Discontinuous Galerkin Time-Domain Method in Electromagnetics
Discontinuous Galerkin Time-Domain Method in Electromagnetics
Discontinuous Galerkin Time-Domain Method in Electromagnetics
Citation Dong, M., Chen, L., Li, P., Jiang, L., & Bagci, H. (2022).
Discontinuous Galerkin Time-Domain Method in
Electromagnetics: From Nanostructure Simulations to
Multiphysics Implementations. Advances in Time-Domain
Computational Electromagnetic Methods, 135–198. Portico.
https://doi.org/10.1002/9781119808404.ch4
DOI 10.1002/9781119808404.ch4
Publisher IEEE
Ming Dong,1 Liang Chen,1 Ping Li,2 Lijun Jiang,3 and Hakan
Bagci1
1
Electrical and Computer Engineering (ECE) Program, Division of Computer, Electrical,
and Mathematical Science and Engineering (CEMSE), King Abdullah University of
Science and Technology (KAUST), Thuwal, Saudi Arabia
2
Department of Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China
3
Department of Electrical and Electronic Engineering, The University of Hongkong,
Hongkong, China
1
implementation to solve Maxwell equations. With further developments in
more recent years, the DGTD method has become one of the preeminent
solutions to tackle a wide variety of challenging large scale electromagnetic
problems including those that require multiphysics modeling.
This chapter starts with a brief introduction to the DGTD method. This
introduction provides the fundamentals of numerical flux, discretization
techniques that rely on vector and nodal basis functions, and incorpora-
tion of absorbing boundary conditions. This is followed by descriptions of
a time-domain boundary integral(TDBI) scheme, which replaces absorb-
ing boundary conditions within the DGTD method, and a multi-step time
integration technique, which uses different time step sizes for the DGTD
and TDBI parts. Numerical results show that both techniques significantly
improve the efficiency, accuracy, and stability of the traditional DGTD
method. Then, the chapter continues with the applications of the DGTD
method to several real-life practical problems. More specifically, it describes
various novel techniques developed to enable the application of the DGTD
method to electromagnetic analysis of nanostructures and graphene-based
devices, and multiphysics simulation of optoelectronic antennas and source
generators. For each application, several numerical examples are provided
to demonstrate the accuracy, efficiency, and robustness of the developed
techniques.
2
1.1. Introduction to the Discontinuous
Galerkin Time-domain Method
Since the discontinuous Galerkin (DG) method was first applied to solve the
neutron transport equation by Reed and Hill in 1973 [Reed and Hill, 1973], it
time-domain (DGTD) methods have been developed rather quickly after that.
These methods can account for unstructured meshes and uses various Runge-
Kutta schemes to carry out the time integration [Hesthaven and Warburton,
2002, Kabakian et al., 2004, Fezoui et al., 2005, Shu, 2016a]. Since then, other
methods have been developed increasing the popularity of DGTD in the field
los and Lee, 2010b, Garcı́a De Abajo, 2013, Li et al., 2014a, Lin et al., 2016,
Harmon et al., 2016, Angulo et al., 2015, Yan et al., 2016, Ren et al., 2017,
Chen et al., 2018, Sirenko et al., 2018, Chen and Bagci, 2020b]. Finally, in more
including those that require multiphysics modeling [Gedney et al., 2007, Jin,
3
method and the finite element time-domain (FETD) method [Hesthaven and
elements is realized through the use of numerical flux. This approach localizes
element-level mass matrix blocks are factorized and stored very efficiently be-
scheme is used, DGTD becomes an efficient solver with very small memory
footprint (compared to classical time domain FEM). In addition, the block di-
agonal mass matrix makes DGTD very suitable for parallelization. It can easily
2015, Lin et al., 2016]. For example, an efficient explicit scheme can be used
solver can be used in subdomains with dense meshes. This approach allows for
the same large time step to be used throughout whole computation domain
As a final word, just like other time domain methods, DGTD offers sev-
broadband data with a single execution of the simulation and it allows for
hybridization with circuit solvers that can account for nonlinear components
4
[Li and Jiang, 2013c, Li et al., 2014a, Li and Jiang, 2013a].
DGTD developed for solving the Maxwell curl equations using vector basis
within the solver using numerical flux. This is followed by the description
functions to expand the three components of the electric and magnetic fields.
Then, the chapter continues with the applications of DGTD to several real-
objects calls for numerical solution of Maxwell equations for electric field E(r, t)
and magnetic field H(r, t) [Dosopoulos and Lee, 2010b] or the wave equation
for E(r, t) or H(r, t) [Sun et al., 2017]. Here, we focus on the formulation of
the DGTD schemes for the former and refer the readers to [Sun et al., 2017]
and references therein for the details for DGTD schemes developed for solving
the wave equation. Let J(r, t) represent the volumetric current density located
in the domain of interest, then Maxwell curl equations can be written as [Jin,
5
2002]
∂E(r, t)
ε(r) − ∇ × H(r, t) + J(r, t) = 0
∂t (1.1)
∂H(r, t)
µ(r) + ∇ × E(r, t) = 0
∂t
where ε(r) and µ(r) denote the permittivity and permeability respectively.
Note that for the sake of simplicity in the presentation, in the rest of the
∂U
+ ∇ · F(U) + S = 0. (1.2)
∂t
This can be achieved by defining field vector U = [E, H]T , excitation vector
−x̂ × H/ε −ŷ × H/ε −ẑ × H/ε
F(U) = x̂ + ŷ + ẑ. (1.3)
x̂ × E/µ ŷ × E/µ ẑ × E/µ
The flux vector F(U) can be further expressed as a multiplication of the purely
F(U) = AU (1.4)
6
where
0 − 1ε Cx 0 − 1ε Cy 0 − 1ε Cz
A= x̂ + ŷ + ẑ
1 1 1
µ Cx 0 µ Cy 0 µ Cz 0
(1.5)
0 0 0 0 0 1 0 −1 0
Cx =
0 0 −1 , Cy = 0 0 0 , Cz = 1 0 0 .
0 1 0 −1 0 0 0 0 0
It is shown later in this section that the eigenvalues of this flux matrix AU
are real (a necessary condition for the hyperbolic system) and they are the
Let Ω and ∂Ω denote the computation domain of interest and its boundary,
(1.1) in space with basis function Λi = [Ψei , Ψhi ] (Ψei and Ψhi are used to expand
Z
∂U
Λi · + ∇ · F(U) + S dr
Ωi ∂t
Z (1.6)
=− {n̂i · [F ∗ (U) − F(U)]} · Λi dr
∂Ωi
where n̂i is the outward pointing unit normal vector of ∂Ωi and n̂i · F ∗ (U) is
the numerical flux. Since the solutions in adjacent elements are allowed to be
solution.
7
The definition of the numerical flux starts by solving the corresponding Rie-
Hugoniot jump condition [Sankaran, 2007, Jeffrey, 2011], the characteristic plot
of the Maxwell system has to be obtained in advance from the flux matrix de-
fined as
εi 0 0 Ci
Ă = n̂i · A = · (1.7)
0 µi −C i 0
0 ni,z −ni,y
Ci =
−ni,z 0 ni,x . (1.8)
ni,y −ni,x 0
The flux matrix Ă has six eigenvalues given by [Sankaran, 2007, Jeffrey, 2011]
λ1,4 = 0
and they contain important information about the solution of the Riemann
problem. For hyperbolic system, the solutions remain unchanged while prop-
8
Sankaran, 2007, Jeffrey, 2011]
where j ∈ N (i), N (i) is the set of elements that are neighboring element i, and
U∗ and U∗∗ represent two intermediate states. The explicit jump relations for
1
n̂i × (E∗ − Ei ) = −ci (H∗ − Hi ) (1.13)
µi
1
n̂i × (Hi − H∗ ) = −ci (E∗ − Ei ). (1.14)
εi
1
n̂i × (Ej − E∗∗ ) = cj (Hj − H∗∗ ) (1.17)
µj
1
n̂i × (H∗∗ − Hj ) = cj (Ej − E∗∗ ). (1.18)
εj
With the above jump conditions, we obtain the explicit expression of the
9
numerical flux n̂i · F ∗ (U) = [−n̂i × H∗ , n̂i × E∗ ]T as
p
where Zi,j = µi,j /εi,j and Yi,j = 1/Zi,j denote the characteristic impedance
and (1.20) is termed “upwind flux”, which is the most commonly used type
the rest of this chapter, it is assumed that upwind numerical flux is unless
otherwise is stated.
Substituting (1.19) and (1.20) into (1.6), we get two separate equations as
Z Z
∂Ei
εi − ∇ × Hi + Ji · Ψei,k dr = Ψei,k · [n̂i × (H∗ − Hi )] dr (1.21)
Ωi ∂t ∂Ω
Z Z i
∂Hi
µi + ∇ × Ei · Ψhi,k dr = Ψhi,k · [n̂i × (Ei − E∗ )] dr. (1.22)
Ωi ∂t ∂Ωi
Next, the (local) electric and magnetic fields in element i, Ei and Hi , are
expanded using vector basis functions Ψei,l and Ψhi,l , respectively as [Jin, 2002]
K
X K
ei,l (t)Ψei,l (r), hi,l (t)Ψhi,l (r).
P
Ei (r, t) = Hi (r, t) = (1.23)
l=1 l=1
where ei,l and hi,l are time-dependent unknown coefficients of Ψei,l and Ψhi,l ,
respectively. Note that the numbers of basis functions used for expanding Ei
and Hi do not have to be the same. Similarly, the number of basis functions
can be different in different elements. But for the sake of simplicity in the
notation, we assume that all elements use the same number of basis functions
10
for both Ei and Hi , and that number is K.
