Engineering 5 (2019) 721–729

Contents lists available at ScienceDirect

Engineering
journal homepage: www.elsevier.com/locate/eng

Research
Additive Manufacturing—Review

Applying Neural-Network-Based Machine Learning to Additive

Manufacturing: Current Applications, Challenges, and Future
Perspectives
Xinbo Qi a,⇑, Guofeng Chen b, Yong Li a, Xuan Cheng b, Changpeng Li b
a
State Key Laboratory of Tribology, Tsinghua University, Beijing 100084, China
b
Corporate Technology, Siemens Ltd., Beijing 100102, China

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

Article history: Additive manufacturing (AM), also known as three-dimensional printing, is gaining increasing attention
Received 29 July 2018 from academia and industry due to the unique advantages it has in comparison with traditional subtrac-
Revised 6 April 2019 tive manufacturing. However, AM processing parameters are difficult to tune, since they can exert a huge
Accepted 9 April 2019
impact on the printed microstructure and on the performance of the subsequent products. It is a difficult
Available online 3 July 2019
task to build a process–structure–property–performance (PSPP) relationship for AM using traditional
numerical and analytical models. Today, the machine learning (ML) method has been demonstrated to
Keywords:
be a valid way to perform complex pattern recognition and regression analysis without an explicit need
Additive manufacturing
3D printing
to construct and solve the underlying physical models. Among ML algorithms, the neural network (NN) is
Neural network the most widely used model due to the large dataset that is currently available, strong computational
Machine learning power, and sophisticated algorithm architecture. This paper overviews the progress of applying the NN
Algorithm algorithm to several aspects of the AM whole chain, including model design, in situ monitoring, and
quality evaluation. Current challenges in applying NNs to AM and potential solutions for these problems
are then outlined. Finally, future trends are proposed in order to provide an overall discussion of this
interdisciplinary area.
Ó 2019 THE AUTHORS. Published by Elsevier LTD on behalf of Chinese Academy of Engineering and
Higher Education Press Limited Company. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction of fusion and gas entrapment, heavily anisotropic microstructure

in both the perpendicular and parallel directions relative to the
Additive manufacturing (AM), as opposed to traditional printing direction, and distortion due to large residual stress intro-
subtractive manufacturing technologies, is a promising digital duced by a high cooling rate and steep temperature gradient [6]. It
approach for the modern industrial paradigm that has gained is thus essential to better understand the complex relationship
widespread interest all over the world [1–4]. By fabricating objects between a powder’s metallurgical parameters, the printing
layer by layer from three-dimensional (3D) computer-aided design process, and the microstructure and mechanical properties of AM
(CAD) models, AM provides several benefits: ① It creates products parts. The AM process always involves many essential parameters
with complex shapes, such as topologically optimized structures, that can determine the final products’ performance. For example,
which are difficult to manufacture with traditional casting or forg- in selective laser melting (SLM), the processing parameters—which
ing processes; ② it can be used to generate novel characteristics of include laser power, scan speed, hatch spacing, and layer thick-
materials, such as dislocation networks [5], which are very attrac- ness—all significantly affect the quality of the produced parts.
tive to academic researchers; and ③ it reduces material waste and Unfortunately, the relationship between these parameters and
thus saves on cost for industry. However, AM parts also present the output quality is too complicated to fully understand, as SLM
dozens of unique defects that differ from those that appear in their is a multi-physics and multiscale process that includes powder-
cast and wrought counterparts; these include porosity due to a lack laser interaction at the microscale, melt pool dynamics and
columnar grain growth at the mesoscale, and thermal–mechanical
⇑ Corresponding author. coupling at the macroscale. Researchers have tried to develop
E-mail address: [email protected] (X. Qi). various physical models in order to classify this relationship in a

https://doi.org/10.1016/j.eng.2019.04.012
2095-8099/Ó 2019 THE AUTHORS. Published by Elsevier LTD on behalf of Chinese Academy of Engineering and Higher Education Press Limited Company.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
722 X. Qi et al. / Engineering 5 (2019) 721–729

