Advanced Engineering Informatics 38 (2018) 712–724

Contents lists available at ScienceDirect

Advanced Engineering Informatics

journal homepage: www.elsevier.com/locate/aei

Full length article

A generative design technique for exploring shape variations T

Shahroz Khan , Muhammad Junaid Awan
⁎

Brainlabs Software, Taxila, Pakistan

ARTICLE INFO ABSTRACT

Keywords: Because innovative and creative design is essential to a successful product, this work brings the benefits of
Generative design generative design in the conceptual phase of the product development process so that designers/engineers can
Computer-aided design effectively explore and create ingenious designs and make better design decisions. We proposed a state-of-the-art
Parametric design generative design technique (GDT), called Space-filling-GDT (Sf-GDT), for the creation of innovative designs. The
Space-filling design
proposed Sf-GDT has the ability to create variant optimal design alternatives for a given computer-aided design
Jaya algorithm
(CAD) model. An effective GDT should generate design alternatives that cover the entire design space. Toward
that end, the criterion of space-filling is utilized, which uniformly distribute designs in the design space thereby
giving a designer a better understanding of possible design options. To avoid creating similar designs, a
weighted-grid-search approach is developed and integrated into the Sf-GDT. One of the core contributions of this
work lies in the ability of Sf-GDT to explore hybrid design spaces consisting of both continuous and discrete
parameters either with or without geometric constraints. A parameter-free optimization technique, called Jaya
algorithm, is integrated into the Sf-GDT to generate optimal designs. Three different design parameterization and
space formulation strategies; explicit, interactive, and autonomous, are proposed to set up a promising search
region(s) for optimization. Two user interfaces; a web-based and a Windows-based, are also developed to utilize
Sf-GDT with the existing CAD software having parametric design abilities. Based on the experiments in this
study, Sf-GDT can generate creative design alternatives for a given model and outperforms existing state-of-the-
art techniques.

1. Introduction preliminary analysis to select a potential design. Computer-aided design

(CAD) is rarely used during this phase; it is primarily utilized later to
Engineering or industrial product design is a complex process in analyze, validate, and fabricate the design [4]. For the most part, design
which a design arrives at its final form after passing through a series of engineers convert a design selected at the conceptual phase into a CAD
design phases. The conceptual phase is an initial and important com- model when they explore a narrow design space in order to analyze the
ponent of these phases; it is recognized as a foundational step in any performance of the design.
product development process. This phase can be complex and time- With recent advancements in artificial intelligence, optimization,
consuming if the appearance of the product under consideration is design simulation, and parametric design techniques, the role of com-
valuable to its target customers. To select an appealing design, de- puters in the field of design is changing. In comparison to traditional
signers often develop a number of design alternatives using two-di- CAD modeling techniques, these new techniques allow designers and
mensional (2D) sketches. However, the formulation of these alter- engineers to iterate through a large number of design alternatives [5].
natives is a critical and time-consuming task, especially for novice Generative design systems use these techniques to provide a promising
designers. To create these alternatives, designers have to develop and way to explore design space to create alternative designs based on the
explore the entire design space effectively within a product’s design specific performance objective defined by a user. A typical generative
requirements or the customer’s preferences. design system takes a problem definition as input and produces a single
Exploration of design alternatives is recognized as a major char- or set of optimal solutions for a given problem. Commercially available
acteristic of the conceptual design phase [1,2]. Pahl et al. [3] cate- generative design systems, such as Altair’s OptiStruct, solidThinking’s
gorized the conceptual phase into two sub-phases. In the first sub- Inspire, Siemen’s Frustum, and efiForm [6], etc., are based on the to-
phase, design alternatives are formalized based on the design require- pology optimization techniques [7]. These techniques are the mathe-
ments. In the second phase, these alternatives are ranked based on a matical methods that optimize a layout of material distribution within a

⁎
Corresponding author.
E-mail address: [email protected] (S. Khan).

https://doi.org/10.1016/j.aei.2018.10.005
Received 2 March 2018; Received in revised form 29 September 2018; Accepted 12 October 2018
Available online 22 October 2018
1474-0346/ © 2018 Elsevier Ltd. All rights reserved.
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

predefined design space. Mostly, in these systems, the objective is to

maximize/minimize compliance, the temperature at a certain point or
globally, or minimize weight under volume, stress or displacement
constraints. Typically, a generative system involves three steps to set up
a problem. First, design engineer transforms 2D sketches into a three-
dimensional (3D) CAD model. Then, various constraints and properties
are defined based on the design specifications. Later, the design en-
gineer executes generations to obtain a single or multiple optimized
solutions. Different researchers have developed GDTs to create archi-
tectural structures [8], site layouts [9], and energy efficient [10] and Fig. 1. Illustration of the outcomes of Sf-GDT.
eco-friendly building designs [11]. However, few studies [12,4] have
investigated how to create generative systems to explore designs based
on their external form appearance. rules [15]. Fig. 1 illustrates the outcome of the proposed technique.
Therefore, it is beneficial to develop a system that can automatically Following major contributions are made to Sf-GDT to enhance its
generate a variety of unique design alternatives for the outer form of a ability for optimal creation of designs.
product based on its design requirements. The prime objective of this
study is to develop a GDT that can effectively explore a design space 1. The search strategy of Jaya algorithm is extended to generate N
and generate optimal aesthetically convincing design alternatives for a optimal designs.
product at the conceptual phase of the design process. To develop the 2. A weighted-grid-search technique is embedded in Sf-GDT to main-
proposed technique, the following points were considered in order to tain diversity between designs.
make it effective. The proposed technique should: 3. The ability of Sf-GDT is enhanced to explore hybrid design spaces
consisting of continuous and discrete design parameters.
1. Have an effective search and generation strategy to generate op- 4. Different design parameterization and space formulation strategies
timum design alternatives. are proposed for an effective creation of design space.
2. Be able to autonomously set up a viable design space for a given 5. Deb’s [15] heuristic constraint handling rules are applied to gen-
model. erate designs in constrained spaces.
3. Be able to work with both continuous and discrete design para- 6. A web-based and Windows-based user interfaces are developed to
meters. utilize Sf-GDT with existing CAD software.
4. Be able to create uniformly distributed and variant designs from the
entire design space. The remainder of this paper is organized as follows: Section 2 gives
5. Have the ability to effectively explore both constrained and un- a comprehensive review of the relevant literature. Section 3 discusses
constrained design spaces. the proposed approach to generating new designs. The numerical re-
sults of the proposed technique are given in Section 4. Section 5 de-
By considering the points mentioned above, the present study pro- scribes the usage of the proposed technique with existing CAD software.
poses a new generative design technique, Space-filling-GDT (Sf-GDT). Sf- Concluding remarks and opportunities for future work are presented in
GDT has the ability to generate variant optimal designs. However, the Section 6.
decision on the selection of appropriate design parameters and setting a
suitable design space for a given problem is critical. Therefore, Sf-GDT 2. Related works
provides different space formulation strategies for users to obtain op-
timal designs. Generative formulation of designs is a high-dimensional The proposed technique is inspired by the prior research in gen-
constrained optimization problem as there are generally a high number erative and space-filling design techniques and is based on Jaya algo-
of design parameters and geometric constraints. Therefore, there is a rithm. Below, we discuss some previous works done by different re-
need for a simple yet effective optimization approach that can search searchers in these fields.
different optimum designs. Among many well-known optimization
techniques, genetic algorithms have been widely used in generative 2.1. Jaya Algorithm
systems. However, the performance of the genetic algorithms ex-
tensively depends on the selection of tuning parameters [13] and the Most of the well-known meta-heuristic optimization techniques re-
proper tuning of these parameters requires an entirely different set of quire algorithm-specific parameters and proper tuning of these para-
expertise, which most designers do not possess [12]. For this reason, we meters is a critical factor, which affects their performance [13]. For
selected a newly proposed simple, effective, and parameter-free opti- example, the genetic algorithm uses selection operator, mutation and
mization approach called Jaya algorithm [14]. crossover probability; particle swarm optimization uses inertia weight,
To generate N optimum design alternatives, the user first para- cognitive and social parameters; artificial bee colony uses the number
meterizes the given CAD model and defines a viable design space based of onlooker bees, scout bees, and employed bees. The Jaya algorithm
on any of the three different proposed space formulation techniques. does not require tuning of specific parameters except common con-
Within the defined design space, Sf-GDT randomly generates an initial trolling parameters like population size and a number of generations.
population of solutions/designs, which consists of further N sub- This simplicity and tuning free nature of the Jaya algorithm make it
populations, one for each design alternative. Afterward, Sf-GDT applies suitable for generative design systems.
the search strategy of the Jaya algorithm to each subpopulation to The optimization process in Jaya starts by randomly generating a
converge the initial solutions to the optimum position in the design population P of initial solutions for a given size s within the defined
space while minimizing a cost/objective function, which ensures uni- design space. In order to achieve an optimum solution during the
form and diverse design exploration. A weighted-grid-search approach search, the algorithm always tries to move towards the best solution
is proposed, which enables Sf-GDT to maintain diversity between de- and moves away from the worst solution. Suppose for a specific pro-
signs. Sf-GDT has also the ability to explore the design space by syn- blem there are n number of design parameters (i.e. j = 1, 2, …, n ) and s
thesizing the design with different style forms, which can be im- is the number of solutions (i.e. k = 1, 2, …, s ). If the value of the jth
plemented as discrete design parameters. To generate designs from the parameter for the kth solution during the ith iteration is represented as
constraint spaces, Sf-GDT uses Deb’s heuristic constrained handling Xj, k, i , then this value is updated according to Eq. (1).

