHF bands. We
Nowadays an accurate characterization of the radio channel compare the path loss prediction with well-known propagation
is constantly required when planning wireless networks. The methods in the literature: Ott’s [6] formulation for HF and a
overwhelming demand for faster data transmission rates and Method of Moments (MoM) model [7] for VHF.
wideband signals increases the number of studies involving
II. P ROPAGATION M ODEL
radio wave coverage prediction. The use of full wave analytic
models is becoming more attractive due to development of A. Moving window algorithm
cheaper and faster personal computers. Nevertheless, the tra- The moving window concept is illustrated by Figure 1,
ditional techniques such as the Method of Moments (MoM), where a typical radio propagation scenario shows a transmitter
Parabolic Equation (PE), transmission line matrix (TLM), (TX) and a receiver (RX) in the presence of an irregular
forward-backward method (FBM) are developed for a limited terrain. The FDTD mesh needs to be long enough to contain
frequency bandwidth, i.e., it do not include all the propagation the dispersed pulse and all of the FDTD computation space
mechanisms specific for each band [1]. outside of this virtual window is not considered. As the pulse
A FDTD application on propagation analysis was presented propagates toward the point of interest, the window moves
by Sevgi et al. [2] and Luebbers et al. [3] based on a bidimen- along with the pulse and the appropriate path parameters
sional moving window (MW) algorithm. Although the FDTD (terrain elevation and constitutive values, atmosphere effects,
has a full wave electromagnetic calculation feature, the directly etc) are added to the mesh [2], [3]. In order to implement the
application on propagation situations becomes impracticable algorithm, the moving window is terminated by PML cells
due the memory and computing time required. Thus, the MW from left, bottom and top. On the right side a PML termination
algorithm simulates a broadband pulse propagating over long is not necessary, once the window must follow the pulse and
distances by a virtual window that moves along the path. The avoid that significant energy cross this interface.
FDTD mesh is constructed only in the virtual window that The window dimensions must be chosen observing two
represents a small portion of the computation domain. As a factors: the initial size and dispersion of the pulse along
time domain method, the FDTD can model signals over a the axis x̂ and a minimum number of Fresnel zones above
large frequency range and may be viewed as a multi-band the higher terrain/obstacle in ŷ direction. To include such
propagator. effects one can vary the window size dynamically along the
In this work, we presented a 2D time-domain propagator propagation path. Basically, the algorithm is performed by the
that adopts the moving window algorithm and, different from following steps: 1) Construct a W S × Wy size FDTD window
[2],[3], we use an unconditionally stable (US) FDTD method initially positioned at the location of the source pulse (position
[4] combined with a material independent perfectly matched m), where W S is the window size; 2) The information of
layer (MI-PML) formulation [5]. The final method has two terrain, obstacles and atmospheric refractive index are added to
important features: the time step used in the simulations are no the mesh of the window via electrical parameters; 3) Solve the
longer restricted by the Courant-Friedrich-Levy (CFL) stability 2-D Maxwell’s equations via an specific FDTD formulation; 4)
condition and the formulation is independent from the material If the pulse is near the right border of the window, the window

978-1-4577-0919-7/12/$26.00 ©2011 IEEE 1178

  n


Ex |n+1
i,j =C4 |i,j Dx |n+1
i,j − C5 |i,j Ex |si,j , (2)
s=1
n s

∂Hz 
Dy |n+1 n
i,j =C1 |i Dy |i,j − C2 |j+1/2,i
s=1
∂x i,j
 n+1 n 
∂Hz  ∂Hz 
− C3 |j+1/2,i + , (3)
∂x i,j ∂x i,j
 n


Ey |n+1
i,j
n+1
=C4 |i,j Dy |i,j − C5 |i,j s
Ey |i,j , (4)
s=1
Fig. 1. FDTD propagator using the moving window concept. n

