Computation of Electromagnetic Fields

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JMMCT.2019.2897490, IEEE Journal
on Multiscale and Multiphysics Computational Techniques
IEEE JOURNAL ON MULTISCALE AND MULTIPHYSICS COMPUTATIONAL TECHNIQUES, VOL. NUMBER, 2018 1
Computation of Electromagnetic Fields Scattered

From Objects With Uncertain Shapes Using
Multilevel Monte Carlo Method
Alexander Litvinenko∗ , Abdulkadir C. Yucel, Hakan Bagci, Senior Member, IEEE,
Jesper Oppelstrup, Eric Michielssen, Fellow, IEEE, and Raúl Tempone
Abstract—Computational tools for characterizing electro- to enhancing Raman spectroscopy using metallic nanopar-
magnetic scattering from objects with uncertain shapes are ticles [3], [4]. When the target size is comparable to or
needed in various applications ranging from remote sensing
larger than the wavelength at the operation frequency, the
at microwave frequencies to Raman spectroscopy at optical
frequencies. Often, such computational tools use the Monte scattered field is a strong function of the target shape.
Carlo (MC) method to sample a parametric space describing However, in many of the applications, whether the target
geometric uncertainties. For each sample, which corresponds is a (vegetated) rough surface or a nanoparticle, its exact
to a realization of the geometry, a deterministic electromagnetic shape may not be known and has to be parameterized using
solver computes the scattered fields. However, for an accurate
a stochastic/statistical model. Consequently, computational
statistical characterization the number of MC samples has
to be large. In this work, to address this challenge, the tools, which are capable of generating statistics of a quantity
continuation multilevel Monte Carlo (CMLMC) method is of interest (QoI) (RCS or SCS in this case) given a geometry
used together with a surface integral equation solver. The described using random variables/parameters, are required.
CMLMC method optimally balances statistical errors due to These computational tools often use the Monte Carlo
sampling of the parametric space, and numerical errors due
to the discretization of the geometry using a hierarchy of
(MC) method [5]–[11]. The MC method is non-intrusive
discretizations, from coarse to fine. The number of realizations and straightforward to implement; therefore, its incorpo-
of finer discretizations can be kept low, with most samples ration with an existing deterministic EM solver is rather
computed on coarser discretizations to minimize computational trivial. However, the traditional MC method has an error
cost. Consequently, the total execution time is significantly convergence rate of O(N −1/2 ) [12], where N is the number
reduced, in comparison to the standard MC scheme.
of samples in the parametric space used for describing the
Index Terms—uncertainty quantification, uncertain geom- geometry. Provided more regularity of the QoI w.r.t. the
etry, numerical methods, multilevel Monte Carlo method geometry parameters, quasi-MC methods may have a better
(MLMC), integral equation, fast multipole method (FMM),
fast Fourier transform (FFT). convergence rate, O(N −1 ) with a multiplicative log-term
that depends on the number of parameters [12], [13]. Both
the traditional MC and quasi-MC methods require large N to
I. I NTRODUCTION yield accurate statistics of the QoI and are computationally
N UMERICAL methods for predicting radar and scat-

tering cross sections (RCS and SCS) of complex tar-
gets find engineering applications ranging from microwave
expensive since the function evaluation at each sampling
point, which corresponds to the computation of the QoI for
one realization of the deterministic problem, requires the
remote sensing soil/ocean surface and vegetation [1], [2] execution of a simulation. For practical real-life scattering
∗ Corresponding author: A. Litvinenko is with the RWTH Aachen,
scenarios, each of these simulations may take a few hours,
Aachen, Germany, (e-mail: litvinenko@uq.rwth-aachen.de). if not a few days.
R. Tempone is the Alexander von Humboldt Professor in Mathematics In recent years, schemes that make use of surrogate
for Uncertainty Quantification, RWTH Aachen University, Germany, (e- models have received significant attention as potential alter-
mail: tempone@uq.rwth-aachen.de).
H. Bagci, and R. Tempone are with the Strategic Research Initiative natives to the MC method [14]–[20]. The surrogate model
- Uncertainty Quantification (SRI-UQ) Center, Division of Computer, is generated using the values of the QoI that are computed
Electrical, and Mathematical Science and Engineering (CEMSE), King by the simulator at a small number of “carefully selected”
Abdullah University of Science and Technology (KAUST), Thuwal, 23955,
Saudi Arabia (e-mails: {hakan.bagci, raul.tempone}@kaust.edu.sa). (collocation) points in the parametric space. The surrogate
A. C. Yucel is with the School of Electrical and Electronics Engineering, model is then probed using the MC method to obtain
Nanyang Technological University, Singapore 639798 (e-mail: abdulka- statistics of the QoI. Consequently, the surrogate model-
dir.yucel@kaust.edu.sa).
J. Oppelstrup is with the Department of Mathematics, KTH Royal based schemes are more efficient than the MC method that
Institute of Technology, Stockholm, Sweden (e-mail: jespero@kth.se). operates directly on the computationally expensive simula-
E. Michielssen is with the Department of Electrical Engineering and tor. The efficiency and accuracy of these schemes leveraging
Computer Science, University of Michigan, Ann Arbor, MI 48109, US
(e-mail: emichiel@umich.edu). generalized polynomial chaos expansion (gPCE) [21]–[23]
Manuscript received Month Day, 2018; revised Month Day, 2018. can be improved using a refinement strategy that adaptively
2379-8793 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JMMCT.2019.2897490, IEEE Journal
divides the parametric space into smaller domains, in each of