713
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

X j, k, i = Xj, k, i + r1, k, i (Xj, best , i Xj, k, i ) r2, k, i (Xj, worst , i Xj, k, i ) (1) as complex city planning [33] and computer pattern design [34]. L-
systems are also based on the design rules applied in the form of a
Xj, best , i and Xj, worst , i are the updated values of the parameter j for the best string. Among these methods, shape syntheses are preferable for
and worst solutions, respectively. X j, k, i is the updated value of Xj, k, i , and creating a higher design variation of a given design. However, these
r1, k, i and r2, k, i are the two random numbers in the range [0,1]. At the end techniques can only be employed for creating variations of existing
of each iteration i if X j, k, i is better than Xj, k, i then it is accepted other-
designs/shapes. In which system is first trained on a large dataset of
wise rejected.
existing designs/shapes that are then synthesized to create variations.
Several improvements have also been made on the Jaya algorithm
in order to improve its performance and to expend its application in
different fields. For example, Huang and Wang [16] introduced an elite 2.3. Space-filling design
opposition-based Jaya algorithm called EO-Jaya. EO-Jaya is a swarm
intelligence based algorithm with no specific parameters to tune its There is a considerable amount of research that has been done on the
performance. The elite opposition learning strategy was incorporated optimal selection of space-filling Design of Experiments (DoE). However,
into EO-Jaya’s solution updating phase, which enhances the solution most works done by researchers are proposed for the unconstrained design
diversity. A hybrid parallel Jaya algorithm for a multi-core environ- spaces. The research problem becomes more complicated when a selection
ment called HHCP was developed by Michailidis [17]. HHCP Jaya has a of designs has to be performed in a constrained and high-dimensional
hierarchical cooperation search mechanism to solve large-scale global design space like in the research of this paper. Fuerle and Sienz [35]
optimization problems. Another version of Jaya called SAMP-Jaya al- proposed a method to produce designs in constrained spaces. However,
gorithm was introduced by Rao and Saroj [18] for solving the con- this method is not feasible for high-dimensional problems more than 3D .
strained and unconstrained numerical and engineering optimization Draguljić et al. [36] proposed a CoNcaD algorithm for constructing non-
problems. collapsing and space-filling designs for bounded nonrectangular design
Jaya algorithm and its variations have also been implemented to spaces. Trosset [37] and Stinstra et al. [38] used maximin criterion for the
different fields of science and engineering such as manufacturing [19], construction of space-filling designs in the constrained 10-dimensional
classification [20], power [21], combinatorial optimization [22] and design space. The technique proposed by Trosset [37] and Stinstra et al.
topology optimization of truss structures [23]. [38] does not guarantee the sampled DoE to be non-collapsing.

3. Proposed technique
2.2. Generative design
This section presents details of the proposed Sf-GDT that explores a
To date, the field of generative design has been passed through the design space to generate N designs. We first outline the core idea behind
various advancements for different applications. Several GDTs have Sf-GDT approach and then the ability of Sf-GDT to explore constrained
been proposed by different researchers for architectural applications spaces with continuous and discrete design parameters will be ex-
and for the creation of a specific class of products. Apart from the plained.
techniques developed for the architectural applications, here, we dis-
cuss some recent studies that are close to the proposed technique.
3.1. The Sf-GDT
An exhaustive searched based GDT was proposed by Krish [12] for
creating design alternatives. In which, designs are randomly searched in
Basic terminologies are described first in relation to problem setting.
the design space and to generate dissimilar designs, the designer defines
A CAD model m can be represented by n number of design parameters
a threshold value, which is set on the Euclidean distance, between the
x m,1, xm,2 , x m,3 , …, xm, n . Each design parameter defines a dimension in
generated designs. A major drawback of this technique lies in its ex-
the design space. To form the design space limits, the upper and lower
haustive search strategy, which hinders designers from exploring and
bounds for each design parameter are set. [x ml , j ] and [x mu , j ] represents the
creating optimum design options. A practical generative design system
lower and upper bounds of the jth design parameter, respectively,
called DreamSketch was developed by Kazi et al. [4] to support gen-
where j = 1, 2, 3, …, n . Therefore, a design space is formed by a set of n
erative design at the conceptual phase. In DreamSketch, a user creates
design parameters along with their lower and upper bounds.
an initial design by sketching and then its alternatives are generated in
To generate an optimal set of N design alternatives with the ap-
the sketched context. In order to benefit from DreamSketch, a user
propriate degree of dissimilarity, designs must be uniformly distributed
requires possessing digital sketching abilities. A shape sampling tech-
with the maximum separating distance within the design space.
nique, similar to ours, have been proposed by Gunpinar and Gunpinar
Therefore, a cost function based on the Audze and Eglais [39] technique
[24], and Khan and Gunpinar [25]. However, these techniques lack the
is utilized, which follows a physical analogy: Molecules in a space exert
ability to work with discrete parameters and present no practical ap-
repulsive forces on each other that lead to potential energy in a space. These
proach to design parametrization and design space formulation. Fur-
molecules are in equilibrium in case of minimum potential energy. The
thermore, the sampling technique of [25] is computationally expensive
analogous potential energy U1 (B ) for the creation of the space-filling
compared to the proposed technique. A biologically motivated algo-
designs is defined as:
rithm was developed by Runions et al. [26] for the generative creation
of leaf venation patterns. Sousa and Xavier proposed a symmetric-based N 1 N
1
generative technique for digital fabrication of geometric shapes like a U1 (B ) = 2
p = 1 q= p + 1
Lpq (2)
triangular prism, cuboctahedron, and rhombicuboctahedron, etc.
In literature, techniques like shape grammars [27], shape syntheses where
[28] and L-systems [29] have been utilized by researchers to develop
n
generative systems. Shape grammars are a generative method for Lpq = (x p, j xq, j ) 2
j=1 (3)
creation of design alternatives by incorporating geometric logics/rules
and have been utilized in different applications such as product design Here, Lpq is the distance between the designs p and q, and x p, j and x q, j
[30], architectural design [27], and embroidery design [31], etc. De- are the scaled parameter values for the jth dimension of these designs,
spite being its usage for different application, shape grammars’ usage is which are computed by scaling parameter values between 0 (i.e., lower
limited to the industry because of its computational complexity and bound for the parameter) and 1 (i.e., upper bound for the parameter).
difficulty in developing user interfaces [32]. L-systems are a variation of The design space formed from these bounds is called scaled design space.
shape-grammars and has been used for different design problems such Recall that N is the number of designs to be generated and n is the

