Download as pdf or txt
Download as pdf or txt
You are on page 1of 36

Res Eng Design (2012) 23:17–52

DOI 10.1007/s00163-011-0112-y

ORIGINAL PAPER

Supporting product architecture design using computational


design synthesis with network structure constraints
David F. Wyatt • David C. Wynn • Jerome P. Jarrett •

P. John Clarkson

Received: 21 July 2010 / Accepted: 26 March 2011 / Published online: 21 April 2011
 Springer-Verlag London Limited 2011

Abstract A product’s architecture can affect many 1 Introduction


aspects of product and process quality, from technical
performance to the design effort required, production costs Product architectures are the abstract conceptual structures
and satisfaction of later lifecycle requirements. This paper underlying the functioning of engineering artefacts, and
explores how computational tools can augment creative their design is an important but difficult task (Ulrich 1995).
methods in product architecture design. Based on an They have a wide-ranging influence on product and
empirical study aiming to understand the context of prod- enterprise success (Crawley et al. 2004; Ulrich 1995;
uct architecture design, a new computational method is Yassine and Wissmann 2007). In particular, the early
proposed to support this activity. In the method, product stages of design commit a large fraction of the overall cost
architectures—networks of components linked by connec- of bringing a product to market (Berliner and Brimson
tions—can be synthesised using constraints on the structure 1988; Port and Schiller 1990). However, it can be hard to
of the network to define the set of ‘realisable’ architectures identify ‘good’ product architectures since they are quali-
for a product. An example illustrates how the method tative and abstracted from reality.
might be used on a real design problem, including the This paper proposes an approach based around compu-
construction of an appropriate set of network structure tational design synthesis, which would allow designers to
constraints and the identification of promising architectures consider possible architectures for a product systemati-
from the synthesis results. Preliminary evaluation of the cally, potentially leading to better results. The approach is
method’s usability, assessed through a laboratory experi- combined with insights from a case study to develop a
ment, and its utility, assessed through application to a real specific method, in which the ‘space of possible architec-
historical design problem, supported by initial validation tures’ is modelled in a declarative formal language. The
by an engineer from the case study company, suggests that model comprises an ontology of the components and
the method has value for engineering design practice. relations from which architectures may be constructed, and
network structure constraints to identify the arrangements
Keywords Product architecture  Conceptual design  of these elements that are ‘realisable’. Product architec-
Design support method  Computational design synthesis  tures that comply with the constraints can then be syn-
Network structure constraints  Cambridge advanced thesised computationally and reviewed by the designer.
modeller (CAM) Preliminary evaluation shows that the method has value for
engineering practice.
The next section provides background on definitions of
‘product architecture’ and informal and formal methods for
D. F. Wyatt (&)  D. C. Wynn  J. P. Jarrett  P. J. Clarkson product architecture design. Section 3 then identifies
Cambridge Engineering Design Centre, requirements for a support method from a case study into
Department of Engineering, University of Cambridge,
the context of product architecture design. Section 4
Trumpington Street, Cambridge CB2 1PZ, UK
e-mail: dw274@cam.ac.uk describes the support method and tool developed to meet
URL: http://www-edc.eng.cam.ac.uk these requirements, and Sect. 5 illustrates the method

123
18 Res Eng Design (2012) 23:17–52

through application to an example problem: the redesign of respects (Ziv-Av and Reich 2005); this is discussed further
a vacuum cleaner. Section 6 describes preliminary evalu- in Sect. 2.2.2.)
ation of the method’s usability, through an experiment
where novice designers applied the method to a simple 2.2 Designing product architectures
design problem, and applicability and usefulness, through
applying the method to a real historical design problem; it 2.2.1 Informal and formal approaches
then describes an initial validation of the method through
its use by an engineer from the case study company. Sec- Methods to support product architecture design may be
tion 7 reflects on the advantages and limitations of the divided into informal methods, which rely on human cre-
method and approach, and Sect. 8 concludes. ativity, and formal methods, involving codifiable rules and
procedures. Frequently used informal methods include
brainstorming (Osborn 1957), the 635 method (Rohrbach
1969) and synectics (Gordon 1961). However, the discon-
2 Background
tinuous and qualitative nature of the ‘design space’ of
product architectures may make it difficult to construct new
2.1 Defining ‘product architecture’
product architectures, or even to be aware that there are
alternatives to a particular architecture. This is in contrast
Ulrich (1995) states that ‘product architecture’ represents a
to certain other classes of design problems, for instance
product’s functions (in the sense of Pahl et al. (2007)), their
numerical optimisation problems where the ‘design space’
mapping to physical components, and the interfaces
is clearly defined with one quantitative dimension for each
between the components (particularly the degree of cou-
design variable. ‘Fixation’ effects (Purcell and Gero 1996)
pling at interfaces and issues of cross-compatibility
may further hinder designers from thinking beyond known
between uncoupled interfaces). Other authors have inclu-
architectures. These two factors may result in reliance by
ded additional elements: the partitioning of subfunctions
the designer on a restricted set of architectures, often those
into ‘modules’ (Stone et al. 2000); the requirements on a
used in previous designs. Since different architectures may
product (Yassine and Wissmann 2007); and ‘principle
have different levels of ‘overall quality’ (in whatever way
solutions’ or ‘organs’, intermediate between functions and
this is judged), a restricted set of architectures may not
components (Andreasen 1992). In the domain of software
include the one(s) which are the ‘best’ in the specific
and other (non-physical) systems, Maier and Rechtin
design context.
(2000) quote a range of definitions for ‘architecture’,
An alternative approach is to use formal (often com-
focusing on structure in terms of components and rela-
putational) methods (Cagan et al. 2005) to synthesise
tionships, but also point out that ‘architecture is what
potential architectures. Formal methods can be made sys-
architects produce, and […] what architects do is help
tematic (Pahl et al. 2007), giving them the potential to
clients make decisions about building systems’.
identify high-quality architectures more reliably than
All these definitions share the concept of product or
informal techniques and reducing the effect of fixation
system architecture as ‘an abstract description of the enti-
(Kurtoglu and Campbell 2009). They also require a product
ties of a system and the relationships between those enti-
architecture design problem to be represented in a form
ties’ (Crawley et al. 2004). Therefore, that definition will
appropriate for such methods, giving advantages from
be used throughout the rest of this paper:
greater transparency (Schmidt and Cagan 1997):
A product architecture is a model of an engineering
• Modelling the problem requires the designer to make
artefact in terms of components linked by relations.
explicit their assumptions about the space of possible
In most engineering design domains, components and architectures, which allows them to be reasoned about
relations are physical, but increasingly products involve and reconsidered if necessary.
mechatronic elements and thus electronic and software • If aspects of ‘overall quality’ can be expressed
components must also be considered. The view of product formally, architectures resulting from formal methods
architecture as a design’s components and their interrela- may be evaluated automatically against them, reducing
tions has been termed ‘topology’ (Shea et al. 1997), effort and ensuring consistency.
‘product structure’ (Lindemann et al. 2008), or ‘configu- • The models of the design problem and the synthesised
ration’ (Murdoch and Wallace 1992; Campbell et al. 2000). solutions can contribute to documentation (particularly
(‘Configuration’ may also be used to mean selection from the rationale for choosing a particular architecture).
predesigned components, a less open-ended problem than This may be important for future design projects and/or
designing the initial ‘concept’ although similar in some certification.

123
Res Eng Design (2012) 23:17–52 19

Fig. 1 A model of the process


by which product architecture is
designed, showing the
conventional approach based on
informal techniques (solid lines)
and the modifications to this
process in order to exploit
formal computational methods
to support design (dashed lines).
The shaded portion is the focus
of the method developed in this
paper

Figure 1 illustrates how these two alternative approa- architectures (6) (which may be the same as past
ches, informal and formal, may be realised in a design solutions).
process. The solid lines represent the informal process for • These solution architectures are evaluated for appro-
designing the (product) architectures of solutions to a priateness to establish whether they would indeed solve
particular design problem, based on Cross’ four-phase the problem, informed by the designer’s understanding
Exploration-Generation-Evaluation-Communication model of the problem. If so, one or more may then be
of design (Cross 1994): communicated in the form of solution architecture
designs (7). If not, the designer may iteratively refine
• Through exploring the problem to be solved (1), the
the solutions.
designer develops an understanding of it (2). At the
• A solution architecture design is then evaluated for
same time they may abstract from past solutions (3)
quality against successively higher-fidelity models of
(products that solved the same or a similar problem) to
the problem, for instance, through simulation or
give mental models of the architectures of those
prototype testing. If the solution architecture design is
solutions (4).
shown to be appropriate, it may be taken forward into
• The designer then generalises their understanding of
detailed design, manufacture and ultimately the pro-
the problem and past solutions to an understanding of
duction of a solution (8). If the designer’s understand-
possible architectures for solutions to the problem (5).
ing is at fault, they may need to refine their
• The understanding of possible solution architectures
understanding of the problem.
allows the designer to generate specific solution

123
20 Res Eng Design (2012) 23:17–52

As discussed above, the ‘Generation’ step in the infor- representing product architectures (Helms et al. 2009;
mal process may be problematic. Dashed lines indicate Kerzhner and Paredis 2009; Kurtoglu and Campbell 2009;
how it may be circumvented using computational design Schmidt and Cagan 1997; Stanković et al. 2009). (Gram-
synthesis, inspired by Cagan et al.’s (2005) framework: mar-based approaches have also been used with geomet-
rical shapes (Stiny 2006).) In such ‘graph grammar’
• The understanding of possible solutions (5) is forma-
approaches, the nodes in a graph represent components
lised into a representation of the design space of
and/or subfunctions of the product and the edges represent
solution architectures (5a).
connections between them. The grammar then defines
• This is used to synthesise possible solution architec-
‘transformation rules’ that modify either the structure of
tures (5b).
the graph or the properties of the elements (components
• These solution architectures must be interpreted as
and/or connection) within it; multiple modifications may be
mental models (6) in order that they can be evaluated
made in a single rule application (e.g. adding a component
for appropriateness against the designer’s understand-
and linking it to other parts of the architecture). Graph
ing of the problem.
grammars can, if appropriately constructed, act on existing
• If the solution architectures are not ‘appropriate’ in
architectures as well as designing ‘from scratch’. This is
terms of the designer’s understanding of the problem,
valuable because in practice engineering design is fre-
the designer may iteratively refine the formalisation.
quently incremental, with many products evolving and/or
having parts of their design reused in later products (Eckert
2.2.2 Formal architecture synthesis et al. 2004). However, formulating the rules for a grammar
can be challenging (Kerzhner and Paredis 2009), since
Methods for formal and/or computational design synthesis typically each rule only determines a small fraction of the
in engineering are many and varied (Antonsson and Cagan resulting structure and thus multiple rules must be applied
2001; Chakrabarti 2002), with applications ranging from in succession to produce valid results. Thus, the designer
load-bearing structures (Xie and Steven 1997) to robotic must construct the grammar such that appropriate sequen-
mechanisms and their controllers (Lipson and Pollack ces of rules can be completed and unintended interaction
2000). Of the methods that specifically address product between rules is prohibited.
architecture design, perhaps the most fundamental is An alternative approach is exemplified by the A-Design
morphological analysis (Zwicky 1969): a design problem is system (Campbell et al. 2000), in which autonomous
decomposed into partial problems, solutions are identified software ‘agents’ modify an emerging architecture
to each partial problem, and combinations of those partial according to their preferences and the desired objectives.
solutions are systematically considered, potentially gov- The agents are of several types, with distinct populations
erned by compatibility constraints (Allwood 2007; Golkar working to construct a function structure out of functional
et al. 2009; Hellenbrand and Lindemann 2008). However, component descriptions and to ‘instantiate’ a design using
morphological analysis requires the lists of partial solutions components from a catalogue. The Schemebuilder project
as input and assumes the relationships between the partial (Bradley et al. 1992) and the FBS Modeller (Umeda et al.
solutions do not vary, unless such variation is explicitly 1996) used a similar model of design as incremental con-
represented by a ‘configuration’ partial problem (Hellen- struction of linked functional and physical structures, but
brand and Lindemann 2008). The Subjective–Objective were intended to be used interactively by a designer rather
System (Ziv-Av and Reich 2005) is based on morpholog- than autonomously. In the former case, this was based on a
ical analysis combined with a definition of ‘overall quality’ function-means tree; the latter was later linked with a
as a weighted sum of different characteristics of the solu- separate product family design system in the IMS/GNOSIS
tion related to the fulfilment of the customer requirements project (Ranta et al. 1996).
(‘objective’) or enterprise-specific context (‘subjective’). Knowledge-Based Engineering (KBE) represents a more
Both the characteristics and the compatibility constraints pragmatic form of design support. Rather than supplying a
are defined numerically, allowing the ‘best’ solution to be formalised problem definition to generic solution-finding
found through optimisation rather than exhaustive enu- mechanisms, KBE involves capturing the steps an expert
meration as long as such detailed information about designer would take to solve the problem in computer-
‘quality’ is available. executable form. KBE tools such as ICAD (now incorpo-
Other methods can handle situations which cannot be rated into CATIA KnowledgeWare), Design ??,
expressed as combinations of partial solutions, or where the Unigraphics NX with KnowledgeFusion and GenWorks
possible ‘configurations’ are not known a priori. One GDL typically provide an environment with geometry
popular approach uses the concept of generative grammars and a notion of product structure and are thus applicable
from linguistics (Chomsky 2002) applied to graphs to architecture design. More general tools include

