Structural Optimization 8, 52-59 Springer-Verlag 1994
Shape design sensitivity analysis of eigenvalues using "exact"
numerical differentiation of finite element matrices
E. L u n d a n d N. O l h o f f
Institute of Mechanical Engineering, Aalborg University, Pontoppidanstraede 101, DK-9220 Aalborg East, Denmark
Abstract As has been shown in recent years, the approximate
numerical differentiation of element stiffness matrices which is in-
herent in the semi-analytical method of finite element based design
sensitivity analysis, may give rise to severely erroneous shape de-
sign sensitivities in static problems involving linearly elastic bend-
ing of beam, plate and shell structures.
This paper demonstrates that the error problem also mani-
fests itself in semi-analytical sensitivity analyses of eigenvalues of
such structures and presents a method for complete elimination of
the error problem. The method, which yields "exact" numerical
sensitivities on the basis of simple first-order numerical differen-
tiation, is computationally inexpensive and easy to implement as
an integral part of the finite element analysis.
The method is presented in terms of semi-analytical shape de-
sign sensitivity analysis of eigenvalues in the form of frequencies
of free transverse vibrations of plates modelled by isoparametric
Mindlin finite elements. Finally, the development is illustrated
via two examples of occurrence of the error phenomenon when
the traditional method is used and it is shown that the problem
is completely eliminated by the application of the new method.
1 Introduction
In a general purpose computer aided environment for inter-
active structural design and shape optimization, it is very
important to have a reliable and efficient method for design
sensitivity analysis. In this paper the sensitivity analysis is
confined to design sensitivities of eigenvalues w . r . t shape de-
sign variables when the finite element method is used as the
analysis tool. The method of semi-analytical design sensi-
tivity analysis of finite element discretized structures where
the derivatives of finite element matrices are calculated by
a first-order finite difference approximation (see Zienkiewicz
and Campell 1973; Cheng and Liu 1987; Haftka and Adel-
man 1989) is very efficient and will be described here. How-
ever, it will be shown that application of the traditional semi-
analytical method may yield completely erroneous results for
eigenvalue sensitivities. In fact, the same inaccuracy prob-
lem that has been reported in the application of the tradi-
tional semi-analytical method in static shape design sensitiv-
ity analysis (see Barthelemy and Haftka 1990; Pedersen et al.
1989; Cheng et al. 1989) can be expected in dynamic shape
design sensitivity analysis as well. The inaccuracy problem
in static design sensitivity analysis is entirely due to errors in
the approximate finite difference calculation of stiffness ma-
trix derivatives (see Olhoff and Rasmussen 1991; Fenyes and
Lust 1991; Cheng and Olhoff 1993), but a new method of
"exact" numerical differentiation of element stiffness matri-
ces (exact except for round-off errors) (see Olhoff el al. 1993)
completely eliminates the inaccuracy problem.
In this paper the method of "exact" numerical differen-
tiation will be illustrated for isoparametric Mindlin plate fi-
nite elements for determining "exact" numerical derivatives
of element stiffness, mass and stress stiffness matrices. Fur-
thermore, two numerical examples will demonstrate that the
method of "exact" numerical differentiation of finite element
matrices should also be implemented in design sensitivity
analysis of structural eigenvalue problems when using the
semi-analytical method.
2 D e s i g n s e n s i t i v i t y a n a l y s i s o f s t r u c t u r a l , eigen-
value p r o b l e m s
The analysis problem considered here is a standard struc-
tural eigenvalue problem where the eigenvalueS are all real
and represent either free vibration frequencies, eigenfrequen-
cies considering stress stiffening effects or linear buckling load
factors. For such a real, symmetric, structural eigenvalue
problem, the finite element formulation is
Kej=,~jMej, j = 1, . . . , n , (1)
where K and M arc symmetric, positive definite matrices,
)~j is the eigenvalue and e j is the corresponding eigenveetor.
The dimension of the problem is denoted by n, so (1) has
n solutions consisting of real eigenvalues )~j and correspond-
ing eigenvectors ~bj. Depending on the type of problem, the
global K and M matrices consist of contributions from either
element stiffness, mass or stress stiffness matrices.
In the following it is assumed that the eigenvalues are
ordered by magnitude and that the eigenvectors have been
M-orthonormalized, i.e.
d~jT Mdpk = 6jk , j, k = 1 . . . . , n , (2)
where 5jk denotes Kronecker's delta.
It is assumed that the shape of the structure is described
by a set of design variables ai, i = 1, . . . , I and that it is the
goal to obtain expressions for eigenvalue sensitivities with
respect to these shape design variables. It is also assumed
that the components of the K and M matrices are smooth
functions of design variables a i.
If the eigenvalue ,~j is distinct (or simple), the eigen-
value sensitivity with respect to a design variable a i can be
found by differentiating (1) and premultiplying by ~bT (see
e.g. Courant and Hilbert 1953)

