Additive Manufacturing 52 (2022) 102691

Contents lists available at ScienceDirect

Additive Manufacturing
journal homepage: www.elsevier.com/locate/addma

Research and application of machine learning for additive manufacturing

Jian Qin a, b, 1, Fu Hu a, 1, Ying Liu a, *, Paul Witherell c, Charlie C.L. Wang d, David W. Rosen e,
Timothy W. Simpson f, Yan Lu g, Qian Tang h
a
Department of Mechanical Engineering, School of Engineering, Cardiff University, Cardiff CF24 3AA, UK
b
Welding Engineering and Laser Processing Centre, Cranfield University, Cranfield, UK
c
Systems Integration Division, National Institute of Standards and Technology (NIST), Gaithersburg, MD 20899, USA
d
Department of Mechanical, Aerospace and Civil Engineering, University of Manchester, Manchester, UK
e
The George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA
f
Mechanical Engineering, The Pennsylvania State University, University Park, PA, 16802, USA
g
Engineering Laboratory, National Institute of Standards and Technology (NIST), Gaithersburg, MD 20899, USA
h
State Key Laboratory of Mechanical Transmissions, Chongqing University, Chongqing, China

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

Keywords: Additive manufacturing (AM) is poised to bring a revolution due to its unique production paradigm. It offers the
Additive manufacturing prospect of mass customization, flexible production, on-demand and decentralized manufacturing. However, a
3D Printing number of challenges stem from not only the complexity of manufacturing systems but the demand for
Rapid prototyping
increasingly complex and high-quality products, in terms of design principles, standardization and quality
Machine learning
Deep learning
control. These challenges build up barriers to the widespread adoption of AM in the industry and the in-depth
Digital manufacturing research of AM in academia. To tackle the challenges, machine learning (ML) technologies rise to play a crit­
Intelligent manufacturing ical role as they are able to provide effective ways to quality control, process optimization, modelling of complex
systems, and energy management. Hence, this paper employs a systematic literature review method as it is a
defined and methodical way of identifying, assessing, and analysing published literature. Then, a keyword co-
occurrence and cluster analysis are employed for analysing relevant literature. Several aspects of AM,
including Design for AM (DfAM), material analytics, in situ monitoring and defect detection, property prediction
and sustainability, have been clustered and summarized to present state-of-the-art research in the scope of ML for
AM. Finally, the challenges and opportunities of ML for AM are uncovered and discussed.

1. Introduction processing parameters, process stability, and working conditions. This

leads to the challenges that are summarized and highlighted as follows.
Additive manufacturing (AM), also known as rapid prototyping, 3D
printing, and freeform fabrication, is capable of depositing, joining or • It is generally difficult to model the mathematical relations of the
solidifying materials to construct physical objects from computer-aided underlying AM process because the correlated factors are from
design (CAD) models [1]. Compared with conventional manufacturing various heterogeneous perspectives and different process stages.
methodologies, such as subtractive manufacturing and formative • High-fidelity physical-based models are generally too complicated
manufacturing, AM systems show higher efficiency and flexibility considering the in-process uncertainties of the AM process, which
within the high-yield production and offer a new perspective for the demand significant computational resources.
design and processing of both parts and materials. However, the AM • It is challenging to integrate AM digital models, pertinent to various
process is well-known as a highly complex system including various phenomena, at multiple scales into one unified framework [2].
technologies that combines material science, mechanics, optics, and
electronics with computer science. As a result, the quality of produced The applications of machine learning (ML) technologies have been
parts is affected by numerous factors, such as material properties, proved effective in a wide range of fields, such as computer science,

* Corresponding author.
E-mail address: [email protected] (Y. Liu).
1
These authors are the first authors and have contributed equally to this work.

https://doi.org/10.1016/j.addma.2022.102691
Received 6 October 2021; Received in revised form 6 February 2022; Accepted 15 February 2022
Available online 18 February 2022
2214-8604/© 2022 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
J. Qin et al. Additive Manufacturing 52 (2022) 102691

aviation, healthcare, and the manufacturing industry [3]. With the using diverse ML technologies. To clarify the significant research chal­
advancement of data acquisition and storage technologies, data-driven lenges and future opportunities of ML for AM, a comprehensive review is
approaches based on ML technologies have been increasingly adopted necessarily crucial to summarise and analyse current research topics.
to discover hidden knowledge and build highly complex relationships in There have been some existing review articles focusing on different
digital manufacturing systems [4]. By using reliable datasets, ML models perspectives of this topic [5], such as the ML for the material develop­
are capable of learning hidden patterns and uncovering latent knowl­ ment of AM [6], and ML applications in laser powder bed fusion pro­
edge to support decision-making, in terms of process optimization, cesses [7]. Additionally, the highlighted papers in these review articles
quality control, and system improvement. As one of the most popular were selected and reviewed based on authors’ research and industry
manufacturing systems in Industry 4.0, AM has been incorporated with experience, which is valuable but subjective. In order to discover the
digital systems and sensor networks where high-volume data can be exhaustive challenges and opportunities in this increasingly growing
obtained. Hence, a growing number of researchers have applied ML research field, a systematic and data-driven review method is needed.
algorithms to tackle challenges in AM, such as design optimization, in This paper aims to review, summarise, analyse and present the latest
situ monitoring, process modelling, and energy management. However, research and applications of ML for AM. Section 2 introduces the state-
different researchers and organisations focus on various AM issues by of-the-art of the AM process, including seven main AM system categories

Table 1
AM process categories.
AM process Working principle Material Material Material State of fusion Represented technology
feedstock distribution

Binder Jetting[15] Polymer Liquid Print head Chemical reaction Binder jetting
bonding

Material extrusion[16] Polymer Filament Deposition Thermal reaction bonding Fused filament
nozzle fabrication (FFF)

Directed energy Metallic Filament/ Deposition Melted state: (electric Wire + arc additive
deposition[17] Powder nozzle beam/arc /laser) manufacturing (WAAM)

Material jetting[20] Polymer Liquid Print head Chemical/Thermal Drop on demand (DOD)
reaction bonding

Powder bed fusion[19] Polymer/ Powder Powder bed Melted state/Laser/ Solid- Selective laser melting
Metallic/ state (SLM)
Ceramic

Sheet lamination[21] Polymer/ Sheet Sheet stack Solid state: (Ultrasound)/ Ultrasonic additive
Metallic Chemical reaction manufacturing (UAM)
bonding

Vat Polymer Liquid Vat Chemical reaction Digital light processing

photopolymerization bonding (DLP)
[22]

2
J. Qin et al. Additive Manufacturing 52 (2022) 102691

and their characteristics. In Section 3, the methodology of systematic are capable of manufacturing large parts with submillimetre details
review is introduced. Based on the results of the systematic review, this and are widely applied in the coating and printing industry [22].
paper also applies a data-driven method to analyse notable articles, a
keyword co-occurrence and clustering method. Several aspects of AM The processing characteristics of each AM process is specific and can
issues are clustered based on these notable articles which are reviewed be allocated into different production scenarios. For example, the
and summarised in Section 4. The challenges and opportunities are WAAM, as a variation of DED technology normally uses an arc-based
discovered and clarified in Section 5. Section 6 concludes. heat source, such as a plasma or Metal Inert Gas deposition [23], to
melt and fuse metal materials for constructing parts layer by layer. It has
2. AM: the state-of-the-art recently attracted great attention in the industry, especially the aero­
space industry, as it is capable of producing large metal parts with a high
The first commercial AM system was recognizably emerged in 1987 deposition rate, low equipment cost, high material utilization [24].
with stereolithography (SL) by 3D Systems [8]. Since then, AM has
become one of the most crucial manufacturing solutions across various 3. Research methodology
industries, such as automobile [9], aerospace [10], and construction
[11]. According to the Wohlers Report 2020, the global market size of To investigate the research and application of ML for AM, a sys­
AM industry is over USD 11 billion in 2019 and will increase to over USD tematic literature review and text mining analysis are adopted for
35 billion by 2024 [12]. There are currently 7 AM process categories, identifying, assessing, and analysing the literature published between
published in the Standard Terminology for Additive Manufacturing 2000 and 2020. The overall methodology is illustrated in Fig. 1. In the
Technologies, which are Binder Jetting, Directed Energy Deposition, first place, a systematic literature review is adopted to search, select, and
Material Extrusion, Material Jetting, Powder Bed Fusion, Sheet Lami­ assess relevant publications. The systematic literature review is defined
nation and Vat Photopolymerization [13]. Table 1 shows the details of as a systematic, explicit and reproducible method for identifying, eval­
these AM process categories. These seven categories focus on the uating, and synthesizing the existing body of completed and recorded
single-step process, and for the multi-step AM process, the system work produced by researchers [25]. The review process typically in­
combines two or more AM processes [14]. volves several main steps [26], including specifying research questions,
identification of research, selecting and assessing the collected publi­
• Binder Jetting. Binder jetting process is one of the earliest AM cations. In the article selection and assessment processes, explicit
processes developed for polymer powder-based material. An inkjet exclusion and inclusion criteria are required to assess each potential
print head is used to spray the liquid binder onto the polymer primary study. After the systematic literature review is conducted, the
powder. The powder material is solidified crossing the section of keyword co-occurrence and clustering analysis for selected publications
produced part layer by layer by the chemical reaction bonding at a are applied to make a comprehensive overview of the main research
reasonable speed [15]. topics and directions in the applications of ML for AM.
• Material Extrusion. Material extrusion (ME) AM process is
currently the most prevailing AM process in the market. The mate­ 3.1. Research questions
rials of extrusion-based AM systems are normally forced out in a
semisolid state via a nozzle where constant pressure is applied. Then The research questions for conducting the systematic literature re­
the extruded materials solidify and bond to the previous extruded view are set up and presented in Table 2 with its motivations.
materials to form a solid structure [16].
• Direct Energy Deposition. Direct energy deposition (DED) pro­ 3.2. Search strategy
cesses utilize focused energy, such as a laser beam, electron beam, or
plasma arc, to melt and fuse simultaneously the substrate and the 3.2.1. Search terms identification
material that is being deposited into the substrate’s melt pool to The search strategy designed in this work includes keyword identi­
construct parts [1]. Powder-based and wire-based materials can be fication, resources for the searching, searching process, and criteria for
used for DED processes [17]. article selection, to collect available and qualified published articles
• Material Jetting. Material jetting (MJ) is another fast AM process, relevant to the topic [26]. The search query using Boolean operators is
which uses ultraviolet (UV) light as the main power to solid-liquid shown in Fig. 2.
photopolymer droplets. The droplets are controlled by the voltage The search terms are modified by reducing synonyms while search­
signal. Through the print head, the liquid or melted material is jetted ing in some databases (e.g., IEEE Xplore) due to the search term
onto the produced part surface [18]. limitation.
• Powder bed fusion. The powder bed fusion (PBF) processes consist
of thin layers of fine powders, which are spread and closely packed 3.2.2. Resources for searching
on a platform. One or two thermal sources are employed in the The searching of relevant articles is conducted by inserting keywords
systems to melt and fuse material powder particles in each layer. (search terms) in five databases, including ACM digital library, IEEE
Subsequent layers of powders are spread across the previous layers Xplore digital library, Science Direct, Springer Link, and Scopus. These
using a roller and then fused together until the entire product is built. databases are the most representative databases of scientific research
Selective laser sintering (SLS), selective laser melting (SLM), and that is closely related to the topic of this review and contains a massive
powder-based electron beam melting (EBM) are the commonly used volume of literature, such as journal papers, conference proceedings,
PBF technologies [19]. and books.
• Sheet lamination. 2 AM technologies are commonly classified into
the class of sheet lamination (SL) AM process, ultrasonic additive • ACM digital library (https://dl.acm.org)
manufacturing (UAM) and laminated object manufacturing (LOM). • IEEE Xplore digital library (http://ieeexplore.ieee.org)
The sheets of the material are bonded together as the part based on • ScienceDirect (http://www.sciencedirect.com)
the fusion resource, like ultrasonic [21]. • SpringerLink (https://link.springer.com)
• Vat Photopolymerization. Vat photopolymerization (VP) processes • Scopus (https://www.scopus.com)
mainly use ultraviolet (UV) to cure or harden materials such as
photopolymers, liquids, resins for building products. These processes 3.2.3. Inclusion and Exclusion Criteria for Article Selection
Based on the directions of exploration and motivations presented in

3
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Fig. 1. The overview of the research methodology.

a) Articles that were focused on other technologies rather than machine

Table 2
learning for tackling issues in AM.
Research questions posed for the systematic literature review.
b) Articles that were focused on other manufacturing systems rather
Research Question Motivation than AM systems by using ML technologies.
RQ1: What kind of issues in AM can be To explore state-of-the-art of AM c) Articles that were not written in English.
currently tackled by using ML challenges that can be tackled by using
technologies? ML technologies. Inclusion Criteria, sources that met the constraints reported below
RQ2: What ML algorithms have been To present the ML algorithms adopted by
adopted to tackle the challenges? previous studies to tackle challenges in
were included from our study:
AM. Answer this question will help
researchers understand and select a) All the articles, written in English, reporting machine learning
appropriate ML algorithms for dealing technologies for tackling AM issues.
with similar issues in AM.
b) Articles that introduce new techniques to improve the performance
RQ3: What are the advantages and To study the reasons and limitations for
limitations of using ML algorithms to adopting a certain ML algorithm to of existing machine learning technologies used for AM.
solve the issues in AM? tackle a specific research challenge in
AM. Answer this question will assist 3.3. Article Selection and Assessment Process
researchers to make the trade-off
between ML-based methods and
conventional methods, as well as
The framework for the article selection process is illustrated in Fig. 3.
exploring modified strategies. The searching process starts with searching publications from the pre­
defined databases using the Boolean operator based on the identified
keywords (10,054 publications were included). Then these publications
were filtered through an initial selection process based on the proposed
exclusion criteria where 4212 publications remained. The publications
were selected by the proposed inclusion criteria and 277 papers were
included. A manual search process is adopted to search for additional
sources related to the review topic [27,28], where the exclusion and
Fig. 2. The search query for searching publications.
inclusion criteria were re-applied. Twenty-five publications were
selected by the manual search process. In the quality assessment stage,
the quality of the selected publications is assessed by the following
Table 2, the listed following are the exclusion and inclusion criteria for
criteria and 228 publications were finally obtained.
the article selection process. The exclusion criteria are applied in the
title, abstract, and keyword list of a publication while inclusion criteria
• In the paper, the focused issues in AM should be clearly defined.
are applied in full text.
• The motivations for employing the ML algorithm to tackle the issues
Exclusion Criteria, sources that met the constraints reported below
should be clarified.
were excluded from this study:

4
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Fig. 3. The framework of the proposed article selection process.

• The evaluation or validation of the proposed methodology should be contents of a study. The two keywords are counted as one co-occurrence
presented. when they occur together in the publications. The clustering approach,
introduced by Waltman et al. [30], is employed to group these keywords
3.4. Keyword co-occurrence and clustering analysis for selected articles based on the co-occurrence matrix [31]. Considering a bibliometric
network of n nodes, the main concept of this clustering approach is to
The analysis of keyword co-occurrence can provide an effective way assign the keywords into n nodes and group these n nodes into k clusters.
to reflect the research hotspots in the research fields through analysing Based on the co-occurrence matrix, the similarity sij, also known as the
the distribution of keywords [29]. In the keyword co-occurrence anal­ association strength [32], between keyword i and j can be calculated as:
ysis, the keywords can be words or phrases that are extracted from the
title, abstract, and keyword list in a publication for representing the core

Fig. 4. The clustering results based on the keyword co-occurrence for the selected publications.

