Upload
hadiep
View
219
Download
2
Embed Size (px)
Citation preview
Acoustical scattering from N spheres using
a multilevel fast multipole method
Nail A. Gumerov Ramani Duraiswami
Institute for Advanced Computer StudiesUniversity of Maryland at College Parkwww.umiacs.umd.edu/~gumerovwww.umiacs.umd.edu/~ramani
This study has been supported by NSF
Outline! Introduction! Problem Formulation! Method of Solution
! Multipole Reexpansion (T-matrix) Method! Iterative Methods! Fast Multipole Method
! Results of Computations! Conclusion
Introduction
Multiple Scattering Problems! Sound propagation in disperse media
(particles, bubbles, etc.)! Modeling of scattering from environment
(humans, animals, fish, etc.)! Electromagnetic scattering problems
(microwaves, optics, etc.)! Efficient parametrization in inverse problems
(tomography, etc.)
Introduction
Why Multipole Methods?
Introduction
BEM Mesh5402 nodes10800 elementsDiscretization # Dn=30
Can be usedto compute the fieldonly for ka < 25 (for human head < 16.5 kHz)
Run Time for one frequencyon Dual Processor1 GHz Pentium III ~ 1 day.
Required Maximum Frequency to Compare with ExperimentalHRIR (22 or 44 kHz), 200 frequencies
Formal Requirement:ka << Dn
Scattering computation with BEM for a single object
Why Multipole Methods?
! Meshless for spherical scatterers
! Fast
! Needs Mesh! Relatively Slow
Introduction
Multipole Methods BEM
Problem Formulation
Equations and Boundary Conditions
Helmholtz Equation
Impedance Boundary Conditions
Field Decomposition
Sommerfield Radiation Condition
4
2
1
6
5
3
Incident Wave
Formulation
Wave EquationFourierTransform
Multipole Reexpansion(T-Matrix) Method
Scattered Field Decomposition
T-Matrix Method
Expansion Coefficients
Singular Basis Functions Hankel Functions
Spherical Harmonics
Vector Form:
dot product
Incident Field Decompositionand T-matrix for a Single Sphere
T-Matrix Method
Regular Basis Functions Bessel Functions
Analytical Solution of the Problem:
T-matrix
Solution of Multiple Scattering Problem
T-Matrix Method
4
2
1
6
53
Incident Wave
Scattered Wave
Coupled System of Equations:
(S|R)-TranslationMatrix
�Effective� Incident Field
Reexpansions/Translations
T-Matrix Method
q
p
M
O
rprq
r�pr�q
r�pq
r q
p
M
O
rprq
r�pr�q
r�pq
r
Two Spheres: Convergence with Respect to Truncation Number
T-matrix Method
-30
-20
-10
0
10
0 10 20 30 40 50Truncation Number
HR
TF (d
B)
ka1=30
5110
20
Two spheres,θ1 = 60o, φ1 = 0o,rmin/a1 = 2.3253.
Three Spheres Comparisons ofBEM & MultisphereHelmholtz
T-matrix Method
BEM: 5184 triangular elementsMH: Ntrunc = 9 (100 coefficients for each sphere)
-12
-9
-6
-3
0
3
6
9
12
-180 -90 0 90 180
Angle φ1 (deg)
HR
TF (d
B)
BEMMultisphereHelmholtz
θ1 = 0o
30o
60o
90o
60o
30o
90o
120o
120o
150o
150o
180o
Three Spheres, ka1 =3.0255.
Conclusions on T-matrix Method! We used recursive computation of translation
matrices (Chew, 1992; Gumerov & Duraiswami, 2001).
! In some cases speed up of computations 103-104
times compared to BEM.! But� Computational Complexity is O(N3P3)= O(N3
Nt6), where P= Nt
2 is the total length of the vector of expansion coefficients. Method is not suitable for large N and ka.
! Details can be found in our paper JASA 112(6), 2002, 2688-2701.
T-matrix Method
Iterative Methods
Reflection Method & Krylov Subspace Method (GMRES)
Reflection (Simple Iteration) Method:
General Formulation (used in GMRES)
Iterative Methods
Convergence of Reflection Iteration Method
Iterative Methods
Exponential Convergence
Conclusions on Iterative Methods
! Both the Reflection Method (RM) and the GMRES converge well, while the RM is simpler and faster;
! Some problems in convergence were found for larger ka and regular spacing of the scatterers;
! In iterative methods fast translation algorithms can be used (weused O(Nt
3)=O(P3/2) fast translation based on sparse matrix decomposition of translation operators). This cost potentially can be reduced further (we are working on O(PlogP) methods).
