ACS6124 Multisensor and Decision Systems

Part II – Decision Systems for Engineering Design

Lecture 6: Multi-disciplinary design optimization

Prof. Robin Purshouse

University of Sheffield

Spring Semester 2019/2020

1 Lecture overview
This lecture examines the topic of multi-disciplinary design optimization (MDO), which
consists of a family of methods that can be used to perform decision support for the
design of complex engineered products, such as vehicles or critical infrastructure.
In this lecture, we will cover:
• The motivation for using MDO within a decision system;
• A detailed look at a specific MDO architecture known as collaborative optimiza-
tion;

• An engineering design example where collaborative optimization has been used

successfully;
• A high-level overview of the MDO architectures available.

2 Motivation for multi-disciplinary design optimization

2.1 The organisation of engineering design activities
The design of a complex engineered product, such as the Boeing Dreamliner or the
Jaguar I-PACE, is itself a complex endeavour, involving hundreds of designers dis-
tributed across teams both within the original equipment manufacturer (OEM) itself
and also across the OEM’s supply chain. These design teams may be geographi-
cally distributed and only communicate with each other on an infrequent basis. Some
teams may use proprietary design and simulation software that is not accessible to the
other teams, and will typically be keen to retain their own independence as designers,
rather than relinquish design responsibilities to a central authority.
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

In this setting it is not realistic for an optimization algorithm (such as those we

have considered in previous lectures) to be set up as a ‘design god’ with control of
all aspects of design and with access to all the simulations required to evaluate the
objectives and constraints. Rather, it is necessary to embed optimization algorithms
within each design team. Such embedding poses a coordination problem – how do we
ensure that all the optimizers are working to a common purpose? MDO architectures
have been developed to help answer this question.
Traditionally, owing to its roots in the aerospace industry, MDO has assumed that
design teams are organised according to engineering ‘disciplines’ – e.g. aerodynam-
ics, propulsion, structures, noise and vibration, control systems, weight and cost engi-
neering. In this set up, the disciplines typically own distinct performance criteria (e.g.
drag, thrust, strength) but share responsibility for aspects of the design. Meanwhile,
other industries (e.g. the automotive industry) tend to organise their teams according
to the components that make up the product – e.g. powertrain, body, brake systems.
In this case, the teams tend to have sole responsibility for an aspect of design, but
share the ownership of the performance criteria.
MDO, as a concept, it agnostic to the particular type of decomposition employed by
the OEM and its supply chain. However particular MDO architectures may be better
suited than others to the configuration of teams in a particular scenario.

2.2 Aims of MDO

There are three main aims that MDO is trying to deliver, in order of priority from highest
to lowest:

• Ensure that the overall design is internally consistent – i.e. that all design teams
are using the same assumptions about the shared design variables and inter-
faces. This is equivalent to saying that the design is, in principle, physically
realisable;

• Ensure that all hard constraints are satisfied;

• Achieve optimal performance against the objectives.

3 Collaborative Optimization
3.1 Design scenario
Consider a generic engineering design problem with two design teams, i  1 and
i  2. We will assume that each design team is responsible for its own local design
variables, x1 and x2 respectively, and that there are additional global design variables
x0 that are shared by the teams.
We will further assume that each design team has a number of inequality con-
straints that its local designs need to satisfy, i.e. g1,2 p.q ¤ 0, and that there is an
overall objective f p.q that the integrated, global design is aiming to minimise.

Part II – Decision Systems for Engineering Design 2

Lecture 6: Multi-disciplinary design optimization
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

System problem
choose x̂ 0, â1, â2
minimise objective f
s.t. consistency constraints

x0,1
* , a*
1 x0,2
* , a*
2
x̂ 0, â1, â2

Subsystem problem 1 Subsystem problem 2

choose x0,1, x1 choose x0,2, x2

minimise discrepancy D1 minimise discrepancy D2
s.t. local constraints s.t. local constraints

Figure 1: Collaborative optimization architecture

Finally, we assume that, in addition to the coupling between the teams introduced
by the shared design variables x0 , there is further coupling relating to parameters
needed by the teams to evaluate their constraints and also the overall objective – i.e.
that team 1 produces interdisciplinary information a1 needed by team 2, and that team
2 produces interdisciplinary information a2 needed by team 1.
Note that, for simplicity, we are considering here only two design teams; however
the following description is generalisable to problems with greater than two teams.

