Advanced Engineering Informatics 54 (2022) 101786

Contents lists available at ScienceDirect

Advanced Engineering Informatics

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

Full length article

Data-driven generative design for mass customization: A case study

Zhoumingju Jiang a, Hui Wen b, Fred Han a, Yunlong Tang c, *, Yi Xiong a, *
a
School of System Design and Intelligent Manufacturing, Southern University of Science and Technology, Shenzhen 518055, China
b
School of Mechanical and Electrical Engineering, and the National Demonstration Center for Experiment Electronic Circuit Education, Guilin University of Electronic
Technology, Guilin 541004, China
c
Mechanical and Aerospace Engineering Department Monash University, Melbourne 3168, Australia

A R T I C L E I N F O A B S T R A C T

Keywords: Generative design provides a promising algorithmic solution for mass customization of products, improving both
Design for additive manufacturing product variety and design efficiency. However, the current designer-driven generative design formulates the
Generative design automated program in a manual manner and has insufficient ability to satisfy the diverse needs of individuals. In
Data-driven design
this work, we propose a data-driven generative design framework by integrating multiple types of data to
Mass customization
Design automation
improve the automation level and performance of detail design to boost design efficiency and improve user
Product design satisfaction. A computational workflow including automated shape synthesis and structure design methods is
established. More specifically, existing designs selected based on user preferences are utilized in the shape
synthesis for creating generative models. For structural design, user-product interaction data gathered by sensors
are used as inputs for controlling the spatial distributions of heterogeneous lattice structures. Finally, the pro­
posed concept and workflow are demonstrated with a bike saddle design with a personalized shape and inner
structures to be manufactured with additive manufacturing.

[6], which is a highly automated and effective approach. However, at

the beginning of the design, designers still need to create a parametric
1. Introduction model [7] and set design objectives, constraints, and variables. Such a
design process will become impractical once the variation of user needs
Mass customization is a product development paradigm that aims to become increasingly large. In addition, one inherent limitation of the
provide users with tailored products and services while balancing designer-driven approach is the large difficulty for designers to capture
product variety and production costs [1,2]. The pioneering work on the actual needs of each user, particularly their implicit needs, clearly,
mass customization, dating back to the early 1990s, focuses on devel­ precisely, and unbiasedly [8]. It is evident that the high involvement of
oping product family architecture [3] and associated product platform designers within the current generative design for mass customization
strategies. Both module-based [4] and scale-based [5] product platforms has hindered the utilization of advanced manufacturing technologies
have been proposed to generate product variants. The first approach moving toward product personalization. The use of data-driven methods
creates family members by adding, substituting, and removing func­ provides a promising solution to resolve this problem.
tional modules, and the latter approach makes products via scaling The use of data-driven approach has become pervasive in product
design variables. The number of modules and variables of these plat­ development, with successful cases reported for various products, e.g.,
forms were mainly constrained by the manufacturing capabilities. In ankle braces [9], sports shoes [10], and vehicle chassis [11]. Data-driven
recent years, the rapid development of additive manufacturing provides design relies on data and the information contained therein to make
a more flexible and agile process to remove these constraints, supporting decisions. The data used include user-end data, e.g., descriptive and
the realization of more product variants at affordable cost. behavior data, production-end data, e.g., design and manufacturing
Emerging computation design approaches, e.g., generative design, data, as well as service-end data, e.g., product usage data and user
have been used in mass customization for fully harnessing the design feedback [12]. Both explicit and implicit information about the user can
opportunities provided by advanced manufacturing technologies for be elicited from these data with data-driven methods. For example,
improving user satisfaction. In generative design, user-defined algo­ design preference is identified from historical purchase data of
rithms synthesize designs via explicit programming or implicit learning

* Corresponding authors.
E-mail addresses: [email protected] (Y. Tang), [email protected] (Y. Xiong).

https://doi.org/10.1016/j.aei.2022.101786
Received 10 July 2022; Received in revised form 21 September 2022; Accepted 15 October 2022
Available online 26 October 2022
1474-0346/© 2022 Elsevier Ltd. All rights reserved.
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