! Complexity of Iterative Methods in this case O(N2Niter Nt3);
! Savings in complexity compared to straightforward T-matrix are O(Nt
3 N/Niter)! For N~200, Niter~20, Nt~10 (P~100) this yields of order 104
times savings.
Iterative Methods
Multilevel Fast Multipole Method
Some Facts on the Fast Multipole Methods (FMM)
! Introduced by Rokhlin & Greengard (1987,1988) for computation of 2D and 3D fields for Laplace Equation;
! Reduces complexity of matrix-vector product from O(N2) to O(N) or O(NlogN) (depends on data structure);
! Hundreds of publications for various 1D, 2D, and 3D problems (Laplace, Helmholtz, Maxwell, Yukawa Potentials, etc.);
! Application to acoustical scattering problems (Koc & Chew, 1998; JASA);
! We taught the first in the country course on FMM fundamentals & application at the University of Maryland (2002);
! Our recent report on fundamentals and data structures is available online (visit our web pages).
MLFMM
Far and Near Fields
MLFMM
Neighborhood(Near Field)
Far Field
Computation of the Far Field (1)
MLFMM
xixc
(n,L)
y
1). Set Data Structure (hierarchically subdivide space with an oct-tree)
2). (S|S)-translate S-expansions for all scatterers in a box at the finest level to the center of the box and sum up (determine contribution to Far Field for each box at the finest level).
3). Recursively (S|S)-translate S-expansions to the center of the parent box and sum up (determine contribution to Far Field for each box at all courser levels).
UpwardPass
(From the finestto the coarsest
level)
Computation of the Far Field (2)
MLFMM
4). (S|R)-translate S-expansions for boxes which are inside the parent neighborhood but outside the box neighborhood to the center of the box (convert S-expansion to R-expansion).
5). (R|R)-translate R-expansions from the center of the box to the center of its child boxes (determine Far Field for each box at all levels).
DownwardPass
(From the coarsest
to the finest level)
Steps 4 and 5 performed one after the other recursively
6). (R|R)-translate R-expansions from the center of the boxes at the finest level to the centers of the spheres.
Complexity of MLFMM
MLFMM
! Savings in complexity compared to the iterative T-matrix method are O(NNt/(logN)).
! For N~104, Nt~10 this yields about 104 times speed up (109
speed up compared to straightforward T-matrix).
(For translation cost O(Nt3)= O(P3/2) without optimization):
(with optimization of the number of scatterers inside the finest level box):
Computable Problems on Desktop PC
MLFMM
Met
hod
Number of Scatterers101 102 103100 104 105
BEM
Multipole Straightforward
Multipole Iterative
MLFMM
Also strongly depends on ka !
Results of Computations
Range of Parameters! Number of Spheres: 1-104;! ka: 10-3-50;! Random and regularly spaced grids of
spheres;! Polydispersity: 0.5-1.5 (ratio to the
mean radius);! Volume fractions: 0.01-0.1
Results
4 spheres (T-matrix straightforward)
Results
Vector of the incidentplane wave
Imaging plane
Scatterers ka=15.2
Incident Field Total Field Scattered Field
100 random spheres (MLFMM)
Results
ka=1.6 ka=4.8ka=2.8
R G B
Convergence for 100 spheres (MLFMM)
Results
1.E-04
1.E-03
1.E-02
1.E-01
1.E+00
1.E+01
1.E+02
0 5 10 15 20 25 30 35Iteration #
Max
Abs
olut
e Er
ror
Iterations with Reflection Method
3D Helmholtz Equation,MLFMM100 Spheres
ka = 4.8
2.8 1.6
1000 random sphereska=1
Conclusions
! We developed, implemented, and tested the Multilevel Fast Multipole Method for computation of multiple scattering problems.
! Performance of the method depends on a number of controlling parameters. At proper selection of these parameters fast and accurate results can be achieved.
! Some convergence problems in iterative methods were observed for short wave propagation in regularly spaced sphere grids. This may be due to some internal resonances, which should be investigated.
Future work! Development of faster translation algorithms;! Extension for non-spherical scaterrers;! Comparisons with continuum (averaging)
theories and theories of wave propagation in random media;
! Computations of acoustic fields in disperse systems (bubbly liquids, particulate systems);
! Comparisons with experimental data.
THANK YOU !