Proceedings of The Twelfth (2002) International Offshore and Polar Engineering Conference Kitakyushu, Japan, May 26 31, 2002

Copyright 2002 by The International Society of Offshore and Polar Engineers ISBN 1-880653-58-3 (Set); ISSN 1098-6189 (Set)

Accelerated Higher Order Boundary Element Method for Wave Diffraction/Radiation Problems and Its Applications
Tomoaki Utsunomiya and Eiichi Watanabe
Dept. of Civil Engineering, Kyoto University Kyoto, Japan

ABSTRACT
This paper presents an accelerated higher order boundary element method for wave diffraction/radiation problems and its applications, especially for wave response analysis of VLFS (Very Large Floating Structures). The Fast Multipole Method (FMM) has been implemented on the higher order boundary element code using 8-node quadrilateral element. The method utilizes an iterative solver, multipole expansion of Greens function, and hierarchical algorithm using quadrant-tree. For solving hydroelastic problem efficiently using iterative solver, a new algorithm, where the equations of motions representing plate vibration are solved at each iterative step, has been introduced. The numerical benchmark calculations have shown the efficiency of the method both in storage requirement of O(N) and computation time of O(N log N), where N is the number of unknowns for the velocity potential.

This paper employs an alternative approach using Fast Multipole Method (FMM) (Rokhlin, 1985; Barnes and Hut, 1986; Greengard and Rokhlin, 1987; Greengard, 1988; Fukui and Katsumoto, 1997, 1998). The FMM is applied here to accelerate the calculation of the Higher Order Boundary Element Method (HOBEM) using 2nd order 8-noded quadrilateral panels. The essential theoretical background of the FMM is based on Graf's addition theorem of Bessel functions which appear in the free surface Green function, introduction of hierarchical algorithm to compute the system equations, and employment of the iterative solver. Although similar approach to accelerate BEM for wave diffraction/radiation problems can be found using the pre-corrected FFT algorithm (Korsmeyer, et al., 1996, 1999; Kring, et al., 2000), it is limited up to now to a low-order panel and a deep-water case. Formulations are presented with the several benchmark calculations including wave response analysis of VLFS. The storage requirement of O(N) and CPU time characteristics of O(N log N) have been confirmed by the benchmark calculations.

KEY WORDS: fast multipole method; higher order boundary element method; diffraction; radiation; very large floating structure; VLFS; hydroelasticity. INTRODUCTION
The Boundary Element Method (BEM) using the free surface Green function has been used as a basic tool for solving diffraction/radiation problems of a floating body. However, when the method is applied to a Very Large Floating Structure (VLFS) such as a Mega-Float, the large requirement of computer resources with O(N2) for storage and with either O(N2) or O(N3) for CPU time, which depends on the selection of the solver either as iterative type or LU-factorization type (N: the number of unknowns), has made the computation of velocity potentials by BEM impractical. Therefore, the computation for a VLFS has been tackled by the finite element method (FEM) where band storage characteristics of the system matrix are utilized, by the semi-analytical approaches, or by the BEM utilizing higher-order panels such as B-spline functions (see e.g. Ohtsubo and Sumi, 2000). However, the higher-order panel method based on B-spline functions is hardly convergent for iterative solvers.

FORMULATIONS The Boundary Value Problem

As shown in Fig. 1, consider a box-like VLFS (Very Large Floating Structure) of the length 2a (=L), the width 2b (=B), the draft d, and floating in the open sea of variable depth (with constant depth h at infinity). The variable depth sea-bottom surface (SB) is defined such that the boundary line (B) touches on the flat bottom base-surface of z=-h and the SB must locate higher than the base surface (z=-h). The coordinate system is defined such that the xy plane locates on the undisturbed free surface and the z axis points upwards. The center of the VLFS is on the z axis as shown in Fig. 1. Consider the long-crested harmonic wave with small amplitude. The amplitude of the incident wave is defined by A, the circular frequency by , and the angle of incidence by (=0 corresponds to the head wave from positive x direction and =/2 to the beam wave from positive y direction).

