Artificial Intelligence for Engineering Design, Analysis and Manufacturing (2010), 24, 49–62.

Copyright # Cambridge University Press, 2010. 0890-0604/10 $25.00

doi:10.1017/S0890060409990163

Automating the conceptual design process: “From black box

to component selection”

TOLGA KURTOGLU,1 ALBERT SWANTNER,2 AND MATTHEW I. CAMPBELL2

1
Mission Critical Technologies, NASA Ames Research Center, Intelligent Systems Division, Moffett Field, California, USA
2
Automated Design Laboratory, Department of Mechanical Engineering, University of Texas at Austin, Austin, Texas, USA
(RECEIVED October 15, 2008; ACCEPTED July 3, 2009)

Abstract
Conceptual design is a vital part of the design process during which designers first envision new ideas and then synthesize
them into physical configurations that meet certain design specifications. In this research, a suite of computational tools is
developed that assists the designers in performing this nontrivial task of navigating the design space for creating conceptual
design solutions. The methodology is based on automating the function-based synthesis paradigm by combining various
computational methods. Accordingly, three nested search algorithms are developed and integrated to capture different
design decisions at various stages of conceptual design. The implemented system provides a method for automatically gen-
erating novel alternative solutions to real design problems. The application of the approach to the design of an electrome-
chanical device shows the method’s range of capabilities and how it serves as a comparison to human conceptual design
generation and as a tool suite to complement the skills of a designer.
Keywords: Automated Design; Concept Generation; Functional Design; Graph Grammars

1. INTRODUCTION The difficulty may hinge on the very nature of conceptual

design, which does not lend itself easily to automation. The
Conceptual design plays the central role in ensuring design
conceptual design process begins with the specification of
quality and a high level of innovation. It is in this phase
the product to be designed and involves the continual cycle
that the architecture of the final design is established, the tech-
of concept generation and evaluation until a design opportu-
nologies are chosen to fulfill the customer needs, and the bulk
nity is transformed into an embodied solution that satisfies a
of the cost for a product is committed. Because of these char-
set of design requirements.
acteristics, conceptual design is often considered the most
This systematic view of conceptual design starts with the
important phase of the product development cycle.
formulation of the overall function of the product to be de-
Yet, the conceptual design process has seen few attempts at
signed. This high level product function is then decomposed
automation. The concept of “automating design” has often
recursively into lower level functions—a process that pro-
been leveraged in later stages of the design process where a
duces a function structure, which is a representation that
to-be-designed artifact accrues numerous parameters but
defines function as transformation between energy, material,
lacks specific dimensions. Automated methods such as opti-
and information (Pahl & Beitz, 1996). The function structure
mization provide a useful framework for managing and deter-
is then used to generate solutions to each of the product sub-
mining details of the final designed artifact. These methods
functions. Here, the designer seeks solutions (a component or
make the design process less tedious and time-consuming,
a set of components that perform a particular function). Next,
and they are used in a wide variety of industries to support
solutions to the subfunctions are synthesized together to
or optimize current design efforts. However, one of the per-
arrive at the final architecture or configuration of a product.
vasive bottlenecks in design is the lack of continuity between
Finally, the design is embodied by the selection of designed
computational design tools and conceptual design methods.
components (Suh, 1990; Ullman, 1995; Ulrich & Eppinger,
1995; Pahl & Beitz, 1996; Otto & Wood, 2001). Using
this systematic view of conceptual design, a broad number
Reprint requests: Tolga Kurtoglu, Mission Critical Technologies, NASA
Ames Research Center, Intelligent Systems Division, Moffett Field, of concepts can be generated by making decisions about
CA 94035, USA. E-mail: [email protected] the decomposition of the overall product function and the
49
50 T. Kurtoglu et al.

