(1) Computer Science Department - University of Alcalá, SPAIN

19
A Numerically Efficient Technique for A Numerically Efficient Technique for Orthogonalizing the Basis Functions Orthogonalizing the Basis Functions Arising in the Solution of Arising in the Solution of Electromagnetic Scattering Problems Electromagnetic Scattering Problems Using the CBFM Using the CBFM Carlos Delgado Carlos Delgado (1) (1) , Felipe Cátedra , Felipe Cátedra (1) (1) and Raj Mittra and Raj Mittra (2) (2) (1) Computer Science Department - University of Alcalá, SPAIN (2) Electromagnetic Communication Laboratory - Penn State University, USA

description

A Numerically Efficient Technique for Orthogonalizing the Basis Functions Arising in the Solution of Electromagnetic Scattering Problems Using the CBFM Carlos Delgado (1) , Felipe Cátedra (1) and Raj Mittra (2). (1) Computer Science Department - University of Alcalá, SPAIN - PowerPoint PPT Presentation

Transcript of (1) Computer Science Department - University of Alcalá, SPAIN

Page 1: (1) Computer Science Department - University of Alcalá, SPAIN

A Numerically Efficient Technique for A Numerically Efficient Technique for Orthogonalizing the Basis Functions Arising Orthogonalizing the Basis Functions Arising

in the Solution of Electromagnetic in the Solution of Electromagnetic

Scattering Problems Using the CBFMScattering Problems Using the CBFM

Carlos DelgadoCarlos Delgado(1)(1), Felipe Cátedra, Felipe Cátedra(1)(1) and Raj Mittra and Raj Mittra(2)(2)

(1) Computer Science Department - University of Alcalá, SPAIN(2) Electromagnetic Communication Laboratory - Penn State University, USA

Page 2: (1) Computer Science Department - University of Alcalá, SPAIN

Outline

• Introduction• Characteristic Basis Function Method (CBFM):

General features• Speeding up the CBF generation with the Physical

Optics approach• CBF orthogonalization via multi-step SVD

computations• Numerical Results• Conclusions

Page 3: (1) Computer Science Department - University of Alcalá, SPAIN

Introduction-I

• When analysing electrically large problems, the high sampling rate required by the conventional Method of Moments (MoM), typically around /10, usually leads to huge linear systems of equations described by dense matrices, which involves serious computational burden.

• Some efficient numerical techniques improve the MoM performance by rigorously computing and storing only the near-field terms of the coupling matrix, highly improving the CPU-time via an efficient evaluation of the matrix-vector product operations in the iterative system solution process and also reducing the memory requirements.

• Other family of approaches rely on the reduction of the total number of unknowns defining macro-basis functions over enlarged domains, in order to avoid the bottleneck imposed by the conventional MoM sampling rate.

Page 4: (1) Computer Science Department - University of Alcalá, SPAIN

Introduction-II

• The Characteristic Basis Function Method (CBFM) introduces a new system of equations related to a set of high-level basis functions defined over relatively large domains.

• Each CBF can be expressed in terms of the conventional MoM subdomain functions (RWG, Rooftops, etc).

• The CBFs do not rely upon predetermined fixed shapes. They are estimated for each problem.

• The application of high-level macro-basis functions introduces the possibility of using direct solvers for the solution of some problems where iterative solvers are the only possible choice if using the low-level discretization.

• The CBFM is a kernel-independent approach.

Page 5: (1) Computer Science Department - University of Alcalá, SPAIN

The Characteristic Basis Function Method (CBFM): General Features-I• What are the motivations for reducing of the number of unknowns via

the CBFM?

-Reduction of the matrix size-Maintain a good condition number

• This high-level basis functions are defined over large subdomains called blocks. It is necessary to separate the geometry into blocks during the preprocessing stage.

• In each of these blocks, the current distribution is reprsented in terms of a superposition of several Characteristic Basis Functions (CBFs) whose weights are to be determined solving the system that describes the problem.- It is important to minimize the number of CBFs, since it is going

to determine the size of the Reduced matrix that represents the coupling between the blocks

Page 6: (1) Computer Science Department - University of Alcalá, SPAIN

• Each CBF is defined in terms of weighted low-level subdomain functions, such as Rooftops, RWGs, etc.

• The reduced matrix represents the coupling between the CBFs

Reduced Matrix:

Diagonal blocks: Self-InteractionsOff-Diagonal blocks: Mutual Coupling

The Characteristic Basis Function Method (CBFM): General Features-II

Page 7: (1) Computer Science Department - University of Alcalá, SPAIN

• The generation of CBFs is very flexible– Primary & secondary CBFs vs. PWS

– Different approaches can be used (MoM, PO, ...)

k1i

k2i

k3i

k4i

k5i

k6i

k7i

k8i

k9i

k10i

k11i

k12i

k13ik14

i

k15i

k16i

k17i

k18i

k19i

k20i

k21i

k22i

k23i

k24i

S

Dq• To avoid the computation of secondary CBFs, the excitation can be derived from several plane waves with different incidence angles- This approach renders the reduced matrix

independent of the excitation

• Orthogonalization techniques are applied in order to reduce the number of basis functions, and to imporve the condition number- The same set of CBFs can be used for different blocks

The Characteristic Basis Function Method (CBFM): General Features-III

Page 8: (1) Computer Science Department - University of Alcalá, SPAIN

• If the MoM is applied to obtain the currents induced by an incident plane wave, each contour segment is extended a distance Δu and Δv, so that there is a physical overlap between thr extended blocks, in order to avoid the artificially introduced edge behaviour at the interface between blocks

Extended block i

Extended bloque j

D u

D u

D u

D u

vD vD

S

i

j

u

vu

v

i

j

• After obtaining the induced currents, only the values contained within the original (non-extended) blocks are retained.