which a separate collocation scheme is used [16], [17]. Ad-
ditionally, if the QoI can be approximated in terms of low- Excitation: (ϑ inc ,ϕ inc ) Observation: (ϑ ,ϕ )
order contributions from only a few of the parameters, one
z
can use high dimensional model representation expansions
û(ϑ ,ϕ )
to accelerate the generation of the surrogate model [18].
The surrogate model-based schemes and their accelerated ûinc (ϑ inc ,ϕ inc ) y
variants have been successfully applied to certain stochastic
EM problems [24]–[32]. On the other hand, for the problem
of scattering from geometries with uncertain shapes, all ε 0 , µ0 ε (r), µ0
parameters and their high-order combinations contribute
significantly to the QoI limiting the efficient use of a
surrogate model [33]–[35]. x
Recently, the multilevel MC (MLMC) methods have
seen increasing use due to their efficiency, robustness, and
simplicity [36]–[42]. The MLMC methods operate on a Fig. 1. Description of the scattering problem.
hierarchy of meshes and perform most of the simulations
using low-fidelity models (coarser meshes) and only a few
simulations using high-fidelity models (finer meshes). By erated using a (parallelized) fast multipole method (FMM)
doing so, their cost becomes significantly smaller than the - fast Fourier transform (FFT) scheme (FMM-FFT) [45]–
cost of the traditional MC methods using only high-fidelity [51]. Numerical results, which demonstrate the accuracy,
models. The continuation MLMC (CMLMC) algorithm [42] efficiency, and convergence of the proposed computational
uses the MLMC method to calibrate the average cost framework, are presented. The developed computational
per sampling point and the corresponding variance using framework is proven effective not only in studying the
Bayesian estimation, taking particular notice of the deepest effects of uncertainty in the geometry on scattered fields
levels of the mesh hierarchy, to minimize the computational but also in increasing the robustness of the FMM-FFT
cost. To balance discretization and statistical errors, the accelerated PMCHWT-SIE solver by testing its convergence
CMLMC method estimates how the discretization error and for a large set of scenarios with deformed geometries and
the computational cost depend on the mesh level and uses varying mesh densities, quadrature rules, iterative solver
this information to select optimal numbers of levels and tolerances, and FMM parameters.
samples on each level. Remark 1: Reasons for this specific choice of SIE
Surrogate-model based schemes leveraging gPCE [14], formulation (namely PMCHWT) can be explained as fol-
[15], [43] (as briefly described above) could be an alternative lows. One can define two types of “convergence” for the
to the CMLMC method in general. If the response/surrogate SIE formulation used within the CMLMC framework: (i)
model is smooth and simple, then the collocation scheme convergence of the iterative solver, and (ii) convergence
used for computing/finding the coefficients of the gPCE of the solution as mesh is made denser. The PMCHWT
requires only a few deterministic solver executions. How- formulation produces a first-kind SIE, and it is well known
ever, gPCE-based surrogate models suffer from several that there are other formulations that produce second-kind
drawbacks: it is not clear how to determine the number of SIEs, which are better behaved under convergence type
terms in the gPCE, how to avoid aliasing effects/errors by (ii). However, results obtained by solving these SIEs are
computing higher order polynomials, and how to control the less accurate than those of the PMCHWT formulation
accuracy of the integration needed for computing the gPCE [52]. Additionally, PMCHWT-SIE is better behaved under
coefficients. In addition, as mentioned above, specifically convergence type (i), i.e., its solution gets more accurately
for the scattering problem, surrogate models lose their consistently as mesh is made denser. Having said that, it is
efficiency. Ideas from CMLMC and gPCE methods could clear that the convergence type (i) might have a significant
be combined, but it goes out of the scope of this work. effect on the efficiency of the CMLMC. Use of second-kind
SIEs within the CMLMC framework will be investigated
In this work, a computational framework, which makes
elsewhere.
use of the CMLMC method to efficiently and accurately
characterize EM wave scattering from dielectric objects with
uncertain shapes, is proposed. The deterministic simulations II. F ORMULATION
required by the CMLMC method to compute the sam- This section describes a computational framework for
ples at different levels are carried out using the Poggio- characterizing scattering from dielectric objects with uncer-
Miller-Chan-Harrington-Wu-Tsai surface integral equation tain shapes (1). It is assumed that the shape of the scatterer
(PMCHWT-SIE) solver [44]. The PMCHWT-SIE is dis- can be parameterized by means of random variables. The
cretized using the method of moments (MoM) and the QoI is the SCS over a user-defined solid angle (i.e., a
iterative solution of the resulting matrix system is accel- measure of far-field scattered power in a cone). Section II-A
describes the MLMC and CMLMC methods, Section II-B themselves would yield large tolerances. The CMLMC algo-
formulates a scheme to parameterize and generate objects rithm assumes that the convergence rates for the mean (weak
with uncertain shapes, and Section II-C outlines the FMM- convergence) and variance (strong convergence) follow
FFT accelerated PMCHWT-SIE solver used for computing
the scattering cross section of a given object. E[g − gℓ ] ≈ QW hqℓ 1 (3a)
Var[gℓ − gℓ−1 ] ≈ QS hqℓ−1
2
(3b)
A. MLMC and CMLMC Methods for QW 6= 0, QS > 0, q1 > 0, and 0 < q2 ≤ 2q1 [42]. Note
This section describes elements of the MLMC and that variables q1 and q2 are solver- and problem-dependent.
CMLMC methods relevant to the characterization of scat- For example, the CMLMC algorithm estimates q1 ≈ 2
tering from objects with uncertain shapes. A more in-depth and q2 ≈ 4, and q1 ≈ 3 and q2 ≈ 5 for the examples in
description of these techniques can be found in [42]. Sections III-B and III-C, respectively. These parameter es-
Let ξ and g(ξ) represent the vector of random param- timates are crucial to optimally distribute the computational
eters/variables and the QoI, respectively. The goal of the effort, as shown below. The total computational cost of the
MLMC method is to approximate the expected value, E[g], adaptive algorithm is close to that of the MLMC method
to a guaranteed tolerance TOL with predefined probability, with correct values of parameters given a priori.
and minimal computational cost. To achieve this, the method For the sake of completeness, the main ingredients of
constructs a telescoping sum, defined over a sequence of the CMLMC algorithm (described in full in [42]) are stated
mesh levels ℓ = 0, . . . , L as described next. here. To estimate the number of samples, the algorithm
invokes of ‘cost per sample’ and ‘total cost’ as described
Let {Pℓ }L L
ℓ=0 and {hℓ }ℓ=0 be sequences of meshes that
next.
discretize the randomly generated object’s surface and their
average element edge sizes, respectively. It is assumed The CMLMC estimator for the QoI, A, can be written
PL ∼
{Pℓ }L −ℓ as A = ℓ=0 Gℓ . Let the average cost of generating one
ℓ=0 are generated hierarchically with hℓ = h0 β for
h0 > 0 and a constant β > 1. A method to such meshes sample of Gℓ (cost of one deterministic simulation for one
for the purpose of illustrating the proposed technique is random realization) be
described in Section II-B.
Let gℓ (ξ) represent the approximation to g(ξ) computed Wℓ ∝ h−dγ
ℓ = h−dγ
0 β ℓdγ (4)
using mesh Pℓ . The MLMC method expresses E[gL ], the
expected value of the most accurate approximation gL , using where d is the spatial dimension and γ is determined by
the computational complexity of the deterministic solver.
L
X For the FMM-FFT-accelerated PMCHWT-SIE solver used
E[gL ] = E[Gℓ ] (1) here d = 2 (only surfaces are discretized) and γ ≈ 1
ℓ=0 (Sections II-C and III-B). Note that this solver calls for
where Gℓ is defined as an iterative solution of the MoM system of equations
( (Section II-C), and that the cost of computing a sample of
g0 if ℓ = 0 Gℓ may fluctuate for different realizations depending on the
Gℓ = . (2)
gℓ − gℓ−1 if ℓ > 0 number of iterations required. Finally, for the method used
for generating {Pℓ }L
l=0 (Section II-B), β = 2.
Note that gℓ and gℓ−1 are computed using the same input The total CMLMC computational cost is
random parameter ξ.
L
In the telescoping sum (1), the expected values in practice X
∼ W = Mℓ W ℓ . (5)
are replaced
P by sample averages, i.e., E[Gℓ ] ≈ Gℓ = ℓ=0
Mℓ
Mℓ−1 m=1 Gℓ,m , where random variable Gℓ,m have the
same distribution as Gℓ and are independent identically dis- The estimator A satisfies a tolerance with a prescribed
tributed (i.i.d.) samples. Gℓ,m represents the m-th random failure probability 0 < ν ≤ 1, i.e.,
sample of the random output Gℓ . Gℓ and Gℓ,m are random P[|E[g] − A| ≤ TOL] ≥ 1 − ν (6)
variables because they depend on random (input) parameters
ξ, which is defined in (17). while minimizing W . The total error is split into bias and
As ℓ increases, the variance of Gℓ decreases. As a result, statistical error,
the total computational cost can be reduced by approximat-
ing E[Gℓ ] with a smaller number of samples. |E[g] − A| ≤ |E[g − A]| + |E[A] − A|
| {z } | {z }
The CMLMC algorithm is an improved version of the Bias Statistical error
MLMC method in that it approximates E[g] with a sequence
where θ ∈ (0, 1) is a splitting parameter, so that
of decreasing tolerances [42]. In doing so, it continuously
improves estimates of several problem-dependent parame- TOL = (1 − θ)TOL + | θTOL
| {z } {z }. (7)
ters, while solving relatively inexpensive problems that by Bias tolerance Statistical error tolerance
The CMLMC algorithm bounds the bias, B = |E[g − A]|, that of the MC method assuming that the cost per sample is
and the statistical error as O (1). In other words, for this case, the CMLMC algorithm
can in effect remove the computational cost required by the
B = |E[g − A]| ≤ (1 − θ)TOL (8) discretization, namely O(TOL−dγ/q1 ).
|E[A] − A| ≤ θTOL (9) Remark 2: Note that the numerical errors relevant to
where the latter bound holds with probability 1 − ν. Note the deterministic solver, such as discretization error, error
that θ itself can be a variable [42]. due to the numerical integration used for computing the
To satisfy condition in (9) we require: MoM matrix entries, and error due to the FMM used
2 for accelerating the matrix-vector product, do not have a
θTOL significant effect on the statistical error, but they contribute
Var[A] ≤ (10)
Cν to the bias error. Convergence of bias and the statistical
for some given confidence parameter, Cν , such that errors are essential for CMLMC.
Φ(Cν ) = 1 − ν2 , (see more in [53]); here, Φ is the cumu-
lative distribution function of a standard normal random B. Random Shape Generation
variable. This section describes a method to generate the sequence
PL E[A]−1= E[gL ],
By construction of the MLMC estimator, of meshes {Pℓ }L ℓ=0 that are used in the numerical exper-
and by independence, Var[A] = ℓ=0 Vℓ Mℓ , where iments (Section III) to demonstrate the properties of the
Vℓ = Var[Gℓ ]. Given L, TOL, and 0 < θ < 1, and by CMLMC scheme. Alternative ways to generating random
minimizing W subject to the statistical constraint (10) for perturbations can be found in [22], [56]–[59]. The method
L
{Mℓ }ℓ=0 gives the following optimal number of samples used here relies on perturbing an initial mesh generated on
per level ℓ (apply ceiling function to Mℓ if necessary): a sphere and refining it as level ℓ is increased as described
!
2 r L next.
Cν Vℓ X p
Mℓ = Vℓ W ℓ . (11) First, the unit sphere is discretized using a mesh P0 with
θTOL Wℓ
ℓ=0 triangular elements, then the perturbations are generated by
Summing the optimal numbers of samples over all levels moving the nodes of this mesh using
yields the following expression for the total optimal com- K
X
putational cost in terms of TOL: v(ϑm , ϕm ) ≈ ṽ(ϑm , ϕm ) + ak κk (ϑm , ϕm ). (15)
2 X L p
!2 k=1
Cν
W (TOL, L) = Vℓ Wℓ . (12) where ϑm and ϕm are angular coordinates of node
θTOL m, v(ϑm , ϕm ) is its (perturbed) radial coordinate, and
ℓ=0
The total cost of the CMLMC algorithm can be estimated ṽ(ϑm , ϕm ) = 1 m is its (unperturbed) radial coordinate
using Theorem 1 below [36], [39], [40], [54], [55]. on the unit sphere (all in spherical coordinate system).
Theorem 1: Let d = {1, 2, 3} denote the problem dimen- Here, κk (ϑ, ϕ) are obtained from spherical harmonics PK by
sion. Suppose there exist positive constants q1 , q2 , γ > 0 re-scaling their arguments and ak , which satisfy k=1 ak <
such that q1 ≥ 12 min(q2 , γd), and |E[gℓ − g]| = O(hqℓ 1 ), 0.5, are uncorrelated random variables.
Var[gℓ − gℓ−1 ] = O(hqℓ 2 ), and Wℓ = O(h−dγ ). Then for For the numerical experiments considered in this work,
ℓ
any accuracy TOL and confidence level ν, 0 < ν ≤ 1, there K = 2, and κ1 (ϑ, ϕ) = cos(α1 ϑ) and κ2 (ϑ, ϕ) =
exist a deepest level L(TOL) and number of realizations sin(α2 ϑ) sin(α3 ϕ), where α1 , α2 , and α3 are positive
{Mℓ (TOL)} such that constants. If a κk (ϑ, ϕ) depends on ϕ, then its dependence
on ϑ must vanish at the poles ϑ = {0, π}. Therefore, α2
lim inf P[(|E (g) − A| ≤ TOL)] ≥ 1 − ν (13) must be an integer. Additionally, α3 must be an integer to
TOL→0
generate a smooth perturbation.
with cost Mesh P0 , which is now after the application of (15), is


 O TOL−2 , q2 > dγ also rotated and scaled using the simple transformation

 −2 −1 2
 ′   
W (TOL) ≤ O TOL log(TOL ) , q2 = dγ . xm xm

 ym′ 
= L̄(lx , ly , lz )R̄x (ϕx )R̄y (ϕy )R̄z (ϕz )  ym  .