Table 1
Nomenclature A literature survey on data-driven product design.
Methods Design stages
DGM deep generative model
References Tools Data types Rqrt. Cnpt. Dtl.
GAN generative adversarial network
Zhou et al. Case reasoning User reviews
VAE variational autoencoder √ – –
[32]
RL reinforcement learning Burnap et al. Quantitative Prior purchase √ – –
Rqrt requirement analysis [13] preference model data
Cnpt conceptual design Feature extraction
Dtl detail design Lin et al. [33] Data mining User experience √ – –
Joung et al. Importance User reviews
NLP natural language processing
√ – –
[31] performance
F a set of feature vectors analysis
fusr feature vector of user’s needs Wang et al. Graph-based Sensor data √ – –
fdsg feature vector of an existing product [34] framework
pusr latent space position of a user identified generated He et al. [36] Natural language Idea data – √ –
processing (NLP)
design Jiang et al. Automatic Patent images – √ –
vusr product feature in latent space [37] vectorization
vcust customized input of shape generator Convolutional
G generator neural network
Liu et al. [38] Machine learning Web documents
D discriminator
– √ –
Concept clustering Literature
z latent variable NLP
GUI graphical user interface Shi et al. [39] Text mining Textual data – √ –
IoT internet of things Ontology network
Wang et al. Context-aware Sensor data – √ –
[40] approach
Bu et al.[41] Hybrid intelligence Sensor data – √ –
approach
automobiles using a feature extraction method [13]. Meanwhile, im­ Nourbakhsh Design synthesis Sensor data – √ √
plicit needs are difficult to obtain due to the nature of being incorpo­ et al.[11] IoT
rated in more subtle forms of exchange. With more products getting Lewis et al. Design analytic User reviews √ √ √
[35] Opinion mining Sensor data
equipped with sensors and being interconnected, it becomes possible to
Embedding sensors
gather product in-service performance [14], e.g., usage frequency and Lee et al.[42] Kansei engineering Semantic √
interaction status, and leverage these data for obtaining implicit needs. method customer
The study aims to integrate data-driven methods into the generative requirements
design-based mass customization process to boost design efficiency and Product form
Shieh et al. Support vector Product form
improve user satisfaction. The main contribution of this work is three­
√
[43] regression Experiment data
fold. Firstly, a data-driven generative design framework, complement­ Multi-objective
ing the current designer-driven manner, is proposed for the detail design evolutionary
stage of mass customization. By mining multiple types of data, the algorithm
automation level and performance of the detail design stage are Wang et al. Machine learning Product form √
[44] Product color
improved for rapidly satisfying a large number of individual needs. Tang et al. Gaussian process Product √
Moreover, a computational workflow including automated shape syn­ [45] regression model structure
thesis and structure design methods is established. For the shape syn­ Simulation data
thesis, existing designs selected based on user preferences are utilized
for creating implicit generative models. For structural design, user-
The grammar-based development is a logical argumentation process
product interaction data gathered by the Internet of Things (IoT) are
based on shapes and their operations. Grammar is set manually by de­
used as inputs for controlling the spatial distributions of heterogeneous
signers considering the design constraints and objectives. Typical ap­
lattice structures. Finally, the proposed concept and workflow are
plications are furniture [15] and architecture [16] design. Inspired by
demonstrated with a bike saddle design that has a personalized shape
shape grammar, the parametric model-based generative design creates
and inner structures to be manufactured with additive manufacturing.
products by manually defining product parameters and their relation­
The remainder of the paper is organized as follows: the next section
ships based on existing design knowledge [17,18].
reviews the related work of generative design and data-driven methods
For implicit methods, deep generative models (DGMs) are used to
in product design. Section 3 proposes a data-driven generative design
create new product designs through shape synthesis with various
framework and its computational workflow. Section 4 introduces a case
methodologies [6], such as generative adversarial networks (GAN) [19],
study of bike saddles to demonstrate the effectiveness of the proposed
variational autoencoders (VAE) [20], and reinforcement learning (RL)
framework. Finally, conclusions and future work are presented in Sec­
[21]. GANs and VAEs synthesize high-quality designs by approximate
tion 5.
complicated, high-dimensional probability distributions based on a
large dataset. GANs have been used to generate a wide range of different
2. Related works
product shapes and support the designers to meet design requirements,
such as aircraft [22] and automobiles [23]. The power of VAE was
2.1. Generative design
shown in 3D CAD model retrieval [24] and generation [25]. Besides,
explicit and implicit methods can be jointly used in product shape
The generative design process consists of a design schema, a gener­
synthesis. For example, RL has been integrated for complementing the
ative model, and an evaluation process [7], which aims to create product
generative capacity of shape grammars [26]. Dworschak et al. mapped
variations to satisfy different user needs automatically. Both explicit and
the design requirements and parametric model of a bike part into a
implicit generative design methods exist. Explicit generative design can
learning environment and explored the design space automatically by
be further divided into grammar-based and parametric model-based.

2
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

Fig. 1. Data-driven generative design framework for mass customization.

using RL [27]. Unlike GANs and VAEs, RL learns from the trial-and-error analysis, user reviews and purchase data are commonly used for
interactions between the agent and the environment instead of given defining user needs [13,31–33]. Meanwhile, sensor data from the usage
datasets. DGMs have demonstrated their powerfulness in improving phase is utilized to elicit potential user requirements by using a graph-
product performance and variation with accumulated design knowledge based framework[34]. For conceptual design, the central task of
[28,29]. In these DGMs, design variables and constraints are defined applying data-driven methods is to facilitate designers’ idea generation
based on data and user inputs. However, the design objective in DGMs is by extracting concepts and discovering their associations from data
still set manually. [35–39]. Besides, the evaluation of concepts is also important for pur­
The above designer-driven generative design approaches have suing rapid iteration of the design in a mass customization context. To
significantly increased the product variety. However, designers still relieve the lag between user requirement changes and the design con­
need to manually set up generative models and define the objective, cepts, user behavior and perceptual data have been used to evaluate
variables, and constraints of design, which is impractical for products concepts with less human intervention through a context-aware
that require personalization. Meanwhile, other well-known generative approach [40]. The automation level in both stages is enhanced with
design methods, like topology optimization, also have their limitations data-driven methods.
for mass customization. For instance, the design objective is pre-defined However, applications of data-driven methods for detail design,
and cannot include in-service information after design. This study ex­ including geometry, structure, and material design, are still worth
plores the use of existing designs, user feedback, and in-service data as a exploring. The application in this stage mainly focuses on linking
solution to resolve the above issues within the designer-driven genera­ product form to the affective needs of customers [42–44]. Though these
tive design. previous researches achieved higher user emotional satisfaction by
using regression models and heuristic algorithms. The user needs in this
stage are more than just visual emotional satisfaction, but also other
2.2. Data-driven product design
physical sensations, such as tactile comfort. Such needs that are not
commonly expressed in writing or orally are difficult to be measured and
Data-driven product design is a decision-making process that trans­
extracted based on user-end data but are feasible by analyzing service-
forms a set of functional requirements into specific product designs
end data [14]. Service-end data can be collected through simulation
based on various data [30]. Previous studies have applied data-driven
[45] or IoT [11,14]. Besides, by incorporating additive manufacturing to
design methods to different stages of product design for multiple pur­
fabricate complex product structures, the product performance is also
poses, as summarized in Table 1. It is found that significant work has
influenced by process planning [9,46]. Thus, data from the production-
been done in requirement analysis and conceptual design. In these
end is valuable for customized production. Considering the manufac­
design stages, user-related information forms the basis of the data-
turability, data from the user-end, service-end, and production-end
driven product design for mass customization. For requirement