∂ei
Mei = Sei hi − ji + Fee ee eh eh
ii ei − Fij ej − Fii hi + Fij hj (1.24)
∂t
∂hi
Mhi = −Shi ei + Fhh hh he he
ii hi − Fij hj + Fii ei − Fij ej (1.25)
∂t
where ei and hi are time-dependent vectors that store unknown basis function
coefficients, Mei and Mhi are the mass matrices, Sei and Shi are the stiffness
ee/eh
matrices, ji is a vector that stores the tested current source, and Fii/ij and
hh/he
Fii/ij are the flux matrices. The entries of these matrices and vectors are given
11
by [Li et al., 2014b, Li and Jiang, 2015]:
Note that the semi-discrete equation (1.24) and (1.25) are integrated in time
for all i = 1, · · · , N . Before the time integration/time marching starts, Mei and
(as described below), the resulting DGTD solver is very efficient with a small
12
Next, we describe how the semi-discrete system in equation (1.24) and
burton, 2008] and the leap-frog scheme that relies on central finite difference
∆t
un+1 = un + (an + 2bn + 2cn + dn ) (1.27)
6
f (un + cn ∆t).
Leap-frog scheme: One can use the central finite difference method to ap-
proximate the two time derivatives in (1.24) and (1.25) at staggered time points
Inserting equation (1.28) and (1.29) into equation (1.24) and (1.25) yields
13
the update equations resulting in the so-called leap frog time marching scheme:
Mei en+1
i = Mei eni
(1.30)
n+ 1 n+ 1 n+ 21 n+ 12
+∆t Sei hi 2 − ji 2 + Fee n
ii ei − Fee n
ij ej − Feh h
ii i + Feh h
ij j
n+ 32 n+ 12
Mhi hi = Mhi hi
1 1
(1.31)
h n+1 hh n+ 2 hh n+ 2 he n+1 he n+1
−∆t Si ei − Fii hi + Fij hj − Fii ei + Fij ej .
are used as examples here because of their popularity. Other explicit schemes
can also be used to integrate (1.24) and (1.25) in time [Liu et al., 2012, Angulo
et al., 2015].
ical flux. The expressions of the upwind flux in equation (1.19) and (1.20) show
that the computation of numerical flux for a given element requires the field
values on that element and its neighbors (this can be thought of as a decom-
position of numerical flux into incoming flux and outgoing flux [Jeffrey, 2011]).
element does not exist and the incoming flux is not immediately available in
the form of field values. In this case, the relevant boundary condition is used
to modify the incoming flux expression and update the numerical flux of the
conditions (ABCs) and the boundary conditions for perfect electrically con-
14
incorporated within DGTD using the numerical flux.
2002]. On the truncation surface, a boundary condition that imitates the radia-
is widely used due to its simplicity and ease of implementation even though
by simply setting the field values Ej and Hj in the numerical flux expres-
sion to zero since all outgoing waves are assumed to be absorbed without any
Zi Hi + n̂i × Ei
n̂i × H∗ = n̂i ×
2Zi
(1.33)
Y E
i i − n̂i × Hi
n̂i × E∗ = n̂i × .
2Yi
domains is described. This approach hybridizes the DGTD scheme with a time
domain boundary integral (TDBI) method, but just like ABCs, the field from
TDBI are coupled to DGTD via the use of the numerical flux. Please read
15
1.1.2.2. Boundary condition on perfect electrically con-
numerical flux to
1
n̂i × H∗ = n̂i × (Hi + n̂i × Ei )
Zi (1.34)
∗
n̂i × E = 0.
and Hj = Hi .
numerical flux to
1
n̂i × E∗ = n̂i × (Ei − n̂i × Hi )
Yi (1.35)
∗
n̂i × H = 0.
and Hj = −Hi .
16
1.1.3. Hybridization with Time-domain Bound-
To this end, one can use mathematically exact absorbing boundary conditions
(EACs) [Hagstrom et al., 2010, Sirenko et al., 2013] and their localized but ap-
proximate versions (ABCs) [Mur, 1981, Higdon, 1986]. EACs are constructed
using the outgoing wave modes defined on the truncation surface and there-
fore they can only be enforced on planar and/or spherical boundaries. Their
accuracy depends on the number of these modes, which cannot be easily esti-
mated for a prescribed accuracy of the solution [Hagstrom et al., 2010, Sirenko
et al., 2013]. Localized ABCs are computationally less expensive but their ac-
surface [Mur, 1981, Higdon, 1986]. Another approach to imitate the radiation
(PMLs) [Berenger, 1994, Gedney, 1996, Chen et al., 2020b]. The attenuation
in PML does not depend on the incidence angle [Fezoui et al., 2005], but it
decreases at low frequencies [Gedney et al., 2012, Gedney and Zhao, 2010]
making the PML less accurate and/or more computationally expensive. Addi-
in the solution [Abarbanel et al., 2002] and reduce its accuracy due to spurious
17
{Einc (r,t), H inc (r,t)}
∂Ω
ε 0 µ0 Jh
∂Γ Je
∂ΩT
ε (r) µ (r)
In this section, we describe a scheme [Li et al., 2014b, Dong et al., 2020,
2022] that combines DGTD with TDBI to truncate computation domains with-
out suffering from the shortcomings of the other methods briefly described
above. TDBI represents the fields on the truncation surface in the form of a
retarded time boundary integral defined over a Huygens surface enclosing the
condition and ensures that the accuracy of the solution is limited only by the
form to the scatterer’s shape and can be located very close to its surface, the
as DGTDBI, is described.
The permittivity and permeability inside the scatterer and in the background
medium are represented by ε(r) and µ(r) and ε0 and µ0 , respectively. The
scatterer is excited by the electromagnetic field Einc (r, t), Hinc (r, t) , and the
18
total and scattered field (TF/SF) technique is used to introduce the excitation
facilitate TDBI. The spatial discretization used in Ω is exactly same as the one
described in Section 1.1.1. DGTD solves for the scattered field in the region
between ∂Ω and ∂ΩT , and solves for the total field in the region enclosed by
∂ΩT .
In this set up, let the upwind numerical flux [Fezoui et al., 2005, Dosopoulos
and Lee, 2010b, Hesthaven and Warburton, 2008, Dosopoulos and Lee, 2010a,
Zi Hi + n̂i,m × Ei + Z̃ H̃ − n̂i,m × Ẽ
n̂i,m × H∗m = n̂i,m ×
Zi + Z̃
(1.36)
Yi Ei − n̂i,m × Hi + Ỹ Ẽ + n̂i,m × H̃
n̂i,m × E∗m = n̂i,m ×
Yi + Ỹ
where n̂i,m is the outward pointing unit normal vector on ∂Ωi,m , which is
the m-th facet of the boundary ∂Ωi of element i. In (1.36), the outgoing flux
involves Ei and Hi (fields in Ωi ) and the incoming flux involves Ẽ and H̃ (fields
between fields in Ωi and other elements “touching” ∂Ωi,m , (ii) enforce the
boundary condition on ∂Ω, and (iii) introduce {Einc , Hinc } on ∂ΩT [Alvarez
19
et al., 2010]. Accordingly, Ẽ and H̃ in (1.36) are selected as
E∂Ω ∂Ωi,m ∈ ∂Ω
i,m
Ẽ =
Ej ± Einc
i,m ∂Ωi,m ∈ ∂ΩT , (1.37)
Ej else
H∂Ω ∂Ωi,m ∈ ∂Ω
i,m
H̃ =
Hj ± Hinc
i,m ∂Ωi,m ∈ ∂ΩT . (1.38)
Hj
else
Here, j is the index of the element that neighbors element i via ∂Ωi,m , E∂Ω
i,m
incident fields computed on ∂Ωi,m ∈ ∂ΩT , and finally, ‘+’/‘−’ sign should be
p
where Z0 = µ0 /ε0 and Y0 = 1/Z0 are the wave impedance and admittance
E∂Ω ∂Ω
i,m and Hi,m [as they appear in (1.37) and (1.38)] on ∂Ωi,m ∈ ∂Ω are
computed using the TDBI method as described next. Let ∂Ωi′ ,m′ represent the
triangular facets that discretize ∂Γ. Here, i′ runs over the indices of elements
that are outside the volume enclosed by ∂Γ and have at least three nodes
residing on ∂Γ, and m′ runs over the indices of each element’s facets that are
described by these nodes. On facet ∂Ωi′ ,m′ , the equivalent electric and magnetic
surface currents, Jhi′ ,m′ and Jei′ ,m′ [Jiao et al., 2001, Shanker et al., 2005] can
20
be expressed as
K
X
Jhi′ ,m′ (r, t) = fih′ ,l′ (t)n̂i′ ,m′ (r) × Ψhi′ ,l′ (r)
l′ =1
(1.40)
XK
Jei′ ,m′ (r, t) = − fie′ ,l′ (t)n̂i′ ,m′ (r) × Ψei′ ,l′ (r).
l′ =1
Then, E∂Ω ∂Ω h
i,m and Hi,m on ∂Ωi,m ∈ ∂Ω are constructed using currents Ji′ ,m′ and
XXh i
E∂Ω
i,m (r, t) = µ0 Li′ ,m′ (Jhi′ ,m′ ) − Ki′ ,m′ (Jei′ ,m′ )
i′ m′
XXh i (1.41)
H∂Ω
i,m (r, t) = ε0 Li′ ,m′ (Jei′ ,m′ ) + Ki′ ,m′ (Jhi′ ,m′ )
i′ m′
Here, the operators Li′ ,m′ and Ki′ ,m′ are given by [Jin, 2002, Dong et al., 2020,
2022]
∂t J(r′ , t − R/c0 ) ′
Z
1
Li′ ,m′ (J) = − dr
4π∂Ωi′ ,m′ R
Z t−R/c0 ′
c20 ∇ · J(r′ , t′ ) ′ ′
Z
+ ∇ dt dr (1.42)
4π ∂Ωi′ ,m′ 0 R
J(r′ , t − R/c0 ) ′
Z
1
Ki′ ,m′ (J) = ∇ × dr (1.43)
4π ∂Ωi′ ,m′ R
background medium.
