Received: 26 December 2022 Revised: 17 March 2023 Accepted: 1 April 2023

DOI: 10.1002/amp2.10159

REVIEW

A review of artificial intelligence applications

in manufacturing operations

Siby Jose Plathottam | Arin Rzonca | Rishi Lakhnori | Chukwunwike O. Iloeje

Energy Systems and Infrastructure

Analysis Division, Argonne National Abstract
Laboratory, Lemont, Illinois, USA Artificial intelligence (AI) and machine learning (ML) can improve
manufacturing efficiency, productivity, and sustainability. However, using
Correspondence
Siby Jose Plathottam and Chukwunwike AI in manufacturing also presents several challenges, including issues with
O. Iloeje, Energy Systems and data acquisition and management, human resources, infrastructure, as well
Infrastructure Analysis Division, Argonne
as security risks, trust, and implementation challenges. For example, get-
National Laboratory, Lemont, IL 60439,
USA. ting the data needed to train AI models can be difficult for rare events or
Email: [email protected] and costly for large datasets that need labeling. AI models can also pose security
[email protected]
risks when integrated into industrial control systems. In addition, some
Funding information industry players may be hesitant to use AI due to a lack of trust or under-
Office of Energy Efficiency and Renewable standing of how it works. Despite these challenges, AI has the potential to
Energy, Grant/Award Number: DE-
AC02-06CH11357 be extremely helpful in manufacturing, particularly in applications such as
predictive maintenance, quality assurance, and process optimization. It is
important to consider the specific needs and capabilities of each
manufacturing scenario when deciding whether and how to use AI in
manufacturing. This review identifies current developments, challenges,
and future directions in AI/ML relevant to manufacturing, with the goal of
improving understanding of AI/ML technologies available for solving
manufacturing problems, providing decision-support for prioritizing and
selecting appropriate AI/ML technologies, and identifying areas where fur-
ther research can yield transformational returns for the industry. Early
experience suggests that AI/ML can have significant cost and efficiency
benefits in manufacturing, especially when combined with the ability to
capture enormous amounts of data from manufacturing systems.

KEYWORDS
AI, AI challenges, industry automation, industry operations, machine learning,
manufacturing industry

Rishi Lakhnori and Arin Rzonca were Research Intern at Argonne National Laboratory at the time of contribution.

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided
the original work is properly cited.
© 2023 UChicago Argonne, LLC, Operator of Argonne National Laboratory. Journal of Advanced Manufacturing and Processing published by Wiley Periodicals LLC on
behalf of American Institute of Chemical Engineers.

J Adv Manuf Process. 2023;5:e10159. wileyonlinelibrary.com/journal/amp2 1 of 19

https://doi.org/10.1002/amp2.10159
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 of 19 PLATHOTTAM ET AL.

1 | INTRODUCTION infrastructure required to collect and process data. Second,

it can be challenging to recruit additional human resources
1.1 | Overview with AI/ML expertise, and train existing personnel for new
roles involving AI/ML solutions. Third, interpreting predic-
The transformation of artificial intelligence (AI) and tive outcomes and deriving and implementing actionable
machine learning (ML) from computer science theory intelligence is nontrivial. Finally, several aspects of AI/ML
into real-world technologies is a key enabler of the fourth technology are not fully mature, and there is a non-zero
industrial revolution (Industry 4.0), to the extent that it probability that implementation might not yield sufficient
integrates AI/ML and other emerging technologies to return on investment to justify it.[4] Additional challenges
transform industry. Governments and industries world- include, (a) unintended security risks could arise when
wide have recognized the strategic implications of AI/ML AI/ML solutions are introduced into industrial control sys-
technologies and launched a host of initiatives seeking to tems, (b) computationally intensive AI/ML models could
explore and capitalize on this new revolution by incorpo- increase the energy and environmental footprint of
rating of AI/ML into manufacturing and industrial pro- manufacturing facilities, (c) AI/ML techniques have the
cesses. These initiatives involve bringing AI/ML onto the capability to take on some of the higher level decision mak-
factory floor[1,2] and integrating information technology ing responsibility, and by doing so, fundamentally alter the
advances (e.g., Internet of Things [IoT], big data analyt- nature of human-machine interaction in a manufacturing
ics, edge computing, and cybersecurity) into the existing plant. However, there is limited trust in the reliability of
process automation infrastructure.[3] With such AI/ML AI/ML techniques, lack of interpretability in the outputs
solutions, the manufacturing industry can leverage the from AI/ML models, and behavioral inertia towards the
vast amounts of data created by measurement devices on culture change that will be brought about by introducing
the factory floor to improve manufacturing efficiency, AI/ML to manufacturing, and (d) the field of AI/ML is con-
productivity, and sustainability.[4–7] stantly evolving, which makes implementation challenging,
Incorporating AI into manufacturing is considered especially for firms that are not computer science
distinct from digitization and integration of information technology-oriented, and without ready access to required
technology. The latter may be seen as a pre-requisite for ML expertise.[9] However, as research continues to develop,
the former, that is, digitization and information technol- these areas of concern can become minimized through, for
ogy provide the infrastructure required to implement example, (i) the use of generative models to supplement
AI/ML-based solutions. In the same vein, AI/ML solu- sparse datasets, (ii) developing power and memory-efficient
tions can provide additional value for established digitiza- computing architectures for IoT devices, (iii) developing
tion and information technologies by extracting new, appropriate metrics to quantify confidence in decisions
actionable intelligence from data, such as better process made through AI/ML models, and (iv) the growth of a wide
control paradigms or optimized preventive maintenance variety of automated AI/ML tools provided as “Software as
schedules which leverage large volumes of historical a Service (SaaS)” that save individual companies the need
operations and failure mode data, as well as better busi- to build their own in-house AI/ML capability.
ness insights from data analytics. Early experience in the
industry demonstrates the potential of AI/ML to bolster
cost, efficiency, and productivity gains for a wide range of 1.2 | Study approach
applications. These applications include predictive main-
tenance to improve real-time monitoring of equipment Literature review on AI/ML in manufacturing primarily
performance to reduce the likelihood of unexpected fail- focused on three sources, namely, academic literature,
ures; quality assurance to identify product imperfections blog posts, and industry reports. In this review, we priori-
and support factory floor error detection; energy forecast- tized academic literature, and used other sources for spe-
ing to improve sustainability and manage energy needs; cific use-case illustrations. The initial survey started with
safety and security to mitigate cybersecurity risks and over 200 sources, which we narrowed down to around
rapidly detect and flag unsafe practices; generative design 100 sources using the following criteria: (a) published
to drive rapid topology optimization in product design, within the last 10 years; (b) focuses on AI/ML applied to
and experimentation to simulate normal and anomalous manufacturing; (c) uses mature AI/ML techniques which
behavior without needing to run disruptive tests on the have found success in non-manufacturing industry appli-
actual manufacturing process.[8] cations. The main takeaway from each technical article is
Incorporating AI into manufacturing processes and summarized and categorized by manufacturing applica-
facilities faces significant challenges. First, it can be capital- tion and AI technique. The review also incorporated dif-
intensive, in terms of the hardware and software fering perspectives on certain applications of AI/ML in
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
PLATHOTTAM ET AL. 3 of 19

F I G U R E 1 (A) The technical articles used for this manuscript were grouped by their year of publication and the number of publications
that were vetted from 2017 to 2021. This 2021 data was collected mid-year, so does not reflect the total count. (B) Word cloud displaying the
125 most common words among all references used. Results are out of a total of 67 different documents and 640 634 total words.

manufacturing to provide additional context for the chal- in manufacturing. This upward trend reflects a convergence
lenges and benefits AI/ML has in manufacturing. For easy of interests from both the research community and the
organization, each reference was summarized and tagged manufacturing industry in exploring opportunities for
according to year (as illustrated in Figure 1A), location of leveraging AI/ML to improve value.
publication, industry, AI technique, and manufacturing The manuscript is organized as follows: Section 2
application. Voyant Tools,[10] a text analysis application, provides a brief description of mature AI/ML algorithms
was used to provide an informal representation of topic and models which are used in manufacturing applica-
prevalence among the reviewed literature targeted at tions. Section 3 describes AI/ML applications in the
AI/ML applications in industry, shown in the world cloud manufacturing verticals such as operations, design, and
in Figure 1B. The relative sizes of each word in the word automation. Section 4 describes the top four challenges
cloud provides a measure of the frequency with which encountered when trying to deploy AI/ML manufactur-
they are encountered. It suggests the centrality of certain ing applications within an existing manufacturing plant.
themes that combine elements such as data, learning, AI, Section 5 provides a brief description of the top four
systems intelligence, process, and research in manufacturing trends in AI/ML research and development which
applications. The article count chart in Figure 1B provides a would aid AI/ML manufacturing applications, followed
proxy indicator of the growing interest in AI/ML applications by conclusion in Section 6.
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
4 of 19 PLATHOTTAM ET AL.

