Transactions on the Built Environment vol 13, © 1995 WIT Press, www.witpress.

com, ISSN 1743-3509

MINLP optimization approach to the synthesis

of large scale mechanical structures
S. Kravanja, Z. Kravanja, B.S. Bedenik
Faculty of Civil Engineering, University ofMaribor,
Smetanova 17, SI-62000 Maribor, Slovenia

Abstract

The paper describes Mixed-Integer Nonlinear Programming approach

(MINLP) to the synthesis of large scale mechanical structures. The MINLP
optimization approach is performed through three steps. The first one is a
generation of a mechanical superstructure of different topology and standard
dimension alternatives, the second one is a development of an MINLP model
formulation and the last one is a solution of the formulated MINLP problem.
The Modified Outer-Approximation/Equality-Relaxation algorithm and a
Linked Two Phased MINLP Strategy have been proposed to solve large scale
mechanical problems. A practical example of the MINLP synthesis of the
three-element hydraulic steel gate structure Intake Gate in Aswan indicates
that significant savings can be obtained by the proposed simultaneous
topology and parameter optimization.

1 Introduction

An activity to generate automatically structural design alternatives and then to

select the best one can be regarded as the synthesis of mechanical structures.
The synthesis of mechanical structures thus comprises the generation and the
selection of different structure alternatives as well as a decision on which
structure elements integrate the structure and how they should be
interconnected. The sizes and other parameters should be determined, too. The
former implies discrete decision making, while the latter implies making
choices inside the continuous space. Thus, from the conceptual standpoint the
synthesis problem correspond to a discrete/continuous optimization problem
 Transactions on the Built Environment vol 13, © 1995 WIT Press, www.witpress.com, ISSN 1743-3509

286 Structural Optimization

which mathematically gives rise to Mixed-Integer Linear Programming

(MILP) or Mixed-Integer Nonlinear Programming (MINLP) problem.
As the majority of problems in the mechanics are nonlinear, the MINLP
optimization approach has been applied to solve problems of the synthesis in
mechanics. The MINLP approach enables both topology and parameter
optimization simultaneously. In the most cases such problems are large,
nonconvex and highly nonlinear. Three steps have been developed for the
solution of this task:
• Generation of mechanical superstructure for different structure/topology
and other design alternatives that are all candidates for a feasible and
optimal solution.
• Development of special MINLP model formulation for the postulated
mechanical superstructure in equation oriented environment. Besides
continuous variables that are used for continuous decisions (dimensions),
a set of integer (binary) variables is declared to denote potential existence
of structure elements and standard dimensions.
• Solution of a defined MINLP model, performed by the Modified Outer-
Approximation/Equality-Relaxation algorithm by Kravanja and
Grossmann[l] and by proposed Linked Two Phased MINLP Strategy
which in the simultaneous MINLP optimization approach yields an
optimal topology, standard dimensions and continuous parameters.
The simultaneous MINLP optimization approach described in this paper gives
essentially better results when compared to the parameter and/or standard
dimension optimization performed at fixed structure/topology.

2 Generation of mechanical superstructures

The synthesis of mechanical structures performed by simultaneous MINLP

optimization approach begins by the generation of an MINLP superstructure
in which different topology and other design alternatives are embedded to
compete for a feasible and optimal solution. While topology alternatives
represent selection and interconnection of corresponding structure elements,
design alternatives usually comprise standard dimensions. Different materials
and loads may also be included as design alternatives.
The superstructure is typically described by the unit representation:
structure elements and their interconnection nodes. Each potential topology
alternative is represented by a special number and a configuration of selected
structure elements and their interconnections, and each structure element may
have in addition different standard dimension (and material) alternatives.
Therefore, the main goal is to find, within the given superstructure, a feasible
structure that is optimal with respect to topology, standard dimensions,
performance and other continuous design parameters.
 Transactions on the Built Environment vol 13, © 1995 WIT Press, www.witpress.com, ISSN 1743-3509