5
J. Qin et al. Additive Manufacturing 52 (2022) 102691

sij =
cij
(1) Table 3
wi wj Clustering results for selected publications.
Clustered Keywords Research Domain
where cij denotes the number of co-occurrences of keyword i and j, and
where wi and wj denote the total number of occurrences of word i and j. Cluster Complex geometry, geometry, layer DfAM
1 thickness, parameter
The clustering method is to minimize V (x1,…, xn): Cluster Powder, metal powder, shape, class Material analytics
∑ ∑ 2
V(x1 , ., xn ) = sij d2ij − dij (2) Cluster Image, defect, detection, classification, Image-based defect
3 convolutional neural network, porosity, detection and monitoring
i<j i<j

⃒⃒ ⃒⃒ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ pore, fusion

⃒⃒ ⃒⃒ ( )2 Sensor, process monitoring, acoustic Sensor signal-based defect
dij = ⃒⃒⃒⃒xi − xj ⃒⃒⃒⃒ = xi − xj (3)
emission, real-time, surface roughness, detection and monitoring
predictive model, crack, defect detection
In Eqs. (2) and (3), xi represents a positive integer that indicates a Cluster Physics, property, microstructure, Process modelling and
cluster to which node i belongs, and dij denotes the distance between 4 simulation, mechanical property, laser control
power
node i and j. By minimizing the Eq. (2), this clustering method can be
Cluster Energy consumption Sustainability
interpreted as that the higher the association strength of two nodes, the 5
stronger the relatedness between the nodes. The results of applying the
clustering algorithm [30] based on the keyword co-occurrence for the
selected publications are obtained by using VOSviewer [31], profes­
Table 4
sional software for bibliometric analysis.
The links and total link strength information of the top 10 occurrence keywords.
In Fig. 4, the research hotspots are clustered into 5 main clusters that
Keywords Cluster Number Occurrences Links Total Link Strength
are represented by different colours (i.e., purple, blue, yellow, green,
and red). According to the clustering algorithm described previously, Parameter 1 87 60 630
the keywords close to each other indicates strong relatedness. Two Image 3 58 46 619
Monitoring 3 43 52 434
keywords connected by a link represents the co-occurrence of them in Defect 3 41 52 415
the publications. For each cluster, the research domain is defined by the Property 4 41 44 329
understanding of the clustered keywords. The keywords in the green Porosity 3 39 53 547
cluster such as complex geometry, geometry, layer thickness, and Classification 3 34 47 324
Sensor 3 34 48 393
parameter are related to the topics of DfAM. However, the keyword (i.e.,
Simulation 4 32 32 178
energy consumption) relevant to the research of AM sustainability is also Detection 3 27 34 187
displayed in this cluster. This may be due to the increasing trend that
researchers study AM sustainability from the design perspective.
The research domain of the cluster in blue is defined as image-based research. In addition, most studies focus on process modelling and defect
defect detection and monitoring, where the keywords are image, defect, detection domain. Based on the keyword co-occurrence and clustering
detection, classification, porosity, pore, etc. Similarly, the cluster in red analysis, the articles under different research domains are reviewed and
is defined as sensor signal-based monitoring, where the keywords are discussed in the next section.
sensor, acoustic emission, process monitoring, defect detection, real-
time, etc. The implementation of process monitoring is based on the 4. ML for AM
success of detecting defects during the manufacturing processes. Hence,
these two clusters (blue and red) can be merged into one cluster, rep­ 4.1. ML for DfAM
resenting the same research domain but with different strategies. The
keywords displayed in the yellow cluster are physics, property, micro­ AM has provided opportunities for innovative designs and advances
structure, simulation, mechanical property, laser power, etc. These in product performance, in terms of geometric freedom and highly in­
keywords are closely related which indicates that the studies mainly tegrated structures [33]. Due to its unique production paradigm, the AM
focus on investigating and modelling the relationships between pro­ processes may involve different batch sizes, production times, and cost
cessing parameters and their resulting performances. For the cluster in drivers compared with conventional processes. It also requires different
purple, the keywords are powder, mental powder, shape, and class approaches to metrology and quality control. Therefore, DfAM has been
which can be defined as material analytics. To present the clustered proposed as a way to provide AM design professionals with a wide range
keywords with their corresponding research domains more reasonably, of design and analysis tools for complex part structures and AM pro­
the keywords in each cluster are polished and re-clustered manually cesses. Typically, DfAM includes two main research topics, part design
based on domain understanding, shown in Table 3. The general termi­ and design optimization [34]. For part design, AM creates free forms and
nologies in the clustered keywords (e.g., additive manufacturing, 3D customized geometries, enabling the creation of complex internal fea­
printing, technology, model, etc.) are removed from each cluster and not tures to increase functionality and improve performance of target parts,
displayed in the table. which provides designers with huge design space. For design optimi­
Table 4 shows the link and total link strength information of the top zation, AM part designers need to determine production path strategies,
10 occurrence keywords (the general terminologies are removed). The part locations, build orientations, and support structures for improving
occurrences represent the total number of a keyword that occurs in the the quality of final printed products. Due to the advances of artificial
selected publications and a link is counted when two keywords occur intelligence and available data, ML technologies have been increasingly
together in a publication. For a keyword, the total link strength is the applied to DfAM in recent years [35].
sum of its association strengths, indicating the relatedness of the
keyword with other keywords. In other words, the higher the total link 4.1.1. Part design
strength, the stronger the relatedness. At the conceptual design phase, most AM designers select appro­
According to Table 4, the quality issues (keywords: monitoring, priate design features based on their knowledge and experience. How­
defect, property, porosity) of the AM produced products is the biggest ever, there is a lack of systematic and intelligent techniques to assist AM
concerns that attract most researchers’ attention. It is worth noting that professionals to explore AM-enabled design space [36,37]. Hence, Yao,
the parameter is the most commonly considered attribute in AM et al. [36] introduced a hybrid ML approach for design features

6
J. Qin et al. Additive Manufacturing 52 (2022) 102691

recommendation at the conceptual design phase in AM. In the paper, the for geometry corrections of the designed lattice infill patterns in FDM
authors classified the functionality-centric design knowledge inherent in systems. In this work, the 3D coordinates of the designed infill structures
AM design features and target components into ‘loadings’, ‘objectives’ were used as the input of the ANN model, while the symmetrical devi­
and ‘properties’, which were coded with numerical digits and saved in ation surface coordinates were the output. After training, the model was
database files. Then hierarchical clustering was carried out on the coded implemented to the STL file for geometric corrections., Considering the
design knowledge to reveal the relationships among design features and manufacturability of AM process, Tang et al., [41] proposed a strategy
target components, resulting in a dendrogram. Previous industrial for lattice structure design and optimization to ensure products quality.
application examples with their design features implementation were To represent the design space of lattice structures, the concept “physical
simplified as a binary classification problem (implemented design fea­ entity” was introduced to include the design information (geometrical
tures denote as ‘+1′ , otherwise ‘− 1′ ) and trained by a support vector information and material information) for each design stage. Then, the
machine (SVM) classifier. The trained SVM model was used to refine the lattice unit cell model was defined and proposed to represent the to­
hierarchical clustering results by an SVM-based progressive dendrogram pology of the elements inside lattice structures. The authors used the
cutting process, which aims at identifying the final sub-cluster con­ concept of “manufacturable element” to include the geometry, material,
taining the recommended AM design features. Through the case study and process information of a lattice strut. ANN was used to bridge the
results, the proposed hybrid ML approach was demonstrated useful in relationship between manufacturability and geometrical data.
identifying appropriate AM design features for inexperienced designers. Integrating the physical and domain knowledge in ML algorithm has
Neural network is another popular ML technology used to improve the been proven which can improve the ML performance significantly [42].
part design in AM processes. Ko et al., [43] introduced a methodology for bridging the gap between
Andrew and Markus [38] introduced a method that uses variational multi-discipline designs and AM capabilities based on knowledge graph
autoencoders (VAE) and ML techniques for the compliance optimization ML. In this work, the framework of the proposed method, shown in
of cantilever design. In this work, the cantilever structures were encoded Fig. 5, consists of 4 main modules, including (1) AM prior knowledge
into a 2D latent space by VAE and the long short-term memory (LSTM) structuration, (2) transformation of knowledge to DfAM ontology, (3)
neural network was adopted to learn the latent space trajectories that extraction of knowledge from AM data using ML, and (4) design rules
correspond to the topology optimization process. The results showed transformation. The design fundamentals, principles, and rules were
that the VAE-LSTM model was capable of generating complex structures obtained by formalizing unstructured AM prior knowledge into struc­
and evolving them. A framework of using neural networks for the tured knowledge and extracting knowledge from AM data based on ML
analysis and design of micro-lattices architectures in AM was introduced algorithms. Then, these designs knowledge was encoded into ontologies
by Nathaniel [39]. In this study, to obtain training datasets, the authors with knowledge graphs. Finally, the design rules were constructed by
used a compact genetic algorithm (cGA) to generate micro-lattice reasoning with prior knowledge and newly discovered knowledge.
structures of which the corresponding mechanical properties were ob­
tained by finite-element analysis (FEA). The graph convolutional net­ 4.1.2. Design optimization
works (GCN) with an asymmetric auto-encoder was adopted and trained To obtain the required production quality, design optimization is a
by the graph representation of the generated micro-lattices. Specifically, critical step before the AM process begins [44]. Many crucial elements
through the training process, the encoder was able to learn the physical and parameters are defined in this step. For instance, the determination
characteristics of micro-lattices and infer their mechanical properties. of build orientation and direction significantly affects process and
Then the decoder was used for generating the micro-lattices structures fabrication attributes [45]. In Ref [44], the authors applied K-means
with specified mechanical properties. According to the empirical results, clustering with Davies–Bouldin Criterion cluster measuring on surface
the encoder had an accuracy of 93.72% in predicting the mechanical models to generate alternative build orientations in a computationally
properties of the given micro-lattices structures. The decoder was efficient way. The K-means clustering method was adopted to decom­
demonstrated to be capable of generating the micro-lattices from the pose stereolithography (STL) models into K facet clusters where the
specified mechanical properties. Another example of applying neural number of clusters was determined by Davies–Bouldin Criterion. The
networks is that Jonnel et al., [40] used artificial neural networks (ANN) central normal vectors of each facet normal cluster were used as

Fig. 5. The proposed data-knowledge-design rule framework. This framework consists of 4 main modules, including AM prior knowledge structuration, trans­
formation of knowledge to DfAM ontology, extraction of knowledge from AM data using ML, design rules transformation, Ref [43].

7
J. Qin et al. Additive Manufacturing 52 (2022) 102691

alternative build orientations where the optimal orientation was ulti­ of the proposed methodology with prediction accuracies over 90%.
mately obtained by a statistical evaluation process. To prevent unsightly An automated geometric shape deviation modelling approach based
surface artefacts or damages of fine surface details when removing on Bayesian neural networks (BNN) and transfer learning techniques for
support structures, a perceptual model of preference in the printing di­ different shapes and AM processes was proposed by Ferreira et al. [52].
rection of AM was proposed by Zhang, et al. [46]. The authors developed In this approach, the geometry shapes are defined under the polar co­
a perceptual model to determine the preference of printing orientation ordinate representation, where each point on a product is identified by
in terms of area of support, visual saliency, preferred viewpoint, and an angle θ. The in-plane and out-of-plane deviations of different shapes
smoothness preservation. and processes are represented by statistical models. A baseline BNN for
Support structures are also important and required in some AM modelling shape deviation was firstly built by training on a small
processes if the designed models contain separated segments or over­ number of product samples under a specific AM process. Then transfer
hang parts in a layer where does not exist solid material underneath. To learning techniques were employed to transfer the baseline model to
find the minimum amount of support structures for successfully fabri­ new shapes and processes. A case study was carried out under different
cating a model, Huang et al. [47] developed a support detection SLA processes, where the proposed model yielded good performance in
approach based on a surfel convolutional neural network (surface an automated manner. This study provides insights of automatically
element - CNN) in AM. In this method, the surfel is the sampling point on leveraging data and models from different processes, addressing the
the surface with normal information, defined through layered challenges of modelling for various shapes produced by distinct AM
depth-normal image (LDNI) [48] sampling method. The LDNI stores processes.
array of rays that are shot to intersect with the CAD model, where the Tootooni et al., [59] introduced a method to classify the dimensional
depth and normal values of intersection points on the rays are included. variation of AM produced products based on spectral graph theory and
Based on LDNI sampling, local surfel images with ground-truth support ML techniques. It is interesting that this work extracted the spectral
regions were obtained and fed into the CNN model for classification. The graph Laplacian eigenvalues from the 3D point cloud data of the man­
experimental results indicated that the proposed methodology out­ ufactured parts and used them as features in ML models for classifica­
performed the normal-based method and image-based method in terms tion. This work provides a solution to reduce the measurement burden
of support detection. It is highlighted in the paper that, due to the for post-process quality assurance. For shape deviation compensation,
topology-preserving and salient feature extracting capability, the surfel Shen et al. [60] introduced a framework for AM using a convolutional
CNN model is more robust for support detection on extreme features neural network (CNN). In this framework, the 3D model was encoded as
than the traditional image-based method. a binary probabilistic distribution in 3D space and fed into the CNN
AM has been increasingly employed for printing composite material model for capturing deformation features. An inverse function network
parts. However, the fiber size, volume fraction and direction are was trained for obtaining the compensated model.
important in determining the properties of the printed part. Kaushik
et al. [49] introduced a method for reversing additive manufactured 4.2. ML on material analytics for AM
composite parts by toolpath reconstruction of the printing process using
the LSTM network. In this method, the CT-scan images of fibre orien­ A variety of materials, such as metals, ceramics, plastics, and their
tation at each layer were sequentially fed into the LSTM model for combinations are used for AM applications and the development of new
predicting the orientation angle of fibres. Then the G-code can be materials is in progress [61]. Using different materials for producing
generated based on the fibre orientation and measured layer thickness. products can result in different performances and properties. A signifi­
The method developed in this study showed the effectiveness of the ML cant amount of data can be generated from material property and con­
model in identifying and tracking any given orientation of the fibres. It ditions. It is essential to analyse and understand the relationships among
also demonstrated the possibility of reconstructing the G-code and material chemistry, material characteristics and final part performances
reverse engineering any composite part for improving the properties of based on the material data. The powder property is one of the key ele­
the printed parts. ments that affect the build process and final part quality in
powder-based AM [62]. During the printing process, the interaction and
4.1.3. Shape deviation consolidation between powder particles are complicated where
Due to the functionality and manufacturing requirements, shape high-quality powders are required to ensure the process reliability and
accuracy measurement is essential and critical in DfAM, aiming to final part property. Thus, qualifying powder materials is critical, and
reduce the geometrical deviations of the final products [50]. In general, some researchers have made efforts in measuring and analysing powder
during an AM process, the geometries of final products are affected by materials by using ML technologies.
various factors, such as material properties, thermal gradients, and build Powder characterization is important for evaluating the quality of
orientations, which lead to the low quality of printed parts. Hence, the powder materials in AM, where computer vision and ML technologies
geometrical inaccuracies of the produced products pose significant have been applied for autonomous characterization [63,64]. DeCost,
challenges to predictive modelling of shape deviations and developing et al. [64] introduced a method that used key-point based computer
error compensation strategies for AM. Several researchers have explored vision for quantitatively characterizing powder materials. In this study,
that ML models are used for tackling geometrical accuracy-relevant is­ eight powders that only differed in their particle size distribution were
sues, such as shape deviation prediction [50–57], classifying and considered. The authors employed a computer graphics suite, called
quantifying geometrical accuracy [58,59], and deviation compensation Blender, to generate synthetic powder micrographs. A bag of visual
[60]. In the studies [53–56], ANN was adopted to model the relationship words (BOVW) image representation was adopted for characterizing the
between process parameters and geometry-related errors in different synthetic powder micrographs, where the images were represented by
AM processes. key-point features and organized into a visual dictionary. Then the
Zhu et al. [50] proposed an ML-based method to model in-plane difference of Gaussians (DoG) and Harris-LaPlace (HL) interest point
deviation and random local variant in AM. A mathematical relation­ detectors were used to select critical key-point features. The regions
ship between the designed shape and the final shape was constructed surrounding the key-point features were characterized by applying scale
from a transformation perspective, aiming at capturing the global trend invariant feature transform (SIFT). Finally, the support vector machine
of shape deviations. Due to unexplained variations with complex pat­ (SVM) algorithm was used for the particle size classification of which the
terns, a multi-task Gaussian process (GP) learning algorithm was accuracy was 0.894. Compared with conventional characterizing
adopted to learn from the unexplained deviation data and model the methods, this study provided an alternative method to characterize
local deviation. The experimental results demonstrated the effectiveness feedstock powders based on ML.