714
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

number of dimensions in the design space. usability. DoE are crucial in physical analyses, which has the major goal
The optimization problem for Sf-GDT can be formulated as the of determining which design parameters have more effect on the si-
minimization of U1 (B ) to generate N optimum solutions (or designs). mulation results. Most analyses are computationally expensive, and
However, standard Jaya algorithm provides a single optimal solution by running the analysis for collapsing designs and non-space-filling would
guiding the initial population of individuals to an optimum position. ultimately result in an unnecessary computational effort [36].
Therefore, the search strategy of Jaya algorithm has to modify in order
to provide N optimum solutions. The optimal design creation process of 3.2. Weight-grid-search technique
Sf-GDT starts by creating the random initial population P consisting of
N subpopulations (i.e., P = [(p1 ) s × n , (p2 ) s × n , (p3 ) s × n , …(pN ) s × n]T ). Minimization of U1 (B ) favors placement of the designs at the max-
pL = [X1 , X2 , …Xs ]T denotes the Lth subpopulation of P and imum separating distance from each other. In the case of high-dimen-
L = 1, 2, 3, …, N . Each subpopulation consists of s solutions and the sth sional design space, this function itself locates some designs at the
solution Xs is comprised of n design parameters (i.e., boundaries of the design space [40]. This will result in the violation of a
Xs = [xs,1, xs,2 , …xs, n]). non-collapsing criterion [36] (i.e., designs not sharing any parameter
For each solution, there is a subpopulation of size s, during con- values within a specific interval), thereby generating similar designs.
vergence all the N subpopulations are guided to their optimum position For generative designs, it is desired to spread designs evenly also in
with Eq. (1) under the consideration of their best and worst solutions. N the inner portions of the design space. Therefore, the non-collapsing
worst and best solutions are selected, one from each subpopulation. The criterion for the generated designs should be satisfied as much as pos-
best and the worst solutions are the individuals that minimize and sible. A weighted-grid-search technique is introduced in order to gen-
maximizes the cost function, receptively. The cost function is calculated erate non-collapsing designs in the design space. A new term, U2 (B ) , is
based on the best solutions of the subpopulations. The division of po- included in the cost function, which is as follows:
pulation P into subpopulations is similar to [18,25]. Let
B = [B1, B2, …, BN ] and W = [W1, W2, …, WN ] are sets of best and worst N 1 N n

solutions, respectively, and BL and WL is the best and worst solution for U2 (B ) = × f yp, j , yq, j
p = 1 q= p + 1 j= 1 (4)
the Lth subpopulation. For the selection of N best and worst initial so-
lutions, there are 2 × s N combinations. This means that the cost func-
1 if yp, j = yq, j
tion has to be evaluated 2 × s N times. For instance, at N = 10 and f yp, j , yq, j =
s = 40 setting, 2 × 10485760000000000 evaluations of cost function has 0 otherwise (5)
to be performed for the selection of B = 10 and W = 10 solutions. This
where
can result in a high computational cost if N or s are assigned to a larger
value. Therefore, an initial-designs-selection strategy is utilized for the if x pe, j x p, j < x pe,+j 1 then yp, j = e
selection of N initial worst and best solutions.
if xqe, j x q, j < xqe,+j 1 then yq, j = e (6)
The initial-designs-selection strategy for the selection of N best initial
solutions is based on the fact that the best individuals have the ability to
The term U2 (B ) is based on the degree of violation for the non-collap-
select other exceptional individuals from a group. Following the similar
sing criterion. Here, is a user-defined parameter adjusting the weight
analogy, N is first set to 2 in the cost function, and two individuals of the
of the U2 (B ) term. yp, j and yq, j in Eq. (4) are the corresponding integer
first two subpopulations that minimize the cost are selected as best solu-
coordinate values for x p, j and x q, j in the jth dimension, respectively. To
tions B1 and B2 . Afterward, a solution that minimizes the cost function is
calculate yp, j and yq, j the range of each design parameter is partitioned
selected as the best solution B3 from the third subpopulation. This solution
is selected under the consideration of the preselected solutions B1 and B2 into N equal intervals (levels) as follows: [x ml , j = xm1 , j , xm2 , j , …, x mN, j = xmu , j]
and an integer coordinate e is assigned to them using Eq. (6), where e
by setting N = 3. The selection process is repeated in a similar manner
ranges from 1 to N. Based on these integer values, the piecewise
until N best solutions from the N number of subpopulations are de-
function f in Eq. (5) decides if the designs p and q are collapsing or non-
termined. Similarly, this selection strategy is utilized to select N worst
collapsing.
solutions, which maximizes the cost function. Note that this selection
N
strategy checks 2 × s 2 + 2 s individuals’ combinations to select each set The maximum value for this term can be n × ( 2 ) . 2 represents the
N N
()
N N!
of N best and worst initial solutions. combinations between designs, which is as follows: ( 2 ) = 2 ! (N 2) ! .
In Sf-GDT, the optimization process in any iteration is completed by Setting the parameter to small values will lead to semi non-collapsing
performing the N number of sub-iterations, one for each subpopulation. designs and larger values will produce more non-collapsing designs.
Each subpopulation moves towards the better position in design space The cost function U (B ) , which is given in Eq. (7), have to minimize to
individually while keeping the best and worst solutions of other sub- create space-filling and non-collapsing designs. This function is overall
populations the same. During optimization, a new position for a solu- composed of a parameter , and U1 (B ) and U2 (B ) for space-filling and
tion is found using Eq. (1). Let Xk and Xk be the current and new po- non-collapsing criteria, receptively. Algorithm 1 summarizes the step-
sitions of a solution in the first subpopulation, respectively. The new wise procedure of Sf-GDT.
position of the solution is accepted if the cost value of
N 1 N N 1 N n
B = [Xk , B2, …, BN ] is less than B = [Xk , B2, …, BN ]. The best B and 1
Minimize U (B ) = + × f yp, j , yq, j
worst W solutions are updated after each sub-iteration for the sub- 2
Lpq
p = 1 q= p + 1 p=1 q=p+1 j =1
population. The best solution is an individual having a minimum cost
value (computed with the best solutions of other subpopulations) (7)
among the other solutions in the same subpopulation. Sub-iterations in Fig. 2(a) shows a 3D CAD model parameterized with two design
other subpopulations are performed in the same way. An iteration is parameters x1 and x2 . The design parameter and their parametric ranges
completed when a sub-iteration for each of the subpopulation is per- ([x1l] x1 [x1u] and [x 2l ] x2 [x 2u]) forms a 2D design space. 20 de-
formed. After Sf-GDT stops the convergence process the best solutions sign alternatives for this CAD model are created using the proposed Sf-
of the subpopulations are regarded as final optimal designs. GDT under the space-filling criterion (Eq. (2)), Non-collapsing criterion
Furthermore, the alternatives obtained from Sf-GDT can work as (Eq. (4)) and combined space-filling and non-collapsing criteria (Eq.
Design of Experiments (DoE) for physics simulations, which can be run (7)), which are shown in Fig. 2(b), (c) and (d), respectively. It should be
for validation of designs’ functionality, structural integrity, and noted that each point/dot in the Fig. 2 represents a position of a design

715
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

in the design space. In Fig. 2(a), it can be seen that the design alter- parameters. Sf-GDT is customized in the following way in order to be
natives are spread evenly in the design space, therefore, space-filling employed for the discrete design parameters.
designs can be obtained using Sf-GDT. However, in this case, there are Suppose, an integer value (round 1, rectangular 2, and tri-
more designs are the boundary of the design space. Therefore, ex- angle 3) is assigned to each of three styles. Let x p, d be the dth discrete
ploration performed only based on this criterion may not produce sa- parameter of design p containing t styles and [x pl , d] and [x pu, d] are the
tisfactory designs because during space exploration designer desires to lower and upper bounds for x p, d , respectively. In the above case,
obtain designs that also evenly covers the inners regions of the design t = 3, [x pl , d] = 1 and [x pu, d] = 3. Instead of dividing this parameter into N
spaces. Furthermore, the design alternatives in Fig. 2(c) are created number of intervals, it should be divided into t number of styles. Now,
with the only non-collapsing criterion and has resulted in designs with the range of x p, d is divided into t equal number of intervals as follows:
the poor space-filling property. It can be observed in the Fig. 2(c) that [x pl , d = x p1, d , xp2, d , …, x pt , d = xpu, d]. After Sf-GDT converges, all the design
the better space exploration is achieved when the when both space- parameters, including x p, d , of each design consists of continuous values.
filling and non-collapsing properties are considered (i.e. when space The parameter x p, d contains the style profiles in the form of discrete
exploration is performed using the Eq. (7), which involves both space- values and required to be converted to discrete values. Otherwise, no
filling and non-collapsing criterion). Later, the 3D CAD model in decision can be made on the selection of the style shape. Therefore,
Fig. 2(e) is parameterize with design parameters x1, x2 and x3 . Here, after generating N designs, continuous values of x p, d for each design will
these three design parameters form a 3D design space. 20 design al- be converted into discrete values by using Eq. (8).
ternatives are created in the 3D space using the Sf-GDT (see Fig. 2(f)).
if x pr , d xp, d < x pr + 1
, d then x p, d = r (8)
Again, the points in this space represent the positions of the 20 design
alternatives. Here, r in an integer number ranging from 1 to t.
Algorithm 1. The pseudo-code of Sf-GDT
3.4. Generation of design from constrained spaces

