You are on page 1of 5

IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS, VOL. 20, NO.

9, SEPTEMBER 2021 1755

A Stochastic Radial Point Interpolation Method for


Wideband Uncertainty Analysis
R Kiran , Student Member, IEEE, and Kalarickaparambil Joseph Vinoy , Senior Member, IEEE

Abstract—In this letter, a time-domain stochastic radial point in- optimum correlation coefficients is a challenge here and an iter-
terpolation method (SRPIM) is developed for uncertainty quantifi- ative technique is proposed in [12] to overcome this. Even though
cation of electromagnetic systems. Derivatives of field quantities in PCE-based FDTD method is extended for microwave circuits
Maxwell’s equations are obtained using radial basis function, and
stochasticity in the dielectric constant of a part of the model space
in [13], inherent limitations of FDTD in representing curved ge-
is incorporated into this formulation. This is validated using the ometries limit its universal applicability. Meshless time-domain
numerical example of a parallel-plate waveguide with dielectrics method allows conformal models with multiscale capabili-
inside, which is implemented using uniaxial perfectly matched ties and has advantages over conventional FDTD [14]–[16].
boundary layers. The stochastic variations in the computed field in Furthermore, greater accuracy can be achieved using fewer
the time domain and the transmission coefficient in the frequency variables compared to conventional FDTD [17], [18], and it is
domain are evaluated. Accuracy of these simulation results is val- also reported that multiphysics models work well in meshless
idated using Kolmogorov Smirnov test, with Monte Carlo (MC)
simulation as the reference. The computation time of the proposed
methods [19]. Radial point interpolation method (RPIM) is one
method is found to be significantly better than MC and superior to of the most popular methods for electromagnetic problems using
stochastic collocation. The proposed method performs better even an unstructured mesh.
for large variations. This letter attempts to combine the advantages of RPIM and
PCE to propose a new UQ formulation for electromagnetics.
Index Terms—Polynomial chaos expansion (PCE), stochastic
radial point interpolation method (SRPIM), time-domain Being an intrusive formulation, the proposed method may be
electromagnetic methods, uncertainty quantification. combined with other approaches to quantify the effects of
several uncertain parameters. After deriving expressions for
I. INTRODUCTION field components incorporating PCE, an approach to evaluate
the required number of PCE terms is demonstrated using a
NCERTAINTIES pose challenges in electromagnetics,
U particularly for next generation wireless communication
and terahertz systems [1]. Monte Carlo (MC) method, stochastic
numerical example. This letter is organized as follows. Section II
is the theoretical development of the stochastic radial point
interpolation method (SRPIM). Section III deals with numerical
collocation (SC), perturbation method, and their variations have examples of the method, and Section IV concludes the letter.
been studied extensively for uncertainty quantification (UQ).
Although MC [2] is the most straightforward approach, large II. MATHEMATICAL FORMULATION
computational time overrides its simplicity. SC [3] is another
nonintrusive approach that produces results over a range, based In RPIM, a function u with respect to its support domain [17]
on computations at a few collocation points. Intrusive meth- is interpolated as
ods based on polynomial chaos expansion (PCE) have major N
 M

computational advantages over MC in efficiently estimating the u(x) = rn (x)an + pm (x)bm (1)
stochastic behavior [4]. n=1 m=1
Uncertainty quantification methods such as spectral stochastic where u(x) is the approximate interpolated function of u,
finite element methods [5], [6], fast PCE methods [7], Neumann rn (x) is the radial basis function (RBF), pm (x) is the monomial
expansion, and other perturbation-based methods [8] have been basis function, and an and bn are the coefficients of interpolation.
developed for frequency domain problems. Time-domain meth- N is the number of points in the support domain and M is the
ods are better suited for challenges in the wideband design of number of terms in the monomial basis function. Gaussian basis
modern electromagnetic systems. A PCE-based time-domain function exp(−c|x/r|2 ) is chosen as the radial basis, with c as
approach is introduced in [9] as a computationally efficient the shape parameter and r as the average distance between nodes.
methods for UQ. Another approach using Delta method to In order to evaluate an and bn , a point matching procedure is
calculate mean and variance of finite-difference time-domain adopted at all N nodes inside the support domain, where the field
(FDTD) solutions is described in [10] and [11]. Finding the values are ue = [u1 , u2 , . . .uN ]T . A moment matrix G can be
used to express u(x) at any point inside the support domain
Manuscript received April 14, 2021; revised June 9, 2021; accepted July 4, as
2021. Date of publication July 9, 2021; date of current version September 3,  
  e
−1 u
2021. (Corresponding author: R Kiran.)
The authors are with the Department of Electrical Communication Engi- T
u(x) = r (x) b (x) G T = Υ(x)ue (2)
neering, Indian Institute of Science, Bengaluru 560012, India (e-mail: ki- 0
ranr@iisc.ac.in; kjvinoy@ieee.org).
Digital Object Identifier 10.1109/LAWP.2021.3095913 where Υ() is called the shape function matrix.
1536-1225 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 05,2021 at 04:38:42 UTC from IEEE Xplore. Restrictions apply
1756 IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS, VOL. 20, NO. 9, SEPTEMBER 2021