8
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Vrábel, et al. [65] also adopted SVM to classify Al alloy powder SLM process. Additionally, a modification of the CNN model developed
materials for the SLM process. In this work, the authors used the by Scime and Beuth [83], called multi-scale CNN (MsCNN), improves
laser-induced breakdown spectroscopy (LIBS) technique to obtain the flexibility and overall classification accuracy of the conventional
spectra from the powder materials. Then the spectra were processed CNN model in autonomous anomaly detection. The proposed MsCNN
through unit vector normalization and principal component analysis methodology has been demonstrated to be robust when analysing builds
(PCA). The PCA model was applied to reduce dimensionality and that were manufactured by different materials, including AlSi10Mg,
remove noise data, where four principal components (PCs) were ob­ bronze, Inconel 625, Inconel 718, stainless steel 316 L, stainless steel
tained. These PCs were fed into the SVM model for material classifica­ 17–4 PH, and Ti-6Al-4 V, in the L-PBF systems. Lee et al. [96] adopted
tion. Powder flows significantly affect the deposition behaviour of the 3D-CNN and CNN-LSTM for the classification of different statuses,
layers on the substrate. Richard et al., [66] used the decision trees (DT) including damaged, cured, and uncured, in the two-photon lithography
algorithm to classify powder flowability based on particle-level physical (TPL) process.
property measurements in cold spray AM, achieving an accuracy of Some researchers have also explored image-based monitoring ap­
98.04% in classification. proaches with their corresponding control strategies [97,98]. Wang
Currently, ML technologies are generally used for classification tasks et al. [98] presented a closed-loop control framework by seamlessly
for powder material analysis. There still need further exploration and integrating vision-based techniques and neural networks (NN) to detect
research to achieve the potential of leveraging ML for material analytics droplet phenomena and accordingly implement control strategies in
in AM, such as analysing material composition for alloy development liquid metal jet printing (LMJP) processes. In this work, a
and modelling the relationship among material chemistry, material charge-coupled device (CCD) camera was employed in the monitoring
properties and final part performances. system to capture jetting images. The key features were extracted from
jet images by a flood-fill algorithm [99] and summarized to the prop­
4.3. Defect detection and in situ monitoring for AM based on ML erties of the droplet patterns (i.e., satellite, ligament, volume, and speed
of droplets). The complex relationship between droplet features and
Lack of quality assurance in AM produced parts is one of the key voltage level was modelled by NN of which the architecture is shown in
technological barriers that prevent manufacturers from adopting AM Fig. 6(a). This NN model enabled the conversion of real-time droplet
technologies, especially for high-value applications where component features to voltage values and a proportional integral derivative (PID)
failure cannot be tolerated [67]. There still lacks effective and mature process was used to adjust the drive voltage by comparing the output
monitoring technologies in AM systems for detecting the onset of defects values of NN with target values for process control. Results from
in real-time and keeping the stability of the process in control [68]. Due empirical tests indicated that more stable jetting processes can be real­
to different material supplies and working principles of different AM ized in real-time through the proposed methodology. Fig. 6(b). shows
processes, the defects or quality issues can be various. For instance, the the control processes for offset jetting stabilization.
issues of porosity, lack-of-fusion, balling, crack are critical in the Besides, Jin et al. [100] developed a real-time monitoring and
powder-based processes [67,69,70] and the geometry deviation [71], autonomous correction system for FDM processes. A CNN classification
shape shrinkage [72], and surface roughness [73–75] in FDM processes model was used for detecting defects and a feedback loop was used to
have been focused by many relevant research groups. Only when these modify processing parameters. In their paper, three categories,
defect issues are detected synchronously and accurately during the AM including ‘Good-quality’, ‘Under-extrusion’, and ‘Over-extrusion’, were
process, the real-time control strategies can be realized. With the adopted to represent the quality of the printed parts and 120,000 images
advancement of data acquisition, communication, and storage technol­ were used for training and testing. The adjusting commands were sent to
ogies, ML technologies have been increasingly used for in situ moni­ the control system to modify the flow rate of the AM system when five
toring in AM systems [76,77]. The ML models are trained by different successive over or under-extrusion judgments were made by the CNN
types of data which are classified into three categories, including model. Experimental results indicated that the proposed model achieved
one-dimensional 1D data (e.g., spectra), 2D data (e.g., images), and 3D above 98% accuracy in predicting the part quality and the response rate
data (e.g., tomography) [78]. Each strategy developed in existing studies of the system from defect-recognition to correction reaches or even
has pros and cons. In general, two main types of strategies, image-based surpasses the human reaction (9 s). However, the key parameters for the
and sensor signal-based, are adopted for defect detection and in situ adjustment are selected based on the authors’ understanding of the
monitoring in AM. Strategies that leverage 3D point cloud data with ML processes in both Ref [98,100]. For instance, in Ref [100], the adjust­
models have also been explored in recent studies [79,80]. ment of flow rate was focused in this paper, while other parameters such
as printing speed and nozzle height were not considered.
4.3.1. Image-based approach Apart from CNN-based models, SVM [101,102], Bayesian classifi­
Visual camera images can present the surface characteristics of every cation [103], deep belief networks (DBN) [104], deep neural networks
build layer to reflect the quality of the AM produced parts. Zhang et al. (DNN) [105], k-means singular value decomposition (K-SVD) [106]
[81] developed a vision system with a high-speed camera to capture the have also been employed to analyse image data for quality inspection in
sequential images for PBF process monitoring. Their research focused on AM systems. However, in Ref [101–103], key features of visual images
detecting the information of melt pool, plume, and spatter. Features of need to be extracted before being fed into the ML models. In Ref [104],
these objects were extracted based on the understanding of the physical Ye et al. proposed an in situ monitoring method based on analysing
mechanisms. These features were then selected by PCA before being plume and spatter signatures during the SLM process. The plume and
used as inputs for SVM classification. The performance of the SVM spatter images were obtained by a high-speed near-infrared (NIR)
model showed an accuracy of 90.1% for quality level classification. This camera and normalized by zero mean and unit variance to capture the
work also demonstrated that CNN is promising to achieve real-time pair-wise interactions between pixel values. These processed data were
monitoring as it has an accuracy of 92.7% without the feature selec­ then fed into the DBN structure with four-level hidden restricted
tion process. CNN is capable of learning fairly representative features Boltzmann machines (RBMs) for classifying 5 distinct melted states (i.e.,
from the raw data automatically and is commonly used for image over-melted, middle-over melted, normal melted, middle-under melted,
analysis. CNN-based defect detection and monitoring methods are and under melted). The experimental results showed that the proposed
developed in Ref [82–95]. As an extension of the CNN model, deep CNN method had a classification accuracy of 83.4%. The DBN structure
(DCNN) with a hierarchical structure that allows multilevel image fea­ developed in this study were demonstrated effective in learning features
tures to be extracted to achieve accurate pattern discovery was from original data with less prior knowledge.
employed by Caggiano et al. [82] for online defect-recognition in the Different ML models for image-based defect detection and

9
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Fig. 6. (a) The developed NN for droplet defects detection based on key droplet patterns (i.e., satellite, ligament, volume, and speed) analysis from jet images; (b)
The real-time control processes for offset jetting stabilization, Ref [98].

monitoring strategies have been developed in existing studies. However, the acoustic features and used as input for spectral convolutional neural
it’s worth noting that CNN-based models are prevailing in image-based networks (SCNN) for classification. Based on the porosity measurement
methods and normally yield superior results than conventional ML results of the printed workpieces, the qualities were categorized into
models. poor, medium, and high. According to the empirical results, the classi­
fication accuracies of SCNN varied from 83% to 89%, indicating the
4.3.2. Sensor signal-based approach feasibility of quality monitoring based on the proposed method. Similar
Sensor signal-based approaches for system monitoring are widely methods using AE signals for quality monitoring on PBF processes can be
applied in the manufacturing industry. In existing studies, different found in Ref [108,109]. In Ref [109], where the reinforcement learning
signals, including AE, optical emission, infrared signal, and multi-sensor (LR) algorithm was adopted for training and classification. However, the
signals are used for defect detection and monitoring in AM systems. corresponding control strategies have not been discussed in these studies
[107–109].
4.3.2.1. Acoustic Emission. Besides developing monitoring strategies In addition, some other researchers have also developed monitoring
based on image analysis, AE sensors have been applied in conventional strategies by analysing acoustic emission data based on various ML
manufacturing processes and are increasingly explored in AM. A quality models, such as SVM [110], hidden semi-Markov model (HSMM) [111],
monitoring approach based on AE for powder bed fusion (PBF) AM Clustering by fast search and find of density peaks (CFSFDP) [112], and
processes was proposed by S.A. Shevchik, et al. [107]. In this paper, the LSTM [113] for FDM processes, DBN [114] for SLM processes and
AM machine was equipped with a fiber Bragg grating (FBG) sensor to k-means [115] for laser metal deposition (LMD) processes. However, the
detect AE signals, shown in Fig. 7. The acoustic features, extracted from studies [110–113] focus more on identifying and detecting the abnormal
signals during the manufacturing process, were the relative energies of machine conditions, such as material run-out, filament breakage,
the narrow frequency bands of the wavelet packet transform. The extruder blockage or incorrect nozzle height, rather than the quality of
spectrograms localized in the time-frequency domain were built from manufactured products. For example, the methodology developed in Ref

10
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Fig. 7. The FBG system for collecting AE signals during the SLM process. The spectrograms localized in the time-frequency domain were built from the acoustic
features and used as input for spectral convolutional neural networks (SCNN) for classification, Ref [107].

[114] is capable of identifying five melted states, including balling, been used to better understand physical mechanisms and is also
slight balling, overheating, slight overheating, and normal phenomena considered promising for real-time monitoring in AM systems
for the SLM process. Compared with conventional ML models such as [116–119]. Mohammad et al. [120] proposed an in-process porosity
multilayer perceptron (MLP) and SVM, the DBN model is highlighted in monitoring approach using optical emission signatures captured by the
this study to be capable of achieving a high defect detection rate without multispectral sensor during the LPBF process. In this study, the
extracting features from raw signal data. Liu et al. [112] also developed line-to-continuum ratio, a measurement of the emission spectrum to
a machine state monitoring platform based on AE sensors for FDM determine the defects, of chromium emission around 520 nm was
machine states identification using unsupervised learning. The moni­ monitored during L-PBF of nickel alloy 718 powder feedstock. Based on
toring platform (a) and procedure for machine condition fault diagnosis emission spectroscopy, Ren et al., [121] introduced a quality monitoring
(b) are presented in Fig. 8 respectively. method for the DED process by applying LSTM-autoencoder and
K-means clustering models. In their work, the LSTM-Autoencoder was
4.3.2.2. Optical emission. Optical emission spectroscopy (OES) has long adopted for extracting features from the spectra collected during the

Fig. 8. (a) The AE sensor-based monitoring platform for the FDM machine; (b) the procedure of fault diagnosis for FDM machine states using unsupervised learning,
Ref [112].

11
J. Qin et al. Additive Manufacturing 52 (2022) 102691

production, and the K-means clustering algorithm was used for the X-ray tomography. In the proposed method, shown in Fig. 9, features
deposition quality classification. from melt pool boundaries were obtained through functional principal
The X-ray computed tomography (CT) techniques were employed to component analysis (FPCA) and used as input in different supervised ML
quantify the pore severity in each layer of a test part. The pore severity models, including k-NN, SVM, decision tree (DT), and linear discrimi­
was classified into four levels, Disc A (160 J/mm3), Disc B (80 J/mm3), nant analysis (DA), for predicting porosity. It was reported from the
Disc C (53 J/mm3), Disc D (107 J/mm3). The graph-theoretic approach experimental results that the k-NN model obtained the best performance
was adopted to extract features, called graph Fourier transform co­ for correctly predicting an abnormal melt pool with an accuracy of
efficients, from the line-to-continuum signatures. These features were 98.44%, while the DT model achieved the best result for false-negative
used as input in various ML models, such as k-nearest neighbours (k- value of only 0.033%.
NN), NN and SVM, for predicting the porosity level in each layer with
the CT data taken as ground truth. The results demonstrated that the 4.3.2.4. Multisensor signal. Recently, researchers have begun to fuse
proposed methodology had an accuracy of 90% (F-score) of classifica­ data acquired from multiple in-process sensors for analysing and
tion in a computation time of less than 0.5 s. developing defect detection strategies as it is unusual that the data
collected from a single source can cover all information of different
4.3.2.3. Infrared signal. Outputs from the visual and simulation-based phenomena. Kim et al. [128] proposed a data-driven method for
porosity detection methods are possibly far from actual yields in some monitoring and fault diagnosis of the FDM process states using two types
cases since they are often incapable of taking into account the uncer­ of sensors, an accelerometer and an AE sensor. The root mean square
tainty that results from material or process parameters [122]. Charac­ (RMS) values were extracted as critical features from time-domain sig­
teristics of a melt pool have been demonstrated to have a strong link nals under healthy and faulty process states. The faulty FDM process
with the formation of defects through existing studies[123–125]. state was realized by considering the loosened bolt which would be
Therefore, some researchers have explored methods for detecting and resulted in shifting layers in the printed specimen due to a feed motion of
predicting porosity by capturing in situ melt pool morphologies using the extruder head. The extracted features were fed into the SVM model
infrared sensors [122,126]. The information of the melt pool can be which had an accuracy of 87.5% for classification according to the
obtained by using various instruments such as infrared sensors, py­ experiment results.
rometers or high-speed cameras. Khanzadeh et al. [127] adopted A heterogeneous sensor-based in situ monitoring approach was
self-organizing maps (SOMs) to analyse 2D melt pool images for developed by Montazeri et al. [129] to detect the occurrence of
detecting anomalies in additively manufactured thin walls in the DED lack-of-fusion defects in titanium alloy (Ti-6Al-4 V) parts manufactured
process. As an extension of the work in Ref [127], a real-time porosity in the DED process. In this study, the data was collected from an optical
prediction method based on morphological characteristics of melt pool emissions spectrometer and a CCD camera with a near-infrared (NIR)
boundaries was proposed by Khanzadeh et al. [122]. In this paper, the filter, aiming at capturing the dynamic phenomena around the melt pool
time-varying melt pool signals were captured by a dual-wavelength region. The authors fused the data into a weighted network graph
pyrometer and categorized as either pores or normal melt pools by developed in Ref [130] and employed the graph Kronecker product

Fig. 9. Illustration of using ML techniques for porosity prediction based on melt pool boundary features that are extracted from thermal images by using FPCA,
Ref [122].

12
J. Qin et al. Additive Manufacturing 52 (2022) 102691

approach to building a dictionary of graph-theoretic features related to relationship in the DED process, illustrated in Fig. 10, where
the severity level of lack-of-fusion defects. These features were used as multi-physics modelling, experimental measurements, and data mining
inputs to the SVM model for classification which achieved F-scores close were integrated. In this method, simulations were carried out based on a
to 85% and 70% for a two-level and three-level classification scenario computational thermal-fluid dynamics (CFD) model to obtain structure
respectively. Compared with traditional statistical signal processing and property results (e.g., melt pool geometry, cooling rate, dilution,
approaches, the merits of the graph Kronecker product method was microhardness, etc.). These results were then validated by actual ex­
demonstrated in the paper. Similarly, strategies of using multi-sensor periments and fed into the SOM (self-organizing map) model with pro­
data for quality monitoring in AM systems are developed in cess parameters (i.e., laser power, mass flow rate, and energy density)
[131–133]. It is worth noting that, in Ref [131], Bastani et al. proposed a for investigating the PSP relationships. It was stated in the paper that
novel supervised classification approach, called online sparse helping researchers visually identify underlying relations among the
estimation-based classification (OSEC), which aimed to improve the features is the main advantage of using the SOM technique.
classification accuracy and reduce the computational burden for The following paragraph reviews and concludes the main areas that
real-time monitoring. In this study, the OSEC method was demonstrated researchers apply ML technologies for process modelling and control in
to be capable of processing raw sensor data directly with F-score per­ AM, including the prediction for mechanical property, shape deviation,
formances between 89% and 95% and being successfully applied to and in-process signature.
multiple sensor monitoring scenarios with a 685 Hz sampling rate.
Photodiode data is considered closely correlated to the properties of 4.4.1. Mechanical property
the melt pool by Okaro [134]. In this paper, the authors used a Masahiro et al. [148] introduced a prediction model for tensile
semi-supervised ML model, Gaussian mixture model (GMM), trained by properties based on analysing the microstructural features of the fabri­
the key features extracted from photodiode data to classify ‘acceptable’ cated parts with post-heat treatments. In this study, the specimens were
and ‘faulty’ AM builds regarding the tensile strength in the L-PBF pro­ fabricated by the SLM process using Ti-6Al-4 V alloy powder. The mi­
cess. randomized singular value decomposition (SVD) was employed as crostructures on the cross-section of the specimens were observed using
a feature extraction method to handle large datasets (approximately scanning electron microscopy (SEM), while the parallel-length part of
400 G per build) collected from photodiode sensors. The experiment the specimens was observed by using a micro-focus X-ray CT. The
results indicated that the GMM model achieved a 77% success rate in averaged maximum and minimum Feret diameters and aspect ratios of
identifying faulty specimens. Additionally, the proposed each α and prior-β grains were extracted by using ML-based image
semi-supervised approach was demonstrated to be able to use data from analysis tools. In addition, the defect features (e.g., the volume fraction
both builds where the resulting components were certified and build of pore) were also taken into consideration. Finally, these features were
where the quality of the resulting components is unknown, which is used for the prediction of tensile properties by multiple linear regression
cost-efficient especially in cases where part certification is costly and analysis with leave-one-out cross-validation. According to the experi­
time-consuming. Monitoring methods for AM systems based on vibra­ mental results, the models showed good performances in predicting the
tion data and ML models were also developed by several researchers yield strength and the ultimate tensile strength by using the selected
[135,136]. microstructural features.
Various monitoring and defect detection methodologies have been Process parameters are considered significant to determine the
developed, such as using pyrometer, AE, optical emission, infrared properties of the printed parts. Therefore, finding optimal combinations
camera, and high-speed visual camera, to detect macroscale or meso­ of process parameters for producing products with desired mechanical
scale defects based on ML models. Typically, the 1D (sensing signal- properties is crucial. ANN is frequently adopted by researchers for
based) data can be processed faster but is normally less informative modelling the complex relationships between process parameters and
while the 2D (image-based) or 3D data is often incapable of taking into part properties such as the strain recovery rates and transformation
account the process and material uncertainties. temperatures [149] and compressive strength [150]. Besides ANN,
Nathan et al. [151] introduced a method that bridged the links between
4.4. ML for process modelling and control in AM process parameters and part properties by using a Bayesian network
(BN). In this study, laser power, scan speed, hatch spacing, and layer
The properties and performances of additively manufactured parts thickness were selected as the parent nodes with different parameter
have long been major concerns of the AM industry as a high degree of settings to govern the casual relationships with child nodes. The yield
quality, performance, reliability, and repeatability is required in aero­ strength, ultimate tensile strength, surface roughness, hardness, and
space, automobile, defence, etc. [137]. This urges the development of density were used as child nodes. The authors used the data collected
robust predicting tools to feedback specific properties and performances from the publications where the parts were manufactured by SLM pro­
under different AM conditions. ML technologies have been increasingly cesses using 316 L to train the BN. By using this BN model, the users can
adopted for bridging the process conditions to final product perfor­ be provided with the probability distribution predictions of the
mances due to their ability of learning and modelling highly complex remaining nodes when they enter a known value for one or more nodes.
relationships. In general, the in-process performances, such as the melt For a specific machine, the developed BN model can be continually
pool geometry or line morphology, are significantly related to the re-trained to improve the accuracy.
quality of the final products. Therefore, some researchers focus on Compared with previous studies that investigated the relationships
modelling in-process performance by adopting various combinations of between static factors and part properties, Zhang et al. [152] presented a
process parameters for understanding the physical phenomena and method taking into account the in-process layer-wise information. The
identifying optimal parameter settings. However, with the development authors proposed a predictive model based on deep learning to improve
of in-process sensing systems and IoT technologies, researchers tend to the quality control regarding the tensile strength of printed parts in the
model the process-structure (PS) [138–143], process-structure-property FDM process. In this model, a merged structure that combines a
(PSP) [144–147], and process-property-performance (PPP) [2] re­ fully-connected neural network with a long short-term memory (LSTM)
lationships directly by exploring the data and information acquired from network was constructed for tensile strength predictions. The LSTM
the printing processes. This information includes processing parameters network was employed to process sensing signals, temperature, and
and processing resultant data during the printing process. By building vibration, which aimed at capturing the process variations and
these models, further studies such as material design, process optimi­ sequential inter-layer interactions of the FDM process. Other relevant
zation, and quality improvement, can be explored. For instance, Gan factors, such as printing speed, layer height, extruder temperature, and
et al. [144] introduced a data-driven method for modelling the PSP material property, were combined with the output of the LSTM network

