Вы находитесь на странице: 1из 4

G. A.

Mohr
Applied ~om~utatiooai

A~~-~~~~it~

Mechanics in BngineeriniS;, 7 Marine Avenue, St Kilda, Metboume, Australia

element models of a

wide range of probkns are aptimixed using the steepest descent

method and merit functions appropriate to each particular problem. Here, in part I, we deal with typic4
structural problems in w&h the merit functions are base4 an the stresses caused by loading of the
structure. In part II we shall deal with problems ia which the structure is loaded by a fluid flow and we
alter the structure shape to improve the efficiency of the ~uid/stt~t~e
interaction.

fn the first two decades since the i~tr~~c~on


of the
finite element method fl] little attention was directed
at op~~z~~g f&&l: element mode& [2& of ~rt~G~ar
note, however, bemg the pro& of GdfatEy and Gaitagher [3j in which &fullystressed design (FSD) and the
steepest descent method is used for this purpose,
In the last decade, however, interest in the optima
ization of finite element models has increased [4-g],
in part because of the av~~ab~lity of high speed
computers f9f.
In the present paper stress criteria and the steepest
descent method are applied to a simple truss probfem
studied by Wang using sequeatial linear prog~rn~
ming {IO], and to the arch problem studied by M&r
using optim~~ity criteria[I1] in both cases yielding
~rnpr~v~ rest&s. The techniques used can readily be
extend& to frame and sheh probiems and, as will be
demonstrated in part ff, to many other types of
problem.

The
prob~m
of op~rnj~~~ the ~r~s~-~ct~o~~
areas of the members of simple truss structures
provides a useful introductory example of the application of the steepest descent method to finite eiem~~t
models. The procedure used here uses the following
strategy:

fi) FSD is first used to obtain an ~rnp~~~ design


which witI usually be close to the optimum sofution.
(2) The steepest descent method is applied to the
FSD solution to obtain the optimum solution, that is
the change in each of the design variables {xi 3 is given
by 021
S{XJ = -iff5~/zkj)

= -I,&],

(XI

where R. is the step length in the direction of tke


gradient vector fg) andfis the merit function which

is here the volume of material in the truss,f = G AILi,


where Ai, 15~are the cross-~ti~na1 areas and lengths
of the truss members.
The gradient vector is ikst determined by a perturbation 6.~~of each design vtiabk
Qn this case the
member areas A,) in turu and ~~~i~t~ng the resubing
change in the merit function, t?& giving

~r~~b~~~o~
requiring a re-analysis of the strutture to determine the associated value of SA.
The o~~irnurn step Iengtb in eqn fij_ that which
minimizes f, is found by a search, using gradually
increased trial values of 2 until the minimum f is
bracketed and then bisection or interpolation (usually
Iinear or quad~~~rl2~
is used to locate the minimum, T&n a new gradient vector is ~a~~u~~~edby
successive ~rtur~~~~s
of the cmrent vafues of the
design variables and this used for the next search and
the procedure is te~inated
when no further reduction in f can be obtained,
Figure I shows a sisraple empIe probkm in which
there are two a~ternat~~~ foad cases, la such situations
the element areas are rescakd in the FSD or gradient
vector calculation using the largest stress ratio
(rr/TLIM or a/CLIM, where TLIM and CLXM are
the tensile and compressive stress limits) given by the
afternative load cases,
fn this type of problem the foI~ow~ng rest~~~t~o~s
are r~m~ende~
is&

(1) In capsulating fgj the element areas are doubled, that is dn, = xi.
(2) Only one pre~~rnj~ar~ FSD step is used (fu~her
Use of FSD until convergence is obtained results in
member 4 vanishing in the present probEem and this
is not the optimum sofution).
(3) fn all but the first steepest descent search, a first
step length of IO-% is used. Then if no decrease inf
is observed with this test step the search direction is

1218

G. A. Mohr

unprofitable and the current solution is taken as the


estimate of the optimum. Then, using these recommendations, the results in Table 1 are obtained.
The present FSD with the steepest descent procedure gives a better result than that obtained by
Wang [lo] using sequential (stepwise) linear programming (SLP). Note, however, that at the end of each
LP step Wang scales all the element areas by the
largest stress factor encountered in any element and
this cannot yield the true optimum solution. Note
also that all solutions given by the present method are
valid solutions because each element is factored according to its stress level at each step.
The present procedure could easily be applied to
the optimization of frame structures. In these optimizations the sizes of the elements of the frame would
follow the same process used for the foregoing truss
problem, except that section limits (that is, a minimum allowable size for each element) must be
specified, but inclusion of these in finite element
programs is a trivial exercise.
Finally, in relation to optimum truss problems,
the classical Hencky-Prandtl
network solutions of
Michell[13] are worth remembering. These are for
very special cases but for the case of line (rather than
point) support Rozvany and Gollub [14] report for
this less restricted case solutions correspond to fields
of constant strain, which give layouts consisting of a
finite number of straight members and such solutions
may have more practical applications.

