c 2012 Tianjian Lu

SIGNAL INTEGRITY ANALYSIS OF HIGH-SPEED MULTILAYER

INTERCONNECTS USING THE FINITE ELEMENT METHOD

BY

TIANJIAN LU

THESIS

Submitted in partial fulfillment of the requirements

for the degree of Master of Science in Electrical and Computer Engineering
in the Graduate College of the
University of Illinois at Urbana-Champaign, 2012

Urbana, Illinois

Adviser:

Professor Jianming Jin

 ABSTRACT

Approaches of modeling high-speed interconnects generally fall into two

categories: circuit models based on circuit-parameter extractions and full-
wave methods based on solving Maxwell’s equations. Even though computa-
tionally efficient, the approaches based on circuit-parameter extractions fail
to provide accurate predictions due to the increased coupling and radiation
effects at high frequencies. This work targets the fast and efficient full-wave
analysis of high-speed multilayer interconnects. To enhance the efficiency of
the full-wave analysis, fast frequency sweep techniques and domain decom-
position schemes are first investigated and then implemented. This work
consists of four major parts. (1) The full-wave analysis is implemented based
on the finite element method. The capability of handling arbitrary geome-
tries of the finite element method is utilized to deal with complex circuit
structures. (2) The efficiency of the full-wave method is enhanced for broad-
band characterization by the incorporation of the solution space projection.
(3) Various domain decomposition schemes are employed to break an ori-
ginally large problem into smaller ones and allow the possibility of parallel
computing. (4) The solution space projection and the domain decomposi-
tion schemes are combined in a mutually beneficial manner which achieves
an even more efficient full-wave analysis.

ii
To my parents and Lin, for their endless love and support.

iii
 ACKNOWLEDGMENTS

I am very grateful to my advisor, Professor Jianming Jin, for his profound

knowledge and expertise in leading me into this field and his constant moti-
vation and inspiration not only toward the completion of this work but also
in my life as a researcher.
I would also like to thank all my colleagues for their help and the fun they
brought into my research life.
Last but not least, I would like to thank my parents and Ms. Lin Feng for
their endless love and support over the years as I pursued my dream.

iv
 TABLE OF CONTENTS

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

CHAPTER 1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . 1

CHAPTER 2 THREE-DIMENSIONAL FINITE ELEMENT FOR-

MULATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 The Boundary Value Problem . . . . . . . . . . . . . . . . . . 4
2.2 Absorbing Boundary Condition . . . . . . . . . . . . . . . . . 5
2.3 Wave Port . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Basis Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 11

CHAPTER 3 SOLUTION SPACE PROJECTION AND FAST

FREQUENCY SWEEP . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1 Solution Space Projection . . . . . . . . . . . . . . . . . . . . 16
3.2 Adaptive Multi-Point Scheme . . . . . . . . . . . . . . . . . . 17
3.3 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . 18

CHAPTER 4 DOMAIN DECOMPOSITION OF THE MULTI-

LAYER INTERCONNECTS . . . . . . . . . . . . . . . . . . . . . . 31
4.1 Domain Decomposition by Layers: The Approximate Modal
Interface Method . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.2 The Approximate Modal Interface Method and the Solu-
tion Space Projection . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Domain Decomposition within Layers: The Finite Element
Tearing and Interconnecting . . . . . . . . . . . . . . . . . . . 41
4.4 The Finite Element Tearing and Interconnecting and the
Solution Space Projection . . . . . . . . . . . . . . . . . . . . 44
4.5 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . 44

CHAPTER 5 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . 55

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

v
 LIST OF FIGURES

2.1 Nodal basis functions on triangular element . . . . . . . . . . 12

2.2 Edge basis functions on a triangular and tetrahedral element . 14

3.1 Coaxial transmission line . . . . . . . . . . . . . . . . . . . . . 19

3.2 S-parameters of lossy coaxial cable . . . . . . . . . . . . . . . 20
3.3 S-parameters of lossy coaxial cable . . . . . . . . . . . . . . . 21
3.4 S-parameters of lossy coaxial cable . . . . . . . . . . . . . . . 21
3.5 S-parameters of lossy coaxial cable . . . . . . . . . . . . . . . 22
3.6 Top view and front view . . . . . . . . . . . . . . . . . . . . . 23
3.7 S-parameters of a metal-loaded H-plane waveguide T-junction 23
3.8 Microstrip low-pass filter . . . . . . . . . . . . . . . . . . . . 24
3.9 Insertion loss of a microstrip low-pass filter . . . . . . . . . . . 25
3.10 Return loss of a microstrip low-pass filter . . . . . . . . . . . . 25
3.11 Microstrip branch line coupler . . . . . . . . . . . . . . . . . 26
3.12 S-parameters of microstrip branch line coupler . . . . . . . . 27
3.13 Single via structure with two dielectric layers . . . . . . . . . . 28
3.14 S-parameters of single via structure with θ = 180◦ . . . . . . 29
3.15 S-parameters of single via structure with θ = 90◦ . . . . . . . 29
3.16 Single via structure with three dielectric layers . . . . . . . . 30
3.17 S-parameters of single via structure with three dielectric
layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4.1 Cut-through view of a multilayer structure . . . . . . . . . . . 33

4.2 S11 and S12 of the through-hole via structure . . . . . . . . . . 45
4.3 S11 and S12 of the through-hole via structure . . . . . . . . . . 45
4.4 Top view of a three-layered interconnect with coupled sig-
nal traces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.5 S-parameters of the three-layered interconnect with cou-
pled signal traces by direct FEM and AMI . . . . . . . . . . . 47
4.6 S-parameters of the three-layered interconnect with cou-
pled signal traces by direct FEM and FETI-DPEM . . . . . . 47
4.7 Three-layered interconnect with stitching vias . . . . . . . . . 48
4.8 S-parameter of the three-layered interconnect with stitch-
ing vias (separation of 120 mil) . . . . . . . . . . . . . . . . . 50

vi
4.9 S-parameter of the three-layered interconnect with stitch-
ing vias (separation of 240 mil) . . . . . . . . . . . . . . . . . 51
4.10 Three-layered interconnect with stitching vias . . . . . . . . . 52
4.11 S-parameter of the three-layered interconnect without stitch-
ing vias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

vii
 CHAPTER 1

INTRODUCTION

A digital integrated circuit (IC) consists of two parts: transistors and inter-
connects. With the increased integration and design complexities and the
continuous demand for higher operating frequency and sharper rising time in
the very large-scale integration (VLSI) technology, the interconnect effects
are no longer negligible [1–3]. From a microscopic point of view, the digi-
tal signal can be considered as electromagnetic waves propagating among
transistors through interconnects. When the gate delay that transistors take
to process the signals keeps being reduced, the interconnect delay that the
signal spends on traveling along the interconnects tends to dominate the
overall performance of an integrated circuit. Interconnects are also believed
to be responsible for many of the signal integrity issues and electromagnetic
interference problems which are the potential causes of malfunctionings of
the assembled circuits and the failures of the distorted signals in meeting
with design requirements. Due to the long period and high expense in the
practical design cycles, it is hence extremely necessary for the designers to be
aware of the interconnect effects through modeling and simulation techniques
in the very early design stages.
Interconnects fall into four categories [4]: on-chip interconnects provide
the on-chip metallization and IC packaging; chip-to-chip interconnects con-
nect pins or pads among chips and/or with other components, and the most
common chip-to-chip interconnect is printed circuit boards (PCBs); module-
to-module interconnects, as the highest level interconnects inside a digital
system, provide connections among subsystem modules, and one of the com-
mon examples on the module-to-module interconnects is the backplane or
motherboard; system-to-system interconnects are responsible for sharing in-
formation at system level and over long distances, such as coaxial cables and
optical fibers. Among all levels of the aforementioned interconnects, multi-
layer interconnects, targeting higher packaging density, shorter signal delay,

1
and smaller chip/board dimension, become the imperative trend in the de-
sign of high-speed systems. However, it is a great challenge to efficiently
and accurately model the high-speed interconnects in a multilayer structure.
This challenge is attributed to the large problem size, the three-dimensional
geometrical complexities, nonuniform material properties in the multilayer
structure, and high aspect ratios of traces and vias dimensions with respect
to the chip/board dimensions [5].
Current approaches of modeling interconnects can be categorized as circuit
models and full-wave models. The circuit models utilizing the circuit para-
meters extracted through quasi-static assumptions are generally computa-
tionally efficient. However, the circuit models are accurate only within a cer-
tain range of operation frequency beyond which the quasi-static assumption
is no longer valid. In that case, in order to retain the accuracy in the inter-
connect modeling we have to consider the fields components in all the direc-
tions as in the full-wave models. Generally, the full-wave approaches for cha-
racterizing multilayer interconnects include the finite element method [6], [7],
the finite difference method [8], [9], and the moment method [10], [11]. Com-
pared with the circuit models based on the circuit parameter extraction, the
full-wave models can be very computational intensive primarily due to the
spatial discretization.
In this work, we choose the finite element method as the foundation for
the full-wave analysis of multilayer interconnects. Owing to the capability of
dealing with arbitrary geometries, the finite element method has earned its
popularity in handling complicated circuit structures. In order to enhance
the efficiency of the full-wave analysis, two major algorithms have been in-
vestigated and incorporated into the finite element analysis, namely, the so-
lution space projection [12] and the domain decomposition schemes [13], [14].
Generally, broadband information is mandatory for the circuit analysis and
a frequency sweep needs to be performed. In other words, the linear sys-
tem of a large dimension resulting from the volumetric discretization has to
be factorized and solved repeatedly for many frequency samples. It can be
imagined that such a frequency sweep will make the full-wave analysis ex-
tremely computationally intensive and sometimes it becomes an impossible
task to tackle. The fast frequency sweep, particularly the method of solution
space projection [12] in this work, will enhance the efficiency of the full-wave
analysis over a broadband. The details of the method of solution space pro-

2
jection will be provided in later chapters and it is worth mentioning that
by employing the solution vectors at a few selected frequency samples and
forming a reduced-order model to accurately represent the original problem,
the frequency sweep can be extremely efficient.
Another approach available to enhance the efficiency of the full-wave ana-
lysis is domain decomposition. In practical application, a problem can be
really large, and sometimes doing the mesh itself would be an impossible
task, not to mention solving the problem. By applying domain decomposi-
tion, instead of tacking the entire problem as a whole, one can divide it into
smaller pieces. The domain decomposition schemes also allow the possibility
of parallel computing which will further enhance the efficiency. Two domain
decomposition schemes are investigated in this work: one is the approxi-
mate modal interface (AMI) method [13], devised for the multilayer structure
and utilizing modal field expansion to perform decomposition and recoupling
among the layers; the other approach, named the finite element tearing and
interconnecting method [14], can be considered as a two-way decomposition,
not only by layers as AMI does but also within each individual layer. There-
fore, the finite element tearing and interconnecting method is more flexible
in terms of decomposing the computational domains.
This work is organized as follows: in Chapter 2 the full-wave analysis based
on the finite element method is formulated to solve circuit problems. The
applications of the absorbing boundary condition and the port boundary con-
dition are also presented in Chapter 2. Chapter 3 elaborates on the method
of solution space projection and validates the implementation with various
numerical examples. Chapter 4 focuses on the two aforementioned domain
decomposition schemes, namely, the approximate modal interface method
and the finite element tearing and interconnecting method. In Chapter 4,
the domain decomposition schemes and the fast frequency sweep technique
are combined in a mutually beneficial manner for a further enhancement of
the full-wave analysis. Numerical examples are also provided in Chapter 4 to
validate the implementation and demonstrate the efficiency for the domain
decomposition schemes. The entire work concludes in Chapter 5.