,4,T OK,4, . A .~T OM ~ .
OAj = ,ej 8-~i,ea - j , e j -8--~iv-a i = 1. . . . , I ,
Oat ~bTM~j '
where the denominator is equal to unity due to the M-
orthonormalization in (2).
If the eigenvalue analysis yields an N-fold multiple eigen-
value
=A j, j=I,...,N,
where, for convenience, the repeated eigenvalues have been
numbered from 1 to N, then the eigenvalue sensitivities # =
~j, j = 1,... , N , of the multiple eigenvalue A with respect
to a small change Aa i of a single design parameter a i can be
found by solving the following sub-eigenvalue problem (see
e.g. Courant and Hilbert 1953; Bratus and Seyranian 1983;
ttaug el al. 1986)
d e t [ ~ T ( ~ a K i - ~ 0 M ' ~ 0i ja ~ k - - t t 6 s k ] =0,
j,s,k,=l,...,N, i=l,...,I.
It should be noted that multiple eigenvalues are not differ-
entiable in the common sense, i.e. not Frdchet-differentiable
(see e.g. Haug et al. 1986). This means that the eigenvalue
sensitivities p j are only valid as directional derivatives.
The only unknown quantities in (3) and (5) are the deriva-
tives of the K and M matrices, and these derivatives are
normally calculated at the element level, i.e.
oK = g - , o k ,
Oa i ~0a i i=1,...,I,
0M = E i = 1, . . , ,,
0at ne
where k and m are element matrices and ne is the number
of finite elements.
If the design sensitivities Ok/Oa i and Om/Oa i are deter-
mined analytically before their numerical evaluation, the ap-
proach is called analytical design sensitivity analysis, and if
they are determined by numerical differentiation, the method
is called semi-analytical design sensitivity analysis
The method of analytical design sensitivity analysis is
very cumbersome to implement in a general purpose shape
design system containing many different kinds of shape de-
sign variables and finite element types. Thus, a large amount
of analytical work and programming will be required in order
to develop analytic expressions for derivatives of various stiff-
ness matrices with respect to possible shape design variables.
It is much more attractive to use the method of semi-
analytic design sensitivity analysis in the present context,
since this method is easy to implement for many different
kinds of shape design variables and finite element types be-
cause simple and computationally inexpensive first-order fi-
nite differences are used, i.e.
0k(a) , Ak(al,... ,aI)
Oa i Aa i
k(al,...,a i+Aa i ..... aI)-k(al,...
Aa i
0 m ( a ) _ A m ( a I . . . . , ai) _
Oa i Aa i
53
m ( a I . . . . , ai + Aa i, . . . , aI) -- m ( a l , ... , a i , . . . , aI)
(3) (9)
Aa i
However, when used in the computation of eigenvalue sensi-
tivities in (3) and (5), as will be demonstrated in this paper,
the approximate finite difference calculation of element ma-
trix derivatives in (8) and (9) can result in severe errors, and
therefore the method for "exact" numerical differentiation of
(4) finite element matrices will be introduced. This method is
efficient and applicable for a wide range of problems and is
described in detail by Olhoff el al. (1993) for static design
sensitivity analyses. Application of the method for eigen-
value problems is described in the sequel, where isoparamet-
ric Mindlin plate elements are used for illustration.
3 "Exact" numerical differentiation of element ma-
trices
The accuracy of the first-order finite difference approxima-
tions in (8) and (9) is strongly dependent on the chosen size
(5) of perturbation A a i as the element matrices generally de-
pend non-linearly on shape design variables, and in some
cases, such small perturbations are needed that computa-
tional round-off errors become the problem. In order to avoid
dependence on the chosen perturbation Aai, the goal is to
construct a method for "exact" numerical differentiation of
element matrices based on computationally inexpensive first-
order finite differences.
This goal may seem unattainable, but a closer study of
(6) the functions that form the element matrices reveals that
the same mathematical forms are common for large groups
of finite elements. For instance, the element matrices for
(7) all isoparametric elements with translational degrees of free-
dom, and isoparametric Mindlin plate and shell elements
depend on the same class of functions. Similarly, the ele-
ment matrices of a large class of finite elements comprising a
Bernoulli-Euler beam, and Kirchhoff plate and shell elements
have a similar mathematical structure. The members of these
classes of matrices in general depend non-linearly on the de-
sign variables, but are defined within a special mathematical
form. The mathematical form implies that their approxi-
mate numerical derivatives, computed by a usual first-order
finite difference scheme, can be upgraded to "exact" deriva-
tives by simple multiplication by appropriate correction fac-
tors. The values of these correction factors can be very easily
pre-computed and be used throughout the procedure of de-
sign sensitivity analysis. It follows as a remarkable side-effect
of the "exactness" that the results become totally indepen-
dent of the magnitude of the perturbation.
As an example consider here isoparametric Mindlin plate
finite elements that include in-plane membrane capabilities.
It is assumed that the element matrices of a particular finite
element only depend on a given sub-set from among the total
set of shape design variables at, i = 1, . . . , I, which in this
context is considered to be the global coordinates of the finite
element nodal points. This sub-set of the design variables is
renumbered and denoted by aj, j = 1, . . . , J, where J < I.
,ai,...,ai) In the case of isoparametric finite elements, the element
, (8)
matrices turn out to depend on functions g, which are incom-
plete polynomia and defined by the following form:
g(al,...,aj)=pj(al, . . . , a j _ l , aj+l, . . . , a j ) +
54