clearer and more accurate way. Acharya et al. [7] developed a com- molten materials are extruded from the nozzle of an FDM printer
putational fluid dynamics (CFD) and phase-field framework to to form layers, as the material hardens immediately after extrusion
simulate grain structure evolution in the as-deposited state for [25]. It can be seen that there are various kinds of AM technologies,
the laser powder-bed fusion (PBF) process; Fergani et al. [8] and that these produce different kinds of data sheets. How to orga-
proposed an analytical model to assess residual stress in the AM nize these data with a unified format and integrate the data-flow
process of metallic materials; and Chen et al. [9] adopted a into the subsequent ML algorithms is a challenging task.
finite-element model to investigate melt pool profiles and bead
shape. As can be seen, the above simulations vary from the powder 2.2. NN algorithm
scale to the part scale, and concentrate on only one or two aspects of
the whole process as a result of the lack of an in-depth understand- NNs are a kind of supervised ML, while other forms of ML are
ing of AM. It is currently impractical to predict the whole AM unsupervised learning. The easiest method to distinguish between
process quickly and accurately via these physics-driven methods these two patterns is to check whether the dataset that they oper-
in a short time. In addition to the abovementioned physics- ate on has labels or not. That is to say, in an NN algorithm, the data
driven models, data-driven models have been widely used in the is labeled—that is, the model has been told the ‘‘answer” to the
field of AM; these models have the unified name of machine inputs. This is suitable for an AM case, since there are always clear
learning (ML) [10,11]. The overwhelming advantage of this kind targets and qualification methods for this manufacturing tech-
of model is that they do not need to construct a long list of nique. An NN has strong evaluating skills for representing complex,
physics-based equations; instead, they automatically learn the highly nonlinear relationships between input and output features,
relationship between the input features and output targets based and it has been shown that a network with only one hidden layer
on previous data. Among ML methods, the neural network (NN) but sufficient neurons can express an arbitrary function. The archi-
algorithm is the most widely used and is currently under rapid tecture or settings of an NN consist of three kinds of layers: the
development, as a result of the massive data available today, the input layer, hidden layer, and output layer [26]. Each layer consists
great availability of computational resources, and its advanced of nodes or neurons, which borrow the idea from neurological
algorithm structure [12]. For example, NNs are the main stimulating sciences. The parameters or coefficients in NN are called weights,
force in these areas: computer vision [13], voice recognition [14], and represent the connection magnitudes between neurons in
natural language processing [15], and autonomous driving [16]. adjacent layers. The values of weights are determined by training
The NN shows its great power in recognizing the underlying the NN iteratively, in order to minimize the loss function between
complicated patterns in the abovementioned tasks, most of which predictions and actual outputs. Within this type of process, the
were once thought to be only possible for human beings. Further- most famous and widely used method for updating weights is
more, there is an obvious trend in that the successful experiences called back propagation, which uses the mathematical chain rule
with utilizing NN in these areas are being transferred into to iteratively compute gradients for each layer [27]. Once training
traditional manufacturing fields (which of course include AM). is achieved, the NN will have the capacity to infer the outputs
The NN has exerted a deep and wide impact on all value chain based on previously unseen inputs. Many types of specific NNs
innovation in industry—from product design, manufacturing, and have been proposed by researchers over the decades of its develop-
qualification to delivery—and it is believed that the impact of NN ment. The following three classes of NNs have proved their value
will be increasingly intensive. This paper provides an overview of and gained wide popularity. ① The multilayer perceptron (MLP)
the current progress achieved by researchers in applying the NN [28] is the most typical NN; its common mathematical operations
algorithm to AM. It is organized as follows: Section 2 gives a brief are linear summation and nonlinear activation (such as the
introduction of AM technologies and the NN algorithm, while sigmoid function). It is widely used in dealing with tabular data.
Section 3 summarizes detailed applications of NN in AM. Section 4 ② The convolutional neural network (CNN) [13] dominates image
outlines challenges and potential solutions, and Section 5 describes processing, since it considers the spatial relationship between
future trends in this area. image pixels. It is named after the mathematical ‘‘convolution”
operation. ③ The recurrent neural network (RNN) [29] plays a
2. Methods key role in dealing with temporal dynamics, since it builds connec-
tions between the nodes in one layer. The most famous RNN is long
2.1. AM technologies short-term memory (LSTM), which accurately reproduces the
finite-element simulation in the following case.
As a terminology, AM is comparable with traditional subtractive
manufacturing (i.e., casting, forging, and computer numerical con-
trol (CNC)); it can be divided into several categories based on dif- 3. Applications
ferent printing technologies [17]. Among them, PBF [18], binder
jetting (BJ) [19], and material extrusion (ME) [20] are three widely AM is a value chain incorporating many aspects: model design,
used technologies. PBF uses a thermal source to build parts layer material selection, manufacturing, and quality evaluation. This sec-
upon layer by sintering or melting fine metal/plastic powders. tion stresses the application of NNs to the following parts of AM:
Based on different application cases, PBF can further be divided design, in situ monitoring, and the process–property–performance
into selective laser sintering (SLS), SLM, electron beam melting linkage.
(EBM), and so on. SLS and SLM both utilize a laser as the thermal
source; however, in SLM, the material is fully melted rather than 3.1. Design for AM
sintered as in SLS [21,22]. In contrast, the thermal source of EBM
is an electron beam, which results in certain advantages such as Design for AM (DfAM) involves building a CAD model of AM
smaller residual stress and less oxidation, in comparison with parts; thus, it is the first and crucial step for the whole processing
laser-based technologies [23]. The BJ process uses two materials: chain. However, there are always deviations between CAD models
a powder-based material and a binder. The binder is selectively and the printed parts, because of residual stress introduced by dis-
deposited onto areas of the powder bed, and bonds these areas tortion in the processing results. Thus, compensation is usually
together to form a solid part one layer at a time [24]. Fused depo- performed in order to obtain an AM part with high accuracy.
sition modeling (FDM) is a kind of ME technology. During printing, Chowdhury and Anand [30] presented an NN algorithm to directly
 X. Qi et al. / Engineering 5 (2019) 721–729 723

compensate the part geometric design, which helps to counter- original 3D models. The CNN model is composed of two convolu-
balance thermal shrinkage and deformation in the manufactured tional layers, two pooling layers, and one fully connected layer.
part. The whole process is as follows: ① A CAD model of the According to the experimental results, the error rate of classifica-
required part is prepared, and its surface 3D coordinates are tion can be reduced to 1.84%.
extracted as the input of the NN model; ② a thermo–mechanical
finite-element analysis software (such as ANSYS or ABAQUS) is 3.2. In situ monitoring
used to simulate the AM process with a defined set of process
parameters. The deformed surface coordinates are extracted as In situ monitoring for data acquisition from multiple sensors
the output of the NN model; ③ an NN model with 14 neurons provides first-hand information regarding product quality during
and mean square error (MSE) as the loss function is trained to learn the AM process. If these real-time data can be analyzed syn-
the difference between the input and output; and ④ the trained chronously and accurately, complete closed-loop control for
network is implemented to STL file to make the required geometric manufacturing is realized. The data source is divided into three
corrections so that manufacturing the part using the modified types, including one-dimensional (1D) data (e.g., spectra), 2D data
geometry results in a dimensional-accurate finished product. (e.g., images), and 3D data (e.g., tomography) [34]. Each data type
Koeppe et al. [31] proposed a framework that combined has its pros and cons. For example, 1D data can be processed faster
experiments, finite-element method (FEM) simulation, and NNs, and its hardware is relatively cheaper; however, it may provide
as illustrated in Fig. 1. First, they conducted actual experiments less useful information than the others. Two examples will be
to validate FEM simulation. Next, FEM was used to run 85 proposed to demonstrate the usage of these different types of
simulation samples based on a different parametric combination signal data. Shevchik et al. [35,36] presented a study on in situ
of global loads, displacement and strut radius, and cell scale. These quality monitoring for SLM using acoustic emission (AE) and
are the NN input features, and the outputs are the maximum Von NNs, which is depicted in Fig. 2. The AE signals are recorded using
Mises and equivalent principal stresses. The NN architecture a fiber Bragg grating sensor, while the selected NN algorithm is a
contains a fully connected layer with 1024 rectified linear units, spectral convolutional neural network (SCNN), which is an
two LSTM cells with 1024 units, respectively, and a fully connected extension of a traditional CNN. The input features of the model
linear output layer. It should be noted here that LSTM is selected are the relative energies of the narrow frequency bands of the
and recommended because of its excellent capacity in dealing with wavelet packet transform. The output feature is a classification of
time series events. After training, an NN can reproduce the loading whether the quality of the printed layer is high, medium, or poor.
history in good agreement with an FEM simulation. From this point It was reported that the classification accuracies using SCNN are as
on, the NN can act as a substitute for traditional numerical high as 83%, 85%, and 89% for high, medium, and poor workpiece
simulation methods with a low operating velocity. qualities, respectively.
Unlike the above two cases, which applied an NN to DfAM, Recently, Zhang et al. [37] built a vision system with a high-
McComb et al. [32] attempted to establish an autoencoder (a kind speed camera for process image acquisition. The system can
of NN that learns from the input and then tries to reconstruct the detect the information of three objects: the melt pool, plume,
input with high accuracy) to learn a low-dimensional representa- and spatter, as illustrated in Fig. 3. The features of these objects
tion of the part design. In addition to this autoencoder, the other are carefully extracted based on the authors’ understanding of
three networks were trained to determine the relationship the physical mechanisms of the process in order to feed them
between the design geometries and three DfAM attributes (i.e., into the traditional ML algorithm. However, the authors stress
part mass, mass of support material, and build time). In this way, that the CNN model does not require this feature-extraction step,
a combination of these four NNs can be utilized to evaluate the as it still has a high accuracy of 92.7% in quality-level identifica-
attributes of parts designed for AM. Another interesting instance tion. It is believed that CNN has great potential to achieve
of applying ML to DfAM concerns the security level of the 3D print- real-time monitoring in industrial applications. The cases men-
ing process. Li et al. [33] trained a CNN to detect and recognize ille- tioned above mainly focus on purely in situ monitoring of the
gal components (e.g., guns) made through AM. After the CNN is AM process; however, the qualification result of the NN model
well constructed, it is integrated into the printers in order to detect cannot affect the real manufacturing in reverse. On the contrary,
gun printing at an early stage and then terminate the manufactur- the following case realizes closed-loop control by seamlessly
ing process in time. The authors collected a dataset of 61 340 two- integrating a vision-based technique and an NN tool for liquid
dimensional (2D) images of ten classes, including guns and other metal jet printing (LMJP) [38]. First, Wang et al. developed a
non-gun objects, corresponding to the projection results of the vision system with a charge-coupled device (CCD) camera to