To account for TDBI, the semi-discrete system in (1.24) and (1.25) (Section
21
1.1.3) is updated using equation (1.36) - (1.43):
X
εi Mei ∂t ei =[Seh eh ee
i − Fii ]hi + Fii ei + [Feh ee
ij hj − Fij ej ]
j
− tee eh ee eh
i + ti ∓ bi ± bi (1.44)
X
µi Mhi ∂t hi =[−She he hh
i + Fii ]ei + Fii hi − [Fhe hh
ij ej + Fij hj ]
j
− thh he hh he
i − ti ∓ bi ∓ bi (1.45)
where the mass matrices, stiffness matrices and flux matrices are the same
as those in Section 1.1.1. The non-zero entries of the remaining matrices and
(K K
)
XX X X
iei,m = µ0 Li′ ,m′ (fih′ ,l′ n̂i′ ,m′ × Ψhi′ ,l′ ) + Ki′ ,m′ (fie′ ,l′ n̂i′ ,m′ × Ψei′ ,k′ )
i′ m′ l′ =1 l′ =1
(K K
)
XX X X
ihi,m = Ki′ ,m′ (fih′ ,l′ n̂i′ ,m′ × Ψhi′ ,l′ ) − ε0 Li′ ,m′ (fie′ ,l′ n̂i′ ,m′ × Ψei′ ,l′ ) .
i′ m′ l′ =1 l′ =1
22
Figure 1.2: The illustration of temporal interpolation process.
to obtain the samples of ei and hi . It should be emphasized that the time step
size used by DGTD and TDBI does not need to be same. Let ∆tDG represent
the time step size of the integration scheme used. To ensure stability, ∆tDG
the order and type of spatial discretization and time integration schemes as de-
scribed in detail in [Fezoui et al., 2005, Dosopoulos and Lee, 2010b, Hesthaven
and Warburton, 2008, Gedney et al., 2007, Dosopoulos and Lee, 2010a, Alvarez
et al., 2010]. On the other hand, the time step size used by TDBI, denoted by
∆tBI , is only required to resolve the maximum frequency of the excitation and
is independent from the spatial discretization [Yilmaz et al., 2004], [Bagci et al.,
2007]. In a typical scenario ∆tBI ≫ ∆tDG ; therefore, if one uses a single time
step size for the hybrid DGTDBI scheme and sets it to ∆tDG , the computa-
the computation of fip′ ,l′ (k∆tBI − R/c0 ). But during marching their samples at
k∆tDG , i.e., fip′ ,l′ (l∆tDG ), are available. Therefore, an interpolation scheme is
needed to compute fip′ ,l′ (k∆tBI − R/c0 ) from fip′ ,l′ (l∆tDG ). From the opposite
side, only samples ipi,m (k∆tBI ) are available during time marching. But time-
23
integration scheme requires samples tpp pp̃
i (l∆tDG ) and ti (l∆tDG ). Therefore,
polynomials [Yilmaz et al., 2004], [Bagci et al., 2007]. Also, time derivative
in operator Li′ ,m′ is moved onto this interpolation function. We should also
emphasize here that, to maintain the explicitness of the time marching, which
is one of the main advantages of DGTD scheme over classical finite element
methods, the minimum distance between any two points on ∂Γ and ∂Ω has to
( −0.1, 0.173, 0 )
!!
( 0.2, 0, 0 )
(0, 0, 0) !!
0.1 0.015
0.015
!!
( −0.1, − 0.173, 0 )
Figure 1.3: Cluster of three PEC spheres. All dimensions are in meters. Repro-
duced with permission from [Li et al., 2014b].
accuracy and the applicability of DGTDBI. In the first example, the scatterer
is a cluster of three PEC spheres residing in free space (Figure 1.3). Use of
24
Figure 1.4: Transient scattered electric field computed at the origin of cluster
of PEC spheres. Reproduced with permission from [Li et al., 2014b].
of three equally sized domains, each of which encloses one of the spheres.
are located 0.03 m and 0.015 m away from the sphere surfaces (Figure 1.3). For
the first simulation, the cluster is excited with a plane wave with parameters
fm = 500 MHz, t0 = 13.66 ns, and τm = 2.6 ns. During the simulation, the
transient scattered electric near field computed at the origin is recorded. Figure
1.4 compares this field to that computed using a time-domain integral equation
(TDIE) solver [Yilmaz et al., 2004], [Bagci et al., 2007]. Results agree well. For
the second simulation, τm = 0.65 ns while other excitation parameters are kept
the Huygens surface. Figure 1.5 and 1.6 plot the RCS on the xy- and xz-
planes computed at 2.53 MHz and 1.003 GHz, respectively. Results agree very
well with those obtained from the solution of an in-house method of moments
(MoM) solver. These two figures clearly demonstrate that the “absorption”
25
(a) (b)
Figure 1.5: RCS on (a) xy- and (b) xz- planes computed at 2.53 MHz from
the solutions of the DGTDBI method and MoM. Reproduced with permission
from [Li et al., 2014b].
1.7). The boundary of the computation domain and the Huygens surface con-
form to the shape of the scatterer and are located 0.02 m and 0.01 m away
from its surface. The scatterer is excited with a plane wave with parameters
fm = 500 MHz, t0 = 13.66 ns, and τm = 2.6 ns. Two simulations are carried
out: (i) The scatterer surface is PEC. (ii) The scatterer is a dielectric body
and monostatic RCS of the scatterers are computed. Figure 1.8 (a) and (b)
and Figure 1.9 (a) and (b) plot the bistatic and monostatic RCS of the PEC
Plots clearly demonstrate that the results agree very well with those obtained
from an MoM solver, which further verifies the accuracy and also low-frequency
absorbing capability of the DGTDBI method. Additionally, Figure 1.10 (a) and
(b) compare the RCS of the PEC scatterer on the xz- and yz- planes computed
at 1 GHz to those obtained by the same DGTD method, which uses ABC in-
stead of TDBI. It is clearly shown that ABC is not accurate for this structure.
26
(a) (b)
Figure 1.6: RCS on (a) xy- and (b) xz- planes computed at 1.003 GHz from
the solutions of the DGTDBI method and MoM. Reproduced with permission
from [Li et al., 2014b].
instead of nodal basis functions to avoid spurious solutions [Jin, 2002]. On the
other hand, for DGTD, it is found that spurious modes dissipate much faster
than physical modes when upwind or penalized fluxes are used, and therefore
they can be avoided even without using vector basis functions [Hesthaven and
functions cannot ensure that the solutions are globally divergence-free. From
this viewpoint, the main advantages of using vector basis functions in FEM do
27
0.05
0.2
0.01
0.01
0.05
0.1
0.3
Figure 1.7: U-shaped scatterer. All dimensions are in meters. Reproduced with
permission from [Li et al., 2014b].
are expanded separately and the continuity between neighboring elements for
All these reasons point out to the fact that nodal basis functions are a
2008]
Np Np
X X
Eiν (r, t) ≃ Eiν (rl , t)Ψl (r) = eνi,l (t)Ψl (r) (1.47)
l=1 l=1
Np Np
X X
Hiν (r, t) ≃ Hiν (rl , t)Ψl (r) = hνi,l (t)Ψl (r) (1.48)
l=1 l=1
28
(a) (b)
Figure 1.8: The bistatic (a) and monostatic (b) RCS of the PEC U-shaped
scatterer computed at a range of frequencies changing from 286.4 KHz to 1.0
GHz. Reproduced with permission from [Li et al., 2014b].
components of vectors in the Cartesian coordinate system, and eνi,l (t) and hνi,l (t)
∂exi
εi = Dyi hzi − Dzi hyi − Fi fie,x − jxi (1.49)
∂t
∂ey
εi i = Dzi hxi − Dxi hzi − Fi fie,y − jyi (1.50)
∂t
∂ezi
εi = Dxi hyi − Dyi hxi − Fi fie,z − jzi (1.51)
∂t
∂hx
µi i = −Dyi ezi + Dzi eyi + Fi fih,x (1.52)
∂t
∂hyi
µi = −Dzi exi + Dxi ezi + Fi fih,y (1.53)
∂t
∂hz
µi i = −Dxi eyi + Dyi exi + Fi fih,z (1.54)
∂t
where eνi = [eνi,1 , ..., eνi,Np ]T and hνi = [hνi,1 , ..., hνi,Np ]T are unknown vectors,
29
(a) (b)
Figure 1.9: The bistatic (a) and monostatic (b) RCS of the dielectric U-shaped
scatterer computed at a range of frequencies changing from 100 MHz to 1.0
GHz. Reproduced with permission from [Li et al., 2014b].
e,ν
jνi = [ji,1
ν , ..., j ν T ν ν
i,Np ] is the current density vector, ji,l = ji (rl , t), and fi and
n̂ × [Ei − E∗i ], respectively. Here, the numerical flux H∗i and E∗i are defined the
matrix defined as
Z
Mi (l, k) = Ψl (r)Ψk (r)dr (1.55)
Ωi
Z
Sνi (l, k) = Ψl (r)∂ν Ψk (r)dr (1.56)
Ωi
I
Li (l, k) = Ψl (r)Ψk (r)dr. (1.57)
∂Ωi
30
(a) (b)
Figure 1.10: The calculated RCS in xoz (a) and yoz (b) planes with DGTDBI
method and its comparisons from MoM and DGTD-ABC at 1.00013 GHz.
Reproduced with permission from [Li et al., 2014b].
plicit schemes just like it is done for the semi-discrete system in (1.30)-(1.31),
last decade due to their remarkable electrical, thermal, and mechanical prop-
thick layer of carbon atoms that are arranged in a honeycomb lattice. In the
last few years, graphene has found various applications in different academic
research fields as well as industry [Sharma et al., 2014, Geim and Novoselov,
31
2007, Schwierz, 2010, Tamagnone et al., 2012, Koppens et al., 2011, Lovat,
vices is one of these research fields. Indeed, several methods, which rely on
method of moments, [Hanson, 2008, Shapoval et al., 2013], the finite differ-
ence time-domain, [Lin et al., 2012, Nayyeri et al., 2013, Feizi et al., 2018],
DGTD [Li et al., 2015, Li and Jiang, 2015, Li et al., 2018], and circuit repre-
the gigahertz-terahertz frequency band. All these methods model the atomi-
(SIBC) to avoid very fine volumetric discretization of the graphene layer, which
would dramatically increase the number of unknowns and decrease the time-
step size (through CFL condition) in the case of explicit time-domain methods
static biasing or a tensor [Hanson, 2008, Lovat, 2012] due to quantum hall effect
ods, which use finite integration and auxiliary differential equation (ADE)
methods, respectively, to take into account the temporal dispersion effects and
anisotropic surface conductivity in time domain [Li et al., 2015, Li and Jiang,
2015].