qj(al .... , aj_l, a j + l , . . . , aj)(aj) rj , 4.1 Derivative of element stiffness matrix

rj N U {0}, for all j = 1, . . . , J . (10) The element stiffness matrix k for a Mindlin plate finite ele-
ment is given by
Thus, a function g is such that for any j , the term pj and the
coefficient qj ~ 0 are independent of aj. The design variable
aj appears in one and only one power rj which belongs to k = fBTDMBIJld~, (15)
a/
the set N U {0} of non-negative integers. Although not stated
explicitly, pj and qj, and therefore the element function g, where D M is the elasticity matrix, B the strain-displacement
will generally depend on the local coordinates within the fi- matrix, IJ] the determinant of the Jacobian matrix J, and w
nite element. Typically, g may represent the determinant or is the domain of the finite element as shown in Fig. 1. The
components of the Jaeobian matrix or components of other matrices 3 and B depend on the coordinates of the nodal
matrices in the definition of an element matrix. points, whereas D M is independent of the shape design vari-
As shown by Olhoff et al. (1993), it is possible to estab- ables.
lish the following relationship of proportionality between the
analytical (and exact) derivative Og/Oaj and its first-order
finite difference approximation Ag/Aaj:
Og Ag Vi
cgaj =cU-~aj' rjNU{O}, j=l,...,Z. (11)

The proportionality factors Crj may be termed correction fac-

tors and they only depend on the relative perturbation r/j
defined as Z ~yt
Aaj Fig. 1. Domain, coordinates and nodal degrees of freedom of
r]j= aj > 0 , j=l,...,J. (12)
isoparametric Mindlin plate element with in-plane membrane ca-
pability
Thus, the correction factors Crj can be expressed as

Before the derivative of this stiffness matrix is found, let

c0 = O , c 1 = 1, c2 = l+~r/j ,
us recall that within the isoparametric formulation of a finite
element with an arbitrarily given number n of nodal points,
% = the shape functions N i only depend on the local, nondimen-
sional coordinates ~ and rj, i.e.

N i=Ni((,rl) , i = l , . . . , n . (16)
[lk_p~=o(kp)
- ~}k-1)-p ] - 1 , (13)
The same set of shape functions N i is used for interpolation
of the global coordinates x, y from nodal values xi, Yi and of
where the latter expression for c k is given in terms of binomial displacement functions u, v from nodal values ui, v i.
coefficients.
In terms of the vector d i of nodal degrees of freedom, cf.
The correction factors in (13) are independent of the ac- Fig. 1,
tual values of the design variables and can therefore be pre-
computed for a selected finite value of rlj > 0. Hence, the d i = { u iv iw iOxiOyi) T, i= 1 , . . . , n , (17)
absolute perturbation Aa i can be eliminated in (11) which
may be rewritten as the relationship between the vector d of generalized displace-
9g Ag Crj ments of element nodal points
Oaj -- Crj -Aaj
- - r~jaj {g((1 + rlj)aj) - g(aj)},
d={d TdT...dT...dT} T, (18)
rj N U { 0 } , j=l,..., J. (14)
and the generalized strain vector takes the standard form
Equations (13) and (14) are the main expressions to be used = B d where B is given by
when calculating derivatives of isoparametric finite elements.
B=[b lb 2 ...b i ...bn], (19)
4 D e r i v a t i v e s o f i s o p a r a m e t r i c M i n d l i n p l a t e finite and
elements
UlZ
In this section the method of "exact" numerical differentia- f V~y
tion is used for determining ':exact" numerical derivatives of U,y nUV,x
element stiffness, mass and initial stress stiffness matrices for ~(*, y) = / 02g~x
isoparametric Mindlin plate finite elements. For more details
concerning the derivation of "exact" numerical derivatives of
other element matrices, the reader is referred to the paper by
Olhoff et al. (1993).
[ Oy,y
Ox,y + Oy,x
Oy -- W~y
0 x -- W,X
 55

