Compurers & Strucrures Vol. 65, No. 5, pp.

71S7719, 1997
0 1997 Elsevier Science Ltd. All rights reserved
Pergamon Printed in Great Britain
Pn: SO@&7949(96)00322-7 00457949/97 917.00 + 0.00

SHAPE DESIGN SENSITIVITY ANALYSIS OF ACOUSTIC

PROIBLEMS USING A BOUNDARY ELEMENT METHOD
B. U. Koa
Department of Mechanical Engineering, Korea Advanced Institute of Science and Technology,
373-1, Kuseong, Yuseong, Taejon, 305-701, South Korea

(Received 16 June 1995)

Abstract-A boundary integral equation for shape design sensitivity analysis of acoustic problems is
presented. In the derivation of the sensitivity equation, the material derivative concept is used. To remove
the singularntiess that appear in the conventional sensitivity equation, an integral identity which represents
the one-dimensional wave propagation is used. Since the sensitivity equation is regularized, one can obtain
the accurate solution using the standard Gaussian scheme. The equation is verified for the pulsating sphere
example where an analytical solution is available. The boundary was discretized by a six-node triangular
element. As a result, one can obtain the accurate sensitivities of acoustic pressure not only on the boundary
but also at the given field points. 0 1997 Elsevier Science Ltd.

INTRODUCHON of the finite difference interval on the convergence

is in question. The direct differentiation of the
There is an important class of structural design boundary integral equation is an effective alternative
problems in which the shape of a structural system is to the finite difference method. Kane and Saigal[4]
to be determined, subject to some constraints. During computed the sensitivities by differentiating the
the last few years, there has been considerable system matrix anaiytically in the elastic problem, but
research activity in the area of computational some singular behaviours may have appeared in the
methods for shape optimization. The goal of the integral equation by analitic differentiation. To avoid
shape optimization is to find design parameters which these singular behaviours, they placed the source
define the shape of the structure. The shape design point outside of the region. Barone and Yang [5,6]
sensitivity which represents the rate of change of located the source point on the boundary to avoid the
object function or constraint values is important ambiguity of Kane’s method and introduced the
information to find the optimum shape. If the recovery process by using rigid body conditions to
sensitivity of a given configuration to the design reduce the singularities.
variables can be determined, iterative process gives In the acoustic problem, the above idea may be
usually more improved design to us. Haug et nl. [I] used to obtain desired sensitivities with a combi-
provided a general method for computation of design nation of potential problems whose fundamental
sensitivity. solution behaves in a similar way to the acoustic
In the numerical implementation of a sensitivity one, but the computational cost may be increased
analysis, either the finite element method or the due to the combination of potential problems and
boundary element method can be used. However, the the method which uses the acoustic equation
boundary element method has some advantages only, has not been announced in the author’s
compared to the finite element method such as knowledge.
accuracy on the boundary and easiness in remesh In this paper, an analytical expression for design
process. Especially, the boundary element method sensitivity of acoustic problems is to be presented by
can easily treat the infinite region which usually differentiating the conventional boundary integral
occurs in acoustic problem. equation and using the material derivative concept.
Some authors have previously investigated the To remove the singularities of integrands in the
computation of design sensitivity for acoustic integral representation, an integral identity which
problems. Cunefare and Koopmann [2] studied the represents one-dimensional wave propagation is
sensitivity of acoustic radiation power to the change used. Finally, one can obtain the regular represen-
of acoustic velocity for a given configuration. Smith tation of sensitivity equation. The equation is verified
and Bernhard [3] compute the sensitivity by differen- for a pulsating sphere example where an analytical
tiating the discretized boundary element equation. In solution is available. One can find that the obtained
this case, the derivative of the system matrix was sensitivity is accurate not only on the boundary but
approximated by a finite difference. Although they also at the given field points by using the derived
were able to obtain the acceptable solutions, the effect equation.

713
714 B. U Koo

MATERIAL DERIVATIVES
where G(r, x) and F(<, x) are the fundamental
solutions of the Helmholtz equation, j = ,/ - 1, O.Iis
Consider a domain 9, shown schematically in
the vibration frequency, p. is the equilibrium density
Fig. 1. Suppose that only one parameter r defines the
of the fluid and v. is the outward normal component
transformation T, as shown in Fig. 1. The mapping
of surface velocity. The fundamental solutions
T: x+x~(x), x ESZ, is given by
G(<, x) and F(r, x) are a well known singular
function such that
x, = T(x, T). (1)

The above mapping may be viewed as a dynamic Jtii' (kr)/4,

process of deforming a continuum, with 7 playing the G(t, xl = exp( -jkr)/4nr,
role of time. Thinking of 7 as time, a design velocity
can be defined as