305

Assuming the water to be perfect fluid with no viscosity and incompressible, and the fluid motion to be irrotational, then the fluid motion can be represented by a velocity potential . Also, we consider the steady-state harmonic motions of the fluid and the structure, with the circular frequency . Then, all of the time-dependent quantities can by represented similarly as follows:

The Boundary Integral Equation

In this study, we employ the following integral equation as a fundamental equation, which is applicable to a floating body with variable-depth sea bottoms and/or breakwaters (Teng and Eatock Taylor, 1995; Utsunomiya, et al., 1998):

( x, y , z; t ) = Re[ ( x, y, z )e it ]

(1)

where i is the imaginary unit (i2=-1) and t is the time. In the following, all time-dependent variables are represented in the frequency domain unless stated explicitly. Assuming also that the fluid and structural motions are small so that the linear potential theory can be applied for formulating the fluid-motion problem. With these assumptions, the following boundary value problem may be formulated:

Incident wave

2 = 0 = K z =0 n =0 z = iw( x, y ) n =0 n lim r (
r

in on S F on S B on B0 on S HB on S HS

(2) (3)

(4) Fig. 1 Configuration of the analytical model. (5)

(6)

(7)

( I ) ik ( I )) = 0 on S r

(8)

I = i

gA cosh k ( z + h ) ik ( x cos + y sin ) e cosh kh

r r r G 2 ( x , ) ( 4 + S dS ) ( x ) z r r r r r G ( x , ) r G 2 ( x , ) (11) + S S S { ( ) ( x) }dS n n r r r ( ) r S S S G ( x , ) dS = 4 I ( x ) n r r Here, x = ( x, y , z ) and = ( , , ) . The symbol S I designates the

I HB HS B HB HS B

(9)

inner plane of z = h inside the boundary B . It is noted that the integral in Eq. (11) is performed for = ( , , ) and the normal derivative is made on the same surface. Using Eq. (11), the evaluations of solid angles and CPV integrals that may needed when usual Fredholm type integral equations are used, can be avoided. The function G ( x , ) represents the Green function representing water waves, and satisfying the boundary conditions on the free surface, on the flat sea bottom (z=-h), and the radiation condition at infinity. The series form can be represented as follows (Newman, 1985; Linton, 1999):
2 K (k R) r r 0 m G( x, ) = cos k m ( z + h ) cos k m ( + h ) m =0 Nm

where I is the incident wave potential. The symbols , S F , S B , and B0 represent the fluid domain, the free surface, the variable depth sea-bottom surface, and the flat-bottom base surface on z = h , respectively. The symbols S HB , S HS , and S represent the bottom surface of the floating body, the wetted side surface of the floating body, and the artificial fluid boundary at infinity, respectively. The 2 symbol K represents the wave number at infinite depth sea ( = / g ; g is the gravitational acceleration), whereas k is the wave number at constant depth h satisfying the following dispersion relation:

r r

k tanh kh = K

(10)

(12) (13) (14) The

The symbol n represents unit normal vector (the positive direction points out of the fluid domain). The complex-valued variable w( x, y ) represents the vertical deflection of the bottom surface of the floating body. The variable r in Eq. (8) represents horizontal distance between the origin and the referred point.

Nm =

sin 2k m h h (1 + ) 2 2k m h

k m tan k m h = K
where k m ( m 1 ) is a positive real number, and k 0 = ik . following relationship is also used:

306

1 K 0 (ikR ) = iH 0( 2 ) ( kR ) 2

(15)
(2)

V = S HB p( x, y ) wdS

(23)

Here, K 0 is the modified Bessel function of the second kind, and H 0 is the Hankel function. The symbol R represents horizontal distance between x and . The function G2 ( x , ) is given as (Teng and Eatock Taylor, 1995):

Substituting Eq. (18), Eqs. (21)-(23) into Eq. (20), and calculating the variations with respect to w, we have

r r

D S HB [

2 w 2 w 2 w 2 w 2 w 2 w 2 w 2 w + 2 + 2 + x 2 x 2 y y 2 x y 2 y 2 x 2 2 w 2 w ]dS + k S HB wwdS 2 S HB wwdS xy xy