- Ni, z 0 0 0 0 OJ f Ni'z } = - J -1 1 { J ( ( l + ~j)aj)--

1 J ~ ~
0 Ni,y 0 0 0 ~aj Ni,y ~jaj
Ni,y Ni, x 0 0 0
0 0 0 Ni, x 0 }{Ni'x}, i = 1 . . . . . n, j=l,...,J. (26)
bi = 0 0 0 0 Ni,y J(aj) _ _Xi,y
0 0 0 Ni,y Ni, x The shape functions N i only depend on the local, nondimen-
0 0 -Ni,y 0 Ni sional coordinates ~ and r/, i.e.
0 0 -Ni, x Ni 0 ONi
----=0, i=l,...,n, j=l,...,J. (27)
i=l,...,n. (20) Oaj
Here, the derivatives of the shape functions N i with respect Now all terms needed for the "exact" numerical derivative of
to x and y are given by the stiffness matrix, el. (23), are found and the same proce-
dure can be used to calculate other element matrix deriva-

{ N-:.,}s
i = 1. . . . , n,
,,,,,"'] :{-,,< }=r { }
Ni, ~
(21)
tives.

4.2 Derivative of element mass matrix

where the matrix r is the inverse of the Jacobian The element mass matrix m for a Mindlin plate finite element
is given by
]= [ Ni,5xi Ni,5yi ] (22)
Lx,~ Y,~ [ Ni31xi Ni,~Yi m =/pNTNialdw, (28)
i=1
Now the terms of the stiffness matrix in (15) are described
and the derivative of the stiffness matrix can be found by where p is the mass density, N the vector of shape functions
differentiating (15) with respect to any of the design variables Ni, [J] the determinant of the Jacobian matrix 3, and w is
aj, j = l , . . . , J the domain of the finite element.
The derivative of the mass matrix can be found by differ-
Ok 7 [OBT T 0 B ] ]Jidw + entiating (28) with respect to any of the design variables aj,
Oaj = k ' DMB + B DMOa---J
|-g~-a~" j = 1, ..., J,i.e.
0m
- - i p i T i O[J]dw, j 1,..., J. (29)
f B T D M B OIJI dw, j= 1,..., J. (23)
Oaj ~ Oaj
O2
Oaj
O2 The derivative of the determinant of the Jacobian is given by
Thus, the derivatives of the determinant of the Jacobian 3 (24) so all terms needed for calculating the "exact" numerical
and of the components in the strain-displacement matrix B derivative of the mass matrix are known.
must be determined.
Applying (10), (13) and (14) on the above terms in (19)- 4.3 Derivative of element initial stress stiffness matrix
(22) yields the "exact" numerical derivatives based on first- In the derivation of element initial stress stiffness matrix
order finite differences. The scalar IJI will be either indepen- derivatives it is convenient to reorder and omit some nodal
dent or a linear function of any of the shape design variables degrees of freedom by introducing the reordered, condensed
aj, and the derivative of [3[ is therefore given by element vector d*, which only contains translational nodal
0]3f degrees of freedom. These translational d.o.f, are reordered
_ 1 { i a ( ( l + vj)aj) I _ 13(aj)l} ' J = 1,... a.(24) so that first all x-direction d.o.f, are given, then y, and then
Oaj ~jaj
z as follows:
The computation of derivatives of components of the strain-
displacement matrix B requires differentiation of b i with re- d*={UlU 2 . . . u i . . . un vl v 2 . . . v i . . . vn wl w 2 . . . w i . . . wn } T .
spect to aj and hence the derivatives of Ni,x, Ni,y and N i. (30)
This involves differentiation of the matrix F, and since the
Relating d.o.f, to the reordered, condensed element vector
components of this matrix are given by Fqp = 131-1 cof(Jpq),
these components cannot be differentiated exactly on the ba- d*, the element initial stress stiffness matrix k~r for a Mindlin
plate finite element is given by
sis of a simple polynomial approximation. This difficulty can
be circumvented by differentiating the identity r 3 = I, where k~ = f GTsGtJfd~, (31)
I is the identity matrix, which gives
OF 0J where G is a matrix obtained by appropriate differentiation of
_-r?-~,r, j = l . . . . . J. (25)
Oaj shape functions Ni, S a matrix of membrane stresses, IJI the
The derivatives of Ni, x and Ni,y can then be found using determinant of the Jaeobian matrix J, and w is 'the domain
(21) and (25), i.e. of the finite element.
The matrix G is given by
Ni,y Ni,~
G = g , (32)
Oa--~O f~ Ni,~ } or { Ni,( } =
~_~jaj FOJF 0
Ni,v = Ni,7 Ni,~ where each submatrix g is given by
56

