See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/340428995

Smart additive manufacturing: Current artificial intelligence-enabled

methods and future perspectives

Article  in  Science China Technological Sciences · March 2020

DOI: 10.1007/s11431-020-1581-2

CITATIONS READS

18 1,266

5 authors, including:

Yuanbin Wang Pai Zheng

Zhejiang University The Hong Kong Polytechnic University
16 PUBLICATIONS   393 CITATIONS    122 PUBLICATIONS   2,271 CITATIONS   

SEE PROFILE SEE PROFILE

Tao Peng Jun Zou

Zhejiang University South China Agricultural University
81 PUBLICATIONS   1,255 CITATIONS    90 PUBLICATIONS   895 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Development of an energy-efficient machining system View project

Cyber-Physical Machine Tools for Smart Manufacturing View project

All content following this page was uploaded by Yuanbin Wang on 04 April 2020.

The user has requested enhancement of the downloaded file.

Smart additive manufacturing: current artificial intelligence-enabled methods and future perspectives
Wang Yuanbin, Zheng Pai, Peng Tao, Yang Huayong and Zou Jun
Citation: SCIENCE CHINA Technological Sciences; doi: 10.1007/s11431-020-1581-2

View online: http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

Published by the Science China Press

Articles you may be interested in

Cryopreservation of mammalian oocytes and embryos: current problems and future perspectives
SCIENCE CHINA Life Sciences 57, 903 (2014);

Intelligent and connected vehicles: Current status and future perspectives

SCIENCE CHINA Technological Sciences 61, 1446 (2018);

Cryo-electron microscopy for structural biology: current status and future perspectives
SCIENCE CHINA Life Sciences 58, 750 (2015);

Artificial intelligence in drug design

SCIENCE CHINA Life Sciences 61, 1191 (2018);

Artificial intelligence in computer-aided diagnosis of abdomen diseases

SCIENCE CHINA Life Sciences 62, 1396 (2019);
 SCIENCE CHINA
Technological Sciences
• Article • Draft

doi: Null

Smart additive manufacturing: current artificial

ed
intelligence-enabled methods and future perspectives

WANG Yuanbin1, ZHENG Pai2, PENG, Tao1, YANG Huayong1 & ZOU Jun1,*

1 State Key Laboratory of Fluid Power and Mechatronic Systems, Zhejiang University, Hangzhou 310027, China;
2 Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, HKSAR, China
pt Received February 10

Additive manufacturing (AM) has been increasingly used in production. Because of its rapid growth, the efficiency
and robustness of AM-based product development processes should be improved. Artificial intelligence (AI) is a
ce
powerful tool that has outperformed humans in numerous complex tasks. AI-enabled intelligent agents can reduce
the workforce required to scale up AM production and achieve higher resource utilization efficiency. This study
provides an introduction of AI techniques. Then, the current development of AI-enabled AM product development is
investigated. Existing intelligent agents are used for problems in product design, process design and production
stages. Based on the review, current research gaps and future research directions are identified. To guide future
development of more efficient and comprehensive intelligent agents, a smart AM framework based on cloud-edge
computing is proposed. Global consideration can be realized in the cloud environment, and a fast response can be
Ac

achieved at the edge nodes.

Keywords: Additive manufacturing, artificial intelligence, product development, cloud-edge computing

Citation: null

1 Introduction

The development of additive manufacturing (AM)

*Corresponding author (email: [email protected])

© Science China Press and Springer-Verlag Berlin Heidelberg 2020 tech.scichina.com link.springer.com

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

184 Wang Y, et al. Sci China Tech Sci February (2020) draft

technologies, also known as 3D printing, has have sufficient knowledge of the rules of each stage.
increased in recent years [1,2]. Because 3D printing Unsuitable decisions may dramatically influence the
can produce geometrically complex parts without results of AM. With the rapid growth of the AM market,
accessories (e.g., fixtures and molds) [3,4], it has the accessibility of this knowledge should be
been applied to a wide range of applications including increased to help people achieve more flexible,
aerospace, medicine, footwear, etc. [5]. AM can adaptive and intelligent AM operations.
realize mass customization and personalization Artificial intelligence (AI), in contrast to natural
without cost penalties. The customization process intelligence, is intelligence demonstrated by machines

ed
requires knowledge and experience to implement and software rather than living systems [8]. The
suitable adjustments. Currently, the success of AM purpose of AI is to develop intelligent agents that can
highly relies on the users’ knowledge and experience perceive its environment and take actions that
to make the right decisions in the product maximize its chance to successfully achieve
development process [6,7]. As a complex process, predefined goals [9]. Empowered by the explosive
product development involves multiple stages, growth of accessible data, AI, especially deep learning
including design, process
pt
planning,
planning, and process monitoring. These stages are
highly interrelated, and the decision makers should
production techniques, has experienced another large increase in
recent years. The capability of AI has been similar to
or surpassed
ce
Ac

Figure 1 The basic structure of intelligent agents

human performance in several areas, including image been referred to as a machine learning technique at
recognition [10] and game playing [11,12]. some places [8,13]; however, it is broader than that. It
Because of the potential of AI, research interests in is a group of techniques that enables machines to
applying it to AM processes have increased to interact with the environment and adaptively solve
improve the product development process. AI has problems. These techniques enable the creation of

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

 Wu W, et al. Sci China Tech Sci February (2016) Vol.59 No.2 185

intelligent agents to perceive the environment, infer 2 Overview of AI

based on a knowledge model, search answers from
solution repositories, learn from historical data, The fundamental goal of AI research is to develop
communicate, take action, etc. intelligent agents that can perform rational actions in a
AM-based product development is a complex dynamic environment [15]. The basic structure of an
process that combines design, parameter selection, intelligent agent is shown in Figure 1. The inputs come
planning, manufacturing, control, and other aspects. from its environment, and the outputs affect the
There are numerous issues that have been dealt with environment. An intelligent agent could be software or

ed
using different AI techniques. In 2017, an initial survey hardware. For software, the human-machine interface
was conducted on AI applications in AM; however, it (HMI) is used to provide inputs and deliver outputs in
focused on the manufacturing preparation process the form of commands, files, suggestions or other
[14]. Other aspects have not been investigated. information. For hardware, sensors are used as input
Numerous new trends have emerged, such as deep devices, including image sensors (e.g. charge coupled
learning-based process monitoring and control. device (CCD)), global positioning system (GPS),
pt
Considering the rapid development of AI-related areas,
the research progress in recent years should be
summarized to guide future development.
sound sensors (e.g., microphone), etc. The layer
between the inputs and outputs contains the core
functions that form problems and generate solutions.
In this study, a survey was conducted to investigate These functions can be based on various structured
ce
the current status of AI-enabled AM. A global view of and unstructured information and knowledge (e.g., the
the product development process was considered, physical model, expert knowledge and historical data).
and the current knowledge gaps and system
limitations were identified. Based on the survey, we 2.1 Types of intelligent agents

provide a vision for future development to achieve a

There are mainly three types of intelligent agents
smarter AM production environment. The remainder of
based on their methods of generating action: a reflex
Ac

this paper is as follows. Section 2 introduces the basic

agent, goal-based agent and utility-based agent [15].
methods and techniques related to AI. Section 3
A reflex agent takes input signals and directly
reviews the current research activities on intelligent
generates reactions accordingly. It could be based on
agent development for AM. A framework for future
simple IF-THEN rules or more complex models (e.g.,
smart AM environment is proposed in Section 4.
physical models, statistical models and knowledge
Section 5 provides the conclusions drawn from this
models). The problems to be solved are often
research.
straightforward and do not involve extra optimization
or complex trade-offs.