Fig. 1. Application of an NN model to predict the deformation of an AM structure. (a) Specimens, which are manufactured and tested under controlled loading conditions;
(b) the FEM, whose simulation results are validated by specimens; (c) the NN, which is trained by the data generated by the FEM, and then used to predict the deformation
history in a faster way than the FEM. FC: fully-connected layers. Reproduced from Ref. [31] with permission of Elsevier, Ó 2018.
724 X. Qi et al. / Engineering 5 (2019) 721–729

Fig. 2. Scheme of the AM quality monitoring and analyzing system. The workflow is as follows: An acoustic signal is emitted during the AM process, and then captured by
sensors; an SCNN model is finally applied to the recorded data in order to distinguish whether the quality of the printed layer is adequate or not. Reproduced from Ref. [35]
with permission of Elsevier, Ó 2018.

Fig. 3. Scheme of the SLM process monitoring configuration. A high-speed camera is used to capture sequential images of the built process; a CNN model is applied to identify
quality anomalies. CMOS: complementary metal-oxide semiconductor; ROI: region of interest. Reproduced from Ref. [37] with permission of Elsevier, Ó 2018.

capture the jetting images, which contain various droplet 3.3. Process–property–performance linkage
patterns. Second, they formulated an NN model to establish a
complex relationship between the voltage level and the droplet From a technological and economic point of view, process
features. Thus, the real-time jetting behavior and the ideal parameter selection for the optimization of the performance of
behavior (in which each pulse of the input signal generates only AM parts is highly desirable. Constructing a direct linkage between
a single droplet with sufficient volume and without satellites process, property, and performance is of great interest to scientists
behind it) can be converted into exact voltage values according and engineers. This linkage is often highly nonlinear, since the
to the NN model. Finally, proportional integral derivative (PID) amount of the input variables is usually greater than three. As a
control technology was used to compare these values in order result, it is very difficult to identify the underlying mathematical
to adjust the drive voltage and stabilize the printing process formula for such a linkage. Because of its intrinsic nonlinear
accordingly. characteristics, NN models have been applied to formulate these
 X. Qi et al. / Engineering 5 (2019) 721–729 725

Table 1
NN application to build process–property–performance linkage.

AM technique Processing parameters Property/performance Ref.

FDM Layer thickness, orientation, raster angle, raster width, air gap Compressive strength [39]
FDM Layer thickness, orientation, raster angle, raster width, air gap Wear volume [40]
FDM Orientation, slice thickness Volumetric error [41]
FDM Layer thickness, orientation, raster angle, raster width, air gap Dimensional accuracy [42]
FDM Layer thickness, orientation, raster angle, raster width, air gap Dimensional accuracy [43]
BJ Layer thickness, printing saturation, heater power ration, drying time Surface roughness [44]
BJ Layer thickness, printing saturation, heater power ration, drying time Shrinkage rate (Y-axis) [44]
BJ Layer thickness, printing saturation, heater power ration, drying time Shrinkage rate (Z-axis) [44]
SLS Laser power, scan speed, scan spacing, layer thickness Density [45]
SLS Laser power, scan speed, scan spacing, layer thickness Dimension [46]
SLS Z height, volume, bounding box Build time [47]
SLS Laser power, scan speed, hatch spacing, layer thickness, scan mode, temperature, interval time Shrinkage ratio [48]
SLS Layer thickness, laser power, scan speed Open porosity [49]
SLS Laser power, scan speed, hatch spacing, layer thickness, powder temperature Tensile strength [50]
SLS Laser power, scan speed, hatch spacing, layer thickness, scan mode, temperature, interval time Density [51]
SL Layer thickness, border overcure, hatch overcure, fill cure depth, fill spacing and hatch spacing Dimensional accuracy [52]
LMD Laser power, scanning speed, powder feeding rate Geometrical accuracy [53]
EBM Spreader translation speed, rotation speed Volume, roughness [54]
WAAM Bead width, height, center distance of adjacent deposition paths Offset distance [55]

SL: stereolithography; LMD: laser metal deposition; WAAM: wire and arc additive manufacturing.