13
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Fig. 10. The structure and property results (e.g., melt pool geometry, cooling rate, dilution, microhardness, etc.) of DED processed parts are obtained from simu­
lations and validated by actual experiments. These results are fed into the SOM (self-organizing map) model with process parameters (i.e., laser power, mass flow
rate, and energy density) for PSP (process-structure-property) relationship modelling, Ref [144].

Fig. 11. The neural network schematic of tensile strength prediction including two full connection layers and relevance propagation methods, Ref [152].

14
J. Qin et al. Additive Manufacturing 52 (2022) 102691

and fed into the fully-connected neural network for the final part areas (d-f) are presented in Fig. 12.
property prediction. A case study was carried out which demonstrated In addition, Mondal et al. [175] also used the predictive melt-pool
the effectiveness of the proposed sequential layer-by-layer modelling. In dimensions to obtain optimal scan strategy based on GP surrogate
addition, a layer-wise relevance propagation (LRP) algorithm was model. Other studies on the prediction of the characteristics of the
adopted in this work to identify the contributions of each input to output melt-pool using ML models can be found in Ref [176–179]. Thermal
based on the LSTM network. The proposed methodology of this study is profiles are able to reflect the interaction between layers, resulting in
illustrated in Fig. 11. residual stress and distortion distribution during the printing processes.
Researchers have also made effort to investigate and study the me­ This drives researchers to investigate the thermal profiles for the
chanical properties of fabricated parts in AM processes using ML tech­ improvement of product quality. Mriganka and Olga [180] introduced a
nologies [145,146,153–162]. For exploring substitute model for data-driven method for the modelling of thermal history based on ANN
traditional numerical simulation methods, Koeppe et al. [153] intro­ in FDM. The authors proposed a unique geometry representation
duced a strategy, combining experiments, finite element (FE) simula­ method that translated G-Code into a set of features. There were three
tions, and deep learning (DL) model, to predict the maximum von Mises types of features, including features related to the deposition time,
and equivalent principal stresses of printed lattice-cell structures in AM. features related to the distances from the cool surfaces and heat sources,
In this study, the FE simulations were validated by empirical experi­ that were used for predicting the thermal profile. A heat influence zone
ments, and the datasets obtained from simulations were used to train the that determined the area significantly affected by the heat source was
LSTM model for prediction. By taking design-related information into defined for feature selection. These selected features were fed into ANN,
account, Baturynska, et al. [156] developed ML-based models for part also called the surrogate model (SM) in the paper, to predict the tem­
property (e.g., tensile modulus, nominal stress, and elongation) pre­ perature profiles. In the case study, the training data was obtained by FE
diction, where part location, orientation, and STL model properties were simulations based on a physical model, and the prediction from the SM
considered as the inputs. Using ANN for fatigue life predictions for had an accuracy higher than 95%. The comparison of the temperature
aerospace alloy parts have been investigated by Zhan et al., [157,158]. distributions obtained from the FE model (a-d) and SM (e-h) is shown in
Fig. 13.
4.4.2. In-process signature Other studies on the prediction of thermal histories in different AM
Many researchers have done many efforts on computational models processes can be found in Ref [181–186], where ML-based models were
to simulate AM processes [2,154,163,164]. However, with the emer­ applied. For instance, Ren et al. [182] introduced a combined model
gence of advanced data mining techniques, ML is considered promising called RNN-DNN (recurrent neural networks and deep neural networks)
for modelling and uncovering the complex correlations among process to model the relationship between laser scanning strategies and their
conditions and resultant signatures. Meng and Zhang [165] proposed a corresponding thermal history distributions in the DED process. From
process modelling method for laser powder bed fusion (LPBF) of stain­ the studies above, it’s apparent that ML technologies have made sig­
less steel. In this work, a processing map of the re-melted depth of single nificant contributions to modelling complex AM processes, largely
tracks in terms of laser power and laser scan speed was developed by facilitating the development of control strategies and improving the
using the GP-based ML model. The GP regression model was trained reliability of the manufacturing processes.
using the datasets obtained from simulations of the computational fluid Lack of physical insights and normally requiring a considerable
dynamics (CFD) model and the predictions were then compared against volume of training data samples are the main drawbacks of most ML
experimental data for validation. For 316 L and 17–4 PH stainless steel, models. Integrating physical knowledge with ML models has great po­
the preferred conduction mode regions of laser power and laser scan tential to provide more explainable results and reduce training samples.
speed were predicted by the GP model to assist the process optimization. Hence, physical-informed ML techniques have risen in recent studies
Based on the process maps, the normalized enthalpy criterion of iden­ [187–190]. The bond formation and mesostructure have strong in­
tifying keyhole modes was modified for these two metals respectively. fluences on the final mechanical properties of the FDM produced
Tapia et al. [166] also adopted the GP model for process modelling in products. Different from the studies that purely rely on data-driven or
LPBF of 316 L stainless steel. A GP-based surrogate modelling frame­ physical models, Berkcan et al., [188] introduced a physical-informed
work was proposed in the paper, which was then used to predict melt ML approach to predict bond quality and porosity of the parts manu­
pool depth of single tracks in terms of laser power, scan speed, and laser factured by the FDM process. In their study, two coupled multi-physics
beam size combination. The operators can benefit from the applications models, thermal model and polymer sintering model, were first con­
of process maps which can provide predictions to reduce the demand for structed to predict the temperature evolution, bond formation, and
experimental or computational studies and obtain optimal process mesostructure evolution. As the multi-physics models were built within
parameter combinations. ML models have also been applied to predict certain assumptions that cannot fully represent the highly complex
in-process signatures such as geometries of deposited metal trace using physical phenomenon, a DNN was adopted to improve the prediction
ANN [167] in laser metal deposition (LMD), the stress distribution of performance. There were three strategies, (1) embed physics constraints
cured layers using CNN [168] in SLA, product magnetic characteristics in the loss function of the DNN, (2) the outputs of physical models are
using XGboost in SLM [169], connection status between printed lines used as extra inputs in the DNN, and (3) the DNN is firstly trained by the
using DNN in FDM [170], and printed line morphology using ANN [171] data generated from physical models and updated by real-world ex­
in SLM and using GP [172] and SVM [173] in aerosol jet printing (AJP). periments data, for integrating physics knowledge with ML algorithms.
Melt-pool geometries or characteristics have closely related to the The experimental results indicated that incorporating physics knowl­
quality of the produced products in metal AM processes. The control and edge with ML models was able to enhance the prediction accuracy even
minimization of the melt-pool variation are crucial to the stability and with a small amount of data.
reliability of the AM processes. Therefore, several studies have been
conducted on the investigation of melt-pool characteristics. A data- 4.5. ML on AM sustainability
driven approach to predict melt-pool area for scan strategy improve­
ment in the PBF process was developed by Yeung et al. [174]. The build Over the past couple of decades, AM technologies have attracted
time, laser power, scan speed, and neighbouring effect factors were extensive attention across the world. Compared with conventional
considered as the input to the polynomial regression model for melt-pool manufacturing, AM shows higher efficiency and flexibility, leading to its
area prediction. Then, the laser power was adopted as the design vari­ increasing adoption in the industry. However, according to the life-cycle
able for controlling melt pool size and reducing its variation. The opti­ analysis (LCA), the energy consumption of AM systems tends to have a
mized laser power distributions (a–c) and the resultant melt pool image significant effect on the environment [191]. This drives AM

15
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Fig. 12. The optimized laser power with corresponding melt pool image areas is predicted by the proposed ML approach. The melt pool area is allocated in high
power density areas as predicted Ref [174].

sustainability to a crucial research topic as the number of AM systems that is used for surgical preparation. Huff et al. [194] demonstrated and
being employed keeps growing. More specifically, cost and energy discussed the feasibility of using ML models to reduce the time and cost
consumption are considered the key indicators to measure the sustain­ of AM for surgical planning. In this work, ML models are applied for
ability of AM [192]. patient selection, optimizing the scanning process of CT scans, and
automated segmentation of medical images. ML technologies can
4.5.1. Cost estimation potentially improve the process of 3D anatomical modelling and
Cost estimation is a crucial task before the manufacturing processes contribute to the sustainability of applying AM in medical practice.
start. Reduction of costs (e.g. material and build time costs) is significant
to AM sustainability in the industry. A data-driven cost estimation 4.5.2. Energy consumption
framework, shown in Fig. 14, was introduced by Chan et al. [193] for Mathematical models for estimating energy consumption have been
AM systems based on big data analytics tools, aiming at reducing the explored and investigated in existing studies [192,195] of various AM
subjectivity of the cost estimation process. In the framework, the auto­ systems. For instance, Verma and Rai [192], developed mathematical
mated cost estimation system is an online service provider where models for estimating energy consumption and material waste in the SLS
manufacturing jobs with 3D models and relevant information, such as system and optimized the AM processes. However, AM systems are
material types, surface textures, and tolerances, were uploaded. Feature complex of which the energy consumption is correlated with various
vectors of the submitted jobs were extracted and clustered which were subsystems and factors, showing a large difference in terms of different
then fed in ML models with their costs as output for training. A working principles and main material supplies. It is difficult to take into
simulation-based cost prediction function was also presented in the account various factors based on conventional methods (e.g., mathe­
framework to simulate the whole manufacturing process for cost esti­ matical formulas) for energy modelling. Hence, ML models have been
mation when there was a small number of relevant jobs in the database. increasingly adopted for analysing and modelling the energy con­
The final cost prediction from the ML model was combining the costs of sumption of AM systems.
similar jobs in the database and the prediction from the simulation. A linear regression (LR) model was adopted by Tian et al. [196] to
When a new 3D model in STL file format was uploaded, g-code was capture the relationships between process parameters, part quality and
generated automatically where feature vectors were able to be extracted energy consumption respectively in the fused filament fabrication (FFF)
to represent geometry information. These feature vectors were used as process. In this paper, the printing resolution, printing speed, and nozzle
the input for the ML models that were built from the closest cluster to temperature were considered as the process parameters. The geometry
predict the cost. However, cost elements such as post-processing and accuracy features, including thickness deviation and average
labour cost are not included in the model, and obtaining sufficient data out-of-tolerance percentage, were selected as the indicators of part
with high quality in the database is also challenging. quality. Based on the linear regression models, the optimal solution for
It is usually time-consuming to build a patient-customized model acquiring energy-efficient process parameters under the specific quality

16
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Fig. 13. The comparison of the FEM (finite element model) and SM (surrogate model) for temperature distribution, Ref [180].

Fig. 14. The automated cost estimation framework for AM systems, Ref [193].

requirements was developed. This work provided a strategy for mini­ layer-level and build-level which were hard to integrate in a direct way
mizing the energy consumption of the AM system while simultaneously for modelling. Hence, a clustering method was carried out on the
ensuring the geometry-related quality of the manufactured parts. layer-level data and then integrated with build-level data in the ANN
Qin et al. [197] proposed a multi-source data analytics method for model. A case study was implemented on an SLS system. The empirical
AM energy consumption modelling based on ANN. In this method, the experiment results indicated that the ANN model had an accuracy of
data generation of an AM process was categorized into four sources, 80.3% for energy consumption prediction when the number of clusters
including design, process operation, working environment, and material was four. Furthermore, as an extension of the study, the authors [198]
condition, which tended to cover the entire production phases during an found out that the design-relevant features, including part design and
AM process. This multi-source data was heterogeneous and classified as design optimization, had significant impacts on AM energy consumption