2 | AI PARADIGMS, TECHNIQUES, features. A trained model can then be used to predict labels
AND WORKFLO WS for a new set of input features for which there are no
labels.[14] Supervised learning is particularly useful for image,
AI is an umbrella term that encompasses a broad range of voice, and object recognition, and in general, for applications
techniques and approaches, which has emerged as a major where large, labeled datasets can be obtained easily.
field in computer science. Almost all AI programs are
meant for solving a single task for which it was specifically
developed,[11] so, it would be apt to use the term artificial 2.1.2 | Unsupervised learning
narrow intelligence (ANI) as opposed to AI. In earlier years,
AI programs were mostly the so-called “expert systems.” In unsupervised learning, the ML model is trained using
These were computer programs that mimicked expert only the input features with no corresponding labels.[15] The
human decisions for a given task. The expert knowledge goal of training in unsupervised learning may be reducing
was hard-coded into a computer program as a set of rules, the dimensionality of input features (e.g., Principal compo-
based on which the program performed logical inference to nent analysis), clustering similar data points (e.g., k-means
provide an output that closely mirrors a human expert. clustering), mimicking the training dataset (e.g., Autoenco-
Then came approaches based on heuristics, such as evolu- ders), or finding anomalous data points (e.g., Anomaly detec-
tionary algorithms, which discover solutions on their own tion).[16] During training, the ML models for unsupervised
while maximizing a performance metric. In recent years, AI learning discover hidden patterns in the data without the
systems based on ML, and specifically deep learning have need for human intervention.
gained popularity. These AI/ML systems need not be mono-
lithic and may be comprised of several different techniques.
2.1.3 | Reinforcement learning

2.1 | Machine learning paradigms In reinforcement learning (RL), there is no pre-existing

training dataset. Instead, an ML agent interacts with an
ML refers to a set of algorithms and models that can learn environment and generates the training data which consists
to identify patterns and make decisions to solve specific of observations (input features), actions taken by the agent
tasks using data related to that task.[12] Software based on when interacting with the environment, and a scalar value
ML is developed by sourcing datasets that are related to that captures the net benefit of agent actions (rewards).
the task (referred to as the training data), selecting a suit- Here the environment refers to software that encapsulates
able ML model, and training the ML model to complete the problem that we want to solve.[17] In reinforcement
the task.[13] ML may be broadly classified into three main learning, the ML agent learns based on feedback from its
learning paradigms, namely supervised learning, unsuper- environment. It takes action to explore its environment and
vised learning, and reinforcement learning. As shown in transitions into a new state after every action. For each
Figure 2, different machine learning models (techniques) state, the agent receives a positive or a negative reward
may combine one or more of these learning paradigms for based on whether its action is desirable for the given state.
a given learning task. The relationship between learning As this process is repeated, the agent gradually learns to
paradigms, learning models, and tasks that is illustrated in seek out positive actions and avoid negative ones and even-
Figure 2 is based on distilling the information from techni- tually determines an optimal (maximum-reward) path to its
cal articles and author experience. A combination of learn- goal.[18–20] Unlike supervised learning algorithms, reinforce-
ing tasks can be used to support a wide range of ment learning does not need correct answers or targets to
applications. Note that there is no single standardized solve a given problem, it only needs information on
approach for categorizing ML paradigms and techniques whether an answer from the ML agent is in the correct
and alternative classifications exist. direction. Reinforcement learning has useful applications in
robotics, as well as optimization.

2.1.1 | Supervised learning

2.2 | Machine learning techniques
In supervised learning, the ML algorithm is trained using
data that has a set of inputs (features) and corresponding Machine learning techniques refer to the methods and
labels (labeled dataset). The labels might either be a discrete algorithms that enable learning from data using one of
class or a continuous value. During training, the model the learning paradigms mentioned in Section 2.1. Most
learns to correctly predict the label associated with the input ML techniques are specifically suited for a single learning
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
PLATHOTTAM ET AL. 5 of 19

F I G U R E 2 Common categories for various aspects of machine learning, grouped into paradigms, techniques, tasks, and relevant
manufacturing industry applications.

paradigm. For example, decision tree models are exclu- on tabular data. In general, ML techniques can take mul-
sively used for supervised learning. However, a few tiple data types with multiple dimensions as input.
models like neural networks can be used for learning all
three paradigms. While multiple techniques can be used
on the same types of problems, there are instances where 2.2.1 | Neural networks
one technique performs better than others. For instance,
neural networks are better suited to computer vision Artificial neural networks (ANNs) are the most powerful
problems while decision trees work better for regression learning model currently available and can be used for
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6 of 19 PLATHOTTAM ET AL.

supervised, unsupervised, and reinforcement learning. hyperplane that has the maximum margin, that is, the
They have a weak analogy to the network in the human maximum distance between data points of the different
brain formed by the biological neurons[6,21] and are com- classes. SVMs have the advantage of quickly classifying
prised of connected layers of artificial neurons. Each neu- and categorizing data, which reduces the costs related to
ron transmits the weighted sum of inputs through a manually sorting the data.
nonlinear activation function (e.g., sigmoid, rectified lin-
ear) to produce an output that proceeds to the next layer
of neurons.[22] An ANN with two or more layer of neu- 2.2.4 | Clustering algorithms
rons is known as a deep neural network (DNN), which is
more widely referred to as deep learning in both industry Clustering algorithms are unsupervised learning algo-
and academia. DNNs can be further classified into convo- rithms that employ an iterative process to sort data into
lutional neural networks (CNNs), which apply a filter at specific categories or groupings known as clusters based
each layer of nodes meant to observe a different feature on the “nearness” (e.g., Euclidean distance) of the data
and preserve spatial information; and Recurrent Neural points to a center of gravity. This machine learning tech-
Networks, whose nodes retain information from previous nique is particularly useful for large sets of data as the
inputs and preserve state information. resulting clusters can give rise to conclusions or previ-
ously undiscovered patterns within sets of data, which
can be visually represented.[28]
2.2.2 | Decision trees

Decision trees fall into the supervised learning category 2.2.5 | Generative adversarial networks
and are a tree-like learning model of decisions that weigh
decisions based on factors like consequences, likelihood, Generative modeling is an application of unsupervised
and associated costs. Decision trees are visual maps that learning to develop a probabilistic model that describes
represent the outcomes of related choices. They start with training datasets. Generative adversarial networks
a single node—analogous to a root of a tree—that, for (GANs) have been one of the most successful learning
instance, represents a decision such as whether to make models for generative modeling. GANS consists of two
a sub-assembly of a product in-house or to outsource the neural networks, a generator that learns from a training
assembly[23] and then branches into possible outcomes dataset to generate new data that could plausibly have
(“Yes” or “No” in this case) based on the weighted benefit come from the original dataset, and a discriminator,
of answers to that question.[24] These possible outcomes which evaluates whether the data point is from the origi-
then lead to more nodes with each having its own dis- nal dataset or created by the generator. During training,
tinct set of possibilities, giving it a tree-like structure. The the generator and discriminator try to outcompete each
tree is constructed such that the decisions taken within other until the generator starts producing more and more
the tree lead to classifications that result in the least plausible results.[21]
entropy. Decision trees are particularly useful because
they allow users to quantitatively weigh the outcomes of
actions based on parameters such as cost, benefits, and 2.2.6 | Scientific machine learning
likelihood,[25] and can be supplemented with algorithms
that objectively demonstrate the best action to take. AI/ML has found increasing use in the domain of scien-
tific computing through an emerging field known as sci-
entific machine learning (SciML). SciML is a data-driven
2.2.3 | Support vector machines approach that utilizes conventional ML models in combi-
nation with known physical laws for a given problem
Support vector machines (SVMs) are a supervised learn- within a scientific domain.[29] SciML can be used to per-
ing method used for regression analysis and for the classi- form high-performance simulations that are several
fication of data. They are discriminative classifiers orders of magnitude faster than those using classical
defined by hyperplanes that divide a potentially large approaches. SciML utilizes the differential equations that
dataset into distinct classes or groupings of data.[26] define the known physics of a process while training the
SVMs, initially, generate a hyperplane onto a graph that ML model. Their major use cases are (a) surrogate
contains all the data. This hyperplane acts as a line of models which are orders of magnitude faster than classi-
separation that separates two or more classes.[27] The cal models. (b) Parameterization of classical models using
SVM uses an optimization algorithm to find the sparse measurement data.[30]
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
PLATHOTTAM ET AL. 7 of 19