32
1.2.1.1. A Resistive Boundary Condition to Represent
and Semlyen, 1999], [Gustavsen, 2006]. Then, using inverse Laplace transform,
obtained. Finally, these matrix equations are solved by time-domain finite in-
tegration technique (FIT). For elements not touching the graphene sheet, the
tions as described in Section 1.1.1. We emphasize here again that the use of
volumetric mesh that would be defined inside the thin graphene sheet is now
avoided. In the rest of the section, we first present the formulation underly-
ing the DGTD scheme extended to account for graphene’s SIBC model and
where ε indicates the energy state, ℏ denotes the reduced Planck constant, −q
33
−1
is the charge of electron, and fd (ε) = e(ε−µc )/kB T + 1
is the Fremi-Dirac
jq 2 kB T
µc
−µc /kB T
σintra =− 2 + 2ln e +1 (1.59)
πℏ (ω − j2Γ) kB T
while the second term corresponds to the interband contribution, which can
be approximated by
jq 2
2 |µc | − (ω − j2Γ)ℏ
σinter =− ln (1.60)
4πℏ 2 |µc | + (ω − j2Γ)ℏ
when kB T ≪ ℏω and |µc |. On the resistive thin sheet with the finite conduc-
n̂ × (E2 − E1 ) = 0
(1.61)
2 1
n̂ × (H − H ) = σg Et
where the superscripts 1 and 2 represent the two sides of the graphene sheet,
n̂ × (E1 + E2 )/2 is the tangential component of the electric field along the
graphene sheet. Note that (1.61) is expressed in frequency domain (or Laplace
domain).
numerical flux (see (1.13)-(1.20)) in Section 1.1.1 for the jump conditions).
More specifically, the jump condition for magnetic field for characteristics xn =
34
0 is updated as
n̂i,fg × (H∗∗ ∗
fg − H fg ) = σ g E t (1.62)
for element facets fg that touch the graphene sheet. Here, i represents the
index of these elements. Note that only one facet can touch the graphene sheet
for a given element. Using (1.62) together with the other jump conditions
counterparts), the upwind numerical fluxes for facets that touch the graphene
The numerical flux in (1.63) and (1.64) cannot be directly converted to time
basis functions in the Laplace domain (using FRVF technique [Gustavsen and
35
and Semlyen, 1999, Gustavsen, 2006]:
P
up s−p
P
u1 s−1 + u2 s−2 + · · · + uP s−P p=1
σg (s) = −1 −2 −Q
= Q . (1.65)
d0 + d1 s + d2 s + · · · + dQ s
s−q
P
dq
q=0
Here, s is the Laplace state variable, up and dq are the pth and the qth coeffi-
cients, and P and Q are the orders of numerator and denominator, respectively.
The tested DG equations for Maxwell curl equations in Laplace domain are
given as
Z 4
X Z
Ψei,k · εi Ei −s−1 ∇ × Hi dr = s−1 Ψei,k · n̂i,f ×(H∗f − Hi ) dr (1.66)
Ωi f =1 ∂Ωi,f
Z 4
X Z
Ψhi,k · µi Hi +s−1 ∇ × Ei dr = s−1 Ψhi,k · n̂i,f ×(Ei − E∗f ) dr. (1.67)
Ωi f =1 ∂Ωi,f
Note that on the right-hand side of (1.66) and (1.67), when f = fg , the nu-
merical flux expression in (1.63) and (1.64) is used. For other facets, numerical
Inserting (1.63) - (1.65) in (1.66) and (1.67), and using the Fourier trans-
form pair
Z Z
−n
s F (s) ↔ ··· | ·{z
f (τ ) dτ · · dτ} (1.68)
| {z } n
n
in the resulting equation yields the following time-domain matrix equations [Li
36
et al., 2015]:
Mei (d0 ei,0 +d1 ei,1 +· · ·+dQ ei,Q )−Sei (d0 hi,1 +d1 hi,2 + · · · + dQ hi,Q+1 )
4
ee,σg σg ee,σ σ
X
= Fee ee eh eh
ii,f ei,f +Fij,f ej,f +Fii,f hi,f +Fij,f hi,f +Fij ej +Fii g ei g (1.69)
f =1 | {z }
Fe,σg
Mhi (d0 hi,0 +d1 hi,1 +· · ·+dQ hi,Q )+Shi (d0 ei,1 +d1 ei,2 +· · · +dQ ei,Q+1 )
4
he,σg σg he,σ σ
X
= Fhh h
ii,f i,f +F hh
h
ij,f i,f +F he
e
ii,f i,f +F he
ij,f i,f +Fij
e ej +Fii g ei g (1.70)
f =1 | {z }
Fh,σg
e/h e/h
where Mi and Si are mass and stiffness matrices and Fee , Feh , Fhh , and
Fhe are the flux matrices. Matrix entries for those matrices can be found in
ee,σg ee,σg he,σg he,σg
Section 1.1.1. The entries of the matrices Fii , Fij Fii , and Fij are
given by:
Zj,fg
Z
ee,σ
[Fii g ]kl = Ψe · n̂i,fg × (n̂i,fg × Ψei,l ) dr
2(Zi + Zj,fg ) ∂Ωi,j i,k
Zj,fg
Z
ee,σg
[Fij ]kl = Ψe · n̂i,fg × (n̂i,fg × Ψej,l ) dr
2(Zi + Zj,fg ) ∂Ωi,j i,k
Z
he,σg 1
[Fii ]kl = Ψh · (n̂i,fg × Ψei,l ) dr
2(Yi + Yj,fg ) ∂Ωi,j i,k
Z
he,σg 1
[Fij ]kl = Ψh · (n̂i,fg × Ψej,l ) dr.
2(Yi + Yj,fg ) ∂Ωi,j i,k
are non-zero only for faces (f = fg ) touching the graphene sheet, otherwise
(f ̸= fg ) they are zero. Moreover, terms ei,q and hi,q are column vectors storing
37
the unknown coefficients eki and hli and are defined as
Z Z
eki,q = ··· eki (τ ) dτ
| ·{z
· · dτ}
| {z } q
q
Z Z (1.71)
hli,q = ··· hli (τ ) dτ
| ·{z
· · dτ}
| {z } q
q
Q
X Q
X
(e/h)i,f = dq (ei,f /hi,f )q+1 , (e/h)j,f = dq (ej,f /hj,f )q+1
q=0 q=0
(1.72)
P P
σ σ
X X
ei g = up ei,p+1 , ej g = up ej,fg ,p+1
p=1 p=1
Note that all terms in (1.72) involve integrals in time. To enable solution
of (1.69) and (1.70) using a time marching scheme, these integrals in time
are discretized using an FIT scheme. This scheme uses the rectangular rule
[Shapoval et al., 2013], which yields the discretized version of (1.71) with order
p or q at t = (n + 1)δt as:
n
X k3 X
X k2
en+1
i,p/q = (δt)p/q ··· ei,k1 +1
kp/q =0 k2 =0 k1 =0
n k3 X
k2
X X (1.73)
hn+1
i,p/q = (δt)p/q ··· hi,k1 +1
kp/q =0 k2 =0 k1 =0
en+1 n+1
i,0 = ei , hn+1 n+1
i,0 = hi .
To maintain the explicit nature of the time marching scheme, the terms
38
ward rectangular rule. This yields
n
X k3 X
X k2
en+1
j,p/q = (δt)p/q ··· ej,k1
kp/q=0 k2 =0 k1 =0
(1.74)
n
X k3 X
X k2
hn+1
j,p/q = (δt)p/q ··· hj,k1
kp/q=0 k2 =0 k1 =0
M̃ei S̃ei en+1
i F̃ei
= . (1.75)
S̃hi M̃hi hn+1
i F̃hi
Q Q P
ee,σg
X X X
M̃ei = dq (δt)q Mei − dq (δt)q+1 Fee
ii − up (δt)p+1 Fii
q=0 q=0 p=1
Q
X Q
X
S̃ei = − dq (δt)q+1 Sei + dq (δt)q+1 Feh
ii
q=0 q=0
(1.76)
Q
X Q
X
M̃hi = dq (δt)q Mhi − dq (δt)q+1 Fhh
ii
q=0 q=0
Q Q P
he,σg
X X X
S̃hi = dq (δt)q Shi − dq (δt)q+1 Fhe
ii − up (δt)p+1 Fii
q=0 q=0 p=1
39
and the column vectors are given by
XQ XQ
F̃ei = − dq (δt)q ẽn+1 e
i,q Mi + dq (δt)q+1 h̃n+1 e
i,q+1 Si
q=0 q=0
X Q X Q
+ dq (δt)q+1 ẽn+1
i,q+1 F ee
ii + d q (δt) q+1 e n+1 ee
j,q+1 Fij
q=0 q=0
X Q X Q
n+1 eh n+1
+ dq (δt)q+1 h̃i,q+1 Fii + dq (δt)q+1 hj,q+1 Feh
ij
q=0 q=0
X P X P
ee,σg ee,σ
+ up (δt)p+1 ẽn+1 F
i,p+1 ii + u p (δt) en+1
p+1 j,p+1 F g
ij
p=1 p=1
XQ XQ
F̃hi = − dq (δt)q h̃n+1 h
i,q Mi − dq (δt)q+1 ẽn+1
i,q+1 Si
h
q=0 q=0
X Q X Q
+ dq (δt)q+1 h̃n+1 F
i,q+1 ii
hh
+ d q (δt) h n+1
q+1 j,q+1 Fhh
ij
q=0 q=0
X Q X Q
+ dq (δt)q+1 ẽn+1 F
i,q+1 ii
he
+ dq (δt) q+1 e n+1
j,q+1 Fij
he
q=0 q=0
X P X P
he,σg he,σ
+ up(δt)p+1 ẽn+1
i,p+1 F ii + u p (δt) p+1 en+1
j,p+1 Fij g
p=1 p=1
where
ẽn+1 n+1
i,0 = 0 , h̃i,0 = 0 .