3
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

should be comprehensively analyzed. However, detail design for mass data in conceptual design can be analyzed by text mining and other NLP
customization lacks a framework that can connect different design methods to synthesize new ideas, such as product descriptions, idea
stages and enable the share of data from different ends across the entire data, patent documents, etc. Besides, visual data like patent images can
design flow. The lack of framework limits detail design for mass cus­ also flow into the framework according to a convolutional neural
tomization in two aspects: automation and personalization levels. On network for inspiring designers. To combine customized design and
the one hand, detail design needs a higher design automation level to flexible manufacturing technologies, the design model generated from
improve its capability in efficiently generating product variants. This the detail design computational workflow deliver the data of geometry
means the data stream from conceptual design should be connected with and structure to manufacture. For securing the manufacturability of
the detail design stage in an efficient manner. Meanwhile, as many customized design, material and process data can also be utilized to
products become more intelligent, IoT can be used to obtain these data drive the process planning. For instance, the printing thickness, speed,
about user-product interaction for evaluating design performance. These and direction are constraints of the additively manufactured product. By
sensor measurements that contain in-service evaluations on design incorporating these data into a data-driven process, designers expect to
[11,23] can be fed back into the early design stage, which can support obtain stable manufacturability, higher performance, and accurate cost
designer with valuable insights to facilitate the new generation product of tailor products in product development. The rest of this section fo­
[41].However, there are no existing tools for utilizing data to improve cuses on the detail design of product geometry to exemplify the imple­
the personalized level of detail design. To fill this gap, the study pro­ mentation of the proposed framework.
poses a data-driven method to utilize multiple types of data for
improving the automation and personalization levels. Compared to 3.1. Customized shape synthesis
designer-driven for mass customization, the design objectives, con­
straints, and variables are determined based on data instead of relying 3.1.1. Data collection
on the designer. In contrast, generative design provides an effective way Design of existing products carries valuable information about the
to explore high-dimensional design space, rather than generate an user preference. The shape synthesis technique applied in this compu­
optimal design for consumers. Various types of data flow in the whole tational workflow attempts to mine these data for accurately trans­
process of product development make this framework very extensible ferring accumulated information about user needs into the creation of
when customizing products. new product geometry. Such a shape synthesis process contains two
main steps: data collection and shape generator creation.
3. Data-driven generative design framework Typical data of existing products include two-dimensional (2D) im­
ages, user reviews, product descriptions, etc. Tools, such as web crawl­
This section proposes a data-driven generative design framework ing, are often used for collecting data. To improve the quality of the
and computational workflow, as shown in Fig. 1. In this framework, training dataset, the images are pre-processed by the designer. Images
multiple types of data, including data from the product, user, and service taken from different angles were discarded from the training set
ends, are used in the detail design stage, i.e., shape synthesis and inner manually and irrelevant parts were removed by using Photoshop. Then
structure generation, through a data-driven approach. This approach three steps were taken in image processing. First, the product in each
aims to complement the designer-driven generative design currently image is detected and boxed using OpenCV. Second, the geometry center
used in detail designs for mass customization. and selected geometry (e.g., saddle nose) were automatically labeled.
In detail design for mass customization, the geometry, structure, and Meanwhile, each image is aligned based on the geometric center and the
material of products are determined based on the individual user needs. selected geometry feature. Third, the images are normalized to the same
This study assumes that users’ needs for shape design can be extracted size based on the architecture of the generative model.
from questionnaires and surveys, while their needs for structure design In data labeling, all possible features of a product are collected and
can be indicated by user-product interaction, which is measured by denoted as a vector f= (f1, f2, …, fn), where n is the total number of
sensors [8]. As shown in Fig. 1, the automated level can be classified into features being considered. Based on the product’s review and de­
three levels: the high level means the element can be automatically scriptions, the value of fdsg_i for the existing product design i is deter­
determined by data; the weak level means the element is determined mined by humans based on its product description and review
manually by the designer; and the moderate level is in between them. In information. Nouns (e.g., comfort), verbs (e.g., comforted), adjectives
the proposed approach, data from different design stages are utilized in (e.g., comfortable), and adverbs (e.g., comfortably) are used to extract
the detail design for various purposes. Firstly, the early design stages’ product features from reviews. The set of product reviews is transformed
information about the user’s needs is fed into detail design through data into a matrix of binary values (0 or 1) through a two-step judgment.
instead of the designer’s direct decisions. For instance, users’ prefer­ Preliminary judgment detects the related words of features, the binary
ences can be extracted from online reviews with data mining and natural value is temporarily set to 1 if the reviews process the related words,
language processing methods. Moreover, existing designs relevant to the otherwise to 0. Secondary judgment defines the evaluation of each
user’s requirements are automatically collected as datasets to create the feature through sentiment analysis. Valence Aware Dictionary for
generative model for detail designs of geometry. Additionally, evalua­ Sentiment Reasoning VADER is a model used for text sentiment analysis
tion information from the service stage, e.g., user-product interactions, that is sensitive to both polarity (positive/negative) and intensity
is used for setting the objective of generative design. These data can be (strength) of emotion [47]. VADER sentiment analyzer in Python is used
collected from simulations or experiments using IoT techniques. Both to construct the sentiment lexicon, which has a [-1, +1] scale, − 1 is the
mapping-based and optimization-based approaches can be used to most negative, and + 1 is the most positive. Sentiment-related infor­
devise a solution satisfying the objectives. Such data-driven methods mation is extracted and analyzed through original online reviews. We
automate the setting of design objectives, variables, and constraints. For assume that the customer will express his emotion regarding a feature
instance, the design objective in a insole design case can be automati­ within a window of words around the feature word. The information is
cally configured based on the measurements of each user, e.g., physical sent to VADER, and a compound score is returned. If the compound
data and usage histories. Meanwhile, the generative models are auto­ score >= 0.5, this feature is considered positive sentiment, and the value
matically created from the existing designs. This is very different from of fdsg_i is kept at 1, otherwise the value is adjusted to 0. The sentiment
the designer-driven method, where the insole shapes are defined by lexicon can be updated by the designer based on product features and
humans through computer-aided design interfaces. emotional expressions for different fields. Vectors fdsg_i for all existing
Besides, this framework seeks to expand the data-driven method to products is collected and denoted as F= (fdsg_1, fdsg_2, …fdsg_i…, fdsg_q), q
other stages in product development for mass customization. Textual is the total number of existing products. Similarly, the user needs for