3.2 Architecture
Collaborative Optimization (CO) is a popular architecture that attempts to deliver the
three aims of MDO described in Section 2.2. CO defines a subsystem problem for
each design team and a further system problem that coordinates the subsystems. A
schematic of the architecture is given in Figure 1.

3.2.1 System problem

The CO system problem attempts to minimise the objective f p.q whilst maintaining
a consistent overall design. To do this it defines targets or master copies of all the
coupled variables in the design – i.e. the shared design variables x0 and the interdis-
ciplinary coupling variables a1,2 between the subsystems. These targets are denoted
using circumflex ( ˆ ) notation:

• x̂0 – targets for shared design variables;

• â1 – targets for interdisciplinary coupling variables produced by subsystem 1;
• â2 – targets for interdisciplinary coupling variables produced by subsystem 2.

Part II – Decision Systems for Engineering Design 3

Lecture 6: Multi-disciplinary design optimization
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

We will also now introduce star notation to indicate the choices presently made by
each subsystem and the corresponding results of subsystem analyses:

• x0,1 – choice made by subsystem 1 for the global design variables;

• x0,2 – choice made by subsystem 2 for the global design variables;
• x1 – choice made by subsystem 1 for its local design variables;
• x2 – choice made by subsystem 2 for its local design variables;
• a1 – results of disciplinary analysis of subsystem 1 using x0,1 , x1 and â2 ;
• a2 – results of disciplinary analysis of subsystem 2 using x0,2 , x2 and â1 .

To evaluate the consistency of the subsystems with respect to the global problem,
CO defines a set of equality constraints, known as consistency constraints based
on discrepancy functions that calculate the squared Euclidean distance between the
subsystem choices/results, and the system-level targets. The overall system problem
formulation is therefore given by:

minimise f px̂, â1 , â2 q

x̂,â1 ,â2

subject to c1 px̂, â1 , â2 q  0

c2 px̂, â1 , â2 q  0 (1)

where:


c1 px̂, â1 , â2 q 
1
2
||x0,1
x̂0 ||2 ||a1
â1 ||2

c2 px̂, â1 , â2 q 
1
2
||x0,2
x̂0 ||2 ||a2
â2 ||2 (2)

3.2.2 Subsystem problems

First consider subsystem 1. In CO, the subsystem is tasked with minimising incon-
sistencies with respect to the system problem concerning: (a) its choices for shared
design variables x0 (denoted x0,1 ); and (b) its interdisciplinary coupling variables a1 .
Further, the subsystem must also satisfy its own local constraints g1 p.q. The sub-
system is free to make any choices for its local design variables x1 , subject to these
considerations. When performing its disciplinary analyses, subsystem 1 uses the copy
of subsystem 2’s interdisciplinary coupling variables â2 supplied by the system node.
The problem formulation for subsystem 1 can be expressed as:


minimise D1
x0,1 ,x1
 12 ||x0,1
x̂0 ||2 ||A1 px0,1 , x1 , â2 q
â1 ||2
subject to g1 px0,1 , x1 , A1 px0,1 , x1 , â2 qq ¤ 0 (3)

Part II – Decision Systems for Engineering Design 4

Lecture 6: Multi-disciplinary design optimization
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

where A1 p.q is the disciplinary analysis for subsystem 1 that provides a1 , together
with any other variables needed to evaluate the local constraints (for notational sim-
plicity, these additional variables are not shown).
There is an analogous problem for the second subsystem. Its problem formulation
can be expressed as:


minimise D2
x0,2 ,x2
 12 ||x0,2
x̂0 ||2 ||A2 px0,2 , x2 , â1 q
â2 ||2
subject to g2 px0,2 , x2 , A2 px0,2 , x2 , â1 qq ¤ 0 (4)

where A2 p.q is the disciplinary analysis for subsystem 2.