3
 CHAPTER 2

THREE-DIMENSIONAL FINITE
ELEMENT FORMULATION

2.1 The Boundary Value Problem

The finite element method is applied to solve the vector wave equation
 
1
∇× ∇×E − k0 2 r E = −jk0 Z0 J (2.1)
µr

where r and µr are the relative permittivity and permeability, respectively;

Z0 is the intrinsic impedance of free space; J is the density of the excitation
current source; and k0 is the frequency at which the source is excited.
In order to solve for E, which satisfies the vector wave equation in Equa-
tion (2.1), one has to apply boundary conditions. The boundary conditions
frequently used in the electromagnetic simulations include the homogeneous
Dirichlet boundary condition

n̂ × E = 0 (2.2)

and the boundary condition of the third kind

1
n̂ × (∇ × E) + γe n̂ × (n̂ × E) = U (2.3)
µr

where n̂ is a normal unit vector to the boundary surface and points outward
from the surface; γe is a constant; and U represents the source that exists
on the boundary surface.
When γe and U are both zero, the boundary condition of the third kind
reduces to the Neumann boundary condition as

n̂ × (∇ × E) = 0. (2.4)

4
The boundary condition of the third kind also includes the absorbing boun-
dary condition (ABC) and the port boundary condition (PBC). The Dirichlet
boundary condition in Equation (2.2) represents the perfect electric conduc-
tors (PECs) and are commonly used to model conductors and side walls
in the circuit structures. The absorbing boundary conditions are typically
employed to truncate the computational domain into a finite one. The port
boundary conditions are attached on port surfaces as excitations to the circuit
structures. It is worth mentioning that when the ports are not excited they
function as the absorbing boundaries to truncate the computational domain.
The details of the absorbing boundary condition and the port boundary con-
dition are provided in the later sections.

2.2 Absorbing Boundary Condition

The finite element analysis can only be applied to a finite computational
domain whereas in most circuit problems one deals with unbounded or open
structures. Therefore, some artificial boundary conditions are required to
truncate an infinite computational domain. The artificial boundary condi-
tions should at the same time be able to simulate the extension of the com-
putational domain into infinity. One such boundary condition is absorbing
boundary condition. It establishes relations among local unknowns attached
on the side walls of the circuit structures and lying on the outer surface
of the computational domain. Because of the localized relation introduced
by absorbing boundary condition, the sparsity of the finite element system
matrices is preserved.

2.2.1 Absorbing Boundary Condition

The simplest absorbing boundary condition is the first-order condition, which
is adopted here. Neglecting terms of order r13 and higher, the scattered field
E sca in a scattering analysis can be written as
r
1 r
n̂ × (∇ × E sca ) + jk0 n̂ × (n̂ × E sca ) = 0. (2.5)
µr µr

5
Note that the incident field E inc , the scattered field E sca , and the total field
E have the relation
E = E inc + E sca . (2.6)

With Equation (2.6), Equation (2.5) can be re-written in terms of the total
field E as r
1 r
n̂ × (∇ × E) + jk0 n̂ × (n̂ × E) = U inc (2.7)
µr µr
where U inc is defined by
r
inc 1 r
U = n̂ × (∇ × E inc ) + jk0 n̂ × (n̂ × E inc ). (2.8)
µr µr

As the circuit excitation is locally constructed and there is no incident field

from side walls, U inc is zero. The absorbing boundary condition becomes
r
1 r
n̂ × (∇ × E) + jk0 n̂ × (n̂ × E) = 0. (2.9)
µr µr

2.2.2 Finite Element Formulation

Applying the Galerkin method to Equation (2.1), the residual R is
Z    Z
1 2
R= W · ∇× ∇×E dv − k0 r W · E dv
V µr V
Z
+jk0 Z0 W · J dv, (2.10)
V

where W is the testing function and V denotes the volume of the computa-
tional domain.
With the vector identity

∇ · (A × B) = B · (∇ × A) − A · (∇ × B), (2.11)

the first term on the right side of Equation (2.10) becomes

Z    Z   
1 1
W · ∇× ∇×E dv = ∇· ∇×E ×W dv
V µr V µr
Z
1
+ (∇ × W ) · (∇ × E) dv.
V µr
(2.12)

6
With the divergence theorem, the first term on the right-hand side of Equa-
tion (2.12) can be written as
Z    Z   
1 1
∇· ∇×E ×W dv = ∇ × E × W · n̂ds
V µr S µr
Z   
1
= W · n̂ × ∇ × E ds.
S µr
(2.13)

With the above derivations, the residual in Equation (2.10) can be written
as
Z
1
R= (∇ × W ) · (∇ × E) dv
V µrZ

−k0 2 r W · E dv
V
Z   
1
+ W · n̂ × ∇ × E ds
S µr
Z
+jk0 Z0 W · J dv, (2.14)
V

where S denotes the surface boundary enclosing the volume. If there exists
an absorbing boundary condition, one can further reduce Equation (2.14).
On the absorbing boundary Sabc , the surface integral becomes
Z    Z r
1 r
W · n̂ × ∇ × E ds = jk0 (n̂ × W ) · (n̂ × E)ds.
S µr Sabc µr
(2.15)

The residual in Equation (2.10) can be finalized as

Z
1
R= (∇ × W ) · (∇ × E) dv
V µrZ

−k0 2 r W · E dv
V
Z r
r
+jk0 (n̂ × W ) · (n̂ × E)ds
Sabc µr
Z
+jk0 Z0 W · J dv. (2.16)
V

Let the residual be zero and assemble all the elemental matrices, we will

7
arrive at Equation (2.17) the matrix equation

(K 0 − k0 K 1 − k0 2 K 2 )x = Kx = b (2.17)

where the elemental matrices are defined as

Z
1
K e0 = e
(∇ × W e ) · (∇ × W e ) dv (2.18)
V e µr
Z r e
e r
K 1 = −j (n̂ × W e ) · (n̂ × W e )ds (2.19)
e
Sabc µer
Z
e
K2 = er W e · W e dv (2.20)
V e
Z
e
b = −jk0 Z0 W e · J e dv. (2.21)
Ve

2.3 Wave Port

The full-wave analysis here requires waveguide port boundary conditions
attached to the port surfaces to launch an incident wave into the circuit struc-
tures and at the same time absorb the reflected wave. Absorbing boundary
condition based on the modal expansion has been proposed for such kind of
port boundary. As in [15] the modal expansion on the port surface has been
successfully incorporated into the finite element analysis.

2.3.1 Port Boundary Condition

Before deriving the port boundary condition, let us derive the relation:

→
−
Z
κm = h tm × →
−
e tm · n̂dΓ
Γ Z
j
= (γm →
−
e tm · →
−
e tm + →
−
e tm · ∇t ezm )dΓ (2.22)
ωµ Γ
→
−
where κm is a constant; h tm and →−
e tm are the transverse components of the
modal magnetic and electric fields on a port surface, respectively; γm is the
propagation constant; and ẑ defines a local unit vector that points inward
to the port such that ẑ = −n̂ where n̂ is the unit normal vector on a port

8
surface.
Starting with Maxwell’s equations and applying the following operations

α
∇ = ∇t + ẑ
αz
Em = →−
e tm − ẑezm
→−
H m = − h tm + ẑhzm , (2.23)

we have

α
∇ × E m = (∇t + ẑ ) × (→ −e tm − ẑezm )
αz
= ∇t × →
−
e tm |ẑ − ∇t × (ẑezm )|ẑ⊥ + γm ẑ × →
−
e tm |ẑ⊥ , (2.24)

where the subscript ẑ ⊥ denotes the direction that is perpendicular to ẑ.

Therefore,
→
− j 
−∇t × (ẑezm ) + γm ẑ × →
− 
h tm = e tm . (2.25)
ωµ
Because

∇t × (ẑezm ) = ezm (∇t × ẑ) − ẑ × ∇t ezm

= −ẑ × ∇t ezm , (2.26)

we have
→
− j
h tm = ẑ × (∇t ezm + γm →
−
e tm ). (2.27)
ωµ
Now, let us go back to the target equation

→
− j 
( h tm × →
− −n̂ × ∇t ezm + γm → −
e tm × →

 −
e tm ) · n̂ = e tm · n̂
ωµ
j −
=− → e tm · n̂ × n̂ × (∇t ezm + γm → −
 
e tm )
ωµ
j →
= (−e tm · ∇t ezm + γm →
−
e tm · →
−e tm ). (2.28)
ωµ

If the port is excited by the incident field E inc on the surface, there will
be reflected field E sca that can be expressed as a linear combination of the
modal fields. Thus the total field E is in the form of
∞
cn (→
−
X
E = E inc + e tn − ẑezn )eγn z . (2.29)
n=1

9
Employing the orthogonality condition for the modal fields,

→
−
Z
h tm × →
−
e tm · n̂dΓ = 0 m 6= n, (2.30)
Γ

we can solve for the modal coefficient cm as

1 −γm z → −
Z
cm = e h tm × (E − E inc ) · n̂dΓ, (2.31)
κm Γ

where κm is defined in Equation (2.22).