mathematical relationships for various AM processes. Table 1 itself. However, the number of neurons for the only hidden
[39–55] summarizes the application of NNs into AM (in fact, NN layer needs to be selected carefully, since it is directly related to
is referred to here as MLP, since all the datasets are the tabular the underfitting and overfitting problems in ML [56]. According
type), and lists the input values of the processing parameters and to Table 2, we suggest 5–10 neurons to be the starting point for
the output values of the property/performance. As can be seen in determining the optimal number of hidden units for AM
Table 1, different AM techniques should select different input fea- applications.
tures, since the key factors in determining the AM part are differ- (3) Activation function. The activation function is the nonlin-
ent. Furthermore, because a large number of parameters can ear transformation over the input signal (x); it decides whether a
exert influence on the final products, determining which parame- neuron should be activated or not. This is of vital importance to
ters to select requires a deep knowledge of the AM process. This the NN, because a network without an activation function is just
topic will be discussed in detail in Section 4.3. a linear regression model, and cannot handle complicated tasks.
The detailed setting of the NN algorithm is summarized in Some popular types of activation functions are as follows:
Table 2. The typical hyperparameters to determine an NN structure
1
usually consist of four parts: the number of hidden layers, number Sigmoid ðxÞ ¼ ð1Þ
1 þ ex
of neurons in one layer, activation function, and loss function.
(1) Number of hidden layers. In the ‘‘Layer/neuron” column of
2
Table 2, ‘‘5-8-1” means that this NN contains three layers: the Tanh ðxÞ ¼ 1 ð2Þ
1 þ e2x
input layer has five neurons, the only hidden layer has eight neu-
rons, and the output layer has one neuron. As can be seen from
ReLU ðxÞ ¼ max ð0; xÞ ð3Þ
the table, one hidden layer is sufficient for a large majority of AM
problems. In a real implementation, the gradient toward either end of the
(2) Number of neurons in one layer. The neuron numbers of sigmoid and tanh functions and at the negative axis of the ReLU
the input layer and output layer are determined by the problem function is going to be small and even zero; as a result, the weights

Table 2
Detailed information on the NN algorithm.

AM technique Layer/neuron Activation function Error function Dataset Error (%) Ref.
FDM 5-8-1 Tanh MAE 32 1.2 [39]
FDM 5-8-1 Tanh — 32 1 [40]
FDM 4-15-12-1 Sigmoid MAE 375 0–7.9 [41]
FDM 5-6-4 — — 27 4.07 [42]
FDM 5-7-3 — — 27 0–0.12 [43]
BJ 4-6-1 Sigmoid MSE 16 0.2–8.5 [44]
BJ 4-20-1 Sigmoid MSE 16 4.0–19.6 [44]
BJ 4-11-1 Sigmoid MSE 16 8.0–29.1 [44]
SLS 4-9-1 Sigmoid SSE 15 7 [45]
SLS 4-6-1 — — 34 1.05–1.360 [46]
SLS 3-7-1 — MSE 130 15 [47]
SLS 7-7-1 — MSE 33 4.35–27.60 [48]
SLS 3-9-1 Tanh RMSE 36 0–9.1 [49]
SLS 5-27-1 Sigmoid MSE 66 0.9–9.2 [50]
SLS 7-8-1 — MAE 32 — [51]
SL 6-20-5 Sigmoid MSE 140 6 [52]
LMD 3-9-3 — RMSE 120 2.0–5.8 [53]
EBM 2-200-2 Sigmoid MAE 45 1.74–2.27 [54]
WAAM 3-12-1 Sigmoid MSE 35 — [55]

MAE: mean absolute error; RMSE: root mean square error; SSE: sum square error.
726 X. Qi et al. / Engineering 5 (2019) 721–729

will not be adjusted during learning. This situation gives rise to the mean l and standard deviation r; when the decoder samples a
vanishing gradient problem. Max–min normalization, which point from this probabilistic distribution, new input data gener-
refines the inputs to a range ð0; 1Þ, is a good supplementary tech- ates. Other generative models, such as generative adversarial nets
nique to avoid this problem. If necessary, batch normalization (GANs) [63] and adversarial autoencoders (AAEs, in a combination
[57] should be used in order to continue to refine input signals in of AE and GAN) [64], can also provide ways to perform data
every layer. augmentation.
(4) Loss function. The loss function should be determined by
the exact problem, and often carries a real-world interpretation. 4.2. Lack of experience in labeling data
For example, both the root mean square error (RMSE) and the
mean absolute error (MAE) are ways of measuring the distance As mentioned before, most of the NN use-cases are supervised
between two vectors: the vector of predictions and the vector of learning, which requires outputs as targets to learn. However,
target values. Their expressions are listed below: sometimes it is very difficult to label data. For example, how can
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi the different objects in Fig. 3 be accurately labeled as melt pool,
Pn 2
i¼1 ðyi  yt Þ plume, or spatters? The authors of Fig. 3 admit that many spatters
RMSE ¼ ð4Þ
n have characteristics that are similar to those of a melt pool in terms
of shape, size, and grey value. In other words, these judgments
Pn heavily rely on the analyst’s deep knowledge of the welding pro-
i¼1 jyi  yt j
MAE ¼ ð5Þ cess. Such a dependency will greatly hinder the development of
n
NNs in the AM area. In other words, the massive application of
where i is the sample index, yi is the predicted value, and yt is the NNs to AM requires deep cooperation between experts in both
targeted value. There are some small variations between them: computer science and material science.
Computing the RMSE corresponds to the L2 norm (i.e., the Euclidean
norm), which is the most common familiar distance; computing 4.3. Lack of knowledge in selecting good features
MAE corresponds to the L1 norm (i.e., the Manhattan norm), which
measures the distance in a rectangular grid from the origin to the Many processing parameters may heavily affect the properties
target. More generally, the Lp norm is expressed by the following: of AM parts, whereas others may have a smaller effect. Meanwhile,
!1=p for a limited dataset, an overabundance of input features can easily
X
n
Lp ¼ jyi  yt jp ð6Þ cause the model to overfit. Therefore, it is of vital importance to
i¼1 ensure that the NN algorithm is operating on a good set of features.
A type of preprocessing on input data named feature engineering
where p is the norm index. The bigger p is, the more sensitive it is can bring considerable benefits to researchers. It can be divided
to large values. For example, since an L2 norm squares the error, into two aspects: ① Feature selection aims to select the most useful
the model will encounter a much larger error than the L1 norm.
features from the existing ones as inputs. For example, people may
If this case is an outlier, the L2 norm will pay more attention to this select ‘‘hatch distance,” ‘‘laser power,” and ‘‘layer thickness” as the
single outlier case, since the errors of many other common cases
most influential factors in determining the properties of parts. In
are smaller. In other words, if it is important to consider any out- this situation, the principles of selection rely on the researchers’
liers, the RMSE method is a better choice. On the other hand, the
experience with and knowledge of AM—that is, in terms of
MAE is more helpful in studies in which outliers may safely and performing in-depth investigation into the mechanisms of the
effectively be ignored. It should be noted that in some special
AM process, not just in terms of doing experiments again and
cases, it may be necessary to consider designing the loss function again. Another useful way is to use statistical tools to perform
in house.
quantitative analysis. The following are some widely used parame-
ters in statistical science. The Pearson correlation coefficient is a
4. Challenges and potential solutions good parameter to measure the linear relationship between two
features; when it is close to 1/1, it indicates that there is a strong
4.1. Small dataset positive/negative correlation between these two inputs. The
Kendall rank correlation coefficient is another parameter to
Since the NN method is data-driven, its performance is directly measure the nonlinear relationship between two features. A scatter
related to the amount of accessible data. Some areas have built matrix is a mathematical tool to plot every numerical attribute
their own big datasets for training, such as ImageNet [58] for image against every other numerical attribute. Through the calculation
recognition, MNIST [59] for optical character recognition, SQuAD of these parameters, it is possible to obtain information on which
[60] for natural language processing, and YouTube-8M [61] for attributes are much more correlated with the targeted property.
video classification. As a result, NNs have demonstrated their great ② Feature combination aims to perform dimensionality reduction
power in these areas. In contrast, AM has no huge dataset, as it is on input features, and thus concentrates on newly produced
always expensive to collect training data. Furthermore, economic features. Once the translation rule is known, manual manipulation
considerations limit the activity of interested parties to create their may be preferable. For example, energy density has been shown to
own open-source dataset. As a result of this dilemma, it is essential have an obvious influence on solidification and metallurgy during
to build up the current small datasets. In fact, certain methods AM processing, as well as on the resulting microstructures and
called generative models can realize data augmentation in order mechanical properties of the fabricated parts [65]. Energy density
to enlarge a dataset artificially. For example, the autoencoder is a (E) is represented in SLM as follows:
representative technology that is capable of randomly generating
P
new data that looks very similar to the training data [11]. It uses E¼ ð7Þ
an encoder to convert the inputs to an internal representation, v hd
and then uses a decoder to generate new outputs that are similar where P is laser power, v is scan speed, h is hatch distance, and d is
to the inputs based on this representation. A famous extension of layer thickness. These four features can then be converted into the
the basic autoencoder is called the variational autoencoder (VAE) novel but influential feature E. Furthermore, it is still possible to use
[62]. It transforms the input into a Gaussian distribution with mathematical tools for assistance, such as applying principle
 X. Qi et al. / Engineering 5 (2019) 721–729 727