Hz |n+1 n
i,j =C6 |i+1/2,j+1/2 Hz |i,j − C7 |i+1/2,j+1/2 Hz |si,j
s=1
 n+1 n
moves W S − (OR + 1) cells in the direction x̂, where OR ∂Ex  ∂Ex 
+ C8 |i+1/2,j+1/2 +
determines an overlap region of the window at position m + 1 ∂y i,j ∂y i,j
on the window at m. This region OR represents the excitation n+1 n 
∂Ey  ∂Ey 
signal of the window m + 1. −  − , (5)
The moving window algorithm can decrease severely the ∂x i,j ∂x i,j
computer effort when applied in long distance propagation here, we simplify the spatial notation with each component
paths, once the FDTD calculation is performed only in the located at (i, j), but the fields are distributed according with
W S × Wy window’s mesh. The window size W S and the Yee’s cell scheme [12]. The coefficients are given by:
overlap region OR define the algorithm accuracy and speed.
P1 (p) 2ΔtP3 (p)
B. US-IPML Formulation C1 |p = , C2 |p,q = P2 (q) ,
P2 (p)
The unconditionally stable FDTD method was proposed ΔtP2 (p)
C3 |p,q = , C4 |p,q = L1 (p, q),
by Zhao [4],[8], and like the well know alternating direction 2P2 (q)
implicit (ADI) method [9],[10] removes the CFL condition P1 (p)P1 (q)
C5 |p,q = L1 (p, q)L2 (p, q), C6 |p,q = P2 (p)P2 (q) ,
restraint. However, the US-FDTD are defined at only two time
steps (n and n+1, demanding one iteration process), and the 4P3 (p)P3 (q) Δt
C7 |p,q = , C8 |p,q = 2μ0 P2 (p)P2 (q) , (6)
left-hand and right-hand sides of the original updating equa- P2 (p)P2 (q)
tions are balanced (with regards to time steps) as accurately as where Δt is the time step, μ0 is permeability of the vacuum.
possible. These lead to a faster and more accurate algorithm The medium parameters σ and r are accounted via L1 and
when compared to the ADI, which requires two iteration L2 terms:
processes and is based on unbalanced time-steps equations.
1
The MI-PML formulation adopted in this work was pro- L1 (p, q) = ,
posed by Sullivan [5] and avoid the estimation of an effective 0 r (p, q) + σ(p, q)Δt
permittivity ef f to use in the equations. This parameter is L2 (p, q) =σ(p, q)Δt, (7)
calculated based on the r values of the medium. The ef f and the PML attenuation is performed by P1 , P2 and P3 :
estimation is not trivial and may become an error source in  m
the moving window context. Thus, the PML is implemented p
P1 (p) =1 − Gmax ,
associating fictitious conductivities with the magnetic field H  npml
  m
and electrical displacement D, instead of the usual association p
with H  and electric field E.  This makes the formulation P2 (p) =1 + Gmax ,
npml
completely independent of the FDTD domain medium and  m
p
separate from real conductivities of the medium [11]. P3 (p) =Gmax , (8)
npml
For the propagation analysis, we consider the 2-D environ-
ment depicted in Figure 1 and adopted a vertical polarized where npml is the total number of PML layers and p =
signal, modeled by a T Mx solution (Ex , Ey and Hz compo- 1, 2, 3..., npml. The parameters Gmax = 0.333 and m = 3
nents). Thus, we used the time balanced scheme proposed by define the PML absorption and we use those values recom-
Zhao [8] to obtain our initial equations: mended by Sullivan [5] for guaranteed stability.
n s The equations (1) and (3) have to be calculated by implicit
 ∂Hz 
n+1 n
Dx |i,j =C1 |j Dx |i,j + C2 |i+1/2,j techniques, once have terms calculated at the same time step
s=1
∂y i,j on both left and right-hands. So, for the Dx component, we
 n+1 n 
∂Hz  ∂Hz  substitute (2) into (5) and the resulting expression into (1). This
+ C3 |i+1/2,j + , (1) leads to a tridiagonal system of equations for each i coordinate,
∂y i,j ∂y i,j
where we vary the j possible values. The final equation is:

1179
 Finally, the equations for the bidimensional US-
IPML-FDTD algorithm must be calculated in order of
A1 |i,j−1,j Dx |n+1 n+1
i,j−1 − A1 |i,j,j + A1 |i,j,j+1 + A2 |j Dx |i,j (9),(2),(10),(4),(5), solved in one iteration process from n
+ A1 |i,j+1,j+1 Dx |n+1 n
i,j+1 = A1 |i,j−1,j Dx |i,j−1
to n + 1. The Dx and Dy components has to be updated
− [A1 |i,j,j + A1 |i,j,j+1 − A3 |j ] Dx |ni,j + A1 |i,j+1,j+1 Dx |ni,j+1 implicitly via linear systems, each one covering a specific
n mesh direction (ŷ or x̂). It is noteworthy that the time sum