Substituting the expression of modal coefficient cm into Equation (2.29),
taking the curl on both sides, and crossing it with n̂, we have

n̂ × ∇ × E + P (E) = U wp , (2.32)

where
∞ Z
1
(γn →
− E · (γn →
−
X
P (E) = e tn + ∇ezn ) e tn + ∇ezn )dΓ (2.33)
n=1
jωµκn Γ
wp inc inc
U = n̂ × ∇ × E + P (E ). (2.34)

If we only have the excitation from the dominant mode, then

E inc = (→
−
e t1 + ẑez1 )e−γ1 z . (2.35)

The wave port excitation vector U wp is reduced to

U wp = −2(γ1 →
−
e t1 + ∇t ez1 ). (2.36)

2.3.2 Finite Element Formulation

Let us get back to Equation (2.14). If there is no additional source other
than the wave port excitation, the term associated with
h the current source
i
J in Equation (2.14) disappears. Substituting the term µ1r n̂ × (∇ × E) by

10
the port boundary condition yields the new residual expression
Z
1
R= (∇ × W ) · (∇ × E) dv
V µrZ

−k0 2 r W · E dv
V
Z r
r
+jk0 (n̂ × W ) · (n̂ × E)ds
Sabc µr
N port Z
X 1
− W · Pn (E)ds
n=1 Spbc
µ r
N port Z
X 1
+ W · U wp
n ds. (2.37)
n=1 Spbc µr

Let the residual be zero and assemble all the elemental matrices, we will
arrive at the matrix equation similar to Equation (2.17)

Nport Nport
" #
X X
2
K 0 − k0 K 1 − k0 K 2 − B n x = Kx = bwp
n (2.38)
n=1 n=1

where
Z
1
Bn = W · Pn (E)ds
Spbc µr
Z
1
bwp
n =− W · U wp
n ds. (2.39)
Spbc µr

2.4 Basis Functions

The finite element method is formulated via Galerkin’s method where the
weighting functions are chosen to be the same as the basis functions. It is
believed that using the same set of functions as basis and weighting functions
to enforce the zero weighted residual will usually lead to the most accurate
solution.
Two types of basis functions are employed in this work, namely, nodal basis
functions and edge basis functions. Due to the scalar nature of the nodal
basis functions, they are usually used to approximate physical quantities
of scalars, such as the electric potential in electrostatic problems and the
axial component of modal electric field in the full-wave modal analysis on a

11
 Figure 2.1: Nodal basis functions on triangular element

waveguide port. Edge basis functions or vector basis functions are used to
approximate vector physical quantities, such as the transverse components of
a modal electric field within triangular elements on a waveguide port surface,
or the electric field in tetrahedral elements in a discretized volume.

2.4.1 Simplex Coordinate

The two types of basis functions are constructed based on simplex coordi-
nates. Consider the triangular element defined in Figure 2.1, where ∆i defines
the area of the subtriangle formed by point P (x, y) and any two of the three
vertices, for example,
 
1 x y 
1  e e

∆2 = 1 x1 y1  . (2.40)

2  
 1 xe y e 
3 3

With these, we can define the area coordinate as L1 = ∆ ∆e

1
, L2 = ∆ 2
∆e
,
∆3
L3 = ∆e where ∆e is the area of the entire triangular element. (L1 , L2 , L3 )
are also recognized as simplex coordinates. The simplex coordinates (L1 , L2 ,
L3 ) depend on the position of P only.

2.4.2 Node Basis Functions

Linear nodal basis functions are defined on the three vertices of the triangle,
in terms of simplex coordinates, and they can be written as

N1 = L1 , N 2 = L2 , N3 = L3 .

12
It is obvious from the definition that Ni becomes 1 at vertex i and 0 at the
other two vertices.
Another set of nodal basis functions that have been widely used in finite
element analysis is quadratic nodal basis functions. They provide one higher
order approximation than the linear set, and only work on scalar quantities
as well. In terms of simplex coordinates, with the numbering of the basis
functions following the node numbers in Figure 2.1, the quadratic nodal basis
functions can be expressed as follows:

N1 = (2L1 − 1)L1 , N2 = (2L2 − 1)L2 , N3 = (2L3 − 1)L3 ,

N4 = 4L1 L2 , N5 = 4L2 L3 , N6 = 4L3 L1 .

Like linear nodal basis functions, Ni will reach 1 at node i and 0 at all other
nodes. This property automatically ensures the requirement of continuity
along the boundaries of adjacent triangular elements within the formulation
of finite element analysis. To be specific, the scalar electric potential or the
axial component of a modal electric field will remain the same or continuous
across the boundaries of adjacent triangular elements upon the expansion by
the nodal basis functions.
Since only first- and second-order nodal basis functions are listed here, the
interested readers are referred to [16] for the unified derivations of nodal basis
functions of arbitrary orders.

2.4.3 Edge Basis Functions

Edge basis or vector basis functions are employed to approximate vector field
quantities. For example, we use edge basis functions to expand the transverse
components of a modal electric field on a waveguide port surface. For the
three-dimensional full-wave analysis, the edge basis functions are used to
approximate the electric field on tetrahedral elements.
As we can see from Figure 2.2, the edge basis functions are defined along
the edges of a triangular or a tetrahedral element, and hence the continuity
across element boundaries is only enforced on the tangential components.
This allows a discontinuity on the normal components. If we use node basis
functions and assign the same value on the boundaries, we are actually en-

13
 Figure 2.2: Edge basis functions on a triangular and tetrahedral element

forcing the continuity of both tangential and normal components of the field,
which causes problems while dealing with dielectric interfaces. Even though
there are remedies to either avoid or fix those problems, it is very incon-
venient compared to the implementation with edge basis functions. Other
than that, the edge basis functions satisfy the divergence condition in the
source free region and eliminate the spurious modes in the modal analysis
with nodal basis functions.
On a triangular element, the first-order edge basis functions are defined as

W1 = (L1 ∇L2 − L2 ∇L1 )e1 , W2 = (L3 ∇L1 − L1 ∇L3 )e2 ,

W3 = (L3 ∇L1 − L3 ∇L1 )e3 ,

where Li is the simplex coordinate, and ei is the length of edge i. Note that
the tangential component of each edge basis functions is constant along the
edge it sits on and vanishes on all the other edges.
Similarly, the second-order edge basis functions on a triangular element
are given by

W1 = (3L1 − 1)(L1 ∇L2 − L2 ∇L1 )e1 , W2 = (3L2 − 1)(L2 ∇L3 − L1 ∇L3 )e1 ,
W3 = (3L2 − 1)(L2 ∇L3 − L3 ∇L2 )e2 , W4 = (3L3 − 1)(L2 ∇L3 − L3 ∇L2 )e2 ,
W5 = (3L3 − 1)(L3 ∇L1 − L1 ∇L3 )e3 , W6 = (3L1 − 1)(L3 ∇L1 − L1 ∇L3 )e3 ,
9 9
W7 = L3 (L1 ∇L2 − L2 ∇L1 )e1 , W8 = L1 (L2 ∇L3 − L3 ∇L2 )e2 .
2 2

As shown in Figure 2.2, the first-order edge basis functions are defined on
the edges of a tetrahedron. In terms of simplex coordinates, they are written

14
as

W1 = (L1 ∇L2 − L2 ∇L1 )e12 , W2 = (L2 ∇L3 − L3 ∇L2 )e23 ,

W3 = (L3 ∇L1 − L1 ∇L3 )e31 , W4 = (L1 ∇L4 − L4 ∇L1 )e14 ,
W5 = (L4 ∇L2 − L2 ∇L4 )e42 , W6 = (L3 ∇L4 − L4 ∇L3 )e34 ,

where eij is the length of the edge from node i to node j.

The second-order edge basis functions consist of two parts, namely, the set
of basis functions defined on edges and those on the faces. The second-order
edge basis functions are given by

W1 = (3L1 − 1)(L1 ∇L2 − L2 ∇L1 )e12 , W2 = (3L2 − 1)(L1 ∇L2 − L2 ∇L1 )e12
W3 = (3L2 − 1)(L2 ∇L3 − L3 ∇L2 )e23 , W4 = (3L3 − 1)(L2 ∇L3 − L3 ∇L2 )e23
W5 = (3L3 − 1)(L3 ∇L1 − L1 ∇L3 )e31 , W6 = (3L1 − 1)(L3 ∇L1 − L1 ∇L3 )e31
W7 = (3L1 − 1)(L1 ∇L4 − L4 ∇L1 )e14 , W8 = (3L4 − 1)(L1 ∇L4 − L4 ∇L1 )e14
W9 = (3L4 − 1)(L4 ∇L2 − L2 ∇L4 )e42 , W10 = (3L2 − 1)(L4 ∇L2 − L2 ∇L4 )e42
W11 = (3L3 − 1)(L3 ∇L4 − L4 ∇L3 )e34 , W12 = (3L4 − 1)(L3 ∇L4 − L4 ∇L3 )e34
W13 = 29 L3 (L1 ∇L2 − L2 ∇L1 )e12 , W14 = 92 L1 (L2 ∇L3 − L3 ∇L2 )e23
W15 = 29 L4 (L1 ∇L2 − L2 ∇L1 )e12 , W16 = 92 L2 (L1 ∇L4 − L4 ∇L1 )e14
W17 = 29 L1 (L3 ∇L4 − L4 ∇L3 )e34 , W18 = 92 L3 (L1 ∇L4 − L4 ∇L1 )e14
W19 = 29 L2 (L3 ∇L4 − L4 ∇L3 )e34 , W20 = 92 L4 (L2 ∇L3 − L3 ∇L2 )e23 .

15
 CHAPTER 3

SOLUTION SPACE PROJECTION AND

FAST FREQUENCY SWEEP

3.1 Solution Space Projection

A volumetric discretization in the finite element analysis leads to a large-
dimensional system of equations. Even though the sparsity of the resulting
linear system can be utilized by the ever-progressing direct or iterative sparse
matrix solvers, it is still considered computational intensive especially when
the broadband information is desired. Take the deterministic problem in
Equation (2.38) for example. The space discretization results in the following
matrix equation

K(k)x(k) = b(k) (3.1)

where k is the wave number. The matrix K, the excitation vector b, and
the solution vector x are all frequency-dependent. Therefore, a broadband
analysis may require solving the system of equations at every single frequency
sampling point, which is likely to result in a huge computation burden.
The solution space projection (SSP) method [12] to be introduced is de-
vised for a fast and robust broadband finite element analysis. It is to con-
struct a new model of the linear system in Equation (3.1) yet with a much
smaller dimension; at the same time, the newly developed reduced-order
model is capable of accurately describing the behaviors of the original linear
system. With the reduced-order model, the broadband analysis requires sol-
ving a linear system of a much smaller dimension at the frequency samples,
which tremendously reduces the computational burden.
The solution space projection possesses an adaptive multi-point nature.
As the name solution space projection conveys, the original finite element
model described by Equation (3.1) is projected onto a subspace spanned by
a set of solution vectors. It is a set of frequency sampling points at which the

16
the original problem are solved and that reflects the nature of the multi-point
scheme of SSP. The number and locations of the frequency samples are critical
in constructing the reduced order model. An inadequate number of frequency
samples will lead to an inaccurate prediction of the original problem while
redundancy will downgrade the efficiency. In order to construct the reduced-
order model in an accurate and efficient way, SSP bears an adaptive scheme
such that the number of samples can be minimized.