(24)

r r 1 1 1 1 + G2 ( x , ) = + + r r1 r21 r31 r = [ R 2 + ( z ) 2 ]1 / 2 r1 = [ R 2 + ( z + ) 2 ]1 / 2 r21 = [ R 2 + ( z 2h ) 2 ]1 / 2 r31 = [ R 2 + ( z + + 2h ) 2 ]1 / 2 Equation of Motion for VLFS Deflections

(16)

+ 2(1 )

= i S HB ( x, y,d )wdS
Here, substituting Eq. (19) and the variation into Eq. (24), and considering the arbitrarity of l , we finally obtain the following equation of motion in the modal coordinates:

l ( K lj 2 M lj ) = i S ( x, y , d ) f j dS
l =1
HB

(25)

Here, It is now widely accepted that VLFS response in terms of the vertical deflection can be captured well by modeling the whole VLFS as an elastic plate. In this study, we employ the thin plate theory on elastic foundations (which model the hydrostatic restoring forces):

K lj = D S HB [

2 2 2 2 fl f j 2 fl f j 2 fl f j + + x 2 x 2 y 2 y 2 x 2 y 2

D 4 w( x, y ) 2 w( x, y ) + kw( x, y ) = p( x, y )

(17)

2 2 2 fl f j 2 fl f j ]dS + k S HB f l f j dS + + 2(1 ) 2 2 y x xy xy

(26)

where D is the plate rigidity, is the mass per unit area, k = g ( : density of fluid), and p (x, y ) is the dynamic pressure on the bottom surface of the VLFS. The dynamic pressure p( x, y ) relates to the velocity potential on the bottom surface of the VLFS from the linearized Bernoulli's equation:

M lj = S HB f l f j dS

(27)

p ( x, y ) = i ( x, y , d )

(18)

Our target is to solve the stationary solution which is governed by Eq. (17) and the free boundary conditions are specified along the edges of the plate. In the following, we firstly represent the deflection w( x, y ) by series of the products of the modal functions f l ( x, y ) and the complex amplitudes l , and then derive the equation of motion for the modal amplitudes using Ritz-Galerkin's method:

are the generalized stiffness matrix and the generalized mass matrix, respectively. The selection of the modal function has in fact some arbitrariness; e.g., dry-modes of the plate may be frequently used. In this study, we choose the modal function made as a tensor product of modal functions of a vibrating free-free beam in the x and y directions (Utsunomiya, et al., 1998).

Formulation for Hydroelastic Analysis

Substituting Eqs. (4), (6) and (7) into Eq. (11), we have

w( x , y ) = l f l ( x , y )
l =1

(19)

The Hamilton's principle is used to derive the equation of motion.

(U T + V ) = 0

r r r G 2 ( x , ) dS ) ( x ) ( 4 + S z r r r r r G ( x , ) r G 2 ( x , ) + S S S { ( ) ( x) }dS n n r r r i S G ( x , ) w( , )dS = 4 I ( x )
I HB HS B HB

(28)

(20) Moreover, substitution of Eq. (19) into Eq. (28) yields

Here, U represents the strain energy, T the kinetic energy, and V the potential energy due to external force.

2w 2w 2w 2w 1 U = S HB [ D{( 2 ) 2 + ( 2 ) 2 + 2 2 2 x y x y 2 + 2(1 )( T= 2w 2 ) } + kw 2 ]dS xy

(21)

r r r G 2 ( x , ) dS ) ( x ) ( 4 + S z r r r r r G ( x , ) r G 2 ( x , ) + S S S { ( ) ( x) }dS n n P r r r i l S G ( x , ) f l ( , )dS = 4 I ( x )
I HB HS B

(29)

l =1

HB

1 2 S HB w 2 dS 2

(22)

If we determine the unknown modal amplitudes l ( l = 1,K, P ) and

the unknown nodal potentials i ( i = 1,K, N ) to be distributed on the

surface S HB , S HS and S B , the problem is solved. This can be done by

307

solving the simultaneous equations composed of (25) and (29). The right-hand side of Eq. (25) may be rewritten as

1 S HB ( x, y,d ) f j dS = S e [ N ,K, N ] f j dS M e =1 8 e
NHB 1 8

and also [ K 2 M ] is a real banded matrix, thus the LU decomposition would be trivial in the total computation time. When we employ Eq. (35) for solving the velocity potential by using an iterative solver, we can obtain the final solution with one iterative procedure. This feature is most important advantageous point over the conventional modal method when an iterative solver is used. (The computation time is reduced by a factor of 1/(number of modes)). (30)

1 NHB = [ L j ] e M = [ L j ]{ } e =1 8 e

( j = 1,K, P )

ACCELERATION BY FAST MULTIPOLE METHOD Mutipole Expansion of the Green Function

In the following, we consider the multipole expansion of the free surface Green function in the series form, Eq. (12). As an alternative expression of the Graf's addition theorem for the Bessel function, the next equation can be obtained (Utsunomiya, et al., 2001).