4
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

Fig. 2. GAN based shape synthesis based on existing product designs.

features of a product are also described by f and denoted as fusr_j for the distribution generated by the generator to be close to the real data dis­
user j. The value of each element in fusr_j is set based on the questionnaire tribution [19]. The network architecture of GAN based shape synthesis
filled. A customized training set is expected to be built with existing for existing product design is shown in Fig. 2. The generator aims to
products that fit user needs. The idea of a recommendation system is capture the data distribution of selected product designs in the training
drawn to construct a customized training set. There are multiple ways to dataset and make the generated data distribution close to the existing
obtain a matching score between textual data. Cosine similarity is the designs. The discriminator attempts to identify whether the input is a
most common approach [48], which defines as the cosine angle between user-preferred design. The output of GAN is customized product designs
the two vectors when text is represented in the form of term vectors. Its that express user preference. The generator and discriminator are
pros include its independence of document length, fitting both contin­ trained for a min–max game as [19]:
uous and categorical variables[49,50]. While the cons of it are not
minmaxV(D, G) = Ex pdata (x)[logD(x) ] + Ez pz (z)[log(1 − D(G(z) ))] (2)
efficient with nominal data. For identifying the most satisfying designs G D

from existing products for a specific user, the cosine similarity is where x is the variable of existing designs, and G is the generator’s
computed between the user need fusr_j and the existing design fdsg_i. The differentiable function with parameters θg. The generator G maps the
cosine similarity is defined as: latent variable z to the space of existing designs by adjusting θg.
( ) 〈f usr j • f dsg i 〉 Inputs from users are needed after a set of customized designs are
cosinesimilarity f usr j , f dsg i = √̅̅̅̅̅̅̅̅̅̅̅√̅̅̅̅̅̅̅̅̅̅̅̅ (1) generated from GAN. For each user-preferred features within fusr_j, user
f usr j 2 f dsg i 2
identifies the top s most satisfying designs from all generated designs.
The cosine similarity is in a range of 0 to 1. If the cosine similarity is The location of a selected design in the two-dimensional latent space is
closer to 1, then this existing product design strongly matches the user’s denoted as pusr_t¼ (pLv1_t,pLv2_t), where t = 1,2, …, s. The geometry
needs and vice versa. For instance, three features are considered for the center of all selected designs in the latent space for the ith user-preferred
design of shoes and are denoted as f = (f1, f2, f3), where f1: respon­ feature is:
siveness, f2: lightweight, and f3 = ankle stability. If the ith shoe only has 1∑ s

the first two features, this design’s feature vector is fdsg_i= (1,1,0). vusr i = p (3)
s t=1 usrt
Meanwhile, if the jth user selects the last two features, the feature vector
of user’s needs is fusr_j= (0,1,1). The cosine similarity(fusr_j,fdsg_i) be­ Finally, the user can generate a single customized design by using
tween the ith design and the jth user is 0.5. The cosine similarity is latent variable arithmetic as follows:
computed for all existing product designs F and sorted in descending
∑
m
order. The top k existing products are selected as the training dataset for vcust = wi *vusr i (4)
creating a customer specific generative design model. i=1

where vcust is the location of the final customized design within the
3.1.2. Modeling shape generator latent space, m is the number nonzero elements in fusr_j, and wi is the
The shape generator for a customized product requires inputs, e.g., weight of ith function. The value of wi is determined by the user and thus
the existing designs, as the initial design information. Such information allows the exploration among various customized designs. The geometry
can be synthesized through deep generative models, e.g., generative of the 3D model is partially set by the designer according to parametric
adversarial networks. The output of these models is a set of designs for modeling in CAD modeling software, such as Grasshopper. In Grass­
the user to select. hopper, the 3D model is represented by boundary representation (Brep),
GAN provides an attractive way taken from game theory to train a which consists of a set of relationships between surfaces, curves, and
generative model by framing the problem as a supervised learning points. These basic elements of Brep can be built with a non-uniform
problem with two sub-models: a generator and a discriminator. GAN is rational b spline (NURBS). To create a basic 3D model, the designer
designed to infer the data generating a distribution by making the model firstly set design parameters for NURBS, such as the number of control