123
Res Eng Design (2012) 23:17–52 21

optimisation environments (such as Isight or modeFRON- easier and more accurate formalisation of a product
TIER) and constraint solvers (e.g. ILOG Solver). KBE tools architecture design problem.
have reduced the design time in domains such as aerospace
(Corallo et al. 2009); however, due to their complexity, the
effort required to configure a tool for a specific application 3 Understanding product architecture design
is significant (Bermell-Garcı́a and Fan 2002), and thus their in practice
use is restricted to largely repeatable design problems. In
addition, the focus of the approach on capturing existing Despite the potential advantages of computational design
knowledge may mean that novel architectures are beyond synthesis-based methods described in the previous section,
the reach of such tools (since they effectively formalise the their take-up in industry has been limited (Wood and Greer
fixation effects described in Sect. 2.2.1). 2001). Acting on the hypothesis that this is due to a mis-
Similar in philosophy to KBE tools are product config- match between the needs of engineering design practice
uration systems, or ‘configurators’ (in the second sense of and the systems developed to date, requirements for a
‘configuration’ discussed in Sect. 2.1). These are used method to support product architecture design were
specifically to guide the selection of combinations of pre- established from a case study at an engineering firm.
designed components for a particular customer’s needs.
One of the earliest, R1/XCON, was used at Digital 3.1 Procedure
Equipment Corporation to configure VAX computers from
1978 (McDermott 1980); more advanced systems such as The basis for selecting the case study was that the design
COCOS (Stumptner et al. 1998) can handle systems context should be representative of other engineering
involving thousands of components. However, such sys- products and that the architecture of the product should
tems assume total knowledge of the product and are thus permit variation (whether or not such variation occurs).
not applicable during the early part of design considered in The target chosen was a UK-based manufacturer of med-
this paper. ium-scale off-highway diesel engines, which in the last
Many of these methods employ constraints, in the widest decade became vertically integrated into a multinational
sense of explicit tests which a design must satisfy. In some engineering group after decades of independent operation.
cases, such as in grammar-based approaches, the con- Diesel engines are comparable to many types of mechan-
straints play a minor role in synthesising solutions; how- ical products in terms of complexity, safety-criticality,
ever, in other methods, the constraints are more significant strong legislative pressure, requirements for high reliability
in determining the resulting solutions. Numerical con- and potentially contradictory market demands (in this case,
straints, restrictions on the possible values of a design increased performance within a fixed physical footprint).
parameter, have been used in parametric design (Medjdoub While the core technology is mature and restricted by many
and Yannou 2000; Mullineux et al. 2005). Alternatively, of these factors, and thus the scope for architecture change
constraints may represent high-level qualitative informa- in the ‘core engine’ was known to be limited, innovation
tion. In this case, they may be part of the problem defini- was thought to be possible in the design of the peripheral
tion, as in the configurators described above; or they may systems (cooling, fuel, air, exhaust handling); this was
be implicit in the method, such as for flow compatibility in borne out during the study. Thus, diesel engine design is an
the FuncSION system for synthesis of mechanisms from a appropriate context from which to infer requirements for a
set of fundamental ‘building blocks’ (Chakrabarti and support method.
Bligh 1996) and in the system described by Ulrich and The questions investigated during the case study inclu-
Seering (1989), for synthesis of input–output dynamic ded not only requirements for product architecture design
systems from specified components. Although not directly support methods but also design processes, organisation
applicable to architecture design, numerical constraints and effects of product architecture on later design; a full
may represent qualitative constraints—for instance, a discussion of the findings may be found in Wyatt et al.
numerical constraint specifying a distance of zero between (2009a).
two walls in a building floorplan effectively defines a Interaction with the company involved 13 semi-struc-
qualitative constraint of connectivity between the corre- tured interviews with eight engineers at the company’s
sponding rooms. In terms of product architecture design, headquarters over a period from May to December, 2008.
qualitative constraints have similar expressive power to Interviews lasted between 45 min and 2 h, with each
graph-grammar-based approaches. However, while graph engineer interviewed 1–3 times. Initial selection of inter-
grammars can express complex transformations of an viewees used contacts from previous research projects
architecture succinctly, constraints represent the ‘bound- carried out at the company by the authors’ research group,
aries’ of the ‘design space’ more directly, which may allow after which further interviews were based on suggestions

123
22 Res Eng Design (2012) 23:17–52

from interviewees (‘snowball’ sampling). All interviews of the relevant subproblem and to consider the effects
were recorded, then transcribed and reviewed. Other of the excluded portions when appropriate.
sources of information included informal comments from
To gain further insights, a specific design problem was
other employees of the company and sample items of
investigated during the case study, involving the design of
design documentation.
the mounting arrangement for the engine components
(particularly, the core engine and ‘cooler group’, a com-
3.2 Findings and inferred requirements for the support
bination radiator and charge air cooler) in electrical gen-
method
eration and pumping applications. Numerical analysis had
suggested that the design used in the previous product
In the case study, design was observed to be primarily
generation could no longer be used due to changes in
incremental: development of a new product line was based
component dimensions, and therefore a redesign was nee-
on the previous generation, and in particular few changes
ded. Five potential new architecture concepts were gener-
were made at the architectural level. To a large extent, this
ated by ‘brainstorming’, although most were drawn from
architecture carry-over was deliberate and planned: the
previous products made by the company or its competitors
large ‘investment’ both in capital equipment and in
rather than from first principles (evidence of the difficulty
organisational processes associated with an architecture
of solution generation as described above). The concepts
meant that the cost of changes to it had to be spread over
involved different arrangements of a small set of possible
multiple product generations, and thus modifications to the
‘components’, many of which (e.g. the core engine) had
architecture only occurred where new requirements could
complicated internal structures which were abstracted
not be met using the existing architecture. Recently,
away while brainstorming (evidence of the need to support
requirement changes have come from emissions legislation
problem decomposition and integration). When asked
(Jarratt et al. 2003), and corresponding architectural
about the concepts, the engineers were readily able to
changes have included adding exhaust after-treatment
explain why particular patterns of elements were workable
systems to the engine. However, in addition to deliberate
while others were not; for instance, the cooler group would
architecture carry-over, there was evidence that architec-
be damaged by vibrations from the engine block and so
ture was carried over involuntarily because the engineers
could not be rigidly mounted to it. These explanations
involved in design were unable to identify alternatives,
often could not be simply formulated using the concept of
reinforcing the difficulty of ‘Generation’ as stated in Sect.
‘function’ as an input–output transformation, used in many
2.2.1. Therefore, a formal method for systematically gen-
of the methods described in Sect. 2.2.2, in some cases due
erating solutions to a design problem could be beneficial in
to the internal structures of the elements. Two further
this context. The need to incorporate planned architecture
requirements arise from these observations:
carry-over represents a formalisation-related requirement
on a support method: • F3–Problem-specific elements: The method should
allow the designer to define the architecture’s compo-
• F1–Incremental design: The method should allow the
nents as fits the problem, rather than at the level of
designer to specify an architecture as a starting point for
elementary machine elements or input–output
synthesis of new solutions.
‘functions’.
Through the course of the interviews, it became clear • F4–Declarative definition: The method should repre-
that there was no single point where the engine’s archi- sent the design space of architectures implicitly (with-
tecture, in the sense of the definition in Sect. 2.1, was out enumeration) and declaratively (through constraints
designed. Instead, those parts of the architecture that did that define allowable architectures).
change were redesigned in separate but interlinked design
The five new engine-mounting architecture concepts
subproblems of limited complexity, such as the core engine
were evaluated against the original using a Pugh analysis
layout or the incorporation of the fuel injector into the
(Pugh 1990) with 19 weighted criteria. In the evaluation, the
cylinder head. This is a characteristic commonly observed
‘technical confidence’ in the solution was only one of the
in engineering artefacts (Alexander 1964), which have
most important criteria—others were high-level business
been termed ‘almost decomposable systems’ (Simon
and customer drivers, e.g., market competitiveness and ease
1996), and represents a further requirement:
of assembly. Again, the engineers were usually able to
• F2–Problem decomposition: Where an overall design articulate the specific features of each architecture concept
problem may be decomposed into interlinked subprob- that caused them to assign it a particular score, based on
lems, the method should allow designers both to their previous experience since the detailed knowledge
abstract away portions of the design outside the scope needed to predict its technical performance was not yet

123
Res Eng Design (2012) 23:17–52 23

available. When the design problem was originally under- • Many methods that use the energy-transformer model
taken, the two highest-scoring architectures were taken of components, such as Campbell et al. (2000) and
forward to more detailed evaluation. Chakrabarti and Bligh (1996), are not appropriate in
This suggests two requirements for the interpretation this context (Requirement F3).
step of the method: • Grammar-based approaches are by definition proce-
dural, rather than declarative, descriptions of the design
• I1–Interpretation support: The method should present
problem (Requirement F4).
synthesised architectures such that they can be inter-
• Few methods explicitly support designers in refining
preted and further evaluated by designers.
the definition of the space of architectures and/or the
• I2–Feature-based evaluation: The method should allow
start point (Requirement P1).
designers to specify structural features of architectures
which have beneficial or detrimental effects on any of The next section presents a method to support product
the lifecycle objectives of the product, and to evaluate architecture design based on these requirements.
synthesised architectures by testing for the presence of
these features.
4 Proposed support method
Finally, in terms of the overall solution-finding proce-
dure, the iterative nature of design and the difficulty of
4.1 Formalisation: representing architectures
constructing an appropriate description of a problem (in
and spaces of architectures
informal terms much less a formal representation) means
that in addition to the three phases of Formalisation, Syn-
In the method, product architecture design problems are
thesis and Interpretation, a method should also support the
represented using a set of constructs which have been
‘Refinement of formalisation’ iteration in Fig. 1:
developed iteratively through a series of examples,
• P1–Refinement of formalisation: The method should including those described in Sects. 5 and 6. A simple
support the modification of the problem formalisation example will be used to illustrate the various constructs,
by the designers in response to their interpretation of based on the problem of transmitting torque through a gear
the synthesised solutions. train. Figure 2, from Starling and Shea (2005), shows an
example gear train containing seven gears indicated by
It is notable that many of the computational synthesis
numerals; ‘[S]’ indicates the shaft of the motor (the input to
methods described in Sect. 2.2.2 do not fulfil these
the gear train) and ‘[W]’ indicates the connection to the
requirements, in particular:
winding mechanism (the output from the gear train).
• Although certain classes of approach are in theory
applicable to incremental design problems (Require-
4.1.1 Modelling product architectures
ment F1), such as graph grammars if appropriately
formulated, the issues associated with this are rarely
A product architecture is represented as a network in which
discussed.
nodes represent components and links represent connec-
tions or other relations, in accordance with the definition in
Sect. 2.1. Both components and connections are assigned a
concrete type (e.g. ‘Input gear’) that indicates where dif-
ferent elements are members of some category and are thus
equivalent at the architectural level; the details of compo-
nents of the same type may be different, for instance, one
‘Idler gear’ may have a different number of teeth from
another. Higher-level similarities may be indicated by a
generalisation-specialisation hierarchy of types, in which
types above the ‘leaf’ level may be marked as abstract to
indicate that they may not be instantiated directly; a given
type can have multiple ‘parents’. Abstract component and
connection types may serve purely a cognitive or aesthetic
role, to indicate the organisation of the hierarchy, but they
can also be used to assign a single network structure
constraint (described in the next subsection) to multiple
Fig. 2 A simple gear train, reproduced from Starling and Shea (2005) concrete types.

123
24 Res Eng Design (2012) 23:17–52

The types of connections modelled are at the discretion modelling notation that will be used throughout this paper:
of the designer, but may include any or all of: components are represented as cubes and connections as
concentric circles, with component and connection types
• Structural relations such as ‘attached to’, as used in
indicated by the corresponding symbol within a border. In
Shea et al. (1997), and ‘contains’;
this case, there are three concrete component types (‘Input
• Behavioural relations such as flows of energy, materials
gear’, ‘Output gear’ and ‘Idler gear’), one concrete con-
and signals and their subtypes, as used in function
nection type (‘Meshes with’, which is symmetric) and one
structures (Pahl et al. 2007; Hirtz et al. 2002);
abstract component type (‘Gear’) from which all three
• Assignment relations such as ‘comprises’, similar to the
concrete component types inherit.
‘component realises organ’ relations of Andreasen
For brevity, in the remainder of this paper the term
(1992);
‘architecture’ will be used to refer to a model of a product
• Geometrical relations such as relative position.
architecture in this form.
This flexibility in defining relations makes it possible to
focus on the essence of different solutions to an architec- 4.1.2 Defining a space of architectures
ture design problem–for example, geometry may be
unimportant. A connection type must be marked as either In the method, a specific product architecture is viewed as a
symmetric (e.g. ‘attached to’) or asymmetric (e.g. ‘con- point in a ‘space of possible architectures’. The extent of a
tains’) to reflect its semantic meaning. In the examples space of architectures is defined by the complete set of
considered, the expressiveness of this representation of the possible arrangements of components and connections, but
relations between components has been found to be not every arrangement of components and connections
equivalent to other representations that equip components ‘makes sense’. Put more formally, an arrangement of
with input and output ports that may be interconnected if components and connections is realisable if it is an archi-
their flow types are compatible, such as Helms et al. (2009). tecture of a product that could fulfil the desired function. Of
While detailed modelling and analysis of an architecture course, an architecture can give rise to a very large or
may require consideration of the connections of individual infinite number of detailed designs, only some of which will
ports of elementary components (such as the three termi- be functional in the desired sense. In keeping with
nals of a transistor) (Paredis et al. 2001), the omission of requirement F4, the set of realisable product architectures is
ports in this method simplifies the modelling language defined implicitly and declaratively by using qualitative
while maintaining flexibility. constraints on the arrangement of the components and the
Figure 3 shows how the architecture of the simple gear connections between them. This is analogous to defining the
train in Fig. 2 may be represented in this form, as a net- feasible region in a parametric design problem through
work (in this case a chain) of components linked by con- numerical constraints. Four types of network structure
nections. The figure uses elements from a graphical constraints (NSCs) are defined in the method:

Fig. 3 The architecture of the gear train in Fig. 2, modelled using the graphical notation introduced in this paper

123
Res Eng Design (2012) 23:17–52 25

• Component number constraints (CNCs) indicate how concept of ‘slot modularity’ (Ulrich 1995) and may also
many components of a given type must be present, for be used to represent alternative embodiments (child
instance ‘There must be one input gear’. concrete component types) for a function (the abstract
• Direct connection constraints (DCCs) indicate which component type). Connection types may also be
component types may be connected together by which similarly grouped, particularly when used in ICCs.
connection types. In addition, they specify the required
Although fairly simple, the set of NSC types has been
cardinality of the connection—how many components
found sufficient to define almost all aspects of realisability.
of the second type must be connected to a component of
However, in some cases this language may not suffice;
the first type, and vice versa. For instance, ‘A gear must
possibilities for extending it are discussed in Sect. 7.1.
mesh with at least one other gear’.
The set of component types, connection types and NSCs
• Fan out constraints (FOCs) indicate how many
for a particular problem, which describes a range of pos-
connections of a certain type that components of a
sible architectures, is termed a schema. This terminology
certain type must have in total, for instance ‘An idler
was chosen by analogy with Extensible Markup Language
gear must mesh with at least two other components’ (to
(XML) schemas (XML Working Group 2004); equivalent
avoid ‘dangling’ idler gears at the end of the gear train).
pairs of terms from other areas of research include
If the connection type is asymmetric, an FOC may
‘instantiated architecture’ and ‘generic architecture’ in
apply to incoming connections only, outgoing connec-
mass customisation (Jiao and Tseng 1999), ‘closed product
tions only or both.
structure’ and ‘open product structure’ in engineer-to-order
• Indirect connection constraints (ICCs) indicate how
systems (Mesihovic and Malmqvist 2000), ‘product model
many continuous paths there must be from every
instance’ and ‘product model’ in knowledge-based engi-
component of one type to every component of another
neering (Brimble and Sellini 2000), and ‘model’ and
type, made up of connections of specified type(s), for
‘metamodel’ in object modelling (Object Management
instance ‘There must be exactly one path of ‘meshes
Group 2006).
with’ connections from the input gear to the output
In many cases, there are multiple ways of abstracting an
gear’ (to transmit torque from the input to the output
architecture design problem into a schema, varying in level
without loops in the gear train). Where asymmetric
of detail and scope:
connections are present, an ICC may be configured to
either obey or ignore the direction of the connections • Level of detail: The component and connection types in
when checking for paths. Optionally, the number of the schema may represent architecture elements at any
components of particular types on the path may also be level of granularity, from high-level assemblies to
specified. individual geometrical features; the level of detail need
not be consistent across all elements.
Numerical aspects of NSCs are specified as ranges with
• Scope: The schema may allow any architecture that
a minimum and a maximum for the quantity concerned. All
could potentially solve the problem, in an unrestricted
NSCs are assumed to apply simultaneously (Boolean
design context, or may be more focused around a
‘AND’) to all components of the corresponding types
particular set of possibilities if, for example, manufac-
(universal quantification). As mentioned in the previous
turing limitations or intellectual property issues are
subsection, NSCs may refer to either concrete or abstract
significant.
component/connection types. This capability may be used
to express constructions which would otherwise require Other forms of variation between schemas are possible:
Boolean combinations of multiple NSCs (the example for instance, the NSCs may be set up such that the number
DCC quoted previously is an example of both of these of components of a particular type is limited by the cor-
constructions): responding CNC or by a cardinality constraint in a DCC.
However, the processes of formalisation and interpretation
• One NSC may apply to all component types within a
are complementary, and thus different schemas may still
higher-level grouping (‘AND’), for instance ‘Every car
lead to realisable architectures when interpreted appropri-
[an abstract component type with child component
ately. Indeed, architectures synthesised from different
types hatchback, saloon, estate and sports car] must
schemas can lead to the same embodiment. This issue is
comprise four wheels’.
discussed further in the context of the usability evaluation
• Alternatively, an abstract component type may express
in Sect. 6.1.
the possibility for alternatives (‘OR’), for instance
To continue the illustrative example, the space of gear
‘Every sports car must comprise one roof [an abstract
trains for the winding mechanism shown in Fig. 2 may be
component type with child component types solid roof
represented by the schema shown in Fig. 4, which indicates
and folding roof]’. This construction is similar to the

123
26 Res Eng Design (2012) 23:17–52

the design problem. In particular, apart from syntactic


errors in the use of the method constructs, there are two
possible problems with a schema: overconstraint, where
realisable architectures are rejected; and underconstraint,
where unrealisable architectures are allowed. A single
schema may be both underconstraining and overcon-
straining in different respects. However, although a schema
is a declarative representation, because it is also implicit it
may be difficult to check its accuracy manually. This stems
from both the implicit nature of the ‘real’ set of possible
architectures and from cognitive limitations in ‘thinking
through’ schemas with more than a few elements.
On the other hand, it is straightforward to test for
consistency between a schema and an architecture by
checking whether the architecture is in the design space
described by the schema (i.e. satisfies the schema’s
NSCs). If the schema accurately represents the space of
possible designs, and if the architecture accurately rep-
Fig. 4 A schema for the space of simple gear trains resents a specific design, the schema and architecture will
be consistent. Conversely, if the architecture and schema
are inconsistent, one or other must be unrepresentative of
how NSCs are represented in the graphical notation. Since reality and must be refined by the designer. Typically,
each component type will have its own component number such refinement will proceed iteratively, guided by
constraint, CNCs are shown by numbers within the com- information about the specific constraints violated, until
ponent type symbols; other types of NSCs (DCCs, FOCs the architecture and schema are mutually consistent—at
and ICCs) are indicated by separate elements whose icon which point the schema will be representative of at least
indicates the type of constraint, linked by arrows to the that segment of the design space that contains the
relevant component and connection types. An unlimited example architecture. While this process does not assess
upper bound in a numerical range is indicated by ‘*’. In whether the schema is accurate outside the example
this case, the schema uses appropriate CNCs, the example concerned, applying the process several times using dif-
DCC, FOC and ICC quoted previously, and an additional ferent architectures can then extend the segment of the
ICC specifying that there must be a single path of con- design space known to be represented by the schema and
nections from the input gear to each idler gear (again to reduce the likelihood of overconstraint. Since it is more
prevent loops in the gear train). Some of these constraints direct to construct a representative model of a real
are not necessary for the case of a single output gear, but product’s architecture than to construct a representative
become important if multiple output gears are allowed. schema, this consistency checking can reduce the diffi-
Finally, the hierarchies of concrete component and con- culty of constructing a representative schema. Under-
nection types have been completed by top-level abstract constraint can be detected more straightforwardly by
component and connection types. inspecting the ‘realisability’ of synthesised architectures,
In cases where continuous arrows between elements would and again can be corrected by iteratively refining the
be inconvenient, the arrows may be split by ‘hyperlinks’. schema.
These are represented by small white circles containing The discussion has focused on ensuring that a schema
numbers; pairs of circles containing the same number repre- accurately represents the real design problem, but the
sent a logical connection. approach of consistency checking can also be applied
The graphical notation is summarised in Fig. 5. in situations in which it is important to obtain an archi-
tecture that is an accurate model of a real product—for
4.1.3 Ensuring consistency between schema instance, if an architecture is to be used as the starting point
and architectures for synthesis as discussed in the next section.

In order to be used for computational synthesis, it is 4.2 Synthesis algorithm


important that a schema accurately represents the space of
solution architectures for a product, taking into account any Based on a schema, sets of solution architectures may
abstractions or modelling conventions made in setting up be generated using computational design synthesis. The

123
Res Eng Design (2012) 23:17–52 27

2. Removing an existing connection.


3. Adding a component of one of the types defined in the
schema;
4. Adding a connection of one of the types defined in the
schema, between two existing components;
While these operations could be represented by rules in
a graph grammar such as those discussed in Sect. 2.2, in
graph grammar methods any architecture that contains only
terminal symbols and has been generated from the starting
point by a sequence of rules is assumed to be realisable. In
contrast, in this method, the evaluation of realisability is
separate from the synthesis process and achieved by testing
whether the current architecture satisfies the NSCs at each
point in the search.
The architecture from which the synthesis starts, the
start point, is supplied by the designer according to the
type of design problem:
• If the problem involves adapting an existing design,
i.e., is a case of planned architecture carry-over, the
start point may be the architecture of the existing
design. This start point may be modified before
synthesis by manually removing components or con-
nections; alternatively, the synthesis algorithm may be
configured to remove different combinations of ele-
ments automatically.
• If the problem involves the relation between a design
and an external context, the start point may be a model
of this context. The synthesis algorithm can then be
restricted to using only those component and connec-
tion types that represent the architecture to be designed,
rather than those which represent the context.
• If there are no preconceived ideas about the form of the
solution (i.e. there is no architecture carry-over except
that embodied in the schema), to reduce computational
expense the synthesis algorithm automatically gener-
ates a start point: an architecture containing the
minimum number of components of each type to
satisfy the corresponding CNC. For instance, in the
case of the schema in Fig. 4, the minimal architecture
will contain one input gear and one output gear but no
idler gears.
Fig. 5 A summary of the graphical notation for product architectures
and schemas The type of the design problem may also determine
whether all types of elementary operations may be used or
whether a specific subset is appropriate. For instance, if the
technique used is state space search (Russell and Norvig
designer wishes to conserve the set of components in an
2003). The space of architectures forms a state space, since
existing architecture and investigate how those components
any architecture, realisable or not, can be reached from any
could be connected in different ways, synthesis may be
other architecture by appropriate sequences of four types of
restricted to adding and removing connections (operation
elementary operations:
types 2 and 4), while in design with no architecture carry-
1. Removing an existing component (in which case all over there are no existing components or connections to
connections associated with it must also be removed); remove and thus operation types 1 and 2 are unavailable.

123
28 Res Eng Design (2012) 23:17–52

State space search problems can be solved exhaustively et al. (1998)); when no unsatisfied DCCs remain, connec-
using conventional graph search algorithms (Knuth 1998). tions are only added if they do not violate any DCCs.
In this case, depth first search is used to reduce runtime Likewise, when adding components architectures are not
memory requirements, with a limit on the maximum search considered for adding connections unless all CNCs are
depth—i.e., the maximum ‘distance’ (number of elemen- satisfied. Lastly, in the case of large problems where
tary operations) between the start point and any generated exhaustive search methods such as depth first search
architecture. The algorithm is based on a first-in last-out require an excessive computation time, stochastic methods
‘stack’ of architectures to be examined and proceeds as can be used to sample from the space of architectures.
follows: Samples are generated by performing a random number
(less than or equal to the maximum search depth) of ran-
1. Initialise the stack to contain the start point.
dom elementary operations with search pruning as descri-
2. Remove one architecture from the top of the stack.
bed above, after which the resulting architecture is tested
3. Test whether this architecture is valid (i.e. satisfies all
for validity. These refinements have proved sufficient to
NSCs in the schema, and thus represents a realisable
allow application to realistic cases, such as the example
product architecture).
described in Sect. 5; however, the computational cost
a. If so, add it to the list of solutions found and return remains considerable. Future avenues for improving syn-
to step 2. thesis efficiency are outlined in Sect. 7.3.
b. If not, proceed to step 4.
4.3 Interpretation of generated architectures
4. Test whether the maximum search depth has been
reached. 4.3.1 Graphical representation
a. If so, proceed to step 5.
b. If not, generate all ‘children’ of the current A single synthesised architecture may be presented to the
architecture (those architectures that can be designer as a two-dimensional image using a force-directed
reached by a single elementary operation), and layout (Eades 1984) to arrange the elements automatically
add them to the top of the stack. in an interactive manner; masses represent components and
springs represent connections. Four examples are shown in
5. Test whether the stack is empty. Fig. 6, representing gear train architectures synthesised
a. If so, terminate and return the list of solutions from the schema in Fig. 4 modified to require two output
found. gears; the layout has been configured to display compo-
b. If not, return to step 2. nents as gear shapes, with light grey representing the input
gear, dark grey the output gears and white the idler gears (if
In step 4b, elementary operation types are applied in present).
sequence: different combinations of components are
removed from the start point; for each of those, different 4.3.2 Quantitative evaluation
combinations of connections are removed; for each of
those, different combinations of components are added; for To support the designer in choosing between a set of
each of those, different combinations of connections are architectures, each architecture may be evaluated quantita-
added. Testing for validity is only performed after adding tively using one or more metrics. In this context, a metric is
connections (which must always occur), and each ele- defined as a function f from a space of architectures A to the
mentary operation type has its own maximum search depth. set of real numbers R, as shown in Eq. 1. Without loss of
The search is most efficient if the maximum search depth is generality, such functions must return positive values with
not significantly greater than the minimum depth necessary smaller values corresponding to ‘better’ architectures, in
to synthesise architectures that satisfy the NSCs. accordance with the conventions of objective functions in
The computational expense of the search can be sig- optimisation. Each metric may assess a particular aspect of
nificant for moderately complex problems. To make the the ‘quality’ of an architecture related to any of the pro-
search more efficient, techniques are used to prune the cesses with which the product is involved during its lifecycle
search space. Each architecture examined is recorded to (Huang 1996). Multiple metrics may be used in combination
ensure each point in the state space is searched only once. to give a greater overview of the solution space.
When adding connections, the set of connections added to
f :A!R ð1Þ
produce the ‘children’ is restricted to those which would
contribute to satisfying a DCC whose lower bound is not However, an architecture as a network of components
met, until all DCCs are satisfied (inspired by Stumptner and connections contains little information about any

123
Res Eng Design (2012) 23:17–52 29

aspects of its quality. This means that many conventional product (D-complexity), the complexity of manufacture
assessment methods for product design cannot be used, (M-complexity) and the complexity of making a change to
with the exception of pure ‘complexity’ metrics (Summers the product (C-complexity).
and Shah 2010). To assess more concrete aspects of Since these metrics are not calibrated against an abso-
quality, it is necessary to incorporate additional lute scale (e.g. cost or time) and are intended to be generic
information about a specific design context via one of across product domains, they are used only as an indication
two routes: of the relative differences between architectures. If infor-
mation about a specific product domain is available,
• Into the architecture, by assigning properties to the
structural complexity metrics may be tailored for that
elements, e.g., the manufacturing operations used to
context; an example of this is given in Sect. 6.2.
produce each component or the materials of which they
are composed, which allows the use of generic metrics;
4.3.3 Constraint-based classification and visualisation
or
• Into the metric, for instance by testing an architecture
In design situations where the design objectives are well-
for the presence of specific structural patterns which are
defined and their meaning in architectural terms is well
deemed to be significant for the corresponding aspect of
understood, quantitative evaluation using metrics may be
the quality of the design (Lindemann et al. 2008).
appropriate. However, in other situations, it may be more
In this work, the architectures resulting from synthesis appropriate for a designer to explore the different concepts
do not include additional information themselves (although within the set of architectures resulting from synthesis, for
the method could potentially be extended to include such instance, by classifying them into categories based on
information). Thus, more specific metrics must be derived specific patterns in the way components are connected
by including information in the metric definition. Three together. Patterns may be ‘interesting’ either because they
such metrics, which assess different aspects of the devel- are known from prior experience to be significant (e.g.
opment process for a product, are described in Table 1. The patterns that have been used in previous products), or
metrics are based on graph-theoretical properties of an simply because they are observed to be shared between
architecture and assess the complexity of designing a architectures.