 dγ−q2
O TOL− 2+ q1

 , q2 < dγ zm′
zm
(14) (16)
This theorem shows that even in the worst case scenario, Here, (xm , ym , zm ), and (x′m , ym ′ ′
, zm ) are the coordinates
the CMLMC algorithm has a lower computational cost than of node m before and after the transformation, matrices
that of the traditional (single level) MC method, which R̄x (ϕx ), R̄y (ϕy ), and R̄z (ϕz ) perform rotations around
scales as O(TOL−2−dγ/q1 ). Furthermore, in the best case x, y, and z axes by angles ϕx , ϕy , and ϕz , and matrix
scenario presented above, the computational
cost of the L̄(lx , ly , lz ) implements (down) scaling along x, y, and z
CMLMC algorithm scales as O TOL−2 , i.e. identical to axes by lx , ly , and lz , respectively.
characteristic
p impedance in Vi , i ∈ {0, 1} are εi , µi , and
ηi = µi /εi , respectively. It is assumed that the scatterer
is excited by an external electromagnetic field with electric
and magnetic components Einc (r) and Hinc (r). Using the
surface equivalence theorem and enforcing the tangential
continuity of total electromagnetic fields on S yield the
PMCHWT SIE [44]:
n
(a) (b)
n̂(r)×Einc (r) = n̂(r) × L0 [J](r) + L1 [J](r) (18)
o
− K0 [M](r) − K1 [M](r) , r ∈ S
n
n̂(r)×Hinc (r) = n̂(r) × K0 [J](r) + K1 [J](r) (19)
o
+ η0−2 L0 [M](r) + η1−2 L1 [M](r) , r ∈ S.
Here, n̂(r) is the outward pointing unit normal vector,

(c) (d) and J(r) and M(r) represent equivalent electric and mag-
netic surface current densities on S. The integral operators
Fig. 2. An example of four nested meshes with (a) 320, (b) 1280, (c) 5120, Li [.](r) and Ki [.](r) in (18) and (19) are
and (d) 20480 triangular elements, which are generated for a perturbed
shape with α1 = 2, α2 = 3, α3 = 2, a1 = 0.04 m, a2 = 0.048 m, Z
ϕx = 0.32 rad, ϕy = 0.88 rad, ϕz = 0.81 rad, lx = 1.06, ly = 1.08, ∇∇′
and lz = 1.07. Li [X](r) = jωµi Ī + · X(r′ )gi (r, r′ )ds′
ki2
ZS
The random variables used in generating the final version Ki [X](r) = ∇ × X(r′ )gi (r, r′ )ds′
S
of P0 are the perturbation weights ak , k = 1, . . . , K, the
rotation angles ϕx , ϕy , and ϕz , and the scaling factors lx , ′
where gi (r, r′ ) = e−jki |r−r | /(4π|r−r′ |) is the Green func-
ly , and lz , making the dimension of the stochastic space
tion of the Helmholtz equation in the unbounded medium
K + 6, i.e., random parameter vector √
with wave number ki = ω εi µi .
ξ = {a1 , . . . , aK , ϕx , ϕy , ϕz , lx , ly , lz }. (17) To numerically solve (18) and (19) for the unknowns J(r)
and M(r), S is discretized by triangular elements, and J(r)
Note that P0 is the coarsest mesh used in CMLMC (ℓ = and M(r) are approximated as
0) (see Fig. 2(a) for an example). The mesh of the next
level (ℓ = 1), P1 , is generated by refining each triangle of N
X
the perturbed P0 into four (by halving all three edges and J(r) = I¯n fn (r) (20)
connecting mid-points) [Fig. 2(b)]. The mesh at level ℓ = 2, n=1
P2 , is generated in the same way from P1 [Fig. 2(c)], and X2N
so on. All meshes Pℓ at all levels ℓ = 1, . . . , L are nested M(r) = I¯n fn−N (r) (21)
discretizations of P0 . This method of refinement results in n=N +1
β = 2 in (4). Note that no uncertainties are added on meshes
Pℓ , ℓ > 0; the uncertainty is introduced only at level ℓ = 0. where fn (r), n = 1, ..., N , represent Rao-Wilton-Glisson
It is assumed that P0 is fine enough to accurately represent basis functions [60], and I = [I1 , . . . , I2N ]T is the vector of
the variations of the highest harmonic in the expansion used unknown coefficients. Substituting (20) and (21) into (18)
in (15). and (19), and testing the resulting equations with fm (r),
m = 1, ..., N , yield the MoM system of equations:
C. Electromagnetic Solver
V̄ = Z̄ I¯ (22)
This section briefly describes the PMCHWT-SIE solver
used for the RCS and SCS of a dielectric scatterer with where the entries of the vector V̄ and the MOM matrix Z̄
surface S. It is assumed that sources and fields are time- are
harmonic, i.e., their time dependence varies as ejωt , where
(

t and ω are the time and angular frequency, respectively. fm , Einc , 1≤m≤N
Let V1 and V0 represent the space internal and exter- V̄m =
(23)
inc
nal to S, respectively. The permittivity, permeability, and fm−N , H , N + 1 ≤ m ≤ 2N

hfm , L0 [fn ] + L1 [fn ]i , experiments in Section III show that the CMLMC algorithm



 1 ≤ m ≤ N, 1 ≤ n ≤ N estimates the computational cost parameter as γ ≈ 1 for



 objects comparable to the wavelength in size.


− hfm−N , K0 [fn ] + K1 [fn ]i , To compute the RCS and SCS, the scatterer is excited

 N + 1 ≤ m ≤ 2N , 1 ≤ n ≤ N inc
by a plane wave: Einc (r) = E0 e−jk0 û ·r and Hinc (r) =
Z̄m,n = . (24) inc


 hfm , K0 [fn−N ] + K1 [fn−N ]i , (ûinc × E0 )/η0 e−jk0 û ·r , where ûinc is the direction of

 1 ≤ n ≤ N , N + 1 ≤ n ≤ 2N

 propagation. The unknown vector I¯ is solved for under this


excitation and the RCS σ rcs (ϑ, ϕ) along direction û(ϑ, ϕ) =
fm−N , η0−2 L0 [fn ] + η1−2 L1 [fn ] ,



 x̂ sin ϑ cos ϕ+ ŷ sin ϑ sin ϕ+ẑ cos ϑ is computed using [70]

N + 1 ≤ m ≤ 2N , N + 1 ≤ n ≤ 2N
R F(ϑ, ϕ)2
Here, the inner product is hfm , ai = S fm (r) · a(r)dr. rcs
σ (ϑ, ϕ) = 2 . (25)
Matrix equation (22) is solved iteratively for I. ¯ The 4π E0
computational cost of multiplying Z̄ with a trial solution Here, F(ϑ, ϕ) is the scattered electric field pattern in the far
vector scales as O(N iter N 2 ), where N iter is the number of field and computed using
iterations required for the residual error to reach the desired
level: typically N iter ≪ N . Likewise, the storage costs of F(ϑ, ϕ) = −jωµ0 û(ϑ, ϕ)
the unaccelerated/classical solution scale as O(N 2 ). N
X Z
I¯n
′
To minimize the computational and storage cost while × û(ϑ, ϕ) × fn (r′ )ejk0 û(ϑ,ϕ)·r dr′
Sn
executing the MLMC algorithm a fast multipole method n=1
(FMM)-fast Fourier transform (FFT) scheme is used. A 2N
X Z
I¯n
′
detailed formulation of FMM and its extension FMM-FTT − jk0 û(ϑ, ϕ) × fn−N (r′ )ejk0 û(ϑ,ϕ)·r dr′ .
n=N +1 Sn−N
can be found in [45]–[51]. The scheme encloses the scatterer
in a fictitious box that is embedded into a uniform grid of The SCS C sca (Ω) is obtained by integrating σ rcs (ϑ, ϕ) over
smaller boxes. Two non-empty boxes (i.e., boxes containing the solid angle Ω [70]:
at least a pair of patches) constitute a far-field pair if there 1
Z
is at least one box between them. Otherwise, they form a C sca (Ω) = σ rcs (ϑ, ϕ) sin ϑdϑdϕ. (26)
4π Ω
near-field pair. Interactions between basis and test functions
in near-field pairs (and same boxes) are computed using (24) The integral in (26) is efficiently computed using the exact
and stored in matrix Z̄ near . The contribution of near- and quadrature rule in [71].
self-interactions to the matrix-vector multiplication Z̄ I¯ is
calculated by simply multiplying Z̄ near with I. ¯ The inter- III. N UMERICAL RESULTS
actions between basis and test functions in far-field pairs In all numerical experiments considered in this section,
are represented using the radiation patterns of basis and the scatterer resides in free space (vacuum) with µ0 =
test functions (sampled over a solid angle of a sphere) 4π × 10−7 H/m and ε0 = c20 /µ0 F/m, where c0 is the
and a translation operator and their contributions to Z̄ I¯ are speed of light in vacuum, and the scatterer’s permittivity and
computed using the FMM-FFT scheme. The computational permeability are µ1 = µ0 and ε1 = 4ε0 , respectively. For the
cost and memory requirement of the FMM-FFT acceler- plane wave excitation, E0 = x̂, ûinc = −ẑ, and ω = 2πf
ated iterative solution scale as O(N iter N 4/3 log2/3 N ) and with f = 300 MHz. The QoI is the scattering cross-section
O(N 4/3 log2/3 N ), respectively [50], [51]. Note that these C sca (Ω) computed over the cone Ω = [ϑ0 , ϑ1 ] × [ϕ0 , ϕ1 ] =
estimates are higher than those of the multilevel FMM [61], [1/6, 11/36]π rad×[5/12, 19/36]π rad. For all random per-
[62]; the FMM-FFT scheme is preferred here since its paral- turbations, the parameters of the quasi-spherical harmonics
lel implementation is significantly simpler [47]–[51], [63]– are α1 = 2, α2 = 3, and α3 = 2. Meshes Pℓ at levels ℓ =
[69]. The parallelization strategy implemented here uses 0, 1, 2, 3, 4, which are generated using the method described
a hybrid message passing interface/open multiprocessing in Section II-B, have {320, 1280, 5120, 20480, 81920} tri-
(MPI/OpenMP) standard to uniformly distribute the memory angles, respectively. Note that not all five mesh levels are
and computational load among processors [47]–[51], [63]– required in every experiment. Since the CMLMC algorithm
[69]. is stochastic, 15 independent CMLMC runs are executed for
Note that the computational cost estimate provided above each experiment to statistically characterize the performance
is obtained under the assumption that the ratio of the of the method.
wavelength to the (average) element edge length stays the The matrix system in (22) is solved iteratively us-
same as N is decreased/increased [47]–[51]. Within the ing the transpose-free quasi-minimal residual (TFQMR)
CMLMC algorithm, the mesh is refined from one level to method [72] with a tolerance (residual) of ǫℓ = ǫ0 β −2ℓ ,
the next while the frequency is kept constant. This means where ℓ = 0, 1, 2, 3, 4 represent mesh levels, ǫ0 = 6.0 ×
that the ratio of the wavelength to the edge length increases 10−4 , and β = 2 (Section II-B). Using a level-dependent
and the computational cost of the FFT-FMM-accelerated tolerance is consistent with the dependence of the dis-
solver is expected to scale differently. Indeed, numerical cretization error in the QoI, which scales at a (heuristically
25
Sphere
20
Perturbed surface
15
(dB)
10
rcs
5
0
-5
-10
3 /4 /2 /4 0 /4 /2 3 /4
(rad)
(a) (b)
(a)
25
Sphere
20
Perturbed surface
15
(dB)
10
rcs
5
0
-5
-10
3 /4 /2 /4 0 /4 /2 3 /4
(c) (d)
(rad)
Fig. 3. Amplitudes of (a) J(r) and (b) M(r) induced on the unit sphere
under excitation by an x̂-polarized plane wave propagating in −ẑ direction (b)
at 300 MHz. Amplitudes of (c) J(r) and (d) M(r) induced on the
perturbed shape shown in Fig. 2(c) under excitation by the same plane Fig. 4. RCS of the unit sphere and the perturbed shape shown in Fig. 2(c)
wave. For all figures, amplitudes are normalized to 1 and plotted in dB computed on (a) xz and (b) yz planes under excitation by an x̂-polarized
scale. plane wave propagating in −ẑ direction at 300 MHz. For (a) ϕ = 0 and
ϕ = π rad in the first and second halves of the horizontal axis, respectively.
For (b), ϕ = π/2 rad and ϕ = 3π/2 rad in the first and second halves of