5
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

Fig. 3. Flow diagram of customized bike saddle design process.

points and the position of control points. With the basic 3D model, the 3.2.2. Modeling structure generator
geometry of the top view or side view of the product is automatically The inner structure of the product is spatially divided into several
determined by the generated designs from GAN. While the other ge­ volumes based on their functions. The division of function is determined
ometry parameters are fixed, such as thickness and curvature. by the product usage data. For example, mapping pressure distribution
on shoes to functional design volumes considering the need for user
3.2. Customized structure design comfort. The structure, e.g., the lattice cell size of individual functional
volumes, is automatically generated by algorithms. In the next section, a
3.2.1. Data collection bike saddle design case is studied to exemplify the implementation of the
The lack of objective user insight causes the limitation of designer- proposed framework.
driven generative design to customize a structure that fits user needs
appropriately. These user needs in specific usage contexts are usually 4. Case study: Bike saddle
expressed in the data of user-product interaction. Traditionally, such
data can be obtained through designer observations and co-design. The saddle is a crucial component of a bike that continuously in­
However, human approaches to seeking user-product interaction are teracts with the rider during cycling. The case study aims to customize
costly and infeasible in some contexts [14], such as running or cycling. both the geometry and inner structure of a bike saddle to fit users’
With the development of intelligent products, IoT techniques have personalized needs. As shown in Fig. 3, the training dataset was built
become a practical way to gather product usage data under controlled through data collection of existing designs using web data extraction
experiments or simulations. For example, force sensors can be used to and preprocessing techniques. The user needs were identified through
characterize pressure distributions between the user and bike saddles or questionaries and utilized for selecting related images to construct the
running shoes. The force sensor data contain user-product interaction dataset. A GAN-based generative model was trained with the selected
information and provide objective input of user insight. dataset to generate a set of shapes that fits the user’s preference. In shape
design, users’ direct participation using GUI gives users more freedom to

6
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

Fig. 4. Data collection processing of exiting saddles.

express their preferences directly, which reduces the cost of communi­ model. The motivation for choosing the pressure distribution as inner
cation between designers and users due to bias. To improve user satis­ structure design input is due to its direct link with riding comfort [51].
faction, the proposed workflow needs objective data of user-product
interaction for analyzing the difference between users. This user- 4.1. Collection of existing saddle designs
involved step is indispensable for achieving customization goals.
Compared to getting users’ feedback through survey or interview, this In this case study, images were collected from Amazon.com and
step provides an effective way to avoid designer bias. Meanwhile, the Bike24.com which fit the need for good training samples of GAN.
pressure distribution of users was gathered from force sensors. Then Because saddle images from these websites were from the top view with
with the input of pressure distribution of each test user, personalized 3D a white background and the image ratio is 1:1 (a square image). Then,
inner structures were automatically generated based on the basic 3D the orientation of the saddle was aligned based on the geometric center

Fig. 5. Saddle shape customization using a GAN based shape synthesis method.

7
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

Fig. 6. The graphical user interface for customized design selections.

and the saddle nose towards the positive direction of the x-axis. Besides, 4.2. Shape synthesis of saddles
all images were normalized to a size of 128 × 128 pixels. This study
invited three users to participate. As shown in Fig. 5, the shape of saddles was synthesized with GAN
The designer’s expertise is collected from bike design magazines, based method. Given the limited size of the dataset, this study utilizes
rider’s forums, and product descriptions on the website. By synthesizing the data-efficient GAN (DEGAN) from reference [52]. Each of the 100
this information, the designer determined ten features of a bike saddle. A training samples was represented by an image tensor (3, 128, 128). The
questionnaire in Appendix A was used to collect the user’s need and latent variable(z) was chosen as a 100-dimensional normally distributed
preference features of a bike saddle. These features reflect user prefer­ random vector. DEGAN has proven that the quality of generated output
ences through aesthetic, functional, and comfort perspectives. The top is acceptable based on 100 training samples[52]. The architecture of the
three features that the user most wanted are denoted as 1, otherwise are GAN models was presented in Appendix B. We used Adam Optimizer,
denoted as 0. As shown in Fig. 4, ten features of a bike saddle are and the learning rate was 0.0002. After training for 300 epochs using the
collected and denoted as a vector f= (f1, f2, …, f10), where the detail of fi GPU offered by Colab, the quality of output designs was acceptable. The
is shown in Appendix A. If the product possesses the feature fn, the value size of the output image was 128 × 128 pixels.
is set to 1 and otherwise to 0. Vectors fdsg_i for all existing products is This customized design set generated from GAN contains the user-
expressed as F= (fdsg_1, fdsg_2, …fdsg_i…, fdsg_600). Then, the user needs on preferred features. As shown in Fig. 5, user 1 chose the three gener­
the product’s features are also described by f and denoted as fusr_j for the ated designs that most satisfied each functions and identified their
user j (j = 1,2, and 3). The value of each element in fusr_j is set based on location in the latent space as pusr_t. The vusr of each feature was
the questionnaire filled (Appendix A). For identifying the most satisfying calculated using Equation (3), where vusr_1:center cutout, vusr_2:support,
designs from existing products for a specific user, the cosine similarity is vusr_3: race style. As shown in Fig. 6, the user determined the weight
computed between the user need fusr_j and the existing design fdsg_i by factors w1 to w3 and simultaneously obtained a customized saddle
using Equation (1). The top 100 related existing saddles were selected to design from the GAN based shape generator. The GAN-based shape
build the training set. synthesis method helps people quickly explore the design space of user
preference saddles. To enhance user centralization, a GUI was provided

Fig. 7. Saddle pressure experiment setup: (a) the distribution of force sensors; and (b) test setup.