3.2 Adaptive Multi-Point Scheme

As mentioned at the beginning of the chapter, if the solution over a given
band is sought, the analysis must be repeated for many frequency sampling
points in the range of interest, resulting in a high computational burden.
Besides, due to the double integration, the port boundary information forms a
dense matrix. It may lead to the degradation of the sparsity of the original K
matrix and an undesired large condition number of the final matrix. Further,
for lossless medium in the circuit we usually expect a real system whereas the
existence of wave ports will stuff complex entries into the system and force
us to store the entire system in terms of complex numbers which actually
doubles the storage. Fortunately, all those mentioned above can be avoided
by the solution space projection method to be introduced in the following.
Assume that the dimension of the linear system in Equation (3.1) is N .
The adaptive scheme starts with solving Equation (3.1) at the two ending
frequency points kstart and kend , and the obtained solution vectors are column-
wise arranged to form X. A set of basis vectors H is then obtained though
the orthonormalization of the solution vectors X. The reduced-order model
is formed by projecting the original linear system onto the subspace spanned
by H as
H H [K(k)] Hy = H H b(k) (3.2)

where the superscript H represents the transpose of the complex conjugate.

It is worth mentioning that the recovery of the solution vector x(k) of the
original problem can simply be done by

x(k) = Hy(k). (3.3)

17
 The adaptive scheme requires the calculation of relative error r at each
step before expanding the set of solution vectors X and consequently the
set of basis vectors H. The residual is calculated at the middle point kmid
of interval [kstart kend ]. At kmid , we calculate the solution of the reduced
order model y(kmid ) and project it back to the original problem denoted as
Hy(kmid ). The relative error is defined as
 
b(kmid ) − [K (k)]Hy (kmid )
r= . (3.4)
|b(kmid )|

The adaptive process terminates if the relative error is below the specified
tolerance, otherwise the original problem is solved at kmid and the obtained
solution vector is employed to expand the set of H. After that, the relative
errors at (kmin + kmid )/2 and (kmid + kmax )/2 are checked. The adaptive
process will terminate only if both of the relative errors at (kmin + kmid )/2
and (kmid +kmax )/2 are below the preset tolerance. The solution vector at the
frequency point of the larger relative error is used to expand the set of H.
The same procedure repeats until the relative errors at the middle points of
all the existing intervals drop below the preset tolerance or the process hits
the maximum number of iterations.
Suppose at the end the solutions at a total number of M sampling points
are employed to construct the reduced-order model, when the frequency
sweep is performed over the band of interest, a linear system of dimension
M × M instead of N × N will be solved at each sampling point, where M is
much smaller than N .

3.3 Numerical Examples

In this section, numerical examples will be provided as verifications of the
three-dimensional FEM formulation with various boundary conditions and
demonstrations of the robustness and efficiency of the method of solution
space projection.

18
 Figure 3.1: Coaxial transmission line

3.3.1 Conventional Waveguide Structures

Traditional metallic waveguide structures are common in microwave and RF
devices. Even though the work targets at multilayer interconnects, examples
of conventional waveguide structures are borrowed to verify the incorporation
of solution space projection into the finite element analysis. Two waveguide
examples, namely, coaxial cable and waveguide T-junction with discontinuity,
are provided in the following.

3.3.1.1 Coaxial Cable

Consider a lossy coaxial cable of 1 m length shown in Figure 3.1. The inner
and outer radii of the coaxial cable are 0.4 mm and 1 mm, respectively. The
relative permittivity of the dielectric between the two conducting walls is
chosen to be  = 1 − 0.2j. S-parameters are calculated over the frequency
range from 10 MHz to 1 GHz at 80 discrete sampling points. Different boun-
dary conditions are applied at the two side surfaces in order to validate the
implementation.
In Figure 3.2 and Figure 3.3, a PMC or PEC surface attached to the end of
the coaxial port acts as open-circuit load or short-circuit load, respectively.
The results obtained from full-wave analysis is compared to the analytical
solutions, where the comparison verifies the correctness of the full-wave ana-
lysis implementation. However, it takes only five frequency points to build
the reduced-order model, by which the solution vectors at all other sampling
points are expanded. Therefore, instead of solving at every single frequency
point out of 80, the solution space projection only employs five of them. It
saves a tremendous amount of time and enables the fast frequency sweep.

19
 1
FEMSSP
0.8 Analytical Solution

|S11|
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
Frequency (GHz)
Phase(S11) (degree)

0
−10 FEMSSP
Analytical Solution
−20
−30
−40
−50

0 0.2 0.4 0.6 0.8 1

Frequency (GHz)
Figure 3.2: S-parameters of lossy coaxial cable
with wave port and open-circuit load

Figure 3.4 and Figure 3.5 are taken as the validations of the implemen-
tations of an absorbing boundary condition (ABC) and a port boundary
condition. As we know, ABC acts as transparent when wave penetrates on
it, and it serves as a matched load for the coaxial cable. Therefore, we see no
reflections at the incident end. As the frequency goes higher, it is observed
that the portion of the transmitted wave reduces since the lossy transmission
line consumes more power at higher frequencies.

3.3.1.2 Waveguide T-Junction

The second example is an H-plane waveguide T-junction with a metal block

loaded inside. The detailed dimensional information can be found in Figure
3.6 and also in [17]. The frequency band of interest ranges over from 8 GHz
to 12 GHz with 200 samples. With the method of solution space projection,
the original problem is solved at five adaptively selected frequency samples
with the predetermined tolerance of 0.001. The obtained results with solution
space projection are compared with those in [15] and the two are identical

20
 1
FEMSSP
0.8
Analytical Solution

|S11|
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
Frequency (GHz)
Phase(S11) (degree)

10 FEMSSP
Analytical Solution

−10

−30

0 0.2 0.4 0.6 0.8 1

Frequency (GHz)
Figure 3.3: S-parameters of lossy coaxial cable
with wave port and short-circuit load

−4
x 10
4.5

3.5

3
|S11|

2.5

1.5

0.5
0 0.2 0.4 0.6 0.8 1
Frequency (GHz)

Figure 3.4: S-parameters of lossy coaxial cable

with wave port and absorbing boundary condition

21
 1
|S11|
|S21|
0.8

0.6
|S|
0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
Frequency (GHz)

Figure 3.5: S-parameters of lossy coaxial cable

with wave ports at both ends

as in Figure 3.7.

3.3.2 Planar Circuit Components

The bulky metallic waveguide structures are gradually replaced by the planar
structures with the development of integrated circuit technology. Comparing
to conventional waveguide structures, planar structures are compact and low
cost. It also allows the possibility to integrate several planar components to-
gether with other active devices such as diodes and transistors, all of which
are held on a single substrate, and the entire system can be mass-produced
in a rather cost-effective manner by photolithography. In this section several
examples of planar circuit components are provided. It is worth mentioning
that the full-wave analysis of these planar structures are more challenging
than for the conventional waveguide structures due to the existence of inho-
mogeneity and open structures.

3.3.2.1 Microstrip Low-Pass Filter

In this example, we calculate the scattering parameters of a low-pass filter

using the solution space projection method. As in Figure 3.8, the detailed

22
 Figure 3.6: Top view and front view
of a metal-loaded H-plane waveguide T-junction

−2

−4

−6
|Sij| (dB)

−8
S11−−FEMSSP
−10
S12−−FEMSSP
−12 S13−−FEMSSP
S11−−FEM
−14 S12−−FEM
S13−−FEM
−16
8 8.5 9 9.5 10 10.5 11 11.5 12
Frequency (GHz)

Figure 3.7: S-parameters of a metal-loaded H-plane waveguide T-junction

23
 Figure 3.8: Microstrip low-pass filter

geometry information are w = 2.413 mm, b = 2.54 mm, a1 = 5.65 mm, a2 =

12.257 mm, the substrate height is 0.794 mm, and the relative permittivity
of the substrate is r = 2.2. The desired frequency range is from 1 GHz
to 20 GHz with 400 samples. The solution space projection is used with
the tolerance for the adaptive procedure set to be 0.005. The reduced-order
model is built with solutions at 11 adaptively selected frequency sampling
points.
The results with both the FDTD method and the measurement are avai-
lable in [18], with which a comparison is performed. Figures 3.9 and 3.10
show the results of comparison of insertion loss and return loss, respectively.
The results with the solution space projection agree well with those from
the FDTD method and measurement in the desired frequency range. It is
worth mentioning that the measurement performed in [18] employed network
analyzer and associated connectors that are rated only to 18 GHz, which
explains the minor discrepancies occur at high frequencies.

3.3.2.2 Microstrip Branch Line Coupler

A microstrip branch coupler is shown in Figure 3.11. The four microstrip

lines that branch out from the square have width w = 2.413 mm, each of
which is taken as a port. Also in Figure 3.11, a = 9.75 mm and b = 3.96

24
 0

|S21| (dB) −10

−20

−30 FEMSSP
MEASUREMENT
FDTD
−40

0.5 2.5 4.5 6.5 8.5 10.5 12.5 14.5 16.5 18.5 20.5
Frequency (GHz)
Figure 3.9: Insertion loss of a microstrip low-pass filter

−5

−10

−15
|S11| (dB)

−20

−25 FEMSSP
MEASUREMENT
−30 FDTD

−35

−40
0.5 2.5 4.5 6.5 8.5 10.5 12.5 14.5 16.5 18.5 20
Frequency (GHz)

Figure 3.10: Return loss of a microstrip low-pass filter

25
 Figure 3.11: Microstrip branch line coupler

mm. The substrate has a height of 0.794 mm and permittivity r = 2.2.

To complete the broadband analysis from 100 MHz to 10 GHz, a number
of 200 frequency samples are chosen over a linear scale. With the solution
space projection method, only 9 out of 200 frequency samples are employed
to build the reduced-order model. The predetermined tolerance of such an
adaptive procedure is 0.002. The same structure has been examined with
the FDTD method in [18]. It is shown in Figure 3.12 that the results here
match well with those from the FDTD method.