components analysis (PCA) to reduce dimensionality based on the it is very easy to access. For application development, AMMD also
feature’s value rather than its attribute. provides a representational state transfer (REST) API for third par-
ties to call.
4.4. The problem of overfitting and underfitting
5.2. Sensing
A good generalization ability is the key goal of an NN algorithm,
and is a measure of how accurately the algorithm is able to predict 5.2.1. Hardware
outputs from previously unknown data. However, a cause of poor As demonstrated in Section 3.2, researchers have developed
performance of an NN algorithm is the overfitting or underfitting several kinds of sensor systems in order to provide real-time
problem. Overfitting means that the NN algorithm tries to fit every information on AM. Sensors are deployed to precisely detect and
data point in the training set; thus, the model is very vulnerable to measure optical, thermal, acoustic, and ultrasonic signals, and to
noises or outliers. In contrast, underfitting means that the NN algo- deliver valuable insights to solidify the understanding of AM. How-
rithm fails to extract the reasonable relationship between the data ever, huge requirements still exist for a reliable sensor system. For
points in the training set. Some techniques to avoid overfitting and example, the sensors installed inside of a printer must survive and
underfitting include regularization [66] and dropout [67]. operate in a harsh environment for a long time. In EBM technology,
the metallic vapor generated by a high-energy electron beam in a
vacuum environment may destroy the camera lens. Furthermore,
5. Future perspectives
the sensor system must be quick enough to capture the central
position of the melt pool, since the laser’s scanning speed is usually
5.1. Data
very fast. From this perspective, a qualified sensor system is highly
desirable in the rapid developing area of AM.
5.1.1. Strengthening the interoperability of APIs for data acquisition
With the rapid development of AM, huge amounts of data are
5.2.2. Software
generated every day. However, the accessibility of these data is
The sensors need to be controlled by powerful operating soft-
not easy across different research groups, since the data in ‘‘these
ware. The basic modes of the control software include monitoring,
isolated islands” usually have inconsistent application program-
recording, analyzing, and storing data. In a typical scenario, such as
ming interfaces (APIs) to call. Thus, a unified API for data acquisi-
during the process of SLM, once the hardware delivers the captured
tion will be beneficial for every stakeholder in this area. The
melt pool image to the software, it may have the capability to com-
paradigm for this kind of qualified API should include well-
pute the temperature profiles and extract thermal and dimensional
defined schema for the thermal–mechanical attributes and pro-
metrics for next-step analysis. Other interesting functional points
cessing parameters of materials, a unified image type for
can be added to the sensing software. For example, it is desirable
microstructure characterization, and the same testing standards
for software to be equipped with the algorithms of detecting voids,
for qualification. In this way, there will be fewer or no barriers to
lack of fusion or porosity, and so forth (especially aided with ML
‘‘fluent” data flow, and a closer e-collaboration will be realized in
methods).
the community.
5.3. Control/optimization
5.1.2. Data preprocessing
Data preprocessing is an essential prerequisite for a data-driven AM builds parts layer by layer, and the quality of every layer
NN algorithm, since it erases ‘‘dirty” data and feeds the correct data exerts a great influence on the properties of the final products.
into the models. However, this step usually includes many cum- As a result, it is necessary to ensure the quality of every layer. Mul-
bersome tasks that need to be accomplished. For example, there tiple types of sensors, such as those capturing photonic, electrical,
is currently a batch of scanning electron microscope (SEM) images sonic, and thermal signals, can provide in situ measurements of the
that contain both grain and porosity information, while the corre- AM process. Closed-loop control can be achieved with the applica-
sponding NN model only requires the crack feature as inputs. The tion of ML in order to analyze this information synchronously and
problem is to accurately extract the crack distribution separately then feed the outputs into the controller of the machine. A possible
from the grain boundaries. It can be a challenge for someone use is to train a CNN to judge whether the quality of a layer is ade-
without solid knowledge and experience in image processing and quate or not based on the layer picture captured by a high-speed
analysis to identify these digital representations of structural camera. In this case, the NN algorithm must quickly respond to
features. One necessary task may be to establish standards and the input picture. Fortunately, some model compression technolo-
best practices for data preprocessing, especially for image features. gies are already available, such as parameter pruning and sharing,
A successful implementation can then be transferred to a broader low-rank factorization, and knowledge distillation [70].
area.
5.4. Whole chain linkage
5.1.3. Database construction
In many material areas, researchers have developed well- Sections 3.2 and 3.3 have shown the great power of NNs in
known databases for organizing/storing/accessing data electroni- building the relationships between structure–property and
cally, such as MatWeb, OQMD, and Citrine [68]. Given the high process–property, respectively. In addition, researchers have con-
complexity and variety of AM, it is necessary to build a unified structed other models to establish the process–structure–
database platform to host the huge amounts of data that are gen- property–performance (PSPP) linkage. For example, Azimi et al.
erated every day by different research groups and different [71] utilized a fully convolutional neural network (FCNN) to
machines. A currently accessible project is the AM Material Data- classify martensite/bainite/pearlite phases in low-carbon steels,
base (AMMD), which was developed by the National Institute of as depicted in Fig. 4. The classification accuracy can reach
Standards and Technology (NIST) [69]. This data management sys- 93.94%, which greatly exceeds the state-of-the-art method with
tem is built with a Not Only Structured Query Language (NoSQL) an accuracy of 48.89%. Although this case is not within the scope
database engine, whose flexible data structure fits the AM case of AM, its concept can easily be transferred to AM; we anticipate
very well. AMMD is web-visualized by the Django framework, so an explosive development in building PSPP linkages using NNs,
728 X. Qi et al. / Engineering 5 (2019) 721–729