+ B1 |i+1/2,j+1 Hz |si,j+1 − B1 |1+1/2,j Hz |si,j terms in the above equations are evaluated on a specific
s=1
virtual window, so it is not necessary to account the time
− B2 |i+1/2,j+1 Hz |ni,j+1 + B2 |i+1/2,j Hz |ni,j steps since the beginning of the entire simulation.
n−1
 III. N UMERICAL S IMULATION
+2 B3 |j,i,j−1 Ex |si,j−1 − [B3 |j,i,j + B3 |j+1,i,j ] Ex |si,j
A. Scattering by a conducting cylinder
s=1
+ B3 |j+1,i,j+1 Ex |si,j+1 + B3 |j,i,j−1 Ex |ni,j−1 In order to validate the proposed US-IPML formulation,
we studied the scattering properties of a perfectly conducting
− [B3 |j,i,j + B3 |j+1,i,j ] Ex |ni,j + B3 |j+1,i,j+1 Ex |ni,j+1 cylinder illuminated by a 2D current line source. The cylinder
+B4 |j+1 Ey |ni,j+1 − Ey |ni−1,j+1 +B4 |j Ey |ni−1,j − Ey |ni,j . radius is a = 5 cm, the computation space size is 42x12 cm and
(9) we used a grid sample resolution Δx=Δy= λmin /30, where
The Dy component is obtained in an analogous manner: λmin is the smallest wavelength in the medium. The source
is a Gaussian pulse with a half-bandwidth of 2.5 GHz.
A1 |i−1,j,i Dy |n+1
i−1,j − A1 |i,j,i + A1 |i,j,i+1 + A2 |i Dy |i,j
n+1
Figure 2 shows the total Hz field along the x̂ axis at
n+1 n
+ A1 |i+1,j,i+1 Dy |i+1,j = A1 |i−1,j,i Dy |i−1,j 800 MHz and 5 GHz, calculated via fast Fourier transform.
We applied the algorithm to two cases without any PML
− [A1 |i,j,i + A1 |i,j,i+1 − A3 |i ] Dy |ni,j + A1 |i+1,j,i+1 Dy |ni+1,j modification: the first uses a background medium defined by
 n
r =1.0 and σ=0 S/m, and the second medium is characterized
− B1 |j+1/2,i+1 Hz |si+1,j − B1 |j+1/2,i Hz |si,j by r =10.0, σ=0.3 S/m. The US-IPML results are obtained
s=1
varying the CFLN parameter:
+ B2 |j+1/2,i+1 Hz |ni+1,j − B2 |j+1/2,i Hz |ni,j
n−1
Δt
 CF LN = , (12)
+2 B3 |i,i−1,j Ey |s
− [B3 |i,i,j + B3 |i+1,i,j ] Ey | s Δt max
i−1,j i,j
s=1 where Δtmax is the maximum time step value limited by the
+ B3 |i+1,i+1,j Ey |si+1,j + B3 |i,i−1,j Ey |ni−1,j Courant-Friedrich-Levy stability condition. The reference is
− [B3 |i,i,j + B3 |i+1,i,j ] Ey |ni,j + B3 |i+1,i+1,j Ey |ni+1,j the analytic solution from Bessel function expansions [13].
B4 |i+1 B. Path loss estimation
+ Ex |n+1 n+1 n n
i+1,j − Ex |i+1,j−1 + Ex |i+1,j − Ex |i+1,j−1
2 The proposed FDTD-based propagation method is analyzed
B4 |i
+ Ex |n+1 n+1 n
i,j−1 − Ex |i,j + Ex |i,j−1 − Ex |i,j ,
n
(10) through two idealized terrain profiles involving HF and VHF
2 signals.The FDTD simulations are performed with a single cell
where: excited by a broadband modulated Gaussian pulse. For the size
L1 (p, q)Δt of the FDTD virtual window, the dimension Wy is defined by
A1 |p,q,r = ,
2μ0 P2 (r) the higher terrain point or the source height above the ground,
2(Δx)2 P2 (p) whichever is higher. The minimum value of the window size
A2 |p = , W S is calculated by the pulse space waveform size, the FDTD
Δt
2
2(Δx) P1 (p) domain discretization Δx and the CFLN parameter.
A3 |p = , The two-dimensional FDTD technique developed in Section
Δt  
P3 (q) II models the fields spread as cylindrical waves. However,
B1 |p,q =4ΔxP3 (p) −1 , in real scenarios the fields have spherical wave spread as
P2 (q)
  they move away from the source. To compare the FDTD
P1 (p)P1 (q)
B2 |p,q =Δx P2 (p) + , directly to measurements or 3D formulations, we need to add
P2 (q) the spherical wave attenuation with distance by considering a
L1 (q, r)L2 (q, r) correction factor CF [14]. Thus, as done in [15], we adjust
B3 |p,q,r = ,
2μ0 P2 (p) the FDTD fields by multiplying by:
Δt 
B4 |p = , (11) 1 λ
μ0 P2 (p) CF = , (13)
2 ρ
and Δx is the spatial increment. We note that the form of
(10) is slightly different from (9) due the definition of the where λ is the wave length and ρ is the distance between the
intermediate fields in the equations substitution process [8]. source and the receiver.