determined) rate O h2ℓ ∝ O β −2ℓ and ensures that the the horizontal axis, respectively.
number of iterations is kept in check when analyzing coarser
meshes. The integrals in (23) and (24) are computed using
sidered in the numerical experiments, i.e., f = 300 MHz,
the Gaussian quadrature rules; the accuracy of the rule, i.e.
ε1 = 4ε0 , and the size of the object is roughly 2 m.
number of quadrature points, adjusts with the mesh level:
More specifically, J(r) and M(r) induced on a (rotated and
it is seven for ℓ = 0, 1, 2 and six for ℓ = 3, 4. For all
scaled) perturbed geometry and its RCS are compared to
levels of meshes, the parameters of the FMM-FFT scheme
the same quantities of the unit sphere. The rotated, scaled,
are selected carefully to ensure that it has six digits of
and perturbed surface is generated using a1 = 0.04 m,
accuracy [61], [62]. All simulations are executed on Intel(R)
a2 = 0.048 m, ϕx = 0.32 rad, ϕy = 0.88 rad, ϕz =
Xeon(R) CPU E5-2680 v2 @ 2.80 GHz DELL workstations
0.81 rad, lx = 1.06, ly = 1.08, and lz = 1.07. The
with 40 cores and 128 GB RAM.
discretization is refined twice, resulting in the mesh with
For the examples in Sections III-B and III-C, the perfor-
5120 triangles [Fig. 2(c)]. A mesh with the same number
mance of the CMLMC is compared to that of the traditional
of triangles is generated on the unit sphere. Figs 3(a) and
MC method for varying values of TOL. Note that the
3(b) show normalized amplitudes of J(r) and M(r) induced
MC runs are not actually executed, but their execution
on this unit sphere. Fig. 3(c) and 3(d) plot the normalized
time is predicted using the results of the CMLMC runs.
amplitudes of J(r) and M(r) on the perturbed shape. It
More specifically, the variance Var(gL ) computed using the
is clear that there is a significant difference between the
CMLMC method is used to predict the number of required
current distributions induced on the sphere and the perturbed
MC samples as NM C ∼ Var(gL )TOL−2 . Also, note that,
shaped. Figs. 4(a) and 4(b) compare σ rcs (ϑ, ϕ) of the unit
since 15 independent CMLMC runs are executed for each
sphere and the perturbed shaped computed on the xz−
experiment, 15 MC curves are obtained and confidence
(ϑ ∈ [0, π] rad, ϕ = 0 and ϕ = π rad) and yz−
intervals are used on plots to visualize the results obtained
(ϑ ∈ [0, π] rad, ϕ = π/2 rad, and ϕ = 3π/2 rad) planes,
from these 15 curves.
respectively.
As expected, the RCS of the perturbed geometry is
A. Single Realization of Random Variables significantly different than that of the unit sphere. The results
This section demonstrates that the scattered fields strongly presented in Figs. 3 and 4 clearly demonstrate the need for a
depend on the shape of the object for the scenarios con- computational tool that can predict EM fields scattered from
objects with uncertain shapes parameterized using random 108

variables. TOL−2
Next convergence of the error in the SCS of the perturbed CMLMC
object is demonstrated. The reference SCS value used in the 107
MC Estimate
rel. error decay
Average Time (s)

1/N=1/h 2
-2
10 106
105
10 -3
Error in SCS
104 −3
10 10−2 10−1 100
10 -4 TOL
(a)
10 -5
106
10 4 10 5
TOL−2
# of Unknowns CMLMC
105
MC Estimate
Fig. 5. Relative norm error in SCS versus N .
Work estimate
104
computation of the relative norm error is obtained using the
finest mesh at the highest level. Figure 5 plots this error
versus N . A good convergence is observed. Having said 103
that, the error convergence depends on the specific random

perturbation, i.e., the realization of the random input vector 102
ξ, defined in (17). For some realizations that result surfaces
with large perturbations, the convergence of the SCS error
might be poor. Large number of such realizations results 101 −3
10 10−2 10−1 100
in non-optimal identification of CMLMC parameters q1 and TOL
q2 , and may slow down the CMLMC method. A possible (b)

remedy is not to allow very large perturbations.
Fig. 6. (a) Computation times required by the CMLMC and MC methods
vs. TOL. (b) Value of the computational cost estimate W [given by (5)] for
B. High variability the CMLMC and MC methods vs. TOL. Computation times are averaged
over 15 repetitions of the experiment.
In this example, the CMLMC algorithm is executed for
random variables a1 , a2 ∼ U [−0.14, 0.14] m, ϕx , ϕy , ϕz ∼
U [0.2, 3] rad, and lx , ly , lz ∼ U [0.9, 1.1]; here U [a, b] is similar computational resources. To achieve higher accuracy,
the uniform distribution between a and b. The CMLMC both the bias and the statistical error should be reduced. The
algorithm is run for TOL values ranging from 0.2 to 0.008. statistical error could be reduced by taking more samples,
At the lowest value of TOL, the CMLMC algorithm requires and the bias by using finer meshes (i.e., increasing ℓ).
five mesh levels, i.e., Pℓ , ℓ = 0, 1, 2, 3, 4. The third zone, 0.008 / TOL / 0.012, corresponds
Fig. 6(a) compares computation times for the CMLMC to the start of the asymptotic regime. The bias becomes
and MC algorithms as a function of TOL. The figure reveals important and finer meshes are required. MC computation
three convergence regimes (TOL zones). time increases rapidly and for the smallest TOL used, the
The first zone, 0.1 / TOL / 0.3, covers the range where CMLMC algorithm becomes almost 10 times faster than the
TOL is larger than the sum of bias and statistical error. No MC method. Note that in the first and second regimes the
additional samples and no additional mesh refinements are MC method may outperform the CMLMC algorithm since
required by the CMLMC algorithm. The MC and CMLMC the latter carries some initial overhead.
methods consume similar computational resources. For example, for TOL ' 0.012, only Pℓ , ℓ = 0, 1, 2, 3
The second zone, 0.012 / TOL / 0.1, corresponds are required, but for TOL / 0.012 an additional finer mesh,
to a pre-asymptotic regime. Both convergence rates are Pℓ , ℓ = 4, is required. Sampling on level ℓ = 4 is more
approximately 2. The statistical error is dominant, and many expensive; therefore, the MC computation time increases
new samples on existing meshes Pℓ , ℓ = 0, 1, 2 (by default) rapidly for TOL / 0.012. On the other hand, the CMLMC
are required. The MC and CMLMC methods again consume algorithm requires only very few samples on level ℓ = 4,
and is, therefore, significantly faster.

