Вы находитесь на странице: 1из 16

Computational Mechanics (1991) 8, 71 86

Computational
Mechanics
9 Springer-Verlag 1991

A boundary element approach for non-linear boundary-value problems


R. Zheng, N. Phan-Thien, and C. J. Coleman*
Department of Mechanical Engineering, The University of Sydney, N.S.W. 2006, Australia

Abstract. A Boundary Element Method has been developed for solving nonlinear boundary-value problems. This method
avoids the tedious calculation of the domain integral contributions to the boundary integral equations. This is achieved by
applying approximate particular solutions which are obtained by expressing the "pseudo-body force" in terms of a linear
combination of radial basis functions. Numerical examples show that the method is efficient and can produce accurate results.

I Introduction

The most successful applications of the Boundary Element Method (BEM) have been in the
solution of linear problems such as potential flows (Laplace equation), linear elasticity (Navier
equations) and viscous creeping flows (Stokes equations). The method replaces the original
problem of solving a set of linear partial differential equations (PDE's) with one of solving a set of
integral equations defined on the boundary, and hence reduces the dimension of the problem by
one. In particular, for the case of an homogeneous PDE, one only needs to consider a boundary
discretization. This is a major advantage of the BEM over other methods requiring full domain
discretization. However, the advantage is generally lost in solving non-linear equations.
When the PDE exhibits some kind of non-linearity, for example, the inertia effect in viscous
flows (Navier-Stokes equations), the BEM can still be used but in an iterative manner (Bush and
Tanner 1983). The non-linear BEM schemes produce linear PDE's at each step of iteration but
these linear PDE's are inhomogeneous in general. It is these equations to which the BEM is
applied. The linearization is achieved by lumping the nonlinearities into a body force term known
as a "pseudo-body force", which is evaluated at the previous iteration. In the conventional
boundary integral formulation, the pseudo-body force produces domain integral contributions
to the boundary integral equations. The domain must then be discretized to allow numerical
evaluation of the domain integral. Although this discretization does not introduce any further
unknowns, the numerical integration process often very much increases the total amount of
computational time.
For some special inhomogeneous PDE's, it is possible to find a particular solution. The
remaining part of the solution will then satisfy a linear homogeneous PDE and hence lead to a set
of boundary integral equations, for instance, potential problems with constant or harmonic source
functions, or elasticity problems with gravitational, centrifugal or thermal loads. Rizzo and Shippy
(1977), Stippes and Rizzo (1977) and Danson (1981) presented a formulation that employs particular
solutions with the aid of the divergence theorem to reduce the domain integrals to boundary
integrals for these problems. However, the shortage of closed form of particular solution make a
general implementation difficult.
More general approaches were developed by Nardini and Brebbia (1982) and Ahmad and
Banerjee (1986) for solving a two-dimensional free vibration problem where a closed form of

* Surveillance Research Laboratory, DSTO, Salisbury, SA, Australia


72 Computational Mechanics 8 (1991)

particular solution is difficult or practically impossible to find. They overcome the restriction by
using global shape functions to aproximate the inhomogeneity. More recently, Banerjee and his
co-workers (Banerjee et al. 1988; Henry and Banerjee 1988a, b; Welson et al. 1990) extended the
technique to thermoelastic stress, elastoplastic and free vibration analyses of two- and three-
dimensional objects, Saigal et al. (1990) treated a thermoelasticity problem using the same idea,
except that the body force was represented by polynomials based on the statistical procedure of
linear regression. Tang (1988) and Azevedo and Brebbia (1988) proposed a different technique
based on expressing the body force in terms of either Fourier series or polynomials and using the
divergence theorem to transfer domain integrals to the boundary. Another approach in terms of
a complex potential was proposed by Coleman (1990), who used Gaussian distributions as basis
functions to construct an approximation to the body force function. The Gaussians in fact belong
to a more general class known as radial basis functions which, in recent years, have become
popular for multidimensional interpolation (Powell 1987). Zheng et al. (1990) carried out a further
study on the use of Gaussians and other radial basis functions in the particular solution technique,
and developed an efficient BEM for solving inhomogeneous potential problems. Zheng et al.'s
method (1990) is similar to the method proposed by Abroad and Banerjee (1986), Henry and
Banerjee (1988a, b), Welson et al. (1990), Saigal et al. (1990), except for the choice of the functional
form of the approximation of the body forces. We prefer the use of radial basis functions as
approximants for the body forces.
The use of the radial basis functions in two- or three-dimensional BEM has several advantages.
The main advantage is that the radial basis function is always a function of a single variable,
independent of the dimension of the problem. Another advantage is that there is no difficulty of
deriving representations in the case of irregular boundaries; this difficulty often occurs when using
Fourier or Chebyshev polynomials. Furthermore, there are some radial basis functions such as
the Gaussians having a decay property that makes the influence of each distribution limited to
points near its maximum and so results in an almost piecewise approximation.
The main purpose of the present work is to extend the method proposed in the previous study
(Zheng et al. 1990) to the problems of Newtonian flows with inertia (Navier-Stokes equations). As
an interim, the Navier equations for elasticity problems with general body forces are considered
first. In the first section of the paper we briefly recall the basic equations governing the elasticity
and viscous flow problems. In subsequent sections we derive the approximate particular solutions
for the governing equations and give details of the numerical treatment and solution scheme used.
Finally we demonstrate the new technique by presenting some numerical examples. The major test
problem chosen is the Hamel flow problem. This is the problem of flow between two converging
plane walls. Since the flow is not unidirectional and hence the inertia terms do not vanish
identically, the problem provides a good test of numerical solution scheme. It was used as a test
example by Gartling et al. (1977) in a finite element convergence study and by Bush and Tanner
(1983) to test the domain integration procedures employed in a BEM. The planar extrusion
problem is then considered next. The results are compared favorably with finite element results.