17
J. Qin et al. Additive Manufacturing 52 (2022) 102691

based on the weights of neurons in the ANN model. Thus, a been implemented as one of the most popular ML technologies for AM
design-relevant feature-based energy consumption prediction model issues. Benefiting from its high compatibility of input data, various types
was established and a particle swarm optimization (PSO) method was of data is collected and used for deep learning models, such as image,
adopted to optimize the design-relevant features for reducing the energy video, acoustic data, process parameters, CAD model, and other sensor
consumption of the target AM system. Hu et al. [199] also analysed the data. Also, comparing ML technologies between deep learning and
impacts of design and working environment attributes on AM energy others, deep learning has generally shown merit. Deep learning tech­
consumption based on the gradient boosting decision tree (GBDT) al­ nologies, such as artificial neural networks, convolutional neural net­
gorithm. In this work, information gain was used to evaluate the works and Long-short term memory neural networks, have been studied
contribution of each attribute to the unit energy consumption in the SLS and applied to many industrial fields as one type of the latest ML algo­
process. rithms. Many researchers tend to compare deep learning with conven­
Geometry feature-based energy consumption estimation methods for tional ML algorithms which deep learning generally shows its merits, in
mask image projection SLA systems using ML technologies are devel­ terms of, high prediction accuracy, data adaptability and the capability
oped by Yang et al. [200,201]. In the Ref [201], three methods, of processing big data [202].
including sensitivity analysis based on Pearson correlation coefficient However, deep learning algorithms are required a huge amount of
(PCC) and Laplacian score, PCA, and stacked autoencoders (SAE), were training data to build the model. For many real issues, collecting high
applied to feature extraction and selection of layer-wise geo­ quality and big size data is challenging. Additionally, the majority of
metry-related indexes. These extracted features were fed into different deep learning technologies require a high-performance computational
ML models for predicting energy consumption. According to the platform to train the model which increases the research or development
experimental results, the neural network has the lowest averaged root budget. The necessity of using deep learning as a solution for the issues
mean square error (RMSE) of 0.75% regarding both training and testing should be considered [203]. Many factors should be discussed before
while the SAE structure has the best testing performance with an RMSE selecting an ML algorithm which depends on the complexity and nature
of 0.85%. The main contribution of these studies [197–201] is pre­ of the issues. Another issue of both deep learning and conventional ML is
senting a design-based method to better manage the energy consump­ the rational interpretations. Many researchers have realised this issue
tion of the target systems before the AM processes start for AM which ‘black box’ solutions are hard to use for revealing the basic
designers, which is a significant improvement for AM sustainability. mechanisms of many research fields [204]. ML technologies are
However, still, the feature extraction method that best represents the increasingly required a reasonable explanation and domain
characteristics of geometry models remains an open challenge. knowledge-based structure. The fundamental knowledge and correc­
Several existing studies have demonstrated the effectiveness and tions of AM processes largely obey the law of physics and material.
superiority of using ML technologies for cost estimation, reducing cost, Knowledge-based ML technologies can integrate the known theory to
improving energy management in different AM systems. However, the discover the potential information from the multi-source datasets [108].
potential of applying ML for AM sustainability have not been fully Many researchers have started to tackle this issue by using
achieved, where further studies of using ML for energy saving, reducing physical-informed and knowledge graph neural networks [43,187,188].
material waste, and improving manufacturing process efficiency need to To present the performance of the ML algorithms, evaluation metrics
be explored. are the most important element which use for measuring the quality of
the proposed ML model in the case studies. Many different metrics can
4.6. Classification of recent research work be applied depending on the purposes of the research [210]. However,
for AM issues, the focuses of the majority of research are on the solution
In the previous section, the recent research of ML for AM is reviewed of the critical issues in which the evaluation metrics theoretically in­
including different AM processes of manufacturing various types of fluence the proposed methods less than other factors of ML technologies,
material, like powder bed fusion (PBF), direct energy deposition (DED), such as the prior parameters, data format, and data variables. Although
material extrusion (ME), vat photopolymerization (VP) and material the evaluation metrics are not included in Table 5, many more details
jetting (MJ), which have been introduced in Section 2. It is obvious that are reported in the ML literature [211,212].
ML technologies have played a critical role in these AM processes. In addition, about half of the reviewed research applied ML tech­
Table 5 has classified the recent research work of ML technologies for nologies to solve the issues in the research domain of defect detection
AM processes during the last five years. and in-situ monitoring. However, for the work on defect detection and
In the table, the eight most representative ML technologies are in-situ monitoring, multiple data processing and modelling are still
included, SVM (I), deep learning (II), decision tree (III), k-NN (IV), crucial problems. While ML is a powerful tool for empirical modelling
Bayesian (V), linear regression (VI), Gaussian process (VII), Markov but it still highly depends on data. From the studies reviewed, various
model (VIII), clustering algorithms (IX) which are the most often used types of data are considered as the input of ML technologies, such as
ML technologies for the AM research perspective. Generally, researchers image data, video data, sensing data and design data. How to integrate
choose ML technologies based on the research or project target, in which multi-source data for ML is still an important research topic. Further­
the most applicable technology is selected. These ML technologies are more, increasingly researchers have moved their concentration from
used in 5 AM research domains, in terms of DfAM (RD1), material an­ process to design and sustainability which ML technologies can also play
alytics (RD2), defect detection and in-situ monitoring (RD3), process a critical role.
modelling and control (RD4), and sustainability (RD5). Interestingly,
over 70% of research focus on RD3 (defect detection and in-situ moni­ 5. Challenges and opportunities
toring), and RD4 (process modelling and control) which data-driven
methods are more intuitive in these two research domains, and ML Although ML technologies have been increasingly employed in dig­
technologies highly reply to the data. However, data is also available ital manufacturing systems, there are still several challenges that remain
and important for other research domains which are lack-of-attention. to be tackled, as well as opportunities to be seized. In this section, four
There are huge potential research opportunities in these domains aspects of challenges are summarized based on the previous literature
compared to RD3 and RD4. Furthermore, many potential research ideas review. In addition, three directions of opportunities are presented to
in RD1 (DfAM), RD2 (material analytics) and RD5 (sustainability) can be herald the future research and development of ML for AM.
inspired by ML approaches in the RD3 and RD4.
The most used ML technology is deep learning which was applied for
over half of the reviewed research and in all five research domains. It has

18
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Table 5
The classification of research work of ML for AM.
Reference AM Research ML Technologies Data Type Research Target Material Used in
Process Domain Studies
I II III IV V VI VII VIII IX

Zhang, et al.[81] PBF RD3 √ √ Image Quality level Steel

Ye, et al.[104] RD3 √ Image Melt pool state
Shevchik, et al. RD3 √ Acoustic data Quality level
[107]
Ye, et al.[114] RD3 √ Acoustic data Melt pool state
Tapia, et al.[166] RD4 √ Process parameters Melt pool state
Meng and Zhang RD4 √ Process parameters Melt pool state
[165]
Chandrika and Ya RD4 √ √ √ Process parameters Melt pool state
[179]
Nathan, et al.[151] RD4 √ Process parameters Quality level
Germán, et al.[205] RD4 √ √ √ √ √ Process parameters Quality level
Aniruddha, et al. RD4 √ Sensor data and image Quality level
[187]
Xin, et al.[106] RD3 √ Thermal imaging Porosity detection
Bugatti, et al.[77] RD3 √ √ √ Video Defect detection
Montazeri, et al. RD3 √ √ √ Optical data Porosity detection Nickel alloy
[120]
Aminzadeh and RD3 √ Image Quality level
Kurfess[103]
Caggiano, et al.[82] RD3 √ Image Defect detection
Seulbi, et al.[178] RD4 √ √ √ Process parameters Melt pool state
Mehrshad, et al. RD4 √ Process parameters Quality level
[149]
Scime and Beuth RD3 √ Image Defect detection
[83]
Fathizadan, et al. RD3 √ √ Image Defect detection
[93]
Okaro, et al.[134] RD3 √ Photodiode data Quality level
Yeung, et al.[174] RD4 √ Process parameters Melt pool state
Mondal, et al.[175] RD4 √ Process parameters Melt pool state
Vrábel, et al.[65] RD2 √ Radiation data Material Aluminium
classification alloy
Yao, et al.[36] RD1 √ CAD model Feature selection
Chen, et al.[171] RD4 √ Process parameters Line morphology
Zhang, et al.[206] RD4 √ Production Manufacturability
information
Ertay, et al.[95] RD3 √ Image Pore prediction
Masahiro, et al. RD4 √ Image Tensile property Titanium alloy
[148]
Shin, et al.[141] RD4 √ Process parameters Density prediction
Zackary, et al.[90] RD3 √ Image Defect detection
Li, et al.[91] RD3 √ Image Quality level Cobalt alloy
Baturynska[156] RD4 √ √ √ Design information Nominal stress Polymer
Qin, et al.[198] RD5 √ CAD model Energy consumption
Hu, et al.[199] RD5 √ CAD model Energy consumption
DeCost, et al.[63] RD2 √ Image Material Multiple
classification materials
Ko, et al.[43] RD1 √ Production Design rule
information construction
Montazeri, et al. DED RD3 √ Optical data and image Quality level Titanium alloy
[129]
Khanzadeh, et al. RD3 √ √ √ Thermal imaging Porosity detection
[122]
Zhang, et al.[84] RD3 √ Acoustic data Porosity detection
Li, et al.[56] RD4 √ Process parameters Offset distance Copper alloy
Ren, et al.[121] RD3 √ √ Emission spectroscopy Quality level Aluminium
Ding, et al.[207] RD4 √ Process parameters Bead modelling alloy
Ren, et al.[182] RD4 √ Scanning patterns Thermal history Steel
Chen, et al.[80] RD3 √ √ √ √ √ √ √ 3D point cloud Defect detection
Lam, et al.[142] RD4 √ Tool paths Void-filling strategy
Ren, et al.[183] RD4 √ Process parameters Temperature field
Richard, et al.[66] RD2 √ Particle-level features Flowability Multiple
classification materials
Jonnel et al.[40] ME RD1 √ 3D coordinates Geometry correction Polymers
Wu, et al.[111] RD3 √ Acoustic data Material condition
Zhang, et al.[152] RD4 √ Process parameter Tensile strength
Wu, et al.[132] RD4 √ √ Sensor data Surface roughness
Jin, et al.[100] RD3 √ Image Quality level
Kim, et al.[128] RD3 √ Sensor signals Defect detection
Koeppe, et al.[153] RD4 √ CAD model Stress prediction
Zhu, et al.[50] RD4 √ Shape variation Geometric deviation
RD4 √ G-Code Thermal history
(continued on next page)

19
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Table 5 (continued )
Reference AM Research ML Technologies Data Type Research Target Material Used in
Process Domain Studies
I II III IV V VI VII VIII IX

Mriganka and Olga

[180]
Jiang, et al.[170] RD4 √ Process parameters Quality level
Kapusuzoglu, et al. RD4 √ Process parameters Porosity prediction
[188]
Mohammad, et al. RD3 √ Image Defect detection
[88]
Li, et al.[79] RD3 √ √ √ 3D point cloud Defect detection
Roach, et al.[208] RD4 √ Image Compression Silicone
response
Aditya, et al.[168] VP RD4 √ Image Stress distribution Liquid resin
Yang, et al.[201] RD5 √ √ CAD model Energy consumption
Lee, et al.[96] RD3 √ Video Quality level
Ferreira, et al.[52] RD4 √ CAD model Geometric deviation
Shen, et al.[60] RD4 √ CAD model Geometric deviation
Wang, et al.[98] MJ RD3 √ Image Droplet behaviour Polymers
Zhang, et al.[173] RD4 √ Process parameters Line morphology Silver
Després, et al.[39] General RD1 √ CAD model mechanical Software-based
properties
Huang, et al.[47] RD1 √ CAD model Structure design
Nathan, et al.[209] RD1 √ Image Design optimization
Zhan et al.[158] RD4 √ √ √ Process parameters Fatigue life Steel
and fatigue loadings

5.1. Challenges different processing parameter combinations are considered. Testing the
properties of the manufactured parts also requires considerable labour.
5.1.1. Data fusion Thus, it is always expensive and impractical to collect a large amount of
AM is considered a complex system, where several subsystems and training data from experiments. Some researchers are investigating the
the manufacturing processes are affected by various correlated factors. use of process simulations for training, in addition to experimental re­
Therefore, it is necessary and significant to integrate or fuse the data sults. Furthermore, the ML models are consequently applied to the real
from multiple sources or modalities to jointly analyse for enhancing AM process and collaborate with some control algorithm, such as
knowledge discovery, as well as improving the modelling accuracy. But adaptive control. The limited data may lead to a high possibility of
it is difficult to fuse the heterogeneous data generated from AM systems failure due to the lack of training. Even if a well-trained ML model is
as it normally has different types, dimensions, and structures (e.g., im­ obtained, it is difficult to be robust and widely applied in the same AM
ages, 3D models, and signals). Several methods have been adopted by process but for different machines where the uncertainties affect the
researchers to tackle the issues, such as extracting features from raw model performance. Therefore, it is challenging to obtain a robust and
data to reduce data dimensions or using decomposition and factorization reliable ML model with a limited amount of training data.
techniques for data concatenation. There is still significant latent or
useful information from raw data which is lost through feature extrac­ 5.1.3. Interpretability
tion processes. Moreover, the data from different modalities or sources Since ML technologies have been increasingly adopted for modelling
may yield conflicting results in ML models. Hence, in what way and how and decision making in several crucial scenarios (e.g. medical use,
to fuse the heterogeneous data for modelling and analysing becomes a defence, and precision machining), people have to make a trade-off
critical challenge when applying ML for AM. Data registration tech­ between transparency and accuracy [217]. The results yielded by ML
niques play an important role in aligning multi-sensory data to ground models need to be interpretable and tractable to assist decision-makers
truth data properly before the data is fused or integrated into ML to understand the results and their rationale. Some conventional ML
models. Some researchers have explored tackling data alignment issues, algorithms such as LR and DT are capable of generating explainable
such as [213–216], that effectively pave the way for data fusion in AM. results. However, it requires profound domain knowledge for variable
selections and is normally hard to yield better results than DNN-based
5.1.2. Training with limited amount of data models. In AM systems, an interpretable ML model can help engineers
According to the selected articles reviewed in the previous section of improve the understanding of manufacturing processes and develop
the paper, ML algorithms play an important role in classification and corresponding control strategies. Hence, still, the explanation of the
regression tasks for AM. However, the performances of ML algorithms inference from black-box ML models remains challenging and require
are influenced by the data available for training. For instance, in the further effort. One possible strategy moving forward is to integrate ML
topic of process monitoring, CNN is the prevailing algorithm to process results with high fidelity simulations of relevant parts, processes, or
image data for defect detection. It is capable of learning useful infor­ process conditions to help manufacturing personnel understand the
mation from raw images directly and automatically. The convolutional outcomes yielded by ML.
layers and pooling layers in the CNN architecture can extract repre­
sentative features and lower feature dimensions. In the meantime, its 5.1.4. Spatial and temporal scales
performances suffer from the data volume. The CNN model often re­ AM processes occur over a wide range of size and time scales. Build
quires a sufficient amount of data for training processes and fine-tuning volumes can be measured in meters, while grains in metal alloys may be
the parameters to yield high accuracy results. Using X-ray computed micrometres in size. Similarly, a build may require dozens of hours,
tomography to detect defects (e.g., porosity, cracks, and lack of fusion) while heating and cooling of feedstock occurs over microseconds. These
of the produced products for labelling data is typically costly and time- wide ranges of spatial and temporal scales lead to significant challenges
consuming. Additionally, when using ML for mechanical property pre­ in AM process monitoring and control and, from an ML perspective,
dictions, researchers need to conduct a number of experiments where relate to the data fusion and latency issues raised earlier. Data fusion

20
J. Qin et al. Additive Manufacturing 52 (2022) 102691

must occur at the appropriate size and time scales. Furthermore, con­ and transmitted. Consequently, unexpected latency will occur and affect
sistency across size and time scales is necessary to help ensure results are the decision making during the manufacturing process. However, it is
meaningful. It seems that networks of ML models will likely need to be also challenging to train ML models directly on the edge devices with
developed to model relationships at the relevant size and time scales and local datasets. Therefore, collaborative learning between the Cloud and
additional relationships across these scales. It is not clear if similar ML edge is a possible and promising solution that leverages the knowledge
network models have been developed for other domains which could from the Cloud to improve the model developed by edge devices.
perhaps inform the AM research community of promising technical Knowledge transfer is applied in the Cloud-Edge learning paradigm.
directions.
5.2.4. Other opportunities
5.1.5. Latency Apart from the above three potential research topics of ML for AM,
Using ML technologies for modelling and data analytics based on there are still some research opportunities in many relevant research
cloud computing often requires considerable computing power and are areas. Although many research pioneers have started to discover these
not ideal where low latency is a major concern (e.g., real-time moni­ topics, the lack of research is still obvious. One of these opportunities is
toring and control). For instance, DNN is a prevailing ML algorithm and on the topic of sustainability. In this paper, the research of the ML on AM
DNN-based models are capable of performing high accuracy or gener­ sustainability is reviewed in the last section. However, these papers
ating reliable inferences in many different tasks. It has been widely mainly focused on the cost and energy consumption where the research
applied in natural language processing, target recognition, and fault of AM sustainability involves more perspectives, in terms of re-
diagnosis. However, training a DNN-based model normally requires manufacturing, waste recycling and process and life-cycle emission
considerable time and computing resources. This is unlikely to be sup­ [221,222]. These research issues are generally complex and affected by
ported by local or edge devices due to the limited computing power. many different variables that ML has the strong advantage of over­
Thus, in the real-time monitoring and control scenario, the collected coming. Moreover, according to the studies reviewed, another research
data normally need to be sent to the Cloud for processing and analysing. opportunity, AM researchers pay less attention to, is the physics
Then the control instructions are sent back from the Cloud to the local informed ML approach. With the rapid development of AM, the current
system. This whole process normally causes latency for real-time defect unsolved research issues tend to be increasingly fundamental. The
detection and control. Moreover, the uncertainties of the network typical ML algorithms including deep learning technologies are hard to
connection and the limitation of the network bandwidth also affect the discover the hidden fundamental knowledge without the assistance of
reaction time. Therefore, it is challenging to apply ML models to practice domain knowledge. The knowledge-based ML methods become more
where low latency is required. and more critical and essential for solving challenging AM issues.
Additionally, the interpretability of ML for engineering problems is al­
5.2. Opportunities ways discussed. The research of physics informed, and knowledge-based
ML models are effectively explained.
5.2.1. Transfer learning
A major assumption in most ML models is that the training and future 6. Conclusions
data must be in the same feature space with the same distribution [218].
However, in real-world industry applications, it is impractical and costly ML has made a significant contribution in many AM aspects, such as
to retrain a model for each machine or system. Therefore, knowledge DfAM, defect detection, process monitoring, modelling and control, and
transfer becomes critical in this scenario. Transfer learning can transfer also has the great potential to become a type of critical solution for the
prior knowledge or information from a domain to a related domain, entire field of AM. The focus of this paper has been on the research and
improving ML models [219]. For example, for process modelling in AM, application of various ML technologies for AM systems. The paper was
transfer learning techniques can be used to transfer a pre-trained model inspired by the current popularity of machine learning for solving AM
for different machines with similar working principles, which requires issues. After reviewing seven categories of AM processes this paper has
less amount of training data than training a new model. Hence, transfer then followed a systemic literature review method which aims to target
learning can enable the rapid development of ML models for different the state-of-the-art relevant research articles. There are over hundreds of
AM machines, materials, and part designs. In the future, more studies on research articles selected for further detailed reviewing and analysis.
transfer learning techniques for AM can be explored to achieve its Based on these research articles, this paper has also applied the keyword
potential. co-occurrence and clustering analysis aiming to identify the hotspots in
the research field. Five aspects of AM issues were clustered by the pro­
5.2.2. Light-weight computing model posed method. Then, the selected research articles were arranged, pre­
As described previously, low latency is the major concern in real- sented and reviewed following these clusters. These articles are also
time monitoring and control. Most studies of applying ML for defect summarised based on the targeted AM processes, applied ML algorithms,
detection or anomaly correction are implemented at the Cloud level research domains, and utilised data. Finally, research challenges and
where latency is not a major concern. In real-world applications, opportunities of ML for AM are highlighted and discussed focusing on
applying analytical models on the local or edge devices is important to various perspectives.
attain low latency. Due to the limited computing power of edge devices,
it is crucial and significant to develop lightweight computing models for CRediT authorship contribution statement
implementing artificial intelligence on the edge. Data fusion techniques
are one of the promising approaches as they can reduce data dimen­ Qian Tang : Writing – review & editing. Charlie C.L. Wang: Writing
sionality in raw data. Additionally, designing and adopting ML algo­ – review & editing, Methodology, Formal analysis. David W. Rosen:
rithms with fast data sampling or sorting strategies such as light gradient Formal analysis, Methodology, Writing – review & editing. Timothy W.
boosting machines (LGBM) [220] also reduce the computing load. Simpson: Writing – review & editing. Lu Yan: Writing – review &
editing. Jian Qin: Writing – original draft, Formal analysis, Conceptu­
5.2.3. Learning by cloud-edge synergy alization, Methodology. Fu Hu: Conceptualization, Formal analysis,
The current state-of-the-art applications of ML in AM are mostly Methodology, Writing – original draft. Ying Liu: Writing – original
performed by centralized services on the Cloud. The information and draft, review & editing, Supervision, Methodology, Formal analysis,
data are transmitted between Cloud servers and local devices which is Conceptualization. Paul Witherell: Formal analysis, Methodology,
unavoidably increase network burden if tremendous data is collected Writing – review & editing.