3.3.3 Multilayer Board Interconnect

Interconnects find their existences on various levels such as printed circuit
boards, packaging, and on-chip. Common interconnect structures include
transitions, wire bonds, meander lines, and via structures. The trend of
high-speed applications in the very large-scale integrated (VLSI) circuits has
been highlighting the role of interconnects. As in a high-speed interconnect,
the time that signals spend on traveling through the interconnect cannot
be neglected. Generally, the interconnect is also responsible for signal in-
tegrity issues such as signal delays, distortions, reflections, and crosstalk.
Therefore, it is essential for the designers to be aware of interconnect effects
through modeling and simulation. However, the existing circuit simulators
such as SPICE that depends on circuit parameter extractions cannot meet
the requirement of accuracy especially as the frequency of operation goes
higher. In this section, a few examples on the full-wave analysis of multi-

26
 0

−5

|S| (dB) −10

−15 S11−−FEMSSP
S11−−FDTD
−20 S21−−FEMSSP
S21−−FDTD
S31−−FEMSSP
−25 S31−−FDTD
S41−−FEMSSP
−30 S41−−FDTD

−35
0 2 4 6 8 10
Frequency (GHz)

Figure 3.12: S-parameters of microstrip branch line coupler

layer interconnects are provided. Together with the method of solution space
projection, the finite element analysis makes us closer to our goal of simu-
lating interconnect structures as efficiently as possible while retaining high
accuracy.

3.3.3.1 Single Via through Two Board Layers

The first via structure is a single via going through two dielectric layers which
are separated by one ground plane as shown in Figure 3.13. The red shaded
area in Figure 3.13(a) is considered as the ground plane that separates the
dielectric layers. It is also shown that the clearance hole and the metal pad
have the same diameter denoted as c. The via rod has a diameter denoted as
d. The top and bottom dielectric layers share the same height h. In Figure
3.13(b), the microstrip with solid color is the strip attached onto the top of
the via structure and the one enclosed by dash lines are the one attached
onto the bottom, where θ is the angle formed by the two microstrips.
Figure 3.14 shows the calculated S-parameters of the single via structure
and the comparison with the result from measurement [19]. The angle be-
tween the top and bottom microstrips is θ = 180◦ . The diameter of the

27
 (a) Side view (b) Top view

Figure 3.13: Single via structure with two dielectric layers

clearance hole (or via pad) is c = 3.9 mm. The diameter of the via rod is
d = 1.5 mn. The width of the microstrip is w = 3.3 mm. The substrate
has a height h = 1.6 mm and permittivity r = 3.4. The broadband analysis
targets at the frequency range from 1 GHz to 25 GHz with 400 samples and
only 14 of them are used to construct the reduced-order model.
Figure 3.15 is the calculated S-parameters of the same structure with θ =
◦
90 . The obtained results with solution space projection are also compared
with the measurement [19]. Everything remains the same except that the
diameter of the via rod is changed to d = 0.7 mm. The broadband analysis
also considers 400 frequency sampling points from 1 GHz to 25 GHz and
again only 14 of them are used to construct the reduced-order model.

3.3.3.2 Single Via through Three Board Layers

The last example of via structures is shown in Figure 3.16. The top and
bottom microstrips are sitting on the same side. The detailed information
of geometry is depicted in Figure 3.16. The frequency range of interest is
from 100 MHz to 15 GHz with 200 samples. The reduced-order model is
constructed with solutions at five frequency samples. Figure 3.17 shows the
results and a comparison with the finite element analysis without the solution
space projection. It can be seen that results from the two ways of simulation
match pretty well.

28
 0

−5

−10
|Sij| (dB)

−15
S11 −− FEMSSP
S11 −− MEASUREMENT
−20
S21 −− FEMSSP
S21 −− MEASUREMENT
−25

−30
0 5 10 15 20 25
Frequency (GHz)

Figure 3.14: S-parameters of single via structure with θ = 180◦

−10
|Sij| (dB)

−20

−30 S11 −− FEMSSP

S11 −− MEASUREMENT
S21 −− FEMSSP
S21 −− MEASUREMENT
−40
0 2 4 6 8 10 12 14 16 18 20 22 24
Frequency (GHz)

Figure 3.15: S-parameters of single via structure with θ = 90◦

29
 Figure 3.16: Single via structure with three dielectric layers

−5

−10
|Sij| (dB)

−15
S11 −− FEMSSP
S11 −− FEM
−20 S21 −− FEMSSP
S21 −− FEM

−25

−30
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Frequency (GHz)

Figure 3.17: S-parameters of single via structure with three dielectric layers

30
 CHAPTER 4

DOMAIN DECOMPOSITION OF THE

MULTILAYER INTERCONNECTS

The modeling of the interconnects falls into two major categories, lumped
models and distributed models. The selection between the lumped and dis-
tributed models depends on the operating frequency and the dimension of
the interconnect itself. At relatively low frequencies, the interconnects can
be modeled by lumped RC and RLC circuit models. As the frequency goes
higher when the electrical length of the interconnect becomes significant com-
paring to the wavelength, the lumped circuit models are inadequate and
transmission line models have to be adopted. Two major assumptions are
made while assuming the transmission line models. One is the assumption
on the transverse electromagnetic mode (TEM) where both electric field and
magnetic field are perpendicular to the direction of propagation. This is valid
only if the dimension of the line cross-section is much smaller than the wave-
length. The other assumption is that the current is uniformly distributed
over the cross-sectional area. However, as the frequency increases, the di-
mension of the cross-sectional area is comparable to or even larger than the
wavelength; at the same time, the current distribution is no longer uniform
over the cross-sectional area and the skin, edge, and proximity effects have
to be taken into account. Besides, the inhomogeneity in most of the prac-
tical interconnect structures will also fail the TEM assumption by bringing
in the E and H along the direction of propagation. All these aforementioned
phenomena will degrade the accuracy of the transmission line models.
Even though people are attempting to modify the transmission line models
by introducing distance-dependant and frequency-dependant R, L, C, G per
unit length parameters to account for high-frequency effects, fields along
the direction of the propagation cannot be neglected on a further increase
of frequency. Consequently, the full-wave model, which takes into account
all the spatial components of the fields, is inevitably the best alternative
in modeling the interconnect structures. However, the full-wave analysis

31
can be considerably computational expensive. In this chapter, we proceed
with the search for possibilities of fast and efficient full-wave analysis for
the multilayer interconnects. We concentrate our work in this chapter on the
methods of domain decomposition. The basic ideas of domain decomposition
is to break the original problem into smaller ones and allow the possibility
of parallel computing. Two schemes of domain decomposition are provided
in this section, namely, the approximate modal interface (AMI) method [13]
and the finite element tearing and interconnecting (FETI) [14] method. The
approximate modal interface method is devised for the multilayer structures.
It postulates the boundary conditions on the interfaces among layers and
separates the original problem into smaller ones with each associated with
one individual layer. The finite element tearing and interconnecting method
is more general and flexible in terms of domain decomposition and it can be
performed further within layers.

4.1 Domain Decomposition by Layers: The

Approximate Modal Interface Method
The approximate modal interface method [13] to be described here is devised
for multilayer interconnects. In a multilayer interconnect structure, layers are
separated by metallic planes. Assume perfect electric conductors (PECs) for
the planes. The tangential component of electric field is zero everywhere on
those PEC planes. If we can postulate the tangential components of electric
fields at the apertures (via-holes), a complete set of boundary conditions
for that particular layer can be obtained and the layer can thus be treated
independently from the rest. If we apply this rationale to all the layers,
information of every single layer can be extracted to form an independent
problem. All these individual problems shall be re-coupled by the apertures
shared by the neighboring layers.
Consider the multilayer structure shown in Figure 4.1. For simplicity,
let us temporarily ignore the terms associated with the absorbing boundary

32
 Figure 4.1: Cut-through view of a multilayer structure

condition. By Galerkin’s method, the weighted residual can be expressed as

Z Z
1
R= (∇ × W ) · (∇ × E) dv − k0 2 r W · E dv
V µr V
Z   
1
+ W · n̂ × ∇ × E ds (4.1)
S µr

where S denotes the boundary surface. Since the absorbing boundary condi-
tion is temporarily ignored, S consists of wave port surfaces and the via-hole
interfaces between two adjacent layers.
Assume that H i is the magnetic field at aperture Si , i = 1, 2, 3, 4 in Figure
4.1, the electric field at the aperture can be obtained via Faraday’s law as

1
n̂ × ( ∇ × E i ) = −jk0 Z0 n̂ × H i (4.2)
µr

where n̂ is the normal vector to Si and points outward from Si as shown

in Figure 4.1. Treating Equation (4.2) as the boundary condition at the
apertures and substituting it into Equation (4.1), one obtains
Z Z
1
R= (∇ × W ) · (∇ × E) dv − k0 2 r W · E dv
V µ r V
Ns Z
X
−jk0 Z0 W · (n̂ × H i )ds. (4.3)
i=1 Si

At aperture Si , the magnetic field can be written as a linear combination

of modal magnetic fields weighted by their corresponding modal coefficients

33
denoted as ci,k (k stands for the k th mode)

Nmode
X
Hi = ci,k H i,k . (4.4)
k=1

Again, by Faraday’s law,

ci,k
H i,k = ∇ × (→
−
e ti,k + ẑezi,k ) (4.5)
−jk0 Z0 µr

where →−
e ti and ẑezi represent the modal electric fields of the tangential direc-
tion and longitudinal direction, respectively. Substituting Equation (4.4) and
Equation (4.5) into Equation (4.3), the third term of the weighted residual
in Equation (4.3) becomes

Ns Z
X
−jk0 Z0 W · (n̂ × H i )ds
i=1 Si
Ns Nmode Z
1
W · n̂ × ∇ × (→
−
X X  
= ci,k e ti,k + ẑezi,k ) ds
i=1 k=1 Si µr
Ns Nmode Z
X X 1 h α →
−
i
= ci,k W · n̂ × (∇t + ẑ ) × ( e ti,k + ẑezi,k ) ds
i=1 k=1 Si µr αz
Ns Nmode Z
1 α−
e ti,k + ∇t × (ẑezi,k ) + ẑ × ( →
W · n̂ × (∇t × →
−
X X h i
= ci,k e ti,k )) ds
i=1 k=1 Si µr αz
Ns Nmode Z
1
W · n̂ × (∇t × →
−
e ti,k − ẑ × ∇t ezi,k − ẑ × γk,i →
−
X X  
= ci,k e ti,k ) ds
i=1 k=1 Si µr
X Nmode
Ns Z
1
W · ∇t ezi,k + γk,i →
−
X  
=− ci,k e ti,k ds.
i=1 k=1 Si µr

To conclude the preceding derivation,

Ns Z Ns Nmode Z
1
W ·(∇t ezi,k +γk,i →
−
X X X
−jk0 Z0 W ·(n̂×H i )ds = − ci,k e ti,k )ds.
i=1 Si i=1 k=1 Si µr
(4.6)
Note that since n̂ and ẑ have opposite signs on all the apertures except those
on the upper interface, a negative sign will appear in the right-hand side of
Equation (4.6) except for the apertures on the upper interface. It is worth

34
mentioning that with Equation (4.6) all the formulations are now in terms
of the electric field.
Enforcing the weighted residual to be zero and applying the finite element
discretization, one obtains the system of equations
    
K S1 S1 0 K S1 V1 E S1 bS1
 0 K S3 S3 K S3 V1  E S3  = bS3  , (4.7)
    

K V1 S1 K V1 S3 K V1 V1 E V1 0

where E S1 , E S3 , and E V1 are the field coefficients of aperture S1 , S3 , and

volume V1 , respectively. The right-hand side excitation vector bSi is given by

Nmode Z
1
W · (∇t ezi,k + γk,i →
−
X
bSi = ci,k e ti,k )ds. (4.8)
k=1 Si µr

Rewrite bSi into the matrix form

 
ci,1
 
 ci,2 
 
i .. 
h  . 
bSi = ai,1 ai,2 . . . ai,k . . . ai,Nmode 
  = A i ci . (4.9)
 ci,k 

 .. 