V(Xc,T)=~=
dx, dT(x,
a7
7)
(2)
F((, x) = w
since the initial point x does not depend on 7. -$#‘(kr)& in 2D
Ignoring higher-order terms, xr can be written as = (5b)
1 +jkr
follows in a neighbourhood of 7 = 0: - ~ ev( -jkr) g, in 3D

xr = T(x, 7) = x + 7 g (x, 0) = x + TV(X), (3) where k = w/c is wave number, c is the speed of
sound, r is the distance between < and x and H(,I)is
the Hankel function of the n-th order of the first kind.
where V(x) = V(x, 0) represents velocity.
Since the second integrand is strongly singular, the
Using the above relations, material derivatives
left-hand-side of eqn (4) is multiplied by n(e), which
of boundary variables can be denoted by design
depends upon the geometry of boundary near 5, for
velocity. The desired expressions are well known and
a boundary point r.
are given in the Appendix.
After taking material derivative of eqn (4) with
respect to design variable, one can obtain conven-
DESIGN SENSITIVITY ANALYSIS tional sensitivity equation as follows:
Consider an acoustic field 0 subject to prescribed
acoustic pressure p or surface velocity fin on the
boundary r. In this case, the boundary integral
equation for acoustic pressure p at point 5 in the
domain 52 can be expressed as follows:
d(5) = -hop0
s I-
G(t,xh(x) dr(x)

~(0 = -&WI
s I-
G(5, x)r,(x) dW)
-
s r
FCC,
xM(x) dW)

-
s I-
05, xl&) dr(x), (4)
-
s
jwp0
r
6th xbn(x) Wx)

-
s I-
e’cr,xl&) dW)

-jopo
s I-
G(t, x)o&)JYx) Wx)

-
s I-
F(L x)~(x)JYx) dr(x), (64

J’(x) = df(x)/dI-(x), (6b)

where (*) denotes material derivative with respect to

Fig. 1. Deformation of domain by the mapping. design variable.
 Shape design sensitivity analysis 715

In eqn (6a), the sensitivity of acoustic pressure at Taking the material derivative of eqns (8a, b) one
any point g is represented by the surface velocity, can obtain the following equations:
acoustic pressure on the boundary and sensitivities of
them. As menticlned before, the integrands of k{ V(r) - P}.a sin{({ - q).a}
eqn (6a) have singularities when x approaches 5 and
it is not easy to obtain the accurate integral value of
the singular function. To remove the strong =k {c(C;, x) + G(5, x)J(x)}n(x).a sin(r.) dT(x)
singularity related to F(c, x) in eqn (6a), one may s I-
combine the uniform potential field equation.
However, this method may increase the compu- +k G(g, x)a.[ti(x)sin(ra)
tational cost and the weak singularity related to s I-
G(c, x) remains. The remaining weakly singular
behavior of the integrand may introduce the +Mx){ V(x) - V”}.a cos(r.)] d&x)
inaccuracy in the numerical procedure and require
some special treatment to obtain the accurate
solution. + {F(:(5,
x) + F(t, x)Js(x)}cos(r,)
dr(x)
In this paper, some different methods are presented sr
as follows to remove every singularity as well as to
avoid the potential analysis. If a and q are arbitrary
unit vectors and coordinates of an arbitrary point, -k f(l, x){ V(x) - P}.a sin(r,) dT(x), (9a)
respectively, the following equation is satisfied for sr
arbitrary point c in the domain S2:
k{ V(5) - P}.acos{k(< - q).a}

ev Mt-tlha) =jk
sI-
G(t, x)n(x)
=k {G(C,X) + (35, x)P(x)}n(x).a cos(r,) dT(x)

‘a exp{jk(x - q).a} dT(x)

+k G(5, x)a.[ti(x)cos(r,)
sI-
- F(c, x)exp(jk(x - yl)*a} dT(x). (7)
sr
- kn(x){ V(x) - V”}.a sin(r,)] dT(x)
Note that the above equation represents the
boundary integral equation of one-dimensional
- {SC, x) + F(5, x)P(x)}sin(r,) Wx)
wave propagation directed to a with unit magnitude.
As a special case of eqn (7), one can obtain the
following cosine or sine wave representation by
adding or subtracting the wave propagation directed
-k F(t, x){ V(x) - V”}.a cos(r,) dT(x), (9b)
to -a:

r, = k(x - q).a, (SC)

cos{k(t - q).a} -= - k G(<, x)n(x)
5r
where ( )” denotes the value at the reference point r]
‘a sin{k(x - q).a} dT(x) on the boundary.
Multiplying acoustic pressure at reference point pa
on eqn (9a) and substituting unit normal vector no
- F(<, x)cos{k(x - ?).a) dT(x),
into a, one can obtain the following sensitivity
(8a)
sr equation by combining it with eqn (6a):

sin{k(< - q).a} =: k
sr
G(<, x)n(x) fit51 = -jw0
sr
G(5, x)bdx) - p’kn”