2 Governing equations

The basic equations of elastostatics, or the Navier equations, can be expressed in terms of the
displacement u as
1 b
V2U -~-] ~ V VV.u + - = O, (1)
- #
where b is the body force, v is the Poisson ratio and # is the shear modulus.
A useful transformation of (1) into a biharmonic equation can be achieved by using the
"Galerkin" form of solution given by

U = V2 G __1 VV'G, (2)


2(1 - v)
R. Zheng et al.: A boundary element approach for nonlinear boundary-value problems 73

where G is the Galerkin vector (see, for example, Little 1973). Substitution of (2) into (1) yields the
following equation:
V2V2G = f , (3)
where
b
f = - -. (4)
#
If biharmonic function G satisfying (3) can be found, then the corresponding displacement can
be found using Eq. (2). This technique will be used in the later section to derive the particular
solution of the governing equation for a given body force function.
The solution of the governing equations of linear elasticity with Poisson's ratio v = 0.5 is also
a solution of the Stokes equations for the creeping flow of an incompressible fluid, if the displace-
ment is interpreted as the velocity, and the shear modulus is interpreted as the shear viscosity.
The full equations describing the motion of an incompressible, viscous Newtonian fluid can be
written as

VZu - 1 - V p + P-t f - u'Vu)= 0, (5)


# #
and
V.u = 0, (6)
where u is the Eulerian velocity vector, f is the body force vector per unit mass, p is the pressure,
p and # are the fluid density and viscosity respectively. Note that we have assumed a steady,
laminar and isothermal flow. The equations are the well-known Navier-Stokes equations.
The pressure can be expressed as
[ 2v/~ ~ "~
p=- lim / ~ v u / . (7)
v-,0.5 \l-zv /
By substituting (7) into (5) an obvious analogy between the Navier-Stokes equation and the Navier
equation can be seen. Thus Eq. (5) can be treated analogously to the standard elastostatic problem,
if the term
p ( f - u" Vu)
is regarded as a pseudo-body force.

3 Solution method

The conventional BEM formation for elasticity or for the fluid flow problems can be written in
the general form:
Couj(x) = S [tj(y)u*(x, y) - uj(y)t*(x, y)] d r ( y ) - S bju*(x, y)dD(y), (8)
F
where D is the region within which a solution is sought, F is the boundary of this region, u* and
t* are fundamental solutions (Brebbia 1980), tj is the traction vector, uj is the displacement vector
(or velocity vector in flow problems), bj represents an actual or a pseudo-body force. Ci~ = 5ij if x
lies in the domain .Q, C 0 = (1/2)5~j for a point on the boundary which is smooth at the point x.
If Eq. (8) is solved numerically by subdividing the boundary F into a number of elements, the
unknown continuous functions uj and tj can be approximated by certain unknown discretized
values on these elements, so that the boundary integral terms in Eq. (8) can be written as a
summation in which only the unknowns on the boundary are involved. For the homogeneous case
( b j - 0), the discretization is required only at the boundary, and this can result in considerable
saving over the equivalent domain solution methods. Obviously, when bj ~ O, the domain integral
74 Computational Mechanics 8 (1991)

in (8) must be calculated over all the domain .(2 and the solutio n procedure becomes rather
cumbersome.
One way of removing the domain integrals is to construct a particular solution to the
inhomogeneous governing differential equations. The remainder of solution will then satisfy a
homogeneous PDE and hence leads to a boundary integral equation only. Let us write the
governing equations given in the previous section in operator notation as
5~(u(r)) = b(r), reI2, (9)
which is subject to a set of boundary conditions
~ i ( u ( r ) ) = tii(r), i = 1, 2,...,N; r on F,
where 5r and ~i(u) are linear operators of the unknown function u and its derivatives. Every
solution of this linear boundary value problem can be written as the sum u = u p + u h, where u p
satisfies
~(uP(r)) = b(r), r612, (10)
and u h satisfies
= O, rcO, = i = 1,2,...,N; r on/-'. (11)
Note that u p, which is a particular solution to (9), does not necessarily satisfy the boundary
conditions of the problem, and that the choice of the particular solution is not unique; whereas
Eq. (11) represents a homogeneous boundary-value problem, and the boundary conditions are
adjusted to ensure that, when u h is added to u p, the total solution satisfies the boundary conditions
of the original problem.
Suppose a particular solution to (9) can be found, the problem will reduce to the solution of
a homogeneous PDE alone, and it can be solved using the standard BEM without the need for
domain integration.
In our previous work (Zheng et al. 1990), we found approximate particular solutions for the
Poisson equation by representing the inhomogeneity in terms of a linear combination of radial
basis functions. The method can be easily extended to elasticity and non-linear flow problems as
well. The major differences are the derivation of the particular solutions corresponding to a given
basis function and the numerical implementation of the particular solutions, which we will
describe in the following sections.

