Materials Today: Proceedings 70 (2022) 95–100

Contents lists available at ScienceDirect

Materials Today: Proceedings

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

Applications of machine learning in 3D printing

Guo Dong Goh a, Wai Yee Yeong b,⇑
a
School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore
b
Singapore Centre for 3D Printing, School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore

a r t i c l e i n f o a b s t r a c t

Article history: The 4th Industrial Revolution integrates the digital revolution into the physical world, opens up new
Available online 13 September 2022 directions for research in several fields, including artificial intelligence and 3D printing. Data-driven mod-
els have been developed thanks to the advancement in computational power and machine learning algo-
Keywords: rithms, which enables machines to learn and execute a certain task through data without the need to
Additive manufacturing formulate them explicitly. Machine learning have been used in 3D printing to enhance the product qual-
Machine learning ity and productivity through in-situ monitoring, to optimize design and process parameters, and to speed
Process optimization
up the prediction microstructure evolution through the development of physic-based data driven surro-
In-situ monitoring
gate model. This article summarizes the work done by our group on the use of machine learning in 3D
printing and provides an outlook on the 3D printing industry.
Copyright Ó 2022 Elsevier Ltd. All rights reserved.
Selection and peer-review under responsibility of the scientific committee of The International Confer-
ence on Additive Manufacturing for a Better World.

1. Introduction A thorough comprehension of the 3D printing method, from the

manufacturability of the raw materials (viscoelastic properties and
3D printing, have developed and caused a change in perspective powder flow property) to the linkage among the process-structure-
on how products are designed and fabricated. The additive manu- properties of the 3D printed parts, is critical. Nevertheless, the 3D
facturing methods allow the creation of products with intricate printing processes encompass many print parameters that would
shapes such as topology optimized structures [1] and sandwich affect the quality of the printed parts. This would involve transdis-
core panels [2] and functionally graded properties. 3D printing is ciplinary knowledge like the material properties, solid–liquid
generally more environmental-friendly as they lessen material interdependence, fluid dynamics, grain-growth development, and
wastage. It has made some amazing progress from being a proto- thermal–mechanical coupling. The development of physics-based
typing device to gradually being used for end-product creation. models are complicated and tedious as it needs a broad under-
Different 3D printing methods have been developed and catego- standing of the multi-scale and multi-physics 3D printing pro-
rized into material extrusion, vat-photo polymerization, binder jet- cesses. Consequently, the small-scaled research usually takes into
ting, material jetting, powder bed fusion (PBF), and direct energy account a limited number of characteristics of the whole fabrica-
deposition. All these techniques are capable of printing actual tion process which limited the fidelity. For example, research like
working parts with different types of materials, which can be in fil- microscale grain structure evolution of the PBF that uses computa-
ament, liquid, or powder forms [3]. Nonetheless, there are some tional fluid dynamics (CFD) [6], analytical modeling of residual
challenges that are unique to the 3D printing techniques, such as stress that involves thermos-mechanic finite element analysis
the formation of void due to poorly optimized print parameters [7], as well as the melt pool profile and bead shape simulation
leading to poor inter-laminar properties [4], inhomogeneity of [8] that requires the knowledge of fluid–solid interaction have
the printed materials, and the distortion caused by the residual been attempted. These kinds of analysis requires a lot of computa-
stress resulting from the rapid cooling characteristic of the 3D tional power and often take days to complete, making real-time
printing processes [5]. modeling of the 3D printing process almost impossible.
ML is an artificial intelligence (AI) technique that enables
machines to learn and execute a certain task through data without
the need to formulate them explicitly. Machine learning is a new
⇑ Corresponding author.
wave of tool that is able to enhance systems through smarter
E-mail address: [email protected] (W.Y. Yeong).

https://doi.org/10.1016/j.matpr.2022.08.551
2214-7853/Copyright Ó 2022 Elsevier Ltd. All rights reserved.
Selection and peer-review under responsibility of the scientific committee of The International Conference on Additive Manufacturing for a Better World.
G.D. Goh and W.Y. Yeong Materials Today: Proceedings 70 (2022) 95–100