1: Create an input CAD model and parameterize it with n design The design space consists of feasible and infeasible regions in the
parameters ( x1, x2 , …, xn ). presence of geometric constraints. Feasible regions consist of feasible
2: Initialize the number of parameters (n), parameter ranges, designs that satisfy the predefined constraints. Infeasible designs are
number of design to be created (N), subpopulation size (p) and located in the infeasible regions. There are different types of constraint
parameter . handling techniques are available in the literature, such as the in-
3: Randomly create an initial population P of feasible solutions/ corporation of static penalties, dynamic penalties, adaptive penalties
designs within the parametric ranges consisting of N etc. In this study, Deb’s heuristic constrained handling method [15] is
subpopulations (pL ) s × n of size s, where 1 L N . adopted in order to avoid Sf-GDT from selecting designs from con-
4: Obtain set of N initial best (B) and worst (W) designs, one from strained spaces. Deb’s method uses a tournament selection operator in
each subpopulation based on the initial-designs-selection which two solutions are selected and compared with each other. A
strategy. design p is said to be constrained-dominate other design q if any of the
B = [B1, B2, …, BN ], W = [W1, W2, …, WN ] following heuristic rules are true:
5: while termination criterion is not satisfied do
6: for L = 1 to N do 1. Design p is feasible and design q is not.
7: for k = 1 to s do 2. Designs p and q both are infeasible but design p violate less number
8: Update the design Xk of (pL ) s × n using Eq. 1 based on the BL of constraints.
and WL and obtain an updated/new design Xk . 3. Designs p and q both are feasible but design p has better cost
9: Calculate the cost value U (B ) and U (B ) using Eq. (7) for function value.
B = [X1 , B2, …, BN ] and B = [X1 , B2, …, BN ].
If design p constrained-dominate design q then design p is selected.
10: if U (B ) < U (B ) then
This domination is checked at the end of each sub-iteration. There can
11: Accept the design Xk .
be the case when both designs, p and q, are infeasible and have the same
12: else
number of constraint violations then the design with better cost value is
13: Accept the design Xk .
selected. In case of constraint space, Sf-GDT generates an initial po-
14: end if
pulation P consisting of only feasible solutions. So, during the selection
15: end for
of the initial best and worst solutions, the initial-designs-selection
16: Obtain the updated (pL ) s × n , which is (pL ) s × n . strategy does not have to check these constrained handling rules.
17: Find the new best BL and worst WL solutions from (pL ) s × j .
18: Replace BL and WL with BL and WL in the initial set 3.5. Design parameterization
(B = [B1 , B2, …, BN ], B = [W1 , W2, …, WN ]).
19: end for An effective design parameterization of a CAD model is required to
20: end while create variant designs. All the important features of the design should
21: Final N optimal designs are obtained. be parameterized with the appropriate number of design parameters.
However, a decision on the suitable set of parameters is a critical step in
the parametrization, which requires the strong understanding of the
design requirement and key attributes. There are different techniques
available in the literature on how to form a well-structured parametric
3.3. Sf-GDT for discrete parameters model [41]. A well-structured model can enable the designer to create a
variety of design alternatives within its design requirement than a
Sf-GDT also gives the ability to the designer to explore design space poorly structured model. The high number of design parameters may
by synthesizing the design with different ”style” profiles (e.g., round, not keep the original form of the design. As mostly designers desire to
triangular, and rectangular, etc). The designer can add an option for keep the common underlying structure of the model while generating
variable base styles that can be implemented as discrete design its alternatives. On the other hand, less number of parameters can

716
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

Fig. 2. Design alternatives for a 3D CAD model

with two design parameters (a) are obtained in
2D spaces considering; (b) only space-filling
criterion, (c) only non-collapsing criterion, and
both space-filling and non-collapsing criteria
using Sf-GDT (d). Design alternatives for the
same CAD model with three parameters (e) are
generated in 3D design space using Sf-GDT
while considering both space-filling and non-
collapsing criteria (f).

narrow down the design space and larger variation of designs may not stage and based on these specifications the designer limits the space.
be achieved. Therefore, the decision on the selection of appropriate Autonomous Formulation: The autonomous formulation helps to
design parameters should be carefully made. coarsely form the design space as a percentage of the initial parameter
One strategy, which the designer can follow, is to first detect the values of the design. This formulation happens when no primary un-
important features of a given model and then these features can be derstanding of the design specifications are available in the conceptual
parametrized with a relatively higher number of parameters and de- phase. The autonomous formulation gives a good initial guess of sui-
signs can be generated with these parameters. Later, after some trials, table space limits. With this formulation, the designer can first in-
the designer can detect quixotic parameters and eliminate them by adequately build up an initial map of promising regions of the design
directly modifying the CAD model. Such capability of the generative space and then explore designs in that space. Afterward, the designer
design system is recognized as ’designerly’ method, which allows de- can further reform the design space based on the previous exploration
signers to modify the model under consideration and use its generative results. There can be some infeasible designs in the autonomously
capabilities at any phase of the design process [12]. After exploring the formalized space, but this can be overridden by implementing geo-
designs based on the important features, later, if required, design space metric constraints.
can be explored based on its nominal features. Interactive Formulation: In the interactive formulation of the
design space, the designer creates multiple spaces and gradually pro-
ceeds to a final design. First, the designer can autonomously form an
3.6. Formulation of design space initial design space around the given CAD model and creates designs in
this space. Afterward, the designer can select a design and then for-
As stated before, the design space for any CAD model is formed by malize an autonomous space around that design. In this way, the de-
the number of the design parameters and their bounds. The di- signer can interactively proceed by selecting designs and forming the
mensionality of the design space depends on the number of design design spaces until he/she achieves a final desired design. For example,
parameter used to define the CAD model and the limits of the design Fig. 3 gives the illustration of the interactive formulation of the design
space are set by defining the upper and lower bounds for each design space. In which initial space (design space 1) is formed around the
parameter. However, formulation of a suitable design space is a decisive initial design. A design (marked in green) is selected from this space
task as the performance of a technique in term of creating better design and then a new space (design space 2) is formed around the previously
alternatives mainly depends on it. Setting up the design space should be selected design. This process continues until the final design is
carefully done in order to achieve the maximum performance of the Sf- achieved. During selection, if the designer selects more than one design,
GDT and should have sufficient high potential region. If design space is then a new design space is created around the centroid of the selected
too narrow then Sf-GDT will result in the creation of similar/same de- designs. The designer can also refine the space after each interaction as
signs. On the other hand, a vast design space can result in the waste of he/she approaches the final design. Once the final design is selected
computational effort in exploring undesirable regions of the design then, if desired, it can be further modified easily due to its parametric
space. Typically, a design space is set up by defining the upper and
lower bounds of the design parameters. Where each parameter re-
presents a dimension in the design space. Defining the upper and lowers
bounds usually done based on the initial design specifications and de-
signers’ understanding of the design.
In Sf-GDT, design space formulation can happen in three different
way; explicit formulation, autonomous formulation, and interactive
formulation.
Explicit Formulation: The explicit formulation of the design space
happens when the design specifications are known at the conceptual Fig. 3. Interactive formulation of design space.

717
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

nature. value for B = [X11, X12 ]:

4. Results and discussion 6.4 T

0.28 T

Scale X11 and X12 between 0 and 1 X11 = 9.8 0.46 X12
In this section, we first demonstrated a step-wise procedure for 5.0 0.29
implementing Sf-GDT on a simple 3D CAD model and then discuss the 13.6 T
0.66 T

performance of Sf-GDT for different test models and settings. The pro- = 2.5 0.08
posed technique has also been compared with the existing state-of-the- 5.3 0.31

1 2
1 1
U1 (B ) = 2 = 2
L12
Lpq
p = 1 q= p + 1

Distance between first and second design of B = L12 = (0.28 0.66)2 + (0.46 0.08)2 + (0.29 0.31)2 = 0.54
1
U1 (B ) = 2
L12
= 3.4 U2 (B ) = 1.0
Cost function U (B ) = U1 (B ) + × U2 (B ) = 8.4

art techniques. Similarly, calculate cost for B = [X11, X22], B = [X21, X12] and
B = [X21 , X22]:
4.1. Implementation of the Sf-GDT
B = [X11, X22] U (B ) = 9.5
To generate design alternatives with Sf-GDT, first, develop an input B = [X21, X12] U (B ) = 0.9
CAD model, shown in Fig. 2(e). This CAD represents a celling lamp and B = [X21, X22] U (B ) = 1.0
parameterized with three design parameters, x1, x2 and x3 . Each design
Solution set [X21, X12] ([X11 , X22]) give lowest (highest) cost, there-
parameter denotes the radius of the circular region of the lamp model.
fore, X21 ( X11) and X12 ( X22 ) are regarded as the best (worst) solu-
Then, form an explicit design space for this model my defining the
tions of p1 and p2 , respectively.
parametric ranges as 1 x1 20, 1 x2 20 and 1 x3 15 , and fi-
2. Under the consideration of X21 ( X11) and X12 ( X22 ) find a best (worst)
nally, perform following steps to generate design alternatives for this
solution of p3 .
CAD model:
Calculate cost for B = [X21 , X12 , X13]:

Step 1: Initialize the following parameters: 3.9 T

0.15 T
13.6 T
0.66 T

– Subpopulation size (s) = 2 X21 = 16.1 0.79 X12 = 2.5 0.08 X13
– Number of designs (N) = 4 13.6 0.90 5.3 0.31
– Weight parameter ( ) = 5
T T
7.4 0.34
– Number of design parameters (n) = 3 = 11.9 0.57
– Ranges of design parameters 3.7 0.19
Step 2: Randomly generate a population P consists of N sub-
populations. Each subpopulation contains s = 2 initial designs/so-
2 3
lutions X1g and X2g . The super script g represents the subpopulation 1 1 1
U1 (B ) = = +
to which these solutions belong. The initial population is shown p = 1 q= p + 1
2
Lpq 2
L12 2
L 23

below:
L12 = (0.15 0.66)2 + (0.79 0.08)2 + (0.90 0.312 = 1.06
P= [(p1 ) 2× 3 (p2 ) 2× 3 (p3 )2 × 3 (p4 )2 × 3 ]T
L 23 = (0.66 0.34)2 + (0.08 0.57)2 + (0.31 0.19)2 = 0.85
1 1
U1 (B ) = 2
L12
+ 2
L23
= 2.3 U2 (B ) = 1.0
X11 6.4 9.8 5.0 X12 U (B ) = U1 (B ) + × U2 (B ) = 7.3
(p1 )2 × 3 = = (p2 ) 2× 3 =
X21 3.9 16.1 13.6 X22 Similarly, calculate cost for B = [X21 , X12 , X23]:
13.6 2.5 5.3
= B = [X21 , X12 , X23] U (B ) = 24.1
19.1 5.4 10.3
X13 7.4 11.9 3.7 X14 The solution X13 ( X23)
give lowest (highest) cost value and thus
(p3 )2× 3 = = (p4 )2 × 3 =
X23 10.9 17.0 12.7 X24 regarded as best solution of p3
3. Select the best (worst) solution of the subpopulation p4
16.6 18.4 5.6
=
19.8 17.0 10.7 B = [X21, X12 , X13 , X14 ] U (B ) = 20.7
B = [X21, X12 , X13 , X24 ] U (B ) = 14.2
Step 3: Select an initial set of best and worst solutions, one from The best (worst) solution of p4 is X24 ( X14 ).
each subpopulation, using the initial-designs-selection strategy de- 4. The initial best (worst) solution set is
scribed in Section 3. This strategy works as follows: B = [B1, B2, B3, B4] = [X21, X12 , X13 , X24 ]
1. Calculate the cost value U (B ) using Eq. (7) for s 2 = 4 combina- (W = [W1, W2, W3, W4] = [X11, X22 , X23 , X14 ]).
tions of solutions in population p1 and p2 . Then select a combi- Step 4: Update solution X11 of p1 based on its best and worst solutions
nation which gives lowest (highest) value of the cost as best using Eq. (1).
(worst) solution. First, calculate the potential energy U1 (B ) using
T
Eq. (2) and number of collapsing designs U2 (B ) using Eq. (4) and 4.9
then input these values in Eq. (7) to calculate U (B ) . Calculate cost X1 1 = X11 + r1 B1 X11 r2 W1 X11 = 12.3
9.6

718
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

Where Table 1
T T
Parameter ranges for the test models.
0.6 0.02
r1 = 0.4 r2 = 0.5 Speaker model
0.5 0.6 5 x1S 150 5 y1S 90 5 x 2S 150 5 y2S 90 5 x 3S 150
5 y3S 90 5 x4S 150 5 y4S 90 5 x5S 150 5 z1S 90

Step 5: Calculate the cost U (B ) and U (B ) for B = [X1 1, B2, …, BN ] 5 x 6S 150 5 z2S 90 5 x 7S 150 5 z3S 90 5 x 8S 150

and B = [X11, B2, …, BN ], respectively. 5 z 4S 90 5.0 x 9S 90 5 z5S 90 5 x1 0S 90 5 z 6S 90

5 x1 1S 90 5 z7S 90
B = [X1 1, B2, B3, B4 ] U (B ) = 23.7
Motorbike model
B = [X11, B2, B3, B4] U (B ) = 64.1 6 r1M 11 20 x1M 25 3.5 y1M 6 20 x 2M 25 3.5 y2M 5

As U (B ) < U (B ) so accept the new solution X1 and reject the old1 1.8 y3M 2.3 1 x 3M 2 5 y4M 7 3 x4M 8 6 y5M 12
solution X11. 2 x5M 6 6 y6M 12 10 x 6M 12 6 y7M 12 4 z2M 6
Step 6: Similarly, update the solution X21 of p1. 4 z3M 6 3 z 4M 5 3.5 z5M 5.5 3 z 6M 6 3.5 z7M 5
T 15 y8M 20 17 y9M 23 15 y1 0 M 20 8 y11M
13 10 M
y12 15
3.8
X2 1 = X21 + r1 B1 X21 r2 W1 X21 = 19.6 3 z 8M 7 3 z 9M 6.5 2 M
z10 6 2 x 7M 3.5 2 M
y13 3.5
12.7 1.5 M
y14 3 2.5 x 8M 4 1 M
y15 2 4 x 9M 6 2 M
x10 5
M M M M M
0.8 y16 3 1 x11 4 2 y17 4 7.5 x12 12 4 x13 7
2 y18M
5 1 r2M 7
Step 7: Calculate the cost U (B ) and U (B ) using B = [X2 , B2, …, BN ]
1

and B = [X21 , B2, …, BN ], respectively. Celling lamp model

1 y1L 10 1 y2L 10 1 y3L 10 1 r1L 20 1 r2L 20
B = [X2 1, B2, B3, B4 ] U (B ) = 14.0
B = [X21, B2, B3, B4] U (B ) = 14.2

U (B ) < U (B ), so accept the new solution X2 1. glass, which are shown in Figs. 4(a), (b), 6(a) and 7(a), respectively.
Step 8: Obtain the updated subpopulation p1 These models were selected based on their aesthetic importance. A wine
glass defines elegance of the wine drinker, an aesthetic ceiling lamp, an
X1 1 4.9 12.3 9.6 elegant speaker box and motorbike design can attract more customers.
(p1 ) 2 × 3 = =
X2 1 3.8 19.6 12.7 Except for the motorbike, these models are single component 3D de-
signs. Where the bike model is composed of several design components.
For the complex test model like a motorbike, a user can first work on
Step 9: Find the new best (B1) and worst (W1) solutions of p1 . the low-level details of the design and then can move to the high-level
4.9 3.8 details. For example, the user can first explore the form outline of the
B1 = X1 1 = 12.3 W1 = X2 1 = 19.6 design using Sf-GDT and once a collection of different initial base forms
9.6 12.7 is selected, the designer can then explore further design details by
keeping the base form constant. The user may also first explore the
design space to create design alternatives for each component and then
Step 10: Replace B1 and W1 with B1 and W1 in the initial set
assembles these alternative parts to create the final design. For the
(B = [B1 , B2, B3], B = [W1 , W2, W3]).
motorbike model, only components for that outer appearance is con-
Step 11: Repeat the steps 4–10 to obtain p2 , p3 and p4 .
sidered to be significant are created such as fuel tank, seat, wheels,
Step 12: Repeat the steps 4–11 until the change in the cost function
headlight, backlight, handlebars and speedometer dock. 3D surfaces of
becomes negligibly small between a few consecutive iterations.
the wine glass, ceiling lamp, speaker, fuel tank and seat of motorbike
After 13th iteration algorithm converges and the best solution of
models are created by interpolating Coons patches between spline
each subpopulation is regarded as final optimum design.
curves and design parameters are defined with these curves. The mo-
Step 13: Obtain final design alternatives, which are shown below:
torbike’s front and rare wheels are the 3D solid models.
B = [B1, B2, B3, B4] The speaker model shown in Fig. 4(a) is represented using 22 design

4.2. Test models parameters (n = 22 ). The speaker model is created using three spline
curves. First, a quarter section of the speaker model is created using
To validate the performance of Sf-GDT we also utilized more test these spline curves. Then, this section is mirrored first along the x-y
models, such as a speaker, a motorbike, a ceiling lamp, and a wine plane and then mirrored along x-z plane. Curve 1 lies in x-y plane and

Fig. 4. CAD models of (a) speaker and (b)

Motorbike with their design parameters.