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

186 Wang Y, et al. Sci China Tech Sci February (2020) draft

ed
pt Figure 2

A goal-based agent predicts the future status of the

Development of AM

These problems are typically knowledge-intensive and

environment and tries to achieve predefined goals. In frequently interact with the users.
this situation, there are usually multiple solutions.
ce
Criteria are used to measure their performances. The 2.2 Typical AI models

goal could be maximizing or minimizing the criteria.

All these agents rely on various models to project
Various advanced methods have been proposed for
input signals to output actions. For AI, the models
this type of searching and planning problem, including
based on knowledge and learning have received
genetic algorithms [16], particle swarm optimization
increasing attention.
[17], and simulated annealing algorithms [18], to find
A knowledge-based model solves problems by
Ac

the optimal solutions. Most problems in practice are

imitating human expertise [22]. It consists of a
complex, thus this method requires large
knowledge model and an inference system. Certain
computational power and a long processing time.
knowledge can be represented by rule-based models
A utility-based agent is used when there are
[23], case-based models, and ontology [24]. Uncertain
multiple goals that may conflict [19]. Trade-offs are
knowledge can be represented by Bayesian networks
required to achieve the best utility. Using a goal state
[25] and fuzzy logic [26].
and non-goal state in the goal-based agents is not
A learning model discovers patterns from historical
enough to measure the level of satisfaction of the
data and creates models for problem solving. This
users on each criterion. Therefore, the utility theory is
type of model has been disruptive in many areas
applied to model the user preference [20]. When
recently, including illness diagnosis and self-driving.
multiple agents are involved, the theory of game [21]
There are various types of learning models, such as
could be applied to model the players’ performance.

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

 Wu W, et al. Sci China Tech Sci February (2016) Vol.59 No.2 187

supervised and unsupervised, discriminative and printing services online. Customers can directly use
generative, deep learning and non-deep learning, etc. AM resources via the internet. However, the
[8]. The learning model has been widely applied to the customers should have enough knowledge to make
manufacturing environment for image recognition the correct decisions and print the expected parts. To
[27,28], equipment monitoring, diagnosis [29–31], etc. fulfill the growing needs for AM knowledge, AI tools
are required to make them accessible via the internet

3 AI-enabled AM and help customers make rational decisions.

At the system level, the CPS has shown great

ed
3.1 ICT-driven AM development potential recently [41]. Researchers have also
developed CPS for 3D printers [7,38,42,43]. A digital
The development of AM has been driven by advanced
twin of the physical 3D printer has to process a lot of
information and communication technologies (ICT). It
data and optimize the printing process in real-time. In
can be summarized into four stages, as shown in
this situation, AI technologies become critical to
Figure 2. The first AM machine was developed by Hull
generate solutions automatically when facing different
pt
in the 1980s [32] and is controlled by local computers.
In the mid-1990s, researchers attempted to connect
AM machines to a network to achieve remote
printer statuses based on digitalized expert knowledge
or patterns from historical data.
At the technology level, the booming ICTs, including
monitoring and control [33,34]. Users could send 3D
cloud computing, edge computing, 5G, and internet of
models to printers remotely and receive the status of
ce
things (IoT), enable large amounts of real-time data to
the printing processes. In the early 2000s, researchers
be efficiently gathered and processed. Human experts
started to treat AM as a service and accept online
cannot deal with the large amounts of data efficiently.
orders from distributed customers [35–37]. These
Thus, learning-based agents have become a popular
methods have developed a bridge between distributed
research area in recent years. These technologies
service demanders and providers. The requirements
have provided advanced solutions for various aspects
of the service demanders could be matched to the
Ac

of intelligent agents development for AM, including

capabilities of the solutions from the service providers.
data acquisition, machine-to-machine communication,
Recently, researchers started to integrate advanced
and efficient computation.
ICTs, such as the cyber-physical system (CPS) [7,38]
and deep learning [39], to automatically process
3.2 AM-based product development
complex information and help users make rational
decisions in product development processes. There are four major stages in the development of 3D

The demand to develop intelligent agents has printed parts: product design, process design,

increased, driven by the increase of connectedness production and service (Figure 3), not necessarily in

and application areas [5], as shown in Figure 2. the same order. Product design and process design

At the market level, cloud-based AM platforms have are often conducted concurrently as the design

been proposed [7,40] and companies (e.g. features and process parameters are interrelated [25].

Shapeways and 3D Hubs) have started to sell 3D The information gathered at different stages can be

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

188 Wang Y, et al. Sci China Tech Sci February (2020) draft

used at other stages to optimize the products.

Various intelligent agents have been proposed to
solve problems in these stages. In general, there are
four types of problems. Knowledge-intensive problems
rely on expert knowledge to quickly assess potential
solutions and find suitable ones. For example, the
design rules, such as a minimum feature size and
Figure 3 AM product development process
heterogeneous properties, should be considered by

ed
designers in the design stage. Otherwise, the design 3.3 Intelligent agents for product design
model may not be printable. Attention-intensive
There are three types of product design methods:
problems require continuous attention on running
opportunistic, restrictive, and dual methods [44].
equipment or other objects, and adjustments are
Opportunistic methods are used to explore the
made adaptively. Printing process monitoring is a
geometric possibilities provided by AM. Topology
typical example of an attention-intensive problem that
pt
requires continuous attention during the printing
process to identify problems as early as possible.
Computation-intensive problems include searching
optimization and elementary shapes are the most
commonly applied methods. Restrictive methods are
used to identify the boundaries of design spaces, in
terms of the geometry and properties. These
and planning problems that require a large amount of
boundaries should be represented as rules to
ce
computational power to find optimal answers. For
automatically check manufacturability. Dual methods
example, the design space for 3D printing products is
combine both types of methods to generate more
large. A large amount of computational power is
practical design proposals. Various intelligent agents
required to explore possible solutions and predict its
have been proposed to solve these problems. The
properties to identify the optimal design model.
details are listed in Table 1.
Preference-intensive problems require modelling the
Initially, a majority of the research focused on a
Ac

users’ preferences to help them make rational

single aspect, opportunistic or restrictive. For
decisions. This type of problem usually occurs in the
opportunistic methods, the optimal design solutions
process design stage to match the preferences of the
searched for specific tasks. These types of searching
properties of the printed part to the process
problems are computation-intensive and various
parameters. In this situation, there may be multiple
goal-based agents have been proposed. Physical
properties that conflict. Therefore, the trade-offs
models are applied to simulate the properties of the
become the central problem.
printed parts. For restrictive methods, the purpose is
to check if a design can be printed with the desired
quality and properties. This is a knowledge-intensive
problem, where the relationships between the design
features, process features, and properties of the