40
cursive schemes can be used [Shapoval et al., 2013]:
(e/h)n+1 n n+1
i/j,p/q = (e/h)i/j,p/q + δt · (e/h)i/j,p−1/q−1
(1.77)
(ẽ/h̃)n+1
i,p/q = (ẽ/h̃)ni,p/q + δt · (ẽ/h̃)n+1
i,p−1/q−1 .
Figure 1.11: Fitted magnitude and phase of the surface conductivity σg from
500 MHz to 10 THz with four poles. Reproduced with permission from [Li
et al., 2015].
demonstrate the accuracy and the applicability of the DGTD scheme described
above. In both examples, the excitation is a plane wave with electric field
In the first example, the scatterer is an infinitely large graphene sheet that
resides in free space on the xy- plane. The temperature used in the expression of
41
Figure 1.12: The calculated magnitudes of coefficients ΓT , ΓR and ΓA by the
proposed algorithm as well as the theoretical data for µc = 0.3eV. Reproduced
with permission from [Li et al., 2015].
σg is set to T = 300K while µc and Γ are changed for two different simulations.
The FRVF scheme with four poles is applied to the samples of σg in a fre-
quency range between microwave and terahertz. In Figure 1.11, the compar-
ison between the fitted values and the original data is shown for frequencies
using the DGTD scheme. The results are shown in Figure 1.12. For compar-
2008] are also presented in the same figure. Good agreement between two sets
of results is observed.
For the second simulation, the transmission of a plane wave through the
42
Figure 1.13: Magnitude of calculated transmission coefficient ΓT by the pro-
posed algorithm as well as theoretical data from terra hertz to near-infrared
region. Reproduced with permission from [Li et al., 2015].
seven poles is used to represent σg in the frequency range between 500 GHz
and 100 THz. Note that we use seven poles here since σg has a jump around
60 THz. The transmission coefficient computed using the DGTD scheme and
the analytical expression are compared in Figure 1.13. Again, very good agree-
ment is observed. We should mention here that, for these two simulations, the
condition [Mur, 1981] (see (1.32) in Section 1.1.2.1), which yields very accurate
(3-D) problem. Two simulations are carried out. In the first simulation, the
scatterer is a 5 × 10 µm2 graphene patch that resides in free space on the xy-
Γ= 1
2τ with τ = 10−13 s and the parameters of excitation are fm = 2.5THz,
k̂ = ẑ, τm = 1.274 × 10−13 s, and t0 = 3τm . The FRVF scheme with three
43
Figure 1.14: Normalized ECS versus frequency for a freestanding graphene
patch in [Llatser et al., 2012] as its counterpart calculated by integral equation
method. Reproduced with permission from [Li et al., 2015].
500 MHz and 5 THz. For comparison, the normalized extinction cross section
between 0.1 and 4 THz are computed using the DGTD scheme as shown in
Figure 1.14. The reference result in the same figure is obtained by using an
integral equation solver [Llatser et al., 2012]. The good agreement between the
results demonstrates the accuracy and applicability of the DGTD scheme for
3D examples.
sections of the graphene patch are computed for different values of µc . In Figure
1.15, the normalized cross sections between 1 and 10 THz are shown for µc =
0.5, 1.0 and 1.5eV. As expected, plasmon resonances are observed at various
addition, the radar cross section between 1 and 10 THz is presented in Figure
1.16. It is found that the peaks in the radar cross section happen at the plasmon
44
Figure 1.15: TSCS and ACS of the 50 × 50 µm2 freestanding graphene patch
corresponding to different chemical potentials µc = 0.5, 1.0, and 1.5 eV. Re-
produced with permission from [Li et al., 2015].
1.17, the normalized near-field patterns for the y-component of the electric
that the far-field patterns resemble those of conventional short dipoles, which
45
Figure 1.16: Calculated forward RCS of the 50 × 50 µm2 freestanding graphene
patch corresponding to different chemical potentials µc = 0.5, 1.0, and 1.5eV.
Reproduced with permission from [Li et al., 2015].
mention here that, for these two simulations for 3D problems, the computation
tostatic biasing. However, when there is a biasing magnetic field, the surface
dition (RBC) (representing the graphene sheet) within DGTD would result a
46
Figure 1.17: Normalized near-field distribution of Ey over the graphene sheet
at (a) f1 = 1.76THz ; (b) f2 = 4.98THz; and (c) f3 = 6.97THz for µc = 1.5eV
case. Reproduced with permission from [Li et al., 2015].
numerical flux in a tensor form. In addition, still, the dispersion of the surface
we introduce another DGTD scheme that can account for both the anisotropy
and the dispersion of the surface conductivity. In this scheme, a surface po-
larization current is introduced over the graphene sheet [Li and Jiang, 2015],
(ADE) that involves the electric field. With the ADE and the polarization
current, the anisotropic and the dispersive RBC is converted an isotropic and
corresponding numerical flux in the presence of the new RBC is made isotropic.
47
Figure 1.18: Normalized far-field pattern in E- and H-plane at f1 = 1.76THz
for µc = 1.5eV. Reproduced with permission from [Li et al., 2015].
of the Lorentz force, the surface conductivity exhibits anisotropy [Sounas and
Caloz, 2012]. To have a fundamental insight into this phenomenon, the motion
accelerated along the x- direction due to the electric force Fe = −eE with −e
denoting the electron charge. Simultaneously, the moving electron with velocity
where σxx and σyx are conductivities parallel and perpendicular to the electric
field E, respectively. Similarly, for the case with an electric field E = ŷEy , two
48
current components will be generated:
For graphene with same properties in all directions, the conductivities must
satisfy σxx = σyy and σyx = σxy . Based on this, the generated currents in (1.78)
and (1.79) can be combined together and re-written in a more compact form
J = σg · E (1.80)
σxx −σyx 0
σg =
σ yx σxx 0 .
(1.81)
0 0 0
The analytical expressions for the surface conductivities σxx and σyx con-
Kubo’s formula [Kubo, 1957] (see (1.59) and (1.60) in Section 1.2.1.1 also).
For frequencies in the terahertz range, the intraband term is dominant over
the interband term [Sounas and Caloz, 2012]. Thus, only the intraband term is
considered here. Based on the fact that the probability of electron transitions
between Landau levels around the Fermi level µc is the strongest over others,
the expressions for σxx and σyx can be approximated by a Drude-like model
49
[Wang et al., 2013, Sounas and Caloz, 2012]:
1 + jωτ
σxx (ω, µc , τ, T, B0 ) = σ0
(ωc τ ) + (1 + jωτ )2
2
(1.82)
ωc τ
σyx (ω, µc , τ, T, B0 ) = σ0
(ωc τ ) + (1 + jωτ )2
2
with
e2 τ kB T
µc
σ0 = + 2ln(e−µc /kB T + 1) (1.83)
πℏ2 kB T
ωc ≈ eB0 vF2 /µc is the cyclotron frequency with vF ≈ 106 m/s denoting the
Fermi velocity.
n̂ × (E2 − E1 ) = 0
(1.84)
n̂ × (H2 − H1 ) = σ g · E
Note that (1.81) is expressed in frequency domain and the only difference
tensor.
50
are entangled with each other). To address these challenges, an auxiliary polar-
as
∂J
= C · J + 2Γσ0 E (1.87)
∂t
with
−2Γ −ωc
C= . (1.88)
ωc −2Γ
With this auxiliary polarization surface current J and the ADE, the anisotropic
one in (1.88). Then, the incorporation of this scalar RBC within DGTD can be
For the DGTD-formulation, the electric field and magnetic field are ap-
51
proximated in the same way as described in the section 1.2.1.1. Other than
φi (r)
K
X
Ji = ci,l (t)φi,l (r) (1.89)
l=1
By applying the DG testing to the Maxwell curl equations and the ADE, we
obtain
Z 4 Z
∂Ei X
Ψei,k Ψei,k · n̂i,f × (H∗f − Hi ) dr
· εi − ∇ × Hi dr =
Ωi ∂t ∂Ωi,f
f =1
Z 4 Z
∂Hi X
Ψhi,k Ψhi,k · n̂i,f × (Ei − E∗f ) dr (1.90)
· µi + ∇ × Ei dr =
Ωi ∂t
f =1 ∂Ωi,f
Z Z Z
∂Ji
φi,k · dr = φi,k · C · Ji dr + 2Γσ0 φi,k · Ei dr.
∂Ωi,fg ∂t ∂Ωi,fg ∂Ωi,fg
The semi-discrete matrix equations for the ADE can then be obtained as
∂ci
J̄i = M̄c ci + 2Γσ0 M̄v ei (1.91)
∂t
where
Z
[Mc ]kl = φi,k · C · φi,l dr
∂Ωi,fg
Z
[Mv ]kl = φi,k · Ψei,l dr (1.92)
∂Ωi,fg
Z
i
J̄ kl = φi,k · φi,l dr.
∂Ωi,fg
We only give the matrix related to the ADE since other matrices are same as
those introduced in the section 1.2.1.1 [Li and Jiang, 2015]. In the remainder of
52
this section, two numerical examples are provided to demonstrate the accuracy
(a) (b)
(c)
Figure 1.19: (a) Total transmission coefficients, (b) Faraday rotation angle,
and (c) cross-polarized transmission coefficients versus frequency for different
magnetostatic biasing. Reproduced with permission from [Li and Jiang, 2015].
As the first example, we use the DGTD scheme to study the Faraday ro-
tation and surface plasmon polarization by varying the chemical potential and
the magnetostatic biasing. An infinitely large graphene sheet placed on the xy-
sheet is excited a plane wave with electric field Einc (r, t) = x̂g(t−k̂·r/c0 ), where
is a Gaussian pulse, where the pulse’s modulation frequency, duration, and de-
53
lay are set to fm = 5THz, τm = 6.37 × 10−13 s, and t0 = 5tm , respectively.