719
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

Fig. 5. Design alternatives generated by Sf-GDT for (a) speaker and (b) motorbike models.

position of its control points is represented by the parameter

x1S , y1S , x 2S , y2S , x3S , y3S , x4S and y4S and Curve 2 lies in x-z plane and
parameters x5S , z1S , x 6S , z2S , x 7S , z3S , x 8S and z 4S denote the position of its
control points. Similarly, x11 S
, z 7S , x10
S
, z 6S , x 9S and z5S represents the
control point position of curve 3. The parameter ranges of the speaker
model are given in Table 1.
Each component of the motorbike model is parameterized sepa- Fig. 6. (a) Parametric CAD model of ceiling lamp. (b) Design alternatives of
rately and consist of total 42 design parameters (n = 42 ), which are ceiling lamp model generated using Sf-GDT with continuous and discrete
shown in Fig. 4(b). The back wheel is parameterized with 6 continuous parameters.
design parameter ( x1M , y1M , x 2M , y2M , y3M and r1M ) and one discrete
parameter (r2M ), and the front wheel is created as copy of the back Table 1.
wheel. x1M , x 2M and y1M , y2M represent the position of control points in x- Sf-GDT was tested for both constrained and unconstrained design
axis and y-axis, respectively, and y3M and r1M are the width of the tire spaces with different algorithm setting and design space formulation.
and radius of the wheel. The discrete parameter, r2M , defines the number Design alternatives for the speaker and motorbike models were created
of spokes. The fuel tank is created using three spline curves, one in the with the application of Sf-GDT in the explicitly formalized un-
y-x plane and two in x-z plane and represented with 14 design para- constrained space and can be seen in Fig. 5. A careful inspection of the
meters. Similarly, the seat of the motorbike is parameterized with 8 designs in Fig. 5 can reveal that the generated alternatives by Sf-GDT
parameters and created using two spline curves, one in 3D space and for each model are distinct from each other to a great extent. This va-
other in the x-y plane. The design parameters x 7M and y13M denotes the lidates the ability of Sf-GDT to create distinct designs for any given CAD
width of the speedometer dock in x-axis and y-axis. The backlight and model. Table 2 provides the algorithm settings and the values of various
headlight are represented with x 8M , x 9M , y14M , y15M and x13M , y18M , recep-
parameters/criteria such as design alternative (N) and design para-
tively, where x 9M and y14M adjusts the length and width of the backlight. meters (n) for the test models, U1 (B ) and U2 (B ) values, computational
The handlebars are also created with spline curves with design para- time and number of iterations (i) performed while creating the designs
meters x10M , x16M , x11M , x17M and x12M representing the position of control
alternatives.
points. The parametric ranges of each design parameter are provided in

4.2.1. Sf-GDT with discrete parameters

Table 2 A ceiling lamp model (see Fig. 6(a)) is used to demonstrate the
Algorithm setting and the results obtained from Sf-GDT for CAD models utilized performance of Sf-GDT for continuous and discrete parameters. The
for experimentation. continuous parameter, y1L , y2L , and y3L , represents the vertical length of
the lamp along the y-axis, and r1L and r2L are the radii of upper and lower
Designs N n U1 (B ) U2 (B ) Maximum value of CT (minutes) i
U2 (B )
circular region of the lamp. Where, discrete parameters, r1d and r2d , each
containing five style forms (t = 5). These style forms, circle 1,
Fig. 5(a) 20 22 46.43 10 4180 3.72 500 square 2, ellipse 3, hexagon 4 and octagon 5, are defined on the
Fig. 5(b) 20 7 151.97 60 1330 1.47 300 profile-1 (P1) and the profile-2 (P2) of the lamp model. Fig. 6(b) shows
Fig. 7(b) 20 10 109.79 0 1900 1.92 300
the design alternatives generated by Sf-GDT based on both discrete and
Fig. 7(c) 20 10 107.94 2 1900 2.10 300
Fig. 7(d) 40 10 487.02 2 7800 21.74 1500 continuous parameters. The ranges of the continuous design parameters
of ceiling lamp model are given in Table 1. It was observed that the

720
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

Fig. 7. (a) Parametric representation of a wine glass model. Design alternatives of glass model generated by Sf-GDT in constrained space with (b) constraint-1 (c)
constraint-2. (d) Design alternatives of glass model generated by utilizing Sf-GDT in an autonomously formed design space.

designs with both continuous and discrete parameters have more var- design. Note that set 2 contains all the plausible designs. From this set,
iation compared to the designs with only continuous parameters. two designs were select and new space was formed around the centroid
of these two designs with 30% extension. Afterward, again 30% exten-
4.2.2. Sf-GDT in constrained design spaces sion was done for the creation of designs in set-4 and from this final
Sf-GDT can generate a variety of designs for a given model in the design was selected and the interactive process was stopped.
constrained and unconstrained design spaces. Both the design specifi-
cations and user preferences can be represented by constraints. To va- 4.3. Computational time (CT)
lidate the performance of Sf-GDT, design specification such as the ca-
pacity of a wine glass to store a certain amount of wine, was given as a A PC having an Intel Core i7-5500 CPU, 2.4 GHz processor and
geometric constraint. The parametric representation of the wine glass 16 GB memory was used for the experiments in this study, and C++
model is shown in Fig. 7(a). 10 design parameters (n = 10 ) are used to programming language along with Siemens’ Parasolid APIs were uti-
represent this model. The design parameter y0G is the vertical length of lized for implementation and testing of Sf-GDT. We measured the CT
glass stem and the design parameter x1G , x 2G , y1G , x 3G , y2G , x4G , y3G , x5G and taken to obtain results in Figs. 5–7, which is shown in Table 2; it varied
y4G represent the 2D position of the control points of spline curve used to between 1.47 and 21.74 min. The study on the effects of these para-
create the profile of the glass. meters on CT is important in order to effectively utilize Sf-GDT. In the
The glass design alternatives in Fig. 7(b) and (c) can store less than proposed approach, CT mainly depends on the number of designs to be
or equal to 200 ( 200 ) and greater than or equal to 700 ( 700 ) mil- generated (N), the dimensionality of the design space (n) and the size of
liliter (ml) of wine, respectively. No design in Fig. 7(b) and (c) have the subpopulations (s). Increase in the values of these parameters will
violated these geometric constraints. increase Sf-GDT’s processing time. As the values of either N or s in-
creases CT for Sf-GDT to create designs increases.
4.2.3. Performance of Sf-GDT in different design space formulations
The performance of Sf-GDT is also validated under different design 4.4. Parameter tuning
space formulation (i.e. explicit, autonomous, and interactive) for the
wine glass. The wine glass designs in Fig. 7(d) are generated by Sf-GDT For the experiments in this study, was set equal to 10 except for
in an autonomously formed design space with 50% extension of initial the motorbike model for which = 20 was utilized because of the high
design. It can be observed from Fig. 7(d) that the underlined designs are number of design parameter (n = 42 ). We recommend the users to set
implausible. These designs may not be feasible as a final market pro- an initial value of equal to n/2 , which can be altered later depending
duct. As mentioned before, one way to overcome this issue is to define on the users’ intention to create complete or semi-non-collapsing de-
geometric constraints. signs.
In interactive space, an initial envelope can be set up either explicitly It is noteworthy that the value of U2 (B ) can be high for the problems
or autonomously. For example, Fig. 8 demonstrate the interactive for- with discrete parameters compared to the same problem with con-
mulation of designs. In Fig. 8, set-1 contains 17 design alternatives for tinuous parameters. Instead of dividing the discrete parameter(s) into N
the wine glass model. These designs were generated in a space that was intervals, we divide these parameters into t intervals, where t is the
created with a 50% autonomous extension of the initial design. This set number of style profiles. If t is less than N, then, the number of col-
contains both plausible and implausible designs. From this set, one de- lapsing designs will increase, which will result in a high value of U2 (B ) .
sign was selected (checked in red). In the next step, this design was In order to have a low number of collapsing designs, t should be greater
considered as new input model and a new set (set 2) of designs were than or equal to N (t N ). However, t < N does not affect the space-
created again with 50% autonomous extension of space around the new filling quality of the designs.