The computational cost estimate W is an indicator of 105
computation time. It depends on how the computational cost 22ℓ
of the deterministic solver changes from level ℓ − 1 to ℓ [as Gℓ
indicated by parameters γ and β in (4)] and on the order of 104
decay for the mean and the variance [parameters q1 , q2 in
(3a)]. Fig. 6(b) plots the values of W vs. TOL. The curve
Time (s)
is similar to that of the CMLMC computation time given in 103
Fig. 6(a) demonstrating that γ ≈ 1, q1 ≈ 3, and q2 ≈ 5 are
reasonably accurate.
Fig. 7(a) shows the time required to compute Gℓ = gℓ − 102
gℓ−1 vs. ℓ. Computation times vary roughly as 22ℓ , which
verifies that γ ≈ 1 [since d = 2 and β = 2 in (4)]. Fig. 7(b)
shows Eℓ = E[Gℓ ] vs. ℓ, revealing that Eℓ ∼ 2−3ℓ (assumed 101
0 1 2 3 4
weak convergence obtained with q1 = 3). Fig. 7(c) shows ℓ
Vℓ = Var[Gℓ ] vs ℓ, demonstrating that Vℓ ∼ 2−5ℓ (assumed
(a)
strong convergence with q2 = 5).
The results presented in Figs. 6(a) and 6(b), and Figs. 7(a),
7(b), and 7(c) confirm the assumptions stated in Section II-A 100
as well as the CMLMC scheme’s quasi-optimality. 2−3ℓ
Fig. 8 shows θ vs. TOL and demonstrates the complex 10 −1 Gℓ
relationship between TOL and the bias and statistical error.
θ decreases from 1 to ≈ 0.7. θ = 1 implies that the ratio 10−2
of bias to the total error is negligible, i.e. that the ratio of
of statistical error to the total error is 1. Such variability in 10−3
Eℓ
θ is one of the differences between the CMLMC algorithm

and the MLMC method where θ = 0.5. The blue dots show 10−4
that for TOL ≈ 0.2, θ ≈ 1, meaning that the impact of the
bias is negligible and that there is no need to further extend 10−5
the mesh hierarchy by adding finer mesh levels.
A higher accuracy can be achieved by decreasing either 10−6
0 1 2 3 4
the bias or the statistical error (i.e., smaller TOL). θ ≈ 1 ℓ
means that the bias is negligible and that the statistical
(b)
error should be decreased to achieve a smaller TOL. Only
when introducing a sufficient number of new samples will
the statistical error decrease and the ratio of bias to TOL 10−1
becomes higher. The bias can be decreased by including an 2−5ℓ
additional mesh level. After that, the ratio of bias to TOL is 10 −2 Gℓ
dropping again, and the statistical error becomes dominant
and should be decreased, etc. For TOL ≈ 0.01, the present 10−3
mesh hierarchy is not sufficient anymore, and the CMLMC
algorithm adds one more mesh level. 10−4
Vℓ
Another way to study the behavior of Gℓ is to look at

its probability density function (pdf). Figs. 9(a) and 9(b) 10−5
plot empirical pdfs of Gℓ from {2000, 400, 50} samples for
ℓ = 1 and ℓ = {2, 3}, respectively. Fig. 9(a) shows that 10−6
Var[G1 ] = Var[g1 − g0 ], where g0 and g1 are computed
using the meshes P0 and P1 , varies roughly in the range 10−7
(0, 2.5), i.e., this variance is large. Note that E[g1 − g0 ] ≈ 0 1 2 3 4
ℓ
1.0, but the following E[gℓ − gℓ−1 ] are close to zero, that is
why we show two plots. Fig. 9(b) shows that G2 = g2 − g1 , (c)
where g2 are g1 are computed using the meshes P2 and P1 , Fig. 7. (a) Time required to compute Gℓ vs. ℓ. (b) Eℓ = E[Gℓ ] vs. ℓ and
varies in the interval (−0.1, 0.1) and E[g2 − g1 ] ≈ 0.02. assumed weak convergence curve 2−3ℓ (q1 = 3). (c) Vℓ = Var[Gℓ ] vs.
Finally, G3 = g3 − g2 , where g3 are g2 are computed using ℓ and assumed strong convergence curve 2−5ℓ (q2 = 5). The experiment
is repeated 15 times independently and the obtained values are shown as
the meshes P3 and P2 , varies in the interval (−0.02, 0.02) error bars on the curves.
and E[G3 ] ≈ 0. The pdfs of Gℓ concentrate more and more
1.0 107
TOL−2
CMLMC
0.8
MC Estimate
106
Average Time (s)

0.6
θ
0.4
105
0.2
0.0 −3 104 −2
10 10−2 10−1 100 10 10−1 100
TOL TOL
(a)
Fig. 8. Value of θ used by the CMLMC algorithm vs. TOL. Computation
time is averaged over 15 repetitions of the experiment, i.e., there are 15
values of θ for a given value of TOL.
104
1
g1-g0 TOL−2
CMLMC
0.8
MC Estimate
0.6 Work estimate 103
0.4
0.2 102
0
-1 0 1 2 3 4
(a)
101 −2
120 10 10−1 100
g2-g1 TOL
g3-g2
100
(b)
80 Fig. 10. (a) Computation times required by the CMLMC and MC methods
vs. TOL. (b) Value of the computational cost estimate W [given by (4)] for
60 the CMLMC and MC methods vs. TOL. Computations times are averaged
over 15 repetitions of the experiment.
40
20 algorithm requires four mesh levels, i.e., Pℓ , ℓ = 0, 1, 2, 3

0
are the only meshes levels used for this experiment.
-0.2 -0.1 0 0.1 0.2 Fig. 10(a) compares the computation time of the CMLMC
(b) and MC methods as a function of TOL. There are two zones
in the figure. The first zone, 0.07 / TOL / 0.15, describes
Fig. 9. CMLMC pdfs of gℓ − gℓ−1 for (a) ℓ = 1 and (b) ℓ = {2, 3}.
the regime when TOL is higher than the sum of the bias
and statistical error. No additional samples or refinements
around zero [with the rates shown in Figs. 7(b), and 7(c)] are needed.
as ℓ increases. The second zone, TOL / 0.07, describes the pre-
asymptotic regime. As TOL gets smaller, the CMLMC
algorithm becomes more efficient than the MC method. For
C. Low Variability
values of TOL close to 0.02, the CMLMC algorithm is
In this example, the CMLMC algorithm is executed for roughly 10 times faster than MC. Fig. 10(b) shows the values
random variables a1 , a2 ∼ U [0.014, 0.49] m, ϕx , ϕy , ϕz ∼ of the computational cost estimate W vs. TOL. The curve
U [0.2, 1.0] rad, and lx , ly , lz ∼ U [0.8, 1.2]. The variability is similar to that of the CMLMC computation time given in
of these random variables is lower than that of the variables Fig. 10(a) validating the model developed for the CMLMC
used in the previous example. The value of TOL is changed algorithm. Figs. 11(a) and 11(b) show Eℓ = E[Gℓ ] and
from 0.2 to 0.02. For the lowest TOL value, the CMLMC Vℓ = Var[Gℓ ] vs. ℓ, respectively, revealing dependencies on
101
nested meshes). Later samples were better than ones used
2 −2ℓ for “training” the CMLMC method. The remedy to get a
Gℓ better fitting of curves in Figs. 11(a) and 11(b) is a more
100 accurate preliminary estimates of q1 and q2 . For this, we
need either to take more samples for training or to generate
10−1 “smoother” perturbations of the original shape.
Eℓ
IV. C ONCLUSION
10−2
A computational framework is developed to efficiently
and accurately characterize EM wave scattering from di-
10−3
electric objects with uncertain shapes. To this end, the
framework uses the CMLMC algorithm, which reduces
10−4
0 1 2
the computational cost of the traditional MC method by
ℓ performing most of the simulations with lower accuracy
(a)
and lower cost (using coarser meshes) and smaller number
of simulations with higher accuracy and higher cost (using
finer meshes). To increase the efficiency further, each of the
10−1 simulations is carried out using the FMM-FFT accelerated
2−4ℓ PMCHWT-SIE solver. Numerical results demonstrate that
Gℓ
10 −2
the CMLMC algorithm can be 10 times faster than the
traditional MC method depending on the amplitude of the
10−3 perturbations used for representing the uncertainties in the
scatterer’s shape. This work confirms that the known advan-
10−4 tages of the CMLMC algorithm can be observed when it is
Vℓ
applied to EM wave scattering: non-intrusiveness, dimension