8
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

Fig. 8. Customized saddle inner structures for different users are generated based on the pressure distribution measurements.

to the user. Users can determine weights according to their preferences, In this case study, design efficiency and user satisfaction were
and users can quickly get visual feedback on the customized saddle. improved. For design efficiency, two main aspects are mentioned. First,
user preference and product features were automatically linked by using
data-driven methods; Second, the process of creating product variants
4.3. Inner structure design for different users is decreased by utilizing GAN. In this case, after 10 h
of training, the GAN-based model is available to generate new designs
The inner structure of the bike saddle was designed based on the for the different users in a few seconds. For user satisfaction, two main
pressure distribution characterized by an array of sensors. These sensors aspects are mentioned. First, GUI gives users a very simple way to
were evenly distributed on the saddle, as shown in Fig. 7. The force participate in the customization of shapes, which increased user
sensor (40 DF9-16, Learnstar, Suzhou, China) is a resistive sensor fixed engagement and freedom in customization. Second, user-product
to the saddle with adhesive tapes. The sampling rate is set to 10 Hz. interaction data provide a chance to generate designs with better per­
Three users (User 1: bodyweight: 60 kg, height: 175 cm; User 2: body­ formance by reducing designer bias in the structure design.
weight: 60 kg, height: 165 cm; User 3: bodyweight: 70 kg, height: 180
cm) participated the measurement. During the test, the height from the 5. Conclusions
saddle to the rear-wheel fixed center was set to 60 cm, and the height
from the bike grip to the rear-wheel fixed center was set to 70 cm. The This paper has proposed a data-driven generative design framework
saddle angle is set to 0. The users were required to maintain a torso angle for mass customization and a computational workflow for the detail
of 60 degrees from the horizontal (±1 degree). Before the test began, design stage in customized product development by using the data-
participants first adjusted their riding position on the seat to their driven method to incorporate multiple types of data into the genera­
comfortable position. During the test, the participants sit on the bike tive design process. This framework shows a way to bridge the gap be­
saddle for 5 min. tween designer-driven generative design and to personalized needs of
The measured pressure distribution is mapped into the design of users in two aspects: automation level and performance. For improving
heterogeneous lattice structures for bike saddles. The pressure mea­ the automation level, the computational workflow delivers the data
surements of three users are shown in Fig. 8. The inner structure’s from the early stages of product development: requirement analysis and
customization is based on the pressure distribution, while the same conceptual design to detail design. The design variables, constraints, and
material and shape of the saddle were used in the experiment. These objectives of product geometry and structure are automatically set by
pressure distributions are used to design the Voronoi-based lattice the data-driven method instead of manually by the designer. For product
structures of a bike saddle. A mapping function from previous research performance, usage data gathered by using IoT techniques provides
[53] is used to proportionally map the pressure level into the size of the objective user insight in the context of product-user interaction. The
lattice cell. Finally, three customized saddle designs that incorporate proposed framework is demonstrated through a bike saddle design with
information about individual users are generated, which can be addi­ a personalized shape and inner structures to be manufactured with
tively manufactured.

9
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

Table 2
User needs and preference feature of bike saddle.
Features ID Features of saddle

1 Cushion
2 Center cutout
3 Support
4 Comfort
5 Breathability
6 Nose less
7 Mountain style
8 Race style
9 Durability
10 Lightweight

Fig. 10. Discriminator architecture.

additive manufacturing.
Although this work presented an approach to automatically gener­
ating product variation by incorporating multiple types of data, it also
brings other challenges to data-driven generative design for mass cus­
tomization. The evaluation tool is needed to evaluate generated product
performance for bridging the gap between visual aspects of design and
Fig. 9. Generator architecture.