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

 Wu W, et al. Sci China Tech Sci February (2016) Vol.59 No.2 189

printed parts should be understood. These rules need only consider simple design rules, such as thin
to be encoded as various knowledge-based models to features and size limitations [57,58]. In practice, the
form reflex agents. The manufacturability can then be design rules are more complex [59]. How to engage
determined for a given design model. More recently, these complex knowledge models in opportunistic
researchers have merged both types of methods as methods is still not understood.
dual methods to generate more manufacturable
design models. In these scenarios, multiple models 3.4 Intelligent agents for process design

are involved, including knowledge-based and physical

ed
In the process design stage, optimal process-related
models, to consider manufacturability in the searching
settings and parameters are chosen to achieve the
process.
desired properties. Process planning and support
structure optimization are involved. In process
Table 1 Intelligent agents for product design
planning, optimal printing strategies are chosen for
Proposed Model
Aspects Problem type Agent type specific tasks. These include material-related settings
agents type
(e.g., material types and ingredients), machine-related

Opportunistic
Design feature Knowledge-inte
database [45]
Topology
optimization
nsive

Computation-int
pt Reflex
Knowledg
e-based

Goal-based Physical
settings (e.g., power energy and layer height) and
design-related settings (e.g., build orientation and

methods ensive
trajectory). As many AM processes require support
[46–48]
structures to avoid failures, the least amount of
Elementary Computation-int
ce
Goal-based Physical support materials should be used to achieve the
shapes [49–52] ensive
Knowledge-bas required quality. The design of the support structures
Knowledge-inte Knowledg
ed system Reflex also influence the printed properties [60]. Therefore,
nsive e-based
Restrictive [25,53]
support structure optimization has become an
methods Key feature Knowledg
Knowledge-inte important topic in recent years [61,62]. Various
recognition Reflex e-based &
nsive
[54,55] physical intelligent agents have been proposed to solve this
Ac

Part Knowledg problem. The details are listed in Table 2.

Computation-int
consolidation Goal-based e-based &
ensive
[56] physical
Table 2 Intelligent agents for process design
Dual methods Topology
Knowledg
optimization for Computation-int Proposed Model
Goal-based e-based & Aspects Problem type Agent type
manufacturabilit ensive agents type
physical
y [57,58] MCDM for preference-inte
Knowledg
parameter nsive & Utility-base
e-based &
selection knowledge-inte d
In the product design stage, the integration of physical
[63–66] nsive
Process
opportunistic and restrictive methods provides a Path planning Computation-int
planning Goal-based Physical
promising solution to fulfil the complex design and [67–71] ensive
Build direction
manufacturing requirements and generate more Computation-int
optimization Goal-based Physical
reliable design proposals. However, current methods ensive
[72–74]

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

190 Wang Y, et al. Sci China Tech Sci February (2020) draft

Topology reduce the mass of the materials and improve the

Computation-int
Support optimization Goal-based physical
ensive ease of removal while ensuring functional
structure [75,76]
performance.
optimization Elementary Computation-int
Goal-based physical
shapes [77,78] ensive In the process design stage, utility-based agents
have been introduced to tackle different user
preferences on the properties of the printed parts.
There are multiple parameters (e.g., layer thickness,
However, the accuracy of the knowledge-based model
printing speed, and base temperature) for the printing
is important for projecting user requirements to

ed
process, and they often have a different impact on
suitable printing parameters. Furthermore, the process
various aspects of the printed parts. Therefore,
parameters, build directions, and support structures
multi-criteria decision making (MCDM) methods are
are interrelated. More comprehensive models are
widely adopted to help users make rational decisions
required to find global optimal solutions.
for parameter selection. As the user preferences
influence the decisions, modeling these preferences
3.5 Intelligent agents for production
pt
becomes crucial. Knowledge also plays a significant
role in mapping the parameters with printed results.
Therefore, these utility-based agents usually have a
In the production stage, there could be multiple
printing tasks simultaneously, especially in cloud
knowledge-based model for inference. manufacturing. Considering the long pre-processing
Path planning of AM processes is also complex. and post-processing times, it is more efficient to print
ce
Two aspects are involved: horizontal trajectory as many parts as possible for each print [79,80]. In
optimization and adaptive layer slicing. The first this situation, production planning becomes important
aspect influences the efficiency of the printing process to improve the overall production efficiency. The
while the second is a trade-off between the printing deadline of each task, design geometries, required
efficiency and part quality, in terms of dimensional parameters, etc. should be considered. The fabrication
accuracy and surface roughness. As a planning process typically requires hours or days to finish.
Ac

problem, various goal-based agents have been During this period, defects may occur, and they may
proposed to determine optimal solutions within shorter cause the part to fail. Therefore, real-time monitoring
computational times. of the printing process is required to quickly identify
Build direction optimization is another important these defects, fix the problem, and avoid unnecessary
topic owing to the heterogeneous properties of the waste of time and money. Various intelligent agents
printed parts. These agents are also goal-based to have been proposed to solve these problems. The
search for the optimal direction for a given design and details are listed in Table 3.
achieve the best quality. They involve physical models
to predict the surface roughness, accuracy, and other Table 3 Intelligent agents for production

properties. Support structure optimization is similar to Proposed Model

Aspects Problem type Agent type
the opportunistic methods of product design, while agents type
Production Single print Computation-intGoal-based Physical
targeting the support structures. The purpose is to

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

 Wu W, et al. Sci China Tech Sci February (2016) Vol.59 No.2 191

planning optimization ensive These agents use learning models to extract complex
[81–84]
features of targeted patterns from historical images
Multiple prints
Computation-int and identify them in real-time images. AR-based
optimization Goal-based Physical
ensive
[79,85–87] methods have also become popular. Users can
Image-based capture the 3D point cloud of the printing part and
Attention-intens
methods reflex Learning
ive compare it to the design model to identify
[39,88–91]
discrepancies. In addition, various sensors (e.g.,
Real-time AR-based
Attention-intens
temperature, vibration, and power sensors) and

ed
control & methods reflex Physical
ive
monitoring [42,92] process parameters can be used to predict the quality
Sensor-based of the printed parts. The learning models are required
Attention-intens Learning
methods reflex
ive & physical to find features from the time-series data.
[93–95]
In the production stage, attention-based problems
have occurred, and learning models have advantages
Production planning starts from single-print
for solving these problems. These agents could
scenarios, where the user attempts to place as many
pt
parts as possible into one printing space. Each part
could rotate and move to any place in the printing
release people from tedious process monitoring tasks.
However, the image-based and AR-based methods
can only identify visible defects. Sensor-based
space to provide infinite possible solutions. The
methods are able to identify invisible defects; however,
genetic algorithm (GA) has been widely used to find a
ce
the relationships of the printing parameters and part
sub-optimal solution with reasonable computation time.
quality must be defined in advance. Furthermore, the
The problem becomes more complex to plan for
available data for the AM processes are limited. How
multiple prints. There are more printers available,
to use the limited data to generate more robust and
which may have different statuses (idle or working).
general models is still a challenge.
The parts may have different deadlines. If the task
space is also dynamic (i.e., a new task could be added
Ac