Fig. 6 A selection of synthesised gear trains with two output gears, arranged using a force-directed layout

123
30 Res Eng Design (2012) 23:17–52

Table 1 Metrics used to evaluate product architectures


D-complexity M-complexity C-complexity

P
n P
n .P
n P
n
DðAÞ ¼ 1n zi MðAÞ ¼ n þ 1n zi CðAÞ ¼ nðn1Þ
n dij
i¼1 i¼1 i¼1 j¼iþ1

Components with more interfaces require The manufacturing cost of a product may be The likelihood of change propagating between
more redesign work (Sosa et al. 2007). divided into the manufacturing cost of its two components in a product is inversely
Thus, a metric for the design complexity, components and the cost of assembling them. related to the graph distance between the two
D(A), of an architecture A containing While the costs of a component and a single components in the architecture (Clarkson
n components is: the mean of the number of assembly operation may be different, the total et al. 2004; Giffin et al. 2009; Keller et al.
connections per component zi for all part and assembly costs per item can be 2006). Therefore, a metric for the change
components i in A comparable (Huang 1996). Therefore, without complexity, C(A), of an architecture
additional information about the relative costs A containing n components is: the reciprocal
of parts and assembly operations, a metric for of the mean of the graph distances dij between
the manufacture complexity, M(A), of an all pairs of components i,j in A
architecture A containing n components is: the
sum of the number of components n and the
number of connections per component zi for
all components i in A (halved to avoid double
counting)

Such patterns can be defined using the same NSCs as in Java programming language (Sun Microsystems 2006).
the construction of the schema. These additional NSCs do The force-directed layout described in Sect. 4.3.1 was
not affect synthesis, but are used only in the exploration of implemented using Touchgraph (Shapiro 2002), and the
the synthesis results. By comparing (computationally) all Euler diagram visualisation described in Sect. 4.3.3 using
of the synthesised architectures against the auxiliary NSCs the Prefuse information visualisation library (Heer 2007).
that define a category, the set of synthesised architectures is Figure 7 shows a screenshot of the software tool in use on
classified into two subsets: those which satisfy the NSCs, the example problem described in the next section.
exhibit the pattern and are thus members of the category;
and those which do not and are not. Likewise, two patterns 4.5 Summary
divide the set of synthesised architectures into four subsets,
and three patterns divide it into eight subsets etc., although A flowchart of the use of the system is shown in Fig. 8. The
some subsets may be empty. The category-membership sequence of activities is derived from Fig. 1, involving the
information may be condensed to a numerical ‘similarity’ three key steps of formalisation, synthesis and interpreta-
value and used for hierarchical clustering (Langdon and tion. As indicated, the designer’s understanding of the
Chakrabarti 2001). Alternatively, the raw information may problem can potentially be enhanced by the synthesised
be presented a visual form that preserves the complex architectures, leading to generation of a wider range of
similarity relationships between architectures, for instance solutions. Alternatively, if the synthesised architectures
through automatically generated Venn/Euler diagrams. directly represent the possible options, they may be eval-
Examples of this are given in Sect. 5.4; more information uated through the methods described in Sect. 4.3.2.
about this technique is presented in Wyatt et al. (2009b).

4.4 Implementation 5 Illustrative example: vacuum cleaner redesign

The method was implemented as a support tool based on Household vacuum cleaners originally used physical fil-
the Cambridge Advanced Modeller (CAM) software tration to separate dust from the air stream, most com-
framework, formerly the P3 Platform (Wynn et al. 2009). monly employing a cloth or disposable paper ‘bag’ that
CAM provides a user interface for constructing network also collects dust for disposal. However, recent vacuum
models conforming to a specified modelling language, with cleaners have been equipped with cyclonic dust separators
an extensible plug-in architecture to allow analysis func- (Dyson 1986). These use centrifugal effects in a cylinder of
tionality to be added to the software. In this case, a mod- rapidly rotating air to move dust to the perimeter leaving
elling language was constructed to specify the eight types clean air in the centre—a technology originally developed
of element described in Sect. 4.1, including the graphical for industrial dust extraction systems (Ogawa 1984).
notation defined in Fig. 4. The synthesis and evaluation Cyclonic systems can be more complex to engineer than
algorithms were then implemented as plug-ins using the systems with physical filters, but have potential

123
Res Eng Design (2012) 23:17–52 31

Fig. 7 A screenshot of the


software tool in use on the
example problem described in
Sect. 5, showing the modelling
window, the synthesis control
panel and a force-directed
layout of a synthesised
architecture

advantages: they are not dependent on pores that can Moreover, this problem is an example of ‘technology
become blocked by the collected dust, meaning that the infusion’ (Suh et al. 2008), which is widespread in engi-
resistance to air flow does not reduce significantly over neering design—the addition of aftertreatment systems to
time; and there is reduced waste and expense since the dust diesel engines in the case study described in Sect. 3 was
collector can be emptied and reused. Thus, cyclonic vac- also an example of this trend. As seen there, technology
uum cleaners are attractive to consumers: in a survey of infusion can cause wide-ranging architectural change but at
features of vacuum cleaners considered by consumers the same time can allow further changes to be made with
when purchasing, ‘bagless’ (i.e. cyclonic dust separation) minimal additional cost. Technology infusion is therefore a
was ranked second after ‘suction power’ (Mintel 2008). It suitable class of problems for architecture design support
is therefore attractive for vacuum cleaner manufacturers to methods, and the procedure used to apply the method to
introduce products that use cyclonic separation. However, this design problem could potentially be used for other
as discussed in Sect. 3.2, rather than designing a cyclonic design problems in this class.
vacuum cleaner ‘from scratch’, a manufacturer may con- This problem was addressed using the method proposed
sider conserving architecture elements from previous non- in Sect. 4; the description follows the sequence of steps in
cyclonic designs to minimise risk and cost. Fig. 1. The time taken to carry out each step was recorded,
Therefore, this example considers a hypothetical case of to give an indication of the effort required.
adapting a bagged cylinder vacuum cleaner to use a
cyclonic dust separator. It is assumed that only the base 5.1 Understanding the problem and possible solutions
unit requires redesign to accommodate the cyclone—the
hose, nozzle etc. are not affected by the change. This In order to understand the different possible architectures
problem is an appropriate example for the method because of vacuum cleaners, four examples were obtained spanning
its level of complexity permits many possible architectural a range of types; they are shown in Table 2. The last of
variants, witnessed by the diversity of vacuum cleaner these, the Sanyo SC-N200, was chosen to fulfil the role of
designs in production, and thus a systematic exploration of the existing product in the example. The four example
possible architectures could be beneficial. While more vacuum cleaners were disassembled, and stages in disas-
complex design problems exist, as argued in Sect. 3.2 they sembly were photographed. After disassembly a written
are often decomposed into sub-problems, each of which description of each vacuum cleaner was prepared, record-
may be of a similar level of complexity to this example. ing the architecture of the vacuum cleaner in terms of the

123
32 Res Eng Design (2012) 23:17–52

Fig. 8 A flowchart showing the use of the method proposed in this paper to support product architecture design

components and the connections between them. The level The descriptions and photographs were used as refer-
of detail of the description was chosen by the authors, ences during formalisation of the problem. Disassembly
trained engineers with no direct experience in vacuum and documentation of each of the four examples took
cleaner design, to include those design elements which approximately 3 h.
were considered necessary to describe different solutions to
the problem. Accordingly:
5.2 Formalising the problem
• Fasteners and internal wiring were recorded as con-
nections rather than components. 5.2.1 Initial modelling
• Components such as the motor and the switch, which
would be designed and assembled in a separate process The understanding gained through the activities described
or bought preassembled from suppliers, were not in Sect. 5.1 was used to construct a schema for cylinder
reduced to individual parts. vacuum cleaner power units, both cyclonic and non-
• Some features of the product that were not physically cyclonic. The schema is shown in Fig. 9, using the
separable components but whose role in the architec- graphical notation presented in Fig. 5. In this case, the
ture would nonetheless be explicitly considered, such as level of detail was chosen to be appropriate to the
the handle, were recorded separately. problem under consideration, as described in Sect. 5.1,

123
Res Eng Design (2012) 23:17–52 33

Table 2 A comparison of the four example vacuum cleaners studied


Manufacturer Lever Industrial Goblin Tesco Sanyo

Product name Ensign 360e AquaVac Industrial 40 VC106 SC-N200


Form factor Upright Wet/dry Cylinder Cylinder
Dust separator Bag Filter Cyclone Bag
Picture

and the scope was intended to include the entire space of flow paths: the primary flow from the hose socket to the
possible devices of this type, including not only the four atmosphere via a bag, filter and fan; and a secondary
vacuum cleaners dismantled but also other designs of flow from the atmosphere to the motor and vice versa,
which the authors were aware. The schema uses 16 which is required in all designs to cool the motor
concrete component types (including ‘Bag’ and windings.
‘Moulding’), nine abstract component types (including In addition, a model was built of the architecture of the
‘Airflow component’ and ‘Environment’), 11 concrete Sanyo SC-N200 vacuum cleaner. The architecture is shown
connection types (including ‘Attached to’ and ‘Air in Fig. 10, again using the graphical notation presented in
[flow]’) and one abstract connection type (‘Structurally Fig. 5. This served two roles during the formalisation of
connected to’). Each component type has an associated the problem:
CNC, in addition to which 19 DCCs and three ICCs
• During modelling it was used as an example architec-
were used. Although needed in other problems, FOCs
ture to check the schema for consistency (as described
were not used here.
in Sect. 4.1.3).
The majority of the entities in the schema were defined
• After modification, it was used as the start point for
by considering the types of physical components, and the
synthesis (as described in Sect. 5.2.2).
connections between them, present in the dismantled vac-
uum cleaners. Most of the abstract component types were The architecture contains 29 components and 50 con-
created in order to enable the same DCCs to apply to nections. The components and connections in the archi-
multiple concrete component types (e.g. ‘Airflow compo- tecture are connected to the corresponding component or
nents can be interconnected [by Air connections]’ and connection type in the schema using hyperlinks, which are
‘Components that must be contained must be contained by also used for clarity within the schema in Fig. 9, to link
one compartment’, two out of 10 such cases). Other NSCs to the connection types with which they are
abstract component types were created to give a ‘complete’ associated.
hierarchy and thereby assist the organisation and read- This stage took approximately 10 h in total:
ability of the diagram (e.g. ‘Vacuum cleaner component’ is
• 8 h to build the schema;
a supertype of all component types except ‘Atmosphere’).
• 2 h to build the SC-N200 architecture.
ICCs were used to ensure the existence of appropriate air

123
34 Res Eng Design (2012) 23:17–52

Fig. 9 The schema for the vacuum cleaner illustrative example. For types that were removed or collapsed when the schema was
clarity, the connection types are grouped separately with hyperlinks to simplified—see text for details
the relevant NSCs. Shading indicates component and connection

5.2.2 Model simplification and refinement architecture that could be completely decoupled were
removed (the ‘Wheels’, ‘Handle’, ‘Pressure relief valve’,
While constructing the complete schema and architecture ‘Mains cable’ and associated components), shown by light
models, it became clear that the incorporation of the grey shading in Fig. 10, and other appropriate modifica-
cyclone could be decoupled from many other parts of the tions were made. The schema was likewise simplified to
architecture such as the wheels or the mains cable storage. remove definitions for unused component and connection
Again this reflects the decomposable nature of many types, shown in Fig. 9 by dark grey shading for those
engineering problems described in Sect. 3.2. In order to component types that were collapsed into ‘Rest of vacuum
focus the computational synthesis on solving the problem cleaner’ and light grey for those that were removed.
at hand, therefore, the architecture of the SC-N200 vacuum In order to set up the problem for solution through
cleaner was simplified by removing or collapsing all ele- computational synthesis, the ‘Bag’ (outlined in black in
ments which, in the view of the authors, were peripheral to Fig. 10) was removed from the architecture and replaced
the problem. In particular, the ‘Airflow components’ after by a single ‘Cyclone’ component. This component repre-
the first filter (the ‘Fan’, ‘Motor’, ‘Rear compartment’ and sented the cyclone of air but not the surrounding casing;
‘Exhaust vent’) and the ‘Atmosphere’ were collapsed into a thus it was defined such that it had to be ‘comprised by’ at
single component labelled ‘Rest of vacuum cleaner’; this is least two ‘Mouldings’ (plastic structural parts; two were
shown by dark grey shading in Fig. 10. Parts of the specified to ensure that it can be disassembled to empty the