[ Nl,x N2,z ... Ni,z ... Nn,x ] where E is the plane stress elasticity matrix, the expression
g = LNl,y N2,y ... Ni,y ... Nn,y .I
d
(33)
for stress sensitivities is found by differentiating (39) with
The stress matrix S is given by respect to any of the design variables aj, j = 1,..., J,

s = [ o!1 s
0
, (34) Oaj - E d+ Oaj] ' j=l'''''J"
The terms on the right-hand side in (40) are given by (!9),
(40)

where each submatrix s is given by

(20), (26) and (38), and having calculated the stress sensitiv-
S ~
[ O'x Vxy ] (35) ities Oo-/aaj, the "exact" numerical derivative of the initial
7xy O'y stress stiffness matrix, cf. (36), can be calculated.
Here az, ~ry and ~rzy are the membrane stresses in the plate Now expressions of "exact" numerical derivatives have
found by an initial static stress analysis. been derived for element stiffness, mass and initial stress
If (31) is differentiated with respect to any of the design stiffness matrices for Mindlin plate finite elements, and in
variables aj, j = 1,..., J, the following expression for the the following section numerical examples will emphasize the
"exact" numerical derivative of the initial stress stiffness ma- importance of using our modified version of the method of
trix is obtained: semi-analytical design sensitivity analysis.

Oaj [ Oaj Oaj G + G T S " ]Jldw +

[GTsG~dw, j= 1,...,J.
t,O
. /
Thus, it is necessary to find the derivatives of the components
in the stress matrix S, of the determinant of the Jacobian J,
and of the components of the matrix G.
The two latter terms are already found as the derivative
of the determinant of the Jacobian J is given by (24) and the
derivatives of the components of the matrix G are given by
(26). The derivatives of the stress components of the stress
matrix S must be found in an initial static design sensitivity
analysis and the method for doing this will be summarized
here.
First, the design sensitivities of the displacement field
must be determined and it is then straightforward to calcu-
late "exact" numerical derivatives of the stresses. The direct
approach to obtain design sensitivities of the displacement
field of a finite element discretized structural design problem
with linearly elastic response is based on implicit differentia-
tion of the global equilibrium equation
KD = F,
where K is the reduced global stiffness matrix, D the global
nodal displacement vector, and F is the global consistent
nodal force vector. If (37) is differentiated with respect to
any of the design variables aj, j = 1, ..., J, and the terms
are rearranged, the following expression for the displacement
sensitivities OD/Oaj is obtained:
0~. 0K(a) 0F
K(a) ~ - Oaj D + Oaj "
Equation (38) is of the same form as (37), so the factorized
stiffness matrix K from (37) can be reused and only the new
right-hand side, termed the pseudo load vector, needs to be
calculated before the displacement sensitivities OD/Oaj can
be found. The derivatives OF/Oaj of the force vector are eas-
ily calculated (note that they are zero for design independent
loads), and the derivative of the global stiffness matrix can be
calculated at the element level using (6) and (23). Then the
displacement sensitivities OD/Oaj can be determined, and as
the membrane stresses can be found from
~r = E B d ,
5 Numerical examples
In this section two numerical plate examples are considered
(36) where the traditional method of semi-analytical design sen-
sitivity analysis can result in severe errors.
5.1 Free lransverse vibralion frequencies of clamped square
plale
The first example concerns a square plate which is clamped at
all edges and modelled by 100 9-node isoparametric Mindlin
plate elements, see Fig. 2. The plate is made of steel with
the following material properties: Young's modulus = 210000
MPa, Poisson's ratio = 0.3 and mass density = 7800 k g / m 3.
(37) Fig. 2. Finite element model of square plate
The side length L of the square plate is 1.0 m and the
thickness 1.0.10 - a m, i.e. the length/thickness ratio is 1000.
The lowest eigenfrequency f l = 8.994 Hz is simple, while
f2 = f3 =18.36 Itz is double.
The shape design variable is the side length L of the plate,
and the sensitivities of the three lowest eigenfrequencies will
be computed for different perturbations. The overall finite
(3S)
difference (OFD) method is used as a reference method whose
limit with respect to design sensitivity accuracy is known to
be set only by the solution procedure, the discretization, and
the usual accuracy capabilities of the applied finite element,
when the design perturbation Aa i is sufficiently small. The
OFD method implies that the design is perturbed, a new
eigenfrequency analysis is performed, and then the eigenfre-
quency sensitivities #j are found from
Afj(al,...,ar)
#j "" Aa i =-
fj(al ..... a i + A a i , . . . , a I ) - fj(al,...,aI)
(39) Aa i (41)
 57