selection and integration of different design solutions to and Seering (1989) developed a mechanical design “com-
elemental subfunctions. piler” to support catalog-based design. Bracewell and Sharpe
In this research, we automate the aforementioned concep- (1996) developed “Schemebuilder,” a software tool using
tual design process starting from a black box1 level product bond graph methodology to support the functional design
specification to the physical embodiment of design compo- of dynamic systems with different energy domains. Chakra-
nents. Accordingly, we develop a suite of automation tools barti and Bligh (1996) model the design problem as a set of
that combine and formalize the systematic view of the con- input–output transformations, where structural solutions to
ceptual design process (Suh, 1990; Ullman, 1995; Ulrich & each of the instantaneous transformation are found, and infea-
Eppinger, 1995; Pahl & Beitz, 1996; Otto & Wood, 2001). sible solutions are filtered according to a set of temporal rea-
The implemented system consists of three nested search algo- soning rules. Bryant et al. (2005) developed a concept genera-
rithms that can capture different design decisions at various tion technique that utilizes a function–component matrix and
stages of conceptual design and serves as a comparison to hu- a filter matrix to generate a morphological matrix of solutions
man conceptual design generation. during conceptual design. The A-Design research (Campbell
The remainder of this paper is broken up into six sections. et al., 2000) is an agent-based system that synthesizes compo-
The next section talks about the different approaches and nents based on the physical interactions between them.
techniques developed to automate the conceptual design pro- Function structure research, in contrast, has found its way
cess. An overview of the proposed method and the three into a number of educational texts since the presentation pro-
search algorithms are described in Section 3. Section 4 pre- vided by Pahl and Beitz (1996). Furthermore, research build-
sents the implementation of the method on a case study in ing upon the concept of function structures has flourished in
which an electromechanical device is synthesized. Various the past 15 years. Numerous publications have extended the
results obtained from this example are analyzed in Section 5. application of function structures and formalized the use of
Finally, the conclusions are presented in Section 6. such structures (Otto & Wood, 1997; Umeda & Tomiyama,
1997; Kirshman & Fadel, 1998; Kitamura & Mizoguchi,
1999; Stone & Wood, 1999; Szykman et al., 1999; Hirtz
2. RELATED WORK et al., 2002). Computational approaches have also been ex-
Various researchers have employed different methods to com- plored that further expand the value of function structures
putationally support the conceptual phase of design. These (Wang & Yan, 2002). One of the interesting implementations
methods include computer techniques such as constraint pro- of automating the function-based design is the work of Srid-
gramming (Kota & Chiou, 1992; Subramanian & Wang, haran and Campbell (2004), which uses graph grammars.
1995), qualitative symbolic algebra (Williams, 1990), expert Graph grammars are composed of rules for manipulating
systems (Mittal et al., 1985), or case-based reasoning (Navin- nodes and arcs within a graph. The rules create a formal lan-
chandra et al., 1991; Bhatta et al., 1994; Qian & Gero, 1996). guage for generating and updating complex designs from an
Among these, one of the most historically significant is the initial graph-based specification. The development of the
expert system formulation described in the PRIDE system rules encapsulate a set of valid operations that can occur in
established by Mittal et al. (1985), which was specifically de- the development of a design. Through the application of
veloped for creating paper roller systems. A subset of expert each grammar rule the design is transformed into a new state,
systems, case-based reasoning techniques apply past knowl- incrementally evolving toward a desired solution. The rules
edge stored in a computational database toward solving prob- are established prior to the design process and capture a cer-
lems in similar contexts. Examples include Qian and Gero tain type of design knowledge that is inherent to the problem.
(1995), who presented a system called FBS, which uses rela- The knowledge captured in the rules offers the option of
tions among function, behavior, and structure to retrieve design exploring different design decisions and thus different design
information to conduct analogy-based design. Similarly, the alternatives.
structure–behavior–function modeling scheme (Bhatta et al., Using this formalism, Sridharan and Campbell (2004) de-
1994) and its computational application KRITIK is a system fined a set of 69 grammar rules that were developed to guide
relying on a design-case memory to conduct computational the design process from an initial functional goal to a detailed
synthesis. function structure. Elsewhere, graph grammars are widely
Apart from expert system formulations, typical examples used in various engineering applications. Agarwal and
of computational synthesis applications start with a set of fun- Cagan’s (1998) coffee maker grammar was one of the first
damental building blocks and some composition rules that examples of using grammars for product design. Their gram-
govern the combination of these building blocks into com- mar described a language that generated a large class of cof-
plete design solutions. Hundal (1990) designed a program fee makers. Shea et al. (1997) presented a parametric shape
for automated conceptual design that associates a database grammar for the design of truss structures that uses recursive
of solutions for each function in a function database. Ward annealing techniques for topology optimization. Other engi-
neering applications include Brown and Cagan (1997), who
1
The black box representation defines the energy, material, and signal presented a lathe grammar, Schmidt and Cagan’s (1995)
flows entering and leaving an artifact. grammar for machine design, Starling and Shea’s (2003)
Automating the conceptual design process 51

grammars for mechanical clocks, and gear trains (Starling & configuration. The customer need analysis and the formula-
Shea, 2005). tion of the initial “black box” steps are performed by the de-
Although these methods are primarily concerned with gen- signer. The computational design synthesis is initiated at the
eration aspects of conceptual design, there are various tech- level of a black box. The first design tool (the function struc-
niques developed to automate the selection of components ture grammar) converts the black box of a device to be de-
for an already generated design configuration. These tech- signed into a set of detailed function structures by using func-
niques include using genetic algorithms, simulated anneal- tional decomposition rules. Based on this functional input,
ing, and integer programming. Weilinga and Schreiber the second design tool (the configuration grammar) synthe-
(1997) classify the component selection problem as category sizes individual or sets of components into a set of conceptual
one within their work, where the set of components as well design configurations. Finally, the third design tool (the tree
their assembly is fixed. Carlson (1996) uses a genetic algo- search algorithm for component selection) instantiates spe-
rithm for component selection given a user-defined system cific components in a design configuration guided by specific
layout, a database of components and a set of design specifi- design constraints and objectives. In the following para-
cations. Carlson et al. (1998) apply a genetic algorithm for graphs, each of the three design tools and their specific search
solving the problem of catalog design. They create an initial algorithms are explained in detail.
set of components types followed by component selection
from the component database. Dallaali and Premaratne
(2004) apply genetic elitism and double string coding to solve 3.1. Research effort I: Function structure grammar
optical component selection problems. A common technique in phrasing the problem as a black box is
In summary, our background research shows that a number useful in engineering design to clarify the goals of the project
of attempts have been made to automate various key elements (Pahl & Beitz, 1996; Otto & Wood, 2001). By removing all
of the design process such as the creation of function struc- unnecessary information, the black box defines only the flows
tures, configuration design, and component selection. How- entering and leaving the product. The black box is often la-
ever, most of these methods have been developed for specific beled with a primary function, which is typically a verb–
applications. These methods have also been restricted to a noun pair. The first automated design tool (Sridharan &
specific phase or task of conceptual design. The method pre- Campbell, 2004) acts on this black box input to automatically
sented here is a generalized technique that follows the gram- create the necessary functions for translating the input flows
mar formalism and integrates it with fundamentals of the into the output flows. To accomplish this, a series of 52 rules
function-based synthesis paradigm to automate the design de- have been created based on the data of 30 black boxes and their
cision making that governs the entire concept generation pro- corresponding function structures. These rules are created to
cess starting at a black box level product specification and capture common function chains used in a variety of artifacts.
finalized by the selection of components that physically For example, an automobile jack performs a function that re-
embody the design. The details of this design automation quires energy from the human; but instead of being applied di-
approach are explained next. rectly, it is first converted to pneumatic energy and then con-
verted to mechanical energy. This series of functions is also
seen in a toy NERF gun among other applications. This infer-
3. RESEARCH APPROACH
ence of common modules can help make better function struc-
This research aims to automate the systematic design process tures, as it is not always easy for a designer to make a connec-
(presented in Pahl and Beitz, 1996). Accordingly, it extends tion between such different products. To ensure consistency
the previous automated design research by developing a suite for the representation of the function structures and the conse-
of computational design tools that transform a high-level, quent rule development, we have adopted the functional basis
functional description of a nonexistent product into a set of taxonomy (Hirtz et al., 2002). The functional basis is a set of
embodied concept variants by following the process shown function and flow terms that combine to form a subfunction
in Figure 1. description (in verb–object format) and that are utilized during
By automating this process, a design is changed from an the generation of a black box and functional model to encap-
abstract set of customer needs to an embodied conceptual sulate the actual or desired functionality of a product.