123
Res Eng Design (2012) 23:17–52 35

Fig. 10 The architecture of the Sanyo SC-N200 vacuum cleaner. Shading and outlining indicates regions that were removed or collapsed when
the schema was simplified—see text for details

collected dust) and required incoming and outgoing ‘Air’ and synthesis start point—before the architectures syn-
connections. In the architecture, the ‘Cyclone’ was given thesised were almost all realisable. However, one synthesis
‘Air’ connections corresponding to those of the ‘Bag’, but run did not necessarily correspond to a single problem with
no other connections. The task for the synthesis algorithm the schema, as multiple runs were often performed to
was therefore to find ways to incorporate the cyclone into resolve a single problem due to errors in the initial solution
the existing architecture. or interactions with other parts of the schema. Almost all
However, when the initial version of the schema and runs were short in duration as flaws in the schema were
architecture were used for synthesis, inspection of the frequently obvious from the first few architectures
results revealed many architectures that were unrealisable synthesised.
(i.e. failed the ‘Appropriateness evaluation’ in Fig. 1), The final versions of the simplified architecture and
indicating that the formalisation of the problem was schema are shown in Figs. 11 and 12. The final schema
incorrect and required refinement. The schema was there- contains seven concrete and five abstract component types,
fore revised (following the ‘Refinement of formalisation’ seven concrete and one abstract connection types, 13
iteration in Fig. 1) and the synthesis repeated. For example, DCCs, two FOCs and two ICCs. The final architecture
some architectures included a ‘Moulding’ that ‘comprised’ contains nine components and 13 connections. This stage
the ‘Cyclone’ without being structurally connected to the took approximately 9 h:
other components of the architecture, which would be
• 1 h to simplify the architecture and schema;
unrealisable under the interpretation used when construct-
• 8 h (spread over seven sessions) to iteratively refine the
ing the schema. To rule out such architectures, a new
simplified architecture and schema.
abstract component type ‘Structurally connected to’ was
introduced into the schema, with children ‘Hinged to’,
‘Clips to’, ‘Attached to’ and ‘Locates against’; a new ICC 5.3 Synthesis
was then defined, requiring every ‘Moulding’ to be (indi-
rectly) ‘Structurally connected to’ the ‘Rest of vacuum Once the schema and architecture were finalised, synthesis
cleaner’. In total, approximately 75 synthesis runs were was carried out to explore the problem of incorporating the
performed—interspersed with modifications to the schema cyclone. The configuration of the synthesis algorithm is

123
36 Res Eng Design (2012) 23:17–52

Fig. 11 The simplified schema for the illustrative example

shown in Table 3, along with some information about the connections, the synthesis algorithm was allowed to add up
results obtained. To establish appropriate maximum search to three components and six connections.
depths, a plausible ‘complete’ architecture was constructed
that satisfied all NSCs in the schema. In this architecture, 5.4 Interpretation of results
the ‘Cyclone’ was ‘comprised by’ two ‘Mouldings’, which
were ‘clipped to’ each other to form a unit that was The 1101 architectures resulting from synthesis were
physically separate from the other ‘Mouldings’ in the evaluated using the D-, C- and M-complexity metrics
vacuum cleaner and ‘contained by’ the ‘Front compart- defined in Sect. 4.3.2. Table 4 shows the minimum value of
ment’. The architecture had 11 components and 18 con- each metric, corresponding to the ‘best’ architecture in
nections. In order to allow architectures that were more each case, and the metric values obtained from the original
complex than this, the maximum search depths were cho- architecture. Figure 13 shows a three-dimensional scatter
sen to be one larger for both components and connections, plot of the individual results, again including the original
i.e., 12 components and 19 connections. Since the start architecture; the ‘best’ architectures are those closest to the
point shown in Fig. 12 has nine components and 13 origin of the plot. Comparing the synthesised architectures

123
Res Eng Design (2012) 23:17–52 37

Fig. 12 The simplified


architecture of the Sanyo
SC-N200, used as the synthesis
start point in the illustrative
example

Table 3 Synthesis configuration for the example problem Table 4 The minimum metric values achieved by the synthesised
architectures and the values obtained from the original architecture
Schema As shown in Fig. 11
Synthesis start point Architecture of Sanyo SC-N200 Metric Minimum value Original
vacuum cleaner, as shown in Fig. 12 obtained from architecture
synthesised architectures
Elementary operations Adding components and
connections of all types D-complexity 1.55 1.56
Maximum search Adding components: 3 C-complexity 0.33 0.37
depths Adding connections: 6 M-complexity 26 23
Search algorithm Stochastic search (500,000 samples)
CPU timea 11 h 0 min 2 s
would therefore be recommended for further investigation for
Number of 1101
architectures synthesised
the new vacuum cleaner. The architecture is shown in
Fig. 15, which is a copy of Fig. 12 with the components and
a
Intel Core 2 Duo P8600 processor at 2.4 GHz, 4 Gb RAM, Java connections added during synthesis shown in dark grey (the
Runtime Environment version 1.6.0_20 on Windows Vista 32-bit
diagram was created manually based on a force-directed
layout visualisation of the synthesis results). One possible
with the initial architecture, it has been possible to obtain a three-dimensional embodiment layout, out of the many
slight improvement in D-complexity and a significant defined by the architecture, is shown in Fig. 16. Interestingly,
improvement in C-complexity while the M-complexity has this bears certain similarities to the Dyson DC08 shown in
deteriorated due to the addition of the cyclone. However, it Fig. 17. Since Dyson have never manufactured bagged
can clearly be seen that there is a tradeoff between the three vacuum cleaners, there will have been little incentive for
criteria, and thus a single ‘best’ architecture does not exist. planned architecture carry-over from such products (although
In such a situation, one method for deciding between involuntary carry-over may still occur), and thus their
alternatives is weighted summation of the individual values architecture may be regarded as approaching the best
to produce a single ‘overall quality’ value (Q(A) in Eq. 2). available for a vacuum cleaner with a cyclone.
For this example, equal importance is assumed for the three In general, however, an architecture found using this
metrics and thus wC = wD = wM = 1 after normalising method may not be usable in practice. This may be because
each metric to return values between 0 and 1; in practice, it is unrealisable due to imperfections in the schema, or
these weights would be determined by the context of the because it violates constraints that have not been modelled,
design problem. for instance related to manufacturing processes. If this
QðAÞ ¼ wC CðAÞ þ wD DðAÞ þ wM MðAÞ ð2Þ occurs, one of several routes may be followed to obtain a
usable architecture:
Figure 14 shows a histogram of the Q values of the
synthesised architectures. In this case, the ‘best’ (lowest) Q • The schema may be modified to correct imperfections
value of 0.47 is exhibited by a single architecture, which (‘Refinement of formalisation’ in Fig. 1);

123
38 Res Eng Design (2012) 23:17–52

Fig. 13 The original


architecture of the vacuum
cleaner and the 1101
synthesised architectures,
evaluated with respect to the
three structural complexity
metrics. Data points
corresponding to synthesised
architectures are plotted as dark
grey circles/spheres. The
original architecture is indicated
by a black square/cube

Fig. 14 A histogram of the Q values of the 1101 synthesised architectures

123
Res Eng Design (2012) 23:17–52 39

Fig. 15 The synthesised


architecture with the lowest Q
value

• The recommended architecture may be modified man-


ually to repair problems;
• Other architectures with a higher (but still low overall)
Q value may be investigated and considered.
The routes require different amounts of effort, and thus
the route followed will depend on the design context.
To further understand the different possibilities pro-
duced by synthesis, the constraint-based classification
method described in Sect. 4.3.3 was applied to investigate
which ‘Mouldings’ ‘comprise’ the ‘Cyclone’ in the syn- Fig. 16 One possible embodiment of the architecture shown in Fig. 13.
The ‘Front moulding’ and ‘Moulding 4’ comprise the cyclone, while the
thesised architectures. As noted above, the schema speci-
‘Front compartment’ and ‘Compartment 3’ (hidden in this view) are
fies that the ‘Cyclone’ must be ‘comprised by’ two empty, and could potentially be used for tool storage
‘Mouldings’. Since the constraint-based classification
method does not distinguish between different instances of Figure 18b shows an Euler diagram of this classification.
a component type, the schema was modified by adding This now represents a complete description of the ‘Moul-
three new component types as children of ‘Moulding’, to dings’ that ‘comprise’ the ‘Cyclone’, since all architectures
represent the three mouldings present in the start point (the are either members of two of the original three categories
‘Front moulding’, the ‘Lower casing’ and the ‘Lid’). This or are members of one of the additional three and the ‘An
allowed the following categories to be defined using DCCs: extra moulding’ category. However, some architectures are
members of the ‘An extra moulding’ category as well as
• Front moulding comprises cyclone
two of the original three categories (e.g. architecture 1),
• Lower casing comprises cyclone
raising the question of what role(s) the extra moulding
• Lid comprises cyclone
plays. Inspecting a selection of these architectures could in
Figure 18a shows an Euler diagram for a set of 25 turn lead to the definition of patterns representing the dif-
architectures generated using the synthesis configuration in ferent roles, potentially leading to further investigation of
Table 3, classified with these three categories. The diagram the possibilities. Iteratively classifying architectures in this
shows that, as expected, in some architectures two of the way can help the designer to explore the solution space and
pre-existing ‘Mouldings’ ‘comprise’ the ‘Cyclone’; how- incrementally expand their understanding of possible
ever, in others only one of the existing ‘Mouldings’ is used, solutions to the original problem.
for instance only the ‘Lower casing’ in architecture 5.
By inspecting such cases individually, it was determined
that synthesis had added an additional ‘Moulding’ which 6 Evaluation
also ‘comprised’ the ‘Cyclone’. Thus, an additional cate-
gory was defined using an auxiliary CNC to identify those To establish the value of the method proposed in Sect. 4 for
architectures where an additional ‘Moulding’ was present. design practice, the Design Research Methodology

123
40 Res Eng Design (2012) 23:17–52

Fig. 17 The Dyson DC08 cyclonic vacuum cleaner: a assembled, and the mating surfaces of b the base unit and c the cyclone/dust collector
module

developed by Blessing and Chakrabarti (2009) recom- simple design scenario, involving generating new concepts
mends evaluation with respect to: for bicycle lighting systems. As source material, the par-
ticipants were supplied with an example set of bicycle
• Usability: the ease with which the method can be used
lights, shown in Fig. 19a, and a sheet of images spanning
for the intended task;
the range of available bicycle light designs, examples from
• Applicability: whether it has the intended direct effect
which are shown in Fig. 19b–d. They were also given a
on a design process; and
summary of the relevant standard (British Standards
• Usefulness: whether the direct effect leads to an
Institute 1986), although they were not required to produce
improvement in a high-level success factor, taking into
designs that conformed with it. The problem was chosen
account possible side effects.
due to participants’ familiarity with the domain; the aim of
In this case, the chosen high-level success factor is the the study was to establish how easily the participants found
quality of the architecture concept selected at the conclu- it to follow the steps of the method rather than to examine
sion of conceptual design. As discussed in Sect. 2.2.1, the the range of architectures produced, and thus the existence
quality of this selected concept will be positively correlated of a clear dominant design (represented by Fig. 19a) is not
with the thoroughness of the evaluation of the solution problematic.
space of product architectures, operationalised as the The six participants comprised four Ph.D. students and
number of candidate architectures examined during con- two postdoctoral researchers. All had backgrounds in
ceptual design. This operational definition has previously design research, but not computational design synthesis
been shown to be correlated with success in student design and all worked or studied at the Cambridge Engineering
projects (Yang 2008). Therefore, the intended direct effect Design Centre. Five out of six participants reported prior
of the method is to increase this number. experience with bicycle lights, all reported experience of
Preliminary evaluation of the method was carried out in mechanical design and modelling products as networks of
two separate studies, the first considering its usability and components, and three out of six had used applications
the second its applicability and usefulness. Since these took based on the CAM framework (though not the support tool
place in somewhat artificial contexts, an initial validation described in Sect. 4.4). Variations in prior experience did
study was subsequently carried out to explore all three not discernibly affect their performance on the example
aspects in a more realistic setting. task.
The experiment took place on a single occasion, after a
6.1 Usability pilot study with a single participant to assess feasibility
(results not included here). All participants were physically
6.1.1 Experiment setup present with access to their own laptop computers. The
participants were given a 1 h 20 min presentation covering
To assess usability, a laboratory study was carried out in the method, the software tool and the example design
which participants were asked to apply the method to a problem. The presentation suggested a procedure for using

123
Res Eng Design (2012) 23:17–52 41

of architectures. The participants then worked individually


on the problem for up to 4 h, with support from the authors
on the modelling language and software tool but no
assistance with decisions about the formalisation of the
problem. One participant subsequently continued working
on the problem unassisted for a further 4 h. Those who
successfully synthesised architectures were asked to pro-
vide sketches of possible embodiments of three of their
results, to establish whether the synthesis results could be
interpreted. Schemas, architectures and embodiment sket-
ches were recorded, and subjective impressions of the
method were elicited through a feedback questionnaire. In
addition, the authors recorded their own observations.