In Eq. (30), the symbol S e represents each element (or panel) on the bottom surface of the floating body S HB , and the number of elements is represented by NHB . That is
NHB

S HB = S e e =1

(31)

K 0 (k m R ) = K n ( k m r )e in I n ( k m )e in
n =

(36)

potential inside each element from the nodal values of , K ,

1

The symbols N 1 ,K, N 8 represent shape functions to interpolate the

8

(when 8 noded panel is used). The complex valued vector { } has elements of the potentials at nodes i ( i = 1,K, N ), and the row vector [ L j ] (the size is N) is composed by a superposition of [ L j ]e at the corresponding nodes. Substituting Eq. (30) into Eq. (25), and expressing it in the matrix form, we have

Here, x = ( r, , z ) and = ( , , ) are the points measured from newly defined cylindrical coordinate systems, of which origin O may be referred to as the multipole expansion point. The multipole expansion point O can be located arbitrarily on the undisturbed free surface (thus on the xy plane of the global rectangular coordinate system) under the restriction of < r . Substituting Eq. (36) into Eq. (12), we have
r r G ( x , ) = M mn K n ( k m r )e in cos k m ( z + h ) m = 0 n =

[ K M ]{ } = i [ L]{ }
2

(32)

(37)

Here, { } is the complex valued vector of which elements are l ( l = 1,K, P ), [ K M ] is the real matrix (the size is P P ) of
2 2 which elements are K lj M lj ( l , j = 1,K, P ), and [ L] is the real

M mn =

2 I n (k m )e in cos k m ( + h ) Nm

(38)

matrix (the size is P N ) of which row is composed by [ L j ] ( j = 1, K , P ).

Thus, we have obtained the expression of the Green function represented in the form of the multipole expansions. The normal derivative of the Green function can be obtained similarly as

Eq. (29) may be represented in a matrix form as

[ A]{ } i [ B ]{ } = 4 { I }

r r G ( x , ) M mn K n ( k m r )e in cos k m ( z + h ) = m = 0 n = n n M mn M mn M mn M mn = n + n + n n r
r

(39)

(33) (40)

where [ A] is a complex valued matrix (the size is N N ), [B ] is a complex valued matrix (the size is N P ), and { I } is a vector (of the size N). From Eq. (32),

Here, ( n , n , n ) represent the coordinate values in the directions of ( , , ) of the normal vector n defined at , where

{ } = i [ K M ] [ L]{ }
2 1

(34)

Substituting Eq. (34) into Eq. (33), we have

n = n cos + n sin and n = n sin + n cos .

(35) The multipole expansion point O can be moved to arbitrary position in the global xy plane if the restriction < r is satisfied. The expansion coefficient M mn with respect to the newly defined multipole expansion point O relates to the original expansion coefficient M mn as follows (Utsunomiya, et al., 2001):
~ M m = M mn I n ( k m )e i ( n ) n =

([ A] 2 [ B ][ K 2 M ] 1 [ L]){ } = 4 { I }

Solving first Eq. (35) for i ( i = 1,K, N ) and then substituting them into Eq. (34), we also have l ( l = 1,K, P ). Thus we have solved the given problem.

In the above equation, we have unknowns only for i ( i = 1,K, N ).