2.3 | Machine learning workflow The traditional approach to implementing ML work-

flows is heavily manual, but AutoML (automated machine
Developing a machine learning solution is an iterative learning) technology provides a domain-agnostic way to
process, hence, the software and hardware infrastructure automate repetitive tasks involved in implementing the ML
(also referred to as the workflow) used to develop the workflow such as Extract-Transform-Load pipelines, feature
solution is as important as the specific machine learning selection, and hyper-parameter optimization.[31,32] These
technique used. Figure 3 illustrates a high-level overview technologies create the potential to reduce the time and
of a typical ML workflow. The relationship between dif- engineering cost of developing ML-based solutions.
ferent components of the workflow is based on software
engineering and data science best practices and from the
authors' own experience in implementing these work- 3 | AI/M L APPLIC ATIONS IN
flows for AI/ML projects. The essential components of MANUFAC TURING
the workflow are (a) databases to store large amounts of
data, such as those from industrial sensors; (b) an Current and emerging industrial applications for AI/ML
Extract-Transform-Load pipeline to preprocess and techniques which play a crucial role in Industry 4.0 include
deliver data that can be used by an ML model; (c) the fea- optimization of manufacturing operations, process and prod-
ture engineering pipeline for selecting features that give uct design, scientific machine learning, computational experi-
the best performance; and (d) the model development mentation, and industrial automation. In Figure 4 we
pipeline to define the train the AI/ML models. The fea- illustrate the scope of AI/ML applications in different
ture engineering pipeline requires analysis of the data manufacturing domains classified based on whether they are
before model training and performing model validation part of operations, design, or automation. Of these, process
after a trained model is available to perform down-select and product design are further along on the industrial dem-
features for the AI/ML model. The model development onstration and adoption curve, while real-time AI-driven
pipeline requires modules to create an instance of the automation and scientific machine learning are at an earlier
model (using an AI/ML software framework like Tensor- stage of adoption. Overall, AI enables companies to gather
Flow) given the architecture and hyperparameters as and analyze copious amounts of data, identify patterns and
input. It also requires hyperparameter optimization to insights, and automate processes, enabling faster and more
find the hyperparameters which give the best perfor- informed decisions that improve operations and product
mance. Finally, there is a need for sufficient computing development. This section will examine the objective, poten-
hardware and software infrastructure necessary to train tial benefits, and challenges of AI strategies in manufacturing
AI/ML models in a reasonable amount of time. applications.

F I G U R E 3 General
workflow for developing the ML
model for an AI/ML solution.
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
8 of 19 PLATHOTTAM ET AL.

FIGURE 4 Representative AI/ML applications in the manufacturing industry.

3.1 | Operations and scheduling maintenance routines to prevent unneces-

sary downtime. It is one of the most common uses for AI/ML
The term “operations” concerns the actual use of production in manufacturing industries, such as aerospace, chemicals,
facilities and resources.[33] On the physical side, it involves electronics, and consumer goods manufacturing.[5–7] The
industrial machinery, sensors and controls, human workers, value to manufacturers is huge, as the ability to anticipate
and facility infrastructure. More abstractly, it deals with logis- equipment failure can avoid significant material and finan-
tics and processes, such as how exactly the necessary cial losses by avoiding unscheduled disruptions and down-
resources will be transported and manufactured into the times in manufacturing operations. Typical manufacturing
final product. Operations may be classified as long-term facilities average 15 h of downtime a week, with losses of
plans for facilities and processes or real-time actions on the about $20 000 per minute of production line downtime for
factory floor, depending on the relevant time horizon. They large automotive companies.[5] Predictive maintenance can
are critical to ensuring that industrial equipment is func- also minimize risks of unplanned shutdowns to workers,
tioning as desired and product quality is maintained. His- communities, and the environment.[34,35] AI/ML strategies
torically, human operators have exclusively managed such as computer vision, regression, classification models,
operations. However, the predictive and analytic capabilities and anomaly detection can be directly applied to predictive
of AI/ML models can provide useful insights to human maintenance to support factory floor error detection. Com-
operators to facilitate planning, support real-time decision- puter vision provides more detailed visual data than the
making, and improve manufacturing efficiency and safety. human eye, and computers can observe factory operations
for longer without interruption or fatigue. Using data from
networked sensors, CNNs, ML-based computer vision
3.1.1 | Planning models, and various supervised learning algorithms can be
trained to predict equipment failure probability.[36,37] Regres-
a. Predictive maintenance sion models may be employed to predict the remaining use-
ful life of a piece of equipment or estimate how much time it
Predictive maintenance involves analyzing sensor data takes until the next failure. Classification models may be
from equipment to anticipate potential equipment failures used to predict whether a piece of equipment will fail within
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
PLATHOTTAM ET AL. 9 of 19

a given time span. Predictive maintenance may also use analytics and real time data analysis to manage their
anomaly detection to determine when equipment is behav- inventory levels and production planning.[48–50] AI/ML
ing outside of normal parameters.[38,39] can use predictive analytics to forecast critical supply
chain variables including (a) demand for a finished indus-
b. Quality assurance trial product and (b) lead times for a critical component
used to make that product. With real time data analysis,
Quality assurance is essential to customer health and AI can provide valuable insights into market trends and
safety. Preventing quality failures can improve customer conditions, helping companies to make timely decisions
satisfaction, and reduce costs and waste.[5] AI/ML models on purchasing feedstock or selling products in response to
have shown promise for augmenting quality control in a spot market signals. AI/ML-based predictive models can
broad set of applications across manufacturing industries. make use of far more historical data and features to pro-
Recent studies have shown that CNNs can match, and in vide predictions that are significantly more accurate than
some cases, outperform traditional methods for detecting classical methods. Natural language processing (NLP) can
manufactured product imperfection.[21,40] This is particu- extract valuable information from news feeds to provide
larly important for additive manufacturing, where fea- market insights and digitize physical data—such as
tures such as density and porosity can have significant invoices—primarily meant for humans faster and more
effects on the final product's mechanical properties.[40] In accurately than an unaided human data entry operator.
semiconductor manufacturing, computer vision models Industry robots and drones using computer vision based
using CNN developed with the aid of AutoML were used on AI/ML models can operate within a warehouse with
to detect random defects in electron microscope images minimal supervision and an unprecedented level of accu-
and wafer maps, which are important predictors of semi- racy. They can be used to both keep track of inventory and
conductor performance.[31] These models have also been aid in recovering items from inventory.[51] Other applica-
used on automotive manufacturing assembly lines to tions include tracking and reducing wastage, real-time
detect defects in LCD screens, optical films, and fabrics, monitoring during logistical operations,[48] as well as auto-
with up to a 6% increase in detection accuracy.[31] mation of routine tasks to reduce errors and improve pro-
ductivity. RL has also been used to streamline production
c. Energy consumption forecasting pathways and for scheduling to minimize delays and opti-
mize productivity.[52]
Predicting the energy consumption of manufacturing
processes is a proactive way of reducing environmental
impact and improving sustainability. Using temperature, 3.1.2 | Near real-time operations
humidity, lighting usage, facility activity, and historical
energy consumption data, regression models can predict a. Process optimization
energy consumption profiles at the facility and specific
process levels. These can be exceptionally useful for Historically, optimization has been applied to individual
energy efficiency and demand-response strategy, espe- manufacturing processes as well as larger-scale opera-
cially in energy-intensive industries such as mining and tions such as facility layout and supply chain manage-
steelmaking,[41–44] as well as in additive manufacturing ment. The increasing diversity and complexity of
applications.[45] DNNs are particularly good at forecast- manufacturing tasks, workflows, and supply chains has
ing when historical time-series data from a large number increased the number of variables and interdependencies.
of devices are available for training (e.g., smart energy Manually finding an optimal solution via experimenta-
meters).[46,47] SVMs are also suitable for short-term elec- tion is time- and resource-intensive, and the effectiveness
tricity consumption forecasting, especially when the fore- of current mathematical approaches such as heuristic or
casting problem involves a small number of samples model-based optimization decreases as the number of
(fewer historical data), and high dimensional inputs.[47] variables and interdependencies grows.[18,43,53] AI/ML
techniques have emerged as supplements or comparable
d. Supply chain management alternatives to classical optimization algorithms for
manufacturing processes and procedures.[5,40] Examples
Supply chain management is one of the most critical include RL for hydrometallurgical separation process
tasks in a manufacturing operation and one of the most design optimization,[18] and hybrid support vector and
complex since supply chains can span multiple countries evolutionary algorithms for multi-objective optimization,
and continents. AI/ML can be used in applications that which was applied to a carbon fiber manufacturing pro-
facilitate and optimize supply chains, leveraging predictive cess to realize a 45% reduction in energy consumption.[43]
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
10 of 19 PLATHOTTAM ET AL.