6.1.2 Results

After spending 3–8 h on the problem, five of the six par-


ticipants had constructed schemas that yielded multiple
realisable architectures when used for synthesis. All five
participants who reached this stage also demonstrated the
interpretation of 1–3 of their synthesised architectures.
Figure 20 shows examples from two participants of sche-
mas (a, d), synthesised architectures (b, e) and sketches of
possible embodiments (c, f). For clarity, the synthesised
architectures are presented using the graphical notation of
Fig. 5 and were manually laid out, although they were
originally presented to the participants using the force-
directed layout visualisation described in Sect. 4.3.1.
Although the participants were given considerable
freedom in choosing the type of bicycle light to model
(indeed as shown in Fig. 19b–d the sample images inclu-
Fig. 18 Automatically generated Euler diagrams for classifications of ded many different types), all five successful formalisations
a set of 26 vacuum cleaner architectures, a according to which of the were based on the type of lights shown in Fig. 19a. How-
pre-existing ‘Mouldings’ ‘comprise’ the ‘Cyclone’, b with an ever, one participant allowed for an external power supply
additional category to identify those architectures where synthesis (such as a dynamo) and another assumed that the lights
added a new ‘Moulding’
would be mounted on an armband rather than fixed to the
bicycle’s handlebars. Most participants used the connection
the method based on Fig. 7: choose a specific bicycle light types ‘contains’ and ‘attached to’ (these were used in an
architecture; model it; construct a schema able to represent example from a different design domain in the initial
it; add flexibility to the schema to represent a wider range presentation) and chose a similar level of detail for their

Fig. 19 a The example bicycle lights given to participants. b–d Examples from the sheet of images of alternative bicycle light designs,
illustrating the diversity in commercially available bicycle lighting systems

123
42 Res Eng Design (2012) 23:17–52

(a)
(b)

(c)

Bulbs 1

Light switch 1

Total casing 1

Power source 1

Fastening device 1

(d)
(e)

(f)

Light 1

Light plate 1

Switch 1

Battery 1

Compartment 1

Strip 1

Fig. 20 a A schema created by one participant as a formalisation of sketched embodiment layout (labels added for clarity). d–f A schema,
the problem. b An example architecture synthesised from the schema. architecture and sketch produced by a second participant
c The participant’s interpretation of this architecture in the form of a

concrete component types (a typical example is ‘Light component of a given type was connected isomorphically
source’, ‘Power supply’, ‘Switch’, ‘Socket’, ‘Casing’, (e.g. varying numbers of ‘Batteries’ ‘contained by’ a
‘Attachment’). Participants obtained 6–17 architectures from ‘compartment’), but only two allowed for true topolog-
synthesis, containing 4–8 components and 3–10 connections. ical variation. The participant who was not able to reach
Four participants employed abstract component types, in two the stage of successfully synthesising architectures also
cases to represent alternative options for components to attempted to allow variations in the numbers of com-
fulfil a certain role and in three cases to apply a constraint ponents, as well as alternative options for specific
to multiple concrete component types. Two participants components.
constructed an architecture manually and used the consis- Aside from user-interface issues, a small number of
tency checking facility described in Sect. 4.1.3 to check it conceptual problems with the method were uncovered
against the schema, while one constructed an architecture through the experiment: one participant initially believed
but did not use the software to check it. In addition, one the abstract component types represented an assembly
participant built a separate small schema in order to test hierarchy rather than generalisation-specialisation, and the
part of the problem. All five completed schemas allowed distinction between components/connections and compo-
variation in the number of some components if every nent/connection types was initially unclear. However, a

123
Res Eng Design (2012) 23:17–52 43

more serious issue was identified in the procedure sug- A common theme was that beneficial effects resulted
gested to the participants. Many participants began by from the activity of formalising the problem, rather than
constructing a schema to represent only their chosen ‘basis the results produced:
architecture’. Only once they had successfully constructed
‘the more valuable bit is probably the thought process
this did they experiment with relaxing the NSCs to gen-
while constructing the schema’
erate other architectures. Thus, although the procedure
enabled them to produce results from the method in many The tool, although a prototype, was usable for the task:
cases, it was not conducive to generating a wide range of
‘intuitive in application, easy to use, once you know
different architectures. Participants perceived this limita-
what you want to design and understand the meaning
tion, and during later discussion expressed interest in
of different connection types’
constructing schemas with wider scopes. This procedure
also led to difficulties on a user-interface level: once they ‘with a small amount of practice, I believe that it
could successfully synthesise architectures, participants often should be an easy tool to use’
tried relaxing the numerical ranges in the CNCs, but numerical
However, some participants felt that the example prob-
ranges elsewhere (typically in the DCCs) also required con-
lem was too simple to properly test the method:
certed changes which were often not immediately perceived
by the participants. For instance, if the basis architecture ‘[I] didn’t come up with anything I wouldn’t have
contained a single LED and the CNC for ‘LED’ was modified been able to come up with just using a piece of paper,
to require two LEDs, the DCC between the battery and the [it] felt a bit like shooting cannons at sparrows’
LEDs also required modification to allow each battery to
Overall, the study showed that the method could be
supply electricity to (at least) two LEDs.
applied by users unfamiliar with computational design
Despite the similar level of detail and similar (narrow)
synthesis to formalise a simple design problem, to syn-
scope chosen by many participants, the schemas con-
thesise solutions, and to interpret them as potential physical
structed demonstrate the potential for variation in formal-
embodiments. However, the participants were students and
ising a given design problem, as discussed in Sect. 4.1.2.
junior researchers rather than engineers with experience in
As an example, in Fig. 20 the embodiment sketches in
the relevant design domain, which may affect the way they
(c) and (f) are extremely similar, but the architectures
conceptualise the problem and thus the range of solutions
(b) and (e) represent these embodiments quite differently
obtained. Further research is necessary, in particular, to
and the schemas in (a) and (d) have little in common.
establish a procedure for formalising a design problem that
Nonetheless, since each participant was able to interpret
allows for innovation while giving the designer the clarity
the architectures synthesised from their schemas, the
and confidence to make rapid progress.
schemas are equally valid as representations of (segments
of) the space of possible designs.
In the feedback questionnaire, opinions differed as to 6.2 Applicability and usefulness
whether the method limited the variety of results:
6.2.1 Method
‘it is quite tempting to think along conventional lines’
‘[the results are] constrained around the chosen To investigate the applicability and usefulness of the pro-
model and mainly tell you what would happen if you posed method, it was applied to the design problem elicited
change [the] number of component[s], but do not during the case study described in Sect. 3 concerning the
yield radically new ideas’ design of the mounting arrangement for the engine com-
ponents in generator and pumping applications. The aim
or enabled more wide-ranging thinking:
was to establish whether the method led to more archi-
‘it helps you overcome your own inhibitions in that it tecture concepts than the original brainstorming (which
makes you consider design concepts that may not be would demonstrate applicability) and whether the best
realised in practice’ architecture concept was better (demonstrating usefulness).
Based on the six original architecture concepts, the first
‘specifying the relationship types makes you think a
author constructed a schema for this design problem,
bit wider [sic]’
shown in Fig. 21. The scope and level of detail were
‘Once the model was created, it was an effective way chosen to match the original architectures. Synthesis was
to think of alternative ideas… The schema also carried out using the configuration in Table 5; here, the
forced the modeller to develop a clear logic of the maximum search depths were chosen to match the size of
interaction between components.’ the original architectures. To evaluate the synthesised

123
44 Res Eng Design (2012) 23:17–52

Fig. 21 The schema used for the engine block mounting problem

architectures, four metrics were developed by the first Table 5 Synthesis configuration for the engine block mounting
author based on the four highest-weighted criteria in the problem
initial Pugh analysis: market competitiveness, product Schema As shown in Fig. 21
price, assembly complexity and technical confidence.
Synthesis start point Automatically generated minimal
These four criteria by themselves gave the same rank architecture: one ‘Cooler group’
ordering of the original architectures as the full set of cri- Elementary operations Adding components and
teria, and thus were considered a representative subset. The connections of all types
metrics were defined by assigning positive or negative scores Maximum search depths Adding components: 5
to specific structural features (components of particular types, Adding connections: 6
specific connections etc.) and evaluated by summing the Search algorithm Exhaustive depth first search
scores of the features present in an architecture. To allow CPU timea 2 min 13 s
comparison with the original design problem, the metrics Number of 256
returned the same scores for the original architectures as the architectures
engineers’ evaluations. This meant that larger metric values synthesised
correspond to ‘better’ architectures. Since these metrics used a
Intel Core 2 Duo P8600 processor at 2.4 GHz, 4 Gb RAM, Java
knowledge about the interpretation of the different component Runtime Environment version 1.6.0_20 on Windows Vista 32-bit

123
Res Eng Design (2012) 23:17–52 45

types in their definition, this is an example of the second Since the formalisation was carried out by the authors
method for assessing quality in a specific design context outside the original design context, to investigate the
described in Sect. 4.3.2. The definitions of the metrics were results further, the highest-ranked synthesised architectures
presented to engineers from the case study company and, after were presented to engineers at the case study company.
improvement, were acknowledged to be appropriate measures Their opinions were mixed. The approach was praised by
of the corresponding criteria. one engineer for bringing systematic approaches (which he
used in parametric design) into conceptual design:
6.2.2 Results ‘‘What I think you’ve come up with which I like, is
that rather than doing our normal type analysis to
A comparison of the ‘quality’ of the six original architec- identify the optimum design you can actually come
tures and the 256 synthesised architectures is shown in up with a more rigorous approach… [It takes]
Fig. 22; note that the original architectures also appear in some of the subjectivity away, but the subjectivity
the synthesis results. Interestingly, all six original archi- comes in the business’ understanding of what’s
tectures are comparatively high in quality. Only 78 out of important.’’
the 256 synthesised architectures (30.5%) had quality
values greater than or equal to that of the ‘worst’ original In addition, the automated evaluation and particularly
architecture (-50.3); if six different architectures were the ability to change metric weightings were viewed as
sampled randomly from the 256 in the design space, the valuable:
probability that they would all lie within the top 78 is ‘‘The weightings can change through a project …the
78 77 76 75 74 73
256  255  254  253  252  251 ¼ 0:07%: This small value market could change, the data that you’ve got to start
suggests that the original concepts were well chosen, with maybe isn’t as good as the data maybe you’ve
although other possible explanations are discussed at the got later on in the project… And therefore if you have
end of this section. Nonetheless, after removing trivial got a really robust model that you’ve generated, you
variations and those that were realisable but would not can actually change the weightings and see how the
serve a useful purpose, exhaustive search found two [rankings change].’’
architectures of a higher quality than the ‘best’ of the six
However, the specific results obtained were viewed as less
original architectures (quality values of 35.0 compared to
significant. One engineer criticised the schema for being
20.3), and four further architectures whose quality equalled
simultaneously underconstrained and overconstrained:
or exceeded that of the second-highest-quality original
architecture (values of 14.7 and -0.6 compared to -0.9). ‘‘I’m reasonably happy that you’ve [formalised the
design problem accurately], I mean it would have
been lovely if you’d come up with something we
haven’t done… there are more constraints that would
go in, I think, and then you probably wouldn’t get
[256] solutions, you may only get eight, but those
eight would potentially be workable solutions.’’
As for the quality of the synthesised architectures, their
rankings were regarded as contingent on the weightings,
perhaps because of their similarity to the original six
architectures:
‘‘So then it’s then a question of weighting, isn’t it? If
you change your weighting…’’
The two latter points reflect the fact that the schema was
constructed by researchers with no direct experience of the
design domain in question. A specific potential issue with
the metrics is that only the structural features that occurred
in the original architectures were incorporated into the
metric definitions; novel features exhibited only by the
Fig. 22 A histogram showing the overall quality of the 256
synthesised architectures, whether beneficial or detrimen-
synthesised architectures (grey) and the six original concepts (black),
using the four highest-weighted criteria from the original Pugh tal, could not be included due to the authors’ lack of
analysis knowledge. In practice, both of these problems would be

123
46 Res Eng Design (2012) 23:17–52

less severe since the engineers would either use the method 7 Reflections on the approach and method
themselves or work closely with those who did.
This evaluation suggests that the proposed method can 7.1 Challenges of formalisation
increase both the number of architecture concepts exam-
ined during conceptual design and the quality of those The method presented in this paper requires an architecture
selected at its conclusion, which leads to a preliminary design problem to be formalised into a schema and starting
conclusion that the method is both applicable and useful. point for synthesis. The first step is to abstract either
physical products or mental concepts into sets of compo-
6.3 Initial validation nents connected by relations, which may not be straight-
forward. For instance, a challenging part of the example in
To give a better perspective on the value of the method in a real Sect. 5 was representing the flow of air through the vacuum
engineering context, particularly ‘side effects’ both positive and cleaner, which required the components ‘Duct’ and
negative, one of the engineers at the case study company (who ‘Compartment’, neither of which exists independently but
participated in the presentation of synthesis results) was asked to is formed (‘comprised’) by a combination of other
apply it to the design problem immediately subsequent to that components.
described in the previous section. After an explanation of the Constructing the schema may also be challenging, in
method and software tool similar to that delivered in the usability particular identifying the ‘correct’ set of NSCs for a par-
experiment, and with assistance from the first author, within ticular product. ‘Hard’ constraints may follow from phys-
52 min the engineer had constructed a simple schema and syn- ical principles or the ‘function’ of the components
thesised 3 architectures which between them represented (at a concerned (such as the need for an electric motor to be
low level of detail) the 9 concepts originally brainstormed. Lack supplied with electricity), or from semantics (such as the
of time prevented the elaboration of the schema to represent the fact that one object cannot simultaneously contain and be
problem at a higher level of detail and with a wider scope, which contained by another). As discussed in Sect. 4.1.2, other
would allow more architectures to be synthesised and evaluated. restrictions limiting the scope of the schema may be
The engineer believed that the method would be useful: derived from the design context. However, the intention in
developing the method has been that the scope should be as
‘‘I think this is, an excellent tool to tease out and make
wide as possible and selection should be made through
people think, perhaps a little differently. […] Because really
quantitative evaluation as described in Sect. 4.3.2, with the
what you’re trying to do is you’re trying to understand what
metrics and their weightings embodying the designers’
used to be tacit knowledge in amongst a few people and
preferences; objective definition of metrics assures their
interpret it into [a] capability that anyone can use.’’
consistent application.
Interestingly, this echoes the comments made in the The applications, particularly the vacuum cleaner
usability study regarding the thought process involved in example in Sect. 5, demonstrated the importance of the
formalising a problem. issues of underconstraint and overconstraint discussed in
He also commented positively on some specific aspects Sect. 4.1.3, and of the methods for overcoming them: when
of the method: due simply to a mismatch between the designer’s intent and
the schema, such errors may be resolved by testing the
• The need to model even ‘common sense’ in NSC form,
schema for consistency against modelled architectures (to
a corollary of giving the designer complete control over
check for overconstraint) and inspecting the realisability of
the syntax and semantics of models to fulfil Require-
synthesised architectures (to check for underconstraint),
ments F2 and F3—particularly helpful for training
and incrementally refining the schema corresponding to the
novice designers;
‘Refinement of formalisation’ iteration in Fig. 1. Possible
• The ability of the method to provide quick interactive
improvements to the support tool to tighten this feedback
feedback to a designer, in contrast to other more accurate
loop could include, respectively:
analysis methods with a higher computational cost;
• The constraint validation mechanism described in Sect.
• Allowing the designer to construct models directly in
4.1.3, which was used during the study to diagnose
the language of component and connection types
modelling problems.
defined by a schema, rather than using generic ‘com-
Overall, despite the limited nature of this validation, it ponents’ and ‘connections’ as in Fig. 3, or by importing
demonstrated that the method was usable by an engineer product models from other software environments.
who also expressed positive opinions of the method’s • Facilitating interpretation and implicit evaluation by
potential usefulness—both in the strict sense of Sect. 6.2 presenting synthesis results in a way that better
and through other, less direct mechanisms. resembles a potential embodiment than the force-