Structural Optimization 287

3 MINLP model formulation for mechanical superstructures

After a mechanical superstructure is defined, the synthesis is followed by a

development of an MINLP model formulation. It is assumed that a general
nonconvex and nonlinear discrete/continuous optimization problem can be
formulated as an MINLP problem (MINLP-G) in the form:
min z = c^y + f(x)
s.t. h(x) = 0
g(x) < 0 (MINLP-G)
By + Cx<b
x e X= {x e R": x^ < x < x^}
^E7={o,ir
where x is a vector of continuous variables specified in the compact set X and
y is a vector of discrete, mostly binary 0-1 variables. Functions fix), h(x) and
g(x) are nonlinear functions involved in the objective function z, equality and
inequality constraints, respectively. Finally, By+Cx<b represents a subset of
mixed linear equality /inequality constraints.
This general MINLP model formulation has been adapted for the
synthesis of mechanical superstructures (MINLP-SMS). The resulted
formulation that is more specific, particularly in variables and constraints, is
given in the form:
min z-c^y-\-f(x)
s.t. h(x) = 0
fW^O
A(x)<a
Ey<e (MINLP-SMS)
Dy*+R(x)<r
Ky*+l(d™}<k

x e X= {x e R": x^ < x < x™}

fEy={o,ir
The MINLP model formulation for mechanical superstructures may be
described as follows:
• Included are continuous variables x={d, p} and discrete binary variables
*}- Continuous variables are partitioned into design variables
,d*} and into performance (nondesign) variables p, where
subvectors d™ and d* stand for continuous and standard dimensions,
respectively. Subvectors of binary variables j^and y** denote potential
 Transactions on the Built Environment vol 13, © 1995 WIT Press, www.witpress.com, ISSN 1743-3509

288 Structural Optimization

existence of structure elements inside the superstructure and potential

selection of their standard dimension alternatives, respectively.
• The economical objective function z involves fixed cost charges in the
linear term c*y and dimension dependant costs in the termX*).
• Parameter nonlinear and linear constraints h(x)=0, g(x)<0 and A(x)<a
represent the rigorous system of the design, loading, stress, deflection,
stability, etc. constraints known from the structural analysis.
• Integer linear constraints Ey<e are proposed to describe relations between
binary variables.
• Mixed linear constraints Dy"+R(x)<r restore interconnection relations
between currently selected or existing structure elements (corresponding
y=l) and cancel relations for currently disappearing or nonexisting
elements (corresponding X=0).
• Mixed linear constraints Ky*+L(d™)<k are proposed to define the
continuous design variables for each existing structure element. The space
is defined only when the corresponding structure element exists (y=l),
otherwise it is empty.
• Mixed linear constraints Py+S(d**)<s define standard design variables d**.
Each standard dimension d* is determined as a scalar product
d* = Z<7/.y? between its vector of standard dimension constants grand its
/e/
vector of binary variables y*. Only one discrete value can be selected for
each standard dimension since £ yf = 1 •
/e/

4 Solving an MINLP problem

After the MINLP model formulation is developed, the defined MINLP

synthesis problem is solved by the use of a suitable MINLP algorithm and
strategies.
A general MINLP class of optimization problems can be solved in
principle by the following algorithms and their extensions: the Nonlinear
Branch and Bound, NBB, proposed and used by many authors, e.g. Beale[2],
Gupta and Ravindran[3]; the Sequential Linear Discrete Programming
method, SLOP, by Olsen and Vanderplaats[4]; the Extended Cutting Plane
method by Westerlund et al.[5]; Generalized Benders Decomposition, GBD,
by Benders[6], Geoffrion[7]; the Outer-Approximation /Equality Relaxation
algorithm, OA/ER, by Kocis and Grossmann[8]; and the Feasibility
Technique by Mawengkang and Murtagh[9].

4.1 The Modified OA/ER algorithm

The OA/ER algorithm seems to be the most efficient algorithm to solve large
scale MINLP problems, when NLP subproblems are expensive and difficult to
 Transactions on the Built Environment vol 13, © 1995 WIT Press, www.witpress.com, ISSN 1743-3509

Structural Optimization 289

solve. It was proved on some numerical examples that the OA/ER algorithm
requires much fewer major iteration than the NBB, ECP or GBD methods.
The OA/ER algorithm consists of solving an alternative sequence of NLP
and MIL? master problem optimization subproblems. The former corresponds
to the optimization of parameters for a mechanical structure with fixed
structure topology and standard dimensions and yields an upper bound to the
objective to be minimized. The latter involves a global approximation to the
superstructure of alternatives in which a new topology and standard
dimensions are identified so that its lower bound does not exceed the current
best upper bound. The search is terminated when the predicted lower bound
exceeds the upper bound.
The OA/ER algorithm as well as all other mentioned MINLP algorithms
do not generally guarantee that the solution found is the global optimum. This
is due to the presence of nonconvex functions in the models that may cut off
the global optimum. In order to reduce undesirable effects of nonconvexities
the Modified OA/ER algorithm was proposed by Kravanja and Grossmann[l]
by which the following modifications can been applied for the master
problem:
• deactivation of linearizations,
• decomposition and deactivation of the objective function linearization,
• use of the penalty function,
• use of the upper bound on the objective function to be maximized, and
• global convexity test and validation of the outer approximations.