The first-order SM-ABC [Dosopoulos et al., 2013] (see also ((1.32)) in Section
which is sufficient for this example since all waves are normally incident to
the truncation boundary. In Fig. 19, the total transmission coefficient, Fara-
the DGTD scheme are presented. For comparison, the analytical results for
total transmission coefficient and Faraday rotation angle are also provided in
Figure 1.19. Very good agreement is observed between the two sets of results.
The results show that the magnetostatic biasing has significant impacts on the
The next example validates the accuracy of the DGTD scheme for 3D
54
(a)
(b)
Figure 1.21: (a) Total transmission coefficients, (b) Faraday rotation angle,
and (c) cross-polarized transmission coefficients versus frequency for different
magnetostatic biasing. Reproduced with permission from [Li and Jiang, 2015].
maintain the accuracy of the solution. In Figure 1.20, the extinction cross
section computed using the DGTD scheme and an integral equation solver
[Llatser et al., 2012] are compared. The figure shows the results agree well.
Then, a 20 × 100 µm2 graphene patch under biased by a static magnetic field
graphene resonance, the scattering and the extinction cross sections obtained
55
Figure 1.22: Normalized ECS for substrates Si, SiO2 , and Si3 N4 . Reproduced
with permission from [Li and Jiang, 2015].
Figure 1.21. It is noted that the resonant frequencies shift upwards with higher
chemical potentials, and the resonance becomes much stronger. Next, to study
the effect of the substrate material on the plasmon resonance, the graphene
are silicon (Si), silicon-nitrate (Si3 N4 ), and silicon-dioxide (SiO2 ) (all of which
Figure 1.22. It is noted that the first two resonant peaks of the extinction cross
section shift to low frequencies with the increasing permittivity due to the
56
1.2.2. Multiphysics simulation of optoelectronic
devices
The several advantages the DG methods comes with (as explained in Section
1.1.6) make them very suitable for multiphysics simulations. Indeed, in the
last decade, they have been applied to many multiphysics problems in various
tor carrier interactions [Homsi et al., 2017, Don, 2019, Dong et al., 2022]. In
order spatial basis functions and high-order time integration schemes. Further-
2006, Hesthaven and Warburton, 2008, Harmon et al., 2016, Shu, 2016b, Yan
namics, and their nonlinear coupling. We note here although the DG frame-
57
phototransistors and photovoltaic devices) with only minor modifications.
Initially, a bias voltage is applied to the electrodes. The resulting static electric
field changes the carrier distribution. The re-distributed carriers, in turn, affect
drift-diffusion equations [Chen and Bagci, 2020c, Vasileska et al., 2010, Chuang,
2012]
Jsc (r) = qµc (Es )Es (r)nsc (r) ± qdc (Es )∇nsc (r) (1.95)
where Es (r) is the static electric field, ε(r) is the dielectric permittivity, q is the
electron charge, C(r) is the doping concentration, the superscript “s” stands
for static, the subscript c ∈ {e, h} represents the carrier type and hereinafter
the upper and lower signs should be selected for electron (c = e) and hole
(c = h), respectively, nsc (r) is the carrier density, R(ne , nh ) is the recombina-
tion rate, and µc (Es ) and dc (Es ) are the mobility and diffusion coefficient, re-
temperature.
Two most common recombination processes are the trap assisted recom-
58
2020c, Vasileska et al., 2010, Chuang, 2012]
RAuger (nse , nsh ) = [nse (r)nsh (r) − ni 2 ][CeA nse (r) + ChA nsh (r)]
Here, ni is the intrinsic carrier concentration, τe and τh are the carrier lifetimes,
ne1 and nh1 are SRH model parameters related to the trap energy level, and CeA
and ChA are the Auger coefficients. Depending on the semiconductor material,
more recombination models can be included [Vasileska et al., 2010]. The net
terms.
proper mobility models with respect to the semiconductor materials and device
operating conditions [Selberherr, 1984, Vasileska et al., 2010, Sil, 2016, Min,
2017, COM, 2017]. For a photoconductive device working under a high bias
for the carrier velocity saturation effect [Vasileska et al., 2010, Sil, 2016, Min,
2017, COM, 2017]. Here, the Caughey-Thomas model [Vasileska et al., 2010,
!βc βc−1
µ0c E∥ (r)
µc (Es ) = µ0c 1 +
Vcsat
where E∥ (r) is amplitude of the electric field intensity parallel to the current
flow, µ0c is the low-field mobility, and Vcsat and βc are fitting parameters ob-
59
tained from experimental data.
cident on the device. The photoconductive material absorbs the optical elec-
tromagnetic wave energy induced on the device and generates carriers. The
carriers are driven by both the bias electric field and the optical electromag-
netic fields. The carrier dynamics and electromagnetic wave interactions are
∂H(r, t)
µ(r) = −∇ × E(r, t) (1.96)
∂t
∂E(r, t)
ε(r) = ∇ × H(r, t) − [Je (r, t) + Jh (r, t)] (1.97)
∂t
∂nc (r, t)
q = ±∇ · Jc (r, t) − q[R(ne , nh ) − G(E, H)] (1.98)
∂t
Jc (r, t) = qµc (E)E(r, t)nc (r, t) ± qdc (E)∇nc (r, t). (1.99)
Here, E(r, t) and H(r, t) are the total electric and magnetic field intensities,
ne (r, t) and nh (r, t) are the total electron and hole densities, Je (r, t) and Jh (r, t)
are the total current densities due to electron and hole movement, ε(r) and
µ(r) are the dielectric permittivity and permeability, R(ne , nh ) and G(E, H)
are the recombination and generation rates, µc (E) and dc (E) are mobility and
Since the bias voltage persists during the transient stage, and the boundary
conditions for Poisson and drift-diffusion equations, e.g., the Dirichlet bound-
and nsc (r) solved from the Poisson-drift-diffusion system are valid through-
out the transient stage. Therefore, the total field intensities and carrier and
60
E(r, t) = Es (r) + Et (r, t), H(r, t) = Hs (r) + Ht (r, t), nc (r, t) = nsc (r) + ntc (r, t),
Jc (r, t) = Jsc (r) + Jtc (r, t), respectively. Here, the superscript “t” stands for
Rt (nte , nth ). The mobility is assumed to be a function of Es (r) only, i.e., µc (E) ≈
static components from the left hand of (1.96)–(1.99), yielding a reduced cou-
∂Ht (r, t)
µ(r) = −∇ × Et (r, t) (1.100)
∂t
∂Et (r, t)
ε(r) = ∇ × Ht (r, t) − [Jte (r, t) + Jth (r, t)] (1.101)
∂t
∂nt (r, t)
q c = ±∇ · Jtc (r, t) − q[Rt (nte , nth ) − G(Et , Ht )] (1.102)
∂t
Jtc (r, t) = qµc (Es )([Es (r) + Et (r, t)]ntc (r, t) (1.103)
Here, the generation rate describes the generation of carriers upon absorption
of optical electromagnetic wave energy [Saleh and Teich, 2019, Chuang, 2012,
Piprek, 2018]
P abs (r, t)
G(Et , Ht ) = η
E ph
erated by each absorbed photon), P abs (r, t) is the absorbed power density of
ν is the frequency of the optical wave. Here, ν must be high enough such that
61
E ph is large enough to excite electrons, e.g., E ph should be larger than the
Chuang, 2012, Piprek, 2018]. Apparently, P abs (r, t) only depends on the op-
Figure 1.23: Time marching scheme. Source: [Chen and Bagci, 2020b].
tem 1.96-1.99. The steady-state solutions are used as inputs to the transient
solver. The steady-state solver uses the Gummel iteration method to take the
nonlinearity into account [Chen and Bagci, 2020c]. At every iteration of this
method, one has to solve two partial differential equations, i.e., the nonlin-
ear Poisson equation and the linearized drift-diffusion equation. The transient
equations using explicit time integration methods and takes into account their
coupling by alternately feeding the updated solutions into each other during
the time marching. Since the carrier response is much slower than the time
62
variation of the electromagnetic waves, the drift-diffusion equations are up-
dated using a larger time-step size (typically 10 times larger). This mixed-step
explicit time integration scheme can greatly reduce the computational cost.
The time marching scheme is illustrated in Fig. 1.23 [Chen and Bagci, 2020b].
twice the step size for the Maxwell equations. Let the time steps of the two
calculated from Et,T and Ht,T and then used to update the carrier densities
′ ′
nt,T
c (from step T ′ − ∆T ′ to T ′ ). Then, nt,T
c are used to compute the cur-
Maxwell equations at steps T and T +∆T to produce {Et,T +∆T , Ht,T +∆T } and
{Et,T +2∆T , Ht,T +2∆T }, respectively. The time steps of the two systems match
Note that G̃T +2∆T is calculated as the average value (GT +∆T + GT +2∆T )/2.
The above multiphysics framework calls for solving four partial differ-
ential equations, namely, the nonlinear Poisson equation and the stationary
than 40 years [Shu, 2016a]. The difficulty stems from the existence of exponen-
63
tial boundary layers in the solution, which makes numerical methods unstable
unless extremely fine meshes are used. To this end, FEM requires additional
Gummel scheme) poses restrictions on the meshes [Chen and Bagci, 2020c].