.a cos{k(x - q)*a} dT(x) ’{i(x) + kn(x) V,”cos(r.)}ijwpO] df(x)

- F(t, x)sin{k(x - q).a} dT(x). @b) - F(5, x)V(x) + p”kV,” sin(r.)] dT(x)
sr
716 B. U. Koo

s
can be obtained with consideration for weak
-TWO {e(& x) + G(5, x)JYx)}Mx) singularity only.
r The remaining weak singularity in eqn (11) can be
removed by the following way. At first, define q” and

q*
=
-p”kno.n(x)sin(r,)ljwpa] dT(x) 4” as the following values:

-
s r
m’(r,xl + or, 4mHP(4
jpocv:
4” f jp,ct$.
(124

Wb)
-p” cos(r.)] dT(x) Secondly, multiply q” and 4” on eqn (9b) and (8b),
respectively, and substituting no into a in both
-p’k{ V(t) - vO).n” sin{k(< - fl).n”), (10a) equations. Lastly, combining the results of the second
step with the weakly singular representation (1 1), the
r, = k(x - q).n”, (lob) following equation can be obtained:

V,. = {V(x) - P}.nO. UOc)

d(5) = -@PO G(5, x)Mx)
In addition, multiplying p” on eqn (8a) and sr
substituting no into a, one can obtain the following
representation of acoustic pressure sensitivity at -ti.” n(x).n” cos(r,) - kn”.{pob(x)sin(r,)
internal point 5 by combining it with eqn (lOa):
+p”kn(x)Vpn cos(r,) + P’n(x)sin(r,)}ljwpo
r

- kn”. {p”i(x)sin(rV) + pOkn(x) V,. cos(r,)

- 05, X)b(X)-_P” cos(r.) + p”kV,” sin(r,)
s I-
+d”u(x)sin(m)}iiwpO] dI’(x)
+q”kVVncos(r,) + 4” sin(r,)] dT(x)
- F(5, x)@(x) - p” cos(r,)
sr
-.bwo {e(t, x) + G(t, x)J”(x)}[v.(x)
s I-
+p’kV, sin(r,)] dT(x)

- v.”n”.n(x)cos(r,)
-&PO c(5, x) + G(4, x).r(x)}[u,(x)
s I- -p”kno.n(x)sin(r,,)ljwpo] dT(x)

-p”/zno.n(x)sin(r,)ljwpo] dT(x)

-p” cos(r.) + q” sin(r,)] dT(x)

--p” cos(r,)] dT(x) - p”k{ V(5) - P} -[pOk{V(<) - vO}.nO+ d’]sin{k(t - q).n”}

.n” sin{k(r - q).u”} +[p” - q’k{ V(5) - P)

+@O
Cos{k(< - q).n*). (11) -n”]cos(k(~ - q)m”}. (13)

Defining v] as the closest boundary point from the Note that all the integrands in the above equation
source point <, one can find that the above equation are regular if the boundary is smooth near the source
is only weakly singular, although < is a point on the point 5 and design velocity is continuous. So any
boundary as well as near the boundary. Since the consideration for singularity is not required to obtain
strong singularity in the sensitivity equation is the solution near the smooth boundary. On the other
removed, the numerical solution of sensitivity is hand, consideration for weak singularity is required
continuous on the boundary and an accurate result near a sharp corner or edge.
 Shape design sensitivity analysis 111

In acoustics, exterior problems usually occur to -v.” no. {n(x)cos(r.)

analyze the acoustic field generated by vibrating
bodies in the infinite region. The sensitivity equation
can be utilized for the exterior problem. In order to - kn(x) V, sin@.)]] dT(x) - F(& x)W)
derive the exterior sensitivity equation, one needs to
know the effects of integration over the infinite
boundary. Applying the Sommerfeld radiation -POCOs(r,)+ p’kl/, sin(r,)
condition, one can know that the integral effect of an
infinite boundary has vanished in the original +q”kV, cos(r.) + 4” sin(r.)] dI’(x)
sensitivity eqn (6a). However, the effect of a
one-dimensional wave which is used to remove the
singularities remains. Because the source point 6
always exists inside the infinite boundary, one can
easily find that the effects are the same with the free
-MPO
sr
@‘(Lx)+ Gtt, x)W)}b.tx)