AI/ML may also be used to optimize larger-scale pro- program's sequence of system calls to evaluate whether it
cesses such as factory layout design, inter- and intra- is malicious or not.[67] Unsupervised learning models may
facility dispatch management, and logistics. Recurrent be combined with expert systems for anomaly detection,
Neural Networks have been used to optimize delivery as data generated by operational technology is predictable.
and dispatch services by autonomous guided vehicles GANs can be used to generate data that can model the
while avoiding conflicts with workers or other vehi- relationship and information flow between cyber and
cles.[54] RL has also been used to optimize dispatch physical systems, using the information to determine
within a facility[20] (such as a factory floor or warehouse), whether security requirements are met.[21]
and for job-shop scheduling—where one product requir-
ing several tasks which must be completed on separate
machines while ensuring an optimal layout of equip- 3.2 | Design
ment.[55] Other use case applications for RL include
improving the ability of robots to identify and pick out an Design involves developing new or modified products
object from specific bins,[56] select the best paths to mini- and processes, modeling them either digitally or physi-
mize unnecessary stops, and avoid obstacles and interfer- cally, and testing them to ensure they are feasible and
ence with human operators.[57] It is also becoming more meet the manufacturer's goals. The typical design process
common to see hybrid applications where machine learn- for manufacturing is iterative, requiring tests of each new
ing techniques are integrated with classical optimization design to determine its effectiveness, and requiring pro-
techniques and process simulation.[58] cess designers to expend considerable time, labor, and
Digital twins represent another application where materials while iterating towards an ideal result. Emerg-
AI/ML can have significant impact on manufacturing, ing AI/ML techniques for aiding process/product design
with a 2020 case study reporting significant benefits from include predictive modeling, generative design, and rein-
the use of digital twins in large scale smart manufactur- forcement learning. AI/ML-based predictive models
ing operations.[59] Digital twins are virtual replicas of a provide significant productivity leverage to product
manufacturing unit or facility within a simulated envi- designers,[68] and when coupled with AutoML to simplify
ronment. One major value proposition of AI/ML-based the predictive analytics workflow, speed up implementa-
digital twins is that they provide several orders of magni- tion while avoiding drawbacks in the traditional
tude reduction in simulation time over conventional approach due to costly experiments and time-consuming
approaches, which makes it feasible to use them for real simulations. Currently, NVIDIA (through the Omniverse
time data analysis and process control. They can also be platform) and ANSYS (through the Twinbuilder tool)
used to conduct experiments and to test minor changes provide some of the most powerful design tools that uti-
to system design, allowing operators to evaluate potential lize by AI/ML models.[69,70]
process responses and behaviors before actuating an
updated control logic on the factory floor. AI/ML based
digital twins may also assist with automation to enable 3.2.1 | Process and product design
intelligent and autonomous manufacturing.[60,61]
Generative design involves the use of AI to explore the
b. Security and safety design space for a product or process based on user-
provided requirements. To achieve this, the AI is first
AI/ML can also improve worker and critical equipment trained on a large corpus of existing designs. New designs
safety within factories through intelligent access control are then generated by interpolating or sampling within
systems. It can also be used to mitigate the cybersecurity the space. This approach permits a wide variety of design
risks introduced by the ever-increasing number of net- options to be explored in a shorter amount of time. The
worked devices within a manufacturing plant. Computer human designer can then focus on selecting from the
vision based on deep learning can visually identify unsafe generated design alternatives.[71] Generative models can
behaviors for employees and identify the presence of also be repurposed to make modifications to an existing
unauthorized personnel within a facility.[62–64] A study of product design to enhance customization, improve per-
process manufacturing in the chemicals industries used formance, or adapt to new situations.[72] GANs are the
deep learning to examine the relationships between pro- most popular because of their ability to generate high-
cess factors to predict potential accidents[65] For industrial grade image data consistently and allow for multi-modal
cybersecurity, AI/ML models are used in intrusion detec- inputs[73] and may be applied to generative design. GANs
tion systems[66] by detecting anomalous patterns in user are most effective when training data is available, and a
behavior or network traffic such as, for example, analyze a basic idea for the design is already known (e.g., developing
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
PLATHOTTAM ET AL. 11 of 19

new models of cars in the same segment[5,21]). Reinforce- trends in the data that model underlying behavior/proper-
ment learning has been used for performance-optimized ties, simplifying datasets ahead of more complex, detailed
layouts for computer chips, given density and congestion treatment with supervised learning to inform material
constraints.[19] selection.[82]
SciML has been used in fluid mechanics modeling
and was shown to reduce the computational time
required to solve these higher dimensional Partial Differ- 3.3 | Automation and human–machine
ential Equations, allowing designers to iterate faster.[74] interaction
In one study, a fluid mechanics-based model of a 2D noz-
zle design was simulated using a surrogate SciML model, Industrial robots are already a mainstay of modern
saving significant labor and expenses otherwise needed manufacturing. Incorporating AI into existing industrial
for the traditional approach that requires multiple design robots has the potential to facilitate a step change in
iterations involving computationally intensive simula- the cooperation between human workers and robots. It
tions.[75] SciML can be used to speed up solving optimiza- could allow robots to quickly adapt to variable human
tion problems in cases where optimization algorithms behavior to maintain safety and efficiency. AI/ML can
need to solve a Partial Differential Equations at every assist with shortages of human expertise by enabling
iteration.[76] robots to copy expert human behavior. This can be
achieved by using supervised learning techniques to
train an ML model to imitate the decision-making skills
3.2.2 | Experimentation of experts. A 2018 study on using deep reinforcement
learning for automating water purification plants pro-
AI and ML can be used for high-fidelity simulations of posed combining supervised learning with RL for
manufacturing processes and to reduce the need for man- greater versatility—allowing the agent to both refer to a
ual experimentation which saves cost and time. This is “manual” of information and make experience-based
especially beneficial in situations where the process or decisions.[84] Other studies have also shown that AI/ML
experiment is complex or expensive. ML models can sim- can enable robots to perform tasks such as support
ulate experiments, optimize independent variables, and removal in metal additive manufacturing,[85] or autono-
predict outcomes with accuracy comparable to traditional mous vehicle management[54] that were dangerous or
experimentation. For instance, techniques such as Bayes- tedious for humans while being too complex for con-
ian optimization have been used to develop autonomous ventional robots. Human flexibility is still a necessity in
experimentation systems to perform mechanical testing manufacturing, and as AI for industrial robots becomes
of additive manufacturing structures towards identifying more common, human-machine interaction becomes
best-performing configurations for different applications, unavoidable. It is necessary to examine methods of
and achieved a 60-fold reduction in the number of experi- facilitating these interactions and permitting machines
ments needed.[77] Hybrid AI/ML workflows integrating to adapt to the nuances of human behavior. This may
statistical methods, machine learning surrogate model- come in the form of NLP—for example, developing a
ing, and Bayesian optimization have also been used suc- language database and then applying a recurrent neural
cessfully in design of experiments for applications such network to understand verbal complaints and assist
as nanomaterial production via flame spray pyrolysis, with maintenance or repair.[86] It may also involve RL
where it reduced number of in-situ particle size measure- as a tool for allowing industrial robots to observe the
ments and improved particle size distribution in the state of the factory floor and take actions corresponding
product.[78] Similar success was reported for a bioprocess to the need.[87]
development study that combined design of experiment
with surrogate modeling using DNNs.[79] ML experimen-
tation involving a range of supervised and unsupervised 4 | CHALLENGES
ML techniques has been used successfully in advanced
materials development to examine how various substances Incorporating AI/ML into manufacturing involves sev-
will behave under extreme heat and pressure[80–82] and eral challenges in the areas of data acquisition, energy
to reduce costs and development time for additive consumption, implementation, security and privacy, and
manufacturing processes for high-strength light-weight decision validation. Figure 5 illustrates some of these
alloys in aerospace components.[83] These studies show broad challenges and their underlying causes. The rest of
that different learning approaches can play complemen- the section discusses their implications for implementing
tary roles. For instance, unsupervised learning can filter AI/ML in manufacturing.
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
12 of 19 PLATHOTTAM ET AL.

F I G U R E 5 A high-level illustration
of the challenges involved in
implementing AI/ML solutions within
the manufacturing industry.

4.1 | Data acquisition addition, taking frequent measurements can be costly, in

part due to the associated cost of data storage and trans-
Large amounts of data are required to train supervised fer which in 2016 was estimated at $3351 on average for
and unsupervised learning models. Acquiring any kind of storing a terabyte of data for 1 year.[90]
data from the premises of an industry can be challenging
because of the proprietary nature of manufacturing
equipment and units. In addition, interfacing with 4.2 | Energy consumption
servers in the plant control room would also require secu-
rity clearances which plant managers and operators may Developing AI/ML-based solutions requires training
be reluctant to provide. Any data acquired from the models that require moving large quantities of data
industry will require a significant amount of pre- (i.e., memory transfer) and large computing operations
processing before it can be used to train an AI/ML model. (e.g., high dimensional matrix multiplication) at each
Labeled data will be even more costly due to the expertise step of a training run. Large training runs can take multi-
necessary to label the data, and the time required to do ple days or weeks and require significant amounts of
so. Additional challenges include the rarity of certain energy—most of the energy used with respect to AI/ML
events of interest such as equipment breakdowns, diffi- models is during training—and consequently, emissions
culty of interfacing between disparate sensors,[88,89] and if non-renewable energy sources are used. Since 2012, the
the impact of production conditions on measurements. computing power used for training AI/ML models has
Exposure to varying or extreme manufacturing environ- doubled approximately every 3.4 months.[91] A 2021
ment conditions like heat release, robot motion, and study on AI-powered analytics for manufacturing
pressure can affect sensor output, and over time, can lead reported that the carbon emitted during the training of a
to sensor drift which may bias collected data.[53] In DNN-based NLP model was about 60% of what a single
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
PLATHOTTAM ET AL. 13 of 19