3.3 Process
CO is a bi-level architecture, in which the system is the upper-level problem and the
subsystems are lower-level problems. There is therefore a requirement to solve the
subsystem problems for every candidate set of targets generated by the system prob-
lem. The sequencing works as follows:

1. Initialise the system iteration counter tsys 0

p0q p0q
2. Initialise candidate target variables x̂0 , â1 and â2
p0q

3. Execute iteration of the system problem:

(a) For each subsystem i, solve the subsystem problem:
i. Initialise the subsystem iteration counter tsub,i 0
p0q p0q
ii. Initialise candidate design variables x0,i and xi
iii. Execute iteration of the subsystem problem:
A. Perform the disciplinary analysis to generate ai sub,i
pt q
B. Evaluate the discrepancy function Di and local constraints gi p.q
C. Generate new candidate design variables x0,isub,i
pt q
1
and xi sub,i
pt q
1

D. Increment the subsystem iteration counter tsub,i  tsub,i 1

E. If a convergence criterion is satisfied then stop the subsystem solver
pt q pt q
and return x0,i  x0,isub,i and ai  ai sub,i ; otherwise go to A
(b) Evaluate the system objective f p.q and consistency constraints cp.q
(c) Generate new candidate target variables x̂0 sys
pt 1 q , âptsys 1q and âptsys 1q
1 2
(d) Increment the system iteration counter tsys  tsys 1
(e) If a convergence criterion is satisfied then stop the system solver; otherwise
go to 3.

Part II – Decision Systems for Engineering Design 5

Lecture 6: Multi-disciplinary design optimization
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

3.4 Strengths and weaknesses

A key strength of CO is that its decomposition reflects the organisational structures
used by many engineering design companies, and this has been reflected in its use
across many types of engineering design problem. A further, related, strength is the
high degree of autonomy afforded to individual design teams in determining their own
local designs, subject to coordination with the wider system.
The core weakness of CO is its use of equality constraints in the system problem
formulation. In general, optimization algorithms struggle to satisfy equality constraints
without expending a large number of function evaluations. This aspect of CO has
meant that the architecture is often regarded as inefficient. To address this issue, two
main measures have been proposed:

• Relaxation of the equality constraint, through the introduction of a small thresh-

old. The threshold may be progressively reduced as the optimization process
unfolds. The use of a threshold has been argued as appropriate for many engi-
neering problems, since modelled values are never precisely realised in practice.
• Use of surrogate models to approximate the disciplinary analyses or entire dis-
crepancy functions. The surrogate models are low-complexity approximations
of the often high-fidelity simulation models used for analysis. If these surrogates
are sufficiently cheap and accurate, they can prevent many computationally ex-
pensive evaluations that would otherwise be required by the architecture.

4 Example
4.1 Overview of the design problem
To illustrate the application of collaborative optimization, consider the design of a long-
span bridge. The organisation tasked with designing the bridge operates with two
design teams:

1. The superstructure team, who will design the towers and supporting cables for
the bridge;
2. The deck team, who will design the deck of the bridge.

The decision problem is to find a consistent design for the bridge which minimises
the build cost, subject to constraints on the strength of the bridge. In this example, the
two design teams do not share any design variables, but they do have interdisciplinary
couplings resulting from the forces that the deck exerts on the superstructure and vice
versa.
The overall configuration is shown in Figure 2; the notation used in the figure is
introduced during the following discussion.