and effective use of products, materials, and services. For the layer- score of >80% using the AP50 metric. Then, two of the models
by-layer manufacturing processes, machine learning can bring (YOLOv3-Tiny 100 and 300 epochs) were enhanced using Open
benefits such as the shorter fabrication time, lower cost, and Neural Network Exchange (ONNX) model conversion and ONNX
enhanced quality through data-driven optimization. ML tech- Runtime to increase the inference speed. A classification accuracy
niques can be commonly grouped into 3 categories: supervised rate of 89.8% and an inference speed of 70 frames per second were
learning, unsupervised learning, and reinforced learning. In super- attained. In order to realize the real-time monitoring system, a cor-
vised learning, the training dataset is a collection of input/output rection algorithm was created to do corrective actions based off of
pairs, and the goal is to learn to predict the output in relation to defect classification. The G-codes of the corrective actions were
the input parameters. In unsupervised learning, there are no data transmitted to the printers in the course of the fabrication. This
being labeled as output, which is unlike supervised learning. The work showcased the real-time monitoring and automated rectifi-
algorithms will learn to identify patterns and extract important cation for the FDM process. This work advances the development
features from the input data and group them into dissimilar of an real-time inspection and correction system via closed-loop
groups. The goal of unsupervised learning is to identify the rela- feedback from other 3D printing techniques.
tionships of the input data which are previously unknown. In rein-
forced learning, the information fed to the algorithm is in between
supervised and unsupervised learning. Rather than providing the 2.2. Machine learning for 3D printed multi-materials tissue-mimicking
correct output based on the provided input in supervised learning, anatomical models
the training data are used to give an suggestion on whether an
action is right or wrong. Polyjet, a material jetting 3D printing technique, has become
In-depth reviews about the use of machine learning in 3D print- the common tool for the production of customized anatomical
ing have been published [9–11], readers are encouraged to refer to models due to layer-by-layer manufacturing technique that does
the review articles to gain more insight about it. This article targets not require tools or molds and its multi-material printing capabil-
to give a brief assessment of uses of machine learning techniques ity. Despite several successful demonstration of production of
in different aspects of 3D printing practices and our outlooks on highly dimensionally accurate anatomical models, 3D printed
some potential research directions that are still under way or anatomical models with tissue-like characteristics have not gained
may transpire ultimately. much success. In this study, we attempted to use a composite lay-
ering design to control the shore hardness and compressive mod-
ulus of the Polyjet-printed parts to simulate the mechanical
2. Applications of machine learning in 3D printing characteristics of the human tissues (Fig. 2). A total of 216 speci-
mens (with 72 combinations of design parameters) were fabricated
2.1. Anomaly detection in fused deposition modeling using machine and tested to build the material library for the anatomical models.
learning We also developed an analytical model to determine the effective
compressive modulus and shore hardness of the composite lami-
Fused deposition modeling (FDM) has matured over the years nate [13]. A neural network was trained to understand and predict
and is being adopted in various industries. Nevertheless, the FDM the multi-faceted connections between the design parameters and
technique exhibits some limitations such as poor part consistency mechanical properties. The 5–33-2 network size was determined
and print repeatability, which are often caused by manufacturing- to be the optimal neural network structure with a mean square
induced defects. We aim to create a real-time monitoring system, error of 0.98% for the compressive modulus, lower than the tradi-
which comprises a camera installed to the print head and the lap- tional response surface method model. A genetic algorithm was
top that handles the video feed, making use of computer vision and then applied to explore the design space for the targeted shore
machine learning models to identify printing anomalies and per- hardness. The best population size and number of generations
form rectifications while printing (Fig. 1) [12]. Image data from were found to be 20 and 200, respectively. One of the drawbacks
defects were gathered and classified into two classes, under extru- of this study is the small sample sizes that do not take into account
sion and over extrusion. Several YOLO object detection models the entire Polyjet’s design windows. Consequently, the effective
were assessed to investigate the capability to identify and recog- shore A hardness range was from 35 to 60A. This could be over-
nize different printing. Four of the trained models, YOLOv3 and come by enlarging dataset’s size to encompass the entire design
YOLOv4 with ‘‘Tiny” variant, attained a mean average precision window and by using knowledge-based transfer learning.

Fig. 1. An illustration of the full in situ monitoring and self-correction system [12].

96
G.D. Goh and W.Y. Yeong Materials Today: Proceedings 70 (2022) 95–100

Fig. 2. The workflow in creating tissue-mimicking anatomical models using machine learning techniques [13].