In order to extend RPIM to quantify uncertainty, a PCE-based where Ex , Ey , Hz are field components in x, y, z directions,
approach is adopted here. PCE represents a random process as a respectively, Dx is electric flux density in x direction, and Jy is
direct sum of orthogonal projections in Hilbert space, called y component of the current density, which is used to introduce
spectral expansions [20]. A random process w(x, ξ) can be the source. [∂y ΥH ] represents spatial derivative of the shape
represented using PCE as function matrix of magnetic field nodes with respect to y, and

 [∂p ΥE ] represent spatial derivative of the shape function matrix
w(x, ξ) = gi (x)Ψi (ξ) (3) of electric field nodes along p ∈ {x, y}. 0 is the permittivity
i=0 of free space, r is the relative permittivity of the medium, μ
is permeability of the medium, n is the time step index, and
where x is a deterministic variable representing space
Δt is the time discretization interval. σx represents the loss
(x ∈ (x, y) for a two-dimensional (2-D) space). Let ξ =
term for absorption inside the UPML region (σx is zero outside
(ξ1 , ξ2 , . . .ξNr ) be the random vector of dimension Nr asso-
the UPML region). Δt is chosen according to the stability
ciated with the stochastic electromagnetic problem, where each
criteria [14]
random variable is assumed to be uncorrelated. In this letter, √
Gaussian distributed random permittivities are considered as Δt ≤ min dmini μi i (10)
i
source of randomness. This may be caused by manufacturing
tolerances. Similar variations may result from variations in where dmini is the distance between node i and its closest
geometry, inhomogeneity, etc. Thus, gi (x) are deterministic neighboring node and μi , i are material properties in the neigh-
functions and Ψi (ξ) are a set of orthogonal functions of the borhood of the ith node.
random variables. Hermite polynomials are chosen as orthogo- Field components everywhere in the model space may have
nal basis for PCE, according to Wiener–Askey scheme [4]. From uncertainties due to randomness anywhere in the domain, and
the orthogonal property of the inner product hence these can be treated as random processes. Thus, (6) to (9)
 ∞ become stochastic partial differential equations. If we consider
2
Ψi (ξ), Ψk (ξ) = Ψi (ξ), Ψk (ξ)e−ξ /2 dξ (8) as a typical example to formulate a PCE model, the random
−∞ process Eyn+1 (x, ξ) can be approximated using (3), truncated to
= Ψ2k (ξ)δi,k (4) m terms as
m−1

where ., . is the inner product space, which uses a Gaussian Eyn+1 (x, ξ) = en+1
yi (x)Ψi (ξ) (11)
function as its weighting function, and δi,k is the Kronecker delta i=0
function. For simplicity, Ψi (ξ) is represented as Ψi below.
where m is obtained from the highest polynomial order q of
Assume that nondispersive, lossless dielectric regions inside
Ψi (ξ) in (11) as [20]
the model space have random permittivity. Let each relative
random permittivity (r (ξj )) inside the model be represented (q + Nr )!
m= . (12)
as q!Nr !
r (ξj ) = N (μ , σ ) = μ + σ ξj (5) A representation similar to (11) can be formulated for
n+1/2
where μ and σ are mean and variance of the relative random Hz (x, ξ) in (8). After omitting x and ξ for simplicity, (8)
permittivity and ξj is assumed N (0, 1), the standard Gaussian can be written as
m−1
  m−1
having zero mean and unit variance. The impact of such vari- n+1 σx Δt  n
ations may be observed at all nodes in the domain, including eyi Ψi = 1 − eyi Ψi
i=0
0 i=0
boundary regions. In this letter, RPIM is adopted, and hence
using (2), the set of field equations in matrix form for 2-D T M z m−1
Δt  h
n+1/2
Ψi Δt
mode with uniaxial perfectly matched layer (UPML) boundary − [∂x ΥH ] zi − Jy .
0 i=0 r 0 r
in x direction are [17]
(13)
Dxn+1 = Dxn + Δt[∂y ΥH ]Hzn+1/2 (6)
Since the matrix [∂x ΥH ] and Jy are invariant under the assumed
Exn+1 = Exn stochastic variations, these terms can be treated as constants
 σx Δt here. Following the Galerkin procedure, (13) can be multiplied
1 1 + 2 1 − σ2x Δt 2
+ 0 n+1
Dx − 0
Dxn−1
(7) with Ψk , e−ξ /2 , and integrated over the domain of ξ to obtain
r 0 0  ∞ m−1
 