10
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

its level of personalization. To generate a customized structure driven by [13] A. Burnap, Y. Pan, Y. Liu, Y. Ren, H. Lee, R. Gonzalez, P.Y. Papalambros, Improving
design preference prediction accuracy using feature learning, J. Mech. Des. 138 (7)
real-time data of user-product interaction, a more comprehensive and
(2016), 071404.
more profound insight into measuring the correlation of user needs with [14] D. Ghosh, A. Olewnik, K. Lewis, Application of Feature-Learning Methods Toward
this product usage data is needed. The limitations of this study also Product Usage Context Identification and Comfort Prediction, J. Comput. Inf. Sci.
include that the design of the internal structure does not consider the Eng. 18 (1) (2018).
[15] M. Barros, J.P. Duarte, B. Chaparro, “Integrated generative design tools for the
interaction of materials and geometrical factors at the same time. Sys­ mass customization of furniture, Design Computing and Cognition’12 Springer
tematic parallel consideration of the impact of these elements on user- (2014) 285–300.
product interaction is something that future research needs to explore. [16] N. Gu, P.A. Behbahani, Shape Grammars: A Key Generative Design Algorithm, in:
B. Sriraman (Ed.), Handbook of the Mathematics of the Arts and Sciences, Springer
Besides, future work will extend the current procedure of measurements International Publishing, Cham, 2018, pp. 1–21.
for other scenarios including outdoor conditions. [17] R.H. Kazi, T. Grossman, H. Cheong, A. Hashemi, G.W. Fitzmaurice, “DreamSketch:
Early Stage 3D Design Explorations with Sketching and Generative Design,” Proc.
UIST, pp. 401-414.
Declaration of Competing Interest [18] E. Gunpinar, S. Gunpinar, A shape sampling technique via particle tracing for CAD
models, Graph. Models 96 (2018) 11–29.
The authors declare the following financial interests/personal re­ [19] I. Goodfellow, “Nips 2016 tutorial: Generative adversarial networks,” arXiv
preprint arXiv:1701.00160. 2016.
lationships which may be considered as potential competing interests: Yi [20] D.P. Kingma, M. Welling, “Auto-encoding variational bayes,” arXiv preprint arXiv:
Xiong reports financial support was provided by the National Natural 1312.6114, 2013.
Science Foundation of China. Yi Xiong reports financial support was [21] L.P. Kaelbling, M.L. Littman, A.W. Moore, Reinforcement learning: A survey,
Journal of artificial intelligence research 4 (1996) 237–285.
provided by the Guangdong Basic and Applied Basic Research Founda­ [22] D. Shu, J. Cunningham, G. Stump, S.W. Miller, M.A. Yukish, T.W. Simpson, C.
tion. Yi Xiong reports financial support was provided by the Shenzhen S. Tucker, 3D Design Using Generative Adversarial Networks and Physics-Based
Science and Technology Innovation Committee. Validation, J. Mech. Des. 142 (7) (2020).
[23] S. Oh, Y. Jung, I. Lee, N. Kang, “Design Automation by Integrating Generative
Adversarial Networks and Topology Optimization,” Proc. ASME 2018 International
Data availability Design Engineering Technical Conferences and Computers and Information in
Engineering ConferenceV02AT03A008.
[24] F. Qin, S. Qiu, S. Gao, J. Bai, 3D CAD model retrieval based on sketch and
Data will be made available on request.
unsupervised variational autoencoder, Adv. Eng. Inf. 51 (2022), 101427.
[25] W. Zhang, Z. Yang, H. Jiang, S. Nigam, S. Yamakawa, T. Furuhata, K. Shimada, L.B.
Acknowledgments Kara, “3D shape synthesis for conceptual design and optimization using variational
autoencoders,” Proc. International Design Engineering Technical Conferences and
Computers and Information in Engineering Conference, American Society of
X.Y. acknowledges the National Natural Science Foundation of China Mechanical Engineers, p. V02AT03A017.
[Grant No. 52105261] and the Guangdong Basic and Applied Basic [26] M. Ruiz-Montiel, J. Boned, J. Gavilanes, E. Jiménez, L. Mandow, J.-L. PéRez-De-La-
Research Foundation [Grant No. 2022A1515010316]. Cruz, Design with shape grammars and reinforcement learning, Adv. Eng. Inf. 27
(2) (2013) 230–245.
[27] F. Dworschak, S. Dietze, M. Wittmann, B. Schleich, S. Wartzack, Reinforcement
Appendix A. Questionnaire for identifying the user needs for the Learning for Engineering Design Automation, Adv. Eng. Inf. 52 (2022), 101612.
bike saddle design [28] W. Chen, F. Ahmed, Padgan: Learning to generate high-quality novel designs,
J. Mech. Des. 143 (3) (2021).
[29] W. Chen, K. Chiu, M. Fuge, “Aerodynamic design optimization and shape
Please select up to 3 features you consider essential for a customized exploration using generative adversarial networks,” Proc. AIAA Scitech 2019
bike saddle in Table 2: Forum, p. 2351.
[30] Y. Feng, Y. Zhao, H. Zheng, Z. Li, J. Tan, Data-driven product design toward
intelligent manufacturing: A review, Int. J. Adv. Rob. Syst. 17 (2) (2020),
Appendix B. Architecture of GAN models 1729881420911257.
[31] J. Joung, H.M. Kim, Approach for Importance-Performance Analysis of Product
Attributes From Online Reviews, J. Mech. Des. 143 (8) (2021), 081705.
See Figs. 9 and 10. [32] F. Zhou, R. Jianxin Jiao, J.S. Linsey, Latent Customer Needs Elicitation by Use Case
Analogical Reasoning From Sentiment Analysis of Online Product Reviews,
References J. Mech. Des. 137 (7) (2015).
[33] K.-Y. Lin, C.-F. Chien, R. Kerh, UNISON framework of data-driven innovation for
extracting user experience of product design of wearable devices, Comput. Ind.
[1] B.J. Pine, S. Davis, “Mass customization: the new frontier in business competition.”
Eng. 99 (2016) 487–502.
(1993).
[34] Z. Wang, C.-H. Chen, P. Zheng, X. Li, L.P. Khoo, A novel data-driven graph-based
[2] J. Jiao, T.W. Simpson, Z. Siddique, Product family design and platform-based
requirement elicitation framework in the smart product-service system context,
product development: a state-of-the-art review,“ J. Intell. Manuf, 18(1) ((2007)) 5-
Adv. Eng. Inf. 42 (2019), 100983.
29.
[35] K. Lewis, D. Van Horn, “Design analytics in consumer product design: A simulated
[3] M.M. Tseng, J. Jiao, M.E. Merchant, Design for mass customization, CIRP Ann. 45
study,” Proc. International Design Engineering Technical Conferences and
(1) (1996) 153–156.
Computers and Information in Engineering Conference, American Society of
[4] K. Ulrich, The role of product architecture in the manufacturing firm, Res. Policy
Mechanical Engineers, p. V03BT03A003.
24 (3) (1995) 419–440.
[36] Y. He, B. Camburn, H. Liu, J. Luo, M. Yang, K. Wood, Mining and Representing the
[5] T.W. Simpson, J.R. Maier, F. Mistree, Product platform design: method and
Concept Space of Existing Ideas for Directed Ideation, J. Mech. Des. 141 (12)
application, Res. Eng. Des. 13 (1) (2001) 2–22.
(2019).
[6] L. Regenwetter, A.H. Nobari, F. Ahmed, Deep generative models in engineering
[37] S. Jiang, J. Luo, G. Ruiz-Pava, J. Hu, C.L. Magee, Deriving design feature vectors
design: A review, J. Mech. Des. 144 (7) (2022), 071704.
for patent images using convolutional neural networks, J. Mech. Des. 143 (6)
[7] S. Krish, A practical generative design method, Comput. Aided Des. 43 (1) (2011)
(2021), 061405.
88–100.
[38] Q. Liu, K. Wang, Y. Li, Y. Liu, Data-Driven Concept Network for Inspiring
[8] D. Ghosh, A. Olewnik, K. Lewis, J. Kim, A. Lakshmanan, Cyber-Empathic Design: A
Designers’ Idea Generation, J. Comput. Inf. Sci. Eng. 20 (3) (2020).
Data-Driven Framework for Product Design, J. Mech. Des. 139 (9) (2017).
[39] F. Shi, L. Chen, J. Han, P. Childs, A Data-Driven Text Mining and Semantic
[9] Y. Xiong, P.L.T. Duong, D. Wang, S.-I. Park, Q. Ge, N. Raghavan, D.W. Rosen, Data-
Network Analysis for Design Information Retrieval, J. Mech. Des. 139 (11) (2017).
driven design space exploration and exploitation for design for additive
[40] Z. Wang, C.-H. Chen, X. Li, P. Zheng, L.P. Khoo, A context-aware concept
manufacturing, J. Mech. Des. 141 (10) (2019).
evaluation approach based on user experiences for smart product-service systems
[10] C. Yuan, T. Marion, M. Moghaddam, Leveraging End-User Data for Enhanced
design iteration, Adv. Eng. Inf. 50 (2021), 101394.
Design Concept Evaluation: A Multimodal Deep Regression Model, J. Mech. Des.
[41] L. Bu, C.-H. Chen, G. Zhang, B. Liu, G. Dong, X. Yuan, A hybrid intelligence
144 (2) (2021).
approach for sustainable service innovation of smart and connected product: A case
[11] M. Nourbakhsh, N. Morris, M. Bergin, F. Iorio, D. Grandi, “Embedded sensors and
study, Adv. Eng. Inf. 46 (2020), 101163.
feedback loops for iterative improvement in design synthesis for additive
[42] Y.-C. Lee, W. Han, Soccer shoe recommendation system based on multitechnology
manufacturing,” Proc. International Design Engineering Technical Conferences and
integration for digital transformation, Adv. Eng. Inf. 51 (2022), 101457.
Computers and Information in Engineering Conference, American Society of
[43] M.-D. Shieh, Y. Li, C.-C. Yang, Comparison of multi-objective evolutionary
Mechanical Engineers, p. V01AT02A031.
algorithms in hybrid Kansei engineering system for product form design, Adv. Eng.
[12] C. Zhang, D. Chen, F. Tao, A. Liu, Data driven smart customization, Procedia CIRP
Inf. 36 (2018) 31–42.
81 (2019) 564–569.