Part II – Decision Systems for Engineering Design 6

Lecture 6: Multi-disciplinary design optimization
 whose value (1 to 4) indicates one of the four concepts The analyses will be valid only if the
from Fig. 11. The vector xsuper is a list of design variable h′ match the outputs v and h.
values (see Table 1) for the concept indicated by isuper .
The scalar fsuper. is the total cost of the superstructure,
and the vector gsuper is a list of design constraint values 3.2
for the concept indicated by isuper . We assume that the Disciplinary optimizations
design constraints are satisfied when each of the elements
in gsuper is less than zero. The scalars and vectors ideck , Figure 6 shows the optimization and
xdeck , fdeck , and gdeck have similar meanings for the deck perstructure group in collaborative
Prof. Robin Purshouse ACS6124 group.
Multisensor and Decision Systems University
thatoftheSheffield
input to superstructure op
of target values for the superstruct
horizontal cable forces h′ , and
force v ′ . The superstructure optimiz
cept isuper and associated design va
satisfy the superstructure design co
and minimize a “discrepancy func
that the discrepancy function cons
squared differences between target
values. The vector squared term (h
crepancy function is a shorthand n
Fig. 4 Coupled analyses
product (h′ − h) · (h′ − h). The opt
Figure 2: Bridge design problem: Design variables, objectives, constraints, discrepancy and function
in- is returned fr
ture optimization. Note that the
terdisciplinary couplings
Notebetween the subsystems.
that the superstructure Taken
and deck fromare
analyses Balling & Rawlings’
alysis in Fig. 6 receives the same
Collaborative optimization with disciplinary conceptual design,
coupled. To analyze the superstructure, the vertical 2000. the same outputs as the superst
forces exerted on the cables by the deck must be known. Fig. 5.
But these vertical forces depend on the weight of the deck
which is unknown to the superstructure group. Thus, the
4.2 Subsystem 1: Superstructure deck group sends to the superstructure group the scalar
v which is the vertical force exerted by the deck on the
4.2.1 Shared design variables x0,1 end of each cable. Since the deck is prismatic and the ca-
H
ble attachment points are equally spaced, this vertical
The problem has no shared
force will bedesign
the same variables, so xThe
on each cable. 0,1 cables. exert ho-
rizontal forces on the deck. But these horizontal forces
depend on the inclinations of the cables which are un-
4.2.2 Local design variables x1 known to the deck group. Thus, the superstructure group
sends to the deck group the vector h which is a list of
The local design variables for the superstructure team are, collectively, the choice of
the horizontal forces exerted by each cable on the deck.
cabling concept (fan, harp,
These semi-harp,
forces will be theandsamesuspension)
for each cableisuper
in the, together
sus- with the cable
diameters and tower geometry x . These variables are
pension and harp concepts, but they will be different
super shown in Figure 3.
in the fan and semi-harp concepts. Both v and h are
4.2.3
known as “coupling variables” between the disciplinary
Interdisciplinary coupling variables p q
groups.a1 and disciplinary analysis A1 x0,1 , x1 , â2 Fig. 6 Superstructure optimization an
The analyses may be parallelized as shown in Fig. 5.
The superstructureNote
generates a horizontal force on the deck at the location of each
that the superstructure and deck groups receive tar-
cable. Since the cables
get inputs v and h′ , by
may ′ incline different rather
respectively, amounts,
than va and
vector
h. coupling The variable
superstructure optimization
(denoted h) is needed to represent the force at each cable location.in To Fig.work
6 mustoutbe modified slightly in
the horizontal forces, the superstructure needs to know the vertical force (denoted
mization. The v)reason for modificati
exerted on the cables due to the mass of the deck. These forces therefore ceivabledepend
that there may exist a value
on decisions taken in the deck subsystem and consequently, according v ′ fortowhich
the COit is impossible for the
timization to satisfy all of its desig
architecture, are provided to the superstructure subsystem as a copy â2 – denoted
here as v 1 .
the only thing returned from the s
mization is the optimal value of th
tion, there is no way of knowing t
4.2.4 Discrepancy function D1 and local constraints g1 . pq straints were not satisfied. The mod
optimization and analysis is shown
According to CO, the superstructure subsystem should aim to minimise introduction a discrepancy of the dummy design v
Fig. 5 Parellel analyses dummy design variable which is sen
function dsuper based on the deviation of the shared variables and interdisciplinary
coupling variables from targets. The bridge example departs slightly from the standard
CO formulation here by having the subsystem also attempt to meet a target for build
cost, in addition to considering how far the horizontal forces are from target. The build
cost of the superstructure acts as a local objective, which needs to be calculated as
part of the disciplinary analysis.