Fig. 4. Workflow of martensite/bainite/pearlite classification approach using FCNNs. H: height; W: width. Reproduced from Ref. [71] with permission of Springer Nature,
Ó 2018.

since the latter holds intrinsic advantages in complex pattern corresponding solutions to these challenges, and outlined our
recognition in comparison with other methods and models. thoughts on future trends in this field.

5.5. Modeling forecasting Compliance with ethics guidelines

As mentioned before, the physics-based model is a traditional Xinbo Qi, Guofeng Chen, Yong Li, Xuan Cheng, and Changpeng Li
computational way to reproduce the AM process. However, it declare that they have no conflict of interest or financial conflicts to
requires substantial computational cost in terms of time, hard- disclose.
ware, and software. As demonstrated in Section 3.1, it is possible
to learn from previously accumulated numerical datasets and References
extract the embedded linkages between inputs and simulated
outputs. In other words, numerical simulations can be a data [1] Lu B, Li D, Tian X. Development trends in additive manufacturing and 3D
printing. Engineering 2015;1(1):85–9.
source for the ML algorithm, and can play the same role as exper- [2] Derby B. Additive manufacture of ceramics components by inkjet printing.
imental data. Popova et al. [72] developed a data science workflow Engineering 2015;1(1):113–23.
to combine ML with simulations. This workflow is then applied to a [3] Gu D, Ma C, Xia M, Dai D, Shi Q. A multiscale understanding of the
thermodynamic and kinetic mechanisms of laser additive manufacturing.
set of AM microstructures obtained using the Potts kinetic Monte
Engineering 2017;3(5):675–84.
Carlo (kMC) approach, which is open-source hosted in the Harvard [4] Herzog D, Seyda V, Wycisk E, Emmelmann C. Additive manufacturing of
Dataverse [73]. Karpatne et al. [74] proposed the concept of metals. Acta Mater 2016;117(15):371–92.
theory-guided data science (TGDS) as a new paradigm for integrat- [5] Liu L, Ding Q, Zhong Y, Zou J, Wu J, Chiu YL, et al. Dislocation network in
additive manufactured steel breaks strength–ductility trade-off. Mater Today
ing physics-based models and data-driven models. They have 2018;21(4):354–61.
identified five broad categories of approaches for combining [6] Gorsse S, Hutchinson C, Gouné M, Banerjee R. Additive manufacturing of
scientific knowledge with data science in diverse disciplines. In metals: a brief review of the characteristic microstructures and properties of
steels, Ti–6Al–4V and high-entropy alloys. Sci Technol Adv Mater 2017;18
the near future, the combination of these two kinds of models will (1):584–610.
certainly provide a way to address the current issues of a lack of [7] Acharya R, Sharon JA, Staroselsky A. Prediction of microstructure in laser
experimental AM data, non-interpretable NN models, and so on. powder bed fusion process. Acta Mater 2017;124:360–71.
[8] Fergani O, Berto F, Welo T, Liang SY. Analytical modelling of residual stress in
additive manufacturing. Fatigue Fract Eng Mater Struct 2017;40(6):971–8.
[9] Chen Q, Guillemot G, Gandin CA, Bellet M. Three-dimensional finite element
6. Conclusion thermomechanical modeling of additive manufacturing by selective laser
melting for ceramic materials. Addit Manuf 2017;16:124–37.
Two explosive developments have recently occurred in the [10] Kohavi R, Provost F. Glossary of terms. Mach Learn 1998;30(2–3):271–4.
[11] Géron A. Hands-on machine learning with Scikit-Learn and TensorFlow:
areas of manufacturing and information technology: AM and concepts, tools, and techniques to build intelligent systems. Boston: O’Reilly
NNs. AM has advantages such as integration with digital CAD mod- Media, Inc.; 2017.
els and the capability to build parts with complex morphology, [12] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521(7553):436–44.
[13] Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep
while NNs excel at avoiding constructing and solving complicated convolutional neural networks. In: Pereira F, Burges CJC, Bottou L, Weinberger
multiscale and multi-physical mathematical models. The combina- KQ, editors. Advances in Neural Information Processing Systems 25:
tion of AM and NN has demonstrated great potential for realizing Proceedings of Neural Information Processing Systems 2012; 2012 Dec 3–6;
Lake Tahoe, NV, USA. p. 1097–105.
the attractive concept of ‘‘agile manufacturing” in industry. This [14] Anusuya MA, Katti SK. Speech recognition by machine, a review. 2010.
paper provided a comprehensive overview of current progress in arXiv:1001.2267.
applying the NN algorithm to the complete AM process chain, from [15] Devlin J, Chang M, Lee K, Toutanova K. Bert: pre-training of deep bidirectional
transformers for language understanding. 2018. arXiv:1810.04805.
design to post-treatment. The scope of this work covers many [16] Ondruska P, Posner I. Deep tracking: seeing beyond seeing using recurrent
variants of NNs in various application scenarios, including: a neural networks. 2016. arXiv: 1602.00991.
traditional MLP for linking the AM process, properties, and perfor- [17] ISO/ASTM52900-15: Standard terminology for additive manufacturing—
general principles—terminology. ASTM standard. West Conshohocken: ASTM
mance; a convolutional NN for AM melt pool recognition; LSTM for
International; 2015.
reproducing finite-element simulation results; and the VAE for [18] King WE, Anderson AT, Ferencz RM, Hodge NE, Kamath C, Khairallah SA, et al.
data augmentation. However, as they say, ‘‘every coin has two Laser powder bed fusion additive manufacturing of metals; physics,
sides”: It is difficult to control the quality of AM parts, while NNs computational, and materials challenges. Appl Phys Rev 2015;2(4):041304.
[19] Gibson I, Rosen D, Stucker B. Binder jetting. In: Gibson I, Rosen D, Stucker B,
rely strongly on data collection. Thus, some challenges remain editors. Additive manufacturing technologies: 3D printing, rapid prototyping,
in this interdisciplinary area. We have proposed potential and direct digital manufacturing. New York: Springer; 2015. p. 205–18.
 X. Qi et al. / Engineering 5 (2019) 721–729 729