2 In Eq. (35), we have to evaluate LU factorization of [ K M ] in advance; however, we have to be aware that for most cases P << N

(41)

308

Here, ( , ) represent the polar coordinate values of the original multipole expansion point O with respect to the newly defined multipole expansion point O. Similarly, the following relationship is satisfied:

(4 + 1
1 e =1

NI

NE G 2 r 1 G | J e |d d 1 2 | J e | d d ) ( x ) = er 1 z n ( x e )

~ M M m mn = I n ( k m )e i ( n ) n = n n

+ 1 {
1 er 1 = ( x e )

NE

G 8 k k G 2 r ( x )} | J e | d d N n k =1 n
k k 1 n N | J e | d d k =1 1

(42)

e =1 r r ( x e ) ( e: near of x )

NE

Implementation to Higher Order Boundary Element Method

In Eq. (29), the distribution of the potential within each element (or panel) may be expressed by using the 2nd order interpolation functions and the potentials at the nodal points of the element as follows:

+
1

m = 0 n = r e =1 r ( x e ) ( e:far from x )

K n (k m r )e in cos k m ( z + h)

NE

k k mn 1 n 1 N | J e | d d k=

( , , ) = N k ( , ) k
k =1

(43)

NHB

e =1 r ( e: near of x )

1 G l f l ( , ) | J e | d d l =1
1 NHB

(47)

where ( , ) represents the coordinate values in the curved plane coordinate system defined on the surface of each element. Substitution of Eq. (43) into Eq. (29) yields

i
1 P

m = 0 n = e =1 r ( e:far from x )

K n (k m r )e in cos k m ( z + h)
r

r G ( 4 + 1 2 | J e |d d ) ( x ) e =1 z NE G 8 k k G 2 r 1 + 1 { ( x )} | J e | d d N e =1 n k =1 n
NI 1 P NHB r 1 i l 1 Gf l ( , ) | J e | d d = 4 I ( x ) l =1 e =1

1 M mn l f l ( , ) | J e | d d = 4 I ( x ) l =1
(44)

Application of the Fast Multipole Method

When Eq. (47) is solved by an iterative method, { } is calculated for an assumed { } using Eq. (34), and then the left-hand side of Eq. (47) is obtained for the assumed { } . The values of { } are then updated to minimize the error between the left-hand side and the right-hand side values of Eq. (47). In the calculation of the left-hand side values of Eq. (47), the diagonal terms of the coefficient matrices are stored in the core memory; on the other hand, the off-diagonal terms are calculated in each step of the iterative process. This enables the drastic reduction of the storage requirement to O(N). However, of course, we have to accelerate the computation concerning the off-diagonal terms of the coefficient matrices. The calculation concerning the off-diagonal terms may be separated r into two parts: the first one is on the calculation for elements near x (3rd and 6th lines in Eq. (47)), and the second one is on the calculation r for elements far from x (4th, 5th, 7th and 8th lines in Eq. (47)). In the following, we explain the accelerated calculation method of the terms in the 4th and 5th lines in Eq. (47) as an example for the second part.
r b( x ) = n m m = 0 n = r e =1 r ( xe ) ( e:far from x ) in K (k r )e cos k m ( z + h ) NE

where J e = J e ( , ) represents the Jacobian defined for each element e. The symbols NHB , NHS , NB , and NI represent the number of elements on the surfaces of S HB , S HS , S B , and S I , respectively, and

NE = NHB + NHS + NB
is satisfied. Eq. (44) may be rearranged as

(45)

( 4 + 1
1 e =1

NI

NE r G 2 1 G | J e |d d 1 2 | J e | d d ) ( x ) e =1 z n r ( xe )

+ 1{
1 e =1 r ( xe ) NE

NE

G 8 k k G 2 r ( x )} | J e | d d N n k =1 n

+ 1
1 e =1 r ( xe ) NHB e =1

G 8 k k N | J e | d d n k =1
P

r 1 i 1 G l f l ( , ) | J e | d d = 4 I ( x )
l =1

(46)

where x e corresponds to the case when the collocation point x is r r included in the element e, and x e to the case when x is outside of the element e. Further substitution of Eqs. (37) and (39) into (46) yields

M mn 8 k k N | J e | d d 1 n k =1
1

(48)

As has already been defined, the values of ( r, , z ) in Eq. (48) define the collocation point x measured from the multipole expansion point O. The multipole expansion point O can be located in an arbitrary horizontal position under the restriction of < r ; in this study, the multipole expansion point O is located on the projection of the origin of the local coordinate system ( , ) of each element to the xy plane of the global coordinate system.