Part II – Decision Systems for Engineering Design 7

Lecture 6: Multi-disciplinary design optimization
 quite remarkable when one considers that the disciplinary
groups are searching over discrete design concepts. It
has been shown previously that a gradient-based NLP
233
algorithm can be used at the system level when opti-
mization with discrete variables is being performed at
s to perform conceptual design with collaborative conceptual
the disciplinarydesign
levelof(Balling
the towers
and isGale
the 1998).
same for allpa-
This four
ization. concepts,
per and tothis
goes further reinforcedthat
demonstrate concrete
discretetower
searchconcept
over
hen disciplinary groups are free to exercise design is shown
different in Fig. can
concepts 2. Suppose that during
be performed at the the course of
disciplinary
omy, we must ensure that the overall objectives of collaborative optimization, the deck group comes up
level.
ystem are met, and that coupling between disci- with
Thethe two isconcepts
paper organizedwhose cross-sections
as follows. areexample
First, an shown in
is adequately considered. This is the role of the Fig. 3. It
problem is be
will assumed fromwhich
described the beginning thatathe
will provide points
context
Prof. Robin
m level in collaborative optimization. Purshouse
To achieve this forwhere ACS6124
cables
discussion are attached
throughout Multisensor
thetopaper.
the deck
Thenarewe and
equally Decision
spaced
will formu- Systems University of Sheffield
the system level solves an optimization problem. at 10
late m.
collaborative optimization for the example problem.
ill demonstrate that a continuous gradient-based Next, the example problem will be solved with undecom-
near programming (NLP) algorithm may be used posed optimization and analysis to provide a baseline for
ve the system-level optimization problem. This is the evaluation of results. Then results from collaborative Fig. 1 Superstructure concepts
remarkable when one considers that the disciplinary optimization will be presented and evaluated. Finally,
s are searching over discrete design concepts. It conclusions will be drawn.
een shown previously that a gradient-based NLP
thm can be used at the system level when opti-
ion with discrete variables is being performed at 2
sciplinary level (Balling and Gale 1998). This pa- The example problem
oes further to demonstrate that discrete search over
ent concepts can be performed at the disciplinary Our example consists of the design of a long-span cable-
supported bridge over a river. The total length of the
he paper is organized as follows. First, an example bridge is 430 meters. We assume that there are two towers
em will be described which will provide a context which divide the bridge into three spans. The end spans
scussion throughout the paper. Then we will formu- from the river banks to the towers are 110 metres and
ollaborative optimization for the example problem. the center span between the towers is 210 metres. The
the example problem will be solved with undecom- width of the bridge is 16 metres to accommodate four
optimization and analysis to provide a baseline for lanes of traffic. Even though wind and seismic loading are
valuation of results. Then results from collaborative critical
Fig. 1for the design ofconcepts
Superstructure long-span bridges, we will only Fig. 2 Reinforced tower concept
ization will be presented and evaluated. Finally, consider gravity loading design
for the sake of simplicity. A lane
usions will be drawn.
Figure 3: Bridge problem subsystem 1: Superstructure design concepts and
load of 21.6 kN/m per lane will be used. Unit weights During collaborative optimization, the two disci-
associated 3 design variables. 3 Taken from Balling & Rawlings’ Collaborative optimiza-
of 24 kN/m for concrete and 77 kN/m for steel will be plinary groups must formulate the design variables and
tion
used. with disciplinary conceptual design, 2000. design constraints for each concept that they consider.
We will assume this design problem will be decom- The design variables formulated by the superstructure
example problem posed into two disciplinary groups. One group will design group are listed in Table 1, and the design variables for-
the superstructure consisting of the towers and support- mulated by the deck group are listed in Table 2. Note
xample consists of the design of a long-span cable- ing cables, and the other group will design the bridge that the number and type of design variables vary from
rted bridge over a river. The total length of the
e is 430 meters. We assume that there are two towers
deck.Subsystem
Another group 1 could
also havehasbeen assigned
local the design constraints,
inequality
of the foundation abutments and piers, but for the sake
concept to concept. pq
denoted gsuper . , to be satisfied,
The design constraints considered by the superstruc-
divide the bridge into three spans. The end spans relating to requirements for axial/flexural
of simplicity, we will assume that these have already been and ture group include cableatstrength,
shear strengths the base of the
combined tower,
axial/flex-
the river banks to the towers are 110 metres and and also cable strength, under two load cases.
designed. ural strength with moment magnification at the base of
enter span between the towers is 210 metres. The Suppose that during the course of collaborative op- the tower, and shear strength at the base of the tower.
of the bridge is 16 metres to accommodate four timization, the superstructure group comes up with the The tower constraints were considered under two loading
of traffic. Even though wind and seismic loading are four concepts shown in Fig. 1. Let us assume that the cases: 1) lane load across entire bridge, and 2) lane load
4.3 Subsystem 2: Deck
al for the design of long-span bridges, we will only Fig. 2 Reinforced tower concept
der gravity loading for the sake of simplicity. A lane
of 21.6 kN/m per lane 4.3.1 Shared
will be used. design variables
Unit weights x0,2
During collaborative optimization, the two disci-
kN/m3 for concrete and 77 kN/m3 for steel will be plinary groups must formulate the design variables and