[20] Chacón JM, Caminero M, García-Plaza E, Núñez P. Additive manufacturing of International Conference on Information Engineering and Computer Science;
PLA structures using fused deposition modelling: effect of process parameters 2009 Dec 19–20; Wuhan, China; 2009.
on mechanical properties and their optimal selection. Mater Des [47] Munguía J, Ciurana J, Riba C. Neural-network-based model for build-time
2017;124:143–57. estimation in selective laser sintering. Proc Inst Mech Eng B J Eng Manuf
[21] Kruth JP, Wang X, Laoui T, Froyen L. Lasers and materials in selective laser 2009;223(8):995–1003.
sintering. Assem Autom 2003;23(4):357–71. [48] Wang RJ, Li XH, Wu QD, Wang LL. Optimizing process parameters for selective
[22] Kruth JP, Mercelis P, Van Vaerenbergh J, Froyen L, Rombouts M. Binding laser sintering based on neural network and genetic algorithm. Int J Adv Manuf
mechanisms in selective laser sintering and selective laser melting. Rapid Technol 2009;42(11–12):1035–42.
Prototyping J 2005;11(1):26–36. [49] Garg A, Tai K, Savalani MM. State-of-the-art in empirical modelling of rapid
[23] Murr LE, Gaytan SM, Ramirez DA, Martinez E, Hernandez J, Amato KN, et al. prototyping processes. Rapid Prototyp J 2014;20(2):164–78.
Metal fabrication by additive manufacturing using laser and electron beam [50] Wang CY, Jiang N, Chen ZL, Chen Y, Dong Q. Prediction of sintering strength for
melting technologies. J Mater Sci Technol 2012;28(1):1–14. selective laser sintering of polystyrene using artificial neural network. J
[24] Bai Y, Williams CB. An exploration of binder jetting of copper. Rapid Donghua Universit 2015;5:825–30.
Prototyping J 2015;21(2):177–85. [51] Wang RJ, Li J, Wang F, Li X, Wu Q. Ann model for the prediction of density in
[25] Sood AK, Ohdar RK, Mahapatra S. Parametric appraisal of mechanical property selective laser sintering. Int J Manuf Res 2009;4(3):362–73.
of fused deposition modelling processed parts. Mater Des 2010;31(1):287–95. [52] Lee SH, Park WS, Cho HS, Zhang W, Leu MC. A neural network approach to the
[26] Goldberg Y. Neural network methods for natural language processing. Synth modelling and analysis of stereolithography processes. Proc Inst Mech Eng B J
Lect Hum Lang Technol 2017;10(1):1–309. Eng Manuf 2001;215(12):1719–33.
[27] Rumerlhart DE, Hinton GE, Williams RJ. Learning representations by back- [53] Caiazzo F, Caggiano A. Laser direct metal deposition of 2024 Al alloy: trace
propagating errors. Nature 1986;323(6088):533–6. geometry prediction via machine learning. Materials 2018;11(3):444.
[28] Gardner MW, Dorling S. Artificial neural networks (the multilayer [54] Zhang W, Mehta A, Desai PS, Higgs III CF. Machine learning enabled powder
perceptron)—a review of applications in the atmospheric sciences. Atmos spreading process map for metal additive manufacturing (AM). In:
Environ 1998;32(14–5):2627–36. Proceedings of the 28th Annual International Solid Freeform Fabrication
[29] Mikolov T, Karafiát M, Burget L, Černocký J, Khudanpur S. Recurrent neural Symposium; 2017 Aug 7–9; Austin, TX, USA. 2017. p. 1235–49.
network based language model. In: Proceedings of the 11th Annual Conference [55] Li Y, Sun Y, Han Q, Zhang G, Horváth I. Enhanced beads overlapping model for
of the International Speech Communication Association; 2010 Sep 26–30; wire and arc additive manufacturing of multi-layer multi-bead metallic parts. J
Makuhari, Japan; 2010. Mater Process Technol 2018;252:838–48.
[30] Chowdhury S, Anand S. Artificial neural network based geometric [56] Xu S, Chen L. A novel approach for determining the optimal number of hidden
compensation for thermal deformation in additive manufacturing processes. layer neurons for FNN’s and its application in data mining. In: Proceedings of
In: Proceedings of the 11th International Manufacturing Science and the 5th International Conference on Information Technology and Applications;
Engineering Conference; 2016 June 27– July 1; Blacksburg, VA, USA; 2016. 2008 June 23–26; Cairns, Australia. 2008. p. 683–6.
[31] Koeppe A, Hernandez Padilla CA, Voshage M, Schleifenbaum JH, Markert B. [57] Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by
Efficient numerical modeling of 3D-printed lattice-cell structures using neural reducing internal covariate shift. 2015. arXiv: 1502.03167.
networks. Manuf Lett 2018;15:147–50. [58] Deng J, Dong W, Socher R, Li LJ, Li K, Li FF. ImageNet: a large-scale hierarchical
[32] McComb C, Meisel N, Murphy C, Simpson TW. Predicting part mass, required image database. In: Proceedings of 2009 IEEE Computer Society Conference on
support material, and build time via autoencoded voxel patterns. EngrXiv. Computer Vision and Pattern Recognition; 2009 June 20–25; Miami, FL, USA.
Epub 2018 Jul 4. 2009. p. 248–55.
[33] Li H, Ma X, Rathore AS, Li Z, An Q, Song C, et al. Image dataset for visual objects [59] LeCun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to
classification in 3D printing. 2018. arXiv:1803.00391. document recognition. Proc IEEE 1998;86(11):2278–324.