10−5 independence, better convergence rates compared to the
classical MC method, and higher immunity to irregularity
10−6 w.r.t. uncertain parameters, than, for example, sparse grid
methods.
10−7 For optimal performance (for a given value of accuracy
0 1 2
ℓ parameter TOL), the CMLMC algorithm requires the mean
and the variance to have reliable convergence rates (i.e., one
(b)
should be able estimate q1 and q2 without much difference
Fig. 11. (a) Eℓ = E[Gℓ ] vs. ℓ and assumed weak convergence curve from one level to next). However, some random pertur-
2−2ℓ (q1 = 2). (b) Vℓ = Var[Gℓ ] vs. ℓ and assumed strong convergence bations may affect the convergence rates. With difficult-
curve 2−4ℓ (q2 = 4). The experiment is repeated 15 times and error bars
are shown on the curves. to-predict convergence rates, it is hard for the CMLMC
algorithm to estimate the computational cost W , the number
of levels L, the number of samples on each level Mℓ , the
ℓ that vary as 2−2ℓ (assuming weak convergence obtained
computation time, and the parameter θ, and the variance
with q1 = 2) and 2−4ℓ (assuming strong convergence ob-
in QoI. All these may result in a sub-optimal performance.
tained with q2 = 4). Note that in the previous example these
Indeed, numerical results demonstrate that there is a sig-
parameters are 3 and 5, respectively, demonstrating that the
nificant pre-asymptotic regime where the performance is
dependence of q1 and q2 on the variability of the random
not optimal. Additionally, it is observed that the settings of
variables used to described the geometry. Convergence in
the the FMM-FFT accelerated PMCHWT-SIE solver, which
Figs. 11(a) and 11(b) is faster than the predicted rates of
regulate the computation time and the accuracy (such as the
2−2ℓ and 2−4ℓ , respectively. A possible reason is that we
iterative solver threshold, the number of quadrature points,
are still in a pre-asymptotic regime and we were excessively
and the FMM-FFT parameters), have a significant effect on
pessimistic with factors q1 = 2 and q2 = 4. Although we
the performance of the CMLMC algorithm.
tried to avoid “outliers” in generating the random sphere-
like geometries, we could not completely avoid them. So,
some shape samples which were used for “training” of the ACKNOWLEDGEMENTS
CMLMC method, were “bad” and the deterministic solver The research reported in this publication was supported
produced solutions which were very different from others. by funding from King Abdullah University of Science and
As a result, we were forced to give pessimistic a priori Technology (KAUST), and by the Alexander von Hum-
estimates for q1 and q2 (we remind that the CMCLMC boldt Foundation. We gratefully acknowledge support from
method computes 10-20 samples on each of the first three Abdul-Lateef Haji-Ali (KAUST and Oxford), developer of
the CMLMC, Erik von Schwerin and Haakon Hoel for [19] L. Giraldi, A. Litvinenko, D. Liu, H. G. Matthies, and A. Nouy, “To
be or not to be intrusive? the solution of parametric and stochastic
valuable comments, concerning the CMLMC algorithm, as equations—the “plain vanilla” Galerkin case,” SIAM J. Sci. Comput.,
well as Ismail Enes Uysal and Hüseyin Arda Ülkü (KAUST) vol. 36, no. 6, pp. A2720–A2744, 2014.
for their initial efforts on this project. [20] A. Litvinenko, H. Matthies, and T. A. El-Moselhy, “Sampling and
low-rank tensor approximation of the response surface,” in Monte
Carlo and Quasi-Monte Carlo Methods 2012, ser. Springer Proceed-
ings in Mathematics & Statistics, J. Dick, F. Y. Kuo, G. W. Peters,
R EFERENCES and I. H. Sloan, Eds. Springer Berlin Heidelberg, 2013, vol. 65, pp.
535–551.
[1] L. Tsang, K. Ding, S. Huang, and X. Xu, “Electromagnetic computa- [21] S. Hosder, R. W. Walters, and M. Balch, “Efficient sampling for non-
tion in scattering of electromagnetic waves by random rough surface intrusive polynomial chaos applications with multiple uncertain input
and dense media in microwave remote sensing of land surfaces,” Proc. variables,” in Proceedings of the 48th AIAA/ASME/ASCE/AHS/ASC
IEEE, vol. 101, no. 2, pp. 255–279, Feb. 2013. Structures, Structural Dynamics, and Materials Conference, number
[2] L. Tsang, T. Liao, S. Tan, H. Huang, T. Qiao, and K. Ding, AIAA-2007-1939, Honolulu, HI, vol. 125, 2007.
“Rough surface and volume scattering of soil surfaces, ocean surfaces, [22] D. Liu, A. Litvinenko, C. Schillings, and V. Schulz, “Quantification
snow, and vegetation based on numerical Maxwell model of 3-D of airfoil geometry-induced aerodynamic uncertainties—comparison
simulations,” IEEE J. Sel. Topics Appl. Earth Observ., vol. 10, no. 11, of approaches,” SIAM/ASA J. Uncertain. Quantif., vol. 5, no. 1, pp.
pp. 4703–4720, Nov. 2017. 334–352, 2017.
[3] H. Wang, C. S. Levin, and N. J. Halas, “Nanosphere arrays with [23] D. Xiu and G. Karniadakis, “The wiener–askey polynomial chaos
controlled sub-10-nm gaps as surface-enhanced Raman spectroscopy for stochastic differential equations,” SIAM Journal on Scientific
substrates,” J. Am. Chem. Soc., vol. 127, no. 43, pp. 14 992–14 993, Computing, vol. 24, no. 2, pp. 619–644, 2002. [Online]. Available:
2005. https://doi.org/10.1137/S1064827501387826
[4] C. E. Talley, J. B. Jackson, C. Oubre, N. K. Grady, C. W. Hollars, [24] Y. Xing, D. Spina, A. Li, T. Dhaene, and W. Bogaerts, “Stochastic
S. M. Lane, T. R. Huser, P. Nordlander, and N. J. Halas, “Surface- collocation for device-level variability analysis in integrated photon-
enhanced Raman scattering from individual Au nanoparticles and ics,” Photon Res., vol. 4, no. 2, pp. 93–100, Apr. 2016.
nanoparticle dimer substrates,” Nano Lett., vol. 5, no. 8, pp. 1569– [25] T.-W. Weng, Z. Zhang, Z. Su, Y. Marzouk, A. Melloni, and L. Daniel,
1574, 2005. “Uncertainty quantification of silicon photonic devices with correlated
[5] N. Garcia and E. Stoll, “Monte Carlo calculation for electromagnetic- and non-Gaussian random parameters,” Opt. Lett., vol. 23, no. 4, pp.
wave scattering from random rough surfaces,” Phys. Rev. Lett., 4242–4254, Feb. 2015.
vol. 52, pp. 1798–1801, May 1984. [26] A. C. M. Austin, N. Sood, J. Siu, and C. D. Sarris, “Application
[6] M. Nieto-Vesperinas and J. M. Soto-Crespo, “Monte Carlo simula- of polynomial chaos to quantify uncertainty in deterministic channel
tions for scattering of electromagnetic waves from perfectly conduc- models,” IEEE Trans. Antennas Propag., vol. 61, no. 11, pp. 5754–
tive random rough surfaces,” Opt. Express, vol. 12, no. 12, pp. 979– 5761, Nov. 2013.
981, July 1987. [27] H. Bagci, A. C. Yucel, J. S. Hesthaven, and E. Michielssen, “A
[7] J. M. Soto-Crespo and M. Nieto-Vesperinas, “Electromagnetic scat- fast Stroud-based collocation method for statistically characterizing
tering from very rough random surfaces and deep reflection gratings,” EMI/EMC phenomena on complex platforms,” IEEE Trans. Electro-
J. Opt. Soc. Am. A, vol. 6, no. 3, pp. 367–384, Mar. 1989. magn. Compat., vol. 51, no. 2, pp. 301–311, May 2009.
[8] K. Sarabandi, P. F. Polatin, and F. T. Ulaby, “Monte Carlo simulation [28] A. C. Yucel, Y. Liu, H. Bagci, and E. Michielssen, “Statistical
of scattering from a layer of vertical cylinders,” IEEE Trans. Antennas characterization of electromagnetic wave propagation in mine envi-
Propag., vol. 41, no. 4, pp. 465–475, Apr. 1993. ronments,” IEEE Antennas Wireless Propag. Lett., vol. 12, pp. 1602–
[9] R. L. Wagner, J. Song, and W. C. Chew, “Monte Carlo simulation 1605, 2013.
of electromagnetic scattering from two-dimensional random rough [29] A. C. Yucel, H. Bagci, and E. Michielssen, “An adaptive multi-
surfaces,” IEEE Trans. Antennas Propag., vol. 45, no. 2, pp. 235– element probabilistic collocation method for statistical EMC/EMI
245, Feb. 1997. characterization,” IEEE Trans. Electromagn. Compat., vol. 55, no. 6,
[10] V. Jandhyala, E. Michielssen, S. Balasubramaniam, and W. C. Chew, pp. 1154–1168, Dec. 2013.
“A combined steepest descent-fast multipole algorithm for the fast [30] J. S. Ochoa and A. C. Cangellaris, “Random-space dimensionality
analysis of three-dimensional scattering by rough surfaces,” IEEE reduction for expedient yield estimation of passive microwave struc-
Trans. Geosci. Remote Sens., vol. 36, no. 3, pp. 738–748, May 1998. tures,” IEEE Trans. Microwave Theory Tech., vol. 61, no. 12, pp.
[11] L. Tsang, K. H. Ding, S. E. Shih, and J. A. Kong, “Scattering of 4313–4321, Dec. 2013.
electromagnetic waves from dense distributions of spheroidal particles [31] A. C. Yucel, H. Bagci, and E. Michielssen, “An ME-PC enhanced
based on Monte Carlo simulations,” J. Opt. Soc. Am. A, vol. 15, HDMR method for efficient statistical analysis of multiconductor
no. 10, pp. 2660–2669, Oct. 1998. transmission line networks,” IEEE Trans. Compon. Packag. Manuf.
[12] W. J. Morokoff and R. E. Caflisch, “Quasi-Monte Carlo integration,” Technol., vol. 5, no. 5, pp. 685–696, May 2015.
J. Comput. Phys., vol. 122, no. 2, pp. 218–230, 1995. [32] L. J. Gomez, A. C. Yucel, L. Hernandez-Garcia, S. F. Taylor, and
[13] F. Kuo, R. Scheichl, C. Schwab, I. Sloan, and E. Ullmann, “Multi- E. Michielssen, “Uncertainty quantification in transcranial magnetic