DG seems to be favorable for this type of problem since it is free from these
electric field components described by Poisson and Maxwell equations are used
in the drift-diffusion equations, nodal basis functions are preferred here. With
these discussions, all the four partial differential equations are discretized using
(LDG) method. Since the electron and hole drift-diffusion equations only differ
by the sign in front of the drift term, we only discuss the electron drift-diffusion
equation. Also note that the subscript “e” (meaning electron) and the super-
script “t” (meaning transient) are dropped from the variables to simplify the
as the three drift terms in (1.103) are treated in the same way, for brevity,
they are combined into one term and is denoted by v(r, t)n(r, t). Under those
64
are expressed as the following initial-boundary value problem
∂n(r, t)
= ∇ · [d(r)q(r, t)] + ∇ · [v(r, t)n(r, t)] − R(r, t), r ∈ Ω (1.104)
∂t
q(r, t) = ∇n(r, t), r ∈ Ω (1.105)
Here, q(r, t) is an auxiliary variable introduced to reduce the order of the spa-
tial derivative in the diffusion term, R(r, t) ≡ Rt (nte , nth ) − G(Et , Ht ), ∂ΩD and
∂ΩR represent the surfaces where Dirichlet and Robin boundary conditions are
enforced and fD and fR are the coefficients associated with these boundary con-
ditions, respectively, and n̂(r) denotes the outward normal vector ∂ΩR . For the
1994].
the previous process described for Maxwell equations. Here, the nodal basis
functions (see Section 1.1.6) are used. Testing (1.104) and (1.105) with Ψl (r)
on element i and applying the divergence theorem yield the weak form
Z Z I
ni (r, t)
Ψl (r)dV =− d(r)qi (r, t) · ∇Ψl (r)dV + n̂(r) · (dq)∗ Ψl (r)dS
Ωi ∂t Ωi ∂Ωi
Z
− v(r, t)ni (r, t)·∇Ψl (r)dV (1.108)
Ωi
I Z
∗
+ n̂(r) · (vn) Ψl (r)dS − R(r, t)Ψl (r)dV
∂Ωi Ωi
Z Z I
ν
qi (r, t)Ψl (r)dV =− ni (r, t)∂ν Ψl (r)dV + n̂ν (r)n∗ Ψl (r)dS. (1.109)
Ωi Ωi ∂Ωi
65
Here, n∗ , (dq)∗ , and (vn)∗ are numerical fluxes defined as
the diffusion term, 1.110 and 1.111 are the so-called LDG alternate flux [Cock-
burn et al., 2000]. The vector β̂ determines the upwind direction of n and (dq).
In LDG, opposite directions are chosen for n and (dq), while the precise direc-
tion of each variable is not important [Cockburn et al., 2000, Hesthaven and
1.112 is the local Lax-Friedrichs flux [Hesthaven and Warburton, 2008], with
Expanding ni (r, t) and qkν (r, t), ν ∈ {x, y, z}, using the set of Lagrange
polynomials
Np Np
X X
ni (r, t) ≃ ni (rl , t)Ψl (r) = ni,l (t)Ψl (r) (1.113)
l=1 l=1
Np Np
X X
qiν (r, t) ≃ qiν (rl , t)Ψl (r) = ν
qi,l (t)Ψl (r) (1.114)
l=1 l=1
66
yields the semi-discretized form
ni
Mi = Ci ni + Cii′ nk′ + Di d¯i qi + Dii′ d¯k′ qi′ − Bni (1.115)
∂t
Mqi qi = Gi nk + Gii′ ni′ + Bqi (1.116)
where the global unknown vectors are defined as ni = [ni,1 , ..., ni,Np ]T and
matrix with blocks Mi defined in 1.55. d¯k is a diagonal matrix with entries
d1 , ..., dN , where di = (dxi , dyi , dzi ) and dνi (i) = di (rl ), l = 1, ..., Np , ν ∈ {x, y, z}.
Matrices Gi and Gii′ , and Di and Dii′ correspond to the gradient and
the divergence operators, respectively. For the LDG flux, Di = −GTi and
Dii′ = −GTii′ . Gi is a 3Np ×Np matrix and it has contributions from the volume
1 + sign(β̂ · n̂)
I
Lνi (l, k) = θi (k) n̂ν (r)Ψl (r)Ψk (r)dS.
2 ∂Ωii′
Here, ∂Ωii′ denotes the interface connecting element i and i′ and θi (k) selects
Matrix Gii′ corresponds to the surface integral term in 1.109, which involves
T
the unknowns from neighboring elements of element i: Gii′ = Lxi′ Lyi′ Lzi′ ,
67
where
1 − sign(β̂ · n̂)
I
Lνi′ (l, k) = θi′ (j) n̂ν (r)Ψl (r)Ψk (r)dS.
2 ∂Ωii′
Similarly, matrix Ci has contributions from the volume integral and the
surface integral related to the drift term on the right hand side of (1.108):
Ci = CV S
i + Ci , where
X Z
CV
i (l, k) =− vν (r, t)∂ν Ψl (r)Ψk (r)dV
ν Ωi
and
I
1X
CSi (l, k) = θi (k) [ n̂ν (r)vν (r, t) + α(r, t)]Ψl (r)Ψk (r)dS.
∂Ωii′ 2 ν
Matrix Cii′ is from the last surface integral term in (1.108), which involves the
I
1X
Cii′ (l, k) = θi′ (k) [ n̂ν (r)vν (r, t) − α(r, t)]Ψl (r)Ψk (r)dS.
∂Ωii′ 2 ν
Matrices Bni and Bqi are contributed from the force term and boundary con-
Z I
Bni (k) = R(r, t)Ψl (r)dV + fR (r)Ψl (r)dS
Ωi ∂Ωi ∩∂ΩR
I
+ n̂(r) · v(r, t)fD (r)Ψl (r)dS
∂Ωi ∩∂ΩD
I
Bq,ν
i (k) = n̂ν (r)fD (r)Ψl (r)dS.
∂Ωi ∩∂ΩD
[Shu and Osher, 1988] is used for integrating 1.115-1.116 in time. With initial
68
value n(r, t = 0) = 0, time samples of the unknown vector ni are obtained step
by step in time.
formulations can be found in [Chen and Bagci, 2020c]. One should note that
differential equations. Solving them calls for solving global linear systems. For-
tunately, for photoconductive device simulation, one only needs to solve the
steady-state once (for each bias voltage), and the solution can be reused in
work, two practical device simulation examples are presented below. We first
1997]. This device has a relatively simple structure and it can be simulated us-
ing the FDTD [Moreno et al., 2014b] or FEM-based [Burford and El-Shenawee,
2016, Bashirpour et al., 2017] approach, where the carrier generation due to the
et al., 2017].
Since the focus of the device study is the photocurrent response, we only
consider the central gap region of the device. The cross section of the device is
shown in Fig. 1.24. The width and height of the semiconductor and substrate
layers are 7 µm and 0.5 µm, respectively. Both electrodes are made of gold and
the permittivity of gold is represented using the Drude model [Gedney et al.,
2012]. The Drude model parameters of gold at the optical frequencies [Olmon
69
et al., 2012] as well as other physical parameters are shown in Table 1.3.
1.5
0.5
y (µm)
–0.5
–1.5
–4.5 –3.5 –2.5 –1.5 –0.5 0.5 1.5 2.5 3.5 4.5
x (µm)
–4.5
Figure 1.24: Cross section of the conventional PCD considered in the first
example. Reproduced with permission from [Chen and Bagci, 2020b].
each of the above DG solvers, and the device should be meshed with respect
equation and the time-dependent Maxwell equations are solved in the whole
which is derived using the local charge neutrality [Schroeder, 1994, Vasileska
70
2. Poisson equation: On electrode surfaces, assuming ideal Ohmic contact,
assumption that the static fields reaching the boundary are small and do
magnetic wavelength at the optical and THz frequencies, the skin depth
2013] constrain the edge length used in the mesh discretizing the com-
p
putation domain. The Debye length lD = 2εVT /(qnc ) is a measure
element and CP = |µc E|/(2Dc ), has to be less than 1 to ensure the sta-
tions [Trangenstein, 2013]. We should note here that the potential distri-
bution is smooth, therefore the mesh using the edge length determined by
these space-scale parameters for the materials used in this example (see
Table 1.3) and the corresponding required edge lengths are provided in
71
Table 1.1: Length scales
Table 1.1. We should note that the last number in Table 1.1 is the small-
est geometry dimension and the mesh used for this example represents
Wang et al., 2015], constrain the time-step sizes used in the transient
solution. Table 1.2 lists the values of the maximum time-step sizes al-
lowed by these three CFL conditions with the edge length obtained from
the characteristic space scale discussion above. We should note that the
time-step sizes required to resolve the periods of the optical and THz
electromagnetic waves are larger than those required by their CFL con-
ditions. Table 1.2 clearly shows that the drift-diffusion equations can be
integrated using a larger time-step size than the one required for the
Using these scales as a guide, the device is discretized with mesh sizes in
the range [10, 200] nm, where the finest meshes are used near the boundaries of
the semiconductor layer and near the metallic electrodes, the maximum mesh
72
Table 1.2: Time scales
size in the semiconductor layer and the substrate is 70 nm, and the maximum
mesh size in the rest of the computation domain is 200 nm. The time-step sizes
for the Maxwell and drift-diffusion equations are selected to be 10−7 ps and
solver and used as inputs in the transient simulation. Here, we focus on the
tical electromagnetic waves propagating from top to bottom. The pulse shape
parameters are given in Table 1.3. The intensity of the electromagnetic field
Figs. 1.25 (a) and (b) show Hzt (r, t) and nte (r, t) at several time instants,
respectively, and Figs. 1.26 (a) and (b) show the time signatures of the optical
electromagnetic wave excitation and the generated THz current and their spec-
trum obtained using Fourier transform. Figs. 1.25 (a) and (b) and Fig. 1.26
a large part of the incident field’s energy is reflected back (see the reflected
wave above the air-semiconductor interface and behind the aperture). At the
same time, in the LT-GaAs layer, the incident field’s energy entering the device
73
Table 1.3: Physical parameters used for the PCD examples
is partially absorbed and carriers are generated near the air-semiconductor in-
terface. At t = 0.25 ps, the incident field reaches its pulse peak and the electron
density increases to ∼ 1011 cm−3 . The short incident field pulse passes quickly
and, after t = 0.5 ps, only some scattered fields reside in the high permittiv-
ity region due to internal reflections. During that time, the electron density
keeps increasing until the excitation pulse decays to of 20% its peak value at
t ≈ 0.4 ps, after t ≈ 0.4 ps, nte (r, t) decays slowly due to the recombination.
can be clearly observed that electrons move toward the anode on the left side.
The picture of holes (not shown) is similar but with holes moving toward the
cathode on the right side with a lower speed. The resulting current shown in
74
t = 0.05 ps t = 0.25 ps
t = 0.5 ps t = 1 ps
(a)
t=0.05 ps
t=0.25 ps
t=0.5 ps
t=0.75 ps
t=1 ps
t=1.25 ps
t=1.5 ps
t=1.75 ps
(b)
Figure 1.25: (a) |Hz (r, t)| computed using the DG scheme at different time
instants. The dotted line shows the aperture of optical EM wave excitation.