In the following, the eigenfrequency sensitivities will be cal- 10 - 1 and 10 - 2 are due to these large values. Furthermore,
culated by the traditional semi-analytical (SA) method, of. the OFD method becomes inaccurate for very small pertur-
(8) and (9), and by the new semi-analytical method. bations due to numerical round-off errors.
The data presented in Table 1 are the sensitivities of the The inaccuracy problem associated with the traditional
lowest eigenfrequency while Table 2 presents sensitivities of SA method can be explained as in the paper by Cheng and
the multiple eigenfrequency. As ]'2 and f3 remain multiple Olhoff (1993), where a geometrical-physical interpretation
with the design change considered, the same values are ob- of the influence of the error associated with the first-order
tained for sensitivities//2 and #3, and therefore only results forward difference approximation Ak/Aaj to the element
for #2 are listed in Table 2. stiffness derivative Ok/Oaj in (8) is presented for static de-
sign sensitivity analysis. The contribution from the stiffness
Table 1. Computed derivative of lowest eigenfrequency f1 with matrix derivative to the element pseudo loads is given by
respect to side length L -(Ok/Oaj)d, cf. (38), where d is the vector of nodal dis-
placements of the element. Now, any displacement vector d
Plate length OFD Traditional New
can be subdivided into three vectors: a vector d t for the rigid
perturbation method SA method SA method body translation, a vector dr for the rigid body rotation, and
.AL all all All a vector for the remaining part of d associated with the actual
L AL AL AL deformation of the finite element. Along the same lines as a
10 - 2 -17.71 12188 -17.99 proper finite element should possess the rigid body motion
capabilities k d t = 0 and k d r = 0 (vanishing of the element
10 - 3 -17.96 1308 -17.99
nodal forces associated with d t and dr), it is clear that all
10 - 4 -17.99 115.7 -17.99 the components of the element pseudo loads associated with a
rigid body translation and rotation, respectively, should van-
10 - 5 -17.99 -4.605 -17.99
ish such that -(Ok/Oaj)d t = 0 and -(Ok/Oaj)dr = 0. It is
10 - 6 -17.99 -16.45 -17:99 shown by Cheng and Olhoff (1993) that the latter conditions
are satisfied for both sizing and shape design variables when
10 - 7 -17.99 -17.86 -17.99
analytical design sensitivities Ok/Oaj are used.
10 - 8 -18.00 -17.98 -17.99 Now, it is shown by Cheng and Olhoff (1993) that if the
10 - 9 -17.97 -17.99 -17.99 design sensitivities are replaced by their forward difference
approximations, then the conditions --(Ak/Aaj)dt = 0 and
10 - l -18.13 -17.97 -17.99 -(Ak/Aaj)dr = 0 are also both satisfied if aj is a sizing de-
sign variable. However, if aj is a shape design variable, only
the former condition is satisfied, i.e. only the approximate
Table 2. Computed sensitivities tt2 (= #3) of double eigenfre- element pseudo loads associated with a rigid body transla-
quency f2 = fa with respect to side length L tion vanish. Hence the latter condition is generally not sat-
isfied if aj is a shape variable, which means that the com-
Plate length OFD Traditional New
ponents of the approximate element pseudo loads that corre-
pertubation method SA method SA method spond to a rigid body rotation do not vanish in general, i.e.
AL -(Ak/Aaj)dr # O. This was also shown by Mlejnek (1992).
L /12 #2 //2 This is the geometrical-physical reason why the tradi-
10 - 2 -36.10 19415 -36.71 tional SA method becomes inaccurate for problems where
10 - 3 -36.63 2074 -36.71
the displacement field is characterized by rigid body rotations
which are large relative to actual deformations of the finite eL
10 - 4 -36.71 176.2 -36.71 ements, i.e. for example in problems involving linearly elastic
bending of long-span, beam-like structures, and of plate and
10 - 5 -36.71 -15.40 -36.71
shell structures. It is important to note that the inaccuracy
10 - 6 -36.71 -34.58 -36.71 problem associated with the traditional SA method is depen-
dent on the kind of structure considered and manifests itself
10 - 7 -36.72 -36.50 -36.71
independently of the type of finite element used to model the
10 - 8 -36.72 -36.69 -36.71 structure.
10 - 9 -36.80 -36.71 -36.71 As the design sensitivity expressions for eigenvalues in-
volve multiplication of stiffness matrix derivatives by the
10 - l -36.19 -36.70 -36.71 eigenvector, see (3) and (5), the same type of inaccuracy
problem will arise when the SA method is applied to design
The results in Table 1 and 2 confirm that the traditional sensitivity analysis of eigenvalues.
SA method is strongly dependent on the size of the rela- In the current example, the small length/thickness ratio
tive perturbation, and that even the sign of the sensitivity will result in eigenvector displacement fields where the rigid
is wrong unless a very small perturbation is used. The new body rotations are comparatively large relative to the actual
SA method, however, yields accurate sensitivities. The OFD deformations of the finite elements, and therefore the tradi-
method is used as a reference, and it should be noted that tional SA method yields completely wrong results unless very
the inaccuracies of the OFD method for the perturbations small perturbations are used.
58