4 Particular solution

We now consider the following inhomogeneous equation

r1~ r~-r r drr r~-r --0(r), (12)

in which the inhomogeneous term 0 is a radial function. When ~ = exp ( - r 2) a particular solution
of (12) is given by
q~(r) = ~6 {(1 + r2)[ln (r 2) + El(r2)] -- r 2 -- exp ( - r2)}. (13)
It should be noted that by adding any null solution of (12) to the expression of 4) we still have
a particular solution. This is useful in some cases to eliminate singularities which may appear in
some expressions of q~ and its derivatives.
We can now approximate the body force term in Eq. (3) in terms of the radial function ~:
N
fi(r) ~ ~ ch,O ( l r - r,l/fl.), (14)
n=l
where r, and ft, are suitably chosen constants. Letting ~, = Iv - r, [/fl,, we can show that a particular
R. Zheng et al.: A boundary element approach for nonlinear boundary-value problems 75

solution corresponding to (14) for the Galerkin vector in (3) will be given by
N
G,(r): y~ ~,./~.~(~.). (15)
n=l
For brevity, we shall use only one term to represent the right-hand side of (15), i.e., we rewrite (15) as
G~(r) = ~,fl4 0(f ). (16)
Derivatives of G i can then be evaluated as follows:

(19, 20)

Gk,jk=~kfl{l(f,6jk+fj6k~+fk6ij)[__~ff),(f)+~dp,,(f)] fifjfk [- 3 3,,

(21)
where the summation convention is adopted, and the comma denotes the partial derivative of a
variable with respect to a Cartesian coordinate, e.g., Gi,j = ~Gi/~xj, i, j = 1, 2.
Substituting these expressions into Eq. (2) yields the desired particular solution for displace-
ment (or velocity in fluid flow problems)
1

The particular solution for the stress will be

tiP--__ -f D:o~kr2(~iJ-~- (1-- V)(O~irJ-J- ~Jri) ] ~2)l -- ~k (yiC~Jk-~- rJ(~ki "Jr fk(~iJ) O~kfif (~)3 , (23)

where

9 1=- ~'(~)+~ 1 4,"(~) + 0~"(~), oh -


1 1 .
~2 ~'(~) + =r~ (~)'
3
oh = ~ ~'(e) - q~"(e)+ r
(24-26)
The particular solution for traction acting on boundary is found using the relation
t~' = ti~nj, (27)
where nj is the unit normal.
In incompressible fluid flow cases we simply set v = 0.5.

5 Numerical implementation

The numerical implementation of the desoribed method mainly consists of two parts. First, the
boundary integral equations representing the homogeneous elasticity or the Stokes flow are
discretized in the usual manner (Brebbia 1980) and the detail will not be repeated here. Second, in
order to compute the particular solutions we need to fit the pseudo-body force into the given
model, i.e., Eq. (14).
In Eq. (14) the parameters ~i., ft,, and r, are to be determined. This is essentially a nonlinear
data-fitting problem. Although some standard nonlinear data-fitting techniques are available
(Stoer and Bulirsh 1976), the calculations can generally be performed only by iterations and
therefore can be time consuming. Alternatively, we can choose//, and r, a priori, while leave ~i, to
76 Computational Mechanics 8 (1991)

be determined using a linear data-fitting scheme. There is a fair degree of freedom in choice of
parameters r,. One could place them on a regular grid for simplicity of data preparation, or,
concentrate them in the region where the function has its most rapid variation for economy.
Constant ft. must be carefully chosen since too small a value of ft, will result in an approximation
of isolated peaks and too large a value of ft, will make the approximation procedure ill-conditioned.
In practice, it was found that [3, should be of the same order as the distance of the closest neighbours
to r n.
To find ~i,, one can use either a collocation method or a least-squares fitting method. Both of
these methods result in a set of linear algebraic equations that determine the unknowns ~i,
(i = 1, 2, n = 1, 2,..., N). The use of localized basis function such as r = exp ( - r E) makes the system
matrix effectively sparse and suitable for using an iterative solution method such as the conjugate
gradient method (CGD) (Stoer and Bulirsh 1976) which is employed here. We have also employed
a singular value decomposition (SVD) method (Stoer and Bulirsh 1976). The SVD is a direct
method and can be quite slow when solving a large system. However, this method is capable of
dealing with singular systems of equations and hence is powerful for solving most linear least
squares problems.
Once the parameters have been determined, the particular solutions can be evaluated. This
involves an evaluation of the exponential integral which is defined by

E I ( X ) = ~ exp ( - t) dr. (28)


x t
In our computer code E1 is evaluated approximately based on the polynomial (when 0 < X < 1)
and rational (when 1 < X < ~ ) approximations reported in Abramowitz and Stegun (1972).
Caution is needed when X = 0 since at this point E~ has a logarithmic singularity. A way round
this problem is simply to calculate the sum of (ln X + E~ (X)) instead of E~ (X) alone. The reason
for this can be seen more clearly by expanding El (X) into a series as follows

Ea(X)=-7-1nX- ~ (-1)"X" (29)


n=t F/'n! '
where y = 0.577216--. is the Euler constant. The logarithmic term will be cancelled when a In X
term is added in, and the value of (lnX + E l ( X ) ) is just what we want in the expression of ~b(r)
(Eq. 13).
The computations of the particular solutions also involve evaluating the following derivatives
of q~(r)

~'(r)=~
1{ [ r l n r 2+El(r2)]q 1-exp(-r2)}
r qS"(r) = ~ In r e + E t ( r 2 ) +
2r z + exp ( - r 2) - 1 -I
/.2 J
(30, 31)
r E -- 2r z exp(-- r 2) -- exp(-- r 2) + 1
~b'"(r) - 4r 3 (32)