2.3. Investigation of relationships between continuous-wave process 2.4. Machine learning of lattice design recommendation for multi-
parameters and physical properties of SLM-manufactured Ti-64 using loading lightweight structures
machine learning
Lightweight structures provide superior mechanical properties
Selective Laser Melting (SLM) technology has numerous print despite requiring less materials. These lightweight structures are
parameters that can affect the physical properties of the printed made through additive manufacturing as it grants freedom to
part. As a result, there is high dimensionality in the dataset con- designers. In additive manufacturing, parts are fabricated layer
taining data of process parameters used to print the parts and by layer [3], allowing for creation of complex designs and internal
the corresponding physical properties. Hence, it is time- cavities which are usually difficult to fabricate using traditional
consuming and costly to utilise conventional statistical approaches manufacturing methods. Some popular additive manufacturing
to identify the underlying parameter-property relationships. In this techniques to develop lightweight structures include Topology
study, a data-driven approach is taken to find out the relationships. Optimization, Parts Consolidation and Latticing [4]. During the pro-
Before the investigation of the parameter-property relationships, cess of building a hierarchical lightweight structure using such
data were collated from the online literature. As each research techniques, numerous variables including loads on the structure
work, from the online literature, focuses on various process param- and various material properties are evaluated to obtain the final
eters and physical properties, the dataset built was not complete. design and its characteristics.
Thus, five imputation techniques were explored in this study to fill Although these techniques are useful in optimizing a single
up the missing values, and the median of the imputed results from lightweight structure design, multivariable problems encountered
the five approaches was taken as the final value. Thereafter, Self- during the designing process tend to be high in complexity and
Organising Map (SOM) and Accumulated Local Effect (ALE) plots computationally expensive to evaluate. Long simulation times for
were used to visualise the relationships. The ALE plots were gener- a single case makes it hard to apply these techniques in explora-
ated based on three different models, namely Random Forest, Gra- tory research and rapid prototyping projects, where simulations
dient Boosting, and neural network. From the SOM plots (Fig. 3), using parameters across the design space are required to under-
UTS, yield strength, Young’s modulus, and density are directly cor- stand the relationships between the material used and the struc-
related. Scan speed is inversely correlated to energy density and tural properties. With a lack of efficient exploration techniques,
microhardness. Layer thickness, laser power, and hatch spacing development phases require more time and computing resources.
are directly correlated. From the ALE plots, increase in energy den- Modifications of a structure can take up to 4 years while develop-
sity and its constituent process parameters tend to increase most ing a new type of material for additive manufacturing can take up
of the physical properties before experiencing a decrease. This to 10 or even 20 years.
observation conforms with the interaction between energy density A machine learning method is proposed to assist in the develop-
and density. Both lack and surplus of energy density would lead to ment of lightweight structures by reducing the computational
drop in density due to LOF and keyhole porosities respectively. resources required in the design process. The method will first
Since many of the physical properties are dependent on density, establish a dataset obtained from actual simulations. This dataset
they move in tandem with density. will be introduced to a surrogate model which will model the rela-

97
G.D. Goh and W.Y. Yeong Materials Today: Proceedings 70 (2022) 95–100

Fig. 3. SOM heat maps for all 15 features.

tionships between the various parameters using Gaussian Process saving a high cost. But numerical modelling has a high computa-
Regression (GPR) (Fig. 4). Using this surrogate model, researchers tional cost and requires well-trained operator to execute. Thus,
will be able to map the response surfaces to help designers visual- the goal of the project is to build a machine learning model to pre-
ize the relationship between the variables across the design space. dict the evolution of the microstructure to increase the efficiency
As this method uses a regression technique to fill in the gaps of the entire process. The model aims to be more efficient in gen-
between the existing datapoints, only a small number of data- eration of microstructure as compared to the conventional simula-
points will be required to understand the characteristics of a struc- tion method which can take up to a day or more. A well-trained
ture. Hence, the implementation of this machine learning method machine learning model with high accuracy is able to predict the
in additive manufacturing will help researchers to save on the evolution of the microstructure within a short period of time with-
number of simulations conducted on each structure, significantly out explicit prior knowledge of the physical mechanisms during
reducing the time spent during the development process. the SLM printing process.
To make sure the good accuracy of the model, a large set of
2.5. Development of machine learning model for 1st layer single track high-quality data is needed to be collected and trained by the
microstructure prediction of SLM manufactured parts machine learning algorithms. Results attained could be used to
identify various behaviour and property of the microstructure.
The ability to examine the microstructure through numerical In this project, SLM simulations on steel will be run under dif-
modelling before printing, will avoid the waste of material thus ferent print parameters. The evolution of the microstructure will
98
G.D. Goh and W.Y. Yeong Materials Today: Proceedings 70 (2022) 95–100

Fig. 4. Lattice design ML recommendation method.

be collected in the form of gif file, and it will be use as training

data for the selected machine learning model which is a next- Fig. 6. Summary of applications of ML in 3D printing.
frame video prediction model known as CNN-LSTMs model. The
model will predict the final microstructure with the earlier
frames of image extracted from the gif file (Fig. 5). The final 3. Conclusions
microstructure attained will be evaluated and compare with a
new simulation result to investigate the accuracy of the predic- In this article, the use of machine learning techniques in 3D
tion by the model. printing are presented. Fig. 6 shows the summary of applications

Fig. 5. Simulated microstructure. (a) Ground truth from simulation. (b) Prediction results from CNN-LSTM model.