5.2 Eigcnfrequencies of transversely vibrating square plate
with in-plane loads
The second example concerns a square plate which is clamped
at one edge and simply supported at two opposite edges.
The plate is subjected to a uniformly distributed in-plane
load of magnitude 200 N / m at the fourth edge. The finite
element model consists of 100 9-node isoparametric Mindlin
plate elements, see Fig. 3. The plate has the same dimensions
and material properties as the plate in the previous example.
are larger for this example due to the boundary conditions
considered.
Now, consider the case where the plate is subjected to the
in-plane load. This implies a lowering of the first eigenfre-
quency from 3.167 ttz to 2.690 ttz. Again design sensitivities
are calculated, and the results are given in Table 4.
Table 4. Computed derivative of lowest eigenfrequency fl with
respect to side length L for plate with in-plane load = 200 N/re.
Plate length/thickness ratio is 1000

Plate length OFD Traditional New

perturbation method SA method SA method
AL Z~fl
L AL
10-2 -6.511 103053 -6.587
10-3 -6.583 11180 -6.587
10-4 -6.597 1122 -6.587
Fig. 3. Finite element model of square plate with in-plane load
10-5 -6.598 106.4 -6.587
For comparison purposes, it is first assumed that the in-
10-6 -6.598 4.714 -6.587
plane load is absent and then the lowest frequency is obtained
as f l = 3.167 ttz. The shape design variable is again taken 10-7 -6.570 -5.451 -6.587
to be the side length L of the plate, and the sensitivities of
10-8 -6,322 -6.383 -6.587
the lowest eigenfrequency are given in Table 3.
10-9 -5,477 -5.847 -6.587
Table 3. Computed derivative of lowest eigenfrequency fl with
respect to side length L for plate without in-plane load. Plate 10-10 -0,384 -7.325 -6.587
length/thickness ratio is 1000
This example emphasizes the importance of using our
Plate length OFD Traditional New
modified version of the SA method. The new SA method
perturbation method SA method SA method is independent of the perturbation used, whereas a compar-
zSL ison of the results obtained by the OFD method and the
-L-- z~L traditional SA method for this example reflects difficulties in
10 - 2 -6.231 73253 -6.325 obtaining reliable sensitivities. This is due to both compara-
tively larger rigid body rotations relative to actual deforma-
10 - 3 -6.315 7951 -6.325
tions of the finite elements, and the necessity of performing
10 - 4 -6.324 796.3 -6.325 two design sensitivity analyses. First the stress sensitivities
are determined in a static sensitivity analysis and then the
10 - 5 -6.325 7.401 -6.325
eigenfrequency sensitivities are calculated including the ini-
10 - 6 -6.325 1.709 -6.325 tim stress stiffening effects. Inaccuracies in the static sensi-
tivity analysis hereby become accumulated in the dynamic
10 - 7 -6.320 -5.521 -6.325
sensitivity analysis and this is the reason why there are small
10 - 8 -6.271 -6.184 -6.325 differences in the results obtained by the OFD method and
the new SA method. The OFD method is based on first-order
10 - 9 -4.948 -5.804 -6.325
forward finite differences and as the problem is strongly non-
10 - 1 -8.880 -6.837 -6.325 linear, this gives small inaccuracies although small perturba-
tions are used.
It is seen from Table 3 that there is a good agreement Finally, the same example is considered but now the thick-
between the results obtained by the OFD method and the ness is increased to 20.0.10 - 3 m, i.e. the length/thickness
new SA method for the perturbations 10 - 4 - 10 - 6 , but for ratio is 50, and the in-plane load is set to 1.5 .106 N/m.
smaller values of the perturbation, the OFD method starts Without the in-plane load the lowest eigenfrequency be-
to diverge due to numerical round-off errors. The traditional comes f l = 62.97 IIz, and with the in-plane load included,
SA method starts to diverge for perturbations smaller than the lowest eigenfrequency f l is reduced to 53.93 Hz. The
10 - 8 and does not reach the same sensitivity value as the shape design variable is still taken to be the side l'ength L of
two other methods. the plate and the design sensitivities obtained are given in
The errors in the sensitivities obtained by the traditional Table 5.
SA method are seen to be larger for this example than for For this example, there is seen to be good agreement be-
the previous one. This is to be expected, as the rigid body tween the OFD method and the new SA method. The tra-
rotations relative to actual deformations of the finite elements ditional SA method is seen to converge to the same value
 59

Table 5. Computed derivative of lowest eigenfrequency fl with as the analytical method.

respect to side length L for plate with in-plane load = 1.5 .108 The new method of semi-analytical design sensitivity
N/re. Plate length/thickness ratio is 50 analysis has been tested for many different finite element
types and problems, and it has always performed excellently.
Plate length OFD Traditional New
It eliminates the original deficiencies of the semi-analytical
perturbation method SA method SA method method, and thus extends the area of applicabifity of the
~L a_4/k ajk method.
'-L- AL AL
10 - 2 -127.9 5024 -129.5 Acknowledgement
10 - 3 -129.5 430.0 -129.5 This work was partly supported by the Danish Technical Research
10 - 4 -129.6 -72.93 -129.5 Council, Programme of Research on Computer Aided Design.