21
J. Qin et al. Additive Manufacturing 52 (2022) 102691

Declaration of Competing Interest [30] L. Waltman, N.J. Van Eck, E.C. Noyons, A unified approach to mapping and
clustering of bibliometric networks, J. Informetr. 4 (4) (2010) 629–635.
[31] N.J. Van Eck, L. Waltman, Software survey: VOSviewer, a computer program for
The authors declare that they have no known competing financial bibliometric mapping, scientometrics 84 (2) (2010) 523–538.
interests or personal relationships that could have appeared to influence [32] N.J. Van Eck, L.J.I.J. o U. Waltman, Fuzziness, K.-B. Systems, Bibliometric
the work reported in this paper. mapping of the computational intelligence field, Int. J. Uncertain. Fuzziness
Knowl. -Based Syst. 15 (05) (2007) 625–645.
[33] M. Kumke, H. Watschke, T. Vietor, A new methodological framework for design
References for additive manufacturing, Virtual Phys. Prototyp. 11 (1) (2016) 3–19.
[34] M.K. Thompson, et al., Design for additive manufacturing: trends, opportunities,
[1] I. Gibson, D. Rosen, B. Stucker, M. Khorasani. Additive Manufacturing considerations, and constraints, CIRP Ann. 65 (2) (2016) 737–760.
Technologies, third ed., Springer, 2021. [35] J. Jiang, Y. Xiong, Z. Zhang, D.W. Rosen, Machine learning integrated design for
[2] J. Smith, et al., Linking process, structure, property, and performance for metal- additive manufacturing, J. Intell. Manuf. (2020) 1–14.
based additive manufacturing: computational approaches with experimental [36] X. Yao, S.K. Moon, G. Bi, A hybrid machine learning approach for additive
support, Comput. Mech. 57 (4) (2016) 583–610. manufacturing design feature recommendation, Rapid Prototyp. J. (2017).
[3] J. Yang, Y. Chen, W. Huang, and Y. Li, Survey on artificial intelligence for [37] S.-W. Hsiao, H.-C. Tsai, Applying a hybrid approach based on fuzzy neural
additive manufacturing, in 2017 23rd International Conference on Automation network and genetic algorithm to product form design, Int. J. Ind. Ergon. 35 (5)
and Computing (ICAC), 2017, pp. 1–6: IEEE. (2005) 411–428.
[4] H. Lasi, P. Fettke, H.-G. Kemper, T. Feld, M. Hoffmann, Industry 4.0, Bus. Inf. [38] A.J. Lew, M.J. Buehler, Encoding and exploring latent design space of optimal
Syst. Eng. 6 (4) (2014) 239–242. material structures via a VAE-LSTM model, Forces Mech. 5 (2021), 100054.
[5] C. Wang, X. Tan, S. Tor, C. Lim, Machine learning in additive manufacturing: [39] N. Després, E. Cyr, P. Setoodeh, M.J.J.-J. o t M. Mohammadi, Metals, and M.
state-of-the-art and perspectives, Addit. Manuf. (2020), 101538. Society, Deep Learning and Design for Additive Manufacturing: A Framework for
[6] N. Johnson, et al., Invited review: Machine learning for materials developments Microlattice Architecture, JOM-J. Miner., Met. Mater. Soc. 72 (6) (2020)
in metals additive manufacturing, Addit. Manuf. (2020), 101641. 2408–2418.
[7] S. Sing, C. Kuo, C. Shih, C. Ho, C. Chua, Perspectives of using machine learning in [40] J.D. Alejandrino, et al., A machine learning approach of lattice infill pattern for
laser powder bed fusion for metal additive manufacturing, Virtual Phys. Prototyp. increasing material efficiency in additive manufacturing processes, Int. J. Mech.
(2021) 1–15. Eng. Robot. Res. 9 (9) (2020) 1253–1263.
[8] T. Wohlers, T. Gornet, History of additive manufacturing, Wohlers Rep. 24 (2014) [41] Y. Tang, G. Dong, Q. Zhou, Y.F. Zhao, Lattice structure design and optimization
(2014) 118. with additive manufacturing constraints, IEEE Trans. Autom. Sci. Eng. 15 (4)
[9] S.G. Sarvankar, S.N. Yewale, Additive manufacturing in automobile industry, Int. (2017) 1546–1562.
J. Res. Aeronaut. Mech. Eng. 7 (4) (2019) 1–10. [42] G.E. Karniadakis, I.G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang, Physics-
[10] M. Kamal, G. Rizza, Design for metal additive manufacturing for aerospace informed machine learning, Nat. Rev. Phys. 3 (6) (2021) 422–440.
applications. Additive Manufacturing for the Aerospace Industry, Elsevier, 2019, [43] H. Ko, P. Witherell, Y. Lu, S. Kim, D.W. Rosen, Machine learning and knowledge
pp. 67–86. graph based design rule construction for additive manufacturing, Addit. Manuf.
[11] A. Paolini, S. Kollmannsberger, E. Rank, Additive manufacturing in construction: 37 (2021), 101620.
a review on processes, applications, and digital planning methods, Addit. Manuf. [44] Y. Zhang, R. Harik, G. Fadel, A. Bernard, A statistical method for build orientation
30 (2019), 100894. determination in additive manufacturing, Rapid Prototyp. J. (2019).
[12] T. Wohlers, Wohlers report 2020: 3D printing and additive manufacturing state of the [45] N. Ahsan, A. Habib, B.J.P.M. Khoda, Geometric analysis for concurrent process
industry: annual worldwide progress report. Wohlers Associates, 2020. optimization of AM, Procedia Manuf. 5 (2016) 974–988.
[13] I. A. 52900:, "Standard Terminology for Additive Manufacturing–General [46] X. Zhang, X. Le, A. Panotopoulou, E. Whiting, C.C.J.A.T. o G. Wang, Perceptual
Principles–Terminology," 2016. models of preference in 3D printing direction, ACM Trans. Graph. 34 (6) (2015)
[14] Z. Quan, et al., Additive manufacturing of multi-directional preforms for 1–12.
composites: opportunities and challenges, Mater. Today 18 (9) (2015) 503–512. [47] J. Huang, T.-H. Kwok, C. Zhou, W. Xu, Surfel convolutional neural network for
[15] P.K. Gokuldoss, S. Kolla, J. Eckert, Additive manufacturing processes: selective support detection in additive manufacturing, Int. J. Adv. Manuf. Technol. 105 (9)
laser melting, electron beam melting and binder jetting—selection guidelines, (2019) 3593–3604.
Materials 10 (6) (2017) 672. [48] C.C. Wang, Y.-S. Leung, Y. Chen, Solid modeling of polyhedral objects by layered
[16] S.-I. Park, D.W. Rosen, S.-K. Choi, C.E. Duty, Effective mechanical properties of depth-normal images on the GPU, Comput. -Aided Des. 42 (6) (2010) 535–544.
lattice material fabricated by material extrusion additive manufacturing, Addit. [49] K. Yanamandra, G.L. Chen, X. Xu, G. Mac, N. Gupta, Reverse engineering of
Manuf. 1 (2014) 12–23. additive manufactured composite part by toolpath reconstruction using imaging
[17] D.-S. Shim, et al., Effect of layer thickness setting on deposition characteristics in and machine learning, Compos. Sci. Technol. (2020), 108318.
direct energy deposition (DED) process, Opt. Laser Technol. 86 (2016) 69–78. [50] Z. Zhu, N. Anwer, Q. Huang, L. Mathieu, Machine learning in tolerancing for
[18] R. Udroiu, I.C. Braga, A. Nedelcu, Evaluating the quality surface performance of additive manufacturing, CIRP Ann. 67 (1) (2018) 157–160.
additive manufacturing systems: methodology and a material jetting case study, [51] Q. Huang, Y. Wang, M. Lyu, W. Lin, Shape deviation generator–a convolution
Materials 12 (6) (2019) 995. framework for learning and predicting 3-D printing shape accuracy, IEEE Trans.
[19] T.D. Ngo, A. Kashani, G. Imbalzano, K.T. Nguyen, D. Hui, Additive manufacturing Autom. Sci. Eng. (2020).
(3D printing): a review of materials, methods, applications and challenges, [52] R. d S.B. Ferreira, A. Sabbaghi, Q. Huang, Automated geometric shape deviation
Compos. Part B: Eng. 143 (2018) 172–196. modeling for additive manufacturing systems via Bayesian neural networks, IEEE
[20] Y.L. Yap, C. Wang, S.L. Sing, V. Dikshit, W.Y. Yeong, J. Wei, Material jetting Trans. Autom. Sci. Eng. 17 (2) (2019) 584–598.
additive manufacturing: an experimental study using designed metrological [53] W. Rong-Ji, L. Xin-Hua, W. Qing-Ding, W. Lingling, Optimizing process
benchmarks, Precis. Eng. 50 (2017) 275–285. parameters for selective laser sintering based on neural network and genetic
[21] P.M. Bhatt, A.M. Kabir, M. Peralta, H.A. Bruck, S.K. Gupta, A robotic cell for algorithm, Int. J. Adv. Manuf. Technol. 42 (11–12) (2009) 1035–1042.
performing sheet lamination-based additive manufacturing, Addit. Manuf. 27 [54] S. Lee, W. Park, H. Cho, W. Zhang, M.-C. Leu, A neural network approach to the
(2019) 278–289. modelling and analysis of stereolithography processes, Proc. Inst. Mech. Eng.,
[22] N.A. Chartrain, C.B. Williams, A.R. Whittington, A review on fabricating tissue Part B: J. Eng. Manuf. 215 (12) (2001) 1719–1733.
scaffolds using vat photopolymerization, Acta Biomater. 74 (2018) 90–111. [55] G. Vosniakos, T. Maroulis, D. Pantelis, A method for optimizing process
[23] A.R. McAndrew, et al., Interpass rolling of Ti-6Al-4V wire+ arc additively parameters in layer-based rapid prototyping, Proc. Inst. Mech. Eng. Part B J. Eng.
manufactured features for microstructural refinement, Addit. Manuf. 21 (2018) Manuf. 221 (8) (2007) 1329–1340.
340–349. [56] Y. Li, Y. Sun, Q. Han, G. Zhang, I. Horváth, Enhanced beads overlapping model
[24] B. Wu, et al., A review of the wire arc additive manufacturing of metals: for wire and arc additive manufacturing of multi-layer multi-bead metallic parts,
properties, defects and quality improvement, J. Manuf. Process. 35 (2018) J. Mater. Process. Technol. 252 (2018) 838–848.
127–139. [57] C. Wacker, et al., Geometry and distortion prediction of multiple layers for wire
[25] A. Fink, Conducting Research Literature Reviews: From the Internet to Paper, arc additive manufacturing with artificial neural networks, Appl. Sci. 11 (10)
Sage Publications, 2019. (2021) 4694.
[26] B. Kitchenham, S. Charters, Guidelines for performing systematic literature [58] M. Khanzadeh, P. Rao, R. Jafari-Marandi, B.K. Smith, M.A. Tschopp, L. Bian,
reviews in software engineering, Sch. Comput. Sci. Math. (2007). Quantifying geometric accuracy with unsupervised machine learning: Using self-
[27] C. Wohlin, Guidelines for snowballing in systematic literature studies and a organizing map on fused filament fabrication additive manufacturing parts,
replication in software engineering, in Proceedings of the 18th international J. Manuf. Sci. Eng. 140 (3) (2018).
conference on evaluation and assessment in software engineering, 2014, pp. [59] M. Samie Tootooni, A. Dsouza, R. Donovan, P.K. Rao, Z.J. Kong, P. Borgesen,
1–10. Classifying the dimensional variation in additive manufactured parts from laser-
[28] M.I. Azeem, F. Palomba, L. Shi, Q. Wang, Machine learning techniques for code scanned three-dimensional point cloud data using machine learning approaches,
smell detection: a systematic literature review and meta-analysis, Inf. Softw. J. Manuf. Sci. Eng. 139 (9) (2017).
Technol. 108 (2019) 115–138. [60] Z. Shen, X. Shang, M. Zhao, X. Dong, G. Xiong, F.-Y. Wang, A learning-based
[29] H.-N. Su, P.-C. Lee, Mapping knowledge structure by keyword co-occurrence: a framework for error compensation in 3D printing, IEEE Trans. Cybern. 49 (11)
first look at journal papers in technology foresight, Scientometrics 85 (1) (2010) (2019) 4042–4050.
65–79. [61] S. Singh, S. Ramakrishna, R. Singh, Material issues in additive manufacturing: a
review, J. Manuf. Process. 25 (2017) 185–200.