OITIMlZATlON

OF AN ARCH

Next we optimize the shape of an arch structure


using the curved quadratic element of Mohr and
Garner [ 153.The element has freedoms M,t, 4 at three
nodes and the extensional strain, flexural curvature
and transverse shear strain are given by
c = &i/ax
X=

+ v/R

(3)

~a~f~~~l~~~
+ iqjak

(4)

y = adjax,-

(5)

4,

418

-I
146

Fig. I. Truss with two alternative loadings. Element numbers are underlined.

where
{j> = ((s2 - s)j2, 1 -s?, ($2+ s)/2)

(7)

{f} = (2s - 1, -4s, 2s $ I}/L

(8)

and a = cos c(, b = sin tl, where CLis the local slope of
the element.
Using two point Gaussian quadrature the element
stiffness matrix is given as

k = CMW~)B:B, + W)B;B, + P(~GA/~)B:BJ,


(9)
where B is a penalty factor and the element is assumed
to have constant cross-section (and hence A and I).
For thick arches fl = 1 but for thin arches (as in the
example studied here) p = 100t2, where t is the element thickness [I 51.
In the example problem of Fig. 2, however, ,8 = I
was used as the arch is very thin (span/
thickness = 160) and hence the shear strains are not
taken into consideration in the optimization process.
The merit function used is the factor weight which
is the material volume scaled according to the stress
ratios in each element, that is we assume unit density
for convenience. This is then calculated as
,f = y LwtF,
i=,

where x is a curvilinear coordinate along the element


and u, u are respectively the tangential and normal
displacements.
Using quadratic interpolation with a dimensionless
coordinate s = 2x/L (s = - I+ l), where L is the
element arc length, for u, v, 4 the strain interpolation matrix is given by

E = 30.000
TLIM = 20
CLIM = 15

(10)

where L, Wand t are the element arc length, breadth


and depth, respectively, and F is the largest vaiue
obtained by dividing the two extreme fibre stresses by
the tensile and compressive stress limits.
In shape optimization applications we use straight
elements, however, as movement of the central node

Finite element optimization of structures-l


Table 1. Optimization solutions for the problem of Fig. 1
using FSD with steepest descent
A,
Initial
FSD {I step)
Search 1 (A =0.09)
Search 2 (A = 10sm)
(A = 0.03)
Search 3 (A = lO-m)
SLP solution

(Wana IlOll

A,

A,

A,

0.261

0.121

0.084

4.837

of an element during

gradient vector calculations


produces a local arching effect which leads to poor
results. Hence only the elevations of nodes 3, 5,7 and

9 in Fig. 2 are specified as shape function variables


and 10% increments in these are used to calculate the
gradient vector.
The results for the caSe of a horizontally distributed load are given in Table 2. After completion
of steepest descent searching the resulting shape is
closely parabolic as expected, largely eliminating the
bending stresses. In this thin arch, however, the
remaining bending stresses still swamp the extensional stresses.
To obtain the final optimum shape, therefore, the
arch shape is taken as parabolic; this is the design
shape for elimination of bending. Scaling of this
shape is then used to obtain the final optimum shape.
For this, a search with a gradually increasing scale
factor is used to minimize the factor weight, which is
now predominantly a function of the extensional
stresses.
The results are in reasonable agreement with the
exact solution [16] and those of the optimality criterion method [Ill given in Table 2. In the latter,
constant strain shell elements are elevated according
to their bending stresses in order to reduce them, and
scaling is used simultaneously to minimize the factor
weight, calculated only from the extensional stresses.
The present method is to be recommended, though,
because it does not presume that bending should be
eliminated.
q=

Table 2. Optimization results for an arch with horizontally


distributed load

1.440 2.000 1.800 2.000 45.040


0.231 0.264 0.104 0.143 4.421
0.296 0.189 0.139 0.085 4.200
0.316 0.224 0.123 0.055 4.152
0.310 0.213 0.126 0.064 4.148
Test step increased f so stop 4.178
0.364

1219

Y3

Initial
2
Search 1 (A = 0.04)
3.06
Search 2 (2 = 0.03)
2.49
Search 3 (A = 0.012)
2.54
Search 4 (A = 0.035)
2.67
Search 5 (1 = 0.005)
254
Search 6 (A = 0.~03)
2.53
Design shape
2.8
Scaled shape (A = 2.19) 6.13
OCM technique [ 1I]
5.6
Theoretical solution [ 161 6.06