e will assume this design problem will be decom-

Thedesign constraints
problem hasfor no
eachshared
concept that
design
The design variables formulated by the superstructure
they consider.
variables, so x0,2 H .
into two disciplinary groups. One group will design group are listed in Table 1, and the design variables for-
uperstructure consisting of the towers and support- mulated by the deck group are listed in Table 2. Note
4.3.2 Local design
ables, and the other group will design the bridge variables
that the number x2 and type of design variables vary from
Another group could have been assigned the design concept to concept.
foundation abutments and piers, but for the sake The The deck subsystem
design has local
constraints considered by thedesign variables relating to the choice of deck con-
superstruc-
plicity, we will assume that these have already been cept ture igroup include cable strength, combined axial/flex-
deck – steel plate girder or reinforced concrete box girder – and corresponding
ned. ural strength with moment magnification at the base of
ppose that during the course of collaborative op- geometrythe tower, and variables xdeckat. the
shear strength The deck
base of thedesign
tower. variables are shown in Figure 4.
ation, the superstructure group comes up with the The tower constraints were considered under two loading

concepts shown in Fig. 1. Let us assume that the
4.3.3 Interdisciplinary coupling variables a2 and disciplinary analysis A2 x0,2 , x2 , â1
p q
cases: 1) lane load across entire bridge, and 2) lane load
As mentioned in Section 4.2.3, the deck generates vertical force on the superstructure.
The force is equal across the superstructure, due to the prismatic geometry of the deck
and (pre-determined) equal spacing of the cables, so can be represented as a scalar
– here denoted v. The deck will also be subjected to horizontal forces that depend on
cable angles arising from the chosen cabling concept – since the choice of concept is
unknown to the deck team, these forces are provided as copies â1 – here denoted as
h1 .
The disciplinary analysis provides the vertical force v, together with the other
forces needed to evaluate the local constraints.