level quasi-monte carlo methods for lognormal diffusion problems,” stimulation via high-dimensional model representation,” IEEE Trans.
Mathematics of Computation, vol. 86, no. 308, pp. 2827–2860, 2017. Biomed. Eng., vol. 62, no. 1, pp. 361–372, Jan. 2015.
[14] D. Xiu, “Efficient collocational approach for parametric uncertainty [33] Z. Zeng and J.-M. Jin, “Efficient calculation of scattering variation
analysis,” Commun. Comput. Phys., vol. 2, no. 2, pp. 293–309, Apr. due to uncertain geometrical deviation,” Electromagn., vol. 27, no. 7,
2007. pp. 387–398, 2007.
[15] D. Xiu and J. S. Hesthaven, “High-order collocation methods for [34] C. Chauviére, J. S. Hesthaven, and L. Lurati, “Computational mod-
differential equations with random inputs,” SIAM J. Sci. Comput., eling of uncertainty in time-domain electromagnetics,” SIAM J. Sci.
vol. 27, no. 3, pp. 1118–1139, 2005. Comput., vol. 28, no. 2, pp. 751–775, 2006.
[16] J. Foo, X. Wan, and G. E. Karniadakis, “The multi-element probabilis- [35] C. Chauviére, J. S. Hesthaven, and L. C. Wilcox, “Efficient com-
tic collocation method (ME-PCM): Error analysis and applications,” putation of RCS from scatterers of uncertain shapes,” IEEE Trans.
J. Comput. Phys., vol. 227, no. 22, pp. 9572–9595, 2008. Electromagn. Compat., vol. 55, no. 5, pp. 1437–1448, May 2007.
[17] J. Foo and G. E. Karniadakis, “Multi-element probabilistic collocation [36] M. B. Giles, “Multilevel Monte Carlo path simulation,” Operations
method in high dimensions,” J. Comput. Phys., vol. 229, no. 5, pp. Research, vol. 56, no. 3, pp. 607–617, 2008.
1536–1557, 2010. [37] ——, “Multilevel Monte Carlo methods,” Acta Numerica, vol. 24,
[18] X. Ma and N. Zabaras, “An adaptive high-dimensional stochastic pp. 259–328, 2015.
model representation technique for the solution of stochastic partial [38] A. Barth, C. Schwab, and N. Zollinger, “Multi-level Monte Carlo
differential equations,” J. Comput. Phys., vol. 229, no. 10, pp. 3884– finite element method for elliptic PDEs with stochastic coefficients,”
3915, 2010. Numer. Math., vol. 119, no. 1, pp. 123–161, 2011.
[39] J. Charrier, R. Scheichl, and A. L. Teckentrup, “Finite element error [59] A. Litvinenko and H. G. Matthies, “Sparse data representation of
analysis of elliptic pdes with random coefficients and its application random fields,” Proc. Appl. Math. Mech., vol. 9, no. 1, pp. 587–588,
to multilevel Monte Carlo methods,” SIAM J. Numer. Anal., vol. 51, 2009.
no. 1, pp. 322–352, 2013. [60] S. Rao, D. Wilton, and A. Glisson, “Electromagnetic scattering by
[40] K. A. Cliffe, M. B. Giles, R. Scheichl, and A. L. Teckentrup, surfaces of arbitrary shape,” IEEE Trans. Antennas Propag., vol. 30,
“Multilevel Monte Carlo methods and applications to elliptic PDEs no. 3, pp. 409–418, May 1982.
with random coefficients,” Comput. Vis. Sci., vol. 14, no. 1, pp. 3–15, [61] J. Song and W. C. Chew, “Multilevel fast multipole algorithm for
2011. solving combined field integral equations of electromagnetic scatter-
[41] A. L. Teckentrup, R. Scheichl, M. B. Giles, and E. Ullmann, “Further ing,” Microw. Opt. Technol. Lett., vol. 10, no. 1, pp. 14–19, 1995.
analysis of multilevel Monte Carlo methods for elliptic PDEs with [62] J. Song, C.-C. Lu, and W. C. Chew, “Multilevel fast multipole
random coefficients,” Numer. Math., vol. 125, no. 3, pp. 569–600, algorithm for electromagnetic scattering by large complex objects,”
Nov. 2013. IEEE Trans. Antennas Propag., vol. 45, no. 10, pp. 1488–1493, Oct.
[42] N. Collier, A.-L. Haji-Ali, F. Nobile, E. von Schwerin, and R. Tem- 1997.
pone, “A continuation multilevel Monte Carlo algorithm,” BIT Numer. [63] O. Ergul and L. Gurel, “Efficient parallelization of the multilevel
Math., vol. 55, no. 2, pp. 399–432, 2015. fast multipole algorithm for the solution of large-scale scattering
problems,” IEEE Trans. Antennas Propag., vol. 56, no. 8, pp. 2335–
[43] D. Xiu and G. E. Karniadakis, “Modeling uncertainty in flow simu- 2345, Aug. 2008.
lations via generalized polynomial chaos,” Journal of Computational [64] ——, “A hierarchical partitioning strategy for an efficient paralleliza-
Physics, vol. 187, no. 1, pp. 137 – 167, 2003. tion of the multilevel fast multipole algorithm,” IEEE Trans. Antennas
[44] L. N. Medgyesi-Mitchang, J. M. Putnam, and M. B. Gedera, “Gener- Propag., vol. 57, no. 6, pp. 1740–1750, June 2009.
alized method of moments for three-dimensional penetrable scatter- [65] J. Fostier and F. Olyslager, “An asynchronous parallel MLFMA
ers,” J. Opt. Soc. Am., vol. A11, pp. 1383–1398, 1994. for scattering at multiple dielectric objects,” IEEE Trans. Antennas
[45] N. Engheta, W. D. Murphy, V. Rokhlin, and M. S. Vassiliou, “The fast Propag., vol. 56, no. 8, pp. 2346–2355, Aug. 2008.
multipole method (FMM) for electromagnetic scattering problems,” [66] B. Michiels, J. Fostier, I. Bogaert, and D. D. Zutter, “Weak scalability
IEEE Trans. Antennas Propag., vol. 40, no. 6, pp. 634–641, June analysis of the distributed-memory parallel MLFMA,” IEEE Trans.
1992. Antennas Propag., vol. 61, no. 11, pp. 5567–5574, Nov. 2013.
[46] R. Coifman, V. Rokhlin, and S. Wandzura, “The fast multipole method [67] Y. Liu, A. C. Yucel, H. Bagci, and E. Michielssen, “A scalable parallel
for the wave equation: a pedestrian prescription,” IEEE Antennas PWTD-accelerated sie solver for analyzing transient scattering from
Propag. Mag., vol. 35, no. 3, pp. 7–12, June 1993. electrically large objects,” IEEE Trans. Antennas Propag., vol. 64,
[47] C. Waltz, K. Sertel, M. A. Carr, B. C. Usner, and J. L. Volakis, no. 2, pp. 663–674, Feb. 2016.
“Massively parallel fast multipole method solutions of large elec- [68] Y. Liu, A. Al-Jarro, H. Bagci, and E. Michielssen, “Parallel PWTD-
tromagnetic scattering problems,” IEEE Trans. Antennas Propag., accelerated explicit solution of the time-domain electric field volume
vol. 55, no. 6, pp. 1810–1816, June 2007. integral equation,” IEEE Trans. Antennas Propag., vol. 64, no. 6, pp.
[48] J. M. Taboada, L. Landesa, F. Obelleiro, J. L. Rodriguez, J. M. 2378–2388, June 2016.
Bertolo, M. G. Araujo, J. C. Mourino, and A. Gomez, “High scala- [69] Y. Liu, A. C. Yucel, H. Bagci, A. C. Gilbert, and E. Michielssen, “A
bility FMM-FFT electromagnetic solver for supercomputer systems,” wavelet-enhanced PWTD-accelerated time-domain integral equation
IEEE Antennas Propag. Mag., vol. 51, pp. 20–28, 2009. solver for analysis of transient scattering from electrically large
[49] J. M. Taboada, M. G. Araujo, F. O. Basteiro, J. L. Rodriguez, and conducting objects,” IEEE Trans. Antennas Propag., vol. 66, no. 5,
L. Landesa, “MLFMA-FFT parallel algorithm for the solution of pp. 2458–2470, May 2018.
extremely large problems in electromagnetics,” Proc. IEEE, vol. 101, [70] A. Ishimaru, Wave Propagation and Scattering in Random Media.
pp. 350–363, 2013. New York, US: Academic Press, 1978.
[71] A. C. Yucel, “Helmholtz and high-frequency Maxwell multilevel
[50] A. C. Yucel, L. J. Gomez, and E. Michielssen, “Compression of fast multipole algorithms with self-tuning library,” Master’s thesis,
translation operator tensors in FMM-FFT-accelerated SIE solvers University of Michigan, Ann Arbor, Michigan, US, 2008.
via Tucker decomposition,” IEEE Antennas Wireless Propag. Lett., [72] R. W. Freund, “A transpose-free quasi-minimal residual algorithm for
vol. 16, pp. 2667–2670, 2017. non-Hermitian linear systems,” SIAM J. Sci. Stat. Comput., vol. 14,
[51] A. C. Yucel, W. Sheng, C. Zhou, Y. Liu, H. Bagci, and E. Michielssen, pp. 470–482, Mar. 1993.
“An FMM-FFT accelerated SIE simulator for analyzing EM wave
propagation in mine environments loaded with conductors,” IEEE J.
Multiscale and Multiphysics Comput. Techn., vol. 3, pp. 3–15, 2018.
[52] P. Ylä-Oijala, M. Taskinen, and S. Järvenpää, “Surface integral
equation formulations for solving electromagnetic scattering problems
with iterative methods,” Radio Science, vol. 40, no. 06, pp. 1–19, Dec
2005.
[53] A.-L. Haji-Ali, F. Nobile, E. von Schwerin, and R. Tempone, “Op-
timization of mesh hierarchies in multilevel Monte Carlo samplers,”
Stoch. Partial Differ. Equ. Anal. Comput., vol. 4, no. 1, pp. 76–112,
Mar. 2016.
[54] H. Hoel, E. Von Schwerin, A. Szepessy, and R. Tempone, “Implemen-
tation and analysis of an adaptive multilevel Monte Carlo algorithm,”
Monte Carlo Methods and Applications, vol. 20, no. 1, pp. 1–41,
2014.
[55] ——, “Adaptive multilevel Monte Carlo simulation,” in Numerical
Analysis of Multiscale Computations. Springer, 2012, pp. 217–234.
[56] A. Litvinenko and H. G. Matthies, “Numerical methods for uncer-
tainty quantification and bayesian update in aerodynamics,” in Man-
agement and Minimisation of Uncertainties and Errors in Numerical
Aerodynamics. Springer, 2013, pp. 265–282.
[57] A. Lang and C. Schwab, “Isotropic Gaussian random fields on the
sphere: Regularity, fast simulation and stochastic partial differential
equations,” ArXiv e-prints, May 2013.
[58] B. N. Khoromskij, A. Litvinenko, and H. G. Matthies, “Application of
hierarchical matrices for computing the Karhunen-Loève expansion,”
Computing, vol. 84, no. 1-2, pp. 49–67, 2009.
Alexander Litvinenko is a group Raul Tempone is an expert in Un-