Fig. 1. Three nested search algorithms automate the conceptual design process.
52 T. Kurtoglu et al.

Starting with the black box, the grammar rules enumerate tive centers are the points where grammar rules can be applied
all possible valid function structures. This can be viewed as and where new functions and/or flows are added if certain cri-
a search tree, where the black box is the seed or start node, teria are met at a specific open connection. In recreating these
and the rules provide the transition operators that lead to the rules (Fig. 2b), the concept is maintained but no longer re-
many leaves of the tree. It is interesting to note that the black quired as GraphSynth includes a more general and powerful
box is a detailed start state for the tree, and the existence (or subgraph recognition procedure than using active centers.
absence) of input and output flows limits the size of the search In either case, the rule provides guidance in developing the
tree significantly. This is shown clearly in the results of this connecting flows to a “remove” function. This rule captures
paper. In the future, the rules may be revisited so as to create the principle that whenever we cut or grind a solid, we need
or modify specified input and output flows of the black box, to supply some mechanical energy, and this results in two
but it is unclear how these will be regulated. or more pieces of the solid. It should also be noted that this
The original function structure grammar (Sridharan & rule cannot be applied again because the active center neces-
Campbell, 2004) was modeled in the same way grammar sary for rule recognition has been eliminated. Care must be
rules are modeled, as independent if–then statements. This taken to define rules that prevent multiple applications of
is often shown graphically with left-hand and right-hand the same rule because this would cause an endless loop where
sides (Fig. 2a). In terms of implementation, this work was no functionality is being added, only duplicates of existing
done ad hoc, resulting in a large and unorganized set of functionality.
java files. Recently, these rules have been rewritten in a The rule shown in Figure 3a captures another common
new graphical environment known as GraphSynth (Camp- principle. When mechanical energy is supplied, the effort
bell, 2007) that allows one to graphically create the rules of the energy, torque, can be amplified and this is represented
and manages the resulting data as a series of portable xml files by the function “Change ME.” This function can be satisfied
(see Fig. 2b). Note that in Figure 2a, the gray and black cir- by using a gear train. This rule looks for a flow of type me-
cles, which are referred to as “active centers,” serve as mark- chanical energy that is open at the tail and is pointing to the
ers to ease the implementation. This concept of active centers function, “Remove Solid.” If applied, this rule adds the func-
can be found in nature in the chemical polymerization pro- tion “Change ME” to the tail and adds another flow open at its
cess. In addition to polymerization, unsaturated molecules tail to the back of the “Change ME” function. Figure 3b
add onto a growing polymer chain one at a time. The last mo- shows a rule where an electric energy flow that is pointing
lecule on the chain is the active center upon which the unsat- from “Import” is recognized and the functions “Transmit
urated molecules can attach. Similarly, during the creation of EE” and “Actuate EE” are added to it. This rule is observed
a function structure, there are many active centers where in- in many products that use electrical energy, as electrical
coming flows and functions can attach themselves. These ac- energy is always transmitted and actuated before being

Fig. 2. An example rule shows the use and propagation of active centers (a) as shown in the original publication (Sridharan & Campbell,
2005) and (b) as recreated in GraphSynth.
Automating the conceptual design process 53

Fig. 3. Three additional examples of function structure grammar rules (Sridharan & Campbell, 2004).