average car produced over its entire lifetime.[51] While been built upon on decades or even centuries of human
energy consumed during inference is orders of magnitude experience on the factory floor. Many of these practices
less than that during training, a study on energy comput- still exist because they are tried-and-true, and not neces-
ing trends reported that the average energy consumption sarily because they are the most efficient. Consequently,
to perform one inference step in DNNs increased from there is no guarantee that any AI/ML based solution—no
0.1 to 20 J during the period from 2012 to 2022.[92] Hence matter how efficient—will be readily accepted on the fac-
energy-intensive AI models can have consequences for tory floor, especially if it requires radical changes to exist-
the environment. Assessments and conclusions from ing industry practice. A 2019 survey of 250 manufacturing
multiple studies suggest that to significantly integrate professionals analyzed the challenges industries have faced
AI/ML-based solutions, manufacturers face a trade-off when implementing AI and identified difficulties in estab-
between complex, detailed algorithms that provide high lishing a clear industry-specific implementation plan.[9]
levels of accuracy on the one hand, and the need to According to the survey responses, pressure from within
reduce training time and consequent energy consump- the industry to use AI has contributed to companies feel-
tion on the other.[40,51,66,82,88,93] ing obligated to pursue AI/ML solutions even when they
lack a concrete plan. Moreover, AI/ML is often deployed
in isolated, specialized situations, and thus does not pick
4.3 | Security and privacy up contextual information that could benefit the process
further.[96]
Developing AI/ML applications requires accessing data Moreover, every AI/ML solution for a given manufactur-
on servers (historians) located within plant control ing problem has its risks and benefits and it differs across
rooms. There is a potential for malicious actors to use companies, across applications, and across specific instances
this opportunity to engage in cyberattacks on industrial of said applications. For instance, a study on AI-powered
control systems which would result in large financial real-time analytics for manufacturing compared three ML-
costs as well as safety concerns due to possibility of seri- based approaches to solving the problem of product collisions
ous equipment malfunctions.[66] In 2020, the global aver- on a conveyor belt.[51] The first, a classification algorithm
age cost of a data breach was reported at $3.86 using classical ML approaches for video data, was easy to
million.[94] In 2021, an IBM report estimated the global implement but lacked adaptability for situations that did not
average cost at $4.35 million, with the United States resemble the sample footage. The second, a CNN trained to
average closer to $8.5 million.[95] While general cyberse- classify objects as “together” or “apart” and alert operators if
curity solutions are also available,[66] and the deployment objects were “together” for too long, was fast but not general-
AI security solutions can potentially reduce the cost of izable. The third approach used two CNNs in succession,
data breaches by up to 70%,[95] the nature of threat tracked multiple objects on the product line, and transferred
advancement means this problem is always changing information between frames about product position and
and requires constant adaptation. Research into safety velocity relative to other products. This was accurate but
systems and human-machine interaction involves using required heavy computational power and a long time to
human data and monitoring employees and there is a train. Thus, in general, the choice of which AI/ML solutions
debate concerning employee privacy.[87] Data involving to implement for specific manufacturing problems is not
employees must be kept secure and anonymous and always trivial and involves a certain degree of trade-offs.
applied in a way that respects their rights. Additionally Manufacturers must be able to decide what drawbacks they
pre-processing should be applied to key performance var- can afford when implementing AI.
iables such that bias is removed.

4.5 | Decision validation

4.4 | Implementation
Decision validation plays a key role in the consideration
Implementing AI solutions, including using mature of AI/ML for manufacturing. The lack of interpretability
AI/ML technologies remains challenging. There are diffi- of the outputs from the AI/ML model makes it difficult
culties with establishing a foundation of infrastructure to use for planning considering that the human, environ-
and personnel, limited consideration of the interrelation- mental, and financial costs of failure in a manufacturing
ships between the complex human and technical systems operation may be significant. Determining the trustwor-
affected, or discrepancies between the most accepted thiness of decisions made by AI/ML is a topic of ongoing
solutions and the solutions that work best in a particular research, particularly since ML models typically take the
setting. Lastly, the manufacturing industry practices have form of black boxes. Historically, human operators
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
14 of 19 PLATHOTTAM ET AL.

gradually learn how much trust to place in new software synthetic data suitable for training AI/ML models or aug-
technologies after observing outputs from these systems menting existing sparse datasets can be obtained using gen-
over time and it would be expected that the same would erative models like GANs,[21,98–100] to compensate for
be true for AI/ML applications.[96] smaller quantities of data from manufacturing operations.
This technique cannot extrapolate or generate results
beyond the extremes of its training data. But it is low-risk
5 | AI/ML TRENDS A ND and addresses a challenge of providing anonymized data to
OPPORTUNITIES IN train AI/ML models.[100] Researchers have used DNNs to
MANUFACTURING deal specifically with sparse functional data.[101] An archi-
tectural framework to unify data acquisition from disparate
The literature that was surveyed in the previous sections networked industry devices, signal processing, and analysis
suggests that currently, AI/ML-based solutions supple- within a robust interface was presented in Serizawa and
ment human labor rather than provide complete automa- Shomura.[89] Second, improving the FLOPS to Watt ratio
tion. This is further corroborated by the survey done in (i.e., energy efficiency) of AI/ML hardware accelerators will
Wee[97] which found that 38% of manufacturers use AI for translate to lowering the capital cost incurred in both devel-
operations related to business continuity, 38% use it to oping and deploying AI/ML solutions in manufacturing.
help employees be more efficient, and 34% found it to be Another approach is reducing the size of trained models. In
helpful for employees overall. In the opinion of the Ding et al.,[61] the efforts to develop a framework for reduc-
authors, we may be currently witnessing a gradual process ing memory use in a variety of DNN architectures are
whereby manufacturing industries steadily develop trust described by removing unused parameters which reduced
and experience with AI/ML solutions starting with high- memory by 96% and computation by 90%. Third, improving
level analytical tasks and ending with automation on the the computing and communication capabilities of edge
factory floor. We illustrate this gradual development in computing hardware (e.g., Jetson NANO which can run
Figure 6. Hence, AI/ML solutions that allow companies to AI/ML models for applications like image classification,
experiment with AI/ML with minimal risk (e.g., obtain object detection, segmentation, and speech processing[102])
high-level analytics useful for plant operators on existing can accelerate the deployment of AI/ML solutions on the
processes generating copious amounts of data) would be factory floor by removing the need to run AI/ML models
most sought after. Applications that support decision-mak- on a server.[103] Fourth, building trust in decisions from
ing, such as design and optimization algorithms would be AI/ML decisions through the concept of Explainable
the next solutions of interest. Finally, AI/ML solutions AI,[104] which involves working to develop a formal deci-
that directly integrate with the automation and robotics sion confidence measure for AI, to improve interpretability
on the factory floor would be implemented after a signifi- by humans. This provides human operators with more
cant level of trust has been generated and in-house exper- detailed information to determine whether to trust a deci-
tise created. In addition, the AI/ML-based analytical and sion made by AI/ML. Another concept is that of “humble
decision-support support applications should have demon- AI”[105] that can understand its limitations and revert to a
strated measurable value. default, safer state of behavior if it is uncertain about its sit-
In terms of fundamental AI, research and development, uation or competence. Another approach to improve
there are four areas where advances would benefit AI/ML explainability of AI predictions is the use of easy-to-
solutions in manufacturing and overcome the challenges understand graphs to interpret decisions from AI/ML
mentioned in the previous section. First, high quality models. For instance, a study on equipment health indicator

F I G U R E 6 The progression of
AI/ML solutions for manufacturing as
trust in AI grows. The arrows at the
bottom represent proximity to the
factory floor for each application, with
analytical tasks being the farthest and
automation being nearest.
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
PLATHOTTAM ET AL. 15 of 19