-21,”nO.n(x)cos(r.)
terms in the right-hand-side of eqn (13). The
sensitivity equation for the exterior problem is
-p”kno.n(x)sin(r.)ljwpo] dT(x)

P(t) = -&WO
5I-
G(<, x)[d.(x) - d,O

4 cos(r.) - kn”.{pob(x)sin(r.)
n(x)
-
stat,
r
x) + F(t, -w(-4)[P(x)

-POCOS(r,)-I- q” sin(r.) dr(x), (15)

+p’Wx)I’, costr,) + P”n(x)sin(r~)~/jwpo
where a = 0, 1 for interior and exterior problems,
respectively.
- 0,”n”~{~(.x)COs(r~)
-kn(x) V, sin(r,,)}] dT(x)

NUMERICAL IMPLEMENTATION
- F(5, a:)&(X) - 0” cos(r.)
1r In this section, some problems encountered in the
numerical implementation of sensitivity equations are
+pOkV, sin(r,) + q”kV,” cos(r,) described. The boundary surface was discretized with
a triangular element six-nodes and 7 points Gaussian
quadrature was used for numerical integration.
+4” sin(r.N dT(x) -jwpo {G(<, x) Choosing an orthonormal basis on the boundary is
sr
a problem in obtaining the rates of change of
boundary surface and normal vector. The normal
+ G(5, x)JYx))[v.(x) - vEnO.n(x>cos(r,)
vector can be chosen uniquely. However, two
-p’knO.n(;r)sin(r,)ljop,] dT(x) orthogonal tangential vectors can be chosen
arbitrarily. The tangential vectors are chosen in the
following way. Suppose that the element was mapped
on the reference coordinate (u, v) as shown in Fig. 2
and a unit tangential vectors was chosen parallel with
-p” cos(r, ) + q” sin(r.)] dr(x). (14) the u-axis in the reference coordinate system. Then
the other unit tangential vector t could be chosen
Here, r denotes a vibrating surface only. uniquely using the orthonormality of s, t and n:
As mentioned previously, the sensitivity of
boundary variables has to be known to obtain the t=nxs. (16)
sensitivity at field. The sensitivities of boundary
variables can be obtained by placing a source on the The tangential derivatives were obtained by the
boundary and defining the reference point q as the chain rule and shape function. Let the auxiliary unit
point identical to ([ in the sensitivity equation for an tangential vector w in global coordinate system be
internal point. The sensitivity equation for the defined such that it is parallel with v-axis in reference
boundary variable is coordinate system. Then the two directional deriva-
tives are obtained as follows:

aDtO= -jwo
sr
x)[h(x)
,G(t, a
_=-- a du
as auds’ (1W
-&! n(x).nO cos(r,) - kllO{pOL(x)sin(rO)

+pOkn(x)V, cos(r.) + p”n(x)sin(r.)}ljopo

718 B. U. Koo

Fig. 2. Mapping of tangential vectors on the boundary.

Using eqns (17a, b), the tangential derivative with point with fixed distance from the surface in the field
respect to t can be represented as follows: and compared to the theoretical solution. For a
vibrating sphere with velocity u., the theoretical
pressure sensitivity at a point with fixed distance from
a
at= -g+.w)$$ (18) the surface is represented as follows when the velocity
u, does not change with the variation of sphere radius:

Finally, the coordinates of the reference point have p,cu,ka{(kar - kd -@)(l +jku) + 2jr}
to be found to obtain the sensitivities at the given field ii(r) = rZ(l + k*a*)(l +jku)
points near the boundary. The first step is to find the
closest element to the source point <. This element x exp{ -jk(r - a)}, (19)
can be obtained among elements on which the closest
node exists. Then the coordinates of the closest point where r is a radial distance from the center of the
can be obtained employing some optimization sphere to a location in the field and a is the radius of
techniques [7, 81. To obtain the robust result, the the sphere.
sequential quadratic programming and inaccurate The sensitivity was evaluated by using the derived
line search (Armijo’s rule) algorithms were used. equation. The boundary surface of a pulsating sphere
with radius, a = 0.1 m, was discretized with 48
triangular elements with six nodes. Sensitivity was
VERIFICATION OF SHAPE DESIGN SENSITIVITY
PROBLEM calculated for ku = 1 at some field points approach-
ing the boundary.
Analytical sphere models were used to verify the Tables 1 and 2 represent the accuracies of predicted
shape design sensitivity formulation. Boundary sensitivities of acoustic pressure. From Table 1, the
conditions were specified such that the sphere was real part of sensitivity is predicted to within 0.64%
modelled to vibrate with a uniform radial velocity. error at a field point and 0.51% error on the
The spherical radius a was designated as the design boundary. The imaginary part of sensitivity is more
variable. The sensitivity of the acoustic pressure due or less inaccurate compared to the real part. From
to the variation of sphere radius was calculated at a Table 2, one can find that the imaginary part of