converted to the required form. The final rule shown in research efforts, the secondary level of the functional basis
Figure 3c is a termination rule where active centers are de- taxonomy (Hirtz et al., 2002) is used as the representational
stroyed. Whenever two flows are recognized such that we language for expressing this input function structure. The
have an open electrical energy flow and energy of any other synthesis process is aimed to perform a graph transformation
kind (represented as XE) that needs to be supplied, we con- of a function structure into a set of configuration-based graphs
vert the electrical energy to the required form and transmit called the Configuration Flow Graphs (Kurtoglu et al., 2005).
it. Termination rules are vital in obtaining a valid function In a configuration flow graph (CFG), nodes represent design
structure. A valid function structure ensures that all of the flows components and arcs represent energy, material, or signal
that go into the system are utilized and then exported in some flows between them. The graph is also similar to an exploded
form. If the termination rules were not called, then energy, view in that components are shown connected to one another
material, or signal chains could dead end and it would be im- through arcs or assembly paths. Using a CFG, designers can
possible to continue with the other steps of the design process. capture components that are present in a design, their connec-
Figure 3 is based on the work of Sridharan and Campbell tivity, and physical interfaces between a design’s compo-
(2004). nents. Figure 4 shows an example of a CFG for a disposable
camera.
The grammar rules for the configuration design are de-
3.2. Research effort II: Configuration design
fined through a knowledge acquisition process that is based
grammar
on the dissection of existing electromechanical devices.
For decisions at the conceptual phase of design, the intercon- Accordingly, for each device that is dissected, a function
nectivity of design elements is more important than parametric structure and a configuration flow graph are generated. Then,
details (Kurtoglu, 2007). In such conceptual design problems, the mapping between the two graphs is captured where
it becomes essential to determine an optimal configuration of each mapping represents a potential grammar rule (Kurtoglu
components prior to tuning individual component parameters. & Campbell, 2005). Some of the rules derived from this anal-
Creating such configurations is the objective of the second ysis are shown in Figure 5. Note that the rules in Figure 5 are
design tool: the configuration design grammar. not simply one-to-one matches of functions to components.
The starting point for the configuration design grammar is The open-endedness of the grammar formulation allows
a function structure. To maintain consistency between the two us to tend to assign single components to single functions.
54 T. Kurtoglu et al.

Fig. 4. The CFG of a disposable camera.

Instead, through multiple node recognition and application different concepts with potentially varying degrees of com-
the grammar provides a more generic approach capable of in- plexity as candidate configurations to the same functional
serting multiple components for a single function, a single specifications.
component for multiple functions, as is the case in function In detail, the transformation from the function structure to a
sharing, or multiple components for multiple functions. CFG is part of a recognize–choose–apply cycle. The recog-
In reality, each rule represents a design decision that shows nize step identifies all possible locations in the function struc-
how a functional requirement was transformed into an em- ture where a grammar rule can be applied. These locations
bodied solution in an actual design. Currently, the rule data- define a set of possible graph transformations that can be ap-
base contains 170 grammar rules derived from the dissection plied at that design stage. This step is followed by choosing
of 23 electromechanical products. These rules and additional one of the valid grammar rule options. In the final apply
information about these products are stored in a Web-based step, the CFG is updated per the instructions provided by
Design Repository (Design Engineering Lab, Missouri the selected rule. This process is repeated until there are no
University of Science and Technology; http://function.basic more rules that can be applied.
eng.umr.edu/repository) that is managed at the Missouri The final configurations obtained at the end of this genera-
University of Science and Technology. The grammar pro- tion process depend on the selection of the rules applied. To
vides an effective method for automatically generating design fully automate the generation process, this selection is made
configurations through a search-based execution of rules. by the computer. The basis and the guidelines to select the
The computational synthesis approach performs a graph rules are embedded in the search algorithm. In the current im-
transformation of the initial function structure of the to-be- plementation, each applicable grammar rule is systematically
designed product into a set of configuration flow graphs. selected by the computer as the configuration space is tra-
Each execution of a rule adds more components to the versed using a breadth-first search (BFS) approach.
design configuration, which incrementally builds to a final At the end, the search process generates a variety of config-
concept. At the end, the computational search process returns urations that are developed from a functional description of
Automating the conceptual design process 55

Fig. 5. Two grammar rules of the configuration grammar in GraphSynth. The left-hand side of the rules captures the functional requirement
and the right-hand side depicts how the functional requirement can be addressed by the use of specific component concepts.

a product by synthesizing component solutions together that that particular component. Each node further down the tree
have been successfully used in the design of past products. replaces a generalized CFG component (e.g., gear) with an
artifact (e.g., steel plain bore 14.58 pressure angle spur gear
with 24 teeth, a pitch of 32, and a face width of 3/16 in.).
3.3. Research effort III: Tree search algorithm
The branching factor2 thus depends on the number of choices
for component selection
for a component, whereas the number of levels in the tree is
The objective of the third design tool is to determine the op- the number of selections to be made as determined by the
timal choice of components for a specified CFG (Tamhankar CFG. Currently there are, on average, six artifacts per compo-
& Campbell 2007). To accomplish this, the components are nent (there is only one electrical cord but 20 gears). As the
chosen from a database, which contains a compilation of search process unfolds, more components are instantiated
real component information for each component abstraction by replacing abstract components in the CFG with actual ar-
(such as electric motor, bearing, shaft, and gear) that can be tifacts from the database.
present in a CFG. This data for each component, or artifact, At this stage in the engineering design process it is possible
has been collected from an online repository created and to include some evaluation of the design decisions. In this re-
maintained by the Missouri University of Science and Tech- search, an objective function was constructed that combined
nology (2007) as well as online product catalogs (2007). various customer needs (such as minimize cost and maxi-
The approach is an iterative process that replaces each com- mize power) with compatibility metrics for adjacent compo-
ponent in the CFG with an artifact that is stored in the data- nents (e.g., how different is the shaft diameter from the
base. Each component in a CFG represents a different level
of the search tree, and the number of options at each level 2
In computing and tree data structures the branching factor is the average
is equivalent to the number of artifacts in the database for number of children from each node in the tree.
56 T. Kurtoglu et al.