T A B L E 1 Challenges and opportunities for AI/ML in and algorithms are iterated upon and provide compound-
manufacturing. ing benefits to the manufacturing industry over the next
Research
decade. At the same time, the AI/ML solutions also
Challenges opportunities require a thorough understanding of the possible trade-
offs involved (e.g., restructuring facilities, energy costs,
Data paucity: Obtaining sufficient Generative models,
data is expensive transfer learning and expertise) and the specific needs and capabilities of
the company and stakeholders. The trends indicate that
Data privacy: Industry data is Edge computing,
sensitive generative models
AI/ML will continue to be applied cooperatively, along-
side human skills, while at the same time gradually
Energy consumption: Large AI/ML Energy efficient AI/ML
increasing the amount of automation. The rapid evolu-
models tend to have higher models
performance but training large tion and advancement of AI/ML algorithms and tech-
AI/ML models is energy intensive niques will drive adoption in manufacturing industry
Implementation: New workflows Edge computing, large applications, to the extent that they keep demonstrating
introduced by AI/ML applications language models improvements in safety, product quality, and operational
may not be readily accepted efficiency. With the increasing adoption of AI/ML in
Decision validation: Higher level Explainable AI industry, trust in the efficacy and productivity potential
decisions from AI/ML of this technology will grow across industrial sectors and
applications may not be trusted among practitioners. However, the rate of adoption will
be constrained by the associated risks, especially as the
application moves from analytical support to AI control
of industrial operations. AI/ML implementation deci-
learning with DRL developed a time-based graph correlated sions must be suited to each company's unique situation
with system health and operating conditions to address the and needs and in the future, this field would benefit from
black box concern in predictive maintenance algorithms.[38] longer-term case studies of manufacturers that have
Table 1 summarizes the challenges and associated research adopted AI/ML versus those that have not.
opportunities for improving the adoption of AI/ML tech-
niques in manufacturing. NOTATION
AI artificial intelligence
ANI artificial narrow intelligence
6 | C ON C L U S I ON ANN artificial neural network
AutoML automated machine learning
The rapid evolution of AI/ML technologies offers CNN convolutional neural network
an unprecedented opportunity to transform the DL deep learning
manufacturing industry. This review covered a broad DNN deep neural network
range of manufacturing applications, detailing the DRL deep reinforcement learning
potential of AI/ML to improve the safety, efficiency, FLOPS floating point operations per second
productivity, and sustainability of manufacturing. It GAN generative adversarial network
examined applications, potential benefits, and chal- IoT internet of things
lenges of integrating AI/ML in the manufacturing pipe- ML machine learning
line, including operations, planning, quality assurance, NLP natural language processing
energy consumption forecasting, process optimization, NN neural networks
security and safety, product design, automation, and RL reinforcement learning
human-machine interaction. Consequently, the review SaaS Software as a Service
identified nascent developments, current challenges, SciML scientific machine learning
and future directions in AI/ML relevant to manufactur- SVM support vector machine
ing, highlighted AI/ML technologies available for solv-
ing manufacturing problems and identified areas where AUTHOR CONTRIBUTIONS
further research can yield transformational returns for Siby Jose Plathottam: Conceptualization (equal); meth-
the industry. odology (equal); supervision (equal); validation (equal);
AI/ML can leverage the large amount of data gener- visualization (lead); writing – original draft (lead);
ated from industrial sensors to derive actionable insights writing – review and editing (lead). Arin Rzonca: Data
as well as take optimal actions independently. AI/ML curation (equal); formal analysis (supporting); investigation
models can improve over time as the data, infrastructure, (supporting); visualization (supporting); writing – original
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
16 of 19 PLATHOTTAM ET AL.

draft (supporting). Rishi Lakhnori: Data curation (equal); [5] P. Brosset, A. L. Thieullent, S. Patsko, P. Ravix, Scaling AI in
formal analysis (supporting); investigation (supporting); visu- Manufacturing Operations: A Practitioners' Perspective, Cap-
alization (supporting); writing – original draft (supporting). gemini Research Institute, Paris 2019. https://www.cap
gemini.com/wp-content/uploads/2019/12/AI-in-manufacturing-
Chukwunwike O. Iloeje: Conceptualization (lead); meth-
operations.pdf (accessed: June, 2021).
odology (lead); project administration (lead); supervision [6] S. Fahle, C. Prinz, B. Kuhlenkötter, Proc. CIRP 2020, 93, 413.
(lead); validation (equal); writing – original draft (equal); [7] R. Cioffi, M. Travaglioni, G. Piscitelli, A. Petrillo, F. De
writing – review and editing (equal). Felice, Sustainability 2020, 12(2), 492.
[8] A. Rizzoli, 7 Out-of-the-Box Applications of AI in Manufactur-
A C K N O WL E D G M E N T S ing, V7 Labs Blog, 2022. https://www.v7labs.com/blog/ai-in-
The submitted manuscript has been created by UChicago manufacturing
[9] Plutoshift, Breaking Ground on Implementing AI: Instituting
Argonne LLC, Operator of Argonne National Laboratory
Strategic AI Programs – From Promise to Productivity, Pluto-
(“Argonne”). Argonne, a U.S. Department of Energy
shift, Palo Alto 2019. https://plutoshift.com/wp-content/
Office of Science laboratory, is operated under Contract uploads/2022/02/plutoshift-breaking-ground-on-implementing-
No. DE-AC02-06CH11357. The U.S. Government retains ai.pdf (accessed: May 2023).
for itself, and others acting on its behalf, a paid-up nonex- [10] Voyant Tools, https://voyant-tools.org/ (accessed: March,
clusive, irrevocable worldwide license in said article to 2023).
reproduce, prepare derivative works, distribute copies to [11] V. Kanade, Narrow AI vs. General AI vs. Super AI: Key Com-
the public, and perform publicly and display publicly, by parisons, SpiceWorks, 2022. https://www.spiceworks.com/
tech/artificial-intelligence/articles/narrow-general-super-ai-
or on behalf of the Government. The Department of
difference/
Energy will provide public access to these results of feder- [12] Machine learning, Wikipedia, Machine learning, 2022.
ally sponsored research in accordance with the DOE Pub- https://en.wikipedia.org/w/index.php?title=Machine_learnin
lic Access Plan. Argonne National Laboratory's work was g&oldid=1084622324 (accessed: April, 2022).
supported by the U.S. Department of Energy, Office of [13] Amazon (AWS), Training ML Models – Amazon Machine Learn-
Energy Efficiency and Renewable Energy, Advanced ing, Amazon (AWS), 2022. https://docs.aws.amazon.com/
Materials Office under contract DE-AC02-06CH11357. machine-learning/latest/dg/training-ml-models.html (accessed:
April, 2022).
[14] IBM Cloud Education, What is Supervised Learning? IBM
CONFLICT OF INTEREST STATEMENT
Cloud Education, 2021. https://www.ibm.com/cloud/learn/
The authors declare no competing interests. supervised-learning (accessed: April, 2022).
[15] javapoint, Unsupervised Machine Learning – Javatpoint, java-
DATA AVAILABILITY STATEMENT point, 2022. https://www.javatpoint.com/unsupervised-
Data sharing is not applicable to this article as no new machine-learning (accessed: April, 2022).
data were created or analyzed in this study. [16] IBM Cloud Team, Supervised vs. Unsupervised Learning: What's
the Difference? IBM Cloud Team, 2021. https://www.ibm.com/
ORCID cloud/blog/supervised-vs-unsupervised-learning (accessed:
Siby Jose Plathottam https://orcid.org/0000-0003-4813- April, 2022).
[17] V. François-Lavet, P. Henderson, R. Islam, M. G. Bellemare,
5724
J. Pineau, Found Trends Mach. Learn. Mach. Learn. 2018,
Chukwunwike O. Iloeje https://orcid.org/0000-0002-
11(3–4), 219.
3426-9425 [18] S. J. Plathottam, B. Richey, G. Curry, J. Cresko, C. O. Iloeje,
J. Adv. Manuf. Process. 2021, 3(2), e10079.
R EF E RE N C E S [19] A. Mirhoseini, A. Goldie, M. Yazgan, J. Jiang, E. Songhori, S.
[1] J. Zhou, P. Li, Y. Zhou, B. Wang, J. Zang, L. Meng, Engineer- Wang, Y.-J. Lee, E. Johnson, O. Pathak, S. Bae, A. Nazi, J.
ing 2018, 4(1), 11. Pak, A. Tong, K. Srinivasa, W. Hang, E. Tuncer, A. Babu,
[2] R. Y. Zhong, X. Xu, E. Klotz, S. T. Newman, Engineering Q. V. Le, J. Laudon, R. Ho, R. Carpenter, J. Dean. Chip Place-
2017, 3(5), 616. ment with Deep Reinforcement Learning. ArXiv200410746 Cs,
[3] S. K. Jagatheesaperumal, M. Rahouti, K. Ahmad, A. Al- 2020. http://arxiv.org/abs/2004.10746 (accessed: June, 2022).
Fuqaha, M. Guizani. The Duo of Artificial Intelligence and Big [20] S. Zheng, C. Gupta, S. Serita. Manufacturing Dispatching
Data for Industry 4.0: Review of Applications, Techniques, using Reinforcement and Transfer Learning, 2019. https://doi.
Challenges, and Future Research Directions. ArXiv210402425 org/10.48550/arXiv.1910.02035
Cs, 2021. http://arxiv.org/abs/2104.02425 (accessed: July, [21] A. Kusiak, Int. J. Prod. Res. 2020, 58(5), 1594.
2021). [22] S. Madhavan, M. T. Jones, Deep Learning Architectures – IBM
[4] R. Geissbauer, S. Schrauf, P. Berttram, F. Cheraghi, Digital Developer, IBM Developer Articles, 2017. https://developer.
Factories 2020: Shaping the Future of Manufacturing, Price- ibm.com/articles/cc-machine-learning-deep-learning-architec
waterhouseCoopers, 2017. https://www.pwc.de/de/digitale- tures/
transformation/digital-factories-2020-shaping-the-future-of- [23] Vortarus Technologies LLC, Evaluating a Manufacturing
manufacturing.pdf (accessed: June, 2021). Decision with a Decision Tree, Vortarus Technologies LLC,
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
PLATHOTTAM ET AL. 17 of 19

