Title: Domain Decomposition Methods for Solving Large Electromagnetic Problems
1Domain Decomposition Methods for Solving Large
Electromagnetic Problems
Marinos Vouvakis, Seung-Cheol Lee, Kezhong Zhao
and Jin-Fa Lee Computational Science Group, ESL,
ECE Dept. The Ohio State University http//esl.eng
.ohio-state.edu/csg
Sponsored jointly by Ansoft Corp. Pittsburgh And
Temasek Lab., NUS, Singapore and Northrop Grumman
Corp.
2Vector Finite Element Methods
- Capability of brute-force FEM in a single PC
using
- p-type MUltiplicative Schwarz (pMUS)
preconditioner
1 Jin-Fa Lee, Din Kow Sun, pMUS (p-type
Multiplicative Schwarz) Method with Vector Finite
Elements for Modeling Three-Dimensional Waveguide
Discontinuities, IEEE Trans. Microwave Theory
Tech., March, 2004. 2 D. K. Sun, J. F. Lee,
and Z. J. Cendes, Construction of Nearly
Orthogonal Nedelec Bases for Rapid Convergence
with Multilevel Preconditioned Solvers, SIAM J.
Sci. Comput., vol. 23, no. 4, pp. 1053-1076.
3Dual band operation
7 elements antenna array, circular polarized
The 7 elements array, coated with Radome, sitting
on top of a large finite ground plane
4Smoothed finite ground plane
Smoothed finite ground plane
5Outline
- Non-overlapping Conformal Domain Decomposition
- Conventional Formulation
- Alternative Formulation with Largrange
Multiplier - Duality Pairing
- Non-Overlapping Non-Conformal Domain
Decomposition - Mortar Technique
- 1st Order Transmission Condition Between Domains
- Fourier Analysis
- Convergence Plots for Dirichlet-Neumann mapping
and Robin Transmission Condition - Numerical Results
- Vivaldi Arrays
- Monopole Antenna Arrays
6Implementing ABC in FEM
B.V.P.
1 Traditional Formulation
Imposing the natural boundary condition results in
72 Alternative Formulation Introduce an
additional set of unknowns on the boundary
, and the BVP reads
8Galerkin Formulation of the Alternative BVP
Note the matrix is Symmetric and There is no zero
diagonal block!!!
9DDM with non-matching grids and Mortar Technique
And,
Do these two BVPs imply the needed Physics,
namely ?
YES, though weakly
Galerkin treatments result in
10Non-overlap and non-matching Grids DDMs with
Dual-Primal Unknowns
- Primal Unknowns (FEM variables) Ei, ei
- Dual Unknowns (Mortar variables) Ji
- Need T.C. (Transmission Conditions) between
domains to enforce tangential continuities of
electric and magnetic fields - Proper T.C. leads to CONVERGING domain
decomspotion method
11Usual Dirichlet-Neumann Mapping
NOT WORKING
12Works for Radition and Scattering Problems
1st Order Transmission Condition Robin-Robin
Mapping
13Conforming DDMs
At nth iteration, let us assume that
are available everywhere. Then,
and
- Gauss-Siedel iteration scheme is adopted, and
therefore denote the most
update values of the variables - The two BVPs are well-posed. Existence and
uniqueness are trivially established - For simplicity, ABC is used for mesh truncation
in the formulation. Other mesh truncation methods
can be subsituted into the formulation with
straightforward modifications.
Can be employed to update variables
- Common variable e is used to strongly enforced
tangential component of electric field being
continuous - Tangential component of magnetic field continuous
is imposed weakly through natural boundary
condition in FEM formulation
14Boundary Value Problem for Single Building Block
?back
?left
?top
?bottom
?front
?right
non-conforming triangular meshes
15Non-conforming DDM and Mortar Technique
And,
Do these two BVPs imply the needed Physics,
namely ?
YES, though weakly
Galerkin treatments result in
16Boundary Value Problem for Two Domains (cont.)
17Boundary Value Problem for Two Domains (cont.)
Mortar Elements (non-Galerkin testing)
18DD Viewed as Operator Splitting
19Mortar DDM
where
Back (3)
Back (3)
Left (0)
Right (1)
Domain (i-1, j)
Left (0)
Right (1)
Domain (i1, j)
Front (2)
Front (2)
20Comparisons With brute-force FEBI
7?7 Vivaldi Array
10?10 Vivaldi Array
21Beam Scanning
15x10 vivaldi array with scanning angle ? 45,?
0
?
?