1180
 Fig. 3. Path loss for the Gaussian-shaped mixed terrain at 1 MHz for various
(a) Background medium: r = 1.0, σ = 0 S/m. values of window size W S and Courant-Friedrich-Levy number CF LN .

for the Ott’s HF integral formulation [6] and the FDTD

propagator. The results were generated with Δx = λ/40,
window an overlap region OR = 0.5, i.e., 0.5 times the
window size, and different values of window size W S and
CF LN . The values of W S are given in percent of the total
terrain length. For W S = 17 the algorithm run with 11 virtual
windows and for W S = 34 with 5 windows. Ott’s technique is
a well known integral equation method that neglects the back-
scattering, modeling the forward propagation. Although the
FDTD formulation computes all the wave propagation mech-
anisms inside a window, mainly models forward propagation
as the window moves along the terrain.
Table I summarizes the effect of different values of WS
and CFLN on the total simulation time. The simulations were
(b) Background medium: r = 10.0, σ = 0.3 S/m. performed on a personal computer with a Intel Core 2 Duo
at 1.83 GHz and 2 Gbytes RAM memory. We can see that
Fig. 2. Total Hz component along the x̂ axis for the cylinder the smaller the window size, smaller is the total simulation
scattering by a 2D current line source. time, however we loose accuracy. As we increase the CFLN
TABLE I parameter, i.e., increase the time step Δt, the computation time
E FFECT OF THE WS AND CFLN VARIATION ON THE SIMULATION is almost reduced by the same factor.
TIME FOR THE G AUSSIAN - SHAPED CASE . The second numeric simulation involves a wedge-shaped
Window Size Computation Relative
terrain profile with a transmitter at x=0 and y=80 m. The
CFLN wedge have a maximum height of 100 m, the terrain length is 1
WS time [s] time [%]
17 1 1424 100 km and the ground constants are r = 15 and σ = 0.012 S/m
17 2 715 51 for land and r = 81 and σ = 0.01 S/m for lake water.
17 4 354 27
The source irradiates a 100 MHz vertically polarized signal.
34 1 2462 173
34 2 1267 89 The path loss is calculated for a 3 m height receiver moving
34 4 676 47 along the terrain. The reference is a Method of Moments
(MoM) prediction technique developed in [7] with a λ/60
discretization. The FDTD simulations used Δx = λ/20,
The first case study is a 10 km mixed path with a Gaussian- CF LN = 2, OR = 0.5 and different values of W S.
shaped terrain profile. The transmitter has 40 m height and The path loss estimation is exhibited in Figure 4-a, where
is vertically polarized with F=1 MHz. The ground constants the FDTD algorithm computed similar values along the entire
are r = 15 and σ = 0.012 S/m for land and r = 81 and path. Figure 4-b shows the absolute errors. The large differ-
σ = 0.01 S/m for lake water. The total path loss is estimated ences occurred at two regions: the end of the wedge and over
for 100 points over the terrain for a 5 m height receiver. the lake. Table II summarizes the performance of the FDTD
Figure 3 shows the terrain profile and the path loss results method for various W S values.