mating gear’s bore diameter). These customer need satisfac- the machine would use mechanical energy to perform the
tion and compatibility equations play a key role in the final “change” function and that the human material would be re-
component selections. There is an element of subjectivity turned. The specification of input and output flows poses con-
and dependence on the designer for the assignment of straints to the design problem and keeps the artifact choices in
weights and penalty that cannot be eliminated. certain domains. By specifying electrical energy as input and
The space of solutions found in the tree is searched using a mechanical energy as output, we limit the functions that can
uniform cost search algorithm. The search begins at the root be called and the consequent variety of components that can
node (a complete CFG) and the traversal of the tree is em- be selected. The specification of the primary function and the
ployed by instantiating one component at a time. During input and output flows ensures that the customer needs are
the search, it is possible to evaluate the design decisions gov- captured before the design process starts and that the comput-
erning the instantiation of components. Accordingly, an ob- er will not end up with solutions that the user did not intend or
jective function is constructed that combines criteria measur- is not interested in.
ing how well various customer needs (such as minimize cost The complete search process is run using the GraphSynth
and maximize power) are met with compatibility metrics for environment (Campbell, 2007). The process starts with the
neighboring components (e.g., how different is the shaft diam- user drawing the black box of the to-be-designed artifact.
eter from the mating gear’s bore diameter). At each node of The function structures and CFGs are then created automati-
the tree, the node expansion is performed after calculating cally using their respective sets of grammar rules. The results
transition costs based on this objective function formulation. for the selection of the components are yet to be implemented
These transition costs are additive in nature, and at each step for the presented design example as we continue working on
only the child with the minimum transition cost is generated. formulating an objective function to evaluate the perfor-
This search process continues until the CFG is fully instanti- mance of the coffee grinder.
ated and an optimum solution is reached.

5. RESULTS AND DISCUSSION

4. CASE STUDY: DESIGN OF A COFFEE
For the implementation of this case study, the algorithms are run
GRINDER
on a Windows PC with 3.25 GB of RAM and a 2.2-GHz pro-
The proposed methodology is demonstrated in this section by cessor. After the input is specified as shown in Figure 6, the
solving a test problem: the design of a coffee grinder. In this computational synthesis of the coffee grinder begins, initially
problem, we start off by creating a simple black box and with the creation of potential function structures for the design.
illustrate how the computer can generate associated function The function structure grammar makes use of three rule
structures and configuration flow graphs. sets. The first rule set encompasses initiation rules and inserts
Figure 6 shows a black box, the primary function of which active centers to the graph for each of the input and output
is “Change Solid.” The primary function, along with the input flows as described in Sridharan and Campbell (2004). The
and output flows, of the black box are determined by the de- second rule set of the grammar, called the propagation rule
signer. These decisions govern the energy domains, materi- set, generates all of the functions in the functional model.
als, and signals, which the product would need to use. For This model is built utilizing two directions starting from
the selected problem, it is envisioned that the designed arti- both the input and the output (left and right as seen in
fact would primarily utilize human and electrical energy for Fig. 6) flows. After the propagation rules are executed some
the input energies and that a user would actuate the device functions may not be fully connected. The final termination
operation. These design decisions are captured by the specif- rule set ensures that these connections are not left dangling
ication of human material and human and electrical energy and that the generated function structure is complete.
input flows as shown in Figure 6. Similar decisions are For the coffee grinder problem, the function structure gram-
made for the output flows. Accordingly, it is specified that mar ran for 4 h and created two unique function structures.

Fig. 6. The black box for a coffee grinder. Note the flows and primary function are simplified so that it is understandable to the computer.
Automating the conceptual design process 57

Fig. 7. The second candidate function structure created automatically from the black box shown in Figure 6.

These are shown in Figure 7 and Figure 8. The generated the simple example given above where the computer cannot
function structures are similar in nature and only differ in their tell that there are three identical copies of the design in the
use of mechanical energy. One of them uses translational me- search tree until later in the process. As the plot shows, the
chanical energy to perform the grinding operation while the number of unique candidates needs to reach a critical point
other one uses rotational mechanical energy. after which the filtering takes effect. This filtering cuts the
These two unique function structures are then posed as in- amount of time that the process took by a factor of 20 and
puts into the second design tool, which takes the function greatly reduces the number of candidates at each level.
structure as a starting point and generates conceptual config- The fourth set of grammar rules (the configuration gram-
urations. Specifically, the computer employs a modified BFS mar) ran for about 5 days and generated 1536 unique solu-
algorithm that includes a filtering mechanism that removes tions. This many unique solutions is still an overwhelming
duplications of previously visited nodes from the search number for a human designer to consider, but the natural
space. This filtering is based on a property in graph grammar reduction from over 50,000 is interesting. Two of the unique
theory known as confluence, which means that the order in solutions are shown in Figure 10 and Figure 11. The CFG in
which a subset of rules is invoked does not affect the final out- Figure 10 was derived from the function structure in Figure 7
come. To better illustrate this concept, let us consider three and the CFG in Figure 11 was derived from the function struc-
rule choices, A, B, and C, that must be called to complete a ture in Figure 8. These CFGs are different from each other in
candidate design. After making the first choice it appears as many ways, but strictly speaking, for a CFG to be unique,
if one may have three unique solutions, A__, B__, and only one component needs to be different. For example, if
C__. After the second choice, the number of possible unique the “transfer mechanical energy” function is accomplished
solutions stays at three, AB__, AC__, BC__, but these solu- by a shaft in one CFG and by a conveyer in another, those
tions are interpreted by the computer using six possible bran- two CFGs are considered to be unique. In the two CFGs
ches because the computer cannot differentiate between the shown here, there are many differences. The first of the
designs AB and BA. Finally, it is not until the last iteration CFGs takes the mechanical energy from the motor and goes
that one can conclude that there is only one unique solution. straight to the shaft to the blade. In the second CFG, this me-
This is shown more clearly in the bar graph of Figure 9, where chanical energy goes from the motor to a gripper then to a
the number of candidates is shown on the y-axis and the num- support, a sprocket, and then the blade. This second chain
ber of rules called is depicted on the x-axis. As the rules are of energy may spark an interesting idea by the designer on
invoked the number of candidates quickly expands only to a way to get rid of the costly shaft, or it may be deemed too
be followed by a decrease. The reason for this is similar to complicated for this application and disregarded. The purpose
58 T. Kurtoglu et al.