[34] Everton SK, Hirsch M, Stravroulakis P, Leach RK, Clare AT. Review of in-situ [60] Rajpurkar P, Zhang J, Lopyrev K, Liang P. SQuAD: 100,000+ questions for
process monitoring and in-situ metrology for metal additive manufacturing. machine comprehension of text. 2016. arXiv:1606.05250.
Mater Des 2016;95:431–45. [61] Abu-El-Haija S, Kothari N, Lee J, Natsev P, Toderici G, Varadarajan B, et al.
[35] Shevchik SA, Kenel C, Leinenbach C, Wasmer K. Acoustic emission for in situ YouTube-8M: a large-scale video classification benchmark. 2016.
quality monitoring in additive manufacturing using spectral convolutional arXiv:1609.08675.
neural networks. Addit Manuf 2018;21:598–604. [62] Kingma DP, Welling M. Auto-encoding variational bayes. 2013. arXiv:1312.6114.
[36] Wasmer K, Le-Quang T, Meylan B, Shevchik SA. In situ quality monitoring in [63] Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, et al.
AM using acoustic emission: a machine learning approach. J Mater Eng Generative adversarial nets. In: Proceedings of Advances in Neural Information
Perform 2019;28(2):666–72. Processing Systems 27; 2014 Dec 8–13; Montreal, QC, Canada. 2014. p.
[37] Zhang Y, Hong GS, Ye D, Zhu K, Fuh JY. Extraction and evaluation of melt pool, 2672–80.
plume and spatter information for powder-bed fusion AM process monitoring. [64] Makhzani A, Shlens J, Jaitly N, Goodfellow I, Frey B. Adversarial autoencoders.
Mater Des 2018;156:458–69. 2015. arXiv:1511.05644.
[38] Wang T, Kwok TH, Zhou C, Vader S. In-situ droplet inspection and closed-loop [65] Yusuf SM, Gao N. Influence of energy density on metallurgy and properties in
control system using machine learning for liquid metal jet printing. J Manuf metal additive manufacturing. Mater Sci Technol 2017;33(11):1269–89.
Syst 2018;47:83–92. [66] Ng AY. Feature selection, L1 vs. L2 regularization, and rotational invariance. In:
[39] Sood AK, Ohdar RK, Mahapatra SS. Experimental investigation and empirical Proceedings of the 21st International Conference on Machine Learning; 2004
modelling of FDM process for compressive strength improvement. J Adv Res July 4–8; Banff, AB, Canada; 2004.
2012;3(1):81–90. [67] Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a
[40] Sood AK, Equbal A, Toppo V, Ohdar R, Mahapatra S. An investigation on sliding simple way to prevent neural networks from overfitting. J Mach Learn Res
wear of FDM built parts. CIRP J Manuf Sci Technol 2012;5(1):48–54. 2014;15(1):1929–58.
[41] Vosniakos GC, Maroulis T, Pantelis D. A method for optimizing process [68] Ramakrishna S, Zhang TY, Lu WC, Qian Q, Low JSC, Yune JHR, et al. Materials
parameters in layer-based rapid prototyping. Proc Inst Mech Eng B J Eng informatics. J Intell Manuf. 2019;30(6):2307–26.
Manuf 2007;221(8):1329–40. [69] Lu Y, Witherell P, Donmez A. A collaborative data management system for
[42] Equbal A, Sood AK, Mahapatra S. Prediction of dimensional accuracy in fused additive manufacturing. In: Proceedings of ASME 2017 International Design
deposition modelling: a fuzzy logic approach. Int J Product Qual Manag 2011;7 Engineering Technical Conferences and Computers and Information in
(1):22–43. Engineering Conference; 2017 Aug 6–9; Cleveland, Oh, USA; 2017.
[43] Sood AK, Ohdar RK, Mahapatra SS. Parametric appraisal of fused deposition [70] Cheng Y, Wang D, Zhou P, Zhang T. A survey of model compression and
modelling process using the grey taguchi method. Proc Inst Mech Eng B J Eng acceleration for deep neural networks. 2017. arXiv:1710.09282.
Manuf 2010;224(1):135–45. [71] Azimi SM, Britz D, Engstler M, Fritz M, Mücklich F. Advanced steel
[44] Chen H, Zhao YF. Learning algorithm based modeling and process parameters microstructural classification by deep learning methods. Sci Rep 2018;8
recommendation system for binder jetting additive manufacturing process. In: (1):2128.
Proceedings of 2015 International Design Engineering Technical Conferences [72] Popova E, Rodgers TM, Gong X, Cecen A, Madison JD, Kalidindi SR. Process-
and Computers and Information in Engineering Conference; 2015 Aug 2–5; structure linkages using a data science approach: application to simulated
Boston, MA, USA; 2015. p. V01AT02A029. additive manufacturing data. Integr Mater Manuf Innov 2017;6(1):54–68.
[45] Shen X, Yao J, Wang Y, Yang J. Density prediction of selective laser sintering [73] Rodgers T. Exploration of process-structure linkages in simulated additive
parts based on artificial neural network. In: Yin FL, Wang J, Guo C, editors. manufacturing microstructures. Harvard Dataverse 2015.
Advances in neural networks—ISNN 2004. Berlin: Springer; 2004. p. 832–40. [74] Karpatne A, Atluri G, Faghmous JH, Steinbach M, Banerjee A, Ganguly A, et al.
[46] Li XF, Dong JH, Zhang YZ. Modeling and applying of RBF neural network based Theory-guided data science: a new paradigm for scientific discovery from
on fuzzy clustering and pseudo-inverse method. In: Proceedings of 2009 data. IEEE Trans Knowl Data Eng 2017;29(10):2318–31.