−ξ2 /2
σx Δt en+1
yi Ψi Ψk e dξ
Eyn+1 = 1− Eyn −∞ i=0
0
   ∞ m−1

1 Δt Δt σx Δt 2

− [∂x ΥH ] Hzn+1/2 + Jy (8) = 1− enyi Ψi Ψk e−ξ /2



r 0 0 0 −∞ i=0
  ∞ m−1
 n+1/2 Ψi Ψk
σx Δt Δt 2
Hzn+1/2 = 1− Hzn−1/2 − [∂x ΥH ] hzi e−ξ /2 dξ
0 0 −∞ i=0 r
  ∞
Δt Δt 2
+ n
[∂y ΥE ] Ex − [∂x ΥE ] Eyn
(9) − Jy Ψk e−ξ /2 dξ. (14)
μ 0 r −∞

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 05,2021 at 04:38:42 UTC from IEEE Xplore. Restrictions apply
KIRAN AND VINOY: STOCHASTIC RADIAL POINT INTERPOLATION METHOD FOR WIDEBAND UNCERTAINTY ANALYSIS 1757

By interchanging summation and integration in these terms and


applying orthogonality property in (4)
n+1/2
en+1 n
yk = C1 eyk − C2 k uzik − C3 k Jy (15)
where k = 0, 1, . . . m − 1. In a similar way, the remaining equa-
tions (6), (7), and (9) can be written as
n+1/2
dn+1 n
xk = dxk + C4 hzk (16)
en+1 n n+1 n
xk = exk + C5 k vxik − C6 k vxik (17)
Fig. 1. Representation of the computational domain, consisting of a parallel-
n+1/2 n−1/2 plate waveguide with two dielectric regions inside. Region-1 extends from x =
hzk = C7 hzk + C8 enxk − C9 enyk . (18) 0.125 to 0.15 m and region-2 from x = 0.15 to 0.175 m with dimensions d1 =
0.025 m, d2 = 0.025 m, d3 = 0.025 m. Other dimensions are l = 0.30 m,
Multiplication factors in (15) to (18) are C1 = 1 − σx0Δt , C2 k = h = 0.035 m, d4 = 0.025 m, and d5 = 0.025 m.
Δt
[∂ Υ ], C3 k = Δt
0 Ψ2  x H 0 r
for k = 0, C3 k = 0 for k = 1, 2, . . .,
k
1+ σ2
x Δt 1− σ2
x Δt
C4 = Δt[∂y ΥH ], C5 k = 0
0 Ψ2k 
, C6 k = 0
0 Ψ2k 
, C7 = 1 −
σx Δt Δt Δt
0 , C8 = μ [∂y ΥE ]
and C9 = μ [∂x ΥE ].
Other factors in
the region with stochastic variation are