Fig. 8. Design alternatives generated for the wine glass model in an interactively formalized design space.

721
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

Sf-GDT with Krish’s GDT [12]. Fig. 10(a) and (b) shows the design
points representing designs generated by Krish’s technique for the
speaker model in Fig. 4(a) in a 2D design space. The designs in the 2D
space give a better perspective to readers on how designs generated by
[12] are spread in the design space. As mentioned in Section 2, Krish
utilized a threshold value, ranging from 0.0 to 1.0 , to create dissimilar
designs. The designs in Fig. 10(a) and (b) are created with threshold
values of 0.5 and 1.0 , respectively. It can be seen from the Fig. 10(a) and
(b) that the design points (N = 30 ) are not uniformly distributed in the
design space especially when the threshold value is 1.0 . The designs in
Fig. 10(a) and (b) have space-filling of 7149.9 and 44015, respectively.
In case of threshold equal to 1.0 designs are clustered at the two corners
Fig. 9. Plot showing the cost values versus number of A-GDT iterations for the
of the design space and approximately more than 90% of space is left
models in Fig. 5(a), (b), Figs. 6 and 7(b), (c) and (d).
empty. In this case, designs are also generated by Sf-GDT, which are
shown in Fig. 10(c). This gives a comparative view to the readers on
The size of the subpopulations (s) also plays an important role in the how Sf-GDT produces design alternatives for the same CAD model in 2D
generation of space-filling designs. High values of s create diverse initial space. Note that the designs generated by Sf-GDT had space-filling of
solutions for Sf-GDT, which facilitates its search for the global optimum 2492.29, which is less than the designs generated by [12].
solutions. In contrast, the application of Sf-GDT with the high values of
s can result in a higher CT. We recommend setting s to a value higher
than n. For the experiments in the current study, s was set equal to 15 4.6.1. User study
except for the designs in Fig. 5(b). For that s = 23 was selected. Fig. 10(d) shows the designs created by Krish’s technique [12] for
the speaker model in Fig. 4(a) within a 22-dimensional design space,
which were created to visually compare the results of Krish’s technique
4.5. Convergence of Sf-GDT
and Sf-GDT. For this visual comparison, a user study was conducted to
obtain the human perception about the quality of designs generated
The quality of any optimization technique mainly depends on its
from the two techniques. This user study included 12 participants to
ability to provide an optimum solution or a solution close to the global
compare the designs in Figs. 10(d) and 5(a), which are obtained using
optimum. The global optimum is a point in search space where the best
Krish’s technique and Sf-GDT, respectively. Six participants had more
solution(s) exists. As the Sf-GDT is based on the optimization technique,
than two years of design experience in product development, and others
therefore, in order to verify the convergence ability to a global optimum
were selected from the Amazon Mechanical Turk platform. The parti-
its performance is observed against the number of iterations i it per-
cipants were asked to rate each design in Figs. 10(d) and 5(a) based on
forms. The convergence ability of Sf-GDT is analyzed on different test
a Likert scale, with anchors ranging from ”very poor” to ”very good” (1:
models shown in Fig. 5. Sf-GDT stops the optimization process when
very poor, 2: very good, 3: fair, 4: good, 5: very good). The participants
there is no improvement in the cost function U (B ) for some consecutive
involved in the study had not any information about the techniques
iterations (i); at this point, the designs being created reach the optimal
used to generate these designs. This was done to minimize the possi-
position, and the algorithm is considered to converge to its optimality.
bility of a bias decision during design rating. In this study different set
Fig. 9 shows the plot for U (B ) versus i for the designs in Figs. 5–7. A
of rules was applied to the participants to ensure the reliability of the
large number of iterations were performed for these models to analyze
obtained results. The designs in Figs. 10(d) and 5(a) were shuffled
the convergence of Sf-GDT. No improvements were observed in U (B )
randomly and presented in two surveys, each with 25 designs. There
after some consecutive iterations. For the designs in Figs. 6(b) and 7(b),
was a repetition of five designs in each survey. For any participant, if
(c) there was no improvement occurred after approximately 300th
there was no consistency in the ratings given to the designs and survey
iteration and for the designs in Figs. 5(a), (b) and 7(d) Sf-GDT con-
was completed in less than five minutes, then that participant’s results
verged at 500, 1200 and 1500 number of iterations, respectively. The
were excluded from the study. Note that the design space utilized for
convergence rate of Sf-GDT depends on the number of designs (N), the
the generation of alternatives in Fig. 10(d) and 5(a) was same.
dimensionality of the design space (n), and the total number of geo-
Table 3 summarizes the user study’s results. From the table, it can be
metric constraints.
observed that the average rating given by the participants to the designs
generated with Krish’s technique is lower than those obtained using Sf-
4.6. Comparison with existing works GDT. Ten out of 12 participants preferred the designs generated using
Sf-GDT, including the experience designers.
We compared the performance of Sf-GDT with the existing state-of- A t-test was utilized to statistically examine the results of the user
the-art techniques in the literature that have been proposed for gen- study. The data obtained from the user study were normally distributed,
erative and space-filling designs. First, we compared the performance of as the skewness value was close to zero and their mean values were

Fig. 10. Design points created in 2D space by utilizing the technique of [12] with threshold values of (a) 0.5 and (b) 1.0. (c) Design points created using Sf-GDT. (d)
Design alternative for speaker model created by utilizing the technique of [12].

722
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

Table 3
Results of the user study.
Average Grade

User 1 2 3 4 5 6 7 8 9 10 11 12

Sf-GDT 4.55 2.70 3.60 3.25 3.50 3.75 4.00 4.15 4.15 3.10 4.10 3.90
Krish [12] 2.10 3.20 3.65 2.25 3.10 2.45 2.95 4.10 3.75 2.55 2.80 2.75

Space-filling µ Skewness p-value

Sf-GDT 46.430 0.630 3.020 0.12 0.00718

Krish [12] 55.039 0.520 3.730 -0.40

6. Conclusions and future works

This paper proposes a state-of-the-art generative design technique

for the automatic search and generation of design variations for a given
CAD model based on its design specification. From these design alter-
natives, users can select a design(s) based on their aesthetic preference.
Sf-GDT has the ability to generate designs in constrained and un-
constrained design spaces. To obtain distinct and uniformly distributed
designs in the design space, designs with space-filling and non-collap-
sing criteria are favored during the search process. To generate N op-
timal designs based on these criteria, Jaya algorithm is utilized and
Fig. 11. Window-based interface to integrate Sf-GDT with SoildWorks. modified. Sf-GDT, first, randomly generate a subpopulation of solutions
and improves these solutions using search approach of Jaya algorithm.
approximately equal. The null hypothesis states that there is no sig- Finally, Sf-GDT is compared with other existing techniques in the lit-
nificant difference between the ratings given to the designs generated erature. The results of this paper show that Sf-GDT outperforms these
using Krish’s technique and Sf-GDT. The p-value obtained from the t- techniques.
test is less than the significance level of 0.05, this indicates a stronger As a future work, we would like to integrate the users’ preference
evidence against the null hypothesis. and aesthetic judgment into Sf-GDT so that they can create designs
From the results of the user study and statistical test, it can also be based on their preference and aesthetic perception. The proposed
concluded that Sf-GDT outperforms the Krish’s technique in term of technique will be extended for the generative creation of complex 3D
creating appealing design alternatives for the users. character models. Finally, the performance of the different optimization
techniques will be studied for this specific problem.