123
Res Eng Design (2012) 23:17–52 47

directed layouts described in Sect. 4.3.1. In particular, method could be modified to allow formalisation of a
although structural connections (e.g. ‘attached to’) are design problem using only architectures, without the con-
represented intuitively in force-directed visualisations, struction of an explicit schema by the designer. For
geometrical relations (e.g. ‘contains’) are not. An instance, the designer might start from an architecture and
alternative would be to create preliminary geometrical indicate on it where variation was allowed (e.g. alternative
embodiments as in Fig. 20c, even if such embodiments ways in which components may be linked together). Such a
are not themselves feasible (the ‘topological solutions’ representation of the problem might be easier to construct,
of Medjdoub and Yannou (2000)). A mechanism for since there would be no need to obtain or generate multiple
doing so would be to use the connections in the alternative concepts before constructing the schema.
architecture to define geometrical constraints in However, it would potentially be less general: where more
embodiment sketches, which may then be resolved sweeping design changes are allowed, it may be difficult to
using a numerical constraint solver (Mullineux et al. show on a single architecture all the changes that would be
2005). necessary to convert it to a very different architecture.
• Allowing the designer to ‘preview’ the impact of Lastly, while the representation has proved sufficiently
possible schema modifications on the synthesis output. powerful to express the problems considered to date, there
may be situations where greater expressive power is
A more serious potential issue arises if the designer’s
required. Particular directions for improvement include
mental model of the space of possible architectures does
extending the language of network structure constraint
not match reality; in particular, if their understanding is
types to include Boolean combinations of constraints and
overconstrained (the ‘fictitious constraints’ of Pahl et al.
non-universal, i.e., existential quantification; and allowing
(2007)). For instance, in the experiment in Sect. 6.1, many
components to have multiple distinct ports to which con-
participants assumed that the bicycle light should be bat-
nections can be made.
tery-powered, although this was not specified (and indeed
the source material included bicycle lights powered by 7.2 Decomposing design problems
some form of dynamo, such as Fig. 19b–d). However, as
stated in Sect. 2.2.1 and reinforced in Sect. 6.3 the use of a An advantage of the approach is that, as shown by the
formal representation makes the assumptions about the applications, the separation between the ‘real world’ and
solution explicit (and thus reviewable) rather than implicit the models (traversed by the formalisation and interpre-
within the thought processes of the designers. More tation processes) make it possible to find solutions to a
importantly, the authors’ experiences in the example specific subpart of a design problem—as illustrated in the
described in Sect. 5 and the participants’ comments in example in Sect. 5, where the arrangement of the motor,
Sects. 6.1 and 6.3 show that constructing the schema may fan etc. were not considered during synthesis. This may
itself stimulate thought on the part of the designer to guide be more practical than complete redesign in a real-life
them towards new solutions. In particular, formalisation context and reduces the scale of the problem, making it
encourages questioning assumptions by attempting to more tractable. Increasing the complexity of the schema
(mentally) construct realisable architectures that do not and the size of the design problem increases the likeli-
conform to the schema; the resulting iterations involve an hood that erroneous assumptions will have been included;
increase in the understanding of the designer. To improve while these may be resolved as discussed in the previous
the efficacy of this mental constraint-challenging process, it section, this may require significant effort. On the other
could potentially be made explicit and documented, by hand, restricted focus will limit the possible solutions,
explicitly questioning each NSC as it is entered into the potentially missing solutions that are significantly ‘better’
support tool, along the lines of systems used to challenge but that require changes outside the scope of consider-
the ‘normal flow’ of events when eliciting use cases ation—for instance, the simplified vacuum cleaner rede-
(Maiden 1998). sign problem could not generate an architecture where the
From a practical point of view, Sect. 6.1 demonstrated fan is located before the dust separator in the airflow path
limitations with the procedure for constructing a schema (the ‘dirty-fan’ configuration found in upright vacuum
based on a single architecture, in terms of the tendency cleaners).
towards overconstraint. Further research is necessary to
identify a more appropriate procedure for formalising 7.3 Exploiting results from the method
design problems that avoids this issue but is sufficiently
prescriptive that experienced designers who have signifi- Once potential architectures have been synthesised, those
cant domain knowledge but are novices to the method are which show most promise and should be taken forward into
able to follow a clear series of steps. Alternatively, the later design must be identified. Previous research by

123
48 Res Eng Design (2012) 23:17–52

Langdon and Chakrabarti (2001) has indicated the diffi- of the numerical relationships defined by an architecture
culty of navigating large numbers of automatically syn- could themselves be parameterised, they would form
thesised abstract solutions using only the solution structure. appropriate criteria for selection when combined with
However, the use of metrics to quantify solution quality, specific performance requirements. Developing methods
potentially in combination with Langdon and Chakrabarti’s for generic performance-based selection is an important,
structure-based clustering, may make navigation of the though challenging, future direction for this research
solution space easier. Interpreting individual solutions may area.
be facilitated by creating rough geometrical embodiments In Sect. 2.2, it was suggested that the models of schemas
as discussed in Sect. 7.1. It may also be important to pro- and architectures could also, in principle, contribute to
vide guidance to designers on how to choose alternative design documentation; however, as seen in the example
architectures if the one ranked highest according to the such schema and architecture diagrams can be complex
metrics used is not realisable. despite their visual nature. Nonetheless, the participant in
In addition, as described in Sect. 3.2, product architec- Sect. 6.3 indicated the possibility for using the models to
tures are often not designed in a single event but as a series communicate previously tacit information. Further research
of linked subproblems, and the architecture may not be is necessary to investigate further how readily they could
finalised before the end of design. Thus, if an architecture be interpreted by someone not involved in their
is chosen based only on its own ‘quality’, subsequent construction.
changes might cause it to become less desirable. Therefore, The computational cost of the method remains an issue
the robustness-to-change of an architecture should be for large problems. Characterising the scaling of the
considered as a selection criterion (Fricke and Schulz 2005; method is not straightforward, since each problem mod-
de Weck et al. 2007). elled has a typical size of the architectures that might
In many situations, the key to architecture choice is solve it, and different problems can have significantly
some aspect of ‘technical performance’. However, as different properties. Figure 23 shows the exponential
stated in Sect. 4.3.2, it is not possible to evaluate the increase in both synthesis time and design space size with
technical performance of an architecture based purely on search depth for a simple problem; however, this over-
the information contained within a model of it. Indeed, estimates the complexity for more complicated problems
an architecture cannot be said to have a particular level with larger numbers of component and connection types,
of technical performance, as that is dependent on the since at present isomorphic graphs with different node
design parameters of the components that comprise it. ordering are regarded as distinct. This source of ineffi-
Nonetheless, the architecture does define the set of ciency could be addressed within the existing state space
relationships that determine the overall performance from search framework through graph canonisation (McKay
individual component parameters (or, equivalently, the 1981). Alternatively, Jackson (2006) proposes that similar
component parameter values required to achieve a par- ‘model-finding’ problems in software engineering may be
ticular overall performance). Inasmuch as the properties addressed using Boolean constraint satisfaction (SAT)

Fig. 23 The relationships


between search depth, number
of architectures synthesised and
time taken for exhaustive
synthesis in a simple design
problem. Intel Core 2 Duo
P8600 processor at 2.4 GHz,
4 Gb RAM, Java Runtime
Environment version 1.6.0_23
on Windows 7 64-bit

123
Res Eng Design (2012) 23:17–52 49

techniques. To assess whether such techniques offer 7.5 Further evaluation


improvements in efficiency over state space search, an
attempt has been made to implement a translator from the While initial evaluation of the method has been carried out
formalism described in this paper to the Alloy language and has proved positive, as described in Sect. 6, further
defined by Jackson. However, using the ‘natural’ repre- evaluation is needed. In particular, while any formal
sentation for an architecture in Alloy (representing com- approach has potential advantages over informal approa-
ponents by instances and connections by relations), ches, it also has an associated cost due to the need to
certain constructs from the method proposed in this paper formalise the problem and interpret the results. In an
are not translatable. Specifically, an ICC may be trans- industrial context, it may be advantageous to consider a
lated using the Alloy reflexive transitive closure construct lower-quality solution (Fricke 1992) if it saves time.
but this is purely a test for the existence of at least one Evaluation should cover both the direct cost and benefit
path and cannot test the size of the set of paths or the of using the method and other, less tangible benefits such as
presence of components or connections on a path. This increasing the designer’s understanding of the set of pos-
means that, while the models can be exported and run in sible solutions, as described in Sect. 7.1.
Alloy, the results do not correspond to those produced
from our tool (typically including additional unrealisable
architectures). Additionally, Alloy cannot remove com- 8 Conclusions
ponents or connections from a start point, preventing
exploration of architectures derived from an existing The design of product architectures is a challenging and
product. Work is continuing to establish whether trans- important part of the design process. This paper has
lation is possible and whether SAT is superior. explored the use of formal computational approaches to
assist product architecture design. Specifically, the paper
7.4 Applicability to different forms of design proposes that alternative architectures for a product may be
synthesised computationally starting from formal declara-
An important issue is the suitability of the method for tive descriptions of the ‘design space’, termed a ‘schema’
different forms of design—original, incremental and for a design problem, and of existing product architectures.
adaptive, in the terminology of Pahl et al. (2007). Since the This method has the advantages of searching systemati-
method requires the possible types of components and cally for possible architectures, making explicit any
connections to be specified during formalisation, original assumptions about the form of the solution, and evaluating
design in the sense described by Pahl et al. (starting from a the generated architectures objectively, while meeting
description of the product’s function) is not supported. At requirements elicited from a case study of product archi-
the same time, the method’s applicability goes beyond the tecture design in a diesel engine manufacturing firm. The
redesign of existing products, since a novel product may be method has been illustrated through an example application
(and frequently is, as discussed in Sect. 2.2.2) constructed to a hypothetical realistic design problem: adapting a
from known ‘partial solutions’. Indeed, as noted in Sect. vacuum cleaner to use a cyclonic dust separator instead of a
4.1.2 it is possible to incorporate a notion of ‘function bag. Preliminary evaluation suggests that the method is
embodiment’ into a schema using abstract component usable, applicable and useful in practice. Usability was
types. A possible extension of the method to increase its assessed through laboratory experiments, in which five out
applicability in original design would be to allow abstract of six participants could use the method to develop new
component types representing functions to be linked to architectures for bicycle lights when provided with
concrete component types from a predefined ‘library’, to appropriate training and support. A historical design
allow the system to suggest function embodiments auto- problem from the case study company showed that the
matically. Potential issues include maintaining the library method was applicable, in that it generated more solutions
and ensuring consistency of terminology (Hirtz et al. than were originally considered by the engineers, and
2002). Furthermore, Requirement F3 and the initial vali- useful, in that a small number of the solutions were ‘better’
dation in Sect. 6.3 indicate that such an approach may not (according to the quality metrics used in the original design
be appropriate in practice. problem) than some of the original architectures. Finally,
The solution of technology infusion problems such as use of the method by an engineer from the case study
that described in Suh et al. (2008), where the infused company supported its usability and overall value, as well
technology is not a single component but rather a system of as additional potential benefits. Future work will consider
multiple components may be facilitated on an implemen- practical procedures for formalising specific design prob-
tation level by allowing multiple schemas and synthesis lems and will investigate the cost-benefit ratio of this
start points to be combined during synthesis. approach compared with that of other methods for