Part II – Decision Systems for Engineering Design 8

Lecture 6: Multi-disciplinary design optimization
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield
234

Fig. 3 Deck concepts

Figure 4: Bridge design problem subsystem 2: Deck design concepts and associ-
ated designTable
variables.
1 DesignTaken
variablesfrom Balling &group
for superstructure Rawlings’ Collaborative optimization
Table 2 Design with
variables for deck group
disciplinary conceptual design, 2000.
Concept Variable Symbol Concept Variable Symbol

fan cable diametres di top flange width btop

(i = 1 to 19) steeltop flange thickness ttop

4.3.4 Discrepancy
harp cable diametre
function D2 and local (same
constraints
for all cables)g2 . pq d plateweb thickness
girder
web depth
tweb
d
semi-harp cable diametres (i = 1 to 19) di bottom flange width bbot
The deck subsystem
suspension attempts
vertical cabletodiametre
minimise the ddeviation
v of its interdisciplinary cou-
bottom flange thickness tbot
parabolic cable diametre dp
pling variable v from a target. Like the superstructure subsystem, stiffener
it also spacing
attempts
crossbeam standard size
sstiff
Scross
to minimise tower
deviation from
(same for
tower height
a target build cost,dHwhich
cross-sectional depth
tower places a requirement
slab thicknesson the tslab
tower
disciplinary analysis
all four to cross-sectional
compute the widthbuild cost w oftower
the deck.reinforced
The overall discrepancy
slab thickness tslab
concepts) d
function is denoted area of main reinforcing steel Atower concrete soffit thickness tsoffit
deck .
spacing of shear stirrups
In the deck subsystem, local constraints gdeck . relate
stower
pq box interior web thickness
to a variety
girder of flexural
exterior web thickness
tint
text
strengths, shear strengths and deflections. section depth d
on centre span only. The design constraints considered deck longitundial reinforcement Aslablong
by the deck group for the steel plate girder concept in- deck transverse reinforcement Aslabtran
clude combined axial/flexural strength in the plate girder soffit longitudinal reinforcement Asoffit
4.4 System problem at locations of maximum positive and negative moment,
shear strength in the plate girder, deflection in the plate
The systemgirder,
problem forinthe
the bridge manages the targets for the interdisciplinary
bars, and $31 150 cou-
/m3 for steel in the plate
h1 andinv 1the
buckling plate girder web and flanges, flex- reinforcing
pling variables:
ural strength . There areflexural
slab, and no shared
strengthdesign
in the variables to manage,
girder, crossbeam, and the
and cables.
crossbeam.
system problem doesThenot design constraints
concern considered
itself with bythe
the subsystem
deck design variables xsuper
group for the reinforced concrete box girder include com-
and xdeck , these being left to the individual teams.
bined axial/flexural strength with moment magnification 3
Unlike inat alocations
conventional
of maximum CO architecture,
positive and negativethe system Collaborative
moment, node in thisoptimization
example does
not directly shear strength
evaluate in the
the webs, transverse
objective to beflexural strength–inin this case, total build cost for
minimised
3.1
the bridge. the slab, and deflection.
Rather, it decomposes the total cost into targets
The overall system objective was the1 minimization
for the superstructure
and deck subsystems to meet,
of cost. We considered given
only by fsuper andcosts.
volume-proportional
1 respectively.
fdeck
Disciplinary analysis
This approach
is close in spirit to the
These costs MDOmaterial
included architecture
costs and known
some labor ascosts.
analytical target cascading
The analyses (ATC) group and the deck
for the superstructure
3 3
The unit costs were $78.5 /m
that is used widely in the automotive industry. for concrete, $4200 /m for group are shown in Fig. 4. The scalar isuper is an integer

Part II – Decision Systems for Engineering Design 9

Lecture 6: Multi-disciplinary design optimization
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

The system problem aims to minimise the sum of the subsystem build cost targets,
1
i.e. fsuper 1 . To produce a consistent bridge design, the discrepancy terms must
fdeck
be reduced to zero in each of the subsystems. The system problem therefore includes
equality constraints dsuper  0 and ddeck  0.