5. Usage of Sf-GDT with existing CAD software Appendix A. Supplementary material

Sf-GDT can be easily utilized with existing CAD software having Supplementary data associated with this article can be found, in the
parametric modeling functionality and can create a design table in the online version, at https://doi.org/10.1016/j.aei.2018.10.005.
form of a spreadsheet such as Microsoft XL. CAD software such as
SolidWorks has the ability to create and read external XL based design References
tables. A user interface called DesignN is developed to integrate Sf-GDT
with such CAD software, which is shown in Fig. 11. The design para- [1] Y.-C. Liu, A. Chakrabarti, T. Bligh, Towards an idealapproach for concept genera-
meters of a model can be stored in the design table using build-in CAD tion, Des. Stud. 24 (4) (2003) 341–355.
functions. These parameter values can be given as input to DesignN to [2] J. Wang, Improved engineering design concept selection using fuzzy sets, Int. J.
Comput. Integr. Manuf. 15 (1) (2002) 18–27.
create design alternatives and their parameter values can be stored in a [3] G. Pahl, W. Beitz, Engineering Design: A Systematic Approach, Springer Science &
CSV file. These parameter values can be transferred to the design table Business Media, 2013.
that can then read by the CAD software to create designs. Data in the [4] R.H. Kazi, T. Grossman, H. Cheong, A. Hashemi, G. Fitzmaurice, Dreamsketch: early
stage 3d design explorations with sketching and generative design, Proceedings of
design table can also be structured in other formats required by the the 30th Annual ACM Symposium on User Interface Software and Technology,
analytical software. ACM, 2017, pp. 401–414.
Fig. 11 demonstrates the steps involved in the creation of designs for [5] N. Umetani, T. Igarashi, N.J. Mitra, Guided exploration of physically valid shapes
for furniture design, ACM Trans. Graph. 31 (4) (2012) 86-1.
the wine glass model in SolidWorks via the window-based interface of [6] K. Shea, R. Aish, M. Gourtovaia, Towards integrated performance-driven generative
DesignN. The wine glass model is first sketched and the design para- design tools, Autom. Constr. 14 (2) (2005) 253–264.
meters are defined on this sketch. 3D surface model of the wine glass is [7] A. Nana, J.-C. Cuillière, V. Francois, Towards adaptive topology optimization, Adv.
Eng. Softw. 100 (2016) 290–307.
created using Swept Surface feature of SolidWorks. Initial design
[8] M. Turrin, P. von Buelow, R. Stouffs, Design explorations of performance driven
parameter values are inputted to DesignN and designs are created in geometry in architectural design using parametric modeling and genetic algorithms,
50% autonomously formed design space. The parameter values of the Adv. Eng. Inform. 25 (4) (2011) 656–675.
[9] J.J.L. Kitchley, A. Srivathsan, Generative methods and the design process: a design
generated designs are stored in the CSV file and are copied to the design
tool for conceptual settlement planning, Appl. Soft Comput. 14 (2014) 634–652.
table. SolidWorks read these parameter values and generate designs [10] L. Caldas, Generation of energy-efficient architecture solutions applying gene_arch:
that are presented to the user within a Design Tree. Through this tree, an evolution-based generative design system, Adv. Eng. Inform. 22 (1) (2008)
each design can be visually inspected by the user for the final selection. 59–70.
[11] V. Granadeiro, J.P. Duarte, J.R. Correia, V.M. Leal, Building envelope shape design
A web-based interface of DesignN is also developed, which can be ac- in early stages of the design process: integrating architectural design systems and
cessed from https://geometric.brainlabsgp.com and a tutorial to use energy simulation, Autom. Constr. 32 (2013) 196–209.
DesignN with SolidWorks can be found at https://youtu.be/ [12] S. Krish, A practical generative design method, Comput. Aided Des. 43 (1) (2011)
88–100.
QDcW2FPvq-Q.

723
S. Khan, M.J. Awan Advanced Engineering Informatics 38 (2018) 712–724

[13] Á.E. Eiben, R. Hinterding, Z. Michalewicz, Parameter control in evolutionary al- Cruz, Design with shape grammars and reinforcement learning, Adv. Eng. Inform.
gorithms, IEEE Trans. Evol. Comput. 3 (2) (1999) 124–141. 27 (2) (2013) 230–245.
[14] R.V. Rao, G. Waghmare, A new optimization algorithm for solving complex con- [28] E. Kalogerakis, S. Chaudhuri, D. Koller, V. Koltun, A probabilistic model for com-
strained design optimization problems, Eng. Optim. 49 (1) (2017) 60–83. ponent-based shape synthesis, ACM Trans. Graphics (TOG) 31 (4) (2012) 55.
[15] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective genetic [29] P. Prusinkiewicz, M. Shirmohammadi, F. Samavati, L-systems in geometric mod-
algorithm: Nsga-ii, Ocean Eng. 136 (2017) 243–259. eling, Int. J. Found. Comput. Sci. 23 (01) (2012) 133–146.
[16] L. Wang, C. Huang, A novel elite opposition-based jaya algorithm for parameter [30] M.C.A.H.H. CHAU, A. MCKAY, A. DE PENNINGTON, Combining evolutionary al-
estimation of photovoltaic cell models, Optik-Int. J. Light Electron Optics 155 gorithms and shape grammars to generate branded product design, Design
(2018) 351–356. Computing and Cognition06, Springer, 2006, pp. 521–539.
[17] P.D. Michailidis, An efficient multi-core implementation of the jaya optimisation [31] J. Cui, M.-X. Tang, Integrating shape grammars into a generative system for zhuang
algorithm, Int. J. Parallel Emergent Distrib. Syst. (2017) 1–33. ethnic embroidery design exploration, Comput. Aided Des. 45 (3) (2013) 591–604.
[18] R.V. Rao, A. Saroj, A self-adaptive multi-population based jaya algorithm for en- [32] S.C. Chase, Generative design tools for novice designers: Issues for selection, Autom.
gineering optimization, Swarm Evol. Comput. 37 (2017) 1–26. Constr. 14 (6) (2005) 689–698.
[19] R.V. Rao, D.P. Rai, J. Balic, Multi-objective optimization of abrasive waterjet ma- [33] G. Kelly, H. McCabe, Interactive generation of cities for real-time applications, ACM
chining process using jaya algorithm and promethee method, J. Intell. Manuf. SIGGRAPH 2006 Research Posters, ACM, 2006, p. 44.
(2017) 1–27. [34] W. Palubicki, K. Horel, S. Longay, A. Runions, B. Lane, R. Měch, P. Prusinkiewicz,
[20] R.R. Kurada, K.P. Kanadam, Automatic unsupervised data classification using jaya Self-organizing tree models for image synthesis, ACM Trans. Graphics (TOG) 28 (3)
evolutionary algorithm, Adv. Comput. Intell.: Int. J 3 (2016) 35–42. (2009) 58.
[21] S.P. Singh, T. Prakash, V. Singh, M.G. Babu, Analytic hierarchy process based au- [35] F. Fuerle, J. Sienz, Formulation of the audze–eglais uniform latin hypercube design
tomatic generation control of multi-area interconnected power system using jaya of experiments for constrained design spaces, Adv. Eng. Softw. 42 (9) (2011)
algorithm, Eng. Appl. Artif. Intell. 60 (2017) 35–44. 680–689.
[22] R. Buddala, S.S. Mahapatra, Improved teaching–learning-based and jaya optimi- [36] D. Draguljić, T.J. Santner, A.M. Dean, Noncollapsing space-filling designs for
zation algorithms for solving flexible flow shop scheduling problems, J. Ind. Eng. bounded nonrectangular regions, Technometrics 54 (2) (2012) 169–178.
Int. (2017) 1–16. [37] M.W. Trosset, Approximate maximin distance designs, in: Proceedings of the
[23] S. Degertekin, L. Lamberti, I. Ugur, Sizing, layout and topology design optimization Section on Physical and Engineering Sciences, 1999, pp. 223–227.
of truss structures using the jaya algorithm, Appl. Soft Comput. 70 (2018) 903–928. [38] E. Stinstra, D. den Hertog, P. Stehouwer, A. Vestjens, Constrained maximin designs
[24] E. Gunpinar, S. Gunpinar, A shape sampling technique via particle tracing for cad for computer experiments, Technometrics 45 (4) (2003) 340–346.
models, Graph. Models 96 (2018) 11–29. [39] P. Audze, V. Eglais, New approach for planning out of experiments, Probl. Dynam.
[25] S. Khan, E. Gunpinar, Sampling cad models via an extended teaching–learning- Strengths 35 (1977) 104–107.
based optimization technique, Comput. Aided Des. 100 (2018) 52–67. [40] V.R. Joseph, E. Gul, Maximum projection designs for computer experiments,
[26] A. Runions, M. Fuhrer, B. Lane, P. Federl, A.-G. Rolland-Lagan, P. Prusinkiewicz, Biometrika 102 (2) (2015) 371–380.
Modeling and visualization of leaf venation patterns, ACM Trans. Graphics (TOG) [41] J.D. Camba, M. Contero, P. Company, Parametric cad modeling: an analysis of
24 (3) (2005) 702–711. strategies for design reusability, Comput. Aided Des. 74 (2016) 18–31.
[27] M. Ruiz-Montiel, J. Boned, J. Gavilanes, E. Jiménez, L. Mandow, J.-L. PéRez-De-La-

724