2017. https://vortarus.com/manufacturing-decision-decision- [42] D. T. Kearns, Machine Learning in the Mining Industry – A

tree/ (accessed: April, 2022). Case Study, Medium, 2017. https://medium.com/sustainable-
[24] Lucidchart, What is a Decision Tree Diagram, Lucidchart, 2022. data/machine-learning-in-the-mining-industry-a-case-study-
https://www.lucidchart.com/pages/decision-tree (accessed: April, 33b771729eb2 (accessed: June, 2021).
2022). [43] G. Golkarnarenji, M. Naebe, K. Badii, A. S. Milani, A. Jamali,
[25] Master's in Data Science, What is a Decision Tree? Master's in A. Bab-Hadiashar, R. N. Jazar, H. Khayyam, IEEE Access
Data Science, 2022. https://www.mastersindatascience.org/ 2019, 7, 67576.
learning/introduction-to-machine-learning-algorithms/decision- [44] M. Willenbacher, C. Kunisch, V. Wohlgemuth, in From Sci-
tree/ (accessed: April, 2022). ence to Society. Progress in IS (Eds: B. Otjacques, P. Hitzelber-
[26] R. Mall, Support Vector Machine, Medium, 2019. https:// ger, S. Naumann, V. Wohlgemuth), Springer International
medium.com/@mallrishabh52/support-vector-machine- Publishing, Cham 2018, p. 225. https://doi.org/10.1007/978-3-
2f4280d8ad18 (accessed: April, 2022). 319-65687-8_20
[27] R. Gandhi, Support Vector Machine – Introduction to Machine [45] J. Qin, Y. Liu, R. Grosvenor, F. Lacan, Z. Jiang, J. Cleaner
Learning Algorithms, Medium, 2018. https://towardsdata Prod. 2020, 245, 118702.
science.com/support-vector-machine-introduction-to-machine [46] N. B. Vanting, Z. Ma, B. N. Jørgensen, Energy Inform. 2021,
-learning-algorithms-934a444fca47 (accessed: April, 2022). 4(2), 49.
[28] D. Xu, Y. Tian, Ann. Data Sci. 2015, 2(2), 165. [47] J. Duan, X. Tian, W. Ma, X. Qiu, P. Wang, L. An, Entropy
[29] Brown University, What is SciML? SciML Research Group, 2019, 21(7), 707.
Providence 2022. https://sites.brown.edu/bergen-lab/research/ [48] P. Helo, Y. Hao, Prod. Plan Control 2022, 33(16), 1573.
what-is-sciml/ (accessed: April, 2022). [49] R. Dash, M. Mcmurtrey, C. Rebman, U. K. Kar, Management
[30] P. Nair, 43rd AIAA/ASME/ASCE/AHS/ASC Structures, Struc- 2019, 14, 43.
tural Dynamics, and Materials Conference, American Institute [50] R. Toorajipour, V. Sohrabpour, A. Nazarpour, P. Oghazi, M.
of Aeronautics and Astronautics, Denver 2002. https://doi. Fischl, J. Bus. Res. 2021, 122, 502.
org/10.2514/6.2002-1586 [51] B. Ross, Practical AI-Powered Real-Time Analytics for
[31] Y. Fei, V. Tirumalashetty, AI and Machine Learning Improve Manufacturing: Lessons Learned From Design to Deployment,
Manufacturing Visual Inspection Process, Google Cloud Blog, AI Trends, 2019. https://www.aitrends.com/ai-software/soft
2022. https://cloud.google.com/blog/products/ai-machine- ware-development/practical-ai-powered-real-time-analytics-for-
learning/ai-and-machine-learning-improve-manufacturing- manufacturing-lessons-learned-from-design-to-deployment/
visual-inspection-process/ (accessed: April, 2022). (accessed: June, 2021).
[32] Dilmegani, Cem, AutoML: In depth Guide to Automated [52] Editorial Team, Reinforcement Learning and its Applications
Machine Learning [2022], Dilmegani, Cem, 2018. https:// in Manufacturing, eeDesignIt.com, 2020. https://www.eede
research.aimultiple.com/auto-ml/ (accessed: April, 2022). signit.com/reinforcement-learning-and-its-applications-in-
[33] Unnamed A removed at request of original, Exploring Busi- manufacturing/ (accessed: April, 2022).
ness, University of Minnesota Libraries Publishing, Minneapo- [53] G. Immerman, Production and Process Optimization in
lis 2016, p. 466. https://open.lib.umn.edu/exploringbusiness/ Manufacturing, MachineMetrics, 2022. https://www.machine
chapter/11-1-operations-management-in-manufacturing/ metrics.com/blog/process-optimization-manufacturing (accessed:
(accessed: March, 2023). April, 2022).
[34] S. Lygren, M. Piantanida, A. Amendola. Unsupervised, Deep [54] D. Li, B. Ouyang, D. Wu, Y. Wang. Artificial Intelligence Empow-
Learning-Based Detection of Failures in Industrial Equip- ered Multi-AGVs in Manufacturing Systems. ArXiv190903373 Cs,
ments: The Future of Predictive Maintenance, 2019. https:// 2019. http://arxiv.org/abs/1909.03373 (accessed: July, 2021).
doi.org/10.2118/197629-MS [55] Intelligent Information Research Department, Hitachi, Ltd.,
[35] L. Leoni, A. BahooToroody, M. M. Abaei, F. De Carlo, N. Tokyo, Japan, R. Kamoshida, J. Ind. Intell. Inf. 2019, 7(1), 12.
Paltrinieri, F. Sgarbossa, Process Saf. Environ. Prot. 2021, 147, 115. https://doi.org/10.18178/jiii.7.1.12-17
[56] C. Nicholson, Deep Reinforcement Learning will Transform
[36] Y. Inoue, H. Nagayoshi. in 2019 IEEE Winter Conference on
Manufacturing as We Know It, TechCrunch, 2022. https://
Applications of Computer Vision (WACV) 2019, 686. https://
social.techcrunch.com/2021/06/17/deep-reinforcement-learni
doi.org/10.1109/WACV.2019.00078
ng-will-transform-manufacturing-as-we-know-it/ (accessed:
[37] R. Kaur, J. Acharya, S. Gaur, Int. J. Comput. Inform. Eng.
April, 2022).
2019, 13(7), 6.
[57] K. Hao, This Robot Taught Itself to Walk Entirely on Its Own,
[38] C. Zhang, C. Gupta, A. Farahat, K. Ristovski, D. Ghosh, in MIT Technology Review, Cambridge 2022. https://www.
Machine Learning and Knowledge Discovery in Databases, technologyreview.com/2020/03/02/905593/ai-robot-learns-to-
Lecture Notes in Computer Science, Vol. 11053 (Eds: U. Bre- walk-autonomously-reinforcement-learning/ (accessed: April,
feld, E. Curry, E. Daly), Springer International Publishing, 2022).
Cham 2019, p. 488. [58] K. Ristovski, C. Gupta, K. Harada, H. K. Tang, Proceedings of
[39] J. Waring, C. Lindvall, R. Umeton, Artif. Intell. Med. 2020, the 23rd ACM SIGKDD International Conference on Knowl-
104, 101822. edge Discovery and Data Mining, Association for Computing
[40] G. D. Goh, S. L. Sing, W. Y. Yeong, Artif. Intell. Rev. 2021, Machinery, New York 2017, p. 1981. https://doi.org/10.1145/
54(1), 63. 3097983.3098178
[41] C. Chen, Y. Liu, M. Kumar, J. Qin, Y. Ren, Comput. Ind. Eng. [59] J. O'Sullivan, D. O'Sullivan, K. Bruton, Proc. Manuf. 2020, 51,
2019, 135, 757. 1523.
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
18 of 19 PLATHOTTAM ET AL.