123
50 Res Eng Design (2012) 23:17–52

designing architectures, through further evaluation in real- Corallo A, Laubacher R, Margherita A, Turrisi G (2009) Enhancing
life settings. product development through knowledge-based engineering
(KBE): a case study in the aerospace industry. J Manuf Tech
Manag 20:1070–1083. doi:10.1108/17410380910997218
Acknowledgments The authors thank the engineers at the case Crawley EF, de Weck OL, Eppinger SD, Magee C, Moses J, Seering
study company and Claudia Eckert of the Open University for her WP, Schindall J, Wallace D, Whitney DE (2004) The influence
assistance with carrying out the interviews. They also thank the of architecture in engineering systems. Massachusetts Institute of
members of the Cambridge Engineering Design Centre who took part Technology Engineering Systems Division monograph
in the usability evaluation described in Sect. 6.1, and the three Cross NG (1994) Engineering design methods: strategies for product
anonymous reviewers for their insightful comments. Figure 2 is design. Wiley, Chichester
reproduced from Starling and Shea (2005), originally published by de Weck OL, Suh ES, Chang D (2007) Flexible product platforms:
ASME, and the images of bicycle lights in Fig. 19b–d are reproduced framework and case study. Res Eng Des 18:67–89
by kind permission of Pedalite International Ltd (http://www. Dyson J (1986) Vacuum cleaning appliance. Patent number 4593429,
pedalite.com), Reelight ApS (http://www.reelight.com) and Brando United States Patent Office
WorkShop (http://www.gadget.brando.com). This research was Eades P (1984) A heuristic for graph drawing. Congressus Numer-
funded by the George and Lilian Schiff Studentship and the UK antium 42:149–160
Engineering and Physical Sciences Research Council (grant reference Eckert CM, Clarkson PJ, Zanker W (2004) Change and customisation
EP/E001777/1). in complex engineering domains. Res Eng Des 15:1–21. doi:
10.1007/s00163-003-0031-7
Fricke G (1992) Experimental investigation of individual processes in
engineering design. In: Cross N, Doorst K, Roozenburg N (eds)
References Research in design thinking. Delft University Press, The
Netherlands, pp 105–109
Alexander C (1964) Notes on the synthesis of form. Harvard Fricke E, Schulz AP (2005) Design for changeability (DfC):
University Press, Cambridge principles to enable changes in systems throughout their entire
Allwood JM (2007) A structured search for novel manufacturing lifecycle. Sys Eng 8:342–359. doi:10.1002/sys.20039
processes leading to a periodic table of ring rolling machines. Giffin ML, de Weck OL, Bounova G, Keller R, Eckert CM, Clarkson
J Mech Des 129:502–511 PJ (2009) Change propagation analysis in complex technical
Andreasen MM (1992) Designing on a ‘designer’s workbench’ systems. J Mech Des 131:081001.1–081001.14. doi:10.1115/
(DWB). In: Proceedings of the 9th WDK workshop. Rigi, 1.3149847
Switzerland Golkar AA, Keller R, Robinson B, de Weck OL, Crawley EF (2009)
Antonsson EK, Cagan J (2001) Formal engineering design synthesis. A methodology for system architecting of offshore oil produc-
Cambridge University Press, Cambridge tion systems. In: Proceedings of the international design
Berliner C, Brimson JA (1988) Cost management for today’s structure matrix conference (DSM’09). Greenville, South Car-
advanced manufacturing: the CAM-I conceptual design. Harvard olina, USA, pp 343–356
Business School Press, Boston Gordon WJJ (1961) Synectics. Harper and Row
Bermell-Garcı́a P, Fan I (2002) A KBE System for the design of wind Heer J (2007) The prefuse information visualisation toolkit.
tunnel models using reusable knowledge components. In: http://www.prefuse.org
Proceedings of the VI Congreso Internacional de Ingenierı́a de Hellenbrand D, Lindemann U (2008) Using the DSM to support the
Proyectos. Barcelona, Spain selection of product concepts. In: Proceedings of the international
Blessing LT, Chakrabarti A (2009) DRM, a design research design structure matrix conference (DSM’08). Stockholm, Sweden
methodology. Springer, London Helms B, Shea K, Hoisl F (2009) A framework for computational
Bradley DA, Bracewell RH, Chaplin RV (1992) Engineering design design synthesis based on graph-grammars and function-
and mechatronics: the Schemebuilder project. Res Eng Des behaviour-structure. In: Proceedings of the ASME international
4:241–248. doi:10.1007/BF02032467 design engineering technical conferences and computers and
Brimble R, Sellini F (2000) The MOKA modelling language. In: information in engineering conference (IDETC/CIE 2009). San
Proceedings of the 12th international conference on knowledge Diego, California
engineering and knowledge management (EKAW2000). Juan- Hirtz J, Stone R, McAdams D, Szykman S, Wood K (2002) A
les-Pins, France functional basis for engineering design: reconciling and evolving
British Standards Institute (1986) Cycles–specification for photomet- previous efforts. Res Eng Des 13:65–82
ric and physical requirements of lighting equipment. British Huang GQ (1996) Design for X—concurrent engineering imperatives,
Standard 6102–3:1986 1st edn. Chapman & Hall, London
Cagan J, Campbell MI, Finger S, Tomiyama T (2005) A framework Jackson D (2006) Software abstractions: logic, language and analysis.
for computational design synthesis: model and applications. MIT Press, Cambridge
J Comput Info Sci Eng 5:171–181. doi:10.1115/1.2013289 Jarratt TA, Eckert CM, Weeks R, Clarkson PJ (2003) Environmental
Campbell MI, Cagan J, Kotovsky K (2000) Agent-based synthesis of legislation as a driver of design. In: Proceedings of the
electromechanical design configurations. J Mech Des 122:61–69. international conference on engineering design (ICED ‘03).
doi:10.1115/1.533546 Stockholm, Sweden, pp 231–232
Chakrabarti A (2002) Engineering design synthesis: understanding, Jiao J, Tseng MM (1999) A methodology of developing product family
approaches and tools. Springer, London architecture for mass customization. J Intell Manuf 10:3–20
Chakrabarti A, Bligh TP (1996) An approach to functional synthesis Keller R, Eckert CM, Clarkson PJ (2006) Heuristics for change
of mechanical design concepts: theory, applications, and prediction. In: Proceedings of the international design confer-
emerging research issues. AIEDAM 10:313–331 ence (DESIGN 2006). Dubrovnik, Croatia, pp 873–880
Chomsky N (2002) Syntactic structures. Mouton de Gruyter Kerzhner AA, Paredis CJ (2009) Using domain specific languages to
Clarkson PJ, Simons CS, Eckert CM (2004) Predicting change capture design synthesis knowledge for model-based systems
propagation in complex design. J Mech Des 126:788–797 engineering. In: Proceedings of the ASME international design

123
Res Eng Design (2012) 23:17–52 51

engineering technical conferences and computers and informa- Shapiro A (2002) TouchGraph version 1.22. http://www.touchgraph.com
tion in engineering conference (IDETC/CIE 2009). San Diego, Shea K, Cagan J, Fenves SJ (1997) A shape annealing approach to
California, USA optimal truss design with dynamic grouping of members. J Mech
Knuth D (1998) The art of computer programming, volume 3: sorting and Des 119:388–394
searching, 2nd edn. Addison-Wesley, Reading, Massachusetts Simon HA (1996) The sciences of the artificial, 3rd edn. MIT Press,
Kurtoglu T, Campbell MI (2009) Automated synthesis of electrome- Cambridge
chanical design configurations from empirical analysis of Sosa ME, Eppinger SD, Rowles CM (2007) A network approach to
function to form mapping. J Eng Des 20:83–104. doi: define modularity of components in complex products. J Mech
10.1080/09544820701546165 Des 129:1118–1129
Langdon PM, Chakrabarti A (2001) Improving access to design Stanković T, Shea K, Štorga M, Marjanović D (2009) Grammatical
solution spaces using visualisation and data reduction tech- evolution of technical processes. In: Proceedings of the ASME
niques. In: Proceedings of the international conference on international design engineering technical conferences and
engineering design (ICED ‘01). Glasgow, UK computers and information in engineering conference (IDETC/
Lindemann U, Maurer MS, Braun T (2008) Structural complexity CIE 2009). San Diego, California
management: an approach for the field of product design. Starling AC, Shea K (2005) A parallel grammar for simulation-driven
Springer, Berlin mechanical design synthesis. In: Proceedings of the ASME
Lipson H, Pollack JB (2000) Automatic design and manufacture of international design engineering technical conferences and
robotic lifeforms. Nature 406:974–978. doi:10.1038/35023115 computers and information in engineering conference (IDETC/
Maiden NA (1998) CREWS-SAVRE: scenarios for acquiring and CIE 2005). Long Beach, California
validating requirements. J Autom Softw Eng 5:419–446 Stiny G (2006) Shape: talking about seeing and doing. MIT Press,
Maier MW, Rechtin E (2000) The art of systems architecting, 2nd Cambridge
edn. CRC Press, Boca Raton Stone RB, Wood KL, Crawford RH (2000) A heuristic method for
McDermott J (1980) R1: an expert in the computer systems domain. identifying modules for product architectures. Des Stud 21:5–31.
In: Proceedings of the first national conference on artificial doi:10.1016/S0142-694X(99)00003-4
intelligence (AAAI-80). Stanford University, California Stumptner M, Friedrich GE, Haselböck A (1998) Generative
Mckay BD (1981) Practical graph isomorphism. Congressus Numer- constraint-based configuration of large technical systems. AIE-
antium 30:45–87 DAM 12:307–320. doi:10.1017/S0890060498124046
Medjdoub B, Yannou B (2000) Separating topology and geometry in Suh ES, Furst MR, Mihalyov KJ, de Weck OL (2008) Technology
space planning. Comput-Aided Des 32:39–61. doi:10.1016/ infusion: an assessment framework and case study. In: Proceed-
S0010-4485(99)00084-6 ings of the ASME international design engineering technical
Mesihovic S, Malmqvist J (2000) PDM system based support for the conferences and computers and information in engineering
sales-delivery process of engineer-to-order products. In: Pro- conference (IDETC/CIE 2008). Brooklyn, New York, USA
ceedings of product models. Linköping, Sweden Summers JD, Shah JJ (2010) Mechanical engineering design com-
Mintel (2008) Vacuum cleaners. Consumer report, Mintel Interna- plexity metrics: size, coupling, and solvability. J Mech Des
tional Group 132:021004.1–02100411. doi:10.1115/1.4000759
Mullineux G, Hicks B, Medland A (2005) Constraint-aided product Sun Microsystems (2006) Java SE version 6. http://java.sun.com
design. Acta Polytechnica 45:31–36 Ulrich KT (1995) The role of product architecture in the manufac-
Murdoch TN, Wallace KM (1992) An approach to configuration turing firm. Res Policy 24:419–440
optimization. J Eng Des 3:99–116 Ulrich KT, Seering WP (1989) Synthesis of schematic descriptions in
Object Management Group (2006) Meta object facility (MOF) Core mechanical design. Res Eng Des 1:3–18
Specification v2.0 Umeda Y, Ishii M, Yoshioka M, Shimomura Y, Tomiyama T (1996)
Ogawa A (1984) Separation of particles from air and gases. CRC Supporting conceptual design based on the function-behavior-
Press, Boca Raton state modeler. AIEDAM 10:275–288
Osborn A (1957) Applied imagination. Scribner, New York Wood KL, Greer JL (2001) Function-based synthesis methods in
Pahl G, Beitz W, Feldhusen J, Grote K-H (2007) Engineering design: engineering design. In: Antonsson EK, Cagan J (eds) Formal
a systematic approach, 3rd edn. Springer, London engineering design synthesis. Cambridge University Press,
Paredis CJJ, Diaz-Calderon A, Sinha R, Khosla PK (2001) Composable Cambridge, pp 170–227
models for simulation-based design. Eng Comput 17:112–128 Wyatt DF, Eckert CM, Clarkson PJ (2009a) Design of product
Port O, Schiller Z (1990) A smarter way to manufacture. Business- architectures in incrementally developed complex products. In:
Week 110–117 Proceedings of the international conference on engineering
Pugh S (1990) Total design: integrated methods for successful design (ICED ‘09). Palo Alto, California
product engineering. Addison, Wesley Wyatt DF, Wynn DC, Clarkson PJ (2009b) Exploring spaces of
Purcell AT, Gero JS (1996) Design and other types of fixation. Des system architectures using constraint-based classification and Euler
Stud 17:363–383. doi:10.1016/S0142-694X(96)00023-3 diagrams. In: Proceedings of the international design structure
Ranta M, Mantyla M, Umeda Y, Tomiyama T (1996) Integration of matrix conference (DSM’09). Greenville, South Carolina
functional and feature-based product modelling—the IMS/ Wynn DC, Nair SM, Clarkson PJ (2009) The P3 platform: an
GNOSIS experience. Comput-Aided Des 28:371–381. doi: approach and software system for developing diagrammatic
10.1016/0010-4485(95)00056-9 model-based methods in design research. In: Norell Bergendahl
Rohrbach B (1969) Kreativ nach Regeln–Methode 635, eine neue M, Grimheden M, Leifer L, Skogstad P, Lindemann U (eds)
Technik zum Lösen von Problemen. Absatzwirtschaft 12:73–75 Proceedings of the international conference on engineering
Russell SJ, Norvig P (2003) Artificial intelligence: a modern design (ICED ‘09). Palo Alto, California, USA, pp 559–570
approach, 2nd edn. Prentice Hall, Englewood Cliffs Xie YM, Steven GP (1997) Evolutionary structural optimisation.
Schmidt LC, Cagan J (1997) GGREADA: a graph grammar-based Springer, London
machine design algorithm. Res Eng Des 9:195–213. doi: XML Schema Working Group (2004) XML Schema. http://
10.1007/BF01589682 www.w3.org/TR/xmlschema-1

123
52 Res Eng Design (2012) 23:17–52

Yang MC (2008) Observations on concept generation and sketching Ziv-Av A, Reich Y (2005) SOS–subjective objective system for
in engineering design. Res Eng Des 20:1–11. doi:10.1007/ generating optimal product concepts. Des Stud 26:509–533
s00163-008-0055-0 Zwicky F (1969) Discovery, invention, research–through the mor-
Yassine AA, Wissmann LA (2007) The implications of product phological approach. Macmillian, Toronto
architecture on the firm. Sys Eng 10:118–137

123

You might also like