 . 

ci,Nnmode

With the matrix form of bSi , the matrix form of Equation (4.7) can also be
written as
   −1  
E S1 K S1 S1 0 K S1 V1 A1 0 " #
 c1
E S3  =  0 K S3 S3 K S3 V1   0 A3  . (4.10)
    
c3
E V1 K V1 S1 K V1 S3 K V1 V1 0 0

Similarly, for the rest two layers, we have

   −1  
E S3 K S3 S3 0 K S3 V2 −A3 0 " #
 c3
E S4  =  0 K S4 S4 K S4 V2   0 A4  (4.11)
    
c4
E V2 K V2 S3 K V2 S4 K V2 V4 0 0

35
    −1  
E S2 K S2 S2 0 K S2 V3 A2 0 " #
 c2
E S4  =  0 K S4 S4 K S4 V3   0 −A4  . (4.12)
    
c4
E V3 K V3 S2 K V3 S4 K V3 V3 0 0

By solving the linear system with the modal field excitations and extracting
the field related to the apertures, one obtains
" # " #" #
E S1 x111 x112 c1
= (4.13)
E S3 x121 x122 c3

" # " #" #

E S3 x211 x212 c3
= (4.14)
E S4 x221 x222 c4

" # " #" #

E S2 x311 x312 c2
= , (4.15)
E S4 x321 x322 c4

where in xki,j , i, j denote the local aperture numbers and k denotes the layer
number.
To determine the four groups of the modal field coefficient for the four
apertures in Figure 4.1, we need four equations. Two neighboring layers share
the same aperture on their interface where the fields have to be identical.
Therefore, we have two equations in the following

x121 c1 + (x122 − x211 )c3 − x212 c4 = 0 (4.16)

x321 c2 − x221 c3 + (−x222 + x322 )c4 = 0. (4.17)

The other two equations come from the port boundary condition

n̂ × ∇ × E + P (E) = U inc . (4.18)

By applying Galerkin’s method and the finite element discretization to the

port boundary condition, we have two equations associated with port 1 and

36
port 2, respectively

−A1 c1 + P 1 E S1 = uS1 (4.19)

−A2 c2 + P 2 E S2 = uS2 . (4.20)

Enforcing the equality of the electric field at ports, we have

(P 1 x111 − A1 )c1 + P 1 x112 c3 = uS1 (4.21)

(P 2 x311 − A2 )c2 + P 2 x312 c4 = uS2 . (4.22)

Till now, we have all the four equations as

(P 1 x111 − A1 )c1 + P 1 x112 c3 = uS1 (4.23)

x121 c1 + (x122 − x211 )c3 − x212 c4 = 0 (4.24)
(P 2 x311 − A2 )c2 + P 2 x312 c4 = uS2 (4.25)
x321 c2 − x221 c3 + (−x222 + x322 )c4 = 0. (4.26)

In the matrix form, the set of Equations (4.23-4.26) is

(P M − A) c = u, (4.27)

37
where
 
P1 0 0 0
 
 0 P 2 0 0
P = 
0
 (4.28)
 0 I 0 
0 0 0 I
 
1 1
x11 0 x12 0
3
x312
 
 0 x11 0 
M = 
x1 1 2 2
 (4.29)
 21 0 x 22 − x 11 −x 12


3 2 2 3
0 x21 −x21 −x22 + x22
 
A1 0 0 0
 
 0 A2 0 0 
A= 
0
 (4.30)
 0 0 0 

0 0 0 0
h iT
c= c1 c2 c3 c4 (4.31)
h iT
u= uS1 uS2 0 0 . (4.32)

H
Multiplying both sides of Equation (4.27) by à in the form of
 
A1 0 0 0
 
 0 A2 0 0
à = 
  (4.33)
 0 0 A 3 0 

0 0 0 A3

results in a square system of modal coefficients

H H
à (P M − A) c = à u. (4.34)

All the modal coefficients can be solved through Equation (4.34), and all
of the aperture fields can be obtained via the interface systems of Equation
(4.13), Equation (4.14), and Equation (4.15).

38
4.2 The Approximate Modal Interface Method and the
Solution Space Projection
It is observed that the most time-consuming part is the process of building
the modal coefficient matrix in Equation (4.34) rather than solving it. In
building Equation (4.34) upon the quest of a broadband analysis, one needs
to solve the linear system of equations associated with each layer at many
frequency sampling points. Moreover, a single mode is generally insufficient
to expand the fields on a aperture, and in order to maintain a certain accu-
racy, it is necessary to utilize multiple modes. Therefore, it requires matrix
factorization for every single layer at every frequency sampling point, and
at the same time solving for multiple right-hand sides ai,k (excitation of k th
mode at aperture i). In this section, we attempt to generate reduced-order
models for individual layers such that the process of constructing the modal
coefficient matrices can be accelerated over a broadband.
The method of solution space projection introduced in Chapter 3 is excita-
tion-dependent, which introduces more than one ways to build the reduced-
order models. One of the approaches is to build the reduced-order model
based on each single-mode excitation at one aperture. For example, there
are apertures S1 and S3 in the top layer as shown in Figure 4.1, and suppose
that the field at each aperture is expanded by three modes. The total number
of six excitations will create six reduced-order models for a single layer. This
process will become cumbersome when we have thousands of via-holes and
we have to deal with thousands of reduced-order models simultaneously. It
seems not the best choice of associating one reduced-order model with a
single-mode excitation. Another approach is to build a single reduced-order
model for each layer by lumping all the solution vectors from all the modal
excitations in that layer. It takes fewer sampling frequency points to build
this kind of reduced-order model since the latter contains more information
per sampling frequency. However, this turns out to be inefficient since it
takes a considerable amount of time and memory in orthonormalizing the
solution vectors to obtain a set of orthonormal basis. After filtering out the
two extremes cases, it ends up with a balanced option. All the via-holes
sitting on the same interface will have one reduced-order model; ports in the
same layer will contribute to the same reduced-order model. In this case,
we will have maximum three reduced-order models per layer, namely, one

39
associated with excitations on the upper interface, one with excitations on
the lower interface, and one with the wave port expiations.
The following part will illustrate the process of the adaptive multi-point
solution space projection scheme under the approximate modal interface
method. Again, taking the top layer of Figure 4.1 as an example and re-
writing Equation 4.10 into a more compact form, we have
   −1  
E S1 K S1 S3 0 K S1 V1 A1 0 " #
 c1
E S3  =  0 K S3 S3 K S3 V1   0 A3 
    
c3
E V1 K V1 S1 K V1 S3 K V1 V1 0 0
" #
h i c
1
= K −1 a1 a3
c3
" #
h i c
1
= x1 x3 . (4.35)
c3

The top layer in Figure 4.10 does not have apertures on the upper interface,
and we will have only two reduced-order models for that layer. There will be
one reduced-order model associated with all the ports in the top layer, to be
specific, the port surface S1 , and another reduced-order model with apertures
sitting on the lower interface, namely, S3 . Certainly for the cases with more
apertures in the layer, the extra apertures will contribute to either the port
reduced-order model or the lower interface reduced-order model accordingly.
The reduced-order model associated with S1 is built upon adaptively by
solving Kx1 = a1 . After a number of sampling frequencies the residual
drops below the preset tolerance and we can arrange the solution vectors of
different frequencies column-wise as
h i
X = x1 (k1 ) x1 (k2 ) . . . x1 (kN s ) . (4.36)

Orthonormalizing X to H and projecting the original problem onto the space

spanned by the column vectors of H, one obtains

H H K(k)Hy = H H a1 (k), (4.37)

and the solution of the original problem can be recovered by x(k) = Hy.
It is worth mentioning that the strategy of checking convergence is slightly
different from that in Chapter 3. Since it is the unknowns on the aperture

40
that are to be used in constructing the modal coefficient matrix, only those
entries need to be checked by the residual defined as

|a1 (k, DOFa ) − K(k, DOFa )x(k, DOFa )|

r= , (4.38)
|a1 (k, DOFa )|

where DOFa denotes the unknowns associated with the apertures.

4.3 Domain Decomposition within Layers: The Finite

Element Tearing and Interconnecting
As the decomposition of domains is performed over layers and through the
expansion and re-coupling of the modal fields of the apertures on the layer
interfaces, the approximate modal interface method is seemingly predestined
as a one-way domain decomposition. Even though the layers can be tho-
roughly decoupled, the approximate modal interface methods find limitations
in a few ways. Generally the size of a single layer in most of the practical
problems can still be considerably large, and factorizing the linear system
of huge dimension may require a tremendous amount of time and memory
resource. Other than that, most of times we encounter unbalanced domains
since the top and bottom layers are usually open structures truncated with
absorbing boundaries, and the thicknesses of the multiple dielectric layers
and metal planes are nonuniform. The existence of unbalanced layers is
likely to degrade the parallel computing efficiency. Therefore, in this section,
we seek the possibility of two-way domain decomposition, which will not
only decompose the computational domain by layers but also within each
layer. The two-way domain decomposition with its generality and flexibility
seems a good remedy to account for the problems appearing in the one-
way domain decomposition. In this section, we use the method of the finite
element tearing and interconnecting to further cut each layer into a number
of subdomains.