leader at the Chair of Mathematics certainty Quantification. He is a Full
for Uncertainty Quantification, RWTH Professor at King Abdullah University
Aachen, Germany. In 2013-2018 he was of Science and Technology in Thuwal,
a research scientist at King Abdullah Saudi-Arabia, which he joined in 2009.
University of Science and Technology Prior to that, he moved between Sweden
(KAUST), where he developed statis- and the United States, taking his doctor-
tical and stochastic methods for solv- ate at the Royal Institute of Technology
ing PDEs with uncertainties. In 2007- in Stockholm and later becoming a Re-
2013, he was a PostDoc at the TU search Fellow and Associate Professor there. In between,
Braunschweig, Germany. During his Ph.D. research, 2002- he worked as a postdoc at the Institute for Computational
2006, at the Max Planck Institute for Mathematics in and Engineering Sciences (ICES) at the University of Texas,
the Sciences in Leipzig he developed hierarchical domain Austin, and as an Assistant Professor at Florida State
decomposition methods for solving multi-scale problems. University, Tallahassee. Raul Tempone is a member of
Alexander earned his B.Sc. degree (2000) and M.Sc. degree various scientific bodies, including the American Mathe-
(2002) in Mathematics at the Novosibirsk State University, matical Society and the Society for Industrial and Applied
Russia. His research interests are low-rank tensor methods Mathematics. In October 2018 he took up his position as a
for uncertainty quantification, Bayesian Update and data Humboldt Professor at RWTH Aachen University.
assimilation. Hakan Bagci received the B.S. de-
Abdulkadir C. Yucel received the gree in Electrical and Electronics En-
B.S. degree in electronics engineering gineering from the Bilkent University,
from Gebze Institute of Technology, Ankara, Turkey, in 2001; and the M.S.
Kocaeli, Turkey, in 2005, and the M.S. and Ph.D. degrees in Electrical and
and Ph.D. degrees in electrical engineer- Computer Engineering from the Univer-
ing from the University of Michigan, sity of Illinois at Urbana- Champaign
Ann Arbor, MI, USA, in 2008 and 2013, (UIUC), Urbana, IL, USA, in 2003 and
respectively. From September 2005 to 2007, respectively. From June 1999 to
August 2006, he worked as a Research July 2001, he worked as an Undergrad-
and Teaching Assistant at Gebze Insti- uate Researcher with the Computational Electromagnetics
tute of Technology. From August 2006 to April 2013, he Group, Bilkent University. From August 2001 to December
was a Graduate Student Research Assistant at the University 2006, he was a Research Assistant with the Center for
of Michigan. Between May 2013 and December 2017, he Computational Electromagnetics and Electromagnetics Lab-
worked as a Postdoctoral Research Fellow at the University oratory, UIUC. From January 2007 to August 2009, he was
of Michigan, Massachusetts Institute of Technology, and a Research Fellow with the Radiation Laboratory, University
King Abdullah University of Science and Technology. Since of Michigan, Ann Arbor, MI, USA. Since August 2009, he
2018, he has been working as an Assistant Professor at the has been with the King Abdullah University of Science and
School of Electrical and Electronic Engineering, Nanyang Technology (KAUST), Thuwal, Saudi Arabia, where he is
Technological University, Singapore. currently an Associate Professor of Electrical Engineering.
Dr. Yucel received the Fulbright Fellowship in 2006, His research interests include various aspects of theoretical
Electrical Engineering and Computer Science Departmental and applied computational electromagnetics with emphasis
Fellowship of University of Michigan in 2007, and Student on well-conditioned frequency and time domain integral
Paper Competition Honorable Mention Award at IEEE Inter- equation formulations and their discretization, hybrid time
national Symposium on Antennas and Propagation Sympo- domain integral and differential equation solvers, accurate,
sium in 2009. He has been serving as an Associate Editor for stable, and efficient marching schemes for time domain
the International Journal of Numerical Modelling: Electronic solvers, stochastic characterization of electromagnetic field
Networks, Devices and Fields and as a reviewer for various and wave interactions on complex geometries, and solu-
technical journals. His research interests include various tion of two and three dimensional electromagnetic inverse
aspects of computational electromagnetics with emphasis scattering problem using signal processing techniques. He
on uncertainty quantification for electromagnetic analysis authored or co-authored more than 90 journal papers and
on complex platforms, electromagnetic compatibility and in- more than 200 papers in conference proceedings. Dr. Bagci
terference analysis, nature-based design of electromagnetic was the recipient of the 2008 International Union of Radio
devices, and integral equation based frequency and time Scientists (URSI) Young Scientist Award and the 2004-
domain solvers and their accelerators. 2005 Interdisciplinary Graduate Fellowship from the Com-
putational Science and Engineering Department, UIUC. His
paper titled “Fast and Rigorous Analysis of EMC/EMI
Phenomena on Electrically Large and Complex Structures
Loaded With Coaxial Cables” was one of the three finalists bust optimizers for the synthesis of electromagnetic/optical
(with honorable mention) for the 2008 Richard B. Schulz devices.
Best Transactions Paper Award given by the IEEE Electro- Jesper Oppelstrup is an emeritus Professor at KTH
magnetic Compatibility Society. He authored (as student) or Royal Institute of Technology, Sweden. Jesper’s research
co-authored (as student and advisor) 17 finalist/honorable interests are computational fluid dynamics, numerical mod-
mention papers in the student paper competitions at the eling, numerical analysis, and engineering thermodynamics.
2005, 2008, 2010, 2014, 2015, 2016, 2017, 2018 IEEE An-
tennas and Propagation Society International Symposiums
and 2013, 2014, 2016, 2017, and 2018 Applied Computa-
tional Electromagnetics Society Conferences. Dr. Bagci is
currently an Associate Editor for the IEEE Transactions on
Antennas and Propagation and IEEE Journal on Multiscale
and Multiphysics Computational Techniques.
Eric Michielssen Eric Michielssen
(F’02) received his M.S. in Electri-
cal Engineering (Summa Cum Laude)
from the Katholieke Universiteit Leu-
ven (KUL, Belgium) in 1987, and his
Ph.D. in Electrical Engineering from
the University of Illinois at Urbana-
Champaign (UIUC) in 1992. He joined
the faculty of the UIUC Department of
Electrical and Computer Engineering in 1993, reaching the
rank of Full Professor in 2002. In 2005, he joined the
University of Michigan (UM) as Professor of Electrical
Engineering and Computer Science. Since 2013, he serves as
University of Michigan’s Associate Vice President for Ad-
vanced Research Computing, and also directs the Michigan
Institute for Computational Discovery and Engineering.
Dr. Michielssen received a Belgian American Educational
Foundation Fellowship in 1988 and a Schlumberger Fel-
lowship in 1990. Furthermore, he was the recipient of a
1994 International Union of Radio Scientists (URSI) Young
Scientist Fellowship, a 1995 National Science Foundation
CAREER Award, and the 1998 Applied Computational
Electromagnetics Society (ACES) Valued Service Award. In
addition, he was named 1999 URSI United States National
Committee Henry G. Booker Fellow and selected as the
recipient of the 1999 URSI Koga Gold Medal. He also
was awarded the UIUC’s 2001 Xerox Award for Faculty
Research, appointed 2002 Beckman Fellow in the UIUC
Center for Advanced Studies, named 2003 Scholar in the
Tel Aviv University Sackler Center for Advanced Studies,
selected as UIUC 2003 University and Sony Scholar. In
2011 he received the UM College of Engineering David
E. Liddle Research Excellence Award and in 2014 he was
the recipient of the IEEE APS Chen-To-Tai Distinguished
Educator Award. He is a Fellow of the IEEE (elected 2002)
and a member of URSI Commission B.
Dr. Michielssen authored or co-authored over 180 journal
papers and book chapters and over 350 papers in conference
proceedings. His research interests include all aspects of
theoretical and applied computational electromagnetics. His
research focuses on the development of fast frequency and
time domain integral-equation-based techniques for analyz-
ing electromagnetic phenomena, and the development of ro-

Computation of Electromagnetic Fields

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Computation of Electromagnetic Fields

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

Computation of Electromagnetic Fields Scattered

N UMERICAL methods for predicting radar and scat-

divides the parametric space into smaller domains, in each of

Here, n̂(r) is the outward pointing unit normal vector,

objects with uncertain shapes parameterized using random 108

rel. error decay

Average Time (s)

that, the error convergence depends on the specific random

q2 , and may slow down the CMLMC method. A possible (b)

and is, therefore, significantly faster.

θ is one of the differences between the CMLMC algorithm

Another way to study the behavior of Gℓ is to look at

Average Time (s)

0.6 Work estimate 103

20 algorithm requires four mesh levels, i.e., Pℓ , ℓ = 0, 1, 2, 3

applied to EM wave scattering: non-intrusiveness, dimension

Alexander Litvinenko is a group Raul Tempone is an expert in Un-

You might also like