4.5 Special considerations

The decision to optimize build cost through the use of cost targets in the subsystems
has knock-on consequences for the CO architecture. Since it is no longer possible
for the system node to know whether a positive discrepancy term reflects failure to
meet the consistency target or failure to meet the cost target, the architecture may
not converge on a consistent overall design. To mitigate this issue, each subsystem
is allowed to deviate from both its own interdisciplinary coupling variable copy and
the copy of the coupling variable for the other subsystem. Deviation from the copy is
then also included in the discrepancy term. Another option – used in ATC – would
be to incorporate penalty weights for the different terms of the discrepancy equation,
where the weights are gradually updated to emphasise the consistency term as the
optimization process progresses.
To reduce the number of coupling variables in the problem, the vector of horizontal
forces h can be replaced with a parameterised function based on distance of the
cable attachment from the tower, which requires only two variables rather than the
10-element force vector. Use of such parameterised functions is a common way of
reducing the complexity of the interactions in an MDO problem.

5 Overview of MDO architectures

There are a large number of MDO architectures available. They can be classified
into two fundamental types, which vary in the approach used for dealing with shared
design variables and interdisciplinary coupling variables: individual discipline feasible
(IDF) MDO and multidisciplinary feasible (MDF) MDO. The specific choice of MDO
architecture will ultimately depend on how the design teams have been organised
and the amount of computing resource that is available to each team.

5.1 Individual discipline feasible architectures

Individual discipline feasible architectures use copies of the coupling variables and
propagate these copies to the individual teams, who are working on their designs in
parallel. Over time, the architectures work towards delivering an overall design that is
fully consistent (or feasible, from the perspective of the physics that is being modelled
and any shared aspects of the design) – i.e. the copies are updated over the course
of the optimization process until all teams agree on their values. A disadvantage of
the IDF approach is that, if the optimization process is terminated early, then there
is no guarantee that the current candidate design will be consistent, or consistently

Part II – Decision Systems for Engineering Design 10

Lecture 6: Multi-disciplinary design optimization
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

modelled, across teams – with potentially serious consequences for the validity of the
objective and constraints assessments.
Collaborative optimization is an example of an IDF architecture. Other examples
include analytical target cascading (mentioned in Section 4.4) and quasi-separable
decomposition (QSD).

5.2 Multidisciplinary feasible architectures

Multidisciplinary feasible architectures impose (or attempt to impose) consistent cou-
pling variables at each stage of the optimization process. An advantage of this ap-
proach is that consistent designs are generated by the architecture, even if the opti-
mization process is terminated earlier than anticipated.
Classical MDF approaches organise the disciplinary analyses in a strict sequence,
where – given a fixed design – coupling variables are fed from the output of one team
to the input of the next. This sequence is iterated until the teams agree on a consistent
set of coupling variables. Since the work of real-world engineering design teams tends
not to be organised according to such a sequential schedule, more modern MDF
approaches allow for more concurrent working. This is achieved by using surrogate
models within each design team that aim to predict the interdisciplinary couplings
arising from other teams’ analyses in response to the choices/analyses of the design
team. In this way, the design team does not have to wait for the ‘true’ interdisciplinary
coupling variable inputs before getting on with their work. Clearly, the success of these
methods will depend on the accuracy of the predictions made by the surrogate model.
Examples of MDF architectures include concurrent subspace optimization (CSSO),
bilevel integrated system synthesis (BLISS), and asymmetric subspace optimization
(ASO).

6 Further reading
• Martins J.R.R.A., Lambe A.B. Multidisciplinary design optimization: A survey of
architectures. AIAA Journal 2013;51:2049–2075.
• Balling R., Rawlings M.R. Collaborative optimization with disciplinary conceptual
design. Structural and Multidisciplinary Optimization 2000;20:232–241.

Part II – Decision Systems for Engineering Design 11

Lecture 6: Multi-disciplinary design optimization