When r becomes small or equal to zero, the above equations become numerically inaccurate or
inapplicable. In this case the Taylor expansions of the above expressions are used to circumvent
the problem.
An iterative scheme must be applied to solve nonlinear problems. The iterative scheme
employed in this work consists of the following steps:
1. Solve the homogeneous boundary integral equations with the boundary conditions given by
the problem, evaluating the homogeneous solutions u h and t h o n the boundary.
2. With the boundary values obtained in the previous step and those prescribed as boundary
conditions, compute the velocity field at the internal nodal points by using the boundary integral
equations again.
3. Compute the pseudo-body force using the results obtained in the previous step.
R. Zheng et al.: A boundary element approach for nonlinear boundary-value problems 77

4. Fit the pseudo-body force data into the given model, finding the appropriate coefficients for the
model.
5. Evaluate the particular solutions u p at both internal and boundary nodal points, and t p at
boundary nodal points.
6. Update the boundary conditions, i.e., subtract the boundary values of the particular solutions
from the original boundary conditions of the problem.
7. With the updated boundary conditions, resolve the homogeneous boundary integral equations
for uh and t h, and recompute velocity field in the domain.
8. Add the homogeneous solution to the particular solution to obtain the total solution.
9. Check the convergence of the results. If the results are not convergent, repeat the previous steps
from step 3.
The program is coded in Fortran 77 and it was run on a MIPS-120 computer.

6 Numerical examples

6.1 Elasticity problem

The simple test problem investigated here is a linear elastic deformation problem with a general
type of distributed loading. This problem does not require any iteration, so that it provides a quick
test of the numerical procedure of computing the particular solution. In this example, both body
and surface forces are applied to an elastic body of square cross section and the problem is assumed
to be in plane strain. The governing equations are

02ux ~2u,, 1 0 (Ou,, Ou,'~+bx=o, (33)


2 -) #
e2u, ~?2u, 1 e (Ou,, ~?u,'~+b,= 0 {34)
Ox2 + 0 ~ --t 1--2v e y \ O x + ~ - y ) .
defined in
O<_x<_L, O<y<L.
The body forces are given by
bx ~ ( 3 - 4 v ) . nx . ny tl nx ny
- --L sm ~ - sm ~ , : by= - ~- cos ~ - cos ~-, (35,36)

and the boundary conditions are


ux = O, ty = O, o n E l : 0--<X--<L, y = O;
ux = 0, ty =/t, o n F2: x < L, 0 < y < L;
ux=0, ty=0, o n F 3" O<x<L, y=L; (37)
ux=0, uy=0, o n F 4 : x=0, O<y<L;
where t i represents the surface traction.
Analytical solutions to this problem for displacements and stresses can be readily obtained, i.e.,
( 1 - 2 v ) L . n x . ny
ux = n2 sin ~ - sm ~-, uy = x, (38, 39)

2or(l-v) nx . ny // 1-2v nx ny'~ nx . ny


Gxx -- n cos-L-SmL-, axy=/t~l+ n sin T c o s T ) , %y = 2#v
n cosTsmT.
(40-42)
For the numerical calculations, we set/l = 1, v = 0.3 and L = 1. Four different uniform meshes,
78 Computational Mechanics 8 (1991)

Table 1. The displacements ux at internal points

ux values at internal points

Grid spacing (0.25, 0.25) (0.25, 0.5) (0.5, 0.25) (0.5, 0.5)

0.25 0.2254207E- 1 0.2892627E- 1 0.3146474E- 1 0.4083337E- 1


(11.2~o) (0.936%) (9.79~) (0.752%)
0.125 0.2126844E- 1 0.2869038E- 1 0.3004501E- 1 0.4055438E- 1
(4.95%) (0.113~o) (4.84~) (0.064%)
0.0625 0.2070199E - 1 0.2863992E- 1 0.2930094E- 1 0.4051669E- 1
(2.16~o) (0.063%) (2.24%) (0.029Vo)
0.0500 0.2060512E- 1 0.2863644E- 1 0.291647tE- 1 0.4052391E- 1
(1.68~) (0.045~o) (1.77~o) (0.011%)

Exact 0.202642E- 1 0.286579E- 1 0.286579E- 1 0.405284E- 1

Table 2. The displacements uy at internal points

uy values at internal points

Grid spacing (0.25, 0.25) (0.25, 0.5) (0.5, 0.25) (0.5, 0.5)

0.25 0.2497096E0 0.2492517E0 0.4999980E0 0.4999950E0


(0.116%) (0.299~o) (0.0004~o) (0.001 ~o)
0.125 0.2497461E0 0.2495649E0 0.5000020E0 0.5000033E0
(0.102%) (0.174~o) (0.0004~o) (0.0007~)
0.0625 0.2498317E0 0.2497578E0 0.5000010E0 0.4999980E0
(0.067~o) (0.097%) (0.0002~o) (0.0004~o)
0.05 0.2498974E0 0.2497996E0 0.4999997E0 0.4999996E0
(0.041%) (0.080%) (0.00006%) (0.00008%)

Exact 0.25 0.25 0.5 0.5

Table 3. The stresses ax~ at internal points

a~x values at internal points

Grid spacing (0.25, 0.25) (0.25, 0.5). (0.5, 0.25) (0.5, 0.5)