n+1/2
 n+1/2
Ψi Ψk
m−1
uzik = hzi (19)
i=0
r
m−1


n+1 Ψi Ψk
vxik = dn+1
xi (20)
i=0
r

where r is given by (5). Elsewhere in the domain r is a constant Fig. 2. (a) Probability distribution of Ey by SRPIM and MC method at t =
and thus (19) and (20) have the form 1.84 ns, σ = 0.2 for different order of PCE terms (q). (b) KS-test statistics
(Dα ) versus q at t = 1.84 ns for various σ .
2
n+1/2 n+1/2 Ψk 
uzk = hzk (21)
r
Ψ2k  and f2 = 3.3 GHz at x = 0.025 in T M 0 mode. The source
n+1
vxk = dn+1
xk . (22) excitation is given by
r  2
nΔt−4σ s
− √
Stochastic estimates of fields, for example, Ey (t) at any point fs (n) = sin (π(f2 + f1 )(nΔt − 4σs )) e 2σs (25)
in the domain, can be obtained from (15) as
where σs = π|f21−f1 | . A uniform node distribution is chosen for
m−1
 primary (H) nodes with a spatial distance of 0.005 m between ad-
Ey (t) = eyi (t)Ψi (23)
jacent nodes along x- and y- directions. Locations of interleaved
i=0
E nodes are generated through the Voronoi algorithm. Time
where eyi (t) represents the time-domain PCE coefficients ob- discretization interval Δt = 8 × 10−12 s is chosen according
tained by (15)-(18). Transmission coefficient (S21 ) at a port with to (10). A UPML of four-layer thickness is included at both
respect to source can be represented as ends along the x- direction. The dielectric region-1 is consid-
m−1 ered to have a normally distributed random relative permittivity
eyi (ω)Ψi N (4, 0.05). The dielectric region-2 has a random relative per-
S21 (ω) = i=0 src (24)
Ey (ω) mittivity N (2, 0.05), uncorrelated to region-1. Since the electric
where eyi (ω) are the frequency domain PCE coefficients ob- field has a component perpendicular to the dielectric disconti-
tained through Fourier transform, and Eysrc (ω) is the Fourier nuity, relative permittivity along the interfaces with free space
transform of the source signal. Stochastic estimates of S21 can are modified to an average N (2.5, 0.025) and N (1.5, 0.025),
be obtained by (24). respectively [24]. Optimal shape parameter (c) value is chosen
as 0.05.
First, the effect of random dielectric property of the regions
III. NUMERICAL EXAMPLES on fields in the domain is investigated. This random behavior
To validate the formulation in the previous section, a 2-D at P1 at some time instant (here t = 1.84 ns) is quantified
parallel-plate waveguide in the xy plane as shown in Fig. 1, hav- as a probability distribution using the proposed formulation
ing length l = 0.30 m, height h = 0.035 m, with two dielectric for different order of PCE terms (q). MC simulation is also
regions, region-1 from x = 0.125 to 0.15 m and region-2 from performed for a comparison in the above case for 10 000 re-
x = 0.15 m to 0.175 m, y = 0 m to 0.025 m is considered. This alizations by separate deterministic RPIM computations. The
waveguide with PEC walls at y = 0 and y = 0.035 m is excited results are shown in Fig. 2(a), which shows a very good match
using a band-limited Gaussian wire source with f1 = 2 GHz when q = 3. A two-sided Kolmogorov-Smirnov (KS) test [25] is

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 05,2021 at 04:38:42 UTC from IEEE Xplore. Restrictions apply
1758 IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS, VOL. 20, NO. 9, SEPTEMBER 2021

TABLE I
ACCURACY(Dα ) AND EXECUTION TIME IN ESTIMATING |S21 |

Fig. 3. 3σ intervals of |S21 | and probability distributions at different frequen-


cies by SRPIM and MC method for q = 5 and σ = 0.05.

performed to assess the goodness of fit between MC and SRPIM


results. A null hypothesis that these two datasets follow the same
distribution is tested with a significance level α. Test statistics
Dαq is evaluated as
 