22
J. Qin et al. Additive Manufacturing 52 (2022) 102691

[62] B. Poorganji, E. Ott, R. Kelkar, A. Wessman, M. Jamshidinia, Materials ecosystem [93] S. Fathizadan, F. Ju, Y. Lu, Deep representation learning for process variation
for additive manufacturing powder bed fusion processes, JOM 72 (1) (2020) management in laser powder bed fusion, Addit. Manuf. 42 (2021), 101961.
561–576. [94] E. Westphal, H. Seitz, A machine learning method for defect detection and
[63] B.L. DeCost, H. Jain, A.D. Rollett, E.A. Holm, Computer vision and machine visualization in selective laser sintering based on convolutional neural networks,
learning for autonomous characterization of am powder feedstocks, Jom 69 (3) Addit. Manuf. 41 (2021), 101965.
(2017) 456–465. [95] D.S. Ertay, et al., Toward sub-surface pore prediction capabilities for laser powder
[64] B.L. DeCost, E.A. Holm, Characterizing powder materials using keypoint-based bed fusion using data science, J. Manuf. Sci. Eng. 143 (7) (2021), 071016.
computer vision methods, Comput. Mater. Sci. 126 (2017) 438–445. [96] X.Y. Lee, S.K. Saha, S. Sarkar, B. Giera, Automated detection of part quality
[65] J. Vrábel, et al., Classification of materials for selective laser melting by laser- during two photon lithography via deep learning, Addit. Manuf. (2020), 101444.
induced breakdown spectroscopy, Chem. Pap. 73 (12) (2019) 2897–2905. [97] Y. Wang, et al., Active disturbance rejection control of layer width in wire arc
[66] R. Valente et al., Classifying Powder Flowability for Cold Spray Additive additive manufacturing based on deep learning, J. Manuf. Process. 67 (2021)
Manufacturing Using Machine Learning, in 2020 IEEE International Conference 364–375.
on Big Data (Big Data), 2020, pp. 2919–2928: IEEE. [98] T. Wang, T.-H. Kwok, C. Zhou, S. Vader, In-situ droplet inspection and closed-
[67] S.K. Everton, M. Hirsch, P. Stravroulakis, R.K. Leach, A.T. Clare, Review of in-situ loop control system using machine learning for liquid metal jet printing,
process monitoring and in-situ metrology for metal additive manufacturing, J. Manuf. Syst. 47 (2018) 83–92.
Mater. Des. 95 (2016) 431–445. [99] P. Soille, Morphological Image Analysis: Principles and Applications, Springer
[68] M. Grasso, B.M. Colosimo, Process defects and in situ monitoring methods in Science & Business Media, 2013.
metal powder bed fusion: a review, Meas. Sci. Technol. 28 (4) (2017), 044005. [100] Z. Jin, Z. Zhang, G.X. Gu, Autonomous in-situ correction of fused deposition
[69] B. Zhang, Y. Li, Q. Bai, Defect formation mechanisms in selective laser melting: a modeling printers using computer vision and deep learning, Manuf. Lett. 22
review, Chin. J. Mech. Eng. 30 (3) (2017) 515–527. (2019) 11–15.
[70] H. Taheri, M. Shoaib, L.W. Koester, T.A. Bigelow, P.C. Collins, L.J. Bond, Powder [101] C. Gobert, E.W. Reutzel, J. Petrich, A.R. Nassar, S.J.A.M. Phoha, Application of
based additive manufacturing-a review of types of defects, generation supervised machine learning for defect detection during metallic powder bed
mechanisms, detection, property evaluation and metrology, Electr. Comput. Eng. fusion additive manufacturing using high resolution imaging, Addit. Manuf. 21
1 (2) (2017) 172–209. (2018) 517–528.
[71] W.-c Lee, C.-c Wei, S.-C. Chung, Development of a hybrid rapid prototyping [102] L. Scime, J. Beuth, Using machine learning to identify in-situ melt pool signatures
system using low-cost fused deposition modeling and five-axis machining, indicative of flaw formation in a laser powder bed fusion additive manufacturing
J. Mater. Process. Technol. 214 (11) (2014) 2366–2374. process, Addit. Manuf. 25 (2019) 151–165.
[72] A. Wang, S. Song, Q. Huang, F. Tsung, In-plane shape-deviation modeling and [103] M. Aminzadeh, T.R. Kurfess, Online quality inspection using Bayesian
compensation for fused deposition modeling processes, IEEE Trans. Autom. Sci. classification in powder-bed additive manufacturing from high-resolution visual
14 (2) (2016) 968–976. camera images, J. Intell. Manuf. 30 (6) (2019) 2505–2523.
[73] D. Ahn, J.-H. Kweon, S. Kwon, J. Song, S. Lee, Representation of surface [104] D. Ye, J.Y.H. Fuh, Y. Zhang, G.S. Hong, K. Zhu, In situ monitoring of selective
roughness in fused deposition modeling, J. Mater. Process. Technol. 209 (15–16) laser melting using plume and spatter signatures by deep belief networks, ISA
(2009) 5593–5600. Trans. 81 (2018) 96–104.
[74] C. Xia, Z. Pan, J. Polden, H. Li, Y. Xu, S. Chen, Modelling and prediction of surface [105] J.E. Siegel, M.F. Beemer, S.M. Shepard, Automated non-destructive inspection of
roughness in wire arc additive manufacturing using machine learning, J. Intell. fused filament fabrication components using thermographic signal
Manuf. (2021) 1–16. reconstruction, Addit. Manuf. 31 (2020), 100923.
[75] N. Gerdes, C. Hoff, J. Hermsdorf, S. Kaierle, L. Overmeyer, Hyperspectral imaging [106] X. Zhang, J. Saniie, A. Heifetz, Detection of defects in additively manufactured
for prediction of surface roughness in laser powder bed fusion, Int. J. Adv. Manuf. stainless steel 316L with compact infrared camera and machine learning
Technol. (2021) 1–10. algorithms, JOM 72 (12) (2020) 4244–4253.
[76] M.N. Bisheh, S.I. Chang, S. Lei, A layer-by-layer quality monitoring framework for [107] S.A. Shevchik, C. Kenel, C. Leinenbach, K. Wasmer, Acoustic emission for in situ
3D printing, Comput. Ind. Eng. 157 (2021), 107314. quality monitoring in additive manufacturing using spectral convolutional neural
[77] M. Bugatti, B.M. Colosimo, Towards real-time in-situ monitoring of hot-spot networks, Addit. Manuf. 21 (2018) 598–604.
defects in L-PBF: a new classification-based method for fast video-imaging data [108] S.A. Shevchik, G. Masinelli, C. Kenel, C. Leinenbach, K. Wasmer, Deep learning
analysis, J. Intell. Manuf. (2021) 1–17. for in situ and real-time quality monitoring in additive manufacturing using
[78] X. Qi, G. Chen, Y. Li, X. Cheng, C. Li, Applying neural-network-based machine acoustic emission, IEEE Trans. Ind. Inform. 15 (9) (2019) 5194–5203.
learning to additive manufacturing: current applications, challenges, and future [109] K. Wasmer, T. Le-Quang, B. Meylan, S. Shevchik, and Performance, “In situ
perspectives, Engineering (2019). quality monitoring in AM using acoustic emission: A reinforcement learning
[79] R. Li, M. Jin, V.C. Paquit, Geometrical defect detection for additive approach,”, J. Mater. Eng. 28 (2) (2019) 666–672.
manufacturing with machine learning models, Mater. Des. 206 (2021), 109726. [110] H. Wu, Y. Wang, Z. Yu, In situ monitoring of FDM machine condition via acoustic
[80] L. Chen, X. Yao, P. Xu, S.K. Moon, G. Bi, Rapid surface defect identification for emission, Int. J. Adv. Manuf. Technol. 84 (5–8) (2016) 1483–1495.
additive manufacturing with in-situ point cloud processing and machine learning, [111] H. Wu, Z. Yu, Y. Wang, Real-time FDM machine condition monitoring and
Virtual Phys. Prototyp. 16 (1) (2021) 50–67. diagnosis based on acoustic emission and hidden semi-Markov model, Int. J. Adv.
[81] Y. Zhang, G.S. Hong, D. Ye, K. Zhu, J.Y.J.M. Fuh, and Design, “Extraction and Manuf. Technol. 90 (5–8) (2017) 2027–2036.
evaluation of melt pool, plume and spatter information for powder-bed fusion AM [112] J. Liu, Y. Hu, B. Wu, Y. Wang, An improved fault diagnosis approach for FDM
process monitoring,”, Mater. Des. 156 (2018) 458–469. process with acoustic emission, J. Manuf. Process. 35 (2018) 570–579.
[82] A. Caggiano, J. Zhang, V. Alfieri, F. Caiazzo, R. Gao, R. Teti, Machine learning- [113] P. Becker, C. Roth, A. Roennau, R. Dillmann, Acoustic anomaly detection in
based image processing for on-line defect recognition in additive manufacturing, additive manufacturing with long short-term memory neural networks, 2020 IEEE
CIRP Ann. 68 (1) (2019) 451–454. 7th Int. Conf. Ind. Eng. Appl. ICIEA (2020) 921–926.
[83] L. Scime, J. Beuth, A multi-scale convolutional neural network for autonomous [114] D. Ye, G.S. Hong, Y. Zhang, K. Zhu, J.Y.H. Fuh, Defect detection in selective laser
anomaly detection and classification in a laser powder bed fusion additive melting technology by acoustic signals with deep belief networks, Int. J. Adv.
manufacturing process, Addit. Manuf. 24 (2018) 273–286. Manuf. Technol. 96 (5–8) (2018) 2791–2801.
[84] B. Zhang, S. Liu, Y.C. Shin, In-Process monitoring of porosity during laser additive [115] H. Gaja, F. Liou, Defects monitoring of laser metal deposition using acoustic
manufacturing process, Addit. Manuf. 28 (2019) 497–505. emission sensor, Int. J. Adv. Manuf. Technol. 90 (1–4) (2017) 561–574.
[85] Y. Zhang, H.G. Soon, D. Ye, J.Y.H. Fuh, K. Zhu, Powder-bed fusion process [116] A. Ancona, V. Spagnolo, P.M. Lugara, M. Ferrara, Optical sensor for real-time
monitoring by machine vision with hybrid convolutional neural networks, IEEE monitoring of CO 2 laser welding process, Appl. Opt. 40 (33) (2001) 6019–6025.
Trans. Ind. Inform. 16 (9) (2019) 5769–5779. [117] L. Song, J. Mazumder, Real time Cr measurement using optical emission
[86] R. Angelone, A. Caggiano, R. Teti, A. Spierings, A. Staub, K. Wegener, Bio- spectroscopy during direct metal deposition process, IEEE Sens. J. 12 (5) (2011)
intelligent selective laser melting system based on convolutional neural networks 958–964.
for in-process fault identification, Procedia CIRP 88 (2020) 612–617. [118] A. Nassar, T. Spurgeon, E. Reutzel, Sensing defects during directed-energy
[87] H. Baumgartl, J. Tomas, R. Buettner, M. Merkel, A deep learning-based model for additive manufacturing of metal parts using optical emissions spectroscopy, Solid
defect detection in laser-powder bed fusion using in-situ thermographic Free. Fabr. Symp. . Proc. (2014) 278–287.
monitoring, Prog. Addit. Manuf. (2020) 1–9. [119] W. Sun, Z. Zhang, W. Ren, J. Mazumder, J.J. Jin, In-situ monitoring of optical
[88] M.F. Khan, et al., Real-time defect detection in 3D printing using machine emission spectra for microscopic pores in metal additive manufacturing
learning, Mater. Today.: Proc. 42 (2021) 521–528. Submitted for Publication, J. Manuf. Sci. Eng. (2021) 1–21.
[89] O. Davtalab, A. Kazemian, X. Yuan, B. Khoshnevis, Automated inspection in [120] M. Montazeri, A.R. Nassar, A.J. Dunbar, P. Rao, In-process monitoring of porosity
robotic additive manufacturing using deep learning for layer deformation in additive manufacturing using optical emission spectroscopy, IISE Trans. 52 (5)
detection, J. Intell. Manuf. (2020) 1–14. (2020) 500–515.
[90] Z. Snow, B. Diehl, E.W. Reutzel, A. Nassar, Toward in-situ flaw detection in laser [121] W. Ren, G. Wen, Z. Zhang, J. Mazumder, Quality monitoring in additive
powder bed fusion additive manufacturing through layerwise imagery and manufacturing using emission spectroscopy and unsupervised deep learning,
machine learning, J. Manuf. Syst. 59 (2021) 12–26. Mater. Manuf. Process. (2021) 1–8.
[91] X. Li, X. Jia, Q. Yang, J. Lee, Quality analysis in metal additive manufacturing [122] M. Khanzadeh, S. Chowdhury, M. Marufuzzaman, M.A. Tschopp, L. Bian, Porosity
with deep learning, J. Intell. Manuf. 31 (8) (2020) 2003–2017. prediction: supervised-learning of thermal history for direct laser deposition,
[92] B. Zhang, P. Jaiswal, R. Rai, P. Guerrier, G. Baggs, Convolutional neural network- J. Manuf. Syst. 47 (2018) 69–82.
based inspection of metal additive manufacturing parts, Rapid Prototyp. J. [123] A.J. Pinkerton, L.J.J. o P.D.A.P. Li, Modelling the geometry of a moving laser melt
(2019). pool and deposition track via energy and mass balances, J. Phys. D: Appl. Phys. 37
(14) (2004) 1885.