1181
 above the CFL stability restraint and the computation speed
can be improved. The MI-IPML formulation can be applied
to any real propagation scenario without any modification.
The method performance was investigated through idealized
study cases at HF and VHF bands, and we verified the effect
of the main parameters variation. The FDTD computed similar
results when compared to well-known propagation methods.
As we adopted a pure 2D-FDTD formulation, the results were
adjusted by a correction factor [15] for comparison with the
3D propagation methods. This may be a source of error and
we plan to investigate it in further publications.
ACKNOWLEDGMENT
This work has been supported by the Brazilian agencies
CAPES, Fapemig and CNPq.
R EFERENCES
(a) Path loss
[1] L. Sevgi, “Groundwave Modeling and Simulation Strategies and Path
Loss Prediction Virtual Tools”, IEEE Trans. Antennas Propagat. , vol.55,
No. 6, pp.1591-1598, June 2007.
[2] L. Sevgi, F. Akleman “A Novel Finite-Difference Time-Domain Wave
Propagator”, IEEE Trans. Antennas Propagat., vol.48, No 3, Mar 2000.
[3] R. Luebbers, J. Schuster, K. Wu, “Full Wave Propagation Model Based on
Moving Window FDTD”, Military Communications Conference, 2003,
MILCOM 2003 IEEE , vol.2, pp.1397-1401, out 2003.
[4] A. Zhao, More accurate and efficient unconditionally stable FDTD
method”, Electronics Letters, vol.38, No. 16, pp.862-864, 2002.
[5] D. Sullivan, “An unsplit step 3-D PML for use with the FDTD method”,
IEEE Microwave Guided Wave Letters vol.7, pp.184-186, Jul. 1997.
[6] R.H.Ott, “A New Method for Predicting HF Ground Wave Attenuation
Over Inhomogeneous, Irregular Terrain”, OT/ITS Research Report No.7,
Jan. 1971.
[7] C. Batista and C. Rego, “An integral equation model for radiowave prop-
agation over inhomogeneous smoothly irregular terrain”, Microwave and
Optical Tech. Letters, vol.54, pp.26-31, 2012. DOI: 10.1002/mop.26473
[8] A. Zhao, “A Novel Implementation For Two-Dimensional Uncondition-
ally Stable FDTD Method”, Microwave and Optical Technology Letters,
vol.38, No. 6, pp.457-462, Sept. 2003.
[9] T. Namiki, “A New FDTD Algorithm Based on Alternating-Direction
Implicit Method”, IEEE Trans. Microwave Theory Tech., vol. 47,
pp.2003-2007, No. 10, Out. 1999.
(b) FDTD absolute error compared to the MoM [7] technique. [10] F. Zheng, Z. Chen, J. Zhang, “A Finite-Difference Time-Domain Method
Without the Courant Stability Conditions”, IEEE Microwave Guided
Fig. 4. Propagation of a 100 MHz signal over a wedge-shaped mixed Wave Letters, vol. 9, No. 11, pp.441-443, 1999.
terrain profile. [11] A. Zhao, M. Rinne, “Theoretical proof of the material independent PML
absorbers used for arbitrary anisotropic media”, Electronics Letters ,
vol.34, pp.48-49, 1998.
TABLE II [12] K. S. Yee, “Numerical solution of initial boundary value problems
P ERFORMANCE OF THE FDTD BASED ALGORITHM COMPARED TO involving maxwell’s equation in isotropic media”, IEEE Trans. Antennas
THE M O M FORMULATION [7] FOR THE WEDGE - SHAPED PROFILE .
Propagat. , vol. ap.14, No. 3, pp.302-307, May 1966.
[13] R. F. Harrington, Time-Harmonic Electromagnetic Fields, New York:
Window Size No. of Mean value Absolute McGraw Hill, 1961.
WS windows error [dB] mean error [dB] [14] R. Luebbers, J. Schuster, K. Wu, R. Ohs, “Application Of Moving
5 40 2.69 3.58 Window FDTD To Modeling The Effects Of Atmospheric Variations And
10 20 2.24 2.99 Foliage On Radio Wave Propagation Over Terrain”, Military Communi-
15 13 2.45 3.14 cations Conference, 2004, MILCOM 2004 IEEE , vol.3, pp.1515-1521,
30 6 1.30 2.11 nov 2004.
40 5 1.98 2.66 [15] Y. Wu, M. Lin and I. Wassell,, “Path loss estimation in 3D environments
using a modified 2D Finite-Difference Time-Domain technique”, IEEE
Trans. Ant. Propagat., CEM 2008, Brighton, UK, pp.98-99, Apr. 2008.

IV. C ONCLUSIONS
A FDTD-based propagator was presented employing a mov-
ing window algorithm [2],[3] with an unconditionally stable
(US) FDTD method [4] combined with a material independent
perfectly matched layer (MI-PML) formulation [5]. As an
unconditionally stable method, we can use time step values

1182