4.2 MINLP strategies for mechanical problems

Recently two different MINLP strategies for medium scale MINLP
optimization problems in the mechanics were developed. The first one is the
Two Phased MINLP Approach which performs topology and standard
dimension optimization separately in two phases, see Kravanja et al.[10]. This
strategy needs double number of MINLP iterations to converge to the final
solution, which may not be optimal since interaction between topology and
standard dimensions is not accounted for in such a separate and a sequential
optimization approach.
Then the Single Phase MINLP Approach was developed to perform
simultaneous topology, parameter and standard dimension optimization
directly by a single MINLP phase, see Kravanja et al.fll]. However, the
initialization scheme is weak since initial standard dimensions selected may
be bad or infeasible. This strategy also exhibits poor convergence when it is
applied to a large scale MINLP problems.
In order to overcome the drawbacks, a Linked Two Phased MINLP
Strategy has been proposed as a combination of the above mentioned MINLP
approaches. It starts with topology optimization, while all standard
dimensions are relaxed into continuous parameters. Topology and continuous
parameter optimization is combinatorically easier to solve and accumulates
 Transactions on the Built Environment vol 13, © 1995 WIT Press, www.witpress.com, ISSN 1743-3509

290 Structural Optimization

much better global linear approximation of the superstructure model

representation in comparison to the Single Phase MINLP Approach. When the
optimal topology is found, standard dimensions are re-established and the
simultaneous topology, parameter and standard dimension optimization is
then continued based on the obtained linear global approximation until the
optimal solution is found.

5 MINLP synthesis of Intake Gate

The synthesis of some constructed vertical-lift hydraulic steel gates structures

has been performed as a comprehensive comparative design research work.
The synthesis of the entire roller gate Intake Gate is presented as an example.
Intake Gate was constructed by the Slovenian firm Metalna Maribor for
Aswan II, Egypt. There were eight identical turbine intakes erected. Each is
regulated by the three-vertical-elements assembled Intake Gate[10]. The
actual self manufacturing and transportation costs are 70518 $ per gate with
actual topology 4-4-4/11 (4 horizontal girders for each of the three main gate
elements and 11 vertical girders for the entire gate), see Figure la.
First, the gate superstructure has been generated in which all possible
structures are embedded by topology variation between 4 to 6 horizontal
girders for each main gate element and 5 to 9 vertical girders for the entire
gate. Also 11 standard dimension alternatives have been proposed for all
structure elements including standard thicknesses of sheet-iron plates from 10
to 40 mm. Actual material St 44-2 has been considered.
A general mathematical optimization model for roller and sliding gate
structures GATOP (GATe Optimization) has been developed according to the
proposed MINLP model formulation for mechanical superstructures. As an
interface for mathematical modelling and data inputs/outputs GAMS (General
Algebraic Modelling System) by Brooke et al.[12], a high level language, has
been used. The self manufacturing and transportation costs have been
accounted for in the economic type of the objective function, subjected to the
given design, material, stress, deflection and stability constraints.
The synthesis has been carried out by the user friendly version of MINLP
computer package TOP (Topology Optimization Program)[10] which is the
computer implementation of the Modified OA/ER algorithm and the Linked
Two Phased MINLP Strategy. MINOS by Murtagh and Saunders[13] has
been used to solve the NLP subproblems and OSL[14] to solve the MILP
master problems.
The optimal solution yields costs of 49783 $ at the optimal topology 4-5-
5/5 (Figure Ib), that represents 29.4 % of net savings when compared to the
costs of the erected gate. The Modified OA/ER algorithm accompanied with
The Linked Two Phased MINLP Strategy converged very fast: only five
major MINLP iterations were needed (4 MILP and 5 NLP subproblems). 869
equations with 3169 nonzero elements, 327 continuous and 15 currently fixed
 Transactions on the Built Environment vol 13, © 1995 WIT Press, www.witpress.com, ISSN 1743-3509

Structural Optimization 291

discrete variables are included in the first NLP subproblem and 2032
equations with 1764 continuous and 54 discrete variables are included in the
MILP master problem at which the optimal solution was found. Only 293
seconds of the CPU time on the computer VAX 4100 were spent.