0.25 0.2337669E0 0.3141703E0 0.8897443E- 5 0.6707008E- 5


(4.91%) (0.298~o)
0.125 0.2309604E0 0.3144996E0 0.1984613E- 5 0.4480759E- 5
(3.65%) (0.194%)
0.0625 0.2271909E0 0.3151509E0 0.8579725E-6 0.2685341E- 5
(1.96~o) (0.013%)
0.05 0.2263626E0 0.3151265E0 0.2718766E - 6 0.6080299E - 6
(1.59~o) (0.005%)

Exact 0.2228169E0 0.3151107 E0 0 0

Table 4. The stresses crxyat internal points

crxyvalues at internal points

Grid spacing (0.25, 0.25) (0.25, 0.5). (0.5, 0.25) (0.5, 0.5)

0.25 0.1056761E+ 1 0.9962829E0 0.1093624E + l 0.9962829E0


(0.649%) (0.372%) (0.330%) (0.372~o)
0.125 0.1057853E+ 1 0.9981886E0 0.1083169E+ 1 0.9978178E0
(0.546%) (0.181 ~o) (0.630%) (0.218~o)
0.0625 0.1061312E+ 1 0.9992106E0 0.1086839E+ 1 0.9992585E0
(0.221%) (0.078%) (0.293%) (0.074%)
0.5 0.1061845E + 1 0.9993890E0 0.1087488E + 1 0.9994526E0
(0.171~o) (0.061~o) (0.233%) (0.055~o)

Exact 0.1063662E + 1 0.1000000E + 1 0.1090032E + l 0.1000000E + 1


R. Zheng et al.: A boundary element approach for nonlinear boundary-value problems 79

Table 5. The stresses ayy at internal points

ayr values at internal points

Grid spacing (0.25, 0.25) (0.25, 0.5) (0.5, 0.25) (0.5, 0.5)
0.25 0.9409946E- 1 0.1380089E0 0.1585108E-5 0.3286759E-5
(1.46~) (2.19~o)
0.125 0.9442429E- 1 0.1360373E0 0.8831925E-6 0.5719154E- 6
(1.12~o) (0.733~o)
0.0625 0.9500687E- 1 0.1351251E0 0.2722053E- 6 0.1411327E-5
(0.509~) (0.058~)
0.05 0.9517094E- 1 0.1351054E0 0.1439845E- 5 02472502E- 5
(0.372~) (0.043~)

Exact 0.954929E-1 0.1350474E0 0 0

Table 6. The displacements ux obtained using a BEM with domain integrals


u~ values at internal points

Grid spacing (0.25, 0.25) (0.25, 0.5) (0.5, 0.25) (0.5, 0.5)
0.0625 0.20578E- 1 0.28482E- 1 0.29116E- 1 0.40273E- 1
(1.55Vo) (0.614~o) (1.60~0) (0.630~o)
0.0500 0.20531E - 1 0.28544E- 1 0.29052E- 1 0.40360E- 1
(1.32~) (0.397~o) (1.38~o) (0.415Vo)

Table 7. A comparison of the CPU time (see) between the present method and
the method with domain integrals
CPU time (s)
Grid spacing The present method Domain int. method
0.0625 77.78 127.87
0.0500 168.11 279.07

i.e., 5 x 5,9 x 9, 17 x 17 and 21 x 21, are used, thus the c o r r e s p o n d i n g grid spacing are 0.25, 0.125,
0.0625 and 0.05, respectively. T h e results are listed in Tables 1 to 5, in which displacements and
stresses at several typical internal points are presented, with the relative errors given in the brackets.
F o r the p u r p o s e of c o m p a r i s o n we also solved the same p r o b l e m using the d o m a i n integral
m e t h o d . T h e 17 x 17 a n d 21 x 21 meshes are used. T h e results for u x are s h o w n in T a b l e 6. A
c o m p a r i s o n of C P U time is given in T a b l e 7. O n e can see that the present m e t h o d is able to save
a b o u t 40}/0 C P U time, c o m p a r e d to the m e t h o d using d o m a i n integrals.

6.2 Hamel flow problem

A steady-state t w o - d i m e n s i o n a l flow in the region between two infinite flat plate set at an angle,
as s h o w n in Fig. 1, is k n o w n as the H a m e l problem. C o m p r e h e n s i v e discussions of the H a m e l
p r o b l e m m a y be f o u n d in R o s e n h e a d (1963) and B a t c h e l o r (1970). T o describe this problem, p o l a r
c o o r d i n a t e s (r, O) are used, with 0 = T c~ at the two plane walls, where r is the distance f r o m the
intersection of the two walls and 0 is m e a s u r e d f r o m centreline. T h e Reynolds n u m b e r for the flow
is defined by

Re - ~u~
# (43)
80 Computational Mechanics 8 (1991)

where/~ is the viscosity, p is the density, c~is the wedge half-angle and uo the velocity along the
centreline at the radial distance r. The exact solution of the problem is based on the assumption
of purely radial flow which yields self-similar velocity profiles at all radii. The product of u0 and r
is constant, and hence the Reynolds number is constant throughout the flow field.
It is convenient to introduce the dimensionless variables
r1 = O/a, f = ur/Uo, (44)
where u, is the radial velocity.
The analysis of Batchelor (1970) leads to the following ordinate differential equation
(f,)2 -~ (1 --/){~a Re ( f 2 _1_f ) + 4 a 2 f + c}, (45)
where c is a constant and the primes denote differentiation with respect to r/.
For the calculations presented in this section the flow is taken to be inflow (towards the origin).
In this case Re < 0 according to the definition. The result of further integration of (45) can be
written as