Ys

Y7

Y9

4
6
8
4.39 5.28 1.35
5.13 4.99 6.94
4.80 5.39 6.67
4.50 5.35 6.25
4.53 5.37 6.19
4.53 5.37 6.19
4.8
6.0
6.4
IO.51 13.14 14.02
9.3 11.6 13.5
10.40 12.99 13.86

f
71.11
26.42
16.30
Ii.94
8.95
8.70
8.69
4.14
2.95
-

An alternative procedure is to base the factor


weight calculation of the present method only on the
extensional stresses from the outset and again apply
steepest descent. This gives an arch of similar height
to that expected but, with the triangular initial shape
of Fig. 2, the solution tends towards a w shape
(with f = 2.73). Such a solution would involve extremely high flexural stresses and is not feasible.
Finally, for the case of a surface distributed load,
the steepest descent/scaling method is applied once
again, giving the results in Table 3. These are reasonable but not as good as those for the horizontally
distributed load case, in part because the true optimum shape is not now parabolic, as assumed prior to
scaling. Overall, however, the effectiveness of steepest
descent techniques in arch and shell problems is
demonstrated, as it will be for fluid loading problems in part II of this paper.
The present procedure can easily be applied to
three-dimensional and shell problems, as can that of
Mohr [I 11. Once again, as noted at the close of the
preceding section, section limits (minimum and perhaps maximum permisssible element thicknesses)
will generally need to be specified. Additional details
such as these are easily added to finite element
programs.
CONCLUSIONS

The steepest descent procedure used in the present


work gives better results than the sequential linear
programming method of Wang for trusses [IO] and
the optimality criterion method of Mohr for
arches 1111.

10

Table 3. Optimization results for an arch with surface


distributed load
E = 20 x 106, v = 0.2
w = 4, t 3 0.2
TLIM = CLIM = 1000

A------

Four elcmcntv @ 4 m

-+

Fig. 2. Initial shape of an arch and design shape after


application of steepest descent method.

Y,

Initial
Search 1 (A = 0.04)
Search 2 (i = 0.03)
Design shape
Scaled shape(I = 1.71)
OCM technique [l I]
Theory (approx) [I I]

2
3.25
2.68
2.8
4.79
4.3
4.27

Y5

Y7

Y9

44.37 65.14 87.28 79.50


32.80
5.09 5.31 6.73 13.63
4.8
6.0
6.4
4.14
8.21 IO.26 10.94 3.53
7.4
9.2 IO.2
7.32 9.15 9.76 -

1220

G. A.

Like the latter the present procedure can be generalized to deal with shell problems, though in these a
large number of elements with thirty DOFs or more
may be involved and calculation of {g) on an element
by element basis may be impractical. This difficulty
can be overcome by using a polynomial shape function z =f(x, y) and using the relatively few coefficients of this as the design variables.
When we also seek to optimize the thickness of the
shell, another shape function t =f(.x, y) and, of
course, section limits will be needed, as they will be
in most structural problems.
There are, of course, many other possible applications of the FEM and optimization, for example to
the optimization of fluid flows coupled with the
optimization of an associated structure, and simple
examples of this type of problem are considered in
part II.

varying depth with stress, deflection and size constraints. Comput. Struct. 20, 947-955 (1985).
5. V. Thevendran and D. P. Thambiratnam, Minimum
weight design of conical concrete water tanks. Comput.
Struck. 29. 699-708 (1988).
6. T. Kam and R. R. Chang; Optimal design of laminated
composite plates with dynamic and static considerations. Compuf. Struct. 32, 387-396 (1989).
7. T. Yao and-K. K. Choi, 3-D shape optimal design and
automatic finite element regridding. Int. J. Numer.
Meth. Engng 28, 369-381 (1989).
8. E. A. Sadek, An optimality criterion method for dynamic optimization of structures. int. J. Numer. Meth.
Engng 28, 579-587 (1989).

9. E. Sikiotis, V. Saouma, M. Long and W. Rogger, Finite


element based optimization of structures on a CrayMP supercomputer. Comput. Struct. 36, 901-909
(1990).

10. C. K. Wang, Computer Methods in Advanced Structural


Analysis. International Textbook Co., New York
(1973).
II. G. A. Mohr, Design of shell shape using finite elements.
Comput. Siruct. 10, 745-749 (1979).

REFERENCES

M. J. Turner, R. W. Clough, H. C. Martin and L. J.


Topp, Stiffness and deflection analysis of complex structures. J. Aero. Sci. 23, 8055823 (1956).
F. 1. Niordson and P. Pedersen, A review of optimum
structural design. Proc. 13th Int. Cong. Theoretjeal
Applied Mech. University of Moscow (1972).
R. A. Gellatly and R. H. Gallagher, A procedure for
automated minimum weight structural design, II: Applications. Aero. Quart. 17, 216224 (1966).
R. J. Allwood and Y. S. Chung, An optimality criteria
method applied to the design of continuous beams of

12. M. J. Box, D. Davies and W. K. Swann, Nonlinear


Optimization Techniques. ICI Monograph No. 5. Oliver
and Boyd, Edinburgh (1969).
13. A. G. M. Michell, The limits of economy in frame
structures. Phil. &fag., series 6, 8, 589-593 (1904).
14. G. I. N. Rozvany and W. Goliub, Michell layouts for
various combinations of line supports-I. Int. J. Mech.
Sci. 32, 1021-1029 (1990).
15. G. A. Mohr and R. Garner, Reduced integration and
penalty factors in an arch element. Int. J. Struct. 3(l),
9-1s (1983).
16. G. A, Mohr, Finite Elements for Solids, Fluids, and
Optimization. Oxford University Press, Oxford (1992).

Вам также может понравиться