Fig. 8. The first candidate function structure created automatically from the black box shown in Figure 6.

of this part of the research is to present the designer with several of possible instantiations is 613 or 13 billion. Taken with the
different ways to solve the problem. For each possible function previous search trees, the total number of embodied config-
the grammar rules are attempting to capture all of the knowl- urations that result from the single black box is estimated at
edge about this function and which parts can perform this 20 trillion [1536 CFGs  (613 instantiations/CFG)  20 
function. 1012 ]. This underscores the need for an approach to eliminate
The next step in the automated process is to invoke the many of these branches to focus in on more beneficial
automated component selection discussed above as research solutions. Research effort III accomplished this in the prior
effort III (see Section 3.3). Current efforts are in place to implementation through the uniform cost search and specific
accomplish this, but a detailed evaluation of component evaluation functions and functions to penalize systems
choices is not completed at this time. Fortunately, we are with incompatible combinations of components. However,
able to approximate the size of the search tree. The average to implement this with the current coffee grinder design prob-
number of components to instantiate is 13 (note that Fig. 10 lem would require us to define functions to evaluate the per-
has 13 components, whereas Fig. 11 has 14). This determines formance of coffee grinders (cost, noise, uniformity of grain
the depth of the tree. The breadth of the tree is 6 (the average size, etc.) as well as new compatibility functions for the set of
number of instantiated components). As a result, the number the original equipment manufacturer components relevant to

Fig. 9. The bar graph showing the number of candidates at each level.
Automating the conceptual design process 59

Fig. 10. The first candidate CFG created automatically from the function structure shown in Figure 8.

Fig. 11. The second candidate CFG created automatically from the function structure shown in Figure 9.
60 T. Kurtoglu et al.

coffee grinders. This may not be possible without significant work lacked an evaluation mechanism, it currently outputs
and groundbreaking development of simulation environ- entirely too many solutions for the user to consider. However,
ments for predicting noise and grain size uniformity. combined with a meaningful evaluation, this large number
could be reduced to the two or three solutions most useful
to the user.
6. DISCUSSION
Without the evaluation, 20 trillion possible candidate solu-
In this paper, three previous research projects are combined tions were found. This large space is not captured explicitly,
to automate the conceptual design process from a black but rather implicitly in the grammar rules (170 rules total) and
box to a configuration of specific components. This research the database of real components (300 in all). The grammar af-
exercise was accomplished to explore the possibility of using fords a representation of the design space as a tree of solutions
the computer to solve a conceptual design problem through built from an initial specification. Each transition in the tree
the manipulation of standard graph representations. It is pos- corresponds to an application of a rule, thus incrementally
sible that this work could lead to an automated design tool building a final design that is represented as one of the leaves
that would present the user with possible connections of com- of the tree. This process is illustrated in Figure 12. As is evi-
ponents to solve a particular design problem. Because this dent from the tree, the result of rule applications generates a

Fig. 12. An illustration of the cascading search trees.