309

As is clear from Eq. (40), the integrand in Eq. (48) includes only the values concerning ( , , ) ; this fact means that the integral can be evaluated beforehand regardless of the position of x . By storing the values of the calculated integral in the core memory, we can obtain an efficient algorithm very simply; however, still this has an efficiency of only O(N2). Next, we introduce the concept of "cell" gathering the influences from many far field elements into an influence from one large cell (Rokhlin, 1985). For this purpose, we reform Eq. (48) as follows:
r b( x ) = Ncell

either of S HB , S HS , or S B . At this stage, we had better calculate from upper-level to lower-level, by selecting larger cells as far as possible. This hierarchical algorithm is known to be O(N log N) in the computation time (Barnes and Hut, 1986). The truncation of the infinite summations in Eq. (49) is made as m max = [6h / R ] (Newman, 1985), n max = 20 ( ka < 5 ), and

n max = [1.2ka + 15] ( ka 5 ) where a is the maximum cell size to be used in the actual computation. NUMERICAL EXAMPLES

m = 0 n =

k =1 r ( k : far from x )

K n (k m r )e in cos k m ( z + h )
(49)

M mn 8 l l 1 N | J e | d d eCell k n l =1
1

In Eq. (49), the summation for elements is replaced by the summation for cells. In Eq. (49), the cell k should be as large as possible (to include elements as many as possible) from the computational efficiency. In Eq. (49), the multipole expansion point O is located on the center of each cell, and ( r, , z ) , the coordinates of the collocation point x , is newly defined correspondingly. In general, as larger the cell size, the number of cells, Ncell, becomes small, and thus leads to shorter computation time. In this study, we employ the quadrant-tree to define the tree structure of cells and the hierarchical computation algorithm (Greengard and Rokhlin, 1987; Greengard, 1988; Fukui and Katsumoto, 1997, 1998). For simplicity, the shape of each cell is chosen to be a square. The hierarchical cell structure is organized on the xy plane (on the undisturbed free surface). Firstly, define the level-0 cell of square shape, which must include all of the analyzed area ( S HB , S HS , and

In order to check the accuracy and the computational efficiency, the wave response analysis of a box-like VLFS has been made by using the fast multipole method developed in this study and by the program using direct LU factorization solver (Utsunomiya, et al., 1998). The specifications of the VLFS are: the length L=1,500m, the beam B=150m, the draft d=1m, the rigidity as an elastic plate D=3.88x107 kNm, and the Poissons ratio =0.3. Number of modal functions employed is 160 (20 in longitudinal and 8 in beam). The water depth is 8m (constant). The wave period is 18 seconds (the corresponding incident wave length is =156.8m), and the angle of wave incidence is =/4. Table 1 shows the performance of the benchmark calculations, where in the developed program, GMRES solver is used (Barrett, et al., 1994) with the residual tolerance =10-4. The CPU time is measured with the use of single CPU of a parallel computer (IBM RS/6000 SP; CPU: POWER 3-II, 375MHz) unless otherwise stated. We note that the fast multipole method (FMM) is efficient both in CPU time and memory allocation compared to the direct method using either LU factorization solver or GMRES solver. We also note that the memory consumption is almost O(N) for the FMM. The CPU time is a little bit worth (slower) than the predicted O(N log N), but much faster than O(N2). Table 2 shows the number of iterations and CPU time to converge in GMRES for various incident wave lengths. We observe that the number of iterations is sensitive to the incident wave length, whereas insensitive to panel divisions as can be seen in Table 1. Although no preconditioning method is applied, convergent results have been obtained even at high frequency of L/=62. The efficiency of the parallel computation is also noteworthy that the computation time for the model D is reduced to 22% with the use of 5 CPUs. The increase of number of iterations with the wave frequency may be due to illconditioning at the short waves. Further studies are needed to confirm this, and to reduce the number of iteration by developing a suitable preconditioner. Fig. 2 shows that the deflection amplitudes by FMM coincide with those by direct method within the graphical accuracy for =10-3. Although we see some differences in the two graphical lines when =10-2, the results by FMM may still be satisfactory. Next, the wave response analysis of the same VLFS model located in the variable depth sea has been made. The contour plot of the variable depth configuration is shown in Fig. 3, where the water depth at infinity is assumed to be 100m. This configuration corresponds to the experiment made in Port and Airport Research Institute, Japan (Shiraishi, et al., 2001). The variable depth surface (SB) is discretized into 20,278 elements of the size of 12.5m x 12.5m, and the number of nodes is 61,721. For meshing the VLFS model, the model B (number of nodes is 5,377) in Table 1 is used; thus the total node number of the