• For large and smooth surfaces, it is possible to use the PO

The Characteristic Basis Function Method (CBFM): General Features-IV

approach instead of the MoM to obtain these currents. This avoids

matrix solutions while generating the CBFs, and saves considerable time.

Page 9: (1) Computer Science Department - University of Alcalá, SPAIN

Generation of CBFs Using Physical Optics-I

• Application of MoM to obtain the currents induced by a set of plane waves entails solving the following matrix problem in a rigorous way.

• When the object geometry is subdivided into a relatively large number of medium-sized blocks, and when the geometries of these blocks do not have fine features, it is considerably more efficient to resort to high-frequency techniques, such as the Physical Optics (PO) approach, for deriving the CBFs. This is because, in contrast to MoM, no matrix solution is needed in this approach.

Page 10: (1) Computer Science Department - University of Alcalá, SPAIN

• The current induced by the incident plane waves on the surface are calculated by applying the PO approach in the middle point of the domain where each low-level basis function is defined.

Generation of CBFs Using Physical Optics-II

Sum of all the low-level basis functions included

in the block Incident field

Low-level basis function “n”

Sampling point over the “n-th” basis function

Incidence direction of the plane wave

Page 11: (1) Computer Science Department - University of Alcalá, SPAIN

Multi-Step SVD Computation for Obtaining the CBFs-I

• The ratio between the number of CBFs and low-level basis functions decreases as the block size is increased, which makes it desirable to develop techniques that calculate the CBFs efficiently over large surfaces.

Page 12: (1) Computer Science Department - University of Alcalá, SPAIN

Multi-Step SVD Computation for Obtaining the CBFs-II

• The Singular Value Decomposition can be used to obtain an orthogonal set of vectors which replace the original induced currents.

A =

………

Sample points (low-level basis functions)

Cur

rent

s in

duce

d by

pla

ne w

aves

= [U][σ][V]*

The rows of [V]* are the new basis vectors

We only retain vectors for which the associated singular value is above a

certain threshold:

Page 13: (1) Computer Science Department - University of Alcalá, SPAIN

Multi-Step SVD Computation for Obtaining the CBFs-III

• When electrically large blocks are considered, the Singular Value Decomposition can present considerable computational burden if the entire Plane Wave Spectrum is considered. Also, this can lead to:– Increase in CPU-time

– Possible inaccuracies due to internal error accumulation, which forces us to increase the numerical precision for the computation

• A better alternative is to perform the computation in multiple steps, limiting the number of “input basis vectors” or rows in the A matrix for each computation. The resulting basis vector from one step serve as the input vectors for the next level. Finally, the resulting set of basis vectors in the last step are retained as the CBFs for the purpose of Reduced Matrix generaion.

Page 14: (1) Computer Science Department - University of Alcalá, SPAIN

Multi-Step SVD Computation for Obtaining the CBFs-IV

• The PWS can be represented as a set of points in the (θ,φ) plane containing all the possible incident plane waves within the visible range.

q

f0 2p

0

p• The plane waves are compartmentalized into several groups.

• After performing the SVD for each group, the resulting basis vectors corresponding to adjacent regions are grouped one again and a new SVD operation is performed.

• Finally, we retain the result of the highest-level SVD computation as the desired CBFs.

Page 15: (1) Computer Science Department - University of Alcalá, SPAIN

Numerical Example-I

• Scattering problem considered: PEC sphere (R=1m) modeled by means of 8 NURBS patches.

• Simulations run on a Pentium IV at 2.8 GHz PC with 1.5 Gbytes of RAM memory.

• Bistatic RCS obtained at a frequency of 1.5 GHz for both polarizations.

• Using a λ/10 discretization, 12324 low-level subdomains are needed for each block. In contrast to this, the CBFM retains only 806 CBFs per block.

Page 16: (1) Computer Science Department - University of Alcalá, SPAIN

Numerical Example-II

• CBFM results have been compared with the analytical solution (MIE).

Page 17: (1) Computer Science Department - University of Alcalá, SPAIN

Numerical Example-III

• CPU times for the CBF generation process have been compared when considering different levels.

NUMBER OF LEVELS CPU-TIME (CBF Generation)

1 * Out of memory *

2 5404 s

3 2804 s

4 1812 sq

f0 2p

0

p

Page 18: (1) Computer Science Department - University of Alcalá, SPAIN

Numerical Example-IV

• The threshold value used for retaining the output vectors can be relaxed after the first-level SVD computation.

The number of basis vectors retained after the SVD computation for level-L is

given by the condition:

Page 19: (1) Computer Science Department - University of Alcalá, SPAIN

Conclusions

• A methodology for enlarging the block sizes in the CBFM, without increasing the associated CPU-time needed to compute the CBFs, has been presented.

• Larger blocks usually yield fewer CBFs, leading, in turn, to a smaller-size Reduced Matrix.

• The PO approach circumvents the matrix solution for generating the CBFs, and is well-suited for smooth parts of the geometries of electrically large blocks.

• The total number of CBFs obtained with the present approach is of the same order as that obtained by carrying out a single (no multi-level) SVD computation.

(u a,v a)

u a

v a

b 01

b 20

b 02

1

0

0 1

u

v

b 00

b 10

b 21

b 22

b 12

b 11

u

v

Control Mesh

Control Points