1
s df (46)
=~(lo - f ) l / z { ~ a R e ( f 2 + f ) + 4 a 2 f + c } 1/2'
and the boundary condition requires

1
[1 df (47)
(1 -- f)1/2 {2a Re (f2 + f ) + 4azf + c} 1/2"
For a given Re, the constant c is determined by (47). The dimensionless velocity profiles can then
be computed from (46). This provides the "exact" solutions which are used in this work for
comparison with the numerical results.
At zero Reynolds number, a close form solution is available, that is,
ur cos(20)- cos(2a)
-- = (28)
Uo 1 - cos (2a)
At large Reynolds number, one has the following asymptotic approximation (Batchelor 1970):

u-L=3 tanh 2 ~ [ - ~a Re] 1- + tanh -1 - 2. (49)


Uo
The centreline velocity uo can be evaluated for a given flux Q using
1
Q = ar S (50)
-1
For the BEM simulation we chose a = 30 ~ # = 1, and the Reynolds number is varied by
changing p. The section of the wedge consider in the numerical calculations is the region between
r = 0.25 and r = 4. We use two meshes as shown in Figs. 2a, b for examining the effect of mesh
refinement. The coarse mesh was designed to have the same element size as the one used in Bush
and Tanner (1983). The meshes represent a half wedge. The horizontal side is the symmetry
plane where the symmetry boundary conditions are prescribed. No-slip velocity conditions are
prescribed on the plate wall. The boundary conditions on the upstream and downstream ends of
the wedge are not known exactly and approximations are needed. On the upstream end the velocity
boundary conditions corresponding to zero Reynolds number is imposed for Re < 21 while the
asymptotic solution corresponding to large Reynolds number is used to obtain approximate
boundary conditions for higher Re cases. Incorrect zero traction boundary conditions are applied
on the downstream end; but its influence region is small compared to the overall flow domain.
In order to observe the iterative convergence of the solution algorithms and the convergence
of the solution with mesh refinement, we have plotted in Fig. 3 the radial velocity Uo at the centreline
and r = 1 as a function of iteration number for both two meshes. Iteration zero is the creeping
(Re = 0) solution. Exact solution and the result of Bush and Tanner (1983) are also given for
R. Zheng et al.: A boundary element approach for nonlinear boundary-value problems 81

Centreline Fig. 1. Hamel flow problem

a b

Fig. 2 a and b. Coarse mesh for Hamel flow problem; b finer mesh for Hamel flow problem

Re = 10.9 9 This work (Coarse mesh)


, 2.9
* This work (Fine mesh)
o 2.8 * Bush & Tanner
~--- Exact . . . .
_~ 2.7
>
..... ! ..... ! .....
~ 2.6
E
r 2.5
o

2.4
o ~ ;, ; ~ 10 Fig. 3. Convergence behaviour of centreline velocity u o at
Iter. No. r=l

1.1
1
09
0.8
0.7
il

0.6
0.5
0.4 . ~ , o 0 \\\
0.3 , ~.4o \\\
0.2 - - Exact "X~
0.1
0 i J i t 1 i J i i I i i

4 8 12 16 20 24 28
Angle (Degrees) Fig. 4. Profile of u~/u o at r = 1
82 Computational Mechanics 8 (1991)

10
9
v Uo (Coarse mesh)
8 Re=29.8 + Uo (Fine mesh)
7 x Uor (Coarse mesh)
6 UEOxra~Finemesh)
5
0
4
3
2
1
0 r

0 1 2 3 4 Fig. 5. Radial variation of centreline velocity uo and the


Radial distance product Uor

Table 8. A comparison of the CPU time (sec) between the


present method and the method with domain integrals

The present method


Domain int.
Mesh SVD CGM method

Coarse 58.6 24.1 54.2


Fine 5389.0 1017.8 1978.3

comparison. This figure indicates that eight iterations yield a well converged solutions with both
meshes at Re = 10.9. It also shows that the finer mesh provides a better approximation to the true
value. At higher Reynolds number (Re > 30) the coarse mesh is no longer sufficient to provide an
adequate approximation to the velocity field because of the velocity boundary layer near the wall
of thickness O(x/~/Re ). We found that convergence becomes slower as Re increases. It was difficult
to get convergent solution on the fine mesh when Re > 40.
The validity of the numerical results was tested by comparing with the exact solutions and
some typical comparisons are presented below. Figure 4 exhibits the ratio of the radial velocity ur
to the centreline velocity u 0 at r = 1 as a function of circumferential angle 0 for different Reynolds
numbers. The results plotted here were obtained using the fine mesh. It can be seen that up to
Re = 40 the numerical results are on the whole close to the exact solution. In Fig. 5 we show the
radial variation of the centreline velocity Uo and the product Uor at Re = 29.8 for different
meshes. Again one can see that the numerical and the theoretical solutions are in good agreement.
To compare the efficiency between the present method and the conventional BEM, the
computational time required for 10 iterations is listed in Table 8. Two optional solvers (i.e. the
singular value decomposition (SVD) and the conjugate gradient method (CGM)) employed in the
present technique are also compared. We see that the CPU time required by the present method
when the CGM is used is about 40~ less than the time required by the conventional domain
integral method.