Automating the conceptual design process 61

design space that requires navigation techniques to enable a mar rules, but it is likely that humans collectively know many
search for a desired or optimal solution. The issue of imple- more, and each likely contains many caveats, exceptions, and
mentation of the grammar then becomes one of controlled useful minutiae. Yet, if all the heuristics about this electrome-
searches through this space of solutions. chanical design domain were captured, the number of alterna-
The visualization in Figure 12 shows cascading search tives at each stage would be even larger. What is required is
trees as created by this knowledge base. At each node in more information to be stored within the rules: information
the process, a recognition process first retrieves the valid op- that defines detailed conditions for when the rule can be
tions. The illustration only shows 2 or 3 options per node, but applied. Capturing these rules is time-consuming and auto-
realistically this varies throughout this design process from 1 mated approaches may be possible, but inevitably an experi-
to 33. Some have claimed that design is essentially a decision- ence human designer must define the conditions under which
making process, and this is captured by this illustration. With a rule is valid. This is a challenge to the scalability of the ap-
each decision, each commitment to follow a particular branch proach, and it would be crucial to extend the system to a wider
of the tree, the design process diverges. One can undo deci- community of users to accomplish a large and useful set of
sions to follow other branches or maintain a small diverse rules for more ambitious design problems.
set of candidate concepts scattered about the search tree. Third, the design process is vague. Approaches to system-
Computationally, no single decision is made in the current ize it like design tools such as creating a black box or a func-
research; rather, we have taken advantage of a large computa- tion structure clarify the design process and make it more sci-
tional memory to follow every path in the search tree to sim- entific. Our work has attempted to transition these design
ply enumerate all possible candidates. This rote approach tools into an even more rigorous language. The results are
(BFS) is complete but unmanageable for the final tree in promising as the computer is capable of creating a configura-
which components are instantiated. Fortunately, this search tion of real components, a coffee grinder in this paper, into a
tree is the first opportunity for us to numerically evaluate real set of connected components. With more rules and a thor-
the quality of each decision, because real components are ough evaluation of concepts, it now seems possible that the
being compared and such components have data available conceptual design process can be solved wholly computa-
about their cost, weight, performance, and so forth. The tionally.
lack of evaluation limits the computer’s ability to decide
which option at each stage of the tree is better. Experienced
designers can make judgments about which paths to follow ACKNOWLEDGMENTS
in these early stages based on intuition or previous experi- This material is based on work supported by the National Science
ences. To take advantage of this, the designer’s feedback Foundation under Grants CMMI-0448806 and IIS-0307665. Any
can be incorporated into the computational system using an opinions, findings, and conclusions or recommendations presented
interactive approach where the computational system presents in this paper are those of the authors and do not necessarily reflect
solutions to the designer for evaluation. The designer feed- the views of the National Science Foundation.
back can then be used to assess the value of generated designs
and to prune inferior designs from the solution space (Kurto-
glu & Campbell, 2009). REFERENCES
Furthermore, the early search trees are complicated by con- Agarwal, M., & Cagan, J. (1998). A blend of different tastes: the language of
fluence in the rules. Confluence clearly happens and essen- coffee makers. Environment and Planning B: Planning and Design
tially reduces the search tree making it easy to manage, but 25(2), 205–226.
Bhatta, S., Goel, A., & Prabhakar, S. (1994). Innovation in analogical design:
it is not clear by examining the rules a priori how much con- a model-based approach. Proc. AI in Design. Dordrecht: Kluwer Aca-
fluence exists or how to manage it. The tree search algorithms demic, pp. 57–74.
used in this paper include a check for common configurations Bracewell, R.H., & Sharpe, J.E.E. (1996). Functional description used in
computer support for qualitative scheme generation—Schemebuilder. Arti-
at each level of the tree to reduce the memory burden; how- ficial Intelligence for Engineering Design, Analysis and Manufacturing
ever, this check is time consuming and accounts for 80% of 10(4), 333–345.
the nearly 5-day span of time required to reach the 1536 can- Brown, K.N., & Cagan, J. (1997). Optimized process planning by generative
simulated annealing. Artificial Intelligence in Engineering Design, Anal-
didates at the bottom of the fourth search tree. ysis and Manufacturing 11, 219–235.
The results of this study provide some interesting insight Bryant, C., Stone, R., McAdams, D., Kurtoglu, T., & Campbell, M. (2005).
when compared to the human activities in accomplishing A computational technique for concept generation. ASME 2005 Int.
Design Engineering Technical Conf., Long Beach, CA.
the design process. First, the stages of the design process Campbell, M., Cagan, J., & Kotovsky, K. (2000). Agent-based synthesis of
can help to reduce the search space by committing to a best electro-mechanical design configurations. Journal of Mechanical Design
candidate at each level and using that as the seed for the 122(1), 61–69.
Campbell, M.I. (2007). The official GraphSynth Site, University of Texas
next. These key decision points provide a moment of evalu- at Austin. Accessed at http://www.graphsynth.com
ation and limit the number of solutions needed to be searched Carlson, S.E. (1996). Genetic algorithm attributes for component selection.
in the future. Second, human designers are capable of com- Research in Engineering Design 8(1), 33–51.
Carlson-Skalak, S., White, M.D., & Teng, Y. (1998). Using an evolutionary
paring only a small number of concepts. The current imple- algorithm for catalog design. Research in Engineering Design 10(2),
mentation contains many heuristics as stored in the 170 gram- 63–83.
62 T. Kurtoglu et al.
Chakrabarti, A., & Bligh, T. (1996). An approach to functional synthesis of Szykman, S., Racz, J., & Sriram, R. (1999). The representation of function in