3.6 Global methods

at any time), the computation speed will become
The intelligent agents mentioned previously are
crucial for a fast response. Here, the GA-based
developed for solving single problems. However,
algorithms cannot meet the efficiency requirement.
these problems are not isolated. For example, the
Therefore, various heuristic methods are proposed to
design of a product will influence the selection of the
use pre-defined rules to narrow the searching space.
printing parameters and production plan. The process
Real-time control and monitoring have been
capability (e.g., minimum feature size and printing
evaluated in recent years owing to the rapid
resolution) will influence the design of the product.
development of machine learning techniques. Image
Therefore, optimal AM solutions can only be achieved
processing is one of the most popular areas in
with global consideration. The first step is to define a
machine learning and has enabled various
uniform data format to carry the information from
image-based monitoring agents for AM processes.
different stages. Lu et al. proposed an integrated data

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

192 Wang Y, et al. Sci China Tech Sci February (2020) draft

schema for AM [96]. A cloud-based automated design different disciplines, and there is a shortage of experts.
and AM platform was proposed to integrate the data in Therefore, these agents can help more users take
the service stage to optimize the product design, advantage of AM. The goal-based agents can find
process setting, and production strategy [97]. The optimal solutions from a large design space. Physical
framework was the focus, and efficient utilization of big models are usually involved in simulation and
data has not been fully studied. A conceptual design determine whether the goals and constraints are
and modelling framework were presented to integrate satisfied. This type of agent can use the computational
various simulation and prediction models for the power to maximize the efficiency of material utilization.

ed
product development process [98]. This framework The waste of materials, energy, and time could be
focused on reflex agents and other types of agents minimized.
were not considered. Majeed et al. proposed a big In the production stage, learning models have
data-driven framework for AM process analysis and printing process control and monitoring advantages.
optimization [99]. Data acquisition, management, Operators no longer need to sit beside the printer and
modelling, and utilization were discussed. A continually check its status. From historical data, this
pt
knowledge-based computer-aided production planning
(CAPP) framework was proposed to integrate the
product and process design stages [100]. Processing
model can identify the features of abnormal situations.
These agents enlarge the production scale without
dramatically expanding the operating team. Deep
big data requires computational resources. How to learning techniques have shown more potential to deal
ce
efficiently process computation-intensive models and with complex problems as they can automatically
provide a fast response for distributed users has not extract the important features in multiple levels, as
been discussed. performed by the human brain [101].
Although successful applications of intelligent
3.7 Discussions agents have been reported, there are still several
knowledge gaps to be explored:
Current studies have shown that AM-based product
Ac

 Scattered and conflicting knowledge management

development could benefit from AI techniques.
is evaluated. AM knowledge is the central part of AM
Various intelligent agents have been developed to
intelligence. As a multi-disciplinary area, the related
provide decision assistance in the product design,
information is widely scattered at different locations
process design, and production stages.
and in different formats (e.g., test data, text, physical
In the first two stages, the main purpose is to
models, statistical models, rules, ontologies, and
determine the optimal design of the product and
graphical databases). They may be described with
process. Therefore, the problems are knowledge and
different standards and conflict with each other.
computation-intensive. Reflex and goal-based agents
Knowledge bases for small applications have been
have been widely applied. The knowledge-based
developed. However, the knowledge in different
models in the reflex agents can gather expert
aspects is interrelated. Informed decisions can only
knowledge and make it available to non-expert users.
be made with global consideration. Therefore,
AM involves a large amount of knowledge from

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

 Wu W, et al. Sci China Tech Sci February (2016) Vol.59 No.2 193

integrating these pieces of information to form a practical agents for complicated scenarios, efficiently
comprehensive and accurate knowledge base is a combining different types of models for better
challenging task. performance should be studied.
 A deeper understanding of AM processes is  A wider application of the learning models is
required. Physical models are also important to possible. Current learning models are mainly
predict the performance of AM processes for developed for real-time control and monitoring. The
goal-based agents. Current models may work well in techniques could potentially tackle more complex
specific circumstances; however, their accuracy and problems. The emerging reinforcement learning

ed
generality should be improved to generate more methods can learn rules in strategical tasks.
realistic results. This type of model can also be Advanced deep learning methods have been
applied in the sensor-based method for remote proposed for 3D models. These emerging
control and monitoring. In this situation, the techniques could develop useful agents for product
computational efficiency of these models becomes and process design stages. The execution of trained
more important to achieve real-time feedback. models could be more efficient than the commonly
pt
 For big data management and utilization, AI
techniques are data-hungry, especially learning
models. Currently, there are no standard databases
used searching algorithms. The first challenge is
standardizing the data model for the design process
to prepare the training data.
that are similar to ImageNet in AM to gather all the  For an efficient computational strategy, a trend of
ce
data available along with the corresponding tags for AM development is to develop online services. In this
researchers to test the generality of their models in a situation, the intelligent agents serve distributed
larger scale. Furthermore, the application of learning users via the internet. The agents with searching and
models in AM is not similar to other areas. Abnormal planning algorithms usually require a large
situations are not common in practice. Therefore, the computational power. In addition, some tasks require
data gathered for training are limited. Using the a fast response. Learning-based real-time control
Ac

limited data to train robust models requires further and monitoring is required for the training and
study. execution stages, respectively. Although
 For integration of different models, most of the cloud-based design and simulation software have
current agents have single models, which may not been developed, real-time tasks cannot be
be adequate for many applications. For example, the performed. Therefore, a new efficient computational
goal-based agents for product design require framework should be studied to fulfil both
physical models to simulate the properties and requirements.
knowledge-based models to examine the
manufacturability of the proposed design. The reflex 4 Framework of smart AM
agents for real-time control and monitoring require
combining learning and knowledge models for From Section 3.7, one problem that hinders the
complex problems. To develop more effective and development of AI for AM is the lack of high-quality