Scaning
? 45,? 0
? 0 plane
broadside
2250?50 Vivaldi Array-Beam Scanning
23Conclusions
- Mortar technique is a powerful machinary to to
implement non-conforming domain decomposition
method - Using Robin transmission condition results in
slower convergence for high frequency Fourier
modes and fast rate of damping for low
frequency Fourier modes. - The MortarDDM can be further improved to model
large electromagnetic problems by employing FETI
algorithm
24Mesh truncation ABC, PML, BIE
Galerkin statement
25Notes
26Mortar Technique for DDM with Non-matching Grids
27Galerkin Finite Element Formulation
28Classical Schwarz Algorithm without Overlap
Classical Schwarz, Dirichlet-to-Neumann Mapping
for 2 Non-overlapping Domains
- Starts with arbitrary initial solutions
- Iterates until the relative changes between
iteration are small ? Converged
Does NOT work.
- Overlap domains It works to damp the evanecent
modes in the error but still not affecting the
propagation modes - Robin conditions It works for over-lapping
domains but the evanecent modes in the error will
not be damped if non-overlapping domains are used
29y
Fourier Analysis
TE modes hz component
N
D
x
Also, the radiation boundary conditions at x?
Take a Fourier transform in the y direction,
namely H(x, ky) components, then
Take into account the radiation B.C. at x?, the
solutions are
30Evaluate H at x0 employing the Dirichlet-Neumann
Mappings, we have
Define the convergence rate for each Fourier
component
Classical Schwarz Dirichlet-Neumann mapping has a
convergence rate of 1 for all evanescent and
propagating modes
31Fourier Analysis of 1st Order Transmission
Condition
Fourier Transform
32(No Transcript)
33Applying Silver-Muller radiation condition in ?1
and ?2
TE
TM
TE and TM Fourier field representations on the
two subdomains
1st order generalized Robin Transmission Condition
34TE
TM
35(No Transcript)
36(No Transcript)
37DDM as an Operator Splitting Preconditioner
38(No Transcript)
39Coarse discretization
Fine discretization
40Periodic Structures
i1M (Many repeated domains)
41Keeping only the variables on interface
42Recovering of the Primal Unknowns
4311?11 Jandayala Patch Array
4411?11 Jandayala Patch Array
E at Near Field (Top)
E at Near Field (Bottom)
Bistatic RCS (xz-plane)
4516?16 Dual Polarized Exponential Tapered
(Vivaldi) Horn Array
0.98 ?
7.89 ?
4616?16 Dual Polarized Exponential Tapered
(Vivaldi) Horn Array
RCS for q0º f0º Incidence
xz-plane RCS
E at Near Zone
4716?16 Dual Polarized Exponential Tapered
(Vivaldi) Horn Array
RCS for q30º f0º Incidence
xz-plane RCS
yz-plane RCS
E at Near Zone
48Complete Flow Chart of ODDMortarFETI Algorithm
Adaptive mesh refinement for every building block
Partitioning the problem domain
FETI using pMUS CG method
Multiple domains glued together through mortar
Construct the domains using the building blocks
Outer loop DDM Iteration using Gauss-Siedel or
Krylov methods
Engineering Info. Antenna Pattern, Impedances, Q
factor, etc
Recover Primal Variables for Each Domain from
Dual-Variable Solution
49DP-FETI vs. Direct DDM
27.23 Million Unknowns
50Comparison of MortarDDM and MortarDDMFETI
Solution time only
51Mortar DDM FETI
52Dual-Primal Finite Element Tearing and
Interconecting like Algorithm (DP-FETI)
Let
Pre-processing
53Large Finite Array
15x15
50x50
5418?18 Dual Polarized Exponential Tapered
(Vivaldi) Horn Array
55PBG Nanocavity
Small PBG laser nanocavity
Few mm at 200THz
567-Layer PBG Nanocavity
5714-Layer PBG Nanocavity
58(No Transcript)
59PBG Optical Channel Waveguidef170 THz
60ODD with Multiple Building Blocks
Partition 1 (9 blocks)
Partition 2 (3 blocks)
61Partitioning
Partition 1 (9 blocks)
Partition 2 (3 blocks)
629?12 Monopole Array
Directivity D022.03 dbi (full FEM-PML) D022.02
dbi (DD-FEM)
Scan direction
63(No Transcript)
64DDM 120 deg 23.46dBi
Azimuthal Gain by DDM
Elevation Gain by DDM
65Rotman Lens for Switch Beam Arrays
66RCS from Large Finite ArraysPatch Antenna Array
67Plane Wave Scattering from a dual-polarized Notch
Antenna Array (16x16)
68Radiation from a dual-polarized Notch Antenna
Array (16x16)
Broad Side
Broad Side
69Active Reflection Coeff. With Broad and 300 Scan
70Electromagnetic Bandgap (EBG) material
Artificial Magnetic Conductor