The gray box indicates the LT-GaAs region. (b) ne (r, t) computed using the
DG scheme at different time instants. Reproduced with permission from [Chen
and Bagci, 2020b].
Figs. 1.26 (a) and (b) match well with the result presented in [Moreno et al.,
2014b], where the latter has been shown to agree with experimental results very
well [Moreno et al., 2014b]. Because of the simplicity of the single interface
75
(a)
(b)
Figure 1.26: (a) The time signature and (b) the spectrum of the optical EM
wave excitation and the generated THz current. Reproduced with permission
from [Chen and Bagci, 2020b].
layer has almost the same pulse shape as the optical excitation signal. This
explains why the analytical generation rate [Moreno et al., 2014b] works very
well.
cross section of the device is shown in Fig. 1.27. For this example, gold nanos-
tructures are added between the two electrodes. The periodicity, the duty cycle,
and the thickness of the nanograting are 180 nm, 5/9, and 190 nm, respec-
surfaces for the stationary state simulation (i.e., for solution of the Poisson
equation) [Chen and Bagci, 2020c, Chen et al., 2020a]. Here, finer meshes are
76
1.69
0.5
y (µm)
–0.5
–1.5
–4.5
–4.5 –3.5
–3.5 –2.5
–2.5 –1.5
–1.5 –0.5
–0.5 0.5
0.5 1.5
1.5 2.5
2.5 3.5
3.5 4.5
4.5
x x(µm)
(µm)
Figure 1.27: Cross section of the plasmonic PCD. Reproduced with permission
from [Chen and Bagci, 2020b].
used near and inside the nanostructures to correctly resolve the geometry and
the exponential decay of the plasmonic fields. In this region, the mesh has
a minimum edge length of 3 nm and the maximum edge length of the mesh
is allowed to reach 70 nm in the GaAs layers and 200 nm in the rest of the
computation domain. The corresponding time-step size used for the time in-
Figs. 1.28 (a) and (b) show Hzt (r, t) and nte (r, t) at several time instants,
respectively. For comparison, Fig. 1.28 uses the same color scale as Fig. 1.25
Fig. 1.28 shows that the transient electromagnetic fields on the plasmonic
Fig. 1.28 is much higher and shows an inhomogeneous pattern. As time march-
ing goes on, electrons drift toward the anode. Figs. 1.29 (a) and (b) compare
77
t = 0.05 ps t = 0.25 ps
t = 0.5 ps t = 1 ps
(a)
t=0.05 ps
t=0.25 ps
t=0.5 ps
t=0.75 ps
t=1 ps
t=1.25 ps
t=1.5 ps
t=1.75 ps
(b)
Figure 1.28: (a) |Hz (r, t)| computed using the DG scheme at different time
instants. The dotted line shows the aperture of optical EM wave excitation.
The gray box indicates the LT-GaAs region. (b) ne (r, t) computed using the
DG scheme at different time instants. Reproduced with permission from [Chen
and Bagci, 2020b].
device and the conventional one (previous example) and their spectrum ob-
tained using Fourier transform. The figures clearly show that the inclusion of
78
500
Conventional PCA
Plasmonic PCA
400 Plasmonic PCA (/7.315)
300
200
100
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (ps)
(a)
500
Amplitude (a.u.)
1
400
0.5
300
0
0 1 2 3 4 5
200 Conventional PCA Frequency (THz)
Plasmonic PCA
0
0 0.5 1 1.5 2 2.5
Time (ps)
(b)
Figure 1.29: (a) The time signature and (b) the spectrum of the THz current
generated on the plasmonic PCD and on the conventional PCD. Reproduced
with permission from [Chen and Bagci, 2020b].
by almost 7 times.
work. However, due to the high computational cost, only the central gap region
scheme models a biased photoconductive device within a unit cell of the nanos-
79
tructure and thus greatly reduces the computational cost. Another challenge
optical and THz waves differs by almost 400 times. Correspondingly, the THz
antenna is typically 400 times larger than the size of the nanostructure. This
dual-mesh scheme [Chen and Bagci, 2020a], where the optical and the THz
waves are treated separately in two DG solvers and their two-way couplings
are taken into account using high-order interpolations. With this dual-mesh
can be strictly modeled in the THz antenna simulation. Thus, the impedance
be strictly analyzed.
Bibliography
ATLAS user’s manual: device simulation software. Silvaco, Int. St. Clara, CA,
2016.
2017.
80
2019.
Luis Angulo, Jesus Alvarez, Mario Pantoja, Salvador Garcia, and A Bretones.
namics: State of the Art. Forum Electromagn. Res. Methods Appl. Technol.,
10, 2015.
Hakan Bagci, Ali E. Yilmaz, Jian Ming Jin, and Eric Michielssen. Fast and
may 2007.
2017.
Jean Pierre Berenger. A perfectly matched layer for the absorption of electro-
81
In Proc. 10th Int. Conf. Comput. methods Appl. Sci. Eng. {Computing}
33(4):748–759, 2016.
Liang Chen and Hakan Bagci. A Dual-mesh Framework for Multiphysics Sim-
Sci. Symp. Int. Union Radio Sci. URSI GASS 2020, aug 2020a.
2020c.
Liang Chen, Ming Dong, and Hakan Bagci. Modeling floating potential con-
2020a.
Liang Chen, Mehmet Burak Ozakin, Shehab Ahmed, and Hakan Bagci.
82
Smoothly-varying Coefficients in Discontinuous Galerkin Time-Domain
Liang Chen, Kostyantyn Sirenko, Ping Li, and Hakan Bagci. Efficient dis-
Pai Yen Chen, Christos Argyropoulos, and Andrea Alu. Terahertz antenna
Shun Lien Chuang. Physics of photonic devices, volume 80. John Wiley and
Sons, 2012.
Ming Dong, Ping Li, and Hakan Bagci. An Explicit Time Domain Finite
Ming Dong, Liang Chen, Lijun Jiang, Ping Li, and Hakan Bagci. An explicit
time domain finite element boundary integral method for analysis of electro-
magnetic scattering. IEEE Trans. Antennas Propag., pages 1–1, 2022. doi:
10.1109/TAP.2022.3142319.
Stylianos Dosopoulos and Jin Fa Lee. Interconnect and lumped elements mod-
83
Stylianos Dosopoulos and Jin Fa Lee. Interior penalty discontinuous galerkin
finite element method for the time-dependent first order maxwell’s equations.
Loula Fezoui, Stéphane Lanteri, Stéphanie Lohrengel, and Serge Piperno. Con-
for the truncation of FDTD lattices. IEEE Trans. Antennas Propag., 44(12):
1630–1639, 1996.
84
tion for the complex-frequency shifted PML. IEEE Trans. Antennas Propag.,
Stephen D. Gedney, Chong Luo, Bryan Guernsey, J. Alan Roden, Robert Craw-
ford, and Jeffrey A. Miller. The Discontinuous Galerkin Finite Element Time
Guernsey, Jeffrey A Miller, Tyler Kramer, and Eric W Lucas. The discon-
1059, 1999.
85
George W. Hanson. Dyadic Green’s functions and guided surface waves for a
2008.
Michael Harmon, Irene M Gamba, and Kui Ren. Numerical algorithms based
NY, 2008.
sciencedirect.com/science/article/pii/S0021999117305363.
96–121, 2006.
86
D. Jiao, A. A. Ergin, B. Shanker, E. Michielssen, and J. M. Jin. A fast higher-
J.M. Jin. The Finite Element Method in Electromagnetics. John Wiley and
87
Ping Li and Li Jun Jiang. Integration of arbitrary lumped multiport circuit
thz range by DGTD with a scalar RBC and an ADE. IEEE Trans. Antennas
Ping Li, Yifei Shi, Li Jun Jiang, and Hakan Bagci. A hybrid time-domain
Ping Li, Li Jun Jiang, and Hakan Bagci. A Resistive Boundary Condition
Ping Li, Li Jun Jiang, and Hakan Bagci. Discontinuous Galerkin Time-Domain
Jian Ming Jin. A DGTD algorithm with dynamic h-adaptation and lo-
Propag. Soc. Int. Symp. APSURSI 2016 - Proc., pages 2079–2080, oct 2016.
Hai Lin, Mario F. Pantoja, Luis D. Angulo, Jesus Alvarez, Rafael G. Martin,
88
conjugate dispersion material model. IEEE Microw. Wirel. Components
ous Galerkin finite element method for highly accurate solution of maxwell
YunXian Liu and Chi-Wang Shu. Analysis of the local discontinuous Galerkin
490–500, 2014b.
89
imation of the Time-Domain Electromagnetic-Field Equations. IEEE Trans.
W.H. Reed and T.R. Hill. Triangular mesh methods for the neutron transport
10.1109/JLT.2017.2760913.
90
Frank Schwierz. Graphene transistors. Nat. Nanotechnol., 5(7):487–496, may
2010.
methods. In Lect. Notes Comput. Sci. Eng., volume 114, pages 369–397.
91
Approaches to Numer. Partial Differ. Equations, pages 371–399. Springer
Chi Wang Shu and Stanley Osher. Efficient implementation of essentially non-
Qingtao Sun, Qiwei Zhan, Qiang Ren, and Qing Huo Liu. Wave Equation-
Masahiko Tani, Shuji Matsuura, Kiyomi Sakai, and Shin-ichi Nakashima. Emis-
92
John A Trangenstein. Numerical solution of elliptic and parabolic partial dif-
Press, 2013.
Haijin Wang, Chi-Wang Shu, and Qiang Zhang. Stability and error estimates of
2015.
Xiang Hua Wang, Wen Yan Yin, and Zhizhang Chen. Matrix exponential
Ali E. Yilmaz, Jian Ming Jin, and Eric Michielssen. Time domain adaptive in-
tegral method for surface integral equations. IEEE Trans. Antennas Propag.,
93