6.3 Plane jet problem

We now apply the numerical method to a plane Newtonian jet flow with non-zero Reynolds
number. This problem is more complicated than the Hamel problem and is analytically interactable
mainly due to the presence of the free surface. The free jet problem has been quite popular in the
field of computational fluid mechanics and there are many numerical solutions available (Tanner
1988). In particular, a finite element program developed by Nickell et al. (1974) has been successfully
used for both the axisymmetric (Gartling et el. 1977) and plane (Omodei 1979)jet problems. We
R. Zheng et al.: A boundary element approach for nonlinear boundary-value problems 83

also have some FEM data provided recently by Beverly (1990). These enable us to assess the
performance of our numerical method by comparing its results with the existing numerical data.
The particular feature of the jet flow that is examined here is the free surface shape and the
die-swell ratio defined by g - hi/ho where ho is the half-width of the channel and hI the final
half-width of the jet. Gravitational force and the surface tension are neglected just for simplicity.
The inclusion of a gravitational force is clearly straightforward; surface tension also can be included
in the boundary element formulation in the manner of Zheng and Tanner (1989) if required. When
these factors are ignored, a dimensional analysis shows that the swelling ratio Z is a function of the
Reynolds number only (Tanner 1988).
The boundary conditions applied correspond to fully developed Poiseuille flow ux =
U[1-(y/ho)2], ur = 0 at a distance 4ho upstream of the exit plane, no-slip velocity conditions on
the wall, and zero normal and tangential traction conditions on the free surface and the outlet
boundary at a distance 6h 0 downstream of the exit plane. It has been pointed out by Tanner (1988)
that an upstream and downstream field of about 2.5 - 3ho is already safe for a plane creeping flow.
However, for the flow with non-zero Reynolds number, one would expect that the suitable field
length will be greater than the creeping flow case, depending on the values of the Reynolds number.
On the free surface a zero normal velocity boundary condition also has to be satisfied. This
kinematic boundary condition is used for updating the free surface. The starting profile of the free
surface was that of a plug flow, hi(x ) = ho throughout. A final mesh for Re = 0 is shown in Fig. 6.
The Reynolds number is defined as
pUho
Re - (51)
/t
We have noted that Omodei (1979) defined Re differently. His results cited for comparison have
been converted according to above definition of Re.
A comparison of the computed swell ratios obtained for Reynolds numbers ranging from 0 to
13 is shown in Fig. 7. The agreement between the present results and the finite-element predictions
(Omoidei 1979; Beverly 1990) is good. For a comparison of the free surface shapes refer to Fig. 8.
Again we see that the results between the different numerical methods are in agreement. We found
that when a downstream field length of less than 4h0 is not adequate for Reynolds number greater
than 13. In this case, the jet shows a slight expansion near the exit and a slight contraction
downstream. It is most likely that the boundary conditions cannot be accommodated when Re is
large for a short downstream field length. Convergence becomes slower as Re increases, and when

. . . . . . . . . . . .
I I I I I I I I J
ll',J',
[ [ F I I I I I I I IIIHIII IIIIII I I I I Fig. 6. Mesh for planejet problem

1.26
1.24
1.22 , This work
1.2 --~____~ • gEM (Beverly.1990)
1.18
1.16
o 1.14

~ 1.08
co 1.06
1.04
1.02
1
0.98
0.96
0.94

Re Fig. 7. Swelling ratios Z as a function of Reynolds number Re


84 Computational Mechanics 8 (1991)

1.17
1.16 o...~ ................................
1.15
1.14
1.13 Re = 4 , , , ' / " ~ . , .
1.12 ,,,~ This Work
1.11
1.1 t f ,~ _,.,.~ ..... FEM (Beverly 1990)
1.09
1.08
>- 1.07
1.06
1.05
1.04
1.03
1.02
1.01 Re = 13 "
1
0.99
0.98
o 6
X Fig. 8. Free jet shapes

80 Level Vx
5 1.64
70 4 1.27
3 0.91
60-
Re=0 2 0.55
1 0.18
5 0 ~2 1 ~3g~-'~__

4.0

3.0 ] R e = 13
20

1.0 ~l-V--=~___~_____~
5
i T i I
-2.5 0 2.5 5O Fig. 9. Contours of velocity ux

8.0 Level Vy
5 0.28
70 4 0.22
3 0.16
6.0 Re=0 2 0.09
1 0 03
5.0

4.0

3.0

2.0 R e = 13

1.0

0 '
-2 5 o 2.5 5.0 Fig. 10. Contours of velocity u r

Re > 16 with the present mesh, the solution for the free jet shape becomes unstable (oscillatory in
iterations).
Finally, in Figs. 9 and 10 typical velocity contours are shown for Re = 0 and Re = 13. For both
cases the velocity distribution in the channel does not change greatly until the fluid approaches
the exit of the channel. When the fluid leaves the channel, in the inertia-less case (Re = 0), the fluid
rapidly approaches a uniform flow; while in the case of Re = 13, the transition region extends
R. Zheng et al.: A boundary element approach for nonlinear boundary-value problems 85

further downstream. The predicted behaviour is physically reasonable. Thus the inertia effects have
been successfully simulated in the numerical analysis.

7 Final remarks

In the present work, a BEM based on approximate particular solutions has been implemented for
solving nonlinear boundary-value problems. The particular solution satisfies the inhomogeneous
PDE and therefore results in a set of integral equations involving boundary values alone. The
technique involves approximating the inhomogeneity in terms of a linear combination of radial
basis functions for which particular solutions are known. The adjustable parameters in the
expression are found by means of collocation or least squares fitting. The method is computationally
efficient by employing the Gaussian distributions as the basis function and using the conjugate
gradient method for the fitting process. The computational time for two-dimensional analyses of
elasticity and fluid flow problems required by the present method is about 40% less than that of
domain integral formulation. Results of example elasticity and fluid flow problems are in agreement
with existing results.
The present technique described in this paper is completely general. It can be used to other
non-linear problems such as non-Newtonian flow problems, provided the nonlinearity in the
problem can be lumped into the body force term. Three-dimensional cases are being investigated.