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

194 Wang Y, et al. Sci China Tech Sci February (2020) draft

and high-volume AM related information. This the rapid development of AI for image processing, the
information includes the data from testing and cloud platforms, such as ImageNet, can be used to
monitoring, expert knowledge, standards, algorithms, gather distributed image data to a
physical models, and other types of information. From

ed
pt
ce
Figure 4 Smart AM framework

central location to allow anyone to access the data be applied to other manufacturing processes with
and develop innovative ideas. Furthermore, minor adjustments.
information processing efficiency is another important
Ac

aspect, as there are numerous scenarios in AM that 4.1 Definition of smart AM

require a real-time response. Existing platforms

Smart manufacturing can be defined as a fully
discussed in Section 3.6 mainly focus on a specific
integrated, collaborative manufacturing system that
aspect of AM, such as product design, fabrication or
responds in real time to meet dynamic changing
CAPP. The efficiency of data processing in the
demands and conditions of the factory, supply network,
platform has not been fully considered. To realize
and customer requirements [102]. Traditional
more intelligent AM with a high efficiency, a
manufacturing processes involve multiple operations
cloud-edge computing-based framework for smart AM
(e.g., molding, heat treatment, milling, grinding, etc.).
is proposed based on AI. The goal is to provide future
Therefore, this definition mainly applies to the
development of intelligent agents for AM. Although the
manufacturing management aspect. The
proposed framework is designed for AM, it could also
manufacturing process of AM is simpler, without extra

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

 Wu W, et al. Sci China Tech Sci February (2016) Vol.59 No.2 195

tools and complex procedures [3]. The focus is to enhance computation efficiency and avoid
changed from the manufacturing and assembly bottlenecks and system failure [106]. A comparison of
aspects to the products. A wider design space can be cloud and edge computing [107] is listed in Table 4.
explored, and faster prototypes can be achieved. Combining the advantages of both methods,
Therefore, AM from the product-centric perspective cloud-edge computing could handle more complex
should be understood. Considering of the whole demands in the industrial environment [108,109].
product lifecycle and integrating different stages to
provide better global product solutions. Hence, Smart Table 4 Comparison of cloud and edge computing

ed
AM is defined in this study as “a fully integrated, Criterion Edge computing Cloud computing

collaborative additive manufacturing system that Computing

On-premises Remote
equipment location
responds in real time to support ubiquitous and
Computation power Limited Unlimited
intelligent design, manufacturing and services of 3D Storage Limited Unlimited
printed products.” Latency Low High
Accessibility Local Public

4.2

Cloud
Cloud-edge computing

computing utilizes
pt
simple centralized
4.3 Proposed framework
architectures with dedicated data centers to offer
several allowances, such as self-service provisioning, The proposed smart AM framework is shown in Figure
ce
elasticity, pay per use, etc. [103]. Big data generated 4. because massive data are generated in the process,

from distributed IoT devices cannot be properly the cloud-edge computing framework is applied to

handled by the remote cloud owing to the large facilitate efficient information transmission and

consumption of time, energy, and bandwidth with a processing.

high response time [104]. Privacy concerns are The information layer contains the source

another critical issue. Services providers may not want information for the platform to form basic knowledge
Ac

to expose their confidential data on the cloud. and models. The information could come from the

Multi-access edge computing (MEC) was standards, expert know-how, test data, historical data,

standardized by the European Telecommunications real-time data, etc. As in a different format, this

Standards Institute (ETSI) and Industry Specification information should be standardized before integrated

Group (ISG) and defined as “MEC provides an IT processing.

service environment and cloud computing capabilities The cloud computing layer is responsible for

at the edge of the mobile network, within the radio distributed information integration, forming intelligent

access network (RAN) and in close proximity to mobile agents for public access and processing

subscribers” [105]. MEC offers cloud computing computationally heavy tasks. For central information

capabilities within the RAN and connects the users storage, the cloud environment can form more

directly to the nearest cloud service-enabled edge comprehensive knowledge-based and learning models

network, by deploying edge nodes at the base stations to developing effective agents. Combining the physical

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

196 Wang Y, et al. Sci China Tech Sci February (2020) draft

models and searching and planning algorithms, future of smart AM.

various agents can be formed for different stages.
Hybrid models can be developed as they are all This work was supported by the National Natural Science Foundation
of China (Grant No. 51890885) and the Natural Science Foundation
available in the cloud environment.
of Zhejiang Province (LY19E050019).
The edge computing layer is used to process light
agents for a fast response and to reduce the burden
References
for the cloud. For example, the learning-based agents
can put the training process in the cloud, while the

ed
1. Mehrpouya M, Dehghanghadikolaei A, Fotovvati B,
execution process at the edge can achieve maximum Vosooghnia A, Emamian SS, Gisario A. The Potential of

efficiency and response time. The reflex agents can Additive Manufacturing in the Smart Factory Industrial 4.0:
A Review. Applied Sciences, 2019, 9: 3865
be deployed at the edge, as they do not require large
2. Gao W, Zhang Y, Ramanujan D, Ramani K, Chen Y,
computation. Conversely, the goal-based agents often Williams CB, Wang CCL, Shin YC, Zhang S, Zavattieri PD.
require a large computational power, thus, it is better The status, challenges, and future of additive
manufacturing in engineering. Computer-Aided Design,
to deploy them in the cloud. The edge nodes can also
pt
synchronize useful local data to the cloud when the
network is not busy for model improvement.
The users can access these agents and resources
3.

4.
2015, 69: 65–89
Gibson I, Rosen DW, Stucker B. Additive Manufacturing
Technologies. New York: Springer, 2010
Zhu Y, Zou J, Yang H yong. Wear performance of metal
parts fabricated by selective laser melting: a literature
from the user layer using various devices. review. Journal of Zhejiang University: Science A, 2018, 19:
ce
95–110
5. Royal Academy of Engineering. Additive manufacturing :
5 Conclusions
opportunities and constraints. 2013
6. Khaleeq uz Zaman U, Siadat A, Rivette M, Baqai AA, Qiao
Because of deep learning development, AI use has L. Integrated product-process design to suggest
appropriate manufacturing technology: a review. The
increased in many fields. The unique capabilities of AI
International Journal of Advanced Manufacturing
have also increased the attention given to the Technology, 2016
Ac

improvement of AM-based product development. This 7. Wang Y, Lin Y, Zhong RY, Xu X. IoT-enabled cloud-based
additive manufacturing platform to support rapid product
study reviewed the current research activities on the
development. International Journal of Production Research,
application of AI in AM, including product design, 2019, 57: 3975–3991
process design, production and services. Several 8. Wang L. From Intelligence Science to Intelligent
Manufacturing. Engineering, 2019, 5: 615–618
research gaps and future directions are summarized.
9. Poole D, Mackworth A, Goebel R. Computational
To develop a more efficient and comprehensive Intelligence: A Logical Approach. Oxford University Press,
environment for AI-enabled AM, a smart AM 1998
10. Rawat W, Wang Z. Deep Convolutional Neural Networks
framework is proposed. Cloud-edge computing is
for Image Classification: A Comprehensive Review. Neural
applied to fulfil the computational requirements of
Computation, 2017, 29: 2352–2449
different types of agents. The aim of this study is to 11. Silver D, Huang A, Maddison CJ, Guez A, Sifre L, Van Den

better understand how AI techniques can help AM Driessche G, Schrittwieser J, Antonoglou I,

Panneershelvam V, Lanctot M, Dieleman S, Grewe D,
product development and provide our view of the
Nham J, Kalchbrenner N, Sutskever I, Lillicrap T, Leach M,