 q 
Dαq = max F̂M C (x) − F̂SRP IM (x) (26)
x

where F̂M C (x) are ordered samples from MC simulation and of SC-RPIM remains above the critical upper bound at resonance
q frequencies even with a large Nq in both the cases. Execution
F̂SRP IM (x) are those from SRPIM with qth order PCE terms.
If the test statistics Dαq is within the bound Dαcritical , the test time of SRPIM is less than SC-RPIM, and yet SRPIM shows
is successful,
√ and the null hypothesis is accepted. Dαcritical = a better convergence to MC. As stochastic dimension of the
1.36/ N for α = 0.05, where N is the number of samples in problem increases, time complexity of SRPIM increases at a rate
each dataset. Fig. 2(b) shows the convergence of the mismatch suggested by (12) and is prone to “the curse of dimensionality.”
between SRPIM and MC with increasing q for different levels The numerical example discussed here is evaluated on an Intel
of variance in permittivity (σ ). SRPIM passes KS-test with Core i7 4790S CPU at 3.20 GHz clock and 16 GB RAM, in
q = 3 even for a high variance level of 25% with average MATLAB R2019a environment.
computation time 0.4889 s, whereas MC with 10 000 realizations
take 111.909 s, which is 230 times higher.
IV. CONCLUSION
Next, an investigation is done to estimate stochasticity in
magnitude of the transmission coefficient, |S21 | using (24). The RPIM can be used for accurate electromagnetic simula-
versatility of RPIM allows achieving sufficient accuracy for a tions without using a well-defined mesh. Being a time-domain
support domain radius of 0.0091 m. Case-1 is considered where method, wideband frequency characteristics can be obtained in
stochasticity is assumed on region 1 and region 2 is kept at its one set of simulations. Intrusive methods such as PCE can be
mean permittivity, and case-2 where both regions are stochastic used for accurate and efficient uncertainty analysis. This letter
and uncorrelated. Fig. 3 is the 3σ interval plot of |S21 | of the attempts to combine the advantages of RPIM and PCE to propose
two cases above, in the frequency range from f1 = 2 GHz to a broadband UQ method. Expressions for field components
f2 = 3.3 GHz for q = 5. Inset graphs in Fig. 3 show probability incorporating PCE in RPIM are derived first. The sufficient PCE
distributions of case-2, calculated by the proposed method and truncation is evaluated for the numerical example by comparing
MC method, for the worst case frequency (3.0823 GHz) and with a large MC simulation.
two other frequencies (2.9793 and 3.1166 GHz). Using (26), A parallel-plate waveguide with dielectric regions of random
the KS-test is performed for q up to 5 at these frequencies and permittivity embedded inside is considered as a numerical ex-
the statistics are included in Table I. It is clear that for q = 5, ample. Statistical estimates of time-domain field variations and
Dα is below Dαcritical for all these frequencies. In this example, transmission coefficient using the proposed method matches MC
execution time of SRPIM is less than MC by 1230 times in simulations, but take significantly less computation time. The
case-1, and 175 times in case-2. Although MC gives consistent convergence of SRPIM for high variations in permittivity up to
and accurate results especially when a large number of samples 25% is studied and found to be accurate. A stochastic collocation
is used, this is not a computationally efficient scheme. Hence, is also implemented in RPIM and used to compare the proposed
the above cases are also evaluated using stochastic collocation method. It is found that SRPIM performs better in terms of
(SC) [3]. SC generates a probability distribution of |S21 | at each computational time and accuracy than SC-RPIM. The proposed
frequency point, from a small number of realizations, by execut- method can be used in situations even with high variations in
ing deterministic RPIM at Nq collocation points obtained from stochasticity and can be extended to stochasticity in proper-
Gauss-Hermite quadrature, and interpolating using Lagrange ties such as permeability, conductivity, boundary conditions,
method. A comparison of accuracy and execution time between geometry, and surface roughness especially for high-frequency
SRPIM, SC-RPIM, and MC is given in Table I. It is clear that Dα electromagnetics.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 05,2021 at 04:38:42 UTC from IEEE Xplore. Restrictions apply
KIRAN AND VINOY: STOCHASTIC RADIAL POINT INTERPOLATION METHOD FOR WIDEBAND UNCERTAINTY ANALYSIS 1759