23
J. Qin et al. Additive Manufacturing 52 (2022) 102691

[124] L. Song, V. Bagavath-Singh, B. Dutta, J.J.T.I.J. o A.M.T. Mazumder, Control of [152] J. Zhang, P. Wang, R.X. Gao, Deep learning-based tensile strength prediction in
melt pool temperature and deposition height during direct metal deposition fused deposition modeling, Comput. Ind. 107 (2019) 11–21.
process, Int. J. Adv. Manuf. Technol. 58 (1–4) (2012) 247–256. [153] A. Koeppe, C.A.H. Padilla, M. Voshage, J.H. Schleifenbaum, B. Markert, Efficient
[125] M. Islam, T. Purtonen, H. Piili, A. Salminen, O. Nyrhilä, Temperature profile and numerical modeling of 3D-printed lattice-cell structures using neural networks,
imaging analysis of laser additive manufacturing of stainless steel, Phys. Procedia Manuf. Lett. 15 (2018) 147–150.
41 (2013) 835–842. [154] W. Yan, et al., Data-driven multi-scale multi-physics models to derive
[126] Q. Tian, S. Guo, E. Melder, L. Bian, W.G. Guo, Deep learning-based data fusion process–structure–property relationships for additive manufacturing, Comput.
method for in situ porosity detection in laser-based additive manufacturing, Mech. 61 (5) (2018) 521–541.
J. Manuf. Sci. Eng. 143 (4) (2021), 041011. [155] X. Zhou, S.-J. Hsieh, J.-C. Wang, Accelerating extrusion-based additive
[127] M. Khanzadeh, S. Chowdhury, M.A. Tschopp, H.R. Doude, M. Marufuzzaman, manufacturing optimization processes with surrogate-based multi-fidelity models,
L. Bian, In-situ monitoring of melt pool images for porosity prediction in directed Int. J. Adv. Manuf. Technol. 103 (9–12) (2019) 4071–4083.
energy deposition processes, IISE Trans. 51 (5) (2019) 437–455. [156] I.J.A.S. Baturynska, Application of machine learning techniques to predict the
[128] J.S. Kim, C.S. Lee, S.-M. Kim, S.W. Lee, Development of data-driven in-situ mechanical properties of polyamide 2200 (PA12) in additive manufacturing,
monitoring and diagnosis system of fused deposition modeling (FDM) process Appl. Sci. 9 (6) (2019) 1060.
based on support vector machine algorithm, Int. J. Precis. Eng. Manuf. Green [157] Z. Zhan, H. Li, A novel approach based on the elastoplastic fatigue damage and
Technol. 5 (4) (2018) 479–486. machine learning models for life prediction of aerospace alloy parts fabricated by
[129] M. Montazeri, A.R. Nassar, C.B. Stutzman, P. Rao, Heterogeneous sensor-based additive manufacturing, Int. J. Fatigue 145 (2021), 106089.
condition monitoring in directed energy deposition, Addit. Manuf. 30 (2019), [158] Z. Zhan, H. Li, Machine learning based fatigue life prediction with effects of
100916. additive manufacturing process parameters for printed SS 316L, Int. J. Fatigue
[130] M. Montazeri, P. Rao, Sensor-based build condition monitoring in laser powder 142 (2021), 105941.
bed fusion additive manufacturing process using a spectral graph theoretic [159] K. Demir, Z. Zhang, A. Ben-Artzy, P. Hosemann, G.X. Gu, Laser scan strategy
approach, J. Manuf. Sci. Eng. 140 (9) (2018). descriptor for defect prognosis in metal additive manufacturing using neural
[131] K. Bastani, P.K. Rao, Z. Kong, An online sparse estimation-based classification networks, J. Manuf. Process. 67 (2021) 628–634.
approach for real-time monitoring in advanced manufacturing processes from [160] H. Hassanin, Y. Zweiri, L. Finet, K. Essa, C. Qiu, M. Attallah, Laser powder bed
heterogeneous sensor data, IIE Trans. 48 (7) (2016) 579–598. fusion of Ti-6Al-2Sn-4Zr-6Mo alloy and properties prediction using deep learning
[132] D. Wu, Y. Wei, J. Terpenny, Predictive modelling of surface roughness in fused approaches, Materials 14 (8) (2021) 2056.
deposition modelling using data fusion, Int. J. Prod. Res. 57 (12) (2019) [161] W. Muhammad, A.P. Brahme, O. Ibragimova, J. Kang, K. Inal, A machine learning
3992–4006. framework to predict local strain distribution and the evolution of plastic
[133] Z. Li, Z. Zhang, J. Shi, D. Wu, Prediction of surface roughness in extrusion-based anisotropy & fracture in additively manufactured alloys, Int. J. Plast. 136 (2021),
additive manufacturing with machine learning, Robot. Comput. Integr. Manuf. 57 102867.
(2019) 488–495. [162] S. Nasiri, M.R. Khosravani, Machine learning in predicting mechanical behavior
[134] I.A. Okaro, S. Jayasinghe, C. Sutcliffe, K. Black, P. Paoletti, P.L. Green, Automatic of additively manufactured parts, J. Mater. Res. Technol. (2021).
fault detection for laser powder-bed fusion using semi-supervised machine [163] Z. Gan, G. Yu, X. He, S. Li, Numerical simulation of thermal behavior and
learning, Addit. Manuf. 27 (2019) 42–53. multicomponent mass transfer in direct laser deposition of Co-base alloy on steel,
[135] C.-T. Yen, P.-C. Chuang, Application of a neural network integrated with the Int. J. Heat. Mass Transf. 104 (2017) 28–38.
internet of things sensing technology for 3D printer fault diagnosis, Microsyst. [164] G.X. Gu, M.J. Buehler, Tunable mechanical properties through texture control of
Technol. (2019) 1–11. polycrystalline additively manufactured materials using adjoint-based gradient
[136] D. Stanisavljevic, D. Cemernek, H. Gursch, G. Urak, G. Lechner, Detection of optimization, Acta Mech. 229 (10) (2018) 4033–4044.
interferences in an additive manufacturing process: an experimental study [165] L. Meng, J. Zhang, Process design of laser powder bed fusion of stainless steel
integrating methods of feature selection and machine learning, Int. J. Prod. Res. using a gaussian process-based machine learning model, JOM 72 (1) (2020)
(2019) 1–23. 420–428.
[137] C.-J. Bae, A.B. Diggs, A. Ramachandran, Quantification and certification of [166] G. Tapia, S. Khairallah, M. Matthews, W.E. King, A. Elwany, Gaussian process-
additive manufacturing materials and processes. Additive Manufacturing, based surrogate modeling framework for process planning in laser powder-bed
Elsevier, 2018, pp. 181–213. fusion additive manufacturing of 316L stainless steel, Int. J. Adv. Manuf. Technol.
[138] E. Popova, T.M. Rodgers, X. Gong, A. Cecen, J.D. Madison, S.R. Kalidindi, Process- 94 (9–12) (2018) 3591–3603.
structure linkages using a data science approach: application to simulated [167] F. Caiazzo, A. Caggiano, Laser direct metal deposition of 2024 Al alloy: trace
additive manufacturing data, Integr. Mater. Manuf. Innov. 6 (1) (2017) 54–68. geometry prediction via machine learning, Materials 11 (3) (2018) 444.
[139] T. Özel, A. Altay, A. Donmez, and R.J. T.I. J. o A.M. T. Leach, "Surface topography [168] A. Khadilkar, J. Wang, R. Rai, Deep learning–based stress prediction for bottom-
investigations on nickel alloy 625 fabricated via laser powder bed fusion," 94, up sla 3d printing process, Int. J. Adv. Manuf. Technol. 102 (5–8) (2019)
9–12, pp. 4451–4458, 2018. 2555–2569.
[140] J. Li, M. Sage, X. Guan, M. Brochu, Y.F. Zhao, Machine learning-enabled [169] T.-W. Chang, K.-W. Liao, C.-C. Lin, M.-C. Tsai, C.-W. Cheng, Predicting magnetic
competitive grain growth behavior study in directed energy deposition fabricated characteristics of additive manufactured soft magnetic composites by machine
Ti6Al4V, JOM 72 (1) (2020) 458–464. learning, Int. J. Adv. Manuf. Technol. 114 (9) (2021) 3177–3184.
[141] C.H. Lee, et al., Optimizing laser powder bed fusion of Ti-5Al-5V–5Mo-3Cr by [170] J. Jiang, C. Yu, X. Xu, Y. Ma, J. Liu, Achieving better connections between
artificial intelligence, J. Alloy. Compd. 862 (2021), 158018. deposited lines in additive manufacturing via machine learning, Math. Biosci.
[142] L. Nguyen, J. Buhl, M. Bambach, Continuous Eulerian tool path strategies for Eng. 17 (4) (2020).
wire-arc additive manufacturing of rib-web structures with machine-learning- [171] Y. Chen, H. Wang, Y. Wu, H. Wang, Predicting the printability in selective laser
based adaptive void filling, Addit. Manuf. 35 (2020), 101265. melting with a supervised machine learning method, Materials 13 (22) (2020)
[143] Y. Han, R.J. Griffiths, Z.Y. Hang, Y. Zhu, Quantitative microstructure analysis for 5063.
solid-state metal additive manufacturing via deep learning, J. Mater. Res. 35 (15) [172] H. Zhang, S.K. Moon, T.H. Ngo, J. Tou, M.A.B.M. Yusoff, Rapid process modeling
(2020) 1936–1948. of the aerosol jet printing based on gaussian process regression with latin
[144] Z. Gan, et al., Data-driven microstructure and microhardness design in additive hypercube sampling, Int. J. Precis. Eng. Manuf. 21 (1) (2020) 127–136.
manufacturing using a self-organizing map, Engineering 5 (4) (2019) 730–735. [173] H. Zhang, S.K. Moon, T.H. Ngo, Hybrid machine learning method to determine
[145] C. Herriott, A.D. Spear, Predicting microstructure-dependent mechanical the optimal operating process window in aerosol jet 3D printing, ACS Appl.
properties in additively manufactured metals with machine-and deep-learning Mater. Interfaces 11 (19) (2019) 17994–18003.
methods, Comput. Mater. Sci. 175 (2020), 109599. [174] H. Yeung, Z. Yang, L. Yan, A meltpool prediction based scan strategy for powder
[146] F. Yan, Y.-C. Chan, A. Saboo, J. Shah, G.B. Olson, W. Chen, Data-driven prediction bed fusion additive manufacturing, Addit. Manuf. (2020), 101383.
of mechanical properties in support of rapid certification of additively [175] S. Mondal, D. Gwynn, A. Ray, A. Basak, Investigation of melt pool geometry
manufactured alloys, Comput. Model. Eng. Sci. 117 (3) (2018) 343–366. control in additive manufacturing using hybrid modeling, Metals 10 (5) (2020)
[147] N. Kouraytem, X. Li, W. Tan, B. Kappes, A. Spear, Modeling process-structure- 683.
property relationships in metal additive manufacturing: A review on physics- [176] O. Kwon, H.G. Kim, W. Kim, G.-H. Kim, K. Kim, A convolutional neural network
driven versus data-driven approaches, J. Phys. Mater. (2020). for prediction of laser power using melt-pool images in laser powder bed fusion,
[148] M. Kusano, et al., Tensile properties prediction by multiple linear regression IEEE Access 8 (2020) 23255–23263.
analysis for selective laser melted and post heat-treated Ti-6Al-4V with [177] Y. Wang, C. Zhang, J. Lu, L. Bai, Z. Zhao, J. Han, Weld reinforcement analysis
microstructural quantification, Mater. Sci. Eng. A (2020), 139549. based on long-term prediction of molten pool image in additive manufacturing,
[149] M. Mehrpouya, A. Gisario, A. Rahimzadeh, M. Nematollahi, K.S. Baghbaderani, IEEE Access 8 (2020) 69908–69918.
M. Elahinia, A prediction model for finding the optimal laser parameters in [178] S. Lee, J. Peng, D. Shin, Y.S.J.S. Choi, t. o. a. materials, Data analytics approach
additive manufacturing of NiTi shape memory alloy, Int. J. Adv. Manuf. Technol. for melt-pool geometries in metal additive manufacturing, Sci. Technol. Adv.
105 (11) (2019) 4691–4699. Mater. 20 (1) (2019) 972–978.
[150] A.K. Sood, R.K. Ohdar, S.S. Mahapatra, Experimental investigation and empirical [179] C. Kamath, Y.J. Fan, Regression with small data sets: a case study using code
modelling of FDM process for compressive strength improvement, J. Adv. Res. 3 surrogates in additive manufacturing, Knowl. Inf. Syst. 57 (2) (2018) 475–493.
(1) (2012) 81–90. [180] M. Roy, O. Wodo, Data-driven modeling of thermal history in additive
[151] N. Hertlein, S. Deshpande, V. Venugopal, M. Kumar, S. Anand, Prediction of manufacturing, Addit. Manuf. 32 (2020), 101017.
selective laser melting part quality using hybrid Bayesian network, Addit. Manuf. [181] M. Mozaffar, et al., Data-driven prediction of the high-dimensional thermal
32 (2020), 101089. history in directed energy deposition processes via recurrent neural networks,
Manuf. Lett. 18 (2018) 35–39.

24
J. Qin et al. Additive Manufacturing 52 (2022) 102691

[182] K. Ren, Y. Chew, Y. Zhang, J. Fuh, G. Bi, Thermal field prediction for laser [203] X. Qi, G. Chen, Y. Li, X. Cheng, C. Li, Applying neural-network-based machine
scanning paths in laser aided additive manufacturing by physics-based machine learning to additive manufacturing: current applications, challenges, and future
learning, Comput. Methods Appl. Mech. Eng. 362 (2020), 112734. perspectives, Engineering 5 (4) (2019) 721–729.
[183] K. Ren, Y. Chew, N. Liu, Y. Zhang, J. Fuh, G. Bi, Integrated numerical modelling [204] L. Zhang, J. Tan, D. Han, H. Zhu, From machine learning to deep learning:
and deep learning for multi-layer cube deposition planning in laser aided additive progress in machine intelligence for rational drug discovery, Drug Discov. Today
manufacturing, Virtual Phys. Prototyp. (2021) 1–15. 22 (11) (2017) 1680–1685.
[184] Z. Zhou, H. Shen, B. Liu, W. Du, J. Jin, Thermal field prediction for welding paths [205] G.O. Barrionuevo, J.A. Ramos-Grez, M. Walczak, C.A. Betancourt, Comparative
in multi-layer gas metal arc welding-based additive manufacturing: a machine evaluation of supervised machine learning algorithms in the prediction of the
learning approach, J. Manuf. Process. 64 (2021) 960–971. relative density of 316L stainless steel fabricated by selective laser melting, Int. J.
[185] H.A. Kumar, S. Kumaraguru, C. Paul, K. Bindra, Faster temperature prediction in Adv. Manuf. Technol. 113 (1) (2021) 419–433.
the powder bed fusion process through the development of a surrogate model, [206] Y. Zhang, S. Yang, G. Dong, Y.F. Zhao, Predictive manufacturability assessment
Opt. Laser Technol. 141 (2021), 107122. system for laser powder bed fusion based on a hybrid machine learning model,
[186] P.K. Nalajam, R. Varadarajan, A hybrid deep learning model for layer-wise melt Addit. Manuf. 41 (2021), 101946.
pool temperature forecasting in wire-arc additive manufacturing process, IEEE [207] D. Ding, F. He, L. Yuan, Z. Pan, L. Wang, M. Ros, The first step towards intelligent
Access (2021). wire arc additive manufacturing: An automatic bead modelling system using
[187] A. Gaikwad, B. Giera, G.M. Guss, J.-B. Forien, M.J. Matthews, P. Rao, machine learning through industrial information integration, J. Ind. Inf. Integr.
Heterogeneous sensing and scientific machine learning for quality assurance in 23 (2021), 100218.
laser powder bed fusion–a single-track study, Addit. Manuf. 36 (2020), 101659. [208] D.J. Roach, et al., Utilizing computer vision and artificial intelligence algorithms
[188] B. Kapusuzoglu, S. Mahadevan, Physics-informed and hybrid machine learning in to predict and design the mechanical compression response of direct ink write 3D
additive manufacturing: application to fused filament fabrication, JOM 72 (12) printed foam replacement structures, Addit. Manuf. 41 (2021), 101950.
(2020) 4695–4705. [209] N. Hertlein, P.R. Buskohl, A. Gillman, K. Vemaganti, S. Anand, Generative
[189] Q. Zhu, Z. Liu, J. Yan, Machine learning for metal additive manufacturing: adversarial network for early-stage design flexibility in topology optimization for
predicting temperature and melt pool fluid dynamics using physics-informed additive manufacturing, J. Manuf. Syst. 59 (2021) 675–685.
neural networks, Comput. Mech. 67 (2) (2021) 619–635. [210] M. Hossin, M.N. Sulaiman, A review on evaluation metrics for data classification
[190] R. Liu, S. Liu, X. Zhang, A physics-informed machine learning model for porosity evaluations, Int. J. Data Min. Knowl. Manag. Process 5 (2) (2015) 1.
analysis in laser powder bed fusion additive manufacturing, Int. J. Adv. Manuf. [211] A. Gunawardana, G. Shani, A survey of accuracy evaluation metrics of
Technol. 113 (7) (2021) 1943–1958. recommendation tasks, J. Mach. Learn. Res. 10 (12) (2009).
[191] R. Huang, et al., Energy and emissions saving potential of additive manufacturing: [212] J. Zhou, A.H. Gandomi, F. Chen, A. Holzinger, Evaluating the quality of machine
the case of lightweight aircraft components, J. Clean. Prod. 135 (2016) learning explanations: a survey on methods and metrics, Electronics 10 (5) (2021)
1559–1570. 593.
[192] A. Verma, R. Rai, Sustainability-induced dual-level optimization of additive [213] J. Petrich, Z. Snow, D. Corbin, E.W. Reutzel, Multi-modal sensor fusion with
manufacturing process, Int. J. Adv. Manuf. Technol. 88 (5–8) (2017) 1945–1959. machine learning for data-driven process monitoring for additive manufacturing,
[193] S.L. Chan, Y. Lu, Y. Wang, Data-driven cost estimation for additive manufacturing Addit. Manuf. 48 (2021), 102364.
in cybermanufacturing, J. Manuf. Syst. 46 (2018) 115–126. [214] J.P. Morgan et al., Selection and Installation of High Resolution Imaging to
[194] T.J. Huff, P.E. Ludwig, J.M. Zuniga, The potential for machine learning Monitor the PBFAM Process, and Synchronization to Post-Build 3D Computed
algorithms to improve and reduce the cost of 3-dimensional printing for surgical Tomography, in Proceedings of the 28th Annual International Solid Freeform
planning, Expert Rev. Med. Devices 15 (5) (2018) 349–356. Fabrication Symposium–An Additive Manufacturing Conference, 2017, pp.
[195] Y. Yang, L. Li, Y. Pan, Z. Sun, Energy consumption modeling of stereolithography- 1382–1399.
based additive manufacturing toward environmental sustainability, J. Ind. Ecol. [215] N. Decker, Q. Huang, Geometric accuracy prediction for additive manufacturing
21 (S1) (2017) S168–S178. through machine learning of triangular mesh data. International Manufacturing
[196] W. Tian, J. Ma, M. Alizadeh, Energy consumption optimization with geometric Science and Engineering Conference, American Society of Mechanical Engineers,
accuracy consideration for fused filament fabrication processes, Int. J. Adv. 2019.
Manuf. Technol. 103 (5–8) (2019) 3223–3233. [216] A. Yaghi, S. Ayvar-Soberanis, S. Moturu, R. Bilkhu, S. Afazov, Design against
[197] J. Qin, Y. Liu, R. Grosvenor, Multi-source data analytics for AM energy distortion for additive manufacturing, Addit. Manuf. 27 (2019) 224–235.
consumption prediction, Adv. Eng. Inform. 38 (2018) 840–850. [217] A.B. Arrieta, et al., Explainable Artificial Intelligence (XAI): Concepts,
[198] J. Qin, Y. Liu, R. Grosvenor, F. Lacan, Z. Jiang, Deep learning-driven particle taxonomies, opportunities and challenges toward responsible AI, Inf. Fusion 58
swarm optimisation for additive manufacturing energy optimisation, J. Clean. (2020) 82–115.
Prod. 245 (2020), 118702. [218] S.J. Pan, Q. Yang, A survey on transfer learning, IEEE Trans. Knowl. Data Eng. 22
[199] F. Hu, Y. Liu, J. Qin, X. Sun, P. Witherell, Feature-level data fusion for energy (10) (2009) 1345–1359.
consumption analytics in additive manufacturing, 2020 IEEE 16th Int. Conf. [219] K. Weiss, T.M. Khoshgoftaar, D.J.J. o B. d Wang, A survey of transfer learning,
Autom. Sci. Eng. (CASE) (2020) 612–617. J. Big Data 3 (1) (2016) 9.
[200] Y. Yang, M. He, L. Li, A new machine learning based geometry feature extraction [220] G. Ke et al., "Lightgbm: A highly efficient gradient boosting decision tree," in
approach for energy consumption estimation in mask image projection Advances in neural information processing systems, 2017, pp. 3146–3154.
stereolithography, Procedia CIRP 80 (2019) 741–745. [221] S. Ford, M. Despeisse, Additive manufacturing and sustainability: an exploratory
[201] Y. Yang, M. He, L.J.J. o C.P. Li, Power consumption estimation for mask image study of the advantages and challenges, J. Clean. Prod. 137 (2016) 1573–1587.
projection stereolithography additive manufacturing using machine learning [222] T. Peng, K. Kellens, R. Tang, C. Chen, G. Chen, Sustainability of additive
based approach, J. Clean. Prod. 251 (2020), 119710. manufacturing: an overview on its energy demand and environmental impact,
[202] A. Korotcov, V. Tkachenko, D.P. Russo, S. Ekins, Comparison of deep learning Addit. Manuf. 21 (2018) 694–704.
with multiple machine learning methods and metrics using diverse drug discovery
data sets, Mol. Pharm. 14 (12) (2017) 4462–4475.

25