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

 Wu W, et al. Sci China Tech Sci February (2016) Vol.59 No.2 197

Kavukcuoglu K, Graepel T, Hassabis D. Mastering the 24. Dinar M, Rosen DW. A Design for Additive Manufacturing
game of Go with deep neural networks and tree search. Ontology. Journal of Computing and Information Science in
Nature, 2016, 529: 484–489 Engineering, 2017, 17: 021013
12. Berner C, Brockman G, Chan B, Cheung V, Dębiak P, 25. Wang Y, Blache R, Zheng P, Xu X. A Knowledge
Dennison C, Farhi D, Fischer Q, Hashme S, Hesse C, Management System to Support Design for Additive
Józefowicz R, Gray S, Olsson C, Pachocki J, Petrov M, Manufacturing Using Bayesian Networks. Journal of
Pinto HP de O, Raiman J, Salimans T, Schlatter J, Mechanical Design, 2018, 140: 051701
Schneider J, Sidor S, Sutskever I, Tang J, Wolski F, Zhang 26. Khrais S, Al-Hawari T, Al-Araidah O. A fuzzy logic
S. Dota 2 with Large Scale Deep Reinforcement Learning. application for selecting layered manufacturing techniques.
arXiv:191206680 [csLG], 2019 Expert Systems with Applications, 2011, 38: 10286–10291

ed
13. Luckow A, Kennedy K, Ziolkowski M, Djerekarov E, Cook 27. Fan YM, Ding M, Cao YF. Vision algorithms for fixed-wing
M, Duffy E, Schleiss M, Vorster B, Weill E, Kulshrestha A, unmanned aerial vehicle landing system. Science China
Smith MC. Artificial Intelligence and Deep Learning Technological Sciences, 2017, 60: 434–443
Applications for Automotive Manufacturing. Proceedings - 28. Yao S, Li H, Zeng L, Zhang XM. Vision-based adaptive
2018 IEEE International Conference on Big Data, Big Data control of a 3-RRR parallel positioning system. Science
2018, 2018, 3144–3152 China Technological Sciences, 2018, 61: 1253–1264
14. Yang J, Chen Y, Huang W, Li Y. Survey on artificial 29. Cao L, Zhang XM, Huang T, Ding H. Online Monitoring
intelligence for additive manufacturing. ICAC 2017 - 2017 Machining Errors of Thin-Walled Workpiece: A Knowledge

15.
Computing: Addressing Global
Automation and Computing, 2017, 7–8
pt
23rd IEEE International Conference on Automation and
Challenges

Russell SJ, Norvig P. Artificial Intelligence: A Modern

through

30.
Embedded Sparse Bayesian Regression
IEEE/ASME Transactions on Mechatronics, 2019, 24:
1259–1270
Yuan Y, Cheng C, Zhou B, Zhao H, Zhang H. A General
Approach.

Approach. 3rd ed. New Jersey, USA: Pearson, 1995 End-to-end Diagnosis Framework for Manufacturing
16. Li YF, Chen MN, Lu XD, Zhao WZ. Research on optimized Systems. National Science Review, 2019, 1–24
ce
GA-SVM vehicle speed prediction model based on 31. Yuan Y, Tang X, Zhou W, Pan W, Li X, Zhang HT, Ding H,
driver-vehicle-road-traffic system. Science China Goncalves J. Data driven discovery of cyber physical
Technological Sciences, 2018, 61: 782–790 systems. Nature Communications, 2019, 10: 1–9
17. Kennedy J, Eberhart R. Particle Swarm Optimization. IEEE 32. Hull CW. Apparatus for production of three-dimensional
International Conference on Neural Networks. Vol 4. 1995. objects by stereolithography. 4,575,330, 1986
1942–1948 33. Bailey MJ. Tele-Manufacturing: Rapid Prototyping on the
18. Eglese RW. Simulated annealing: A tool for operational Internet. IEEE Computer Graphics and Applications, 1995,
research. European Journal of Operational Research, 1990, 15: 20–26
Ac

46: 271–281 34. Luo RC, Tzou JH, Chang YC. Desktop Rapid Prototyping
19. Fernandez MG, Seepersad CC, Rosen DW, Allen JK, System With Supervisory Control and Monitoring Through
Mistree F. Decision Support in Concurrent Engineering - Internet. IEEE/ASME TRANSACTIONS ON
The Utility-Based Selection Decision Support Problem. MECHATRONICS, 2001, 6: 399–409
Concurrent Engineering, 2005, 13: 13–27 35. Liu X, Jin Y, Xi J. Development of a web-based system for
20. Mongin P. Expected utility theory. The Handbook of rapid prototyping. Journal of Manufacturing Technology
Economic Methodology. London, 1997. 342–350 Management, 2006, 17: 303–314
21. Neumann J von, Morgenstern O. Theory of games and 36. Luo RC, Lan CC, Tzou JH, Chen CC. The development of
economic behavior. Princeton University Press, 2007 WEB based e-commerce platform for rapid prototyping
22. Leo Kumar SP. Knowledge-based expert system in system. Proceedings of the 2004 IEEE International
manufacturing planning: state-of-the-art review. Conference on Networking, Sensing and Control. Taipei,
International Journal of Production Research, 2019, 57: Taiwan: IEEE, 2004. 122–127
4766–4790 37. Jiang P, Fukuda S. TeleRP - an Internet web-based
23. Bibb RJ, Taha Z, Brown R, Wright D. Development of a solution for remote rapid prototyping service and
rapid prototyping design advice system. Journal of maintenance. International Journal of Computer Integrated
Intelligent Manufacturing, 1999, 10: 331–339 Manufacturing, 2001, 14: 83–94

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

198 Wang Y, et al. Sci China Tech Sci February (2020) draft

38. Liu XF, Sunny SMN Al, Shahriar MR, Leu MC, Cheng M, Design and Applications, 2007, 4: 585–594
Hu L. Implementation of MTConnect for open source 3D 51. Tang Y, Dong G, Zhao YF. A hybrid geometric modeling
printers in cyber physical manufacturing cloud. method for lattice structures fabricated by additive
Proceedings of the ASME 2016 International Design manufacturing. International Journal of Advanced
Engineering Technical Conferences and Computers and Manufacturing Technology, 2019, 102: 4011–4030
Information in Engineering Conference (IDETC/CIE 2016). 52. Tang Y, Zhao YF. Multifunctional design of heterogeneous
Charlotte, North Carolina: ASME, 2016. 1–10 cellular structures. Structural and Multidisciplinary
39. Jin Z, Zhang Z, Gu GX. Autonomous in-situ correction of Optimization, 2018, 58: 1121–1138
fused deposition modeling printers using computer vision 53. Kim S, Rosen DW, Witherell P, Ko H. A Design for Additive
and deep learning. Manufacturing Letters, 2019, 22: 11–15 Manufacturing Ontology to Support Manufacturability

ed
40. Luo X, Zhang L, Ren L, Lali Y. A dynamic and static data Analysis. Journal of Computing and Information Science in
based matching method for cloud 3D printing. Robotics Engineering, 2019, 19: 1–10
and Computer-Integrated Manufacturing, 2020, 61: 101858 54. Liu S, Li Q, Chen W, Tong L, Cheng G. An identification
41. Tao F, Zhang H, Liu A, Nee AYC. Digital Twin in Industry: method for enclosed voids restriction in manufacturability
State-of-the-Art. IEEE Transactions on Industrial design for additive manufacturing structures. Frontiers of
Informatics, 2018, 15: 2405–2415 Mechanical Engineering, 2015, 10: 126–137
42. Lhachemi H, Malik A, Shorten R. Augmented Reality, 55. Campana G, Mele M. An application to Stereolithography
Cyber-Physical Systems, and Feedback Control for of a feature recognition algorithm for manufacturability