10 - 5 -129.7 -123.7 -129.5

References
10 - 6 -129.7 -128.8 -129.5
Barthelemy, B.; tiaftka, R.T. 1990: Accuracy analysis of the semi-
10 - 7 -129.7 -129.3 -129.5 analytical method for shape sensitivity analysis. Mech. Struct.
10 - 8 -129.7 -129.4 -129.5 Mach. 18, 407-432

10 - 9 -129.7 -129.4 -129.5 Bratus, A.S.; Seyranian, A.P. 1983 : Bimodal solutions in eigen-
value optimization problems. Prikl. Matem. Mekhan. 47, 546-554
10 - l -131.1 -129.5 -129.5 Cheng, G.; Gu, Y.; Zhou, Y. 1989: Accuracy of semi-analytical
sensitivity analysis. Finite Elements in Analysis and Design 6,
as obtained by the other two methods, but it is still depen- 113-128
dent on the chosen value of the perturbation. The sensitiv- Cheng, G.; Liu, Y. 1987: A new computational scheme for sensi-
ities obtained by the traditional S-A method are much bet- tivity analysis. Eng. Opt. 12, 219-235
ter for this example than for the previous one, because the
length/thickness ratio has been decreased. This results in Cheng, G.; Olhoff, N. 1993: Rigid body motion test against error
in semi-analytical sensitivity analysis. Comp. CJ Struct. 46, 515-
smaller rigid body rotations relative to actual deformations
527
of the finite elements, and therefore the errors associated with
the fact that - ( A k / A a j ) d r ~ O, are reduced. Courant, R.; Hilbert, D. 1953: Methods of mathematical physics,
(Volume 1). New York: Interseience Publishers

6 Conclusions Fenyes, P.A.; Lust, R.V. 1991: Error analysis of semi-analytical

The method of semi-analytical design sensitivity analysis
where the derivatives of finite element matrices are approx-
imated by first-order finite differences is fully reliable for
eigenvalue problems where the eigenvector displacement field
entails small rigid body rotations relative to actual deforma-
tions of the finite elements.
However, for problems involving beam, plate or shell
structures, rotations are generally not negligible relative
to actual deformations and the traditional semi-analytical
method may become inaccurate as illustrated by the two ex-
amples in this paper. This inaccuracy problem is entirely due
to the approximate numerical differentiation of the element
matrices, but application of a method of "exact" numerical
differentiation, which is based on simple first order finite dif-
ferences, completely eliminates the error problem. Similar
types of error behaviour and method of elimination has ear-
lier been discussed for shape design sensitivity analysis of
beam, plate and shell structures subjected to static loads by
Cheng and Olhoff (1993) and Olhoff et al. (1993).
The new method of semi-analytical design sensitivity
analysis may be considered as a hybrid between the semi-
analytical and the analytical method. Hence, it possesses all
important advantages of the semi-analytical method regard-
ing relative ease of implementation, applicability for a wide
range of different finite elements, adaptivity to existing finite
element software, and computer cost efficiency, while it also
possesses the important property of being equally as accurate
displacement derivatives for shape and sizing variables. AIAA J.
29, 271-279
Haftka, R.T.; Adelman, H.M. 1989: Recent developments in struc-
tural sensitivity analysis. Struct. Optim. 1, 137-151
Hang, E.J.; Choi, K.K.; Komkov, V. 1986: Design sensitivity anal-
ysis of structural systems. New York: Academic Press
Mlejnek, H.P. 1992: Accuracy of seml-analytical sensitivities and
its improvement by the "natural method". Struct. Optim. 4,
128-131
Olhoff, N.; Rasmussen, 3. 1991: Study of inaccuracy in semi-
analytical sensitivity analysis - a model problem. Struct. Optim.
3, 203-213
Olhoff, N.; Rasmussen, J.; Lund, E. 1993: Method of "exact" nu-
merical differentiation for error elimination in finite element based
semi-analytical shape sensitivity analysis. Mech. Struct. Mach.
21, 1-66
Pedersen, P.; Cheng, G.; Rasmussen, J. 1989: On accuracy prob-
lems for semi-analytical sensitivity analysis. Mech. Struct. Mach.
17, 373-384
Zienkiewicz, O.C.; Campbell, J.S. 1973: Shape optimization and
sequential linear programming. In: Gallagher, R.H.; Zienkiewicz,
O.C. (eds.) Optimum structural design, theory and applications,
pp. 109-126. London: Wiley and Sons
Received Nov. 22, 1993