11
Z. Jiang et al. Advanced Engineering Informatics 54 (2022) 101786

[44] Z. Wang, W. Liu, M. Yang, Data-driven multi-objective affective product design [49] A. Huang, “Similarity measures for text document clustering,” Proc. Proceedings of
integrating three-dimensional form and color, Neural Comput. Appl. (2022) 1–27. the sixth new zealand computer science research student conference
[45] Y. Tang, G. Dong, Y. Xiong, Q. Wang, Data-driven design of customized porous (NZCSRSC2008), Christchurch, New Zealand, pp. 9-56.
lattice sole fabricated by additive manufacturing, Procedia Manuf. 53 (2021) [50] M. Goswami, A. Babu, B. Purkayastha, A comparative analysis of similarity
318–326. measures to find coherent documents, Appl Sci Manag 8 (11) (2018) 786–797.
[46] Y. Tang, G. Dong, Q. Zhou, Y.F. Zhao, Lattice structure design and optimization [51] A.S. Larsen, F.G. Larsen, F.F. Sørensen, M. Hedegaard, N. Støttrup, E.A. Hansen,
with additive manufacturing constraints, IEEE Trans. Autom. Sci. Eng. 15 (4) P. Madeleine, The effect of saddle nose width and cutout on saddle pressure
(2017) 1546–1562. distribution and perceived discomfort in women during ergometer cycling, Appl.
[47] C. Hutto, E. Gilbert, “Vader: A parsimonious rule-based model for sentiment Ergon. 70 (2018) 175–181.
analysis of social media text,” Proc. Proceedings of the international AAAI [52] S. Zhao, Z. Liu, J. Lin, J.-Y. Zhu, S. Han, Differentiable augmentation for data-
conference on web and social media, pp. 216-225. efficient gan training, Advances in Neural Information Processing Systems 33
[48] M. Daibo, “Toroidal vector-potential transformer,” Proc. 2017 Eleventh (2020) 7559–7570.
International Conference on Sensing Technology (ICST), IEEE, pp. 1-4. [53] Y. Tang, G. Dong, Y.F. Zhao, A hybrid geometric modeling method for lattice
structures fabricated by additive manufacturing, The International Journal of
Advanced Manufacturing Technology 102 (9) (2019) 4011–4030.

12