43.
50119–50135
pt
Additive Manufacturing: A Review. IEEE Access, 2019, 7:

DebRoy T, Zhang W, Turner J, Babu SS. Building digital

twins of 3D printing machines. Scripta Materialia, 2017,
56.
evaluation. Journal of Intelligent Manufacturing, 2018
Yang S, Santoro F, Sulthan MA,
numerical-based part consolidation candidate detection
approach with modularization considerations. Research in
Zhao YF. A

135: 119–124 Engineering Design, 2018, 0: 0

44. Laverne F, Segonds F, Anwer N, Le Coq M. Assembly 57. Mhapsekar K, McConaha M, Anand S. Additive
ce
Based Methods to Support Product Innovation in Design Manufacturing Constraints in Topology Optimization for
for Additive Manufacturing: An Exploratory Case Study. Improved Manufacturability. Journal of Manufacturing
Journal of Mechanical Design, 2015, 137: 121701 Science and Engineering, 2018, 140: 1–16
45. Bin Maidin S, Campbell I, Pei E. Development of a design 58. Zhao J, Zhang M, Zhu Y, Li X, Wang L, Hu J. A novel
feature database to support design for additive optimization design method of additive manufacturing
manufacturing. Assembly Automation, 2012, 32: 235–244 oriented porous structures and experimental validation.
46. Orme ME, Gschweitl M, Ferrari M, Madera I, Mouriaux F. Materials & Design, 2018, 163: 107550
Designing for additive manufacturing: Lightweighting 59. Adam GAO, Zimmer D. On design for additive
Ac

through topology optimization enables lunar spacecraft. manufacturing: Evaluating geometrical limitations. Rapid
Journal of Mechanical Design, Transactions of the ASME, Prototyping Journal, 2015, 21: 662–670
2017, 139: 1–6 60. Jiang J, Lou J, Hu G. Effect of Support on Printed
47. Wang X, Xu S, Zhou S, Xu W, Leary M, Choong P, Qian M, Properties in Fused Deposition Modelling Processes.
Brandt M, Xie YM. Topological design and additive Virtual and Physical Prototyping, 2019
manufacturing of porous metals for bone scaffolds and 61. Jiang J, Xu X, Stringer J. Optimization of Process Planning
orthopaedic implants: A review. Biomaterials, 2016, 83: for Reducing Material Waste in Extrusion Based Additive
127–141 Manufacturing. Robotics and Computer-integrated
48. Mirzendehdel AM, Suresh K. Support structure constrained Manufacturing, 2019, 59: 317–325
topology optimization for additive manufacturing. CAD 62. Jiang J, Xu X, Stringer J. Support Structures for Additive
Computer Aided Design, 2016, 81: 1–13 Manufacturing: A Review. Journal of Manufacturing and
49. Lebaal N, Zhang Y, Demoly F, Roth S, Gomes S, Bernard Materials Processing, 2018, 2: 64
A. Optimised lattice structure configuration for additive 63. Wang Y, Zhong RY, Xu X. A decision support system for
manufacturing. CIRP Annals, 2019, 68: 117–120 additive manufacturing process selection using a hybrid
50. Rosen DW. Computer-aided design for additive multiple criteria decision-making method. Rapid
manufacturing of cellular structures. Computer-Aided Prototyping Journal, 2018, 24: 1544–1553

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

 Wu W, et al. Sci China Tech Sci February (2016) Vol.59 No.2 199

64. Qin Y, Qi Q, Scott PJ, Jiang X. An additive manufacturing 2018, 57: 183–195
process selection approach based on fuzzy Archimedean 77. Vaissier B, Pernot JP, Chougrani L, Véron P.
weighted power Bonferroni aggregation operators. Genetic-algorithm based framework for lattice support
Robotics and Computer Integrated Manufacturing, 2020, structure optimization in additive manufacturing. CAD
64: 101926 Computer Aided Design, 2019, 110: 11–23
65. Zheng P, Wang Y, Xu X, Xie SQ. A weighted rough set 78. Jiang J, Stringer J, Xu X. Support Optimization for Flat
based fuzzy axiomatic design approach for the selection of Features via Path Planning in Additive Manufacturing. 3D
AM processes. International Journal of Advanced Printing and Additive Manufacturing, 2019, 6: 171–179
Manufacturing Technology, 2017, 91: 1977–1990 79. Wang Y, Zheng P, Xu X, Yang H, Zou J. Production
66. KEK V, S. V, P. B, R. M. Rapid prototyping process planning for cloud-based additive manufacturing—A

ed
selection using multi criteria decision making considering computer vision-based approach. Robotics and
environmental criteria and its decision support system. Computer-Integrated Manufacturing, 2019, 58: 145–157
Rapid Prototyping Journal, 2016, 22: 225–250 80. Gogate AS, Pande SS. Intelligent layout planning for rapid
67. Dreifus G, Goodrick K, Giles S, Patel M, Foster RM, prototyping. International Journal of Production Research,
Williams C, Lindahl J, Post B, Roschli A, Love L, Kunc V. 2008, 46: 5607–5631
Path Optimization Along Lattices in Additive Manufacturing 81. Zhang Y, Bernard A, Harik R, Fadel G. A new method for
Using the Chinese Postman Problem. 3D Printing and single-layer-part nesting in additive manufacturing. Rapid
Additive Manufacturing, 2017, 4: 98–104 Prototyping Journal, 2018
68.

69.
Path Planner for 3D Printing
pt
Fok KY, Cheng CT, Ganganath N, Iu H, Tse CK. A Nozzle
Applications.
Transactions on Industrial Informatics, 2019, 3203: 1–11
Xu K, Li Y, Chen L, Tang K. Curved layer based process
IEEE
82.