\LLLL IJL-lLli-L

r 1
2900 a) 2921 b)

8000 2700 8000 2589

2400 2490

-r \I

8200 mm -8200mm-

u i i i i i i 1 1 1

Figure 1: Intake Gate structure, vertical and horizontal cross-sections:

a) actual topology 4-4-4/11
b) optimal topology 4-5-5/5, 29.4% of net savings

6 Conclusions

The proposed MINLP approach to the synthesis of mechanical structures

performs simultaneous topology, parameter and standard dimension
optimization in one uniform calculation process. The Modified OA/ER
algorithm accompanied with the Linked Two Phased MINLP Strategy is used
to solve nonconvex large scale MINLP problems in the mechanics. Presented
is the example of the synthesis of roller gate Intake Gate, Aswan. The
obtained results shows that this method enables 29.4% savings of investment
costs when compared to the actual design obtained by the classical method.
 Transactions on the Built Environment vol 13, © 1995 WIT Press, www.witpress.com, ISSN 1743-3509

292 Structural Optimization

References

[1] Kravanja, Z. and Grossmann, I.E., New Developments and Capabilities

in PROSYN - An Automated Topology and Parameter Process Synthesizer,
Computers chem. Engng , 1994,18(11/12), pp. 1097-1114.
[2] Beale, E.M.L., Integer Programming. The State of the Art in Numerical
Analysis, ed D. Jacobs, pp. 409-448, Academic Press, London, 1977.
[3] Gupta, O.K. and Ravindran, A., Nonlinear Mixed Integer Programming
and Discrete Optimization, Progress in Engineering Optimization, ed R.W.
Mayne and K.M. Ragsdell, pp. 297-520, New York, 1984.
[4] Olsen, G.R. and Vanderplaats, G.N., Method for Nonlinear
Optimization with Discrete Design Variables, AIAA Journal, 1989, 27(11),
pp. 1584-1589.
[5] Westerlund, T., Pettersson, F. and Grossmann, I.E., Optimization of
pump configurations as a MINLP problem, Computers chem. Engng., 1994,
Vol. 18(9), pp. 845-858.
[6] Benders, J.F., Partitioning Procedures for Solving Mixed-variables
Programming Problems, Numerische Mathematik, 1962, 4, pp. 238-252.
[7] Geoffrion, A.M., Generalized Benders Decomposition, Journal of
Optimization Theory and Applications, 1972, 10(4), pp. 237-260.
[8] Kocis, G.R. and Grossmann, I.E., Relaxation Strategy for the Structural
Optimization of Process Flowsheets, Ind. Engng Chem. Res., 1987, 26, pp.
1869-1880.
[9] Mawengkang, H. and Murtagh, B.A., Solving Nonlinear Integer
Programs with Large-Scale Optimization Software, Annals of Operations
Research, 1986, 5, pp. 425-437.
[10] Kravanja, S., Kravanja, Z., Bedenik, B.S. and Faith, S., Simultaneous
Topology and Parameter Optimization of Mechanical Structures, in Numerical
Methods in Engineering '92 (ed Ch. Hirsch et al.), pp. 487-495, Proceedings
of the First European Conference on Numerical Methods in Engineering,
Brussels, Belgium, 1992, Elsevier, Amsterdam, 1992.
[11] Kravanja, S., Kravanja, Z. and Bedenik, B.S., MINLP Optimization of
Mechanical Structures, in Structural Optimization 93 (ed J. Herskovits), pp.
21-28, Proceedings, Volume I, The World Congress on Optimal Design of
Structural Systems, Rio de Janeiro, Brazil, 1993, Federal University of Rio de
Janeiro, 1993.
[12] Brooke, A., Kendrick, D. and Meeraus, A., GAMS - A User's Guide,
Scientific Press, Redwood City, CA, 1988.
[13] Murtagh, B.A. and Saunders, M.A., MINOS User's Guide, Technical
Report SOL 83-20, System Optimization Laboratory, Department of
Operations Research, Stanford University, 1985.
[14] OSL, Optimization Subroutine Library, From IBM, Release 2.