[60] G. Zhou, C. Zhang, Z. Li, K. Ding, C. Wang, Int. J. Prod. Res. [80] N. S. Johnson, P. S. Vulimiri, A. C. To, X. Zhang, C. A. Brice,
2020, 58(4), 1034. B. B. Kappes, A. P. Stebner. Machine Learning for Materials
[61] K. Ding, F. T. S. Chan, X. Zhang, G. Zhou, F. Zhang, Int. Developments in Metals Additive Manufacturing. ArXiv20
J. Prod. Res. 2019, 57(20), 6315. 0505235 Cond-Mat Physicsphysics, 2020. http://arxiv.org/abs/
[62] Worker's Compensation Costs are the Tip of a Massive 2005.05235 (accessed: July, 2021).
Iceberg – But Solutions Exist, Intenseye Blog, 2019. https://blog. [81] M. C. Messner, J. Mech. Des. 2019, 142(2), 024503. https://doi.
intenseye.com/worker-compansation-cost/ (accessed: July, 2021). org/10.1115/1.4045040
[63] Get a Grip: Prevent Slip-and-falls with AI-powered Solutions, [82] D. Merayo, A. Rodríguez-Prieto, A. M. Camacho, Proc.
HGS, 2021. https://hgs.cx/blog/get-a-grip-prevent-slip-and- Manuf. 2019, 41, 42.
falls-with-ai-powered-solutions/ (accessed: July, 2021). [83] K. Harston, AMRC J. 2021, 13, 41.
[64] S. Sen, M. Ravikiran. in 2019 IEEE Applied Imagery Pattern [84] P. Nguyen, E. Takashi. Automating Water Purification Plant
Recognition Workshop (AIPR) 2019, 1. https://doi.org/10. Operations Using Deep Deterministic Policy Gradient 2018.
1109/AIPR47015.2019.9174567 [85] K. Harston, AMRC J. 2020, 11, 32.
[65] S. Mao, B. Wang, Y. Tang, F. Qian, Engineering 2019, 5(6), 995. [86] W. Shalaby, A. Arantes, T. GonzalezDiaz, C. Gupta, 2020
[66] A. Bécue, I. Praça, J. Gama, Artif. Intell. Rev. 2021, 54(5), 3849. IEEE International Conference on Prognostics and Health
[67] C. Wolf. Breaking Down the Pros and Cons of AI in Cybersecu- Management (ICPHM), IEEE, Detroit 2020, p. 1. https://doi.
rity, 2022. http://www.asisonline.org/security-management- org/10.1109/ICPHM49022.2020.9187036
magazine/monthly-issues/security-technology/archive/2021/ [87] H. Oliff, Y. Liu, M. Kumar, M. Williams, Proc. CIRP 2020, 93,
april/breaking-down-the-pros-and-cons-of-ai-in-cybersecurity/ 1364.
(accessed: April, 2022). [88] K. Bartsch, A. Pettke, A. Hübert, J. Lakämper, F. Lange,
[68] R. Fujimaki, The Expanding Role of Predictive Analytics in J. Phys. Mater. 2021, 4(3), 032005.
Manufacturing, Manufacturing.net, 2020. https://www.manu [89] Y. Serizawa, Y. Shomura, Sens. Trand. J. 2019, 238(11), 8.
facturing.net/technology/blog/21205966/the-expanding-role-of- [90] StorageCraft, File Storage Cost, By The Numbers, StorageCraft
predictive-analytics-in-manufacturing (accessed: April, 2022). Technology, LLC, 2016. https://blog.storagecraft.com/file-
[69] Omniverse Platform for Virtual Collaboration. NVIDIA. storage-cost-statistics/ (accessed: April, 2022).
https://www.nvidia.com/en-us/omniverse/ (accessed: March, [91] R. Perrault, Y. Shoham, E. Brynjolfsson, J. Clark, J.
2023). Etchemendy, B. Grosz, T. Lyons, J. Manyika, J. C. Niebles, The
[70] Ansys Twin Builder j Create and Deploy Digital Twin Models. 2019 AI Index Report, Stanford University, Stanford 2019.
https://www.ansys.com/products/digital-twin/ansys-twin-builder [92] R. Desislavov, F. Martínez-Plumed, J. Hernandez-Orallo.
(accessed: March, 2023). Compute and Energy Consumption Trends in Deep Learning
[71] The Next Wave of Intelligent Design Automation, Harvard Inference. ArXiv210905472 Cs, 2021. http://arxiv.org/abs/
Business School, Boston 2018, p. 12. 2109.05472 (accessed: April, 2022).
[72] R. Matheson, Design Tool Reveals a Product's Many Possible [93] A. Carlson, T. Sakao, Proc. CIRP. 2020, 90, 171.
Performance Tradeoffs, MIT News j Massachusetts Institute [94] Embroker Team, How Much Does a Data Breach Cost?
of Technology, Cambridge 2018. https://news.mit.edu/2018/ Embroker, 2022. https://www.embroker.com/blog/cost-of-a-
interactive-design-tool-product-performance-tradeoffs-0815 data-breach/ (accessed: April, 2022).
(accessed: June, 2021). [95] IBM Security, Cost of a Data Breach Report 2022, IBM, 2022.
[73] M. Walla Singh, Generative Adversarial Networks j GANs for https://www.ibm.com/reports/data-breach (accessed: March,
Image Data, Analytics Vidhya, 2021. https://www.analyt 2023).
icsvidhya.com/blog/2021/03/why-are-generative-adversarial- [96] P. Trakadas, P. Simoens, P. Gkonis, L. Sarakis, A.
networksgans-so-famous-and-how-will-gans-be-in-the-future/ Angelopoulos, A. P. Ramallo-Gonzalez, A. Skarmeta, C.
(accessed: April, 2022). Trochoutsos, D. Calvο, T. Pariente, K. Chintamani, I.
[74] C. Rackauckas, Y. Ma, J. Martensen, C. Warner, K. Zubov, R. Fernandez, A. A. Irigaray, J. X. Parreira, P. Petrali, N.
Supekar, D. Skinner, A. Ramadhan, A. Edelman. Universal Leligou, P. Karkazis, Sensors 2020, 20(19), 5480.
Differential Equations for Scientific Machine Learning. [97] D. Wee, Research on AI Trends in Manufacturing, Google
ArXiv200104385 Cs Math Q-Bio Stat, 2021. http://arxiv.org/ Cloud Blog, 2022. https://cloud.google.com/blog/products/ai-
abs/2001.04385 (accessed: April, 2022). machine-learning/research-on-ai-trends-in-manufacturing/
[75] D. Schmidt, R. Maulik, K. Lyras, Phys. Fluids 2021, 33(12), (accessed: April, 2022).
127104. [98] S. Zheng, A. Farahat, C. Gupta, in Machine Learning and
[76] C. J. Moore, A. J. K. Chua, C. P. L. Berry, J. R. Gair, R. Soc. Knowledge Discovery in Databases. Lecture Notes in Computer
Open Sci. 2016, 3(5), 160125. Science (Eds: U. Brefeld, E. Fromont, A. Hotho, A. Knobbe,
[77] A. E. Gongora, B. Xu, W. Perry, C. Okoye, P. Riley, K. G. M. Maathuis, C. Robardet), Springer International Publish-
Reyes, E. F. Morgan, K. A. Brown, Sci. Adv. 2020, 6, ing, Cham 2020, p. 621. https://doi.org/10.1007/978-3-030-
eaaz1708. 46133-1_37
[78] N. H. Paulson, J. A. Libera, M. Stan, Mater. Des. 2020, 196, [99] Y. O. Lee, J. Jo, J. Hwang. in 2017 IEEE International Confer-
108972. ence on Big Data (Big Data) 2017, 3248. https://doi.org/10.
[79] D. Rodriguez-Granrose, A. Jones, H. Loftus, T. Tandeski, W. 1109/BigData.2017.8258307
Heaton, K. T. Foley, L. Silverman, Bioprocess Biosyst. Eng. [100] C. Bowles, L. Chen, R. Guerrero, P. Bentley, R. Gunn, A.
2021, 44(6), 1301. Hammers, D. A. Dickie, M. V. Hernandez, J. Wardlaw, D.
 2637403x, 2023, 3, Downloaded from https://aiche.onlinelibrary.wiley.com/doi/10.1002/amp2.10159 by Cochrane Croatia, Wiley Online Library on [21/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
PLATHOTTAM ET AL. 19 of 19

Rueckert. GAN Augmentation: Augmenting Training Data [104] J. v. d. Waa, T. Schoonderwoerd, J. v. Diggelen, M. Neerincx,
using Generative Adversarial Networks. ArXiv181010863 Cs, Int. J. Hum-Comput. Stud. 2020, 144, 102493.
2018. http://arxiv.org/abs/1810.10863 (accessed: August, [105] F. Guterl, Judgment Call: Why GE Is Experimenting With
2021). ‘Humble AI’, GE News, 2021. https://www.ge.com/news/
[101] Q. Wang, S. Zheng, A. Farahat, S. Serita, T. Saeki, C. reports/judgment-call-ge-experimenting-humble-ai (accessed:
Gupta. in 2019 International Joint Conference on Neural Net- July, 2021).
works (IJCNN) 2019, 1. https://doi.org/10.1109/IJCNN.2019.
8851700
[102] NVIDIA, Jetson Nano Developer Kit, NVIDIA Developer,
How to cite this article: S. J. Plathottam,
2019. https://developer.nvidia.com/embedded/jetson-nano-
A. Rzonca, R. Lakhnori, C. O. Iloeje, J. Adv. Manuf.
developer-kit (accessed: April, 2022).
[103] Red Hat. Understanding Edge Computing for Manufacturing, Process. 2023, 5(3), e10159. https://doi.org/10.1002/
2022. https://www.redhat.com/en/topics/edge-computing/ amp2.10159
manufacturing (accessed: April, 2022).