83.
Canellidis V, Giannatsis J, Dedoussis V. Efficient parts
nesting schemes for improving stereolithography utilization.
CAD Computer Aided Design, 2013, 45: 875–886
Yao M, Chen Z, Luo L, Wang R, Wang H. Level-set-based
planning for multi-axis volume printing of freeform parts. partitioning and packing optimization of a printable model.
CAD Computer Aided Design, 2019, 114: 51–63 ACM Transactions on Graphics, 2015, 34: 1–11
ce
70. Mao H, Kwok TH, Chen Y, Wang CCL. Adaptive slicing 84. Jiang J, Xu X, Stringer J. Optimisation of multi-part
based on efficient profile analysis. CAD Computer Aided production in additive manufacturing for reducing support
Design, 2019, 107: 89–101 waste. Virtual and Physical Prototyping, 2019
71. Minetto R, Volpato N, Stolfi J, Gregori RMMH, da Silva 85. Chergui A, Hadj-Hamou K, Vignat F. Production scheduling
MVG. An optimal algorithm for 3D triangle mesh slicing. and nesting in additive manufacturing. Computers and
CAD Computer Aided Design, 2017, 92: 1–10 Industrial Engineering, 2018, 126: 292–301
72. Zhang Y, Harik R, Fadel G, Bernard A. A statistical method 86. Griffiths V, Scanlan JP, Eres MH, Martinez-Sykora A,
for build orientation determination in additive manufacturing. Chinchapatnam P. Cost-driven build orientation and bin
Ac

Rapid Prototyping Journal, 2019, 25: 187–207 packing of parts in Selective Laser Melting (SLM).
73. Qin Y, Qi Q, Scott PJ, Jiang X. Determination of optimal European Journal of Operational Research, 2019, 273:
build orientation for additive manufacturing using Muirhead 334–352
mean and prioritised average operators. Journal of 87. Li Q, Kucukkoc I, Zhang DZ. Production planning in
Intelligent Manufacturing, 2019, 30: 3015–3034 additive manufacturing and 3D printing. Computers and
74. Tnullows DM, Pii S. Optimized Build Orientation of Additive Operations Research, 2017, 83: 157–172
Manufactured Parts for Improved Surface Quality and Build 88. Shen H, Sun W, Fu J. Multi-view online vision detection
Time. Additive Manufacturing, 2016, 12: 314–320 based on robot fused deposit modeling 3D printing
75. Cheng L, Liang X, Bai J, Chen Q, Lemon J, To A. On technology. Rapid Prototyping Journal, 2019, 25: 343–355
utilizing topology optimization to design support structure to 89. Zhang B, Jaiswal P, Rai R, Guerrier P, Baggs G.
prevent residual stress induced build failure in laser Convolutional neural network-based inspection of metal
powder bed metal additive manufacturing. Additive additive manufacturing parts. Rapid Prototyping Journal,
Manufacturing, 2019, 27: 290–304 2019
76. Kuo YH, Cheng CC, Lin YS, San CH. Support structure 90. Zhang Y, Hong GS, Ye D, Zhu K, Fuh JYH. Extraction and
design in additive manufacturing based on topology evaluation of melt pool, plume and spatter information for
optimization. Structural and Multidisciplinary Optimization, powder-bed fusion AM process monitoring. Materials and

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

200 Wang Y, et al. Sci China Tech Sci February (2020) draft

Design, 2018, 156: 458–469 RPJ-04-2017-0075

91. Wu D, Xu C. Predictive modeling of droplet formation 100. Zhang Y, Bernard A. A KBE CAPP framework for qualified
processes in inkjet-based bioprinting. Journal of additive manufacturing. CIRP Annals - Manufacturing
Manufacturing Science and Engineering, Transactions of Technology, 2018, 4–7
the ASME, 2018, 140: 1–9 101. Voulodimos A, Doulamis N, Doulamis A, Protopapadakis E.
92. Ceruti A, Liverani A, Bombardi T. Augmented vision and Deep Learning for Computer Vision: A Brief Review.
interactive monitoring in 3D printing process. International Computational Intelligence and Neuroscience, 2018, 2018:
Journal on Interactive Design and Manufacturing, 2017, 11: 1–13
385–395 102. Kusiak A. Smart manufacturing. International Journal of
93. Mukherjee T, DebRoy T. A digital twin for rapid Production Research, 2018, 56: 508–517

ed
qualification of 3D printed metallic components. Applied 103. Satpathy S, Sahoo B, Turuk AK. Sensing and Actuation as
Materials Today, 2019, 14: 59–65 a Service Delivery Model in Cloud Edge centric Internet of
94. Tlegenov Y, Hong GS, Lu WF. Nozzle condition monitoring Things. Future Generation Computer Systems, 2018, 86:
in 3D printing. Robotics and Computer-Integrated 281–296
Manufacturing, 2018, 54: 45–55 104. Aujla GS, Kumar N, Zomaya AY, Rajan R. Optimal
95. Li Z, Zhang Z, Shi J, Wu D. Prediction of surface Decision Making for Big Data Processing at Edge-Cloud
roughness in extrusion-based additive manufacturing with Environment: An SDN Perspective. IEEE Transactions on
machine learning. Robotics and Computer-Integrated Industrial Informatics, 2017, 14: 1–1

96.
Manufacturing, 2019, 57: 488–495
pt
Lu Y, Choi S, Witherell P. Towards an Integrated Data
Schema Design for Additive Manufacturing: Conceptual
Modeling. Proceedings of the ASME 2015 International
105.

106.
ETSI. Multi-access Edge Computing
https://www.etsi.org/technologies/multi-access-edge-comp
uting, accessed 30 January 2020
Satria D, Park D, Jo M. Recovery for overloaded mobile
(MEC).

Design Engineering Technical Conferences & Computers edge computing. Future Generation Computer Systems,
and Information in Engineering Conference (IDETC/CIE 2017, 70: 138–147
ce
2015). Boston, Massachusetts, USA, 2015. 1–11 107. Mach P, Becvar Z. Mobile Edge Computing: A Survey on
97. Lehmhus D, Wuest T, Wellsandt S, Bosse S, Kaihara T, Architecture and Computation Offloading. IEEE
Thoben K-D, Busse M. Cloud-based Automated Design Communications Surveys and Tutorials, 2017, 19:
and Additive Manufacturing: A Usage Data-enabled 1628–1656
Paradigm Shift. Sensors, 2015, 15: 1–17 108. Wang Y, Hong K, Zou J, Peng T, Yang H. A CNN-based
98. Mokhtarian H, Coatanéa E, Paris H, Mbow MM, Pourroy F, visual sorting system with cloud-edge computing for
Marin PR, Vihinen J, Ellman A. A Conceptual Design and flexible manufacturing systems. IEEE Transactions on
Modeling Framework for Integrated Additive Manufacturing. Industrial Informatics, 2019, 1–10
Ac

Journal of Mechanical Design, Transactions of the ASME, 109. Wang Y, Liu M, Zheng P, Yang H, Zou J. A smart surface
2018, 140: 1–13 inspection system using faster R-CNN in cloud-edge
99. Majeed A, Lv J, Peng T. A framework for big data driven computing environment. Advanced Engineering Informatics,
process analysis and optimization for additive 2020, 43: 101037
manufacturing. Rapid Prototyping Journal, 2018,

Downloaded to IP: 10.159.164.174 On: 2020-04-01 06:16:03 http://engine.scichina.com/doi/10.1007/s11431-020-1581-2

View publication stats