S B ). Then, divide the level-0 square cell into four even square cells, and define them to be level-1 cells. Similarly, define level-2 cells by dividing each level-1 cell into four even square cells. This process is repeated until the bottom level cells (leaf cells) include some appropriate number of elements. In the dividing process, the cell that includes no elements is deleted from the definition list of cells.
The method to define "far" cells and "near" cells is made similarly as Greengard and Rokhlin (1987). That is, the cell i that includes the r collocation point x inside and the cells that enclose the cell i adjacent to it are defined as "near" cells; otherwise the cells are categorized to be "far" cells. Note that this categorization into "near" and "far" cells should be made on the same level of the tree structure. In the calculation of Eq. (47), the computation for "near" cells is made for each element (corresponding to the lines 3 and 6 in Eq. (47)). In the actual computation, firstly calculate the integral of Eq. (49) around the center of each element as the multipole expansion point. Then using Eq. (42), the multipole expansion coefficients are converted to the coefficients for the center of a leaf cell. By gathering the coefficients of all elements included in the leaf cell, the multipole expansion coefficient for the leaf cell can be obtained. The calculation of the multipole expansion coefficients for upper level cells can be made similarly, by using Eq. (42) and gathering the coefficients included in the upper level cell. With these procedures, all of the necessary multipole coefficients can be set up, without specifying the r collocation point x . Next, we specify the collocation point x to coincide a nodal point on

310

Table 1 Performance of the fast multipole method and the direct method. Typical panel size Number of nodes Fast multipole method Number of iter. CPU time per iter. CPU time Memory allocation Direct method CPU time LU GMRES factorization solver 1.55 min 3.28 min 28.5 min 36.4 min (907 min) (474 min) (708 hr) (113 hr) (1658 days) (71 days)

Model

Memory allocation 54 MB 489 MB (6 GB) (85 GB) (1.2 TB)

A 25.0m 1,609 31 3 sec 1.8 min 27 MB B 12.5m 5,377 32 15 sec 10.4 min 89 MB C 6.25m 19,393 32 86 sec 77.4 min 315 MB D 3.125m 73,345 32 570 sec 775 min 1.15 GB E 1.5625m 284,929 31 1081 sec* 1689 min* 1.75 GB* * Parallel computation using 6 CPUs; number of modal function is 120.

Table 2 Number of iterations of GMRES and CPU times for various L/. Number of iterations Wave period = 10 3 = 10 4 C 9.57 18 sec 25 32 C 17.9 10 sec 112 148 C 33.2 6 sec 252 497 D 62.0 4 sec 287 D 62.0 4 sec 287 * Parallel computation using 5 CPUs. Model L/ analyzed model becomes 67,098. The computation time was 38.4 hour using 5 CPUs for the wave period of 18 seconds, the wave direction of =/4, and the residual tolerance =10-2. Fig. 4 shows the deflection amplitudes in the variable depth sea and in the constant depth sea (h=8m). We observe a considerable difference of the response characteristics when the VLFS is located in the variable depth sea from those in the constant depth sea, showing the importance to include the effect of variable depth under the VLFS. Fig. 5 shows a snapshot of the surface elevation around the floating body, where the diffraction and the radiation waves are both included. As seen in Fig. 5, the refraction of waves, and the change of wave length (about =157m inside the reef corresponding to 8m water depth) by dispersion are clearly observed. CPU time

= 10 3
1.15hr 3.46hr 8.23hr 63.8hr 14.5hr*

= 10 4
1.29hr 4.40hr 15.7hr

CONCLUSIONS
The accelerated higher order boundary element method using fast multiople algorithm has been presented. The analysis has been made for a VLFS on the variable depth sea configuration as an example to show the effectiveness of the method. The proposed method utilizes the multipole expansion of the Greens function of the series form, introduction of the hierarchical algorithm based on the quadrant-tree, and a modern iterative solver such as GMRES. The theoretical requirements of the memory allocation of O(N) and the computation time of O(N log N) have been demonstrated by the developed program through numerical examples. The large scale hydroelastic problems with about 300,000 unknowns have been demonstrated to be solvable by the proposed accelerated BEM program within about one day, showing the advantage of the proposed method for large scale analysis such as VLFS problems. This study was supported by the Program for Promoting Fundamental Transport Technology Research from the Corporation for Advanced Transport & Technology (CATT).