REFERENCES [14] T. Kaufmann, Y. Yu, C. Engstrüm, Z. Chen, and C. Fumeaux, “Re-


cent developments of the meshless radial point interpolation method for
[1] H. Sarieddeen, N. Saeed, T. Y. Al-Naffouri, and M. Alouini, “Next gen- time-domain electromagnetics,” Int. J. Numer. Model., vol. 25, no. 5,
eration terahertz communications: A rendezvous of sensing, imaging, and pp. 468–489, 2012.
localization,” IEEE Commun. Mag., vol. 58, no. 5, pp. 69–75, May 2020. [15] T. Kaufmann, C. Engström, C. Fumeaux, and R. Vahldieck, “Eigenvalue
[2] M. N. O. Sadiku, Monte Carlo Methods for Electromagnetics. Boca Raton, analysis and longtime stability of resonant structures for the meshless
FL, USA: CRC Press, 2009. radial point interpolation method in time domain,” IEEE Trans. Microw.
[3] L. Mathelin and M. Y. Hussaini, “A stochastic collocation algorithm for Theory Techn., vol. 58, no. 12, pp. 3399–3408, Dec. 2010.
uncertainty analysis,” NASA/CR-2003-212153, 2003. [16] S. A. Sivaram and K. J. Vinoy, “Inverse multiquadric radial basis functions
[4] D. Xiu and G. E. Karniadakis, “The Wiener–Askey polynomial chaos for in eigenvalue analysis of a circular waveguide using radial point interpo-
stochastic differential equations,” SIAM J. Sci. Comput., vol. 24, no. 2, lation method,” IEEE Microw. Wireless Compon. Lett., vol. 30, no. 6,
pp. 619–644, 2002. pp. 537–540, Jun. 2020.
[5] B. N. Abhijith and K. J. Vinoy, “Spectral stochastic FEM for uncertainty [17] T. Kaufmann, C. Fumeaux, and R. Vahldieck, “The meshless radial point
quantification due to multiple dielectric variabilities,” IEEE Antennas interpolation method for time domain electromagnetics,” in IEEE MTT-S
Wireless Propag. Lett., vol. 18, no. 10, pp. 1961–1965, Oct. 2019. Int. Microw. Symp. Dig., Atlanta, GA, USA, 2008, pp. 61–64.
[6] G. Hemanth, K. J. Vinoy, and S. Gopalakrishnan, “Spectral stochastic [18] S. J. Lai, B. Z. Wang, and Y. Duan, “Meshless radial basis function method
finite element method for periodic structure,” in Proc. IEEE Int. Microw. for transient electromagnetic computations,” IEEE Trans. Magn., vol. 44,
RF Conf., Bangalore, India, 2014, pp. 9–12. no. 10, pp. 2288–2295, Oct. 2008.
[7] K. T. J. Gladwin and K. J. Vinoy, “A fast polynomial chaos expansion for [19] D. J. Price and J. J. Monaghan, “Smoothed particle magnetohydrody
uncertainty quantification in stochastic electromagnetic problems,” IEEE namics—I. Algorithm and tests in one dimension,” Mon. Notices Roy.
Antennas Wireless Propag. Lett., vol. 18, no. 10, pp. 2120–2124, Oct. 2019. Astron., vol 348, no. 1, pp. 123–138, Feb. 2004.
[8] G. J. K. Tomy and K. J. Vinoy, “Neumann-expansion based FEM for uncer- [20] R. G. Ghanen and P. D. Spanos, Stochastic Finite Elements: A Spectral
tainty quantification of permittivity variations,” IEEE Antennas Wireless Approach, New York, NY, USA: Springer-Verlag, 1991.
Propag. Lett., vol. 19, no. 4, pp. 561–565, Apr. 2020. [21] S. D. Gedney, “An anisotropic perfectly matched layer absorbing medium
[9] R. S. Edwards, A. C. Marvin, and S. J. Porter, “Uncertainty analyses for the truncation of FDTD lattices,” IEEE Trans. Antennas Propag.,
in the finite difference time domain method,” IEEE Trans. Electromagn. vol. 44, no. 12, pp. 1630–1639, Dec. 1996.
Compat., vol. 52, no. 1, pp. 155–163, Feb. 2010. [22] A. Taflove and S. C. Hagness, Computational Electrodynamics: The Finite
[10] S. M. Smith and C. Furse, “Stochastic FDTD for analysis of statistical vari- Difference Time Domain Method. Norwood, MA, USA: Artech House,
ation in electromagnetic fields,” IEEE Trans. Antennas Propag., vol. 60, 2005.
no. 7, pp. 3343–3350, Jul. 2012. [23] J. G. Wang and G. R. Liu, “A point interpolation meshless method based
[11] B. T. Nguyen, C. Furse, and J. J. Simpson, “A 3-D stochastic FDTD model on radial basis functions,” Int. J. Numer. Methods Eng., vol. 54, no. 11,
of electromagnetic wave propagation in magnetized ionosphere plasma,” pp. 1623–1648, 2002.
IEEE Trans. Antennas Propag., vol. 63, no. 1, pp. 304–313, Jan. 2015. [24] Y. Yu and Z. Chen, “Implementation of material interface conditions in
[12] T. Tan, A. Taflove, and V. Backman, “Single realization stochastic FDTD the radial point interpolation meshless method,” IEEE Trans. Antennas
for weak scattering waves in biological random media,” IEEE Trans. Propag., vol. 59, no. 8, pp. 2916–2923, Aug. 2011.
Antennas Propag., vol. 61, no. 2, pp. 818–828, Feb. 2013. [25] F. J. Massey, “The Kolmogorov–Smirnov test for goodness of fit,” J. Amer.
[13] A. C. M. Austin and C. D. Sarris, “Efficient analysis of geometrical Statist. Assoc., vol. 46, no. 253, pp. 68–78, Mar. 1951.
uncertainty in the FDTD method using polynomial chaos with application
to microwave circuits,” IEEE Trans. Microw. Theory Techn., vol. 61, no. 12,
pp. 4293–4301, Dec. 2013.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on September 05,2021 at 04:38:42 UTC from IEEE Xplore. Restrictions apply

You might also like