99
G.D. Goh and W.Y. Yeong Materials Today: Proceedings 70 (2022) 95–100

of ML in 3D printing. Several types of machine learning techniques funding scheme. This work was supported by the Singapore Centre
are discussed, and the challenges and potential are highlighted. for 3D Printing, Nanyang Technological University, Singapore,
Future research can concentrate on integration of the various through the use of its additive manufacturing facilities.
aspects of 3D printing from the design stage to the final fabrication
stage through big data and machine learning, which could consid- References
erably enhance the trustworthiness of the data-driven model. Con-
sequently, the designers can evaluate the estimated performance [1] G. Goh et al., Additively manufactured continuous carbon fiber-reinforced
thermoplastic for topology optimized unmanned aerial vehicle structures,
of 3D printed parts prior to actual fabrication. The development Compos. B Eng. 216 (2021) 108840.
of the predictive models will further expedite the creation of digi- [2] G.D. Goh, S.J.C. Neo, V. Dikshit, W.Y. Yeong, Quasi-static indentation and
tal twins for 3D printing. It provides an immense opportunity for sound-absorbing properties of 3D printed sandwich core panels, J. Sandwich
Struct. Mater. 24 (2) (2022) 1206–1225.
machine learning techniques to develop and be applied in 3D [3] V. Dikshit, G.D. Goh, A.P. Nagalingam, G.L. Goh, W.Y. Yeong, Recent progress in
printing applications. 3D printing of fiber-reinforced composite and nanocomposites, in: Fiber-
Reinforced Nanocomposites: Fundamentals and Applications, Elsevier, 2020,
pp. 371–394.
CRediT authorship contribution statement [4] G.D. Goh, W.Y. Yeong, Mode I interlaminar fracture toughness of additively
manufactured carbon fibre thermoplastic, Proceedings of the 3rd International
Guo Dong Goh: Conceptualization, Writing – original draft. Wai Conference on Progress in Additive Manufacturing, 2018.
[5] A. Armillotta, M. Bellotti, M. Cavallaro, Warpage of FDM parts: Experimental
Yee Yeong: Writing – review & editing, Supervision, Funding tests and analytic model, Rob. Comput. Integr. Manuf. 50 (2018) 140–152.
acquisition. [6] R. Acharya, J.A. Sharon, A. Staroselsky, Prediction of microstructure in laser
powder bed fusion process, Acta Mater. 124 (2017) 360–371.
[7] O. Fergani, F. Berto, T. Welo, S.Y. Liang, Analytical modelling of residual stress in
Data availability additive manufacturing, Fatigue Fract. Eng. Mater. Struct. 40 (6) (2017) 971–978.
[8] Q. Chen, G. Guillemot, C.-A. Gandin, M. Bellet, Three-dimensional finite
element thermomechanical modeling of additive manufacturing by selective
Data will be made available on request.
laser melting for ceramic materials, Addit. Manuf. 16 (2017) 124–137.
[9] X. Qi, G. Chen, Y. Li, X. Cheng, C. Li, Applying Neural-Network-Based Machine
Declaration of Competing Interest Learning to Additive Manufacturing: Current Applications, Challenges, and
Future Perspectives, Engineering 5 (4) (2019) 721–729.
[10] C. Yu, J. Jiang, A perspective on using machine learning in 3D bioprinting, Int. J.
The authors declare that they have no known competing finan- Bioprint. 6 (1) (2020).
cial interests or personal relationships that could have appeared [11] G.D. Goh, S.L. Sing, W.Y. Yeong, A review on machine learning in 3D printing:
to influence the work reported in this paper. applications, potential, and challenges, Artif. Intell. Rev. 54 (1) (2020) 63–94.
[12] G.D. Goh, N.M.B. Hamzah, W.Y. Yeong, Anomaly detection in fused filament
fabrication using machine learning, 3D Print. Addit. Manuf. (2022).
Acknowledgements [13] G.D. Goh et al., Machine learning for 3D printed multi-materials tissue-
mimicking anatomical models, Mater. Des. 211 (2021) 110125.

This research is supported by the National Research Foundation,

Prime Minister’s Office, Singapore under its Medium-Sized Centre

100