4.3.1 Finite Element Tearing and Interconnecting

The finite element tearing and interconnecting method has been applied to
solve large-scale three-dimensional electromagnetic problems in [14]. Fol-

41
lowing the convention in [14], the finite element tearing and interconnecting
will be denoted as FETI-DPEM. By applying the FETI-DPEM method, one
single layer is torn into Ns non-overlapping subdomains. The neighboring
subdomains Ωi and Ωj share the interface denoted as Γij . To enforce the
tangential continuity of the electric field at the interfaces, the Dirichlet-type
transmission condition

n̂i × n̂i × E i = n̂j × n̂j × E j (4.39)

and the Neumann-type transmission condition

1 1
n̂i × i
∇ × E i = −n̂j × j ∇ × E j = Λ (4.40)
µ µ

are enforced simultaneously. In Equation (4.39) and Equation (4.40), n̂i is

the unit normal vector at the interface Γij pointing from subdomain i to the
exterior region, and Λ is the Neumann boundary data, which is unknown.
The finite element discretization is applied to individual subdomains, and
the discretization of the Neumann-type transmission condition results in the
term Z
s T
(B ) λ = N s · ΛdS, (4.41)
Γs
s
where B is a signed boolean matrix that extracts the interface DOFs of the
sth subdomain. Also, λ is the dual variable, called the Lagrange multiplier.
By using the Lagrange multiplier, the continuity is enforced at the sub-
domain interfaces. Taking account of other boundary conditions, the linear
system of individual subdomain becomes

K s + (B s )T λ = f s , (4.42)

where (B s )T λ includes the information of the subdomain interfaces of the

sth subdomain, K s includes the volumetric discretization information of the
sth subdomain other than the subdomain interfaces, and f s contains the
information of the excitation in the sth subdomain. Similarly we can have
the system equations for all the subdomains and rearrange them into a global
system as follows

42
  1  1  
K ... 0 (B 1 )T E f1
 . ... .. ..   .   . 
 .. . .   .   . 
   .  =  . . (4.43)
 0 . . . K Ns (B Ns )T  E Ns  f Ns 
    

B1 . . . B Ns 0 λ 0

It is worth mentioning that the last equation in Equation (4.43) is the

contribution from the Dirichlet-type boundary condition. By eliminating
E s , s = 1, 2, ..., Ns , we arrive at an interface system of λ

F λ = d, (4.44)

where
Ns
X
F = B s (K s )−1 (B s )T (4.45)
s=1
XNs
d= B s (K s )−1 f s . (4.46)
s=1

The interface system in Equation (4.44) can be solved using Krylov sub-
space methods such as the generalized minimum residual (GMRES) method
and biconjugate gradient stabilized (BiCGSTAB) method together with the
Dirichlet preconditioner [20]. The Dirichlet preconditioner offers a good
approximation of the inverse of F in Equation (4.44) as

Ns
" #
X 0 0
F −1 = Bs s
(B s )T , (4.47)
s=1
0 S II
where

S II = K sII − K sIV (K sV V )−1 (K sIV )T (4.48)

and the subscripts V and I denote the volume and interface unknowns,
respectively.
The original problem is reduced to an interface problem of a much smaller
dimension. After iteratively solving the interface problem for the dual varia-
bles, the field unknowns can be extracted with the obtained Neumann-type

43
boundary condition for individual subdomains.

4.4 The Finite Element Tearing and Interconnecting

and the Solution Space Projection
The incorporation of the solution space projection into the FETI-DPEM
method is straightforward. Since the original problem is solved interme-
diately through the interface problem, the field coefficient can be recovered
through the known dual variables and all the other given boundary condi-
tions. The field solutions from all the subdomains can be lumped into one
global solution vector. With the global solution vector, one can refer to the
steps in Chapter 2 to build the reduced-order models. Therefore, the only
difference of the FETI-DPEM and SSP from the FEM and SSP is the inter-
mediate step of recovering the field coefficients from the subdomain interface
problem.

4.5 Numerical Examples

4.5.1 Through-Hole Via Structure

The first example is the same as that shown in Figure 3.16, which is a
through-hole via structure of three dielectric layers separated by two conduc-
ting planes. Owing to the existence of the two metal planes, the multilayer
interconnect can be divided into three individual subdomains each associa-
ted with one layer. Other than directly using the finite element method, this
section employs the method of approximate modal interface and the method
of finite element tearing and interconnecting. The method of solution space
projection has also been incorporated to obtain the results in Figures 4.2
and 4.3. Figures 4.2 and 4.3 show a good match of S-parameters among
the results obtained by directly using the finite element method, AMI, and
FETI-DPEM. This example validates the implementation of the approximate
modal interface method and the finite element tearing and interconnecting
method.

44
 0

−5

|Sij| (dB) −10

−15 S11 −− AMISSP

S11 −− FEMSSP
S21 −− AMISSP
−20
S21 −− FEMSSP

−25

−30
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Frequency (GHz)

Figure 4.2: S11 and S12 of the through-hole via structure

−5

−10
|Sij| (dB)

−15
S11 −− FETISSP
S11 −− FEMSSP
−20 S21 −− FETISSP
S21 −− FEMSSP
−25

−30
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Frequency (GHz)

Figure 4.3: S11 and S12 of the through-hole via structure

45
Figure 4.4: Top view of a three-layered interconnect with coupled signal
traces

4.5.2 Multilayer Interconnect with Bending Signal Traces

In this example, coupled signal traces come in on the first layer and branch
out at the bottom layer where the planar directions of in- and out-signal
traces are perpendicular to each other. The stack-up of this example is
the same as that shown in Figure 3.16. Figure 4.4 illustrates the detailed
information of the layout: the width of all the microstrip lines is 0.254 mm;
the antipad shares the same diameter with the pad, both of which are 0.508
mm; the diameter of the drilled via hole is 0.254 mm; and the separation of
the two coupled vias is 1.49 mm.
Similar to the previous example, the method of approximate interface
method cuts the entire structure into three layers. In the simulation with the
method of finite element tearing and interconnecting, each layer is further
cut into two subdomains, each of which forms a square of 2 mm × 2 mm.
It is also worth mentioning that the port 1 and port 2 are on the top layer
and port 3 and port 4 are on the bottom layer. The S-parameters of this
three-layered structure is shown in Figures 4.5 and 4.6 with directly using the
finite element method, applying AMI to cut into layers, and two-way domain
decomposition with FETI-DPEM. It can be seen that the results from all the
three method are identical.

46
 0
S11 −− AMISSP
S11 −− FEMSSP
−5
S21 −− AMISSP
S21 −− FEMSSP
−10 S31 −− AMISSP
S31 −− FEMSSP
S41 −− AMISSP
−15
S41 −− FEMSSP
|Sij| (dB)

−20

−25

−30

−35

−40
0 1 2 3 4 5 6 7 8 9 10
Frequency (GHz)

Figure 4.5: S-parameters of the three-layered interconnect with coupled

signal traces by direct FEM and AMI

0
S11 −− FETISSP
S11 −− FEMSSP
−5
S21 −− FETISSP
S21 −− FEMSSP
−10 S31 −− FETISSP
S31 −− FEMSSP
S41 −− FETISSP
−15
S41 −− FEMSSP
|Sij| (dB)

−20

−25

−30

−35

−40
0 1 2 3 4 5 6 7 8 9 10
Frequency (GHz)

Figure 4.6: S-parameters of the three-layered interconnect with coupled

signal traces by direct FEM and FETI-DPEM

47
 (a) top view

(b) side view

Figure 4.7: Three-layered interconnect with stitching vias

4.5.3 Multilayer Interconnect with and without Stitching Vias

The examples in this subsection investigate the via stitching impact on the
signal integrity of the multilayer interconnects. As we know the multiple
power and ground planes in the multilayer interconnet act as microstip-patch
antennas and result in radiations due to the fringing fields at the board
edges. The radiations are essential to some undesirable couplings inside the
structure and most likely to create signal integrity issues. To mitigate this
kind of radiation, stitching the ground planes together at the periphery of
the board using closely spaced vias has been proposed. It has been shown
that the via stitching can effectively reduce the undesirable radiation from
the fringing fields at the edge of the board [21]. As shown in Figure 4.7,
coupled microstrip lines on the top and bottom layers are connected through
vias. The radii of barrel, pad, and antipad are 5 mil, 10 mil, and 15 mil,

48
respectively. As in Figure 4.7(b), the width of the microstrip lines is 5 mil.
The metal layer has a thickness of d = 0.7 mil. The top and bottom dielectric
layers have a thickness of h1 = 3 mil, and the middle dielectric layer has a
thickness of h2 = 20 mil. The relative permittivity of the dielectric layers is
r = 4.2 with a loss tangent of 0.01.

4.5.3.1 Stitching Vias with a Separation of 120 mil

In this case, the distances shown in Figure 4.7(a) are as follows: d1 = 20

mil, d2 = 60 mil, and d3 = 120 mil. The S-parameters of this structure
are shown in Figure 4.8(a) with directly using FEM, Figure 4.8(b) with the
method of AMI, and Figure 4.8(c) with the method of FETI-DPEM. By using
the method of AMI, the entire structure is separated into three layers. With
the method of FETI-DPEM, each layer is further cut into four subdomains.
The frequency band has a range from 0.5 GHz to 40 GHz, and 400 samples
are employed. The results from all the three methods match very well with
each other. With the solution space projection, only seven frequency points
are used to construct the reduced-order model with both direct FEM and
FETI-DPEM.

4.5.3.2 Stitching Vias with a Separation of 240 mil

In this case, the distances shown in Figure 4.7(a) are as follows: d1 = 20

mil, d2 = 120 mil, and d3 = 160 mil. The S-parameters of this structure
are shown in Figure 4.9(a) with directly using FEM, Figure 4.9(b) with the
method of AMI, and Figure 4.9(c) with the method of FETI-DPEM. Similar
to the previous case, with the method of FETI-DPEM, each layer is further
cut into four subdomains. The frequency band has a range from 0.5 GHz
to 40 GHz, and 400 samples are employed. The results from all the three
methods seems identical. With the solution space projection, only seven
frequency points are used to construct the reduced-order model with both
direct FEM and FETIDP.