REFERENCES
Barnes, J, and Hut, P (1986). "A Hierarchical O(N log N) ForceCalculation Algorithm," Nature, Vol 324, pp 446-449.

Barrett, R, et al (1994). Templates for the Solution of Linear Systems: Building Blocks for Iterative Methods, SIAM. Fukui, T, and Katsumoto, J (1997). "Fast Multipole Algorithm for Two Dimensional Helmholz Equation and Its Application to Boundary Element Method," Proc 14th Japan National Symp on Boundary Element Methods, pp 81-86 (in Japanese). Greengard, L, and Rokhlin, V (1987). "A Fast Algorithm for Particle Simulations," J Comput Phys, Vol 73, pp 325-348. Greengard, L (1988). "The Rapid Evaluation of Potential Fields in Particle Systems," MIT Press. Korsmeyer, T, Phillips, J, and White, J (1996). "A Precorrected-FFT Algorithm for Accelerating Surface Wave Problems," Proc 11th Int Workshop on Water Waves and Floating Bodies, Hamburg. Korsmeyer, T, Klemas, T, White, J, and Phillips, J (1999). "Fast Hydrodynamic Analysis of Large Offshore Structures," Proc 9th Int Offshore and Polar Eng Conf, pp 27-34. Kring, D, Korsmeyer, T, Singer, J, and White, J (2000). "Analyzing Mobile Offshore Bases Using Accelerated Boundary Element Methods," Marine Structures, Vol 13, pp 301-313. Linton, CM (1999). "Rapidly Convergent Representations for Green's Functions for Laplace's Equation," Proc R Soc Lond A, Vol 455, pp 1767-1797. Newman, JN (1985). "Algorithms for the Free-surface Green Function," J Eng Maths, Vol 19, pp 57-67. Ohtsubo, H, and Sumi, Y (2000). Proc 14th Int Ship & Offshore Structures Congress (ISSC2000), Vol 1, pp 195-252. Rokhlin, V (1985). "Rapid Solution of Integral Equations of Classical Potential Theory," J Comput Phys, Vol 60, pp 187-207. Shiraishi, S, Harasaki, K, Yoneyama, H, Iijima, K, and Hiraishi, T (2001). "Experimental Study on Elastic Response and Mooring Forces of Very Large Floating Structures Moored inside Reef, " Proc 20th Int Conf on Offshore Mech & Arctic Eng, OMAE01-5203. Teng, B, and Eatock Taylor, R (1995). "New Higher-order Boundary Element Methods for Wave Diffraction/radiation," Applied Ocean Research, Vol 17, pp 71-77. Utsunomiya, T, Watanabe, E, and Eatock Taylor, R (1998). "Wave Response Analysis of a Box-like VLFS Close to a Breakwater," Proc 17th Int Conf on Offshore Mech & Arctic Eng, OMAE98-4331. Utsunomiya, T, Watanabe, E, and Nishimura, N (2001). "Fast Multipole Algorithm for Wave Diffraction/radiation Problems and

311

Its Application to VLFS in Variable Water Depth and Topography," Proc 20th Int Conf on Offshore Mech & Arctic Eng, OMAE01-5202.

(a) =10-3

(a) in constant depth sea (h=8m).

(b) =10-2

(b) in variable depth sea. Fig. 4 Deflection amplitude at T=18sec. Fig. 2 Deflection amplitudes of VLFS in constant depth sea for model B (solid line: FMM, dash-dotted line: direct method).

Fig. 3 Contour plot of the variable depth configuration.

Fig. 5 Snapshot of the surface elevation around the floating body in the variable depth sea.

312