Table 1. Real part of the sensitivity of acoustic pressure Table 2. Imaginary part of the sensitivity of acoustic
computed by non-singular integral equation at internal pressure computed by non-singular integral equation at
points and accuracy of the computation internal points and accuracy of the computation
Log[(r - ~)/a] Analytical BIE Accuracy (%) Log[(r - a)/~] Analytical BIE Accuracy (X)
0.0 1177.880 1170.358 99.3614 0.0 -3019.874 - 3030.806 100.3620
-0.5 3456.480 3464.455 100.2308 -0.5 - 1681.072 - 1700.266 101.1418
-1.0 3757.275 3773.295 100.4264 -1.0 - 1188.952 - 1198.175 100.7758
-1.5 3788.069 3806.298 100.4812 -1.5 -1115.255 -1118.489 100.2900
-2.0 3791.125 3809.995 100.4977 -2.0 - 1106.823 -1110.822 100.3613
-2.5 3791.429 3810.504 100.5031 -2.5 - 1105.942 - 1110.095 100.3754
-3.0 3791.459 3810.600 100.5048 -3.0 - 1105.853 - 1109.988 100.3739
-3.5 3791.462 3810.623 100.5054 -3.5 - 1105.844 - 1109.964 100.3726
-4.0 3791.462 3810.630 100.5056 -4.0 - 1105.843 - 1109.958 100.3721
-4.5 3791.462 3810.632 100.5056 -4.5 -1105.843 - 1109.956 100.3719
-5.0 3791.462 3810.633 100.5056 -5.0 - 1105.843 - 1109.955 100.3719
--co 3791.462 3810.633 100.5056 -CC - 1105.843 - 1109.955 100.3718
 Shape design sensitivity analysis 719

sensitivity is predicted to within 1.15% error in the 4. J. H. Kane and S. Saigal, Trans. ASCE J. Engng Mech.
field and 0.37% error on the boundary. 114, 1703-1721 (1988).
5. M. R. Barone and R. J. Yang, AIAA J. 26, 589-594
(1988).
6. k4. R. Barone and R. J. Yang, Comput. Meth. Appl.
CONCLUSION Mech. Ennnn 74. 69-82 (19891.
7. D. G. L&berg&, Linea; andonlinear Programming,
A shape design sensitivity equation has been 2nd Edn. Addison-Wesley, (1984).
derived for acoustic problems. The equation is 8. M. S. Bazaraa and C. M. Shetty, Nonlinear Program-
ming. Wiley, New York (1979).
derived by using .the material derivative concept in
continuum mecha:nics and singularities are removed
by using the integral representation of one-dimen- APPENDIX
sional wave propagation mode. Since the derived
The material derivatives of point vectors are written as
equation uses only the wave equation to remove the
singularities, the Ipotential field is not required to ir = V(x), tj = V(q), 5 = V(l). (A.1)
remove the strong singularity. In addition, the
integrands in the equation are regularized and an The material derivatives of displacement vector and
accurate solution can be obtained near the boundary distance are
using the standard Gaussian scheme. A sphere model
was used to verify the shape design sensitivity i(x, r) = V(x) - v(r),
formulation. As a result, accurate sensitivity was
obtained near the boundary as well as on the i(x, r) = {V(x) - V(5)}& O/r, (A.2)
boundary.
where i(x, t;) denotes the vector from < to x and r(x, r)
denotes distance between 5 and x.
The expressions of change of boundary and normal vector
REFERENCES are also needed and expressed as
1. E. J. Haug, K. K. Choi, and V. Komkov, Design
Sensitivity Analysis of Structural Systems. Academic
df = (Vi.,S, + K.,t,) dl-,
Press, New York (1986).
2. K. A. Cunefare .snd G. H. Koopmann, Trans. ASME ri, = - V,.$n,a - V,,,njt,, (A.31
J. Vibr. Acoustics 114, 178-186.
3. D. C. Smith and R. J. Bernard, Trans. ASME J. Vibr. where s and t are two unit tangential vector and s, t and n
, .
Acoustics 114, 12’7-132 are mutually orthonormal.

CAS 65/S-D