49
 0

−5
S11−FEMSSP
S21−FEMSSP
−10
S31−FEMSSP
−15 S41−FEMSSP

−20

|Sij| (dB)
−25

−30

−35

−40

−45

−50
0 5 10 15 20 25 30 35 40
Frequency (GHz)

(a) result with FEM and SSP

0

−5 S11−AMISSP
S21−AMISSP
−10
S31−AMISSP
S41−AMISSP
−15

−20
|Sij| (dB)

−25

−30

−35

−40

−45

−50
0 5 10 15 20 25 30 35 40
Frequency (GHz)

(b) result with AMI and SSP

0

−5 S11−FETISSP
S21−FETISSP
−10
S31−FETISSP
−15 S41−FETISSP

−20
|Sij| (dB)

−25

−30

−35

−40

−45

−50
0 5 10 15 20 25 30 35 40
Frequency (GHz)

(c) result with FETI-DPEM and SSP

Figure 4.8: S-parameter of the three-layered interconnect with stitching

vias (separation of 120 mil)

50
 0

−5 S11−FEMSSP
S21−FEMSSP
−10 S31−FEMSSP
S41−FEMSSP
−15

|Sij| (dB)
−20

−25

−30

−35

−40
0 5 10 15 20 25 30 35 40
Frequency (GHz)

(a) result with direct FEM and SSP

0

−5 S11−AMISSP
S21−AMISSP
−10 S31−AMISSP
S41−AMISSP

−15
|Sij| (dB)

−20

−25

−30

−35

−40
0 5 10 15 20 25 30 35 40
Frequency (GHz)

(b) result with AMI and SSP

0

−5 S11−FETISSP
S21−FETISSP
−10 S31−FETISSP
S41−FETISSP
−15
|Sij| (dB)

−20

−25

−30

−35

−40
0 5 10 15 20 25 30 35 40
Frequency (GHz)

(c) result with FETI-DPEM and SSP

Figure 4.9: S-parameter of the three-layered interconnect with stitching

vias (separation of 240 mil)

51
 (a) top view

(b) side view

Figure 4.10: Three-layered interconnect with stitching vias

52
4.5.4 No Stitching Via
The stitching vias are removed in this case. The distances shown in Figure
4.10(a) are d1 = 20 mil and d2 = 60 mil. The S-parameters of this structure
are shown in Figure 4.11(a) with directly using FEM, Figure 4.11(b) with
the method of AMI, and Figure 4.11(c) with the method of FETI-DPEM.
By using the method of AMI, the entire structure is separated into three
layers. With the method of FETI-DPEM, each layer is further cut into two
subdomains. The frequency band has a range from 0.5 GHz to 40 GHz, and
400 samples are employed. The results from all the three methods match
very well with each other. With the solution space projection, only seven
frequency points are used to construct the reduced-order model with both
direct FEM and FETI-DPEM.

53
 0

S11−FEMSSP
−5
S21−FEMSSP
S31−FEMSSP
S41−FEMSSP
−10

|Sij| (dB)
−15

−20

−25
0 5 10 15 20 25 30 35 40
Frequency (GHz)

(a) result with direct FEM and SSP

0

−5
S11−AMISSP
S21−AMISSP
−10 S31−AMISSP
S41−AMISSP
|Sij| (dB)

−15

−20

−25

−30
0 5 10 15 20 25 30 35 40
Frequency (GHz)

(b) result with AMI and SSP

0

S11−FETISSP
−5
S21−FETISSP
S31−FETISSP
S41−FETISSP
−10
|Sij| (dB)

−15

−20

−25
0 5 10 15 20 25 30 35 40
Frequency (GHz)

(c) result with FETI-DPEM and SSP

Figure 4.11: S-parameter of the three-layered interconnect without stitching

vias

54
 CHAPTER 5

CONCLUSION

In this work, a three-dimensional full-wave solver based on the finite element

method is implemented as a tool to perform the signal integrity analysis
of high-speed multilayer interconnects. Wave ports are incorporated into
the full-wave solver to mimic the circuit excitations and truncate an infi-
nite computational domain. The wave port is based on a two-dimensional
eigen-solver at the port surfaces which is responsible of extracting a certain
number of propagation modes and evanescent modes upon request. The
three-dimensional full-wave solver is capable of solving various circuit prob-
lems as demonstrated in the numerical examples.
There are mainly two categories of approaches of modeling the high-speed
multilayer interconnects. Compared to the circuit models based on circuit-
parameter extractions, full-wave methods are generally computational inten-
sive. In order to speed up the design process and maintain certain accuracies,
fast and efficient full-wave solvers are most desirable. Therefore, methods to
enhance the efficiency of the full-wave solvers are investigated and imple-
mented in this work, namely, the method of solution space projection and
the domain decomposition schemes.
The method of solution space projection assists the full-wave solver in
achieving a fast frequency sweep. Over a broadband analysis of the multi-
layer interconnects, the same problem has to be solved repeatedly at many
frequency samples, leading to a huge computation burden. By utilizing the
information at only a few frequency samples and projecting the original prob-
lem into the subspace spanned by the information at few selected frequency
samples, the method of solution space projection produces a reduced-order
model which has a much smaller dimension than the original problem but is
capable of yielding accurate predictions as the original problem does. The
adaptive and multi-point nature of the solution space projection prevents
the redundancy in the information collected to form the reduced-order mo-

55
del. The efficiency of the full-wave solver with the solution space projection
has been demonstrated with various numerical examples.
Domain decomposition schemes uncover another possibility on enhancing
the efficiency of the full-wave analysis. The original problem is broken into
a few smaller ones and this idea allows the possibility of parallel compu-
ting. Two domain decomposition schemes are investigated here, namely, the
method of approximate modal interface and the method of finite element
tearing and interconnecting. The approximate modal interface method is
a one-way domain decomposition scheme and it performs the domain de-
composition by layers, to be specific, one layer forms one subdomain. The
approximate interface method utilizes the modal solutions at the apertures
shared by two adjacent layers to realize the decomposition and re-couple
among the subdomains. The other domain decomposition scheme presen-
ted in this work is the method of finite element tearing and interconnecting.
This method is able to perform domain decomposition even within layers
and is considered as a two-way domain decomposition. By introducing the
Dirichlet-type and Neumann-type boundaries on the subdomain interfaces,
the original problem is first divided into a number of subdomains and then
transformed into a subdomain interface problem of a much smaller dimen-
sion. The solutions of the original problems can be recovered in subdomain
wise after solving the subdomain interface problem. Finally, the proposed
solution space projection and the domain decomposition schemes are com-
bined in a mutually beneficial manner by which we may be able to come up
with an even faster and more efficient full-wave method.

56
 REFERENCES

[1] A. E. Ruhli and A. C. Cagellaris, “Progress in the methodologies for

the electrical modeling of interconnects and electronic packages,” Proc.
IEEE, vol. 89, pp. 740–771, 2001.

[2] R. Achar and M. S. Nakhla, “Simulation of high-speed interconnects,”

Proc. IEEE, vol. 89, pp. 693–728, 2001.

[3] E. P. Li, E. X. Liu, L. W. Li, and M. S. Leong, “A coupled efficient and

systematic full-wave time-domain macromodeling and circuit simulation
method for signal integrity analysis of high-speed interconnects,” IEEE
Trans Advanced Packaging, vol. 27, pp. 213–223, 2004.

[4] H. J. Liaw, “Characterization and modeling of multilayer interconnec-

tions for high speed digital systems,” Ph.D. dissertation, University of
Illinois at Urbana-Champaign, 1996.

[5] Y. Shao, Z. Peng, and J. F. Lee, “Signal integrity analysis of high-speed

interconnects by using nonconformal domain decomposition method,”
IEEE Trans Components, Packaging and Manufactoring Technology,
vol. 2, pp. 122–130, 2012.

[6] J. Wang and R. Mittra, “Finite element analysis of MMIC structures

and electronic packages using absorbing boundary conditions,” IEEE
Trans. Microw. Theory Tech, vol. 42, no. 3, pp. 441–449, 1994.

[7] G. W. Pan, J. Tan, and B. K. Gilbert, “Full wave edge element based
analysis of 3D metal-dielectric structures for high clock rate digital and
microwave applications,” IEEE Proc. Microw. Antennas Propag, vol.
147, no. 5, pp. 391–397, 2000.

[8] W. D. Becker, P. H. Harms, and R. Mittra, “Time-domain electromag-

netic analysis of interconnects in a computer chip package,” IEEE Trans.
Microw. Theory Tech, vol. 40, no. 12, pp. 2155–2163, 1992.

[9] P. Cherry and M. F. Iskander, “FDTD analysis of high frequency

electronic interconnection effects,” IEEE Trans. Microw. Theory Tech,
vol. 43, no. 10, pp. 2445–2451, 1995.

57
[10] M. R. Abdul-Gaffoor, H. K. Smith, A. A. Kishk, and A. W. Glisson,
“Simple and efficient full-wave modeling of electromagnetic coupling
in realistic RF multilayer PCB layouts,” IEEE Trans. Microw. Theory
Tech, vol. 50, no. 6, pp. 1445–1457, 2002.
[11] S. G. Hsu and R. B. Wu, “Full-wave characterization of a through
hole via in multilayered packaging,” IEEE Trans. Microw. Theory Tech,
vol. 43, no. 5, pp. 1073–1081, 1995.
[12] S. H. Lee and J. M. Jin, “Adaptive solution space projection for fast and
robust wideband finite-element simulation of microwave components,”
IEEE Microwave Wireless Component Letter, vol. 17, no. 7, pp. 474–476,
2007.
[13] S. H. Lee, K. Mao, and J. M. Jin, “A complete finite element analysis of
multilayer anisotropic transmission lines from DC to teraherz frequen-
cies,” IEEE Trans. Adv. Packag., vol. 31, no. 2, pp. 326–338, 2008.
[14] Y. Li and J. M. Jin, “A vector dual-primal finite element tearing and in-
terconnecting method for solving 3-D large-scale electromagnetic prob-
lems,” IEEE Trans Antennas Propag, vol. 54, no. 10, pp. 3000–3009,
2006.
[15] J. Liu, J.-M. Jin, E. K. N. Yung, and R. S. Chen, “A fast, higher order
three-dimensional finite-element analysis of microwave waveguide de-
vices,” Microwave Optic. Tech. Lett., vol. 32, no. 5, pp. 344–352, 2005.
[16] J. M. Jin, The Finite Element Method in Electromagnetics. New York:
John Wiley & Sons, 2002.
[17] K. L. Wu and H. Y. Wang, “A rigorous modal analysis of h-plane wave-
guide t-junction loaded with a partial-height post for wideband appli-
cations,” IEEE Trans Microwave Theory Tech, vol. 46, pp. 1721–1734,
1998.
[18] D. M. Sheen, S. M. Ali, M. D. Abouzahra, and J. A. Kong, “Application
of the three-dimensional finite-difference time-domain method to the
analysis of planar microstrip circuits,” IEEE Trans Microwave Theory
Tech, vol. 38, no. 7, pp. 1712–1717, 1990.
[19] S. Maeda, T. Kashiwa, and I. Fukai, “Full wave analysis of propagation
characteristics of a through hole using the finite-difference time-domain
method,” IEEE Trans Microwave Theory Tech, vol. 39, pp. 2154–2159,
1991.
[20] D. Rixen and C. Farhat, “A simple and efficient extension of a class of
substructure based preconditioners to heterogeneous structural mecha-
nics problems,” Int. J. Numer. Methods Eng., vol. 44, pp. 489–516, Feb.
1999.

58
[21] X. Ye, D. M. Hochanson, M. Li, Y. Ren, W. Cui, J. L. Drewniak, and
R. E. EuBroff, “EMI mitigation with multilayer power-bus stachs and
via stitching of reference planes,” IEEE Trans Electromagnetic compa-
tibility, vol. 43, no. 4, pp. 538–548, 2001.

59