Acknowledgements

We are grateful to Professor R. I. Tanner for many fruitful discussions. We wish to thank Mr. C. Beverly for supplying the
finite element results for the planar extrusion problem. This research is funded by an Australian Research Council Grant. The
support is gratefully acknowledged. R. Zheng also wishes to thank the University of Sydney for a Postgraduate Research Award
(UPRA).

References

Abramowitz, M.; Stegun, A. (1972): Handbook of mathematical functions, New York: Dover Publications, Inc.
Ahmad, S.: Banerjee, P. K. (1986): Free vibration analysis by BEM using particular integrals. J. Eng. Mech. ASCE, 112, 682 695
Azevedo, J. P. S.; Brebbia, C. A. (1988): An efficient technique for reducing domain integral to the boundary. Proc. of the X int.
Conf. on BEM in Eng. 347-361
Bachelor, G. K. (1967): An introduction to fluid dynamics, Cambridge University Press, Great Britain
Banerjee, P. K.; Ahmad, S.; Wang, H. C. (1988): A new BEM formulation for the acoustic eigenfrequency analysis. Int. J. Num.
Meth. Engng. 26, 1299-1309
Beverly, C. (1990): Private communication
Brebbia, C. A. (ed) (1980): The boundary element method for engineers, 2nd ed., Pentech Press, Great Britain
Bush, M. B.; Tanner, R. I. (1983): Numerical solution of viscous flows using integral equation methods. Int. J. Num. Meth.
Fluids. 3, 71 92
Coleman, C. J. (1990): A boundary element approach to some nonlinear equations from fluid mechanics. To appear in
Computational Mechanics.
Danson, D. J. (1981): A boundary element formulation of problems in linear isotropic elasticity with boday forces. In: Brebbia,
C. A. (Ed): Boundary element methods. Berlin, Heidelberg, New York: Springer
Gartling, D. K.; Nickell, R. E.; Tanner, R. I. (1977): A finite element convergence study for accelerating flow problems. Int. J.
Num. Meth. Eng. 11, 1155-1174
Herry, D. P. Jr.; Banerjee, P. K. (1988a): A new boundary element formulation for two- and three-dimensional thermoelasticity
using particular integrals. Int. J. Numer. Meth. Engng, 26, 2061 2078
Herry, D. P. Jr.; Banerjee, P. K. (1988b): A new BEM formulation for two- and three-dimensional elastoplasticity using
particular integrals. Int. J. Numer. Meth. Engng, 26, 2079-2098
Little, R. (1973): Elasticity. Prentice-Hall, U.S.A.
Nardini, D.: Brebbia, C. A. (1982): A new approach to free viberation analysis using boundary elements. In: Brebbia, C. A. (ed):
Boundary element methods in engineering. Berlin, Heidelberg, New York: Springer
Nickell, R. E.; Tanner, R. I.; Caswell, B. (1974): The solution of viscous incompressible jet and free-surface flows using finite
element methods. J. Fluid Mech. 65, 189 206
Omodei, B. J. (1979): Computer solution of a plane Newtonian jet with surface tension. Computer and Fluids 7. 79-96
Powell, M. J. D. (1987): Radial basis functions for multivariate interpolation. In: Mason, J. C.; Cox, M. G. (eds): Algorithms for
approximation. Oxford: Clarendon Press
86 Computational Mechanics 8 (1991)

Rizzo, F. J.; Shippy, D. J. (1977): An advanced boundary integral equation method for three-dimensional thermoelasticity, Int.
J. Num. Meth. Engng. 11, 1753-1768
Rosenhea& L. (ed) (1963): Laminar boundary layers. Oxford University Press, Great Britain
Saigal, S.; Gupta, A.; Cheng, J. (1990): Stepwlse linear regression particular integrals for uncoupled thermoelasticity with
boundary elements. Int. J. Solids Struct. 26, 471-482
Stippes, M.; Rizzo, F. J. (1977): A note on the body force integral of classical elastostatics. ZAMP, 28, 339-341
Stoer, J.; Bulirsch, R. (1976): Introduction to numerical analysis. Berlin, Heidelberg, New York: Springer
Tang, W. (1988): Transforming domain into boundary integrals in BEM, vol. 35, New York: Springer-Verlag
Tanner, R. I. (1988): Engineering rheology. Revised paperback edition. Oxford: Clarendon Press
Wilson, R. B.; Miller, N. M.; Banerjee, P. K. (1990): Free-vibration analysis of three-dimensional solids by BEM. Int. J. Numer,
Meth. Engng. 29, 1737-1757
Zheng, R.; Coleman, C. J.; Phan-Thien, N. (1990): A boundary element approach for non-homogeneous potential problems. To
appear in Computational Mechanics
Zheng, R.; Tanner, R. L. (1989): A boundary element method for free surface flow problems. In: Hogarth, W. L.; Noye, B. J.
(eds): Computational techniques and applications: CTAC-89. Hemisphere Publishing Corporation, New York

Communicated by S. N. Atluri, December 7, 1990

Вам также может понравиться