mechanical design concepts: theory, applications and emerging research computer-based design. Proc. Design Engineering Technical Conf.,
issues. Artificial Intelligence for Engineering Design, Analysis and DETC99/DTM-8742, Las Vegas, NV.
Manufacturing 10, 313–331. Tamhankar, M.S., & Campbell, M.I. (2007). An intelligent and efficient tree
Dallaali, M.A., & Premaratne, M. (2004). Double-constrained optimization search algorithm for computer-aided component selection. ASME Design
of optical component selection problem using genetic elitism and double Engineering Technical Conf. Computers in Engineering, Las Vegas, NV.
string coding. Int. Conf. Numerical Simulation of Ootoelectronic Devices Ullman, D. (1995). The Mechanical Design Process. New York: McGraw–
’04. Hill.
Hirtz, J., Stone, R., McAdams, D., Szykman, S., & Wood, K. (2002). A func- Ulrich, K., & Eppinger, S. (1995). Product Design and Development.
tional basis for engineering design: reconciling and evolving previous New York: McGraw–Hill.
efforts. Research in Engineering Design 13(2), 65–82. Umeda, Y., & Tomiyama, T. (1997). Functional reasoning in design. IEEE
Hundal, M. (1990). A systematic method for developing function structures, Expert March–April, 42–48.
solutions and concept variants. Mechanism and Machine Theory 25(3), Wang, K., & Yan, J. (2002). An analytical approach to functional design.
243–256. Proc. ASME Design Engineering Technical Conf., Montreal, CA.
Kirschman, C., & Fadel, G. (1998). Classifying functions for mechanical Ward, A.C., & Seering, W.P. (1989). The performance of a mechanical
design. Journal of Mechanical Design, Transactions of the ASME design compiler. ASME, Design Engineering 17, 89–97.
120(3), 475–482. Wielinga, B.J., & Schreiber, G. (1997). Configuration design problem solv-
Kitamura, Y., & Mizoguchi, R. (1999). Metafunctions of artifacts. Proc. 13th ing. Technical Report, University of Amsterdam, Department of Social
Int. Workshop on Qualitative Reasoning (QR-99), Loch Awe, Scotland, Science Informatics.
pp. 136–145. Williams, B.C. (1990). Interaction-based invention: designing novel devices
Kota, S., & Chiou, S.-J. (1992). Conceptual design of mechanisms based from first principles. AAAI-90 Proc., 8th National Conf. Artificial Intelli-
on computational synthesis and simulation of kinematic building blocks. gence, Vol. 1, pp. 349–356.
Research in Engineering Design, 4, 75–87.
Kurtoglu, T. (2007). An computational approach to innovative conceptual
design. PhD dissertation. Austin, TX: University of Texas at Austin Press.
Kurtoglu, T., & Campbell, M. (2009). An evaluation scheme for assessing Tolga Kurtoglu is a Research Scientist with MCT at the In-
the worth of automatically generated design alternatives. Journal of Re-
search in Engineering Design 20, 59–76. telligent Systems Division of the NASA Ames Research Cen-
Kurtoglu, T., Campbell, M., Gonzales, J., Bryant, C., Stone, R., & ter. He received his PhD in mechanical engineering from the
McAdams, D. (2005). Capturing empirically derived design knowledge University of Texas at Austin in 2007. Dr. Kurtoglu has pub-
for creating conceptual design configurations. ASME 2005 Int. Design
Engineering Technical Conf., Long Beach, CA. lished over 30 articles and papers in various journals and con-
Mittal, S., Dym, C., & Morjara, M. (1985). “PRIDE: an expert system for the ferences and is an active member of ASME, AIAA, AAAI,
design of paper handling systems. IEEE Computer 19(7), 102–114. ASEE, and the Design Society. His research focuses on the
Navinchandra, D., Sycara, K.P., & Narasimhan, S. (1991). A transforma-
tional approach to case-based synthesis. Artificial Intelligence for Engi- development of prognostic and health management technolo-
neering Design, Analysis and Manufacturing 5(1), 31–45. gies for complex systems, model-based diagnosis, computa-
Otto, K., & Wood, K. (1997). Conceptual and configuration design of pro- tional design tools and optimization, automated reasoning,
ducts and assemblies. ASM Handbook, Materials Selection and Design
(Vol. 20). Materials Park, OH: ASM International. conceptual design theory, artificial intelligence in design,
Otto, K., & Wood, K. (2001). Product Design: Techniques in Reverse Engi- and risk and reliability-based design.
neering and New Product Development. Englewood Cliffs, NJ: Prentice–
Hall. Albert Swantner is a Design Engineer at Flextronics Medi-
Pahl, G., & Beitz, W. (1996). Engineering Design: A Systematic Approach. cal–Avail in Irving, Texas. He graduated from the University
Berlin: Springer–Verlag.
Qian, L., & Gero, J.S. (1996). Function–behavior–structure paths and their of Texas at Austin with a BS in mechanical engineering in
role in analogy-based design. Artificial Intelligence for Engineering 2007 and an MS in mechanical engineering in May of 2009.
Design, Analysis and Manufacturing 10, 289–312. His thesis focused on automating and optimizing gear train
Schmidt, L., & Cagan, J. (1995). Recursive annealing: a computational model
for machine design. Research in Engineering Design 7(2), 102–125. topologies.
Shea, K., Cagan, J., & Fenves, S.J. (1997). A shape annealing approach to
optimal truss design with dynamic grouping of members. ASME Journal Matthew Campbell is an Associate Professor in the Mechan-
of Mechanical Design 119(3), 388–394. ical Engineering Department at the University of Texas at Aus-
Sridharan, P., & Campbell, M.I. (2004). A grammar for function structures. tin. He received his PhD in mechanical engineering from Car-
Proc. ASME DETC, Salt Lake, UT.
Starling, A.C., & Shea, K. (2003). A grammatical approach to computational negie Mellon University in 2000. Dr. Campbell has been
generation of mechanical clock designs. Proc. ICED ’03 Int. Conf. Engi- acknowledged with best paper awards at conferences sponsored
neering Design, Stockholm, Sweden. by the ASME, the ASEE, and the Design Society. His research
Starling, A.C., & Shea, K. (2005). Virtual synthesizers for mechanical gear
systems. Proc. ICED’05 Int. Conf. Engineering Design, Melbourne, Australia. focuses on computational methods that aid the engineering de-
Stone, R., & Wood, K. (1999). Development of a functional basis for design. signer earlier in the design process than traditional optimization
Proc. Design Engineering Technical Conf., Paper No. DETC99/DTM- would. To date, he has been awarded $1.57 million in research
8765, Las Vegas, NV.
Subramanian, D., & Wang, C.-S. (1995). Kinematic synthesis with config- funding, including the CAREER award for research into a ge-
uration spaces. Research in Engineering Design 7(3), 193–213. neric graph topology optimization method.
Suh, N. (1990). The Principles of Design. New York: Oxford University Press.